Use libjpeg-turbo for improved jpg compatibility and speed

Co-authored-by: Rémi Verschelde <[email protected]>

COPYRIGHT.txt

@@ -324,16 +324,18 @@ Comment: Jolt Physics
 Copyright: 2021, Jorrit Rouwe
 License: Expat
 
-Files: thirdparty/jpeg-compressor/*
-Comment: jpeg-compressor
-Copyright: 2012, Rich Geldreich
-License: public-domain or Apache-2.0
-
 Files: thirdparty/libbacktrace/*
 Comment: libbacktrace
 Copyright: 2012-2021, Free Software Foundation, Inc.
 License: BSD-3-clause
 
+Files: thirdparty/libjpeg-turbo/*
+Comment: libjpeg-turbo
+Copyright: 2009-2024, D. R. Commander
+ 2015, Viktor Szathmáry.
+ 1991-2020, Thomas G. Lane, Guido Vollbeding
+License: BSD-3-clause and IJG
+
 Files: thirdparty/libktx/*
 Comment: KTX
 Copyright: 2013-2020, Mark Callow
@@ -1683,6 +1685,42 @@ License: HarfBuzz
  ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO
  PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
 
+License: IJG
+ The authors make NO WARRANTY or representation, either express or implied,
+ with respect to this software, its quality, accuracy, merchantability, or
+ fitness for a particular purpose.  This software is provided "AS IS", and you,
+ its user, assume the entire risk as to its quality and accuracy.
+ .
+ This software is copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
+ All Rights Reserved except as specified below.
+ .
+ Permission is hereby granted to use, copy, modify, and distribute this
+ software (or portions thereof) for any purpose, without fee, subject to these
+ conditions:
+ (1) If any part of the source code for this software is distributed, then this
+ README file must be included, with this copyright and no-warranty notice
+ unaltered; and any additions, deletions, or changes to the original files
+ must be clearly indicated in accompanying documentation.
+ (2) If only executable code is distributed, then the accompanying
+ documentation must state that "this software is based in part on the work of
+ the Independent JPEG Group".
+ (3) Permission for use of this software is granted only if the user accepts
+ full responsibility for any undesirable consequences; the authors accept
+ NO LIABILITY for damages of any kind.
+ .
+ These conditions apply to any software derived from or based on the IJG code,
+ not just to the unmodified library.  If you use our work, you ought to
+ acknowledge us.
+ .
+ Permission is NOT granted for the use of any IJG author's name or company name
+ in advertising or publicity relating to this software or products derived from
+ it.  This software may be referred to only as "the Independent JPEG Group's
+ software".
+ .
+ We specifically permit and encourage the use of this software as the basis of
+ commercial products, provided that all warranty or liability claims are
+ assumed by the product vendor.
+
 License: MPL-2.0
  Mozilla Public License Version 2.0
  ==================================

SConstruct

@@ -276,6 +276,7 @@ opts.Add(BoolVariable("builtin_glslang", "Use the built-in glslang library", Tru
 opts.Add(BoolVariable("builtin_graphite", "Use the built-in Graphite library", True))
 opts.Add(BoolVariable("builtin_harfbuzz", "Use the built-in HarfBuzz library", True))
 opts.Add(BoolVariable("builtin_icu4c", "Use the built-in ICU library", True))
+opts.Add(BoolVariable("builtin_libjpeg_turbo", "Use the built-in libjpeg-turbo library", True))
 opts.Add(BoolVariable("builtin_libogg", "Use the built-in libogg library", True))
 opts.Add(BoolVariable("builtin_libpng", "Use the built-in libpng library", True))
 opts.Add(BoolVariable("builtin_libtheora", "Use the built-in libtheora library", True))

misc/error_suppressions/tsan.txt

@@ -6,3 +6,4 @@ deadlock:modules/text_server_adv/text_server_adv.cpp
 deadlock:modules/text_server_fb/text_server_fb.cpp
 race:modules/navigation_2d/nav_map_2d.cpp
 race:modules/navigation_3d/nav_map_3d.cpp
+race:thirdparty/thorvg/src/loaders/external_jpg/tvgJpgLoader.cpp

misc/error_suppressions/ubsan.txt

@@ -6,6 +6,7 @@ float-divide-by-zero:thirdparty/thorvg/src/renderer/sw_engine/tvgSwFill.cpp
 function:thirdparty/embree/common/sys/thread.cpp
 function:thirdparty/embree/kernels/common/accel.h
 function:thirdparty/xatlas/xatlas.cpp
+implicit-integer-sign-change:thirdparty/basis_universal/encoder/jpgd.cpp
 implicit-integer-sign-change:thirdparty/basis_universal/transcoder/basisu_astc_helpers.h
 implicit-integer-sign-change:thirdparty/embree/common/lexers/../sys/ref.h
 implicit-integer-sign-change:thirdparty/embree/common/lexers/tokenstream.cpp
@@ -38,7 +39,6 @@ implicit-integer-sign-change:thirdparty/icu4c/common/unicode/unistr.h
 implicit-integer-sign-change:thirdparty/icu4c/common/unistr.cpp
 implicit-integer-sign-change:thirdparty/icu4c/common/uresbund.cpp
 implicit-integer-sign-change:thirdparty/icu4c/common/ustrtrns.cpp
-implicit-integer-sign-change:thirdparty/jpeg-compressor/jpgd.cpp
 implicit-integer-sign-change:thirdparty/libogg/bitwise.c
 implicit-integer-sign-change:thirdparty/libvorbis/info.c
 implicit-integer-sign-change:thirdparty/libvorbis/sharedbook.c

modules/basis_universal/SCsub

@@ -38,6 +38,7 @@ if basisu_encoder:
         "basisu_ssim.cpp",
         "basisu_uastc_enc.cpp",
         "basisu_uastc_hdr_4x4_enc.cpp",
+        "jpgd.cpp",
         "pvpngreader.cpp",
     ]
     encoder_sources = [thirdparty_dir + "encoder/" + file for file in encoder_sources]
@@ -47,7 +48,6 @@ transcoder_sources = [thirdparty_dir + "transcoder/basisu_transcoder.cpp"]
 env_basisu.Prepend(CPPEXTPATH=[thirdparty_dir])
 
 if basisu_encoder:
-    env_basisu.Prepend(CPPEXTPATH=["#thirdparty/jpeg-compressor"])
     env_basisu.Prepend(CPPEXTPATH=["#thirdparty/tinyexr"])
 
 if env["builtin_zstd"]:

modules/basis_universal/config.py

@@ -1,6 +1,6 @@
 def can_build(env, platform):
     if env.editor_build:  # Encoder dependencies
-        env.module_add_dependencies("basis_universal", ["jpg", "tinyexr"])
+        env.module_add_dependencies("basis_universal", ["tinyexr"])
     return True
 
 

modules/jpg/SCsub

@@ -6,23 +6,99 @@ Import("env_modules")
 
 env_jpg = env_modules.Clone()
 
-# Thirdparty source files
-
 thirdparty_obj = []
 
-# Not unbundled for now as they are not commonly available as shared library
-thirdparty_dir = "#thirdparty/jpeg-compressor/"
-thirdparty_sources = [
-    "jpgd.cpp",
-    "jpge.cpp",
+thirdparty_dir = "#thirdparty/libjpeg-turbo"
+
+thirdparty_sources_common = [
+    "jaricom.c",
+    "jcapimin.c",
+    "jcarith.c",
+    "jchuff.c",
+    "jcicc.c",
+    "jcinit.c",
+    "jcmarker.c",
+    "jcmaster.c",
+    "jcomapi.c",
+    "jcparam.c",
+    "jcphuff.c",
+    "jctrans.c",
+    "jdapimin.c",
+    "jdarith.c",
+    "jdatadst.c",
+    "jdatadst-tj.c",
+    "jdatasrc.c",
+    "jdatasrc-tj.c",
+    "jdhuff.c",
+    "jdicc.c",
+    "jdinput.c",
+    "jdmarker.c",
+    "jdmaster.c",
+    "jdphuff.c",
+    "jdtrans.c",
+    "jerror.c",
+    "jfdctflt.c",
+    "jmemmgr.c",
+    "jmemnobs.c",
+    "jpeg_nbits.c",
+    "transupp.c",
+    "turbojpeg.c",
+]
+
+thirdparty_sources_bit_dependent = [
+    "jcapistd.c",
+    "jccoefct.c",
+    "jccolor.c",
+    "jcdctmgr.c",
+    "jcmainct.c",
+    "jcprepct.c",
+    "jcsample.c",
+    "jdcoefct.c",
+    "jdcolor.c",
+    "jdapistd.c",
+    "jddctmgr.c",
+    "jdmainct.c",
+    "jdmerge.c",
+    "jdpostct.c",
+    "jdsample.c",
+    "jfdctfst.c",
+    "jfdctint.c",
+    "jidctflt.c",
+    "jidctfst.c",
+    "jidctint.c",
+    "jidctred.c",
+    "jutils.c",
+    "jquant1.c",
+    "jquant2.c",
 ]
-thirdparty_sources = [thirdparty_dir + file for file in thirdparty_sources]
 
-env_jpg.Prepend(CPPEXTPATH=[thirdparty_dir])
+thirdparty_sources_by_bits = {
+    8: list(thirdparty_sources_bit_dependent),
+    12: list(thirdparty_sources_bit_dependent),
+}
+
+
+def source_paths(files):
+    return [thirdparty_dir + "/src/" + f for f in files]
+
+
+env_jpg.Prepend(CPPEXTPATH=[thirdparty_dir + "/src"])
+
+
+def add_bit_depth(bit_depth: int):
+    env_bit_depth = env_jpg.Clone()
+    env_bit_depth.disable_warnings()
+    env_bit_depth["OBJSUFFIX"] = f"_{bit_depth}{env_bit_depth['OBJSUFFIX']}"
+    env_bit_depth.Append(CPPDEFINES=[f"BITS_IN_JSAMPLE={bit_depth}"])
+    env_bit_depth.add_source_files(thirdparty_obj, source_paths(thirdparty_sources_by_bits[bit_depth]))
+
+
+add_bit_depth(8)
+add_bit_depth(12)
 
 env_thirdparty = env_jpg.Clone()
 env_thirdparty.disable_warnings()
-env_thirdparty.add_source_files(thirdparty_obj, thirdparty_sources)
+env_thirdparty.add_source_files(thirdparty_obj, source_paths(thirdparty_sources_common))
 env.modules_sources += thirdparty_obj
 
 # Godot source files

modules/jpg/image_loader_jpegd.cpp

@@ -1,209 +0,0 @@
-/**************************************************************************/
-/*  image_loader_jpegd.cpp                                                */
-/**************************************************************************/
-/*                         This file is part of:                          */
-/*                             GODOT ENGINE                               */
-/*                        https://godotengine.org                         */
-/**************************************************************************/
-/* Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md). */
-/* Copyright (c) 2007-2014 Juan Linietsky, Ariel Manzur.                  */
-/*                                                                        */
-/* Permission is hereby granted, free of charge, to any person obtaining  */
-/* a copy of this software and associated documentation files (the        */
-/* "Software"), to deal in the Software without restriction, including    */
-/* without limitation the rights to use, copy, modify, merge, publish,    */
-/* distribute, sublicense, and/or sell copies of the Software, and to     */
-/* permit persons to whom the Software is furnished to do so, subject to  */
-/* the following conditions:                                              */
-/*                                                                        */
-/* The above copyright notice and this permission notice shall be         */
-/* included in all copies or substantial portions of the Software.        */
-/*                                                                        */
-/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,        */
-/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF     */
-/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. */
-/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY   */
-/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,   */
-/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE      */
-/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                 */
-/**************************************************************************/
-
-#include "image_loader_jpegd.h"
-
-#include <jpgd.h>
-#include <jpge.h>
-
-#include <string.h>
-
-Error jpeg_load_image_from_buffer(Image *p_image, const uint8_t *p_buffer, int p_buffer_len) {
-	jpgd::jpeg_decoder_mem_stream mem_stream(p_buffer, p_buffer_len);
-
-	jpgd::jpeg_decoder decoder(&mem_stream);
-
-	if (decoder.get_error_code() != jpgd::JPGD_SUCCESS) {
-		return ERR_CANT_OPEN;
-	}
-
-	const int image_width = decoder.get_width();
-	const int image_height = decoder.get_height();
-	const int comps = decoder.get_num_components();
-	if (comps != 1 && comps != 3) {
-		return ERR_FILE_CORRUPT;
-	}
-
-	if (decoder.begin_decoding() != jpgd::JPGD_SUCCESS) {
-		return ERR_FILE_CORRUPT;
-	}
-
-	const int dst_bpl = image_width * comps;
-
-	Vector<uint8_t> data;
-
-	data.resize(dst_bpl * image_height);
-
-	uint8_t *dw = data.ptrw();
-
-	jpgd::uint8 *pImage_data = (jpgd::uint8 *)dw;
-
-	for (int y = 0; y < image_height; y++) {
-		const jpgd::uint8 *pScan_line;
-		jpgd::uint scan_line_len;
-		if (decoder.decode((const void **)&pScan_line, &scan_line_len) != jpgd::JPGD_SUCCESS) {
-			return ERR_FILE_CORRUPT;
-		}
-
-		jpgd::uint8 *pDst = pImage_data + y * dst_bpl;
-
-		if (comps == 1) {
-			memcpy(pDst, pScan_line, dst_bpl);
-		} else {
-			// For images with more than 1 channel pScan_line will always point to a buffer
-			// containing 32-bit RGBA pixels. Alpha is always 255 and we ignore it.
-			for (int x = 0; x < image_width; x++) {
-				pDst[0] = pScan_line[x * 4 + 0];
-				pDst[1] = pScan_line[x * 4 + 1];
-				pDst[2] = pScan_line[x * 4 + 2];
-				pDst += 3;
-			}
-		}
-	}
-
-	//all good
-
-	Image::Format fmt;
-	if (comps == 1) {
-		fmt = Image::FORMAT_L8;
-	} else {
-		fmt = Image::FORMAT_RGB8;
-	}
-
-	p_image->set_data(image_width, image_height, false, fmt, data);
-
-	return OK;
-}
-
-Error ImageLoaderJPG::load_image(Ref<Image> p_image, Ref<FileAccess> f, BitField<ImageFormatLoader::LoaderFlags> p_flags, float p_scale) {
-	Vector<uint8_t> src_image;
-	uint64_t src_image_len = f->get_length();
-	ERR_FAIL_COND_V(src_image_len == 0, ERR_FILE_CORRUPT);
-	src_image.resize(src_image_len);
-
-	uint8_t *w = src_image.ptrw();
-
-	f->get_buffer(&w[0], src_image_len);
-
-	Error err = jpeg_load_image_from_buffer(p_image.ptr(), w, src_image_len);
-
-	return err;
-}
-
-void ImageLoaderJPG::get_recognized_extensions(List<String> *p_extensions) const {
-	p_extensions->push_back("jpg");
-	p_extensions->push_back("jpeg");
-}
-
-static Ref<Image> _jpegd_mem_loader_func(const uint8_t *p_png, int p_size) {
-	Ref<Image> img;
-	img.instantiate();
-	Error err = jpeg_load_image_from_buffer(img.ptr(), p_png, p_size);
-	ERR_FAIL_COND_V(err, Ref<Image>());
-	return img;
-}
-
-class ImageLoaderJPGOSFile : public jpge::output_stream {
-public:
-	Ref<FileAccess> f;
-
-	virtual bool put_buf(const void *Pbuf, int len) {
-		f->store_buffer((const uint8_t *)Pbuf, len);
-		return true;
-	}
-};
-
-class ImageLoaderJPGOSBuffer : public jpge::output_stream {
-public:
-	Vector<uint8_t> *buffer = nullptr;
-	virtual bool put_buf(const void *Pbuf, int len) {
-		uint32_t base = buffer->size();
-		buffer->resize(base + len);
-		memcpy(buffer->ptrw() + base, Pbuf, len);
-		return true;
-	}
-};
-
-static Error _jpgd_save_to_output_stream(jpge::output_stream *p_output_stream, const Ref<Image> &p_img, float p_quality) {
-	ERR_FAIL_COND_V(p_img.is_null() || p_img->is_empty(), ERR_INVALID_PARAMETER);
-	Ref<Image> image = p_img->duplicate();
-	if (image->is_compressed()) {
-		Error error = image->decompress();
-		ERR_FAIL_COND_V_MSG(error != OK, error, "Couldn't decompress image.");
-	}
-	if (image->get_format() != Image::FORMAT_RGB8) {
-		image = image->duplicate();
-		image->convert(Image::FORMAT_RGB8);
-	}
-
-	jpge::params p;
-	p.m_quality = CLAMP(p_quality * 100, 1, 100);
-
-	jpge::jpeg_encoder enc;
-	enc.init(p_output_stream, image->get_width(), image->get_height(), 3, p);
-
-	const uint8_t *src_data = image->get_data().ptr();
-	for (int i = 0; i < image->get_height(); i++) {
-		if (!enc.process_scanline(&src_data[i * image->get_width() * 3])) {
-			return FAILED;
-		}
-	}
-
-	if (enc.process_scanline(nullptr)) {
-		return OK;
-	} else {
-		return FAILED;
-	}
-}
-
-static Vector<uint8_t> _jpgd_buffer_save_func(const Ref<Image> &p_img, float p_quality) {
-	Vector<uint8_t> output;
-	ImageLoaderJPGOSBuffer ob;
-	ob.buffer = &output;
-	if (_jpgd_save_to_output_stream(&ob, p_img, p_quality) != OK) {
-		return Vector<uint8_t>();
-	}
-	return output;
-}
-
-static Error _jpgd_save_func(const String &p_path, const Ref<Image> &p_img, float p_quality) {
-	Error err;
-	Ref<FileAccess> file = FileAccess::open(p_path, FileAccess::WRITE, &err);
-	ERR_FAIL_COND_V_MSG(err, err, vformat("Can't save JPG at path: '%s'.", p_path));
-	ImageLoaderJPGOSFile ob;
-	ob.f = file;
-	return _jpgd_save_to_output_stream(&ob, p_img, p_quality);
-}
-
-ImageLoaderJPG::ImageLoaderJPG() {
-	Image::_jpg_mem_loader_func = _jpegd_mem_loader_func;
-	Image::save_jpg_func = _jpgd_save_func;
-	Image::save_jpg_buffer_func = _jpgd_buffer_save_func;
-}

modules/jpg/image_loader_libjpeg_turbo.cpp

@@ -0,0 +1,191 @@
+/**************************************************************************/
+/*  image_loader_libjpeg_turbo.cpp                                        */
+/**************************************************************************/
+/*                         This file is part of:                          */
+/*                             GODOT ENGINE                               */
+/*                        https://godotengine.org                         */
+/**************************************************************************/
+/* Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md). */
+/* Copyright (c) 2007-2014 Juan Linietsky, Ariel Manzur.                  */
+/*                                                                        */
+/* Permission is hereby granted, free of charge, to any person obtaining  */
+/* a copy of this software and associated documentation files (the        */
+/* "Software"), to deal in the Software without restriction, including    */
+/* without limitation the rights to use, copy, modify, merge, publish,    */
+/* distribute, sublicense, and/or sell copies of the Software, and to     */
+/* permit persons to whom the Software is furnished to do so, subject to  */
+/* the following conditions:                                              */
+/*                                                                        */
+/* The above copyright notice and this permission notice shall be         */
+/* included in all copies or substantial portions of the Software.        */
+/*                                                                        */
+/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,        */
+/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF     */
+/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. */
+/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY   */
+/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,   */
+/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE      */
+/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                 */
+/**************************************************************************/
+
+#include "image_loader_libjpeg_turbo.h"
+
+#include <turbojpeg.h>
+
+#include <string.h>
+
+Error jpeg_turbo_load_image_from_buffer(Image *p_image, const uint8_t *p_buffer, int p_buffer_len) {
+	tjhandle tj_instance = tj3Init(TJINIT_DECOMPRESS);
+	if (tj_instance == NULL) {
+		return FAILED;
+	}
+
+	if (tj3DecompressHeader(tj_instance, p_buffer, p_buffer_len) < 0) {
+		tj3Destroy(tj_instance);
+		return ERR_FILE_CORRUPT;
+	}
+
+	const unsigned int width = tj3Get(tj_instance, TJPARAM_JPEGWIDTH);
+	const unsigned int height = tj3Get(tj_instance, TJPARAM_JPEGHEIGHT);
+	const TJCS colorspace = (TJCS)tj3Get(tj_instance, TJPARAM_COLORSPACE);
+
+	if (tj3Get(tj_instance, TJPARAM_PRECISION) > 8) {
+		// Proceed anyway and convert to rgb8?
+		tj3Destroy(tj_instance);
+		return ERR_UNAVAILABLE;
+	}
+
+	TJPF tj_pixel_format;
+	Image::Format gd_pixel_format;
+	if (colorspace == TJCS_GRAY) {
+		tj_pixel_format = TJPF_GRAY;
+		gd_pixel_format = Image::FORMAT_L8;
+	} else {
+		// Force everything else (RGB, CMYK etc) into RGB8.
+		tj_pixel_format = TJPF_RGB;
+		gd_pixel_format = Image::FORMAT_RGB8;
+	}
+
+	Vector<uint8_t> data;
+	data.resize(width * height * tjPixelSize[tj_pixel_format]);
+
+	if (tj3Decompress8(tj_instance, p_buffer, p_buffer_len, data.ptrw(), 0, tj_pixel_format) < 0) {
+		tj3Destroy(tj_instance);
+		return ERR_FILE_CORRUPT;
+	}
+
+	tj3Destroy(tj_instance);
+	p_image->set_data(width, height, false, gd_pixel_format, data);
+	return OK;
+}
+
+Error ImageLoaderLibJPEGTurbo::load_image(Ref<Image> p_image, Ref<FileAccess> f, BitField<ImageFormatLoader::LoaderFlags> p_flags, float p_scale) {
+	Vector<uint8_t> src_image;
+	uint64_t src_image_len = f->get_length();
+	ERR_FAIL_COND_V(src_image_len == 0, ERR_FILE_CORRUPT);
+	src_image.resize(src_image_len);
+
+	uint8_t *w = src_image.ptrw();
+
+	f->get_buffer(&w[0], src_image_len);
+
+	Error err = jpeg_turbo_load_image_from_buffer(p_image.ptr(), w, src_image_len);
+
+	return err;
+}
+
+void ImageLoaderLibJPEGTurbo::get_recognized_extensions(List<String> *p_extensions) const {
+	p_extensions->push_back("jpg");
+	p_extensions->push_back("jpeg");
+}
+
+static Ref<Image> _jpeg_turbo_mem_loader_func(const uint8_t *p_png, int p_size) {
+	Ref<Image> img;
+	img.instantiate();
+	Error err = jpeg_turbo_load_image_from_buffer(img.ptr(), p_png, p_size);
+	ERR_FAIL_COND_V(err, Ref<Image>());
+	return img;
+}
+
+static Vector<uint8_t> _jpeg_turbo_buffer_save_func(const Ref<Image> &p_img, float p_quality) {
+	Vector<uint8_t> output;
+
+	ERR_FAIL_COND_V(p_img.is_null() || p_img->is_empty(), output);
+
+	Ref<Image> image = p_img->duplicate();
+	if (image->is_compressed()) {
+		Error error = image->decompress();
+		ERR_FAIL_COND_V_MSG(error != OK, output, "Couldn't decompress image.");
+	}
+
+	if (image->get_format() != Image::FORMAT_RGB8) {
+		// Allow grayscale L8?
+		image = image->duplicate();
+		image->convert(Image::FORMAT_RGB8);
+	}
+
+	tjhandle tj_instance = tj3Init(TJINIT_COMPRESS);
+	ERR_FAIL_COND_V_MSG(tj_instance == NULL, output, "Couldn't create tjhandle");
+
+	if (tj3Set(tj_instance, TJPARAM_QUALITY, (int)(p_quality * 100)) < 0) {
+		tj3Destroy(tj_instance);
+		ERR_FAIL_V_MSG(output, "Couldn't set jpg quality");
+	}
+
+	if (tj3Set(tj_instance, TJPARAM_PRECISION, 8) < 0) {
+		tj3Destroy(tj_instance);
+		ERR_FAIL_V_MSG(output, "Couldn't set jpg precision");
+	}
+
+	if (tj3Set(tj_instance, TJPARAM_SUBSAMP, TJSAMP_420) < 0) {
+		tj3Destroy(tj_instance);
+		ERR_FAIL_V_MSG(output, "Couldn't set jpg subsamples");
+	}
+
+	// If the godot image format is `Image::FORMAT_L8` we could set the appropriate
+	// color space here rather than defaulting to RGB.
+
+	unsigned char *jpeg_buff = NULL;
+	size_t jpeg_size = 0;
+	int code = tj3Compress8(
+			tj_instance,
+			image->get_data().ptr(),
+			image->get_width(),
+			0,
+			image->get_height(),
+			TJPF_RGB,
+			&jpeg_buff,
+			&jpeg_size);
+
+	if (code < 0) {
+		tj3Destroy(tj_instance);
+		tj3Free(jpeg_buff);
+		ERR_FAIL_V_MSG(output, "Couldn't compress jpg");
+	}
+
+	output.resize(jpeg_size);
+	memcpy(output.ptrw(), jpeg_buff, jpeg_size);
+
+	tj3Destroy(tj_instance);
+	tj3Free(jpeg_buff);
+
+	return output;
+}
+
+static Error _jpeg_turbo_save_func(const String &p_path, const Ref<Image> &p_img, float p_quality) {
+	Error err;
+	Ref<FileAccess> file = FileAccess::open(p_path, FileAccess::WRITE, &err);
+	ERR_FAIL_COND_V_MSG(err, err, vformat("Can't save JPG at path: '%s'.", p_path));
+
+	Vector<uint8_t> data = _jpeg_turbo_buffer_save_func(p_img, p_quality);
+	ERR_FAIL_COND_V(data.size() == 0, FAILED);
+	ERR_FAIL_COND_V_MSG(!file->store_buffer(data.ptr(), data.size()), FAILED, "Failed writing jpg to file");
+
+	return OK;
+}
+
+ImageLoaderLibJPEGTurbo::ImageLoaderLibJPEGTurbo() {
+	Image::_jpg_mem_loader_func = _jpeg_turbo_mem_loader_func;
+	Image::save_jpg_func = _jpeg_turbo_save_func;
+	Image::save_jpg_buffer_func = _jpeg_turbo_buffer_save_func;
+}

modules/jpg/image_loader_jpegd.h → modules/jpg/image_loader_libjpeg_turbo.h

@@ -1,5 +1,5 @@
 /**************************************************************************/
-/*  image_loader_jpegd.h                                                  */
+/*  image_loader_libjpeg_turbo.h                                          */
 /**************************************************************************/
 /*                         This file is part of:                          */
 /*                             GODOT ENGINE                               */
@@ -32,9 +32,9 @@
 
 #include "core/io/image_loader.h"
 
-class ImageLoaderJPG : public ImageFormatLoader {
+class ImageLoaderLibJPEGTurbo : public ImageFormatLoader {
 public:
 	virtual Error load_image(Ref<Image> p_image, Ref<FileAccess> f, BitField<ImageFormatLoader::LoaderFlags> p_flags, float p_scale);
 	virtual void get_recognized_extensions(List<String> *p_extensions) const;
-	ImageLoaderJPG();
+	ImageLoaderLibJPEGTurbo();
 };

modules/jpg/register_types.cpp

@@ -30,17 +30,17 @@
 
 #include "register_types.h"
 
-#include "image_loader_jpegd.h"
+#include "image_loader_libjpeg_turbo.h"
 
-static Ref<ImageLoaderJPG> image_loader_jpg;
+static Ref<ImageLoaderLibJPEGTurbo> image_loader_libjpeg_turbo;
 
 void initialize_jpg_module(ModuleInitializationLevel p_level) {
 	if (p_level != MODULE_INITIALIZATION_LEVEL_SCENE) {
 		return;
 	}
 
-	image_loader_jpg.instantiate();
-	ImageLoader::add_image_format_loader(image_loader_jpg);
+	image_loader_libjpeg_turbo.instantiate();
+	ImageLoader::add_image_format_loader(image_loader_libjpeg_turbo);
 }
 
 void uninitialize_jpg_module(ModuleInitializationLevel p_level) {
@@ -48,6 +48,6 @@ void uninitialize_jpg_module(ModuleInitializationLevel p_level) {
 		return;
 	}
 
-	ImageLoader::remove_image_format_loader(image_loader_jpg);
-	image_loader_jpg.unref();
+	ImageLoader::remove_image_format_loader(image_loader_libjpeg_turbo);
+	image_loader_libjpeg_turbo.unref();
 }

modules/svg/SCsub

@@ -26,8 +26,6 @@ thirdparty_sources = [
     "src/loaders/raw/tvgRawLoader.cpp",
     # image loaders
     "src/loaders/external_png/tvgPngLoader.cpp",
-    "src/loaders/jpg/tvgJpgd.cpp",
-    "src/loaders/jpg/tvgJpgLoader.cpp",
     # renderer common
     "src/renderer/tvgAccessor.cpp",
     # "src/renderer/tvgAnimation.cpp",
@@ -62,6 +60,9 @@ thirdparty_sources = [
 if env["module_webp_enabled"]:
     thirdparty_sources += ["src/loaders/external_webp/tvgWebpLoader.cpp"]
     env_svg.Append(CPPDEFINES=["THORVG_WEBP_LOADER_SUPPORT"])
+if env["module_jpg_enabled"]:
+    thirdparty_sources += ["src/loaders/external_jpg/tvgJpgLoader.cpp"]
+    env_svg.Append(CPPDEFINES=["THORVG_JPG_LOADER_SUPPORT"])
 
 thirdparty_sources = [thirdparty_dir + file for file in thirdparty_sources]
 
@@ -82,7 +83,6 @@ env_thirdparty.Prepend(
         thirdparty_dir + "src/renderer/sw_engine",
         thirdparty_dir + "src/loaders/raw",
         thirdparty_dir + "src/loaders/external_png",
-        thirdparty_dir + "src/loaders/jpg",
     ]
 )
 if env["builtin_libpng"]:
@@ -91,6 +91,10 @@ if env["module_webp_enabled"]:
     env_thirdparty.Prepend(CPPEXTPATH=[thirdparty_dir + "src/loaders/external_webp"])
     if env["builtin_libwebp"]:
         env_thirdparty.Prepend(CPPEXTPATH=["#thirdparty/libwebp/src"])
+if env["module_jpg_enabled"]:
+    env_thirdparty.Prepend(CPPEXTPATH=[thirdparty_dir + "src/loaders/external_jpg"])
+    if env["builtin_libjpeg_turbo"]:
+        env_thirdparty.Prepend(CPPEXTPATH=["#thirdparty/libjpeg-turbo/src"])
 
 env_thirdparty.add_source_files(thirdparty_obj, thirdparty_sources)
 env.modules_sources += thirdparty_obj

modules/svg/config.py

@@ -1,4 +1,5 @@
 def can_build(env, platform):
+    env.module_add_dependencies("svg", ["jpg", "webp"], True)
     return True
 
 

tests/core/io/test_image.h

@@ -133,6 +133,34 @@ TEST_CASE("[Image] Saving and loading") {
 	CHECK_MESSAGE(
 			image_jpg->load_jpg_from_buffer(data_jpg) == OK,
 			"The JPG image should load successfully.");
+
+	Ref<Image> image_grayscale_jpg = memnew(Image());
+	Ref<FileAccess> f_grayscale_jpg = FileAccess::open(TestUtils::get_data_path("images/grayscale.jpg"), FileAccess::READ, &err);
+	REQUIRE(f_grayscale_jpg.is_valid());
+	PackedByteArray data_grayscale_jpg;
+	data_grayscale_jpg.resize(f_grayscale_jpg->get_length() + 1);
+	f_grayscale_jpg->get_buffer(data_grayscale_jpg.ptrw(), f_grayscale_jpg->get_length());
+	CHECK_MESSAGE(
+			image_jpg->load_jpg_from_buffer(data_grayscale_jpg) == OK,
+			"The grayscale JPG image should load successfully.");
+
+	// Save JPG
+	const String save_path_jpg = TestUtils::get_temp_path("image.jpg");
+	CHECK_MESSAGE(image->save_jpg(save_path_jpg) == OK,
+			"The image should be saved successfully as a .jpg file.");
+
+#ifdef MODULE_SVG_ENABLED
+	// Load SVG with embedded jpg image
+	Ref<Image> image_svg = memnew(Image());
+	Ref<FileAccess> f_svg = FileAccess::open(TestUtils::get_data_path("images/embedded_jpg.svg"), FileAccess::READ, &err);
+	REQUIRE(f_svg.is_valid());
+	PackedByteArray data_svg;
+	data_svg.resize(f_svg->get_length() + 1);
+	f_svg->get_buffer(data_svg.ptrw(), f_svg->get_length());
+	CHECK_MESSAGE(
+			image_svg->load_svg_from_buffer(data_svg) == OK,
+			"The SVG image should load successfully.");
+#endif // MODULE_SVG_ENABLED
 #endif // MODULE_JPG_ENABLED
 
 #ifdef MODULE_WEBP_ENABLED

thirdparty/README.md

@@ -80,17 +80,15 @@ Files extracted from upstream source:
 Files extracted from upstream source:
 
 - `encoder/` and `transcoder/` folders, with the following files removed from `encoder`:
-  `jpgd.{cpp,h}`, `3rdparty/{qoi.h,tinydds.h,tinyexr.cpp,tinyexr.h}`
+  `3rdparty/{qoi.h,tinydds.h,tinyexr.cpp,tinyexr.h}`
 - `LICENSE`
 
 Patches:
 
 - `0001-external-zstd-pr344.patch` (GH-73441)
-- `0002-external-jpgd.patch` (GH-88508)
-- `0003-external-tinyexr.patch` (GH-97582)
-- `0004-remove-tinydds-qoi.patch` (GH-97582)
-- `0005-ambiguous-calls.patch` (GH-103968)
-
+- `0002-external-tinyexr.patch` (GH-97582)
+- `0003-remove-tinydds-qoi.patch` (GH-97582)
+- `0004-ambiguous-calls.patch` (GH-103968)
 
 ## brotli
 
@@ -469,22 +467,6 @@ Files extracted from upstream source:
 - `LICENSE`
 
 
-## jpeg-compressor
-
-- Upstream: https://github.com/richgel999/jpeg-compressor
-- Version: 2.00 (aeb7d3b463aa8228b87a28013c15ee50a7e6fcf3, 2020)
-- License: Public domain or MIT
-
-Files extracted from upstream source:
-
-- `jpgd*.{c,h}`
-- `jpge*.{c,h}`
-
-Patches:
-
-- `0001-clang-fortify-fix.patch` (GH-101927)
-
-
 ## libbacktrace
 
 - Upstream: https://github.com/ianlancetaylor/libbacktrace
@@ -503,6 +485,25 @@ Patches:
 - `0001-big-files-support.patch` (GH-100281)
 
 
+## libjpeg-turbo
+
+- Upstream: https://github.com/libjpeg-turbo/libjpeg-turbo
+- Version: git (20ade4dea9589515a69793e447a6c6220b464535, 2024)
+- License: BSD-3-Clause and IJG
+
+Files extracted from upstream source:
+
+- `src/*.{c,h}` except for:
+  * `cdjpeg.c cjpeg.c djpeg.c example.c jcdiffct.c jclhuff.c jclossls.c jcstest.c jddiffct.c jdlhuff.c jdlossls.c jlossls.h jpegtran.c rdbmp.c rdcolmap.c rdgif.c rdjpgcom.c rdppm.c rdswitch.c rdtarga.c strtest.c tjbench.c tjcomp.c tjdecomp.c tjtran.c tjunittest.c tjutil.c wrbmp.c wrgif.c wrjpgcom.c wrppm.c wrtarga.c`
+- `LICENSE.md`
+- `README.ijg`
+
+Patches:
+
+- `0001-cmake-generated-headers.patch` (GH-104347)
+- `0002-disable-16bitlossless.patch` (GH-104347)
+- `0003-remove-bmp-ppm-support.patch` (GH-104347)
+
 ## libktx
 
 - Upstream: https://github.com/KhronosGroup/KTX-Software

thirdparty/basis_universal/encoder/basisu_enc.cpp

@@ -492,7 +492,7 @@ namespace basisu
 	bool load_jpg(const char *pFilename, image& img)
 	{
 		int width = 0, height = 0, actual_comps = 0;
-		uint8_t *pImage_data = jpgd::decompress_jpeg_image_from_file(pFilename, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagBoxChromaFiltering);
+		uint8_t *pImage_data = jpgd::decompress_jpeg_image_from_file(pFilename, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
 		if (!pImage_data)
 			return false;
 		
@@ -512,7 +512,7 @@ namespace basisu
 		}
 
 		int width = 0, height = 0, actual_comps = 0;
-		uint8_t* pImage_data = jpgd::decompress_jpeg_image_from_memory(pBuf, (int)buf_size, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagBoxChromaFiltering);
+		uint8_t* pImage_data = jpgd::decompress_jpeg_image_from_memory(pBuf, (int)buf_size, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
 		if (!pImage_data)
 			return false;
 

thirdparty/jpeg-compressor/jpgd.cpp → thirdparty/basis_universal/encoder/jpgd.cpp

@@ -23,9 +23,7 @@
 // v1.04, May. 19, 2012: Code tweaks to fix VS2008 static code analysis warnings
 // v2.00, March 20, 2020: Fuzzed with zzuf and afl. Fixed several issues, converted most assert()'s to run-time checks. Added chroma upsampling. Removed freq. domain upsampling. gcc/clang warnings.
 //
-// Important:
-// #define JPGD_USE_SSE2 to 0 to completely disable SSE2 usage.
-//
+
 #include "jpgd.h"
 #include <string.h>
 #include <algorithm>
@@ -35,20 +33,6 @@
 #pragma warning (disable : 4611) // warning C4611: interaction between '_setjmp' and C++ object destruction is non-portable
 #endif
 
-#ifndef JPGD_USE_SSE2
-
-	#if defined(__GNUC__)
-		#if defined(__SSE2__)
-			#define JPGD_USE_SSE2 (1)
-		#endif
-	#elif defined(_MSC_VER)
-		#if defined(_M_X64)
-			#define JPGD_USE_SSE2 (1)
-		#endif
-	#endif
-
-#endif
-
 #define JPGD_TRUE (1)
 #define JPGD_FALSE (0)
 
@@ -74,10 +58,6 @@ namespace jpgd {
 
 	enum JPEG_SUBSAMPLING { JPGD_GRAYSCALE = 0, JPGD_YH1V1, JPGD_YH2V1, JPGD_YH1V2, JPGD_YH2V2 };
 
-#if JPGD_USE_SSE2
-#include "jpgd_idct.h"
-#endif
-
 #define CONST_BITS  13
 #define PASS1_BITS  2
 #define SCALEDONE ((int32)1)
@@ -111,7 +91,7 @@ namespace jpgd {
 	template <int NONZERO_COLS>
 	struct Row
 	{
-		static void idct(int* pTemp, const jpgd_block_coeff_t* pSrc)
+		static void idct(int* pTemp, const jpgd_block_t* pSrc)
 		{
 			// ACCESS_COL() will be optimized at compile time to either an array access, or 0. Good compilers will then optimize out muls against 0.
 #define ACCESS_COL(x) (((x) < NONZERO_COLS) ? (int)pSrc[x] : 0)
@@ -156,7 +136,7 @@ namespace jpgd {
 	template <>
 	struct Row<0>
 	{
-		static void idct(int* pTemp, const jpgd_block_coeff_t* pSrc)
+		static void idct(int* pTemp, const jpgd_block_t* pSrc)
 		{
 			(void)pTemp; 
 			(void)pSrc;
@@ -166,7 +146,7 @@ namespace jpgd {
 	template <>
 	struct Row<1>
 	{
-		static void idct(int* pTemp, const jpgd_block_coeff_t* pSrc)
+		static void idct(int* pTemp, const jpgd_block_t* pSrc)
 		{
 			const int dcval = left_shifti(pSrc[0], PASS1_BITS);
 
@@ -280,13 +260,11 @@ namespace jpgd {
 	};
 
 	// Scalar "fast pathing" IDCT.
-	static void idct(const jpgd_block_coeff_t* pSrc_ptr, uint8* pDst_ptr, int block_max_zag, bool use_simd)
+	static void idct(const jpgd_block_t* pSrc_ptr, uint8* pDst_ptr, int block_max_zag)
 	{
-		(void)use_simd;
-
 		assert(block_max_zag >= 1);
 		assert(block_max_zag <= 64);
-				
+
 		if (block_max_zag <= 1)
 		{
 			int k = ((pSrc_ptr[0] + 4) >> 3) + 128;
@@ -303,19 +281,9 @@ namespace jpgd {
 			return;
 		}
 
-#if JPGD_USE_SSE2
-		if (use_simd)
-		{
-			assert((((uintptr_t)pSrc_ptr) & 15) == 0);
-			assert((((uintptr_t)pDst_ptr) & 15) == 0);
-			idctSSEShortU8(pSrc_ptr, pDst_ptr);
-			return;
-		}
-#endif
-
 		int temp[64];
 
-		const jpgd_block_coeff_t* pSrc = pSrc_ptr;
+		const jpgd_block_t* pSrc = pSrc_ptr;
 		int* pTemp = temp;
 
 		const uint8* pRow_tab = &s_idct_row_table[(block_max_zag - 1) * 8];
@@ -642,7 +610,7 @@ namespace jpgd {
 		free_all_blocks();
 		longjmp(m_jmp_state, status);
 	}
-		
+
 	void* jpeg_decoder::alloc(size_t nSize, bool zero)
 	{
 		nSize = (JPGD_MAX(nSize, 1) + 3) & ~3;
@@ -658,7 +626,7 @@ namespace jpgd {
 		}
 		if (!rv)
 		{
-			int capacity = JPGD_MAX(32768 - 256, (nSize + 2047) & ~2047);
+			int capacity = JPGD_MAX(32768 - 256, ((int)nSize + 2047) & ~2047);
 			mem_block* b = (mem_block*)jpgd_malloc(sizeof(mem_block) + capacity);
 			if (!b)
 			{
@@ -675,14 +643,6 @@ namespace jpgd {
 		return rv;
 	}
 
-	void* jpeg_decoder::alloc_aligned(size_t nSize, uint32_t align, bool zero)
-	{
-		assert((align >= 1U) && ((align & (align - 1U)) == 0U));
-		void *p = alloc(nSize + align - 1U, zero);
-		p = (void *)( ((uintptr_t)p + (align - 1U)) & ~((uintptr_t)(align - 1U)) );
-		return p;
-	}
-
 	void jpeg_decoder::word_clear(void* p, uint16 c, uint n)
 	{
 		uint8* pD = (uint8*)p;
@@ -987,15 +947,15 @@ namespace jpgd {
 	// Finds the next marker.
 	int jpeg_decoder::next_marker()
 	{
-		uint c, bytes;
+		uint c;// , bytes;
 
-		bytes = 0;
+		//bytes = 0;
 
 		do
 		{
 			do
 			{
-				bytes++;
+				//bytes++;
 				c = get_bits(8);
 			} while (c != 0xFF);
 
@@ -1191,7 +1151,7 @@ namespace jpgd {
 		m_image_x_size = m_image_y_size = 0;
 		m_pStream = pStream;
 		m_progressive_flag = JPGD_FALSE;
-				
+
 		memset(m_huff_ac, 0, sizeof(m_huff_ac));
 		memset(m_huff_num, 0, sizeof(m_huff_num));
 		memset(m_huff_val, 0, sizeof(m_huff_val));
@@ -1280,19 +1240,6 @@ namespace jpgd {
 
 		for (int i = 0; i < JPGD_MAX_BLOCKS_PER_MCU; i++)
 			m_mcu_block_max_zag[i] = 64;
-
-		m_has_sse2 = false;
-
-#if JPGD_USE_SSE2
-#ifdef _MSC_VER
-		int cpu_info[4];
-		__cpuid(cpu_info, 1);
-		const int cpu_info3 = cpu_info[3];
-		m_has_sse2 = ((cpu_info3 >> 26U) & 1U) != 0U;
-#else
-		m_has_sse2 = true;
-#endif
-#endif
 	}
 
 #define SCALEBITS 16
@@ -1335,7 +1282,7 @@ namespace jpgd {
 
 	void jpeg_decoder::transform_mcu(int mcu_row)
 	{
-		jpgd_block_coeff_t* pSrc_ptr = m_pMCU_coefficients;
+		jpgd_block_t* pSrc_ptr = m_pMCU_coefficients;
 		if (mcu_row * m_blocks_per_mcu >= m_max_blocks_per_row)
 			stop_decoding(JPGD_DECODE_ERROR);
 
@@ -1343,7 +1290,7 @@ namespace jpgd {
 
 		for (int mcu_block = 0; mcu_block < m_blocks_per_mcu; mcu_block++)
 		{
-			idct(pSrc_ptr, pDst_ptr, m_mcu_block_max_zag[mcu_block], ((m_flags & cFlagDisableSIMD) == 0) && m_has_sse2);
+			idct(pSrc_ptr, pDst_ptr, m_mcu_block_max_zag[mcu_block]);
 			pSrc_ptr += 64;
 			pDst_ptr += 64;
 		}
@@ -1354,9 +1301,9 @@ namespace jpgd {
 	void jpeg_decoder::load_next_row()
 	{
 		int i;
-		jpgd_block_coeff_t* p;
+		jpgd_block_t* p;
 		jpgd_quant_t* q;
-		int mcu_row, mcu_block, row_block = 0;
+		int mcu_row, mcu_block;// , row_block = 0;
 		int component_num, component_id;
 		int block_x_mcu[JPGD_MAX_COMPONENTS];
 
@@ -1376,10 +1323,10 @@ namespace jpgd {
 
 				p = m_pMCU_coefficients + 64 * mcu_block;
 
-				jpgd_block_coeff_t* pAC = coeff_buf_getp(m_ac_coeffs[component_id], block_x_mcu[component_id] + block_x_mcu_ofs, m_block_y_mcu[component_id] + block_y_mcu_ofs);
-				jpgd_block_coeff_t* pDC = coeff_buf_getp(m_dc_coeffs[component_id], block_x_mcu[component_id] + block_x_mcu_ofs, m_block_y_mcu[component_id] + block_y_mcu_ofs);
+				jpgd_block_t* pAC = coeff_buf_getp(m_ac_coeffs[component_id], block_x_mcu[component_id] + block_x_mcu_ofs, m_block_y_mcu[component_id] + block_y_mcu_ofs);
+				jpgd_block_t* pDC = coeff_buf_getp(m_dc_coeffs[component_id], block_x_mcu[component_id] + block_x_mcu_ofs, m_block_y_mcu[component_id] + block_y_mcu_ofs);
 				p[0] = pDC[0];
-				memcpy(&p[1], &pAC[1], 63 * sizeof(jpgd_block_coeff_t));
+				memcpy(&p[1], &pAC[1], 63 * sizeof(jpgd_block_t));
 
 				for (i = 63; i > 0; i--)
 					if (p[g_ZAG[i]])
@@ -1389,9 +1336,9 @@ namespace jpgd {
 
 				for (; i >= 0; i--)
 					if (p[g_ZAG[i]])
-						p[g_ZAG[i]] = static_cast<jpgd_block_coeff_t>(p[g_ZAG[i]] * q[i]);
+						p[g_ZAG[i]] = static_cast<jpgd_block_t>(p[g_ZAG[i]] * q[i]);
 
-				row_block++;
+				//row_block++;
 
 				if (m_comps_in_scan == 1)
 					block_x_mcu[component_id]++;
@@ -1478,14 +1425,14 @@ namespace jpgd {
 	// Decodes and dequantizes the next row of coefficients.
 	void jpeg_decoder::decode_next_row()
 	{
-		int row_block = 0;
+		//int row_block = 0;
 
 		for (int mcu_row = 0; mcu_row < m_mcus_per_row; mcu_row++)
 		{
 			if ((m_restart_interval) && (m_restarts_left == 0))
 				process_restart();
 
-			jpgd_block_coeff_t* p = m_pMCU_coefficients;
+			jpgd_block_t* p = m_pMCU_coefficients;
 			for (int mcu_block = 0; mcu_block < m_blocks_per_mcu; mcu_block++, p += 64)
 			{
 				int component_id = m_mcu_org[mcu_block];
@@ -1503,7 +1450,7 @@ namespace jpgd {
 
 				m_last_dc_val[component_id] = (s += m_last_dc_val[component_id]);
 
-				p[0] = static_cast<jpgd_block_coeff_t>(s * q[0]);
+				p[0] = static_cast<jpgd_block_t>(s * q[0]);
 
 				int prev_num_set = m_mcu_block_max_zag[mcu_block];
 
@@ -1541,7 +1488,7 @@ namespace jpgd {
 						if (k >= 64)
 							stop_decoding(JPGD_DECODE_ERROR);
 
-						p[g_ZAG[k]] = static_cast<jpgd_block_coeff_t>(dequantize_ac(s, q[k])); //s * q[k];
+						p[g_ZAG[k]] = static_cast<jpgd_block_t>(dequantize_ac(s, q[k])); //s * q[k];
 					}
 					else
 					{
@@ -1581,7 +1528,7 @@ namespace jpgd {
 
 				m_mcu_block_max_zag[mcu_block] = k;
 
-				row_block++;
+				//row_block++;
 			}
 
 			transform_mcu(mcu_row);
@@ -1669,7 +1616,7 @@ namespace jpgd {
 		int row = m_max_mcu_y_size - m_mcu_lines_left;
 		uint8* d0 = m_pScan_line_0;
 
-		const int half_image_x_size = (m_image_x_size == 1) ? 0 : (m_image_x_size >> 1) - 1;
+		const int half_image_x_size = (m_image_x_size >> 1) - 1;
 		const int row_x8 = row * 8;
 
 		for (int x = 0; x < m_image_x_size; x++)
@@ -1762,7 +1709,7 @@ namespace jpgd {
 		int y = m_image_y_size - m_total_lines_left;
 		int row = y & 15;
 
-		const int half_image_y_size = (m_image_y_size == 1) ? 0 : (m_image_y_size >> 1) - 1;
+		const int half_image_y_size = (m_image_y_size >> 1) - 1;
 
 		uint8* d0 = m_pScan_line_0;
 
@@ -1891,7 +1838,7 @@ namespace jpgd {
 		int y = m_image_y_size - m_total_lines_left;
 		int row = y & 15;
 
-		const int half_image_y_size = (m_image_y_size == 1) ? 0 : (m_image_y_size >> 1) - 1;
+		const int half_image_y_size = (m_image_y_size >> 1) - 1;
 
 		uint8* d0 = m_pScan_line_0;
 
@@ -1915,7 +1862,7 @@ namespace jpgd {
 		const int y0_base = (c_y0 & 7) * 8 + 256;
 		const int y1_base = (c_y1 & 7) * 8 + 256;
 
-		const int half_image_x_size = (m_image_x_size == 1) ? 0 : (m_image_x_size >> 1) - 1;
+		const int half_image_x_size = (m_image_x_size >> 1) - 1;
 
 		static const uint8_t s_muls[2][2][4] =
 		{
@@ -2124,10 +2071,10 @@ namespace jpgd {
 
 	int jpeg_decoder::decode_next_mcu_row()
 	{
-		if (::setjmp(m_jmp_state))
+		if (setjmp(m_jmp_state))
 			return JPGD_FAILED;
 
-		const bool chroma_y_filtering = ((m_flags & cFlagBoxChromaFiltering) == 0) && ((m_scan_type == JPGD_YH2V2) || (m_scan_type == JPGD_YH1V2));
+		const bool chroma_y_filtering = (m_flags & cFlagLinearChromaFiltering) && ((m_scan_type == JPGD_YH2V2) || (m_scan_type == JPGD_YH1V2)) && (m_image_x_size >= 2) && (m_image_y_size >= 2);
 		if (chroma_y_filtering)
 		{
 			std::swap(m_pSample_buf, m_pSample_buf_prev);
@@ -2156,7 +2103,7 @@ namespace jpgd {
 		if (m_total_lines_left == 0)
 			return JPGD_DONE;
 
-		const bool chroma_y_filtering = ((m_flags & cFlagBoxChromaFiltering) == 0) && ((m_scan_type == JPGD_YH2V2) || (m_scan_type == JPGD_YH1V2));
+		const bool chroma_y_filtering = (m_flags & cFlagLinearChromaFiltering) && ((m_scan_type == JPGD_YH2V2) || (m_scan_type == JPGD_YH1V2)) && (m_image_x_size >= 2) && (m_image_y_size >= 2);
 
 		bool get_another_mcu_row = false;
 		bool got_mcu_early = false;
@@ -2186,7 +2133,7 @@ namespace jpgd {
 		{
 		case JPGD_YH2V2:
 		{
-			if ((m_flags & cFlagBoxChromaFiltering) == 0)
+			if ((m_flags & cFlagLinearChromaFiltering) && (m_image_x_size >= 2) && (m_image_y_size >= 2))
 			{
 				if (m_num_buffered_scanlines == 1)
 				{
@@ -2215,7 +2162,7 @@ namespace jpgd {
 		}
 		case JPGD_YH2V1:
 		{
-			if ((m_flags & cFlagBoxChromaFiltering) == 0)
+			if ((m_flags & cFlagLinearChromaFiltering) && (m_image_x_size >= 2) && (m_image_y_size >= 2))
 				H2V1ConvertFiltered();
 			else
 				H2V1Convert();
@@ -2630,9 +2577,9 @@ namespace jpgd {
 		m_real_dest_bytes_per_scan_line = (m_image_x_size * m_dest_bytes_per_pixel);
 
 		// Initialize two scan line buffers.
-		m_pScan_line_0 = (uint8*)alloc_aligned(m_dest_bytes_per_scan_line, true);
+		m_pScan_line_0 = (uint8*)alloc(m_dest_bytes_per_scan_line, true);
 		if ((m_scan_type == JPGD_YH1V2) || (m_scan_type == JPGD_YH2V2))
-			m_pScan_line_1 = (uint8*)alloc_aligned(m_dest_bytes_per_scan_line, true);
+			m_pScan_line_1 = (uint8*)alloc(m_dest_bytes_per_scan_line, true);
 
 		m_max_blocks_per_row = m_max_mcus_per_row * m_max_blocks_per_mcu;
 
@@ -2641,13 +2588,13 @@ namespace jpgd {
 			stop_decoding(JPGD_DECODE_ERROR);
 
 		// Allocate the coefficient buffer, enough for one MCU
-		m_pMCU_coefficients = (jpgd_block_coeff_t *)alloc_aligned(m_max_blocks_per_mcu * 64 * sizeof(jpgd_block_coeff_t));
-				
+		m_pMCU_coefficients = (jpgd_block_t*)alloc(m_max_blocks_per_mcu * 64 * sizeof(jpgd_block_t));
+
 		for (i = 0; i < m_max_blocks_per_mcu; i++)
 			m_mcu_block_max_zag[i] = 64;
 
-		m_pSample_buf = (uint8*)alloc_aligned(m_max_blocks_per_row * 64);
-		m_pSample_buf_prev = (uint8*)alloc_aligned(m_max_blocks_per_row * 64);
+		m_pSample_buf = (uint8*)alloc(m_max_blocks_per_row * 64);
+		m_pSample_buf_prev = (uint8*)alloc(m_max_blocks_per_row * 64);
 
 		m_total_lines_left = m_image_y_size;
 
@@ -2668,17 +2615,17 @@ namespace jpgd {
 		cb->block_num_y = block_num_y;
 		cb->block_len_x = block_len_x;
 		cb->block_len_y = block_len_y;
-		cb->block_size = (block_len_x * block_len_y) * sizeof(jpgd_block_coeff_t);
+		cb->block_size = (block_len_x * block_len_y) * sizeof(jpgd_block_t);
 		cb->pData = (uint8*)alloc(cb->block_size * block_num_x * block_num_y, true);
 		return cb;
 	}
 
-	inline jpgd_block_coeff_t* jpeg_decoder::coeff_buf_getp(coeff_buf* cb, int block_x, int block_y)
+	inline jpgd_block_t* jpeg_decoder::coeff_buf_getp(coeff_buf* cb, int block_x, int block_y)
 	{
 		if ((block_x >= cb->block_num_x) || (block_y >= cb->block_num_y))
 			stop_decoding(JPGD_DECODE_ERROR);
 
-		return (jpgd_block_coeff_t*)(cb->pData + block_x * cb->block_size + block_y * (cb->block_size * cb->block_num_x));
+		return (jpgd_block_t*)(cb->pData + block_x * cb->block_size + block_y * (cb->block_size * cb->block_num_x));
 	}
 
 	// The following methods decode the various types of m_blocks encountered
@@ -2686,7 +2633,7 @@ namespace jpgd {
 	void jpeg_decoder::decode_block_dc_first(jpeg_decoder* pD, int component_id, int block_x, int block_y)
 	{
 		int s, r;
-		jpgd_block_coeff_t* p = pD->coeff_buf_getp(pD->m_dc_coeffs[component_id], block_x, block_y);
+		jpgd_block_t* p = pD->coeff_buf_getp(pD->m_dc_coeffs[component_id], block_x, block_y);
 
 		if ((s = pD->huff_decode(pD->m_pHuff_tabs[pD->m_comp_dc_tab[component_id]])) != 0)
 		{
@@ -2699,14 +2646,14 @@ namespace jpgd {
 
 		pD->m_last_dc_val[component_id] = (s += pD->m_last_dc_val[component_id]);
 
-		p[0] = static_cast<jpgd_block_coeff_t>(s << pD->m_successive_low);
+		p[0] = static_cast<jpgd_block_t>(s << pD->m_successive_low);
 	}
 
 	void jpeg_decoder::decode_block_dc_refine(jpeg_decoder* pD, int component_id, int block_x, int block_y)
 	{
 		if (pD->get_bits_no_markers(1))
 		{
-			jpgd_block_coeff_t* p = pD->coeff_buf_getp(pD->m_dc_coeffs[component_id], block_x, block_y);
+			jpgd_block_t* p = pD->coeff_buf_getp(pD->m_dc_coeffs[component_id], block_x, block_y);
 
 			p[0] |= (1 << pD->m_successive_low);
 		}
@@ -2722,7 +2669,7 @@ namespace jpgd {
 			return;
 		}
 
-		jpgd_block_coeff_t* p = pD->coeff_buf_getp(pD->m_ac_coeffs[component_id], block_x, block_y);
+		jpgd_block_t* p = pD->coeff_buf_getp(pD->m_ac_coeffs[component_id], block_x, block_y);
 
 		for (k = pD->m_spectral_start; k <= pD->m_spectral_end; k++)
 		{
@@ -2743,7 +2690,7 @@ namespace jpgd {
 				r = pD->get_bits_no_markers(s);
 				s = JPGD_HUFF_EXTEND(r, s);
 
-				p[g_ZAG[k]] = static_cast<jpgd_block_coeff_t>(s << pD->m_successive_low);
+				p[g_ZAG[k]] = static_cast<jpgd_block_t>(s << pD->m_successive_low);
 			}
 			else
 			{
@@ -2776,7 +2723,7 @@ namespace jpgd {
 		//int m1 = (-1) << pD->m_successive_low;
 		int m1 = static_cast<int>((UINT32_MAX << pD->m_successive_low));
 
-		jpgd_block_coeff_t* p = pD->coeff_buf_getp(pD->m_ac_coeffs[component_id], block_x, block_y);
+		jpgd_block_t* p = pD->coeff_buf_getp(pD->m_ac_coeffs[component_id], block_x, block_y);
 		if (pD->m_spectral_end > 63)
 			pD->stop_decoding(JPGD_DECODE_ERROR);
 
@@ -2820,7 +2767,7 @@ namespace jpgd {
 
 				do
 				{
-					jpgd_block_coeff_t* this_coef = p + g_ZAG[k & 63];
+					jpgd_block_t* this_coef = p + g_ZAG[k & 63];
 
 					if (*this_coef != 0)
 					{
@@ -2829,9 +2776,9 @@ namespace jpgd {
 							if ((*this_coef & p1) == 0)
 							{
 								if (*this_coef >= 0)
-									*this_coef = static_cast<jpgd_block_coeff_t>(*this_coef + p1);
+									*this_coef = static_cast<jpgd_block_t>(*this_coef + p1);
 								else
-									*this_coef = static_cast<jpgd_block_coeff_t>(*this_coef + m1);
+									*this_coef = static_cast<jpgd_block_t>(*this_coef + m1);
 							}
 						}
 					}
@@ -2847,7 +2794,7 @@ namespace jpgd {
 
 				if ((s) && (k < 64))
 				{
-					p[g_ZAG[k]] = static_cast<jpgd_block_coeff_t>(s);
+					p[g_ZAG[k]] = static_cast<jpgd_block_t>(s);
 				}
 			}
 		}
@@ -2856,7 +2803,7 @@ namespace jpgd {
 		{
 			for (; k <= pD->m_spectral_end; k++)
 			{
-				jpgd_block_coeff_t* this_coef = p + g_ZAG[k & 63]; // logical AND to shut up static code analysis
+				jpgd_block_t* this_coef = p + g_ZAG[k & 63]; // logical AND to shut up static code analysis
 
 				if (*this_coef != 0)
 				{
@@ -2865,9 +2812,9 @@ namespace jpgd {
 						if ((*this_coef & p1) == 0)
 						{
 							if (*this_coef >= 0)
-								*this_coef = static_cast<jpgd_block_coeff_t>(*this_coef + p1);
+								*this_coef = static_cast<jpgd_block_t>(*this_coef + p1);
 							else
-								*this_coef = static_cast<jpgd_block_coeff_t>(*this_coef + m1);
+								*this_coef = static_cast<jpgd_block_t>(*this_coef + m1);
 						}
 					}
 				}
@@ -3040,7 +2987,7 @@ namespace jpgd {
 
 	jpeg_decoder::jpeg_decoder(jpeg_decoder_stream* pStream, uint32_t flags)
 	{
-		if (::setjmp(m_jmp_state))
+		if (setjmp(m_jmp_state))
 			return;
 		decode_init(pStream, flags);
 	}
@@ -3053,7 +3000,7 @@ namespace jpgd {
 		if (m_error_code)
 			return JPGD_FAILED;
 
-		if (::setjmp(m_jmp_state))
+		if (setjmp(m_jmp_state))
 			return JPGD_FAILED;
 
 		decode_start();

thirdparty/jpeg-compressor/jpgd.h → thirdparty/basis_universal/encoder/jpgd.h

@@ -1,6 +1,5 @@
 // jpgd.h - C++ class for JPEG decompression.
-// Richard Geldreich <[email protected]>
-// See jpgd.cpp for license (Public Domain or Apache 2.0).
+// Public domain, Rich Geldreich <[email protected]>
 #ifndef JPEG_DECODER_H
 #define JPEG_DECODER_H
 
@@ -119,20 +118,19 @@ namespace jpgd
 	};
 
 	typedef int16 jpgd_quant_t;
-	typedef int16 jpgd_block_coeff_t;
+	typedef int16 jpgd_block_t;
 
 	class jpeg_decoder
 	{
 	public:
 		enum
 		{
-			cFlagBoxChromaFiltering = 1,
-			cFlagDisableSIMD = 2
+			cFlagLinearChromaFiltering = 1
 		};
 
 		// Call get_error_code() after constructing to determine if the stream is valid or not. You may call the get_width(), get_height(), etc.
 		// methods after the constructor is called. You may then either destruct the object, or begin decoding the image by calling begin_decoding(), then decode() on each scanline.
-		jpeg_decoder(jpeg_decoder_stream* pStream, uint32_t flags = 0);
+		jpeg_decoder(jpeg_decoder_stream* pStream, uint32_t flags = cFlagLinearChromaFiltering);
 
 		~jpeg_decoder();
 
@@ -257,7 +255,7 @@ namespace jpgd
 
 		int m_max_mcus_per_col;
 		uint m_last_dc_val[JPGD_MAX_COMPONENTS];
-		jpgd_block_coeff_t* m_pMCU_coefficients;
+		jpgd_block_t* m_pMCU_coefficients;
 		int m_mcu_block_max_zag[JPGD_MAX_BLOCKS_PER_MCU];
 		uint8* m_pSample_buf;
 		uint8* m_pSample_buf_prev;
@@ -273,13 +271,11 @@ namespace jpgd
 		bool m_ready_flag;
 		bool m_eof_flag;
 		bool m_sample_buf_prev_valid;
-		bool m_has_sse2;
 
 		inline int check_sample_buf_ofs(int ofs) const { assert(ofs >= 0); assert(ofs < m_max_blocks_per_row * 64); return ofs; }
 		void free_all_blocks();
 		JPGD_NORETURN void stop_decoding(jpgd_status status);
 		void* alloc(size_t n, bool zero = false);
-		void* alloc_aligned(size_t nSize, uint32_t align = 16, bool zero = false);
 		void word_clear(void* p, uint16 c, uint n);
 		void prep_in_buffer();
 		void read_dht_marker();
@@ -298,7 +294,7 @@ namespace jpgd
 		void fix_in_buffer();
 		void transform_mcu(int mcu_row);
 		coeff_buf* coeff_buf_open(int block_num_x, int block_num_y, int block_len_x, int block_len_y);
-		inline jpgd_block_coeff_t* coeff_buf_getp(coeff_buf* cb, int block_x, int block_y);
+		inline jpgd_block_t* coeff_buf_getp(coeff_buf* cb, int block_x, int block_y);
 		void load_next_row();
 		void decode_next_row();
 		void make_huff_table(int index, huff_tables* pH);

thirdparty/basis_universal/patches/0002-external-jpgd.patch

@@ -1,22 +0,0 @@
-diff --git a/thirdparty/basis_universal/encoder/basisu_enc.cpp b/thirdparty/basis_universal/encoder/basisu_enc.cpp
-index 1f870c5de4..1cc982b134 100644
---- a/thirdparty/basis_universal/encoder/basisu_enc.cpp
-+++ b/thirdparty/basis_universal/encoder/basisu_enc.cpp
-@@ -504,7 +504,7 @@ namespace basisu
- 	bool load_jpg(const char *pFilename, image& img)
- 	{
- 		int width = 0, height = 0, actual_comps = 0;
--		uint8_t *pImage_data = jpgd::decompress_jpeg_image_from_file(pFilename, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
-+		uint8_t *pImage_data = jpgd::decompress_jpeg_image_from_file(pFilename, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagBoxChromaFiltering);
- 		if (!pImage_data)
- 			return false;
- 		
-@@ -524,7 +524,7 @@ namespace basisu
- 		}
- 
- 		int width = 0, height = 0, actual_comps = 0;
--		uint8_t* pImage_data = jpgd::decompress_jpeg_image_from_memory(pBuf, (int)buf_size, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
-+		uint8_t* pImage_data = jpgd::decompress_jpeg_image_from_memory(pBuf, (int)buf_size, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagBoxChromaFiltering);
- 		if (!pImage_data)
- 			return false;
- 

thirdparty/jpeg-compressor/jpgd_idct.h

@@ -1,462 +0,0 @@
-// Copyright 2009 Intel Corporation
-// All Rights Reserved
-//
-// Permission is granted to use, copy, distribute and prepare derivative works of this
-// software for any purpose and without fee, provided, that the above copyright notice
-// and this statement appear in all copies.  Intel makes no representations about the
-// suitability of this software for any purpose.  THIS SOFTWARE IS PROVIDED "AS IS."
-// INTEL SPECIFICALLY DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, AND ALL LIABILITY,
-// INCLUDING CONSEQUENTIAL AND OTHER INDIRECT DAMAGES, FOR THE USE OF THIS SOFTWARE,
-// INCLUDING LIABILITY FOR INFRINGEMENT OF ANY PROPRIETARY RIGHTS, AND INCLUDING THE
-// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.  Intel does not
-// assume any responsibility for any errors which may appear in this software nor any
-// responsibility to update it.
-//
-// From:
-// https://software.intel.com/sites/default/files/m/d/4/1/d/8/UsingIntelAVXToImplementIDCT-r1_5.pdf
-// https://software.intel.com/file/29048
-//
-// Requires SSE
-//
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-#include <immintrin.h>
-
-#ifdef _MSC_VER
-	#define JPGD_SIMD_ALIGN(type, name) __declspec(align(16)) type name
-#else
-	#define JPGD_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
-#endif
-
-#define BITS_INV_ACC 4
-#define SHIFT_INV_ROW 16 - BITS_INV_ACC
-#define SHIFT_INV_COL 1 + BITS_INV_ACC
-const short IRND_INV_ROW = 1024 * (6 - BITS_INV_ACC);	//1 << (SHIFT_INV_ROW-1)
-const short IRND_INV_COL = 16 * (BITS_INV_ACC - 3);		// 1 << (SHIFT_INV_COL-1)
-const short IRND_INV_CORR = IRND_INV_COL - 1;			// correction -1.0 and round
-
-JPGD_SIMD_ALIGN(short, shortM128_one_corr[8]) = {1, 1, 1, 1, 1, 1, 1, 1};
-JPGD_SIMD_ALIGN(short, shortM128_round_inv_row[8]) = {IRND_INV_ROW, 0, IRND_INV_ROW, 0, IRND_INV_ROW, 0, IRND_INV_ROW, 0};
-JPGD_SIMD_ALIGN(short, shortM128_round_inv_col[8]) = {IRND_INV_COL, IRND_INV_COL, IRND_INV_COL, IRND_INV_COL, IRND_INV_COL, IRND_INV_COL, IRND_INV_COL, IRND_INV_COL};
-JPGD_SIMD_ALIGN(short, shortM128_round_inv_corr[8])= {IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR, IRND_INV_CORR};
-JPGD_SIMD_ALIGN(short, shortM128_tg_1_16[8]) = {13036, 13036, 13036, 13036, 13036, 13036, 13036, 13036}; // tg * (2<<16) + 0.5
-JPGD_SIMD_ALIGN(short, shortM128_tg_2_16[8]) = {27146, 27146, 27146, 27146, 27146, 27146, 27146, 27146}; // tg * (2<<16) + 0.5
-JPGD_SIMD_ALIGN(short, shortM128_tg_3_16[8]) = {-21746, -21746, -21746, -21746, -21746, -21746, -21746, -21746}; // tg * (2<<16) + 0.5
-JPGD_SIMD_ALIGN(short, shortM128_cos_4_16[8]) = {-19195, -19195, -19195, -19195, -19195, -19195, -19195, -19195};// cos * (2<<16) + 0.5
-
-//-----------------------------------------------------------------------------
-// Table for rows 0,4 - constants are multiplied on cos_4_16
-// w15 w14 w11 w10 w07 w06 w03 w02
-// w29 w28 w25 w24 w21 w20 w17 w16
-// w31 w30 w27 w26 w23 w22 w19 w18
-//movq -> w05 w04 w01 w00
-JPGD_SIMD_ALIGN(short, shortM128_tab_i_04[]) = {
-	16384, 21407, 16384, 8867,
-	16384, -8867, 16384, -21407, // w13 w12 w09 w08
-	16384, 8867, -16384, -21407, // w07 w06 w03 w02
-	-16384, 21407, 16384, -8867, // w15 w14 w11 w10
-	22725, 19266, 19266, -4520, // w21 w20 w17 w16
-	12873, -22725, 4520, -12873, // w29 w28 w25 w24
-	12873, 4520, -22725, -12873, // w23 w22 w19 w18
-	4520, 19266, 19266, -22725}; // w31 w30 w27 w26
-
-	// Table for rows 1,7 - constants are multiplied on cos_1_16
-//movq -> w05 w04 w01 w00
-JPGD_SIMD_ALIGN(short, shortM128_tab_i_17[]) = {
-	22725, 29692, 22725, 12299,
-	22725, -12299, 22725, -29692, // w13 w12 w09 w08
-	22725, 12299, -22725, -29692, // w07 w06 w03 w02
-	-22725, 29692, 22725, -12299, // w15 w14 w11 w10
-	31521, 26722, 26722, -6270, // w21 w20 w17 w16
-	17855, -31521, 6270, -17855, // w29 w28 w25 w24
-	17855, 6270, -31521, -17855, // w23 w22 w19 w18
-	6270, 26722, 26722, -31521}; // w31 w30 w27 w26
-
-// Table for rows 2,6 - constants are multiplied on cos_2_16
-//movq -> w05 w04 w01 w00
-JPGD_SIMD_ALIGN(short, shortM128_tab_i_26[]) = {
-	21407, 27969, 21407, 11585,
-	21407, -11585, 21407, -27969, // w13 w12 w09 w08
-	21407, 11585, -21407, -27969, // w07 w06 w03 w02
-	-21407, 27969, 21407, -11585, // w15 w14 w11 w10
-	29692, 25172, 25172, -5906,	// w21 w20 w17 w16
-	16819, -29692, 5906, -16819, // w29 w28 w25 w24
-	16819, 5906, -29692, -16819, // w23 w22 w19 w18
-	5906, 25172, 25172, -29692}; // w31 w30 w27 w26
-// Table for rows 3,5 - constants are multiplied on cos_3_16
-//movq -> w05 w04 w01 w00
-JPGD_SIMD_ALIGN(short, shortM128_tab_i_35[]) = {
-	19266, 25172, 19266, 10426,
-	19266, -10426, 19266, -25172, // w13 w12 w09 w08
-	19266, 10426, -19266, -25172, // w07 w06 w03 w02
-	-19266, 25172, 19266, -10426, // w15 w14 w11 w10
-	26722, 22654, 22654, -5315, // w21 w20 w17 w16
-	15137, -26722, 5315, -15137, // w29 w28 w25 w24
-	15137, 5315, -26722, -15137, // w23 w22 w19 w18
-	5315, 22654, 22654, -26722}; // w31 w30 w27 w26
-
-JPGD_SIMD_ALIGN(short, shortM128_128[8]) = { 128, 128, 128, 128, 128, 128, 128, 128 };
-
-void idctSSEShortU8(const short *pInput, uint8_t * pOutputUB)
-{
-	__m128i r_xmm0, r_xmm4;
-	__m128i r_xmm1, r_xmm2, r_xmm3, r_xmm5, r_xmm6, r_xmm7;
-	__m128i row0, row1, row2, row3, row4, row5, row6, row7;
-	short * pTab_i_04 = shortM128_tab_i_04;
-	short * pTab_i_26 = shortM128_tab_i_26;
-
-	//Get pointers for this input and output
-	pTab_i_04 = shortM128_tab_i_04;
-	pTab_i_26 = shortM128_tab_i_26;
-
-	//Row 1 and Row 3
-	r_xmm0 = _mm_load_si128((__m128i *) pInput);
-	r_xmm4 = _mm_load_si128((__m128i *) (&pInput[2*8]));
-
-	// *** Work on the data in xmm0
-	//low shuffle mask = 0xd8 = 11 01 10 00
-	//get short 2 and short 0 into ls 32-bits
-	r_xmm0 = _mm_shufflelo_epi16(r_xmm0, 0xd8);
-
-	// copy short 2 and short 0 to all locations
-	r_xmm1 = _mm_shuffle_epi32(r_xmm0, 0);
-		
-	// add to those copies
-	r_xmm1 = _mm_madd_epi16(r_xmm1, *((__m128i *) pTab_i_04));
-
-	// shuffle mask = 0x55 = 01 01 01 01
-	// copy short 3 and short 1 to all locations
-	r_xmm3 = _mm_shuffle_epi32(r_xmm0, 0x55);
-		
-	// high shuffle mask = 0xd8 = 11 01 10 00
-	// get short 6 and short 4 into bit positions 64-95
-	// get short 7 and short 5 into bit positions 96-127
-	r_xmm0 = _mm_shufflehi_epi16(r_xmm0, 0xd8);
-		
-	// add to short 3 and short 1
-	r_xmm3 = _mm_madd_epi16(r_xmm3, *((__m128i *) &pTab_i_04[16]));
-		
-	// shuffle mask = 0xaa = 10 10 10 10
-	// copy short 6 and short 4 to all locations
-	r_xmm2 = _mm_shuffle_epi32(r_xmm0, 0xaa);
-		
-	// shuffle mask = 0xaa = 11 11 11 11
-	// copy short 7 and short 5 to all locations
-	r_xmm0 = _mm_shuffle_epi32(r_xmm0, 0xff);
-		
-	// add to short 6 and short 4
-	r_xmm2 = _mm_madd_epi16(r_xmm2, *((__m128i *) &pTab_i_04[8])); 
-		
-	// *** Work on the data in xmm4
-	// high shuffle mask = 0xd8 11 01 10 00
-	// get short 6 and short 4 into bit positions 64-95
-	// get short 7 and short 5 into bit positions 96-127
-	r_xmm4 = _mm_shufflehi_epi16(r_xmm4, 0xd8);
-		
-	// (xmm0 short 2 and short 0 plus pSi) + some constants
-	r_xmm1 = _mm_add_epi32(r_xmm1, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_shufflelo_epi16(r_xmm4, 0xd8);
-	r_xmm0 = _mm_madd_epi16(r_xmm0, *((__m128i *) &pTab_i_04[24]));
-	r_xmm5 = _mm_shuffle_epi32(r_xmm4, 0);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm4, 0xaa);
-	r_xmm5 = _mm_madd_epi16(r_xmm5, *((__m128i *) &shortM128_tab_i_26[0]));
-	r_xmm1 = _mm_add_epi32(r_xmm1, r_xmm2);
-	r_xmm2 = r_xmm1;
-	r_xmm7 = _mm_shuffle_epi32(r_xmm4, 0x55);
-	r_xmm6 = _mm_madd_epi16(r_xmm6, *((__m128i *) &shortM128_tab_i_26[8])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm3);
-	r_xmm4 = _mm_shuffle_epi32(r_xmm4, 0xff);
-	r_xmm2 = _mm_sub_epi32(r_xmm2, r_xmm0);
-	r_xmm7 = _mm_madd_epi16(r_xmm7, *((__m128i *) &shortM128_tab_i_26[16])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm1);
-	r_xmm2 = _mm_srai_epi32(r_xmm2, 12);
-	r_xmm5 = _mm_add_epi32(r_xmm5, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_madd_epi16(r_xmm4, *((__m128i *) &shortM128_tab_i_26[24]));
-	r_xmm5 = _mm_add_epi32(r_xmm5, r_xmm6);
-	r_xmm6 = r_xmm5;
-	r_xmm0 = _mm_srai_epi32(r_xmm0, 12);
-	r_xmm2 = _mm_shuffle_epi32(r_xmm2, 0x1b);
-	row0 = _mm_packs_epi32(r_xmm0, r_xmm2);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm7);
-	r_xmm6 = _mm_sub_epi32(r_xmm6, r_xmm4);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm5);
-	r_xmm6 = _mm_srai_epi32(r_xmm6, 12);
-	r_xmm4 = _mm_srai_epi32(r_xmm4, 12);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm6, 0x1b);
-	row2 = _mm_packs_epi32(r_xmm4, r_xmm6);
-
-	//Row 5 and row 7
-	r_xmm0 = _mm_load_si128((__m128i *) (&pInput[4*8]));
-	r_xmm4 = _mm_load_si128((__m128i *) (&pInput[6*8]));
-
-	r_xmm0 = _mm_shufflelo_epi16(r_xmm0, 0xd8);
-	r_xmm1 = _mm_shuffle_epi32(r_xmm0, 0);
-	r_xmm1 = _mm_madd_epi16(r_xmm1, *((__m128i *) pTab_i_04));
-	r_xmm3 = _mm_shuffle_epi32(r_xmm0, 0x55);
-	r_xmm0 = _mm_shufflehi_epi16(r_xmm0, 0xd8);
-	r_xmm3 = _mm_madd_epi16(r_xmm3, *((__m128i *) &pTab_i_04[16]));
-	r_xmm2 = _mm_shuffle_epi32(r_xmm0, 0xaa);
-	r_xmm0 = _mm_shuffle_epi32(r_xmm0, 0xff);
-	r_xmm2 = _mm_madd_epi16(r_xmm2, *((__m128i *) &pTab_i_04[8])); 
-	r_xmm4 = _mm_shufflehi_epi16(r_xmm4, 0xd8);
-	r_xmm1 = _mm_add_epi32(r_xmm1, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_shufflelo_epi16(r_xmm4, 0xd8);
-	r_xmm0 = _mm_madd_epi16(r_xmm0, *((__m128i *) &pTab_i_04[24]));
-	r_xmm5 = _mm_shuffle_epi32(r_xmm4, 0);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm4, 0xaa);
-	r_xmm5 = _mm_madd_epi16(r_xmm5, *((__m128i *) &shortM128_tab_i_26[0]));
-	r_xmm1 = _mm_add_epi32(r_xmm1, r_xmm2);
-	r_xmm2 = r_xmm1;
-	r_xmm7 = _mm_shuffle_epi32(r_xmm4, 0x55);
-	r_xmm6 = _mm_madd_epi16(r_xmm6, *((__m128i *) &shortM128_tab_i_26[8])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm3);
-	r_xmm4 = _mm_shuffle_epi32(r_xmm4, 0xff);
-	r_xmm2 = _mm_sub_epi32(r_xmm2, r_xmm0);
-	r_xmm7 = _mm_madd_epi16(r_xmm7, *((__m128i *) &shortM128_tab_i_26[16])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm1);
-	r_xmm2 = _mm_srai_epi32(r_xmm2, 12);
-	r_xmm5 = _mm_add_epi32(r_xmm5, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_madd_epi16(r_xmm4, *((__m128i *) &shortM128_tab_i_26[24]));
-	r_xmm5 = _mm_add_epi32(r_xmm5, r_xmm6);
-	r_xmm6 = r_xmm5;
-	r_xmm0 = _mm_srai_epi32(r_xmm0, 12);
-	r_xmm2 = _mm_shuffle_epi32(r_xmm2, 0x1b);
-	row4 = _mm_packs_epi32(r_xmm0, r_xmm2);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm7);
-	r_xmm6 = _mm_sub_epi32(r_xmm6, r_xmm4);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm5);
-	r_xmm6 = _mm_srai_epi32(r_xmm6, 12);
-	r_xmm4 = _mm_srai_epi32(r_xmm4, 12);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm6, 0x1b);
-	row6 = _mm_packs_epi32(r_xmm4, r_xmm6);
-
-	//Row 4 and row 2
-	pTab_i_04 = shortM128_tab_i_35;
-	pTab_i_26 = shortM128_tab_i_17;
-	r_xmm0 = _mm_load_si128((__m128i *) (&pInput[3*8]));
-	r_xmm4 = _mm_load_si128((__m128i *) (&pInput[1*8]));
-
-	r_xmm0 = _mm_shufflelo_epi16(r_xmm0, 0xd8);
-	r_xmm1 = _mm_shuffle_epi32(r_xmm0, 0);
-	r_xmm1 = _mm_madd_epi16(r_xmm1, *((__m128i *) pTab_i_04));
-	r_xmm3 = _mm_shuffle_epi32(r_xmm0, 0x55);
-	r_xmm0 = _mm_shufflehi_epi16(r_xmm0, 0xd8);
-	r_xmm3 = _mm_madd_epi16(r_xmm3, *((__m128i *) &pTab_i_04[16]));
-	r_xmm2 = _mm_shuffle_epi32(r_xmm0, 0xaa);
-	r_xmm0 = _mm_shuffle_epi32(r_xmm0, 0xff);
-	r_xmm2 = _mm_madd_epi16(r_xmm2, *((__m128i *) &pTab_i_04[8])); 
-	r_xmm4 = _mm_shufflehi_epi16(r_xmm4, 0xd8);
-	r_xmm1 = _mm_add_epi32(r_xmm1, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_shufflelo_epi16(r_xmm4, 0xd8);
-	r_xmm0 = _mm_madd_epi16(r_xmm0, *((__m128i *) &pTab_i_04[24]));
-	r_xmm5 = _mm_shuffle_epi32(r_xmm4, 0);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm4, 0xaa);
-	r_xmm5 = _mm_madd_epi16(r_xmm5, *((__m128i *) &pTab_i_26[0]));
-	r_xmm1 = _mm_add_epi32(r_xmm1, r_xmm2);
-	r_xmm2 = r_xmm1;
-	r_xmm7 = _mm_shuffle_epi32(r_xmm4, 0x55);
-	r_xmm6 = _mm_madd_epi16(r_xmm6, *((__m128i *) &pTab_i_26[8])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm3);
-	r_xmm4 = _mm_shuffle_epi32(r_xmm4, 0xff);
-	r_xmm2 = _mm_sub_epi32(r_xmm2, r_xmm0);
-	r_xmm7 = _mm_madd_epi16(r_xmm7, *((__m128i *) &pTab_i_26[16])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm1);
-	r_xmm2 = _mm_srai_epi32(r_xmm2, 12);
-	r_xmm5 = _mm_add_epi32(r_xmm5, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_madd_epi16(r_xmm4, *((__m128i *) &pTab_i_26[24]));
-	r_xmm5 = _mm_add_epi32(r_xmm5, r_xmm6);
-	r_xmm6 = r_xmm5;
-	r_xmm0 = _mm_srai_epi32(r_xmm0, 12);
-	r_xmm2 = _mm_shuffle_epi32(r_xmm2, 0x1b);
-	row3 = _mm_packs_epi32(r_xmm0, r_xmm2);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm7);
-	r_xmm6 = _mm_sub_epi32(r_xmm6, r_xmm4);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm5);
-	r_xmm6 = _mm_srai_epi32(r_xmm6, 12);
-	r_xmm4 = _mm_srai_epi32(r_xmm4, 12);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm6, 0x1b);
-	row1 = _mm_packs_epi32(r_xmm4, r_xmm6);
-
-	//Row 6 and row 8
-	r_xmm0 = _mm_load_si128((__m128i *) (&pInput[5*8]));
-	r_xmm4 = _mm_load_si128((__m128i *) (&pInput[7*8]));
-
-	r_xmm0 = _mm_shufflelo_epi16(r_xmm0, 0xd8);
-	r_xmm1 = _mm_shuffle_epi32(r_xmm0, 0);
-	r_xmm1 = _mm_madd_epi16(r_xmm1, *((__m128i *) pTab_i_04));
-	r_xmm3 = _mm_shuffle_epi32(r_xmm0, 0x55);
-	r_xmm0 = _mm_shufflehi_epi16(r_xmm0, 0xd8);
-	r_xmm3 = _mm_madd_epi16(r_xmm3, *((__m128i *) &pTab_i_04[16]));
-	r_xmm2 = _mm_shuffle_epi32(r_xmm0, 0xaa);
-	r_xmm0 = _mm_shuffle_epi32(r_xmm0, 0xff);
-	r_xmm2 = _mm_madd_epi16(r_xmm2, *((__m128i *) &pTab_i_04[8])); 
-	r_xmm4 = _mm_shufflehi_epi16(r_xmm4, 0xd8);
-	r_xmm1 = _mm_add_epi32(r_xmm1, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_shufflelo_epi16(r_xmm4, 0xd8);
-	r_xmm0 = _mm_madd_epi16(r_xmm0, *((__m128i *) &pTab_i_04[24]));
-	r_xmm5 = _mm_shuffle_epi32(r_xmm4, 0);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm4, 0xaa);
-	r_xmm5 = _mm_madd_epi16(r_xmm5, *((__m128i *) &pTab_i_26[0]));
-	r_xmm1 = _mm_add_epi32(r_xmm1, r_xmm2);
-	r_xmm2 = r_xmm1;
-	r_xmm7 = _mm_shuffle_epi32(r_xmm4, 0x55);
-	r_xmm6 = _mm_madd_epi16(r_xmm6, *((__m128i *) &pTab_i_26[8])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm3);
-	r_xmm4 = _mm_shuffle_epi32(r_xmm4, 0xff);
-	r_xmm2 = _mm_sub_epi32(r_xmm2, r_xmm0);
-	r_xmm7 = _mm_madd_epi16(r_xmm7, *((__m128i *) &pTab_i_26[16])); 
-	r_xmm0 = _mm_add_epi32(r_xmm0, r_xmm1);
-	r_xmm2 = _mm_srai_epi32(r_xmm2, 12);
-	r_xmm5 = _mm_add_epi32(r_xmm5, *((__m128i *) shortM128_round_inv_row));
-	r_xmm4 = _mm_madd_epi16(r_xmm4, *((__m128i *) &pTab_i_26[24]));
-	r_xmm5 = _mm_add_epi32(r_xmm5, r_xmm6);
-	r_xmm6 = r_xmm5;
-	r_xmm0 = _mm_srai_epi32(r_xmm0, 12);
-	r_xmm2 = _mm_shuffle_epi32(r_xmm2, 0x1b);
-	row5 = _mm_packs_epi32(r_xmm0, r_xmm2);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm7);
-	r_xmm6 = _mm_sub_epi32(r_xmm6, r_xmm4);
-	r_xmm4 = _mm_add_epi32(r_xmm4, r_xmm5);
-	r_xmm6 = _mm_srai_epi32(r_xmm6, 12);
-	r_xmm4 = _mm_srai_epi32(r_xmm4, 12);
-	r_xmm6 = _mm_shuffle_epi32(r_xmm6, 0x1b);
-	row7 = _mm_packs_epi32(r_xmm4, r_xmm6);
-
-	r_xmm1 = _mm_load_si128((__m128i *) shortM128_tg_3_16);
-	r_xmm2 = row5;
-	r_xmm3 = row3;
-	r_xmm0 = _mm_mulhi_epi16(row5, r_xmm1);
-
-	r_xmm1 = _mm_mulhi_epi16(r_xmm1, r_xmm3);
-	r_xmm5 = _mm_load_si128((__m128i *) shortM128_tg_1_16);
-	r_xmm6 = row7;
-	r_xmm4 = _mm_mulhi_epi16(row7, r_xmm5);
-
-	r_xmm0 = _mm_adds_epi16(r_xmm0, r_xmm2);
-	r_xmm5 = _mm_mulhi_epi16(r_xmm5, row1);
-	r_xmm1 = _mm_adds_epi16(r_xmm1, r_xmm3);
-	r_xmm7 = row6;
-
-	r_xmm0 = _mm_adds_epi16(r_xmm0, r_xmm3);
-	r_xmm3 = _mm_load_si128((__m128i *) shortM128_tg_2_16);
-	r_xmm2 = _mm_subs_epi16(r_xmm2, r_xmm1);
-	r_xmm7 = _mm_mulhi_epi16(r_xmm7, r_xmm3);
-	r_xmm1 = r_xmm0;
-	r_xmm3 = _mm_mulhi_epi16(r_xmm3, row2);
-	r_xmm5 = _mm_subs_epi16(r_xmm5, r_xmm6);
-	r_xmm4 = _mm_adds_epi16(r_xmm4, row1);
-	r_xmm0 = _mm_adds_epi16(r_xmm0, r_xmm4);
-	r_xmm0 = _mm_adds_epi16(r_xmm0, *((__m128i *) shortM128_one_corr));
-	r_xmm4 = _mm_subs_epi16(r_xmm4, r_xmm1);
-	r_xmm6 = r_xmm5;
-	r_xmm5 = _mm_subs_epi16(r_xmm5, r_xmm2);
-	r_xmm5 = _mm_adds_epi16(r_xmm5, *((__m128i *) shortM128_one_corr));
-	r_xmm6 = _mm_adds_epi16(r_xmm6, r_xmm2);
-
-	//Intermediate results, needed later
-	__m128i temp3, temp7;
-	temp7 = r_xmm0;
-
-	r_xmm1 = r_xmm4;
-	r_xmm0 = _mm_load_si128((__m128i *) shortM128_cos_4_16);
-	r_xmm4 = _mm_adds_epi16(r_xmm4, r_xmm5);
-	r_xmm2 = _mm_load_si128((__m128i *) shortM128_cos_4_16);
-	r_xmm2 = _mm_mulhi_epi16(r_xmm2, r_xmm4);
-
-	//Intermediate results, needed later
-	temp3 = r_xmm6;
-
-	r_xmm1 = _mm_subs_epi16(r_xmm1, r_xmm5);
-	r_xmm7 = _mm_adds_epi16(r_xmm7, row2);
-	r_xmm3 = _mm_subs_epi16(r_xmm3, row6);
-	r_xmm6 = row0;
-	r_xmm0 = _mm_mulhi_epi16(r_xmm0, r_xmm1);
-	r_xmm5 = row4;
-	r_xmm5 = _mm_adds_epi16(r_xmm5, r_xmm6);
-	r_xmm6 = _mm_subs_epi16(r_xmm6, row4);
-	r_xmm4 = _mm_adds_epi16(r_xmm4, r_xmm2);
-
-	r_xmm4 = _mm_or_si128(r_xmm4, *((__m128i *) shortM128_one_corr));
-	r_xmm0 = _mm_adds_epi16(r_xmm0, r_xmm1);
-	r_xmm0 = _mm_or_si128(r_xmm0, *((__m128i *) shortM128_one_corr));
-
-	r_xmm2 = r_xmm5;
-	r_xmm5 = _mm_adds_epi16(r_xmm5, r_xmm7);
-	r_xmm1 = r_xmm6;
-	r_xmm5 = _mm_adds_epi16(r_xmm5, *((__m128i *) shortM128_round_inv_col));
-	r_xmm2 = _mm_subs_epi16(r_xmm2, r_xmm7);
-	r_xmm7 = temp7;
-	r_xmm6 = _mm_adds_epi16(r_xmm6, r_xmm3);
-	r_xmm6 = _mm_adds_epi16(r_xmm6, *((__m128i *) shortM128_round_inv_col));
-	r_xmm7 = _mm_adds_epi16(r_xmm7, r_xmm5);
-	r_xmm7 = _mm_srai_epi16(r_xmm7, SHIFT_INV_COL);
-	r_xmm1 = _mm_subs_epi16(r_xmm1, r_xmm3);
-	r_xmm1 = _mm_adds_epi16(r_xmm1, *((__m128i *) shortM128_round_inv_corr));
-	r_xmm3 = r_xmm6;
-	r_xmm2 = _mm_adds_epi16(r_xmm2, *((__m128i *) shortM128_round_inv_corr));
-	r_xmm6 = _mm_adds_epi16(r_xmm6, r_xmm4);
-
-	//Store results for row 0
-	//_mm_store_si128((__m128i *) pOutput, r_xmm7);
-	__m128i r0 = r_xmm7;
-
-	r_xmm6 = _mm_srai_epi16(r_xmm6, SHIFT_INV_COL);
-	r_xmm7 = r_xmm1;
-	r_xmm1 = _mm_adds_epi16(r_xmm1, r_xmm0);
-
-	//Store results for row 1
-	//_mm_store_si128((__m128i *) (&pOutput[1*8]), r_xmm6); 
-	__m128i r1 = r_xmm6;
-
-	r_xmm1 = _mm_srai_epi16(r_xmm1, SHIFT_INV_COL);
-	r_xmm6 = temp3;
-	r_xmm7 = _mm_subs_epi16(r_xmm7, r_xmm0);
-	r_xmm7 = _mm_srai_epi16(r_xmm7, SHIFT_INV_COL);
-
-	//Store results for row 2
-	//_mm_store_si128((__m128i *) (&pOutput[2*8]), r_xmm1); 
-	__m128i r2 = r_xmm1;
-
-	r_xmm5 = _mm_subs_epi16(r_xmm5, temp7); 
-	r_xmm5 = _mm_srai_epi16(r_xmm5, SHIFT_INV_COL);
-
-	//Store results for row 7
-	//_mm_store_si128((__m128i *) (&pOutput[7*8]), r_xmm5); 
-	__m128i r7 = r_xmm5;
-
-	r_xmm3 = _mm_subs_epi16(r_xmm3, r_xmm4);
-	r_xmm6 = _mm_adds_epi16(r_xmm6, r_xmm2);
-	r_xmm2 = _mm_subs_epi16(r_xmm2, temp3); 
-	r_xmm6 = _mm_srai_epi16(r_xmm6, SHIFT_INV_COL);
-	r_xmm2 = _mm_srai_epi16(r_xmm2, SHIFT_INV_COL);
-
-	//Store results for row 3
-	//_mm_store_si128((__m128i *) (&pOutput[3*8]), r_xmm6); 
-	__m128i r3 = r_xmm6;
-
-	r_xmm3 = _mm_srai_epi16(r_xmm3, SHIFT_INV_COL);
-
-	//Store results for rows 4, 5, and 6
-	//_mm_store_si128((__m128i *) (&pOutput[4*8]), r_xmm2);
-	//_mm_store_si128((__m128i *) (&pOutput[5*8]), r_xmm7);
-	//_mm_store_si128((__m128i *) (&pOutput[6*8]), r_xmm3);
-
-	__m128i r4 = r_xmm2;
-	__m128i r5 = r_xmm7;
-	__m128i r6 = r_xmm3;
-
-	r0 = _mm_add_epi16(*(const __m128i *)shortM128_128, r0);
-	r1 = _mm_add_epi16(*(const __m128i *)shortM128_128, r1);
-	r2 = _mm_add_epi16(*(const __m128i *)shortM128_128, r2);
-	r3 = _mm_add_epi16(*(const __m128i *)shortM128_128, r3);
-	r4 = _mm_add_epi16(*(const __m128i *)shortM128_128, r4);
-	r5 = _mm_add_epi16(*(const __m128i *)shortM128_128, r5);
-	r6 = _mm_add_epi16(*(const __m128i *)shortM128_128, r6);
-	r7 = _mm_add_epi16(*(const __m128i *)shortM128_128, r7);
-
-	((__m128i *)pOutputUB)[0] = _mm_packus_epi16(r0, r1);
-	((__m128i *)pOutputUB)[1] = _mm_packus_epi16(r2, r3);
-	((__m128i *)pOutputUB)[2] = _mm_packus_epi16(r4, r5);
-	((__m128i *)pOutputUB)[3] = _mm_packus_epi16(r6, r7);
-}

thirdparty/jpeg-compressor/jpge.cpp

@@ -1,1076 +0,0 @@
-// jpge.cpp - C++ class for JPEG compression. Richard Geldreich <[email protected]>
-// Supports grayscale, H1V1, H2V1, and H2V2 chroma subsampling factors, one or two pass Huffman table optimization, libjpeg-style quality 1-100 quality factors.
-// Also supports using luma quantization tables for chroma.
-//
-// Released under two licenses. You are free to choose which license you want:
-// License 1: 
-// Public Domain
-//
-// License 2:
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//    http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// v1.01, Dec. 18, 2010 - Initial release
-// v1.02, Apr. 6, 2011 - Removed 2x2 ordered dither in H2V1 chroma subsampling method load_block_16_8_8(). (The rounding factor was 2, when it should have been 1. Either way, it wasn't helping.)
-// v1.03, Apr. 16, 2011 - Added support for optimized Huffman code tables, optimized dynamic memory allocation down to only 1 alloc.
-//                        Also from Alex Evans: Added RGBA support, linear memory allocator (no longer needed in v1.03).
-// v1.04, May. 19, 2012: Forgot to set m_pFile ptr to NULL in cfile_stream::close(). Thanks to Owen Kaluza for reporting this bug.
-//                       Code tweaks to fix VS2008 static code analysis warnings (all looked harmless).
-//                       Code review revealed method load_block_16_8_8() (used for the non-default H2V1 sampling mode to downsample chroma) somehow didn't get the rounding factor fix from v1.02.
-// v1.05, March 25, 2020: Added Apache 2.0 alternate license
-
-#include "jpge.h"
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define JPGE_MAX(a,b) (((a)>(b))?(a):(b))
-#define JPGE_MIN(a,b) (((a)<(b))?(a):(b))
-
-namespace jpge {
-
-	static inline void* jpge_malloc(size_t nSize) { return malloc(nSize); }
-	static inline void jpge_free(void* p) { free(p); }
-
-	// Various JPEG enums and tables.
-	enum { M_SOF0 = 0xC0, M_DHT = 0xC4, M_SOI = 0xD8, M_EOI = 0xD9, M_SOS = 0xDA, M_DQT = 0xDB, M_APP0 = 0xE0 };
-	enum { DC_LUM_CODES = 12, AC_LUM_CODES = 256, DC_CHROMA_CODES = 12, AC_CHROMA_CODES = 256, MAX_HUFF_SYMBOLS = 257, MAX_HUFF_CODESIZE = 32 };
-
-	static uint8 s_zag[64] = { 0,1,8,16,9,2,3,10,17,24,32,25,18,11,4,5,12,19,26,33,40,48,41,34,27,20,13,6,7,14,21,28,35,42,49,56,57,50,43,36,29,22,15,23,30,37,44,51,58,59,52,45,38,31,39,46,53,60,61,54,47,55,62,63 };
-	static int16 s_std_lum_quant[64] = { 16,11,12,14,12,10,16,14,13,14,18,17,16,19,24,40,26,24,22,22,24,49,35,37,29,40,58,51,61,60,57,51,56,55,64,72,92,78,64,68,87,69,55,56,80,109,81,87,95,98,103,104,103,62,77,113,121,112,100,120,92,101,103,99 };
-	static int16 s_std_croma_quant[64] = { 17,18,18,24,21,24,47,26,26,47,99,66,56,66,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99 };
-
-	// Table from http://www.imagemagick.org/discourse-server/viewtopic.php?f=22&t=20333&p=98008#p98008
-	// This is mozjpeg's default table, in zag order.
-	static int16 s_alt_quant[64] = { 16,16,16,16,17,16,18,20,20,18,25,27,24,27,25,37,34,31,31,34,37,56,40,43,40,43,40,56,85,53,62,53,53,62,53,85,75,91,74,69,74,91,75,135,106,94,94,106,135,156,131,124,131,156,189,169,169,189,238,226,238,311,311,418 };
-
-	static uint8 s_dc_lum_bits[17] = { 0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0 };
-	static uint8 s_dc_lum_val[DC_LUM_CODES] = { 0,1,2,3,4,5,6,7,8,9,10,11 };
-	static uint8 s_ac_lum_bits[17] = { 0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d };
-	static uint8 s_ac_lum_val[AC_LUM_CODES] =
-	{
-	  0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08,0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,
-	  0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28,0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,
-	  0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89,
-	  0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,
-	  0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,
-	  0xf9,0xfa
-	};
-	static uint8 s_dc_chroma_bits[17] = { 0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0 };
-	static uint8 s_dc_chroma_val[DC_CHROMA_CODES] = { 0,1,2,3,4,5,6,7,8,9,10,11 };
-	static uint8 s_ac_chroma_bits[17] = { 0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77 };
-	static uint8 s_ac_chroma_val[AC_CHROMA_CODES] =
-	{
-	  0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91,0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,
-	  0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26,0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,
-	  0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87,
-	  0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,
-	  0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,
-	  0xf9,0xfa
-	};
-
-	// Low-level helper functions.
-	template <class T> inline void clear_obj(T& obj) { memset(&obj, 0, sizeof(obj)); }
-
-	const int YR = 19595, YG = 38470, YB = 7471, CB_R = -11059, CB_G = -21709, CB_B = 32768, CR_R = 32768, CR_G = -27439, CR_B = -5329;
-	static inline uint8 clamp(int i) { if (static_cast<uint>(i) > 255U) { if (i < 0) i = 0; else if (i > 255) i = 255; } return static_cast<uint8>(i); }
-
-	static inline int left_shifti(int val, uint32 bits)
-	{
-		return static_cast<int>(static_cast<uint32>(val) << bits);
-	}
-
-	static void RGB_to_YCC(uint8* pDst, const uint8* pSrc, int num_pixels)
-	{
-		for (; num_pixels; pDst += 3, pSrc += 3, num_pixels--)
-		{
-			const int r = pSrc[0], g = pSrc[1], b = pSrc[2];
-			pDst[0] = static_cast<uint8>((r * YR + g * YG + b * YB + 32768) >> 16);
-			pDst[1] = clamp(128 + ((r * CB_R + g * CB_G + b * CB_B + 32768) >> 16));
-			pDst[2] = clamp(128 + ((r * CR_R + g * CR_G + b * CR_B + 32768) >> 16));
-		}
-	}
-
-	static void RGB_to_Y(uint8* pDst, const uint8* pSrc, int num_pixels)
-	{
-		for (; num_pixels; pDst++, pSrc += 3, num_pixels--)
-			pDst[0] = static_cast<uint8>((pSrc[0] * YR + pSrc[1] * YG + pSrc[2] * YB + 32768) >> 16);
-	}
-
-	static void RGBA_to_YCC(uint8* pDst, const uint8* pSrc, int num_pixels)
-	{
-		for (; num_pixels; pDst += 3, pSrc += 4, num_pixels--)
-		{
-			const int r = pSrc[0], g = pSrc[1], b = pSrc[2];
-			pDst[0] = static_cast<uint8>((r * YR + g * YG + b * YB + 32768) >> 16);
-			pDst[1] = clamp(128 + ((r * CB_R + g * CB_G + b * CB_B + 32768) >> 16));
-			pDst[2] = clamp(128 + ((r * CR_R + g * CR_G + b * CR_B + 32768) >> 16));
-		}
-	}
-
-	static void RGBA_to_Y(uint8* pDst, const uint8* pSrc, int num_pixels)
-	{
-		for (; num_pixels; pDst++, pSrc += 4, num_pixels--)
-			pDst[0] = static_cast<uint8>((pSrc[0] * YR + pSrc[1] * YG + pSrc[2] * YB + 32768) >> 16);
-	}
-
-	static void Y_to_YCC(uint8* pDst, const uint8* pSrc, int num_pixels)
-	{
-		for (; num_pixels; pDst += 3, pSrc++, num_pixels--) { pDst[0] = pSrc[0]; pDst[1] = 128; pDst[2] = 128; }
-	}
-
-	// Forward DCT - DCT derived from jfdctint.
-	enum { CONST_BITS = 13, ROW_BITS = 2 };
-#define DCT_DESCALE(x, n) (((x) + (((int32)1) << ((n) - 1))) >> (n))
-#define DCT_MUL(var, c) (static_cast<int16>(var) * static_cast<int32>(c))
-#define DCT1D(s0, s1, s2, s3, s4, s5, s6, s7) \
-  int32 t0 = s0 + s7, t7 = s0 - s7, t1 = s1 + s6, t6 = s1 - s6, t2 = s2 + s5, t5 = s2 - s5, t3 = s3 + s4, t4 = s3 - s4; \
-  int32 t10 = t0 + t3, t13 = t0 - t3, t11 = t1 + t2, t12 = t1 - t2; \
-  int32 u1 = DCT_MUL(t12 + t13, 4433); \
-  s2 = u1 + DCT_MUL(t13, 6270); \
-  s6 = u1 + DCT_MUL(t12, -15137); \
-  u1 = t4 + t7; \
-  int32 u2 = t5 + t6, u3 = t4 + t6, u4 = t5 + t7; \
-  int32 z5 = DCT_MUL(u3 + u4, 9633); \
-  t4 = DCT_MUL(t4, 2446); t5 = DCT_MUL(t5, 16819); \
-  t6 = DCT_MUL(t6, 25172); t7 = DCT_MUL(t7, 12299); \
-  u1 = DCT_MUL(u1, -7373); u2 = DCT_MUL(u2, -20995); \
-  u3 = DCT_MUL(u3, -16069); u4 = DCT_MUL(u4, -3196); \
-  u3 += z5; u4 += z5; \
-  s0 = t10 + t11; s1 = t7 + u1 + u4; s3 = t6 + u2 + u3; s4 = t10 - t11; s5 = t5 + u2 + u4; s7 = t4 + u1 + u3;
-
-	static void DCT2D(int32* p)
-	{
-		int32 c, * q = p;
-		for (c = 7; c >= 0; c--, q += 8)
-		{
-			int32 s0 = q[0], s1 = q[1], s2 = q[2], s3 = q[3], s4 = q[4], s5 = q[5], s6 = q[6], s7 = q[7];
-			DCT1D(s0, s1, s2, s3, s4, s5, s6, s7);
-			q[0] = left_shifti(s0, ROW_BITS); q[1] = DCT_DESCALE(s1, CONST_BITS - ROW_BITS); q[2] = DCT_DESCALE(s2, CONST_BITS - ROW_BITS); q[3] = DCT_DESCALE(s3, CONST_BITS - ROW_BITS);
-			q[4] = left_shifti(s4, ROW_BITS); q[5] = DCT_DESCALE(s5, CONST_BITS - ROW_BITS); q[6] = DCT_DESCALE(s6, CONST_BITS - ROW_BITS); q[7] = DCT_DESCALE(s7, CONST_BITS - ROW_BITS);
-		}
-		for (q = p, c = 7; c >= 0; c--, q++)
-		{
-			int32 s0 = q[0 * 8], s1 = q[1 * 8], s2 = q[2 * 8], s3 = q[3 * 8], s4 = q[4 * 8], s5 = q[5 * 8], s6 = q[6 * 8], s7 = q[7 * 8];
-			DCT1D(s0, s1, s2, s3, s4, s5, s6, s7);
-			q[0 * 8] = DCT_DESCALE(s0, ROW_BITS + 3); q[1 * 8] = DCT_DESCALE(s1, CONST_BITS + ROW_BITS + 3); q[2 * 8] = DCT_DESCALE(s2, CONST_BITS + ROW_BITS + 3); q[3 * 8] = DCT_DESCALE(s3, CONST_BITS + ROW_BITS + 3);
-			q[4 * 8] = DCT_DESCALE(s4, ROW_BITS + 3); q[5 * 8] = DCT_DESCALE(s5, CONST_BITS + ROW_BITS + 3); q[6 * 8] = DCT_DESCALE(s6, CONST_BITS + ROW_BITS + 3); q[7 * 8] = DCT_DESCALE(s7, CONST_BITS + ROW_BITS + 3);
-		}
-	}
-
-	struct sym_freq { uint m_key, m_sym_index; };
-
-	// Radix sorts sym_freq[] array by 32-bit key m_key. Returns ptr to sorted values.
-	static inline sym_freq* radix_sort_syms(uint num_syms, sym_freq* pSyms0, sym_freq* pSyms1)
-	{
-		const uint cMaxPasses = 4;
-		uint32 hist[256 * cMaxPasses]; clear_obj(hist);
-		for (uint i = 0; i < num_syms; i++) { uint freq = pSyms0[i].m_key; hist[freq & 0xFF]++; hist[256 + ((freq >> 8) & 0xFF)]++; hist[256 * 2 + ((freq >> 16) & 0xFF)]++; hist[256 * 3 + ((freq >> 24) & 0xFF)]++; }
-		sym_freq* pCur_syms = pSyms0, * pNew_syms = pSyms1;
-		uint total_passes = cMaxPasses; while ((total_passes > 1) && (num_syms == hist[(total_passes - 1) * 256])) total_passes--;
-		for (uint pass_shift = 0, pass = 0; pass < total_passes; pass++, pass_shift += 8)
-		{
-			const uint32* pHist = &hist[pass << 8];
-			uint offsets[256], cur_ofs = 0;
-			for (uint i = 0; i < 256; i++) { offsets[i] = cur_ofs; cur_ofs += pHist[i]; }
-			for (uint i = 0; i < num_syms; i++)
-				pNew_syms[offsets[(pCur_syms[i].m_key >> pass_shift) & 0xFF]++] = pCur_syms[i];
-			sym_freq* t = pCur_syms; pCur_syms = pNew_syms; pNew_syms = t;
-		}
-		return pCur_syms;
-	}
-
-	// calculate_minimum_redundancy() originally written by: Alistair Moffat, [email protected], Jyrki Katajainen, [email protected], November 1996.
-	static void calculate_minimum_redundancy(sym_freq* A, int n)
-	{
-		int root, leaf, next, avbl, used, dpth;
-		if (n == 0) return; else if (n == 1) { A[0].m_key = 1; return; }
-		A[0].m_key += A[1].m_key; root = 0; leaf = 2;
-		for (next = 1; next < n - 1; next++)
-		{
-			if (leaf >= n || A[root].m_key < A[leaf].m_key) { A[next].m_key = A[root].m_key; A[root++].m_key = next; }
-			else A[next].m_key = A[leaf++].m_key;
-			if (leaf >= n || (root < next && A[root].m_key < A[leaf].m_key)) { A[next].m_key += A[root].m_key; A[root++].m_key = next; }
-			else A[next].m_key += A[leaf++].m_key;
-		}
-		A[n - 2].m_key = 0;
-		for (next = n - 3; next >= 0; next--) A[next].m_key = A[A[next].m_key].m_key + 1;
-		avbl = 1; used = dpth = 0; root = n - 2; next = n - 1;
-		while (avbl > 0)
-		{
-			while (root >= 0 && (int)A[root].m_key == dpth) { used++; root--; }
-			while (avbl > used) { A[next--].m_key = dpth; avbl--; }
-			avbl = 2 * used; dpth++; used = 0;
-		}
-	}
-
-	// Limits canonical Huffman code table's max code size to max_code_size.
-	static void huffman_enforce_max_code_size(int* pNum_codes, int code_list_len, int max_code_size)
-	{
-		if (code_list_len <= 1) return;
-
-		for (int i = max_code_size + 1; i <= MAX_HUFF_CODESIZE; i++) pNum_codes[max_code_size] += pNum_codes[i];
-
-		uint32 total = 0;
-		for (int i = max_code_size; i > 0; i--)
-			total += (((uint32)pNum_codes[i]) << (max_code_size - i));
-
-		while (total != (1UL << max_code_size))
-		{
-			pNum_codes[max_code_size]--;
-			for (int i = max_code_size - 1; i > 0; i--)
-			{
-				if (pNum_codes[i]) { pNum_codes[i]--; pNum_codes[i + 1] += 2; break; }
-			}
-			total--;
-		}
-	}
-
-	// Generates an optimized offman table.
-	void jpeg_encoder::optimize_huffman_table(int table_num, int table_len)
-	{
-		sym_freq syms0[MAX_HUFF_SYMBOLS], syms1[MAX_HUFF_SYMBOLS];
-		syms0[0].m_key = 1; syms0[0].m_sym_index = 0;  // dummy symbol, assures that no valid code contains all 1's
-		int num_used_syms = 1;
-		const uint32* pSym_count = &m_huff_count[table_num][0];
-		for (int i = 0; i < table_len; i++)
-			if (pSym_count[i]) { syms0[num_used_syms].m_key = pSym_count[i]; syms0[num_used_syms++].m_sym_index = i + 1; }
-		sym_freq* pSyms = radix_sort_syms(num_used_syms, syms0, syms1);
-		calculate_minimum_redundancy(pSyms, num_used_syms);
-
-		// Count the # of symbols of each code size.
-		int num_codes[1 + MAX_HUFF_CODESIZE]; clear_obj(num_codes);
-		for (int i = 0; i < num_used_syms; i++)
-			num_codes[pSyms[i].m_key]++;
-
-		const uint JPGE_CODE_SIZE_LIMIT = 16; // the maximum possible size of a JPEG Huffman code (valid range is [9,16] - 9 vs. 8 because of the dummy symbol)
-		huffman_enforce_max_code_size(num_codes, num_used_syms, JPGE_CODE_SIZE_LIMIT);
-
-		// Compute m_huff_bits array, which contains the # of symbols per code size.
-		clear_obj(m_huff_bits[table_num]);
-		for (int i = 1; i <= (int)JPGE_CODE_SIZE_LIMIT; i++)
-			m_huff_bits[table_num][i] = static_cast<uint8>(num_codes[i]);
-
-		// Remove the dummy symbol added above, which must be in largest bucket.
-		for (int i = JPGE_CODE_SIZE_LIMIT; i >= 1; i--)
-		{
-			if (m_huff_bits[table_num][i]) { m_huff_bits[table_num][i]--; break; }
-		}
-
-		// Compute the m_huff_val array, which contains the symbol indices sorted by code size (smallest to largest).
-		for (int i = num_used_syms - 1; i >= 1; i--)
-			m_huff_val[table_num][num_used_syms - 1 - i] = static_cast<uint8>(pSyms[i].m_sym_index - 1);
-	}
-
-	// JPEG marker generation.
-	void jpeg_encoder::emit_byte(uint8 i)
-	{
-		m_all_stream_writes_succeeded = m_all_stream_writes_succeeded && m_pStream->put_obj(i);
-	}
-
-	void jpeg_encoder::emit_word(uint i)
-	{
-		emit_byte(uint8(i >> 8)); emit_byte(uint8(i & 0xFF));
-	}
-
-	void jpeg_encoder::emit_marker(int marker)
-	{
-		emit_byte(uint8(0xFF)); emit_byte(uint8(marker));
-	}
-
-	// Emit JFIF marker
-	void jpeg_encoder::emit_jfif_app0()
-	{
-		emit_marker(M_APP0);
-		emit_word(2 + 4 + 1 + 2 + 1 + 2 + 2 + 1 + 1);
-		emit_byte(0x4A); emit_byte(0x46); emit_byte(0x49); emit_byte(0x46); /* Identifier: ASCII "JFIF" */
-		emit_byte(0);
-		emit_byte(1);      /* Major version */
-		emit_byte(1);      /* Minor version */
-		emit_byte(0);      /* Density unit */
-		emit_word(1);
-		emit_word(1);
-		emit_byte(0);      /* No thumbnail image */
-		emit_byte(0);
-	}
-
-	// Emit quantization tables
-	void jpeg_encoder::emit_dqt()
-	{
-		for (int i = 0; i < ((m_num_components == 3) ? 2 : 1); i++)
-		{
-			emit_marker(M_DQT);
-			emit_word(64 + 1 + 2);
-			emit_byte(static_cast<uint8>(i));
-			for (int j = 0; j < 64; j++)
-				emit_byte(static_cast<uint8>(m_quantization_tables[i][j]));
-		}
-	}
-
-	// Emit start of frame marker
-	void jpeg_encoder::emit_sof()
-	{
-		emit_marker(M_SOF0);                           /* baseline */
-		emit_word(3 * m_num_components + 2 + 5 + 1);
-		emit_byte(8);                                  /* precision */
-		emit_word(m_image_y);
-		emit_word(m_image_x);
-		emit_byte(m_num_components);
-		for (int i = 0; i < m_num_components; i++)
-		{
-			emit_byte(static_cast<uint8>(i + 1));                                   /* component ID     */
-			emit_byte((m_comp_h_samp[i] << 4) + m_comp_v_samp[i]);  /* h and v sampling */
-			emit_byte(i > 0);                                   /* quant. table num */
-		}
-	}
-
-	// Emit Huffman table.
-	void jpeg_encoder::emit_dht(uint8* bits, uint8* val, int index, bool ac_flag)
-	{
-		emit_marker(M_DHT);
-
-		int length = 0;
-		for (int i = 1; i <= 16; i++)
-			length += bits[i];
-
-		emit_word(length + 2 + 1 + 16);
-		emit_byte(static_cast<uint8>(index + (ac_flag << 4)));
-
-		for (int i = 1; i <= 16; i++)
-			emit_byte(bits[i]);
-
-		for (int i = 0; i < length; i++)
-			emit_byte(val[i]);
-	}
-
-	// Emit all Huffman tables.
-	void jpeg_encoder::emit_dhts()
-	{
-		emit_dht(m_huff_bits[0 + 0], m_huff_val[0 + 0], 0, false);
-		emit_dht(m_huff_bits[2 + 0], m_huff_val[2 + 0], 0, true);
-		if (m_num_components == 3)
-		{
-			emit_dht(m_huff_bits[0 + 1], m_huff_val[0 + 1], 1, false);
-			emit_dht(m_huff_bits[2 + 1], m_huff_val[2 + 1], 1, true);
-		}
-	}
-
-	// emit start of scan
-	void jpeg_encoder::emit_sos()
-	{
-		emit_marker(M_SOS);
-		emit_word(2 * m_num_components + 2 + 1 + 3);
-		emit_byte(m_num_components);
-		for (int i = 0; i < m_num_components; i++)
-		{
-			emit_byte(static_cast<uint8>(i + 1));
-			if (i == 0)
-				emit_byte((0 << 4) + 0);
-			else
-				emit_byte((1 << 4) + 1);
-		}
-		emit_byte(0);     /* spectral selection */
-		emit_byte(63);
-		emit_byte(0);
-	}
-
-	// Emit all markers at beginning of image file.
-	void jpeg_encoder::emit_markers()
-	{
-		emit_marker(M_SOI);
-		emit_jfif_app0();
-		emit_dqt();
-		emit_sof();
-		emit_dhts();
-		emit_sos();
-	}
-
-	// Compute the actual canonical Huffman codes/code sizes given the JPEG huff bits and val arrays.
-	void jpeg_encoder::compute_huffman_table(uint* codes, uint8* code_sizes, uint8* bits, uint8* val)
-	{
-		int i, l, last_p, si;
-		uint8 huff_size[257];
-		uint huff_code[257];
-		uint code;
-
-		int p = 0;
-		for (l = 1; l <= 16; l++)
-			for (i = 1; i <= bits[l]; i++)
-				huff_size[p++] = (char)l;
-
-		huff_size[p] = 0; last_p = p; // write sentinel
-
-		code = 0; si = huff_size[0]; p = 0;
-
-		while (huff_size[p])
-		{
-			while (huff_size[p] == si)
-				huff_code[p++] = code++;
-			code <<= 1;
-			si++;
-		}
-
-		memset(codes, 0, sizeof(codes[0]) * 256);
-		memset(code_sizes, 0, sizeof(code_sizes[0]) * 256);
-		for (p = 0; p < last_p; p++)
-		{
-			codes[val[p]] = huff_code[p];
-			code_sizes[val[p]] = huff_size[p];
-		}
-	}
-
-	// Quantization table generation.
-	void jpeg_encoder::compute_quant_table(int32* pDst, int16* pSrc)
-	{
-		int32 q;
-		if (m_params.m_quality < 50)
-			q = 5000 / m_params.m_quality;
-		else
-			q = 200 - m_params.m_quality * 2;
-		for (int i = 0; i < 64; i++)
-		{
-			int32 j = *pSrc++; j = (j * q + 50L) / 100L;
-			*pDst++ = JPGE_MIN(JPGE_MAX(j, 1), 255);
-		}
-	}
-
-	// Higher-level methods.
-	void jpeg_encoder::first_pass_init()
-	{
-		m_bit_buffer = 0; m_bits_in = 0;
-		memset(m_last_dc_val, 0, 3 * sizeof(m_last_dc_val[0]));
-		m_mcu_y_ofs = 0;
-		m_pass_num = 1;
-	}
-
-	bool jpeg_encoder::second_pass_init()
-	{
-		compute_huffman_table(&m_huff_codes[0 + 0][0], &m_huff_code_sizes[0 + 0][0], m_huff_bits[0 + 0], m_huff_val[0 + 0]);
-		compute_huffman_table(&m_huff_codes[2 + 0][0], &m_huff_code_sizes[2 + 0][0], m_huff_bits[2 + 0], m_huff_val[2 + 0]);
-		if (m_num_components > 1)
-		{
-			compute_huffman_table(&m_huff_codes[0 + 1][0], &m_huff_code_sizes[0 + 1][0], m_huff_bits[0 + 1], m_huff_val[0 + 1]);
-			compute_huffman_table(&m_huff_codes[2 + 1][0], &m_huff_code_sizes[2 + 1][0], m_huff_bits[2 + 1], m_huff_val[2 + 1]);
-		}
-		first_pass_init();
-		emit_markers();
-		m_pass_num = 2;
-		return true;
-	}
-
-	bool jpeg_encoder::jpg_open(int p_x_res, int p_y_res, int src_channels)
-	{
-		m_num_components = 3;
-		switch (m_params.m_subsampling)
-		{
-		case Y_ONLY:
-		{
-			m_num_components = 1;
-			m_comp_h_samp[0] = 1; m_comp_v_samp[0] = 1;
-			m_mcu_x = 8; m_mcu_y = 8;
-			break;
-		}
-		case H1V1:
-		{
-			m_comp_h_samp[0] = 1; m_comp_v_samp[0] = 1;
-			m_comp_h_samp[1] = 1; m_comp_v_samp[1] = 1;
-			m_comp_h_samp[2] = 1; m_comp_v_samp[2] = 1;
-			m_mcu_x = 8; m_mcu_y = 8;
-			break;
-		}
-		case H2V1:
-		{
-			m_comp_h_samp[0] = 2; m_comp_v_samp[0] = 1;
-			m_comp_h_samp[1] = 1; m_comp_v_samp[1] = 1;
-			m_comp_h_samp[2] = 1; m_comp_v_samp[2] = 1;
-			m_mcu_x = 16; m_mcu_y = 8;
-			break;
-		}
-		case H2V2:
-		{
-			m_comp_h_samp[0] = 2; m_comp_v_samp[0] = 2;
-			m_comp_h_samp[1] = 1; m_comp_v_samp[1] = 1;
-			m_comp_h_samp[2] = 1; m_comp_v_samp[2] = 1;
-			m_mcu_x = 16; m_mcu_y = 16;
-		}
-		}
-
-		m_image_x = p_x_res; m_image_y = p_y_res;
-		m_image_bpp = src_channels;
-		m_image_bpl = m_image_x * src_channels;
-		m_image_x_mcu = (m_image_x + m_mcu_x - 1) & (~(m_mcu_x - 1));
-		m_image_y_mcu = (m_image_y + m_mcu_y - 1) & (~(m_mcu_y - 1));
-		m_image_bpl_xlt = m_image_x * m_num_components;
-		m_image_bpl_mcu = m_image_x_mcu * m_num_components;
-		m_mcus_per_row = m_image_x_mcu / m_mcu_x;
-
-		if ((m_mcu_lines[0] = static_cast<uint8*>(jpge_malloc(m_image_bpl_mcu * m_mcu_y))) == NULL) return false;
-		for (int i = 1; i < m_mcu_y; i++)
-			m_mcu_lines[i] = m_mcu_lines[i - 1] + m_image_bpl_mcu;
-
-		if (m_params.m_use_std_tables)
-		{
-			compute_quant_table(m_quantization_tables[0], s_std_lum_quant);
-			compute_quant_table(m_quantization_tables[1], m_params.m_no_chroma_discrim_flag ? s_std_lum_quant : s_std_croma_quant);
-		}
-		else
-		{
-			compute_quant_table(m_quantization_tables[0], s_alt_quant);
-			memcpy(m_quantization_tables[1], m_quantization_tables[0], sizeof(m_quantization_tables[1]));
-		}
-
-		m_out_buf_left = JPGE_OUT_BUF_SIZE;
-		m_pOut_buf = m_out_buf;
-
-		if (m_params.m_two_pass_flag)
-		{
-			clear_obj(m_huff_count);
-			first_pass_init();
-		}
-		else
-		{
-			memcpy(m_huff_bits[0 + 0], s_dc_lum_bits, 17);    memcpy(m_huff_val[0 + 0], s_dc_lum_val, DC_LUM_CODES);
-			memcpy(m_huff_bits[2 + 0], s_ac_lum_bits, 17);    memcpy(m_huff_val[2 + 0], s_ac_lum_val, AC_LUM_CODES);
-			memcpy(m_huff_bits[0 + 1], s_dc_chroma_bits, 17); memcpy(m_huff_val[0 + 1], s_dc_chroma_val, DC_CHROMA_CODES);
-			memcpy(m_huff_bits[2 + 1], s_ac_chroma_bits, 17); memcpy(m_huff_val[2 + 1], s_ac_chroma_val, AC_CHROMA_CODES);
-			if (!second_pass_init()) return false;   // in effect, skip over the first pass
-		}
-		return m_all_stream_writes_succeeded;
-	}
-
-	void jpeg_encoder::load_block_8_8_grey(int x)
-	{
-		uint8* pSrc;
-		sample_array_t* pDst = m_sample_array;
-		x <<= 3;
-		for (int i = 0; i < 8; i++, pDst += 8)
-		{
-			pSrc = m_mcu_lines[i] + x;
-			pDst[0] = pSrc[0] - 128; pDst[1] = pSrc[1] - 128; pDst[2] = pSrc[2] - 128; pDst[3] = pSrc[3] - 128;
-			pDst[4] = pSrc[4] - 128; pDst[5] = pSrc[5] - 128; pDst[6] = pSrc[6] - 128; pDst[7] = pSrc[7] - 128;
-		}
-	}
-
-	void jpeg_encoder::load_block_8_8(int x, int y, int c)
-	{
-		uint8* pSrc;
-		sample_array_t* pDst = m_sample_array;
-		x = (x * (8 * 3)) + c;
-		y <<= 3;
-		for (int i = 0; i < 8; i++, pDst += 8)
-		{
-			pSrc = m_mcu_lines[y + i] + x;
-			pDst[0] = pSrc[0 * 3] - 128; pDst[1] = pSrc[1 * 3] - 128; pDst[2] = pSrc[2 * 3] - 128; pDst[3] = pSrc[3 * 3] - 128;
-			pDst[4] = pSrc[4 * 3] - 128; pDst[5] = pSrc[5 * 3] - 128; pDst[6] = pSrc[6 * 3] - 128; pDst[7] = pSrc[7 * 3] - 128;
-		}
-	}
-
-	void jpeg_encoder::load_block_16_8(int x, int c)
-	{
-		uint8* pSrc1, * pSrc2;
-		sample_array_t* pDst = m_sample_array;
-		x = (x * (16 * 3)) + c;
-		for (int i = 0; i < 16; i += 2, pDst += 8)
-		{
-			pSrc1 = m_mcu_lines[i + 0] + x;
-			pSrc2 = m_mcu_lines[i + 1] + x;
-			pDst[0] = ((pSrc1[0 * 3] + pSrc1[1 * 3] + pSrc2[0 * 3] + pSrc2[1 * 3] + 2) >> 2) - 128; pDst[1] = ((pSrc1[2 * 3] + pSrc1[3 * 3] + pSrc2[2 * 3] + pSrc2[3 * 3] + 2) >> 2) - 128;
-			pDst[2] = ((pSrc1[4 * 3] + pSrc1[5 * 3] + pSrc2[4 * 3] + pSrc2[5 * 3] + 2) >> 2) - 128; pDst[3] = ((pSrc1[6 * 3] + pSrc1[7 * 3] + pSrc2[6 * 3] + pSrc2[7 * 3] + 2) >> 2) - 128;
-			pDst[4] = ((pSrc1[8 * 3] + pSrc1[9 * 3] + pSrc2[8 * 3] + pSrc2[9 * 3] + 2) >> 2) - 128; pDst[5] = ((pSrc1[10 * 3] + pSrc1[11 * 3] + pSrc2[10 * 3] + pSrc2[11 * 3] + 2) >> 2) - 128;
-			pDst[6] = ((pSrc1[12 * 3] + pSrc1[13 * 3] + pSrc2[12 * 3] + pSrc2[13 * 3] + 2) >> 2) - 128; pDst[7] = ((pSrc1[14 * 3] + pSrc1[15 * 3] + pSrc2[14 * 3] + pSrc2[15 * 3] + 2) >> 2) - 128;
-		}
-	}
-
-	void jpeg_encoder::load_block_16_8_8(int x, int c)
-	{
-		uint8* pSrc1;
-		sample_array_t* pDst = m_sample_array;
-		x = (x * (16 * 3)) + c;
-		for (int i = 0; i < 8; i++, pDst += 8)
-		{
-			pSrc1 = m_mcu_lines[i + 0] + x;
-			pDst[0] = ((pSrc1[0 * 3] + pSrc1[1 * 3] + 1) >> 1) - 128; pDst[1] = ((pSrc1[2 * 3] + pSrc1[3 * 3] + 1) >> 1) - 128;
-			pDst[2] = ((pSrc1[4 * 3] + pSrc1[5 * 3] + 1) >> 1) - 128; pDst[3] = ((pSrc1[6 * 3] + pSrc1[7 * 3] + 1) >> 1) - 128;
-			pDst[4] = ((pSrc1[8 * 3] + pSrc1[9 * 3] + 1) >> 1) - 128; pDst[5] = ((pSrc1[10 * 3] + pSrc1[11 * 3] + 1) >> 1) - 128;
-			pDst[6] = ((pSrc1[12 * 3] + pSrc1[13 * 3] + 1) >> 1) - 128; pDst[7] = ((pSrc1[14 * 3] + pSrc1[15 * 3] + 1) >> 1) - 128;
-		}
-	}
-
-	void jpeg_encoder::load_quantized_coefficients(int component_num)
-	{
-		int32* q = m_quantization_tables[component_num > 0];
-		int16* pDst = m_coefficient_array;
-		for (int i = 0; i < 64; i++)
-		{
-			sample_array_t j = m_sample_array[s_zag[i]];
-			if (j < 0)
-			{
-				if ((j = -j + (*q >> 1)) < *q)
-					*pDst++ = 0;
-				else
-					*pDst++ = static_cast<int16>(-(j / *q));
-			}
-			else
-			{
-				if ((j = j + (*q >> 1)) < *q)
-					*pDst++ = 0;
-				else
-					*pDst++ = static_cast<int16>((j / *q));
-			}
-			q++;
-		}
-	}
-
-	void jpeg_encoder::flush_output_buffer()
-	{
-		if (m_out_buf_left != JPGE_OUT_BUF_SIZE)
-			m_all_stream_writes_succeeded = m_all_stream_writes_succeeded && m_pStream->put_buf(m_out_buf, JPGE_OUT_BUF_SIZE - m_out_buf_left);
-		m_pOut_buf = m_out_buf;
-		m_out_buf_left = JPGE_OUT_BUF_SIZE;
-	}
-
-	void jpeg_encoder::put_bits(uint bits, uint len)
-	{
-		m_bit_buffer |= ((uint32)bits << (24 - (m_bits_in += len)));
-		while (m_bits_in >= 8)
-		{
-			uint8 c;
-#define JPGE_PUT_BYTE(c) { *m_pOut_buf++ = (c); if (--m_out_buf_left == 0) flush_output_buffer(); }
-			JPGE_PUT_BYTE(c = (uint8)((m_bit_buffer >> 16) & 0xFF));
-			if (c == 0xFF) JPGE_PUT_BYTE(0);
-			m_bit_buffer <<= 8;
-			m_bits_in -= 8;
-		}
-	}
-
-	void jpeg_encoder::code_coefficients_pass_one(int component_num)
-	{
-		if (component_num >= 3) return; // just to shut up static analysis
-		int i, run_len, nbits, temp1;
-		int16* src = m_coefficient_array;
-		uint32* dc_count = component_num ? m_huff_count[0 + 1] : m_huff_count[0 + 0], * ac_count = component_num ? m_huff_count[2 + 1] : m_huff_count[2 + 0];
-
-		temp1 = src[0] - m_last_dc_val[component_num];
-		m_last_dc_val[component_num] = src[0];
-		if (temp1 < 0) temp1 = -temp1;
-
-		nbits = 0;
-		while (temp1)
-		{
-			nbits++; temp1 >>= 1;
-		}
-
-		dc_count[nbits]++;
-		for (run_len = 0, i = 1; i < 64; i++)
-		{
-			if ((temp1 = m_coefficient_array[i]) == 0)
-				run_len++;
-			else
-			{
-				while (run_len >= 16)
-				{
-					ac_count[0xF0]++;
-					run_len -= 16;
-				}
-				if (temp1 < 0) temp1 = -temp1;
-				nbits = 1;
-				while (temp1 >>= 1) nbits++;
-				ac_count[(run_len << 4) + nbits]++;
-				run_len = 0;
-			}
-		}
-		if (run_len) ac_count[0]++;
-	}
-
-	void jpeg_encoder::code_coefficients_pass_two(int component_num)
-	{
-		int i, j, run_len, nbits, temp1, temp2;
-		int16* pSrc = m_coefficient_array;
-		uint* codes[2];
-		uint8* code_sizes[2];
-
-		if (component_num == 0)
-		{
-			codes[0] = m_huff_codes[0 + 0]; codes[1] = m_huff_codes[2 + 0];
-			code_sizes[0] = m_huff_code_sizes[0 + 0]; code_sizes[1] = m_huff_code_sizes[2 + 0];
-		}
-		else
-		{
-			codes[0] = m_huff_codes[0 + 1]; codes[1] = m_huff_codes[2 + 1];
-			code_sizes[0] = m_huff_code_sizes[0 + 1]; code_sizes[1] = m_huff_code_sizes[2 + 1];
-		}
-
-		temp1 = temp2 = pSrc[0] - m_last_dc_val[component_num];
-		m_last_dc_val[component_num] = pSrc[0];
-
-		if (temp1 < 0)
-		{
-			temp1 = -temp1; temp2--;
-		}
-
-		nbits = 0;
-		while (temp1)
-		{
-			nbits++; temp1 >>= 1;
-		}
-
-		put_bits(codes[0][nbits], code_sizes[0][nbits]);
-		if (nbits) put_bits(temp2 & ((1 << nbits) - 1), nbits);
-
-		for (run_len = 0, i = 1; i < 64; i++)
-		{
-			if ((temp1 = m_coefficient_array[i]) == 0)
-				run_len++;
-			else
-			{
-				while (run_len >= 16)
-				{
-					put_bits(codes[1][0xF0], code_sizes[1][0xF0]);
-					run_len -= 16;
-				}
-				if ((temp2 = temp1) < 0)
-				{
-					temp1 = -temp1;
-					temp2--;
-				}
-				nbits = 1;
-				while (temp1 >>= 1)
-					nbits++;
-				j = (run_len << 4) + nbits;
-				put_bits(codes[1][j], code_sizes[1][j]);
-				put_bits(temp2 & ((1 << nbits) - 1), nbits);
-				run_len = 0;
-			}
-		}
-		if (run_len)
-			put_bits(codes[1][0], code_sizes[1][0]);
-	}
-
-	void jpeg_encoder::code_block(int component_num)
-	{
-		DCT2D(m_sample_array);
-		load_quantized_coefficients(component_num);
-		if (m_pass_num == 1)
-			code_coefficients_pass_one(component_num);
-		else
-			code_coefficients_pass_two(component_num);
-	}
-
-	void jpeg_encoder::process_mcu_row()
-	{
-		if (m_num_components == 1)
-		{
-			for (int i = 0; i < m_mcus_per_row; i++)
-			{
-				load_block_8_8_grey(i); code_block(0);
-			}
-		}
-		else if ((m_comp_h_samp[0] == 1) && (m_comp_v_samp[0] == 1))
-		{
-			for (int i = 0; i < m_mcus_per_row; i++)
-			{
-				load_block_8_8(i, 0, 0); code_block(0); load_block_8_8(i, 0, 1); code_block(1); load_block_8_8(i, 0, 2); code_block(2);
-			}
-		}
-		else if ((m_comp_h_samp[0] == 2) && (m_comp_v_samp[0] == 1))
-		{
-			for (int i = 0; i < m_mcus_per_row; i++)
-			{
-				load_block_8_8(i * 2 + 0, 0, 0); code_block(0); load_block_8_8(i * 2 + 1, 0, 0); code_block(0);
-				load_block_16_8_8(i, 1); code_block(1); load_block_16_8_8(i, 2); code_block(2);
-			}
-		}
-		else if ((m_comp_h_samp[0] == 2) && (m_comp_v_samp[0] == 2))
-		{
-			for (int i = 0; i < m_mcus_per_row; i++)
-			{
-				load_block_8_8(i * 2 + 0, 0, 0); code_block(0); load_block_8_8(i * 2 + 1, 0, 0); code_block(0);
-				load_block_8_8(i * 2 + 0, 1, 0); code_block(0); load_block_8_8(i * 2 + 1, 1, 0); code_block(0);
-				load_block_16_8(i, 1); code_block(1); load_block_16_8(i, 2); code_block(2);
-			}
-		}
-	}
-
-	bool jpeg_encoder::terminate_pass_one()
-	{
-		optimize_huffman_table(0 + 0, DC_LUM_CODES); optimize_huffman_table(2 + 0, AC_LUM_CODES);
-		if (m_num_components > 1)
-		{
-			optimize_huffman_table(0 + 1, DC_CHROMA_CODES); optimize_huffman_table(2 + 1, AC_CHROMA_CODES);
-		}
-		return second_pass_init();
-	}
-
-	bool jpeg_encoder::terminate_pass_two()
-	{
-		put_bits(0x7F, 7);
-		flush_output_buffer();
-		emit_marker(M_EOI);
-		m_pass_num++; // purposely bump up m_pass_num, for debugging
-		return true;
-	}
-
-	bool jpeg_encoder::process_end_of_image()
-	{
-		if (m_mcu_y_ofs)
-		{
-			if (m_mcu_y_ofs < 16) // check here just to shut up static analysis
-			{
-				for (int i = m_mcu_y_ofs; i < m_mcu_y; i++)
-					memcpy(m_mcu_lines[i], m_mcu_lines[m_mcu_y_ofs - 1], m_image_bpl_mcu);
-			}
-
-			process_mcu_row();
-		}
-
-		if (m_pass_num == 1)
-			return terminate_pass_one();
-		else
-			return terminate_pass_two();
-	}
-
-	void jpeg_encoder::load_mcu(const void* pSrc)
-	{
-		const uint8* Psrc = reinterpret_cast<const uint8*>(pSrc);
-
-		uint8* pDst = m_mcu_lines[m_mcu_y_ofs]; // OK to write up to m_image_bpl_xlt bytes to pDst
-
-		if (m_num_components == 1)
-		{
-			if (m_image_bpp == 4)
-				RGBA_to_Y(pDst, Psrc, m_image_x);
-			else if (m_image_bpp == 3)
-				RGB_to_Y(pDst, Psrc, m_image_x);
-			else
-				memcpy(pDst, Psrc, m_image_x);
-		}
-		else
-		{
-			if (m_image_bpp == 4)
-				RGBA_to_YCC(pDst, Psrc, m_image_x);
-			else if (m_image_bpp == 3)
-				RGB_to_YCC(pDst, Psrc, m_image_x);
-			else
-				Y_to_YCC(pDst, Psrc, m_image_x);
-		}
-
-		// Possibly duplicate pixels at end of scanline if not a multiple of 8 or 16
-		if (m_num_components == 1)
-			memset(m_mcu_lines[m_mcu_y_ofs] + m_image_bpl_xlt, pDst[m_image_bpl_xlt - 1], m_image_x_mcu - m_image_x);
-		else
-		{
-			const uint8 y = pDst[m_image_bpl_xlt - 3 + 0], cb = pDst[m_image_bpl_xlt - 3 + 1], cr = pDst[m_image_bpl_xlt - 3 + 2];
-			uint8* q = m_mcu_lines[m_mcu_y_ofs] + m_image_bpl_xlt;
-			for (int i = m_image_x; i < m_image_x_mcu; i++)
-			{
-				*q++ = y; *q++ = cb; *q++ = cr;
-			}
-		}
-
-		if (++m_mcu_y_ofs == m_mcu_y)
-		{
-			process_mcu_row();
-			m_mcu_y_ofs = 0;
-		}
-	}
-
-	void jpeg_encoder::clear()
-	{
-		m_mcu_lines[0] = NULL;
-		m_pass_num = 0;
-		m_all_stream_writes_succeeded = true;
-	}
-
-	jpeg_encoder::jpeg_encoder()
-	{
-		clear();
-	}
-
-	jpeg_encoder::~jpeg_encoder()
-	{
-		deinit();
-	}
-
-	bool jpeg_encoder::init(output_stream* pStream, int width, int height, int src_channels, const params& comp_params)
-	{
-		deinit();
-		if (((!pStream) || (width < 1) || (height < 1)) || ((src_channels != 1) && (src_channels != 3) && (src_channels != 4)) || (!comp_params.check())) return false;
-		m_pStream = pStream;
-		m_params = comp_params;
-		return jpg_open(width, height, src_channels);
-	}
-
-	void jpeg_encoder::deinit()
-	{
-		jpge_free(m_mcu_lines[0]);
-		clear();
-	}
-
-	bool jpeg_encoder::process_scanline(const void* pScanline)
-	{
-		if ((m_pass_num < 1) || (m_pass_num > 2)) return false;
-		if (m_all_stream_writes_succeeded)
-		{
-			if (!pScanline)
-			{
-				if (!process_end_of_image()) return false;
-			}
-			else
-			{
-				load_mcu(pScanline);
-			}
-		}
-		return m_all_stream_writes_succeeded;
-	}
-
-	// Higher level wrappers/examples (optional).
-
-	class cfile_stream : public output_stream
-	{
-		cfile_stream(const cfile_stream&);
-		cfile_stream& operator= (const cfile_stream&);
-
-		FILE* m_pFile;
-		bool m_bStatus;
-
-	public:
-		cfile_stream() : m_pFile(NULL), m_bStatus(false) { }
-
-		virtual ~cfile_stream()
-		{
-			close();
-		}
-
-		bool open(const char* pFilename)
-		{
-			close();
-			m_pFile = fopen(pFilename, "wb");
-			m_bStatus = (m_pFile != NULL);
-			return m_bStatus;
-		}
-
-		bool close()
-		{
-			if (m_pFile)
-			{
-				if (fclose(m_pFile) == EOF)
-				{
-					m_bStatus = false;
-				}
-				m_pFile = NULL;
-			}
-			return m_bStatus;
-		}
-
-		virtual bool put_buf(const void* pBuf, int len)
-		{
-			m_bStatus = m_bStatus && (fwrite(pBuf, len, 1, m_pFile) == 1);
-			return m_bStatus;
-		}
-
-		uint get_size() const
-		{
-			return m_pFile ? ftell(m_pFile) : 0;
-		}
-	};
-
-	// Writes JPEG image to file.
-	bool compress_image_to_jpeg_file(const char* pFilename, int width, int height, int num_channels, const uint8* pImage_data, const params& comp_params)
-	{
-		cfile_stream dst_stream;
-		if (!dst_stream.open(pFilename))
-			return false;
-
-		jpge::jpeg_encoder dst_image;
-		if (!dst_image.init(&dst_stream, width, height, num_channels, comp_params))
-			return false;
-
-		for (uint pass_index = 0; pass_index < dst_image.get_total_passes(); pass_index++)
-		{
-			for (int i = 0; i < height; i++)
-			{
-				const uint8* pBuf = pImage_data + i * width * num_channels;
-				if (!dst_image.process_scanline(pBuf))
-					return false;
-			}
-			if (!dst_image.process_scanline(NULL))
-				return false;
-		}
-
-		dst_image.deinit();
-
-		return dst_stream.close();
-	}
-
-	class memory_stream : public output_stream
-	{
-		memory_stream(const memory_stream&);
-		memory_stream& operator= (const memory_stream&);
-
-		uint8* m_pBuf;
-		uint m_buf_size, m_buf_ofs;
-
-	public:
-		memory_stream(void* pBuf, uint buf_size) : m_pBuf(static_cast<uint8*>(pBuf)), m_buf_size(buf_size), m_buf_ofs(0) { }
-
-		virtual ~memory_stream() { }
-
-		virtual bool put_buf(const void* pBuf, int len)
-		{
-			uint buf_remaining = m_buf_size - m_buf_ofs;
-			if ((uint)len > buf_remaining)
-				return false;
-			memcpy(m_pBuf + m_buf_ofs, pBuf, len);
-			m_buf_ofs += len;
-			return true;
-		}
-
-		uint get_size() const
-		{
-			return m_buf_ofs;
-		}
-	};
-
-	bool compress_image_to_jpeg_file_in_memory(void* pDstBuf, int& buf_size, int width, int height, int num_channels, const uint8* pImage_data, const params& comp_params)
-	{
-		if ((!pDstBuf) || (!buf_size))
-			return false;
-
-		memory_stream dst_stream(pDstBuf, buf_size);
-
-		buf_size = 0;
-
-		jpge::jpeg_encoder dst_image;
-		if (!dst_image.init(&dst_stream, width, height, num_channels, comp_params))
-			return false;
-
-		for (uint pass_index = 0; pass_index < dst_image.get_total_passes(); pass_index++)
-		{
-			for (int i = 0; i < height; i++)
-			{
-				const uint8* pScanline = pImage_data + i * width * num_channels;
-				if (!dst_image.process_scanline(pScanline))
-					return false;
-			}
-			if (!dst_image.process_scanline(NULL))
-				return false;
-		}
-
-		dst_image.deinit();
-
-		buf_size = dst_stream.get_size();
-		return true;
-	}
-
-} // namespace jpge
-

thirdparty/jpeg-compressor/jpge.h

@@ -1,174 +0,0 @@
-// jpge.h - C++ class for JPEG compression.
-// Public Domain or Apache 2.0, Richard Geldreich <[email protected]>
-// Alex Evans: Added RGBA support, linear memory allocator.
-#ifndef JPEG_ENCODER_H
-#define JPEG_ENCODER_H
-
-namespace jpge
-{
-	typedef unsigned char  uint8;
-	typedef signed short   int16;
-	typedef signed int     int32;
-	typedef unsigned short uint16;
-	typedef unsigned int   uint32;
-	typedef unsigned int   uint;
-
-	// JPEG chroma subsampling factors. Y_ONLY (grayscale images) and H2V2 (color images) are the most common.
-	enum subsampling_t { Y_ONLY = 0, H1V1 = 1, H2V1 = 2, H2V2 = 3 };
-
-	// JPEG compression parameters structure.
-	struct params
-	{
-		inline params() : m_quality(85), m_subsampling(H2V2), m_no_chroma_discrim_flag(false), m_two_pass_flag(false), m_use_std_tables(false) { }
-
-		inline bool check() const
-		{
-			if ((m_quality < 1) || (m_quality > 100)) return false;
-			if ((uint)m_subsampling > (uint)H2V2) return false;
-			return true;
-		}
-
-		// Quality: 1-100, higher is better. Typical values are around 50-95.
-		int m_quality;
-
-		// m_subsampling:
-		// 0 = Y (grayscale) only
-		// 1 = YCbCr, no subsampling (H1V1, YCbCr 1x1x1, 3 blocks per MCU)
-		// 2 = YCbCr, H2V1 subsampling (YCbCr 2x1x1, 4 blocks per MCU)
-		// 3 = YCbCr, H2V2 subsampling (YCbCr 4x1x1, 6 blocks per MCU-- very common)
-		subsampling_t m_subsampling;
-
-		// Disables CbCr discrimination - only intended for testing.
-		// If true, the Y quantization table is also used for the CbCr channels.
-		bool m_no_chroma_discrim_flag;
-
-		bool m_two_pass_flag;
-
-		// By default we use the same quantization tables as mozjpeg's default. 
-		// Set to true to use the traditional tables from JPEG Annex K.
-		bool m_use_std_tables;
-	};
-
-	// Writes JPEG image to a file. 
-	// num_channels must be 1 (Y) or 3 (RGB), image pitch must be width*num_channels.
-	bool compress_image_to_jpeg_file(const char* pFilename, int width, int height, int num_channels, const uint8* pImage_data, const params& comp_params = params());
-
-	// Writes JPEG image to memory buffer. 
-	// On entry, buf_size is the size of the output buffer pointed at by pBuf, which should be at least ~1024 bytes. 
-	// If return value is true, buf_size will be set to the size of the compressed data.
-	bool compress_image_to_jpeg_file_in_memory(void* pBuf, int& buf_size, int width, int height, int num_channels, const uint8* pImage_data, const params& comp_params = params());
-
-	// Output stream abstract class - used by the jpeg_encoder class to write to the output stream. 
-	// put_buf() is generally called with len==JPGE_OUT_BUF_SIZE bytes, but for headers it'll be called with smaller amounts.
-	class output_stream
-	{
-	public:
-		virtual ~output_stream() { };
-		virtual bool put_buf(const void* Pbuf, int len) = 0;
-		template<class T> inline bool put_obj(const T& obj) { return put_buf(&obj, sizeof(T)); }
-	};
-
-	// Lower level jpeg_encoder class - useful if more control is needed than the above helper functions.
-	class jpeg_encoder
-	{
-	public:
-		jpeg_encoder();
-		~jpeg_encoder();
-
-		// Initializes the compressor.
-		// pStream: The stream object to use for writing compressed data.
-		// params - Compression parameters structure, defined above.
-		// width, height  - Image dimensions.
-		// channels - May be 1, or 3. 1 indicates grayscale, 3 indicates RGB source data.
-		// Returns false on out of memory or if a stream write fails.
-		bool init(output_stream* pStream, int width, int height, int src_channels, const params& comp_params = params());
-
-		const params& get_params() const { return m_params; }
-
-		// Deinitializes the compressor, freeing any allocated memory. May be called at any time.
-		void deinit();
-
-		uint get_total_passes() const { return m_params.m_two_pass_flag ? 2 : 1; }
-		inline uint get_cur_pass() { return m_pass_num; }
-
-		// Call this method with each source scanline.
-		// width * src_channels bytes per scanline is expected (RGB or Y format).
-		// You must call with NULL after all scanlines are processed to finish compression.
-		// Returns false on out of memory or if a stream write fails.
-		bool process_scanline(const void* pScanline);
-
-	private:
-		jpeg_encoder(const jpeg_encoder&);
-		jpeg_encoder& operator =(const jpeg_encoder&);
-
-		typedef int32 sample_array_t;
-
-		output_stream* m_pStream;
-		params m_params;
-		uint8 m_num_components;
-		uint8 m_comp_h_samp[3], m_comp_v_samp[3];
-		int m_image_x, m_image_y, m_image_bpp, m_image_bpl;
-		int m_image_x_mcu, m_image_y_mcu;
-		int m_image_bpl_xlt, m_image_bpl_mcu;
-		int m_mcus_per_row;
-		int m_mcu_x, m_mcu_y;
-		uint8* m_mcu_lines[16];
-		uint8 m_mcu_y_ofs;
-		sample_array_t m_sample_array[64];
-		int16 m_coefficient_array[64];
-		int32 m_quantization_tables[2][64];
-		uint m_huff_codes[4][256];
-		uint8 m_huff_code_sizes[4][256];
-		uint8 m_huff_bits[4][17];
-		uint8 m_huff_val[4][256];
-		uint32 m_huff_count[4][256];
-		int m_last_dc_val[3];
-		enum { JPGE_OUT_BUF_SIZE = 2048 };
-		uint8 m_out_buf[JPGE_OUT_BUF_SIZE];
-		uint8* m_pOut_buf;
-		uint m_out_buf_left;
-		uint32 m_bit_buffer;
-		uint m_bits_in;
-		uint8 m_pass_num;
-		bool m_all_stream_writes_succeeded;
-
-		void optimize_huffman_table(int table_num, int table_len);
-		void emit_byte(uint8 i);
-		void emit_word(uint i);
-		void emit_marker(int marker);
-		void emit_jfif_app0();
-		void emit_dqt();
-		void emit_sof();
-		void emit_dht(uint8* bits, uint8* val, int index, bool ac_flag);
-		void emit_dhts();
-		void emit_sos();
-		void emit_markers();
-		void compute_huffman_table(uint* codes, uint8* code_sizes, uint8* bits, uint8* val);
-		void compute_quant_table(int32* dst, int16* src);
-		void adjust_quant_table(int32* dst, int32* src);
-		void first_pass_init();
-		bool second_pass_init();
-		bool jpg_open(int p_x_res, int p_y_res, int src_channels);
-		void load_block_8_8_grey(int x);
-		void load_block_8_8(int x, int y, int c);
-		void load_block_16_8(int x, int c);
-		void load_block_16_8_8(int x, int c);
-		void load_quantized_coefficients(int component_num);
-		void flush_output_buffer();
-		void put_bits(uint bits, uint len);
-		void code_coefficients_pass_one(int component_num);
-		void code_coefficients_pass_two(int component_num);
-		void code_block(int component_num);
-		void process_mcu_row();
-		bool terminate_pass_one();
-		bool terminate_pass_two();
-		bool process_end_of_image();
-		void load_mcu(const void* src);
-		void clear();
-		void init();
-	};
-
-} // namespace jpge
-
-#endif // JPEG_ENCODER
-

thirdparty/jpeg-compressor/patches/0001-clang-fortify-fix.patch

@@ -1,20 +0,0 @@
-diff --git a/thirdparty/jpeg-compressor/jpge.cpp b/thirdparty/jpeg-compressor/jpge.cpp
-index 5a36c19653..bb0c54bbf0 100644
---- a/thirdparty/jpeg-compressor/jpge.cpp
-+++ b/thirdparty/jpeg-compressor/jpge.cpp
-@@ -30,6 +30,7 @@
- 
- #include "jpge.h"
- 
-+#include <stdio.h>
- #include <stdlib.h>
- #include <string.h>
- 
-@@ -933,7 +934,6 @@ namespace jpge {
- 	}
- 
- 	// Higher level wrappers/examples (optional).
--#include <stdio.h>
- 
- 	class cfile_stream : public output_stream
- 	{

thirdparty/libjpeg-turbo/LICENSE.md

@@ -0,0 +1,135 @@
+libjpeg-turbo Licenses
+======================
+
+libjpeg-turbo is covered by two compatible BSD-style open source licenses:
+
+- The IJG (Independent JPEG Group) License, which is listed in
+  [README.ijg](README.ijg)
+
+  This license applies to the libjpeg API library and associated programs,
+  including any code inherited from libjpeg and any modifications to that
+  code.  Note that the libjpeg-turbo SIMD source code bears the
+  [zlib License](https://opensource.org/licenses/Zlib), but in the context of
+  the overall libjpeg API library, the terms of the zlib License are subsumed
+  by the terms of the IJG License.
+
+- The Modified (3-clause) BSD License, which is listed below
+
+  This license applies to the TurboJPEG API library and associated programs, as
+  well as the build system.  Note that the TurboJPEG API library wraps the
+  libjpeg API library, so in the context of the overall TurboJPEG API library,
+  both the terms of the IJG License and the terms of the Modified (3-clause)
+  BSD License apply.
+
+
+Complying with the libjpeg-turbo Licenses
+=========================================
+
+This section provides a roll-up of the libjpeg-turbo licensing terms, to the
+best of our understanding.  This is not a license in and of itself.  It is
+intended solely for clarification.
+
+1.  If you are distributing a modified version of the libjpeg-turbo source,
+    then:
+
+    1.  You cannot alter or remove any existing copyright or license notices
+        from the source.
+
+        **Origin**
+        - Clause 1 of the IJG License
+        - Clause 1 of the Modified BSD License
+        - Clauses 1 and 3 of the zlib License
+
+    2.  You must add your own copyright notice to the header of each source
+        file you modified, so others can tell that you modified that file.  (If
+        there is not an existing copyright header in that file, then you can
+        simply add a notice stating that you modified the file.)
+
+        **Origin**
+        - Clause 1 of the IJG License
+        - Clause 2 of the zlib License
+
+    3.  You must include the IJG README file, and you must not alter any of the
+        copyright or license text in that file.
+
+        **Origin**
+        - Clause 1 of the IJG License
+
+2.  If you are distributing only libjpeg-turbo binaries without the source, or
+    if you are distributing an application that statically links with
+    libjpeg-turbo, then:
+
+    1.  Your product documentation must include a message stating:
+
+        This software is based in part on the work of the Independent JPEG
+        Group.
+
+        **Origin**
+        - Clause 2 of the IJG license
+
+    2.  If your binary distribution includes or uses the TurboJPEG API, then
+        your product documentation must include the text of the Modified BSD
+        License (see below.)
+
+        **Origin**
+        - Clause 2 of the Modified BSD License
+
+3.  You cannot use the name of the IJG or The libjpeg-turbo Project or the
+    contributors thereof in advertising, publicity, etc.
+
+    **Origin**
+    - IJG License
+    - Clause 3 of the Modified BSD License
+
+4.  The IJG and The libjpeg-turbo Project do not warrant libjpeg-turbo to be
+    free of defects, nor do we accept any liability for undesirable
+    consequences resulting from your use of the software.
+
+    **Origin**
+    - IJG License
+    - Modified BSD License
+    - zlib License
+
+
+The Modified (3-clause) BSD License
+===================================
+
+Copyright (C)2009-2024 D. R. Commander.  All Rights Reserved.<br>
+Copyright (C)2015 Viktor Szathmáry.  All Rights Reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+- Redistributions of source code must retain the above copyright notice,
+  this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+- Neither the name of the libjpeg-turbo Project nor the names of its
+  contributors may be used to endorse or promote products derived from this
+  software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS",
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+
+
+Why Two Licenses?
+=================
+
+The zlib License could have been used instead of the Modified (3-clause) BSD
+License, and since the IJG License effectively subsumes the distribution
+conditions of the zlib License, this would have effectively placed
+libjpeg-turbo binary distributions under the IJG License.  However, the IJG
+License specifically refers to the Independent JPEG Group and does not extend
+attribution and endorsement protections to other entities.  Thus, it was
+desirable to choose a license that granted us the same protections for new code
+that were granted to the IJG for code derived from their software.

thirdparty/libjpeg-turbo/README.ijg

@@ -0,0 +1,260 @@
+libjpeg-turbo note:  This file has been modified by The libjpeg-turbo Project
+to include only information relevant to libjpeg-turbo, to wordsmith certain
+sections, and to remove impolitic language that existed in the libjpeg v8
+README.  It is included only for reference.  Please see README.md for
+information specific to libjpeg-turbo.
+
+
+The Independent JPEG Group's JPEG software
+==========================================
+
+This distribution contains a release of the Independent JPEG Group's free JPEG
+software.  You are welcome to redistribute this software and to use it for any
+purpose, subject to the conditions under LEGAL ISSUES, below.
+
+This software is the work of Tom Lane, Guido Vollbeding, Philip Gladstone,
+Bill Allombert, Jim Boucher, Lee Crocker, Bob Friesenhahn, Ben Jackson,
+Julian Minguillon, Luis Ortiz, George Phillips, Davide Rossi, Ge' Weijers,
+and other members of the Independent JPEG Group.
+
+IJG is not affiliated with the ISO/IEC JTC1/SC29/WG1 standards committee
+(also known as JPEG, together with ITU-T SG16).
+
+
+DOCUMENTATION ROADMAP
+=====================
+
+This file contains the following sections:
+
+OVERVIEW            General description of JPEG and the IJG software.
+LEGAL ISSUES        Copyright, lack of warranty, terms of distribution.
+REFERENCES          Where to learn more about JPEG.
+ARCHIVE LOCATIONS   Where to find newer versions of this software.
+FILE FORMAT WARS    Software *not* to get.
+TO DO               Plans for future IJG releases.
+
+Other documentation files in the distribution are:
+
+User documentation:
+  doc/usage.txt         Usage instructions for cjpeg, djpeg, jpegtran,
+                        rdjpgcom, and wrjpgcom.
+  doc/*.1               Unix-style man pages for programs (same info as
+                        usage.txt).
+  doc/wizard.txt        Advanced usage instructions for JPEG wizards only.
+  doc/change.log        Version-to-version change highlights.
+Programmer and internal documentation:
+  doc/libjpeg.txt       How to use the JPEG library in your own programs.
+  src/example.c         Sample code for calling the JPEG library.
+  doc/structure.txt     Overview of the JPEG library's internal structure.
+  doc/coderules.txt     Coding style rules --- please read if you contribute
+                        code.
+
+Please read at least usage.txt.  Some information can also be found in the JPEG
+FAQ (Frequently Asked Questions) article.  See ARCHIVE LOCATIONS below to find
+out where to obtain the FAQ article.
+
+If you want to understand how the JPEG code works, we suggest reading one or
+more of the REFERENCES, then looking at the documentation files (in roughly
+the order listed) before diving into the code.
+
+
+OVERVIEW
+========
+
+This package contains C software to implement JPEG image encoding, decoding,
+and transcoding.  JPEG (pronounced "jay-peg") is a standardized compression
+method for full-color and grayscale images.  JPEG's strong suit is compressing
+photographic images or other types of images that have smooth color and
+brightness transitions between neighboring pixels.  Images with sharp lines or
+other abrupt features may not compress well with JPEG, and a higher JPEG
+quality may have to be used to avoid visible compression artifacts with such
+images.
+
+JPEG is normally lossy, meaning that the output pixels are not necessarily
+identical to the input pixels.  However, on photographic content and other
+"smooth" images, very good compression ratios can be obtained with no visible
+compression artifacts, and extremely high compression ratios are possible if
+you are willing to sacrifice image quality (by reducing the "quality" setting
+in the compressor.)
+
+This software implements JPEG baseline, extended-sequential, progressive, and
+lossless compression processes.  Provision is made for supporting all variants
+of these processes, although some uncommon parameter settings aren't
+implemented yet.  We have made no provision for supporting the hierarchical
+processes defined in the standard.
+
+We provide a set of library routines for reading and writing JPEG image files,
+plus two sample applications "cjpeg" and "djpeg", which use the library to
+perform conversion between JPEG and some other popular image file formats.
+The library is intended to be reused in other applications.
+
+In order to support file conversion and viewing software, we have included
+considerable functionality beyond the bare JPEG coding/decoding capability;
+for example, the color quantization modules are not strictly part of JPEG
+decoding, but they are essential for output to colormapped file formats.  These
+extra functions can be compiled out of the library if not required for a
+particular application.
+
+We have also included "jpegtran", a utility for lossless transcoding between
+different JPEG processes, and "rdjpgcom" and "wrjpgcom", two simple
+applications for inserting and extracting textual comments in JFIF files.
+
+The emphasis in designing this software has been on achieving portability and
+flexibility, while also making it fast enough to be useful.  In particular,
+the software is not intended to be read as a tutorial on JPEG.  (See the
+REFERENCES section for introductory material.)  Rather, it is intended to
+be reliable, portable, industrial-strength code.  We do not claim to have
+achieved that goal in every aspect of the software, but we strive for it.
+
+We welcome the use of this software as a component of commercial products.
+No royalty is required, but we do ask for an acknowledgement in product
+documentation, as described under LEGAL ISSUES.
+
+
+LEGAL ISSUES
+============
+
+In plain English:
+
+1. We don't promise that this software works.  (But if you find any bugs,
+   please let us know!)
+2. You can use this software for whatever you want.  You don't have to pay us.
+3. You may not pretend that you wrote this software.  If you use it in a
+   program, you must acknowledge somewhere in your documentation that
+   you've used the IJG code.
+
+In legalese:
+
+The authors make NO WARRANTY or representation, either express or implied,
+with respect to this software, its quality, accuracy, merchantability, or
+fitness for a particular purpose.  This software is provided "AS IS", and you,
+its user, assume the entire risk as to its quality and accuracy.
+
+This software is copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
+All Rights Reserved except as specified below.
+
+Permission is hereby granted to use, copy, modify, and distribute this
+software (or portions thereof) for any purpose, without fee, subject to these
+conditions:
+(1) If any part of the source code for this software is distributed, then this
+README file must be included, with this copyright and no-warranty notice
+unaltered; and any additions, deletions, or changes to the original files
+must be clearly indicated in accompanying documentation.
+(2) If only executable code is distributed, then the accompanying
+documentation must state that "this software is based in part on the work of
+the Independent JPEG Group".
+(3) Permission for use of this software is granted only if the user accepts
+full responsibility for any undesirable consequences; the authors accept
+NO LIABILITY for damages of any kind.
+
+These conditions apply to any software derived from or based on the IJG code,
+not just to the unmodified library.  If you use our work, you ought to
+acknowledge us.
+
+Permission is NOT granted for the use of any IJG author's name or company name
+in advertising or publicity relating to this software or products derived from
+it.  This software may be referred to only as "the Independent JPEG Group's
+software".
+
+We specifically permit and encourage the use of this software as the basis of
+commercial products, provided that all warranty or liability claims are
+assumed by the product vendor.
+
+
+REFERENCES
+==========
+
+We recommend reading one or more of these references before trying to
+understand the innards of the JPEG software.
+
+The best short technical introduction to the JPEG compression algorithm is
+        Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+        Communications of the ACM, April 1991 (vol. 34 no. 4), pp. 30-44.
+(Adjacent articles in that issue discuss MPEG motion picture compression,
+applications of JPEG, and related topics.)  If you don't have the CACM issue
+handy, a PDF file containing a revised version of Wallace's article is
+available at http://www.ijg.org/files/Wallace.JPEG.pdf.  The file (actually
+a preprint for an article that appeared in IEEE Trans. Consumer Electronics)
+omits the sample images that appeared in CACM, but it includes corrections
+and some added material.  Note: the Wallace article is copyright ACM and IEEE,
+and it may not be used for commercial purposes.
+
+A somewhat less technical, more leisurely introduction to JPEG can be found in
+"The Data Compression Book" by Mark Nelson and Jean-loup Gailly, published by
+M&T Books (New York), 2nd ed. 1996, ISBN 1-55851-434-1.  This book provides
+good explanations and example C code for a multitude of compression methods
+including JPEG.  It is an excellent source if you are comfortable reading C
+code but don't know much about data compression in general.  The book's JPEG
+sample code is far from industrial-strength, but when you are ready to look
+at a full implementation, you've got one here...
+
+The best currently available description of JPEG is the textbook "JPEG Still
+Image Data Compression Standard" by William B. Pennebaker and Joan L.
+Mitchell, published by Van Nostrand Reinhold, 1993, ISBN 0-442-01272-1.
+Price US$59.95, 638 pp.  The book includes the complete text of the ISO JPEG
+standards (DIS 10918-1 and draft DIS 10918-2).
+
+The original JPEG standard is divided into two parts, Part 1 being the actual
+specification, while Part 2 covers compliance testing methods.  Part 1 is
+titled "Digital Compression and Coding of Continuous-tone Still Images,
+Part 1: Requirements and guidelines" and has document numbers ISO/IEC IS
+10918-1, ITU-T T.81.  Part 2 is titled "Digital Compression and Coding of
+Continuous-tone Still Images, Part 2: Compliance testing" and has document
+numbers ISO/IEC IS 10918-2, ITU-T T.83.
+
+The JPEG standard does not specify all details of an interchangeable file
+format.  For the omitted details, we follow the "JFIF" conventions, revision
+1.02.  JFIF version 1 has been adopted as ISO/IEC 10918-5 (05/2013) and
+Recommendation ITU-T T.871 (05/2011): Information technology - Digital
+compression and coding of continuous-tone still images: JPEG File Interchange
+Format (JFIF).  It is available as a free download in PDF file format from
+https://www.iso.org/standard/54989.html and http://www.itu.int/rec/T-REC-T.871.
+A PDF file of the older JFIF 1.02 specification is available at
+http://www.w3.org/Graphics/JPEG/jfif3.pdf.
+
+The TIFF 6.0 file format specification can be obtained from
+http://mirrors.ctan.org/graphics/tiff/TIFF6.ps.gz.  The JPEG incorporation
+scheme found in the TIFF 6.0 spec of 3-June-92 has a number of serious
+problems.  IJG does not recommend use of the TIFF 6.0 design (TIFF Compression
+tag 6).  Instead, we recommend the JPEG design proposed by TIFF Technical Note
+#2 (Compression tag 7).  Copies of this Note can be obtained from
+http://www.ijg.org/files/.  It is expected that the next revision
+of the TIFF spec will replace the 6.0 JPEG design with the Note's design.
+Although IJG's own code does not support TIFF/JPEG, the free libtiff library
+uses our library to implement TIFF/JPEG per the Note.
+
+
+ARCHIVE LOCATIONS
+=================
+
+The "official" archive site for this software is www.ijg.org.
+The most recent released version can always be found there in
+directory "files".
+
+The JPEG FAQ (Frequently Asked Questions) article is a source of some
+general information about JPEG.  It is available at
+http://www.faqs.org/faqs/jpeg-faq.
+
+
+FILE FORMAT COMPATIBILITY
+=========================
+
+This software implements ITU T.81 | ISO/IEC 10918 with some extensions from
+ITU T.871 | ISO/IEC 10918-5 (JPEG File Interchange Format-- see REFERENCES).
+Informally, the term "JPEG image" or "JPEG file" most often refers to JFIF or
+a subset thereof, but there are other formats containing the name "JPEG" that
+are incompatible with the original JPEG standard or with JFIF (for instance,
+JPEG 2000 and JPEG XR).  This software therefore does not support these
+formats.  Indeed, one of the original reasons for developing this free software
+was to help force convergence on a common, interoperable format standard for
+JPEG files.
+
+JFIF is a minimal or "low end" representation.  TIFF/JPEG (TIFF revision 6.0 as
+modified by TIFF Technical Note #2) can be used for "high end" applications
+that need to record a lot of additional data about an image.
+
+
+TO DO
+=====
+
+Please send bug reports, offers of help, etc. to [email protected].

thirdparty/libjpeg-turbo/patches/0001-cmake-generated-headers.patch

@@ -0,0 +1,629 @@
+diff --git a/thirdparty/libjpeg-turbo/src/jconfig.h b/thirdparty/libjpeg-turbo/src/jconfig.h
+new file mode 100644
+index 0000000000..22ed4b4c32
+--- /dev/null
++++ b/thirdparty/libjpeg-turbo/src/jconfig.h
+@@ -0,0 +1,62 @@
++// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
++
++/* Version ID for the JPEG library.
++ * Might be useful for tests like "#if JPEG_LIB_VERSION >= 60".
++ */
++#define JPEG_LIB_VERSION  62
++
++/* libjpeg-turbo version */
++#define LIBJPEG_TURBO_VERSION  3.1.0
++
++/* libjpeg-turbo version in integer form */
++#define LIBJPEG_TURBO_VERSION_NUMBER  3001000
++
++/* Support arithmetic encoding when using 8-bit samples */
++#define C_ARITH_CODING_SUPPORTED 1
++
++/* Support arithmetic decoding when using 8-bit samples */
++#define D_ARITH_CODING_SUPPORTED 1
++
++/* Support in-memory source/destination managers */
++#define MEM_SRCDST_SUPPORTED  1
++
++/* Use accelerated SIMD routines when using 8-bit samples */
++//#define WITH_SIMD 1
++
++/* This version of libjpeg-turbo supports run-time selection of data precision,
++ * so BITS_IN_JSAMPLE is no longer used to specify the data precision at build
++ * time.  However, some downstream software expects the macro to be defined.
++ * Since 12-bit data precision is an opt-in feature that requires explicitly
++ * calling 12-bit-specific libjpeg API functions and using 12-bit-specific data
++ * types, the unmodified portion of the libjpeg API still behaves as if it were
++ * built for 8-bit precision, and JSAMPLE is still literally an 8-bit data
++ * type.  Thus, it is correct to define BITS_IN_JSAMPLE to 8 here.
++ */
++#ifndef BITS_IN_JSAMPLE
++#define BITS_IN_JSAMPLE  8
++#endif
++
++#ifdef _WIN32
++
++#undef RIGHT_SHIFT_IS_UNSIGNED
++
++/* Define "boolean" as unsigned char, not int, per Windows custom */
++#ifndef __RPCNDR_H__            /* don't conflict if rpcndr.h already read */
++typedef unsigned char boolean;
++#endif
++#define HAVE_BOOLEAN            /* prevent jmorecfg.h from redefining it */
++
++/* Define "INT32" as int, not long, per Windows custom */
++#if !(defined(_BASETSD_H_) || defined(_BASETSD_H))   /* don't conflict if basetsd.h already read */
++typedef short INT16;
++typedef signed int INT32;
++#endif
++#define XMD_H                   /* prevent jmorecfg.h from redefining it */
++
++#else
++
++/* Define if your (broken) compiler shifts signed values as if they were
++   unsigned. */
++/* #undef RIGHT_SHIFT_IS_UNSIGNED */
++
++#endif
+diff --git a/thirdparty/libjpeg-turbo/src/jconfigint.h b/thirdparty/libjpeg-turbo/src/jconfigint.h
+new file mode 100644
+index 0000000000..f6171bf846
+--- /dev/null
++++ b/thirdparty/libjpeg-turbo/src/jconfigint.h
+@@ -0,0 +1,94 @@
++// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
++
++/* libjpeg-turbo build number */
++#define BUILD  "20250317"
++
++/* How to hide global symbols. */
++#ifndef HIDDEN
++	#if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
++		#define HIDDEN  __attribute__((visibility("hidden")))
++	#else
++		#define HIDDEN
++	#endif
++#endif
++
++/* Compiler's inline keyword */
++#undef inline
++
++/* How to obtain function inlining. */
++#if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
++	#define INLINE  __inline__ __attribute__((always_inline))
++#else
++	#define INLINE inline
++#endif
++
++/* How to obtain thread-local storage */
++#if defined(_MSC_VER)
++#define THREAD_LOCAL  __declspec(thread)
++#else
++#define THREAD_LOCAL  __thread
++#endif
++
++/* Define to the full name of this package. */
++#define PACKAGE_NAME  "libjpeg-turbo"
++
++/* Version number of package */
++#define VERSION  "3.1.1"
++
++/* The size of `size_t', as computed by sizeof. */
++#define SIZEOF_SIZE_T  8
++
++/* Define if your compiler has __builtin_ctzl() and sizeof(unsigned long) == sizeof(size_t). */
++#if defined(__GNUC__)
++	#define HAVE_BUILTIN_CTZL
++#endif
++
++/* Define to 1 if you have the <intrin.h> header file. */
++/* #undef HAVE_INTRIN_H */
++
++#if defined(_MSC_VER) && defined(HAVE_INTRIN_H)
++#if (SIZEOF_SIZE_T == 8)
++#define HAVE_BITSCANFORWARD64
++#elif (SIZEOF_SIZE_T == 4)
++#define HAVE_BITSCANFORWARD
++#endif
++#endif
++
++#if defined(__has_attribute)
++#if __has_attribute(fallthrough)
++#define FALLTHROUGH  __attribute__((fallthrough));
++#else
++#define FALLTHROUGH
++#endif
++#else
++#define FALLTHROUGH
++#endif
++
++/*
++ * Define BITS_IN_JSAMPLE as either
++ *   8   for 8-bit sample values (the usual setting)
++ *   12  for 12-bit sample values
++ * Only 8 and 12 are legal data precisions for lossy JPEG according to the
++ * JPEG standard, and the IJG code does not support anything else!
++ */
++
++#ifndef BITS_IN_JSAMPLE
++#define BITS_IN_JSAMPLE  8      /* use 8 or 12 */
++#endif
++
++#undef C_ARITH_CODING_SUPPORTED
++#undef D_ARITH_CODING_SUPPORTED
++#undef WITH_SIMD
++
++#if BITS_IN_JSAMPLE == 8
++
++/* Support arithmetic encoding */
++#define C_ARITH_CODING_SUPPORTED 1
++
++/* Support arithmetic decoding */
++#define D_ARITH_CODING_SUPPORTED 1
++
++/* Use accelerated SIMD routines. */
++//#define WITH_SIMD 1
++
++#endif
+diff --git a/thirdparty/libjpeg-turbo/src/jmorecfg.h b/thirdparty/libjpeg-turbo/src/jmorecfg.h
+new file mode 100644
+index 0000000000..f7de737edb
+--- /dev/null
++++ b/thirdparty/libjpeg-turbo/src/jmorecfg.h
+@@ -0,0 +1,389 @@
++// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
++
++/*
++ * jmorecfg.h
++ *
++ * This file was part of the Independent JPEG Group's software:
++ * Copyright (C) 1991-1997, Thomas G. Lane.
++ * Modified 1997-2009 by Guido Vollbeding.
++ * Lossless JPEG Modifications:
++ * Copyright (C) 1999, Ken Murchison.
++ * libjpeg-turbo Modifications:
++ * Copyright (C) 2009, 2011, 2014-2015, 2018, 2020, 2022, D. R. Commander.
++ * Godot modifications:
++ * Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md).
++ * For conditions of distribution and use, see the accompanying README.ijg
++ * file.
++ *
++ * This file contains additional configuration options that customize the
++ * JPEG software for special applications or support machine-dependent
++ * optimizations.  Most users will not need to touch this file.
++ */
++
++
++/*
++ * Maximum number of components (color channels) allowed in JPEG image.
++ * To meet the letter of Rec. ITU-T T.81 | ISO/IEC 10918-1, set this to 255.
++ * However, darn few applications need more than 4 channels (maybe 5 for CMYK +
++ * alpha mask).  We recommend 10 as a reasonable compromise; use 4 if you are
++ * really short on memory.  (Each allowed component costs a hundred or so
++ * bytes of storage, whether actually used in an image or not.)
++ */
++
++#define MAX_COMPONENTS  10      /* maximum number of image components */
++
++
++/*
++ * Basic data types.
++ * You may need to change these if you have a machine with unusual data
++ * type sizes; for example, "char" not 8 bits, "short" not 16 bits,
++ * or "long" not 32 bits.  We don't care whether "int" is 16 or 32 bits,
++ * but it had better be at least 16.
++ */
++
++/* Representation of a single sample (pixel element value).
++ * We frequently allocate large arrays of these, so it's important to keep
++ * them small.  But if you have memory to burn and access to char or short
++ * arrays is very slow on your hardware, you might want to change these.
++ */
++
++/* JSAMPLE should be the smallest type that will hold the values 0..255. */
++
++typedef unsigned char JSAMPLE;
++#define GETJSAMPLE(value)  ((int)(value))
++
++#define MAXJSAMPLE       255
++#define CENTERJSAMPLE    128
++
++
++/* J12SAMPLE should be the smallest type that will hold the values 0..4095. */
++
++typedef short J12SAMPLE;
++
++#define MAXJ12SAMPLE     4095
++#define CENTERJ12SAMPLE  2048
++
++
++/* J16SAMPLE should be the smallest type that will hold the values 0..65535. */
++
++typedef unsigned short J16SAMPLE;
++
++#define MAXJ16SAMPLE     65535
++#define CENTERJ16SAMPLE  32768
++
++
++/* Representation of a DCT frequency coefficient.
++ * This should be a signed value of at least 16 bits; "short" is usually OK.
++ * Again, we allocate large arrays of these, but you can change to int
++ * if you have memory to burn and "short" is really slow.
++ */
++
++typedef short JCOEF;
++
++
++/* Compressed datastreams are represented as arrays of JOCTET.
++ * These must be EXACTLY 8 bits wide, at least once they are written to
++ * external storage.  Note that when using the stdio data source/destination
++ * managers, this is also the data type passed to fread/fwrite.
++ */
++
++typedef unsigned char JOCTET;
++#define GETJOCTET(value)  (value)
++
++
++/* These typedefs are used for various table entries and so forth.
++ * They must be at least as wide as specified; but making them too big
++ * won't cost a huge amount of memory, so we don't provide special
++ * extraction code like we did for JSAMPLE.  (In other words, these
++ * typedefs live at a different point on the speed/space tradeoff curve.)
++ */
++
++/* UINT8 must hold at least the values 0..255. */
++
++typedef unsigned char UINT8;
++
++/* UINT16 must hold at least the values 0..65535. */
++
++typedef unsigned short UINT16;
++
++/* INT16 must hold at least the values -32768..32767. */
++
++#ifndef XMD_H                   /* X11/xmd.h correctly defines INT16 */
++typedef short INT16;
++#endif
++
++/* INT32 must hold at least signed 32-bit values.
++ *
++ * NOTE: The INT32 typedef dates back to libjpeg v5 (1994.)  Integers were
++ * sometimes 16-bit back then (MS-DOS), which is why INT32 is typedef'd to
++ * long.  It also wasn't common (or at least as common) in 1994 for INT32 to be
++ * defined by platform headers.  Since then, however, INT32 is defined in
++ * several other common places:
++ *
++ * Xmd.h (X11 header) typedefs INT32 to int on 64-bit platforms and long on
++ * 32-bit platforms (i.e always a 32-bit signed type.)
++ *
++ * basetsd.h (Win32 header) typedefs INT32 to int (always a 32-bit signed type
++ * on modern platforms.)
++ *
++ * qglobal.h (Qt header) typedefs INT32 to int (always a 32-bit signed type on
++ * modern platforms.)
++ *
++ * This is a recipe for conflict, since "long" and "int" aren't always
++ * compatible types.  Since the definition of INT32 has technically been part
++ * of the libjpeg API for more than 20 years, we can't remove it, but we do not
++ * use it internally any longer.  We instead define a separate type (JLONG)
++ * for internal use, which ensures that internal behavior will always be the
++ * same regardless of any external headers that may be included.
++ */
++
++#ifndef XMD_H                   /* X11/xmd.h correctly defines INT32 */
++#ifndef _BASETSD_H_             /* Microsoft defines it in basetsd.h */
++#ifndef _BASETSD_H              /* MinGW is slightly different */
++#ifndef QGLOBAL_H               /* Qt defines it in qglobal.h */
++typedef long INT32;
++#endif
++#endif
++#endif
++#endif
++
++/* Datatype used for image dimensions.  The JPEG standard only supports
++ * images up to 64K*64K due to 16-bit fields in SOF markers.  Therefore
++ * "unsigned int" is sufficient on all machines.  However, if you need to
++ * handle larger images and you don't mind deviating from the spec, you
++ * can change this datatype.  (Note that changing this datatype will
++ * potentially require modifying the SIMD code.  The x86-64 SIMD extensions,
++ * in particular, assume a 32-bit JDIMENSION.)
++ */
++
++typedef unsigned int JDIMENSION;
++
++#define JPEG_MAX_DIMENSION  65500L  /* a tad under 64K to prevent overflows */
++
++
++/* These macros are used in all function definitions and extern declarations.
++ * You could modify them if you need to change function linkage conventions;
++ * in particular, you'll need to do that to make the library a Windows DLL.
++ * Another application is to make all functions global for use with debuggers
++ * or code profilers that require it.
++ */
++
++/* a function called through method pointers: */
++#define METHODDEF(type)         static type
++/* a function used only in its module: */
++#define LOCAL(type)             static type
++/* a function referenced thru EXTERNs: */
++#define GLOBAL(type)            type
++/* a reference to a GLOBAL function: */
++#define EXTERN(type)            extern type
++
++
++/* Originally, this macro was used as a way of defining function prototypes
++ * for both modern compilers as well as older compilers that did not support
++ * prototype parameters.  libjpeg-turbo has never supported these older,
++ * non-ANSI compilers, but the macro is still included because there is some
++ * software out there that uses it.
++ */
++
++#define JMETHOD(type, methodname, arglist)  type (*methodname) arglist
++
++
++/* libjpeg-turbo no longer supports platforms that have far symbols (MS-DOS),
++ * but again, some software relies on this macro.
++ */
++
++#undef FAR
++#define FAR
++
++
++/*
++ * On a few systems, type boolean and/or its values FALSE, TRUE may appear
++ * in standard header files.  Or you may have conflicts with application-
++ * specific header files that you want to include together with these files.
++ * Defining HAVE_BOOLEAN before including jpeglib.h should make it work.
++ */
++
++#ifndef HAVE_BOOLEAN
++typedef int boolean;
++#endif
++#ifndef FALSE                   /* in case these macros already exist */
++#define FALSE   0               /* values of boolean */
++#endif
++#ifndef TRUE
++#define TRUE    1
++#endif
++
++
++/*
++ * The remaining options affect code selection within the JPEG library,
++ * but they don't need to be visible to most applications using the library.
++ * To minimize application namespace pollution, the symbols won't be
++ * defined unless JPEG_INTERNALS or JPEG_INTERNAL_OPTIONS has been defined.
++ */
++
++#ifdef JPEG_INTERNALS
++#define JPEG_INTERNAL_OPTIONS
++#endif
++
++#ifdef JPEG_INTERNAL_OPTIONS
++
++
++/*
++ * These defines indicate whether to include various optional functions.
++ * Undefining some of these symbols will produce a smaller but less capable
++ * library.  Note that you can leave certain source files out of the
++ * compilation/linking process if you've #undef'd the corresponding symbols.
++ * (You may HAVE to do that if your compiler doesn't like null source files.)
++ */
++
++/* Capability options common to encoder and decoder: */
++
++#define DCT_ISLOW_SUPPORTED     /* accurate integer method */
++#define DCT_IFAST_SUPPORTED     /* less accurate int method [legacy feature] */
++#define DCT_FLOAT_SUPPORTED     /* floating-point method [legacy feature] */
++
++/* Encoder capability options: */
++
++#define C_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
++#define C_PROGRESSIVE_SUPPORTED     /* Progressive JPEG? (Requires MULTISCAN)*/
++//#define C_LOSSLESS_SUPPORTED        /* Lossless JPEG? */
++#define ENTROPY_OPT_SUPPORTED       /* Optimization of entropy coding parms? */
++/* Note: if you selected 12-bit data precision, it is dangerous to turn off
++ * ENTROPY_OPT_SUPPORTED.  The standard Huffman tables are only good for 8-bit
++ * precision, so jchuff.c normally uses entropy optimization to compute
++ * usable tables for higher precision.  If you don't want to do optimization,
++ * you'll have to supply different default Huffman tables.
++ * The exact same statements apply for progressive and lossless JPEG:
++ * the default tables don't work for progressive mode or lossless mode.
++ * (This may get fixed, however.)
++ */
++#define INPUT_SMOOTHING_SUPPORTED   /* Input image smoothing option? */
++
++/* Decoder capability options: */
++
++#define D_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
++#define D_PROGRESSIVE_SUPPORTED     /* Progressive JPEG? (Requires MULTISCAN)*/
++//#define D_LOSSLESS_SUPPORTED        /* Lossless JPEG? */
++#define SAVE_MARKERS_SUPPORTED      /* jpeg_save_markers() needed? */
++#define BLOCK_SMOOTHING_SUPPORTED   /* Block smoothing? (Progressive only) */
++#define IDCT_SCALING_SUPPORTED      /* Output rescaling via IDCT? */
++#undef  UPSAMPLE_SCALING_SUPPORTED  /* Output rescaling at upsample stage? */
++#define UPSAMPLE_MERGING_SUPPORTED  /* Fast path for sloppy upsampling? */
++#define QUANT_1PASS_SUPPORTED       /* 1-pass color quantization? */
++#define QUANT_2PASS_SUPPORTED       /* 2-pass color quantization? */
++
++/* more capability options later, no doubt */
++
++
++/*
++ * The RGB_RED, RGB_GREEN, RGB_BLUE, and RGB_PIXELSIZE macros are a vestigial
++ * feature of libjpeg.  The idea was that, if an application developer needed
++ * to compress from/decompress to a BGR/BGRX/RGBX/XBGR/XRGB buffer, they could
++ * change these macros, rebuild libjpeg, and link their application statically
++ * with it.  In reality, few people ever did this, because there were some
++ * severe restrictions involved (cjpeg and djpeg no longer worked properly,
++ * compressing/decompressing RGB JPEGs no longer worked properly, and the color
++ * quantizer wouldn't work with pixel sizes other than 3.)  Furthermore, since
++ * all of the O/S-supplied versions of libjpeg were built with the default
++ * values of RGB_RED, RGB_GREEN, RGB_BLUE, and RGB_PIXELSIZE, many applications
++ * have come to regard these values as immutable.
++ *
++ * The libjpeg-turbo colorspace extensions provide a much cleaner way of
++ * compressing from/decompressing to buffers with arbitrary component orders
++ * and pixel sizes.  Thus, we do not support changing the values of RGB_RED,
++ * RGB_GREEN, RGB_BLUE, or RGB_PIXELSIZE.  In addition to the restrictions
++ * listed above, changing these values will also break the SIMD extensions and
++ * the regression tests.
++ */
++
++#define RGB_RED         0       /* Offset of Red in an RGB scanline element */
++#define RGB_GREEN       1       /* Offset of Green */
++#define RGB_BLUE        2       /* Offset of Blue */
++#define RGB_PIXELSIZE   3       /* JSAMPLEs per RGB scanline element */
++
++#define JPEG_NUMCS  17
++
++#define EXT_RGB_RED         0
++#define EXT_RGB_GREEN       1
++#define EXT_RGB_BLUE        2
++#define EXT_RGB_PIXELSIZE   3
++
++#define EXT_RGBX_RED        0
++#define EXT_RGBX_GREEN      1
++#define EXT_RGBX_BLUE       2
++#define EXT_RGBX_PIXELSIZE  4
++
++#define EXT_BGR_RED         2
++#define EXT_BGR_GREEN       1
++#define EXT_BGR_BLUE        0
++#define EXT_BGR_PIXELSIZE   3
++
++#define EXT_BGRX_RED        2
++#define EXT_BGRX_GREEN      1
++#define EXT_BGRX_BLUE       0
++#define EXT_BGRX_PIXELSIZE  4
++
++#define EXT_XBGR_RED        3
++#define EXT_XBGR_GREEN      2
++#define EXT_XBGR_BLUE       1
++#define EXT_XBGR_PIXELSIZE  4
++
++#define EXT_XRGB_RED        1
++#define EXT_XRGB_GREEN      2
++#define EXT_XRGB_BLUE       3
++#define EXT_XRGB_PIXELSIZE  4
++
++static const int rgb_red[JPEG_NUMCS] = {
++  -1, -1, RGB_RED, -1, -1, -1, EXT_RGB_RED, EXT_RGBX_RED,
++  EXT_BGR_RED, EXT_BGRX_RED, EXT_XBGR_RED, EXT_XRGB_RED,
++  EXT_RGBX_RED, EXT_BGRX_RED, EXT_XBGR_RED, EXT_XRGB_RED,
++  -1
++};
++
++static const int rgb_green[JPEG_NUMCS] = {
++  -1, -1, RGB_GREEN, -1, -1, -1, EXT_RGB_GREEN, EXT_RGBX_GREEN,
++  EXT_BGR_GREEN, EXT_BGRX_GREEN, EXT_XBGR_GREEN, EXT_XRGB_GREEN,
++  EXT_RGBX_GREEN, EXT_BGRX_GREEN, EXT_XBGR_GREEN, EXT_XRGB_GREEN,
++  -1
++};
++
++static const int rgb_blue[JPEG_NUMCS] = {
++  -1, -1, RGB_BLUE, -1, -1, -1, EXT_RGB_BLUE, EXT_RGBX_BLUE,
++  EXT_BGR_BLUE, EXT_BGRX_BLUE, EXT_XBGR_BLUE, EXT_XRGB_BLUE,
++  EXT_RGBX_BLUE, EXT_BGRX_BLUE, EXT_XBGR_BLUE, EXT_XRGB_BLUE,
++  -1
++};
++
++static const int rgb_pixelsize[JPEG_NUMCS] = {
++  -1, -1, RGB_PIXELSIZE, -1, -1, -1, EXT_RGB_PIXELSIZE, EXT_RGBX_PIXELSIZE,
++  EXT_BGR_PIXELSIZE, EXT_BGRX_PIXELSIZE, EXT_XBGR_PIXELSIZE, EXT_XRGB_PIXELSIZE,
++  EXT_RGBX_PIXELSIZE, EXT_BGRX_PIXELSIZE, EXT_XBGR_PIXELSIZE, EXT_XRGB_PIXELSIZE,
++  -1
++};
++
++/* Definitions for speed-related optimizations. */
++
++/* On some machines (notably 68000 series) "int" is 32 bits, but multiplying
++ * two 16-bit shorts is faster than multiplying two ints.  Define MULTIPLIER
++ * as short on such a machine.  MULTIPLIER must be at least 16 bits wide.
++ */
++
++#ifndef MULTIPLIER
++#ifndef WITH_SIMD
++#define MULTIPLIER  int         /* type for fastest integer multiply */
++#else
++#define MULTIPLIER  short       /* prefer 16-bit with SIMD for parellelism */
++#endif
++#endif
++
++
++/* FAST_FLOAT should be either float or double, whichever is done faster
++ * by your compiler.  (Note that this type is only used in the floating point
++ * DCT routines, so it only matters if you've defined DCT_FLOAT_SUPPORTED.)
++ */
++
++#ifndef FAST_FLOAT
++#define FAST_FLOAT  float
++#endif
++
++#endif /* JPEG_INTERNAL_OPTIONS */
+diff --git a/thirdparty/libjpeg-turbo/src/jversion.h b/thirdparty/libjpeg-turbo/src/jversion.h
+new file mode 100644
+index 0000000000..8e4f4ef749
+--- /dev/null
++++ b/thirdparty/libjpeg-turbo/src/jversion.h
+@@ -0,0 +1,60 @@
++// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
++
++/*
++ * jversion.h
++ *
++ * This file was part of the Independent JPEG Group's software:
++ * Copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
++ * libjpeg-turbo Modifications:
++ * Copyright (C) 2010, 2012-2024, D. R. Commander.
++ * Godot modifications:
++ * Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md).
++ * For conditions of distribution and use, see the accompanying README.ijg
++ * file.
++ *
++ * This file contains software version identification.
++ */
++
++
++#if JPEG_LIB_VERSION >= 80
++
++#define JVERSION        "8d  15-Jan-2012"
++
++#elif JPEG_LIB_VERSION >= 70
++
++#define JVERSION        "7  27-Jun-2009"
++
++#else
++
++#define JVERSION        "6b  27-Mar-1998"
++
++#endif
++
++/*
++ * NOTE: It is our convention to place the authors in the following order:
++ * - libjpeg-turbo authors (2009-) in descending order of the date of their
++ *   most recent contribution to the project, then in ascending order of the
++ *   date of their first contribution to the project, then in alphabetical
++ *   order
++ * - Upstream authors in descending order of the date of the first inclusion of
++ *   their code
++ */
++
++#define JCOPYRIGHT1 \
++  "Copyright (C) 2009-2024 D. R. Commander\n" \
++  "Copyright (C) 2015, 2020 Google, Inc.\n" \
++  "Copyright (C) 2019-2020 Arm Limited\n" \
++  "Copyright (C) 2015-2016, 2018 Matthieu Darbois\n" \
++  "Copyright (C) 2011-2016 Siarhei Siamashka\n" \
++  "Copyright (C) 2015 Intel Corporation\n"
++#define JCOPYRIGHT2 \
++  "Copyright (C) 2013-2014 Linaro Limited\n" \
++  "Copyright (C) 2013-2014 MIPS Technologies, Inc.\n" \
++  "Copyright (C) 2009, 2012 Pierre Ossman for Cendio AB\n" \
++  "Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies)\n" \
++  "Copyright (C) 1999-2006 MIYASAKA Masaru\n" \
++  "Copyright (C) 1999 Ken Murchison\n" \
++  "Copyright (C) 1991-2020 Thomas G. Lane, Guido Vollbeding\n"
++
++#define JCOPYRIGHT_SHORT \
++  "Copyright (C) 1991-2024 The libjpeg-turbo Project and many others"

thirdparty/libjpeg-turbo/patches/0002-disable-16bitlossless.patch

@@ -0,0 +1,20 @@
+commit 462c1cd875ae8f6b5f6406dda01881fb173ac30c
+Author: Daniel Kinsman <[email protected]>
+Date:   Thu Mar 20 12:21:28 2025 +1100
+
+    remove unneeded source files and lossless jpeg support
+
+diff --git a/thirdparty/libjpeg-turbo/src/turbojpeg.c b/thirdparty/libjpeg-turbo/src/turbojpeg.c
+index 389aea55d3..eec8e2a616 100644
+--- a/thirdparty/libjpeg-turbo/src/turbojpeg.c
++++ b/thirdparty/libjpeg-turbo/src/turbojpeg.c
+@@ -1200,9 +1200,6 @@ bailout:
+ #define BITS_IN_JSAMPLE  12
+ #include "turbojpeg-mp.c"
+ #undef BITS_IN_JSAMPLE
+-#define BITS_IN_JSAMPLE  16
+-#include "turbojpeg-mp.c"
+-#undef BITS_IN_JSAMPLE
+ 
+ /* TurboJPEG 1.2+ */
+ DLLEXPORT int tjCompress2(tjhandle handle, const unsigned char *srcBuf,

thirdparty/libjpeg-turbo/patches/0003-remove-bmp-ppm-support.patch

@@ -0,0 +1,329 @@
+diff --git a/thirdparty/libjpeg-turbo/src/turbojpeg-mp.c b/thirdparty/libjpeg-turbo/src/turbojpeg-mp.c
+index 1fa63b8185..72f99e236a 100644
+--- a/thirdparty/libjpeg-turbo/src/turbojpeg-mp.c
++++ b/thirdparty/libjpeg-turbo/src/turbojpeg-mp.c
+@@ -286,271 +286,6 @@ bailout:
+   return retval;
+ }
+ 
+-
+-/*************************** Packed-Pixel Image I/O **************************/
+-
+-/* TurboJPEG 3.0+ */
+-DLLEXPORT _JSAMPLE *GET_NAME(tj3LoadImage, BITS_IN_JSAMPLE)
+-  (tjhandle handle, const char *filename, int *width, int align, int *height,
+-   int *pixelFormat)
+-{
+-  static const char FUNCTION_NAME[] =
+-    GET_STRING(tj3LoadImage, BITS_IN_JSAMPLE);
+-
+-#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+-
+-  int retval = 0, tempc;
+-  size_t pitch;
+-  tjhandle handle2 = NULL;
+-  tjinstance *this2;
+-  j_compress_ptr cinfo = NULL;
+-  cjpeg_source_ptr src;
+-  _JSAMPLE *dstBuf = NULL;
+-  FILE *file = NULL;
+-  boolean invert;
+-
+-  GET_TJINSTANCE(handle, NULL)
+-
+-  if (!filename || !width || align < 1 || !height || !pixelFormat ||
+-      *pixelFormat < TJPF_UNKNOWN || *pixelFormat >= TJ_NUMPF)
+-    THROW("Invalid argument");
+-  if ((align & (align - 1)) != 0)
+-    THROW("Alignment must be a power of 2");
+-
+-  /* The instance handle passed to this function is used only for parameter
+-     retrieval.  Create a new temporary instance to avoid interfering with the
+-     libjpeg state of the primary instance. */
+-  if ((handle2 = tj3Init(TJINIT_COMPRESS)) == NULL) return NULL;
+-  this2 = (tjinstance *)handle2;
+-  cinfo = &this2->cinfo;
+-
+-#ifdef _MSC_VER
+-  if (fopen_s(&file, filename, "rb") || file == NULL)
+-#else
+-  if ((file = fopen(filename, "rb")) == NULL)
+-#endif
+-    THROW_UNIX("Cannot open input file");
+-
+-  if ((tempc = getc(file)) < 0 || ungetc(tempc, file) == EOF)
+-    THROW_UNIX("Could not read input file")
+-  else if (tempc == EOF)
+-    THROW("Input file contains no data");
+-
+-  if (setjmp(this2->jerr.setjmp_buffer)) {
+-    /* If we get here, the JPEG code has signaled an error. */
+-    retval = -1;  goto bailout;
+-  }
+-
+-  cinfo->data_precision = BITS_IN_JSAMPLE;
+-  if (*pixelFormat == TJPF_UNKNOWN) cinfo->in_color_space = JCS_UNKNOWN;
+-  else cinfo->in_color_space = pf2cs[*pixelFormat];
+-  if (tempc == 'B') {
+-    if ((src = jinit_read_bmp(cinfo, FALSE)) == NULL)
+-      THROW("Could not initialize bitmap loader");
+-    invert = !this->bottomUp;
+-  } else if (tempc == 'P') {
+-#if BITS_IN_JSAMPLE == 8
+-    if (this->precision >= 2 && this->precision <= BITS_IN_JSAMPLE)
+-#else
+-    if (this->precision >= BITS_IN_JSAMPLE - 3 &&
+-        this->precision <= BITS_IN_JSAMPLE)
+-#endif
+-      cinfo->data_precision = this->precision;
+-    if ((src = _jinit_read_ppm(cinfo)) == NULL)
+-      THROW("Could not initialize PPM loader");
+-    invert = this->bottomUp;
+-  } else
+-    THROW("Unsupported file type");
+-
+-  cinfo->mem->max_memory_to_use = (long)this->maxMemory * 1048576L;
+-
+-  src->input_file = file;
+-  /* Refuse to load images larger than the specified size. */
+-  src->max_pixels = this->maxPixels;
+-  (*src->start_input) (cinfo, src);
+-  if (tempc == 'B') {
+-    if (cinfo->X_density && cinfo->Y_density) {
+-      this->xDensity = cinfo->X_density;
+-      this->yDensity = cinfo->Y_density;
+-      this->densityUnits = cinfo->density_unit;
+-    }
+-  }
+-  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr)cinfo);
+-
+-  *width = cinfo->image_width;  *height = cinfo->image_height;
+-  *pixelFormat = cs2pf[cinfo->in_color_space];
+-
+-  pitch = PAD((*width) * tjPixelSize[*pixelFormat], align);
+-  if (
+-#if ULLONG_MAX > SIZE_MAX
+-      (unsigned long long)pitch * (unsigned long long)(*height) >
+-      (unsigned long long)((size_t)-1) ||
+-#endif
+-      (dstBuf = (_JSAMPLE *)malloc(pitch * (*height) *
+-                                   sizeof(_JSAMPLE))) == NULL)
+-    THROW("Memory allocation failure");
+-
+-  if (setjmp(this2->jerr.setjmp_buffer)) {
+-    /* If we get here, the JPEG code has signaled an error. */
+-    retval = -1;  goto bailout;
+-  }
+-
+-  while (cinfo->next_scanline < cinfo->image_height) {
+-    int i, nlines = (*src->get_pixel_rows) (cinfo, src);
+-
+-    for (i = 0; i < nlines; i++) {
+-      _JSAMPLE *dstptr;
+-      int row;
+-
+-      row = cinfo->next_scanline + i;
+-      if (invert) dstptr = &dstBuf[((*height) - row - 1) * pitch];
+-      else dstptr = &dstBuf[row * pitch];
+-      memcpy(dstptr, src->_buffer[i],
+-             (*width) * tjPixelSize[*pixelFormat] * sizeof(_JSAMPLE));
+-    }
+-    cinfo->next_scanline += nlines;
+-  }
+-
+-  (*src->finish_input) (cinfo, src);
+-
+-bailout:
+-  tj3Destroy(handle2);
+-  if (file) fclose(file);
+-  if (retval < 0) { free(dstBuf);  dstBuf = NULL; }
+-  return dstBuf;
+-
+-#else /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */
+-
+-  static const char ERROR_MSG[] =
+-    "16-bit data precision requires lossless JPEG,\n"
+-    "which was disabled at build time.";
+-  _JSAMPLE *retval = NULL;
+-
+-  GET_TJINSTANCE(handle, NULL)
+-  SNPRINTF(this->errStr, JMSG_LENGTH_MAX, "%s(): %s", FUNCTION_NAME,
+-           ERROR_MSG);
+-  this->isInstanceError = TRUE;  THROWG(ERROR_MSG, NULL)
+-
+-bailout:
+-  return retval;
+-
+-#endif
+-}
+-
+-
+-/* TurboJPEG 3.0+ */
+-DLLEXPORT int GET_NAME(tj3SaveImage, BITS_IN_JSAMPLE)
+-  (tjhandle handle, const char *filename, const _JSAMPLE *buffer, int width,
+-   int pitch, int height, int pixelFormat)
+-{
+-  static const char FUNCTION_NAME[] =
+-    GET_STRING(tj3SaveImage, BITS_IN_JSAMPLE);
+-  int retval = 0;
+-
+-#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+-
+-  tjhandle handle2 = NULL;
+-  tjinstance *this2;
+-  j_decompress_ptr dinfo = NULL;
+-  djpeg_dest_ptr dst;
+-  FILE *file = NULL;
+-  char *ptr = NULL;
+-  boolean invert;
+-
+-  GET_TJINSTANCE(handle, -1)
+-
+-  if (!filename || !buffer || width < 1 || pitch < 0 || height < 1 ||
+-      pixelFormat < 0 || pixelFormat >= TJ_NUMPF)
+-    THROW("Invalid argument");
+-
+-  /* The instance handle passed to this function is used only for parameter
+-     retrieval.  Create a new temporary instance to avoid interfering with the
+-     libjpeg state of the primary instance. */
+-  if ((handle2 = tj3Init(TJINIT_DECOMPRESS)) == NULL)
+-    return -1;
+-  this2 = (tjinstance *)handle2;
+-  dinfo = &this2->dinfo;
+-
+-#ifdef _MSC_VER
+-  if (fopen_s(&file, filename, "wb") || file == NULL)
+-#else
+-  if ((file = fopen(filename, "wb")) == NULL)
+-#endif
+-    THROW_UNIX("Cannot open output file");
+-
+-  if (setjmp(this2->jerr.setjmp_buffer)) {
+-    /* If we get here, the JPEG code has signaled an error. */
+-    retval = -1;  goto bailout;
+-  }
+-
+-  this2->dinfo.out_color_space = pf2cs[pixelFormat];
+-  dinfo->image_width = width;  dinfo->image_height = height;
+-  dinfo->global_state = DSTATE_READY;
+-  dinfo->scale_num = dinfo->scale_denom = 1;
+-  dinfo->data_precision = BITS_IN_JSAMPLE;
+-
+-  ptr = strrchr(filename, '.');
+-  if (ptr && !strcasecmp(ptr, ".bmp")) {
+-    if ((dst = jinit_write_bmp(dinfo, FALSE, FALSE)) == NULL)
+-      THROW("Could not initialize bitmap writer");
+-    invert = !this->bottomUp;
+-    dinfo->X_density = (UINT16)this->xDensity;
+-    dinfo->Y_density = (UINT16)this->yDensity;
+-    dinfo->density_unit = (UINT8)this->densityUnits;
+-  } else {
+-#if BITS_IN_JSAMPLE == 8
+-    if (this->precision >= 2 && this->precision <= BITS_IN_JSAMPLE)
+-#else
+-    if (this->precision >= BITS_IN_JSAMPLE - 3 &&
+-        this->precision <= BITS_IN_JSAMPLE)
+-#endif
+-      dinfo->data_precision = this->precision;
+-    if ((dst = _jinit_write_ppm(dinfo)) == NULL)
+-      THROW("Could not initialize PPM writer");
+-    invert = this->bottomUp;
+-  }
+-
+-  dinfo->mem->max_memory_to_use = (long)this->maxMemory * 1048576L;
+-
+-  dst->output_file = file;
+-  (*dst->start_output) (dinfo, dst);
+-  (*dinfo->mem->realize_virt_arrays) ((j_common_ptr)dinfo);
+-
+-  if (pitch == 0) pitch = width * tjPixelSize[pixelFormat];
+-
+-  while (dinfo->output_scanline < dinfo->output_height) {
+-    _JSAMPLE *rowptr;
+-
+-    if (invert)
+-      rowptr =
+-        (_JSAMPLE *)&buffer[(height - dinfo->output_scanline - 1) * pitch];
+-    else
+-      rowptr = (_JSAMPLE *)&buffer[dinfo->output_scanline * pitch];
+-    memcpy(dst->_buffer[0], rowptr,
+-           width * tjPixelSize[pixelFormat] * sizeof(_JSAMPLE));
+-    (*dst->put_pixel_rows) (dinfo, dst, 1);
+-    dinfo->output_scanline++;
+-  }
+-
+-  (*dst->finish_output) (dinfo, dst);
+-
+-bailout:
+-  tj3Destroy(handle2);
+-  if (file) fclose(file);
+-  return retval;
+-
+-#else /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */
+-
+-  GET_TJINSTANCE(handle, -1)
+-  THROW("16-bit data precision requires lossless JPEG,\n"
+-        "which was disabled at build time.")
+-bailout:
+-  return retval;
+-
+-#endif
+-}
+-
+-
+ #undef _JSAMPLE
+ #undef _JSAMPROW
+ #undef _buffer
+diff --git a/thirdparty/libjpeg-turbo/src/turbojpeg.c b/thirdparty/libjpeg-turbo/src/turbojpeg.c
+index eec8e2a616..8ce446148a 100644
+--- a/thirdparty/libjpeg-turbo/src/turbojpeg.c
++++ b/thirdparty/libjpeg-turbo/src/turbojpeg.c
+@@ -3095,48 +3095,3 @@ bailout:
+   free(sizes);
+   return retval;
+ }
+-
+-
+-/*************************** Packed-Pixel Image I/O **************************/
+-
+-/* tj3LoadImage*() is implemented in turbojpeg-mp.c */
+-
+-/* TurboJPEG 2.0+ */
+-DLLEXPORT unsigned char *tjLoadImage(const char *filename, int *width,
+-                                     int align, int *height,
+-                                     int *pixelFormat, int flags)
+-{
+-  tjhandle handle = NULL;
+-  unsigned char *dstBuf = NULL;
+-
+-  if ((handle = tj3Init(TJINIT_COMPRESS)) == NULL) return NULL;
+-
+-  processFlags(handle, flags, COMPRESS);
+-
+-  dstBuf = tj3LoadImage8(handle, filename, width, align, height, pixelFormat);
+-
+-  tj3Destroy(handle);
+-  return dstBuf;
+-}
+-
+-
+-/* tj3SaveImage*() is implemented in turbojpeg-mp.c */
+-
+-/* TurboJPEG 2.0+ */
+-DLLEXPORT int tjSaveImage(const char *filename, unsigned char *buffer,
+-                          int width, int pitch, int height, int pixelFormat,
+-                          int flags)
+-{
+-  tjhandle handle = NULL;
+-  int retval = -1;
+-
+-  if ((handle = tj3Init(TJINIT_DECOMPRESS)) == NULL) return -1;
+-
+-  processFlags(handle, flags, DECOMPRESS);
+-
+-  retval = tj3SaveImage8(handle, filename, buffer, width, pitch, height,
+-                         pixelFormat);
+-
+-  tj3Destroy(handle);
+-  return retval;
+-}

thirdparty/libjpeg-turbo/src/cderror.h

@@ -0,0 +1,114 @@
+/*
+ * cderror.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 2009-2017 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2021, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file defines the error and message codes for the cjpeg/djpeg
+ * applications.  These strings are not needed as part of the JPEG library
+ * proper.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef CDERROR_H
+#define CDERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code, string)
+#endif /* CDERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code, string)  code,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_FIRSTADDONCODE = 1000, NULL) /* Must be first entry! */
+
+JMESSAGE(JERR_BMP_BADCMAP, "Unsupported BMP colormap format")
+JMESSAGE(JERR_BMP_BADDEPTH, "Only 8-, 24-, and 32-bit BMP files are supported")
+JMESSAGE(JERR_BMP_BADHEADER, "Invalid BMP file: bad header length")
+JMESSAGE(JERR_BMP_BADPLANES, "Invalid BMP file: biPlanes not equal to 1")
+JMESSAGE(JERR_BMP_COLORSPACE, "BMP output must be grayscale or RGB")
+JMESSAGE(JERR_BMP_COMPRESSED, "Sorry, compressed BMPs not yet supported")
+JMESSAGE(JERR_BMP_EMPTY, "Empty BMP image")
+JMESSAGE(JERR_BMP_NOT, "Not a BMP file - does not start with BM")
+JMESSAGE(JERR_BMP_OUTOFRANGE, "Numeric value out of range in BMP file")
+JMESSAGE(JTRC_BMP, "%ux%u %d-bit BMP image")
+JMESSAGE(JTRC_BMP_MAPPED, "%ux%u 8-bit colormapped BMP image")
+JMESSAGE(JTRC_BMP_OS2, "%ux%u %d-bit OS2 BMP image")
+JMESSAGE(JTRC_BMP_OS2_MAPPED, "%ux%u 8-bit colormapped OS2 BMP image")
+
+JMESSAGE(JERR_GIF_BUG, "GIF output got confused")
+JMESSAGE(JERR_GIF_CODESIZE, "Bogus GIF codesize %d")
+JMESSAGE(JERR_GIF_COLORSPACE, "GIF output must be grayscale or RGB")
+JMESSAGE(JERR_GIF_EMPTY, "Empty GIF image")
+JMESSAGE(JERR_GIF_IMAGENOTFOUND, "Too few images in GIF file")
+JMESSAGE(JERR_GIF_NOT, "Not a GIF file")
+JMESSAGE(JTRC_GIF, "%ux%ux%d GIF image")
+JMESSAGE(JTRC_GIF_BADVERSION,
+         "Warning: unexpected GIF version number '%c%c%c'")
+JMESSAGE(JTRC_GIF_EXTENSION, "Ignoring GIF extension block of type 0x%02x")
+JMESSAGE(JTRC_GIF_NONSQUARE, "Caution: nonsquare pixels in input")
+JMESSAGE(JWRN_GIF_BADDATA, "Corrupt data in GIF file")
+JMESSAGE(JWRN_GIF_CHAR, "Bogus char 0x%02x in GIF file, ignoring")
+JMESSAGE(JWRN_GIF_ENDCODE, "Premature end of GIF image")
+JMESSAGE(JWRN_GIF_NOMOREDATA, "Ran out of GIF bits")
+
+JMESSAGE(JERR_PPM_COLORSPACE, "PPM output must be grayscale or RGB")
+JMESSAGE(JERR_PPM_NONNUMERIC, "Nonnumeric data in PPM file")
+JMESSAGE(JERR_PPM_NOT, "Not a PPM/PGM file")
+JMESSAGE(JERR_PPM_OUTOFRANGE, "Numeric value out of range in PPM file")
+JMESSAGE(JTRC_PGM, "%ux%u PGM image (maximum color value = %u)")
+JMESSAGE(JTRC_PGM_TEXT, "%ux%u text PGM image (maximum color value = %u)")
+JMESSAGE(JTRC_PPM, "%ux%u PPM image (maximum color value = %u)")
+JMESSAGE(JTRC_PPM_TEXT, "%ux%u text PPM image (maximum color value = %u)")
+
+JMESSAGE(JERR_TGA_BADCMAP, "Unsupported Targa colormap format")
+JMESSAGE(JERR_TGA_BADPARMS, "Invalid or unsupported Targa file")
+JMESSAGE(JERR_TGA_COLORSPACE, "Targa output must be grayscale or RGB")
+JMESSAGE(JTRC_TGA, "%ux%u RGB Targa image")
+JMESSAGE(JTRC_TGA_GRAY, "%ux%u grayscale Targa image")
+JMESSAGE(JTRC_TGA_MAPPED, "%ux%u colormapped Targa image")
+JMESSAGE(JERR_TGA_NOTCOMP, "Targa support was not compiled")
+
+JMESSAGE(JERR_BAD_CMAP_FILE,
+         "Color map file is invalid or of unsupported format")
+JMESSAGE(JERR_TOO_MANY_COLORS,
+         "Output file format cannot handle %d colormap entries")
+JMESSAGE(JERR_UNGETC_FAILED, "ungetc failed")
+#ifdef TARGA_SUPPORTED
+JMESSAGE(JERR_UNKNOWN_FORMAT,
+         "Unrecognized input file format --- perhaps you need -targa")
+#else
+JMESSAGE(JERR_UNKNOWN_FORMAT, "Unrecognized input file format")
+#endif
+JMESSAGE(JERR_UNSUPPORTED_FORMAT, "Unsupported output file format")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTADDONCODE
+} ADDON_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE

thirdparty/libjpeg-turbo/src/cdjpeg.h

@@ -0,0 +1,175 @@
+/*
+ * cdjpeg.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 2019 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2017, 2019, 2021-2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains common declarations for the sample applications
+ * cjpeg and djpeg.  It is NOT used by the core JPEG library.
+ */
+
+#define JPEG_CJPEG_DJPEG        /* define proper options in jconfig.h */
+#define JPEG_INTERNAL_OPTIONS   /* cjpeg.c,djpeg.c need to see xxx_SUPPORTED */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"             /* get library error codes too */
+#include "cderror.h"            /* get application-specific error codes */
+
+
+/*
+ * Object interface for cjpeg's source file decoding modules
+ */
+
+typedef struct cjpeg_source_struct *cjpeg_source_ptr;
+
+struct cjpeg_source_struct {
+  void (*start_input) (j_compress_ptr cinfo, cjpeg_source_ptr sinfo);
+  JDIMENSION (*get_pixel_rows) (j_compress_ptr cinfo, cjpeg_source_ptr sinfo);
+  void (*finish_input) (j_compress_ptr cinfo, cjpeg_source_ptr sinfo);
+
+  FILE *input_file;
+
+  JSAMPARRAY buffer;
+  J12SAMPARRAY buffer12;
+#ifdef C_LOSSLESS_SUPPORTED
+  J16SAMPARRAY buffer16;
+#endif
+  JDIMENSION buffer_height;
+  JDIMENSION max_pixels;
+};
+
+
+/*
+ * Object interface for djpeg's output file encoding modules
+ */
+
+typedef struct djpeg_dest_struct *djpeg_dest_ptr;
+
+struct djpeg_dest_struct {
+  /* start_output is called after jpeg_start_decompress finishes.
+   * The color map will be ready at this time, if one is needed.
+   */
+  void (*start_output) (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo);
+  /* Emit the specified number of pixel rows from the buffer. */
+  void (*put_pixel_rows) (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+                          JDIMENSION rows_supplied);
+  /* Finish up at the end of the image. */
+  void (*finish_output) (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo);
+  /* Re-calculate buffer dimensions based on output dimensions (for use with
+     partial image decompression.)  If this is NULL, then the output format
+     does not support partial image decompression (BMP, in particular, cannot
+     support partial decompression because it uses an inversion buffer to write
+     the image in bottom-up order.) */
+  void (*calc_buffer_dimensions) (j_decompress_ptr cinfo,
+                                  djpeg_dest_ptr dinfo);
+
+
+  /* Target file spec; filled in by djpeg.c after object is created. */
+  FILE *output_file;
+
+  /* Output pixel-row buffer.  Created by module init or start_output.
+   * Width is cinfo->output_width * cinfo->output_components;
+   * height is buffer_height.
+   */
+  JSAMPARRAY buffer;
+  J12SAMPARRAY buffer12;
+#ifdef D_LOSSLESS_SUPPORTED
+  J16SAMPARRAY buffer16;
+#endif
+  JDIMENSION buffer_height;
+};
+
+
+/*
+ * cjpeg/djpeg may need to perform extra passes to convert to or from
+ * the source/destination file format.  The JPEG library does not know
+ * about these passes, but we'd like them to be counted by the progress
+ * monitor.  We use an expanded progress monitor object to hold the
+ * additional pass count.
+ */
+
+struct cdjpeg_progress_mgr {
+  struct jpeg_progress_mgr pub; /* fields known to JPEG library */
+  int completed_extra_passes;   /* extra passes completed */
+  int total_extra_passes;       /* total extra */
+  JDIMENSION max_scans;         /* abort if the number of scans exceeds this
+                                   value and the value is non-zero */
+  boolean report;               /* whether or not to report progress */
+  /* last printed percentage stored here to avoid multiple printouts */
+  int percent_done;
+};
+
+typedef struct cdjpeg_progress_mgr *cd_progress_ptr;
+
+
+/* Module selection routines for I/O modules. */
+
+EXTERN(cjpeg_source_ptr) jinit_read_bmp(j_compress_ptr cinfo,
+                                        boolean use_inversion_array);
+EXTERN(djpeg_dest_ptr) jinit_write_bmp(j_decompress_ptr cinfo, boolean is_os2,
+                                       boolean use_inversion_array);
+EXTERN(cjpeg_source_ptr) jinit_read_gif(j_compress_ptr cinfo);
+EXTERN(djpeg_dest_ptr) jinit_write_gif(j_decompress_ptr cinfo, boolean is_lzw);
+EXTERN(djpeg_dest_ptr) j12init_write_gif(j_decompress_ptr cinfo,
+                                         boolean is_lzw);
+EXTERN(cjpeg_source_ptr) jinit_read_ppm(j_compress_ptr cinfo);
+EXTERN(cjpeg_source_ptr) j12init_read_ppm(j_compress_ptr cinfo);
+#ifdef C_LOSSLESS_SUPPORTED
+EXTERN(cjpeg_source_ptr) j16init_read_ppm(j_compress_ptr cinfo);
+#endif
+EXTERN(djpeg_dest_ptr) jinit_write_ppm(j_decompress_ptr cinfo);
+EXTERN(djpeg_dest_ptr) j12init_write_ppm(j_decompress_ptr cinfo);
+#ifdef D_LOSSLESS_SUPPORTED
+EXTERN(djpeg_dest_ptr) j16init_write_ppm(j_decompress_ptr cinfo);
+#endif
+EXTERN(cjpeg_source_ptr) jinit_read_targa(j_compress_ptr cinfo);
+EXTERN(djpeg_dest_ptr) jinit_write_targa(j_decompress_ptr cinfo);
+
+/* cjpeg support routines (in rdswitch.c) */
+
+EXTERN(boolean) read_quant_tables(j_compress_ptr cinfo, char *filename,
+                                  boolean force_baseline);
+EXTERN(boolean) read_scan_script(j_compress_ptr cinfo, char *filename);
+EXTERN(boolean) set_quality_ratings(j_compress_ptr cinfo, char *arg,
+                                    boolean force_baseline);
+EXTERN(boolean) set_quant_slots(j_compress_ptr cinfo, char *arg);
+EXTERN(boolean) set_sample_factors(j_compress_ptr cinfo, char *arg);
+
+/* djpeg support routines (in rdcolmap.c) */
+
+EXTERN(void) read_color_map(j_decompress_ptr cinfo, FILE *infile);
+EXTERN(void) read_color_map_12(j_decompress_ptr cinfo, FILE *infile);
+
+/* common support routines (in cdjpeg.c) */
+
+EXTERN(void) start_progress_monitor(j_common_ptr cinfo,
+                                    cd_progress_ptr progress);
+EXTERN(void) end_progress_monitor(j_common_ptr cinfo);
+EXTERN(boolean) keymatch(char *arg, const char *keyword, int minchars);
+EXTERN(FILE *) read_stdin(void);
+EXTERN(FILE *) write_stdout(void);
+
+/* miscellaneous useful macros */
+
+#ifdef DONT_USE_B_MODE          /* define mode parameters for fopen() */
+#define READ_BINARY     "r"
+#define WRITE_BINARY    "w"
+#else
+#define READ_BINARY     "rb"
+#define WRITE_BINARY    "wb"
+#endif
+
+#ifndef EXIT_FAILURE            /* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+#ifndef EXIT_SUCCESS
+#define EXIT_SUCCESS  0
+#endif
+#ifndef EXIT_WARNING
+#define EXIT_WARNING  2
+#endif

thirdparty/libjpeg-turbo/src/cmyk.h

@@ -0,0 +1,61 @@
+/*
+ * cmyk.h
+ *
+ * Copyright (C) 2017-2018, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains convenience functions for performing quick & dirty
+ * CMYK<->RGB conversion.  This algorithm is suitable for testing purposes
+ * only.  Properly converting between CMYK and RGB requires a color management
+ * system.
+ */
+
+#ifndef CMYK_H
+#define CMYK_H
+
+#include <jinclude.h>
+#define JPEG_INTERNALS
+#include <jpeglib.h>
+#include "jsamplecomp.h"
+
+
+/* Fully reversible */
+
+INLINE
+LOCAL(void)
+rgb_to_cmyk(int maxval, _JSAMPLE r, _JSAMPLE g, _JSAMPLE b,
+            _JSAMPLE *c, _JSAMPLE *m, _JSAMPLE *y, _JSAMPLE *k)
+{
+  double ctmp = 1.0 - ((double)r / (double)maxval);
+  double mtmp = 1.0 - ((double)g / (double)maxval);
+  double ytmp = 1.0 - ((double)b / (double)maxval);
+  double ktmp = MIN(MIN(ctmp, mtmp), ytmp);
+
+  if (ktmp == 1.0) ctmp = mtmp = ytmp = 0.0;
+  else {
+    ctmp = (ctmp - ktmp) / (1.0 - ktmp);
+    mtmp = (mtmp - ktmp) / (1.0 - ktmp);
+    ytmp = (ytmp - ktmp) / (1.0 - ktmp);
+  }
+  *c = (_JSAMPLE)((double)maxval - ctmp * (double)maxval + 0.5);
+  *m = (_JSAMPLE)((double)maxval - mtmp * (double)maxval + 0.5);
+  *y = (_JSAMPLE)((double)maxval - ytmp * (double)maxval + 0.5);
+  *k = (_JSAMPLE)((double)maxval - ktmp * (double)maxval + 0.5);
+}
+
+
+/* Fully reversible only for C/M/Y/K values generated with rgb_to_cmyk() */
+
+INLINE
+LOCAL(void)
+cmyk_to_rgb(int maxval, _JSAMPLE c, _JSAMPLE m, _JSAMPLE y, _JSAMPLE k,
+            _JSAMPLE *r, _JSAMPLE *g, _JSAMPLE *b)
+{
+  *r = (_JSAMPLE)((double)c * (double)k / (double)maxval + 0.5);
+  *g = (_JSAMPLE)((double)m * (double)k / (double)maxval + 0.5);
+  *b = (_JSAMPLE)((double)y * (double)k / (double)maxval + 0.5);
+}
+
+
+#endif /* CMYK_H */

thirdparty/libjpeg-turbo/src/jaricom.c

@@ -0,0 +1,157 @@
+/*
+ * jaricom.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Developed 1997-2009 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2018, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains probability estimation tables for common use in
+ * arithmetic entropy encoding and decoding routines.
+ *
+ * This data represents Table D.2 in
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994 and Table 24 in
+ * Recommendation ITU-T T.82 (1993) | ISO/IEC 11544:1993.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+/* The following #define specifies the packing of the four components
+ * into the compact JLONG representation.
+ * Note that this formula must match the actual arithmetic encoder
+ * and decoder implementation.  The implementation has to be changed
+ * if this formula is changed.
+ * The current organization is leaned on Markus Kuhn's JBIG
+ * implementation (jbig_tab.c).
+ */
+
+#define V(i, a, b, c, d) \
+  (((JLONG)a << 16) | ((JLONG)c << 8) | ((JLONG)d << 7) | b)
+
+const JLONG jpeg_aritab[113 + 1] = {
+/*
+ * Index, Qe_Value, Next_Index_LPS, Next_Index_MPS, Switch_MPS
+ */
+  V(   0, 0x5a1d,   1,   1, 1 ),
+  V(   1, 0x2586,  14,   2, 0 ),
+  V(   2, 0x1114,  16,   3, 0 ),
+  V(   3, 0x080b,  18,   4, 0 ),
+  V(   4, 0x03d8,  20,   5, 0 ),
+  V(   5, 0x01da,  23,   6, 0 ),
+  V(   6, 0x00e5,  25,   7, 0 ),
+  V(   7, 0x006f,  28,   8, 0 ),
+  V(   8, 0x0036,  30,   9, 0 ),
+  V(   9, 0x001a,  33,  10, 0 ),
+  V(  10, 0x000d,  35,  11, 0 ),
+  V(  11, 0x0006,   9,  12, 0 ),
+  V(  12, 0x0003,  10,  13, 0 ),
+  V(  13, 0x0001,  12,  13, 0 ),
+  V(  14, 0x5a7f,  15,  15, 1 ),
+  V(  15, 0x3f25,  36,  16, 0 ),
+  V(  16, 0x2cf2,  38,  17, 0 ),
+  V(  17, 0x207c,  39,  18, 0 ),
+  V(  18, 0x17b9,  40,  19, 0 ),
+  V(  19, 0x1182,  42,  20, 0 ),
+  V(  20, 0x0cef,  43,  21, 0 ),
+  V(  21, 0x09a1,  45,  22, 0 ),
+  V(  22, 0x072f,  46,  23, 0 ),
+  V(  23, 0x055c,  48,  24, 0 ),
+  V(  24, 0x0406,  49,  25, 0 ),
+  V(  25, 0x0303,  51,  26, 0 ),
+  V(  26, 0x0240,  52,  27, 0 ),
+  V(  27, 0x01b1,  54,  28, 0 ),
+  V(  28, 0x0144,  56,  29, 0 ),
+  V(  29, 0x00f5,  57,  30, 0 ),
+  V(  30, 0x00b7,  59,  31, 0 ),
+  V(  31, 0x008a,  60,  32, 0 ),
+  V(  32, 0x0068,  62,  33, 0 ),
+  V(  33, 0x004e,  63,  34, 0 ),
+  V(  34, 0x003b,  32,  35, 0 ),
+  V(  35, 0x002c,  33,   9, 0 ),
+  V(  36, 0x5ae1,  37,  37, 1 ),
+  V(  37, 0x484c,  64,  38, 0 ),
+  V(  38, 0x3a0d,  65,  39, 0 ),
+  V(  39, 0x2ef1,  67,  40, 0 ),
+  V(  40, 0x261f,  68,  41, 0 ),
+  V(  41, 0x1f33,  69,  42, 0 ),
+  V(  42, 0x19a8,  70,  43, 0 ),
+  V(  43, 0x1518,  72,  44, 0 ),
+  V(  44, 0x1177,  73,  45, 0 ),
+  V(  45, 0x0e74,  74,  46, 0 ),
+  V(  46, 0x0bfb,  75,  47, 0 ),
+  V(  47, 0x09f8,  77,  48, 0 ),
+  V(  48, 0x0861,  78,  49, 0 ),
+  V(  49, 0x0706,  79,  50, 0 ),
+  V(  50, 0x05cd,  48,  51, 0 ),
+  V(  51, 0x04de,  50,  52, 0 ),
+  V(  52, 0x040f,  50,  53, 0 ),
+  V(  53, 0x0363,  51,  54, 0 ),
+  V(  54, 0x02d4,  52,  55, 0 ),
+  V(  55, 0x025c,  53,  56, 0 ),
+  V(  56, 0x01f8,  54,  57, 0 ),
+  V(  57, 0x01a4,  55,  58, 0 ),
+  V(  58, 0x0160,  56,  59, 0 ),
+  V(  59, 0x0125,  57,  60, 0 ),
+  V(  60, 0x00f6,  58,  61, 0 ),
+  V(  61, 0x00cb,  59,  62, 0 ),
+  V(  62, 0x00ab,  61,  63, 0 ),
+  V(  63, 0x008f,  61,  32, 0 ),
+  V(  64, 0x5b12,  65,  65, 1 ),
+  V(  65, 0x4d04,  80,  66, 0 ),
+  V(  66, 0x412c,  81,  67, 0 ),
+  V(  67, 0x37d8,  82,  68, 0 ),
+  V(  68, 0x2fe8,  83,  69, 0 ),
+  V(  69, 0x293c,  84,  70, 0 ),
+  V(  70, 0x2379,  86,  71, 0 ),
+  V(  71, 0x1edf,  87,  72, 0 ),
+  V(  72, 0x1aa9,  87,  73, 0 ),
+  V(  73, 0x174e,  72,  74, 0 ),
+  V(  74, 0x1424,  72,  75, 0 ),
+  V(  75, 0x119c,  74,  76, 0 ),
+  V(  76, 0x0f6b,  74,  77, 0 ),
+  V(  77, 0x0d51,  75,  78, 0 ),
+  V(  78, 0x0bb6,  77,  79, 0 ),
+  V(  79, 0x0a40,  77,  48, 0 ),
+  V(  80, 0x5832,  80,  81, 1 ),
+  V(  81, 0x4d1c,  88,  82, 0 ),
+  V(  82, 0x438e,  89,  83, 0 ),
+  V(  83, 0x3bdd,  90,  84, 0 ),
+  V(  84, 0x34ee,  91,  85, 0 ),
+  V(  85, 0x2eae,  92,  86, 0 ),
+  V(  86, 0x299a,  93,  87, 0 ),
+  V(  87, 0x2516,  86,  71, 0 ),
+  V(  88, 0x5570,  88,  89, 1 ),
+  V(  89, 0x4ca9,  95,  90, 0 ),
+  V(  90, 0x44d9,  96,  91, 0 ),
+  V(  91, 0x3e22,  97,  92, 0 ),
+  V(  92, 0x3824,  99,  93, 0 ),
+  V(  93, 0x32b4,  99,  94, 0 ),
+  V(  94, 0x2e17,  93,  86, 0 ),
+  V(  95, 0x56a8,  95,  96, 1 ),
+  V(  96, 0x4f46, 101,  97, 0 ),
+  V(  97, 0x47e5, 102,  98, 0 ),
+  V(  98, 0x41cf, 103,  99, 0 ),
+  V(  99, 0x3c3d, 104, 100, 0 ),
+  V( 100, 0x375e,  99,  93, 0 ),
+  V( 101, 0x5231, 105, 102, 0 ),
+  V( 102, 0x4c0f, 106, 103, 0 ),
+  V( 103, 0x4639, 107, 104, 0 ),
+  V( 104, 0x415e, 103,  99, 0 ),
+  V( 105, 0x5627, 105, 106, 1 ),
+  V( 106, 0x50e7, 108, 107, 0 ),
+  V( 107, 0x4b85, 109, 103, 0 ),
+  V( 108, 0x5597, 110, 109, 0 ),
+  V( 109, 0x504f, 111, 107, 0 ),
+  V( 110, 0x5a10, 110, 111, 1 ),
+  V( 111, 0x5522, 112, 109, 0 ),
+  V( 112, 0x59eb, 112, 111, 1 ),
+/*
+ * This last entry is used for fixed probability estimate of 0.5
+ * as recommended in Section 10.3 Table 5 of ITU-T Rec. T.851.
+ */
+  V( 113, 0x5a1d, 113, 113, 0 )
+};

thirdparty/libjpeg-turbo/src/jcapimin.c

@@ -0,0 +1,318 @@
+/*
+ * jcapimin.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-compression case or the transcoding-only
+ * case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jcapistd.c.  But also see jcparam.c for
+ * parameter-setup helper routines, jcomapi.c for routines shared by
+ * compression and decompression, and jctrans.c for the transcoding case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jcmaster.h"
+
+
+/*
+ * Initialization of a JPEG compression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateCompress(j_compress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;            /* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != sizeof(struct jpeg_compress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE,
+             (int)sizeof(struct jpeg_compress_struct), (int)structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr *err = cinfo->err;
+    void *client_data = cinfo->client_data; /* ignore Purify complaint here */
+    memset(cinfo, 0, sizeof(struct jpeg_compress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = FALSE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr)cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->dest = NULL;
+
+  cinfo->comp_info = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    cinfo->quant_tbl_ptrs[i] = NULL;
+#if JPEG_LIB_VERSION >= 70
+    cinfo->q_scale_factor[i] = 100;
+#endif
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+#if JPEG_LIB_VERSION >= 80
+  /* Must do it here for emit_dqt in case jpeg_write_tables is used */
+  cinfo->block_size = DCTSIZE;
+  cinfo->natural_order = jpeg_natural_order;
+  cinfo->lim_Se = DCTSIZE2 - 1;
+#endif
+
+  cinfo->script_space = NULL;
+
+  cinfo->input_gamma = 1.0;     /* in case application forgets */
+
+  cinfo->data_precision = BITS_IN_JSAMPLE;
+
+  /* OK, I'm ready */
+  cinfo->global_state = CSTATE_START;
+
+  /* The master struct is used to store extension parameters, so we allocate it
+   * here.
+   */
+  cinfo->master = (struct jpeg_comp_master *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(my_comp_master));
+  memset(cinfo->master, 0, sizeof(my_comp_master));
+}
+
+
+/*
+ * Destruction of a JPEG compression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_compress(j_compress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr)cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG compression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_compress(j_compress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr)cinfo); /* use common routine */
+}
+
+
+/*
+ * Forcibly suppress or un-suppress all quantization and Huffman tables.
+ * Marks all currently defined tables as already written (if suppress)
+ * or not written (if !suppress).  This will control whether they get emitted
+ * by a subsequent jpeg_start_compress call.
+ *
+ * This routine is exported for use by applications that want to produce
+ * abbreviated JPEG datastreams.  It logically belongs in jcparam.c, but
+ * since it is called by jpeg_start_compress, we put it here --- otherwise
+ * jcparam.o would be linked whether the application used it or not.
+ */
+
+GLOBAL(void)
+jpeg_suppress_tables(j_compress_ptr cinfo, boolean suppress)
+{
+  int i;
+  JQUANT_TBL *qtbl;
+  JHUFF_TBL *htbl;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if ((qtbl = cinfo->quant_tbl_ptrs[i]) != NULL)
+      qtbl->sent_table = suppress;
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    if ((htbl = cinfo->dc_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+    if ((htbl = cinfo->ac_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+  }
+}
+
+
+/*
+ * Finish JPEG compression.
+ *
+ * If a multipass operating mode was selected, this may do a great deal of
+ * work including most of the actual output.
+ */
+
+GLOBAL(void)
+jpeg_finish_compress(j_compress_ptr cinfo)
+{
+  JDIMENSION iMCU_row;
+
+  if (cinfo->global_state == CSTATE_SCANNING ||
+      cinfo->global_state == CSTATE_RAW_OK) {
+    /* Terminate first pass */
+    if (cinfo->next_scanline < cinfo->image_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_pass) (cinfo);
+  } else if (cinfo->global_state != CSTATE_WRCOEFS)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any remaining passes */
+  while (!cinfo->master->is_last_pass) {
+    (*cinfo->master->prepare_for_pass) (cinfo);
+    for (iMCU_row = 0; iMCU_row < cinfo->total_iMCU_rows; iMCU_row++) {
+      if (cinfo->progress != NULL) {
+        cinfo->progress->pass_counter = (long)iMCU_row;
+        cinfo->progress->pass_limit = (long)cinfo->total_iMCU_rows;
+        (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+      }
+      /* We bypass the main controller and invoke coef controller directly;
+       * all work is being done from the coefficient buffer.
+       */
+      if (cinfo->data_precision <= 8) {
+        if (!(*cinfo->coef->compress_data) (cinfo, (JSAMPIMAGE)NULL))
+          ERREXIT(cinfo, JERR_CANT_SUSPEND);
+      } else if (cinfo->data_precision <= 12) {
+        if (!(*cinfo->coef->compress_data_12) (cinfo, (J12SAMPIMAGE)NULL))
+          ERREXIT(cinfo, JERR_CANT_SUSPEND);
+      } else {
+#ifdef C_LOSSLESS_SUPPORTED
+        if (!(*cinfo->coef->compress_data_16) (cinfo, (J16SAMPIMAGE)NULL))
+          ERREXIT(cinfo, JERR_CANT_SUSPEND);
+#else
+        ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+      }
+    }
+    (*cinfo->master->finish_pass) (cinfo);
+  }
+  /* Write EOI, do final cleanup */
+  (*cinfo->marker->write_file_trailer) (cinfo);
+  (*cinfo->dest->term_destination) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr)cinfo);
+}
+
+
+/*
+ * Write a special marker.
+ * This is only recommended for writing COM or APPn markers.
+ * Must be called after jpeg_start_compress() and before
+ * first call to jpeg_write_scanlines() or jpeg_write_raw_data().
+ */
+
+GLOBAL(void)
+jpeg_write_marker(j_compress_ptr cinfo, int marker, const JOCTET *dataptr,
+                  unsigned int datalen)
+{
+  void (*write_marker_byte) (j_compress_ptr info, int val);
+
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+  write_marker_byte = cinfo->marker->write_marker_byte; /* copy for speed */
+  while (datalen--) {
+    (*write_marker_byte) (cinfo, *dataptr);
+    dataptr++;
+  }
+}
+
+/* Same, but piecemeal. */
+
+GLOBAL(void)
+jpeg_write_m_header(j_compress_ptr cinfo, int marker, unsigned int datalen)
+{
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+}
+
+GLOBAL(void)
+jpeg_write_m_byte(j_compress_ptr cinfo, int val)
+{
+  (*cinfo->marker->write_marker_byte) (cinfo, val);
+}
+
+
+/*
+ * Alternate compression function: just write an abbreviated table file.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * To produce a pair of files containing abbreviated tables and abbreviated
+ * image data, one would proceed as follows:
+ *
+ *              initialize JPEG object
+ *              set JPEG parameters
+ *              set destination to table file
+ *              jpeg_write_tables(cinfo);
+ *              set destination to image file
+ *              jpeg_start_compress(cinfo, FALSE);
+ *              write data...
+ *              jpeg_finish_compress(cinfo);
+ *
+ * jpeg_write_tables has the side effect of marking all tables written
+ * (same as jpeg_suppress_tables(..., TRUE)).  Thus a subsequent start_compress
+ * will not re-emit the tables unless it is passed write_all_tables=TRUE.
+ */
+
+GLOBAL(void)
+jpeg_write_tables(j_compress_ptr cinfo)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr)cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Initialize the marker writer ... bit of a crock to do it here. */
+  jinit_marker_writer(cinfo);
+  /* Write them tables! */
+  (*cinfo->marker->write_tables_only) (cinfo);
+  /* And clean up. */
+  (*cinfo->dest->term_destination) (cinfo);
+  /*
+   * In library releases up through v6a, we called jpeg_abort() here to free
+   * any working memory allocated by the destination manager and marker
+   * writer.  Some applications had a problem with that: they allocated space
+   * of their own from the library memory manager, and didn't want it to go
+   * away during write_tables.  So now we do nothing.  This will cause a
+   * memory leak if an app calls write_tables repeatedly without doing a full
+   * compression cycle or otherwise resetting the JPEG object.  However, that
+   * seems less bad than unexpectedly freeing memory in the normal case.
+   * An app that prefers the old behavior can call jpeg_abort for itself after
+   * each call to jpeg_write_tables().
+   */
+}

thirdparty/libjpeg-turbo/src/jcapistd.c

@@ -0,0 +1,200 @@
+/*
+ * jcapistd.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-compression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_compress, it will end up linking in the entire compressor.
+ * We thus must separate this file from jcapimin.c to avoid linking the
+ * whole compression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE == 8
+
+/*
+ * Compression initialization.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * We require a write_all_tables parameter as a failsafe check when writing
+ * multiple datastreams from the same compression object.  Since prior runs
+ * will have left all the tables marked sent_table=TRUE, a subsequent run
+ * would emit an abbreviated stream (no tables) by default.  This may be what
+ * is wanted, but for safety's sake it should not be the default behavior:
+ * programmers should have to make a deliberate choice to emit abbreviated
+ * images.  Therefore the documentation and examples should encourage people
+ * to pass write_all_tables=TRUE; then it will take active thought to do the
+ * wrong thing.
+ */
+
+GLOBAL(void)
+jpeg_start_compress(j_compress_ptr cinfo, boolean write_all_tables)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (write_all_tables)
+    jpeg_suppress_tables(cinfo, FALSE); /* mark all tables to be written */
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr)cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  jinit_compress_master(cinfo);
+  /* Set up for the first pass */
+  (*cinfo->master->prepare_for_pass) (cinfo);
+  /* Ready for application to drive first pass through _jpeg_write_scanlines
+   * or _jpeg_write_raw_data.
+   */
+  cinfo->next_scanline = 0;
+  cinfo->global_state = (cinfo->raw_data_in ? CSTATE_RAW_OK : CSTATE_SCANNING);
+}
+
+#endif
+
+
+/*
+ * Write some scanlines of data to the JPEG compressor.
+ *
+ * The return value will be the number of lines actually written.
+ * This should be less than the supplied num_lines only in case that
+ * the data destination module has requested suspension of the compressor,
+ * or if more than image_height scanlines are passed in.
+ *
+ * Note: we warn about excess calls to _jpeg_write_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * excess scanlines passed in the last valid call are *silently* ignored,
+ * so that the application need not adjust num_lines for end-of-image
+ * when using a multiple-scanline buffer.
+ */
+
+GLOBAL(JDIMENSION)
+_jpeg_write_scanlines(j_compress_ptr cinfo, _JSAMPARRAY scanlines,
+                      JDIMENSION num_lines)
+{
+#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+  JDIMENSION row_ctr, rows_left;
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  if (cinfo->global_state != CSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height)
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long)cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long)cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * _jpeg_write_scanlines.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and _jpeg_write_scanlines.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Ignore any extra scanlines at bottom of image. */
+  rows_left = cinfo->image_height - cinfo->next_scanline;
+  if (num_lines > rows_left)
+    num_lines = rows_left;
+
+  row_ctr = 0;
+  (*cinfo->main->_process_data) (cinfo, scanlines, &row_ctr, num_lines);
+  cinfo->next_scanline += row_ctr;
+  return row_ctr;
+#else
+  ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  return 0;
+#endif
+}
+
+
+#if BITS_IN_JSAMPLE != 16
+
+/*
+ * Alternate entry point to write raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+_jpeg_write_raw_data(j_compress_ptr cinfo, _JSAMPIMAGE data,
+                     JDIMENSION num_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if (cinfo->global_state != CSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long)cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long)cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * _jpeg_write_raw_data.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and _jpeg_write_raw_data.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Verify that at least one iMCU row has been passed. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * DCTSIZE;
+  if (num_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Directly compress the row. */
+  if (!(*cinfo->coef->_compress_data) (cinfo, data)) {
+    /* If compressor did not consume the whole row, suspend processing. */
+    return 0;
+  }
+
+  /* OK, we processed one iMCU row. */
+  cinfo->next_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 */

thirdparty/libjpeg-turbo/src/jcarith.c

@@ -0,0 +1,932 @@
+/*
+ * jcarith.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Developed 1997-2009 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2018, 2021-2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains portable arithmetic entropy encoding routines for JPEG
+ * (implementing Recommendation ITU-T T.81 | ISO/IEC 10918-1).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Expanded entropy encoder object for arithmetic encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  JLONG c; /* C register, base of coding interval, layout as in sec. D.1.3 */
+  JLONG a;               /* A register, normalized size of coding interval */
+  JLONG sc;        /* counter for stacked 0xFF values which might overflow */
+  JLONG zc;          /* counter for pending 0x00 output values which might *
+                          * be discarded at the end ("Pacman" termination) */
+  int ct;  /* bit shift counter, determines when next byte will be written */
+  int buffer;                /* buffer for most recent output byte != 0xFF */
+
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+  int next_restart_num;         /* next restart number to write (0-7) */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char *dc_stats[NUM_ARITH_TBLS];
+  unsigned char *ac_stats[NUM_ARITH_TBLS];
+
+  /* Statistics bin for coding with fixed probability 0.5 */
+  unsigned char fixed_bin[4];
+} arith_entropy_encoder;
+
+typedef arith_entropy_encoder *arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS  64
+#define AC_STAT_BINS  256
+
+/* NOTE: Uncomment the following #define if you want to use the
+ * given formula for calculating the AC conditioning parameter Kx
+ * for spectral selection progressive coding in section G.1.3.2
+ * of the spec (Kx = Kmin + SRL (8 + Se - Kmin) 4).
+ * Although the spec and P&M authors claim that this "has proven
+ * to give good results for 8 bit precision samples", I'm not
+ * convinced yet that this is really beneficial.
+ * Early tests gave only very marginal compression enhancements
+ * (a few - around 5 or so - bytes even for very large files),
+ * which would turn out rather negative if we'd suppress the
+ * DAC (Define Arithmetic Conditioning) marker segments for
+ * the default parameters in the future.
+ * Note that currently the marker writing module emits 12-byte
+ * DAC segments for a full-component scan in a color image.
+ * This is not worth worrying about IMHO. However, since the
+ * spec defines the default values to be used if the tables
+ * are omitted (unlike Huffman tables, which are required
+ * anyway), one might optimize this behaviour in the future,
+ * and then it would be disadvantageous to use custom tables if
+ * they don't provide sufficient gain to exceed the DAC size.
+ *
+ * On the other hand, I'd consider it as a reasonable result
+ * that the conditioning has no significant influence on the
+ * compression performance. This means that the basic
+ * statistical model is already rather stable.
+ *
+ * Thus, at the moment, we use the default conditioning values
+ * anyway, and do not use the custom formula.
+ *
+#define CALCULATE_SPECTRAL_CONDITIONING
+ */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than JLONG.
+ * We assume that int right shift is unsigned if JLONG right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS    int ishift_temp;
+#define IRIGHT_SHIFT(x, shft) \
+  ((ishift_temp = (x)) < 0 ? \
+   (ishift_temp >> (shft)) | ((~0) << (16 - (shft))) : \
+   (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x, shft)   ((x) >> (shft))
+#endif
+
+
+LOCAL(void)
+emit_byte(int val, j_compress_ptr cinfo)
+/* Write next output byte; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr *dest = cinfo->dest;
+
+  *dest->next_output_byte++ = (JOCTET)val;
+  if (--dest->free_in_buffer == 0)
+    if (!(*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+}
+
+
+/*
+ * Finish up at the end of an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass(j_compress_ptr cinfo)
+{
+  arith_entropy_ptr e = (arith_entropy_ptr)cinfo->entropy;
+  JLONG temp;
+
+  /* Section D.1.8: Termination of encoding */
+
+  /* Find the e->c in the coding interval with the largest
+   * number of trailing zero bits */
+  if ((temp = (e->a - 1 + e->c) & 0xFFFF0000UL) < e->c)
+    e->c = temp + 0x8000L;
+  else
+    e->c = temp;
+  /* Send remaining bytes to output */
+  e->c <<= e->ct;
+  if (e->c & 0xF8000000UL) {
+    /* One final overflow has to be handled */
+    if (e->buffer >= 0) {
+      if (e->zc)
+        do emit_byte(0x00, cinfo);
+        while (--e->zc);
+      emit_byte(e->buffer + 1, cinfo);
+      if (e->buffer + 1 == 0xFF)
+        emit_byte(0x00, cinfo);
+    }
+    e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+    e->sc = 0;
+  } else {
+    if (e->buffer == 0)
+      ++e->zc;
+    else if (e->buffer >= 0) {
+      if (e->zc)
+        do emit_byte(0x00, cinfo);
+        while (--e->zc);
+      emit_byte(e->buffer, cinfo);
+    }
+    if (e->sc) {
+      if (e->zc)
+        do emit_byte(0x00, cinfo);
+        while (--e->zc);
+      do {
+        emit_byte(0xFF, cinfo);
+        emit_byte(0x00, cinfo);
+      } while (--e->sc);
+    }
+  }
+  /* Output final bytes only if they are not 0x00 */
+  if (e->c & 0x7FFF800L) {
+    if (e->zc)  /* output final pending zero bytes */
+      do emit_byte(0x00, cinfo);
+      while (--e->zc);
+    emit_byte((e->c >> 19) & 0xFF, cinfo);
+    if (((e->c >> 19) & 0xFF) == 0xFF)
+      emit_byte(0x00, cinfo);
+    if (e->c & 0x7F800L) {
+      emit_byte((e->c >> 11) & 0xFF, cinfo);
+      if (((e->c >> 11) & 0xFF) == 0xFF)
+        emit_byte(0x00, cinfo);
+    }
+  }
+}
+
+
+/*
+ * The core arithmetic encoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Parameter 'val' to be encoded may be 0 or 1 (binary decision).
+ *
+ * Note: I've added full "Pacman" termination support to the
+ * byte output routines, which is equivalent to the optional
+ * Discard_final_zeros procedure (Figure D.15) in the spec.
+ * Thus, we always produce the shortest possible output
+ * stream compliant to the spec (no trailing zero bytes,
+ * except for FF stuffing).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(void)
+arith_encode(j_compress_ptr cinfo, unsigned char *st, int val)
+{
+  register arith_entropy_ptr e = (arith_entropy_ptr)cinfo->entropy;
+  register unsigned char nl, nm;
+  register JLONG qe, temp;
+  register int sv;
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jpeg_aritab[sv & 0x7F];  /* => Qe_Value */
+  nl = qe & 0xFF;  qe >>= 8;    /* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF;  qe >>= 8;    /* Next_Index_MPS */
+
+  /* Encode & estimation procedures per sections D.1.4 & D.1.5 */
+  e->a -= qe;
+  if (val != (sv >> 7)) {
+    /* Encode the less probable symbol */
+    if (e->a >= qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency, otherwise code the LPS
+       * as usual: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nl;     /* Estimate_after_LPS */
+  } else {
+    /* Encode the more probable symbol */
+    if (e->a >= 0x8000L)
+      return;  /* A >= 0x8000 -> ready, no renormalization required */
+    if (e->a < qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nm;     /* Estimate_after_MPS */
+  }
+
+  /* Renormalization & data output per section D.1.6 */
+  do {
+    e->a <<= 1;
+    e->c <<= 1;
+    if (--e->ct == 0) {
+      /* Another byte is ready for output */
+      temp = e->c >> 19;
+      if (temp > 0xFF) {
+        /* Handle overflow over all stacked 0xFF bytes */
+        if (e->buffer >= 0) {
+          if (e->zc)
+            do emit_byte(0x00, cinfo);
+            while (--e->zc);
+          emit_byte(e->buffer + 1, cinfo);
+          if (e->buffer + 1 == 0xFF)
+            emit_byte(0x00, cinfo);
+        }
+        e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+        e->sc = 0;
+        /* Note: The 3 spacer bits in the C register guarantee
+         * that the new buffer byte can't be 0xFF here
+         * (see page 160 in the P&M JPEG book). */
+        e->buffer = temp & 0xFF;  /* new output byte, might overflow later */
+      } else if (temp == 0xFF) {
+        ++e->sc;  /* stack 0xFF byte (which might overflow later) */
+      } else {
+        /* Output all stacked 0xFF bytes, they will not overflow any more */
+        if (e->buffer == 0)
+          ++e->zc;
+        else if (e->buffer >= 0) {
+          if (e->zc)
+            do emit_byte(0x00, cinfo);
+            while (--e->zc);
+          emit_byte(e->buffer, cinfo);
+        }
+        if (e->sc) {
+          if (e->zc)
+            do emit_byte(0x00, cinfo);
+            while (--e->zc);
+          do {
+            emit_byte(0xFF, cinfo);
+            emit_byte(0x00, cinfo);
+          } while (--e->sc);
+        }
+        e->buffer = temp & 0xFF;  /* new output byte (can still overflow) */
+      }
+      e->c &= 0x7FFFFL;
+      e->ct += 8;
+    }
+  } while (e->a < 0x8000L);
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart(j_compress_ptr cinfo, int restart_num)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  int ci;
+  jpeg_component_info *compptr;
+
+  finish_pass(cinfo);
+
+  emit_byte(0xFF, cinfo);
+  emit_byte(JPEG_RST0 + restart_num, cinfo);
+
+  /* Re-initialize statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      memset(entropy->dc_stats[compptr->dc_tbl_no], 0, DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    /* AC needs no table when not present */
+    if (cinfo->progressive_mode == 0 || cinfo->Se) {
+      memset(entropy->ac_stats[compptr->ac_tbl_no], 0, AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl;
+  int v, v2, m;
+  ISHIFT_TEMPS
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    m = IRIGHT_SHIFT((int)((*block)[0]), cinfo->Al);
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = m - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;      /* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = m;
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+        arith_encode(cinfo, st + 1, 0); /* Table F.4: SS = S0 + 1 */
+        st += 2;                        /* Table F.4: SP = S0 + 2 */
+        entropy->dc_context[ci] = 4;    /* small positive diff category */
+      } else {
+        v = -v;
+        arith_encode(cinfo, st + 1, 1); /* Table F.4: SS = S0 + 1 */
+        st += 3;                        /* Table F.4: SN = S0 + 3 */
+        entropy->dc_context[ci] = 8;    /* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+        arith_encode(cinfo, st, 1);
+        m = 1;
+        v2 = v;
+        st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+        while (v2 >>= 1) {
+          arith_encode(cinfo, st, 1);
+          m <<= 1;
+          st += 1;
+        }
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int)((1L << cinfo->arith_dc_L[tbl]) >> 1))
+        entropy->dc_context[ci] = 0;    /* zero diff category */
+      else if (m > (int)((1L << cinfo->arith_dc_U[tbl]) >> 1))
+        entropy->dc_context[ci] += 8;   /* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke;
+  int v, v2, m;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se; ke > 0; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[jpeg_natural_order[ke]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Figure F.5: Encode_AC_Coefficients */
+  for (k = cinfo->Ss; k <= ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 0);         /* EOB decision */
+    for (;;) {
+      if ((v = (*block)[jpeg_natural_order[k]]) >= 0) {
+        if (v >>= cinfo->Al) {
+          arith_encode(cinfo, st + 1, 1);
+          arith_encode(cinfo, entropy->fixed_bin, 0);
+          break;
+        }
+      } else {
+        v = -v;
+        if (v >>= cinfo->Al) {
+          arith_encode(cinfo, st + 1, 1);
+          arith_encode(cinfo, entropy->fixed_bin, 1);
+          break;
+        }
+      }
+      arith_encode(cinfo, st + 1, 0);  st += 3;  k++;
+    }
+    st += 2;
+    /* Figure F.8: Encoding the magnitude category of v */
+    m = 0;
+    if (v -= 1) {
+      arith_encode(cinfo, st, 1);
+      m = 1;
+      v2 = v;
+      if (v2 >>= 1) {
+        arith_encode(cinfo, st, 1);
+        m <<= 1;
+        st = entropy->ac_stats[tbl] +
+             (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+        while (v2 >>= 1) {
+          arith_encode(cinfo, st, 1);
+          m <<= 1;
+          st += 1;
+        }
+      }
+    }
+    arith_encode(cinfo, st, 0);
+    /* Figure F.9: Encoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      arith_encode(cinfo, st, (m & v) ? 1 : 0);
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  unsigned char *st;
+  int Al, blkn;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  st = entropy->fixed_bin;      /* use fixed probability estimation */
+  Al = cinfo->Al;
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    arith_encode(cinfo, st, (MCU_data[blkn][0][0] >> Al) & 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_refine(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke, kex;
+  int v;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Section G.1.3.3: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se; ke > 0; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[jpeg_natural_order[ke]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = ke; kex > 0; kex--)
+    if ((v = (*block)[jpeg_natural_order[kex]]) >= 0) {
+      if (v >>= cinfo->Ah) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Ah) break;
+    }
+
+  /* Figure G.10: Encode_AC_Coefficients_SA */
+  for (k = cinfo->Ss; k <= ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k > kex)
+      arith_encode(cinfo, st, 0);       /* EOB decision */
+    for (;;) {
+      if ((v = (*block)[jpeg_natural_order[k]]) >= 0) {
+        if (v >>= cinfo->Al) {
+          if (v >> 1)                   /* previously nonzero coef */
+            arith_encode(cinfo, st + 2, (v & 1));
+          else {                        /* newly nonzero coef */
+            arith_encode(cinfo, st + 1, 1);
+            arith_encode(cinfo, entropy->fixed_bin, 0);
+          }
+          break;
+        }
+      } else {
+        v = -v;
+        if (v >>= cinfo->Al) {
+          if (v >> 1)                   /* previously nonzero coef */
+            arith_encode(cinfo, st + 2, (v & 1));
+          else {                        /* newly nonzero coef */
+            arith_encode(cinfo, st + 1, 1);
+            arith_encode(cinfo, entropy->fixed_bin, 1);
+          }
+          break;
+        }
+      }
+      arith_encode(cinfo, st + 1, 0);  st += 3;  k++;
+    }
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  jpeg_component_info *compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, k, ke;
+  int v, v2, m;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = (*block)[0] - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;      /* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = (*block)[0];
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+        arith_encode(cinfo, st + 1, 0); /* Table F.4: SS = S0 + 1 */
+        st += 2;                        /* Table F.4: SP = S0 + 2 */
+        entropy->dc_context[ci] = 4;    /* small positive diff category */
+      } else {
+        v = -v;
+        arith_encode(cinfo, st + 1, 1); /* Table F.4: SS = S0 + 1 */
+        st += 3;                        /* Table F.4: SN = S0 + 3 */
+        entropy->dc_context[ci] = 8;    /* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+        arith_encode(cinfo, st, 1);
+        m = 1;
+        v2 = v;
+        st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+        while (v2 >>= 1) {
+          arith_encode(cinfo, st, 1);
+          m <<= 1;
+          st += 1;
+        }
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int)((1L << cinfo->arith_dc_L[tbl]) >> 1))
+        entropy->dc_context[ci] = 0;    /* zero diff category */
+      else if (m > (int)((1L << cinfo->arith_dc_U[tbl]) >> 1))
+        entropy->dc_context[ci] += 8;   /* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+
+    /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Establish EOB (end-of-block) index */
+    for (ke = DCTSIZE2 - 1; ke > 0; ke--)
+      if ((*block)[jpeg_natural_order[ke]]) break;
+
+    /* Figure F.5: Encode_AC_Coefficients */
+    for (k = 1; k <= ke; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 0);       /* EOB decision */
+      while ((v = (*block)[jpeg_natural_order[k]]) == 0) {
+        arith_encode(cinfo, st + 1, 0);  st += 3;  k++;
+      }
+      arith_encode(cinfo, st + 1, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+        arith_encode(cinfo, entropy->fixed_bin, 0);
+      } else {
+        v = -v;
+        arith_encode(cinfo, entropy->fixed_bin, 1);
+      }
+      st += 2;
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+        arith_encode(cinfo, st, 1);
+        m = 1;
+        v2 = v;
+        if (v2 >>= 1) {
+          arith_encode(cinfo, st, 1);
+          m <<= 1;
+          st = entropy->ac_stats[tbl] +
+               (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+          while (v2 >>= 1) {
+            arith_encode(cinfo, st, 1);
+            m <<= 1;
+            st += 1;
+          }
+        }
+      }
+      arith_encode(cinfo, st, 0);
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+    /* Encode EOB decision only if k <= DCTSIZE2 - 1 */
+    if (k <= DCTSIZE2 - 1) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 1);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass(j_compress_ptr cinfo, boolean gather_statistics)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info *compptr;
+
+  if (gather_statistics)
+    /* Make sure to avoid that in the master control logic!
+     * We are fully adaptive here and need no extra
+     * statistics gathering pass!
+     */
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  /* We assume jcmaster.c already validated the progressive scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->progressive_mode) {
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+        entropy->pub.encode_mcu = encode_mcu_DC_first;
+      else
+        entropy->pub.encode_mcu = encode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+        entropy->pub.encode_mcu = encode_mcu_DC_refine;
+      else
+        entropy->pub.encode_mcu = encode_mcu_AC_refine;
+    }
+  } else
+    entropy->pub.encode_mcu = encode_mcu;
+
+  /* Allocate & initialize requested statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+        ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+        entropy->dc_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
+          ((j_common_ptr)cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      memset(entropy->dc_stats[tbl], 0, DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    /* AC needs no table when not present */
+    if (cinfo->progressive_mode == 0 || cinfo->Se) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+        ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+        entropy->ac_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
+          ((j_common_ptr)cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      memset(entropy->ac_stats[tbl], 0, AC_STAT_BINS);
+#ifdef CALCULATE_SPECTRAL_CONDITIONING
+      if (cinfo->progressive_mode)
+        /* Section G.1.3.2: Set appropriate arithmetic conditioning value Kx */
+        cinfo->arith_ac_K[tbl] = cinfo->Ss +
+                                 ((8 + cinfo->Se - cinfo->Ss) >> 4);
+#endif
+    }
+  }
+
+  /* Initialize arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_arith_encoder(j_compress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(arith_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *)entropy;
+  entropy->pub.start_pass = start_pass;
+  entropy->pub.finish_pass = finish_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+
+  /* Initialize index for fixed probability estimation */
+  entropy->fixed_bin[0] = 113;
+}

thirdparty/libjpeg-turbo/src/jccoefct.c

@@ -0,0 +1,454 @@
+/*
+ * jccoefct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the coefficient buffer controller for compression.
+ * This controller is the top level of the lossy JPEG compressor proper.
+ * The coefficient buffer lies between forward-DCT and entropy encoding steps.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+/* We use a full-image coefficient buffer when doing Huffman optimization,
+ * and also for writing multiple-scan JPEG files.  In all cases, the DCT
+ * step is run during the first pass, and subsequent passes need only read
+ * the buffered coefficients.
+ */
+#ifdef ENTROPY_OPT_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#else
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#endif
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;      /* iMCU row # within image */
+  JDIMENSION mcu_ctr;           /* counts MCUs processed in current row */
+  int MCU_vert_offset;          /* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;    /* number of such rows needed */
+
+  /* For single-pass compression, it's sufficient to buffer just one MCU
+   * (although this may prove a bit slow in practice).  We allocate a
+   * workspace of C_MAX_BLOCKS_IN_MCU coefficient blocks, and reuse it for each
+   * MCU constructed and sent.  In multi-pass modes, this array points to the
+   * current MCU's blocks within the virtual arrays.
+   */
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+} my_coef_controller;
+
+typedef my_coef_controller *my_coef_ptr;
+
+
+/* Forward declarations */
+METHODDEF(boolean) compress_data(j_compress_ptr cinfo, _JSAMPIMAGE input_buf);
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+METHODDEF(boolean) compress_first_pass(j_compress_ptr cinfo,
+                                       _JSAMPIMAGE input_buf);
+METHODDEF(boolean) compress_output(j_compress_ptr cinfo,
+                                   _JSAMPIMAGE input_buf);
+#endif
+
+
+LOCAL(void)
+start_iMCU_row(j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows - 1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef(j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (coef->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub._compress_data = compress_data;
+    break;
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub._compress_data = compress_first_pass;
+    break;
+  case JBUF_CRANK_DEST:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub._compress_data = compress_output;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data in the single-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(boolean)
+compress_data(j_compress_ptr cinfo, _JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION MCU_col_num;       /* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, bi, ci, yindex, yoffset, blockcnt;
+  JDIMENSION ypos, xpos;
+  jpeg_component_info *compptr;
+
+  /* Loop to write as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num <= last_MCU_col;
+         MCU_col_num++) {
+      /* Determine where data comes from in input_buf and do the DCT thing.
+       * Each call on forward_DCT processes a horizontal row of DCT blocks
+       * as wide as an MCU; we rely on having allocated the MCU_buffer[] blocks
+       * sequentially.  Dummy blocks at the right or bottom edge are filled in
+       * specially.  The data in them does not matter for image reconstruction,
+       * so we fill them with values that will encode to the smallest amount of
+       * data, viz: all zeroes in the AC entries, DC entries equal to previous
+       * block's DC value.  (Thanks to Thomas Kinsman for this idea.)
+       */
+      blkn = 0;
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+        compptr = cinfo->cur_comp_info[ci];
+        blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width :
+                                                  compptr->last_col_width;
+        xpos = MCU_col_num * compptr->MCU_sample_width;
+        ypos = yoffset * DCTSIZE; /* ypos == (yoffset+yindex) * DCTSIZE */
+        for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+          if (coef->iMCU_row_num < last_iMCU_row ||
+              yoffset + yindex < compptr->last_row_height) {
+            (*cinfo->fdct->_forward_DCT) (cinfo, compptr,
+                                          input_buf[compptr->component_index],
+                                          coef->MCU_buffer[blkn],
+                                          ypos, xpos, (JDIMENSION)blockcnt);
+            if (blockcnt < compptr->MCU_width) {
+              /* Create some dummy blocks at the right edge of the image. */
+              jzero_far((void *)coef->MCU_buffer[blkn + blockcnt],
+                        (compptr->MCU_width - blockcnt) * sizeof(JBLOCK));
+              for (bi = blockcnt; bi < compptr->MCU_width; bi++) {
+                coef->MCU_buffer[blkn + bi][0][0] =
+                  coef->MCU_buffer[blkn + bi - 1][0][0];
+              }
+            }
+          } else {
+            /* Create a row of dummy blocks at the bottom of the image. */
+            jzero_far((void *)coef->MCU_buffer[blkn],
+                      compptr->MCU_width * sizeof(JBLOCK));
+            for (bi = 0; bi < compptr->MCU_width; bi++) {
+              coef->MCU_buffer[blkn + bi][0][0] =
+                coef->MCU_buffer[blkn - 1][0][0];
+            }
+          }
+          blkn += compptr->MCU_width;
+          ypos += DCTSIZE;
+        }
+      }
+      /* Try to write the MCU.  In event of a suspension failure, we will
+       * re-DCT the MCU on restart (a bit inefficient, could be fixed...)
+       */
+      if (!(*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+        /* Suspension forced; update state counters and exit */
+        coef->MCU_vert_offset = yoffset;
+        coef->mcu_ctr = MCU_col_num;
+        return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+
+/*
+ * Process some data in the first pass of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * This amount of data is read from the source buffer, DCT'd and quantized,
+ * and saved into the virtual arrays.  We also generate suitable dummy blocks
+ * as needed at the right and lower edges.  (The dummy blocks are constructed
+ * in the virtual arrays, which have been padded appropriately.)  This makes
+ * it possible for subsequent passes not to worry about real vs. dummy blocks.
+ *
+ * We must also emit the data to the entropy encoder.  This is conveniently
+ * done by calling compress_output() after we've loaded the current strip
+ * of the virtual arrays.
+ *
+ * NB: input_buf contains a plane for each component in image.  All
+ * components are DCT'd and loaded into the virtual arrays in this pass.
+ * However, it may be that only a subset of the components are emitted to
+ * the entropy encoder during this first pass; be careful about looking
+ * at the scan-dependent variables (MCU dimensions, etc).
+ */
+
+METHODDEF(boolean)
+compress_first_pass(j_compress_ptr cinfo, _JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION blocks_across, MCUs_across, MCUindex;
+  int bi, ci, h_samp_factor, block_row, block_rows, ndummy;
+  JCOEF lastDC;
+  jpeg_component_info *compptr;
+  JBLOCKARRAY buffer;
+  JBLOCKROW thisblockrow, lastblockrow;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr)cinfo, coef->whole_image[ci],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION)compptr->v_samp_factor, TRUE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (coef->iMCU_row_num < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here, since may not be set! */
+      block_rows = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    blocks_across = compptr->width_in_blocks;
+    h_samp_factor = compptr->h_samp_factor;
+    /* Count number of dummy blocks to be added at the right margin. */
+    ndummy = (int)(blocks_across % h_samp_factor);
+    if (ndummy > 0)
+      ndummy = h_samp_factor - ndummy;
+    /* Perform DCT for all non-dummy blocks in this iMCU row.  Each call
+     * on forward_DCT processes a complete horizontal row of DCT blocks.
+     */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      thisblockrow = buffer[block_row];
+      (*cinfo->fdct->_forward_DCT) (cinfo, compptr,
+                                    input_buf[ci], thisblockrow,
+                                    (JDIMENSION)(block_row * DCTSIZE),
+                                    (JDIMENSION)0, blocks_across);
+      if (ndummy > 0) {
+        /* Create dummy blocks at the right edge of the image. */
+        thisblockrow += blocks_across; /* => first dummy block */
+        jzero_far((void *)thisblockrow, ndummy * sizeof(JBLOCK));
+        lastDC = thisblockrow[-1][0];
+        for (bi = 0; bi < ndummy; bi++) {
+          thisblockrow[bi][0] = lastDC;
+        }
+      }
+    }
+    /* If at end of image, create dummy block rows as needed.
+     * The tricky part here is that within each MCU, we want the DC values
+     * of the dummy blocks to match the last real block's DC value.
+     * This squeezes a few more bytes out of the resulting file...
+     */
+    if (coef->iMCU_row_num == last_iMCU_row) {
+      blocks_across += ndummy;  /* include lower right corner */
+      MCUs_across = blocks_across / h_samp_factor;
+      for (block_row = block_rows; block_row < compptr->v_samp_factor;
+           block_row++) {
+        thisblockrow = buffer[block_row];
+        lastblockrow = buffer[block_row - 1];
+        jzero_far((void *)thisblockrow,
+                  (size_t)(blocks_across * sizeof(JBLOCK)));
+        for (MCUindex = 0; MCUindex < MCUs_across; MCUindex++) {
+          lastDC = lastblockrow[h_samp_factor - 1][0];
+          for (bi = 0; bi < h_samp_factor; bi++) {
+            thisblockrow[bi][0] = lastDC;
+          }
+          thisblockrow += h_samp_factor; /* advance to next MCU in row */
+          lastblockrow += h_samp_factor;
+        }
+      }
+    }
+  }
+  /* NB: compress_output will increment iMCU_row_num if successful.
+   * A suspension return will result in redoing all the work above next time.
+   */
+
+  /* Emit data to the entropy encoder, sharing code with subsequent passes */
+  return compress_output(cinfo, input_buf);
+}
+
+
+/*
+ * Process some data in subsequent passes of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output(j_compress_ptr cinfo, _JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION MCU_col_num;       /* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan.
+   * NB: during first pass, this is safe only because the buffers will
+   * already be aligned properly, so jmemmgr.c won't need to do any I/O.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr)cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION)compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+         MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;                 /* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+        compptr = cinfo->cur_comp_info[ci];
+        start_col = MCU_col_num * compptr->MCU_width;
+        for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+          buffer_ptr = buffer[ci][yindex + yoffset] + start_col;
+          for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+            coef->MCU_buffer[blkn++] = buffer_ptr++;
+          }
+        }
+      }
+      /* Try to write the MCU. */
+      if (!(*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+        /* Suspension forced; update state counters and exit */
+        coef->MCU_vert_offset = yoffset;
+        coef->mcu_ctr = MCU_col_num;
+        return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+#endif /* FULL_COEF_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+_jinit_c_coef_controller(j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *)coef;
+  coef->pub.start_pass = start_pass_coef;
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    int ci;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE, FALSE,
+         (JDIMENSION)jround_up((long)compptr->width_in_blocks,
+                               (long)compptr->h_samp_factor),
+         (JDIMENSION)jround_up((long)compptr->height_in_blocks,
+                               (long)compptr->v_samp_factor),
+         (JDIMENSION)compptr->v_samp_factor);
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  C_MAX_BLOCKS_IN_MCU * sizeof(JBLOCK));
+    for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->whole_image[0] = NULL; /* flag for no virtual arrays */
+  }
+}

thirdparty/libjpeg-turbo/src/jccolext.c

@@ -0,0 +1,152 @@
+/*
+ * jccolext.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009-2012, 2015, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains input colorspace conversion routines.
+ */
+
+
+/* This file is included by jccolor.c */
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ *
+ * Note that we change from the application's interleaved-pixel format
+ * to our internal noninterleaved, one-plane-per-component format.
+ * The input buffer is therefore three times as wide as the output buffer.
+ *
+ * A starting row offset is provided only for the output buffer.  The caller
+ * can easily adjust the passed input_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+INLINE
+LOCAL(void)
+rgb_ycc_convert_internal(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                         _JSAMPIMAGE output_buf, JDIMENSION output_row,
+                         int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int r, g, b;
+  register JLONG *ctab = cconvert->rgb_ycc_tab;
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr0, outptr1, outptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = RANGE_LIMIT(inptr[RGB_RED]);
+      g = RANGE_LIMIT(inptr[RGB_GREEN]);
+      b = RANGE_LIMIT(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* If the inputs are 0.._MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (_JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
+                                 ctab[b + B_Y_OFF]) >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (_JSAMPLE)((ctab[r + R_CB_OFF] + ctab[g + G_CB_OFF] +
+                                 ctab[b + B_CB_OFF]) >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (_JSAMPLE)((ctab[r + R_CR_OFF] + ctab[g + G_CR_OFF] +
+                                 ctab[b + B_CR_OFF]) >> SCALEBITS);
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/**************** Cases other than RGB -> YCbCr **************/
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles RGB->grayscale conversion, which is the same
+ * as the RGB->Y portion of RGB->YCbCr.
+ * We assume rgb_ycc_start has been called (we only use the Y tables).
+ */
+
+INLINE
+LOCAL(void)
+rgb_gray_convert_internal(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                          _JSAMPIMAGE output_buf, JDIMENSION output_row,
+                          int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int r, g, b;
+  register JLONG *ctab = cconvert->rgb_ycc_tab;
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = RANGE_LIMIT(inptr[RGB_RED]);
+      g = RANGE_LIMIT(inptr[RGB_GREEN]);
+      b = RANGE_LIMIT(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* Y */
+      outptr[col] = (_JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
+                                ctab[b + B_Y_OFF]) >> SCALEBITS);
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles extended RGB->plain RGB conversion
+ */
+
+INLINE
+LOCAL(void)
+rgb_rgb_convert_internal(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                         _JSAMPIMAGE output_buf, JDIMENSION output_row,
+                         int num_rows)
+{
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr0, outptr1, outptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      outptr0[col] = inptr[RGB_RED];
+      outptr1[col] = inptr[RGB_GREEN];
+      outptr2[col] = inptr[RGB_BLUE];
+      inptr += RGB_PIXELSIZE;
+    }
+  }
+}

thirdparty/libjpeg-turbo/src/jccolor.c

@@ -0,0 +1,755 @@
+/*
+ * jccolor.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2009-2012, 2015, 2022, 2024, D. R. Commander.
+ * Copyright (C) 2014, MIPS Technologies, Inc., California.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains input colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_converter pub; /* public fields */
+
+#if BITS_IN_JSAMPLE != 16
+  /* Private state for RGB->YCC conversion */
+  JLONG *rgb_ycc_tab;           /* => table for RGB to YCbCr conversion */
+#endif
+} my_color_converter;
+
+typedef my_color_converter *my_cconvert_ptr;
+
+
+/**************** RGB -> YCbCr conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0.._MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *      Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *      Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + _CENTERJSAMPLE
+ *      Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + _CENTERJSAMPLE
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ * Note: older versions of the IJG code used a zero offset of _MAXJSAMPLE/2,
+ * rather than _CENTERJSAMPLE, for Cb and Cr.  This gave equal positive and
+ * negative swings for Cb/Cr, but meant that grayscale values (Cb=Cr=0)
+ * were not represented exactly.  Now we sacrifice exact representation of
+ * maximum red and maximum blue in order to get exact grayscales.
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times R,G,B for all possible values.
+ * For 8-bit samples this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The _CENTERJSAMPLE offsets and the rounding fudge-factor of 0.5 are included
+ * in the tables to save adding them separately in the inner loop.
+ */
+
+#define SCALEBITS       16      /* speediest right-shift on some machines */
+#define CBCR_OFFSET     ((JLONG)_CENTERJSAMPLE << SCALEBITS)
+#define ONE_HALF        ((JLONG)1 << (SCALEBITS - 1))
+#define FIX(x)          ((JLONG)((x) * (1L << SCALEBITS) + 0.5))
+
+/* We allocate one big table and divide it up into eight parts, instead of
+ * doing eight alloc_small requests.  This lets us use a single table base
+ * address, which can be held in a register in the inner loops on many
+ * machines (more than can hold all eight addresses, anyway).
+ */
+
+#define R_Y_OFF         0                       /* offset to R => Y section */
+#define G_Y_OFF         (1 * (_MAXJSAMPLE + 1)) /* offset to G => Y section */
+#define B_Y_OFF         (2 * (_MAXJSAMPLE + 1)) /* etc. */
+#define R_CB_OFF        (3 * (_MAXJSAMPLE + 1))
+#define G_CB_OFF        (4 * (_MAXJSAMPLE + 1))
+#define B_CB_OFF        (5 * (_MAXJSAMPLE + 1))
+#define R_CR_OFF        B_CB_OFF                /* B=>Cb, R=>Cr are the same */
+#define G_CR_OFF        (6 * (_MAXJSAMPLE + 1))
+#define B_CR_OFF        (7 * (_MAXJSAMPLE + 1))
+#define TABLE_SIZE      (8 * (_MAXJSAMPLE + 1))
+
+/* 12-bit samples use a 16-bit data type, so it is possible to pass
+ * out-of-range sample values (< 0 or > 4095) to jpeg_write_scanlines().
+ * Thus, we mask the incoming 12-bit samples to guard against overrunning
+ * or underrunning the conversion tables.
+ */
+
+#if BITS_IN_JSAMPLE == 12
+#define RANGE_LIMIT(value)  ((value) & 0xFFF)
+#else
+#define RANGE_LIMIT(value)  (value)
+#endif
+
+
+/* Include inline routines for colorspace extensions */
+
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define rgb_ycc_convert_internal  extrgb_ycc_convert_internal
+#define rgb_gray_convert_internal  extrgb_gray_convert_internal
+#define rgb_rgb_convert_internal  extrgb_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define rgb_ycc_convert_internal  extrgbx_ycc_convert_internal
+#define rgb_gray_convert_internal  extrgbx_gray_convert_internal
+#define rgb_rgb_convert_internal  extrgbx_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define rgb_ycc_convert_internal  extbgr_ycc_convert_internal
+#define rgb_gray_convert_internal  extbgr_gray_convert_internal
+#define rgb_rgb_convert_internal  extbgr_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define rgb_ycc_convert_internal  extbgrx_ycc_convert_internal
+#define rgb_gray_convert_internal  extbgrx_gray_convert_internal
+#define rgb_rgb_convert_internal  extbgrx_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define rgb_ycc_convert_internal  extxbgr_ycc_convert_internal
+#define rgb_gray_convert_internal  extxbgr_gray_convert_internal
+#define rgb_rgb_convert_internal  extxbgr_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define rgb_ycc_convert_internal  extxrgb_ycc_convert_internal
+#define rgb_gray_convert_internal  extxrgb_gray_convert_internal
+#define rgb_rgb_convert_internal  extxrgb_rgb_convert_internal
+#include "jccolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef rgb_ycc_convert_internal
+#undef rgb_gray_convert_internal
+#undef rgb_rgb_convert_internal
+
+
+/*
+ * Initialize for RGB->YCC colorspace conversion.
+ */
+
+METHODDEF(void)
+rgb_ycc_start(j_compress_ptr cinfo)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  JLONG *rgb_ycc_tab;
+  JLONG i;
+
+  /* Allocate and fill in the conversion tables. */
+  cconvert->rgb_ycc_tab = rgb_ycc_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (TABLE_SIZE * sizeof(JLONG)));
+
+  for (i = 0; i <= _MAXJSAMPLE; i++) {
+    rgb_ycc_tab[i + R_Y_OFF] = FIX(0.29900) * i;
+    rgb_ycc_tab[i + G_Y_OFF] = FIX(0.58700) * i;
+    rgb_ycc_tab[i + B_Y_OFF] = FIX(0.11400) * i   + ONE_HALF;
+    rgb_ycc_tab[i + R_CB_OFF] = (-FIX(0.16874)) * i;
+    rgb_ycc_tab[i + G_CB_OFF] = (-FIX(0.33126)) * i;
+    /* We use a rounding fudge-factor of 0.5-epsilon for Cb and Cr.
+     * This ensures that the maximum output will round to _MAXJSAMPLE
+     * not _MAXJSAMPLE+1, and thus that we don't have to range-limit.
+     */
+    rgb_ycc_tab[i + B_CB_OFF] = FIX(0.50000) * i  + CBCR_OFFSET + ONE_HALF - 1;
+/*  B=>Cb and R=>Cr tables are the same
+    rgb_ycc_tab[i + R_CR_OFF] = FIX(0.50000) * i  + CBCR_OFFSET + ONE_HALF - 1;
+*/
+    rgb_ycc_tab[i + G_CR_OFF] = (-FIX(0.41869)) * i;
+    rgb_ycc_tab[i + B_CR_OFF] = (-FIX(0.08131)) * i;
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ */
+
+METHODDEF(void)
+rgb_ycc_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    extrgb_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    extrgbx_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGR:
+    extbgr_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    extbgrx_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    extxbgr_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    extxrgb_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  default:
+    rgb_ycc_convert_internal(cinfo, input_buf, output_buf, output_row,
+                             num_rows);
+    break;
+  }
+}
+
+
+/**************** Cases other than RGB -> YCbCr **************/
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ */
+
+METHODDEF(void)
+rgb_gray_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                 _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    extrgb_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    extrgbx_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                  num_rows);
+    break;
+  case JCS_EXT_BGR:
+    extbgr_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    extbgrx_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                  num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    extxbgr_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                  num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    extxrgb_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                  num_rows);
+    break;
+  default:
+    rgb_gray_convert_internal(cinfo, input_buf, output_buf, output_row,
+                              num_rows);
+    break;
+  }
+}
+
+
+/*
+ * Extended RGB to plain RGB conversion
+ */
+
+METHODDEF(void)
+rgb_rgb_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    extrgb_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    extrgbx_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGR:
+    extbgr_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    extbgrx_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    extxbgr_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    extxrgb_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                                 num_rows);
+    break;
+  default:
+    rgb_rgb_convert_internal(cinfo, input_buf, output_buf, output_row,
+                             num_rows);
+    break;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles Adobe-style CMYK->YCCK conversion,
+ * where we convert R=1-C, G=1-M, and B=1-Y to YCbCr using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume rgb_ycc_start has been called.
+ */
+
+METHODDEF(void)
+cmyk_ycck_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                  _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int r, g, b;
+  register JLONG *ctab = cconvert->rgb_ycc_tab;
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr0, outptr1, outptr2, outptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    outptr3 = output_buf[3][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = _MAXJSAMPLE - RANGE_LIMIT(inptr[0]);
+      g = _MAXJSAMPLE - RANGE_LIMIT(inptr[1]);
+      b = _MAXJSAMPLE - RANGE_LIMIT(inptr[2]);
+      /* K passes through as-is */
+      outptr3[col] = inptr[3];
+      inptr += 4;
+      /* If the inputs are 0.._MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (_JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
+                                 ctab[b + B_Y_OFF]) >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (_JSAMPLE)((ctab[r + R_CB_OFF] + ctab[g + G_CB_OFF] +
+                                 ctab[b + B_CB_OFF]) >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (_JSAMPLE)((ctab[r + R_CR_OFF] + ctab[g + G_CR_OFF] +
+                                 ctab[b + B_CR_OFF]) >> SCALEBITS);
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles grayscale output with no conversion.
+ * The source can be either plain grayscale or YCbCr (since Y == gray).
+ */
+
+METHODDEF(void)
+grayscale_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                  _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+  int instride = cinfo->input_components;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[col] = inptr[0];
+      inptr += instride;
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles multi-component colorspaces without conversion.
+ * We assume input_components == num_components.
+ */
+
+METHODDEF(void)
+null_convert(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+             _JSAMPIMAGE output_buf, JDIMENSION output_row, int num_rows)
+{
+  register _JSAMPROW inptr;
+  register _JSAMPROW outptr, outptr0, outptr1, outptr2, outptr3;
+  register JDIMENSION col;
+  register int ci;
+  int nc = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  if (nc == 3) {
+    while (--num_rows >= 0) {
+      inptr = *input_buf++;
+      outptr0 = output_buf[0][output_row];
+      outptr1 = output_buf[1][output_row];
+      outptr2 = output_buf[2][output_row];
+      output_row++;
+      for (col = 0; col < num_cols; col++) {
+        outptr0[col] = *inptr++;
+        outptr1[col] = *inptr++;
+        outptr2[col] = *inptr++;
+      }
+    }
+  } else if (nc == 4) {
+    while (--num_rows >= 0) {
+      inptr = *input_buf++;
+      outptr0 = output_buf[0][output_row];
+      outptr1 = output_buf[1][output_row];
+      outptr2 = output_buf[2][output_row];
+      outptr3 = output_buf[3][output_row];
+      output_row++;
+      for (col = 0; col < num_cols; col++) {
+        outptr0[col] = *inptr++;
+        outptr1[col] = *inptr++;
+        outptr2[col] = *inptr++;
+        outptr3[col] = *inptr++;
+      }
+    }
+  } else {
+    while (--num_rows >= 0) {
+      /* It seems fastest to make a separate pass for each component. */
+      for (ci = 0; ci < nc; ci++) {
+        inptr = *input_buf;
+        outptr = output_buf[ci][output_row];
+        for (col = 0; col < num_cols; col++) {
+          outptr[col] = inptr[ci];
+          inptr += nc;
+        }
+      }
+      input_buf++;
+      output_row++;
+    }
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+null_method(j_compress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for input colorspace conversion.
+ */
+
+GLOBAL(void)
+_jinit_color_converter(j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_color_converter));
+  cinfo->cconvert = (struct jpeg_color_converter *)cconvert;
+  /* set start_pass to null method until we find out differently */
+  cconvert->pub.start_pass = null_method;
+
+  /* Make sure input_components agrees with in_color_space */
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->input_components != 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+  case JCS_EXT_RGBA:
+  case JCS_EXT_BGRA:
+  case JCS_EXT_ABGR:
+  case JCS_EXT_ARGB:
+    if (cinfo->input_components != rgb_pixelsize[cinfo->in_color_space])
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_YCbCr:
+    if (cinfo->input_components != 3)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->input_components != 4)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  default:                      /* JCS_UNKNOWN can be anything */
+    if (cinfo->input_components < 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+  }
+
+  /* Check num_components, set conversion method based on requested space.
+   * NOTE: We do not allow any lossy color conversion algorithms in lossless
+   * mode.
+   */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->in_color_space != cinfo->jpeg_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_GRAYSCALE)
+      cconvert->pub._color_convert = grayscale_convert;
+    else if (IsExtRGB(cinfo->in_color_space)) {
+#ifdef WITH_SIMD
+      if (jsimd_can_rgb_gray())
+        cconvert->pub._color_convert = jsimd_rgb_gray_convert;
+      else
+#endif
+      {
+        cconvert->pub.start_pass = rgb_ycc_start;
+        cconvert->pub._color_convert = rgb_gray_convert;
+      }
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub._color_convert = grayscale_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless && !IsExtRGB(cinfo->in_color_space))
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (rgb_red[cinfo->in_color_space] == 0 &&
+        rgb_green[cinfo->in_color_space] == 1 &&
+        rgb_blue[cinfo->in_color_space] == 2 &&
+        rgb_pixelsize[cinfo->in_color_space] == 3) {
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_c_can_null_convert())
+        cconvert->pub._color_convert = jsimd_c_null_convert;
+      else
+#endif
+        cconvert->pub._color_convert = null_convert;
+    } else if (IsExtRGB(cinfo->in_color_space))
+      cconvert->pub._color_convert = rgb_rgb_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCbCr:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->in_color_space != cinfo->jpeg_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (IsExtRGB(cinfo->in_color_space)) {
+#ifdef WITH_SIMD
+      if (jsimd_can_rgb_ycc())
+        cconvert->pub._color_convert = jsimd_rgb_ycc_convert;
+      else
+#endif
+      {
+        cconvert->pub.start_pass = rgb_ycc_start;
+        cconvert->pub._color_convert = rgb_ycc_convert;
+      }
+    } else if (cinfo->in_color_space == JCS_YCbCr) {
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_c_can_null_convert())
+        cconvert->pub._color_convert = jsimd_c_null_convert;
+      else
+#endif
+        cconvert->pub._color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->in_color_space != cinfo->jpeg_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK) {
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_c_can_null_convert())
+        cconvert->pub._color_convert = jsimd_c_null_convert;
+      else
+#endif
+        cconvert->pub._color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCCK:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->in_color_space != cinfo->jpeg_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub._color_convert = cmyk_ycck_convert;
+    } else if (cinfo->in_color_space == JCS_YCCK) {
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_c_can_null_convert())
+        cconvert->pub._color_convert = jsimd_c_null_convert;
+      else
+#endif
+        cconvert->pub._color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:                      /* allow null conversion of JCS_UNKNOWN */
+    if (cinfo->jpeg_color_space != cinfo->in_color_space ||
+        cinfo->num_components != cinfo->input_components)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#if defined(WITH_SIMD) && defined(__mips__)
+    if (jsimd_c_can_null_convert())
+      cconvert->pub._color_convert = jsimd_c_null_convert;
+    else
+#endif
+      cconvert->pub._color_convert = null_convert;
+    break;
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jcdctmgr.c

@@ -0,0 +1,748 @@
+/*
+ * jcdctmgr.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2011, 2014-2015, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the forward-DCT management logic.
+ * This code selects a particular DCT implementation to be used,
+ * and it performs related housekeeping chores including coefficient
+ * quantization.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+#include "jsimddct.h"
+
+
+/* Private subobject for this module */
+
+typedef void (*forward_DCT_method_ptr) (DCTELEM *data);
+typedef void (*float_DCT_method_ptr) (FAST_FLOAT *data);
+
+typedef void (*convsamp_method_ptr) (_JSAMPARRAY sample_data,
+                                     JDIMENSION start_col,
+                                     DCTELEM *workspace);
+typedef void (*float_convsamp_method_ptr) (_JSAMPARRAY sample_data,
+                                           JDIMENSION start_col,
+                                           FAST_FLOAT *workspace);
+
+typedef void (*quantize_method_ptr) (JCOEFPTR coef_block, DCTELEM *divisors,
+                                     DCTELEM *workspace);
+typedef void (*float_quantize_method_ptr) (JCOEFPTR coef_block,
+                                           FAST_FLOAT *divisors,
+                                           FAST_FLOAT *workspace);
+
+METHODDEF(void) quantize(JCOEFPTR, DCTELEM *, DCTELEM *);
+
+typedef struct {
+  struct jpeg_forward_dct pub;  /* public fields */
+
+  /* Pointer to the DCT routine actually in use */
+  forward_DCT_method_ptr dct;
+  convsamp_method_ptr convsamp;
+  quantize_method_ptr quantize;
+
+  /* The actual post-DCT divisors --- not identical to the quant table
+   * entries, because of scaling (especially for an unnormalized DCT).
+   * Each table is given in normal array order.
+   */
+  DCTELEM *divisors[NUM_QUANT_TBLS];
+
+  /* work area for FDCT subroutine */
+  DCTELEM *workspace;
+
+#ifdef DCT_FLOAT_SUPPORTED
+  /* Same as above for the floating-point case. */
+  float_DCT_method_ptr float_dct;
+  float_convsamp_method_ptr float_convsamp;
+  float_quantize_method_ptr float_quantize;
+  FAST_FLOAT *float_divisors[NUM_QUANT_TBLS];
+  FAST_FLOAT *float_workspace;
+#endif
+} my_fdct_controller;
+
+typedef my_fdct_controller *my_fdct_ptr;
+
+
+#if BITS_IN_JSAMPLE == 8
+
+/*
+ * Find the highest bit in an integer through binary search.
+ */
+
+LOCAL(int)
+flss(UINT16 val)
+{
+  int bit;
+
+  bit = 16;
+
+  if (!val)
+    return 0;
+
+  if (!(val & 0xff00)) {
+    bit -= 8;
+    val <<= 8;
+  }
+  if (!(val & 0xf000)) {
+    bit -= 4;
+    val <<= 4;
+  }
+  if (!(val & 0xc000)) {
+    bit -= 2;
+    val <<= 2;
+  }
+  if (!(val & 0x8000)) {
+    bit -= 1;
+    val <<= 1;
+  }
+
+  return bit;
+}
+
+
+/*
+ * Compute values to do a division using reciprocal.
+ *
+ * This implementation is based on an algorithm described in
+ *   "Optimizing subroutines in assembly language:
+ *   An optimization guide for x86 platforms" (https://agner.org/optimize).
+ * More information about the basic algorithm can be found in
+ * the paper "Integer Division Using Reciprocals" by Robert Alverson.
+ *
+ * The basic idea is to replace x/d by x * d^-1. In order to store
+ * d^-1 with enough precision we shift it left a few places. It turns
+ * out that this algoright gives just enough precision, and also fits
+ * into DCTELEM:
+ *
+ *   b = (the number of significant bits in divisor) - 1
+ *   r = (word size) + b
+ *   f = 2^r / divisor
+ *
+ * f will not be an integer for most cases, so we need to compensate
+ * for the rounding error introduced:
+ *
+ *   no fractional part:
+ *
+ *       result = input >> r
+ *
+ *   fractional part of f < 0.5:
+ *
+ *       round f down to nearest integer
+ *       result = ((input + 1) * f) >> r
+ *
+ *   fractional part of f > 0.5:
+ *
+ *       round f up to nearest integer
+ *       result = (input * f) >> r
+ *
+ * This is the original algorithm that gives truncated results. But we
+ * want properly rounded results, so we replace "input" with
+ * "input + divisor/2".
+ *
+ * In order to allow SIMD implementations we also tweak the values to
+ * allow the same calculation to be made at all times:
+ *
+ *   dctbl[0] = f rounded to nearest integer
+ *   dctbl[1] = divisor / 2 (+ 1 if fractional part of f < 0.5)
+ *   dctbl[2] = 1 << ((word size) * 2 - r)
+ *   dctbl[3] = r - (word size)
+ *
+ * dctbl[2] is for stupid instruction sets where the shift operation
+ * isn't member wise (e.g. MMX).
+ *
+ * The reason dctbl[2] and dctbl[3] reduce the shift with (word size)
+ * is that most SIMD implementations have a "multiply and store top
+ * half" operation.
+ *
+ * Lastly, we store each of the values in their own table instead
+ * of in a consecutive manner, yet again in order to allow SIMD
+ * routines.
+ */
+
+LOCAL(int)
+compute_reciprocal(UINT16 divisor, DCTELEM *dtbl)
+{
+  UDCTELEM2 fq, fr;
+  UDCTELEM c;
+  int b, r;
+
+  if (divisor == 1) {
+    /* divisor == 1 means unquantized, so these reciprocal/correction/shift
+     * values will cause the C quantization algorithm to act like the
+     * identity function.  Since only the C quantization algorithm is used in
+     * these cases, the scale value is irrelevant.
+     */
+    dtbl[DCTSIZE2 * 0] = (DCTELEM)1;                        /* reciprocal */
+    dtbl[DCTSIZE2 * 1] = (DCTELEM)0;                        /* correction */
+    dtbl[DCTSIZE2 * 2] = (DCTELEM)1;                        /* scale */
+    dtbl[DCTSIZE2 * 3] = -(DCTELEM)(sizeof(DCTELEM) * 8);   /* shift */
+    return 0;
+  }
+
+  b = flss(divisor) - 1;
+  r  = sizeof(DCTELEM) * 8 + b;
+
+  fq = ((UDCTELEM2)1 << r) / divisor;
+  fr = ((UDCTELEM2)1 << r) % divisor;
+
+  c = divisor / 2;                      /* for rounding */
+
+  if (fr == 0) {                        /* divisor is power of two */
+    /* fq will be one bit too large to fit in DCTELEM, so adjust */
+    fq >>= 1;
+    r--;
+  } else if (fr <= (divisor / 2U)) {    /* fractional part is < 0.5 */
+    c++;
+  } else {                              /* fractional part is > 0.5 */
+    fq++;
+  }
+
+  dtbl[DCTSIZE2 * 0] = (DCTELEM)fq;     /* reciprocal */
+  dtbl[DCTSIZE2 * 1] = (DCTELEM)c;      /* correction + roundfactor */
+#ifdef WITH_SIMD
+  dtbl[DCTSIZE2 * 2] = (DCTELEM)(1 << (sizeof(DCTELEM) * 8 * 2 - r)); /* scale */
+#else
+  dtbl[DCTSIZE2 * 2] = 1;
+#endif
+  dtbl[DCTSIZE2 * 3] = (DCTELEM)r - sizeof(DCTELEM) * 8; /* shift */
+
+  if (r <= 16) return 0;
+  else return 1;
+}
+
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ * Verify that all referenced Q-tables are present, and set up
+ * the divisor table for each one.
+ * In the current implementation, DCT of all components is done during
+ * the first pass, even if only some components will be output in the
+ * first scan.  Hence all components should be examined here.
+ */
+
+METHODDEF(void)
+start_pass_fdctmgr(j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
+  int ci, qtblno, i;
+  jpeg_component_info *compptr;
+  JQUANT_TBL *qtbl;
+  DCTELEM *dtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    qtblno = compptr->quant_tbl_no;
+    /* Make sure specified quantization table is present */
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+        cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    qtbl = cinfo->quant_tbl_ptrs[qtblno];
+    /* Compute divisors for this quant table */
+    /* We may do this more than once for same table, but it's not a big deal */
+    switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+    case JDCT_ISLOW:
+      /* For LL&M IDCT method, divisors are equal to raw quantization
+       * coefficients multiplied by 8 (to counteract scaling).
+       */
+      if (fdct->divisors[qtblno] == NULL) {
+        fdct->divisors[qtblno] = (DCTELEM *)
+          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                      (DCTSIZE2 * 4) * sizeof(DCTELEM));
+      }
+      dtbl = fdct->divisors[qtblno];
+      for (i = 0; i < DCTSIZE2; i++) {
+#if BITS_IN_JSAMPLE == 8
+#ifdef WITH_SIMD
+        if (!compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i]) &&
+            fdct->quantize == jsimd_quantize)
+          fdct->quantize = quantize;
+#else
+        compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i]);
+#endif
+#else
+        dtbl[i] = ((DCTELEM)qtbl->quantval[i]) << 3;
+#endif
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+        /* For AA&N IDCT method, divisors are equal to quantization
+         * coefficients scaled by scalefactor[row]*scalefactor[col], where
+         *   scalefactor[0] = 1
+         *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+         * We apply a further scale factor of 8.
+         */
+#define CONST_BITS  14
+        static const INT16 aanscales[DCTSIZE2] = {
+          /* precomputed values scaled up by 14 bits */
+          16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+          22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+          21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+          19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+          16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+          12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+           8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+           4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+        };
+        SHIFT_TEMPS
+
+        if (fdct->divisors[qtblno] == NULL) {
+          fdct->divisors[qtblno] = (DCTELEM *)
+            (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                        (DCTSIZE2 * 4) * sizeof(DCTELEM));
+        }
+        dtbl = fdct->divisors[qtblno];
+        for (i = 0; i < DCTSIZE2; i++) {
+#if BITS_IN_JSAMPLE == 8
+#ifdef WITH_SIMD
+          if (!compute_reciprocal(
+                DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
+                                      (JLONG)aanscales[i]),
+                        CONST_BITS - 3), &dtbl[i]) &&
+              fdct->quantize == jsimd_quantize)
+            fdct->quantize = quantize;
+#else
+          compute_reciprocal(
+            DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
+                                  (JLONG)aanscales[i]),
+                    CONST_BITS-3), &dtbl[i]);
+#endif
+#else
+          dtbl[i] = (DCTELEM)
+            DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
+                                  (JLONG)aanscales[i]),
+                    CONST_BITS - 3);
+#endif
+        }
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+        /* For float AA&N IDCT method, divisors are equal to quantization
+         * coefficients scaled by scalefactor[row]*scalefactor[col], where
+         *   scalefactor[0] = 1
+         *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+         * We apply a further scale factor of 8.
+         * What's actually stored is 1/divisor so that the inner loop can
+         * use a multiplication rather than a division.
+         */
+        FAST_FLOAT *fdtbl;
+        int row, col;
+        static const double aanscalefactor[DCTSIZE] = {
+          1.0, 1.387039845, 1.306562965, 1.175875602,
+          1.0, 0.785694958, 0.541196100, 0.275899379
+        };
+
+        if (fdct->float_divisors[qtblno] == NULL) {
+          fdct->float_divisors[qtblno] = (FAST_FLOAT *)
+            (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                        DCTSIZE2 * sizeof(FAST_FLOAT));
+        }
+        fdtbl = fdct->float_divisors[qtblno];
+        i = 0;
+        for (row = 0; row < DCTSIZE; row++) {
+          for (col = 0; col < DCTSIZE; col++) {
+            fdtbl[i] = (FAST_FLOAT)
+              (1.0 / (((double)qtbl->quantval[i] *
+                       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
+            i++;
+          }
+        }
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Load data into workspace, applying unsigned->signed conversion.
+ */
+
+METHODDEF(void)
+convsamp(_JSAMPARRAY sample_data, JDIMENSION start_col, DCTELEM *workspace)
+{
+  register DCTELEM *workspaceptr;
+  register _JSAMPROW elemptr;
+  register int elemr;
+
+  workspaceptr = workspace;
+  for (elemr = 0; elemr < DCTSIZE; elemr++) {
+    elemptr = sample_data[elemr] + start_col;
+
+#if DCTSIZE == 8                /* unroll the inner loop */
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+#else
+    {
+      register int elemc;
+      for (elemc = DCTSIZE; elemc > 0; elemc--)
+        *workspaceptr++ = (*elemptr++) - _CENTERJSAMPLE;
+    }
+#endif
+  }
+}
+
+
+/*
+ * Quantize/descale the coefficients, and store into coef_blocks[].
+ */
+
+METHODDEF(void)
+quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
+{
+  int i;
+  DCTELEM temp;
+  JCOEFPTR output_ptr = coef_block;
+
+#if BITS_IN_JSAMPLE == 8
+
+  UDCTELEM recip, corr;
+  int shift;
+  UDCTELEM2 product;
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = workspace[i];
+    recip = divisors[i + DCTSIZE2 * 0];
+    corr =  divisors[i + DCTSIZE2 * 1];
+    shift = divisors[i + DCTSIZE2 * 3];
+
+    if (temp < 0) {
+      temp = -temp;
+      product = (UDCTELEM2)(temp + corr) * recip;
+      product >>= shift + sizeof(DCTELEM) * 8;
+      temp = (DCTELEM)product;
+      temp = -temp;
+    } else {
+      product = (UDCTELEM2)(temp + corr) * recip;
+      product >>= shift + sizeof(DCTELEM) * 8;
+      temp = (DCTELEM)product;
+    }
+    output_ptr[i] = (JCOEF)temp;
+  }
+
+#else
+
+  register DCTELEM qval;
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    qval = divisors[i];
+    temp = workspace[i];
+    /* Divide the coefficient value by qval, ensuring proper rounding.
+     * Since C does not specify the direction of rounding for negative
+     * quotients, we have to force the dividend positive for portability.
+     *
+     * In most files, at least half of the output values will be zero
+     * (at default quantization settings, more like three-quarters...)
+     * so we should ensure that this case is fast.  On many machines,
+     * a comparison is enough cheaper than a divide to make a special test
+     * a win.  Since both inputs will be nonnegative, we need only test
+     * for a < b to discover whether a/b is 0.
+     * If your machine's division is fast enough, define FAST_DIVIDE.
+     */
+#ifdef FAST_DIVIDE
+#define DIVIDE_BY(a, b)  a /= b
+#else
+#define DIVIDE_BY(a, b)  if (a >= b) a /= b;  else a = 0
+#endif
+    if (temp < 0) {
+      temp = -temp;
+      temp += qval >> 1;        /* for rounding */
+      DIVIDE_BY(temp, qval);
+      temp = -temp;
+    } else {
+      temp += qval >> 1;        /* for rounding */
+      DIVIDE_BY(temp, qval);
+    }
+    output_ptr[i] = (JCOEF)temp;
+  }
+
+#endif
+
+}
+
+
+/*
+ * Perform forward DCT on one or more blocks of a component.
+ *
+ * The input samples are taken from the sample_data[] array starting at
+ * position start_row/start_col, and moving to the right for any additional
+ * blocks. The quantized coefficients are returned in coef_blocks[].
+ */
+
+METHODDEF(void)
+forward_DCT(j_compress_ptr cinfo, jpeg_component_info *compptr,
+            _JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+            JDIMENSION start_row, JDIMENSION start_col, JDIMENSION num_blocks)
+/* This version is used for integer DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
+  DCTELEM *divisors = fdct->divisors[compptr->quant_tbl_no];
+  DCTELEM *workspace;
+  JDIMENSION bi;
+
+  /* Make sure the compiler doesn't look up these every pass */
+  forward_DCT_method_ptr do_dct = fdct->dct;
+  convsamp_method_ptr do_convsamp = fdct->convsamp;
+  quantize_method_ptr do_quantize = fdct->quantize;
+  workspace = fdct->workspace;
+
+  sample_data += start_row;     /* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    (*do_convsamp) (sample_data, start_col, workspace);
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    (*do_quantize) (coef_blocks[bi], divisors, workspace);
+  }
+}
+
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+METHODDEF(void)
+convsamp_float(_JSAMPARRAY sample_data, JDIMENSION start_col,
+               FAST_FLOAT *workspace)
+{
+  register FAST_FLOAT *workspaceptr;
+  register _JSAMPROW elemptr;
+  register int elemr;
+
+  workspaceptr = workspace;
+  for (elemr = 0; elemr < DCTSIZE; elemr++) {
+    elemptr = sample_data[elemr] + start_col;
+#if DCTSIZE == 8                /* unroll the inner loop */
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+#else
+    {
+      register int elemc;
+      for (elemc = DCTSIZE; elemc > 0; elemc--)
+        *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - _CENTERJSAMPLE);
+    }
+#endif
+  }
+}
+
+
+METHODDEF(void)
+quantize_float(JCOEFPTR coef_block, FAST_FLOAT *divisors,
+               FAST_FLOAT *workspace)
+{
+  register FAST_FLOAT temp;
+  register int i;
+  register JCOEFPTR output_ptr = coef_block;
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    /* Apply the quantization and scaling factor */
+    temp = workspace[i] * divisors[i];
+
+    /* Round to nearest integer.
+     * Since C does not specify the direction of rounding for negative
+     * quotients, we have to force the dividend positive for portability.
+     * The maximum coefficient size is +-16K (for 12-bit data), so this
+     * code should work for either 16-bit or 32-bit ints.
+     */
+    output_ptr[i] = (JCOEF)((int)(temp + (FAST_FLOAT)16384.5) - 16384);
+  }
+}
+
+
+METHODDEF(void)
+forward_DCT_float(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                  _JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+                  JDIMENSION start_row, JDIMENSION start_col,
+                  JDIMENSION num_blocks)
+/* This version is used for floating-point DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
+  FAST_FLOAT *divisors = fdct->float_divisors[compptr->quant_tbl_no];
+  FAST_FLOAT *workspace;
+  JDIMENSION bi;
+
+
+  /* Make sure the compiler doesn't look up these every pass */
+  float_DCT_method_ptr do_dct = fdct->float_dct;
+  float_convsamp_method_ptr do_convsamp = fdct->float_convsamp;
+  float_quantize_method_ptr do_quantize = fdct->float_quantize;
+  workspace = fdct->float_workspace;
+
+  sample_data += start_row;     /* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    (*do_convsamp) (sample_data, start_col, workspace);
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    (*do_quantize) (coef_blocks[bi], divisors, workspace);
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
+
+
+/*
+ * Initialize FDCT manager.
+ */
+
+GLOBAL(void)
+_jinit_forward_dct(j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct;
+  int i;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  fdct = (my_fdct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_fdct_controller));
+  cinfo->fdct = (struct jpeg_forward_dct *)fdct;
+  fdct->pub.start_pass = start_pass_fdctmgr;
+
+  /* First determine the DCT... */
+  switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+  case JDCT_ISLOW:
+    fdct->pub._forward_DCT = forward_DCT;
+#ifdef WITH_SIMD
+    if (jsimd_can_fdct_islow())
+      fdct->dct = jsimd_fdct_islow;
+    else
+#endif
+      fdct->dct = _jpeg_fdct_islow;
+    break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  case JDCT_IFAST:
+    fdct->pub._forward_DCT = forward_DCT;
+#ifdef WITH_SIMD
+    if (jsimd_can_fdct_ifast())
+      fdct->dct = jsimd_fdct_ifast;
+    else
+#endif
+      fdct->dct = _jpeg_fdct_ifast;
+    break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  case JDCT_FLOAT:
+    fdct->pub._forward_DCT = forward_DCT_float;
+#ifdef WITH_SIMD
+    if (jsimd_can_fdct_float())
+      fdct->float_dct = jsimd_fdct_float;
+    else
+#endif
+      fdct->float_dct = jpeg_fdct_float;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+
+  /* ...then the supporting stages. */
+  switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+  case JDCT_ISLOW:
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  case JDCT_IFAST:
+#endif
+#if defined(DCT_ISLOW_SUPPORTED) || defined(DCT_IFAST_SUPPORTED)
+#ifdef WITH_SIMD
+    if (jsimd_can_convsamp())
+      fdct->convsamp = jsimd_convsamp;
+    else
+#endif
+      fdct->convsamp = convsamp;
+#ifdef WITH_SIMD
+    if (jsimd_can_quantize())
+      fdct->quantize = jsimd_quantize;
+    else
+#endif
+      fdct->quantize = quantize;
+    break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  case JDCT_FLOAT:
+#ifdef WITH_SIMD
+    if (jsimd_can_convsamp_float())
+      fdct->float_convsamp = jsimd_convsamp_float;
+    else
+#endif
+      fdct->float_convsamp = convsamp_float;
+#ifdef WITH_SIMD
+    if (jsimd_can_quantize_float())
+      fdct->float_quantize = jsimd_quantize_float;
+    else
+#endif
+      fdct->float_quantize = quantize_float;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+
+  /* Allocate workspace memory */
+#ifdef DCT_FLOAT_SUPPORTED
+  if (cinfo->dct_method == JDCT_FLOAT)
+    fdct->float_workspace = (FAST_FLOAT *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(FAST_FLOAT) * DCTSIZE2);
+  else
+#endif
+    fdct->workspace = (DCTELEM *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(DCTELEM) * DCTSIZE2);
+
+  /* Mark divisor tables unallocated */
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    fdct->divisors[i] = NULL;
+#ifdef DCT_FLOAT_SUPPORTED
+    fdct->float_divisors[i] = NULL;
+#endif
+  }
+}

thirdparty/libjpeg-turbo/src/jchuff.c

@@ -0,0 +1,1175 @@
+/*
+ * jchuff.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009-2011, 2014-2016, 2018-2024, D. R. Commander.
+ * Copyright (C) 2015, Matthieu Darbois.
+ * Copyright (C) 2018, Matthias Räncker.
+ * Copyright (C) 2020, Arm Limited.
+ * Copyright (C) 2022, Felix Hanau.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains Huffman entropy encoding routines.
+ *
+ * Much of the complexity here has to do with supporting output suspension.
+ * If the data destination module demands suspension, we want to be able to
+ * back up to the start of the current MCU.  To do this, we copy state
+ * variables into local working storage, and update them back to the
+ * permanent JPEG objects only upon successful completion of an MCU.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#ifdef WITH_SIMD
+#include "jsimd.h"
+#else
+#include "jchuff.h"             /* Declarations shared with jc*huff.c */
+#endif
+#include <limits.h>
+#include "jpeg_nbits.h"
+
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+#if defined(__x86_64__) && defined(__ILP32__)
+typedef unsigned long long bit_buf_type;
+#else
+typedef size_t bit_buf_type;
+#endif
+
+/* NOTE: The more optimal Huffman encoding algorithm is only used by the
+ * intrinsics implementation of the Arm Neon SIMD extensions, which is why we
+ * retain the old Huffman encoder behavior when using the GAS implementation.
+ */
+#if defined(WITH_SIMD) && !(defined(__arm__) || defined(__aarch64__) || \
+                            defined(_M_ARM) || defined(_M_ARM64))
+typedef unsigned long long simd_bit_buf_type;
+#else
+typedef bit_buf_type simd_bit_buf_type;
+#endif
+
+#if (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 8) || defined(_WIN64) || \
+    (defined(__x86_64__) && defined(__ILP32__))
+#define BIT_BUF_SIZE  64
+#elif (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 4) || defined(_WIN32)
+#define BIT_BUF_SIZE  32
+#else
+#error Cannot determine word size
+#endif
+#define SIMD_BIT_BUF_SIZE  (sizeof(simd_bit_buf_type) * 8)
+
+typedef struct {
+  union {
+    bit_buf_type c;
+#ifdef WITH_SIMD
+    simd_bit_buf_type simd;
+#endif
+  } put_buffer;                         /* current bit accumulation buffer */
+  int free_bits;                        /* # of bits available in it */
+                                        /* (Neon GAS: # of bits now in it) */
+  int last_dc_val[MAX_COMPS_IN_SCAN];   /* last DC coef for each component */
+} savable_state;
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  savable_state saved;          /* Bit buffer & DC state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+  int next_restart_num;         /* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  c_derived_tbl *dc_derived_tbls[NUM_HUFF_TBLS];
+  c_derived_tbl *ac_derived_tbls[NUM_HUFF_TBLS];
+
+#ifdef ENTROPY_OPT_SUPPORTED    /* Statistics tables for optimization */
+  long *dc_count_ptrs[NUM_HUFF_TBLS];
+  long *ac_count_ptrs[NUM_HUFF_TBLS];
+#endif
+
+#ifdef WITH_SIMD
+  int simd;
+#endif
+} huff_entropy_encoder;
+
+typedef huff_entropy_encoder *huff_entropy_ptr;
+
+/* Working state while writing an MCU.
+ * This struct contains all the fields that are needed by subroutines.
+ */
+
+typedef struct {
+  JOCTET *next_output_byte;     /* => next byte to write in buffer */
+  size_t free_in_buffer;        /* # of byte spaces remaining in buffer */
+  savable_state cur;            /* Current bit buffer & DC state */
+  j_compress_ptr cinfo;         /* dump_buffer needs access to this */
+#ifdef WITH_SIMD
+  int simd;
+#endif
+} working_state;
+
+
+/* Forward declarations */
+METHODDEF(boolean) encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data);
+METHODDEF(void) finish_pass_huff(j_compress_ptr cinfo);
+#ifdef ENTROPY_OPT_SUPPORTED
+METHODDEF(boolean) encode_mcu_gather(j_compress_ptr cinfo,
+                                     JBLOCKROW *MCU_data);
+METHODDEF(void) finish_pass_gather(j_compress_ptr cinfo);
+#endif
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ * If gather_statistics is TRUE, we do not output anything during the scan,
+ * just count the Huffman symbols used and generate Huffman code tables.
+ */
+
+METHODDEF(void)
+start_pass_huff(j_compress_ptr cinfo, boolean gather_statistics)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info *compptr;
+
+  if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->pub.encode_mcu = encode_mcu_gather;
+    entropy->pub.finish_pass = finish_pass_gather;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    entropy->pub.encode_mcu = encode_mcu_huff;
+    entropy->pub.finish_pass = finish_pass_huff;
+  }
+
+#ifdef WITH_SIMD
+  entropy->simd = jsimd_can_huff_encode_one_block();
+#endif
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+      /* Check for invalid table indexes */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
+      if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->dc_count_ptrs[dctbl] == NULL)
+        entropy->dc_count_ptrs[dctbl] = (long *)
+          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                      257 * sizeof(long));
+      memset(entropy->dc_count_ptrs[dctbl], 0, 257 * sizeof(long));
+      if (entropy->ac_count_ptrs[actbl] == NULL)
+        entropy->ac_count_ptrs[actbl] = (long *)
+          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                      257 * sizeof(long));
+      memset(entropy->ac_count_ptrs[actbl], 0, 257 * sizeof(long));
+#endif
+    } else {
+      /* Compute derived values for Huffman tables */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
+                              &entropy->dc_derived_tbls[dctbl]);
+      jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
+                              &entropy->ac_derived_tbls[actbl]);
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bit buffer to empty */
+#ifdef WITH_SIMD
+  if (entropy->simd) {
+    entropy->saved.put_buffer.simd = 0;
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+    entropy->saved.free_bits = 0;
+#else
+    entropy->saved.free_bits = SIMD_BIT_BUF_SIZE;
+#endif
+  } else
+#endif
+  {
+    entropy->saved.put_buffer.c = 0;
+    entropy->saved.free_bits = BIT_BUF_SIZE;
+  }
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jcphuff.c and jclhuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_c_derived_tbl(j_compress_ptr cinfo, boolean isDC, int tblno,
+                        c_derived_tbl **pdtbl)
+{
+  JHUFF_TBL *htbl;
+  c_derived_tbl *dtbl;
+  int p, i, l, lastp, si, maxsymbol;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (c_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(c_derived_tbl));
+  dtbl = *pdtbl;
+
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int)htbl->bits[l];
+    if (i < 0 || p + i > 256)   /* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char)l;
+  }
+  huffsize[p] = 0;
+  lastp = p;
+
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int)huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((JLONG)code) >= (((JLONG)1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+
+  /* Figure C.3: generate encoding tables */
+  /* These are code and size indexed by symbol value */
+
+  /* Set all codeless symbols to have code length 0;
+   * this lets us detect duplicate VAL entries here, and later
+   * allows emit_bits to detect any attempt to emit such symbols.
+   */
+  memset(dtbl->ehufco, 0, sizeof(dtbl->ehufco));
+  memset(dtbl->ehufsi, 0, sizeof(dtbl->ehufsi));
+
+  /* This is also a convenient place to check for out-of-range and duplicated
+   * VAL entries.  We allow 0..255 for AC symbols but only 0..15 for DC in
+   * lossy mode and 0..16 for DC in lossless mode.  (We could constrain them
+   * further based on data depth and mode, but this seems enough.)
+   */
+  maxsymbol = isDC ? (cinfo->master->lossless ? 16 : 15) : 255;
+
+  for (p = 0; p < lastp; p++) {
+    i = htbl->huffval[p];
+    if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    dtbl->ehufco[i] = huffcode[p];
+    dtbl->ehufsi[i] = huffsize[p];
+  }
+}
+
+
+/* Outputting bytes to the file */
+
+/* Emit a byte, taking 'action' if must suspend. */
+#define emit_byte(state, val, action) { \
+  *(state)->next_output_byte++ = (JOCTET)(val); \
+  if (--(state)->free_in_buffer == 0) \
+    if (!dump_buffer(state)) \
+      { action; } \
+}
+
+
+LOCAL(boolean)
+dump_buffer(working_state *state)
+/* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
+{
+  struct jpeg_destination_mgr *dest = state->cinfo->dest;
+
+  if (!(*dest->empty_output_buffer) (state->cinfo))
+    return FALSE;
+  /* After a successful buffer dump, must reset buffer pointers */
+  state->next_output_byte = dest->next_output_byte;
+  state->free_in_buffer = dest->free_in_buffer;
+  return TRUE;
+}
+
+
+/* Outputting bits to the file */
+
+/* Output byte b and, speculatively, an additional 0 byte.  0xFF must be
+ * encoded as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the
+ * byte is 0xFF.  Otherwise, the output buffer pointer is advanced by 1, and
+ * the speculative 0 byte will be overwritten by the next byte.
+ */
+#define EMIT_BYTE(b) { \
+  buffer[0] = (JOCTET)(b); \
+  buffer[1] = 0; \
+  buffer -= -2 + ((JOCTET)(b) < 0xFF); \
+}
+
+/* Output the entire bit buffer.  If there are no 0xFF bytes in it, then write
+ * directly to the output buffer.  Otherwise, use the EMIT_BYTE() macro to
+ * encode 0xFF as 0xFF 0x00.
+ */
+#if BIT_BUF_SIZE == 64
+
+#define FLUSH() { \
+  if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \
+    EMIT_BYTE(put_buffer >> 56) \
+    EMIT_BYTE(put_buffer >> 48) \
+    EMIT_BYTE(put_buffer >> 40) \
+    EMIT_BYTE(put_buffer >> 32) \
+    EMIT_BYTE(put_buffer >> 24) \
+    EMIT_BYTE(put_buffer >> 16) \
+    EMIT_BYTE(put_buffer >>  8) \
+    EMIT_BYTE(put_buffer      ) \
+  } else { \
+    buffer[0] = (JOCTET)(put_buffer >> 56); \
+    buffer[1] = (JOCTET)(put_buffer >> 48); \
+    buffer[2] = (JOCTET)(put_buffer >> 40); \
+    buffer[3] = (JOCTET)(put_buffer >> 32); \
+    buffer[4] = (JOCTET)(put_buffer >> 24); \
+    buffer[5] = (JOCTET)(put_buffer >> 16); \
+    buffer[6] = (JOCTET)(put_buffer >> 8); \
+    buffer[7] = (JOCTET)(put_buffer); \
+    buffer += 8; \
+  } \
+}
+
+#else
+
+#define FLUSH() { \
+  if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \
+    EMIT_BYTE(put_buffer >> 24) \
+    EMIT_BYTE(put_buffer >> 16) \
+    EMIT_BYTE(put_buffer >>  8) \
+    EMIT_BYTE(put_buffer      ) \
+  } else { \
+    buffer[0] = (JOCTET)(put_buffer >> 24); \
+    buffer[1] = (JOCTET)(put_buffer >> 16); \
+    buffer[2] = (JOCTET)(put_buffer >> 8); \
+    buffer[3] = (JOCTET)(put_buffer); \
+    buffer += 4; \
+  } \
+}
+
+#endif
+
+/* Fill the bit buffer to capacity with the leading bits from code, then output
+ * the bit buffer and put the remaining bits from code into the bit buffer.
+ */
+#define PUT_AND_FLUSH(code, size) { \
+  put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \
+  FLUSH() \
+  free_bits += BIT_BUF_SIZE; \
+  put_buffer = code; \
+}
+
+/* Insert code into the bit buffer and output the bit buffer if needed.
+ * NOTE: We can't flush with free_bits == 0, since the left shift in
+ * PUT_AND_FLUSH() would have undefined behavior.
+ */
+#define PUT_BITS(code, size) { \
+  free_bits -= size; \
+  if (free_bits < 0) \
+    PUT_AND_FLUSH(code, size) \
+  else \
+    put_buffer = (put_buffer << size) | code; \
+}
+
+#define PUT_CODE(code, size) { \
+  temp &= (((JLONG)1) << nbits) - 1; \
+  temp |= code << nbits; \
+  nbits += size; \
+  PUT_BITS(temp, nbits) \
+}
+
+
+/* Although it is exceedingly rare, it is possible for a Huffman-encoded
+ * coefficient block to be larger than the 128-byte unencoded block.  For each
+ * of the 64 coefficients, PUT_BITS is invoked twice, and each invocation can
+ * theoretically store 16 bits (for a maximum of 2048 bits or 256 bytes per
+ * encoded block.)  If, for instance, one artificially sets the AC
+ * coefficients to alternating values of 32767 and -32768 (using the JPEG
+ * scanning order-- 1, 8, 16, etc.), then this will produce an encoded block
+ * larger than 200 bytes.
+ */
+#define BUFSIZE  (DCTSIZE2 * 8)
+
+#define LOAD_BUFFER() { \
+  if (state->free_in_buffer < BUFSIZE) { \
+    localbuf = 1; \
+    buffer = _buffer; \
+  } else \
+    buffer = state->next_output_byte; \
+}
+
+#define STORE_BUFFER() { \
+  if (localbuf) { \
+    size_t bytes, bytestocopy; \
+    bytes = buffer - _buffer; \
+    buffer = _buffer; \
+    while (bytes > 0) { \
+      bytestocopy = MIN(bytes, state->free_in_buffer); \
+      memcpy(state->next_output_byte, buffer, bytestocopy); \
+      state->next_output_byte += bytestocopy; \
+      buffer += bytestocopy; \
+      state->free_in_buffer -= bytestocopy; \
+      if (state->free_in_buffer == 0) \
+        if (!dump_buffer(state)) return FALSE; \
+      bytes -= bytestocopy; \
+    } \
+  } else { \
+    state->free_in_buffer -= (buffer - state->next_output_byte); \
+    state->next_output_byte = buffer; \
+  } \
+}
+
+
+LOCAL(boolean)
+flush_bits(working_state *state)
+{
+  JOCTET _buffer[BUFSIZE], *buffer, temp;
+  simd_bit_buf_type put_buffer;  int put_bits;
+  int localbuf = 0;
+
+#ifdef WITH_SIMD
+  if (state->simd) {
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+    put_bits = state->cur.free_bits;
+#else
+    put_bits = SIMD_BIT_BUF_SIZE - state->cur.free_bits;
+#endif
+    put_buffer = state->cur.put_buffer.simd;
+  } else
+#endif
+  {
+    put_bits = BIT_BUF_SIZE - state->cur.free_bits;
+    put_buffer = state->cur.put_buffer.c;
+  }
+
+  LOAD_BUFFER()
+
+  while (put_bits >= 8) {
+    put_bits -= 8;
+    temp = (JOCTET)(put_buffer >> put_bits);
+    EMIT_BYTE(temp)
+  }
+  if (put_bits) {
+    /* fill partial byte with ones */
+    temp = (JOCTET)((put_buffer << (8 - put_bits)) | (0xFF >> put_bits));
+    EMIT_BYTE(temp)
+  }
+
+#ifdef WITH_SIMD
+  if (state->simd) {                    /* and reset bit buffer to empty */
+    state->cur.put_buffer.simd = 0;
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+    state->cur.free_bits = 0;
+#else
+    state->cur.free_bits = SIMD_BIT_BUF_SIZE;
+#endif
+  } else
+#endif
+  {
+    state->cur.put_buffer.c = 0;
+    state->cur.free_bits = BIT_BUF_SIZE;
+  }
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
+
+#ifdef WITH_SIMD
+
+/* Encode a single block's worth of coefficients */
+
+LOCAL(boolean)
+encode_one_block_simd(working_state *state, JCOEFPTR block, int last_dc_val,
+                      c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  JOCTET _buffer[BUFSIZE], *buffer;
+  int localbuf = 0;
+
+#ifdef ZERO_BUFFERS
+  memset(_buffer, 0, sizeof(_buffer));
+#endif
+
+  LOAD_BUFFER()
+
+  buffer = jsimd_huff_encode_one_block(state, buffer, block, last_dc_val,
+                                       dctbl, actbl);
+
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
+#endif
+
+LOCAL(boolean)
+encode_one_block(working_state *state, JCOEFPTR block, int last_dc_val,
+                 c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  int temp, nbits, free_bits;
+  bit_buf_type put_buffer;
+  JOCTET _buffer[BUFSIZE], *buffer;
+  int localbuf = 0;
+  int max_coef_bits = state->cinfo->data_precision + 2;
+
+  free_bits = state->cur.free_bits;
+  put_buffer = state->cur.put_buffer.c;
+  LOAD_BUFFER()
+
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+
+  temp = block[0] - last_dc_val;
+
+  /* This is a well-known technique for obtaining the absolute value without a
+   * branch.  It is derived from an assembly language technique presented in
+   * "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by
+   * Agner Fog.  This code assumes we are on a two's complement machine.
+   */
+  nbits = temp >> (CHAR_BIT * sizeof(int) - 1);
+  temp += nbits;
+  nbits ^= temp;
+
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = JPEG_NBITS(nbits);
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > max_coef_bits + 1)
+    ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
+
+  /* Emit the Huffman-coded symbol for the number of bits.
+   * Emit that number of bits of the value, if positive,
+   * or the complement of its magnitude, if negative.
+   */
+  PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits])
+
+  /* Encode the AC coefficients per section F.1.2.2 */
+
+  {
+    int r = 0;                  /* r = run length of zeros */
+
+/* Manually unroll the k loop to eliminate the counter variable.  This
+ * improves performance greatly on systems with a limited number of
+ * registers (such as x86.)
+ */
+#define kloop(jpeg_natural_order_of_k) { \
+  if ((temp = block[jpeg_natural_order_of_k]) == 0) { \
+    r += 16; \
+  } else { \
+    /* Branch-less absolute value, bitwise complement, etc., same as above */ \
+    nbits = temp >> (CHAR_BIT * sizeof(int) - 1); \
+    temp += nbits; \
+    nbits ^= temp; \
+    nbits = JPEG_NBITS_NONZERO(nbits); \
+    /* Check for out-of-range coefficient values */ \
+    if (nbits > max_coef_bits) \
+      ERREXIT(state->cinfo, JERR_BAD_DCT_COEF); \
+    /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
+    while (r >= 16 * 16) { \
+      r -= 16 * 16; \
+      PUT_BITS(actbl->ehufco[0xf0], actbl->ehufsi[0xf0]) \
+    } \
+    /* Emit Huffman symbol for run length / number of bits */ \
+    r += nbits; \
+    PUT_CODE(actbl->ehufco[r], actbl->ehufsi[r]) \
+    r = 0; \
+  } \
+}
+
+    /* One iteration for each value in jpeg_natural_order[] */
+    kloop(1);   kloop(8);   kloop(16);  kloop(9);   kloop(2);   kloop(3);
+    kloop(10);  kloop(17);  kloop(24);  kloop(32);  kloop(25);  kloop(18);
+    kloop(11);  kloop(4);   kloop(5);   kloop(12);  kloop(19);  kloop(26);
+    kloop(33);  kloop(40);  kloop(48);  kloop(41);  kloop(34);  kloop(27);
+    kloop(20);  kloop(13);  kloop(6);   kloop(7);   kloop(14);  kloop(21);
+    kloop(28);  kloop(35);  kloop(42);  kloop(49);  kloop(56);  kloop(57);
+    kloop(50);  kloop(43);  kloop(36);  kloop(29);  kloop(22);  kloop(15);
+    kloop(23);  kloop(30);  kloop(37);  kloop(44);  kloop(51);  kloop(58);
+    kloop(59);  kloop(52);  kloop(45);  kloop(38);  kloop(31);  kloop(39);
+    kloop(46);  kloop(53);  kloop(60);  kloop(61);  kloop(54);  kloop(47);
+    kloop(55);  kloop(62);  kloop(63);
+
+    /* If the last coef(s) were zero, emit an end-of-block code */
+    if (r > 0) {
+      PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+    }
+  }
+
+  state->cur.put_buffer.c = put_buffer;
+  state->cur.free_bits = free_bits;
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(boolean)
+emit_restart(working_state *state, int restart_num)
+{
+  int ci;
+
+  if (!flush_bits(state))
+    return FALSE;
+
+  emit_byte(state, 0xFF, return FALSE);
+  emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
+    state->cur.last_dc_val[ci] = 0;
+
+  /* The restart counter is not updated until we successfully write the MCU. */
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of Huffman-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  working_state state;
+  int blkn, ci;
+  jpeg_component_info *compptr;
+
+  /* Load up working state */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  state.cur = entropy->saved;
+  state.cinfo = cinfo;
+#ifdef WITH_SIMD
+  state.simd = entropy->simd;
+#endif
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!emit_restart(&state, entropy->next_restart_num))
+        return FALSE;
+  }
+
+  /* Encode the MCU data blocks */
+#ifdef WITH_SIMD
+  if (entropy->simd) {
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      if (!encode_one_block_simd(&state,
+                                 MCU_data[blkn][0], state.cur.last_dc_val[ci],
+                                 entropy->dc_derived_tbls[compptr->dc_tbl_no],
+                                 entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+        return FALSE;
+      /* Update last_dc_val */
+      state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+    }
+  } else
+#endif
+  {
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      if (!encode_one_block(&state,
+                            MCU_data[blkn][0], state.cur.last_dc_val[ci],
+                            entropy->dc_derived_tbls[compptr->dc_tbl_no],
+                            entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+        return FALSE;
+      /* Update last_dc_val */
+      state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+    }
+  }
+
+  /* Completed MCU, so update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  entropy->saved = state.cur;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass_huff(j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  working_state state;
+
+  /* Load up working state ... flush_bits needs it */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  state.cur = entropy->saved;
+  state.cinfo = cinfo;
+#ifdef WITH_SIMD
+  state.simd = entropy->simd;
+#endif
+
+  /* Flush out the last data */
+  if (!flush_bits(&state))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  /* Update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  entropy->saved = state.cur;
+}
+
+
+/*
+ * Huffman coding optimization.
+ *
+ * We first scan the supplied data and count the number of uses of each symbol
+ * that is to be Huffman-coded. (This process MUST agree with the code above.)
+ * Then we build a Huffman coding tree for the observed counts.
+ * Symbols which are not needed at all for the particular image are not
+ * assigned any code, which saves space in the DHT marker as well as in
+ * the compressed data.
+ */
+
+#ifdef ENTROPY_OPT_SUPPORTED
+
+
+/* Process a single block's worth of coefficients */
+
+LOCAL(void)
+htest_one_block(j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
+                long dc_counts[], long ac_counts[])
+{
+  register int temp;
+  register int nbits;
+  register int k, r;
+  int max_coef_bits = cinfo->data_precision + 2;
+
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+
+  temp = block[0] - last_dc_val;
+  if (temp < 0)
+    temp = -temp;
+
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > max_coef_bits + 1)
+    ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+  /* Count the Huffman symbol for the number of bits */
+  dc_counts[nbits]++;
+
+  /* Encode the AC coefficients per section F.1.2.2 */
+
+  r = 0;                        /* r = run length of zeros */
+
+  for (k = 1; k < DCTSIZE2; k++) {
+    if ((temp = block[jpeg_natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+        ac_counts[0xF0]++;
+        r -= 16;
+      }
+
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      if (temp < 0)
+        temp = -temp;
+
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;                /* there must be at least one 1 bit */
+      while ((temp >>= 1))
+        nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > max_coef_bits)
+        ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+      /* Count Huffman symbol for run length / number of bits */
+      ac_counts[(r << 4) + nbits]++;
+
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    ac_counts[0]++;
+}
+
+
+/*
+ * Trial-encode one MCU's worth of Huffman-compressed coefficients.
+ * No data is actually output, so no suspension return is possible.
+ */
+
+METHODDEF(boolean)
+encode_mcu_gather(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int blkn, ci;
+  jpeg_component_info *compptr;
+
+  /* Take care of restart intervals if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      /* Re-initialize DC predictions to 0 */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+        entropy->saved.last_dc_val[ci] = 0;
+      /* Update restart state */
+      entropy->restarts_to_go = cinfo->restart_interval;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
+                    entropy->dc_count_ptrs[compptr->dc_tbl_no],
+                    entropy->ac_count_ptrs[compptr->ac_tbl_no]);
+    entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Generate the best Huffman code table for the given counts, fill htbl.
+ * Note this is also used by jcphuff.c and jclhuff.c.
+ *
+ * The JPEG standard requires that no symbol be assigned a codeword of all
+ * one bits (so that padding bits added at the end of a compressed segment
+ * can't look like a valid code).  Because of the canonical ordering of
+ * codewords, this just means that there must be an unused slot in the
+ * longest codeword length category.  Annex K (Clause K.2) of
+ * Rec. ITU-T T.81 (1992) | ISO/IEC 10918-1:1994 suggests reserving such a slot
+ * by pretending that symbol 256 is a valid symbol with count 1.  In theory
+ * that's not optimal; giving it count zero but including it in the symbol set
+ * anyway should give a better Huffman code.  But the theoretically better code
+ * actually seems to come out worse in practice, because it produces more
+ * all-ones bytes (which incur stuffed zero bytes in the final file).  In any
+ * case the difference is tiny.
+ *
+ * The JPEG standard requires Huffman codes to be no more than 16 bits long.
+ * If some symbols have a very small but nonzero probability, the Huffman tree
+ * must be adjusted to meet the code length restriction.  We currently use
+ * the adjustment method suggested in JPEG section K.2.  This method is *not*
+ * optimal; it may not choose the best possible limited-length code.  But
+ * typically only very-low-frequency symbols will be given less-than-optimal
+ * lengths, so the code is almost optimal.  Experimental comparisons against
+ * an optimal limited-length-code algorithm indicate that the difference is
+ * microscopic --- usually less than a hundredth of a percent of total size.
+ * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
+ */
+
+GLOBAL(void)
+jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl, long freq[])
+{
+#define MAX_CLEN  32            /* assumed maximum initial code length */
+  UINT8 bits[MAX_CLEN + 1];     /* bits[k] = # of symbols with code length k */
+  int bit_pos[MAX_CLEN + 1];    /* # of symbols with smaller code length */
+  int codesize[257];            /* codesize[k] = code length of symbol k */
+  int nz_index[257];            /* index of nonzero symbol in the original freq
+                                   array */
+  int others[257];              /* next symbol in current branch of tree */
+  int c1, c2;
+  int p, i, j;
+  int num_nz_symbols;
+  long v, v2;
+
+  /* This algorithm is explained in section K.2 of the JPEG standard */
+
+  memset(bits, 0, sizeof(bits));
+  memset(codesize, 0, sizeof(codesize));
+  for (i = 0; i < 257; i++)
+    others[i] = -1;             /* init links to empty */
+
+  freq[256] = 1;                /* make sure 256 has a nonzero count */
+  /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
+   * that no real symbol is given code-value of all ones, because 256
+   * will be placed last in the largest codeword category.
+   */
+
+  /* Group nonzero frequencies together so we can more easily find the
+   * smallest.
+   */
+  num_nz_symbols = 0;
+  for (i = 0; i < 257; i++) {
+    if (freq[i]) {
+      nz_index[num_nz_symbols] = i;
+      freq[num_nz_symbols] = freq[i];
+      num_nz_symbols++;
+    }
+  }
+
+  /* Huffman's basic algorithm to assign optimal code lengths to symbols */
+
+  for (;;) {
+    /* Find the two smallest nonzero frequencies; set c1, c2 = their symbols */
+    /* In case of ties, take the larger symbol number.  Since we have grouped
+     * the nonzero symbols together, checking for zero symbols is not
+     * necessary.
+     */
+    c1 = -1;
+    c2 = -1;
+    v = 1000000000L;
+    v2 = 1000000000L;
+    for (i = 0; i < num_nz_symbols; i++) {
+      if (freq[i] <= v2) {
+        if (freq[i] <= v) {
+          c2 = c1;
+          v2 = v;
+          v = freq[i];
+          c1 = i;
+        } else {
+          v2 = freq[i];
+          c2 = i;
+        }
+      }
+    }
+
+    /* Done if we've merged everything into one frequency */
+    if (c2 < 0)
+      break;
+
+    /* Else merge the two counts/trees */
+    freq[c1] += freq[c2];
+    /* Set the frequency to a very high value instead of zero, so we don't have
+     * to check for zero values.
+     */
+    freq[c2] = 1000000001L;
+
+    /* Increment the codesize of everything in c1's tree branch */
+    codesize[c1]++;
+    while (others[c1] >= 0) {
+      c1 = others[c1];
+      codesize[c1]++;
+    }
+
+    others[c1] = c2;            /* chain c2 onto c1's tree branch */
+
+    /* Increment the codesize of everything in c2's tree branch */
+    codesize[c2]++;
+    while (others[c2] >= 0) {
+      c2 = others[c2];
+      codesize[c2]++;
+    }
+  }
+
+  /* Now count the number of symbols of each code length */
+  for (i = 0; i < num_nz_symbols; i++) {
+    /* The JPEG standard seems to think that this can't happen, */
+    /* but I'm paranoid... */
+    if (codesize[i] > MAX_CLEN)
+      ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
+
+    bits[codesize[i]]++;
+  }
+
+  /* Count the number of symbols with a length smaller than i bits, so we can
+   * construct the symbol table more efficiently.  Note that this includes the
+   * pseudo-symbol 256, but since it is the last symbol, it will not affect the
+   * table.
+   */
+  p = 0;
+  for (i = 1; i <= MAX_CLEN; i++) {
+    bit_pos[i] = p;
+    p += bits[i];
+  }
+
+  /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
+   * Huffman procedure assigned any such lengths, we must adjust the coding.
+   * Here is what Rec. ITU-T T.81 | ISO/IEC 10918-1 says about how this next
+   * bit works: Since symbols are paired for the longest Huffman code, the
+   * symbols are removed from this length category two at a time.  The prefix
+   * for the pair (which is one bit shorter) is allocated to one of the pair;
+   * then, skipping the BITS entry for that prefix length, a code word from the
+   * next shortest nonzero BITS entry is converted into a prefix for two code
+   * words one bit longer.
+   */
+
+  for (i = MAX_CLEN; i > 16; i--) {
+    while (bits[i] > 0) {
+      j = i - 2;                /* find length of new prefix to be used */
+      while (bits[j] == 0)
+        j--;
+
+      bits[i] -= 2;             /* remove two symbols */
+      bits[i - 1]++;            /* one goes in this length */
+      bits[j + 1] += 2;         /* two new symbols in this length */
+      bits[j]--;                /* symbol of this length is now a prefix */
+    }
+  }
+
+  /* Remove the count for the pseudo-symbol 256 from the largest codelength */
+  while (bits[i] == 0)          /* find largest codelength still in use */
+    i--;
+  bits[i]--;
+
+  /* Return final symbol counts (only for lengths 0..16) */
+  memcpy(htbl->bits, bits, sizeof(htbl->bits));
+
+  /* Return a list of the symbols sorted by code length */
+  /* It's not real clear to me why we don't need to consider the codelength
+   * changes made above, but Rec. ITU-T T.81 | ISO/IEC 10918-1 seems to think
+   * this works.
+   */
+  for (i = 0; i < num_nz_symbols - 1; i++) {
+    htbl->huffval[bit_pos[codesize[i]]] = (UINT8)nz_index[i];
+    bit_pos[codesize[i]]++;
+  }
+
+  /* Set sent_table FALSE so updated table will be written to JPEG file. */
+  htbl->sent_table = FALSE;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather(j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info *compptr;
+  JHUFF_TBL **htblptr;
+  boolean did_dc[NUM_HUFF_TBLS];
+  boolean did_ac[NUM_HUFF_TBLS];
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  memset(did_dc, 0, sizeof(did_dc));
+  memset(did_ac, 0, sizeof(did_ac));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (!did_dc[dctbl]) {
+      htblptr = &cinfo->dc_huff_tbl_ptrs[dctbl];
+      if (*htblptr == NULL)
+        *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
+      did_dc[dctbl] = TRUE;
+    }
+    if (!did_ac[actbl]) {
+      htblptr = &cinfo->ac_huff_tbl_ptrs[actbl];
+      if (*htblptr == NULL)
+        *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
+      did_ac[actbl] = TRUE;
+    }
+  }
+}
+
+
+#endif /* ENTROPY_OPT_SUPPORTED */
+
+
+/*
+ * Module initialization routine for Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_huff_encoder(j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(huff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *)entropy;
+  entropy->pub.start_pass = start_pass_huff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
+#endif
+  }
+}

thirdparty/libjpeg-turbo/src/jchuff.h

@@ -0,0 +1,44 @@
+/*
+ * jchuff.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains declarations for Huffman entropy encoding routines
+ * that are shared between the sequential encoder (jchuff.c) and the
+ * progressive encoder (jcphuff.c).  No other modules need to see these.
+ */
+
+/* The legal range of a DCT coefficient is
+ *  -1024 .. +1023  for 8-bit data;
+ * -16384 .. +16383 for 12-bit data.
+ * Hence the magnitude should always fit in 10 or 14 bits respectively.
+ */
+
+/* The progressive Huffman encoder uses an unsigned 16-bit data type to store
+ * absolute values of coefficients, because it is possible to inject a
+ * coefficient value of -32768 into the encoder by attempting to transform a
+ * malformed 12-bit JPEG image, and the absolute value of -32768 would overflow
+ * a signed 16-bit integer.
+ */
+typedef unsigned short UJCOEF;
+
+/* Derived data constructed for each Huffman table */
+
+typedef struct {
+  unsigned int ehufco[256];     /* code for each symbol */
+  char ehufsi[256];             /* length of code for each symbol */
+  /* If no code has been allocated for a symbol S, ehufsi[S] contains 0 */
+} c_derived_tbl;
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_c_derived_tbl(j_compress_ptr cinfo, boolean isDC,
+                                     int tblno, c_derived_tbl **pdtbl);
+
+/* Generate an optimal table definition given the specified counts */
+EXTERN(void) jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl,
+                                    long freq[]);

thirdparty/libjpeg-turbo/src/jcicc.c

@@ -0,0 +1,105 @@
+/*
+ * jcicc.c
+ *
+ * Copyright (C) 1997-1998, Thomas G. Lane, Todd Newman.
+ * Copyright (C) 2017, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file provides code to write International Color Consortium (ICC) device
+ * profiles embedded in JFIF JPEG image files.  The ICC has defined a standard
+ * for including such data in JPEG "APP2" markers.  The code given here does
+ * not know anything about the internal structure of the ICC profile data; it
+ * just knows how to embed the profile data in a JPEG file while writing it.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/*
+ * Since an ICC profile can be larger than the maximum size of a JPEG marker
+ * (64K), we need provisions to split it into multiple markers.  The format
+ * defined by the ICC specifies one or more APP2 markers containing the
+ * following data:
+ *      Identifying string      ASCII "ICC_PROFILE\0"  (12 bytes)
+ *      Marker sequence number  1 for first APP2, 2 for next, etc (1 byte)
+ *      Number of markers       Total number of APP2's used (1 byte)
+ *      Profile data            (remainder of APP2 data)
+ * Decoders should use the marker sequence numbers to reassemble the profile,
+ * rather than assuming that the APP2 markers appear in the correct sequence.
+ */
+
+#define ICC_MARKER  (JPEG_APP0 + 2)     /* JPEG marker code for ICC */
+#define ICC_OVERHEAD_LEN  14            /* size of non-profile data in APP2 */
+#define MAX_BYTES_IN_MARKER  65533      /* maximum data len of a JPEG marker */
+#define MAX_DATA_BYTES_IN_MARKER  (MAX_BYTES_IN_MARKER - ICC_OVERHEAD_LEN)
+
+
+/*
+ * This routine writes the given ICC profile data into a JPEG file.  It *must*
+ * be called AFTER calling jpeg_start_compress() and BEFORE the first call to
+ * jpeg_write_scanlines().  (This ordering ensures that the APP2 marker(s) will
+ * appear after the SOI and JFIF or Adobe markers, but before all else.)
+ */
+
+GLOBAL(void)
+jpeg_write_icc_profile(j_compress_ptr cinfo, const JOCTET *icc_data_ptr,
+                       unsigned int icc_data_len)
+{
+  unsigned int num_markers;     /* total number of markers we'll write */
+  int cur_marker = 1;           /* per spec, counting starts at 1 */
+  unsigned int length;          /* number of bytes to write in this marker */
+
+  if (icc_data_ptr == NULL || icc_data_len == 0)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  if (cinfo->global_state < CSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Calculate the number of markers we'll need, rounding up of course */
+  num_markers = icc_data_len / MAX_DATA_BYTES_IN_MARKER;
+  if (num_markers * MAX_DATA_BYTES_IN_MARKER != icc_data_len)
+    num_markers++;
+
+  while (icc_data_len > 0) {
+    /* length of profile to put in this marker */
+    length = icc_data_len;
+    if (length > MAX_DATA_BYTES_IN_MARKER)
+      length = MAX_DATA_BYTES_IN_MARKER;
+    icc_data_len -= length;
+
+    /* Write the JPEG marker header (APP2 code and marker length) */
+    jpeg_write_m_header(cinfo, ICC_MARKER,
+                        (unsigned int)(length + ICC_OVERHEAD_LEN));
+
+    /* Write the marker identifying string "ICC_PROFILE" (null-terminated).  We
+     * code it in this less-than-transparent way so that the code works even if
+     * the local character set is not ASCII.
+     */
+    jpeg_write_m_byte(cinfo, 0x49);
+    jpeg_write_m_byte(cinfo, 0x43);
+    jpeg_write_m_byte(cinfo, 0x43);
+    jpeg_write_m_byte(cinfo, 0x5F);
+    jpeg_write_m_byte(cinfo, 0x50);
+    jpeg_write_m_byte(cinfo, 0x52);
+    jpeg_write_m_byte(cinfo, 0x4F);
+    jpeg_write_m_byte(cinfo, 0x46);
+    jpeg_write_m_byte(cinfo, 0x49);
+    jpeg_write_m_byte(cinfo, 0x4C);
+    jpeg_write_m_byte(cinfo, 0x45);
+    jpeg_write_m_byte(cinfo, 0x0);
+
+    /* Add the sequencing info */
+    jpeg_write_m_byte(cinfo, cur_marker);
+    jpeg_write_m_byte(cinfo, (int)num_markers);
+
+    /* Add the profile data */
+    while (length--) {
+      jpeg_write_m_byte(cinfo, *icc_data_ptr);
+      icc_data_ptr++;
+    }
+    cur_marker++;
+  }
+}

thirdparty/libjpeg-turbo/src/jcinit.c

@@ -0,0 +1,149 @@
+/*
+ * jcinit.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains initialization logic for the JPEG compressor.
+ * This routine is in charge of selecting the modules to be executed and
+ * making an initialization call to each one.
+ *
+ * Logically, this code belongs in jcmaster.c.  It's split out because
+ * linking this routine implies linking the entire compression library.
+ * For a transcoding-only application, we want to be able to use jcmaster.c
+ * without linking in the whole library.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+
+
+/*
+ * Master selection of compression modules.
+ * This is done once at the start of processing an image.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ */
+
+GLOBAL(void)
+jinit_compress_master(j_compress_ptr cinfo)
+{
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, FALSE /* full compression */);
+
+  /* Preprocessing */
+  if (!cinfo->raw_data_in) {
+    if (cinfo->data_precision <= 8) {
+      jinit_color_converter(cinfo);
+      jinit_downsampler(cinfo);
+      jinit_c_prep_controller(cinfo, FALSE /* never need full buffer here */);
+    } else if (cinfo->data_precision <= 12) {
+      j12init_color_converter(cinfo);
+      j12init_downsampler(cinfo);
+      j12init_c_prep_controller(cinfo,
+                                FALSE /* never need full buffer here */);
+    } else {
+#ifdef C_LOSSLESS_SUPPORTED
+      j16init_color_converter(cinfo);
+      j16init_downsampler(cinfo);
+      j16init_c_prep_controller(cinfo,
+                                FALSE /* never need full buffer here */);
+#else
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+    }
+  }
+
+  if (cinfo->master->lossless) {
+#ifdef C_LOSSLESS_SUPPORTED
+    /* Prediction, sample differencing, and point transform */
+    if (cinfo->data_precision <= 8)
+      jinit_lossless_compressor(cinfo);
+    else if (cinfo->data_precision <= 12)
+      j12init_lossless_compressor(cinfo);
+    else
+      j16init_lossless_compressor(cinfo);
+    /* Entropy encoding: either Huffman or arithmetic coding. */
+    if (cinfo->arith_code) {
+      ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+    } else {
+      jinit_lhuff_encoder(cinfo);
+    }
+
+    /* Need a full-image difference buffer in any multi-pass mode. */
+    if (cinfo->data_precision <= 8)
+      jinit_c_diff_controller(cinfo, (boolean)(cinfo->num_scans > 1 ||
+                                               cinfo->optimize_coding));
+    else if (cinfo->data_precision <= 12)
+      j12init_c_diff_controller(cinfo, (boolean)(cinfo->num_scans > 1 ||
+                                                 cinfo->optimize_coding));
+    else
+      j16init_c_diff_controller(cinfo, (boolean)(cinfo->num_scans > 1 ||
+                                                 cinfo->optimize_coding));
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* Forward DCT */
+    if (cinfo->data_precision == 8)
+      jinit_forward_dct(cinfo);
+    else if (cinfo->data_precision == 12)
+      j12init_forward_dct(cinfo);
+    else
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+    /* Entropy encoding: either Huffman or arithmetic coding. */
+    if (cinfo->arith_code) {
+#ifdef C_ARITH_CODING_SUPPORTED
+      jinit_arith_encoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+#endif
+    } else {
+      if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+        jinit_phuff_encoder(cinfo);
+#else
+        ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+      } else
+        jinit_huff_encoder(cinfo);
+    }
+
+    /* Need a full-image coefficient buffer in any multi-pass mode. */
+    if (cinfo->data_precision == 12)
+      j12init_c_coef_controller(cinfo, (boolean)(cinfo->num_scans > 1 ||
+                                                 cinfo->optimize_coding));
+    else
+      jinit_c_coef_controller(cinfo, (boolean)(cinfo->num_scans > 1 ||
+                                               cinfo->optimize_coding));
+  }
+
+  if (cinfo->data_precision <= 8)
+    jinit_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+  else if (cinfo->data_precision <= 12)
+    j12init_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+  else
+#ifdef C_LOSSLESS_SUPPORTED
+    j16init_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+#else
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr)cinfo);
+
+  /* Write the datastream header (SOI) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}

thirdparty/libjpeg-turbo/src/jcmainct.c

@@ -0,0 +1,186 @@
+/*
+ * jcmainct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the main buffer controller for compression.
+ * The main buffer lies between the pre-processor and the JPEG
+ * compressor proper; it holds downsampled data in the JPEG colorspace.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_main_controller pub; /* public fields */
+
+  JDIMENSION cur_iMCU_row;      /* number of current iMCU row */
+  JDIMENSION rowgroup_ctr;      /* counts row groups received in iMCU row */
+  boolean suspended;            /* remember if we suspended output */
+  J_BUF_MODE pass_mode;         /* current operating mode */
+
+  /* If using just a strip buffer, this points to the entire set of buffers
+   * (we allocate one for each component).  In the full-image case, this
+   * points to the currently accessible strips of the virtual arrays.
+   */
+  _JSAMPARRAY buffer[MAX_COMPONENTS];
+} my_main_controller;
+
+typedef my_main_controller *my_main_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main(j_compress_ptr cinfo,
+                                         _JSAMPARRAY input_buf,
+                                         JDIMENSION *in_row_ctr,
+                                         JDIMENSION in_rows_avail);
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main(j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+
+  /* Do nothing in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  if (pass_mode != JBUF_PASS_THRU)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  main_ptr->cur_iMCU_row = 0;   /* initialize counters */
+  main_ptr->rowgroup_ctr = 0;
+  main_ptr->suspended = FALSE;
+  main_ptr->pass_mode = pass_mode;      /* save mode for use by process_data */
+  main_ptr->pub._process_data = process_data_simple_main;
+}
+
+
+/*
+ * Process some data.
+ * This routine handles the simple pass-through mode,
+ * where we have only a strip buffer.
+ */
+
+METHODDEF(void)
+process_data_simple_main(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                         JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  JDIMENSION data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  while (main_ptr->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Read input data if we haven't filled the main buffer yet */
+    if (main_ptr->rowgroup_ctr < data_unit)
+      (*cinfo->prep->_pre_process_data) (cinfo, input_buf, in_row_ctr,
+                                         in_rows_avail, main_ptr->buffer,
+                                         &main_ptr->rowgroup_ctr, data_unit);
+
+    /* If we don't have a full iMCU row buffered, return to application for
+     * more data.  Note that preprocessor will always pad to fill the iMCU row
+     * at the bottom of the image.
+     */
+    if (main_ptr->rowgroup_ctr != data_unit)
+      return;
+
+    /* Send the completed row to the compressor */
+    if (!(*cinfo->coef->_compress_data) (cinfo, main_ptr->buffer)) {
+      /* If compressor did not consume the whole row, then we must need to
+       * suspend processing and return to the application.  In this situation
+       * we pretend we didn't yet consume the last input row; otherwise, if
+       * it happened to be the last row of the image, the application would
+       * think we were done.
+       */
+      if (!main_ptr->suspended) {
+        (*in_row_ctr)--;
+        main_ptr->suspended = TRUE;
+      }
+      return;
+    }
+    /* We did finish the row.  Undo our little suspension hack if a previous
+     * call suspended; then mark the main buffer empty.
+     */
+    if (main_ptr->suspended) {
+      (*in_row_ctr)++;
+      main_ptr->suspended = FALSE;
+    }
+    main_ptr->rowgroup_ctr = 0;
+    main_ptr->cur_iMCU_row++;
+  }
+}
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+_jinit_c_main_controller(j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main_ptr;
+  int ci;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  main_ptr = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_main_controller));
+  cinfo->main = (struct jpeg_c_main_controller *)main_ptr;
+  main_ptr->pub.start_pass = start_pass_main;
+
+  /* We don't need to create a buffer in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  /* Create the buffer.  It holds downsampled data, so each component
+   * may be of a different size.
+   */
+  if (need_full_buffer) {
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+  } else {
+    /* Allocate a strip buffer for each component */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      main_ptr->buffer[ci] = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+         compptr->width_in_blocks * data_unit,
+         (JDIMENSION)(compptr->v_samp_factor * data_unit));
+    }
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jcmarker.c

@@ -0,0 +1,670 @@
+/*
+ * jcmarker.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains routines to write JPEG datastream markers.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+
+
+typedef enum {                  /* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+
+  M_DHT   = 0xc4,
+
+  M_DAC   = 0xcc,
+
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+
+  M_TEM   = 0x01,
+
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_writer pub; /* public fields */
+
+  unsigned int last_restart_interval; /* last DRI value emitted; 0 after SOI */
+} my_marker_writer;
+
+typedef my_marker_writer *my_marker_ptr;
+
+
+/*
+ * Basic output routines.
+ *
+ * Note that we do not support suspension while writing a marker.
+ * Therefore, an application using suspension must ensure that there is
+ * enough buffer space for the initial markers (typ. 600-700 bytes) before
+ * calling jpeg_start_compress, and enough space to write the trailing EOI
+ * (a few bytes) before calling jpeg_finish_compress.  Multipass compression
+ * modes are not supported at all with suspension, so those two are the only
+ * points where markers will be written.
+ */
+
+LOCAL(void)
+emit_byte(j_compress_ptr cinfo, int val)
+/* Emit a byte */
+{
+  struct jpeg_destination_mgr *dest = cinfo->dest;
+
+  *(dest->next_output_byte)++ = (JOCTET)val;
+  if (--dest->free_in_buffer == 0) {
+    if (!(*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  }
+}
+
+
+LOCAL(void)
+emit_marker(j_compress_ptr cinfo, JPEG_MARKER mark)
+/* Emit a marker code */
+{
+  emit_byte(cinfo, 0xFF);
+  emit_byte(cinfo, (int)mark);
+}
+
+
+LOCAL(void)
+emit_2bytes(j_compress_ptr cinfo, int value)
+/* Emit a 2-byte integer; these are always MSB first in JPEG files */
+{
+  emit_byte(cinfo, (value >> 8) & 0xFF);
+  emit_byte(cinfo, value & 0xFF);
+}
+
+
+/*
+ * Routines to write specific marker types.
+ */
+
+LOCAL(int)
+emit_dqt(j_compress_ptr cinfo, int index)
+/* Emit a DQT marker */
+/* Returns the precision used (0 = 8bits, 1 = 16bits) for baseline checking */
+{
+  JQUANT_TBL *qtbl = cinfo->quant_tbl_ptrs[index];
+  int prec;
+  int i;
+
+  if (qtbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, index);
+
+  prec = 0;
+  for (i = 0; i < DCTSIZE2; i++) {
+    if (qtbl->quantval[i] > 255)
+      prec = 1;
+  }
+
+  if (!qtbl->sent_table) {
+    emit_marker(cinfo, M_DQT);
+
+    emit_2bytes(cinfo, prec ? DCTSIZE2 * 2 + 1 + 2 : DCTSIZE2 + 1 + 2);
+
+    emit_byte(cinfo, index + (prec << 4));
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      /* The table entries must be emitted in zigzag order. */
+      unsigned int qval = qtbl->quantval[jpeg_natural_order[i]];
+      if (prec)
+        emit_byte(cinfo, (int)(qval >> 8));
+      emit_byte(cinfo, (int)(qval & 0xFF));
+    }
+
+    qtbl->sent_table = TRUE;
+  }
+
+  return prec;
+}
+
+
+LOCAL(void)
+emit_dht(j_compress_ptr cinfo, int index, boolean is_ac)
+/* Emit a DHT marker */
+{
+  JHUFF_TBL *htbl;
+  int length, i;
+
+  if (is_ac) {
+    htbl = cinfo->ac_huff_tbl_ptrs[index];
+    index += 0x10;              /* output index has AC bit set */
+  } else {
+    htbl = cinfo->dc_huff_tbl_ptrs[index];
+  }
+
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, index);
+
+  if (!htbl->sent_table) {
+    emit_marker(cinfo, M_DHT);
+
+    length = 0;
+    for (i = 1; i <= 16; i++)
+      length += htbl->bits[i];
+
+    emit_2bytes(cinfo, length + 2 + 1 + 16);
+    emit_byte(cinfo, index);
+
+    for (i = 1; i <= 16; i++)
+      emit_byte(cinfo, htbl->bits[i]);
+
+    for (i = 0; i < length; i++)
+      emit_byte(cinfo, htbl->huffval[i]);
+
+    htbl->sent_table = TRUE;
+  }
+}
+
+
+LOCAL(void)
+emit_dac(j_compress_ptr cinfo)
+/* Emit a DAC marker */
+/* Since the useful info is so small, we want to emit all the tables in */
+/* one DAC marker.  Therefore this routine does its own scan of the table. */
+{
+#ifdef C_ARITH_CODING_SUPPORTED
+  char dc_in_use[NUM_ARITH_TBLS];
+  char ac_in_use[NUM_ARITH_TBLS];
+  int length, i;
+  jpeg_component_info *compptr;
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    dc_in_use[i] = ac_in_use[i] = 0;
+
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0)
+      dc_in_use[compptr->dc_tbl_no] = 1;
+    /* AC needs no table when not present */
+    if (cinfo->Se)
+      ac_in_use[compptr->ac_tbl_no] = 1;
+  }
+
+  length = 0;
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    length += dc_in_use[i] + ac_in_use[i];
+
+  if (length) {
+    emit_marker(cinfo, M_DAC);
+
+    emit_2bytes(cinfo, length * 2 + 2);
+
+    for (i = 0; i < NUM_ARITH_TBLS; i++) {
+      if (dc_in_use[i]) {
+        emit_byte(cinfo, i);
+        emit_byte(cinfo, cinfo->arith_dc_L[i] + (cinfo->arith_dc_U[i] << 4));
+      }
+      if (ac_in_use[i]) {
+        emit_byte(cinfo, i + 0x10);
+        emit_byte(cinfo, cinfo->arith_ac_K[i]);
+      }
+    }
+  }
+#endif /* C_ARITH_CODING_SUPPORTED */
+}
+
+
+LOCAL(void)
+emit_dri(j_compress_ptr cinfo)
+/* Emit a DRI marker */
+{
+  emit_marker(cinfo, M_DRI);
+
+  emit_2bytes(cinfo, 4);        /* fixed length */
+
+  emit_2bytes(cinfo, (int)cinfo->restart_interval);
+}
+
+
+LOCAL(void)
+emit_sof(j_compress_ptr cinfo, JPEG_MARKER code)
+/* Emit a SOF marker */
+{
+  int ci;
+  jpeg_component_info *compptr;
+
+  emit_marker(cinfo, code);
+
+  emit_2bytes(cinfo, 3 * cinfo->num_components + 2 + 5 + 1); /* length */
+
+  /* Make sure image isn't bigger than SOF field can handle */
+  if ((long)cinfo->_jpeg_height > 65535L || (long)cinfo->_jpeg_width > 65535L)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int)65535);
+
+  emit_byte(cinfo, cinfo->data_precision);
+  emit_2bytes(cinfo, (int)cinfo->_jpeg_height);
+  emit_2bytes(cinfo, (int)cinfo->_jpeg_width);
+
+  emit_byte(cinfo, cinfo->num_components);
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    emit_byte(cinfo, compptr->component_id);
+    emit_byte(cinfo, (compptr->h_samp_factor << 4) + compptr->v_samp_factor);
+    emit_byte(cinfo, compptr->quant_tbl_no);
+  }
+}
+
+
+LOCAL(void)
+emit_sos(j_compress_ptr cinfo)
+/* Emit a SOS marker */
+{
+  int i, td, ta;
+  jpeg_component_info *compptr;
+
+  emit_marker(cinfo, M_SOS);
+
+  emit_2bytes(cinfo, 2 * cinfo->comps_in_scan + 2 + 1 + 3); /* length */
+
+  emit_byte(cinfo, cinfo->comps_in_scan);
+
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    emit_byte(cinfo, compptr->component_id);
+
+    /* We emit 0 for unused field(s); this is recommended by the P&M text
+     * but does not seem to be specified in the standard.
+     */
+
+    /* DC needs no table for refinement scan */
+    td = cinfo->Ss == 0 && cinfo->Ah == 0 ? compptr->dc_tbl_no : 0;
+    /* AC needs no table when not present */
+    ta = cinfo->Se ? compptr->ac_tbl_no : 0;
+
+    emit_byte(cinfo, (td << 4) + ta);
+  }
+
+  emit_byte(cinfo, cinfo->Ss);
+  emit_byte(cinfo, cinfo->Se);
+  emit_byte(cinfo, (cinfo->Ah << 4) + cinfo->Al);
+}
+
+
+LOCAL(void)
+emit_jfif_app0(j_compress_ptr cinfo)
+/* Emit a JFIF-compliant APP0 marker */
+{
+  /*
+   * Length of APP0 block       (2 bytes)
+   * Block ID                   (4 bytes - ASCII "JFIF")
+   * Zero byte                  (1 byte to terminate the ID string)
+   * Version Major, Minor       (2 bytes - major first)
+   * Units                      (1 byte - 0x00 = none, 0x01 = inch, 0x02 = cm)
+   * Xdpu                       (2 bytes - dots per unit horizontal)
+   * Ydpu                       (2 bytes - dots per unit vertical)
+   * Thumbnail X size           (1 byte)
+   * Thumbnail Y size           (1 byte)
+   */
+
+  emit_marker(cinfo, M_APP0);
+
+  emit_2bytes(cinfo, 2 + 4 + 1 + 2 + 1 + 2 + 2 + 1 + 1); /* length */
+
+  emit_byte(cinfo, 0x4A);       /* Identifier: ASCII "JFIF" */
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0x49);
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0);
+  emit_byte(cinfo, cinfo->JFIF_major_version); /* Version fields */
+  emit_byte(cinfo, cinfo->JFIF_minor_version);
+  emit_byte(cinfo, cinfo->density_unit); /* Pixel size information */
+  emit_2bytes(cinfo, (int)cinfo->X_density);
+  emit_2bytes(cinfo, (int)cinfo->Y_density);
+  emit_byte(cinfo, 0);          /* No thumbnail image */
+  emit_byte(cinfo, 0);
+}
+
+
+LOCAL(void)
+emit_adobe_app14(j_compress_ptr cinfo)
+/* Emit an Adobe APP14 marker */
+{
+  /*
+   * Length of APP14 block      (2 bytes)
+   * Block ID                   (5 bytes - ASCII "Adobe")
+   * Version Number             (2 bytes - currently 100)
+   * Flags0                     (2 bytes - currently 0)
+   * Flags1                     (2 bytes - currently 0)
+   * Color transform            (1 byte)
+   *
+   * Although Adobe TN 5116 mentions Version = 101, all the Adobe files
+   * now in circulation seem to use Version = 100, so that's what we write.
+   *
+   * We write the color transform byte as 1 if the JPEG color space is
+   * YCbCr, 2 if it's YCCK, 0 otherwise.  Adobe's definition has to do with
+   * whether the encoder performed a transformation, which is pretty useless.
+   */
+
+  emit_marker(cinfo, M_APP14);
+
+  emit_2bytes(cinfo, 2 + 5 + 2 + 2 + 2 + 1); /* length */
+
+  emit_byte(cinfo, 0x41);       /* Identifier: ASCII "Adobe" */
+  emit_byte(cinfo, 0x64);
+  emit_byte(cinfo, 0x6F);
+  emit_byte(cinfo, 0x62);
+  emit_byte(cinfo, 0x65);
+  emit_2bytes(cinfo, 100);      /* Version */
+  emit_2bytes(cinfo, 0);        /* Flags0 */
+  emit_2bytes(cinfo, 0);        /* Flags1 */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_YCbCr:
+    emit_byte(cinfo, 1);        /* Color transform = 1 */
+    break;
+  case JCS_YCCK:
+    emit_byte(cinfo, 2);        /* Color transform = 2 */
+    break;
+  default:
+    emit_byte(cinfo, 0);        /* Color transform = 0 */
+    break;
+  }
+}
+
+
+/*
+ * These routines allow writing an arbitrary marker with parameters.
+ * The only intended use is to emit COM or APPn markers after calling
+ * write_file_header and before calling write_frame_header.
+ * Other uses are not guaranteed to produce desirable results.
+ * Counting the parameter bytes properly is the caller's responsibility.
+ */
+
+METHODDEF(void)
+write_marker_header(j_compress_ptr cinfo, int marker, unsigned int datalen)
+/* Emit an arbitrary marker header */
+{
+  if (datalen > (unsigned int)65533)            /* safety check */
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  emit_marker(cinfo, (JPEG_MARKER)marker);
+
+  emit_2bytes(cinfo, (int)(datalen + 2));       /* total length */
+}
+
+METHODDEF(void)
+write_marker_byte(j_compress_ptr cinfo, int val)
+/* Emit one byte of marker parameters following write_marker_header */
+{
+  emit_byte(cinfo, val);
+}
+
+
+/*
+ * Write datastream header.
+ * This consists of an SOI and optional APPn markers.
+ * We recommend use of the JFIF marker, but not the Adobe marker,
+ * when using YCbCr or grayscale data.  The JFIF marker should NOT
+ * be used for any other JPEG colorspace.  The Adobe marker is helpful
+ * to distinguish RGB, CMYK, and YCCK colorspaces.
+ * Note that an application can write additional header markers after
+ * jpeg_start_compress returns.
+ */
+
+METHODDEF(void)
+write_file_header(j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+
+  emit_marker(cinfo, M_SOI);    /* first the SOI */
+
+  /* SOI is defined to reset restart interval to 0 */
+  marker->last_restart_interval = 0;
+
+  if (cinfo->write_JFIF_header) /* next an optional JFIF APP0 */
+    emit_jfif_app0(cinfo);
+  if (cinfo->write_Adobe_marker) /* next an optional Adobe APP14 */
+    emit_adobe_app14(cinfo);
+}
+
+
+/*
+ * Write frame header.
+ * This consists of DQT and SOFn markers.
+ * Note that we do not emit the SOF until we have emitted the DQT(s).
+ * This avoids compatibility problems with incorrect implementations that
+ * try to error-check the quant table numbers as soon as they see the SOF.
+ */
+
+METHODDEF(void)
+write_frame_header(j_compress_ptr cinfo)
+{
+  int ci, prec = 0;
+  boolean is_baseline;
+  jpeg_component_info *compptr;
+
+  if (!cinfo->master->lossless) {
+    /* Emit DQT for each quantization table.
+     * Note that emit_dqt() suppresses any duplicate tables.
+     */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      prec += emit_dqt(cinfo, compptr->quant_tbl_no);
+    }
+    /* now prec is nonzero iff there are any 16-bit quant tables. */
+  }
+
+  /* Check for a non-baseline specification.
+   * Note we assume that Huffman table numbers won't be changed later.
+   */
+  if (cinfo->arith_code || cinfo->progressive_mode ||
+      cinfo->master->lossless || cinfo->data_precision != 8) {
+    is_baseline = FALSE;
+  } else {
+    is_baseline = TRUE;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      if (compptr->dc_tbl_no > 1 || compptr->ac_tbl_no > 1)
+        is_baseline = FALSE;
+    }
+    if (prec && is_baseline) {
+      is_baseline = FALSE;
+      /* If it's baseline except for quantizer size, warn the user */
+      TRACEMS(cinfo, 0, JTRC_16BIT_TABLES);
+    }
+  }
+
+  /* Emit the proper SOF marker */
+  if (cinfo->arith_code) {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF10); /* SOF code for progressive arithmetic */
+    else
+      emit_sof(cinfo, M_SOF9);  /* SOF code for sequential arithmetic */
+  } else {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF2);  /* SOF code for progressive Huffman */
+    else if (cinfo->master->lossless)
+      emit_sof(cinfo, M_SOF3);  /* SOF code for lossless Huffman */
+    else if (is_baseline)
+      emit_sof(cinfo, M_SOF0);  /* SOF code for baseline implementation */
+    else
+      emit_sof(cinfo, M_SOF1);  /* SOF code for non-baseline Huffman file */
+  }
+}
+
+
+/*
+ * Write scan header.
+ * This consists of DHT or DAC markers, optional DRI, and SOS.
+ * Compressed data will be written following the SOS.
+ */
+
+METHODDEF(void)
+write_scan_header(j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+  int i;
+  jpeg_component_info *compptr;
+
+  if (cinfo->arith_code) {
+    /* Emit arith conditioning info.  We may have some duplication
+     * if the file has multiple scans, but it's so small it's hardly
+     * worth worrying about.
+     */
+    emit_dac(cinfo);
+  } else {
+    /* Emit Huffman tables.
+     * Note that emit_dht() suppresses any duplicate tables.
+     */
+    for (i = 0; i < cinfo->comps_in_scan; i++) {
+      compptr = cinfo->cur_comp_info[i];
+      /* DC needs no table for refinement scan */
+      if ((cinfo->Ss == 0 && cinfo->Ah == 0) || cinfo->master->lossless)
+        emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+      /* AC needs no table when not present, and lossless mode uses only DC
+         tables. */
+      if (cinfo->Se && !cinfo->master->lossless)
+        emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+    }
+  }
+
+  /* Emit DRI if required --- note that DRI value could change for each scan.
+   * We avoid wasting space with unnecessary DRIs, however.
+   */
+  if (cinfo->restart_interval != marker->last_restart_interval) {
+    emit_dri(cinfo);
+    marker->last_restart_interval = cinfo->restart_interval;
+  }
+
+  emit_sos(cinfo);
+}
+
+
+/*
+ * Write datastream trailer.
+ */
+
+METHODDEF(void)
+write_file_trailer(j_compress_ptr cinfo)
+{
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Write an abbreviated table-specification datastream.
+ * This consists of SOI, DQT and DHT tables, and EOI.
+ * Any table that is defined and not marked sent_table = TRUE will be
+ * emitted.  Note that all tables will be marked sent_table = TRUE at exit.
+ */
+
+METHODDEF(void)
+write_tables_only(j_compress_ptr cinfo)
+{
+  int i;
+
+  emit_marker(cinfo, M_SOI);
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if (cinfo->quant_tbl_ptrs[i] != NULL)
+      (void)emit_dqt(cinfo, i);
+  }
+
+  if (!cinfo->arith_code) {
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      if (cinfo->dc_huff_tbl_ptrs[i] != NULL)
+        emit_dht(cinfo, i, FALSE);
+      if (cinfo->ac_huff_tbl_ptrs[i] != NULL)
+        emit_dht(cinfo, i, TRUE);
+    }
+  }
+
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Initialize the marker writer module.
+ */
+
+GLOBAL(void)
+jinit_marker_writer(j_compress_ptr cinfo)
+{
+  my_marker_ptr marker;
+
+  /* Create the subobject */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_marker_writer));
+  cinfo->marker = (struct jpeg_marker_writer *)marker;
+  /* Initialize method pointers */
+  marker->pub.write_file_header = write_file_header;
+  marker->pub.write_frame_header = write_frame_header;
+  marker->pub.write_scan_header = write_scan_header;
+  marker->pub.write_file_trailer = write_file_trailer;
+  marker->pub.write_tables_only = write_tables_only;
+  marker->pub.write_marker_header = write_marker_header;
+  marker->pub.write_marker_byte = write_marker_byte;
+  /* Initialize private state */
+  marker->last_restart_interval = 0;
+}

thirdparty/libjpeg-turbo/src/jcmaster.c

@@ -0,0 +1,801 @@
+/*
+ * jcmaster.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2016, 2018, 2022-2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains master control logic for the JPEG compressor.
+ * These routines are concerned with parameter validation, initial setup,
+ * and inter-pass control (determining the number of passes and the work
+ * to be done in each pass).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+#include "jcmaster.h"
+
+
+/*
+ * Support routines that do various essential calculations.
+ */
+
+#if JPEG_LIB_VERSION >= 70
+/*
+ * Compute JPEG image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+GLOBAL(void)
+jpeg_calc_jpeg_dimensions(j_compress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  /* Hardwire it to "no scaling" */
+  cinfo->jpeg_width = cinfo->image_width;
+  cinfo->jpeg_height = cinfo->image_height;
+  cinfo->min_DCT_h_scaled_size = data_unit;
+  cinfo->min_DCT_v_scaled_size = data_unit;
+}
+#endif
+
+
+LOCAL(boolean)
+using_std_huff_tables(j_compress_ptr cinfo)
+{
+  int i;
+
+  static const UINT8 bits_dc_luminance[17] = {
+    /* 0-base */ 0, 0, 1, 5, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0
+  };
+  static const UINT8 val_dc_luminance[] = {
+    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
+  };
+
+  static const UINT8 bits_dc_chrominance[17] = {
+    /* 0-base */ 0, 0, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0
+  };
+  static const UINT8 val_dc_chrominance[] = {
+    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
+  };
+
+  static const UINT8 bits_ac_luminance[17] = {
+    /* 0-base */ 0, 0, 2, 1, 3, 3, 2, 4, 3, 5, 5, 4, 4, 0, 0, 1, 0x7d
+  };
+  static const UINT8 val_ac_luminance[] = {
+    0x01, 0x02, 0x03, 0x00, 0x04, 0x11, 0x05, 0x12,
+    0x21, 0x31, 0x41, 0x06, 0x13, 0x51, 0x61, 0x07,
+    0x22, 0x71, 0x14, 0x32, 0x81, 0x91, 0xa1, 0x08,
+    0x23, 0x42, 0xb1, 0xc1, 0x15, 0x52, 0xd1, 0xf0,
+    0x24, 0x33, 0x62, 0x72, 0x82, 0x09, 0x0a, 0x16,
+    0x17, 0x18, 0x19, 0x1a, 0x25, 0x26, 0x27, 0x28,
+    0x29, 0x2a, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39,
+    0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x49,
+    0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59,
+    0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69,
+    0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79,
+    0x7a, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89,
+    0x8a, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+    0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7,
+    0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6,
+    0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3, 0xc4, 0xc5,
+    0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2, 0xd3, 0xd4,
+    0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xe1, 0xe2,
+    0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea,
+    0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+    0xf9, 0xfa
+  };
+
+  static const UINT8 bits_ac_chrominance[17] = {
+    /* 0-base */ 0, 0, 2, 1, 2, 4, 4, 3, 4, 7, 5, 4, 4, 0, 1, 2, 0x77
+  };
+  static const UINT8 val_ac_chrominance[] = {
+    0x00, 0x01, 0x02, 0x03, 0x11, 0x04, 0x05, 0x21,
+    0x31, 0x06, 0x12, 0x41, 0x51, 0x07, 0x61, 0x71,
+    0x13, 0x22, 0x32, 0x81, 0x08, 0x14, 0x42, 0x91,
+    0xa1, 0xb1, 0xc1, 0x09, 0x23, 0x33, 0x52, 0xf0,
+    0x15, 0x62, 0x72, 0xd1, 0x0a, 0x16, 0x24, 0x34,
+    0xe1, 0x25, 0xf1, 0x17, 0x18, 0x19, 0x1a, 0x26,
+    0x27, 0x28, 0x29, 0x2a, 0x35, 0x36, 0x37, 0x38,
+    0x39, 0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48,
+    0x49, 0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58,
+    0x59, 0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+    0x69, 0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78,
+    0x79, 0x7a, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+    0x88, 0x89, 0x8a, 0x92, 0x93, 0x94, 0x95, 0x96,
+    0x97, 0x98, 0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5,
+    0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4,
+    0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3,
+    0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2,
+    0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda,
+    0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9,
+    0xea, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+    0xf9, 0xfa
+  };
+
+  if (cinfo->dc_huff_tbl_ptrs[0] == NULL ||
+      cinfo->ac_huff_tbl_ptrs[0] == NULL ||
+      cinfo->dc_huff_tbl_ptrs[1] == NULL ||
+      cinfo->ac_huff_tbl_ptrs[1] == NULL)
+    return FALSE;
+
+  for (i = 2; i < NUM_HUFF_TBLS; i++) {
+    if (cinfo->dc_huff_tbl_ptrs[i] != NULL ||
+        cinfo->ac_huff_tbl_ptrs[i] != NULL)
+      return FALSE;
+  }
+
+  if (memcmp(cinfo->dc_huff_tbl_ptrs[0]->bits, bits_dc_luminance,
+             sizeof(bits_dc_luminance)) ||
+      memcmp(cinfo->dc_huff_tbl_ptrs[0]->huffval, val_dc_luminance,
+             sizeof(val_dc_luminance)) ||
+      memcmp(cinfo->ac_huff_tbl_ptrs[0]->bits, bits_ac_luminance,
+             sizeof(bits_ac_luminance)) ||
+      memcmp(cinfo->ac_huff_tbl_ptrs[0]->huffval, val_ac_luminance,
+             sizeof(val_ac_luminance)) ||
+      memcmp(cinfo->dc_huff_tbl_ptrs[1]->bits, bits_dc_chrominance,
+             sizeof(bits_dc_chrominance)) ||
+      memcmp(cinfo->dc_huff_tbl_ptrs[1]->huffval, val_dc_chrominance,
+             sizeof(val_dc_chrominance)) ||
+      memcmp(cinfo->ac_huff_tbl_ptrs[1]->bits, bits_ac_chrominance,
+             sizeof(bits_ac_chrominance)) ||
+      memcmp(cinfo->ac_huff_tbl_ptrs[1]->huffval, val_ac_chrominance,
+             sizeof(val_ac_chrominance)))
+    return FALSE;
+
+  return TRUE;
+}
+
+
+LOCAL(void)
+initial_setup(j_compress_ptr cinfo, boolean transcode_only)
+/* Do computations that are needed before master selection phase */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+#if JPEG_LIB_VERSION >= 70
+#if JPEG_LIB_VERSION >= 80
+  if (!transcode_only)
+#endif
+    jpeg_calc_jpeg_dimensions(cinfo);
+#endif
+
+  /* Sanity check on image dimensions */
+  if (cinfo->_jpeg_height <= 0 || cinfo->_jpeg_width <= 0 ||
+      cinfo->num_components <= 0 || cinfo->input_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long)cinfo->_jpeg_height > (long)JPEG_MAX_DIMENSION ||
+      (long)cinfo->_jpeg_width > (long)JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int)JPEG_MAX_DIMENSION);
+
+  /* Width of an input scanline must be representable as JDIMENSION. */
+  samplesperrow = (long)cinfo->image_width * (long)cinfo->input_components;
+  jd_samplesperrow = (JDIMENSION)samplesperrow;
+  if ((long)jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* Lossy JPEG images must have 8 or 12 bits per sample.  Lossless JPEG images
+   * can have 2 to 16 bits per sample.
+   */
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+    if (cinfo->data_precision < 2 || cinfo->data_precision > 16)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != 8 && cinfo->data_precision != 12)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+             MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor <= 0 ||
+        compptr->h_samp_factor > MAX_SAMP_FACTOR ||
+        compptr->v_samp_factor <= 0 ||
+        compptr->v_samp_factor > MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+                                   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+                                   compptr->v_samp_factor);
+  }
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Fill in the correct component_index value; don't rely on application */
+    compptr->component_index = ci;
+    /* For compression, we never do DCT scaling. */
+#if JPEG_LIB_VERSION >= 70
+    compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = data_unit;
+#else
+    compptr->DCT_scaled_size = data_unit;
+#endif
+    /* Size in data units */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_width * (long)compptr->h_samp_factor,
+                    (long)(cinfo->max_h_samp_factor * data_unit));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_height * (long)compptr->v_samp_factor,
+                    (long)(cinfo->max_v_samp_factor * data_unit));
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_width * (long)compptr->h_samp_factor,
+                    (long)cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_height * (long)compptr->v_samp_factor,
+                    (long)cinfo->max_v_samp_factor);
+    /* Mark component needed (this flag isn't actually used for compression) */
+    compptr->component_needed = TRUE;
+  }
+
+  /* Compute number of fully interleaved MCU rows (number of times that
+   * main controller will call coefficient or difference controller).
+   */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long)cinfo->_jpeg_height,
+                  (long)(cinfo->max_v_samp_factor * data_unit));
+}
+
+
+#if defined(C_MULTISCAN_FILES_SUPPORTED) || defined(C_LOSSLESS_SUPPORTED)
+#define NEED_SCAN_SCRIPT
+#endif
+
+#ifdef NEED_SCAN_SCRIPT
+
+LOCAL(void)
+validate_script(j_compress_ptr cinfo)
+/* Verify that the scan script in cinfo->scan_info[] is valid; also
+ * determine whether it uses progressive JPEG, and set cinfo->progressive_mode.
+ */
+{
+  const jpeg_scan_info *scanptr;
+  int scanno, ncomps, ci, coefi, thisi;
+  int Ss, Se, Ah, Al;
+  boolean component_sent[MAX_COMPONENTS];
+#ifdef C_PROGRESSIVE_SUPPORTED
+  int *last_bitpos_ptr;
+  int last_bitpos[MAX_COMPONENTS][DCTSIZE2];
+  /* -1 until that coefficient has been seen; then last Al for it */
+#endif
+
+  if (cinfo->num_scans <= 0)
+    ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, 0);
+
+#ifndef C_MULTISCAN_FILES_SUPPORTED
+  if (cinfo->num_scans > 1)
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+
+  scanptr = cinfo->scan_info;
+  if (scanptr->Ss != 0 && scanptr->Se == 0) {
+#ifdef C_LOSSLESS_SUPPORTED
+    cinfo->master->lossless = TRUE;
+    cinfo->progressive_mode = FALSE;
+    for (ci = 0; ci < cinfo->num_components; ci++)
+      component_sent[ci] = FALSE;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  }
+  /* For sequential JPEG, all scans must have Ss=0, Se=DCTSIZE2-1;
+   * for progressive JPEG, no scan can have this.
+   */
+  else if (scanptr->Ss != 0 || scanptr->Se != DCTSIZE2 - 1) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    cinfo->progressive_mode = TRUE;
+    cinfo->master->lossless = FALSE;
+    last_bitpos_ptr = &last_bitpos[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++)
+      for (coefi = 0; coefi < DCTSIZE2; coefi++)
+        *last_bitpos_ptr++ = -1;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = cinfo->master->lossless = FALSE;
+    for (ci = 0; ci < cinfo->num_components; ci++)
+      component_sent[ci] = FALSE;
+  }
+
+  for (scanno = 1; scanno <= cinfo->num_scans; scanptr++, scanno++) {
+    /* Validate component indexes */
+    ncomps = scanptr->comps_in_scan;
+    if (ncomps <= 0 || ncomps > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, ncomps, MAX_COMPS_IN_SCAN);
+    for (ci = 0; ci < ncomps; ci++) {
+      thisi = scanptr->component_index[ci];
+      if (thisi < 0 || thisi >= cinfo->num_components)
+        ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+      /* Components must appear in SOF order within each scan */
+      if (ci > 0 && thisi <= scanptr->component_index[ci - 1])
+        ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+    }
+    /* Validate progression parameters */
+    Ss = scanptr->Ss;
+    Se = scanptr->Se;
+    Ah = scanptr->Ah;
+    Al = scanptr->Al;
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      /* Rec. ITU-T T.81 | ISO/IEC 10918-1 simply gives the ranges 0..13 for Ah
+       * and Al, but that seems wrong: the upper bound ought to depend on data
+       * precision.  Perhaps they really meant 0..N+1 for N-bit precision.
+       * Here we allow 0..10 for 8-bit data; Al larger than 10 results in
+       * out-of-range reconstructed DC values during the first DC scan,
+       * which might cause problems for some decoders.
+       */
+      int max_Ah_Al = cinfo->data_precision == 12 ? 13 : 10;
+
+      if (Ss < 0 || Ss >= DCTSIZE2 || Se < Ss || Se >= DCTSIZE2 ||
+          Ah < 0 || Ah > max_Ah_Al || Al < 0 || Al > max_Ah_Al)
+        ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      if (Ss == 0) {
+        if (Se != 0)            /* DC and AC together not OK */
+          ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      } else {
+        if (ncomps != 1)        /* AC scans must be for only one component */
+          ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      }
+      for (ci = 0; ci < ncomps; ci++) {
+        last_bitpos_ptr = &last_bitpos[scanptr->component_index[ci]][0];
+        if (Ss != 0 && last_bitpos_ptr[0] < 0) /* AC without prior DC scan */
+          ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+        for (coefi = Ss; coefi <= Se; coefi++) {
+          if (last_bitpos_ptr[coefi] < 0) {
+            /* first scan of this coefficient */
+            if (Ah != 0)
+              ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+          } else {
+            /* not first scan */
+            if (Ah != last_bitpos_ptr[coefi] || Al != Ah - 1)
+              ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+          }
+          last_bitpos_ptr[coefi] = Al;
+        }
+      }
+#endif
+    } else {
+#ifdef C_LOSSLESS_SUPPORTED
+      if (cinfo->master->lossless) {
+        /* The JPEG spec simply gives the range 0..15 for Al (Pt), but that
+         * seems wrong: the upper bound ought to depend on data precision.
+         * Perhaps they really meant 0..N-1 for N-bit precision, which is what
+         * we allow here.  Values greater than or equal to the data precision
+         * will result in a blank image.
+         */
+        if (Ss < 1 || Ss > 7 ||         /* predictor selection value */
+            Se != 0 || Ah != 0 ||
+            Al < 0 || Al >= cinfo->data_precision) /* point transform */
+          ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      } else
+#endif
+      {
+        /* For sequential JPEG, all progression parameters must be these: */
+        if (Ss != 0 || Se != DCTSIZE2 - 1 || Ah != 0 || Al != 0)
+          ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      }
+      /* Make sure components are not sent twice */
+      for (ci = 0; ci < ncomps; ci++) {
+        thisi = scanptr->component_index[ci];
+        if (component_sent[thisi])
+          ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+        component_sent[thisi] = TRUE;
+      }
+    }
+  }
+
+  /* Now verify that everything got sent. */
+  if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    /* For progressive mode, we only check that at least some DC data
+     * got sent for each component; the spec does not require that all bits
+     * of all coefficients be transmitted.  Would it be wiser to enforce
+     * transmission of all coefficient bits??
+     */
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (last_bitpos[ci][0] < 0)
+        ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+#endif
+  } else {
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (!component_sent[ci])
+        ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+  }
+}
+
+#endif /* NEED_SCAN_SCRIPT */
+
+
+LOCAL(void)
+select_scan_parameters(j_compress_ptr cinfo)
+/* Set up the scan parameters for the current scan */
+{
+  int ci;
+
+#ifdef NEED_SCAN_SCRIPT
+  if (cinfo->scan_info != NULL) {
+    /* Prepare for current scan --- the script is already validated */
+    my_master_ptr master = (my_master_ptr)cinfo->master;
+    const jpeg_scan_info *scanptr = cinfo->scan_info + master->scan_number;
+
+    cinfo->comps_in_scan = scanptr->comps_in_scan;
+    for (ci = 0; ci < scanptr->comps_in_scan; ci++) {
+      cinfo->cur_comp_info[ci] =
+        &cinfo->comp_info[scanptr->component_index[ci]];
+    }
+    cinfo->Ss = scanptr->Ss;
+    cinfo->Se = scanptr->Se;
+    cinfo->Ah = scanptr->Ah;
+    cinfo->Al = scanptr->Al;
+  } else
+#endif
+  {
+    /* Prepare for single sequential-JPEG scan containing all components */
+    if (cinfo->num_components > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+               MAX_COMPS_IN_SCAN);
+    cinfo->comps_in_scan = cinfo->num_components;
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      cinfo->cur_comp_info[ci] = &cinfo->comp_info[ci];
+    }
+    if (!cinfo->master->lossless) {
+      cinfo->Ss = 0;
+      cinfo->Se = DCTSIZE2 - 1;
+      cinfo->Ah = 0;
+      cinfo->Al = 0;
+    }
+  }
+}
+
+
+LOCAL(void)
+per_scan_setup(j_compress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] are already set */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  if (cinfo->comps_in_scan == 1) {
+
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = data_unit;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+
+  } else {
+
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+               MAX_COMPS_IN_SCAN);
+
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_width,
+                    (long)(cinfo->max_h_samp_factor * data_unit));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long)cinfo->_jpeg_height,
+                    (long)(cinfo->max_v_samp_factor * data_unit));
+
+    cinfo->blocks_in_MCU = 0;
+
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * data_unit;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int)(compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int)(compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > C_MAX_BLOCKS_IN_MCU)
+        ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+        cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+
+  }
+
+  /* Convert restart specified in rows to actual MCU count. */
+  /* Note that count must fit in 16 bits, so we provide limiting. */
+  if (cinfo->restart_in_rows > 0) {
+    long nominal = (long)cinfo->restart_in_rows * (long)cinfo->MCUs_per_row;
+    cinfo->restart_interval = (unsigned int)MIN(nominal, 65535L);
+  }
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each pass.  We determine which modules
+ * will be active during this pass and give them appropriate start_pass calls.
+ * We also set is_last_pass to indicate whether any more passes will be
+ * required.
+ */
+
+METHODDEF(void)
+prepare_for_pass(j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  switch (master->pass_type) {
+  case main_pass:
+    /* Initial pass: will collect input data, and do either Huffman
+     * optimization or data output for the first scan.
+     */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (!cinfo->raw_data_in) {
+      (*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->downsample->start_pass) (cinfo);
+      (*cinfo->prep->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+    (*cinfo->fdct->start_pass) (cinfo);
+    (*cinfo->entropy->start_pass) (cinfo, cinfo->optimize_coding);
+    (*cinfo->coef->start_pass) (cinfo,
+                                (master->total_passes > 1 ?
+                                 JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+    (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    if (cinfo->optimize_coding) {
+      /* No immediate data output; postpone writing frame/scan headers */
+      master->pub.call_pass_startup = FALSE;
+    } else {
+      /* Will write frame/scan headers at first jpeg_write_scanlines call */
+      master->pub.call_pass_startup = TRUE;
+    }
+    break;
+#ifdef ENTROPY_OPT_SUPPORTED
+  case huff_opt_pass:
+    /* Do Huffman optimization for a scan after the first one. */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (cinfo->Ss != 0 || cinfo->Ah == 0 || cinfo->arith_code ||
+        cinfo->master->lossless) {
+      (*cinfo->entropy->start_pass) (cinfo, TRUE);
+      (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+      master->pub.call_pass_startup = FALSE;
+      break;
+    }
+    /* Special case: Huffman DC refinement scans need no Huffman table
+     * and therefore we can skip the optimization pass for them.
+     */
+    master->pass_type = output_pass;
+    master->pass_number++;
+#endif
+    FALLTHROUGH                 /*FALLTHROUGH*/
+  case output_pass:
+    /* Do a data-output pass. */
+    /* We need not repeat per-scan setup if prior optimization pass did it. */
+    if (!cinfo->optimize_coding) {
+      select_scan_parameters(cinfo);
+      per_scan_setup(cinfo);
+    }
+    (*cinfo->entropy->start_pass) (cinfo, FALSE);
+    (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+    /* We emit frame/scan headers now */
+    if (master->scan_number == 0)
+      (*cinfo->marker->write_frame_header) (cinfo);
+    (*cinfo->marker->write_scan_header) (cinfo);
+    master->pub.call_pass_startup = FALSE;
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+  }
+
+  master->pub.is_last_pass = (master->pass_number == master->total_passes - 1);
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->total_passes;
+  }
+}
+
+
+/*
+ * Special start-of-pass hook.
+ * This is called by jpeg_write_scanlines if call_pass_startup is TRUE.
+ * In single-pass processing, we need this hook because we don't want to
+ * write frame/scan headers during jpeg_start_compress; we want to let the
+ * application write COM markers etc. between jpeg_start_compress and the
+ * jpeg_write_scanlines loop.
+ * In multi-pass processing, this routine is not used.
+ */
+
+METHODDEF(void)
+pass_startup(j_compress_ptr cinfo)
+{
+  cinfo->master->call_pass_startup = FALSE; /* reset flag so call only once */
+
+  (*cinfo->marker->write_frame_header) (cinfo);
+  (*cinfo->marker->write_scan_header) (cinfo);
+}
+
+
+/*
+ * Finish up at end of pass.
+ */
+
+METHODDEF(void)
+finish_pass_master(j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  /* The entropy coder always needs an end-of-pass call,
+   * either to analyze statistics or to flush its output buffer.
+   */
+  (*cinfo->entropy->finish_pass) (cinfo);
+
+  /* Update state for next pass */
+  switch (master->pass_type) {
+  case main_pass:
+    /* next pass is either output of scan 0 (after optimization)
+     * or output of scan 1 (if no optimization).
+     */
+    master->pass_type = output_pass;
+    if (!cinfo->optimize_coding)
+      master->scan_number++;
+    break;
+  case huff_opt_pass:
+    /* next pass is always output of current scan */
+    master->pass_type = output_pass;
+    break;
+  case output_pass:
+    /* next pass is either optimization or output of next scan */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    master->scan_number++;
+    break;
+  }
+
+  master->pass_number++;
+}
+
+
+/*
+ * Initialize master compression control.
+ */
+
+GLOBAL(void)
+jinit_c_master_control(j_compress_ptr cinfo, boolean transcode_only)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+  boolean empty_huff_tables = TRUE;
+  int i;
+
+  master->pub.prepare_for_pass = prepare_for_pass;
+  master->pub.pass_startup = pass_startup;
+  master->pub.finish_pass = finish_pass_master;
+  master->pub.is_last_pass = FALSE;
+
+  if (cinfo->scan_info != NULL) {
+#ifdef NEED_SCAN_SCRIPT
+    validate_script(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    cinfo->num_scans = 1;
+  }
+
+#ifdef C_LOSSLESS_SUPPORTED
+  /* Disable smoothing and subsampling in lossless mode, since those are lossy
+   * algorithms.  Set the JPEG colorspace to the input colorspace.  Disable raw
+   * (downsampled) data input, because it isn't particularly useful without
+   * subsampling and has not been tested in lossless mode.
+   */
+  if (cinfo->master->lossless) {
+    int ci;
+    jpeg_component_info *compptr;
+
+    cinfo->raw_data_in = FALSE;
+    cinfo->smoothing_factor = 0;
+    jpeg_default_colorspace(cinfo);
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++)
+      compptr->h_samp_factor = compptr->v_samp_factor = 1;
+  }
+#endif
+
+  /* Validate parameters, determine derived values */
+  initial_setup(cinfo, transcode_only);
+
+  if (cinfo->arith_code)
+    cinfo->optimize_coding = FALSE;
+  else {
+    if (cinfo->master->lossless ||      /*  TEMPORARY HACK ??? */
+        cinfo->progressive_mode)
+      cinfo->optimize_coding = TRUE; /* assume default tables no good for
+                                        progressive mode or lossless mode */
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      if (cinfo->dc_huff_tbl_ptrs[i] != NULL ||
+          cinfo->ac_huff_tbl_ptrs[i] != NULL) {
+        empty_huff_tables = FALSE;
+        break;
+      }
+    }
+    if (cinfo->data_precision == 12 && !cinfo->optimize_coding &&
+        (empty_huff_tables || using_std_huff_tables(cinfo)))
+      cinfo->optimize_coding = TRUE; /* assume default tables no good for
+                                        12-bit data precision */
+  }
+
+  /* Initialize my private state */
+  if (transcode_only) {
+    /* no main pass in transcoding */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    else
+      master->pass_type = output_pass;
+  } else {
+    /* for normal compression, first pass is always this type: */
+    master->pass_type = main_pass;
+  }
+  master->scan_number = 0;
+  master->pass_number = 0;
+  if (cinfo->optimize_coding)
+    master->total_passes = cinfo->num_scans * 2;
+  else
+    master->total_passes = cinfo->num_scans;
+
+  master->jpeg_version = PACKAGE_NAME " version " VERSION " (build " BUILD ")";
+}

thirdparty/libjpeg-turbo/src/jcmaster.h

@@ -0,0 +1,43 @@
+/*
+ * jcmaster.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1995, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2016, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains master control structure for the JPEG compressor.
+ */
+
+/* Private state */
+
+typedef enum {
+  main_pass,                    /* input data, also do first output step */
+  huff_opt_pass,                /* Huffman code optimization pass */
+  output_pass                   /* data output pass */
+} c_pass_type;
+
+typedef struct {
+  struct jpeg_comp_master pub;  /* public fields */
+
+  c_pass_type pass_type;        /* the type of the current pass */
+
+  int pass_number;              /* # of passes completed */
+  int total_passes;             /* total # of passes needed */
+
+  int scan_number;              /* current index in scan_info[] */
+
+  /*
+   * This is here so we can add libjpeg-turbo version/build information to the
+   * global string table without introducing a new global symbol.  Adding this
+   * information to the global string table allows one to examine a binary
+   * object and determine which version of libjpeg-turbo it was built from or
+   * linked against.
+   */
+  const char *jpeg_version;
+
+} my_comp_master;
+
+typedef my_comp_master *my_master_ptr;

thirdparty/libjpeg-turbo/src/jcomapi.c

@@ -0,0 +1,110 @@
+/*
+ * jcomapi.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains application interface routines that are used for both
+ * compression and decompression.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Abort processing of a JPEG compression or decompression operation,
+ * but don't destroy the object itself.
+ *
+ * For this, we merely clean up all the nonpermanent memory pools.
+ * Note that temp files (virtual arrays) are not allowed to belong to
+ * the permanent pool, so we will be able to close all temp files here.
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_abort(j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Do nothing if called on a not-initialized or destroyed JPEG object. */
+  if (cinfo->mem == NULL)
+    return;
+
+  /* Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS - 1; pool > JPOOL_PERMANENT; pool--) {
+    (*cinfo->mem->free_pool) (cinfo, pool);
+  }
+
+  /* Reset overall state for possible reuse of object */
+  if (cinfo->is_decompressor) {
+    cinfo->global_state = DSTATE_START;
+    /* Try to keep application from accessing now-deleted marker list.
+     * A bit kludgy to do it here, but this is the most central place.
+     */
+    ((j_decompress_ptr)cinfo)->marker_list = NULL;
+    ((j_decompress_ptr)cinfo)->master->marker_list_end = NULL;
+  } else {
+    cinfo->global_state = CSTATE_START;
+  }
+}
+
+
+/*
+ * Destruction of a JPEG object.
+ *
+ * Everything gets deallocated except the master jpeg_compress_struct itself
+ * and the error manager struct.  Both of these are supplied by the application
+ * and must be freed, if necessary, by the application.  (Often they are on
+ * the stack and so don't need to be freed anyway.)
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_destroy(j_common_ptr cinfo)
+{
+  /* We need only tell the memory manager to release everything. */
+  /* NB: mem pointer is NULL if memory mgr failed to initialize. */
+  if (cinfo->mem != NULL)
+    (*cinfo->mem->self_destruct) (cinfo);
+  cinfo->mem = NULL;            /* be safe if jpeg_destroy is called twice */
+  cinfo->global_state = 0;      /* mark it destroyed */
+}
+
+
+/*
+ * Convenience routines for allocating quantization and Huffman tables.
+ * (Would jutils.c be a more reasonable place to put these?)
+ */
+
+GLOBAL(JQUANT_TBL *)
+jpeg_alloc_quant_table(j_common_ptr cinfo)
+{
+  JQUANT_TBL *tbl;
+
+  tbl = (JQUANT_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, sizeof(JQUANT_TBL));
+  tbl->sent_table = FALSE;      /* make sure this is false in any new table */
+  return tbl;
+}
+
+
+GLOBAL(JHUFF_TBL *)
+jpeg_alloc_huff_table(j_common_ptr cinfo)
+{
+  JHUFF_TBL *tbl;
+
+  tbl = (JHUFF_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, sizeof(JHUFF_TBL));
+  tbl->sent_table = FALSE;      /* make sure this is false in any new table */
+  return tbl;
+}

thirdparty/libjpeg-turbo/src/jconfig.h

@@ -0,0 +1,62 @@
+// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
+
+/* Version ID for the JPEG library.
+ * Might be useful for tests like "#if JPEG_LIB_VERSION >= 60".
+ */
+#define JPEG_LIB_VERSION  62
+
+/* libjpeg-turbo version */
+#define LIBJPEG_TURBO_VERSION  3.1.0
+
+/* libjpeg-turbo version in integer form */
+#define LIBJPEG_TURBO_VERSION_NUMBER  3001000
+
+/* Support arithmetic encoding when using 8-bit samples */
+#define C_ARITH_CODING_SUPPORTED 1
+
+/* Support arithmetic decoding when using 8-bit samples */
+#define D_ARITH_CODING_SUPPORTED 1
+
+/* Support in-memory source/destination managers */
+#define MEM_SRCDST_SUPPORTED  1
+
+/* Use accelerated SIMD routines when using 8-bit samples */
+//#define WITH_SIMD 1
+
+/* This version of libjpeg-turbo supports run-time selection of data precision,
+ * so BITS_IN_JSAMPLE is no longer used to specify the data precision at build
+ * time.  However, some downstream software expects the macro to be defined.
+ * Since 12-bit data precision is an opt-in feature that requires explicitly
+ * calling 12-bit-specific libjpeg API functions and using 12-bit-specific data
+ * types, the unmodified portion of the libjpeg API still behaves as if it were
+ * built for 8-bit precision, and JSAMPLE is still literally an 8-bit data
+ * type.  Thus, it is correct to define BITS_IN_JSAMPLE to 8 here.
+ */
+#ifndef BITS_IN_JSAMPLE
+#define BITS_IN_JSAMPLE  8
+#endif
+
+#ifdef _WIN32
+
+#undef RIGHT_SHIFT_IS_UNSIGNED
+
+/* Define "boolean" as unsigned char, not int, per Windows custom */
+#ifndef __RPCNDR_H__            /* don't conflict if rpcndr.h already read */
+typedef unsigned char boolean;
+#endif
+#define HAVE_BOOLEAN            /* prevent jmorecfg.h from redefining it */
+
+/* Define "INT32" as int, not long, per Windows custom */
+#if !(defined(_BASETSD_H_) || defined(_BASETSD_H))   /* don't conflict if basetsd.h already read */
+typedef short INT16;
+typedef signed int INT32;
+#endif
+#define XMD_H                   /* prevent jmorecfg.h from redefining it */
+
+#else
+
+/* Define if your (broken) compiler shifts signed values as if they were
+   unsigned. */
+/* #undef RIGHT_SHIFT_IS_UNSIGNED */
+
+#endif

thirdparty/libjpeg-turbo/src/jconfigint.h

@@ -0,0 +1,94 @@
+// Originally generated by libjpeg-turbo's cmake build, then modified to support multiple platforms.
+
+/* libjpeg-turbo build number */
+#define BUILD  "20250317"
+
+/* How to hide global symbols. */
+#ifndef HIDDEN
+	#if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+		#define HIDDEN  __attribute__((visibility("hidden")))
+	#else
+		#define HIDDEN
+	#endif
+#endif
+
+/* Compiler's inline keyword */
+#undef inline
+
+/* How to obtain function inlining. */
+#if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+	#define INLINE  __inline__ __attribute__((always_inline))
+#else
+	#define INLINE inline
+#endif
+
+/* How to obtain thread-local storage */
+#if defined(_MSC_VER)
+#define THREAD_LOCAL  __declspec(thread)
+#else
+#define THREAD_LOCAL  __thread
+#endif
+
+/* Define to the full name of this package. */
+#define PACKAGE_NAME  "libjpeg-turbo"
+
+/* Version number of package */
+#define VERSION  "3.1.1"
+
+/* The size of `size_t', as computed by sizeof. */
+#define SIZEOF_SIZE_T  8
+
+/* Define if your compiler has __builtin_ctzl() and sizeof(unsigned long) == sizeof(size_t). */
+#if defined(__GNUC__)
+	#define HAVE_BUILTIN_CTZL
+#endif
+
+/* Define to 1 if you have the <intrin.h> header file. */
+/* #undef HAVE_INTRIN_H */
+
+#if defined(_MSC_VER) && defined(HAVE_INTRIN_H)
+#if (SIZEOF_SIZE_T == 8)
+#define HAVE_BITSCANFORWARD64
+#elif (SIZEOF_SIZE_T == 4)
+#define HAVE_BITSCANFORWARD
+#endif
+#endif
+
+#if defined(__has_attribute)
+#if __has_attribute(fallthrough)
+#define FALLTHROUGH  __attribute__((fallthrough));
+#else
+#define FALLTHROUGH
+#endif
+#else
+#define FALLTHROUGH
+#endif
+
+/*
+ * Define BITS_IN_JSAMPLE as either
+ *   8   for 8-bit sample values (the usual setting)
+ *   12  for 12-bit sample values
+ * Only 8 and 12 are legal data precisions for lossy JPEG according to the
+ * JPEG standard, and the IJG code does not support anything else!
+ */
+
+#ifndef BITS_IN_JSAMPLE
+#define BITS_IN_JSAMPLE  8      /* use 8 or 12 */
+#endif
+
+#undef C_ARITH_CODING_SUPPORTED
+#undef D_ARITH_CODING_SUPPORTED
+#undef WITH_SIMD
+
+#if BITS_IN_JSAMPLE == 8
+
+/* Support arithmetic encoding */
+#define C_ARITH_CODING_SUPPORTED 1
+
+/* Support arithmetic decoding */
+#define D_ARITH_CODING_SUPPORTED 1
+
+/* Use accelerated SIMD routines. */
+//#define WITH_SIMD 1
+
+#endif

thirdparty/libjpeg-turbo/src/jcparam.c

@@ -0,0 +1,592 @@
+/*
+ * jcparam.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2008 by Guido Vollbeding.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009-2011, 2018, 2023-2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains optional default-setting code for the JPEG compressor.
+ * Applications do not have to use this file, but those that don't use it
+ * must know a lot more about the innards of the JPEG code.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jstdhuff.c"
+
+
+/*
+ * Quantization table setup routines
+ */
+
+GLOBAL(void)
+jpeg_add_quant_table(j_compress_ptr cinfo, int which_tbl,
+                     const unsigned int *basic_table, int scale_factor,
+                     boolean force_baseline)
+/* Define a quantization table equal to the basic_table times
+ * a scale factor (given as a percentage).
+ * If force_baseline is TRUE, the computed quantization table entries
+ * are limited to 1..255 for JPEG baseline compatibility.
+ */
+{
+  JQUANT_TBL **qtblptr;
+  int i;
+  long temp;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (which_tbl < 0 || which_tbl >= NUM_QUANT_TBLS)
+    ERREXIT1(cinfo, JERR_DQT_INDEX, which_tbl);
+
+  qtblptr = &cinfo->quant_tbl_ptrs[which_tbl];
+
+  if (*qtblptr == NULL)
+    *qtblptr = jpeg_alloc_quant_table((j_common_ptr)cinfo);
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = ((long)basic_table[i] * scale_factor + 50L) / 100L;
+    /* limit the values to the valid range */
+    if (temp <= 0L) temp = 1L;
+    if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
+    if (force_baseline && temp > 255L)
+      temp = 255L;              /* limit to baseline range if requested */
+    (*qtblptr)->quantval[i] = (UINT16)temp;
+  }
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*qtblptr)->sent_table = FALSE;
+}
+
+
+/* These are the sample quantization tables given in Annex K (Clause K.1) of
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ * The spec says that the values given produce "good" quality, and
+ * when divided by 2, "very good" quality.
+ */
+static const unsigned int std_luminance_quant_tbl[DCTSIZE2] = {
+  16,  11,  10,  16,  24,  40,  51,  61,
+  12,  12,  14,  19,  26,  58,  60,  55,
+  14,  13,  16,  24,  40,  57,  69,  56,
+  14,  17,  22,  29,  51,  87,  80,  62,
+  18,  22,  37,  56,  68, 109, 103,  77,
+  24,  35,  55,  64,  81, 104, 113,  92,
+  49,  64,  78,  87, 103, 121, 120, 101,
+  72,  92,  95,  98, 112, 100, 103,  99
+};
+static const unsigned int std_chrominance_quant_tbl[DCTSIZE2] = {
+  17,  18,  24,  47,  99,  99,  99,  99,
+  18,  21,  26,  66,  99,  99,  99,  99,
+  24,  26,  56,  99,  99,  99,  99,  99,
+  47,  66,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99
+};
+
+
+#if JPEG_LIB_VERSION >= 70
+GLOBAL(void)
+jpeg_default_qtables(j_compress_ptr cinfo, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and straight percentage-scaling quality scales.
+ * This entry point allows different scalings for luminance and chrominance.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+                       cinfo->q_scale_factor[0], force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+                       cinfo->q_scale_factor[1], force_baseline);
+}
+#endif
+
+
+GLOBAL(void)
+jpeg_set_linear_quality(j_compress_ptr cinfo, int scale_factor,
+                        boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and a straight percentage-scaling quality scale.  In most cases it's better
+ * to use jpeg_set_quality (below); this entry point is provided for
+ * applications that insist on a linear percentage scaling.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+                       scale_factor, force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+                       scale_factor, force_baseline);
+}
+
+
+GLOBAL(int)
+jpeg_quality_scaling(int quality)
+/* Convert a user-specified quality rating to a percentage scaling factor
+ * for an underlying quantization table, using our recommended scaling curve.
+ * The input 'quality' factor should be 0 (terrible) to 100 (very good).
+ */
+{
+  /* Safety limit on quality factor.  Convert 0 to 1 to avoid zero divide. */
+  if (quality <= 0) quality = 1;
+  if (quality > 100) quality = 100;
+
+  /* The basic table is used as-is (scaling 100) for a quality of 50.
+   * Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
+   * note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
+   * to make all the table entries 1 (hence, minimum quantization loss).
+   * Qualities 1..50 are converted to scaling percentage 5000/Q.
+   */
+  if (quality < 50)
+    quality = 5000 / quality;
+  else
+    quality = 200 - quality * 2;
+
+  return quality;
+}
+
+
+GLOBAL(void)
+jpeg_set_quality(j_compress_ptr cinfo, int quality, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables.
+ * This is the standard quality-adjusting entry point for typical user
+ * interfaces; only those who want detailed control over quantization tables
+ * would use the preceding three routines directly.
+ */
+{
+  /* Convert user 0-100 rating to percentage scaling */
+  quality = jpeg_quality_scaling(quality);
+
+  /* Set up standard quality tables */
+  jpeg_set_linear_quality(cinfo, quality, force_baseline);
+}
+
+
+/*
+ * Default parameter setup for compression.
+ *
+ * Applications that don't choose to use this routine must do their
+ * own setup of all these parameters.  Alternately, you can call this
+ * to establish defaults and then alter parameters selectively.  This
+ * is the recommended approach since, if we add any new parameters,
+ * your code will still work (they'll be set to reasonable defaults).
+ */
+
+GLOBAL(void)
+jpeg_set_defaults(j_compress_ptr cinfo)
+{
+  int i;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Allocate comp_info array large enough for maximum component count.
+   * Array is made permanent in case application wants to compress
+   * multiple images at same param settings.
+   */
+  if (cinfo->comp_info == NULL)
+    cinfo->comp_info = (jpeg_component_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  MAX_COMPONENTS * sizeof(jpeg_component_info));
+
+  /* Initialize everything not dependent on the color space */
+
+#if JPEG_LIB_VERSION >= 70
+  cinfo->scale_num = 1;         /* 1:1 scaling */
+  cinfo->scale_denom = 1;
+#endif
+  /* Set up two quantization tables using default quality of 75 */
+  jpeg_set_quality(cinfo, 75, TRUE);
+  /* Set up two Huffman tables */
+  std_huff_tables((j_common_ptr)cinfo);
+
+  /* Initialize default arithmetic coding conditioning */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+
+  /* Default is no multiple-scan output */
+  cinfo->scan_info = NULL;
+  cinfo->num_scans = 0;
+
+  /* Default is lossy output */
+  cinfo->master->lossless = FALSE;
+
+  /* Expect normal source image, not raw downsampled data */
+  cinfo->raw_data_in = FALSE;
+
+  /* Use Huffman coding, not arithmetic coding, by default */
+  cinfo->arith_code = FALSE;
+
+  /* By default, don't do extra passes to optimize entropy coding */
+  cinfo->optimize_coding = FALSE;
+  /* The standard Huffman tables are only valid for 8-bit data precision.
+   * If the precision is higher, force optimization on so that usable
+   * tables will be computed.  This test can be removed if default tables
+   * are supplied that are valid for the desired precision.
+   */
+  if (cinfo->data_precision == 12)
+    cinfo->optimize_coding = TRUE;
+
+  /* By default, use the simpler non-cosited sampling alignment */
+  cinfo->CCIR601_sampling = FALSE;
+
+#if JPEG_LIB_VERSION >= 70
+  /* By default, apply fancy downsampling */
+  cinfo->do_fancy_downsampling = TRUE;
+#endif
+
+  /* No input smoothing */
+  cinfo->smoothing_factor = 0;
+
+  /* DCT algorithm preference */
+  cinfo->dct_method = JDCT_DEFAULT;
+
+  /* No restart markers */
+  cinfo->restart_interval = 0;
+  cinfo->restart_in_rows = 0;
+
+  /* Fill in default JFIF marker parameters.  Note that whether the marker
+   * will actually be written is determined by jpeg_set_colorspace.
+   *
+   * By default, the library emits JFIF version code 1.01.
+   * An application that wants to emit JFIF 1.02 extension markers should set
+   * JFIF_minor_version to 2.  We could probably get away with just defaulting
+   * to 1.02, but there may still be some decoders in use that will complain
+   * about that; saying 1.01 should minimize compatibility problems.
+   */
+  cinfo->JFIF_major_version = 1; /* Default JFIF version = 1.01 */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;      /* Pixel size is unknown by default */
+  cinfo->X_density = 1;         /* Pixel aspect ratio is square by default */
+  cinfo->Y_density = 1;
+
+  /* Choose JPEG colorspace based on input space, set defaults accordingly */
+
+  jpeg_default_colorspace(cinfo);
+}
+
+
+/*
+ * Select an appropriate JPEG colorspace for in_color_space.
+ */
+
+GLOBAL(void)
+jpeg_default_colorspace(j_compress_ptr cinfo)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    jpeg_set_colorspace(cinfo, JCS_GRAYSCALE);
+    break;
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+  case JCS_EXT_RGBA:
+  case JCS_EXT_BGRA:
+  case JCS_EXT_ABGR:
+  case JCS_EXT_ARGB:
+#ifdef C_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless)
+      jpeg_set_colorspace(cinfo, JCS_RGB);
+    else
+#endif
+      jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_YCbCr:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_CMYK:
+    jpeg_set_colorspace(cinfo, JCS_CMYK); /* By default, no translation */
+    break;
+  case JCS_YCCK:
+    jpeg_set_colorspace(cinfo, JCS_YCCK);
+    break;
+  case JCS_UNKNOWN:
+    jpeg_set_colorspace(cinfo, JCS_UNKNOWN);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+  }
+}
+
+
+/*
+ * Set the JPEG colorspace, and choose colorspace-dependent default values.
+ */
+
+GLOBAL(void)
+jpeg_set_colorspace(j_compress_ptr cinfo, J_COLOR_SPACE colorspace)
+{
+  jpeg_component_info *compptr;
+  int ci;
+
+#define SET_COMP(index, id, hsamp, vsamp, quant, dctbl, actbl) \
+  (compptr = &cinfo->comp_info[index], \
+   compptr->component_id = (id), \
+   compptr->h_samp_factor = (hsamp), \
+   compptr->v_samp_factor = (vsamp), \
+   compptr->quant_tbl_no = (quant), \
+   compptr->dc_tbl_no = (dctbl), \
+   compptr->ac_tbl_no = (actbl) )
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* For all colorspaces, we use Q and Huff tables 0 for luminance components,
+   * tables 1 for chrominance components.
+   */
+
+  cinfo->jpeg_color_space = colorspace;
+
+  cinfo->write_JFIF_header = FALSE; /* No marker for non-JFIF colorspaces */
+  cinfo->write_Adobe_marker = FALSE; /* write no Adobe marker by default */
+
+  switch (colorspace) {
+  case JCS_GRAYSCALE:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 1;
+    /* JFIF specifies component ID 1 */
+    SET_COMP(0, 1, 1, 1, 0, 0, 0);
+    break;
+  case JCS_RGB:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag RGB */
+    cinfo->num_components = 3;
+    SET_COMP(0, 0x52 /* 'R' */, 1, 1, 0, 0, 0);
+    SET_COMP(1, 0x47 /* 'G' */, 1, 1, 0, 0, 0);
+    SET_COMP(2, 0x42 /* 'B' */, 1, 1, 0, 0, 0);
+    break;
+  case JCS_YCbCr:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 3;
+    /* JFIF specifies component IDs 1,2,3 */
+    /* We default to 2x2 subsamples of chrominance */
+    SET_COMP(0, 1, 2, 2, 0, 0, 0);
+    SET_COMP(1, 2, 1, 1, 1, 1, 1);
+    SET_COMP(2, 3, 1, 1, 1, 1, 1);
+    break;
+  case JCS_CMYK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag CMYK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 0x43 /* 'C' */, 1, 1, 0, 0, 0);
+    SET_COMP(1, 0x4D /* 'M' */, 1, 1, 0, 0, 0);
+    SET_COMP(2, 0x59 /* 'Y' */, 1, 1, 0, 0, 0);
+    SET_COMP(3, 0x4B /* 'K' */, 1, 1, 0, 0, 0);
+    break;
+  case JCS_YCCK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag YCCK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 1, 2, 2, 0, 0, 0);
+    SET_COMP(1, 2, 1, 1, 1, 1, 1);
+    SET_COMP(2, 3, 1, 1, 1, 1, 1);
+    SET_COMP(3, 4, 2, 2, 0, 0, 0);
+    break;
+  case JCS_UNKNOWN:
+    cinfo->num_components = cinfo->input_components;
+    if (cinfo->num_components < 1 || cinfo->num_components > MAX_COMPONENTS)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+               MAX_COMPONENTS);
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      SET_COMP(ci, ci, 1, 1, 0, 0, 0);
+    }
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+  }
+}
+
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+LOCAL(jpeg_scan_info *)
+fill_a_scan(jpeg_scan_info *scanptr, int ci, int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for specified component */
+{
+  scanptr->comps_in_scan = 1;
+  scanptr->component_index[0] = ci;
+  scanptr->Ss = Ss;
+  scanptr->Se = Se;
+  scanptr->Ah = Ah;
+  scanptr->Al = Al;
+  scanptr++;
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_scans(jpeg_scan_info *scanptr, int ncomps, int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for each component */
+{
+  int ci;
+
+  for (ci = 0; ci < ncomps; ci++) {
+    scanptr->comps_in_scan = 1;
+    scanptr->component_index[0] = ci;
+    scanptr->Ss = Ss;
+    scanptr->Se = Se;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  }
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_dc_scans(jpeg_scan_info *scanptr, int ncomps, int Ah, int Al)
+/* Support routine: generate interleaved DC scan if possible, else N scans */
+{
+  int ci;
+
+  if (ncomps <= MAX_COMPS_IN_SCAN) {
+    /* Single interleaved DC scan */
+    scanptr->comps_in_scan = ncomps;
+    for (ci = 0; ci < ncomps; ci++)
+      scanptr->component_index[ci] = ci;
+    scanptr->Ss = scanptr->Se = 0;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  } else {
+    /* Noninterleaved DC scan for each component */
+    scanptr = fill_scans(scanptr, ncomps, 0, 0, Ah, Al);
+  }
+  return scanptr;
+}
+
+
+/*
+ * Create a recommended progressive-JPEG script.
+ * cinfo->num_components and cinfo->jpeg_color_space must be correct.
+ */
+
+GLOBAL(void)
+jpeg_simple_progression(j_compress_ptr cinfo)
+{
+  int ncomps = cinfo->num_components;
+  int nscans;
+  jpeg_scan_info *scanptr;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+    cinfo->master->lossless = FALSE;
+    jpeg_default_colorspace(cinfo);
+  }
+#endif
+
+  /* Figure space needed for script.  Calculation must match code below! */
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    nscans = 10;
+  } else {
+    /* All-purpose script for other color spaces. */
+    if (ncomps > MAX_COMPS_IN_SCAN)
+      nscans = 6 * ncomps;      /* 2 DC + 4 AC scans per component */
+    else
+      nscans = 2 + 4 * ncomps;  /* 2 DC scans; 4 AC scans per component */
+  }
+
+  /* Allocate space for script.
+   * We need to put it in the permanent pool in case the application performs
+   * multiple compressions without changing the settings.  To avoid a memory
+   * leak if jpeg_simple_progression is called repeatedly for the same JPEG
+   * object, we try to re-use previously allocated space, and we allocate
+   * enough space to handle YCbCr even if initially asked for grayscale.
+   */
+  if (cinfo->script_space == NULL || cinfo->script_space_size < nscans) {
+    cinfo->script_space_size = MAX(nscans, 10);
+    cinfo->script_space = (jpeg_scan_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                        cinfo->script_space_size * sizeof(jpeg_scan_info));
+  }
+  scanptr = cinfo->script_space;
+  cinfo->scan_info = scanptr;
+  cinfo->num_scans = nscans;
+
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    /* Initial DC scan */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    /* Initial AC scan: get some luma data out in a hurry */
+    scanptr = fill_a_scan(scanptr, 0, 1, 5, 0, 2);
+    /* Chroma data is too small to be worth expending many scans on */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 0, 1);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 0, 1);
+    /* Complete spectral selection for luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 6, 63, 0, 2);
+    /* Refine next bit of luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 2, 1);
+    /* Finish DC successive approximation */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    /* Finish AC successive approximation */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 1, 0);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 1, 0);
+    /* Luma bottom bit comes last since it's usually largest scan */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 1, 0);
+  } else {
+    /* All-purpose script for other color spaces. */
+    /* Successive approximation first pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    scanptr = fill_scans(scanptr, ncomps, 1, 5, 0, 2);
+    scanptr = fill_scans(scanptr, ncomps, 6, 63, 0, 2);
+    /* Successive approximation second pass */
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 2, 1);
+    /* Successive approximation final pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 1, 0);
+  }
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */
+
+
+#ifdef C_LOSSLESS_SUPPORTED
+
+/*
+ * Enable lossless mode.
+ */
+
+GLOBAL(void)
+jpeg_enable_lossless(j_compress_ptr cinfo, int predictor_selection_value,
+                     int point_transform)
+{
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  cinfo->master->lossless = TRUE;
+  cinfo->Ss = predictor_selection_value;
+  cinfo->Se = 0;
+  cinfo->Ah = 0;
+  cinfo->Al = point_transform;
+
+  /* The JPEG spec simply gives the range 0..15 for Al (Pt), but that seems
+   * wrong: the upper bound ought to depend on data precision.  Perhaps they
+   * really meant 0..N-1 for N-bit precision, which is what we allow here.
+   * Values greater than or equal to the data precision will result in a blank
+   * image.
+   */
+  if (cinfo->Ss < 1 || cinfo->Ss > 7 ||
+      cinfo->Al < 0 || cinfo->Al >= cinfo->data_precision)
+    ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+             cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+}
+
+#endif /* C_LOSSLESS_SUPPORTED */

thirdparty/libjpeg-turbo/src/jcphuff.c

@@ -0,0 +1,1102 @@
+/*
+ * jcphuff.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2011, 2015, 2018, 2021-2022, 2024, D. R. Commander.
+ * Copyright (C) 2016, 2018, 2022, Matthieu Darbois.
+ * Copyright (C) 2020, Arm Limited.
+ * Copyright (C) 2021, Alex Richardson.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains Huffman entropy encoding routines for progressive JPEG.
+ *
+ * We do not support output suspension in this module, since the library
+ * currently does not allow multiple-scan files to be written with output
+ * suspension.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#ifdef WITH_SIMD
+#include "jsimd.h"
+#else
+#include "jchuff.h"             /* Declarations shared with jc*huff.c */
+#endif
+#include <limits.h>
+
+#ifdef HAVE_INTRIN_H
+#include <intrin.h>
+#ifdef _MSC_VER
+#ifdef HAVE_BITSCANFORWARD64
+#pragma intrinsic(_BitScanForward64)
+#endif
+#ifdef HAVE_BITSCANFORWARD
+#pragma intrinsic(_BitScanForward)
+#endif
+#endif
+#endif
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+#include "jpeg_nbits.h"
+
+
+/* Expanded entropy encoder object for progressive Huffman encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  /* Pointer to routine to prepare data for encode_mcu_AC_first() */
+  void (*AC_first_prepare) (const JCOEF *block,
+                            const int *jpeg_natural_order_start, int Sl,
+                            int Al, UJCOEF *values, size_t *zerobits);
+  /* Pointer to routine to prepare data for encode_mcu_AC_refine() */
+  int (*AC_refine_prepare) (const JCOEF *block,
+                            const int *jpeg_natural_order_start, int Sl,
+                            int Al, UJCOEF *absvalues, size_t *bits);
+
+  /* Mode flag: TRUE for optimization, FALSE for actual data output */
+  boolean gather_statistics;
+
+  /* Bit-level coding status.
+   * next_output_byte/free_in_buffer are local copies of cinfo->dest fields.
+   */
+  JOCTET *next_output_byte;     /* => next byte to write in buffer */
+  size_t free_in_buffer;        /* # of byte spaces remaining in buffer */
+  size_t put_buffer;            /* current bit-accumulation buffer */
+  int put_bits;                 /* # of bits now in it */
+  j_compress_ptr cinfo;         /* link to cinfo (needed for dump_buffer) */
+
+  /* Coding status for DC components */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+
+  /* Coding status for AC components */
+  int ac_tbl_no;                /* the table number of the single component */
+  unsigned int EOBRUN;          /* run length of EOBs */
+  unsigned int BE;              /* # of buffered correction bits before MCU */
+  char *bit_buffer;             /* buffer for correction bits (1 per char) */
+  /* packing correction bits tightly would save some space but cost time... */
+
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+  int next_restart_num;         /* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan).
+   * Since any one scan codes only DC or only AC, we only need one set
+   * of tables, not one for DC and one for AC.
+   */
+  c_derived_tbl *derived_tbls[NUM_HUFF_TBLS];
+
+  /* Statistics tables for optimization; again, one set is enough */
+  long *count_ptrs[NUM_HUFF_TBLS];
+} phuff_entropy_encoder;
+
+typedef phuff_entropy_encoder *phuff_entropy_ptr;
+
+/* MAX_CORR_BITS is the number of bits the AC refinement correction-bit
+ * buffer can hold.  Larger sizes may slightly improve compression, but
+ * 1000 is already well into the realm of overkill.
+ * The minimum safe size is 64 bits.
+ */
+
+#define MAX_CORR_BITS  1000     /* Max # of correction bits I can buffer */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than JLONG.
+ * We assume that int right shift is unsigned if JLONG right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS    int ishift_temp;
+#define IRIGHT_SHIFT(x, shft) \
+  ((ishift_temp = (x)) < 0 ? \
+   (ishift_temp >> (shft)) | ((~0) << (16 - (shft))) : \
+   (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x, shft)   ((x) >> (shft))
+#endif
+
+#define PAD(v, p)  ((v + (p) - 1) & (~((p) - 1)))
+
+/* Forward declarations */
+METHODDEF(boolean) encode_mcu_DC_first(j_compress_ptr cinfo,
+                                       JBLOCKROW *MCU_data);
+METHODDEF(void) encode_mcu_AC_first_prepare
+  (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+   UJCOEF *values, size_t *zerobits);
+METHODDEF(boolean) encode_mcu_AC_first(j_compress_ptr cinfo,
+                                       JBLOCKROW *MCU_data);
+METHODDEF(boolean) encode_mcu_DC_refine(j_compress_ptr cinfo,
+                                        JBLOCKROW *MCU_data);
+METHODDEF(int) encode_mcu_AC_refine_prepare
+  (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+   UJCOEF *absvalues, size_t *bits);
+METHODDEF(boolean) encode_mcu_AC_refine(j_compress_ptr cinfo,
+                                        JBLOCKROW *MCU_data);
+METHODDEF(void) finish_pass_phuff(j_compress_ptr cinfo);
+METHODDEF(void) finish_pass_gather_phuff(j_compress_ptr cinfo);
+
+
+/* Count bit loop zeroes */
+INLINE
+METHODDEF(int)
+count_zeroes(size_t *x)
+{
+#if defined(HAVE_BUILTIN_CTZL)
+  int result;
+  result = __builtin_ctzl(*x);
+  *x >>= result;
+#elif defined(HAVE_BITSCANFORWARD64)
+  unsigned long result;
+  _BitScanForward64(&result, *x);
+  *x >>= result;
+#elif defined(HAVE_BITSCANFORWARD)
+  unsigned long result;
+  _BitScanForward(&result, *x);
+  *x >>= result;
+#else
+  int result = 0;
+  while ((*x & 1) == 0) {
+    ++result;
+    *x >>= 1;
+  }
+#endif
+  return (int)result;
+}
+
+
+/*
+ * Initialize for a Huffman-compressed scan using progressive JPEG.
+ */
+
+METHODDEF(void)
+start_pass_phuff(j_compress_ptr cinfo, boolean gather_statistics)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  boolean is_DC_band;
+  int ci, tbl;
+  jpeg_component_info *compptr;
+
+  entropy->cinfo = cinfo;
+  entropy->gather_statistics = gather_statistics;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* We assume jcmaster.c already validated the scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_first;
+    else
+      entropy->pub.encode_mcu = encode_mcu_AC_first;
+#ifdef WITH_SIMD
+    if (jsimd_can_encode_mcu_AC_first_prepare())
+      entropy->AC_first_prepare = jsimd_encode_mcu_AC_first_prepare;
+    else
+#endif
+      entropy->AC_first_prepare = encode_mcu_AC_first_prepare;
+  } else {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_refine;
+    else {
+      entropy->pub.encode_mcu = encode_mcu_AC_refine;
+#ifdef WITH_SIMD
+      if (jsimd_can_encode_mcu_AC_refine_prepare())
+        entropy->AC_refine_prepare = jsimd_encode_mcu_AC_refine_prepare;
+      else
+#endif
+        entropy->AC_refine_prepare = encode_mcu_AC_refine_prepare;
+      /* AC refinement needs a correction bit buffer */
+      if (entropy->bit_buffer == NULL)
+        entropy->bit_buffer = (char *)
+          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                      MAX_CORR_BITS * sizeof(char));
+    }
+  }
+  if (gather_statistics)
+    entropy->pub.finish_pass = finish_pass_gather_phuff;
+  else
+    entropy->pub.finish_pass = finish_pass_phuff;
+
+  /* Only DC coefficients may be interleaved, so cinfo->comps_in_scan = 1
+   * for AC coefficients.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Initialize DC predictions to 0 */
+    entropy->last_dc_val[ci] = 0;
+    /* Get table index */
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)       /* DC refinement needs no table */
+        continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      entropy->ac_tbl_no = tbl = compptr->ac_tbl_no;
+    }
+    if (gather_statistics) {
+      /* Check for invalid table index */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (tbl < 0 || tbl >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->count_ptrs[tbl] == NULL)
+        entropy->count_ptrs[tbl] = (long *)
+          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                      257 * sizeof(long));
+      memset(entropy->count_ptrs[tbl], 0, 257 * sizeof(long));
+    } else {
+      /* Compute derived values for Huffman table */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, is_DC_band, tbl,
+                              &entropy->derived_tbls[tbl]);
+    }
+  }
+
+  /* Initialize AC stuff */
+  entropy->EOBRUN = 0;
+  entropy->BE = 0;
+
+  /* Initialize bit buffer to empty */
+  entropy->put_buffer = 0;
+  entropy->put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/* Outputting bytes to the file.
+ * NB: these must be called only when actually outputting,
+ * that is, entropy->gather_statistics == FALSE.
+ */
+
+/* Emit a byte */
+#define emit_byte(entropy, val) { \
+  *(entropy)->next_output_byte++ = (JOCTET)(val); \
+  if (--(entropy)->free_in_buffer == 0) \
+    dump_buffer(entropy); \
+}
+
+
+LOCAL(void)
+dump_buffer(phuff_entropy_ptr entropy)
+/* Empty the output buffer; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr *dest = entropy->cinfo->dest;
+
+  if (!(*dest->empty_output_buffer) (entropy->cinfo))
+    ERREXIT(entropy->cinfo, JERR_CANT_SUSPEND);
+  /* After a successful buffer dump, must reset buffer pointers */
+  entropy->next_output_byte = dest->next_output_byte;
+  entropy->free_in_buffer = dest->free_in_buffer;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+LOCAL(void)
+emit_bits(phuff_entropy_ptr entropy, unsigned int code, int size)
+/* Emit some bits, unless we are in gather mode */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  register size_t put_buffer = (size_t)code;
+  register int put_bits = entropy->put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+  if (entropy->gather_statistics)
+    return;                     /* do nothing if we're only getting stats */
+
+  put_buffer &= (((size_t)1) << size) - 1; /* mask off any extra bits in code */
+
+  put_bits += size;             /* new number of bits in buffer */
+
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  put_buffer |= entropy->put_buffer; /* and merge with old buffer contents */
+
+  while (put_bits >= 8) {
+    int c = (int)((put_buffer >> 16) & 0xFF);
+
+    emit_byte(entropy, c);
+    if (c == 0xFF) {            /* need to stuff a zero byte? */
+      emit_byte(entropy, 0);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  entropy->put_buffer = put_buffer; /* update variables */
+  entropy->put_bits = put_bits;
+}
+
+
+LOCAL(void)
+flush_bits(phuff_entropy_ptr entropy)
+{
+  emit_bits(entropy, 0x7F, 7); /* fill any partial byte with ones */
+  entropy->put_buffer = 0;     /* and reset bit-buffer to empty */
+  entropy->put_bits = 0;
+}
+
+
+/*
+ * Emit (or just count) a Huffman symbol.
+ */
+
+LOCAL(void)
+emit_symbol(phuff_entropy_ptr entropy, int tbl_no, int symbol)
+{
+  if (entropy->gather_statistics)
+    entropy->count_ptrs[tbl_no][symbol]++;
+  else {
+    c_derived_tbl *tbl = entropy->derived_tbls[tbl_no];
+    emit_bits(entropy, tbl->ehufco[symbol], tbl->ehufsi[symbol]);
+  }
+}
+
+
+/*
+ * Emit bits from a correction bit buffer.
+ */
+
+LOCAL(void)
+emit_buffered_bits(phuff_entropy_ptr entropy, char *bufstart,
+                   unsigned int nbits)
+{
+  if (entropy->gather_statistics)
+    return;                     /* no real work */
+
+  while (nbits > 0) {
+    emit_bits(entropy, (unsigned int)(*bufstart), 1);
+    bufstart++;
+    nbits--;
+  }
+}
+
+
+/*
+ * Emit any pending EOBRUN symbol.
+ */
+
+LOCAL(void)
+emit_eobrun(phuff_entropy_ptr entropy)
+{
+  register int temp, nbits;
+
+  if (entropy->EOBRUN > 0) {    /* if there is any pending EOBRUN */
+    temp = entropy->EOBRUN;
+    nbits = JPEG_NBITS_NONZERO(temp) - 1;
+    /* safety check: shouldn't happen given limited correction-bit buffer */
+    if (nbits > 14)
+      ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+    emit_symbol(entropy, entropy->ac_tbl_no, nbits << 4);
+    if (nbits)
+      emit_bits(entropy, entropy->EOBRUN, nbits);
+
+    entropy->EOBRUN = 0;
+
+    /* Emit any buffered correction bits */
+    emit_buffered_bits(entropy, entropy->bit_buffer, entropy->BE);
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart(phuff_entropy_ptr entropy, int restart_num)
+{
+  int ci;
+
+  emit_eobrun(entropy);
+
+  if (!entropy->gather_statistics) {
+    flush_bits(entropy);
+    emit_byte(entropy, 0xFF);
+    emit_byte(entropy, JPEG_RST0 + restart_num);
+  }
+
+  if (entropy->cinfo->Ss == 0) {
+    /* Re-initialize DC predictions to 0 */
+    for (ci = 0; ci < entropy->cinfo->comps_in_scan; ci++)
+      entropy->last_dc_val[ci] = 0;
+  } else {
+    /* Re-initialize all AC-related fields to 0 */
+    entropy->EOBRUN = 0;
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  register int temp, temp2, temp3;
+  register int nbits;
+  int blkn, ci;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  jpeg_component_info *compptr;
+  ISHIFT_TEMPS
+  int max_coef_bits = cinfo->data_precision + 2;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    temp2 = IRIGHT_SHIFT((int)((*block)[0]), Al);
+
+    /* DC differences are figured on the point-transformed values. */
+    temp = temp2 - entropy->last_dc_val[ci];
+    entropy->last_dc_val[ci] = temp2;
+
+    /* Encode the DC coefficient difference per section G.1.2.1 */
+
+    /* This is a well-known technique for obtaining the absolute value without
+     * a branch.  It is derived from an assembly language technique presented
+     * in "How to Optimize for the Pentium Processors", Copyright (c) 1996,
+     * 1997 by Agner Fog.
+     */
+    temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
+    temp ^= temp3;
+    temp -= temp3;              /* temp is abs value of input */
+    /* For a negative input, want temp2 = bitwise complement of abs(input) */
+    temp2 = temp ^ temp3;
+
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = JPEG_NBITS(temp);
+    /* Check for out-of-range coefficient values.
+     * Since we're encoding a difference, the range limit is twice as much.
+     */
+    if (nbits > max_coef_bits + 1)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+    /* Count/emit the Huffman-coded symbol for the number of bits */
+    emit_symbol(entropy, compptr->dc_tbl_no, nbits);
+
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    if (nbits)                  /* emit_bits rejects calls with size 0 */
+      emit_bits(entropy, (unsigned int)temp2, nbits);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Data preparation for encode_mcu_AC_first().
+ */
+
+#define COMPUTE_ABSVALUES_AC_FIRST(Sl) { \
+  for (k = 0; k < Sl; k++) { \
+    temp = block[jpeg_natural_order_start[k]]; \
+    if (temp == 0) \
+      continue; \
+    /* We must apply the point transform by Al.  For AC coefficients this \
+     * is an integer division with rounding towards 0.  To do this portably \
+     * in C, we shift after obtaining the absolute value; so the code is \
+     * interwoven with finding the abs value (temp) and output bits (temp2). \
+     */ \
+    temp2 = temp >> (CHAR_BIT * sizeof(int) - 1); \
+    temp ^= temp2; \
+    temp -= temp2;              /* temp is abs value of input */ \
+    temp >>= Al;                /* apply the point transform */ \
+    /* Watch out for case that nonzero coef is zero after point transform */ \
+    if (temp == 0) \
+      continue; \
+    /* For a negative coef, want temp2 = bitwise complement of abs(coef) */ \
+    temp2 ^= temp; \
+    values[k] = (UJCOEF)temp; \
+    values[k + DCTSIZE2] = (UJCOEF)temp2; \
+    zerobits |= ((size_t)1U) << k; \
+  } \
+}
+
+METHODDEF(void)
+encode_mcu_AC_first_prepare(const JCOEF *block,
+                            const int *jpeg_natural_order_start, int Sl,
+                            int Al, UJCOEF *values, size_t *bits)
+{
+  register int k, temp, temp2;
+  size_t zerobits = 0U;
+  int Sl0 = Sl;
+
+#if SIZEOF_SIZE_T == 4
+  if (Sl0 > 32)
+    Sl0 = 32;
+#endif
+
+  COMPUTE_ABSVALUES_AC_FIRST(Sl0);
+
+  bits[0] = zerobits;
+#if SIZEOF_SIZE_T == 4
+  zerobits = 0U;
+
+  if (Sl > 32) {
+    Sl -= 32;
+    jpeg_natural_order_start += 32;
+    values += 32;
+
+    COMPUTE_ABSVALUES_AC_FIRST(Sl);
+  }
+  bits[1] = zerobits;
+#endif
+}
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+#define ENCODE_COEFS_AC_FIRST(label) { \
+  while (zerobits) { \
+    r = count_zeroes(&zerobits); \
+    cvalue += r; \
+label \
+    temp  = cvalue[0]; \
+    temp2 = cvalue[DCTSIZE2]; \
+    \
+    /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
+    while (r > 15) { \
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0); \
+      r -= 16; \
+    } \
+    \
+    /* Find the number of bits needed for the magnitude of the coefficient */ \
+    nbits = JPEG_NBITS_NONZERO(temp);  /* there must be at least one 1 bit */ \
+    /* Check for out-of-range coefficient values */ \
+    if (nbits > max_coef_bits) \
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF); \
+    \
+    /* Count/emit Huffman symbol for run length / number of bits */ \
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + nbits); \
+    \
+    /* Emit that number of bits of the value, if positive, */ \
+    /* or the complement of its magnitude, if negative. */ \
+    emit_bits(entropy, (unsigned int)temp2, nbits); \
+    \
+    cvalue++; \
+    zerobits >>= 1; \
+  } \
+}
+
+METHODDEF(boolean)
+encode_mcu_AC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  register int temp, temp2;
+  register int nbits, r;
+  int Sl = cinfo->Se - cinfo->Ss + 1;
+  int Al = cinfo->Al;
+  UJCOEF values_unaligned[2 * DCTSIZE2 + 15];
+  UJCOEF *values;
+  const UJCOEF *cvalue;
+  size_t zerobits;
+  size_t bits[8 / SIZEOF_SIZE_T];
+  int max_coef_bits = cinfo->data_precision + 2;
+
+#ifdef ZERO_BUFFERS
+  memset(values_unaligned, 0, sizeof(values_unaligned));
+  memset(bits, 0, sizeof(bits));
+#endif
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+#ifdef WITH_SIMD
+  cvalue = values = (UJCOEF *)PAD((JUINTPTR)values_unaligned, 16);
+#else
+  /* Not using SIMD, so alignment is not needed */
+  cvalue = values = values_unaligned;
+#endif
+
+  /* Prepare data */
+  entropy->AC_first_prepare(MCU_data[0][0], jpeg_natural_order + cinfo->Ss,
+                            Sl, Al, values, bits);
+
+  zerobits = bits[0];
+#if SIZEOF_SIZE_T == 4
+  zerobits |= bits[1];
+#endif
+
+  /* Emit any pending EOBRUN */
+  if (zerobits && (entropy->EOBRUN > 0))
+    emit_eobrun(entropy);
+
+#if SIZEOF_SIZE_T == 4
+  zerobits = bits[0];
+#endif
+
+  /* Encode the AC coefficients per section G.1.2.2, fig. G.3 */
+
+  ENCODE_COEFS_AC_FIRST((void)0;);
+
+#if SIZEOF_SIZE_T == 4
+  zerobits = bits[1];
+  if (zerobits) {
+    int diff = ((values + DCTSIZE2 / 2) - cvalue);
+    r = count_zeroes(&zerobits);
+    r += diff;
+    cvalue += r;
+    goto first_iter_ac_first;
+  }
+
+  ENCODE_COEFS_AC_FIRST(first_iter_ac_first:);
+#endif
+
+  if (cvalue < (values + Sl)) { /* If there are trailing zeroes, */
+    entropy->EOBRUN++;          /* count an EOB */
+    if (entropy->EOBRUN == 0x7FFF)
+      emit_eobrun(entropy);     /* force it out to avoid overflow */
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  register int temp;
+  int blkn;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    temp = (*block)[0];
+    emit_bits(entropy, (unsigned int)(temp >> Al), 1);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Data preparation for encode_mcu_AC_refine().
+ */
+
+#define COMPUTE_ABSVALUES_AC_REFINE(Sl, koffset) { \
+  /* It is convenient to make a pre-pass to determine the transformed \
+   * coefficients' absolute values and the EOB position. \
+   */ \
+  for (k = 0; k < Sl; k++) { \
+    temp = block[jpeg_natural_order_start[k]]; \
+    /* We must apply the point transform by Al.  For AC coefficients this \
+     * is an integer division with rounding towards 0.  To do this portably \
+     * in C, we shift after obtaining the absolute value. \
+     */ \
+    temp2 = temp >> (CHAR_BIT * sizeof(int) - 1); \
+    temp ^= temp2; \
+    temp -= temp2;              /* temp is abs value of input */ \
+    temp >>= Al;                /* apply the point transform */ \
+    if (temp != 0) { \
+      zerobits |= ((size_t)1U) << k; \
+      signbits |= ((size_t)(temp2 + 1)) << k; \
+    } \
+    absvalues[k] = (UJCOEF)temp; /* save abs value for main pass */ \
+    if (temp == 1) \
+      EOB = k + koffset;        /* EOB = index of last newly-nonzero coef */ \
+  } \
+}
+
+METHODDEF(int)
+encode_mcu_AC_refine_prepare(const JCOEF *block,
+                             const int *jpeg_natural_order_start, int Sl,
+                             int Al, UJCOEF *absvalues, size_t *bits)
+{
+  register int k, temp, temp2;
+  int EOB = 0;
+  size_t zerobits = 0U, signbits = 0U;
+  int Sl0 = Sl;
+
+#if SIZEOF_SIZE_T == 4
+  if (Sl0 > 32)
+    Sl0 = 32;
+#endif
+
+  COMPUTE_ABSVALUES_AC_REFINE(Sl0, 0);
+
+  bits[0] = zerobits;
+#if SIZEOF_SIZE_T == 8
+  bits[1] = signbits;
+#else
+  bits[2] = signbits;
+
+  zerobits = 0U;
+  signbits = 0U;
+
+  if (Sl > 32) {
+    Sl -= 32;
+    jpeg_natural_order_start += 32;
+    absvalues += 32;
+
+    COMPUTE_ABSVALUES_AC_REFINE(Sl, 32);
+  }
+
+  bits[1] = zerobits;
+  bits[3] = signbits;
+#endif
+
+  return EOB;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+#define ENCODE_COEFS_AC_REFINE(label) { \
+  while (zerobits) { \
+    idx = count_zeroes(&zerobits); \
+    r += idx; \
+    cabsvalue += idx; \
+    signbits >>= idx; \
+label \
+    /* Emit any required ZRLs, but not if they can be folded into EOB */ \
+    while (r > 15 && (cabsvalue <= EOBPTR)) { \
+      /* emit any pending EOBRUN and the BE correction bits */ \
+      emit_eobrun(entropy); \
+      /* Emit ZRL */ \
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0); \
+      r -= 16; \
+      /* Emit buffered correction bits that must be associated with ZRL */ \
+      emit_buffered_bits(entropy, BR_buffer, BR); \
+      BR_buffer = entropy->bit_buffer; /* BE bits are gone now */ \
+      BR = 0; \
+    } \
+    \
+    temp = *cabsvalue++; \
+    \
+    /* If the coef was previously nonzero, it only needs a correction bit. \
+     * NOTE: a straight translation of the spec's figure G.7 would suggest \
+     * that we also need to test r > 15.  But if r > 15, we can only get here \
+     * if k > EOB, which implies that this coefficient is not 1. \
+     */ \
+    if (temp > 1) { \
+      /* The correction bit is the next bit of the absolute value. */ \
+      BR_buffer[BR++] = (char)(temp & 1); \
+      signbits >>= 1; \
+      zerobits >>= 1; \
+      continue; \
+    } \
+    \
+    /* Emit any pending EOBRUN and the BE correction bits */ \
+    emit_eobrun(entropy); \
+    \
+    /* Count/emit Huffman symbol for run length / number of bits */ \
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + 1); \
+    \
+    /* Emit output bit for newly-nonzero coef */ \
+    temp = signbits & 1; /* ((*block)[jpeg_natural_order_start[k]] < 0) ? 0 : 1 */ \
+    emit_bits(entropy, (unsigned int)temp, 1); \
+    \
+    /* Emit buffered correction bits that must be associated with this code */ \
+    emit_buffered_bits(entropy, BR_buffer, BR); \
+    BR_buffer = entropy->bit_buffer; /* BE bits are gone now */ \
+    BR = 0; \
+    r = 0;                      /* reset zero run length */ \
+    signbits >>= 1; \
+    zerobits >>= 1; \
+  } \
+}
+
+METHODDEF(boolean)
+encode_mcu_AC_refine(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  register int temp, r, idx;
+  char *BR_buffer;
+  unsigned int BR;
+  int Sl = cinfo->Se - cinfo->Ss + 1;
+  int Al = cinfo->Al;
+  UJCOEF absvalues_unaligned[DCTSIZE2 + 15];
+  UJCOEF *absvalues;
+  const UJCOEF *cabsvalue, *EOBPTR;
+  size_t zerobits, signbits;
+  size_t bits[16 / SIZEOF_SIZE_T];
+
+#ifdef ZERO_BUFFERS
+  memset(absvalues_unaligned, 0, sizeof(absvalues_unaligned));
+  memset(bits, 0, sizeof(bits));
+#endif
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+#ifdef WITH_SIMD
+  cabsvalue = absvalues = (UJCOEF *)PAD((JUINTPTR)absvalues_unaligned, 16);
+#else
+  /* Not using SIMD, so alignment is not needed */
+  cabsvalue = absvalues = absvalues_unaligned;
+#endif
+
+  /* Prepare data */
+  EOBPTR = absvalues +
+    entropy->AC_refine_prepare(MCU_data[0][0], jpeg_natural_order + cinfo->Ss,
+                               Sl, Al, absvalues, bits);
+
+  /* Encode the AC coefficients per section G.1.2.3, fig. G.7 */
+
+  r = 0;                        /* r = run length of zeros */
+  BR = 0;                       /* BR = count of buffered bits added now */
+  BR_buffer = entropy->bit_buffer + entropy->BE; /* Append bits to buffer */
+
+  zerobits = bits[0];
+#if SIZEOF_SIZE_T == 8
+  signbits = bits[1];
+#else
+  signbits = bits[2];
+#endif
+  ENCODE_COEFS_AC_REFINE((void)0;);
+
+#if SIZEOF_SIZE_T == 4
+  zerobits = bits[1];
+  signbits = bits[3];
+
+  if (zerobits) {
+    int diff = ((absvalues + DCTSIZE2 / 2) - cabsvalue);
+    idx = count_zeroes(&zerobits);
+    signbits >>= idx;
+    idx += diff;
+    r += idx;
+    cabsvalue += idx;
+    goto first_iter_ac_refine;
+  }
+
+  ENCODE_COEFS_AC_REFINE(first_iter_ac_refine:);
+#endif
+
+  r |= (int)((absvalues + Sl) - cabsvalue);
+
+  if (r > 0 || BR > 0) {        /* If there are trailing zeroes, */
+    entropy->EOBRUN++;          /* count an EOB */
+    entropy->BE += BR;          /* concat my correction bits to older ones */
+    /* We force out the EOB if we risk either:
+     * 1. overflow of the EOB counter;
+     * 2. overflow of the correction bit buffer during the next MCU.
+     */
+    if (entropy->EOBRUN == 0x7FFF ||
+        entropy->BE > (MAX_CORR_BITS - DCTSIZE2 + 1))
+      emit_eobrun(entropy);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed progressive scan.
+ */
+
+METHODDEF(void)
+finish_pass_phuff(j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Flush out any buffered data */
+  emit_eobrun(entropy);
+  flush_bits(entropy);
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather_phuff(j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  boolean is_DC_band;
+  int ci, tbl;
+  jpeg_component_info *compptr;
+  JHUFF_TBL **htblptr;
+  boolean did[NUM_HUFF_TBLS];
+
+  /* Flush out buffered data (all we care about is counting the EOB symbol) */
+  emit_eobrun(entropy);
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  memset(did, 0, sizeof(did));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)       /* DC refinement needs no table */
+        continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      tbl = compptr->ac_tbl_no;
+    }
+    if (!did[tbl]) {
+      if (is_DC_band)
+        htblptr = &cinfo->dc_huff_tbl_ptrs[tbl];
+      else
+        htblptr = &cinfo->ac_huff_tbl_ptrs[tbl];
+      if (*htblptr == NULL)
+        *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->count_ptrs[tbl]);
+      did[tbl] = TRUE;
+    }
+  }
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_encoder(j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(phuff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *)entropy;
+  entropy->pub.start_pass = start_pass_phuff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+    entropy->count_ptrs[i] = NULL;
+  }
+  entropy->bit_buffer = NULL;   /* needed only in AC refinement scan */
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */

thirdparty/libjpeg-turbo/src/jcprepct.c

@@ -0,0 +1,378 @@
+/*
+ * jcprepct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the compression preprocessing controller.
+ * This controller manages the color conversion, downsampling,
+ * and edge expansion steps.
+ *
+ * Most of the complexity here is associated with buffering input rows
+ * as required by the downsampler.  See the comments at the head of
+ * jcsample.c for the downsampler's needs.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+
+/* At present, jcsample.c can request context rows only for smoothing.
+ * In the future, we might also need context rows for CCIR601 sampling
+ * or other more-complex downsampling procedures.  The code to support
+ * context rows should be compiled only if needed.
+ */
+#ifdef INPUT_SMOOTHING_SUPPORTED
+#define CONTEXT_ROWS_SUPPORTED
+#endif
+
+
+/*
+ * For the simple (no-context-row) case, we just need to buffer one
+ * row group's worth of pixels for the downsampling step.  At the bottom of
+ * the image, we pad to a full row group by replicating the last pixel row.
+ * The downsampler's last output row is then replicated if needed to pad
+ * out to a full iMCU row.
+ *
+ * When providing context rows, we must buffer three row groups' worth of
+ * pixels.  Three row groups are physically allocated, but the row pointer
+ * arrays are made five row groups high, with the extra pointers above and
+ * below "wrapping around" to point to the last and first real row groups.
+ * This allows the downsampler to access the proper context rows.
+ * At the top and bottom of the image, we create dummy context rows by
+ * copying the first or last real pixel row.  This copying could be avoided
+ * by pointer hacking as is done in jdmainct.c, but it doesn't seem worth the
+ * trouble on the compression side.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_prep_controller pub; /* public fields */
+
+  /* Downsampling input buffer.  This buffer holds color-converted data
+   * until we have enough to do a downsample step.
+   */
+  _JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  JDIMENSION rows_to_go;        /* counts rows remaining in source image */
+  int next_buf_row;             /* index of next row to store in color_buf */
+
+#ifdef CONTEXT_ROWS_SUPPORTED   /* only needed for context case */
+  int this_row_group;           /* starting row index of group to process */
+  int next_buf_stop;            /* downsample when we reach this index */
+#endif
+} my_prep_controller;
+
+typedef my_prep_controller *my_prep_ptr;
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_prep(j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_prep_ptr prep = (my_prep_ptr)cinfo->prep;
+
+  if (pass_mode != JBUF_PASS_THRU)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Initialize total-height counter for detecting bottom of image */
+  prep->rows_to_go = cinfo->image_height;
+  /* Mark the conversion buffer empty */
+  prep->next_buf_row = 0;
+#ifdef CONTEXT_ROWS_SUPPORTED
+  /* Preset additional state variables for context mode.
+   * These aren't used in non-context mode, so we needn't test which mode.
+   */
+  prep->this_row_group = 0;
+  /* Set next_buf_stop to stop after two row groups have been read in. */
+  prep->next_buf_stop = 2 * cinfo->max_v_samp_factor;
+#endif
+}
+
+
+/*
+ * Expand an image vertically from height input_rows to height output_rows,
+ * by duplicating the bottom row.
+ */
+
+LOCAL(void)
+expand_bottom_edge(_JSAMPARRAY image_data, JDIMENSION num_cols, int input_rows,
+                   int output_rows)
+{
+  register int row;
+
+  for (row = input_rows; row < output_rows; row++) {
+    _jcopy_sample_rows(image_data, input_rows - 1, image_data, row, 1,
+                       num_cols);
+  }
+}
+
+
+/*
+ * Process some data in the simple no-context case.
+ *
+ * Preprocessor output data is counted in "row groups".  A row group
+ * is defined to be v_samp_factor sample rows of each component.
+ * Downsampling will produce this much data from each max_v_samp_factor
+ * input rows.
+ */
+
+METHODDEF(void)
+pre_process_data(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                 JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail,
+                 _JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+                 JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr)cinfo->prep;
+  int numrows, ci;
+  JDIMENSION inrows;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  while (*in_row_ctr < in_rows_avail &&
+         *out_row_group_ctr < out_row_groups_avail) {
+    /* Do color conversion to fill the conversion buffer. */
+    inrows = in_rows_avail - *in_row_ctr;
+    numrows = cinfo->max_v_samp_factor - prep->next_buf_row;
+    numrows = (int)MIN((JDIMENSION)numrows, inrows);
+    (*cinfo->cconvert->_color_convert) (cinfo, input_buf + *in_row_ctr,
+                                        prep->color_buf,
+                                        (JDIMENSION)prep->next_buf_row,
+                                        numrows);
+    *in_row_ctr += numrows;
+    prep->next_buf_row += numrows;
+    prep->rows_to_go -= numrows;
+    /* If at bottom of image, pad to fill the conversion buffer. */
+    if (prep->rows_to_go == 0 &&
+        prep->next_buf_row < cinfo->max_v_samp_factor) {
+      for (ci = 0; ci < cinfo->num_components; ci++) {
+        expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+                           prep->next_buf_row, cinfo->max_v_samp_factor);
+      }
+      prep->next_buf_row = cinfo->max_v_samp_factor;
+    }
+    /* If we've filled the conversion buffer, empty it. */
+    if (prep->next_buf_row == cinfo->max_v_samp_factor) {
+      (*cinfo->downsample->_downsample) (cinfo,
+                                         prep->color_buf, (JDIMENSION)0,
+                                         output_buf, *out_row_group_ctr);
+      prep->next_buf_row = 0;
+      (*out_row_group_ctr)++;
+    }
+    /* If at bottom of image, pad the output to a full iMCU height.
+     * Note we assume the caller is providing a one-iMCU-height output buffer!
+     */
+    if (prep->rows_to_go == 0 && *out_row_group_ctr < out_row_groups_avail) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+           ci++, compptr++) {
+        expand_bottom_edge(output_buf[ci],
+                           compptr->width_in_blocks * data_unit,
+                           (int)(*out_row_group_ctr * compptr->v_samp_factor),
+                           (int)(out_row_groups_avail * compptr->v_samp_factor));
+      }
+      *out_row_group_ctr = out_row_groups_avail;
+      break;                    /* can exit outer loop without test */
+    }
+  }
+}
+
+
+#ifdef CONTEXT_ROWS_SUPPORTED
+
+/*
+ * Process some data in the context case.
+ */
+
+METHODDEF(void)
+pre_process_context(j_compress_ptr cinfo, _JSAMPARRAY input_buf,
+                    JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail,
+                    _JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+                    JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr)cinfo->prep;
+  int numrows, ci;
+  int buf_height = cinfo->max_v_samp_factor * 3;
+  JDIMENSION inrows;
+
+  while (*out_row_group_ctr < out_row_groups_avail) {
+    if (*in_row_ctr < in_rows_avail) {
+      /* Do color conversion to fill the conversion buffer. */
+      inrows = in_rows_avail - *in_row_ctr;
+      numrows = prep->next_buf_stop - prep->next_buf_row;
+      numrows = (int)MIN((JDIMENSION)numrows, inrows);
+      (*cinfo->cconvert->_color_convert) (cinfo, input_buf + *in_row_ctr,
+                                          prep->color_buf,
+                                          (JDIMENSION)prep->next_buf_row,
+                                          numrows);
+      /* Pad at top of image, if first time through */
+      if (prep->rows_to_go == cinfo->image_height) {
+        for (ci = 0; ci < cinfo->num_components; ci++) {
+          int row;
+          for (row = 1; row <= cinfo->max_v_samp_factor; row++) {
+            _jcopy_sample_rows(prep->color_buf[ci], 0, prep->color_buf[ci],
+                               -row, 1, cinfo->image_width);
+          }
+        }
+      }
+      *in_row_ctr += numrows;
+      prep->next_buf_row += numrows;
+      prep->rows_to_go -= numrows;
+    } else {
+      /* Return for more data, unless we are at the bottom of the image. */
+      if (prep->rows_to_go != 0)
+        break;
+      /* When at bottom of image, pad to fill the conversion buffer. */
+      if (prep->next_buf_row < prep->next_buf_stop) {
+        for (ci = 0; ci < cinfo->num_components; ci++) {
+          expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+                             prep->next_buf_row, prep->next_buf_stop);
+        }
+        prep->next_buf_row = prep->next_buf_stop;
+      }
+    }
+    /* If we've gotten enough data, downsample a row group. */
+    if (prep->next_buf_row == prep->next_buf_stop) {
+      (*cinfo->downsample->_downsample) (cinfo, prep->color_buf,
+                                         (JDIMENSION)prep->this_row_group,
+                                         output_buf, *out_row_group_ctr);
+      (*out_row_group_ctr)++;
+      /* Advance pointers with wraparound as necessary. */
+      prep->this_row_group += cinfo->max_v_samp_factor;
+      if (prep->this_row_group >= buf_height)
+        prep->this_row_group = 0;
+      if (prep->next_buf_row >= buf_height)
+        prep->next_buf_row = 0;
+      prep->next_buf_stop = prep->next_buf_row + cinfo->max_v_samp_factor;
+    }
+  }
+}
+
+
+/*
+ * Create the wrapped-around downsampling input buffer needed for context mode.
+ */
+
+LOCAL(void)
+create_context_buffer(j_compress_ptr cinfo)
+{
+  my_prep_ptr prep = (my_prep_ptr)cinfo->prep;
+  int rgroup_height = cinfo->max_v_samp_factor;
+  int ci, i;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY true_buffer, fake_buffer;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  /* Grab enough space for fake row pointers for all the components;
+   * we need five row groups' worth of pointers for each component.
+   */
+  fake_buffer = (_JSAMPARRAY)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (cinfo->num_components * 5 * rgroup_height) *
+                                sizeof(_JSAMPROW));
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate the actual buffer space (3 row groups) for this component.
+     * We make the buffer wide enough to allow the downsampler to edge-expand
+     * horizontally within the buffer, if it so chooses.
+     */
+    true_buffer = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+      ((j_common_ptr)cinfo, JPOOL_IMAGE,
+       (JDIMENSION)(((long)compptr->width_in_blocks * data_unit *
+                     cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+       (JDIMENSION)(3 * rgroup_height));
+    /* Copy true buffer row pointers into the middle of the fake row array */
+    memcpy(fake_buffer + rgroup_height, true_buffer,
+           3 * rgroup_height * sizeof(_JSAMPROW));
+    /* Fill in the above and below wraparound pointers */
+    for (i = 0; i < rgroup_height; i++) {
+      fake_buffer[i] = true_buffer[2 * rgroup_height + i];
+      fake_buffer[4 * rgroup_height + i] = true_buffer[i];
+    }
+    prep->color_buf[ci] = fake_buffer + rgroup_height;
+    fake_buffer += 5 * rgroup_height; /* point to space for next component */
+  }
+}
+
+#endif /* CONTEXT_ROWS_SUPPORTED */
+
+
+/*
+ * Initialize preprocessing controller.
+ */
+
+GLOBAL(void)
+_jinit_c_prep_controller(j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_prep_ptr prep;
+  int ci;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  if (need_full_buffer)         /* safety check */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  prep = (my_prep_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_prep_controller));
+  cinfo->prep = (struct jpeg_c_prep_controller *)prep;
+  prep->pub.start_pass = start_pass_prep;
+
+  /* Allocate the color conversion buffer.
+   * We make the buffer wide enough to allow the downsampler to edge-expand
+   * horizontally within the buffer, if it so chooses.
+   */
+  if (cinfo->downsample->need_context_rows) {
+    /* Set up to provide context rows */
+#ifdef CONTEXT_ROWS_SUPPORTED
+    prep->pub._pre_process_data = pre_process_context;
+    create_context_buffer(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* No context, just make it tall enough for one row group */
+    prep->pub._pre_process_data = pre_process_data;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      prep->color_buf[ci] = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+         (JDIMENSION)(((long)compptr->width_in_blocks * data_unit *
+                       cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+         (JDIMENSION)cinfo->max_v_samp_factor);
+    }
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jcsample.c

@@ -0,0 +1,556 @@
+/*
+ * jcsample.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2014, MIPS Technologies, Inc., California.
+ * Copyright (C) 2015, 2019, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains downsampling routines.
+ *
+ * Downsampling input data is counted in "row groups".  A row group
+ * is defined to be max_v_samp_factor pixel rows of each component,
+ * from which the downsampler produces v_samp_factor sample rows.
+ * A single row group is processed in each call to the downsampler module.
+ *
+ * The downsampler is responsible for edge-expansion of its output data
+ * to fill an integral number of DCT blocks horizontally.  The source buffer
+ * may be modified if it is helpful for this purpose (the source buffer is
+ * allocated wide enough to correspond to the desired output width).
+ * The caller (the prep controller) is responsible for vertical padding.
+ *
+ * The downsampler may request "context rows" by setting need_context_rows
+ * during startup.  In this case, the input arrays will contain at least
+ * one row group's worth of pixels above and below the passed-in data;
+ * the caller will create dummy rows at image top and bottom by replicating
+ * the first or last real pixel row.
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ *
+ * The downsampling algorithm used here is a simple average of the source
+ * pixels covered by the output pixel.  The hi-falutin sampling literature
+ * refers to this as a "box filter".  In general the characteristics of a box
+ * filter are not very good, but for the specific cases we normally use (1:1
+ * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
+ * nearly so bad.  If you intend to use other sampling ratios, you'd be well
+ * advised to improve this code.
+ *
+ * A simple input-smoothing capability is provided.  This is mainly intended
+ * for cleaning up color-dithered GIF input files (if you find it inadequate,
+ * we suggest using an external filtering program such as pnmconvol).  When
+ * enabled, each input pixel P is replaced by a weighted sum of itself and its
+ * eight neighbors.  P's weight is 1-8*SF and each neighbor's weight is SF,
+ * where SF = (smoothing_factor / 1024).
+ * Currently, smoothing is only supported for 2h2v sampling factors.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
+
+/* Pointer to routine to downsample a single component */
+typedef void (*downsample1_ptr) (j_compress_ptr cinfo,
+                                 jpeg_component_info *compptr,
+                                 _JSAMPARRAY input_data,
+                                 _JSAMPARRAY output_data);
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_downsampler pub;  /* public fields */
+
+  /* Downsampling method pointers, one per component */
+  downsample1_ptr methods[MAX_COMPONENTS];
+} my_downsampler;
+
+typedef my_downsampler *my_downsample_ptr;
+
+
+/*
+ * Initialize for a downsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_downsample(j_compress_ptr cinfo)
+{
+  /* no work for now */
+}
+
+
+/*
+ * Expand a component horizontally from width input_cols to width output_cols,
+ * by duplicating the rightmost samples.
+ */
+
+LOCAL(void)
+expand_right_edge(_JSAMPARRAY image_data, int num_rows, JDIMENSION input_cols,
+                  JDIMENSION output_cols)
+{
+  register _JSAMPROW ptr;
+  register _JSAMPLE pixval;
+  register int count;
+  int row;
+  int numcols = (int)(output_cols - input_cols);
+
+  if (numcols > 0) {
+    for (row = 0; row < num_rows; row++) {
+      ptr = image_data[row] + input_cols;
+      pixval = ptr[-1];
+      for (count = numcols; count > 0; count--)
+        *ptr++ = pixval;
+    }
+  }
+}
+
+
+/*
+ * Do downsampling for a whole row group (all components).
+ *
+ * In this version we simply downsample each component independently.
+ */
+
+METHODDEF(void)
+sep_downsample(j_compress_ptr cinfo, _JSAMPIMAGE input_buf,
+               JDIMENSION in_row_index, _JSAMPIMAGE output_buf,
+               JDIMENSION out_row_group_index)
+{
+  my_downsample_ptr downsample = (my_downsample_ptr)cinfo->downsample;
+  int ci;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY in_ptr, out_ptr;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    in_ptr = input_buf[ci] + in_row_index;
+    out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor);
+    (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * One row group is processed per call.
+ * This version handles arbitrary integral sampling ratios, without smoothing.
+ * Note that this version is not actually used for customary sampling ratios.
+ */
+
+METHODDEF(void)
+int_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+               _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
+  JDIMENSION outcol, outcol_h;  /* outcol_h == outcol*h_expand */
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+  JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
+  _JSAMPROW inptr, outptr;
+  JLONG outvalue;
+
+  h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor;
+  v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor;
+  numpix = h_expand * v_expand;
+  numpix2 = numpix / 2;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
+                    output_cols * h_expand);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    for (outcol = 0, outcol_h = 0; outcol < output_cols;
+         outcol++, outcol_h += h_expand) {
+      outvalue = 0;
+      for (v = 0; v < v_expand; v++) {
+        inptr = input_data[inrow + v] + outcol_h;
+        for (h = 0; h < h_expand; h++) {
+          outvalue += (JLONG)(*inptr++);
+        }
+      }
+      *outptr++ = (_JSAMPLE)((outvalue + numpix2) / numpix);
+    }
+    inrow += v_expand;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+fullsize_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                    _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  /* Copy the data */
+  _jcopy_sample_rows(input_data, 0, output_data, 0, cinfo->max_v_samp_factor,
+                     cinfo->image_width);
+  /* Edge-expand */
+  expand_right_edge(output_data, cinfo->max_v_samp_factor, cinfo->image_width,
+                    compptr->width_in_blocks * data_unit);
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION outcol;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+  JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
+  register _JSAMPROW inptr, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
+                    output_cols * 2);
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    bias = 0;                   /* bias = 0,1,0,1,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (_JSAMPLE)((inptr[0] + inptr[1] + bias) >> 1);
+      bias ^= 1;                /* 0=>1, 1=>0 */
+      inptr += 2;
+    }
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+h2v2_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION outcol;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+  JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
+  register _JSAMPROW inptr0, inptr1, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
+                    output_cols * 2);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow + 1];
+    bias = 1;                   /* bias = 1,2,1,2,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (_JSAMPLE)
+        ((inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1] + bias) >> 2);
+      bias ^= 3;                /* 1=>2, 2=>1 */
+      inptr0 += 2;  inptr1 += 2;
+    }
+    inrow += 2;
+  }
+}
+
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+h2v2_smooth_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                       _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION colctr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+  JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
+  register _JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
+  JLONG membersum, neighsum, memberscale, neighscale;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+                    cinfo->image_width, output_cols * 2);
+
+  /* We don't bother to form the individual "smoothed" input pixel values;
+   * we can directly compute the output which is the average of the four
+   * smoothed values.  Each of the four member pixels contributes a fraction
+   * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
+   * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
+   * output.  The four corner-adjacent neighbor pixels contribute a fraction
+   * SF to just one smoothed pixel, or SF/4 to the final output; while the
+   * eight edge-adjacent neighbors contribute SF to each of two smoothed
+   * pixels, or SF/2 overall.  In order to use integer arithmetic, these
+   * factors are scaled by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
+  neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow + 1];
+    above_ptr = input_data[inrow - 1];
+    below_ptr = input_data[inrow + 2];
+
+    /* Special case for first column: pretend column -1 is same as column 0 */
+    membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
+    neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+               inptr0[0] + inptr0[2] + inptr1[0] + inptr1[2];
+    neighsum += neighsum;
+    neighsum += above_ptr[0] + above_ptr[2] + below_ptr[0] + below_ptr[2];
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
+    inptr0 += 2;  inptr1 += 2;  above_ptr += 2;  below_ptr += 2;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      /* sum of pixels directly mapped to this output element */
+      membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
+      /* sum of edge-neighbor pixels */
+      neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+                 inptr0[-1] + inptr0[2] + inptr1[-1] + inptr1[2];
+      /* The edge-neighbors count twice as much as corner-neighbors */
+      neighsum += neighsum;
+      /* Add in the corner-neighbors */
+      neighsum += above_ptr[-1] + above_ptr[2] + below_ptr[-1] + below_ptr[2];
+      /* form final output scaled up by 2^16 */
+      membersum = membersum * memberscale + neighsum * neighscale;
+      /* round, descale and output it */
+      *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
+      inptr0 += 2;  inptr1 += 2;  above_ptr += 2;  below_ptr += 2;
+    }
+
+    /* Special case for last column */
+    membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
+    neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+               inptr0[-1] + inptr0[1] + inptr1[-1] + inptr1[1];
+    neighsum += neighsum;
+    neighsum += above_ptr[-1] + above_ptr[1] + below_ptr[-1] + below_ptr[1];
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (_JSAMPLE)((membersum + 32768) >> 16);
+
+    inrow += 2;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+fullsize_smooth_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                           _JSAMPARRAY input_data, _JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION colctr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+  JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
+  register _JSAMPROW inptr, above_ptr, below_ptr, outptr;
+  JLONG membersum, neighsum, memberscale, neighscale;
+  int colsum, lastcolsum, nextcolsum;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+                    cinfo->image_width, output_cols);
+
+  /* Each of the eight neighbor pixels contributes a fraction SF to the
+   * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
+   * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
+  neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    above_ptr = input_data[outrow - 1];
+    below_ptr = input_data[outrow + 1];
+
+    /* Special case for first column */
+    colsum = (*above_ptr++) + (*below_ptr++) + inptr[0];
+    membersum = *inptr++;
+    nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
+    neighsum = colsum + (colsum - membersum) + nextcolsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
+    lastcolsum = colsum;  colsum = nextcolsum;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      membersum = *inptr++;
+      above_ptr++;  below_ptr++;
+      nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
+      neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
+      membersum = membersum * memberscale + neighsum * neighscale;
+      *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
+      lastcolsum = colsum;  colsum = nextcolsum;
+    }
+
+    /* Special case for last column */
+    membersum = *inptr;
+    neighsum = lastcolsum + (colsum - membersum) + colsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (_JSAMPLE)((membersum + 32768) >> 16);
+
+  }
+}
+
+#endif /* INPUT_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Module initialization routine for downsampling.
+ * Note that we must select a routine for each component.
+ */
+
+GLOBAL(void)
+_jinit_downsampler(j_compress_ptr cinfo)
+{
+  my_downsample_ptr downsample;
+  int ci;
+  jpeg_component_info *compptr;
+  boolean smoothok = TRUE;
+
+#ifdef C_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  downsample = (my_downsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_downsampler));
+  cinfo->downsample = (struct jpeg_downsampler *)downsample;
+  downsample->pub.start_pass = start_pass_downsample;
+  downsample->pub._downsample = sep_downsample;
+  downsample->pub.need_context_rows = FALSE;
+
+  if (cinfo->CCIR601_sampling)
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* Verify we can handle the sampling factors, and set up method pointers */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
+        compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+        downsample->methods[ci] = fullsize_smooth_downsample;
+        downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+        downsample->methods[ci] = fullsize_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+               compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+      smoothok = FALSE;
+#ifdef WITH_SIMD
+      if (jsimd_can_h2v1_downsample())
+        downsample->methods[ci] = jsimd_h2v1_downsample;
+      else
+#endif
+        downsample->methods[ci] = h2v1_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+               compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+#if defined(WITH_SIMD) && defined(__mips__)
+        if (jsimd_can_h2v2_smooth_downsample())
+          downsample->methods[ci] = jsimd_h2v2_smooth_downsample;
+        else
+#endif
+          downsample->methods[ci] = h2v2_smooth_downsample;
+        downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+      {
+#ifdef WITH_SIMD
+        if (jsimd_can_h2v2_downsample())
+          downsample->methods[ci] = jsimd_h2v2_downsample;
+        else
+#endif
+          downsample->methods[ci] = h2v2_downsample;
+      }
+    } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
+               (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
+      smoothok = FALSE;
+      downsample->methods[ci] = int_downsample;
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+  }
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+  if (cinfo->smoothing_factor && !smoothok)
+    TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
+#endif
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jctrans.c

@@ -0,0 +1,415 @@
+/*
+ * jctrans.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1995-1998, Thomas G. Lane.
+ * Modified 2000-2009 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains library routines for transcoding compression,
+ * that is, writing raw DCT coefficient arrays to an output JPEG file.
+ * The routines in jcapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+
+
+/* Forward declarations */
+LOCAL(void) transencode_master_selection(j_compress_ptr cinfo,
+                                         jvirt_barray_ptr *coef_arrays);
+LOCAL(void) transencode_coef_controller(j_compress_ptr cinfo,
+                                        jvirt_barray_ptr *coef_arrays);
+
+
+/*
+ * Compression initialization for writing raw-coefficient data.
+ * Before calling this, all parameters and a data destination must be set up.
+ * Call jpeg_finish_compress() to actually write the data.
+ *
+ * The number of passed virtual arrays must match cinfo->num_components.
+ * Note that the virtual arrays need not be filled or even realized at
+ * the time write_coefficients is called; indeed, if the virtual arrays
+ * were requested from this compression object's memory manager, they
+ * typically will be realized during this routine and filled afterwards.
+ */
+
+GLOBAL(void)
+jpeg_write_coefficients(j_compress_ptr cinfo, jvirt_barray_ptr *coef_arrays)
+{
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Mark all tables to be written */
+  jpeg_suppress_tables(cinfo, FALSE);
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr)cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  transencode_master_selection(cinfo, coef_arrays);
+  /* Wait for jpeg_finish_compress() call */
+  cinfo->next_scanline = 0;     /* so jpeg_write_marker works */
+  cinfo->global_state = CSTATE_WRCOEFS;
+}
+
+
+/*
+ * Initialize the compression object with default parameters,
+ * then copy from the source object all parameters needed for lossless
+ * transcoding.  Parameters that can be varied without loss (such as
+ * scan script and Huffman optimization) are left in their default states.
+ */
+
+GLOBAL(void)
+jpeg_copy_critical_parameters(j_decompress_ptr srcinfo, j_compress_ptr dstinfo)
+{
+  JQUANT_TBL **qtblptr;
+  jpeg_component_info *incomp, *outcomp;
+  JQUANT_TBL *c_quant, *slot_quant;
+  int tblno, ci, coefi;
+
+  if (srcinfo->master->lossless)
+    ERREXIT(dstinfo, JERR_NOTIMPL);
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (dstinfo->global_state != CSTATE_START)
+    ERREXIT1(dstinfo, JERR_BAD_STATE, dstinfo->global_state);
+  /* Copy fundamental image dimensions */
+  dstinfo->image_width = srcinfo->image_width;
+  dstinfo->image_height = srcinfo->image_height;
+  dstinfo->input_components = srcinfo->num_components;
+  dstinfo->in_color_space = srcinfo->jpeg_color_space;
+#if JPEG_LIB_VERSION >= 70
+  dstinfo->jpeg_width = srcinfo->output_width;
+  dstinfo->jpeg_height = srcinfo->output_height;
+  dstinfo->min_DCT_h_scaled_size = srcinfo->min_DCT_h_scaled_size;
+  dstinfo->min_DCT_v_scaled_size = srcinfo->min_DCT_v_scaled_size;
+#endif
+  /* Initialize all parameters to default values */
+  jpeg_set_defaults(dstinfo);
+  /* jpeg_set_defaults may choose wrong colorspace, eg YCbCr if input is RGB.
+   * Fix it to get the right header markers for the image colorspace.
+   */
+  jpeg_set_colorspace(dstinfo, srcinfo->jpeg_color_space);
+  dstinfo->data_precision = srcinfo->data_precision;
+  dstinfo->CCIR601_sampling = srcinfo->CCIR601_sampling;
+  /* Copy the source's quantization tables. */
+  for (tblno = 0; tblno < NUM_QUANT_TBLS; tblno++) {
+    if (srcinfo->quant_tbl_ptrs[tblno] != NULL) {
+      qtblptr = &dstinfo->quant_tbl_ptrs[tblno];
+      if (*qtblptr == NULL)
+        *qtblptr = jpeg_alloc_quant_table((j_common_ptr)dstinfo);
+      memcpy((*qtblptr)->quantval, srcinfo->quant_tbl_ptrs[tblno]->quantval,
+             sizeof((*qtblptr)->quantval));
+      (*qtblptr)->sent_table = FALSE;
+    }
+  }
+  /* Copy the source's per-component info.
+   * Note we assume jpeg_set_defaults has allocated the dest comp_info array.
+   */
+  dstinfo->num_components = srcinfo->num_components;
+  if (dstinfo->num_components < 1 || dstinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(dstinfo, JERR_COMPONENT_COUNT, dstinfo->num_components,
+             MAX_COMPONENTS);
+  for (ci = 0, incomp = srcinfo->comp_info, outcomp = dstinfo->comp_info;
+       ci < dstinfo->num_components; ci++, incomp++, outcomp++) {
+    outcomp->component_id = incomp->component_id;
+    outcomp->h_samp_factor = incomp->h_samp_factor;
+    outcomp->v_samp_factor = incomp->v_samp_factor;
+    outcomp->quant_tbl_no = incomp->quant_tbl_no;
+    /* Make sure saved quantization table for component matches the qtable
+     * slot.  If not, the input file re-used this qtable slot.
+     * IJG encoder currently cannot duplicate this.
+     */
+    tblno = outcomp->quant_tbl_no;
+    if (tblno < 0 || tblno >= NUM_QUANT_TBLS ||
+        srcinfo->quant_tbl_ptrs[tblno] == NULL)
+      ERREXIT1(dstinfo, JERR_NO_QUANT_TABLE, tblno);
+    slot_quant = srcinfo->quant_tbl_ptrs[tblno];
+    c_quant = incomp->quant_table;
+    if (c_quant != NULL) {
+      for (coefi = 0; coefi < DCTSIZE2; coefi++) {
+        if (c_quant->quantval[coefi] != slot_quant->quantval[coefi])
+          ERREXIT1(dstinfo, JERR_MISMATCHED_QUANT_TABLE, tblno);
+      }
+    }
+    /* Note: we do not copy the source's Huffman table assignments;
+     * instead we rely on jpeg_set_colorspace to have made a suitable choice.
+     */
+  }
+  /* Also copy JFIF version and resolution information, if available.
+   * Strictly speaking this isn't "critical" info, but it's nearly
+   * always appropriate to copy it if available.  In particular,
+   * if the application chooses to copy JFIF 1.02 extension markers from
+   * the source file, we need to copy the version to make sure we don't
+   * emit a file that has 1.02 extensions but a claimed version of 1.01.
+   * We will *not*, however, copy version info from mislabeled "2.01" files.
+   */
+  if (srcinfo->saw_JFIF_marker) {
+    if (srcinfo->JFIF_major_version == 1) {
+      dstinfo->JFIF_major_version = srcinfo->JFIF_major_version;
+      dstinfo->JFIF_minor_version = srcinfo->JFIF_minor_version;
+    }
+    dstinfo->density_unit = srcinfo->density_unit;
+    dstinfo->X_density = srcinfo->X_density;
+    dstinfo->Y_density = srcinfo->Y_density;
+  }
+}
+
+
+/*
+ * Master selection of compression modules for transcoding.
+ * This substitutes for jcinit.c's initialization of the full compressor.
+ */
+
+LOCAL(void)
+transencode_master_selection(j_compress_ptr cinfo,
+                             jvirt_barray_ptr *coef_arrays)
+{
+  /* Although we don't actually use input_components for transcoding,
+   * jcmaster.c's initial_setup will complain if input_components is 0.
+   */
+  cinfo->input_components = 1;
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, TRUE /* transcode only */);
+
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+#ifdef C_ARITH_CODING_SUPPORTED
+    jinit_arith_encoder(cinfo);
+#else
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+#endif
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      jinit_phuff_encoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_encoder(cinfo);
+  }
+
+  /* We need a special coefficient buffer controller. */
+  transencode_coef_controller(cinfo, coef_arrays);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr)cinfo);
+
+  /* Write the datastream header (SOI, JFIF) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}
+
+
+/*
+ * The rest of this file is a special implementation of the coefficient
+ * buffer controller.  This is similar to jccoefct.c, but it handles only
+ * output from presupplied virtual arrays.  Furthermore, we generate any
+ * dummy padding blocks on-the-fly rather than expecting them to be present
+ * in the arrays.
+ */
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;      /* iMCU row # within image */
+  JDIMENSION mcu_ctr;           /* counts MCUs processed in current row */
+  int MCU_vert_offset;          /* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;    /* number of such rows needed */
+
+  /* Virtual block array for each component. */
+  jvirt_barray_ptr *whole_image;
+
+  /* Workspace for constructing dummy blocks at right/bottom edges. */
+  JBLOCKROW dummy_buffer[C_MAX_BLOCKS_IN_MCU];
+} my_coef_controller;
+
+typedef my_coef_controller *my_coef_ptr;
+
+
+LOCAL(void)
+start_iMCU_row(j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows - 1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef(j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  if (pass_mode != JBUF_CRANK_DEST)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Process some data.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output(j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION MCU_col_num;       /* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, blockcnt;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr)cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION)compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+         MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;                 /* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+        compptr = cinfo->cur_comp_info[ci];
+        start_col = MCU_col_num * compptr->MCU_width;
+        blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width :
+                                                  compptr->last_col_width;
+        for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+          if (coef->iMCU_row_num < last_iMCU_row ||
+              yindex + yoffset < compptr->last_row_height) {
+            /* Fill in pointers to real blocks in this row */
+            buffer_ptr = buffer[ci][yindex + yoffset] + start_col;
+            for (xindex = 0; xindex < blockcnt; xindex++)
+              MCU_buffer[blkn++] = buffer_ptr++;
+          } else {
+            /* At bottom of image, need a whole row of dummy blocks */
+            xindex = 0;
+          }
+          /* Fill in any dummy blocks needed in this row.
+           * Dummy blocks are filled in the same way as in jccoefct.c:
+           * all zeroes in the AC entries, DC entries equal to previous
+           * block's DC value.  The init routine has already zeroed the
+           * AC entries, so we need only set the DC entries correctly.
+           */
+          for (; xindex < compptr->MCU_width; xindex++) {
+            MCU_buffer[blkn] = coef->dummy_buffer[blkn];
+            MCU_buffer[blkn][0][0] = MCU_buffer[blkn - 1][0][0];
+            blkn++;
+          }
+        }
+      }
+      /* Try to write the MCU. */
+      if (!(*cinfo->entropy->encode_mcu) (cinfo, MCU_buffer)) {
+        /* Suspension forced; update state counters and exit */
+        coef->MCU_vert_offset = yoffset;
+        coef->mcu_ctr = MCU_col_num;
+        return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+METHODDEF(boolean)
+compress_output_12(j_compress_ptr cinfo, J12SAMPIMAGE input_buf)
+{
+  return compress_output(cinfo, (JSAMPIMAGE)input_buf);
+}
+
+
+/*
+ * Initialize coefficient buffer controller.
+ *
+ * Each passed coefficient array must be the right size for that
+ * coefficient: width_in_blocks wide and height_in_blocks high,
+ * with unitheight at least v_samp_factor.
+ */
+
+LOCAL(void)
+transencode_coef_controller(j_compress_ptr cinfo,
+                            jvirt_barray_ptr *coef_arrays)
+{
+  my_coef_ptr coef;
+  JBLOCKROW buffer;
+  int i;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *)coef;
+  coef->pub.start_pass = start_pass_coef;
+  coef->pub.compress_data = compress_output;
+  coef->pub.compress_data_12 = compress_output_12;
+
+  /* Save pointer to virtual arrays */
+  coef->whole_image = coef_arrays;
+
+  /* Allocate and pre-zero space for dummy DCT blocks. */
+  buffer = (JBLOCKROW)
+    (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                C_MAX_BLOCKS_IN_MCU * sizeof(JBLOCK));
+  jzero_far((void *)buffer, C_MAX_BLOCKS_IN_MCU * sizeof(JBLOCK));
+  for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+    coef->dummy_buffer[i] = buffer + i;
+  }
+}

thirdparty/libjpeg-turbo/src/jdapimin.c

@@ -0,0 +1,421 @@
+/*
+ * jdapimin.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2016, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-decompression case or the
+ * transcoding-only case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jdapistd.c.  But also see jcomapi.c for routines
+ * shared by compression and decompression, and jdtrans.c for the transcoding
+ * case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdmaster.h"
+
+
+/*
+ * Initialization of a JPEG decompression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateDecompress(j_decompress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;            /* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != sizeof(struct jpeg_decompress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE,
+             (int)sizeof(struct jpeg_decompress_struct), (int)structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr *err = cinfo->err;
+    void *client_data = cinfo->client_data; /* ignore Purify complaint here */
+    memset(cinfo, 0, sizeof(struct jpeg_decompress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = TRUE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr)cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->src = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++)
+    cinfo->quant_tbl_ptrs[i] = NULL;
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  /* Initialize marker processor so application can override methods
+   * for COM, APPn markers before calling jpeg_read_header.
+   */
+  cinfo->marker_list = NULL;
+  jinit_marker_reader(cinfo);
+
+  /* And initialize the overall input controller. */
+  jinit_input_controller(cinfo);
+
+  cinfo->data_precision = BITS_IN_JSAMPLE;
+
+  /* OK, I'm ready */
+  cinfo->global_state = DSTATE_START;
+
+  /* The master struct is used to store extension parameters, so we allocate it
+   * here.
+   */
+  cinfo->master = (struct jpeg_decomp_master *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                sizeof(my_decomp_master));
+  memset(cinfo->master, 0, sizeof(my_decomp_master));
+}
+
+
+/*
+ * Destruction of a JPEG decompression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_decompress(j_decompress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr)cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG decompression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_decompress(j_decompress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr)cinfo); /* use common routine */
+}
+
+
+/*
+ * Set default decompression parameters.
+ */
+
+LOCAL(void)
+default_decompress_parms(j_decompress_ptr cinfo)
+{
+  /* Guess the input colorspace, and set output colorspace accordingly. */
+  /* (Wish JPEG committee had provided a real way to specify this...) */
+  /* Note application may override our guesses. */
+  switch (cinfo->num_components) {
+  case 1:
+    cinfo->jpeg_color_space = JCS_GRAYSCALE;
+    cinfo->out_color_space = JCS_GRAYSCALE;
+    break;
+
+  case 3:
+    if (cinfo->saw_JFIF_marker) {
+      cinfo->jpeg_color_space = JCS_YCbCr; /* JFIF implies YCbCr */
+    } else if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+        cinfo->jpeg_color_space = JCS_RGB;
+        break;
+      case 1:
+        cinfo->jpeg_color_space = JCS_YCbCr;
+        break;
+      default:
+        WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+        cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+        break;
+      }
+    } else {
+      /* Saw no special markers, try to guess from the component IDs */
+      int cid0 = cinfo->comp_info[0].component_id;
+      int cid1 = cinfo->comp_info[1].component_id;
+      int cid2 = cinfo->comp_info[2].component_id;
+
+      if (cid0 == 1 && cid1 == 2 && cid2 == 3) {
+#ifdef D_LOSSLESS_SUPPORTED
+        if (cinfo->master->lossless)
+          cinfo->jpeg_color_space = JCS_RGB; /* assume RGB w/out marker */
+        else
+#endif
+          cinfo->jpeg_color_space = JCS_YCbCr; /* assume JFIF w/out marker */
+      } else if (cid0 == 82 && cid1 == 71 && cid2 == 66)
+        cinfo->jpeg_color_space = JCS_RGB; /* ASCII 'R', 'G', 'B' */
+      else {
+        TRACEMS3(cinfo, 1, JTRC_UNKNOWN_IDS, cid0, cid1, cid2);
+#ifdef D_LOSSLESS_SUPPORTED
+        if (cinfo->master->lossless)
+          cinfo->jpeg_color_space = JCS_RGB; /* assume it's RGB */
+        else
+#endif
+          cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+      }
+    }
+    /* Always guess RGB is proper output colorspace. */
+    cinfo->out_color_space = JCS_RGB;
+    break;
+
+  case 4:
+    if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+        cinfo->jpeg_color_space = JCS_CMYK;
+        break;
+      case 2:
+        cinfo->jpeg_color_space = JCS_YCCK;
+        break;
+      default:
+        WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+        cinfo->jpeg_color_space = JCS_YCCK; /* assume it's YCCK */
+        break;
+      }
+    } else {
+      /* No special markers, assume straight CMYK. */
+      cinfo->jpeg_color_space = JCS_CMYK;
+    }
+    cinfo->out_color_space = JCS_CMYK;
+    break;
+
+  default:
+    cinfo->jpeg_color_space = JCS_UNKNOWN;
+    cinfo->out_color_space = JCS_UNKNOWN;
+    break;
+  }
+
+  /* Set defaults for other decompression parameters. */
+  cinfo->scale_num = 1;         /* 1:1 scaling */
+  cinfo->scale_denom = 1;
+  cinfo->output_gamma = 1.0;
+  cinfo->buffered_image = FALSE;
+  cinfo->raw_data_out = FALSE;
+  cinfo->dct_method = JDCT_DEFAULT;
+  cinfo->do_fancy_upsampling = TRUE;
+  cinfo->do_block_smoothing = TRUE;
+  cinfo->quantize_colors = FALSE;
+  /* We set these in case application only sets quantize_colors. */
+  cinfo->dither_mode = JDITHER_FS;
+#ifdef QUANT_2PASS_SUPPORTED
+  cinfo->two_pass_quantize = TRUE;
+#else
+  cinfo->two_pass_quantize = FALSE;
+#endif
+  cinfo->desired_number_of_colors = 256;
+  cinfo->colormap = NULL;
+  /* Initialize for no mode change in buffered-image mode. */
+  cinfo->enable_1pass_quant = FALSE;
+  cinfo->enable_external_quant = FALSE;
+  cinfo->enable_2pass_quant = FALSE;
+}
+
+
+/*
+ * Decompression startup: read start of JPEG datastream to see what's there.
+ * Need only initialize JPEG object and supply a data source before calling.
+ *
+ * This routine will read as far as the first SOS marker (ie, actual start of
+ * compressed data), and will save all tables and parameters in the JPEG
+ * object.  It will also initialize the decompression parameters to default
+ * values, and finally return JPEG_HEADER_OK.  On return, the application may
+ * adjust the decompression parameters and then call jpeg_start_decompress.
+ * (Or, if the application only wanted to determine the image parameters,
+ * the data need not be decompressed.  In that case, call jpeg_abort or
+ * jpeg_destroy to release any temporary space.)
+ * If an abbreviated (tables only) datastream is presented, the routine will
+ * return JPEG_HEADER_TABLES_ONLY upon reaching EOI.  The application may then
+ * re-use the JPEG object to read the abbreviated image datastream(s).
+ * It is unnecessary (but OK) to call jpeg_abort in this case.
+ * The JPEG_SUSPENDED return code only occurs if the data source module
+ * requests suspension of the decompressor.  In this case the application
+ * should load more source data and then re-call jpeg_read_header to resume
+ * processing.
+ * If a non-suspending data source is used and require_image is TRUE, then the
+ * return code need not be inspected since only JPEG_HEADER_OK is possible.
+ *
+ * This routine is now just a front end to jpeg_consume_input, with some
+ * extra error checking.
+ */
+
+GLOBAL(int)
+jpeg_read_header(j_decompress_ptr cinfo, boolean require_image)
+{
+  int retcode;
+
+  if (cinfo->global_state != DSTATE_START &&
+      cinfo->global_state != DSTATE_INHEADER)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  retcode = jpeg_consume_input(cinfo);
+
+  switch (retcode) {
+  case JPEG_REACHED_SOS:
+    retcode = JPEG_HEADER_OK;
+    break;
+  case JPEG_REACHED_EOI:
+    if (require_image)          /* Complain if application wanted an image */
+      ERREXIT(cinfo, JERR_NO_IMAGE);
+    /* Reset to start state; it would be safer to require the application to
+     * call jpeg_abort, but we can't change it now for compatibility reasons.
+     * A side effect is to free any temporary memory (there shouldn't be any).
+     */
+    jpeg_abort((j_common_ptr)cinfo); /* sets state = DSTATE_START */
+    retcode = JPEG_HEADER_TABLES_ONLY;
+    break;
+  case JPEG_SUSPENDED:
+    /* no work */
+    break;
+  }
+
+  return retcode;
+}
+
+
+/*
+ * Consume data in advance of what the decompressor requires.
+ * This can be called at any time once the decompressor object has
+ * been created and a data source has been set up.
+ *
+ * This routine is essentially a state machine that handles a couple
+ * of critical state-transition actions, namely initial setup and
+ * transition from header scanning to ready-for-start_decompress.
+ * All the actual input is done via the input controller's consume_input
+ * method.
+ */
+
+GLOBAL(int)
+jpeg_consume_input(j_decompress_ptr cinfo)
+{
+  int retcode = JPEG_SUSPENDED;
+
+  /* NB: every possible DSTATE value should be listed in this switch */
+  switch (cinfo->global_state) {
+  case DSTATE_START:
+    /* Start-of-datastream actions: reset appropriate modules */
+    (*cinfo->inputctl->reset_input_controller) (cinfo);
+    /* Initialize application's data source module */
+    (*cinfo->src->init_source) (cinfo);
+    cinfo->global_state = DSTATE_INHEADER;
+    FALLTHROUGH                 /*FALLTHROUGH*/
+  case DSTATE_INHEADER:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    if (retcode == JPEG_REACHED_SOS) { /* Found SOS, prepare to decompress */
+      /* Set up default parameters based on header data */
+      default_decompress_parms(cinfo);
+      /* Set global state: ready for start_decompress */
+      cinfo->global_state = DSTATE_READY;
+    }
+    break;
+  case DSTATE_READY:
+    /* Can't advance past first SOS until start_decompress is called */
+    retcode = JPEG_REACHED_SOS;
+    break;
+  case DSTATE_PRELOAD:
+  case DSTATE_PRESCAN:
+  case DSTATE_SCANNING:
+  case DSTATE_RAW_OK:
+  case DSTATE_BUFIMAGE:
+  case DSTATE_BUFPOST:
+  case DSTATE_STOPPING:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    break;
+  default:
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  return retcode;
+}
+
+
+/*
+ * Have we finished reading the input file?
+ */
+
+GLOBAL(boolean)
+jpeg_input_complete(j_decompress_ptr cinfo)
+{
+  /* Check for valid jpeg object */
+  if (cinfo->global_state < DSTATE_START ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->eoi_reached;
+}
+
+
+/*
+ * Is there more than one scan?
+ */
+
+GLOBAL(boolean)
+jpeg_has_multiple_scans(j_decompress_ptr cinfo)
+{
+  /* Only valid after jpeg_read_header completes */
+  if (cinfo->global_state < DSTATE_READY ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->has_multiple_scans;
+}
+
+
+/*
+ * Finish JPEG decompression.
+ *
+ * This will normally just verify the file trailer and release temp storage.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_decompress(j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && !cinfo->buffered_image) {
+    /* Terminate final pass of non-buffered mode */
+    if (cinfo->output_scanline < cinfo->output_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state == DSTATE_BUFIMAGE) {
+    /* Finishing after a buffered-image operation */
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state != DSTATE_STOPPING) {
+    /* STOPPING = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read until EOI */
+  while (!cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;             /* Suspend, come back later */
+  }
+  /* Do final cleanup */
+  (*cinfo->src->term_source) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr)cinfo);
+  return TRUE;
+}

thirdparty/libjpeg-turbo/src/jdapistd.c

@@ -0,0 +1,764 @@
+/*
+ * jdapistd.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2015-2020, 2022-2024, D. R. Commander.
+ * Copyright (C) 2015, Google, Inc.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-decompression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_decompress, it will end up linking in the entire decompressor.
+ * We thus must separate this file from jdapimin.c to avoid linking the
+ * whole decompression library into a transcoder.
+ */
+
+#include "jinclude.h"
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+#include "jdmainct.h"
+#include "jdcoefct.h"
+#else
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+#endif
+#include "jdmaster.h"
+#include "jdmerge.h"
+#include "jdsample.h"
+#include "jmemsys.h"
+
+#if BITS_IN_JSAMPLE == 8
+
+/* Forward declarations */
+LOCAL(boolean) output_pass_setup(j_decompress_ptr cinfo);
+
+
+/*
+ * Decompression initialization.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * If a multipass operating mode was selected, this will do all but the
+ * last pass, and thus may take a great deal of time.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_start_decompress(j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize master control, select active modules */
+    jinit_master_decompress(cinfo);
+    if (cinfo->buffered_image) {
+      /* No more work here; expecting jpeg_start_output next */
+      cinfo->global_state = DSTATE_BUFIMAGE;
+      return TRUE;
+    }
+    cinfo->global_state = DSTATE_PRELOAD;
+  }
+  if (cinfo->global_state == DSTATE_PRELOAD) {
+    /* If file has multiple scans, absorb them all into the coef buffer */
+    if (cinfo->inputctl->has_multiple_scans) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+      for (;;) {
+        int retcode;
+        /* Call progress monitor hook if present */
+        if (cinfo->progress != NULL)
+          (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+        /* Absorb some more input */
+        retcode = (*cinfo->inputctl->consume_input) (cinfo);
+        if (retcode == JPEG_SUSPENDED)
+          return FALSE;
+        if (retcode == JPEG_REACHED_EOI)
+          break;
+        /* Advance progress counter if appropriate */
+        if (cinfo->progress != NULL &&
+            (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+          if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+            /* jdmaster underestimated number of scans; ratchet up one scan */
+            cinfo->progress->pass_limit += (long)cinfo->total_iMCU_rows;
+          }
+        }
+      }
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+    }
+    cinfo->output_scan_number = cinfo->input_scan_number;
+  } else if (cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any dummy output passes, and set up for the final pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Set up for an output pass, and perform any dummy pass(es) needed.
+ * Common subroutine for jpeg_start_decompress and jpeg_start_output.
+ * Entry: global_state = DSTATE_PRESCAN only if previously suspended.
+ * Exit: If done, returns TRUE and sets global_state for proper output mode.
+ *       If suspended, returns FALSE and sets global_state = DSTATE_PRESCAN.
+ */
+
+LOCAL(boolean)
+output_pass_setup(j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state != DSTATE_PRESCAN) {
+    /* First call: do pass setup */
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+    cinfo->global_state = DSTATE_PRESCAN;
+  }
+  /* Loop over any required dummy passes */
+  while (cinfo->master->is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Crank through the dummy pass */
+    while (cinfo->output_scanline < cinfo->output_height) {
+      JDIMENSION last_scanline;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL) {
+        cinfo->progress->pass_counter = (long)cinfo->output_scanline;
+        cinfo->progress->pass_limit = (long)cinfo->output_height;
+        (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+      }
+      /* Process some data */
+      last_scanline = cinfo->output_scanline;
+      if (cinfo->data_precision <= 8)
+        (*cinfo->main->process_data) (cinfo, (JSAMPARRAY)NULL,
+                                      &cinfo->output_scanline, (JDIMENSION)0);
+      else if (cinfo->data_precision <= 12)
+        (*cinfo->main->process_data_12) (cinfo, (J12SAMPARRAY)NULL,
+                                         &cinfo->output_scanline,
+                                         (JDIMENSION)0);
+#ifdef D_LOSSLESS_SUPPORTED
+      else
+        (*cinfo->main->process_data_16) (cinfo, (J16SAMPARRAY)NULL,
+                                         &cinfo->output_scanline,
+                                         (JDIMENSION)0);
+#endif
+      if (cinfo->output_scanline == last_scanline)
+        return FALSE;           /* No progress made, must suspend */
+    }
+    /* Finish up dummy pass, and set up for another one */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  }
+  /* Ready for application to drive output pass through
+   * _jpeg_read_scanlines or _jpeg_read_raw_data.
+   */
+  cinfo->global_state = cinfo->raw_data_out ? DSTATE_RAW_OK : DSTATE_SCANNING;
+  return TRUE;
+}
+
+#endif /* BITS_IN_JSAMPLE == 8 */
+
+
+#if BITS_IN_JSAMPLE != 16
+
+/*
+ * Enable partial scanline decompression
+ *
+ * Must be called after jpeg_start_decompress() and before any calls to
+ * _jpeg_read_scanlines() or _jpeg_skip_scanlines().
+ *
+ * Refer to libjpeg.txt for more information.
+ */
+
+GLOBAL(void)
+_jpeg_crop_scanline(j_decompress_ptr cinfo, JDIMENSION *xoffset,
+                    JDIMENSION *width)
+{
+  int ci, align, orig_downsampled_width;
+  JDIMENSION input_xoffset;
+  boolean reinit_upsampler = FALSE;
+  jpeg_component_info *compptr;
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+#endif
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if ((cinfo->global_state != DSTATE_SCANNING &&
+       cinfo->global_state != DSTATE_BUFIMAGE) || cinfo->output_scanline != 0)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (!xoffset || !width)
+    ERREXIT(cinfo, JERR_BAD_CROP_SPEC);
+
+  /* xoffset and width must fall within the output image dimensions. */
+  if (*width == 0 ||
+      (unsigned long long)(*xoffset) + *width > cinfo->output_width)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* No need to do anything if the caller wants the entire width. */
+  if (*width == cinfo->output_width)
+    return;
+
+  /* Ensuring the proper alignment of xoffset is tricky.  At minimum, it
+   * must align with an MCU boundary, because:
+   *
+   *   (1) The IDCT is performed in blocks, and it is not feasible to modify
+   *       the algorithm so that it can transform partial blocks.
+   *   (2) Because of the SIMD extensions, any input buffer passed to the
+   *       upsampling and color conversion routines must be aligned to the
+   *       SIMD word size (for instance, 128-bit in the case of SSE2.)  The
+   *       easiest way to accomplish this without copying data is to ensure
+   *       that upsampling and color conversion begin at the start of the
+   *       first MCU column that will be inverse transformed.
+   *
+   * In practice, we actually impose a stricter alignment requirement.  We
+   * require that xoffset be a multiple of the maximum MCU column width of all
+   * of the components (the "iMCU column width.")  This is to simplify the
+   * single-pass decompression case, allowing us to use the same MCU column
+   * width for all of the components.
+   */
+  if (cinfo->comps_in_scan == 1 && cinfo->num_components == 1)
+    align = cinfo->_min_DCT_scaled_size;
+  else
+    align = cinfo->_min_DCT_scaled_size * cinfo->max_h_samp_factor;
+
+  /* Adjust xoffset to the nearest iMCU boundary <= the requested value */
+  input_xoffset = *xoffset;
+  *xoffset = (input_xoffset / align) * align;
+
+  /* Adjust the width so that the right edge of the output image is as
+   * requested (only the left edge is altered.)  It is important that calling
+   * programs check this value after this function returns, so that they can
+   * allocate an output buffer with the appropriate size.
+   */
+  *width = *width + input_xoffset - *xoffset;
+  cinfo->output_width = *width;
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+    my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+    upsample->out_row_width =
+      cinfo->output_width * cinfo->out_color_components;
+  }
+#endif
+
+  /* Set the first and last iMCU columns that we must decompress.  These values
+   * will be used in single-scan decompressions.
+   */
+  cinfo->master->first_iMCU_col = (JDIMENSION)(long)(*xoffset) / (long)align;
+  cinfo->master->last_iMCU_col =
+    (JDIMENSION)jdiv_round_up((long)(*xoffset + cinfo->output_width),
+                              (long)align) - 1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    int hsf = (cinfo->comps_in_scan == 1 && cinfo->num_components == 1) ?
+              1 : compptr->h_samp_factor;
+
+    /* Set downsampled_width to the new output width. */
+    orig_downsampled_width = compptr->downsampled_width;
+    compptr->downsampled_width =
+      (JDIMENSION)jdiv_round_up((long)cinfo->output_width *
+                                (long)(compptr->h_samp_factor *
+                                       compptr->_DCT_scaled_size),
+                                (long)(cinfo->max_h_samp_factor *
+                                       cinfo->_min_DCT_scaled_size));
+    if (compptr->downsampled_width < 2 && orig_downsampled_width >= 2)
+      reinit_upsampler = TRUE;
+
+    /* Set the first and last iMCU columns that we must decompress.  These
+     * values will be used in multi-scan decompressions.
+     */
+    cinfo->master->first_MCU_col[ci] =
+      (JDIMENSION)(long)(*xoffset * hsf) / (long)align;
+    cinfo->master->last_MCU_col[ci] =
+      (JDIMENSION)jdiv_round_up((long)((*xoffset + cinfo->output_width) * hsf),
+                                (long)align) - 1;
+  }
+
+  if (reinit_upsampler) {
+    cinfo->master->jinit_upsampler_no_alloc = TRUE;
+    _jinit_upsampler(cinfo);
+    cinfo->master->jinit_upsampler_no_alloc = FALSE;
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 */
+
+
+/*
+ * Read some scanlines of data from the JPEG decompressor.
+ *
+ * The return value will be the number of lines actually read.
+ * This may be less than the number requested in several cases,
+ * including bottom of image, data source suspension, and operating
+ * modes that emit multiple scanlines at a time.
+ *
+ * Note: we warn about excess calls to _jpeg_read_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * an oversize buffer (max_lines > scanlines remaining) is not an error.
+ */
+
+GLOBAL(JDIMENSION)
+_jpeg_read_scanlines(j_decompress_ptr cinfo, _JSAMPARRAY scanlines,
+                     JDIMENSION max_lines)
+{
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+  JDIMENSION row_ctr;
+
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  if (cinfo->global_state != DSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long)cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long)cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+  }
+
+  /* Process some data */
+  row_ctr = 0;
+  (*cinfo->main->_process_data) (cinfo, scanlines, &row_ctr, max_lines);
+  cinfo->output_scanline += row_ctr;
+  return row_ctr;
+#else
+  ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  return 0;
+#endif
+}
+
+
+#if BITS_IN_JSAMPLE != 16
+
+/* Dummy color convert function used by _jpeg_skip_scanlines() */
+LOCAL(void)
+noop_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+             JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+}
+
+
+/* Dummy quantize function used by _jpeg_skip_scanlines() */
+LOCAL(void)
+noop_quantize(j_decompress_ptr cinfo, _JSAMPARRAY input_buf,
+              _JSAMPARRAY output_buf, int num_rows)
+{
+}
+
+
+/*
+ * In some cases, it is best to call _jpeg_read_scanlines() and discard the
+ * output, rather than skipping the scanlines, because this allows us to
+ * maintain the internal state of the context-based upsampler.  In these cases,
+ * we set up and tear down a dummy color converter in order to avoid valgrind
+ * errors and to achieve the best possible performance.
+ */
+
+LOCAL(void)
+read_and_discard_scanlines(j_decompress_ptr cinfo, JDIMENSION num_lines)
+{
+  JDIMENSION n;
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+#endif
+  _JSAMPLE dummy_sample[1] = { 0 };
+  _JSAMPROW dummy_row = dummy_sample;
+  _JSAMPARRAY scanlines = NULL;
+  void (*color_convert) (j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                         JDIMENSION input_row, _JSAMPARRAY output_buf,
+                         int num_rows) = NULL;
+  void (*color_quantize) (j_decompress_ptr cinfo, _JSAMPARRAY input_buf,
+                          _JSAMPARRAY output_buf, int num_rows) = NULL;
+
+  if (cinfo->cconvert && cinfo->cconvert->_color_convert) {
+    color_convert = cinfo->cconvert->_color_convert;
+    cinfo->cconvert->_color_convert = noop_convert;
+    /* This just prevents UBSan from complaining about adding 0 to a NULL
+     * pointer.  The pointer isn't actually used.
+     */
+    scanlines = &dummy_row;
+  }
+
+  if (cinfo->cquantize && cinfo->cquantize->_color_quantize) {
+    color_quantize = cinfo->cquantize->_color_quantize;
+    cinfo->cquantize->_color_quantize = noop_quantize;
+  }
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+    my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+    scanlines = &upsample->spare_row;
+  }
+#endif
+
+  for (n = 0; n < num_lines; n++)
+    _jpeg_read_scanlines(cinfo, scanlines, 1);
+
+  if (color_convert)
+    cinfo->cconvert->_color_convert = color_convert;
+
+  if (color_quantize)
+    cinfo->cquantize->_color_quantize = color_quantize;
+}
+
+
+/*
+ * Called by _jpeg_skip_scanlines().  This partially skips a decompress block
+ * by incrementing the rowgroup counter.
+ */
+
+LOCAL(void)
+increment_simple_rowgroup_ctr(j_decompress_ptr cinfo, JDIMENSION rows)
+{
+  JDIMENSION rows_left;
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+    read_and_discard_scanlines(cinfo, rows);
+    return;
+  }
+
+  /* Increment the counter to the next row group after the skipped rows. */
+  main_ptr->rowgroup_ctr += rows / cinfo->max_v_samp_factor;
+
+  /* Partially skipping a row group would involve modifying the internal state
+   * of the upsampler, so read the remaining rows into a dummy buffer instead.
+   */
+  rows_left = rows % cinfo->max_v_samp_factor;
+  cinfo->output_scanline += rows - rows_left;
+
+  read_and_discard_scanlines(cinfo, rows_left);
+}
+
+/*
+ * Skips some scanlines of data from the JPEG decompressor.
+ *
+ * The return value will be the number of lines actually skipped.  If skipping
+ * num_lines would move beyond the end of the image, then the actual number of
+ * lines remaining in the image is returned.  Otherwise, the return value will
+ * be equal to num_lines.
+ *
+ * Refer to libjpeg.txt for more information.
+ */
+
+GLOBAL(JDIMENSION)
+_jpeg_skip_scanlines(j_decompress_ptr cinfo, JDIMENSION num_lines)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+  my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+  JDIMENSION i, x;
+  int y;
+  JDIMENSION lines_per_iMCU_row, lines_left_in_iMCU_row, lines_after_iMCU_row;
+  JDIMENSION lines_to_skip, lines_to_read;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  /* Two-pass color quantization is not supported. */
+  if (cinfo->quantize_colors && cinfo->two_pass_quantize)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if (cinfo->global_state != DSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Do not skip past the bottom of the image. */
+  if ((unsigned long long)cinfo->output_scanline + num_lines >=
+      cinfo->output_height) {
+    num_lines = cinfo->output_height - cinfo->output_scanline;
+    cinfo->output_scanline = cinfo->output_height;
+    (*cinfo->inputctl->finish_input_pass) (cinfo);
+    cinfo->inputctl->eoi_reached = TRUE;
+    return num_lines;
+  }
+
+  if (num_lines == 0)
+    return 0;
+
+  lines_per_iMCU_row = cinfo->_min_DCT_scaled_size * cinfo->max_v_samp_factor;
+  lines_left_in_iMCU_row =
+    (lines_per_iMCU_row - (cinfo->output_scanline % lines_per_iMCU_row)) %
+    lines_per_iMCU_row;
+  lines_after_iMCU_row = num_lines - lines_left_in_iMCU_row;
+
+  /* Skip the lines remaining in the current iMCU row.  When upsampling
+   * requires context rows, we need the previous and next rows in order to read
+   * the current row.  This adds some complexity.
+   */
+  if (cinfo->upsample->need_context_rows) {
+    /* If the skipped lines would not move us past the current iMCU row, we
+     * read the lines and ignore them.  There might be a faster way of doing
+     * this, but we are facing increasing complexity for diminishing returns.
+     * The increasing complexity would be a by-product of meddling with the
+     * state machine used to skip context rows.  Near the end of an iMCU row,
+     * the next iMCU row may have already been entropy-decoded.  In this unique
+     * case, we will read the next iMCU row if we cannot skip past it as well.
+     */
+    if ((num_lines < lines_left_in_iMCU_row + 1) ||
+        (lines_left_in_iMCU_row <= 1 && main_ptr->buffer_full &&
+         lines_after_iMCU_row < lines_per_iMCU_row + 1)) {
+      read_and_discard_scanlines(cinfo, num_lines);
+      return num_lines;
+    }
+
+    /* If the next iMCU row has already been entropy-decoded, make sure that
+     * we do not skip too far.
+     */
+    if (lines_left_in_iMCU_row <= 1 && main_ptr->buffer_full) {
+      cinfo->output_scanline += lines_left_in_iMCU_row + lines_per_iMCU_row;
+      lines_after_iMCU_row -= lines_per_iMCU_row;
+    } else {
+      cinfo->output_scanline += lines_left_in_iMCU_row;
+    }
+
+    /* If we have just completed the first block, adjust the buffer pointers */
+    if (main_ptr->iMCU_row_ctr == 0 ||
+        (main_ptr->iMCU_row_ctr == 1 && lines_left_in_iMCU_row > 2))
+      set_wraparound_pointers(cinfo);
+    main_ptr->buffer_full = FALSE;
+    main_ptr->rowgroup_ctr = 0;
+    main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
+    if (!master->using_merged_upsample) {
+      upsample->next_row_out = cinfo->max_v_samp_factor;
+      upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+    }
+  }
+
+  /* Skipping is much simpler when context rows are not required. */
+  else {
+    if (num_lines < lines_left_in_iMCU_row) {
+      increment_simple_rowgroup_ctr(cinfo, num_lines);
+      return num_lines;
+    } else {
+      cinfo->output_scanline += lines_left_in_iMCU_row;
+      main_ptr->buffer_full = FALSE;
+      main_ptr->rowgroup_ctr = 0;
+      if (!master->using_merged_upsample) {
+        upsample->next_row_out = cinfo->max_v_samp_factor;
+        upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+      }
+    }
+  }
+
+  /* Calculate how many full iMCU rows we can skip. */
+  if (cinfo->upsample->need_context_rows)
+    lines_to_skip = ((lines_after_iMCU_row - 1) / lines_per_iMCU_row) *
+                    lines_per_iMCU_row;
+  else
+    lines_to_skip = (lines_after_iMCU_row / lines_per_iMCU_row) *
+                    lines_per_iMCU_row;
+  /* Calculate the number of lines that remain to be skipped after skipping all
+   * of the full iMCU rows that we can.  We will not read these lines unless we
+   * have to.
+   */
+  lines_to_read = lines_after_iMCU_row - lines_to_skip;
+
+  /* For images requiring multiple scans (progressive, non-interleaved, etc.),
+   * all of the entropy decoding occurs in jpeg_start_decompress(), assuming
+   * that the input data source is non-suspending.  This makes skipping easy.
+   */
+  if (cinfo->inputctl->has_multiple_scans || cinfo->buffered_image) {
+    if (cinfo->upsample->need_context_rows) {
+      cinfo->output_scanline += lines_to_skip;
+      cinfo->output_iMCU_row += lines_to_skip / lines_per_iMCU_row;
+      main_ptr->iMCU_row_ctr += lines_to_skip / lines_per_iMCU_row;
+      /* It is complex to properly move to the middle of a context block, so
+       * read the remaining lines instead of skipping them.
+       */
+      read_and_discard_scanlines(cinfo, lines_to_read);
+    } else {
+      cinfo->output_scanline += lines_to_skip;
+      cinfo->output_iMCU_row += lines_to_skip / lines_per_iMCU_row;
+      increment_simple_rowgroup_ctr(cinfo, lines_to_read);
+    }
+    if (!master->using_merged_upsample)
+      upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+    return num_lines;
+  }
+
+  /* Skip the iMCU rows that we can safely skip. */
+  for (i = 0; i < lines_to_skip; i += lines_per_iMCU_row) {
+    for (y = 0; y < coef->MCU_rows_per_iMCU_row; y++) {
+      for (x = 0; x < cinfo->MCUs_per_row; x++) {
+        /* Calling decode_mcu() with a NULL pointer causes it to discard the
+         * decoded coefficients.  This is ~5% faster for large subsets, but
+         * it's tough to tell a difference for smaller images.
+         */
+        if (!cinfo->entropy->insufficient_data)
+          cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
+        (*cinfo->entropy->decode_mcu) (cinfo, NULL);
+      }
+    }
+    cinfo->input_iMCU_row++;
+    cinfo->output_iMCU_row++;
+    if (cinfo->input_iMCU_row < cinfo->total_iMCU_rows)
+      start_iMCU_row(cinfo);
+    else
+      (*cinfo->inputctl->finish_input_pass) (cinfo);
+  }
+  cinfo->output_scanline += lines_to_skip;
+
+  if (cinfo->upsample->need_context_rows) {
+    /* Context-based upsampling keeps track of iMCU rows. */
+    main_ptr->iMCU_row_ctr += lines_to_skip / lines_per_iMCU_row;
+
+    /* It is complex to properly move to the middle of a context block, so
+     * read the remaining lines instead of skipping them.
+     */
+    read_and_discard_scanlines(cinfo, lines_to_read);
+  } else {
+    increment_simple_rowgroup_ctr(cinfo, lines_to_read);
+  }
+
+  /* Since skipping lines involves skipping the upsampling step, the value of
+   * "rows_to_go" will become invalid unless we set it here.  NOTE: This is a
+   * bit odd, since "rows_to_go" seems to be redundantly keeping track of
+   * output_scanline.
+   */
+  if (!master->using_merged_upsample)
+    upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+
+  /* Always skip the requested number of lines. */
+  return num_lines;
+}
+
+/*
+ * Alternate entry point to read raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+_jpeg_read_raw_data(j_decompress_ptr cinfo, _JSAMPIMAGE data,
+                    JDIMENSION max_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if (cinfo->global_state != DSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long)cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long)cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+  }
+
+  /* Verify that at least one iMCU row can be returned. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * cinfo->_min_DCT_scaled_size;
+  if (max_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Decompress directly into user's buffer. */
+  if (!(*cinfo->coef->_decompress_data) (cinfo, data))
+    return 0;                   /* suspension forced, can do nothing more */
+
+  /* OK, we processed one iMCU row. */
+  cinfo->output_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 */
+
+
+#if BITS_IN_JSAMPLE == 8
+
+/* Additional entry points for buffered-image mode. */
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Initialize for an output pass in buffered-image mode.
+ */
+
+GLOBAL(boolean)
+jpeg_start_output(j_decompress_ptr cinfo, int scan_number)
+{
+  if (cinfo->global_state != DSTATE_BUFIMAGE &&
+      cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Limit scan number to valid range */
+  if (scan_number <= 0)
+    scan_number = 1;
+  if (cinfo->inputctl->eoi_reached && scan_number > cinfo->input_scan_number)
+    scan_number = cinfo->input_scan_number;
+  cinfo->output_scan_number = scan_number;
+  /* Perform any dummy output passes, and set up for the real pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Finish up after an output pass in buffered-image mode.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_output(j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && cinfo->buffered_image) {
+    /* Terminate this pass. */
+    /* We do not require the whole pass to have been completed. */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_BUFPOST;
+  } else if (cinfo->global_state != DSTATE_BUFPOST) {
+    /* BUFPOST = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read markers looking for SOS or EOI */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+         !cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;             /* Suspend, come back later */
+  }
+  cinfo->global_state = DSTATE_BUFIMAGE;
+  return TRUE;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+#endif /* BITS_IN_JSAMPLE == 8 */

thirdparty/libjpeg-turbo/src/jdarith.c

@@ -0,0 +1,782 @@
+/*
+ * jdarith.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Developed 1997-2015 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015-2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains portable arithmetic entropy encoding routines for JPEG
+ * (implementing Recommendation ITU-T T.81 | ISO/IEC 10918-1).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+#define NEG_1  ((unsigned int)-1)
+
+
+/* Expanded entropy decoder object for arithmetic decoding. */
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  JLONG c;       /* C register, base of coding interval + input bit buffer */
+  JLONG a;               /* A register, normalized size of coding interval */
+  int ct;     /* bit shift counter, # of bits left in bit buffer part of C */
+                                                         /* init: ct = -16 */
+                                                         /* run: ct = 0..7 */
+                                                         /* error: ct = -1 */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char *dc_stats[NUM_ARITH_TBLS];
+  unsigned char *ac_stats[NUM_ARITH_TBLS];
+
+  /* Statistics bin for coding with fixed probability 0.5 */
+  unsigned char fixed_bin[4];
+} arith_entropy_decoder;
+
+typedef arith_entropy_decoder *arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS  64
+#define AC_STAT_BINS  256
+
+
+LOCAL(int)
+get_byte(j_decompress_ptr cinfo)
+/* Read next input byte; we do not support suspension in this module. */
+{
+  struct jpeg_source_mgr *src = cinfo->src;
+
+  if (src->bytes_in_buffer == 0)
+    if (!(*src->fill_input_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  src->bytes_in_buffer--;
+  return *src->next_input_byte++;
+}
+
+
+/*
+ * The core arithmetic decoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Return value is 0 or 1 (binary decision).
+ *
+ * Note: I've changed the handling of the code base & bit
+ * buffer register C compared to other implementations
+ * based on the standards layout & procedures.
+ * While it also contains both the actual base of the
+ * coding interval (16 bits) and the next-bits buffer,
+ * the cut-point between these two parts is floating
+ * (instead of fixed) with the bit shift counter CT.
+ * Thus, we also need only one (variable instead of
+ * fixed size) shift for the LPS/MPS decision, and
+ * we can do away with any renormalization update
+ * of C (except for new data insertion, of course).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(int)
+arith_decode(j_decompress_ptr cinfo, unsigned char *st)
+{
+  register arith_entropy_ptr e = (arith_entropy_ptr)cinfo->entropy;
+  register unsigned char nl, nm;
+  register JLONG qe, temp;
+  register int sv, data;
+
+  /* Renormalization & data input per section D.2.6 */
+  while (e->a < 0x8000L) {
+    if (--e->ct < 0) {
+      /* Need to fetch next data byte */
+      if (cinfo->unread_marker)
+        data = 0;               /* stuff zero data */
+      else {
+        data = get_byte(cinfo); /* read next input byte */
+        if (data == 0xFF) {     /* zero stuff or marker code */
+          do data = get_byte(cinfo);
+          while (data == 0xFF); /* swallow extra 0xFF bytes */
+          if (data == 0)
+            data = 0xFF;        /* discard stuffed zero byte */
+          else {
+            /* Note: Different from the Huffman decoder, hitting
+             * a marker while processing the compressed data
+             * segment is legal in arithmetic coding.
+             * The convention is to supply zero data
+             * then until decoding is complete.
+             */
+            cinfo->unread_marker = data;
+            data = 0;
+          }
+        }
+      }
+      e->c = (e->c << 8) | data; /* insert data into C register */
+      if ((e->ct += 8) < 0)      /* update bit shift counter */
+        /* Need more initial bytes */
+        if (++e->ct == 0)
+          /* Got 2 initial bytes -> re-init A and exit loop */
+          e->a = 0x8000L; /* => e->a = 0x10000L after loop exit */
+    }
+    e->a <<= 1;
+  }
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jpeg_aritab[sv & 0x7F];  /* => Qe_Value */
+  nl = qe & 0xFF;  qe >>= 8;    /* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF;  qe >>= 8;    /* Next_Index_MPS */
+
+  /* Decode & estimation procedures per sections D.2.4 & D.2.5 */
+  temp = e->a - qe;
+  e->a = temp;
+  temp <<= e->ct;
+  if (e->c >= temp) {
+    e->c -= temp;
+    /* Conditional LPS (less probable symbol) exchange */
+    if (e->a < qe) {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nm;   /* Estimate_after_MPS */
+    } else {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nl;   /* Estimate_after_LPS */
+      sv ^= 0x80;               /* Exchange LPS/MPS */
+    }
+  } else if (e->a < 0x8000L) {
+    /* Conditional MPS (more probable symbol) exchange */
+    if (e->a < qe) {
+      *st = (sv & 0x80) ^ nl;   /* Estimate_after_LPS */
+      sv ^= 0x80;               /* Exchange LPS/MPS */
+    } else {
+      *st = (sv & 0x80) ^ nm;   /* Estimate_after_MPS */
+    }
+  }
+
+  return sv >> 7;
+}
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ */
+
+LOCAL(void)
+process_restart(j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Advance past the RSTn marker */
+  if (!(*cinfo->marker->read_restart_marker) (cinfo))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  /* Re-initialize statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (!cinfo->progressive_mode || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      memset(entropy->dc_stats[compptr->dc_tbl_no], 0, DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (!cinfo->progressive_mode || cinfo->Ss) {
+      memset(entropy->ac_stats[compptr->ac_tbl_no], 0, AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;    /* force reading 2 initial bytes to fill C */
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Arithmetic MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * arithmetic-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;   /* if error do nothing */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2;  st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+        st = entropy->dc_stats[tbl] + 20;       /* Table F.4: X1 = 20 */
+        while (arith_decode(cinfo, st)) {
+          if ((m <<= 1) == 0x8000) {
+            WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+            entropy->ct = -1;                   /* magnitude overflow */
+            return TRUE;
+          }
+          st += 1;
+        }
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int)((1L << cinfo->arith_dc_L[tbl]) >> 1))
+        entropy->dc_context[ci] = 0;               /* zero diff category */
+      else if (m > (int)((1L << cinfo->arith_dc_U[tbl]) >> 1))
+        entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+        entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        if (arith_decode(cinfo, st)) v |= m;
+      v += 1;  if (sign) v = -v;
+      entropy->last_dc_val[ci] = (entropy->last_dc_val[ci] + v) & 0xffff;
+    }
+
+    /* Scale and output the DC coefficient (assumes jpeg_natural_order[0]=0) */
+    (*block)[0] = (JCOEF)LEFT_SHIFT(entropy->last_dc_val[ci], cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, sign, k;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;   /* if error do nothing */
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+  /* Figure F.20: Decode_AC_coefficients */
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (arith_decode(cinfo, st)) break;         /* EOB flag */
+    while (arith_decode(cinfo, st + 1) == 0) {
+      st += 3;  k++;
+      if (k > cinfo->Se) {
+        WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+        entropy->ct = -1;                       /* spectral overflow */
+        return TRUE;
+      }
+    }
+    /* Figure F.21: Decoding nonzero value v */
+    /* Figure F.22: Decoding the sign of v */
+    sign = arith_decode(cinfo, entropy->fixed_bin);
+    st += 2;
+    /* Figure F.23: Decoding the magnitude category of v */
+    if ((m = arith_decode(cinfo, st)) != 0) {
+      if (arith_decode(cinfo, st)) {
+        m <<= 1;
+        st = entropy->ac_stats[tbl] +
+             (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+        while (arith_decode(cinfo, st)) {
+          if ((m <<= 1) == 0x8000) {
+            WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+            entropy->ct = -1;                   /* magnitude overflow */
+            return TRUE;
+          }
+          st += 1;
+        }
+      }
+    }
+    v = m;
+    /* Figure F.24: Decoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      if (arith_decode(cinfo, st)) v |= m;
+    v += 1;  if (sign) v = -v;
+    /* Scale and output coefficient in natural (dezigzagged) order */
+    (*block)[jpeg_natural_order[k]] = (JCOEF)((unsigned)v << cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  unsigned char *st;
+  int p1, blkn;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  st = entropy->fixed_bin;      /* use fixed probability estimation */
+  p1 = 1 << cinfo->Al;          /* 1 in the bit position being coded */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    if (arith_decode(cinfo, st))
+      MCU_data[blkn][0][0] |= p1;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  unsigned char *st;
+  int tbl, k, kex;
+  int p1, m1;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;   /* if error do nothing */
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  p1 = 1 << cinfo->Al;          /* 1 in the bit position being coded */
+  m1 = (NEG_1) << cinfo->Al;    /* -1 in the bit position being coded */
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = cinfo->Se; kex > 0; kex--)
+    if ((*block)[jpeg_natural_order[kex]]) break;
+
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k > kex)
+      if (arith_decode(cinfo, st)) break;       /* EOB flag */
+    for (;;) {
+      thiscoef = *block + jpeg_natural_order[k];
+      if (*thiscoef) {                          /* previously nonzero coef */
+        if (arith_decode(cinfo, st + 2)) {
+          if (*thiscoef < 0)
+            *thiscoef += (JCOEF)m1;
+          else
+            *thiscoef += (JCOEF)p1;
+        }
+        break;
+      }
+      if (arith_decode(cinfo, st + 1)) {        /* newly nonzero coef */
+        if (arith_decode(cinfo, entropy->fixed_bin))
+          *thiscoef = (JCOEF)m1;
+        else
+          *thiscoef = (JCOEF)p1;
+        break;
+      }
+      st += 3;  k++;
+      if (k > cinfo->Se) {
+        WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+        entropy->ct = -1;                       /* spectral overflow */
+        return TRUE;
+      }
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Decode one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+decode_mcu(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  jpeg_component_info *compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign, k;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;   /* if error do nothing */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data ? MCU_data[blkn] : NULL;
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2;  st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+        st = entropy->dc_stats[tbl] + 20;       /* Table F.4: X1 = 20 */
+        while (arith_decode(cinfo, st)) {
+          if ((m <<= 1) == 0x8000) {
+            WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+            entropy->ct = -1;                   /* magnitude overflow */
+            return TRUE;
+          }
+          st += 1;
+        }
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int)((1L << cinfo->arith_dc_L[tbl]) >> 1))
+        entropy->dc_context[ci] = 0;               /* zero diff category */
+      else if (m > (int)((1L << cinfo->arith_dc_U[tbl]) >> 1))
+        entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+        entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        if (arith_decode(cinfo, st)) v |= m;
+      v += 1;  if (sign) v = -v;
+      entropy->last_dc_val[ci] = (entropy->last_dc_val[ci] + v) & 0xffff;
+    }
+
+    if (block)
+      (*block)[0] = (JCOEF)entropy->last_dc_val[ci];
+
+    /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Figure F.20: Decode_AC_coefficients */
+    for (k = 1; k <= DCTSIZE2 - 1; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      if (arith_decode(cinfo, st)) break;       /* EOB flag */
+      while (arith_decode(cinfo, st + 1) == 0) {
+        st += 3;  k++;
+        if (k > DCTSIZE2 - 1) {
+          WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+          entropy->ct = -1;                     /* spectral overflow */
+          return TRUE;
+        }
+      }
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, entropy->fixed_bin);
+      st += 2;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+        if (arith_decode(cinfo, st)) {
+          m <<= 1;
+          st = entropy->ac_stats[tbl] +
+               (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+          while (arith_decode(cinfo, st)) {
+            if ((m <<= 1) == 0x8000) {
+              WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+              entropy->ct = -1;                 /* magnitude overflow */
+              return TRUE;
+            }
+            st += 1;
+          }
+        }
+      }
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+        if (arith_decode(cinfo, st)) v |= m;
+      v += 1;  if (sign) v = -v;
+      if (block)
+        (*block)[jpeg_natural_order[k]] = (JCOEF)v;
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass(j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr)cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info *compptr;
+
+  if (cinfo->progressive_mode) {
+    /* Validate progressive scan parameters */
+    if (cinfo->Ss == 0) {
+      if (cinfo->Se != 0)
+        goto bad;
+    } else {
+      /* need not check Ss/Se < 0 since they came from unsigned bytes */
+      if (cinfo->Se < cinfo->Ss || cinfo->Se > DCTSIZE2 - 1)
+        goto bad;
+      /* AC scans may have only one component */
+      if (cinfo->comps_in_scan != 1)
+        goto bad;
+    }
+    if (cinfo->Ah != 0) {
+      /* Successive approximation refinement scan: must have Al = Ah-1. */
+      if (cinfo->Ah - 1 != cinfo->Al)
+        goto bad;
+    }
+    if (cinfo->Al > 13) {       /* need not check for < 0 */
+bad:
+      ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+               cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+    }
+    /* Update progression status, and verify that scan order is legal.
+     * Note that inter-scan inconsistencies are treated as warnings
+     * not fatal errors ... not clear if this is right way to behave.
+     */
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      int coefi, cindex = cinfo->cur_comp_info[ci]->component_index;
+      int *coef_bit_ptr = &cinfo->coef_bits[cindex][0];
+      int *prev_coef_bit_ptr =
+        &cinfo->coef_bits[cindex + cinfo->num_components][0];
+      if (cinfo->Ss && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+        WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+      for (coefi = MIN(cinfo->Ss, 1); coefi <= MAX(cinfo->Se, 9); coefi++) {
+        if (cinfo->input_scan_number > 1)
+          prev_coef_bit_ptr[coefi] = coef_bit_ptr[coefi];
+        else
+          prev_coef_bit_ptr[coefi] = 0;
+      }
+      for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+        int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+        if (cinfo->Ah != expected)
+          WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+        coef_bit_ptr[coefi] = cinfo->Al;
+      }
+    }
+    /* Select MCU decoding routine */
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+        entropy->pub.decode_mcu = decode_mcu_DC_first;
+      else
+        entropy->pub.decode_mcu = decode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+        entropy->pub.decode_mcu = decode_mcu_DC_refine;
+      else
+        entropy->pub.decode_mcu = decode_mcu_AC_refine;
+    }
+  } else {
+    /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+     * This ought to be an error condition, but we make it a warning.
+     */
+    if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2 - 1 ||
+        cinfo->Ah != 0 || cinfo->Al != 0)
+      WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+    /* Select MCU decoding routine */
+    entropy->pub.decode_mcu = decode_mcu;
+  }
+
+  /* Allocate & initialize requested statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (!cinfo->progressive_mode || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+        ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+        entropy->dc_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
+          ((j_common_ptr)cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      memset(entropy->dc_stats[tbl], 0, DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (!cinfo->progressive_mode || cinfo->Ss) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+        ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+        entropy->ac_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
+          ((j_common_ptr)cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      memset(entropy->ac_stats[tbl], 0, AC_STAT_BINS);
+    }
+  }
+
+  /* Initialize arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;    /* force reading 2 initial bytes to fill C */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_arith_decoder(j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(arith_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *)entropy;
+  entropy->pub.start_pass = start_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+
+  /* Initialize index for fixed probability estimation */
+  entropy->fixed_bin[0] = 113;
+
+  if (cinfo->progressive_mode) {
+    /* Create progression status table */
+    int *coef_bit_ptr, ci;
+    cinfo->coef_bits = (int (*)[DCTSIZE2])
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  cinfo->num_components * 2 * DCTSIZE2 *
+                                  sizeof(int));
+    coef_bit_ptr = &cinfo->coef_bits[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++)
+      for (i = 0; i < DCTSIZE2; i++)
+        *coef_bit_ptr++ = -1;
+  }
+}

thirdparty/libjpeg-turbo/src/jdatadst-tj.c

@@ -0,0 +1,199 @@
+/*
+ * jdatadst-tj.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2012 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2011, 2014, 2016, 2019, 2022-2023, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains compression data destination routines for the case of
+ * emitting JPEG data to memory or to a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different destination manager.
+ * IMPORTANT: we assume that fwrite() will correctly transcribe an array of
+ * JOCTETs into 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+void jpeg_mem_dest_tj(j_compress_ptr cinfo, unsigned char **outbuffer,
+                      size_t *outsize, boolean alloc);
+
+
+#define OUTPUT_BUF_SIZE  4096   /* choose an efficiently fwrite'able size */
+
+
+/* Expanded data destination object for memory output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  unsigned char **outbuffer;    /* target buffer */
+  size_t *outsize;
+  unsigned char *newbuffer;     /* newly allocated buffer */
+  JOCTET *buffer;               /* start of buffer */
+  size_t bufsize;
+  boolean alloc;
+} my_mem_destination_mgr;
+
+typedef my_mem_destination_mgr *my_mem_dest_ptr;
+
+
+/*
+ * Initialize destination --- called by jpeg_start_compress
+ * before any data is actually written.
+ */
+
+METHODDEF(void)
+init_mem_destination(j_compress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Empty the output buffer --- called whenever buffer fills up.
+ *
+ * In typical applications, this should write the entire output buffer
+ * (ignoring the current state of next_output_byte & free_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been dumped.
+ *
+ * In applications that need to be able to suspend compression due to output
+ * overrun, a FALSE return indicates that the buffer cannot be emptied now.
+ * In this situation, the compressor will return to its caller (possibly with
+ * an indication that it has not accepted all the supplied scanlines).  The
+ * application should resume compression after it has made more room in the
+ * output buffer.  Note that there are substantial restrictions on the use of
+ * suspension --- see the documentation.
+ *
+ * When suspending, the compressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_output_byte & free_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point will be regenerated after resumption, so do not
+ * write it out when emptying the buffer externally.
+ */
+
+METHODDEF(boolean)
+empty_mem_output_buffer(j_compress_ptr cinfo)
+{
+  size_t nextsize;
+  JOCTET *nextbuffer;
+  my_mem_dest_ptr dest = (my_mem_dest_ptr)cinfo->dest;
+
+  if (!dest->alloc) ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Try to allocate new buffer with double size */
+  nextsize = dest->bufsize * 2;
+  nextbuffer = (JOCTET *)MALLOC(nextsize);
+
+  if (nextbuffer == NULL)
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+
+  memcpy(nextbuffer, dest->buffer, dest->bufsize);
+
+  free(dest->newbuffer);
+
+  dest->newbuffer = nextbuffer;
+
+  dest->pub.next_output_byte = nextbuffer + dest->bufsize;
+  dest->pub.free_in_buffer = dest->bufsize;
+
+  dest->buffer = nextbuffer;
+  dest->bufsize = nextsize;
+
+  return TRUE;
+}
+
+
+/*
+ * Terminate destination --- called by jpeg_finish_compress
+ * after all data has been written.  Usually needs to flush buffer.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_mem_destination(j_compress_ptr cinfo)
+{
+  my_mem_dest_ptr dest = (my_mem_dest_ptr)cinfo->dest;
+
+  if (dest->alloc) *dest->outbuffer = dest->buffer;
+  *dest->outsize = dest->bufsize - dest->pub.free_in_buffer;
+}
+
+
+/*
+ * Prepare for output to a memory buffer.
+ * The caller may supply an own initial buffer with appropriate size.
+ * Otherwise, or when the actual data output exceeds the given size,
+ * the library adapts the buffer size as necessary.
+ * The standard library functions malloc/free are used for allocating
+ * larger memory, so the buffer is available to the application after
+ * finishing compression, and then the application is responsible for
+ * freeing the requested memory.
+ */
+
+GLOBAL(void)
+jpeg_mem_dest_tj(j_compress_ptr cinfo, unsigned char **outbuffer,
+                 size_t *outsize, boolean alloc)
+{
+  boolean reused = FALSE;
+  my_mem_dest_ptr dest;
+
+  if (outbuffer == NULL || outsize == NULL)     /* sanity check */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same buffer without re-executing jpeg_mem_dest.
+   */
+  if (cinfo->dest == NULL) {    /* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(my_mem_destination_mgr));
+    dest = (my_mem_dest_ptr)cinfo->dest;
+    dest->newbuffer = NULL;
+    dest->buffer = NULL;
+  } else if (cinfo->dest->init_destination != init_mem_destination) {
+    /* It is unsafe to reuse the existing destination manager unless it was
+     * created by this function.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  dest = (my_mem_dest_ptr)cinfo->dest;
+  dest->pub.init_destination = init_mem_destination;
+  dest->pub.empty_output_buffer = empty_mem_output_buffer;
+  dest->pub.term_destination = term_mem_destination;
+  if (dest->buffer == *outbuffer && *outbuffer != NULL && alloc)
+    reused = TRUE;
+  dest->outbuffer = outbuffer;
+  dest->outsize = outsize;
+  dest->alloc = alloc;
+
+  if (*outbuffer == NULL || *outsize == 0) {
+    if (alloc) {
+      /* Allocate initial buffer */
+      dest->newbuffer = *outbuffer = (unsigned char *)MALLOC(OUTPUT_BUF_SIZE);
+      if (dest->newbuffer == NULL)
+        ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+      *outsize = OUTPUT_BUF_SIZE;
+    } else
+      ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  dest->pub.next_output_byte = dest->buffer = *outbuffer;
+  if (!reused)
+    dest->bufsize = *outsize;
+  dest->pub.free_in_buffer = dest->bufsize;
+}

thirdparty/libjpeg-turbo/src/jdatadst.c

@@ -0,0 +1,277 @@
+/*
+ * jdatadst.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2012 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2013, 2016, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains compression data destination routines for the case of
+ * emitting JPEG data to memory or to a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different destination manager.
+ * IMPORTANT: we assume that fwrite() will correctly transcribe an array of
+ * JOCTETs into 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data destination object for stdio output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  FILE *outfile;                /* target stream */
+  JOCTET *buffer;               /* start of buffer */
+} my_destination_mgr;
+
+typedef my_destination_mgr *my_dest_ptr;
+
+#define OUTPUT_BUF_SIZE  4096   /* choose an efficiently fwrite'able size */
+
+
+/* Expanded data destination object for memory output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  unsigned char **outbuffer;    /* target buffer */
+  unsigned long *outsize;
+  unsigned char *newbuffer;     /* newly allocated buffer */
+  JOCTET *buffer;               /* start of buffer */
+  size_t bufsize;
+} my_mem_destination_mgr;
+
+typedef my_mem_destination_mgr *my_mem_dest_ptr;
+
+
+/*
+ * Initialize destination --- called by jpeg_start_compress
+ * before any data is actually written.
+ */
+
+METHODDEF(void)
+init_destination(j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr)cinfo->dest;
+
+  /* Allocate the output buffer --- it will be released when done with image */
+  dest->buffer = (JOCTET *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                OUTPUT_BUF_SIZE * sizeof(JOCTET));
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+}
+
+METHODDEF(void)
+init_mem_destination(j_compress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Empty the output buffer --- called whenever buffer fills up.
+ *
+ * In typical applications, this should write the entire output buffer
+ * (ignoring the current state of next_output_byte & free_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been dumped.
+ *
+ * In applications that need to be able to suspend compression due to output
+ * overrun, a FALSE return indicates that the buffer cannot be emptied now.
+ * In this situation, the compressor will return to its caller (possibly with
+ * an indication that it has not accepted all the supplied scanlines).  The
+ * application should resume compression after it has made more room in the
+ * output buffer.  Note that there are substantial restrictions on the use of
+ * suspension --- see the documentation.
+ *
+ * When suspending, the compressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_output_byte & free_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point will be regenerated after resumption, so do not
+ * write it out when emptying the buffer externally.
+ */
+
+METHODDEF(boolean)
+empty_output_buffer(j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr)cinfo->dest;
+
+  if (fwrite(dest->buffer, 1, OUTPUT_BUF_SIZE, dest->outfile) !=
+      (size_t)OUTPUT_BUF_SIZE)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+
+  return TRUE;
+}
+
+METHODDEF(boolean)
+empty_mem_output_buffer(j_compress_ptr cinfo)
+{
+  size_t nextsize;
+  JOCTET *nextbuffer;
+  my_mem_dest_ptr dest = (my_mem_dest_ptr)cinfo->dest;
+
+  /* Try to allocate new buffer with double size */
+  nextsize = dest->bufsize * 2;
+  nextbuffer = (JOCTET *)malloc(nextsize);
+
+  if (nextbuffer == NULL)
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+
+  memcpy(nextbuffer, dest->buffer, dest->bufsize);
+
+  free(dest->newbuffer);
+
+  dest->newbuffer = nextbuffer;
+
+  dest->pub.next_output_byte = nextbuffer + dest->bufsize;
+  dest->pub.free_in_buffer = dest->bufsize;
+
+  dest->buffer = nextbuffer;
+  dest->bufsize = nextsize;
+
+  return TRUE;
+}
+
+
+/*
+ * Terminate destination --- called by jpeg_finish_compress
+ * after all data has been written.  Usually needs to flush buffer.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_destination(j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr)cinfo->dest;
+  size_t datacount = OUTPUT_BUF_SIZE - dest->pub.free_in_buffer;
+
+  /* Write any data remaining in the buffer */
+  if (datacount > 0) {
+    if (fwrite(dest->buffer, 1, datacount, dest->outfile) != datacount)
+      ERREXIT(cinfo, JERR_FILE_WRITE);
+  }
+  fflush(dest->outfile);
+  /* Make sure we wrote the output file OK */
+  if (ferror(dest->outfile))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+METHODDEF(void)
+term_mem_destination(j_compress_ptr cinfo)
+{
+  my_mem_dest_ptr dest = (my_mem_dest_ptr)cinfo->dest;
+
+  *dest->outbuffer = dest->buffer;
+  *dest->outsize = (unsigned long)(dest->bufsize - dest->pub.free_in_buffer);
+}
+
+
+/*
+ * Prepare for output to a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing compression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_dest(j_compress_ptr cinfo, FILE *outfile)
+{
+  my_dest_ptr dest;
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same file without re-executing jpeg_stdio_dest.
+   */
+  if (cinfo->dest == NULL) {    /* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(my_destination_mgr));
+  } else if (cinfo->dest->init_destination != init_destination) {
+    /* It is unsafe to reuse the existing destination manager unless it was
+     * created by this function.  Otherwise, there is no guarantee that the
+     * opaque structure is the right size.  Note that we could just create a
+     * new structure, but the old structure would not be freed until
+     * jpeg_destroy_compress() was called.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  dest = (my_dest_ptr)cinfo->dest;
+  dest->pub.init_destination = init_destination;
+  dest->pub.empty_output_buffer = empty_output_buffer;
+  dest->pub.term_destination = term_destination;
+  dest->outfile = outfile;
+}
+
+
+/*
+ * Prepare for output to a memory buffer.
+ * The caller may supply an own initial buffer with appropriate size.
+ * Otherwise, or when the actual data output exceeds the given size,
+ * the library adapts the buffer size as necessary.
+ * The standard library functions malloc/free are used for allocating
+ * larger memory, so the buffer is available to the application after
+ * finishing compression, and then the application is responsible for
+ * freeing the requested memory.
+ * Note:  An initial buffer supplied by the caller is expected to be
+ * managed by the application.  The library does not free such buffer
+ * when allocating a larger buffer.
+ */
+
+GLOBAL(void)
+jpeg_mem_dest(j_compress_ptr cinfo, unsigned char **outbuffer,
+              unsigned long *outsize)
+{
+  my_mem_dest_ptr dest;
+
+  if (outbuffer == NULL || outsize == NULL)     /* sanity check */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same buffer without re-executing jpeg_mem_dest.
+   */
+  if (cinfo->dest == NULL) {    /* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(my_mem_destination_mgr));
+  } else if (cinfo->dest->init_destination != init_mem_destination) {
+    /* It is unsafe to reuse the existing destination manager unless it was
+     * created by this function.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  dest = (my_mem_dest_ptr)cinfo->dest;
+  dest->pub.init_destination = init_mem_destination;
+  dest->pub.empty_output_buffer = empty_mem_output_buffer;
+  dest->pub.term_destination = term_mem_destination;
+  dest->outbuffer = outbuffer;
+  dest->outsize = outsize;
+  dest->newbuffer = NULL;
+
+  if (*outbuffer == NULL || *outsize == 0) {
+    /* Allocate initial buffer */
+    dest->newbuffer = *outbuffer = (unsigned char *)malloc(OUTPUT_BUF_SIZE);
+    if (dest->newbuffer == NULL)
+      ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+    *outsize = OUTPUT_BUF_SIZE;
+  }
+
+  dest->pub.next_output_byte = dest->buffer = *outbuffer;
+  dest->pub.free_in_buffer = dest->bufsize = *outsize;
+}

thirdparty/libjpeg-turbo/src/jdatasrc-tj.c

@@ -0,0 +1,194 @@
+/*
+ * jdatasrc-tj.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2011 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2011, 2016, 2019, 2023, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains decompression data source routines for the case of
+ * reading JPEG data from memory or from a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different source manager.
+ * IMPORTANT: we assume that fread() will correctly transcribe an array of
+ * JOCTETs from 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+void jpeg_mem_src_tj(j_decompress_ptr cinfo, const unsigned char *inbuffer,
+                     size_t insize);
+
+
+/*
+ * Initialize source --- called by jpeg_read_header
+ * before any data is actually read.
+ */
+
+METHODDEF(void)
+init_mem_source(j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Fill the input buffer --- called whenever buffer is emptied.
+ *
+ * In typical applications, this should read fresh data into the buffer
+ * (ignoring the current state of next_input_byte & bytes_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been reloaded.  It is not necessary to
+ * fill the buffer entirely, only to obtain at least one more byte.
+ *
+ * There is no such thing as an EOF return.  If the end of the file has been
+ * reached, the routine has a choice of ERREXIT() or inserting fake data into
+ * the buffer.  In most cases, generating a warning message and inserting a
+ * fake EOI marker is the best course of action --- this will allow the
+ * decompressor to output however much of the image is there.  However,
+ * the resulting error message is misleading if the real problem is an empty
+ * input file, so we handle that case specially.
+ *
+ * In applications that need to be able to suspend compression due to input
+ * not being available yet, a FALSE return indicates that no more data can be
+ * obtained right now, but more may be forthcoming later.  In this situation,
+ * the decompressor will return to its caller (with an indication of the
+ * number of scanlines it has read, if any).  The application should resume
+ * decompression after it has loaded more data into the input buffer.  Note
+ * that there are substantial restrictions on the use of suspension --- see
+ * the documentation.
+ *
+ * When suspending, the decompressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_input_byte & bytes_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point must be rescanned after resumption, so move it to
+ * the front of the buffer rather than discarding it.
+ */
+
+METHODDEF(boolean)
+fill_mem_input_buffer(j_decompress_ptr cinfo)
+{
+  static const JOCTET mybuffer[4] = {
+    (JOCTET)0xFF, (JOCTET)JPEG_EOI, 0, 0
+  };
+
+  /* The whole JPEG data is expected to reside in the supplied memory
+   * buffer, so any request for more data beyond the given buffer size
+   * is treated as an error.
+   */
+  WARNMS(cinfo, JWRN_JPEG_EOF);
+
+  /* Insert a fake EOI marker */
+
+  cinfo->src->next_input_byte = mybuffer;
+  cinfo->src->bytes_in_buffer = 2;
+
+  return TRUE;
+}
+
+
+/*
+ * Skip data --- used to skip over a potentially large amount of
+ * uninteresting data (such as an APPn marker).
+ *
+ * Writers of suspendable-input applications must note that skip_input_data
+ * is not granted the right to give a suspension return.  If the skip extends
+ * beyond the data currently in the buffer, the buffer can be marked empty so
+ * that the next read will cause a fill_input_buffer call that can suspend.
+ * Arranging for additional bytes to be discarded before reloading the input
+ * buffer is the application writer's problem.
+ */
+
+METHODDEF(void)
+skip_input_data(j_decompress_ptr cinfo, long num_bytes)
+{
+  struct jpeg_source_mgr *src = cinfo->src;
+
+  /* Just a dumb implementation for now.  Could use fseek() except
+   * it doesn't work on pipes.  Not clear that being smart is worth
+   * any trouble anyway --- large skips are infrequent.
+   */
+  if (num_bytes > 0) {
+    while (num_bytes > (long)src->bytes_in_buffer) {
+      num_bytes -= (long)src->bytes_in_buffer;
+      (void)(*src->fill_input_buffer) (cinfo);
+      /* note we assume that fill_input_buffer will never return FALSE,
+       * so suspension need not be handled.
+       */
+    }
+    src->next_input_byte += (size_t)num_bytes;
+    src->bytes_in_buffer -= (size_t)num_bytes;
+  }
+}
+
+
+/*
+ * An additional method that can be provided by data source modules is the
+ * resync_to_restart method for error recovery in the presence of RST markers.
+ * For the moment, this source module just uses the default resync method
+ * provided by the JPEG library.  That method assumes that no backtracking
+ * is possible.
+ */
+
+
+/*
+ * Terminate source --- called by jpeg_finish_decompress
+ * after all data has been read.  Often a no-op.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_source(j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Prepare for input from a supplied memory buffer.
+ * The buffer must contain the whole JPEG data.
+ */
+
+GLOBAL(void)
+jpeg_mem_src_tj(j_decompress_ptr cinfo, const unsigned char *inbuffer,
+                size_t insize)
+{
+  struct jpeg_source_mgr *src;
+
+  if (inbuffer == NULL || insize == 0)  /* Treat empty input as fatal error */
+    ERREXIT(cinfo, JERR_INPUT_EMPTY);
+
+  /* The source object is made permanent so that a series of JPEG images
+   * can be read from the same buffer by calling jpeg_mem_src only before
+   * the first one.
+   */
+  if (cinfo->src == NULL) {     /* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(struct jpeg_source_mgr));
+  } else if (cinfo->src->init_source != init_mem_source) {
+    /* It is unsafe to reuse the existing source manager unless it was created
+     * by this function.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  src = cinfo->src;
+  src->init_source = init_mem_source;
+  src->fill_input_buffer = fill_mem_input_buffer;
+  src->skip_input_data = skip_input_data;
+  src->resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->term_source = term_source;
+  src->bytes_in_buffer = insize;
+  src->next_input_byte = (const JOCTET *)inbuffer;
+}

thirdparty/libjpeg-turbo/src/jdatasrc.c

@@ -0,0 +1,289 @@
+/*
+ * jdatasrc.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2011 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2013, 2016, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains decompression data source routines for the case of
+ * reading JPEG data from memory or from a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different source manager.
+ * IMPORTANT: we assume that fread() will correctly transcribe an array of
+ * JOCTETs from 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data source object for stdio input */
+
+typedef struct {
+  struct jpeg_source_mgr pub;   /* public fields */
+
+  FILE *infile;                 /* source stream */
+  JOCTET *buffer;               /* start of buffer */
+  boolean start_of_file;        /* have we gotten any data yet? */
+} my_source_mgr;
+
+typedef my_source_mgr *my_src_ptr;
+
+#define INPUT_BUF_SIZE  4096    /* choose an efficiently fread'able size */
+
+
+/*
+ * Initialize source --- called by jpeg_read_header
+ * before any data is actually read.
+ */
+
+METHODDEF(void)
+init_source(j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr)cinfo->src;
+
+  /* We reset the empty-input-file flag for each image,
+   * but we don't clear the input buffer.
+   * This is correct behavior for reading a series of images from one source.
+   */
+  src->start_of_file = TRUE;
+}
+
+METHODDEF(void)
+init_mem_source(j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Fill the input buffer --- called whenever buffer is emptied.
+ *
+ * In typical applications, this should read fresh data into the buffer
+ * (ignoring the current state of next_input_byte & bytes_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been reloaded.  It is not necessary to
+ * fill the buffer entirely, only to obtain at least one more byte.
+ *
+ * There is no such thing as an EOF return.  If the end of the file has been
+ * reached, the routine has a choice of ERREXIT() or inserting fake data into
+ * the buffer.  In most cases, generating a warning message and inserting a
+ * fake EOI marker is the best course of action --- this will allow the
+ * decompressor to output however much of the image is there.  However,
+ * the resulting error message is misleading if the real problem is an empty
+ * input file, so we handle that case specially.
+ *
+ * In applications that need to be able to suspend compression due to input
+ * not being available yet, a FALSE return indicates that no more data can be
+ * obtained right now, but more may be forthcoming later.  In this situation,
+ * the decompressor will return to its caller (with an indication of the
+ * number of scanlines it has read, if any).  The application should resume
+ * decompression after it has loaded more data into the input buffer.  Note
+ * that there are substantial restrictions on the use of suspension --- see
+ * the documentation.
+ *
+ * When suspending, the decompressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_input_byte & bytes_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point must be rescanned after resumption, so move it to
+ * the front of the buffer rather than discarding it.
+ */
+
+METHODDEF(boolean)
+fill_input_buffer(j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr)cinfo->src;
+  size_t nbytes;
+
+  nbytes = fread(src->buffer, 1, INPUT_BUF_SIZE, src->infile);
+
+  if (nbytes <= 0) {
+    if (src->start_of_file)     /* Treat empty input file as fatal error */
+      ERREXIT(cinfo, JERR_INPUT_EMPTY);
+    WARNMS(cinfo, JWRN_JPEG_EOF);
+    /* Insert a fake EOI marker */
+    src->buffer[0] = (JOCTET)0xFF;
+    src->buffer[1] = (JOCTET)JPEG_EOI;
+    nbytes = 2;
+  }
+
+  src->pub.next_input_byte = src->buffer;
+  src->pub.bytes_in_buffer = nbytes;
+  src->start_of_file = FALSE;
+
+  return TRUE;
+}
+
+METHODDEF(boolean)
+fill_mem_input_buffer(j_decompress_ptr cinfo)
+{
+  static const JOCTET mybuffer[4] = {
+    (JOCTET)0xFF, (JOCTET)JPEG_EOI, 0, 0
+  };
+
+  /* The whole JPEG data is expected to reside in the supplied memory
+   * buffer, so any request for more data beyond the given buffer size
+   * is treated as an error.
+   */
+  WARNMS(cinfo, JWRN_JPEG_EOF);
+
+  /* Insert a fake EOI marker */
+
+  cinfo->src->next_input_byte = mybuffer;
+  cinfo->src->bytes_in_buffer = 2;
+
+  return TRUE;
+}
+
+
+/*
+ * Skip data --- used to skip over a potentially large amount of
+ * uninteresting data (such as an APPn marker).
+ *
+ * Writers of suspendable-input applications must note that skip_input_data
+ * is not granted the right to give a suspension return.  If the skip extends
+ * beyond the data currently in the buffer, the buffer can be marked empty so
+ * that the next read will cause a fill_input_buffer call that can suspend.
+ * Arranging for additional bytes to be discarded before reloading the input
+ * buffer is the application writer's problem.
+ */
+
+METHODDEF(void)
+skip_input_data(j_decompress_ptr cinfo, long num_bytes)
+{
+  struct jpeg_source_mgr *src = cinfo->src;
+
+  /* Just a dumb implementation for now.  Could use fseek() except
+   * it doesn't work on pipes.  Not clear that being smart is worth
+   * any trouble anyway --- large skips are infrequent.
+   */
+  if (num_bytes > 0) {
+    while (num_bytes > (long)src->bytes_in_buffer) {
+      num_bytes -= (long)src->bytes_in_buffer;
+      (void)(*src->fill_input_buffer) (cinfo);
+      /* note we assume that fill_input_buffer will never return FALSE,
+       * so suspension need not be handled.
+       */
+    }
+    src->next_input_byte += (size_t)num_bytes;
+    src->bytes_in_buffer -= (size_t)num_bytes;
+  }
+}
+
+
+/*
+ * An additional method that can be provided by data source modules is the
+ * resync_to_restart method for error recovery in the presence of RST markers.
+ * For the moment, this source module just uses the default resync method
+ * provided by the JPEG library.  That method assumes that no backtracking
+ * is possible.
+ */
+
+
+/*
+ * Terminate source --- called by jpeg_finish_decompress
+ * after all data has been read.  Often a no-op.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_source(j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Prepare for input from a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing decompression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_src(j_decompress_ptr cinfo, FILE *infile)
+{
+  my_src_ptr src;
+
+  /* The source object and input buffer are made permanent so that a series
+   * of JPEG images can be read from the same file by calling jpeg_stdio_src
+   * only before the first one.  (If we discarded the buffer at the end of
+   * one image, we'd likely lose the start of the next one.)
+   */
+  if (cinfo->src == NULL) {     /* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(my_source_mgr));
+    src = (my_src_ptr)cinfo->src;
+    src->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  INPUT_BUF_SIZE * sizeof(JOCTET));
+  } else if (cinfo->src->init_source != init_source) {
+    /* It is unsafe to reuse the existing source manager unless it was created
+     * by this function.  Otherwise, there is no guarantee that the opaque
+     * structure is the right size.  Note that we could just create a new
+     * structure, but the old structure would not be freed until
+     * jpeg_destroy_decompress() was called.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  src = (my_src_ptr)cinfo->src;
+  src->pub.init_source = init_source;
+  src->pub.fill_input_buffer = fill_input_buffer;
+  src->pub.skip_input_data = skip_input_data;
+  src->pub.resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->pub.term_source = term_source;
+  src->infile = infile;
+  src->pub.bytes_in_buffer = 0; /* forces fill_input_buffer on first read */
+  src->pub.next_input_byte = NULL; /* until buffer loaded */
+}
+
+
+/*
+ * Prepare for input from a supplied memory buffer.
+ * The buffer must contain the whole JPEG data.
+ */
+
+GLOBAL(void)
+jpeg_mem_src(j_decompress_ptr cinfo, const unsigned char *inbuffer,
+             unsigned long insize)
+{
+  struct jpeg_source_mgr *src;
+
+  if (inbuffer == NULL || insize == 0)  /* Treat empty input as fatal error */
+    ERREXIT(cinfo, JERR_INPUT_EMPTY);
+
+  /* The source object is made permanent so that a series of JPEG images
+   * can be read from the same buffer by calling jpeg_mem_src only before
+   * the first one.
+   */
+  if (cinfo->src == NULL) {     /* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                  sizeof(struct jpeg_source_mgr));
+  } else if (cinfo->src->init_source != init_mem_source) {
+    /* It is unsafe to reuse the existing source manager unless it was created
+     * by this function.
+     */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  }
+
+  src = cinfo->src;
+  src->init_source = init_mem_source;
+  src->fill_input_buffer = fill_mem_input_buffer;
+  src->skip_input_data = skip_input_data;
+  src->resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->term_source = term_source;
+  src->bytes_in_buffer = (size_t)insize;
+  src->next_input_byte = (const JOCTET *)inbuffer;
+}

thirdparty/libjpeg-turbo/src/jdcoefct.c

@@ -0,0 +1,885 @@
+/*
+ * jdcoefct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2010, 2015-2016, 2019-2020, 2022-2023, D. R. Commander.
+ * Copyright (C) 2015, 2020, Google, Inc.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the coefficient buffer controller for decompression.
+ * This controller is the top level of the lossy JPEG decompressor proper.
+ * The coefficient buffer lies between entropy decoding and inverse-DCT steps.
+ *
+ * In buffered-image mode, this controller is the interface between
+ * input-oriented processing and output-oriented processing.
+ * Also, the input side (only) is used when reading a file for transcoding.
+ */
+
+#include "jinclude.h"
+#include "jdcoefct.h"
+#include "jpegapicomp.h"
+#include "jsamplecomp.h"
+
+
+/* Forward declarations */
+METHODDEF(int) decompress_onepass(j_decompress_ptr cinfo,
+                                  _JSAMPIMAGE output_buf);
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+METHODDEF(int) decompress_data(j_decompress_ptr cinfo, _JSAMPIMAGE output_buf);
+#endif
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+LOCAL(boolean) smoothing_ok(j_decompress_ptr cinfo);
+METHODDEF(int) decompress_smooth_data(j_decompress_ptr cinfo,
+                                      _JSAMPIMAGE output_buf);
+#endif
+
+
+/*
+ * Initialize for an input processing pass.
+ */
+
+METHODDEF(void)
+start_input_pass(j_decompress_ptr cinfo)
+{
+  cinfo->input_iMCU_row = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Initialize for an output processing pass.
+ */
+
+METHODDEF(void)
+start_output_pass(j_decompress_ptr cinfo)
+{
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  /* If multipass, check to see whether to use block smoothing on this pass */
+  if (coef->pub.coef_arrays != NULL) {
+    if (cinfo->do_block_smoothing && smoothing_ok(cinfo))
+      coef->pub._decompress_data = decompress_smooth_data;
+    else
+      coef->pub._decompress_data = decompress_data;
+  }
+#endif
+  cinfo->output_iMCU_row = 0;
+}
+
+
+/*
+ * Decompress and return some data in the single-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Input and output must run in lockstep since we have only a one-MCU buffer.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(int)
+decompress_onepass(j_decompress_ptr cinfo, _JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION MCU_col_num;       /* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, useful_width;
+  _JSAMPARRAY output_ptr;
+  JDIMENSION start_col, output_col;
+  jpeg_component_info *compptr;
+  _inverse_DCT_method_ptr inverse_DCT;
+
+  /* Loop to process as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num <= last_MCU_col;
+         MCU_col_num++) {
+      /* Try to fetch an MCU.  Entropy decoder expects buffer to be zeroed. */
+      jzero_far((void *)coef->MCU_buffer[0],
+                (size_t)(cinfo->blocks_in_MCU * sizeof(JBLOCK)));
+      if (!cinfo->entropy->insufficient_data)
+        cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
+      if (!(*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+        /* Suspension forced; update state counters and exit */
+        coef->MCU_vert_offset = yoffset;
+        coef->MCU_ctr = MCU_col_num;
+        return JPEG_SUSPENDED;
+      }
+
+      /* Only perform the IDCT on blocks that are contained within the desired
+       * cropping region.
+       */
+      if (MCU_col_num >= cinfo->master->first_iMCU_col &&
+          MCU_col_num <= cinfo->master->last_iMCU_col) {
+        /* Determine where data should go in output_buf and do the IDCT thing.
+         * We skip dummy blocks at the right and bottom edges (but blkn gets
+         * incremented past them!).  Note the inner loop relies on having
+         * allocated the MCU_buffer[] blocks sequentially.
+         */
+        blkn = 0;               /* index of current DCT block within MCU */
+        for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+          compptr = cinfo->cur_comp_info[ci];
+          /* Don't bother to IDCT an uninteresting component. */
+          if (!compptr->component_needed) {
+            blkn += compptr->MCU_blocks;
+            continue;
+          }
+          inverse_DCT = cinfo->idct->_inverse_DCT[compptr->component_index];
+          useful_width = (MCU_col_num < last_MCU_col) ?
+                         compptr->MCU_width : compptr->last_col_width;
+          output_ptr = output_buf[compptr->component_index] +
+                       yoffset * compptr->_DCT_scaled_size;
+          start_col = (MCU_col_num - cinfo->master->first_iMCU_col) *
+                      compptr->MCU_sample_width;
+          for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+            if (cinfo->input_iMCU_row < last_iMCU_row ||
+                yoffset + yindex < compptr->last_row_height) {
+              output_col = start_col;
+              for (xindex = 0; xindex < useful_width; xindex++) {
+                (*inverse_DCT) (cinfo, compptr,
+                                (JCOEFPTR)coef->MCU_buffer[blkn + xindex],
+                                output_ptr, output_col);
+                output_col += compptr->_DCT_scaled_size;
+              }
+            }
+            blkn += compptr->MCU_width;
+            output_ptr += compptr->_DCT_scaled_size;
+          }
+        }
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  cinfo->output_iMCU_row++;
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Dummy consume-input routine for single-pass operation.
+ */
+
+METHODDEF(int)
+dummy_consume_data(j_decompress_ptr cinfo)
+{
+  return JPEG_SUSPENDED;        /* Always indicate nothing was done */
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Consume input data and store it in the full-image coefficient buffer.
+ * We read as much as one fully interleaved MCU row ("iMCU" row) per call,
+ * ie, v_samp_factor block rows for each component in the scan.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ */
+
+METHODDEF(int)
+consume_data(j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION MCU_col_num;       /* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr)cinfo, coef->whole_image[compptr->component_index],
+       cinfo->input_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION)compptr->v_samp_factor, TRUE);
+    /* Note: entropy decoder expects buffer to be zeroed,
+     * but this is handled automatically by the memory manager
+     * because we requested a pre-zeroed array.
+     */
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num < cinfo->MCUs_per_row;
+         MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;                 /* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+        compptr = cinfo->cur_comp_info[ci];
+        start_col = MCU_col_num * compptr->MCU_width;
+        for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+          buffer_ptr = buffer[ci][yindex + yoffset] + start_col;
+          for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+            coef->MCU_buffer[blkn++] = buffer_ptr++;
+          }
+        }
+      }
+      if (!cinfo->entropy->insufficient_data)
+        cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
+      /* Try to fetch the MCU. */
+      if (!(*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+        /* Suspension forced; update state counters and exit */
+        coef->MCU_vert_offset = yoffset;
+        coef->MCU_ctr = MCU_col_num;
+        return JPEG_SUSPENDED;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Decompress and return some data in the multi-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image.
+ */
+
+METHODDEF(int)
+decompress_data(j_decompress_ptr cinfo, _JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num;
+  int ci, block_row, block_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr;
+  _JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  _inverse_DCT_method_ptr inverse_DCT;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number < cinfo->output_scan_number ||
+         (cinfo->input_scan_number == cinfo->output_scan_number &&
+          cinfo->input_iMCU_row <= cinfo->output_iMCU_row)) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (!compptr->component_needed)
+      continue;
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr)cinfo, coef->whole_image[ci],
+       cinfo->output_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION)compptr->v_samp_factor, FALSE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    inverse_DCT = cinfo->idct->_inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row] + cinfo->master->first_MCU_col[ci];
+      output_col = 0;
+      for (block_num = cinfo->master->first_MCU_col[ci];
+           block_num <= cinfo->master->last_MCU_col[ci]; block_num++) {
+        (*inverse_DCT) (cinfo, compptr, (JCOEFPTR)buffer_ptr, output_ptr,
+                        output_col);
+        buffer_ptr++;
+        output_col += compptr->_DCT_scaled_size;
+      }
+      output_ptr += compptr->_DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+
+/*
+ * This code applies interblock smoothing; the first 9 AC coefficients are
+ * estimated from the DC values of a DCT block and its 24 neighboring blocks.
+ * We apply smoothing only for progressive JPEG decoding, and only if
+ * the coefficients it can estimate are not yet known to full precision.
+ */
+
+/* Natural-order array positions of the first 9 zigzag-order coefficients */
+#define Q01_POS  1
+#define Q10_POS  8
+#define Q20_POS  16
+#define Q11_POS  9
+#define Q02_POS  2
+#define Q03_POS  3
+#define Q12_POS  10
+#define Q21_POS  17
+#define Q30_POS  24
+
+/*
+ * Determine whether block smoothing is applicable and safe.
+ * We also latch the current states of the coef_bits[] entries for the
+ * AC coefficients; otherwise, if the input side of the decompressor
+ * advances into a new scan, we might think the coefficients are known
+ * more accurately than they really are.
+ */
+
+LOCAL(boolean)
+smoothing_ok(j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  boolean smoothing_useful = FALSE;
+  int ci, coefi;
+  jpeg_component_info *compptr;
+  JQUANT_TBL *qtable;
+  int *coef_bits, *prev_coef_bits;
+  int *coef_bits_latch, *prev_coef_bits_latch;
+
+  if (!cinfo->progressive_mode || cinfo->coef_bits == NULL)
+    return FALSE;
+
+  /* Allocate latch area if not already done */
+  if (coef->coef_bits_latch == NULL)
+    coef->coef_bits_latch = (int *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  cinfo->num_components * 2 *
+                                  (SAVED_COEFS * sizeof(int)));
+  coef_bits_latch = coef->coef_bits_latch;
+  prev_coef_bits_latch =
+    &coef->coef_bits_latch[cinfo->num_components * SAVED_COEFS];
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* All components' quantization values must already be latched. */
+    if ((qtable = compptr->quant_table) == NULL)
+      return FALSE;
+    /* Verify DC & first 9 AC quantizers are nonzero to avoid zero-divide. */
+    if (qtable->quantval[0] == 0 ||
+        qtable->quantval[Q01_POS] == 0 ||
+        qtable->quantval[Q10_POS] == 0 ||
+        qtable->quantval[Q20_POS] == 0 ||
+        qtable->quantval[Q11_POS] == 0 ||
+        qtable->quantval[Q02_POS] == 0 ||
+        qtable->quantval[Q03_POS] == 0 ||
+        qtable->quantval[Q12_POS] == 0 ||
+        qtable->quantval[Q21_POS] == 0 ||
+        qtable->quantval[Q30_POS] == 0)
+      return FALSE;
+    /* DC values must be at least partly known for all components. */
+    coef_bits = cinfo->coef_bits[ci];
+    prev_coef_bits = cinfo->coef_bits[ci + cinfo->num_components];
+    if (coef_bits[0] < 0)
+      return FALSE;
+    coef_bits_latch[0] = coef_bits[0];
+    /* Block smoothing is helpful if some AC coefficients remain inaccurate. */
+    for (coefi = 1; coefi < SAVED_COEFS; coefi++) {
+      if (cinfo->input_scan_number > 1)
+        prev_coef_bits_latch[coefi] = prev_coef_bits[coefi];
+      else
+        prev_coef_bits_latch[coefi] = -1;
+      coef_bits_latch[coefi] = coef_bits[coefi];
+      if (coef_bits[coefi] != 0)
+        smoothing_useful = TRUE;
+    }
+    coef_bits_latch += SAVED_COEFS;
+    prev_coef_bits_latch += SAVED_COEFS;
+  }
+
+  return smoothing_useful;
+}
+
+
+/*
+ * Variant of decompress_data for use when doing block smoothing.
+ */
+
+METHODDEF(int)
+decompress_smooth_data(j_decompress_ptr cinfo, _JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num, last_block_column;
+  int ci, block_row, block_rows, access_rows, image_block_row,
+    image_block_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr, prev_prev_block_row, prev_block_row;
+  JBLOCKROW next_block_row, next_next_block_row;
+  _JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  _inverse_DCT_method_ptr inverse_DCT;
+  boolean change_dc;
+  JCOEF *workspace;
+  int *coef_bits;
+  JQUANT_TBL *quanttbl;
+  JLONG Q00, Q01, Q02, Q03 = 0, Q10, Q11, Q12 = 0, Q20, Q21 = 0, Q30 = 0, num;
+  int DC01, DC02, DC03, DC04, DC05, DC06, DC07, DC08, DC09, DC10, DC11, DC12,
+      DC13, DC14, DC15, DC16, DC17, DC18, DC19, DC20, DC21, DC22, DC23, DC24,
+      DC25;
+  int Al, pred;
+
+  /* Keep a local variable to avoid looking it up more than once */
+  workspace = coef->workspace;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+         !cinfo->inputctl->eoi_reached) {
+    if (cinfo->input_scan_number == cinfo->output_scan_number) {
+      /* If input is working on current scan, we ordinarily want it to
+       * have completed the current row.  But if input scan is DC,
+       * we want it to keep two rows ahead so that next two block rows' DC
+       * values are up to date.
+       */
+      JDIMENSION delta = (cinfo->Ss == 0) ? 2 : 0;
+      if (cinfo->input_iMCU_row > cinfo->output_iMCU_row + delta)
+        break;
+    }
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (!compptr->component_needed)
+      continue;
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row + 1 < last_iMCU_row) {
+      block_rows = compptr->v_samp_factor;
+      access_rows = block_rows * 3; /* this and next two iMCU rows */
+    } else if (cinfo->output_iMCU_row < last_iMCU_row) {
+      block_rows = compptr->v_samp_factor;
+      access_rows = block_rows * 2; /* this and next iMCU row */
+    } else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+      access_rows = block_rows; /* this iMCU row only */
+    }
+    /* Align the virtual buffer for this component. */
+    if (cinfo->output_iMCU_row > 1) {
+      access_rows += 2 * compptr->v_samp_factor; /* prior two iMCU rows too */
+      buffer = (*cinfo->mem->access_virt_barray)
+        ((j_common_ptr)cinfo, coef->whole_image[ci],
+         (cinfo->output_iMCU_row - 2) * compptr->v_samp_factor,
+         (JDIMENSION)access_rows, FALSE);
+      buffer += 2 * compptr->v_samp_factor; /* point to current iMCU row */
+    } else if (cinfo->output_iMCU_row > 0) {
+      access_rows += compptr->v_samp_factor; /* prior iMCU row too */
+      buffer = (*cinfo->mem->access_virt_barray)
+        ((j_common_ptr)cinfo, coef->whole_image[ci],
+         (cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
+         (JDIMENSION)access_rows, FALSE);
+      buffer += compptr->v_samp_factor; /* point to current iMCU row */
+    } else {
+      buffer = (*cinfo->mem->access_virt_barray)
+        ((j_common_ptr)cinfo, coef->whole_image[ci],
+         (JDIMENSION)0, (JDIMENSION)access_rows, FALSE);
+    }
+    /* Fetch component-dependent info.
+     * If the current scan is incomplete, then we use the component-dependent
+     * info from the previous scan.
+     */
+    if (cinfo->output_iMCU_row > cinfo->master->last_good_iMCU_row)
+      coef_bits =
+        coef->coef_bits_latch + ((ci + cinfo->num_components) * SAVED_COEFS);
+    else
+      coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+
+    /* We only do DC interpolation if no AC coefficient data is available. */
+    change_dc =
+      coef_bits[1] == -1 && coef_bits[2] == -1 && coef_bits[3] == -1 &&
+      coef_bits[4] == -1 && coef_bits[5] == -1 && coef_bits[6] == -1 &&
+      coef_bits[7] == -1 && coef_bits[8] == -1 && coef_bits[9] == -1;
+
+    quanttbl = compptr->quant_table;
+    Q00 = quanttbl->quantval[0];
+    Q01 = quanttbl->quantval[Q01_POS];
+    Q10 = quanttbl->quantval[Q10_POS];
+    Q20 = quanttbl->quantval[Q20_POS];
+    Q11 = quanttbl->quantval[Q11_POS];
+    Q02 = quanttbl->quantval[Q02_POS];
+    if (change_dc) {
+      Q03 = quanttbl->quantval[Q03_POS];
+      Q12 = quanttbl->quantval[Q12_POS];
+      Q21 = quanttbl->quantval[Q21_POS];
+      Q30 = quanttbl->quantval[Q30_POS];
+    }
+    inverse_DCT = cinfo->idct->_inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    image_block_rows = block_rows * cinfo->total_iMCU_rows;
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      image_block_row = cinfo->output_iMCU_row * block_rows + block_row;
+      buffer_ptr = buffer[block_row] + cinfo->master->first_MCU_col[ci];
+
+      if (image_block_row > 0)
+        prev_block_row =
+          buffer[block_row - 1] + cinfo->master->first_MCU_col[ci];
+      else
+        prev_block_row = buffer_ptr;
+
+      if (image_block_row > 1)
+        prev_prev_block_row =
+          buffer[block_row - 2] + cinfo->master->first_MCU_col[ci];
+      else
+        prev_prev_block_row = prev_block_row;
+
+      if (image_block_row < image_block_rows - 1)
+        next_block_row =
+          buffer[block_row + 1] + cinfo->master->first_MCU_col[ci];
+      else
+        next_block_row = buffer_ptr;
+
+      if (image_block_row < image_block_rows - 2)
+        next_next_block_row =
+          buffer[block_row + 2] + cinfo->master->first_MCU_col[ci];
+      else
+        next_next_block_row = next_block_row;
+
+      /* We fetch the surrounding DC values using a sliding-register approach.
+       * Initialize all 25 here so as to do the right thing on narrow pics.
+       */
+      DC01 = DC02 = DC03 = DC04 = DC05 = (int)prev_prev_block_row[0][0];
+      DC06 = DC07 = DC08 = DC09 = DC10 = (int)prev_block_row[0][0];
+      DC11 = DC12 = DC13 = DC14 = DC15 = (int)buffer_ptr[0][0];
+      DC16 = DC17 = DC18 = DC19 = DC20 = (int)next_block_row[0][0];
+      DC21 = DC22 = DC23 = DC24 = DC25 = (int)next_next_block_row[0][0];
+      output_col = 0;
+      last_block_column = compptr->width_in_blocks - 1;
+      for (block_num = cinfo->master->first_MCU_col[ci];
+           block_num <= cinfo->master->last_MCU_col[ci]; block_num++) {
+        /* Fetch current DCT block into workspace so we can modify it. */
+        jcopy_block_row(buffer_ptr, (JBLOCKROW)workspace, (JDIMENSION)1);
+        /* Update DC values */
+        if (block_num == cinfo->master->first_MCU_col[ci] &&
+            block_num < last_block_column) {
+          DC04 = DC05 = (int)prev_prev_block_row[1][0];
+          DC09 = DC10 = (int)prev_block_row[1][0];
+          DC14 = DC15 = (int)buffer_ptr[1][0];
+          DC19 = DC20 = (int)next_block_row[1][0];
+          DC24 = DC25 = (int)next_next_block_row[1][0];
+        }
+        if (block_num + 1 < last_block_column) {
+          DC05 = (int)prev_prev_block_row[2][0];
+          DC10 = (int)prev_block_row[2][0];
+          DC15 = (int)buffer_ptr[2][0];
+          DC20 = (int)next_block_row[2][0];
+          DC25 = (int)next_next_block_row[2][0];
+        }
+        /* If DC interpolation is enabled, compute coefficient estimates using
+         * a Gaussian-like kernel, keeping the averages of the DC values.
+         *
+         * If DC interpolation is disabled, compute coefficient estimates using
+         * an algorithm similar to the one described in Section K.8 of the JPEG
+         * standard, except applied to a 5x5 window rather than a 3x3 window.
+         *
+         * An estimate is applied only if the coefficient is still zero and is
+         * not known to be fully accurate.
+         */
+        /* AC01 */
+        if ((Al = coef_bits[1]) != 0 && workspace[1] == 0) {
+          num = Q00 * (change_dc ?
+                (-DC01 - DC02 + DC04 + DC05 - 3 * DC06 + 13 * DC07 -
+                 13 * DC09 + 3 * DC10 - 3 * DC11 + 38 * DC12 - 38 * DC14 +
+                 3 * DC15 - 3 * DC16 + 13 * DC17 - 13 * DC19 + 3 * DC20 -
+                 DC21 - DC22 + DC24 + DC25) :
+                (-7 * DC11 + 50 * DC12 - 50 * DC14 + 7 * DC15));
+          if (num >= 0) {
+            pred = (int)(((Q01 << 7) + num) / (Q01 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+          } else {
+            pred = (int)(((Q01 << 7) - num) / (Q01 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+            pred = -pred;
+          }
+          workspace[1] = (JCOEF)pred;
+        }
+        /* AC10 */
+        if ((Al = coef_bits[2]) != 0 && workspace[8] == 0) {
+          num = Q00 * (change_dc ?
+                (-DC01 - 3 * DC02 - 3 * DC03 - 3 * DC04 - DC05 - DC06 +
+                 13 * DC07 + 38 * DC08 + 13 * DC09 - DC10 + DC16 -
+                 13 * DC17 - 38 * DC18 - 13 * DC19 + DC20 + DC21 +
+                 3 * DC22 + 3 * DC23 + 3 * DC24 + DC25) :
+                (-7 * DC03 + 50 * DC08 - 50 * DC18 + 7 * DC23));
+          if (num >= 0) {
+            pred = (int)(((Q10 << 7) + num) / (Q10 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+          } else {
+            pred = (int)(((Q10 << 7) - num) / (Q10 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+            pred = -pred;
+          }
+          workspace[8] = (JCOEF)pred;
+        }
+        /* AC20 */
+        if ((Al = coef_bits[3]) != 0 && workspace[16] == 0) {
+          num = Q00 * (change_dc ?
+                (DC03 + 2 * DC07 + 7 * DC08 + 2 * DC09 - 5 * DC12 - 14 * DC13 -
+                 5 * DC14 + 2 * DC17 + 7 * DC18 + 2 * DC19 + DC23) :
+                (-DC03 + 13 * DC08 - 24 * DC13 + 13 * DC18 - DC23));
+          if (num >= 0) {
+            pred = (int)(((Q20 << 7) + num) / (Q20 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+          } else {
+            pred = (int)(((Q20 << 7) - num) / (Q20 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+            pred = -pred;
+          }
+          workspace[16] = (JCOEF)pred;
+        }
+        /* AC11 */
+        if ((Al = coef_bits[4]) != 0 && workspace[9] == 0) {
+          num = Q00 * (change_dc ?
+                (-DC01 + DC05 + 9 * DC07 - 9 * DC09 - 9 * DC17 +
+                 9 * DC19 + DC21 - DC25) :
+                (DC10 + DC16 - 10 * DC17 + 10 * DC19 - DC02 - DC20 + DC22 -
+                 DC24 + DC04 - DC06 + 10 * DC07 - 10 * DC09));
+          if (num >= 0) {
+            pred = (int)(((Q11 << 7) + num) / (Q11 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+          } else {
+            pred = (int)(((Q11 << 7) - num) / (Q11 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+            pred = -pred;
+          }
+          workspace[9] = (JCOEF)pred;
+        }
+        /* AC02 */
+        if ((Al = coef_bits[5]) != 0 && workspace[2] == 0) {
+          num = Q00 * (change_dc ?
+                (2 * DC07 - 5 * DC08 + 2 * DC09 + DC11 + 7 * DC12 - 14 * DC13 +
+                 7 * DC14 + DC15 + 2 * DC17 - 5 * DC18 + 2 * DC19) :
+                (-DC11 + 13 * DC12 - 24 * DC13 + 13 * DC14 - DC15));
+          if (num >= 0) {
+            pred = (int)(((Q02 << 7) + num) / (Q02 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+          } else {
+            pred = (int)(((Q02 << 7) - num) / (Q02 << 8));
+            if (Al > 0 && pred >= (1 << Al))
+              pred = (1 << Al) - 1;
+            pred = -pred;
+          }
+          workspace[2] = (JCOEF)pred;
+        }
+        if (change_dc) {
+          /* AC03 */
+          if ((Al = coef_bits[6]) != 0 && workspace[3] == 0) {
+            num = Q00 * (DC07 - DC09 + 2 * DC12 - 2 * DC14 + DC17 - DC19);
+            if (num >= 0) {
+              pred = (int)(((Q03 << 7) + num) / (Q03 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+            } else {
+              pred = (int)(((Q03 << 7) - num) / (Q03 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+              pred = -pred;
+            }
+            workspace[3] = (JCOEF)pred;
+          }
+          /* AC12 */
+          if ((Al = coef_bits[7]) != 0 && workspace[10] == 0) {
+            num = Q00 * (DC07 - 3 * DC08 + DC09 - DC17 + 3 * DC18 - DC19);
+            if (num >= 0) {
+              pred = (int)(((Q12 << 7) + num) / (Q12 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+            } else {
+              pred = (int)(((Q12 << 7) - num) / (Q12 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+              pred = -pred;
+            }
+            workspace[10] = (JCOEF)pred;
+          }
+          /* AC21 */
+          if ((Al = coef_bits[8]) != 0 && workspace[17] == 0) {
+            num = Q00 * (DC07 - DC09 - 3 * DC12 + 3 * DC14 + DC17 - DC19);
+            if (num >= 0) {
+              pred = (int)(((Q21 << 7) + num) / (Q21 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+            } else {
+              pred = (int)(((Q21 << 7) - num) / (Q21 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+              pred = -pred;
+            }
+            workspace[17] = (JCOEF)pred;
+          }
+          /* AC30 */
+          if ((Al = coef_bits[9]) != 0 && workspace[24] == 0) {
+            num = Q00 * (DC07 + 2 * DC08 + DC09 - DC17 - 2 * DC18 - DC19);
+            if (num >= 0) {
+              pred = (int)(((Q30 << 7) + num) / (Q30 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+            } else {
+              pred = (int)(((Q30 << 7) - num) / (Q30 << 8));
+              if (Al > 0 && pred >= (1 << Al))
+                pred = (1 << Al) - 1;
+              pred = -pred;
+            }
+            workspace[24] = (JCOEF)pred;
+          }
+          /* coef_bits[0] is non-negative.  Otherwise this function would not
+           * be called.
+           */
+          num = Q00 *
+                (-2 * DC01 - 6 * DC02 - 8 * DC03 - 6 * DC04 - 2 * DC05 -
+                 6 * DC06 + 6 * DC07 + 42 * DC08 + 6 * DC09 - 6 * DC10 -
+                 8 * DC11 + 42 * DC12 + 152 * DC13 + 42 * DC14 - 8 * DC15 -
+                 6 * DC16 + 6 * DC17 + 42 * DC18 + 6 * DC19 - 6 * DC20 -
+                 2 * DC21 - 6 * DC22 - 8 * DC23 - 6 * DC24 - 2 * DC25);
+          if (num >= 0) {
+            pred = (int)(((Q00 << 7) + num) / (Q00 << 8));
+          } else {
+            pred = (int)(((Q00 << 7) - num) / (Q00 << 8));
+            pred = -pred;
+          }
+          workspace[0] = (JCOEF)pred;
+        }  /* change_dc */
+
+        /* OK, do the IDCT */
+        (*inverse_DCT) (cinfo, compptr, (JCOEFPTR)workspace, output_ptr,
+                        output_col);
+        /* Advance for next column */
+        DC01 = DC02;  DC02 = DC03;  DC03 = DC04;  DC04 = DC05;
+        DC06 = DC07;  DC07 = DC08;  DC08 = DC09;  DC09 = DC10;
+        DC11 = DC12;  DC12 = DC13;  DC13 = DC14;  DC14 = DC15;
+        DC16 = DC17;  DC17 = DC18;  DC18 = DC19;  DC19 = DC20;
+        DC21 = DC22;  DC22 = DC23;  DC23 = DC24;  DC24 = DC25;
+        buffer_ptr++, prev_block_row++, next_block_row++,
+          prev_prev_block_row++, next_next_block_row++;
+        output_col += compptr->_DCT_scaled_size;
+      }
+      output_ptr += compptr->_DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* BLOCK_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+_jinit_d_coef_controller(j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_coef_controller));
+  cinfo->coef = (struct jpeg_d_coef_controller *)coef;
+  coef->pub.start_input_pass = start_input_pass;
+  coef->pub.start_output_pass = start_output_pass;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  coef->coef_bits_latch = NULL;
+#endif
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    /* Note we ask for a pre-zeroed array. */
+    int ci, access_rows;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      access_rows = compptr->v_samp_factor;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+      /* If block smoothing could be used, need a bigger window */
+      if (cinfo->progressive_mode)
+        access_rows *= 5;
+#endif
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE, TRUE,
+         (JDIMENSION)jround_up((long)compptr->width_in_blocks,
+                               (long)compptr->h_samp_factor),
+         (JDIMENSION)jround_up((long)compptr->height_in_blocks,
+                               (long)compptr->v_samp_factor),
+         (JDIMENSION)access_rows);
+    }
+    coef->pub.consume_data = consume_data;
+    coef->pub._decompress_data = decompress_data;
+    coef->pub.coef_arrays = coef->whole_image; /* link to virtual arrays */
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  D_MAX_BLOCKS_IN_MCU * sizeof(JBLOCK));
+    for (i = 0; i < D_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->pub.consume_data = dummy_consume_data;
+    coef->pub._decompress_data = decompress_onepass;
+    coef->pub.coef_arrays = NULL; /* flag for no virtual arrays */
+  }
+
+  /* Allocate the workspace buffer */
+  coef->workspace = (JCOEF *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(JCOEF) * DCTSIZE2);
+}

thirdparty/libjpeg-turbo/src/jdcoefct.h

@@ -0,0 +1,88 @@
+/*
+ * jdcoefct.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2020, Google, Inc.
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/* Block smoothing is only applicable for progressive JPEG, so: */
+#ifndef D_PROGRESSIVE_SUPPORTED
+#undef BLOCK_SMOOTHING_SUPPORTED
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_coef_controller pub; /* public fields */
+
+  /* These variables keep track of the current location of the input side. */
+  /* cinfo->input_iMCU_row is also used for this. */
+  JDIMENSION MCU_ctr;           /* counts MCUs processed in current row */
+  int MCU_vert_offset;          /* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;    /* number of such rows needed */
+
+  /* The output side's location is represented by cinfo->output_iMCU_row. */
+
+  /* In single-pass modes, it's sufficient to buffer just one MCU.
+   * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
+   * and let the entropy decoder write into that workspace each time.
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays; it is used only by the input side.
+   */
+  JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
+
+  /* Temporary workspace for one MCU */
+  JCOEF *workspace;
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+#endif
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  /* When doing block smoothing, we latch coefficient Al values here */
+  int *coef_bits_latch;
+#define SAVED_COEFS  10         /* we save coef_bits[0..9] */
+#endif
+} my_coef_controller;
+
+typedef my_coef_controller *my_coef_ptr;
+
+
+LOCAL(void)
+start_iMCU_row(j_decompress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row (input side) */
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows - 1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->MCU_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdcol565.c

@@ -0,0 +1,392 @@
+/*
+ * jdcol565.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modifications:
+ * Copyright (C) 2013, Linaro Limited.
+ * Copyright (C) 2014-2015, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+/* This file is included by jdcolor.c */
+
+
+INLINE
+LOCAL(void)
+ycc_rgb565_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                            JDIMENSION input_row, _JSAMPARRAY output_buf,
+                            int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int y, cb, cr;
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  register int *Crrtab = cconvert->Cr_r_tab;
+  register int *Cbbtab = cconvert->Cb_b_tab;
+  register JLONG *Crgtab = cconvert->Cr_g_tab;
+  register JLONG *Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int r, g, b;
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[y + Crrtab[cr]];
+      g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                            SCALEBITS))];
+      b = range_limit[y + Cbbtab[cb]];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[y + Crrtab[cr]];
+      g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                            SCALEBITS))];
+      b = range_limit[y + Cbbtab[cb]];
+      rgb = PACK_SHORT_565(r, g, b);
+
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[y + Crrtab[cr]];
+      g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                            SCALEBITS))];
+      b = range_limit[y + Cbbtab[cb]];
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      y  = *inptr0;
+      cb = *inptr1;
+      cr = *inptr2;
+      r = range_limit[y + Crrtab[cr]];
+      g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                            SCALEBITS))];
+      b = range_limit[y + Cbbtab[cb]];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+INLINE
+LOCAL(void)
+ycc_rgb565D_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                             JDIMENSION input_row, _JSAMPARRAY output_buf,
+                             int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int y, cb, cr;
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  register int *Crrtab = cconvert->Cr_r_tab;
+  register int *Cbbtab = cconvert->Cb_b_tab;
+  register JLONG *Crgtab = cconvert->Cr_g_tab;
+  register JLONG *Cbgtab = cconvert->Cb_g_tab;
+  JLONG d0 = dither_matrix[cinfo->output_scanline & DITHER_MASK];
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int r, g, b;
+
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
+      g = range_limit[DITHER_565_G(y +
+                                   ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                     SCALEBITS)), d0)];
+      b = range_limit[DITHER_565_B(y + Cbbtab[cb], d0)];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
+      g = range_limit[DITHER_565_G(y +
+                                   ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                     SCALEBITS)), d0)];
+      b = range_limit[DITHER_565_B(y + Cbbtab[cb], d0)];
+      d0 = DITHER_ROTATE(d0);
+      rgb = PACK_SHORT_565(r, g, b);
+
+      y  = *inptr0++;
+      cb = *inptr1++;
+      cr = *inptr2++;
+      r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
+      g = range_limit[DITHER_565_G(y +
+                                   ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                     SCALEBITS)), d0)];
+      b = range_limit[DITHER_565_B(y + Cbbtab[cb], d0)];
+      d0 = DITHER_ROTATE(d0);
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      y  = *inptr0;
+      cb = *inptr1;
+      cr = *inptr2;
+      r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
+      g = range_limit[DITHER_565_G(y +
+                                   ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                     SCALEBITS)), d0)];
+      b = range_limit[DITHER_565_B(y + Cbbtab[cb], d0)];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+INLINE
+LOCAL(void)
+rgb_rgb565_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                            JDIMENSION input_row, _JSAMPARRAY output_buf,
+                            int num_rows)
+{
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int r, g, b;
+
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      r = *inptr0++;
+      g = *inptr1++;
+      b = *inptr2++;
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      r = *inptr0++;
+      g = *inptr1++;
+      b = *inptr2++;
+      rgb = PACK_SHORT_565(r, g, b);
+
+      r = *inptr0++;
+      g = *inptr1++;
+      b = *inptr2++;
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      r = *inptr0;
+      g = *inptr1;
+      b = *inptr2;
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+}
+
+
+INLINE
+LOCAL(void)
+rgb_rgb565D_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                             JDIMENSION input_row, _JSAMPARRAY output_buf,
+                             int num_rows)
+{
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  JDIMENSION num_cols = cinfo->output_width;
+  JLONG d0 = dither_matrix[cinfo->output_scanline & DITHER_MASK];
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int r, g, b;
+
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      r = range_limit[DITHER_565_R(*inptr0++, d0)];
+      g = range_limit[DITHER_565_G(*inptr1++, d0)];
+      b = range_limit[DITHER_565_B(*inptr2++, d0)];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      r = range_limit[DITHER_565_R(*inptr0++, d0)];
+      g = range_limit[DITHER_565_G(*inptr1++, d0)];
+      b = range_limit[DITHER_565_B(*inptr2++, d0)];
+      d0 = DITHER_ROTATE(d0);
+      rgb = PACK_SHORT_565(r, g, b);
+
+      r = range_limit[DITHER_565_R(*inptr0++, d0)];
+      g = range_limit[DITHER_565_G(*inptr1++, d0)];
+      b = range_limit[DITHER_565_B(*inptr2++, d0)];
+      d0 = DITHER_ROTATE(d0);
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      r = range_limit[DITHER_565_R(*inptr0, d0)];
+      g = range_limit[DITHER_565_G(*inptr1, d0)];
+      b = range_limit[DITHER_565_B(*inptr2, d0)];
+      rgb = PACK_SHORT_565(r, g, b);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+}
+
+
+INLINE
+LOCAL(void)
+gray_rgb565_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                             JDIMENSION input_row, _JSAMPARRAY output_buf,
+                             int num_rows)
+{
+  register _JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int g;
+
+    inptr = input_buf[0][input_row++];
+    outptr = *output_buf++;
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      g = *inptr++;
+      rgb = PACK_SHORT_565(g, g, g);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      g = *inptr++;
+      rgb = PACK_SHORT_565(g, g, g);
+      g = *inptr++;
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(g, g, g));
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      g = *inptr;
+      rgb = PACK_SHORT_565(g, g, g);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+}
+
+
+INLINE
+LOCAL(void)
+gray_rgb565D_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                              JDIMENSION input_row, _JSAMPARRAY output_buf,
+                              int num_rows)
+{
+  register _JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  JDIMENSION num_cols = cinfo->output_width;
+  JLONG d0 = dither_matrix[cinfo->output_scanline & DITHER_MASK];
+
+  while (--num_rows >= 0) {
+    JLONG rgb;
+    unsigned int g;
+
+    inptr = input_buf[0][input_row++];
+    outptr = *output_buf++;
+    if (PACK_NEED_ALIGNMENT(outptr)) {
+      g = *inptr++;
+      g = range_limit[DITHER_565_R(g, d0)];
+      rgb = PACK_SHORT_565(g, g, g);
+      *(INT16 *)outptr = (INT16)rgb;
+      outptr += 2;
+      num_cols--;
+    }
+    for (col = 0; col < (num_cols >> 1); col++) {
+      g = *inptr++;
+      g = range_limit[DITHER_565_R(g, d0)];
+      rgb = PACK_SHORT_565(g, g, g);
+      d0 = DITHER_ROTATE(d0);
+
+      g = *inptr++;
+      g = range_limit[DITHER_565_R(g, d0)];
+      rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(g, g, g));
+      d0 = DITHER_ROTATE(d0);
+
+      WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
+      outptr += 4;
+    }
+    if (num_cols & 1) {
+      g = *inptr;
+      g = range_limit[DITHER_565_R(g, d0)];
+      rgb = PACK_SHORT_565(g, g, g);
+      *(INT16 *)outptr = (INT16)rgb;
+    }
+  }
+}

thirdparty/libjpeg-turbo/src/jdcolext.c

@@ -0,0 +1,145 @@
+/*
+ * jdcolext.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009, 2011, 2015, 2022-2023, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+
+/* This file is included by jdcolor.c */
+
+
+/*
+ * Convert some rows of samples to the output colorspace.
+ *
+ * Note that we change from noninterleaved, one-plane-per-component format
+ * to interleaved-pixel format.  The output buffer is therefore three times
+ * as wide as the input buffer.
+ * A starting row offset is provided only for the input buffer.  The caller
+ * can easily adjust the passed output_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+INLINE
+LOCAL(void)
+ycc_rgb_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                         JDIMENSION input_row, _JSAMPARRAY output_buf,
+                         int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int y, cb, cr;
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  register int *Crrtab = cconvert->Cr_r_tab;
+  register int *Cbbtab = cconvert->Cb_b_tab;
+  register JLONG *Crgtab = cconvert->Cr_g_tab;
+  register JLONG *Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = inptr0[col];
+      cb = inptr1[col];
+      cr = inptr2[col];
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[RGB_RED] =   range_limit[y + Crrtab[cr]];
+      outptr[RGB_GREEN] = range_limit[y +
+                              ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                SCALEBITS))];
+      outptr[RGB_BLUE] =  range_limit[y + Cbbtab[cb]];
+      /* Set unused byte to _MAXJSAMPLE so it can be interpreted as an */
+      /* opaque alpha channel value */
+#ifdef RGB_ALPHA
+      outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert grayscale to RGB: just duplicate the graylevel three times.
+ * This is provided to support applications that don't want to cope
+ * with grayscale as a separate case.
+ */
+
+INLINE
+LOCAL(void)
+gray_rgb_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                          JDIMENSION input_row, _JSAMPARRAY output_buf,
+                          int num_rows)
+{
+  register _JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    inptr = input_buf[0][input_row++];
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[RGB_RED] = outptr[RGB_GREEN] = outptr[RGB_BLUE] = inptr[col];
+      /* Set unused byte to _MAXJSAMPLE so it can be interpreted as an */
+      /* opaque alpha channel value */
+#ifdef RGB_ALPHA
+      outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}
+
+
+/*
+ * Convert RGB to extended RGB: just swap the order of source pixels
+ */
+
+INLINE
+LOCAL(void)
+rgb_rgb_convert_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                         JDIMENSION input_row, _JSAMPARRAY output_buf,
+                         int num_rows)
+{
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register _JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[RGB_RED] = inptr0[col];
+      outptr[RGB_GREEN] = inptr1[col];
+      outptr[RGB_BLUE] = inptr2[col];
+      /* Set unused byte to _MAXJSAMPLE so it can be interpreted as an */
+      /* opaque alpha channel value */
+#ifdef RGB_ALPHA
+      outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}

thirdparty/libjpeg-turbo/src/jdcolor.c

@@ -0,0 +1,946 @@
+/*
+ * jdcolor.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2011 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2009, 2011-2012, 2014-2015, 2022, 2024, D. R. Commander.
+ * Copyright (C) 2013, Linaro Limited.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_deconverter pub; /* public fields */
+
+#if BITS_IN_JSAMPLE != 16
+  /* Private state for YCC->RGB conversion */
+  int *Cr_r_tab;                /* => table for Cr to R conversion */
+  int *Cb_b_tab;                /* => table for Cb to B conversion */
+  JLONG *Cr_g_tab;              /* => table for Cr to G conversion */
+  JLONG *Cb_g_tab;              /* => table for Cb to G conversion */
+
+  /* Private state for RGB->Y conversion */
+  JLONG *rgb_y_tab;             /* => table for RGB to Y conversion */
+#endif
+} my_color_deconverter;
+
+typedef my_color_deconverter *my_cconvert_ptr;
+
+
+/**************** YCbCr -> RGB conversion: most common case **************/
+/****************   RGB -> Y   conversion: less common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0.._MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *
+ *      R = Y                + 1.40200 * Cr
+ *      G = Y - 0.34414 * Cb - 0.71414 * Cr
+ *      B = Y + 1.77200 * Cb
+ *
+ *      Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
+ *
+ * where Cb and Cr represent the incoming values less _CENTERJSAMPLE.
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ * Notice that Y, being an integral input, does not contribute any fraction
+ * so it need not participate in the rounding.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times Cb and Cr for all possible values.
+ * For 8-bit samples this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The Cr=>R and Cb=>B values can be rounded to integers in advance; the
+ * values for the G calculation are left scaled up, since we must add them
+ * together before rounding.
+ */
+
+#define SCALEBITS       16      /* speediest right-shift on some machines */
+#define ONE_HALF        ((JLONG)1 << (SCALEBITS - 1))
+#define FIX(x)          ((JLONG)((x) * (1L << SCALEBITS) + 0.5))
+
+/* We allocate one big table for RGB->Y conversion and divide it up into
+ * three parts, instead of doing three alloc_small requests.  This lets us
+ * use a single table base address, which can be held in a register in the
+ * inner loops on many machines (more than can hold all three addresses,
+ * anyway).
+ */
+
+#define R_Y_OFF         0                       /* offset to R => Y section */
+#define G_Y_OFF         (1 * (_MAXJSAMPLE + 1)) /* offset to G => Y section */
+#define B_Y_OFF         (2 * (_MAXJSAMPLE + 1)) /* etc. */
+#define TABLE_SIZE      (3 * (_MAXJSAMPLE + 1))
+
+
+/* Include inline routines for colorspace extensions */
+
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extrgb_convert_internal
+#define gray_rgb_convert_internal  gray_extrgb_convert_internal
+#define rgb_rgb_convert_internal  rgb_extrgb_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extrgbx_convert_internal
+#define gray_rgb_convert_internal  gray_extrgbx_convert_internal
+#define rgb_rgb_convert_internal  rgb_extrgbx_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extbgr_convert_internal
+#define gray_rgb_convert_internal  gray_extbgr_convert_internal
+#define rgb_rgb_convert_internal  rgb_extbgr_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extbgrx_convert_internal
+#define gray_rgb_convert_internal  gray_extbgrx_convert_internal
+#define rgb_rgb_convert_internal  rgb_extbgrx_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extxbgr_convert_internal
+#define gray_rgb_convert_internal  gray_extxbgr_convert_internal
+#define rgb_rgb_convert_internal  rgb_extxbgr_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define ycc_rgb_convert_internal  ycc_extxrgb_convert_internal
+#define gray_rgb_convert_internal  gray_extxrgb_convert_internal
+#define rgb_rgb_convert_internal  rgb_extxrgb_convert_internal
+#include "jdcolext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef ycc_rgb_convert_internal
+#undef gray_rgb_convert_internal
+#undef rgb_rgb_convert_internal
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table(j_decompress_ptr cinfo)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  int i;
+  JLONG x;
+  SHIFT_TEMPS
+
+  cconvert->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(int));
+  cconvert->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(int));
+  cconvert->Cr_g_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(JLONG));
+  cconvert->Cb_g_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(JLONG));
+
+  for (i = 0, x = -_CENTERJSAMPLE; i <= _MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0.._MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - _CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    cconvert->Cr_r_tab[i] = (int)
+                    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    cconvert->Cb_b_tab[i] = (int)
+                    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    cconvert->Cr_g_tab[i] = (-FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    cconvert->Cb_g_tab[i] = (-FIX(0.34414)) * x + ONE_HALF;
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert some rows of samples to the output colorspace.
+ */
+
+METHODDEF(void)
+ycc_rgb_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    ycc_extrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    ycc_extrgbx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGR:
+    ycc_extbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    ycc_extbgrx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    ycc_extxbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    ycc_extxrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  default:
+    ycc_rgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                             num_rows);
+    break;
+  }
+}
+
+
+/**************** Cases other than YCbCr -> RGB **************/
+
+
+/*
+ * Initialize for RGB->grayscale colorspace conversion.
+ */
+
+LOCAL(void)
+build_rgb_y_table(j_decompress_ptr cinfo)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  JLONG *rgb_y_tab;
+  JLONG i;
+
+  /* Allocate and fill in the conversion tables. */
+  cconvert->rgb_y_tab = rgb_y_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (TABLE_SIZE * sizeof(JLONG)));
+
+  for (i = 0; i <= _MAXJSAMPLE; i++) {
+    rgb_y_tab[i + R_Y_OFF] = FIX(0.29900) * i;
+    rgb_y_tab[i + G_Y_OFF] = FIX(0.58700) * i;
+    rgb_y_tab[i + B_Y_OFF] = FIX(0.11400) * i + ONE_HALF;
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Convert RGB to grayscale.
+ */
+
+METHODDEF(void)
+rgb_gray_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                 JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int r, g, b;
+  register JLONG *ctab = cconvert->rgb_y_tab;
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      r = inptr0[col];
+      g = inptr1[col];
+      b = inptr2[col];
+      /* Y */
+      outptr[col] = (_JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
+                                ctab[b + B_Y_OFF]) >> SCALEBITS);
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * Color conversion for no colorspace change: just copy the data,
+ * converting from separate-planes to interleaved representation.
+ */
+
+METHODDEF(void)
+null_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+             JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  register _JSAMPROW inptr, inptr0, inptr1, inptr2, inptr3, outptr;
+  register JDIMENSION col;
+  register int num_components = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->output_width;
+  int ci;
+
+  if (num_components == 3) {
+    while (--num_rows >= 0) {
+      inptr0 = input_buf[0][input_row];
+      inptr1 = input_buf[1][input_row];
+      inptr2 = input_buf[2][input_row];
+      input_row++;
+      outptr = *output_buf++;
+      for (col = 0; col < num_cols; col++) {
+        *outptr++ = inptr0[col];
+        *outptr++ = inptr1[col];
+        *outptr++ = inptr2[col];
+      }
+    }
+  } else if (num_components == 4) {
+    while (--num_rows >= 0) {
+      inptr0 = input_buf[0][input_row];
+      inptr1 = input_buf[1][input_row];
+      inptr2 = input_buf[2][input_row];
+      inptr3 = input_buf[3][input_row];
+      input_row++;
+      outptr = *output_buf++;
+      for (col = 0; col < num_cols; col++) {
+        *outptr++ = inptr0[col];
+        *outptr++ = inptr1[col];
+        *outptr++ = inptr2[col];
+        *outptr++ = inptr3[col];
+      }
+    }
+  } else {
+    while (--num_rows >= 0) {
+      for (ci = 0; ci < num_components; ci++) {
+        inptr = input_buf[ci][input_row];
+        outptr = *output_buf;
+        for (col = 0; col < num_cols; col++) {
+          outptr[ci] = inptr[col];
+          outptr += num_components;
+        }
+      }
+      output_buf++;
+      input_row++;
+    }
+  }
+}
+
+
+/*
+ * Color conversion for grayscale: just copy the data.
+ * This also works for YCbCr -> grayscale conversion, in which
+ * we just copy the Y (luminance) component and ignore chrominance.
+ */
+
+METHODDEF(void)
+grayscale_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                  JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  _jcopy_sample_rows(input_buf[0], (int)input_row, output_buf, 0, num_rows,
+                     cinfo->output_width);
+}
+
+
+/*
+ * Convert grayscale to RGB
+ */
+
+METHODDEF(void)
+gray_rgb_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                 JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    gray_extrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    gray_extrgbx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                  num_rows);
+    break;
+  case JCS_EXT_BGR:
+    gray_extbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    gray_extbgrx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                  num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    gray_extxbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                  num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    gray_extxrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                  num_rows);
+    break;
+  default:
+    gray_rgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                              num_rows);
+    break;
+  }
+}
+
+
+/*
+ * Convert plain RGB to extended RGB
+ */
+
+METHODDEF(void)
+rgb_rgb_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    rgb_extrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                num_rows);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    rgb_extrgbx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_BGR:
+    rgb_extbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                num_rows);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    rgb_extbgrx_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    rgb_extxbgr_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    rgb_extxrgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                                 num_rows);
+    break;
+  default:
+    rgb_rgb_convert_internal(cinfo, input_buf, input_row, output_buf,
+                             num_rows);
+    break;
+  }
+}
+
+
+/*
+ * Adobe-style YCCK->CMYK conversion.
+ * We convert YCbCr to R=1-C, G=1-M, and B=1-Y using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume build_ycc_rgb_table has been called.
+ */
+
+METHODDEF(void)
+ycck_cmyk_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                  JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+#if BITS_IN_JSAMPLE != 16
+  my_cconvert_ptr cconvert = (my_cconvert_ptr)cinfo->cconvert;
+  register int y, cb, cr;
+  register _JSAMPROW outptr;
+  register _JSAMPROW inptr0, inptr1, inptr2, inptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  register int *Crrtab = cconvert->Cr_r_tab;
+  register int *Cbbtab = cconvert->Cb_b_tab;
+  register JLONG *Crgtab = cconvert->Cr_g_tab;
+  register JLONG *Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    inptr3 = input_buf[3][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = inptr0[col];
+      cb = inptr1[col];
+      cr = inptr2[col];
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[0] = range_limit[_MAXJSAMPLE - (y + Crrtab[cr])];  /* red */
+      outptr[1] = range_limit[_MAXJSAMPLE - (y +                /* green */
+                              ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+                                                 SCALEBITS)))];
+      outptr[2] = range_limit[_MAXJSAMPLE - (y + Cbbtab[cb])];  /* blue */
+      /* K passes through unchanged */
+      outptr[3] = inptr3[col];
+      outptr += 4;
+    }
+  }
+#else
+  ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+}
+
+
+/*
+ * RGB565 conversion
+ */
+
+#define PACK_SHORT_565_LE(r, g, b) \
+  ((((r) << 8) & 0xF800) | (((g) << 3) & 0x7E0) | ((b) >> 3))
+#define PACK_SHORT_565_BE(r, g, b) \
+  (((r) & 0xF8) | ((g) >> 5) | (((g) << 11) & 0xE000) | (((b) << 5) & 0x1F00))
+
+#define PACK_TWO_PIXELS_LE(l, r)    ((r << 16) | l)
+#define PACK_TWO_PIXELS_BE(l, r)    ((l << 16) | r)
+
+#define PACK_NEED_ALIGNMENT(ptr)    (((size_t)(ptr)) & 3)
+
+#define WRITE_TWO_ALIGNED_PIXELS(addr, pixels)  ((*(int *)(addr)) = pixels)
+
+#define DITHER_565_R(r, dither)  ((r) + ((dither) & 0xFF))
+#define DITHER_565_G(g, dither)  ((g) + (((dither) & 0xFF) >> 1))
+#define DITHER_565_B(b, dither)  ((b) + ((dither) & 0xFF))
+
+
+/* Declarations for ordered dithering
+ *
+ * We use a 4x4 ordered dither array packed into 32 bits.  This array is
+ * sufficient for dithering RGB888 to RGB565.
+ */
+
+#define DITHER_MASK       0x3
+#define DITHER_ROTATE(x)  ((((x) & 0xFF) << 24) | (((x) >> 8) & 0x00FFFFFF))
+static const JLONG dither_matrix[4] = {
+  0x0008020A,
+  0x0C040E06,
+  0x030B0109,
+  0x0F070D05
+};
+
+
+static INLINE boolean is_big_endian(void)
+{
+  int test_value = 1;
+  if (*(char *)&test_value != 1)
+    return TRUE;
+  return FALSE;
+}
+
+
+/* Include inline routines for RGB565 conversion */
+
+#define PACK_SHORT_565  PACK_SHORT_565_LE
+#define PACK_TWO_PIXELS  PACK_TWO_PIXELS_LE
+#define ycc_rgb565_convert_internal  ycc_rgb565_convert_le
+#define ycc_rgb565D_convert_internal  ycc_rgb565D_convert_le
+#define rgb_rgb565_convert_internal  rgb_rgb565_convert_le
+#define rgb_rgb565D_convert_internal  rgb_rgb565D_convert_le
+#define gray_rgb565_convert_internal  gray_rgb565_convert_le
+#define gray_rgb565D_convert_internal  gray_rgb565D_convert_le
+#include "jdcol565.c"
+#undef PACK_SHORT_565
+#undef PACK_TWO_PIXELS
+#undef ycc_rgb565_convert_internal
+#undef ycc_rgb565D_convert_internal
+#undef rgb_rgb565_convert_internal
+#undef rgb_rgb565D_convert_internal
+#undef gray_rgb565_convert_internal
+#undef gray_rgb565D_convert_internal
+
+#define PACK_SHORT_565  PACK_SHORT_565_BE
+#define PACK_TWO_PIXELS  PACK_TWO_PIXELS_BE
+#define ycc_rgb565_convert_internal  ycc_rgb565_convert_be
+#define ycc_rgb565D_convert_internal  ycc_rgb565D_convert_be
+#define rgb_rgb565_convert_internal  rgb_rgb565_convert_be
+#define rgb_rgb565D_convert_internal  rgb_rgb565D_convert_be
+#define gray_rgb565_convert_internal  gray_rgb565_convert_be
+#define gray_rgb565D_convert_internal  gray_rgb565D_convert_be
+#include "jdcol565.c"
+#undef PACK_SHORT_565
+#undef PACK_TWO_PIXELS
+#undef ycc_rgb565_convert_internal
+#undef ycc_rgb565D_convert_internal
+#undef rgb_rgb565_convert_internal
+#undef rgb_rgb565D_convert_internal
+#undef gray_rgb565_convert_internal
+#undef gray_rgb565D_convert_internal
+
+
+METHODDEF(void)
+ycc_rgb565_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    ycc_rgb565_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    ycc_rgb565_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+METHODDEF(void)
+ycc_rgb565D_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                    JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    ycc_rgb565D_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    ycc_rgb565D_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+METHODDEF(void)
+rgb_rgb565_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    rgb_rgb565_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    rgb_rgb565_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+METHODDEF(void)
+rgb_rgb565D_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                    JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    rgb_rgb565D_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    rgb_rgb565D_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+METHODDEF(void)
+gray_rgb565_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                    JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    gray_rgb565_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    gray_rgb565_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+METHODDEF(void)
+gray_rgb565D_convert(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                     JDIMENSION input_row, _JSAMPARRAY output_buf, int num_rows)
+{
+  if (is_big_endian())
+    gray_rgb565D_convert_be(cinfo, input_buf, input_row, output_buf, num_rows);
+  else
+    gray_rgb565D_convert_le(cinfo, input_buf, input_row, output_buf, num_rows);
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+start_pass_dcolor(j_decompress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for output colorspace conversion.
+ */
+
+GLOBAL(void)
+_jinit_color_deconverter(j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+  int ci;
+
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_color_deconverter));
+  cinfo->cconvert = (struct jpeg_color_deconverter *)cconvert;
+  cconvert->pub.start_pass = start_pass_dcolor;
+
+  /* Make sure num_components agrees with jpeg_color_space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  default:                      /* JCS_UNKNOWN can be anything */
+    if (cinfo->num_components < 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+  }
+
+  /* Set out_color_components and conversion method based on requested space.
+   * Also clear the component_needed flags for any unused components,
+   * so that earlier pipeline stages can avoid useless computation.
+   * NOTE: We do not allow any lossy color conversion algorithms in lossless
+   * mode.
+   */
+
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+#ifdef D_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->jpeg_color_space != cinfo->out_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    cinfo->out_color_components = 1;
+    if (cinfo->jpeg_color_space == JCS_GRAYSCALE ||
+        cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub._color_convert = grayscale_convert;
+      /* For color->grayscale conversion, only the Y (0) component is needed */
+      for (ci = 1; ci < cinfo->num_components; ci++)
+        cinfo->comp_info[ci].component_needed = FALSE;
+    } else if (cinfo->jpeg_color_space == JCS_RGB) {
+      cconvert->pub._color_convert = rgb_gray_convert;
+      build_rgb_y_table(cinfo);
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+  case JCS_EXT_RGBA:
+  case JCS_EXT_BGRA:
+  case JCS_EXT_ABGR:
+  case JCS_EXT_ARGB:
+#ifdef D_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless && cinfo->jpeg_color_space != JCS_RGB)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    cinfo->out_color_components = rgb_pixelsize[cinfo->out_color_space];
+    if (cinfo->jpeg_color_space == JCS_YCbCr) {
+#ifdef WITH_SIMD
+      if (jsimd_can_ycc_rgb())
+        cconvert->pub._color_convert = jsimd_ycc_rgb_convert;
+      else
+#endif
+      {
+        cconvert->pub._color_convert = ycc_rgb_convert;
+        build_ycc_rgb_table(cinfo);
+      }
+    } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+      cconvert->pub._color_convert = gray_rgb_convert;
+    } else if (cinfo->jpeg_color_space == JCS_RGB) {
+      if (rgb_red[cinfo->out_color_space] == 0 &&
+          rgb_green[cinfo->out_color_space] == 1 &&
+          rgb_blue[cinfo->out_color_space] == 2 &&
+          rgb_pixelsize[cinfo->out_color_space] == 3)
+        cconvert->pub._color_convert = null_convert;
+      else
+        cconvert->pub._color_convert = rgb_rgb_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB565:
+#ifdef D_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    cinfo->out_color_components = 3;
+    if (cinfo->dither_mode == JDITHER_NONE) {
+      if (cinfo->jpeg_color_space == JCS_YCbCr) {
+#ifdef WITH_SIMD
+        if (jsimd_can_ycc_rgb565())
+          cconvert->pub._color_convert = jsimd_ycc_rgb565_convert;
+        else
+#endif
+        {
+          cconvert->pub._color_convert = ycc_rgb565_convert;
+          build_ycc_rgb_table(cinfo);
+        }
+      } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+        cconvert->pub._color_convert = gray_rgb565_convert;
+      } else if (cinfo->jpeg_color_space == JCS_RGB) {
+        cconvert->pub._color_convert = rgb_rgb565_convert;
+      } else
+        ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    } else {
+      /* only ordered dithering is supported */
+      if (cinfo->jpeg_color_space == JCS_YCbCr) {
+        cconvert->pub._color_convert = ycc_rgb565D_convert;
+        build_ycc_rgb_table(cinfo);
+      } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+        cconvert->pub._color_convert = gray_rgb565D_convert;
+      } else if (cinfo->jpeg_color_space == JCS_RGB) {
+        cconvert->pub._color_convert = rgb_rgb565D_convert;
+      } else
+        ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    }
+    break;
+
+  case JCS_CMYK:
+#ifdef D_LOSSLESS_SUPPORTED
+    if (cinfo->master->lossless &&
+        cinfo->jpeg_color_space != cinfo->out_color_space)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+#endif
+    cinfo->out_color_components = 4;
+    if (cinfo->jpeg_color_space == JCS_YCCK) {
+      cconvert->pub._color_convert = ycck_cmyk_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_CMYK) {
+      cconvert->pub._color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:
+    /* Permit null conversion to same output space */
+    if (cinfo->out_color_space == cinfo->jpeg_color_space) {
+      cinfo->out_color_components = cinfo->num_components;
+      cconvert->pub._color_convert = null_convert;
+    } else                      /* unsupported non-null conversion */
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+  }
+
+  if (cinfo->quantize_colors)
+    cinfo->output_components = 1; /* single colormapped output component */
+  else
+    cinfo->output_components = cinfo->out_color_components;
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdct.h

@@ -0,0 +1,221 @@
+/*
+ * jdct.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This include file contains common declarations for the forward and
+ * inverse DCT modules.  These declarations are private to the DCT managers
+ * (jcdctmgr.c, jddctmgr.c) and the individual DCT algorithms.
+ * The individual DCT algorithms are kept in separate files to ease
+ * machine-dependent tuning (e.g., assembly coding).
+ */
+
+#include "jsamplecomp.h"
+
+
+/*
+ * A forward DCT routine is given a pointer to a work area of type DCTELEM[];
+ * the DCT is to be performed in-place in that buffer.  Type DCTELEM is int
+ * for 8-bit samples, JLONG for 12-bit samples.  (NOTE: Floating-point DCT
+ * implementations use an array of type FAST_FLOAT, instead.)
+ * The DCT inputs are expected to be signed (range +-_CENTERJSAMPLE).
+ * The DCT outputs are returned scaled up by a factor of 8; they therefore
+ * have a range of +-8K for 8-bit data, +-128K for 12-bit data.  This
+ * convention improves accuracy in integer implementations and saves some
+ * work in floating-point ones.
+ * Quantization of the output coefficients is done by jcdctmgr.c. This
+ * step requires an unsigned type and also one with twice the bits.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#ifndef WITH_SIMD
+typedef int DCTELEM;            /* 16 or 32 bits is fine */
+typedef unsigned int UDCTELEM;
+typedef unsigned long long UDCTELEM2;
+#else
+typedef short DCTELEM;          /* prefer 16 bit with SIMD for parellelism */
+typedef unsigned short UDCTELEM;
+typedef unsigned int UDCTELEM2;
+#endif
+#else
+typedef JLONG DCTELEM;          /* must have 32 bits */
+typedef unsigned long long UDCTELEM2;
+#endif
+
+
+/*
+ * An inverse DCT routine is given a pointer to the input JBLOCK and a pointer
+ * to an output sample array.  The routine must dequantize the input data as
+ * well as perform the IDCT; for dequantization, it uses the multiplier table
+ * pointed to by compptr->dct_table.  The output data is to be placed into the
+ * sample array starting at a specified column.  (Any row offset needed will
+ * be applied to the array pointer before it is passed to the IDCT code.)
+ * Note that the number of samples emitted by the IDCT routine is
+ * DCT_scaled_size * DCT_scaled_size.
+ */
+
+/* typedef inverse_DCT_method_ptr is declared in jpegint.h */
+
+/*
+ * Each IDCT routine has its own ideas about the best dct_table element type.
+ */
+
+typedef MULTIPLIER ISLOW_MULT_TYPE;  /* short or int, whichever is faster */
+#if BITS_IN_JSAMPLE == 8
+typedef MULTIPLIER IFAST_MULT_TYPE;  /* 16 bits is OK, use short if faster */
+#define IFAST_SCALE_BITS  2          /* fractional bits in scale factors */
+#else
+typedef JLONG IFAST_MULT_TYPE;       /* need 32 bits for scaled quantizers */
+#define IFAST_SCALE_BITS  13         /* fractional bits in scale factors */
+#endif
+typedef FAST_FLOAT FLOAT_MULT_TYPE;  /* preferred floating type */
+
+
+/*
+ * Each IDCT routine is responsible for range-limiting its results and
+ * converting them to unsigned form (0.._MAXJSAMPLE).  The raw outputs could
+ * be quite far out of range if the input data is corrupt, so a bulletproof
+ * range-limiting step is required.  We use a mask-and-table-lookup method
+ * to do the combined operations quickly.  See the comments with
+ * prepare_range_limit_table (in jdmaster.c) for more info.
+ */
+
+#define IDCT_range_limit(cinfo) \
+  ((_JSAMPLE *)((cinfo)->sample_range_limit) + _CENTERJSAMPLE)
+
+#define RANGE_MASK  (_MAXJSAMPLE * 4 + 3) /* 2 bits wider than legal samples */
+
+
+/* Extern declarations for the forward and inverse DCT routines. */
+
+EXTERN(void) _jpeg_fdct_islow(DCTELEM *data);
+EXTERN(void) _jpeg_fdct_ifast(DCTELEM *data);
+EXTERN(void) jpeg_fdct_float(FAST_FLOAT *data);
+
+EXTERN(void) _jpeg_idct_islow(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_ifast(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_float(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_7x7(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_6x6(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_5x5(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_4x4(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_3x3(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_2x2(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_1x1(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_9x9(j_decompress_ptr cinfo,
+                            jpeg_component_info *compptr, JCOEFPTR coef_block,
+                            _JSAMPARRAY output_buf, JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_10x10(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_11x11(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_12x12(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_13x13(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_14x14(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_15x15(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+EXTERN(void) _jpeg_idct_16x16(j_decompress_ptr cinfo,
+                              jpeg_component_info *compptr,
+                              JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                              JDIMENSION output_col);
+
+
+/*
+ * Macros for handling fixed-point arithmetic; these are used by many
+ * but not all of the DCT/IDCT modules.
+ *
+ * All values are expected to be of type JLONG.
+ * Fractional constants are scaled left by CONST_BITS bits.
+ * CONST_BITS is defined within each module using these macros,
+ * and may differ from one module to the next.
+ */
+
+#define ONE          ((JLONG)1)
+#define CONST_SCALE  (ONE << CONST_BITS)
+
+/* Convert a positive real constant to an integer scaled by CONST_SCALE.
+ * Caution: some C compilers fail to reduce "FIX(constant)" at compile time,
+ * thus causing a lot of useless floating-point operations at run time.
+ */
+
+#define FIX(x)  ((JLONG)((x) * CONST_SCALE + 0.5))
+
+/* Descale and correctly round a JLONG value that's scaled by N bits.
+ * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
+ * the fudge factor is correct for either sign of X.
+ */
+
+#define DESCALE(x, n)  RIGHT_SHIFT((x) + (ONE << ((n) - 1)), n)
+
+/* Multiply a JLONG variable by a JLONG constant to yield a JLONG result.
+ * This macro is used only when the two inputs will actually be no more than
+ * 16 bits wide, so that a 16x16->32 bit multiply can be used instead of a
+ * full 32x32 multiply.  This provides a useful speedup on many machines.
+ * Unfortunately there is no way to specify a 16x16->32 multiply portably
+ * in C, but some C compilers will do the right thing if you provide the
+ * correct combination of casts.
+ */
+
+#ifdef SHORTxSHORT_32           /* may work if 'int' is 32 bits */
+#define MULTIPLY16C16(var, const)  (((INT16)(var)) * ((INT16)(const)))
+#endif
+#ifdef SHORTxLCONST_32          /* known to work with Microsoft C 6.0 */
+#define MULTIPLY16C16(var, const)  (((INT16)(var)) * ((JLONG)(const)))
+#endif
+
+#ifndef MULTIPLY16C16           /* default definition */
+#define MULTIPLY16C16(var, const)  ((var) * (const))
+#endif
+
+/* Same except both inputs are variables. */
+
+#ifdef SHORTxSHORT_32           /* may work if 'int' is 32 bits */
+#define MULTIPLY16V16(var1, var2)  (((INT16)(var1)) * ((INT16)(var2)))
+#endif
+
+#ifndef MULTIPLY16V16           /* default definition */
+#define MULTIPLY16V16(var1, var2)  ((var1) * (var2))
+#endif

thirdparty/libjpeg-turbo/src/jddctmgr.c

@@ -0,0 +1,365 @@
+/*
+ * jddctmgr.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2002-2010 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2010, 2015, 2022, D. R. Commander.
+ * Copyright (C) 2013, MIPS Technologies, Inc., California.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the inverse-DCT management logic.
+ * This code selects a particular IDCT implementation to be used,
+ * and it performs related housekeeping chores.  No code in this file
+ * is executed per IDCT step, only during output pass setup.
+ *
+ * Note that the IDCT routines are responsible for performing coefficient
+ * dequantization as well as the IDCT proper.  This module sets up the
+ * dequantization multiplier table needed by the IDCT routine.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+#include "jsimddct.h"
+#include "jpegapicomp.h"
+
+
+/*
+ * The decompressor input side (jdinput.c) saves away the appropriate
+ * quantization table for each component at the start of the first scan
+ * involving that component.  (This is necessary in order to correctly
+ * decode files that reuse Q-table slots.)
+ * When we are ready to make an output pass, the saved Q-table is converted
+ * to a multiplier table that will actually be used by the IDCT routine.
+ * The multiplier table contents are IDCT-method-dependent.  To support
+ * application changes in IDCT method between scans, we can remake the
+ * multiplier tables if necessary.
+ * In buffered-image mode, the first output pass may occur before any data
+ * has been seen for some components, and thus before their Q-tables have
+ * been saved away.  To handle this case, multiplier tables are preset
+ * to zeroes; the result of the IDCT will be a neutral gray level.
+ */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_inverse_dct pub;  /* public fields */
+
+  /* This array contains the IDCT method code that each multiplier table
+   * is currently set up for, or -1 if it's not yet set up.
+   * The actual multiplier tables are pointed to by dct_table in the
+   * per-component comp_info structures.
+   */
+  int cur_method[MAX_COMPONENTS];
+} my_idct_controller;
+
+typedef my_idct_controller *my_idct_ptr;
+
+
+/* Allocated multiplier tables: big enough for any supported variant */
+
+typedef union {
+  ISLOW_MULT_TYPE islow_array[DCTSIZE2];
+#ifdef DCT_IFAST_SUPPORTED
+  IFAST_MULT_TYPE ifast_array[DCTSIZE2];
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  FLOAT_MULT_TYPE float_array[DCTSIZE2];
+#endif
+} multiplier_table;
+
+
+/* The current scaled-IDCT routines require ISLOW-style multiplier tables,
+ * so be sure to compile that code if either ISLOW or SCALING is requested.
+ */
+#ifdef DCT_ISLOW_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#else
+#ifdef IDCT_SCALING_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#endif
+#endif
+
+
+/*
+ * Prepare for an output pass.
+ * Here we select the proper IDCT routine for each component and build
+ * a matching multiplier table.
+ */
+
+METHODDEF(void)
+start_pass(j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct = (my_idct_ptr)cinfo->idct;
+  int ci, i;
+  jpeg_component_info *compptr;
+  int method = 0;
+  _inverse_DCT_method_ptr method_ptr = NULL;
+  JQUANT_TBL *qtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Select the proper IDCT routine for this component's scaling */
+    switch (compptr->_DCT_scaled_size) {
+#ifdef IDCT_SCALING_SUPPORTED
+    case 1:
+      method_ptr = _jpeg_idct_1x1;
+      method = JDCT_ISLOW;      /* jidctred uses islow-style table */
+      break;
+    case 2:
+#ifdef WITH_SIMD
+      if (jsimd_can_idct_2x2())
+        method_ptr = jsimd_idct_2x2;
+      else
+#endif
+        method_ptr = _jpeg_idct_2x2;
+      method = JDCT_ISLOW;      /* jidctred uses islow-style table */
+      break;
+    case 3:
+      method_ptr = _jpeg_idct_3x3;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 4:
+#ifdef WITH_SIMD
+      if (jsimd_can_idct_4x4())
+        method_ptr = jsimd_idct_4x4;
+      else
+#endif
+        method_ptr = _jpeg_idct_4x4;
+      method = JDCT_ISLOW;      /* jidctred uses islow-style table */
+      break;
+    case 5:
+      method_ptr = _jpeg_idct_5x5;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 6:
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_can_idct_6x6())
+        method_ptr = jsimd_idct_6x6;
+      else
+#endif
+      method_ptr = _jpeg_idct_6x6;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 7:
+      method_ptr = _jpeg_idct_7x7;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+#endif
+    case DCTSIZE:
+      switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+      case JDCT_ISLOW:
+#ifdef WITH_SIMD
+        if (jsimd_can_idct_islow())
+          method_ptr = jsimd_idct_islow;
+        else
+#endif
+          method_ptr = _jpeg_idct_islow;
+        method = JDCT_ISLOW;
+        break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+      case JDCT_IFAST:
+#ifdef WITH_SIMD
+        if (jsimd_can_idct_ifast())
+          method_ptr = jsimd_idct_ifast;
+        else
+#endif
+          method_ptr = _jpeg_idct_ifast;
+        method = JDCT_IFAST;
+        break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+      case JDCT_FLOAT:
+#ifdef WITH_SIMD
+        if (jsimd_can_idct_float())
+          method_ptr = jsimd_idct_float;
+        else
+#endif
+          method_ptr = _jpeg_idct_float;
+        method = JDCT_FLOAT;
+        break;
+#endif
+      default:
+        ERREXIT(cinfo, JERR_NOT_COMPILED);
+        break;
+      }
+      break;
+#ifdef IDCT_SCALING_SUPPORTED
+    case 9:
+      method_ptr = _jpeg_idct_9x9;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 10:
+      method_ptr = _jpeg_idct_10x10;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 11:
+      method_ptr = _jpeg_idct_11x11;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 12:
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_can_idct_12x12())
+        method_ptr = jsimd_idct_12x12;
+      else
+#endif
+      method_ptr = _jpeg_idct_12x12;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 13:
+      method_ptr = _jpeg_idct_13x13;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 14:
+      method_ptr = _jpeg_idct_14x14;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 15:
+      method_ptr = _jpeg_idct_15x15;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+    case 16:
+      method_ptr = _jpeg_idct_16x16;
+      method = JDCT_ISLOW;      /* jidctint uses islow-style table */
+      break;
+#endif
+    default:
+      ERREXIT1(cinfo, JERR_BAD_DCTSIZE, compptr->_DCT_scaled_size);
+      break;
+    }
+    idct->pub._inverse_DCT[ci] = method_ptr;
+    /* Create multiplier table from quant table.
+     * However, we can skip this if the component is uninteresting
+     * or if we already built the table.  Also, if no quant table
+     * has yet been saved for the component, we leave the
+     * multiplier table all-zero; we'll be reading zeroes from the
+     * coefficient controller's buffer anyway.
+     */
+    if (!compptr->component_needed || idct->cur_method[ci] == method)
+      continue;
+    qtbl = compptr->quant_table;
+    if (qtbl == NULL)           /* happens if no data yet for component */
+      continue;
+    idct->cur_method[ci] = method;
+    switch (method) {
+#ifdef PROVIDE_ISLOW_TABLES
+    case JDCT_ISLOW:
+      {
+        /* For LL&M IDCT method, multipliers are equal to raw quantization
+         * coefficients, but are stored as ints to ensure access efficiency.
+         */
+        ISLOW_MULT_TYPE *ismtbl = (ISLOW_MULT_TYPE *)compptr->dct_table;
+        for (i = 0; i < DCTSIZE2; i++) {
+          ismtbl[i] = (ISLOW_MULT_TYPE)qtbl->quantval[i];
+        }
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+        /* For AA&N IDCT method, multipliers are equal to quantization
+         * coefficients scaled by scalefactor[row]*scalefactor[col], where
+         *   scalefactor[0] = 1
+         *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+         * For integer operation, the multiplier table is to be scaled by
+         * IFAST_SCALE_BITS.
+         */
+        IFAST_MULT_TYPE *ifmtbl = (IFAST_MULT_TYPE *)compptr->dct_table;
+#define CONST_BITS  14
+        static const INT16 aanscales[DCTSIZE2] = {
+          /* precomputed values scaled up by 14 bits */
+          16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+          22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+          21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+          19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+          16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+          12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+           8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+           4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+        };
+        SHIFT_TEMPS
+
+        for (i = 0; i < DCTSIZE2; i++) {
+          ifmtbl[i] = (IFAST_MULT_TYPE)
+            DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
+                                  (JLONG)aanscales[i]),
+                    CONST_BITS - IFAST_SCALE_BITS);
+        }
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+        /* For float AA&N IDCT method, multipliers are equal to quantization
+         * coefficients scaled by scalefactor[row]*scalefactor[col], where
+         *   scalefactor[0] = 1
+         *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+         */
+        FLOAT_MULT_TYPE *fmtbl = (FLOAT_MULT_TYPE *)compptr->dct_table;
+        int row, col;
+        static const double aanscalefactor[DCTSIZE] = {
+          1.0, 1.387039845, 1.306562965, 1.175875602,
+          1.0, 0.785694958, 0.541196100, 0.275899379
+        };
+
+        i = 0;
+        for (row = 0; row < DCTSIZE; row++) {
+          for (col = 0; col < DCTSIZE; col++) {
+            fmtbl[i] = (FLOAT_MULT_TYPE)
+              ((double)qtbl->quantval[i] *
+               aanscalefactor[row] * aanscalefactor[col]);
+            i++;
+          }
+        }
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Initialize IDCT manager.
+ */
+
+GLOBAL(void)
+_jinit_inverse_dct(j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct;
+  int ci;
+  jpeg_component_info *compptr;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  idct = (my_idct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_idct_controller));
+  cinfo->idct = (struct jpeg_inverse_dct *)idct;
+  idct->pub.start_pass = start_pass;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate and pre-zero a multiplier table for each component */
+    compptr->dct_table =
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(multiplier_table));
+    memset(compptr->dct_table, 0, sizeof(multiplier_table));
+    /* Mark multiplier table not yet set up for any method */
+    idct->cur_method[ci] = -1;
+  }
+}

thirdparty/libjpeg-turbo/src/jdhuff.c

@@ -0,0 +1,836 @@
+/*
+ * jdhuff.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009-2011, 2016, 2018-2019, 2022, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains Huffman entropy decoding routines.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"             /* Declarations shared with jd*huff.c */
+#include "jpegapicomp.h"
+#include "jstdhuff.c"
+
+
+/*
+ * Expanded entropy decoder object for Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;  /* Bit buffer at start of MCU */
+  savable_state saved;          /* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl *dc_derived_tbls[NUM_HUFF_TBLS];
+  d_derived_tbl *ac_derived_tbls[NUM_HUFF_TBLS];
+
+  /* Precalculated info set up by start_pass for use in decode_mcu: */
+
+  /* Pointers to derived tables to be used for each block within an MCU */
+  d_derived_tbl *dc_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  d_derived_tbl *ac_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  /* Whether we care about the DC and AC coefficient values for each block */
+  boolean dc_needed[D_MAX_BLOCKS_IN_MCU];
+  boolean ac_needed[D_MAX_BLOCKS_IN_MCU];
+} huff_entropy_decoder;
+
+typedef huff_entropy_decoder *huff_entropy_ptr;
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_huff_decoder(j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int ci, blkn, dctbl, actbl;
+  d_derived_tbl **pdtbl;
+  jpeg_component_info *compptr;
+
+  /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+   * This ought to be an error condition, but we make it a warning because
+   * there are some baseline files out there with all zeroes in these bytes.
+   */
+  if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2 - 1 ||
+      cinfo->Ah != 0 || cinfo->Al != 0)
+    WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    /* Compute derived values for Huffman tables */
+    /* We may do this more than once for a table, but it's not expensive */
+    pdtbl = (d_derived_tbl **)(entropy->dc_derived_tbls) + dctbl;
+    jpeg_make_d_derived_tbl(cinfo, TRUE, dctbl, pdtbl);
+    pdtbl = (d_derived_tbl **)(entropy->ac_derived_tbls) + actbl;
+    jpeg_make_d_derived_tbl(cinfo, FALSE, actbl, pdtbl);
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Precalculate decoding info for each block in an MCU of this scan */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    /* Precalculate which table to use for each block */
+    entropy->dc_cur_tbls[blkn] = entropy->dc_derived_tbls[compptr->dc_tbl_no];
+    entropy->ac_cur_tbls[blkn] = entropy->ac_derived_tbls[compptr->ac_tbl_no];
+    /* Decide whether we really care about the coefficient values */
+    if (compptr->component_needed) {
+      entropy->dc_needed[blkn] = TRUE;
+      /* we don't need the ACs if producing a 1/8th-size image */
+      entropy->ac_needed[blkn] = (compptr->_DCT_scaled_size > 1);
+    } else {
+      entropy->dc_needed[blkn] = entropy->ac_needed[blkn] = FALSE;
+    }
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jdphuff.c and jdlhuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_d_derived_tbl(j_decompress_ptr cinfo, boolean isDC, int tblno,
+                        d_derived_tbl **pdtbl)
+{
+  JHUFF_TBL *htbl;
+  d_derived_tbl *dtbl;
+  int p, i, l, si, numsymbols;
+  int lookbits, ctr;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (d_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(d_derived_tbl));
+  dtbl = *pdtbl;
+  dtbl->pub = htbl;             /* fill in back link */
+
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int)htbl->bits[l];
+    if (i < 0 || p + i > 256)   /* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char)l;
+  }
+  huffsize[p] = 0;
+  numsymbols = p;
+
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int)huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((JLONG)code) >= (((JLONG)1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+
+  /* Figure F.15: generate decoding tables for bit-sequential decoding */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    if (htbl->bits[l]) {
+      /* valoffset[l] = huffval[] index of 1st symbol of code length l,
+       * minus the minimum code of length l
+       */
+      dtbl->valoffset[l] = (JLONG)p - (JLONG)huffcode[p];
+      p += htbl->bits[l];
+      dtbl->maxcode[l] = huffcode[p - 1]; /* maximum code of length l */
+    } else {
+      dtbl->maxcode[l] = -1;    /* -1 if no codes of this length */
+    }
+  }
+  dtbl->valoffset[17] = 0;
+  dtbl->maxcode[17] = 0xFFFFFL; /* ensures jpeg_huff_decode terminates */
+
+  /* Compute lookahead tables to speed up decoding.
+   * First we set all the table entries to 0, indicating "too long";
+   * then we iterate through the Huffman codes that are short enough and
+   * fill in all the entries that correspond to bit sequences starting
+   * with that code.
+   */
+
+  for (i = 0; i < (1 << HUFF_LOOKAHEAD); i++)
+    dtbl->lookup[i] = (HUFF_LOOKAHEAD + 1) << HUFF_LOOKAHEAD;
+
+  p = 0;
+  for (l = 1; l <= HUFF_LOOKAHEAD; l++) {
+    for (i = 1; i <= (int)htbl->bits[l]; i++, p++) {
+      /* l = current code's length, p = its index in huffcode[] & huffval[]. */
+      /* Generate left-justified code followed by all possible bit sequences */
+      lookbits = huffcode[p] << (HUFF_LOOKAHEAD - l);
+      for (ctr = 1 << (HUFF_LOOKAHEAD - l); ctr > 0; ctr--) {
+        dtbl->lookup[lookbits] = (l << HUFF_LOOKAHEAD) | htbl->huffval[p];
+        lookbits++;
+      }
+    }
+  }
+
+  /* Validate symbols as being reasonable.
+   * For AC tables, we make no check, but accept all byte values 0..255.
+   * For DC tables, we require the symbols to be in range 0..15 in lossy mode
+   * and 0..16 in lossless mode.  (Tighter bounds could be applied depending on
+   * the data depth and mode, but this is sufficient to ensure safe decoding.)
+   */
+  if (isDC) {
+    for (i = 0; i < numsymbols; i++) {
+      int sym = htbl->huffval[i];
+      if (sym < 0 || sym > (cinfo->master->lossless ? 16 : 15))
+        ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    }
+  }
+}
+
+
+/*
+ * Out-of-line code for bit fetching (shared with jdphuff.c and jdlhuff.c).
+ * See jdhuff.h for info about usage.
+ * Note: current values of get_buffer and bits_left are passed as parameters,
+ * but are returned in the corresponding fields of the state struct.
+ *
+ * On most machines MIN_GET_BITS should be 25 to allow the full 32-bit width
+ * of get_buffer to be used.  (On machines with wider words, an even larger
+ * buffer could be used.)  However, on some machines 32-bit shifts are
+ * quite slow and take time proportional to the number of places shifted.
+ * (This is true with most PC compilers, for instance.)  In this case it may
+ * be a win to set MIN_GET_BITS to the minimum value of 15.  This reduces the
+ * average shift distance at the cost of more calls to jpeg_fill_bit_buffer.
+ */
+
+#ifdef SLOW_SHIFT_32
+#define MIN_GET_BITS  15        /* minimum allowable value */
+#else
+#define MIN_GET_BITS  (BIT_BUF_SIZE - 7)
+#endif
+
+
+GLOBAL(boolean)
+jpeg_fill_bit_buffer(bitread_working_state *state,
+                     register bit_buf_type get_buffer, register int bits_left,
+                     int nbits)
+/* Load up the bit buffer to a depth of at least nbits */
+{
+  /* Copy heavily used state fields into locals (hopefully registers) */
+  register const JOCTET *next_input_byte = state->next_input_byte;
+  register size_t bytes_in_buffer = state->bytes_in_buffer;
+  j_decompress_ptr cinfo = state->cinfo;
+
+  /* Attempt to load at least MIN_GET_BITS bits into get_buffer. */
+  /* (It is assumed that no request will be for more than that many bits.) */
+  /* We fail to do so only if we hit a marker or are forced to suspend. */
+
+  if (cinfo->unread_marker == 0) {      /* cannot advance past a marker */
+    while (bits_left < MIN_GET_BITS) {
+      register int c;
+
+      /* Attempt to read a byte */
+      if (bytes_in_buffer == 0) {
+        if (!(*cinfo->src->fill_input_buffer) (cinfo))
+          return FALSE;
+        next_input_byte = cinfo->src->next_input_byte;
+        bytes_in_buffer = cinfo->src->bytes_in_buffer;
+      }
+      bytes_in_buffer--;
+      c = *next_input_byte++;
+
+      /* If it's 0xFF, check and discard stuffed zero byte */
+      if (c == 0xFF) {
+        /* Loop here to discard any padding FF's on terminating marker,
+         * so that we can save a valid unread_marker value.  NOTE: we will
+         * accept multiple FF's followed by a 0 as meaning a single FF data
+         * byte.  This data pattern is not valid according to the standard.
+         */
+        do {
+          if (bytes_in_buffer == 0) {
+            if (!(*cinfo->src->fill_input_buffer) (cinfo))
+              return FALSE;
+            next_input_byte = cinfo->src->next_input_byte;
+            bytes_in_buffer = cinfo->src->bytes_in_buffer;
+          }
+          bytes_in_buffer--;
+          c = *next_input_byte++;
+        } while (c == 0xFF);
+
+        if (c == 0) {
+          /* Found FF/00, which represents an FF data byte */
+          c = 0xFF;
+        } else {
+          /* Oops, it's actually a marker indicating end of compressed data.
+           * Save the marker code for later use.
+           * Fine point: it might appear that we should save the marker into
+           * bitread working state, not straight into permanent state.  But
+           * once we have hit a marker, we cannot need to suspend within the
+           * current MCU, because we will read no more bytes from the data
+           * source.  So it is OK to update permanent state right away.
+           */
+          cinfo->unread_marker = c;
+          /* See if we need to insert some fake zero bits. */
+          goto no_more_bytes;
+        }
+      }
+
+      /* OK, load c into get_buffer */
+      get_buffer = (get_buffer << 8) | c;
+      bits_left += 8;
+    } /* end while */
+  } else {
+no_more_bytes:
+    /* We get here if we've read the marker that terminates the compressed
+     * data segment.  There should be enough bits in the buffer register
+     * to satisfy the request; if so, no problem.
+     */
+    if (nbits > bits_left) {
+      /* Uh-oh.  Report corrupted data to user and stuff zeroes into
+       * the data stream, so that we can produce some kind of image.
+       * We use a nonvolatile flag to ensure that only one warning message
+       * appears per data segment.
+       */
+      if (!cinfo->entropy->insufficient_data) {
+        WARNMS(cinfo, JWRN_HIT_MARKER);
+        cinfo->entropy->insufficient_data = TRUE;
+      }
+      /* Fill the buffer with zero bits */
+      get_buffer <<= MIN_GET_BITS - bits_left;
+      bits_left = MIN_GET_BITS;
+    }
+  }
+
+  /* Unload the local registers */
+  state->next_input_byte = next_input_byte;
+  state->bytes_in_buffer = bytes_in_buffer;
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  return TRUE;
+}
+
+
+/* Macro version of the above, which performs much better but does not
+   handle markers.  We have to hand off any blocks with markers to the
+   slower routines. */
+
+#define GET_BYTE { \
+  register int c0, c1; \
+  c0 = *buffer++; \
+  c1 = *buffer; \
+  /* Pre-execute most common case */ \
+  get_buffer = (get_buffer << 8) | c0; \
+  bits_left += 8; \
+  if (c0 == 0xFF) { \
+    /* Pre-execute case of FF/00, which represents an FF data byte */ \
+    buffer++; \
+    if (c1 != 0) { \
+      /* Oops, it's actually a marker indicating end of compressed data. */ \
+      cinfo->unread_marker = c1; \
+      /* Back out pre-execution and fill the buffer with zero bits */ \
+      buffer -= 2; \
+      get_buffer &= ~0xFF; \
+    } \
+  } \
+}
+
+#if SIZEOF_SIZE_T == 8 || defined(_WIN64) || (defined(__x86_64__) && defined(__ILP32__))
+
+/* Pre-fetch 48 bytes, because the holding register is 64-bit */
+#define FILL_BIT_BUFFER_FAST \
+  if (bits_left <= 16) { \
+    GET_BYTE GET_BYTE GET_BYTE GET_BYTE GET_BYTE GET_BYTE \
+  }
+
+#else
+
+/* Pre-fetch 16 bytes, because the holding register is 32-bit */
+#define FILL_BIT_BUFFER_FAST \
+  if (bits_left <= 16) { \
+    GET_BYTE GET_BYTE \
+  }
+
+#endif
+
+
+/*
+ * Out-of-line code for Huffman code decoding.
+ * See jdhuff.h for info about usage.
+ */
+
+GLOBAL(int)
+jpeg_huff_decode(bitread_working_state *state,
+                 register bit_buf_type get_buffer, register int bits_left,
+                 d_derived_tbl *htbl, int min_bits)
+{
+  register int l = min_bits;
+  register JLONG code;
+
+  /* HUFF_DECODE has determined that the code is at least min_bits */
+  /* bits long, so fetch that many bits in one swoop. */
+
+  CHECK_BIT_BUFFER(*state, l, return -1);
+  code = GET_BITS(l);
+
+  /* Collect the rest of the Huffman code one bit at a time. */
+  /* This is per Figure F.16. */
+
+  while (code > htbl->maxcode[l]) {
+    code <<= 1;
+    CHECK_BIT_BUFFER(*state, 1, return -1);
+    code |= GET_BITS(1);
+    l++;
+  }
+
+  /* Unload the local registers */
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  /* With garbage input we may reach the sentinel value l = 17. */
+
+  if (l > 16) {
+    WARNMS(state->cinfo, JWRN_HUFF_BAD_CODE);
+    return 0;                   /* fake a zero as the safest result */
+  }
+
+  return htbl->pub->huffval[(int)(code + htbl->valoffset[l])];
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#define AVOID_TABLES
+#ifdef AVOID_TABLES
+
+#define NEG_1  ((unsigned int)-1)
+#define HUFF_EXTEND(x, s) \
+  ((x) + ((((x) - (1 << ((s) - 1))) >> 31) & (((NEG_1) << (s)) + 1)))
+
+#else
+
+#define HUFF_EXTEND(x, s) \
+  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] = {   /* entry n is 2**(n-1) */
+  0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+  0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000
+};
+
+static const int extend_offset[16] = { /* entry n is (-1 << n) + 1 */
+  0, ((-1) << 1) + 1, ((-1) << 2) + 1, ((-1) << 3) + 1, ((-1) << 4) + 1,
+  ((-1) << 5) + 1, ((-1) << 6) + 1, ((-1) << 7) + 1, ((-1) << 8) + 1,
+  ((-1) << 9) + 1, ((-1) << 10) + 1, ((-1) << 11) + 1, ((-1) << 12) + 1,
+  ((-1) << 13) + 1, ((-1) << 14) + 1, ((-1) << 15) + 1
+};
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(boolean)
+process_restart(j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (!(*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+#if defined(__has_feature)
+#if __has_feature(undefined_behavior_sanitizer)
+__attribute__((no_sanitize("signed-integer-overflow"),
+               no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
+LOCAL(boolean)
+decode_mcu_slow(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  BITREAD_STATE_VARS;
+  int blkn;
+  savable_state state;
+  /* Outer loop handles each block in the MCU */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+  state = entropy->saved;
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    JBLOCKROW block = MCU_data ? MCU_data[blkn] : NULL;
+    d_derived_tbl *dctbl = entropy->dc_cur_tbls[blkn];
+    d_derived_tbl *actbl = entropy->ac_cur_tbls[blkn];
+    register int s, k, r;
+
+    /* Decode a single block's worth of coefficients */
+
+    /* Section F.2.2.1: decode the DC coefficient difference */
+    HUFF_DECODE(s, br_state, dctbl, return FALSE, label1);
+    if (s) {
+      CHECK_BIT_BUFFER(br_state, s, return FALSE);
+      r = GET_BITS(s);
+      s = HUFF_EXTEND(r, s);
+    }
+
+    if (entropy->dc_needed[blkn]) {
+      /* Convert DC difference to actual value, update last_dc_val */
+      int ci = cinfo->MCU_membership[blkn];
+      /* Certain malformed JPEG images produce repeated DC coefficient
+       * differences of 2047 or -2047, which causes state.last_dc_val[ci] to
+       * grow until it overflows or underflows a 32-bit signed integer.  This
+       * behavior is, to the best of our understanding, innocuous, and it is
+       * unclear how to work around it without potentially affecting
+       * performance.  Thus, we (hopefully temporarily) suppress UBSan integer
+       * overflow errors for this function and decode_mcu_fast().
+       */
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      if (block) {
+        /* Output the DC coefficient (assumes jpeg_natural_order[0] = 0) */
+        (*block)[0] = (JCOEF)s;
+      }
+    }
+
+    if (entropy->ac_needed[blkn] && block) {
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* Since zeroes are skipped, output area must be cleared beforehand */
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE(s, br_state, actbl, return FALSE, label2);
+
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          CHECK_BIT_BUFFER(br_state, s, return FALSE);
+          r = GET_BITS(s);
+          s = HUFF_EXTEND(r, s);
+          /* Output coefficient in natural (dezigzagged) order.
+           * Note: the extra entries in jpeg_natural_order[] will save us
+           * if k >= DCTSIZE2, which could happen if the data is corrupted.
+           */
+          (*block)[jpeg_natural_order[k]] = (JCOEF)s;
+        } else {
+          if (r != 15)
+            break;
+          k += 15;
+        }
+      }
+
+    } else {
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* In this path we just discard the values */
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE(s, br_state, actbl, return FALSE, label3);
+
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          CHECK_BIT_BUFFER(br_state, s, return FALSE);
+          DROP_BITS(s);
+        } else {
+          if (r != 15)
+            break;
+          k += 15;
+        }
+      }
+    }
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+  entropy->saved = state;
+  return TRUE;
+}
+
+
+#if defined(__has_feature)
+#if __has_feature(undefined_behavior_sanitizer)
+__attribute__((no_sanitize("signed-integer-overflow"),
+               no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
+LOCAL(boolean)
+decode_mcu_fast(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  BITREAD_STATE_VARS;
+  JOCTET *buffer;
+  int blkn;
+  savable_state state;
+  /* Outer loop handles each block in the MCU */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+  buffer = (JOCTET *)br_state.next_input_byte;
+  state = entropy->saved;
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    JBLOCKROW block = MCU_data ? MCU_data[blkn] : NULL;
+    d_derived_tbl *dctbl = entropy->dc_cur_tbls[blkn];
+    d_derived_tbl *actbl = entropy->ac_cur_tbls[blkn];
+    register int s, k, r, l;
+
+    HUFF_DECODE_FAST(s, l, dctbl);
+    if (s) {
+      FILL_BIT_BUFFER_FAST
+      r = GET_BITS(s);
+      s = HUFF_EXTEND(r, s);
+    }
+
+    if (entropy->dc_needed[blkn]) {
+      int ci = cinfo->MCU_membership[blkn];
+      /* Refer to the comment in decode_mcu_slow() regarding the supression of
+       * a UBSan integer overflow error in this line of code.
+       */
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      if (block)
+        (*block)[0] = (JCOEF)s;
+    }
+
+    if (entropy->ac_needed[blkn] && block) {
+
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE_FAST(s, l, actbl);
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          FILL_BIT_BUFFER_FAST
+          r = GET_BITS(s);
+          s = HUFF_EXTEND(r, s);
+          (*block)[jpeg_natural_order[k]] = (JCOEF)s;
+        } else {
+          if (r != 15) break;
+          k += 15;
+        }
+      }
+
+    } else {
+
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE_FAST(s, l, actbl);
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          FILL_BIT_BUFFER_FAST
+          DROP_BITS(s);
+        } else {
+          if (r != 15) break;
+          k += 15;
+        }
+      }
+    }
+  }
+
+  if (cinfo->unread_marker != 0) {
+    cinfo->unread_marker = 0;
+    return FALSE;
+  }
+
+  br_state.bytes_in_buffer -= (buffer - br_state.next_input_byte);
+  br_state.next_input_byte = buffer;
+  BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+  entropy->saved = state;
+  return TRUE;
+}
+
+
+/*
+ * Decode and return one MCU's worth of Huffman-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA HAS BEEN ZEROED BY THE CALLER.
+ * (Wholesale zeroing is usually a little faster than retail...)
+ *
+ * Returns FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * this module, since we'll just re-assign them on the next call.)
+ */
+
+#define BUFSIZE  (DCTSIZE2 * 8)
+
+METHODDEF(boolean)
+decode_mcu(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy;
+  int usefast = 1;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!process_restart(cinfo))
+        return FALSE;
+    usefast = 0;
+  }
+
+  if (cinfo->src->bytes_in_buffer < BUFSIZE * (size_t)cinfo->blocks_in_MCU ||
+      cinfo->unread_marker != 0)
+    usefast = 0;
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (!entropy->pub.insufficient_data) {
+
+    if (usefast) {
+      if (!decode_mcu_fast(cinfo, MCU_data)) goto use_slow;
+    } else {
+use_slow:
+      if (!decode_mcu_slow(cinfo, MCU_data)) return FALSE;
+    }
+
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  if (cinfo->restart_interval)
+    entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * Module initialization routine for Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_huff_decoder(j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  /* Motion JPEG frames typically do not include the Huffman tables if they
+     are the default tables.  Thus, if the tables are not set by the time
+     the Huffman decoder is initialized (usually within the body of
+     jpeg_start_decompress()), we set them to default values. */
+  std_huff_tables((j_common_ptr)cinfo);
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(huff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *)entropy;
+  entropy->pub.start_pass = start_pass_huff_decoder;
+  entropy->pub.decode_mcu = decode_mcu;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+  }
+}

thirdparty/libjpeg-turbo/src/jdhuff.h

@@ -0,0 +1,250 @@
+/*
+ * jdhuff.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010-2011, 2015-2016, 2021, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains declarations for Huffman entropy decoding routines
+ * that are shared between the sequential decoder (jdhuff.c), the progressive
+ * decoder (jdphuff.c), and the lossless decoder (jdlhuff.c).  No other modules
+ * need to see these.
+ */
+
+#include "jconfigint.h"
+
+
+/* Derived data constructed for each Huffman table */
+
+#define HUFF_LOOKAHEAD  8       /* # of bits of lookahead */
+
+typedef struct {
+  /* Basic tables: (element [0] of each array is unused) */
+  JLONG maxcode[18];            /* largest code of length k (-1 if none) */
+  /* (maxcode[17] is a sentinel to ensure jpeg_huff_decode terminates) */
+  JLONG valoffset[18];          /* huffval[] offset for codes of length k */
+  /* valoffset[k] = huffval[] index of 1st symbol of code length k, less
+   * the smallest code of length k; so given a code of length k, the
+   * corresponding symbol is huffval[code + valoffset[k]]
+   */
+
+  /* Link to public Huffman table (needed only in jpeg_huff_decode) */
+  JHUFF_TBL *pub;
+
+  /* Lookahead table: indexed by the next HUFF_LOOKAHEAD bits of
+   * the input data stream.  If the next Huffman code is no more
+   * than HUFF_LOOKAHEAD bits long, we can obtain its length and
+   * the corresponding symbol directly from this tables.
+   *
+   * The lower 8 bits of each table entry contain the number of
+   * bits in the corresponding Huffman code, or HUFF_LOOKAHEAD + 1
+   * if too long.  The next 8 bits of each entry contain the
+   * symbol.
+   */
+  int lookup[1 << HUFF_LOOKAHEAD];
+} d_derived_tbl;
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_d_derived_tbl(j_decompress_ptr cinfo, boolean isDC,
+                                     int tblno, d_derived_tbl **pdtbl);
+
+
+/*
+ * Fetching the next N bits from the input stream is a time-critical operation
+ * for the Huffman decoders.  We implement it with a combination of inline
+ * macros and out-of-line subroutines.  Note that N (the number of bits
+ * demanded at one time) never exceeds 15 for JPEG use.
+ *
+ * We read source bytes into get_buffer and dole out bits as needed.
+ * If get_buffer already contains enough bits, they are fetched in-line
+ * by the macros CHECK_BIT_BUFFER and GET_BITS.  When there aren't enough
+ * bits, jpeg_fill_bit_buffer is called; it will attempt to fill get_buffer
+ * as full as possible (not just to the number of bits needed; this
+ * prefetching reduces the overhead cost of calling jpeg_fill_bit_buffer).
+ * Note that jpeg_fill_bit_buffer may return FALSE to indicate suspension.
+ * On TRUE return, jpeg_fill_bit_buffer guarantees that get_buffer contains
+ * at least the requested number of bits --- dummy zeroes are inserted if
+ * necessary.
+ */
+
+#if !defined(_WIN32) && !defined(SIZEOF_SIZE_T)
+#error Cannot determine word size
+#endif
+
+#if SIZEOF_SIZE_T == 8 || defined(_WIN64)
+
+typedef size_t bit_buf_type;            /* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  64                /* size of buffer in bits */
+
+#elif defined(__x86_64__) && defined(__ILP32__)
+
+typedef unsigned long long bit_buf_type; /* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  64                 /* size of buffer in bits */
+
+#else
+
+typedef unsigned long bit_buf_type;     /* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  32                /* size of buffer in bits */
+
+#endif
+
+/* If long is > 32 bits on your machine, and shifting/masking longs is
+ * reasonably fast, making bit_buf_type be long and setting BIT_BUF_SIZE
+ * appropriately should be a win.  Unfortunately we can't define the size
+ * with something like  #define BIT_BUF_SIZE (sizeof(bit_buf_type)*8)
+ * because not all machines measure sizeof in 8-bit bytes.
+ */
+
+typedef struct {                /* Bitreading state saved across MCUs */
+  bit_buf_type get_buffer;      /* current bit-extraction buffer */
+  int bits_left;                /* # of unused bits in it */
+} bitread_perm_state;
+
+typedef struct {                /* Bitreading working state within an MCU */
+  /* Current data source location */
+  /* We need a copy, rather than munging the original, in case of suspension */
+  const JOCTET *next_input_byte; /* => next byte to read from source */
+  size_t bytes_in_buffer;       /* # of bytes remaining in source buffer */
+  /* Bit input buffer --- note these values are kept in register variables,
+   * not in this struct, inside the inner loops.
+   */
+  bit_buf_type get_buffer;      /* current bit-extraction buffer */
+  int bits_left;                /* # of unused bits in it */
+  /* Pointer needed by jpeg_fill_bit_buffer. */
+  j_decompress_ptr cinfo;       /* back link to decompress master record */
+} bitread_working_state;
+
+/* Macros to declare and load/save bitread local variables. */
+#define BITREAD_STATE_VARS \
+  register bit_buf_type get_buffer; \
+  register int bits_left; \
+  bitread_working_state br_state
+
+#define BITREAD_LOAD_STATE(cinfop, permstate) \
+  br_state.cinfo = cinfop; \
+  br_state.next_input_byte = cinfop->src->next_input_byte; \
+  br_state.bytes_in_buffer = cinfop->src->bytes_in_buffer; \
+  get_buffer = permstate.get_buffer; \
+  bits_left = permstate.bits_left;
+
+#define BITREAD_SAVE_STATE(cinfop, permstate) \
+  cinfop->src->next_input_byte = br_state.next_input_byte; \
+  cinfop->src->bytes_in_buffer = br_state.bytes_in_buffer; \
+  permstate.get_buffer = get_buffer; \
+  permstate.bits_left = bits_left
+
+/*
+ * These macros provide the in-line portion of bit fetching.
+ * Use CHECK_BIT_BUFFER to ensure there are N bits in get_buffer
+ * before using GET_BITS, PEEK_BITS, or DROP_BITS.
+ * The variables get_buffer and bits_left are assumed to be locals,
+ * but the state struct might not be (jpeg_huff_decode needs this).
+ *      CHECK_BIT_BUFFER(state, n, action);
+ *              Ensure there are N bits in get_buffer; if suspend, take action.
+ *      val = GET_BITS(n);
+ *              Fetch next N bits.
+ *      val = PEEK_BITS(n);
+ *              Fetch next N bits without removing them from the buffer.
+ *      DROP_BITS(n);
+ *              Discard next N bits.
+ * The value N should be a simple variable, not an expression, because it
+ * is evaluated multiple times.
+ */
+
+#define CHECK_BIT_BUFFER(state, nbits, action) { \
+  if (bits_left < (nbits)) { \
+    if (!jpeg_fill_bit_buffer(&(state), get_buffer, bits_left, nbits)) \
+      { action; } \
+    get_buffer = (state).get_buffer;  bits_left = (state).bits_left; \
+  } \
+}
+
+#define GET_BITS(nbits) \
+  (((int)(get_buffer >> (bits_left -= (nbits)))) & ((1 << (nbits)) - 1))
+
+#define PEEK_BITS(nbits) \
+  (((int)(get_buffer >> (bits_left -  (nbits)))) & ((1 << (nbits)) - 1))
+
+#define DROP_BITS(nbits) \
+  (bits_left -= (nbits))
+
+/* Load up the bit buffer to a depth of at least nbits */
+EXTERN(boolean) jpeg_fill_bit_buffer(bitread_working_state *state,
+                                     register bit_buf_type get_buffer,
+                                     register int bits_left, int nbits);
+
+
+/*
+ * Code for extracting next Huffman-coded symbol from input bit stream.
+ * Again, this is time-critical and we make the main paths be macros.
+ *
+ * We use a lookahead table to process codes of up to HUFF_LOOKAHEAD bits
+ * without looping.  Usually, more than 95% of the Huffman codes will be 8
+ * or fewer bits long.  The few overlength codes are handled with a loop,
+ * which need not be inline code.
+ *
+ * Notes about the HUFF_DECODE macro:
+ * 1. Near the end of the data segment, we may fail to get enough bits
+ *    for a lookahead.  In that case, we do it the hard way.
+ * 2. If the lookahead table contains no entry, the next code must be
+ *    more than HUFF_LOOKAHEAD bits long.
+ * 3. jpeg_huff_decode returns -1 if forced to suspend.
+ */
+
+#define HUFF_DECODE(result, state, htbl, failaction, slowlabel) { \
+  register int nb, look; \
+  if (bits_left < HUFF_LOOKAHEAD) { \
+    if (!jpeg_fill_bit_buffer(&state, get_buffer, bits_left, 0)) \
+      { failaction; } \
+    get_buffer = state.get_buffer;  bits_left = state.bits_left; \
+    if (bits_left < HUFF_LOOKAHEAD) { \
+      nb = 1;  goto slowlabel; \
+    } \
+  } \
+  look = PEEK_BITS(HUFF_LOOKAHEAD); \
+  if ((nb = (htbl->lookup[look] >> HUFF_LOOKAHEAD)) <= HUFF_LOOKAHEAD) { \
+    DROP_BITS(nb); \
+    result = htbl->lookup[look] & ((1 << HUFF_LOOKAHEAD) - 1); \
+  } else { \
+slowlabel: \
+    if ((result = \
+         jpeg_huff_decode(&state, get_buffer, bits_left, htbl, nb)) < 0) \
+      { failaction; } \
+    get_buffer = state.get_buffer;  bits_left = state.bits_left; \
+  } \
+}
+
+#define HUFF_DECODE_FAST(s, nb, htbl) \
+  FILL_BIT_BUFFER_FAST; \
+  s = PEEK_BITS(HUFF_LOOKAHEAD); \
+  s = htbl->lookup[s]; \
+  nb = s >> HUFF_LOOKAHEAD; \
+  /* Pre-execute the common case of nb <= HUFF_LOOKAHEAD */ \
+  DROP_BITS(nb); \
+  s = s & ((1 << HUFF_LOOKAHEAD) - 1); \
+  if (nb > HUFF_LOOKAHEAD) { \
+    /* Equivalent of jpeg_huff_decode() */ \
+    /* Don't use GET_BITS() here because we don't want to modify bits_left */ \
+    s = (get_buffer >> bits_left) & ((1 << (nb)) - 1); \
+    while (s > htbl->maxcode[nb]) { \
+      s <<= 1; \
+      s |= GET_BITS(1); \
+      nb++; \
+    } \
+    if (nb > 16) \
+      s = 0; \
+    else \
+      s = htbl->pub->huffval[(int)(s + htbl->valoffset[nb]) & 0xFF]; \
+  }
+
+/* Out-of-line case for Huffman code fetching */
+EXTERN(int) jpeg_huff_decode(bitread_working_state *state,
+                             register bit_buf_type get_buffer,
+                             register int bits_left, d_derived_tbl *htbl,
+                             int min_bits);

thirdparty/libjpeg-turbo/src/jdicc.c

@@ -0,0 +1,167 @@
+/*
+ * jdicc.c
+ *
+ * Copyright (C) 1997-1998, Thomas G. Lane, Todd Newman.
+ * Copyright (C) 2017, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file provides code to read International Color Consortium (ICC) device
+ * profiles embedded in JFIF JPEG image files.  The ICC has defined a standard
+ * for including such data in JPEG "APP2" markers.  The code given here does
+ * not know anything about the internal structure of the ICC profile data; it
+ * just knows how to get the profile data from a JPEG file while reading it.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+#define ICC_MARKER  (JPEG_APP0 + 2)     /* JPEG marker code for ICC */
+#define ICC_OVERHEAD_LEN  14            /* size of non-profile data in APP2 */
+
+
+/*
+ * Handy subroutine to test whether a saved marker is an ICC profile marker.
+ */
+
+LOCAL(boolean)
+marker_is_icc(jpeg_saved_marker_ptr marker)
+{
+  return
+    marker->marker == ICC_MARKER &&
+    marker->data_length >= ICC_OVERHEAD_LEN &&
+    /* verify the identifying string */
+    marker->data[0] == 0x49 &&
+    marker->data[1] == 0x43 &&
+    marker->data[2] == 0x43 &&
+    marker->data[3] == 0x5F &&
+    marker->data[4] == 0x50 &&
+    marker->data[5] == 0x52 &&
+    marker->data[6] == 0x4F &&
+    marker->data[7] == 0x46 &&
+    marker->data[8] == 0x49 &&
+    marker->data[9] == 0x4C &&
+    marker->data[10] == 0x45 &&
+    marker->data[11] == 0x0;
+}
+
+
+/*
+ * See if there was an ICC profile in the JPEG file being read; if so,
+ * reassemble and return the profile data.
+ *
+ * TRUE is returned if an ICC profile was found, FALSE if not.  If TRUE is
+ * returned, *icc_data_ptr is set to point to the returned data, and
+ * *icc_data_len is set to its length.
+ *
+ * IMPORTANT: the data at *icc_data_ptr is allocated with malloc() and must be
+ * freed by the caller with free() when the caller no longer needs it.
+ * (Alternatively, we could write this routine to use the IJG library's memory
+ * allocator, so that the data would be freed implicitly when
+ * jpeg_finish_decompress() is called.  But it seems likely that many
+ * applications will prefer to have the data stick around after decompression
+ * finishes.)
+ */
+
+GLOBAL(boolean)
+jpeg_read_icc_profile(j_decompress_ptr cinfo, JOCTET **icc_data_ptr,
+                      unsigned int *icc_data_len)
+{
+  jpeg_saved_marker_ptr marker;
+  int num_markers = 0;
+  int seq_no;
+  JOCTET *icc_data;
+  unsigned int total_length;
+#define MAX_SEQ_NO  255         /* sufficient since marker numbers are bytes */
+  char marker_present[MAX_SEQ_NO + 1];      /* 1 if marker found */
+  unsigned int data_length[MAX_SEQ_NO + 1]; /* size of profile data in marker */
+  unsigned int data_offset[MAX_SEQ_NO + 1]; /* offset for data in marker */
+
+  if (icc_data_ptr == NULL || icc_data_len == NULL)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+  if (cinfo->global_state < DSTATE_READY)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  *icc_data_ptr = NULL;         /* avoid confusion if FALSE return */
+  *icc_data_len = 0;
+
+  /* This first pass over the saved markers discovers whether there are
+   * any ICC markers and verifies the consistency of the marker numbering.
+   */
+
+  for (seq_no = 1; seq_no <= MAX_SEQ_NO; seq_no++)
+    marker_present[seq_no] = 0;
+
+  for (marker = cinfo->marker_list; marker != NULL; marker = marker->next) {
+    if (marker_is_icc(marker)) {
+      if (num_markers == 0)
+        num_markers = marker->data[13];
+      else if (num_markers != marker->data[13]) {
+        WARNMS(cinfo, JWRN_BOGUS_ICC);  /* inconsistent num_markers fields */
+        return FALSE;
+      }
+      seq_no = marker->data[12];
+      if (seq_no <= 0 || seq_no > num_markers) {
+        WARNMS(cinfo, JWRN_BOGUS_ICC);  /* bogus sequence number */
+        return FALSE;
+      }
+      if (marker_present[seq_no]) {
+        WARNMS(cinfo, JWRN_BOGUS_ICC);  /* duplicate sequence numbers */
+        return FALSE;
+      }
+      marker_present[seq_no] = 1;
+      data_length[seq_no] = marker->data_length - ICC_OVERHEAD_LEN;
+    }
+  }
+
+  if (num_markers == 0)
+    return FALSE;
+
+  /* Check for missing markers, count total space needed,
+   * compute offset of each marker's part of the data.
+   */
+
+  total_length = 0;
+  for (seq_no = 1; seq_no <= num_markers; seq_no++) {
+    if (marker_present[seq_no] == 0) {
+      WARNMS(cinfo, JWRN_BOGUS_ICC);  /* missing sequence number */
+      return FALSE;
+    }
+    data_offset[seq_no] = total_length;
+    total_length += data_length[seq_no];
+  }
+
+  if (total_length == 0) {
+    WARNMS(cinfo, JWRN_BOGUS_ICC);  /* found only empty markers? */
+    return FALSE;
+  }
+
+  /* Allocate space for assembled data */
+  icc_data = (JOCTET *)malloc(total_length * sizeof(JOCTET));
+  if (icc_data == NULL)
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 11);  /* oops, out of memory */
+
+  /* and fill it in */
+  for (marker = cinfo->marker_list; marker != NULL; marker = marker->next) {
+    if (marker_is_icc(marker)) {
+      JOCTET FAR *src_ptr;
+      JOCTET *dst_ptr;
+      unsigned int length;
+      seq_no = marker->data[12];
+      dst_ptr = icc_data + data_offset[seq_no];
+      src_ptr = marker->data + ICC_OVERHEAD_LEN;
+      length = data_length[seq_no];
+      while (length--) {
+        *dst_ptr++ = *src_ptr++;
+      }
+    }
+  }
+
+  *icc_data_ptr = icc_data;
+  *icc_data_len = total_length;
+
+  return TRUE;
+}

thirdparty/libjpeg-turbo/src/jdinput.c

@@ -0,0 +1,424 @@
+/*
+ * jdinput.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2016, 2018, 2022, 2024, D. R. Commander.
+ * Copyright (C) 2015, Google, Inc.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains input control logic for the JPEG decompressor.
+ * These routines are concerned with controlling the decompressor's input
+ * processing (marker reading and coefficient/difference decoding).
+ * The actual input reading is done in jdmarker.c, jdhuff.c, jdphuff.c,
+ * and jdlhuff.c.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_input_controller pub; /* public fields */
+
+  boolean inheaders;            /* TRUE until first SOS is reached */
+} my_input_controller;
+
+typedef my_input_controller *my_inputctl_ptr;
+
+
+/* Forward declarations */
+METHODDEF(int) consume_markers(j_decompress_ptr cinfo);
+
+
+/*
+ * Routines to calculate various quantities related to the size of the image.
+ */
+
+LOCAL(void)
+initial_setup(j_decompress_ptr cinfo)
+/* Called once, when first SOS marker is reached */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long)cinfo->image_height > (long)JPEG_MAX_DIMENSION ||
+      (long)cinfo->image_width > (long)JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int)JPEG_MAX_DIMENSION);
+
+  /* Lossy JPEG images must have 8 or 12 bits per sample.  Lossless JPEG images
+   * can have 2 to 16 bits per sample.
+   */
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+    if (cinfo->data_precision < 2 || cinfo->data_precision > 16)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != 8 && cinfo->data_precision != 12)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+             MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor <= 0 ||
+        compptr->h_samp_factor > MAX_SAMP_FACTOR ||
+        compptr->v_samp_factor <= 0 ||
+        compptr->v_samp_factor > MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+                                   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+                                   compptr->v_samp_factor);
+  }
+
+#if JPEG_LIB_VERSION >= 80
+  cinfo->block_size = data_unit;
+  cinfo->natural_order = jpeg_natural_order;
+  cinfo->lim_Se = DCTSIZE2 - 1;
+#endif
+
+  /* We initialize DCT_scaled_size and min_DCT_scaled_size to DCTSIZE in lossy
+   * mode.  In the full decompressor, this will be overridden by jdmaster.c;
+   * but in the transcoder, jdmaster.c is not used, so we must do it here.
+   */
+#if JPEG_LIB_VERSION >= 70
+  cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = data_unit;
+#else
+  cinfo->min_DCT_scaled_size = data_unit;
+#endif
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+#if JPEG_LIB_VERSION >= 70
+    compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = data_unit;
+#else
+    compptr->DCT_scaled_size = data_unit;
+#endif
+    /* Size in data units */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_width * (long)compptr->h_samp_factor,
+                    (long)(cinfo->max_h_samp_factor * data_unit));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_height * (long)compptr->v_samp_factor,
+                    (long)(cinfo->max_v_samp_factor * data_unit));
+    /* Set the first and last MCU columns to decompress from multi-scan images.
+     * By default, decompress all of the MCU columns.
+     */
+    cinfo->master->first_MCU_col[ci] = 0;
+    cinfo->master->last_MCU_col[ci] = compptr->width_in_blocks - 1;
+    /* downsampled_width and downsampled_height will also be overridden by
+     * jdmaster.c if we are doing full decompression.  The transcoder library
+     * doesn't use these values, but the calling application might.
+     */
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_width * (long)compptr->h_samp_factor,
+                    (long)cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_height * (long)compptr->v_samp_factor,
+                    (long)cinfo->max_v_samp_factor);
+    /* Mark component needed, until color conversion says otherwise */
+    compptr->component_needed = TRUE;
+    /* Mark no quantization table yet saved for component */
+    compptr->quant_table = NULL;
+  }
+
+  /* Compute number of fully interleaved MCU rows. */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long)cinfo->image_height,
+                  (long)(cinfo->max_v_samp_factor * data_unit));
+
+  /* Decide whether file contains multiple scans */
+  if (cinfo->comps_in_scan < cinfo->num_components || cinfo->progressive_mode)
+    cinfo->inputctl->has_multiple_scans = TRUE;
+  else
+    cinfo->inputctl->has_multiple_scans = FALSE;
+}
+
+
+LOCAL(void)
+per_scan_setup(j_decompress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] were set from SOS marker */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
+
+  if (cinfo->comps_in_scan == 1) {
+
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+
+    /* For noninterleaved scan, always one data unit per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = compptr->_DCT_scaled_size;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of data unit rows present in the last iMCU row.
+     */
+    tmp = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+
+  } else {
+
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+               MAX_COMPS_IN_SCAN);
+
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_width,
+                    (long)(cinfo->max_h_samp_factor * data_unit));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long)cinfo->image_height,
+                    (long)(cinfo->max_v_samp_factor * data_unit));
+
+    cinfo->blocks_in_MCU = 0;
+
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of data units of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width *
+                                  compptr->_DCT_scaled_size;
+      /* Figure number of non-dummy data units in last MCU column & row */
+      tmp = (int)(compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int)(compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > D_MAX_BLOCKS_IN_MCU)
+        ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+        cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+
+  }
+}
+
+
+/*
+ * Save away a copy of the Q-table referenced by each component present
+ * in the current scan, unless already saved during a prior scan.
+ *
+ * In a multiple-scan JPEG file, the encoder could assign different components
+ * the same Q-table slot number, but change table definitions between scans
+ * so that each component uses a different Q-table.  (The IJG encoder is not
+ * currently capable of doing this, but other encoders might.)  Since we want
+ * to be able to dequantize all the components at the end of the file, this
+ * means that we have to save away the table actually used for each component.
+ * We do this by copying the table at the start of the first scan containing
+ * the component.
+ * Rec. ITU-T T.81 | ISO/IEC 10918-1 prohibits the encoder from changing the
+ * contents of a Q-table slot between scans of a component using that slot.  If
+ * the encoder does so anyway, this decoder will simply use the Q-table values
+ * that were current at the start of the first scan for the component.
+ *
+ * The decompressor output side looks only at the saved quant tables,
+ * not at the current Q-table slots.
+ */
+
+LOCAL(void)
+latch_quant_tables(j_decompress_ptr cinfo)
+{
+  int ci, qtblno;
+  jpeg_component_info *compptr;
+  JQUANT_TBL *qtbl;
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* No work if we already saved Q-table for this component */
+    if (compptr->quant_table != NULL)
+      continue;
+    /* Make sure specified quantization table is present */
+    qtblno = compptr->quant_tbl_no;
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+        cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    /* OK, save away the quantization table */
+    qtbl = (JQUANT_TBL *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(JQUANT_TBL));
+    memcpy(qtbl, cinfo->quant_tbl_ptrs[qtblno], sizeof(JQUANT_TBL));
+    compptr->quant_table = qtbl;
+  }
+}
+
+
+/*
+ * Initialize the input modules to read a scan of compressed data.
+ * The first call to this is done by jdmaster.c after initializing
+ * the entire decompressor (during jpeg_start_decompress).
+ * Subsequent calls come from consume_markers, below.
+ */
+
+METHODDEF(void)
+start_input_pass(j_decompress_ptr cinfo)
+{
+  per_scan_setup(cinfo);
+  if (!cinfo->master->lossless)
+    latch_quant_tables(cinfo);
+  (*cinfo->entropy->start_pass) (cinfo);
+  (*cinfo->coef->start_input_pass) (cinfo);
+  cinfo->inputctl->consume_input = cinfo->coef->consume_data;
+}
+
+
+/*
+ * Finish up after inputting a compressed-data scan.
+ * This is called by the coefficient or difference controller after it's read
+ * all the expected data of the scan.
+ */
+
+METHODDEF(void)
+finish_input_pass(j_decompress_ptr cinfo)
+{
+  cinfo->inputctl->consume_input = consume_markers;
+}
+
+
+/*
+ * Read JPEG markers before, between, or after compressed-data scans.
+ * Change state as necessary when a new scan is reached.
+ * Return value is JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ *
+ * The consume_input method pointer points either here or to the
+ * coefficient or difference controller's consume_data routine, depending on
+ * whether we are reading a compressed data segment or inter-segment markers.
+ */
+
+METHODDEF(int)
+consume_markers(j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr)cinfo->inputctl;
+  int val;
+
+  if (inputctl->pub.eoi_reached) /* After hitting EOI, read no further */
+    return JPEG_REACHED_EOI;
+
+  val = (*cinfo->marker->read_markers) (cinfo);
+
+  switch (val) {
+  case JPEG_REACHED_SOS:        /* Found SOS */
+    if (inputctl->inheaders) {  /* 1st SOS */
+      initial_setup(cinfo);
+      inputctl->inheaders = FALSE;
+      /* Note: start_input_pass must be called by jdmaster.c
+       * before any more input can be consumed.  jdapimin.c is
+       * responsible for enforcing this sequencing.
+       */
+    } else {                    /* 2nd or later SOS marker */
+      if (!inputctl->pub.has_multiple_scans)
+        ERREXIT(cinfo, JERR_EOI_EXPECTED); /* Oops, I wasn't expecting this! */
+      start_input_pass(cinfo);
+    }
+    break;
+  case JPEG_REACHED_EOI:        /* Found EOI */
+    inputctl->pub.eoi_reached = TRUE;
+    if (inputctl->inheaders) {  /* Tables-only datastream, apparently */
+      if (cinfo->marker->saw_SOF)
+        ERREXIT(cinfo, JERR_SOF_NO_SOS);
+    } else {
+      /* Prevent infinite loop in coef ctlr's decompress_data routine
+       * if user set output_scan_number larger than number of scans.
+       */
+      if (cinfo->output_scan_number > cinfo->input_scan_number)
+        cinfo->output_scan_number = cinfo->input_scan_number;
+    }
+    break;
+  case JPEG_SUSPENDED:
+    break;
+  }
+
+  return val;
+}
+
+
+/*
+ * Reset state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_input_controller(j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr)cinfo->inputctl;
+
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+  /* Reset other modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr)cinfo);
+  (*cinfo->marker->reset_marker_reader) (cinfo);
+  /* Reset progression state -- would be cleaner if entropy decoder did this */
+  cinfo->coef_bits = NULL;
+}
+
+
+/*
+ * Initialize the input controller module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_input_controller(j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl;
+
+  /* Create subobject in permanent pool */
+  inputctl = (my_inputctl_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                sizeof(my_input_controller));
+  cinfo->inputctl = (struct jpeg_input_controller *)inputctl;
+  /* Initialize method pointers */
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.reset_input_controller = reset_input_controller;
+  inputctl->pub.start_input_pass = start_input_pass;
+  inputctl->pub.finish_input_pass = finish_input_pass;
+  /* Initialize state: can't use reset_input_controller since we don't
+   * want to try to reset other modules yet.
+   */
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+}

thirdparty/libjpeg-turbo/src/jdmainct.c

@@ -0,0 +1,482 @@
+/*
+ * jdmainct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2016, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the main buffer controller for decompression.
+ * The main buffer lies between the JPEG decompressor proper and the
+ * post-processor; it holds downsampled data in the JPEG colorspace.
+ *
+ * Note that this code is bypassed in raw-data mode, since the application
+ * supplies the equivalent of the main buffer in that case.
+ */
+
+#include "jinclude.h"
+#include "jdmainct.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/*
+ * In the current system design, the main buffer need never be a full-image
+ * buffer; any full-height buffers will be found inside the coefficient,
+ * difference, or postprocessing controllers.  Nonetheless, the main controller
+ * is not trivial.  Its responsibility is to provide context rows for
+ * upsampling/rescaling, and doing this in an efficient fashion is a bit
+ * tricky.
+ *
+ * Postprocessor input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  (We require DCT_scaled_size values to be
+ * chosen such that these numbers are integers.  In practice DCT_scaled_size
+ * values will likely be powers of two, so we actually have the stronger
+ * condition that DCT_scaled_size / min_DCT_scaled_size is an integer.)
+ * Upsampling will typically produce max_v_samp_factor pixel rows from each
+ * row group (times any additional scale factor that the upsampler is
+ * applying).
+ *
+ * The coefficient or difference controller will deliver data to us one iMCU
+ * row at a time; each iMCU row contains v_samp_factor * DCT_scaled_size sample
+ * rows, or exactly min_DCT_scaled_size row groups.  (This amount of data
+ * corresponds to one row of MCUs when the image is fully interleaved.)  Note
+ * that the number of sample rows varies across components, but the number of
+ * row groups does not.  Some garbage sample rows may be included in the last
+ * iMCU row at the bottom of the image.
+ *
+ * Depending on the vertical scaling algorithm used, the upsampler may need
+ * access to the sample row(s) above and below its current input row group.
+ * The upsampler is required to set need_context_rows TRUE at global selection
+ * time if so.  When need_context_rows is FALSE, this controller can simply
+ * obtain one iMCU row at a time from the coefficient or difference controller
+ * and dole it out as row groups to the postprocessor.
+ *
+ * When need_context_rows is TRUE, this controller guarantees that the buffer
+ * passed to postprocessing contains at least one row group's worth of samples
+ * above and below the row group(s) being processed.  Note that the context
+ * rows "above" the first passed row group appear at negative row offsets in
+ * the passed buffer.  At the top and bottom of the image, the required
+ * context rows are manufactured by duplicating the first or last real sample
+ * row; this avoids having special cases in the upsampling inner loops.
+ *
+ * The amount of context is fixed at one row group just because that's a
+ * convenient number for this controller to work with.  The existing
+ * upsamplers really only need one sample row of context.  An upsampler
+ * supporting arbitrary output rescaling might wish for more than one row
+ * group of context when shrinking the image; tough, we don't handle that.
+ * (This is justified by the assumption that downsizing will be handled mostly
+ * by adjusting the DCT_scaled_size values, so that the actual scale factor at
+ * the upsample step needn't be much less than one.)
+ *
+ * To provide the desired context, we have to retain the last two row groups
+ * of one iMCU row while reading in the next iMCU row.  (The last row group
+ * can't be processed until we have another row group for its below-context,
+ * and so we have to save the next-to-last group too for its above-context.)
+ * We could do this most simply by copying data around in our buffer, but
+ * that'd be very slow.  We can avoid copying any data by creating a rather
+ * strange pointer structure.  Here's how it works.  We allocate a workspace
+ * consisting of M+2 row groups (where M = min_DCT_scaled_size is the number
+ * of row groups per iMCU row).  We create two sets of redundant pointers to
+ * the workspace.  Labeling the physical row groups 0 to M+1, the synthesized
+ * pointer lists look like this:
+ *                   M+1                          M-1
+ * master pointer --> 0         master pointer --> 0
+ *                    1                            1
+ *                   ...                          ...
+ *                   M-3                          M-3
+ *                   M-2                           M
+ *                   M-1                          M+1
+ *                    M                           M-2
+ *                   M+1                          M-1
+ *                    0                            0
+ * We read alternate iMCU rows using each master pointer; thus the last two
+ * row groups of the previous iMCU row remain un-overwritten in the workspace.
+ * The pointer lists are set up so that the required context rows appear to
+ * be adjacent to the proper places when we pass the pointer lists to the
+ * upsampler.
+ *
+ * The above pictures describe the normal state of the pointer lists.
+ * At top and bottom of the image, we diddle the pointer lists to duplicate
+ * the first or last sample row as necessary (this is cheaper than copying
+ * sample rows around).
+ *
+ * This scheme breaks down if M < 2, ie, min_DCT_scaled_size is 1.  In that
+ * situation each iMCU row provides only one row group so the buffering logic
+ * must be different (eg, we must read two iMCU rows before we can emit the
+ * first row group).  For now, we simply do not support providing context
+ * rows when min_DCT_scaled_size is 1.  That combination seems unlikely to
+ * be worth providing --- if someone wants a 1/8th-size preview, they probably
+ * want it quick and dirty, so a context-free upsampler is sufficient.
+ */
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main(j_decompress_ptr cinfo,
+                                         _JSAMPARRAY output_buf,
+                                         JDIMENSION *out_row_ctr,
+                                         JDIMENSION out_rows_avail);
+METHODDEF(void) process_data_context_main(j_decompress_ptr cinfo,
+                                          _JSAMPARRAY output_buf,
+                                          JDIMENSION *out_row_ctr,
+                                          JDIMENSION out_rows_avail);
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) process_data_crank_post(j_decompress_ptr cinfo,
+                                        _JSAMPARRAY output_buf,
+                                        JDIMENSION *out_row_ctr,
+                                        JDIMENSION out_rows_avail);
+#endif
+
+
+LOCAL(void)
+alloc_funny_pointers(j_decompress_ptr cinfo)
+/* Allocate space for the funny pointer lists.
+ * This is done only once, not once per pass.
+ */
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  int ci, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY xbuf;
+
+  /* Get top-level space for component array pointers.
+   * We alloc both arrays with one call to save a few cycles.
+   */
+  main_ptr->xbuffer[0] = (_JSAMPIMAGE)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                cinfo->num_components * 2 *
+                                sizeof(_JSAMPARRAY));
+  main_ptr->xbuffer[1] = main_ptr->xbuffer[0] + cinfo->num_components;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    /* Get space for pointer lists --- M+4 row groups in each list.
+     * We alloc both pointer lists with one call to save a few cycles.
+     */
+    xbuf = (_JSAMPARRAY)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  2 * (rgroup * (M + 4)) * sizeof(_JSAMPROW));
+    xbuf += rgroup;             /* want one row group at negative offsets */
+    main_ptr->xbuffer[0][ci] = xbuf;
+    xbuf += rgroup * (M + 4);
+    main_ptr->xbuffer[1][ci] = xbuf;
+  }
+}
+
+
+LOCAL(void)
+make_funny_pointers(j_decompress_ptr cinfo)
+/* Create the funny pointer lists discussed in the comments above.
+ * The actual workspace is already allocated (in main_ptr->buffer),
+ * and the space for the pointer lists is allocated too.
+ * This routine just fills in the curiously ordered lists.
+ * This will be repeated at the beginning of each pass.
+ */
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY buf, xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main_ptr->xbuffer[0][ci];
+    xbuf1 = main_ptr->xbuffer[1][ci];
+    /* First copy the workspace pointers as-is */
+    buf = main_ptr->buffer[ci];
+    for (i = 0; i < rgroup * (M + 2); i++) {
+      xbuf0[i] = xbuf1[i] = buf[i];
+    }
+    /* In the second list, put the last four row groups in swapped order */
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf1[rgroup * (M - 2) + i] = buf[rgroup * M + i];
+      xbuf1[rgroup * M + i] = buf[rgroup * (M - 2) + i];
+    }
+    /* The wraparound pointers at top and bottom will be filled later
+     * (see set_wraparound_pointers, below).  Initially we want the "above"
+     * pointers to duplicate the first actual data line.  This only needs
+     * to happen in xbuffer[0].
+     */
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[0];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_bottom_pointers(j_decompress_ptr cinfo)
+/* Change the pointer lists to duplicate the last sample row at the bottom
+ * of the image.  whichptr indicates which xbuffer holds the final iMCU row.
+ * Also sets rowgroups_avail to indicate number of nondummy row groups in row.
+ */
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  int ci, i, rgroup, iMCUheight, rows_left;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY xbuf;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Count sample rows in one iMCU row and in one row group */
+    iMCUheight = compptr->v_samp_factor * compptr->_DCT_scaled_size;
+    rgroup = iMCUheight / cinfo->_min_DCT_scaled_size;
+    /* Count nondummy sample rows remaining for this component */
+    rows_left = (int)(compptr->downsampled_height % (JDIMENSION)iMCUheight);
+    if (rows_left == 0) rows_left = iMCUheight;
+    /* Count nondummy row groups.  Should get same answer for each component,
+     * so we need only do it once.
+     */
+    if (ci == 0) {
+      main_ptr->rowgroups_avail = (JDIMENSION)((rows_left - 1) / rgroup + 1);
+    }
+    /* Duplicate the last real sample row rgroup*2 times; this pads out the
+     * last partial rowgroup and ensures at least one full rowgroup of context.
+     */
+    xbuf = main_ptr->xbuffer[main_ptr->whichptr][ci];
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf[rows_left + i] = xbuf[rows_left - 1];
+    }
+  }
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main(j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->upsample->need_context_rows) {
+      main_ptr->pub._process_data = process_data_context_main;
+      make_funny_pointers(cinfo); /* Create the xbuffer[] lists */
+      main_ptr->whichptr = 0;   /* Read first iMCU row into xbuffer[0] */
+      main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
+      main_ptr->iMCU_row_ctr = 0;
+    } else {
+      /* Simple case with no context needed */
+      main_ptr->pub._process_data = process_data_simple_main;
+    }
+    main_ptr->buffer_full = FALSE;      /* Mark buffer empty */
+    main_ptr->rowgroup_ctr = 0;
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_CRANK_DEST:
+    /* For last pass of 2-pass quantization, just crank the postprocessor */
+    main_ptr->pub._process_data = process_data_crank_post;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the simple case where no context is required.
+ */
+
+METHODDEF(void)
+process_data_simple_main(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
+                         JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  JDIMENSION rowgroups_avail;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (!main_ptr->buffer_full) {
+    if (!(*cinfo->coef->_decompress_data) (cinfo, main_ptr->buffer))
+      return;                   /* suspension forced, can do nothing more */
+    main_ptr->buffer_full = TRUE;       /* OK, we have an iMCU row to work with */
+  }
+
+  /* There are always min_DCT_scaled_size row groups in an iMCU row. */
+  rowgroups_avail = (JDIMENSION)cinfo->_min_DCT_scaled_size;
+  /* Note: at the bottom of the image, we may pass extra garbage row groups
+   * to the postprocessor.  The postprocessor has to check for bottom
+   * of image anyway (at row resolution), so no point in us doing it too.
+   */
+
+  /* Feed the postprocessor */
+  (*cinfo->post->_post_process_data) (cinfo, main_ptr->buffer,
+                                      &main_ptr->rowgroup_ctr, rowgroups_avail,
+                                      output_buf, out_row_ctr, out_rows_avail);
+
+  /* Has postprocessor consumed all the data yet? If so, mark buffer empty */
+  if (main_ptr->rowgroup_ctr >= rowgroups_avail) {
+    main_ptr->buffer_full = FALSE;
+    main_ptr->rowgroup_ctr = 0;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the case where context rows must be provided.
+ */
+
+METHODDEF(void)
+process_data_context_main(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
+                          JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (!main_ptr->buffer_full) {
+    if (!(*cinfo->coef->_decompress_data) (cinfo,
+                                           main_ptr->xbuffer[main_ptr->whichptr]))
+      return;                   /* suspension forced, can do nothing more */
+    main_ptr->buffer_full = TRUE;       /* OK, we have an iMCU row to work with */
+    main_ptr->iMCU_row_ctr++;   /* count rows received */
+  }
+
+  /* Postprocessor typically will not swallow all the input data it is handed
+   * in one call (due to filling the output buffer first).  Must be prepared
+   * to exit and restart.  This switch lets us keep track of how far we got.
+   * Note that each case falls through to the next on successful completion.
+   */
+  switch (main_ptr->context_state) {
+  case CTX_POSTPONED_ROW:
+    /* Call postprocessor using previously set pointers for postponed row */
+    (*cinfo->post->_post_process_data) (cinfo,
+                                        main_ptr->xbuffer[main_ptr->whichptr],
+                                        &main_ptr->rowgroup_ctr,
+                                        main_ptr->rowgroups_avail, output_buf,
+                                        out_row_ctr, out_rows_avail);
+    if (main_ptr->rowgroup_ctr < main_ptr->rowgroups_avail)
+      return;                   /* Need to suspend */
+    main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
+    if (*out_row_ctr >= out_rows_avail)
+      return;                   /* Postprocessor exactly filled output buf */
+    FALLTHROUGH                 /*FALLTHROUGH*/
+  case CTX_PREPARE_FOR_IMCU:
+    /* Prepare to process first M-1 row groups of this iMCU row */
+    main_ptr->rowgroup_ctr = 0;
+    main_ptr->rowgroups_avail = (JDIMENSION)(cinfo->_min_DCT_scaled_size - 1);
+    /* Check for bottom of image: if so, tweak pointers to "duplicate"
+     * the last sample row, and adjust rowgroups_avail to ignore padding rows.
+     */
+    if (main_ptr->iMCU_row_ctr == cinfo->total_iMCU_rows)
+      set_bottom_pointers(cinfo);
+    main_ptr->context_state = CTX_PROCESS_IMCU;
+    FALLTHROUGH                 /*FALLTHROUGH*/
+  case CTX_PROCESS_IMCU:
+    /* Call postprocessor using previously set pointers */
+    (*cinfo->post->_post_process_data) (cinfo,
+                                        main_ptr->xbuffer[main_ptr->whichptr],
+                                        &main_ptr->rowgroup_ctr,
+                                        main_ptr->rowgroups_avail, output_buf,
+                                        out_row_ctr, out_rows_avail);
+    if (main_ptr->rowgroup_ctr < main_ptr->rowgroups_avail)
+      return;                   /* Need to suspend */
+    /* After the first iMCU, change wraparound pointers to normal state */
+    if (main_ptr->iMCU_row_ctr == 1)
+      set_wraparound_pointers(cinfo);
+    /* Prepare to load new iMCU row using other xbuffer list */
+    main_ptr->whichptr ^= 1;    /* 0=>1 or 1=>0 */
+    main_ptr->buffer_full = FALSE;
+    /* Still need to process last row group of this iMCU row, */
+    /* which is saved at index M+1 of the other xbuffer */
+    main_ptr->rowgroup_ctr = (JDIMENSION)(cinfo->_min_DCT_scaled_size + 1);
+    main_ptr->rowgroups_avail = (JDIMENSION)(cinfo->_min_DCT_scaled_size + 2);
+    main_ptr->context_state = CTX_POSTPONED_ROW;
+  }
+}
+
+
+/*
+ * Process some data.
+ * Final pass of two-pass quantization: just call the postprocessor.
+ * Source data will be the postprocessor controller's internal buffer.
+ */
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+METHODDEF(void)
+process_data_crank_post(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
+                        JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  (*cinfo->post->_post_process_data) (cinfo, (_JSAMPIMAGE)NULL,
+                                      (JDIMENSION *)NULL, (JDIMENSION)0,
+                                      output_buf, out_row_ctr, out_rows_avail);
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+_jinit_d_main_controller(j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main_ptr;
+  int ci, rgroup, ngroups;
+  jpeg_component_info *compptr;
+
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  main_ptr = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_main_controller));
+  cinfo->main = (struct jpeg_d_main_controller *)main_ptr;
+  main_ptr->pub.start_pass = start_pass_main;
+
+  if (need_full_buffer)         /* shouldn't happen */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Allocate the workspace.
+   * ngroups is the number of row groups we need.
+   */
+  if (cinfo->upsample->need_context_rows) {
+    if (cinfo->_min_DCT_scaled_size < 2) /* unsupported, see comments above */
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    alloc_funny_pointers(cinfo); /* Alloc space for xbuffer[] lists */
+    ngroups = cinfo->_min_DCT_scaled_size + 2;
+  } else {
+    ngroups = cinfo->_min_DCT_scaled_size;
+  }
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    main_ptr->buffer[ci] = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+                        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                         compptr->width_in_blocks * compptr->_DCT_scaled_size,
+                         (JDIMENSION)(rgroup * ngroups));
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdmainct.h

@@ -0,0 +1,78 @@
+/*
+ * jdmainct.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_main_controller pub; /* public fields */
+
+  /* Pointer to allocated workspace (M or M+2 row groups). */
+  _JSAMPARRAY buffer[MAX_COMPONENTS];
+
+  boolean buffer_full;          /* Have we gotten an iMCU row from decoder? */
+  JDIMENSION rowgroup_ctr;      /* counts row groups output to postprocessor */
+
+  /* Remaining fields are only used in the context case. */
+
+  /* These are the master pointers to the funny-order pointer lists. */
+  _JSAMPIMAGE xbuffer[2];       /* pointers to weird pointer lists */
+
+  int whichptr;                 /* indicates which pointer set is now in use */
+  int context_state;            /* process_data state machine status */
+  JDIMENSION rowgroups_avail;   /* row groups available to postprocessor */
+  JDIMENSION iMCU_row_ctr;      /* counts iMCU rows to detect image top/bot */
+} my_main_controller;
+
+typedef my_main_controller *my_main_ptr;
+
+
+/* context_state values: */
+#define CTX_PREPARE_FOR_IMCU    0       /* need to prepare for MCU row */
+#define CTX_PROCESS_IMCU        1       /* feeding iMCU to postprocessor */
+#define CTX_POSTPONED_ROW       2       /* feeding postponed row group */
+
+
+LOCAL(void)
+set_wraparound_pointers(j_decompress_ptr cinfo)
+/* Set up the "wraparound" pointers at top and bottom of the pointer lists.
+ * This changes the pointer list state from top-of-image to the normal state.
+ */
+{
+  my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  _JSAMPARRAY xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main_ptr->xbuffer[0][ci];
+    xbuf1 = main_ptr->xbuffer[1][ci];
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[rgroup * (M + 1) + i];
+      xbuf1[i - rgroup] = xbuf1[rgroup * (M + 1) + i];
+      xbuf0[rgroup * (M + 2) + i] = xbuf0[i];
+      xbuf1[rgroup * (M + 2) + i] = xbuf1[i];
+    }
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdmarker.c

@@ -0,0 +1,1384 @@
+/*
+ * jdmarker.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2012, 2015, 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains routines to decode JPEG datastream markers.
+ * Most of the complexity arises from our desire to support input
+ * suspension: if not all of the data for a marker is available,
+ * we must exit back to the application.  On resumption, we reprocess
+ * the marker.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {                  /* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+
+  M_DHT   = 0xc4,
+
+  M_DAC   = 0xcc,
+
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+
+  M_TEM   = 0x01,
+
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_reader pub; /* public fields */
+
+  /* Application-overridable marker processing methods */
+  jpeg_marker_parser_method process_COM;
+  jpeg_marker_parser_method process_APPn[16];
+
+  /* Limit on marker data length to save for each marker type */
+  unsigned int length_limit_COM;
+  unsigned int length_limit_APPn[16];
+
+  /* Status of COM/APPn marker saving */
+  jpeg_saved_marker_ptr cur_marker;     /* NULL if not processing a marker */
+  unsigned int bytes_read;              /* data bytes read so far in marker */
+  /* Note: cur_marker is not linked into marker_list until it's all read. */
+} my_marker_reader;
+
+typedef my_marker_reader *my_marker_ptr;
+
+
+/*
+ * Macros for fetching data from the data source module.
+ *
+ * At all times, cinfo->src->next_input_byte and ->bytes_in_buffer reflect
+ * the current restart point; we update them only when we have reached a
+ * suitable place to restart if a suspension occurs.
+ */
+
+/* Declare and initialize local copies of input pointer/count */
+#define INPUT_VARS(cinfo) \
+  struct jpeg_source_mgr *datasrc = (cinfo)->src; \
+  const JOCTET *next_input_byte = datasrc->next_input_byte; \
+  size_t bytes_in_buffer = datasrc->bytes_in_buffer
+
+/* Unload the local copies --- do this only at a restart boundary */
+#define INPUT_SYNC(cinfo) \
+  ( datasrc->next_input_byte = next_input_byte, \
+    datasrc->bytes_in_buffer = bytes_in_buffer )
+
+/* Reload the local copies --- used only in MAKE_BYTE_AVAIL */
+#define INPUT_RELOAD(cinfo) \
+  ( next_input_byte = datasrc->next_input_byte, \
+    bytes_in_buffer = datasrc->bytes_in_buffer )
+
+/* Internal macro for INPUT_BYTE and INPUT_2BYTES: make a byte available.
+ * Note we do *not* do INPUT_SYNC before calling fill_input_buffer,
+ * but we must reload the local copies after a successful fill.
+ */
+#define MAKE_BYTE_AVAIL(cinfo, action) \
+  if (bytes_in_buffer == 0) { \
+    if (!(*datasrc->fill_input_buffer) (cinfo)) \
+      { action; } \
+    INPUT_RELOAD(cinfo); \
+  }
+
+/* Read a byte into variable V.
+ * If must suspend, take the specified action (typically "return FALSE").
+ */
+#define INPUT_BYTE(cinfo, V, action) \
+  MAKESTMT( MAKE_BYTE_AVAIL(cinfo, action); \
+            bytes_in_buffer--; \
+            V = *next_input_byte++; )
+
+/* As above, but read two bytes interpreted as an unsigned 16-bit integer.
+ * V should be declared unsigned int or perhaps JLONG.
+ */
+#define INPUT_2BYTES(cinfo, V, action) \
+  MAKESTMT( MAKE_BYTE_AVAIL(cinfo, action); \
+            bytes_in_buffer--; \
+            V = ((unsigned int)(*next_input_byte++)) << 8; \
+            MAKE_BYTE_AVAIL(cinfo, action); \
+            bytes_in_buffer--; \
+            V += *next_input_byte++; )
+
+
+/*
+ * Routines to process JPEG markers.
+ *
+ * Entry condition: JPEG marker itself has been read and its code saved
+ *   in cinfo->unread_marker; input restart point is just after the marker.
+ *
+ * Exit: if return TRUE, have read and processed any parameters, and have
+ *   updated the restart point to point after the parameters.
+ *   If return FALSE, was forced to suspend before reaching end of
+ *   marker parameters; restart point has not been moved.  Same routine
+ *   will be called again after application supplies more input data.
+ *
+ * This approach to suspension assumes that all of a marker's parameters
+ * can fit into a single input bufferload.  This should hold for "normal"
+ * markers.  Some COM/APPn markers might have large parameter segments
+ * that might not fit.  If we are simply dropping such a marker, we use
+ * skip_input_data to get past it, and thereby put the problem on the
+ * source manager's shoulders.  If we are saving the marker's contents
+ * into memory, we use a slightly different convention: when forced to
+ * suspend, the marker processor updates the restart point to the end of
+ * what it's consumed (ie, the end of the buffer) before returning FALSE.
+ * On resumption, cinfo->unread_marker still contains the marker code,
+ * but the data source will point to the next chunk of marker data.
+ * The marker processor must retain internal state to deal with this.
+ *
+ * Note that we don't bother to avoid duplicate trace messages if a
+ * suspension occurs within marker parameters.  Other side effects
+ * require more care.
+ */
+
+
+LOCAL(boolean)
+get_soi(j_decompress_ptr cinfo)
+/* Process an SOI marker */
+{
+  int i;
+
+  TRACEMS(cinfo, 1, JTRC_SOI);
+
+  if (cinfo->marker->saw_SOI)
+    ERREXIT(cinfo, JERR_SOI_DUPLICATE);
+
+  /* Reset all parameters that are defined to be reset by SOI */
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+  cinfo->restart_interval = 0;
+
+  /* Set initial assumptions for colorspace etc */
+
+  cinfo->jpeg_color_space = JCS_UNKNOWN;
+  cinfo->CCIR601_sampling = FALSE; /* Assume non-CCIR sampling??? */
+
+  cinfo->saw_JFIF_marker = FALSE;
+  cinfo->JFIF_major_version = 1; /* set default JFIF APP0 values */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;
+  cinfo->X_density = 1;
+  cinfo->Y_density = 1;
+  cinfo->saw_Adobe_marker = FALSE;
+  cinfo->Adobe_transform = 0;
+
+  cinfo->marker->saw_SOI = TRUE;
+
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sof(j_decompress_ptr cinfo, boolean is_prog, boolean is_lossless,
+        boolean is_arith)
+/* Process a SOFn marker */
+{
+  JLONG length;
+  int c, ci;
+  jpeg_component_info *compptr;
+  INPUT_VARS(cinfo);
+
+  if (cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOF_DUPLICATE);
+
+  cinfo->progressive_mode = is_prog;
+  cinfo->master->lossless = is_lossless;
+  cinfo->arith_code = is_arith;
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, cinfo->data_precision, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_height, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_width, return FALSE);
+  INPUT_BYTE(cinfo, cinfo->num_components, return FALSE);
+
+  length -= 8;
+
+  TRACEMS4(cinfo, 1, JTRC_SOF, cinfo->unread_marker,
+           (int)cinfo->image_width, (int)cinfo->image_height,
+           cinfo->num_components);
+
+  /* We don't support files in which the image height is initially specified */
+  /* as 0 and is later redefined by DNL.  As long as we have to check that,  */
+  /* might as well have a general sanity check. */
+  if (cinfo->image_height <= 0 || cinfo->image_width <= 0 ||
+      cinfo->num_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  if (length != (cinfo->num_components * 3))
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  if (cinfo->comp_info == NULL) /* do only once, even if suspend */
+    cinfo->comp_info = (jpeg_component_info *)(*cinfo->mem->alloc_small)
+                        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                         cinfo->num_components * sizeof(jpeg_component_info));
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->component_index = ci;
+    INPUT_BYTE(cinfo, compptr->component_id, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    compptr->h_samp_factor = (c >> 4) & 15;
+    compptr->v_samp_factor = (c     ) & 15;
+    INPUT_BYTE(cinfo, compptr->quant_tbl_no, return FALSE);
+
+    TRACEMS4(cinfo, 1, JTRC_SOF_COMPONENT,
+             compptr->component_id, compptr->h_samp_factor,
+             compptr->v_samp_factor, compptr->quant_tbl_no);
+  }
+
+  cinfo->marker->saw_SOF = TRUE;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sos(j_decompress_ptr cinfo)
+/* Process a SOS marker */
+{
+  JLONG length;
+  int i, ci, n, c, cc, pi;
+  jpeg_component_info *compptr;
+  INPUT_VARS(cinfo);
+
+  if (!cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOS_NO_SOF);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, n, return FALSE); /* Number of components */
+
+  TRACEMS1(cinfo, 1, JTRC_SOS, n);
+
+  if (length != (n * 2 + 6) || n < 1 || n > MAX_COMPS_IN_SCAN)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  cinfo->comps_in_scan = n;
+
+  /* Collect the component-spec parameters */
+
+  for (i = 0; i < MAX_COMPS_IN_SCAN; i++)
+    cinfo->cur_comp_info[i] = NULL;
+
+  for (i = 0; i < n; i++) {
+    INPUT_BYTE(cinfo, cc, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+
+    for (ci = 0, compptr = cinfo->comp_info;
+         ci < cinfo->num_components && ci < MAX_COMPS_IN_SCAN;
+         ci++, compptr++) {
+      if (cc == compptr->component_id && !cinfo->cur_comp_info[ci])
+        goto id_found;
+    }
+
+    ERREXIT1(cinfo, JERR_BAD_COMPONENT_ID, cc);
+
+id_found:
+
+    cinfo->cur_comp_info[i] = compptr;
+    compptr->dc_tbl_no = (c >> 4) & 15;
+    compptr->ac_tbl_no = (c     ) & 15;
+
+    TRACEMS3(cinfo, 1, JTRC_SOS_COMPONENT, cc,
+             compptr->dc_tbl_no, compptr->ac_tbl_no);
+
+    /* This CSi (cc) should differ from the previous CSi */
+    for (pi = 0; pi < i; pi++) {
+      if (cinfo->cur_comp_info[pi] == compptr) {
+        ERREXIT1(cinfo, JERR_BAD_COMPONENT_ID, cc);
+      }
+    }
+  }
+
+  /* Collect the additional scan parameters Ss, Se, Ah/Al. */
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ss = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Se = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ah = (c >> 4) & 15;
+  cinfo->Al = (c     ) & 15;
+
+  TRACEMS4(cinfo, 1, JTRC_SOS_PARAMS, cinfo->Ss, cinfo->Se,
+           cinfo->Ah, cinfo->Al);
+
+  /* Prepare to scan data & restart markers */
+  cinfo->marker->next_restart_num = 0;
+
+  /* Count another SOS marker */
+  cinfo->input_scan_number++;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+#ifdef D_ARITH_CODING_SUPPORTED
+
+LOCAL(boolean)
+get_dac(j_decompress_ptr cinfo)
+/* Process a DAC marker */
+{
+  JLONG length;
+  int index, val;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 0) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+    INPUT_BYTE(cinfo, val, return FALSE);
+
+    length -= 2;
+
+    TRACEMS2(cinfo, 1, JTRC_DAC, index, val);
+
+    if (index < 0 || index >= (2 * NUM_ARITH_TBLS))
+      ERREXIT1(cinfo, JERR_DAC_INDEX, index);
+
+    if (index >= NUM_ARITH_TBLS) { /* define AC table */
+      cinfo->arith_ac_K[index - NUM_ARITH_TBLS] = (UINT8)val;
+    } else {                    /* define DC table */
+      cinfo->arith_dc_L[index] = (UINT8)(val & 0x0F);
+      cinfo->arith_dc_U[index] = (UINT8)(val >> 4);
+      if (cinfo->arith_dc_L[index] > cinfo->arith_dc_U[index])
+        ERREXIT1(cinfo, JERR_DAC_VALUE, val);
+    }
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+#else /* !D_ARITH_CODING_SUPPORTED */
+
+#define get_dac(cinfo)  skip_variable(cinfo)
+
+#endif /* D_ARITH_CODING_SUPPORTED */
+
+
+LOCAL(boolean)
+get_dht(j_decompress_ptr cinfo)
+/* Process a DHT marker */
+{
+  JLONG length;
+  UINT8 bits[17];
+  UINT8 huffval[256];
+  int i, index, count;
+  JHUFF_TBL **htblptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 16) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+
+    TRACEMS1(cinfo, 1, JTRC_DHT, index);
+
+    bits[0] = 0;
+    count = 0;
+    for (i = 1; i <= 16; i++) {
+      INPUT_BYTE(cinfo, bits[i], return FALSE);
+      count += bits[i];
+    }
+
+    length -= 1 + 16;
+
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+             bits[1], bits[2], bits[3], bits[4],
+             bits[5], bits[6], bits[7], bits[8]);
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+             bits[9], bits[10], bits[11], bits[12],
+             bits[13], bits[14], bits[15], bits[16]);
+
+    /* Here we just do minimal validation of the counts to avoid walking
+     * off the end of our table space.  jdhuff.c will check more carefully.
+     */
+    if (count > 256 || ((JLONG)count) > length)
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+    for (i = 0; i < count; i++)
+      INPUT_BYTE(cinfo, huffval[i], return FALSE);
+
+    memset(&huffval[count], 0, (256 - count) * sizeof(UINT8));
+
+    length -= count;
+
+    if (index & 0x10) {         /* AC table definition */
+      index -= 0x10;
+      if (index < 0 || index >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_DHT_INDEX, index);
+      htblptr = &cinfo->ac_huff_tbl_ptrs[index];
+    } else {                    /* DC table definition */
+      if (index < 0 || index >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_DHT_INDEX, index);
+      htblptr = &cinfo->dc_huff_tbl_ptrs[index];
+    }
+
+    if (*htblptr == NULL)
+      *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo);
+
+    memcpy((*htblptr)->bits, bits, sizeof((*htblptr)->bits));
+    memcpy((*htblptr)->huffval, huffval, sizeof((*htblptr)->huffval));
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dqt(j_decompress_ptr cinfo)
+/* Process a DQT marker */
+{
+  JLONG length;
+  int n, i, prec;
+  unsigned int tmp;
+  JQUANT_TBL *quant_ptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 0) {
+    INPUT_BYTE(cinfo, n, return FALSE);
+    prec = n >> 4;
+    n &= 0x0F;
+
+    TRACEMS2(cinfo, 1, JTRC_DQT, n, prec);
+
+    if (n >= NUM_QUANT_TBLS)
+      ERREXIT1(cinfo, JERR_DQT_INDEX, n);
+
+    if (cinfo->quant_tbl_ptrs[n] == NULL)
+      cinfo->quant_tbl_ptrs[n] = jpeg_alloc_quant_table((j_common_ptr)cinfo);
+    quant_ptr = cinfo->quant_tbl_ptrs[n];
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      if (prec)
+        INPUT_2BYTES(cinfo, tmp, return FALSE);
+      else
+        INPUT_BYTE(cinfo, tmp, return FALSE);
+      /* We convert the zigzag-order table to natural array order. */
+      quant_ptr->quantval[jpeg_natural_order[i]] = (UINT16)tmp;
+    }
+
+    if (cinfo->err->trace_level >= 2) {
+      for (i = 0; i < DCTSIZE2; i += 8) {
+        TRACEMS8(cinfo, 2, JTRC_QUANTVALS,
+                 quant_ptr->quantval[i],     quant_ptr->quantval[i + 1],
+                 quant_ptr->quantval[i + 2], quant_ptr->quantval[i + 3],
+                 quant_ptr->quantval[i + 4], quant_ptr->quantval[i + 5],
+                 quant_ptr->quantval[i + 6], quant_ptr->quantval[i + 7]);
+      }
+    }
+
+    length -= DCTSIZE2 + 1;
+    if (prec) length -= DCTSIZE2;
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dri(j_decompress_ptr cinfo)
+/* Process a DRI marker */
+{
+  JLONG length;
+  unsigned int tmp;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  if (length != 4)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_2BYTES(cinfo, tmp, return FALSE);
+
+  TRACEMS1(cinfo, 1, JTRC_DRI, tmp);
+
+  cinfo->restart_interval = tmp;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Routines for processing APPn and COM markers.
+ * These are either saved in memory or discarded, per application request.
+ * APP0 and APP14 are specially checked to see if they are
+ * JFIF and Adobe markers, respectively.
+ */
+
+#define APP0_DATA_LEN   14      /* Length of interesting data in APP0 */
+#define APP14_DATA_LEN  12      /* Length of interesting data in APP14 */
+#define APPN_DATA_LEN   14      /* Must be the largest of the above!! */
+
+
+LOCAL(void)
+examine_app0(j_decompress_ptr cinfo, JOCTET *data, unsigned int datalen,
+             JLONG remaining)
+/* Examine first few bytes from an APP0.
+ * Take appropriate action if it is a JFIF marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  JLONG totallen = (JLONG)datalen + remaining;
+
+  if (datalen >= APP0_DATA_LEN &&
+      data[0] == 0x4A &&
+      data[1] == 0x46 &&
+      data[2] == 0x49 &&
+      data[3] == 0x46 &&
+      data[4] == 0) {
+    /* Found JFIF APP0 marker: save info */
+    cinfo->saw_JFIF_marker = TRUE;
+    cinfo->JFIF_major_version = data[5];
+    cinfo->JFIF_minor_version = data[6];
+    cinfo->density_unit = data[7];
+    cinfo->X_density = (data[8] << 8) + data[9];
+    cinfo->Y_density = (data[10] << 8) + data[11];
+    /* Check version.
+     * Major version must be 1, anything else signals an incompatible change.
+     * (We used to treat this as an error, but now it's a nonfatal warning,
+     * because some bozo at Hijaak couldn't read the spec.)
+     * Minor version should be 0..2, but process anyway if newer.
+     */
+    if (cinfo->JFIF_major_version != 1)
+      WARNMS2(cinfo, JWRN_JFIF_MAJOR,
+              cinfo->JFIF_major_version, cinfo->JFIF_minor_version);
+    /* Generate trace messages */
+    TRACEMS5(cinfo, 1, JTRC_JFIF,
+             cinfo->JFIF_major_version, cinfo->JFIF_minor_version,
+             cinfo->X_density, cinfo->Y_density, cinfo->density_unit);
+    /* Validate thumbnail dimensions and issue appropriate messages */
+    if (data[12] | data[13])
+      TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL, data[12], data[13]);
+    totallen -= APP0_DATA_LEN;
+    if (totallen != ((JLONG)data[12] * (JLONG)data[13] * (JLONG)3))
+      TRACEMS1(cinfo, 1, JTRC_JFIF_BADTHUMBNAILSIZE, (int)totallen);
+  } else if (datalen >= 6 &&
+             data[0] == 0x4A &&
+             data[1] == 0x46 &&
+             data[2] == 0x58 &&
+             data[3] == 0x58 &&
+             data[4] == 0) {
+    /* Found JFIF "JFXX" extension APP0 marker */
+    /* The library doesn't actually do anything with these,
+     * but we try to produce a helpful trace message.
+     */
+    switch (data[5]) {
+    case 0x10:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_JPEG, (int)totallen);
+      break;
+    case 0x11:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_PALETTE, (int)totallen);
+      break;
+    case 0x13:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_RGB, (int)totallen);
+      break;
+    default:
+      TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION, data[5], (int)totallen);
+      break;
+    }
+  } else {
+    /* Start of APP0 does not match "JFIF" or "JFXX", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP0, (int)totallen);
+  }
+}
+
+
+LOCAL(void)
+examine_app14(j_decompress_ptr cinfo, JOCTET *data, unsigned int datalen,
+              JLONG remaining)
+/* Examine first few bytes from an APP14.
+ * Take appropriate action if it is an Adobe marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  unsigned int version, flags0, flags1, transform;
+
+  if (datalen >= APP14_DATA_LEN &&
+      data[0] == 0x41 &&
+      data[1] == 0x64 &&
+      data[2] == 0x6F &&
+      data[3] == 0x62 &&
+      data[4] == 0x65) {
+    /* Found Adobe APP14 marker */
+    version = (data[5] << 8) + data[6];
+    flags0 = (data[7] << 8) + data[8];
+    flags1 = (data[9] << 8) + data[10];
+    transform = data[11];
+    TRACEMS4(cinfo, 1, JTRC_ADOBE, version, flags0, flags1, transform);
+    cinfo->saw_Adobe_marker = TRUE;
+    cinfo->Adobe_transform = (UINT8)transform;
+  } else {
+    /* Start of APP14 does not match "Adobe", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP14, (int)(datalen + remaining));
+  }
+}
+
+
+METHODDEF(boolean)
+get_interesting_appn(j_decompress_ptr cinfo)
+/* Process an APP0 or APP14 marker without saving it */
+{
+  JLONG length;
+  JOCTET b[APPN_DATA_LEN];
+  unsigned int i, numtoread;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  /* get the interesting part of the marker data */
+  if (length >= APPN_DATA_LEN)
+    numtoread = APPN_DATA_LEN;
+  else if (length > 0)
+    numtoread = (unsigned int)length;
+  else
+    numtoread = 0;
+  for (i = 0; i < numtoread; i++)
+    INPUT_BYTE(cinfo, b[i], return FALSE);
+  length -= numtoread;
+
+  /* process it */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, (JOCTET *)b, numtoread, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, (JOCTET *)b, numtoread, length);
+    break;
+  default:
+    /* can't get here unless jpeg_save_markers chooses wrong processor */
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long)length);
+
+  return TRUE;
+}
+
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+METHODDEF(boolean)
+save_marker(j_decompress_ptr cinfo)
+/* Save an APPn or COM marker into the marker list */
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+  jpeg_saved_marker_ptr cur_marker = marker->cur_marker;
+  unsigned int bytes_read, data_length;
+  JOCTET *data;
+  JLONG length = 0;
+  INPUT_VARS(cinfo);
+
+  if (cur_marker == NULL) {
+    /* begin reading a marker */
+    INPUT_2BYTES(cinfo, length, return FALSE);
+    length -= 2;
+    if (length >= 0) {          /* watch out for bogus length word */
+      /* figure out how much we want to save */
+      unsigned int limit;
+      if (cinfo->unread_marker == (int)M_COM)
+        limit = marker->length_limit_COM;
+      else
+        limit = marker->length_limit_APPn[cinfo->unread_marker - (int)M_APP0];
+      if ((unsigned int)length < limit)
+        limit = (unsigned int)length;
+      /* allocate and initialize the marker item */
+      cur_marker = (jpeg_saved_marker_ptr)
+        (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                    sizeof(struct jpeg_marker_struct) + limit);
+      cur_marker->next = NULL;
+      cur_marker->marker = (UINT8)cinfo->unread_marker;
+      cur_marker->original_length = (unsigned int)length;
+      cur_marker->data_length = limit;
+      /* data area is just beyond the jpeg_marker_struct */
+      data = cur_marker->data = (JOCTET *)(cur_marker + 1);
+      marker->cur_marker = cur_marker;
+      marker->bytes_read = 0;
+      bytes_read = 0;
+      data_length = limit;
+    } else {
+      /* deal with bogus length word */
+      bytes_read = data_length = 0;
+      data = NULL;
+    }
+  } else {
+    /* resume reading a marker */
+    bytes_read = marker->bytes_read;
+    data_length = cur_marker->data_length;
+    data = cur_marker->data + bytes_read;
+  }
+
+  while (bytes_read < data_length) {
+    INPUT_SYNC(cinfo);          /* move the restart point to here */
+    marker->bytes_read = bytes_read;
+    /* If there's not at least one byte in buffer, suspend */
+    MAKE_BYTE_AVAIL(cinfo, return FALSE);
+    /* Copy bytes with reasonable rapidity */
+    while (bytes_read < data_length && bytes_in_buffer > 0) {
+      *data++ = *next_input_byte++;
+      bytes_in_buffer--;
+      bytes_read++;
+    }
+  }
+
+  /* Done reading what we want to read */
+  if (cur_marker != NULL) {     /* will be NULL if bogus length word */
+    /* Add new marker to end of list */
+    if (cinfo->marker_list == NULL || cinfo->master->marker_list_end == NULL) {
+      cinfo->marker_list = cinfo->master->marker_list_end = cur_marker;
+    } else {
+      cinfo->master->marker_list_end->next = cur_marker;
+      cinfo->master->marker_list_end = cur_marker;
+    }
+    /* Reset pointer & calc remaining data length */
+    data = cur_marker->data;
+    length = cur_marker->original_length - data_length;
+  }
+  /* Reset to initial state for next marker */
+  marker->cur_marker = NULL;
+
+  /* Process the marker if interesting; else just make a generic trace msg */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, data, data_length, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, data, data_length, length);
+    break;
+  default:
+    TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker,
+             (int)(data_length + length));
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);            /* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long)length);
+
+  return TRUE;
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+METHODDEF(boolean)
+skip_variable(j_decompress_ptr cinfo)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  JLONG length;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker, (int)length);
+
+  INPUT_SYNC(cinfo);            /* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long)length);
+
+  return TRUE;
+}
+
+
+/*
+ * Find the next JPEG marker, save it in cinfo->unread_marker.
+ * Returns FALSE if had to suspend before reaching a marker;
+ * in that case cinfo->unread_marker is unchanged.
+ *
+ * Note that the result might not be a valid marker code,
+ * but it will never be 0 or FF.
+ */
+
+LOCAL(boolean)
+next_marker(j_decompress_ptr cinfo)
+{
+  int c;
+  INPUT_VARS(cinfo);
+
+  for (;;) {
+    INPUT_BYTE(cinfo, c, return FALSE);
+    /* Skip any non-FF bytes.
+     * This may look a bit inefficient, but it will not occur in a valid file.
+     * We sync after each discarded byte so that a suspending data source
+     * can discard the byte from its buffer.
+     */
+    while (c != 0xFF) {
+      cinfo->marker->discarded_bytes++;
+      INPUT_SYNC(cinfo);
+      INPUT_BYTE(cinfo, c, return FALSE);
+    }
+    /* This loop swallows any duplicate FF bytes.  Extra FFs are legal as
+     * pad bytes, so don't count them in discarded_bytes.  We assume there
+     * will not be so many consecutive FF bytes as to overflow a suspending
+     * data source's input buffer.
+     */
+    do {
+      INPUT_BYTE(cinfo, c, return FALSE);
+    } while (c == 0xFF);
+    if (c != 0)
+      break;                    /* found a valid marker, exit loop */
+    /* Reach here if we found a stuffed-zero data sequence (FF/00).
+     * Discard it and loop back to try again.
+     */
+    cinfo->marker->discarded_bytes += 2;
+    INPUT_SYNC(cinfo);
+  }
+
+  if (cinfo->marker->discarded_bytes != 0) {
+    WARNMS2(cinfo, JWRN_EXTRANEOUS_DATA, cinfo->marker->discarded_bytes, c);
+    cinfo->marker->discarded_bytes = 0;
+  }
+
+  cinfo->unread_marker = c;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+first_marker(j_decompress_ptr cinfo)
+/* Like next_marker, but used to obtain the initial SOI marker. */
+/* For this marker, we do not allow preceding garbage or fill; otherwise,
+ * we might well scan an entire input file before realizing it ain't JPEG.
+ * If an application wants to process non-JFIF files, it must seek to the
+ * SOI before calling the JPEG library.
+ */
+{
+  int c, c2;
+  INPUT_VARS(cinfo);
+
+  INPUT_BYTE(cinfo, c, return FALSE);
+  INPUT_BYTE(cinfo, c2, return FALSE);
+  if (c != 0xFF || c2 != (int)M_SOI)
+    ERREXIT2(cinfo, JERR_NO_SOI, c, c2);
+
+  cinfo->unread_marker = c2;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Read markers until SOS or EOI.
+ *
+ * Returns same codes as are defined for jpeg_consume_input:
+ * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ */
+
+METHODDEF(int)
+read_markers(j_decompress_ptr cinfo)
+{
+  /* Outer loop repeats once for each marker. */
+  for (;;) {
+    /* Collect the marker proper, unless we already did. */
+    /* NB: first_marker() enforces the requirement that SOI appear first. */
+    if (cinfo->unread_marker == 0) {
+      if (!cinfo->marker->saw_SOI) {
+        if (!first_marker(cinfo))
+          return JPEG_SUSPENDED;
+      } else {
+        if (!next_marker(cinfo))
+          return JPEG_SUSPENDED;
+      }
+    }
+    /* At this point cinfo->unread_marker contains the marker code and the
+     * input point is just past the marker proper, but before any parameters.
+     * A suspension will cause us to return with this state still true.
+     */
+    switch (cinfo->unread_marker) {
+    case M_SOI:
+      if (!get_soi(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF0:                /* Baseline */
+    case M_SOF1:                /* Extended sequential, Huffman */
+      if (!get_sof(cinfo, FALSE, FALSE, FALSE))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF2:                /* Progressive, Huffman */
+      if (!get_sof(cinfo, TRUE, FALSE, FALSE))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF3:                /* Lossless, Huffman */
+      if (!get_sof(cinfo, FALSE, TRUE, FALSE))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF9:                /* Extended sequential, arithmetic */
+      if (!get_sof(cinfo, FALSE, FALSE, TRUE))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF10:               /* Progressive, arithmetic */
+      if (!get_sof(cinfo, TRUE, FALSE, TRUE))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF11:               /* Lossless, arithmetic */
+      if (!get_sof(cinfo, FALSE, TRUE, TRUE))
+        return JPEG_SUSPENDED;
+      break;
+
+    /* Currently unsupported SOFn types */
+    case M_SOF5:                /* Differential sequential, Huffman */
+    case M_SOF6:                /* Differential progressive, Huffman */
+    case M_SOF7:                /* Differential lossless, Huffman */
+    case M_JPG:                 /* Reserved for JPEG extensions */
+    case M_SOF13:               /* Differential sequential, arithmetic */
+    case M_SOF14:               /* Differential progressive, arithmetic */
+    case M_SOF15:               /* Differential lossless, arithmetic */
+      ERREXIT1(cinfo, JERR_SOF_UNSUPPORTED, cinfo->unread_marker);
+      break;
+
+    case M_SOS:
+      if (!get_sos(cinfo))
+        return JPEG_SUSPENDED;
+      cinfo->unread_marker = 0; /* processed the marker */
+      return JPEG_REACHED_SOS;
+
+    case M_EOI:
+      TRACEMS(cinfo, 1, JTRC_EOI);
+      cinfo->unread_marker = 0; /* processed the marker */
+      return JPEG_REACHED_EOI;
+
+    case M_DAC:
+      if (!get_dac(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_DHT:
+      if (!get_dht(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_DQT:
+      if (!get_dqt(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_DRI:
+      if (!get_dri(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_APP0:
+    case M_APP1:
+    case M_APP2:
+    case M_APP3:
+    case M_APP4:
+    case M_APP5:
+    case M_APP6:
+    case M_APP7:
+    case M_APP8:
+    case M_APP9:
+    case M_APP10:
+    case M_APP11:
+    case M_APP12:
+    case M_APP13:
+    case M_APP14:
+    case M_APP15:
+      if (!(*((my_marker_ptr)cinfo->marker)->process_APPn[
+               cinfo->unread_marker - (int)M_APP0]) (cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_COM:
+      if (!(*((my_marker_ptr)cinfo->marker)->process_COM) (cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    case M_RST0:                /* these are all parameterless */
+    case M_RST1:
+    case M_RST2:
+    case M_RST3:
+    case M_RST4:
+    case M_RST5:
+    case M_RST6:
+    case M_RST7:
+    case M_TEM:
+      TRACEMS1(cinfo, 1, JTRC_PARMLESS_MARKER, cinfo->unread_marker);
+      break;
+
+    case M_DNL:                 /* Ignore DNL ... perhaps the wrong thing */
+      if (!skip_variable(cinfo))
+        return JPEG_SUSPENDED;
+      break;
+
+    default:                    /* must be DHP, EXP, JPGn, or RESn */
+      /* For now, we treat the reserved markers as fatal errors since they are
+       * likely to be used to signal incompatible JPEG Part 3 extensions.
+       * Once the JPEG 3 version-number marker is well defined, this code
+       * ought to change!
+       */
+      ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+      break;
+    }
+    /* Successfully processed marker, so reset state variable */
+    cinfo->unread_marker = 0;
+  } /* end loop */
+}
+
+
+/*
+ * Read a restart marker, which is expected to appear next in the datastream;
+ * if the marker is not there, take appropriate recovery action.
+ * Returns FALSE if suspension is required.
+ *
+ * This is called by the entropy decoder after it has read an appropriate
+ * number of MCUs.  cinfo->unread_marker may be nonzero if the entropy decoder
+ * has already read a marker from the data source.  Under normal conditions
+ * cinfo->unread_marker will be reset to 0 before returning; if not reset,
+ * it holds a marker which the decoder will be unable to read past.
+ */
+
+METHODDEF(boolean)
+read_restart_marker(j_decompress_ptr cinfo)
+{
+  /* Obtain a marker unless we already did. */
+  /* Note that next_marker will complain if it skips any data. */
+  if (cinfo->unread_marker == 0) {
+    if (!next_marker(cinfo))
+      return FALSE;
+  }
+
+  if (cinfo->unread_marker ==
+      ((int)M_RST0 + cinfo->marker->next_restart_num)) {
+    /* Normal case --- swallow the marker and let entropy decoder continue */
+    TRACEMS1(cinfo, 3, JTRC_RST, cinfo->marker->next_restart_num);
+    cinfo->unread_marker = 0;
+  } else {
+    /* Uh-oh, the restart markers have been messed up. */
+    /* Let the data source manager determine how to resync. */
+    if (!(*cinfo->src->resync_to_restart) (cinfo,
+                                           cinfo->marker->next_restart_num))
+      return FALSE;
+  }
+
+  /* Update next-restart state */
+  cinfo->marker->next_restart_num = (cinfo->marker->next_restart_num + 1) & 7;
+
+  return TRUE;
+}
+
+
+/*
+ * This is the default resync_to_restart method for data source managers
+ * to use if they don't have any better approach.  Some data source managers
+ * may be able to back up, or may have additional knowledge about the data
+ * which permits a more intelligent recovery strategy; such managers would
+ * presumably supply their own resync method.
+ *
+ * read_restart_marker calls resync_to_restart if it finds a marker other than
+ * the restart marker it was expecting.  (This code is *not* used unless
+ * a nonzero restart interval has been declared.)  cinfo->unread_marker is
+ * the marker code actually found (might be anything, except 0 or FF).
+ * The desired restart marker number (0..7) is passed as a parameter.
+ * This routine is supposed to apply whatever error recovery strategy seems
+ * appropriate in order to position the input stream to the next data segment.
+ * Note that cinfo->unread_marker is treated as a marker appearing before
+ * the current data-source input point; usually it should be reset to zero
+ * before returning.
+ * Returns FALSE if suspension is required.
+ *
+ * This implementation is substantially constrained by wanting to treat the
+ * input as a data stream; this means we can't back up.  Therefore, we have
+ * only the following actions to work with:
+ *   1. Simply discard the marker and let the entropy decoder resume at next
+ *      byte of file.
+ *   2. Read forward until we find another marker, discarding intervening
+ *      data.  (In theory we could look ahead within the current bufferload,
+ *      without having to discard data if we don't find the desired marker.
+ *      This idea is not implemented here, in part because it makes behavior
+ *      dependent on buffer size and chance buffer-boundary positions.)
+ *   3. Leave the marker unread (by failing to zero cinfo->unread_marker).
+ *      This will cause the entropy decoder to process an empty data segment,
+ *      inserting dummy zeroes, and then we will reprocess the marker.
+ *
+ * #2 is appropriate if we think the desired marker lies ahead, while #3 is
+ * appropriate if the found marker is a future restart marker (indicating
+ * that we have missed the desired restart marker, probably because it got
+ * corrupted).
+ * We apply #2 or #3 if the found marker is a restart marker no more than
+ * two counts behind or ahead of the expected one.  We also apply #2 if the
+ * found marker is not a legal JPEG marker code (it's certainly bogus data).
+ * If the found marker is a restart marker more than 2 counts away, we do #1
+ * (too much risk that the marker is erroneous; with luck we will be able to
+ * resync at some future point).
+ * For any valid non-restart JPEG marker, we apply #3.  This keeps us from
+ * overrunning the end of a scan.  An implementation limited to single-scan
+ * files might find it better to apply #2 for markers other than EOI, since
+ * any other marker would have to be bogus data in that case.
+ */
+
+GLOBAL(boolean)
+jpeg_resync_to_restart(j_decompress_ptr cinfo, int desired)
+{
+  int marker = cinfo->unread_marker;
+  int action = 1;
+
+  /* Always put up a warning. */
+  WARNMS2(cinfo, JWRN_MUST_RESYNC, marker, desired);
+
+  /* Outer loop handles repeated decision after scanning forward. */
+  for (;;) {
+    if (marker < (int)M_SOF0)
+      action = 2;               /* invalid marker */
+    else if (marker < (int)M_RST0 || marker > (int)M_RST7)
+      action = 3;               /* valid non-restart marker */
+    else {
+      if (marker == ((int)M_RST0 + ((desired + 1) & 7)) ||
+          marker == ((int)M_RST0 + ((desired + 2) & 7)))
+        action = 3;             /* one of the next two expected restarts */
+      else if (marker == ((int)M_RST0 + ((desired - 1) & 7)) ||
+               marker == ((int)M_RST0 + ((desired - 2) & 7)))
+        action = 2;             /* a prior restart, so advance */
+      else
+        action = 1;             /* desired restart or too far away */
+    }
+    TRACEMS2(cinfo, 4, JTRC_RECOVERY_ACTION, marker, action);
+    switch (action) {
+    case 1:
+      /* Discard marker and let entropy decoder resume processing. */
+      cinfo->unread_marker = 0;
+      return TRUE;
+    case 2:
+      /* Scan to the next marker, and repeat the decision loop. */
+      if (!next_marker(cinfo))
+        return FALSE;
+      marker = cinfo->unread_marker;
+      break;
+    case 3:
+      /* Return without advancing past this marker. */
+      /* Entropy decoder will be forced to process an empty segment. */
+      return TRUE;
+    }
+  } /* end loop */
+}
+
+
+/*
+ * Reset marker processing state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_marker_reader(j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+
+  cinfo->comp_info = NULL;              /* until allocated by get_sof */
+  cinfo->input_scan_number = 0;         /* no SOS seen yet */
+  cinfo->unread_marker = 0;             /* no pending marker */
+  marker->pub.saw_SOI = FALSE;          /* set internal state too */
+  marker->pub.saw_SOF = FALSE;
+  marker->pub.discarded_bytes = 0;
+  marker->cur_marker = NULL;
+}
+
+
+/*
+ * Initialize the marker reader module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_marker_reader(j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker;
+  int i;
+
+  /* Create subobject in permanent pool */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
+                                sizeof(my_marker_reader));
+  cinfo->marker = (struct jpeg_marker_reader *)marker;
+  /* Initialize public method pointers */
+  marker->pub.reset_marker_reader = reset_marker_reader;
+  marker->pub.read_markers = read_markers;
+  marker->pub.read_restart_marker = read_restart_marker;
+  /* Initialize COM/APPn processing.
+   * By default, we examine and then discard APP0 and APP14,
+   * but simply discard COM and all other APPn.
+   */
+  marker->process_COM = skip_variable;
+  marker->length_limit_COM = 0;
+  for (i = 0; i < 16; i++) {
+    marker->process_APPn[i] = skip_variable;
+    marker->length_limit_APPn[i] = 0;
+  }
+  marker->process_APPn[0] = get_interesting_appn;
+  marker->process_APPn[14] = get_interesting_appn;
+  /* Reset marker processing state */
+  reset_marker_reader(cinfo);
+}
+
+
+/*
+ * Control saving of COM and APPn markers into marker_list.
+ */
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+GLOBAL(void)
+jpeg_save_markers(j_decompress_ptr cinfo, int marker_code,
+                  unsigned int length_limit)
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+  long maxlength;
+  jpeg_marker_parser_method processor;
+
+  /* Length limit mustn't be larger than what we can allocate
+   * (should only be a concern in a 16-bit environment).
+   */
+  maxlength = cinfo->mem->max_alloc_chunk - sizeof(struct jpeg_marker_struct);
+  if (((long)length_limit) > maxlength)
+    length_limit = (unsigned int)maxlength;
+
+  /* Choose processor routine to use.
+   * APP0/APP14 have special requirements.
+   */
+  if (length_limit) {
+    processor = save_marker;
+    /* If saving APP0/APP14, save at least enough for our internal use. */
+    if (marker_code == (int)M_APP0 && length_limit < APP0_DATA_LEN)
+      length_limit = APP0_DATA_LEN;
+    else if (marker_code == (int)M_APP14 && length_limit < APP14_DATA_LEN)
+      length_limit = APP14_DATA_LEN;
+  } else {
+    processor = skip_variable;
+    /* If discarding APP0/APP14, use our regular on-the-fly processor. */
+    if (marker_code == (int)M_APP0 || marker_code == (int)M_APP14)
+      processor = get_interesting_appn;
+  }
+
+  if (marker_code == (int)M_COM) {
+    marker->process_COM = processor;
+    marker->length_limit_COM = length_limit;
+  } else if (marker_code >= (int)M_APP0 && marker_code <= (int)M_APP15) {
+    marker->process_APPn[marker_code - (int)M_APP0] = processor;
+    marker->length_limit_APPn[marker_code - (int)M_APP0] = length_limit;
+  } else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+/*
+ * Install a special processing method for COM or APPn markers.
+ */
+
+GLOBAL(void)
+jpeg_set_marker_processor(j_decompress_ptr cinfo, int marker_code,
+                          jpeg_marker_parser_method routine)
+{
+  my_marker_ptr marker = (my_marker_ptr)cinfo->marker;
+
+  if (marker_code == (int)M_COM)
+    marker->process_COM = routine;
+  else if (marker_code >= (int)M_APP0 && marker_code <= (int)M_APP15)
+    marker->process_APPn[marker_code - (int)M_APP0] = routine;
+  else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}

thirdparty/libjpeg-turbo/src/jdmaster.c

@@ -0,0 +1,893 @@
+/*
+ * jdmaster.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2002-2009 by Guido Vollbeding.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009-2011, 2016, 2019, 2022-2024, D. R. Commander.
+ * Copyright (C) 2013, Linaro Limited.
+ * Copyright (C) 2015, Google, Inc.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains master control logic for the JPEG decompressor.
+ * These routines are concerned with selecting the modules to be executed
+ * and with determining the number of passes and the work to be done in each
+ * pass.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+#include "jdmaster.h"
+
+
+/*
+ * Determine whether merged upsample/color conversion should be used.
+ * CRUCIAL: this must match the actual capabilities of jdmerge.c!
+ */
+
+LOCAL(boolean)
+use_merged_upsample(j_decompress_ptr cinfo)
+{
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  /* Colorspace conversion is not supported with lossless JPEG images */
+  if (cinfo->master->lossless)
+    return FALSE;
+  /* Merging is the equivalent of plain box-filter upsampling */
+  if (cinfo->do_fancy_upsampling || cinfo->CCIR601_sampling)
+    return FALSE;
+  /* jdmerge.c only supports YCC=>RGB and YCC=>RGB565 color conversion */
+  if (cinfo->jpeg_color_space != JCS_YCbCr || cinfo->num_components != 3 ||
+      (cinfo->out_color_space != JCS_RGB &&
+       cinfo->out_color_space != JCS_RGB565 &&
+       cinfo->out_color_space != JCS_EXT_RGB &&
+       cinfo->out_color_space != JCS_EXT_RGBX &&
+       cinfo->out_color_space != JCS_EXT_BGR &&
+       cinfo->out_color_space != JCS_EXT_BGRX &&
+       cinfo->out_color_space != JCS_EXT_XBGR &&
+       cinfo->out_color_space != JCS_EXT_XRGB &&
+       cinfo->out_color_space != JCS_EXT_RGBA &&
+       cinfo->out_color_space != JCS_EXT_BGRA &&
+       cinfo->out_color_space != JCS_EXT_ABGR &&
+       cinfo->out_color_space != JCS_EXT_ARGB))
+    return FALSE;
+  if ((cinfo->out_color_space == JCS_RGB565 &&
+       cinfo->out_color_components != 3) ||
+      (cinfo->out_color_space != JCS_RGB565 &&
+       cinfo->out_color_components != rgb_pixelsize[cinfo->out_color_space]))
+    return FALSE;
+  /* and it only handles 2h1v or 2h2v sampling ratios */
+  if (cinfo->comp_info[0].h_samp_factor != 2 ||
+      cinfo->comp_info[1].h_samp_factor != 1 ||
+      cinfo->comp_info[2].h_samp_factor != 1 ||
+      cinfo->comp_info[0].v_samp_factor >  2 ||
+      cinfo->comp_info[1].v_samp_factor != 1 ||
+      cinfo->comp_info[2].v_samp_factor != 1)
+    return FALSE;
+  /* furthermore, it doesn't work if we've scaled the IDCTs differently */
+  if (cinfo->comp_info[0]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
+      cinfo->comp_info[1]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
+      cinfo->comp_info[2]._DCT_scaled_size != cinfo->_min_DCT_scaled_size)
+    return FALSE;
+  /* ??? also need to test for upsample-time rescaling, when & if supported */
+  return TRUE;                  /* by golly, it'll work... */
+#else
+  return FALSE;
+#endif
+}
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+#if JPEG_LIB_VERSION >= 80
+GLOBAL(void)
+#else
+LOCAL(void)
+#endif
+jpeg_core_output_dimensions(j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase.
+ * This function is used for transcoding and full decompression.
+ */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+
+  if (!cinfo->master->lossless) {
+    /* Compute actual output image dimensions and DCT scaling choices. */
+    if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom) {
+      /* Provide 1/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 1;
+      cinfo->_min_DCT_v_scaled_size = 1;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 2) {
+      /* Provide 2/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 2L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 2L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 2;
+      cinfo->_min_DCT_v_scaled_size = 2;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 3) {
+      /* Provide 3/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 3L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 3L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 3;
+      cinfo->_min_DCT_v_scaled_size = 3;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 4) {
+      /* Provide 4/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 4L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 4L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 4;
+      cinfo->_min_DCT_v_scaled_size = 4;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 5) {
+      /* Provide 5/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 5L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 5L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 5;
+      cinfo->_min_DCT_v_scaled_size = 5;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 6) {
+      /* Provide 6/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 6L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 6L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 6;
+      cinfo->_min_DCT_v_scaled_size = 6;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 7) {
+      /* Provide 7/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 7L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 7L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 7;
+      cinfo->_min_DCT_v_scaled_size = 7;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 8) {
+      /* Provide 8/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 8L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 8L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 8;
+      cinfo->_min_DCT_v_scaled_size = 8;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 9) {
+      /* Provide 9/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 9L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 9L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 9;
+      cinfo->_min_DCT_v_scaled_size = 9;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 10) {
+      /* Provide 10/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 10L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 10L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 10;
+      cinfo->_min_DCT_v_scaled_size = 10;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 11) {
+      /* Provide 11/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 11L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 11L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 11;
+      cinfo->_min_DCT_v_scaled_size = 11;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 12) {
+      /* Provide 12/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 12L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 12L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 12;
+      cinfo->_min_DCT_v_scaled_size = 12;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 13) {
+      /* Provide 13/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 13L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 13L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 13;
+      cinfo->_min_DCT_v_scaled_size = 13;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 14) {
+      /* Provide 14/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 14L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 14L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 14;
+      cinfo->_min_DCT_v_scaled_size = 14;
+    } else if (cinfo->scale_num * DCTSIZE <= cinfo->scale_denom * 15) {
+      /* Provide 15/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 15L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 15L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 15;
+      cinfo->_min_DCT_v_scaled_size = 15;
+    } else {
+      /* Provide 16/block_size scaling */
+      cinfo->output_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width * 16L, (long)DCTSIZE);
+      cinfo->output_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height * 16L, (long)DCTSIZE);
+      cinfo->_min_DCT_h_scaled_size = 16;
+      cinfo->_min_DCT_v_scaled_size = 16;
+    }
+
+    /* Recompute dimensions of components */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      compptr->_DCT_h_scaled_size = cinfo->_min_DCT_h_scaled_size;
+      compptr->_DCT_v_scaled_size = cinfo->_min_DCT_v_scaled_size;
+    }
+  } else
+#endif /* !IDCT_SCALING_SUPPORTED */
+  {
+    /* Hardwire it to "no scaling" */
+    cinfo->output_width = cinfo->image_width;
+    cinfo->output_height = cinfo->image_height;
+    /* jdinput.c has already initialized DCT_scaled_size,
+     * and has computed unscaled downsampled_width and downsampled_height.
+     */
+  }
+}
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ * Also note that it may be called before the master module is initialized!
+ */
+
+GLOBAL(void)
+jpeg_calc_output_dimensions(j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+#endif
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_READY)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Compute core output image dimensions and DCT scaling choices. */
+  jpeg_core_output_dimensions(cinfo);
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+  if (!cinfo->master->lossless) {
+    /* In selecting the actual DCT scaling for each component, we try to
+     * scale up the chroma components via IDCT scaling rather than upsampling.
+     * This saves time if the upsampler gets to use 1:1 scaling.
+     * Note this code adapts subsampling ratios which are powers of 2.
+     */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      int ssize = cinfo->_min_DCT_scaled_size;
+      while (ssize < DCTSIZE &&
+             ((cinfo->max_h_samp_factor * cinfo->_min_DCT_scaled_size) %
+              (compptr->h_samp_factor * ssize * 2) == 0) &&
+             ((cinfo->max_v_samp_factor * cinfo->_min_DCT_scaled_size) %
+              (compptr->v_samp_factor * ssize * 2) == 0)) {
+        ssize = ssize * 2;
+      }
+#if JPEG_LIB_VERSION >= 70
+      compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = ssize;
+#else
+      compptr->DCT_scaled_size = ssize;
+#endif
+    }
+
+    /* Recompute downsampled dimensions of components;
+     * application needs to know these if using raw downsampled data.
+     */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      /* Size in samples, after IDCT scaling */
+      compptr->downsampled_width = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_width *
+                      (long)(compptr->h_samp_factor *
+                             compptr->_DCT_scaled_size),
+                      (long)(cinfo->max_h_samp_factor * DCTSIZE));
+      compptr->downsampled_height = (JDIMENSION)
+        jdiv_round_up((long)cinfo->image_height *
+                      (long)(compptr->v_samp_factor *
+                             compptr->_DCT_scaled_size),
+                      (long)(cinfo->max_v_samp_factor * DCTSIZE));
+    }
+  } else
+#endif /* IDCT_SCALING_SUPPORTED */
+  {
+    /* Hardwire it to "no scaling" */
+    cinfo->output_width = cinfo->image_width;
+    cinfo->output_height = cinfo->image_height;
+    /* jdinput.c has already initialized DCT_scaled_size to DCTSIZE,
+     * and has computed unscaled downsampled_width and downsampled_height.
+     */
+  }
+
+  /* Report number of components in selected colorspace. */
+  /* Probably this should be in the color conversion module... */
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    break;
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+  case JCS_EXT_RGBA:
+  case JCS_EXT_BGRA:
+  case JCS_EXT_ABGR:
+  case JCS_EXT_ARGB:
+    cinfo->out_color_components = rgb_pixelsize[cinfo->out_color_space];
+    break;
+  case JCS_YCbCr:
+  case JCS_RGB565:
+    cinfo->out_color_components = 3;
+    break;
+  case JCS_CMYK:
+  case JCS_YCCK:
+    cinfo->out_color_components = 4;
+    break;
+  default:                      /* else must be same colorspace as in file */
+    cinfo->out_color_components = cinfo->num_components;
+    break;
+  }
+  cinfo->output_components = (cinfo->quantize_colors ? 1 :
+                              cinfo->out_color_components);
+
+  /* See if upsampler will want to emit more than one row at a time */
+  if (use_merged_upsample(cinfo))
+    cinfo->rec_outbuf_height = cinfo->max_v_samp_factor;
+  else
+    cinfo->rec_outbuf_height = 1;
+}
+
+
+/*
+ * Several decompression processes need to range-limit values to the range
+ * 0..MAXJSAMPLE; the input value may fall somewhat outside this range
+ * due to noise introduced by quantization, roundoff error, etc.  These
+ * processes are inner loops and need to be as fast as possible.  On most
+ * machines, particularly CPUs with pipelines or instruction prefetch,
+ * a (subscript-check-less) C table lookup
+ *              x = sample_range_limit[x];
+ * is faster than explicit tests
+ *              if (x < 0)  x = 0;
+ *              else if (x > MAXJSAMPLE)  x = MAXJSAMPLE;
+ * These processes all use a common table prepared by the routine below.
+ *
+ * For most steps we can mathematically guarantee that the initial value
+ * of x is within MAXJSAMPLE+1 of the legal range, so a table running from
+ * -(MAXJSAMPLE+1) to 2*MAXJSAMPLE+1 is sufficient.  But for the initial
+ * limiting step (just after the IDCT), a wildly out-of-range value is
+ * possible if the input data is corrupt.  To avoid any chance of indexing
+ * off the end of memory and getting a bad-pointer trap, we perform the
+ * post-IDCT limiting thus:
+ *              x = range_limit[x & MASK];
+ * where MASK is 2 bits wider than legal sample data, ie 10 bits for 8-bit
+ * samples.  Under normal circumstances this is more than enough range and
+ * a correct output will be generated; with bogus input data the mask will
+ * cause wraparound, and we will safely generate a bogus-but-in-range output.
+ * For the post-IDCT step, we want to convert the data from signed to unsigned
+ * representation by adding CENTERJSAMPLE at the same time that we limit it.
+ * So the post-IDCT limiting table ends up looking like this:
+ *   CENTERJSAMPLE,CENTERJSAMPLE+1,...,MAXJSAMPLE,
+ *   MAXJSAMPLE (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0          (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0,1,...,CENTERJSAMPLE-1
+ * Negative inputs select values from the upper half of the table after
+ * masking.
+ *
+ * We can save some space by overlapping the start of the post-IDCT table
+ * with the simpler range limiting table.  The post-IDCT table begins at
+ * sample_range_limit + CENTERJSAMPLE.
+ */
+
+LOCAL(void)
+prepare_range_limit_table(j_decompress_ptr cinfo)
+/* Allocate and fill in the sample_range_limit table */
+{
+  JSAMPLE *table;
+  J12SAMPLE *table12;
+#ifdef D_LOSSLESS_SUPPORTED
+  J16SAMPLE *table16;
+#endif
+  int i;
+
+  if (cinfo->data_precision <= 8) {
+    table = (JSAMPLE *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                  (5 * (MAXJSAMPLE + 1) + CENTERJSAMPLE) * sizeof(JSAMPLE));
+    table += (MAXJSAMPLE + 1);  /* allow negative subscripts of simple table */
+    cinfo->sample_range_limit = table;
+    /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+    memset(table - (MAXJSAMPLE + 1), 0, (MAXJSAMPLE + 1) * sizeof(JSAMPLE));
+    /* Main part of "simple" table: limit[x] = x */
+    for (i = 0; i <= MAXJSAMPLE; i++)
+      table[i] = (JSAMPLE)i;
+    table += CENTERJSAMPLE;     /* Point to where post-IDCT table starts */
+    /* End of simple table, rest of first half of post-IDCT table */
+    for (i = CENTERJSAMPLE; i < 2 * (MAXJSAMPLE + 1); i++)
+      table[i] = MAXJSAMPLE;
+    /* Second half of post-IDCT table */
+    memset(table + (2 * (MAXJSAMPLE + 1)), 0,
+           (2 * (MAXJSAMPLE + 1) - CENTERJSAMPLE) * sizeof(JSAMPLE));
+    memcpy(table + (4 * (MAXJSAMPLE + 1) - CENTERJSAMPLE),
+           cinfo->sample_range_limit, CENTERJSAMPLE * sizeof(JSAMPLE));
+  } else if (cinfo->data_precision <= 12) {
+    table12 = (J12SAMPLE *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                  (5 * (MAXJ12SAMPLE + 1) + CENTERJ12SAMPLE) *
+                  sizeof(J12SAMPLE));
+    table12 += (MAXJ12SAMPLE + 1);  /* allow negative subscripts of simple
+                                       table */
+    cinfo->sample_range_limit = (JSAMPLE *)table12;
+    /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+    memset(table12 - (MAXJ12SAMPLE + 1), 0,
+           (MAXJ12SAMPLE + 1) * sizeof(J12SAMPLE));
+    /* Main part of "simple" table: limit[x] = x */
+    for (i = 0; i <= MAXJ12SAMPLE; i++)
+      table12[i] = (J12SAMPLE)i;
+    table12 += CENTERJ12SAMPLE; /* Point to where post-IDCT table starts */
+    /* End of simple table, rest of first half of post-IDCT table */
+    for (i = CENTERJ12SAMPLE; i < 2 * (MAXJ12SAMPLE + 1); i++)
+      table12[i] = MAXJ12SAMPLE;
+    /* Second half of post-IDCT table */
+    memset(table12 + (2 * (MAXJ12SAMPLE + 1)), 0,
+           (2 * (MAXJ12SAMPLE + 1) - CENTERJ12SAMPLE) * sizeof(J12SAMPLE));
+    memcpy(table12 + (4 * (MAXJ12SAMPLE + 1) - CENTERJ12SAMPLE),
+           cinfo->sample_range_limit, CENTERJ12SAMPLE * sizeof(J12SAMPLE));
+  } else {
+#ifdef D_LOSSLESS_SUPPORTED
+    table16 = (J16SAMPLE *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                  (5 * (MAXJ16SAMPLE + 1) + CENTERJ16SAMPLE) *
+                  sizeof(J16SAMPLE));
+    table16 += (MAXJ16SAMPLE + 1);  /* allow negative subscripts of simple
+                                       table */
+    cinfo->sample_range_limit = (JSAMPLE *)table16;
+    /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+    memset(table16 - (MAXJ16SAMPLE + 1), 0,
+           (MAXJ16SAMPLE + 1) * sizeof(J16SAMPLE));
+    /* Main part of "simple" table: limit[x] = x */
+    for (i = 0; i <= MAXJ16SAMPLE; i++)
+      table16[i] = (J16SAMPLE)i;
+    table16 += CENTERJ16SAMPLE; /* Point to where post-IDCT table starts */
+    /* End of simple table, rest of first half of post-IDCT table */
+    for (i = CENTERJ16SAMPLE; i < 2 * (MAXJ16SAMPLE + 1); i++)
+      table16[i] = MAXJ16SAMPLE;
+    /* Second half of post-IDCT table */
+    memset(table16 + (2 * (MAXJ16SAMPLE + 1)), 0,
+           (2 * (MAXJ16SAMPLE + 1) - CENTERJ16SAMPLE) * sizeof(J16SAMPLE));
+    memcpy(table16 + (4 * (MAXJ16SAMPLE + 1) - CENTERJ16SAMPLE),
+           cinfo->sample_range_limit, CENTERJ16SAMPLE * sizeof(J16SAMPLE));
+#else
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+  }
+}
+
+
+/*
+ * Master selection of decompression modules.
+ * This is done once at jpeg_start_decompress time.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ * We also initialize the decompressor input side to begin consuming data.
+ *
+ * Since jpeg_read_header has finished, we know what is in the SOF
+ * and (first) SOS markers.  We also have all the application parameter
+ * settings.
+ */
+
+LOCAL(void)
+master_selection(j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+  boolean use_c_buffer;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  /* Disable IDCT scaling and raw (downsampled) data output in lossless mode.
+   * IDCT scaling is not useful in lossless mode, and it must be disabled in
+   * order to properly calculate the output dimensions.  Raw data output isn't
+   * particularly useful without subsampling and has not been tested in
+   * lossless mode.
+   */
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+    cinfo->raw_data_out = FALSE;
+    cinfo->scale_num = cinfo->scale_denom = 1;
+  }
+#endif
+
+  /* Initialize dimensions and other stuff */
+  jpeg_calc_output_dimensions(cinfo);
+  prepare_range_limit_table(cinfo);
+
+  /* Width of an output scanline must be representable as JDIMENSION. */
+  samplesperrow = (long)cinfo->output_width *
+                  (long)cinfo->out_color_components;
+  jd_samplesperrow = (JDIMENSION)samplesperrow;
+  if ((long)jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* Initialize my private state */
+  master->pass_number = 0;
+  master->using_merged_upsample = use_merged_upsample(cinfo);
+
+  /* Color quantizer selection */
+  master->quantizer_1pass = NULL;
+  master->quantizer_2pass = NULL;
+  /* No mode changes if not using buffered-image mode. */
+  if (!cinfo->quantize_colors || !cinfo->buffered_image) {
+    cinfo->enable_1pass_quant = FALSE;
+    cinfo->enable_external_quant = FALSE;
+    cinfo->enable_2pass_quant = FALSE;
+  }
+  if (cinfo->quantize_colors) {
+    if (cinfo->raw_data_out)
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    /* 2-pass quantizer only works in 3-component color space. */
+    if (cinfo->out_color_components != 3 ||
+        cinfo->out_color_space == JCS_RGB565) {
+      cinfo->enable_1pass_quant = TRUE;
+      cinfo->enable_external_quant = FALSE;
+      cinfo->enable_2pass_quant = FALSE;
+      cinfo->colormap = NULL;
+    } else if (cinfo->colormap != NULL) {
+      cinfo->enable_external_quant = TRUE;
+    } else if (cinfo->two_pass_quantize) {
+      cinfo->enable_2pass_quant = TRUE;
+    } else {
+      cinfo->enable_1pass_quant = TRUE;
+    }
+
+    if (cinfo->enable_1pass_quant) {
+#ifdef QUANT_1PASS_SUPPORTED
+      if (cinfo->data_precision == 8)
+        jinit_1pass_quantizer(cinfo);
+      else if (cinfo->data_precision == 12)
+        j12init_1pass_quantizer(cinfo);
+      else
+        ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+      master->quantizer_1pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+
+    /* We use the 2-pass code to map to external colormaps. */
+    if (cinfo->enable_2pass_quant || cinfo->enable_external_quant) {
+#ifdef QUANT_2PASS_SUPPORTED
+      if (cinfo->data_precision == 8)
+        jinit_2pass_quantizer(cinfo);
+      else if (cinfo->data_precision == 12)
+        j12init_2pass_quantizer(cinfo);
+      else
+        ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+      master->quantizer_2pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+    /* If both quantizers are initialized, the 2-pass one is left active;
+     * this is necessary for starting with quantization to an external map.
+     */
+  }
+
+  /* Post-processing: in particular, color conversion first */
+  if (!cinfo->raw_data_out) {
+    if (master->using_merged_upsample) {
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+      if (cinfo->data_precision == 8)
+        jinit_merged_upsampler(cinfo); /* does color conversion too */
+      else if (cinfo->data_precision == 12)
+        j12init_merged_upsampler(cinfo); /* does color conversion too */
+      else
+        ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else {
+      if (cinfo->data_precision <= 8) {
+        jinit_color_deconverter(cinfo);
+        jinit_upsampler(cinfo);
+      } else if (cinfo->data_precision <= 12) {
+        j12init_color_deconverter(cinfo);
+        j12init_upsampler(cinfo);
+      } else {
+#ifdef D_LOSSLESS_SUPPORTED
+        j16init_color_deconverter(cinfo);
+        j16init_upsampler(cinfo);
+#else
+        ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+      }
+    }
+    if (cinfo->data_precision <= 8)
+      jinit_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+    else if (cinfo->data_precision <= 12)
+      j12init_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+    else
+#ifdef D_LOSSLESS_SUPPORTED
+      j16init_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+#else
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+  }
+
+  if (cinfo->master->lossless) {
+#ifdef D_LOSSLESS_SUPPORTED
+    /* Prediction, sample undifferencing, point transform, and sample size
+     * scaling
+     */
+    if (cinfo->data_precision <= 8)
+      jinit_lossless_decompressor(cinfo);
+    else if (cinfo->data_precision <= 12)
+      j12init_lossless_decompressor(cinfo);
+    else
+      j16init_lossless_decompressor(cinfo);
+    /* Entropy decoding: either Huffman or arithmetic coding. */
+    if (cinfo->arith_code) {
+      ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+    } else {
+      jinit_lhuff_decoder(cinfo);
+    }
+
+    /* Initialize principal buffer controllers. */
+    use_c_buffer = cinfo->inputctl->has_multiple_scans ||
+                   cinfo->buffered_image;
+    if (cinfo->data_precision <= 8)
+      jinit_d_diff_controller(cinfo, use_c_buffer);
+    else if (cinfo->data_precision <= 12)
+      j12init_d_diff_controller(cinfo, use_c_buffer);
+    else
+      j16init_d_diff_controller(cinfo, use_c_buffer);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* Inverse DCT */
+    if (cinfo->data_precision == 8)
+      jinit_inverse_dct(cinfo);
+    else if (cinfo->data_precision == 12)
+      j12init_inverse_dct(cinfo);
+    else
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+    /* Entropy decoding: either Huffman or arithmetic coding. */
+    if (cinfo->arith_code) {
+#ifdef D_ARITH_CODING_SUPPORTED
+      jinit_arith_decoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+#endif
+    } else {
+      if (cinfo->progressive_mode) {
+#ifdef D_PROGRESSIVE_SUPPORTED
+        jinit_phuff_decoder(cinfo);
+#else
+        ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+      } else
+        jinit_huff_decoder(cinfo);
+    }
+
+    /* Initialize principal buffer controllers. */
+    use_c_buffer = cinfo->inputctl->has_multiple_scans ||
+                   cinfo->buffered_image;
+    if (cinfo->data_precision == 12)
+      j12init_d_coef_controller(cinfo, use_c_buffer);
+    else
+      jinit_d_coef_controller(cinfo, use_c_buffer);
+  }
+
+  if (!cinfo->raw_data_out) {
+    if (cinfo->data_precision <= 8)
+      jinit_d_main_controller(cinfo, FALSE /* never need full buffer here */);
+    else if (cinfo->data_precision <= 12)
+      j12init_d_main_controller(cinfo,
+                                FALSE /* never need full buffer here */);
+    else
+#ifdef D_LOSSLESS_SUPPORTED
+      j16init_d_main_controller(cinfo,
+                                FALSE /* never need full buffer here */);
+#else
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+#endif
+  }
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr)cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+  /* Set the first and last iMCU columns to decompress from single-scan images.
+   * By default, decompress all of the iMCU columns.
+   */
+  cinfo->master->first_iMCU_col = 0;
+  cinfo->master->last_iMCU_col = cinfo->MCUs_per_row - 1;
+  cinfo->master->last_good_iMCU_row = 0;
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* If jpeg_start_decompress will read the whole file, initialize
+   * progress monitoring appropriately.  The input step is counted
+   * as one pass.
+   */
+  if (cinfo->progress != NULL && !cinfo->buffered_image &&
+      cinfo->inputctl->has_multiple_scans) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long)cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = (cinfo->enable_2pass_quant ? 3 : 2);
+    /* Count the input pass as done */
+    master->pass_number++;
+  }
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each output pass.  We determine which
+ * modules will be active during this pass and give them appropriate
+ * start_pass calls.  We also set is_dummy_pass to indicate whether this
+ * is a "real" output pass or a dummy pass for color quantization.
+ * (In the latter case, jdapistd.c will crank the pass to completion.)
+ */
+
+METHODDEF(void)
+prepare_for_output_pass(j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  if (master->pub.is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Final pass of 2-pass quantization */
+    master->pub.is_dummy_pass = FALSE;
+    (*cinfo->cquantize->start_pass) (cinfo, FALSE);
+    (*cinfo->post->start_pass) (cinfo, JBUF_CRANK_DEST);
+    (*cinfo->main->start_pass) (cinfo, JBUF_CRANK_DEST);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  } else {
+    if (cinfo->quantize_colors && cinfo->colormap == NULL) {
+      /* Select new quantization method */
+      if (cinfo->two_pass_quantize && cinfo->enable_2pass_quant) {
+        cinfo->cquantize = master->quantizer_2pass;
+        master->pub.is_dummy_pass = TRUE;
+      } else if (cinfo->enable_1pass_quant) {
+        cinfo->cquantize = master->quantizer_1pass;
+      } else {
+        ERREXIT(cinfo, JERR_MODE_CHANGE);
+      }
+    }
+    (*cinfo->idct->start_pass) (cinfo);
+    (*cinfo->coef->start_output_pass) (cinfo);
+    if (!cinfo->raw_data_out) {
+      if (!master->using_merged_upsample)
+        (*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->upsample->start_pass) (cinfo);
+      if (cinfo->quantize_colors)
+        (*cinfo->cquantize->start_pass) (cinfo, master->pub.is_dummy_pass);
+      (*cinfo->post->start_pass) (cinfo,
+            (master->pub.is_dummy_pass ? JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+      (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+  }
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->pass_number +
+                                    (master->pub.is_dummy_pass ? 2 : 1);
+    /* In buffered-image mode, we assume one more output pass if EOI not
+     * yet reached, but no more passes if EOI has been reached.
+     */
+    if (cinfo->buffered_image && !cinfo->inputctl->eoi_reached) {
+      cinfo->progress->total_passes += (cinfo->enable_2pass_quant ? 2 : 1);
+    }
+  }
+}
+
+
+/*
+ * Finish up at end of an output pass.
+ */
+
+METHODDEF(void)
+finish_output_pass(j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  if (cinfo->quantize_colors)
+    (*cinfo->cquantize->finish_pass) (cinfo);
+  master->pass_number++;
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+GLOBAL(void)
+jpeg_new_colormap(j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_BUFIMAGE)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (cinfo->quantize_colors && cinfo->enable_external_quant &&
+      cinfo->colormap != NULL) {
+    /* Select 2-pass quantizer for external colormap use */
+    cinfo->cquantize = master->quantizer_2pass;
+    /* Notify quantizer of colormap change */
+    (*cinfo->cquantize->new_color_map) (cinfo);
+    master->pub.is_dummy_pass = FALSE; /* just in case */
+  } else
+    ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+/*
+ * Initialize master decompression control and select active modules.
+ * This is performed at the start of jpeg_start_decompress.
+ */
+
+GLOBAL(void)
+jinit_master_decompress(j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr)cinfo->master;
+
+  master->pub.prepare_for_output_pass = prepare_for_output_pass;
+  master->pub.finish_output_pass = finish_output_pass;
+
+  master->pub.is_dummy_pass = FALSE;
+  master->pub.jinit_upsampler_no_alloc = FALSE;
+
+  master_selection(cinfo);
+}

thirdparty/libjpeg-turbo/src/jdmaster.h

@@ -0,0 +1,28 @@
+/*
+ * jdmaster.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1995, Thomas G. Lane.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the master control structure for the JPEG decompressor.
+ */
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_decomp_master pub; /* public fields */
+
+  int pass_number;              /* # of passes completed */
+
+  boolean using_merged_upsample; /* TRUE if using merged upsample/cconvert */
+
+  /* Saved references to initialized quantizer modules,
+   * in case we need to switch modes.
+   */
+  struct jpeg_color_quantizer *quantizer_1pass;
+  struct jpeg_color_quantizer *quantizer_2pass;
+} my_decomp_master;
+
+typedef my_decomp_master *my_master_ptr;

thirdparty/libjpeg-turbo/src/jdmerge.c

@@ -0,0 +1,594 @@
+/*
+ * jdmerge.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2009, 2011, 2014-2015, 2020, 2022, D. R. Commander.
+ * Copyright (C) 2013, Linaro Limited.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ *
+ * This file combines functions from jdsample.c and jdcolor.c;
+ * read those files first to understand what's going on.
+ *
+ * When the chroma components are to be upsampled by simple replication
+ * (ie, box filtering), we can save some work in color conversion by
+ * calculating all the output pixels corresponding to a pair of chroma
+ * samples at one time.  In the conversion equations
+ *      R = Y           + K1 * Cr
+ *      G = Y + K2 * Cb + K3 * Cr
+ *      B = Y + K4 * Cb
+ * only the Y term varies among the group of pixels corresponding to a pair
+ * of chroma samples, so the rest of the terms can be calculated just once.
+ * At typical sampling ratios, this eliminates half or three-quarters of the
+ * multiplications needed for color conversion.
+ *
+ * This file currently provides implementations for the following cases:
+ *      YCbCr => RGB color conversion only.
+ *      Sampling ratios of 2h1v or 2h2v.
+ *      No scaling needed at upsample time.
+ *      Corner-aligned (non-CCIR601) sampling alignment.
+ * Other special cases could be added, but in most applications these are
+ * the only common cases.  (For uncommon cases we fall back on the more
+ * general code in jdsample.c and jdcolor.c.)
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdmerge.h"
+#include "jsimd.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+#define SCALEBITS       16      /* speediest right-shift on some machines */
+#define ONE_HALF        ((JLONG)1 << (SCALEBITS - 1))
+#define FIX(x)          ((JLONG)((x) * (1L << SCALEBITS) + 0.5))
+
+
+/* Include inline routines for colorspace extensions */
+
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define h2v1_merged_upsample_internal  extrgb_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extrgb_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define h2v1_merged_upsample_internal  extrgbx_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extrgbx_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define h2v1_merged_upsample_internal  extbgr_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extbgr_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define h2v1_merged_upsample_internal  extbgrx_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extbgrx_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define h2v1_merged_upsample_internal  extxbgr_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extxbgr_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define h2v1_merged_upsample_internal  extxrgb_h2v1_merged_upsample_internal
+#define h2v2_merged_upsample_internal  extxrgb_h2v2_merged_upsample_internal
+#include "jdmrgext.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef h2v1_merged_upsample_internal
+#undef h2v2_merged_upsample_internal
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ * This is taken directly from jdcolor.c; see that file for more info.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table(j_decompress_ptr cinfo)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  int i;
+  JLONG x;
+  SHIFT_TEMPS
+
+  upsample->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(int));
+  upsample->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(int));
+  upsample->Cr_g_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(JLONG));
+  upsample->Cb_g_tab = (JLONG *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                (_MAXJSAMPLE + 1) * sizeof(JLONG));
+
+  for (i = 0, x = -_CENTERJSAMPLE; i <= _MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0.._MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - _CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    upsample->Cr_r_tab[i] = (int)
+                    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    upsample->Cb_b_tab[i] = (int)
+                    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    upsample->Cr_g_tab[i] = (-FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    upsample->Cb_g_tab[i] = (-FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_merged_upsample(j_decompress_ptr cinfo)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+
+  /* Mark the spare buffer empty */
+  upsample->spare_full = FALSE;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * The control routine just handles the row buffering considerations.
+ */
+
+METHODDEF(void)
+merged_2v_upsample(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION *in_row_group_ctr,
+                   JDIMENSION in_row_groups_avail, _JSAMPARRAY output_buf,
+                   JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+/* 2:1 vertical sampling case: may need a spare row. */
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  _JSAMPROW work_ptrs[2];
+  JDIMENSION num_rows;          /* number of rows returned to caller */
+
+  if (upsample->spare_full) {
+    /* If we have a spare row saved from a previous cycle, just return it. */
+    JDIMENSION size = upsample->out_row_width;
+    if (cinfo->out_color_space == JCS_RGB565)
+      size = cinfo->output_width * 2;
+    _jcopy_sample_rows(&upsample->spare_row, 0, output_buf + *out_row_ctr, 0,
+                       1, size);
+    num_rows = 1;
+    upsample->spare_full = FALSE;
+  } else {
+    /* Figure number of rows to return to caller. */
+    num_rows = 2;
+    /* Not more than the distance to the end of the image. */
+    if (num_rows > upsample->rows_to_go)
+      num_rows = upsample->rows_to_go;
+    /* And not more than what the client can accept: */
+    out_rows_avail -= *out_row_ctr;
+    if (num_rows > out_rows_avail)
+      num_rows = out_rows_avail;
+    /* Create output pointer array for upsampler. */
+    work_ptrs[0] = output_buf[*out_row_ctr];
+    if (num_rows > 1) {
+      work_ptrs[1] = output_buf[*out_row_ctr + 1];
+    } else {
+      work_ptrs[1] = upsample->spare_row;
+      upsample->spare_full = TRUE;
+    }
+    /* Now do the upsampling. */
+    (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr, work_ptrs);
+  }
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (!upsample->spare_full)
+    (*in_row_group_ctr)++;
+}
+
+
+METHODDEF(void)
+merged_1v_upsample(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION *in_row_group_ctr,
+                   JDIMENSION in_row_groups_avail, _JSAMPARRAY output_buf,
+                   JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+/* 1:1 vertical sampling case: much easier, never need a spare row. */
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+
+  /* Just do the upsampling. */
+  (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr,
+                         output_buf + *out_row_ctr);
+  /* Adjust counts */
+  (*out_row_ctr)++;
+  (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by the control routines to do
+ * the actual upsampling/conversion.  One row group is processed per call.
+ *
+ * Note: since we may be writing directly into application-supplied buffers,
+ * we have to be honest about the output width; we can't assume the buffer
+ * has been rounded up to an even width.
+ */
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+METHODDEF(void)
+h2v1_merged_upsample(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                     JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    extrgb_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                         output_buf);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    extrgbx_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_BGR:
+    extbgr_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                         output_buf);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    extbgrx_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    extxbgr_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    extxrgb_h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  default:
+    h2v1_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                  output_buf);
+    break;
+  }
+}
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ */
+
+METHODDEF(void)
+h2v2_merged_upsample(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                     JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    extrgb_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                         output_buf);
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    extrgbx_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_BGR:
+    extbgr_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                         output_buf);
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    extbgrx_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    extxbgr_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    extxrgb_h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                          output_buf);
+    break;
+  default:
+    h2v2_merged_upsample_internal(cinfo, input_buf, in_row_group_ctr,
+                                  output_buf);
+    break;
+  }
+}
+
+
+/*
+ * RGB565 conversion
+ */
+
+#define PACK_SHORT_565_LE(r, g, b) \
+  ((((r) << 8) & 0xF800) | (((g) << 3) & 0x7E0) | ((b) >> 3))
+#define PACK_SHORT_565_BE(r, g, b) \
+  (((r) & 0xF8) | ((g) >> 5) | (((g) << 11) & 0xE000) | (((b) << 5) & 0x1F00))
+
+#define PACK_TWO_PIXELS_LE(l, r)    ((r << 16) | l)
+#define PACK_TWO_PIXELS_BE(l, r)    ((l << 16) | r)
+
+#define WRITE_TWO_PIXELS_LE(addr, pixels) { \
+  ((INT16 *)(addr))[0] = (INT16)(pixels); \
+  ((INT16 *)(addr))[1] = (INT16)((pixels) >> 16); \
+}
+#define WRITE_TWO_PIXELS_BE(addr, pixels) { \
+  ((INT16 *)(addr))[1] = (INT16)(pixels); \
+  ((INT16 *)(addr))[0] = (INT16)((pixels) >> 16); \
+}
+
+#define DITHER_565_R(r, dither)  ((r) + ((dither) & 0xFF))
+#define DITHER_565_G(g, dither)  ((g) + (((dither) & 0xFF) >> 1))
+#define DITHER_565_B(b, dither)  ((b) + ((dither) & 0xFF))
+
+
+/* Declarations for ordered dithering
+ *
+ * We use a 4x4 ordered dither array packed into 32 bits.  This array is
+ * sufficient for dithering RGB888 to RGB565.
+ */
+
+#define DITHER_MASK       0x3
+#define DITHER_ROTATE(x)  ((((x) & 0xFF) << 24) | (((x) >> 8) & 0x00FFFFFF))
+static const JLONG dither_matrix[4] = {
+  0x0008020A,
+  0x0C040E06,
+  0x030B0109,
+  0x0F070D05
+};
+
+
+/* Include inline routines for RGB565 conversion */
+
+#define PACK_SHORT_565  PACK_SHORT_565_LE
+#define PACK_TWO_PIXELS  PACK_TWO_PIXELS_LE
+#define WRITE_TWO_PIXELS  WRITE_TWO_PIXELS_LE
+#define h2v1_merged_upsample_565_internal  h2v1_merged_upsample_565_le
+#define h2v1_merged_upsample_565D_internal  h2v1_merged_upsample_565D_le
+#define h2v2_merged_upsample_565_internal  h2v2_merged_upsample_565_le
+#define h2v2_merged_upsample_565D_internal  h2v2_merged_upsample_565D_le
+#include "jdmrg565.c"
+#undef PACK_SHORT_565
+#undef PACK_TWO_PIXELS
+#undef WRITE_TWO_PIXELS
+#undef h2v1_merged_upsample_565_internal
+#undef h2v1_merged_upsample_565D_internal
+#undef h2v2_merged_upsample_565_internal
+#undef h2v2_merged_upsample_565D_internal
+
+#define PACK_SHORT_565  PACK_SHORT_565_BE
+#define PACK_TWO_PIXELS  PACK_TWO_PIXELS_BE
+#define WRITE_TWO_PIXELS  WRITE_TWO_PIXELS_BE
+#define h2v1_merged_upsample_565_internal  h2v1_merged_upsample_565_be
+#define h2v1_merged_upsample_565D_internal  h2v1_merged_upsample_565D_be
+#define h2v2_merged_upsample_565_internal  h2v2_merged_upsample_565_be
+#define h2v2_merged_upsample_565D_internal  h2v2_merged_upsample_565D_be
+#include "jdmrg565.c"
+#undef PACK_SHORT_565
+#undef PACK_TWO_PIXELS
+#undef WRITE_TWO_PIXELS
+#undef h2v1_merged_upsample_565_internal
+#undef h2v1_merged_upsample_565D_internal
+#undef h2v2_merged_upsample_565_internal
+#undef h2v2_merged_upsample_565D_internal
+
+
+static INLINE boolean is_big_endian(void)
+{
+  int test_value = 1;
+  if (*(char *)&test_value != 1)
+    return TRUE;
+  return FALSE;
+}
+
+
+METHODDEF(void)
+h2v1_merged_upsample_565(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                         JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  if (is_big_endian())
+    h2v1_merged_upsample_565_be(cinfo, input_buf, in_row_group_ctr,
+                                output_buf);
+  else
+    h2v1_merged_upsample_565_le(cinfo, input_buf, in_row_group_ctr,
+                                output_buf);
+}
+
+
+METHODDEF(void)
+h2v1_merged_upsample_565D(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                          JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  if (is_big_endian())
+    h2v1_merged_upsample_565D_be(cinfo, input_buf, in_row_group_ctr,
+                                 output_buf);
+  else
+    h2v1_merged_upsample_565D_le(cinfo, input_buf, in_row_group_ctr,
+                                 output_buf);
+}
+
+
+METHODDEF(void)
+h2v2_merged_upsample_565(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                         JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  if (is_big_endian())
+    h2v2_merged_upsample_565_be(cinfo, input_buf, in_row_group_ctr,
+                                output_buf);
+  else
+    h2v2_merged_upsample_565_le(cinfo, input_buf, in_row_group_ctr,
+                                output_buf);
+}
+
+
+METHODDEF(void)
+h2v2_merged_upsample_565D(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                          JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf)
+{
+  if (is_big_endian())
+    h2v2_merged_upsample_565D_be(cinfo, input_buf, in_row_group_ctr,
+                                 output_buf);
+  else
+    h2v2_merged_upsample_565D_le(cinfo, input_buf, in_row_group_ctr,
+                                 output_buf);
+}
+
+
+/*
+ * Module initialization routine for merged upsampling/color conversion.
+ *
+ * NB: this is called under the conditions determined by use_merged_upsample()
+ * in jdmaster.c.  That routine MUST correspond to the actual capabilities
+ * of this module; no safety checks are made here.
+ */
+
+GLOBAL(void)
+_jinit_merged_upsampler(j_decompress_ptr cinfo)
+{
+  my_merged_upsample_ptr upsample;
+
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  upsample = (my_merged_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_merged_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *)upsample;
+  upsample->pub.start_pass = start_pass_merged_upsample;
+  upsample->pub.need_context_rows = FALSE;
+
+  upsample->out_row_width = cinfo->output_width * cinfo->out_color_components;
+
+  if (cinfo->max_v_samp_factor == 2) {
+    upsample->pub._upsample = merged_2v_upsample;
+#ifdef WITH_SIMD
+    if (jsimd_can_h2v2_merged_upsample())
+      upsample->upmethod = jsimd_h2v2_merged_upsample;
+    else
+#endif
+      upsample->upmethod = h2v2_merged_upsample;
+    if (cinfo->out_color_space == JCS_RGB565) {
+      if (cinfo->dither_mode != JDITHER_NONE) {
+        upsample->upmethod = h2v2_merged_upsample_565D;
+      } else {
+        upsample->upmethod = h2v2_merged_upsample_565;
+      }
+    }
+    /* Allocate a spare row buffer */
+    upsample->spare_row = (_JSAMPROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                (size_t)(upsample->out_row_width * sizeof(_JSAMPLE)));
+  } else {
+    upsample->pub._upsample = merged_1v_upsample;
+#ifdef WITH_SIMD
+    if (jsimd_can_h2v1_merged_upsample())
+      upsample->upmethod = jsimd_h2v1_merged_upsample;
+    else
+#endif
+      upsample->upmethod = h2v1_merged_upsample;
+    if (cinfo->out_color_space == JCS_RGB565) {
+      if (cinfo->dither_mode != JDITHER_NONE) {
+        upsample->upmethod = h2v1_merged_upsample_565D;
+      } else {
+        upsample->upmethod = h2v1_merged_upsample_565;
+      }
+    }
+    /* No spare row needed */
+    upsample->spare_row = NULL;
+  }
+
+  build_ycc_rgb_table(cinfo);
+}
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */

thirdparty/libjpeg-turbo/src/jdmerge.h

@@ -0,0 +1,48 @@
+/*
+ * jdmerge.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;    /* public fields */
+
+  /* Pointer to routine to do actual upsampling/conversion of one row group */
+  void (*upmethod) (j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                    JDIMENSION in_row_group_ctr, _JSAMPARRAY output_buf);
+
+  /* Private state for YCC->RGB conversion */
+  int *Cr_r_tab;                /* => table for Cr to R conversion */
+  int *Cb_b_tab;                /* => table for Cb to B conversion */
+  JLONG *Cr_g_tab;              /* => table for Cr to G conversion */
+  JLONG *Cb_g_tab;              /* => table for Cb to G conversion */
+
+  /* For 2:1 vertical sampling, we produce two output rows at a time.
+   * We need a "spare" row buffer to hold the second output row if the
+   * application provides just a one-row buffer; we also use the spare
+   * to discard the dummy last row if the image height is odd.
+   */
+  _JSAMPROW spare_row;
+  boolean spare_full;           /* T if spare buffer is occupied */
+
+  JDIMENSION out_row_width;     /* samples per output row */
+  JDIMENSION rows_to_go;        /* counts rows remaining in image */
+} my_merged_upsampler;
+
+typedef my_merged_upsampler *my_merged_upsample_ptr;
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */

thirdparty/libjpeg-turbo/src/jdmrg565.c

@@ -0,0 +1,355 @@
+/*
+ * jdmrg565.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2013, Linaro Limited.
+ * Copyright (C) 2014-2015, 2018, 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ */
+
+
+INLINE
+LOCAL(void)
+h2v1_merged_upsample_565_internal(j_decompress_ptr cinfo,
+                                  _JSAMPIMAGE input_buf,
+                                  JDIMENSION in_row_group_ctr,
+                                  _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr;
+  _JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  unsigned int r, g, b;
+  JLONG rgb;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = *inptr0++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr0++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr, rgb);
+    outptr += 4;
+  }
+
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = *inptr0;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr = (INT16)rgb;
+  }
+}
+
+
+INLINE
+LOCAL(void)
+h2v1_merged_upsample_565D_internal(j_decompress_ptr cinfo,
+                                   _JSAMPIMAGE input_buf,
+                                   JDIMENSION in_row_group_ctr,
+                                   _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr;
+  _JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  JLONG d0 = dither_matrix[cinfo->output_scanline & DITHER_MASK];
+  unsigned int r, g, b;
+  JLONG rgb;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = *inptr0++;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    d0 = DITHER_ROTATE(d0);
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr0++;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    d0 = DITHER_ROTATE(d0);
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr, rgb);
+    outptr += 4;
+  }
+
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = *inptr0;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr = (INT16)rgb;
+  }
+}
+
+
+INLINE
+LOCAL(void)
+h2v2_merged_upsample_565_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                                  JDIMENSION in_row_group_ctr,
+                                  _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr0, outptr1;
+  _JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  unsigned int r, g, b;
+  JLONG rgb;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr * 2];
+  inptr01 = input_buf[0][in_row_group_ctr * 2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = *inptr00++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr00++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr0, rgb);
+    outptr0 += 4;
+
+    y  = *inptr01++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr01++;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr1, rgb);
+    outptr1 += 4;
+  }
+
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    y  = *inptr00;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr0 = (INT16)rgb;
+
+    y  = *inptr01;
+    r = range_limit[y + cred];
+    g = range_limit[y + cgreen];
+    b = range_limit[y + cblue];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr1 = (INT16)rgb;
+  }
+}
+
+
+INLINE
+LOCAL(void)
+h2v2_merged_upsample_565D_internal(j_decompress_ptr cinfo,
+                                   _JSAMPIMAGE input_buf,
+                                   JDIMENSION in_row_group_ctr,
+                                   _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr0, outptr1;
+  _JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  JLONG d0 = dither_matrix[cinfo->output_scanline & DITHER_MASK];
+  JLONG d1 = dither_matrix[(cinfo->output_scanline + 1) & DITHER_MASK];
+  unsigned int r, g, b;
+  JLONG rgb;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr * 2];
+  inptr01 = input_buf[0][in_row_group_ctr * 2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = *inptr00++;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    d0 = DITHER_ROTATE(d0);
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr00++;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    d0 = DITHER_ROTATE(d0);
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr0, rgb);
+    outptr0 += 4;
+
+    y  = *inptr01++;
+    r = range_limit[DITHER_565_R(y + cred, d1)];
+    g = range_limit[DITHER_565_G(y + cgreen, d1)];
+    b = range_limit[DITHER_565_B(y + cblue, d1)];
+    d1 = DITHER_ROTATE(d1);
+    rgb = PACK_SHORT_565(r, g, b);
+
+    y  = *inptr01++;
+    r = range_limit[DITHER_565_R(y + cred, d1)];
+    g = range_limit[DITHER_565_G(y + cgreen, d1)];
+    b = range_limit[DITHER_565_B(y + cblue, d1)];
+    d1 = DITHER_ROTATE(d1);
+    rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
+
+    WRITE_TWO_PIXELS(outptr1, rgb);
+    outptr1 += 4;
+  }
+
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+
+    y  = *inptr00;
+    r = range_limit[DITHER_565_R(y + cred, d0)];
+    g = range_limit[DITHER_565_G(y + cgreen, d0)];
+    b = range_limit[DITHER_565_B(y + cblue, d0)];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr0 = (INT16)rgb;
+
+    y  = *inptr01;
+    r = range_limit[DITHER_565_R(y + cred, d1)];
+    g = range_limit[DITHER_565_G(y + cgreen, d1)];
+    b = range_limit[DITHER_565_B(y + cblue, d1)];
+    rgb = PACK_SHORT_565(r, g, b);
+    *(INT16 *)outptr1 = (INT16)rgb;
+  }
+}

thirdparty/libjpeg-turbo/src/jdmrgext.c

@@ -0,0 +1,184 @@
+/*
+ * jdmrgext.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2011, 2015, 2020, 2022-2023, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ */
+
+
+/* This file is included by jdmerge.c */
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+INLINE
+LOCAL(void)
+h2v1_merged_upsample_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                              JDIMENSION in_row_group_ctr,
+                              _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr;
+  _JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = *inptr0++;
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr += RGB_PIXELSIZE;
+    y  = *inptr0++;
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = *inptr0;
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+  }
+}
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ */
+
+INLINE
+LOCAL(void)
+h2v2_merged_upsample_internal(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                              JDIMENSION in_row_group_ctr,
+                              _JSAMPARRAY output_buf)
+{
+  my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register _JSAMPROW outptr0, outptr1;
+  _JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int *Crrtab = upsample->Cr_r_tab;
+  int *Cbbtab = upsample->Cb_b_tab;
+  JLONG *Crgtab = upsample->Cr_g_tab;
+  JLONG *Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr * 2];
+  inptr01 = input_buf[0][in_row_group_ctr * 2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = *inptr1++;
+    cr = *inptr2++;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = *inptr00++;
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr0[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr0 += RGB_PIXELSIZE;
+    y  = *inptr00++;
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr0[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr0 += RGB_PIXELSIZE;
+    y  = *inptr01++;
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr1[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr1 += RGB_PIXELSIZE;
+    y  = *inptr01++;
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr1[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    outptr1 += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = *inptr1;
+    cr = *inptr2;
+    cred = Crrtab[cr];
+    cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = *inptr00;
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr0[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+    y  = *inptr01;
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+#ifdef RGB_ALPHA
+    outptr1[RGB_ALPHA] = _MAXJSAMPLE;
+#endif
+  }
+}

thirdparty/libjpeg-turbo/src/jdphuff.c

@@ -0,0 +1,681 @@
+/*
+ * jdphuff.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015-2016, 2018-2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains Huffman entropy decoding routines for progressive JPEG.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"             /* Declarations shared with jd*huff.c */
+#include <limits.h>
+
+
+#ifdef D_PROGRESSIVE_SUPPORTED
+
+/*
+ * Expanded entropy decoder object for progressive Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  unsigned int EOBRUN;                  /* remaining EOBs in EOBRUN */
+  int last_dc_val[MAX_COMPS_IN_SCAN];   /* last DC coef for each component */
+} savable_state;
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;  /* Bit buffer at start of MCU */
+  savable_state saved;          /* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;  /* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl *derived_tbls[NUM_HUFF_TBLS];
+
+  d_derived_tbl *ac_derived_tbl; /* active table during an AC scan */
+} phuff_entropy_decoder;
+
+typedef phuff_entropy_decoder *phuff_entropy_ptr;
+
+/* Forward declarations */
+METHODDEF(boolean) decode_mcu_DC_first(j_decompress_ptr cinfo,
+                                       JBLOCKROW *MCU_data);
+METHODDEF(boolean) decode_mcu_AC_first(j_decompress_ptr cinfo,
+                                       JBLOCKROW *MCU_data);
+METHODDEF(boolean) decode_mcu_DC_refine(j_decompress_ptr cinfo,
+                                        JBLOCKROW *MCU_data);
+METHODDEF(boolean) decode_mcu_AC_refine(j_decompress_ptr cinfo,
+                                        JBLOCKROW *MCU_data);
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_phuff_decoder(j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  boolean is_DC_band, bad;
+  int ci, coefi, tbl;
+  d_derived_tbl **pdtbl;
+  int *coef_bit_ptr, *prev_coef_bit_ptr;
+  jpeg_component_info *compptr;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* Validate scan parameters */
+  bad = FALSE;
+  if (is_DC_band) {
+    if (cinfo->Se != 0)
+      bad = TRUE;
+  } else {
+    /* need not check Ss/Se < 0 since they came from unsigned bytes */
+    if (cinfo->Ss > cinfo->Se || cinfo->Se >= DCTSIZE2)
+      bad = TRUE;
+    /* AC scans may have only one component */
+    if (cinfo->comps_in_scan != 1)
+      bad = TRUE;
+  }
+  if (cinfo->Ah != 0) {
+    /* Successive approximation refinement scan: must have Al = Ah-1. */
+    if (cinfo->Al != cinfo->Ah - 1)
+      bad = TRUE;
+  }
+  if (cinfo->Al > 13)           /* need not check for < 0 */
+    bad = TRUE;
+  /* Arguably the maximum Al value should be less than 13 for 8-bit precision,
+   * but the spec doesn't say so, and we try to be liberal about what we
+   * accept.  Note: large Al values could result in out-of-range DC
+   * coefficients during early scans, leading to bizarre displays due to
+   * overflows in the IDCT math.  But we won't crash.
+   */
+  if (bad)
+    ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+             cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+  /* Update progression status, and verify that scan order is legal.
+   * Note that inter-scan inconsistencies are treated as warnings
+   * not fatal errors ... not clear if this is right way to behave.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    int cindex = cinfo->cur_comp_info[ci]->component_index;
+    coef_bit_ptr = &cinfo->coef_bits[cindex][0];
+    prev_coef_bit_ptr = &cinfo->coef_bits[cindex + cinfo->num_components][0];
+    if (!is_DC_band && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+      WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+    for (coefi = MIN(cinfo->Ss, 1); coefi <= MAX(cinfo->Se, 9); coefi++) {
+      if (cinfo->input_scan_number > 1)
+        prev_coef_bit_ptr[coefi] = coef_bit_ptr[coefi];
+      else
+        prev_coef_bit_ptr[coefi] = 0;
+    }
+    for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+      int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+      if (cinfo->Ah != expected)
+        WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+      coef_bit_ptr[coefi] = cinfo->Al;
+    }
+  }
+
+  /* Select MCU decoding routine */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_first;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_first;
+  } else {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_refine;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_refine;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Make sure requested tables are present, and compute derived tables.
+     * We may build same derived table more than once, but it's not expensive.
+     */
+    if (is_DC_band) {
+      if (cinfo->Ah == 0) {     /* DC refinement needs no table */
+        tbl = compptr->dc_tbl_no;
+        pdtbl = (d_derived_tbl **)(entropy->derived_tbls) + tbl;
+        jpeg_make_d_derived_tbl(cinfo, TRUE, tbl, pdtbl);
+      }
+    } else {
+      tbl = compptr->ac_tbl_no;
+      pdtbl = (d_derived_tbl **)(entropy->derived_tbls) + tbl;
+      jpeg_make_d_derived_tbl(cinfo, FALSE, tbl, pdtbl);
+      /* remember the single active table */
+      entropy->ac_derived_tbl = entropy->derived_tbls[tbl];
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize private state variables */
+  entropy->saved.EOBRUN = 0;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#define AVOID_TABLES
+#ifdef AVOID_TABLES
+
+#define NEG_1  ((unsigned)-1)
+#define HUFF_EXTEND(x, s) \
+  ((x) < (1 << ((s) - 1)) ? (x) + (((NEG_1) << (s)) + 1) : (x))
+
+#else
+
+#define HUFF_EXTEND(x, s) \
+  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] = {   /* entry n is 2**(n-1) */
+  0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+  0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000
+};
+
+static const int extend_offset[16] = { /* entry n is (-1 << n) + 1 */
+  0, ((-1) << 1) + 1, ((-1) << 2) + 1, ((-1) << 3) + 1, ((-1) << 4) + 1,
+  ((-1) << 5) + 1, ((-1) << 6) + 1, ((-1) << 7) + 1, ((-1) << 8) + 1,
+  ((-1) << 9) + 1, ((-1) << 10) + 1, ((-1) << 11) + 1, ((-1) << 12) + 1,
+  ((-1) << 13) + 1, ((-1) << 14) + 1, ((-1) << 15) + 1
+};
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(boolean)
+process_restart(j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (!(*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+  /* Re-init EOB run count, too */
+  entropy->saved.EOBRUN = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Huffman MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * Huffman-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ *
+ * We return FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * spectral selection, since we'll just re-assign them on the next call.
+ * Successive approximation AC refinement has to be more careful, however.)
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  int Al = cinfo->Al;
+  register int s, r;
+  int blkn, ci;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  savable_state state;
+  d_derived_tbl *tbl;
+  jpeg_component_info *compptr;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!process_restart(cinfo))
+        return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (!entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+    state = entropy->saved;
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      block = MCU_data[blkn];
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      tbl = entropy->derived_tbls[compptr->dc_tbl_no];
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      HUFF_DECODE(s, br_state, tbl, return FALSE, label1);
+      if (s) {
+        CHECK_BIT_BUFFER(br_state, s, return FALSE);
+        r = GET_BITS(s);
+        s = HUFF_EXTEND(r, s);
+      }
+
+      /* Convert DC difference to actual value, update last_dc_val */
+      if ((state.last_dc_val[ci] >= 0 &&
+           s > INT_MAX - state.last_dc_val[ci]) ||
+          (state.last_dc_val[ci] < 0 && s < INT_MIN - state.last_dc_val[ci]))
+        ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      /* Scale and output the coefficient (assumes jpeg_natural_order[0]=0) */
+      (*block)[0] = (JCOEF)LEFT_SHIFT(s, Al);
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+    entropy->saved = state;
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  if (cinfo->restart_interval)
+    entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  register int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  d_derived_tbl *tbl;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!process_restart(cinfo))
+        return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (!entropy->pub.insufficient_data) {
+
+    /* Load up working state.
+     * We can avoid loading/saving bitread state if in an EOB run.
+     */
+    EOBRUN = entropy->saved.EOBRUN;     /* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+
+    if (EOBRUN > 0)             /* if it's a band of zeroes... */
+      EOBRUN--;                 /* ...process it now (we do nothing) */
+    else {
+      BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+      block = MCU_data[0];
+      tbl = entropy->ac_derived_tbl;
+
+      for (k = cinfo->Ss; k <= Se; k++) {
+        HUFF_DECODE(s, br_state, tbl, return FALSE, label2);
+        r = s >> 4;
+        s &= 15;
+        if (s) {
+          k += r;
+          CHECK_BIT_BUFFER(br_state, s, return FALSE);
+          r = GET_BITS(s);
+          s = HUFF_EXTEND(r, s);
+          /* Scale and output coefficient in natural (dezigzagged) order */
+          (*block)[jpeg_natural_order[k]] = (JCOEF)LEFT_SHIFT(s, Al);
+        } else {
+          if (r == 15) {        /* ZRL */
+            k += 15;            /* skip 15 zeroes in band */
+          } else {              /* EOBr, run length is 2^r + appended bits */
+            EOBRUN = 1 << r;
+            if (r) {            /* EOBr, r > 0 */
+              CHECK_BIT_BUFFER(br_state, r, return FALSE);
+              r = GET_BITS(r);
+              EOBRUN += r;
+            }
+            EOBRUN--;           /* this band is processed at this moment */
+            break;              /* force end-of-band */
+          }
+        }
+      }
+
+      BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+    }
+
+    /* Completed MCU, so update state */
+    entropy->saved.EOBRUN = EOBRUN;     /* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  if (cinfo->restart_interval)
+    entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  int p1 = 1 << cinfo->Al;      /* 1 in the bit position being coded */
+  int blkn;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!process_restart(cinfo))
+        return FALSE;
+  }
+
+  /* Not worth the cycles to check insufficient_data here,
+   * since we will not change the data anyway if we read zeroes.
+   */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    CHECK_BIT_BUFFER(br_state, 1, return FALSE);
+    if (GET_BITS(1))
+      (*block)[0] |= p1;
+    /* Note: since we use |=, repeating the assignment later is safe */
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+
+  /* Account for restart interval (no-op if not using restarts) */
+  if (cinfo->restart_interval)
+    entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr)cinfo->entropy;
+  int Se = cinfo->Se;
+  int p1 = 1 << cinfo->Al;        /* 1 in the bit position being coded */
+  int m1 = (NEG_1) << cinfo->Al;  /* -1 in the bit position being coded */
+  register int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  BITREAD_STATE_VARS;
+  d_derived_tbl *tbl;
+  int num_newnz;
+  int newnz_pos[DCTSIZE2];
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (!process_restart(cinfo))
+        return FALSE;
+  }
+
+  /* If we've run out of data, don't modify the MCU.
+   */
+  if (!entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
+    EOBRUN = entropy->saved.EOBRUN; /* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+    block = MCU_data[0];
+    tbl = entropy->ac_derived_tbl;
+
+    /* If we are forced to suspend, we must undo the assignments to any newly
+     * nonzero coefficients in the block, because otherwise we'd get confused
+     * next time about which coefficients were already nonzero.
+     * But we need not undo addition of bits to already-nonzero coefficients;
+     * instead, we can test the current bit to see if we already did it.
+     */
+    num_newnz = 0;
+
+    /* initialize coefficient loop counter to start of band */
+    k = cinfo->Ss;
+
+    if (EOBRUN == 0) {
+      for (; k <= Se; k++) {
+        HUFF_DECODE(s, br_state, tbl, goto undoit, label3);
+        r = s >> 4;
+        s &= 15;
+        if (s) {
+          if (s != 1)           /* size of new coef should always be 1 */
+            WARNMS(cinfo, JWRN_HUFF_BAD_CODE);
+          CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+          if (GET_BITS(1))
+            s = p1;             /* newly nonzero coef is positive */
+          else
+            s = m1;             /* newly nonzero coef is negative */
+        } else {
+          if (r != 15) {
+            EOBRUN = 1 << r;    /* EOBr, run length is 2^r + appended bits */
+            if (r) {
+              CHECK_BIT_BUFFER(br_state, r, goto undoit);
+              r = GET_BITS(r);
+              EOBRUN += r;
+            }
+            break;              /* rest of block is handled by EOB logic */
+          }
+          /* note s = 0 for processing ZRL */
+        }
+        /* Advance over already-nonzero coefs and r still-zero coefs,
+         * appending correction bits to the nonzeroes.  A correction bit is 1
+         * if the absolute value of the coefficient must be increased.
+         */
+        do {
+          thiscoef = *block + jpeg_natural_order[k];
+          if (*thiscoef != 0) {
+            CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+            if (GET_BITS(1)) {
+              if ((*thiscoef & p1) == 0) { /* do nothing if already set it */
+                if (*thiscoef >= 0)
+                  *thiscoef += (JCOEF)p1;
+                else
+                  *thiscoef += (JCOEF)m1;
+              }
+            }
+          } else {
+            if (--r < 0)
+              break;            /* reached target zero coefficient */
+          }
+          k++;
+        } while (k <= Se);
+        if (s) {
+          int pos = jpeg_natural_order[k];
+          /* Output newly nonzero coefficient */
+          (*block)[pos] = (JCOEF)s;
+          /* Remember its position in case we have to suspend */
+          newnz_pos[num_newnz++] = pos;
+        }
+      }
+    }
+
+    if (EOBRUN > 0) {
+      /* Scan any remaining coefficient positions after the end-of-band
+       * (the last newly nonzero coefficient, if any).  Append a correction
+       * bit to each already-nonzero coefficient.  A correction bit is 1
+       * if the absolute value of the coefficient must be increased.
+       */
+      for (; k <= Se; k++) {
+        thiscoef = *block + jpeg_natural_order[k];
+        if (*thiscoef != 0) {
+          CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+          if (GET_BITS(1)) {
+            if ((*thiscoef & p1) == 0) { /* do nothing if already changed it */
+              if (*thiscoef >= 0)
+                *thiscoef += (JCOEF)p1;
+              else
+                *thiscoef += (JCOEF)m1;
+            }
+          }
+        }
+      }
+      /* Count one block completed in EOB run */
+      EOBRUN--;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
+    entropy->saved.EOBRUN = EOBRUN; /* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  if (cinfo->restart_interval)
+    entropy->restarts_to_go--;
+
+  return TRUE;
+
+undoit:
+  /* Re-zero any output coefficients that we made newly nonzero */
+  while (num_newnz > 0)
+    (*block)[newnz_pos[--num_newnz]] = 0;
+
+  return FALSE;
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_decoder(j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int *coef_bit_ptr;
+  int ci, i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(phuff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *)entropy;
+  entropy->pub.start_pass = start_pass_phuff_decoder;
+
+  /* Mark derived tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+  }
+
+  /* Create progression status table */
+  cinfo->coef_bits = (int (*)[DCTSIZE2])
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                cinfo->num_components * 2 * DCTSIZE2 *
+                                sizeof(int));
+  coef_bit_ptr = &cinfo->coef_bits[0][0];
+  for (ci = 0; ci < cinfo->num_components; ci++)
+    for (i = 0; i < DCTSIZE2; i++)
+      *coef_bit_ptr++ = -1;
+}
+
+#endif /* D_PROGRESSIVE_SUPPORTED */

thirdparty/libjpeg-turbo/src/jdpostct.c

@@ -0,0 +1,328 @@
+/*
+ * jdpostct.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022-2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains the decompression postprocessing controller.
+ * This controller manages the upsampling, color conversion, and color
+ * quantization/reduction steps; specifically, it controls the buffering
+ * between upsample/color conversion and color quantization/reduction.
+ *
+ * If no color quantization/reduction is required, then this module has no
+ * work to do, and it just hands off to the upsample/color conversion code.
+ * An integrated upsample/convert/quantize process would replace this module
+ * entirely.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_post_controller pub; /* public fields */
+
+  /* Color quantization source buffer: this holds output data from
+   * the upsample/color conversion step to be passed to the quantizer.
+   * For two-pass color quantization, we need a full-image buffer;
+   * for one-pass operation, a strip buffer is sufficient.
+   */
+  jvirt_sarray_ptr whole_image; /* virtual array, or NULL if one-pass */
+  _JSAMPARRAY buffer;           /* strip buffer, or current strip of virtual */
+  JDIMENSION strip_height;      /* buffer size in rows */
+  /* for two-pass mode only: */
+  JDIMENSION starting_row;      /* row # of first row in current strip */
+  JDIMENSION next_row;          /* index of next row to fill/empty in strip */
+} my_post_controller;
+
+typedef my_post_controller *my_post_ptr;
+
+
+/* Forward declarations */
+#if BITS_IN_JSAMPLE != 16
+METHODDEF(void) post_process_1pass(j_decompress_ptr cinfo,
+                                   _JSAMPIMAGE input_buf,
+                                   JDIMENSION *in_row_group_ctr,
+                                   JDIMENSION in_row_groups_avail,
+                                   _JSAMPARRAY output_buf,
+                                   JDIMENSION *out_row_ctr,
+                                   JDIMENSION out_rows_avail);
+#endif
+#if defined(QUANT_2PASS_SUPPORTED) && BITS_IN_JSAMPLE != 16
+METHODDEF(void) post_process_prepass(j_decompress_ptr cinfo,
+                                     _JSAMPIMAGE input_buf,
+                                     JDIMENSION *in_row_group_ctr,
+                                     JDIMENSION in_row_groups_avail,
+                                     _JSAMPARRAY output_buf,
+                                     JDIMENSION *out_row_ctr,
+                                     JDIMENSION out_rows_avail);
+METHODDEF(void) post_process_2pass(j_decompress_ptr cinfo,
+                                   _JSAMPIMAGE input_buf,
+                                   JDIMENSION *in_row_group_ctr,
+                                   JDIMENSION in_row_groups_avail,
+                                   _JSAMPARRAY output_buf,
+                                   JDIMENSION *out_row_ctr,
+                                   JDIMENSION out_rows_avail);
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_dpost(j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_post_ptr post = (my_post_ptr)cinfo->post;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+#if BITS_IN_JSAMPLE != 16
+    if (cinfo->quantize_colors) {
+      /* Single-pass processing with color quantization. */
+      post->pub._post_process_data = post_process_1pass;
+      /* We could be doing buffered-image output before starting a 2-pass
+       * color quantization; in that case, jinit_d_post_controller did not
+       * allocate a strip buffer.  Use the virtual-array buffer as workspace.
+       */
+      if (post->buffer == NULL) {
+        post->buffer = (_JSAMPARRAY)(*cinfo->mem->access_virt_sarray)
+          ((j_common_ptr)cinfo, post->whole_image,
+           (JDIMENSION)0, post->strip_height, TRUE);
+      }
+    } else
+#endif
+    {
+      /* For single-pass processing without color quantization,
+       * I have no work to do; just call the upsampler directly.
+       */
+      post->pub._post_process_data = cinfo->upsample->_upsample;
+    }
+    break;
+#if defined(QUANT_2PASS_SUPPORTED) && BITS_IN_JSAMPLE != 16
+  case JBUF_SAVE_AND_PASS:
+    /* First pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub._post_process_data = post_process_prepass;
+    break;
+  case JBUF_CRANK_DEST:
+    /* Second pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub._post_process_data = post_process_2pass;
+    break;
+#endif /* defined(QUANT_2PASS_SUPPORTED) && BITS_IN_JSAMPLE != 16 */
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+  post->starting_row = post->next_row = 0;
+}
+
+
+/*
+ * Process some data in the one-pass (strip buffer) case.
+ * This is used for color precision reduction as well as one-pass quantization.
+ */
+
+#if BITS_IN_JSAMPLE != 16
+
+METHODDEF(void)
+post_process_1pass(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION *in_row_group_ctr,
+                   JDIMENSION in_row_groups_avail, _JSAMPARRAY output_buf,
+                   JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr)cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Fill the buffer, but not more than what we can dump out in one go. */
+  /* Note we rely on the upsampler to detect bottom of image. */
+  max_rows = out_rows_avail - *out_row_ctr;
+  if (max_rows > post->strip_height)
+    max_rows = post->strip_height;
+  num_rows = 0;
+  (*cinfo->upsample->_upsample) (cinfo, input_buf, in_row_group_ctr,
+                                 in_row_groups_avail, post->buffer, &num_rows,
+                                 max_rows);
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->_color_quantize) (cinfo, post->buffer,
+                                        output_buf + *out_row_ctr,
+                                        (int)num_rows);
+  *out_row_ctr += num_rows;
+}
+
+#endif
+
+
+#if defined(QUANT_2PASS_SUPPORTED) && BITS_IN_JSAMPLE != 16
+
+/*
+ * Process some data in the first pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_prepass(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                     JDIMENSION *in_row_group_ctr,
+                     JDIMENSION in_row_groups_avail, _JSAMPARRAY output_buf,
+                     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr)cinfo->post;
+  JDIMENSION old_next_row, num_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (_JSAMPARRAY)(*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr)cinfo, post->whole_image,
+         post->starting_row, post->strip_height, TRUE);
+  }
+
+  /* Upsample some data (up to a strip height's worth). */
+  old_next_row = post->next_row;
+  (*cinfo->upsample->_upsample) (cinfo, input_buf, in_row_group_ctr,
+                                 in_row_groups_avail, post->buffer,
+                                 &post->next_row, post->strip_height);
+
+  /* Allow quantizer to scan new data.  No data is emitted, */
+  /* but we advance out_row_ctr so outer loop can tell when we're done. */
+  if (post->next_row > old_next_row) {
+    num_rows = post->next_row - old_next_row;
+    (*cinfo->cquantize->_color_quantize) (cinfo, post->buffer + old_next_row,
+                                          (_JSAMPARRAY)NULL, (int)num_rows);
+    *out_row_ctr += num_rows;
+  }
+
+  /* Advance if we filled the strip. */
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+
+/*
+ * Process some data in the second pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_2pass(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+                   JDIMENSION *in_row_group_ctr,
+                   JDIMENSION in_row_groups_avail, _JSAMPARRAY output_buf,
+                   JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr)cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (_JSAMPARRAY)(*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr)cinfo, post->whole_image,
+         post->starting_row, post->strip_height, FALSE);
+  }
+
+  /* Determine number of rows to emit. */
+  num_rows = post->strip_height - post->next_row; /* available in strip */
+  max_rows = out_rows_avail - *out_row_ctr; /* available in output area */
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+  /* We have to check bottom of image here, can't depend on upsampler. */
+  max_rows = cinfo->output_height - post->starting_row;
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->_color_quantize) (cinfo, post->buffer + post->next_row,
+                                        output_buf + *out_row_ctr,
+                                        (int)num_rows);
+  *out_row_ctr += num_rows;
+
+  /* Advance if we filled the strip. */
+  post->next_row += num_rows;
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+#endif /* defined(QUANT_2PASS_SUPPORTED) && BITS_IN_JSAMPLE != 16 */
+
+
+/*
+ * Initialize postprocessing controller.
+ */
+
+GLOBAL(void)
+_jinit_d_post_controller(j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_post_ptr post;
+
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  post = (my_post_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                sizeof(my_post_controller));
+  cinfo->post = (struct jpeg_d_post_controller *)post;
+  post->pub.start_pass = start_pass_dpost;
+  post->whole_image = NULL;     /* flag for no virtual arrays */
+  post->buffer = NULL;          /* flag for no strip buffer */
+
+  /* Create the quantization buffer, if needed */
+  if (cinfo->quantize_colors) {
+#if BITS_IN_JSAMPLE != 16
+    /* The buffer strip height is max_v_samp_factor, which is typically
+     * an efficient number of rows for upsampling to return.
+     * (In the presence of output rescaling, we might want to be smarter?)
+     */
+    post->strip_height = (JDIMENSION)cinfo->max_v_samp_factor;
+    if (need_full_buffer) {
+      /* Two-pass color quantization: need full-image storage. */
+      /* We round up the number of rows to a multiple of the strip height. */
+#ifdef QUANT_2PASS_SUPPORTED
+      post->whole_image = (*cinfo->mem->request_virt_sarray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE, FALSE,
+         cinfo->output_width * cinfo->out_color_components,
+         (JDIMENSION)jround_up((long)cinfo->output_height,
+                               (long)post->strip_height),
+         post->strip_height);
+#else
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif /* QUANT_2PASS_SUPPORTED */
+    } else {
+      /* One-pass color quantization: just make a strip buffer. */
+      post->buffer = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+         cinfo->output_width * cinfo->out_color_components,
+         post->strip_height);
+    }
+#else
+    ERREXIT(cinfo, JERR_NOTIMPL);
+#endif
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdsample.c

@@ -0,0 +1,553 @@
+/*
+ * jdsample.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
+ * Copyright (C) 2010, 2015-2016, 2022, 2024, D. R. Commander.
+ * Copyright (C) 2014, MIPS Technologies, Inc., California.
+ * Copyright (C) 2015, Google, Inc.
+ * Copyright (C) 2019-2020, Arm Limited.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains upsampling routines.
+ *
+ * Upsampling input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  Upsampling will normally produce
+ * max_v_samp_factor pixel rows from each row group (but this could vary
+ * if the upsampler is applying a scale factor of its own).
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ */
+
+#include "jinclude.h"
+#include "jdsample.h"
+#include "jsimd.h"
+#include "jpegapicomp.h"
+
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_upsample(j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+
+  /* Mark the conversion buffer empty */
+  upsample->next_row_out = cinfo->max_v_samp_factor;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * In this version we upsample each component independently.
+ * We upsample one row group into the conversion buffer, then apply
+ * color conversion a row at a time.
+ */
+
+METHODDEF(void)
+sep_upsample(j_decompress_ptr cinfo, _JSAMPIMAGE input_buf,
+             JDIMENSION *in_row_group_ctr, JDIMENSION in_row_groups_avail,
+             _JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+             JDIMENSION out_rows_avail)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+  int ci;
+  jpeg_component_info *compptr;
+  JDIMENSION num_rows;
+
+  /* Fill the conversion buffer, if it's empty */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor) {
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+         ci++, compptr++) {
+      /* Invoke per-component upsample method.  Notice we pass a POINTER
+       * to color_buf[ci], so that fullsize_upsample can change it.
+       */
+      (*upsample->methods[ci]) (cinfo, compptr,
+        input_buf[ci] + (*in_row_group_ctr * upsample->rowgroup_height[ci]),
+        upsample->color_buf + ci);
+    }
+    upsample->next_row_out = 0;
+  }
+
+  /* Color-convert and emit rows */
+
+  /* How many we have in the buffer: */
+  num_rows = (JDIMENSION)(cinfo->max_v_samp_factor - upsample->next_row_out);
+  /* Not more than the distance to the end of the image.  Need this test
+   * in case the image height is not a multiple of max_v_samp_factor:
+   */
+  if (num_rows > upsample->rows_to_go)
+    num_rows = upsample->rows_to_go;
+  /* And not more than what the client can accept: */
+  out_rows_avail -= *out_row_ctr;
+  if (num_rows > out_rows_avail)
+    num_rows = out_rows_avail;
+
+  (*cinfo->cconvert->_color_convert) (cinfo, upsample->color_buf,
+                                      (JDIMENSION)upsample->next_row_out,
+                                      output_buf + *out_row_ctr,
+                                      (int)num_rows);
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  upsample->next_row_out += num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor)
+    (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by sep_upsample to upsample pixel values
+ * of a single component.  One row group is processed per call.
+ */
+
+
+/*
+ * For full-size components, we just make color_buf[ci] point at the
+ * input buffer, and thus avoid copying any data.  Note that this is
+ * safe only because sep_upsample doesn't declare the input row group
+ * "consumed" until we are done color converting and emitting it.
+ */
+
+METHODDEF(void)
+fullsize_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                  _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  *output_data_ptr = input_data;
+}
+
+
+/*
+ * This is a no-op version used for "uninteresting" components.
+ * These components will not be referenced by color conversion.
+ */
+
+METHODDEF(void)
+noop_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+              _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  *output_data_ptr = NULL;      /* safety check */
+}
+
+
+/*
+ * This version handles any integral sampling ratios.
+ * This is not used for typical JPEG files, so it need not be fast.
+ * Nor, for that matter, is it particularly accurate: the algorithm is
+ * simple replication of the input pixel onto the corresponding output
+ * pixels.  The hi-falutin sampling literature refers to this as a
+ * "box filter".  A box filter tends to introduce visible artifacts,
+ * so if you are actually going to use 3:1 or 4:1 sampling ratios
+ * you would be well advised to improve this code.
+ */
+
+METHODDEF(void)
+int_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+             _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+  _JSAMPARRAY output_data = *output_data_ptr;
+  register _JSAMPROW inptr, outptr;
+  register _JSAMPLE invalue;
+  register int h;
+  _JSAMPROW outend;
+  int h_expand, v_expand;
+  int inrow, outrow;
+
+  h_expand = upsample->h_expand[compptr->component_index];
+  v_expand = upsample->v_expand[compptr->component_index];
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    /* Generate one output row with proper horizontal expansion */
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;
+      for (h = h_expand; h > 0; h--) {
+        *outptr++ = invalue;
+      }
+    }
+    /* Generate any additional output rows by duplicating the first one */
+    if (v_expand > 1) {
+      _jcopy_sample_rows(output_data, outrow, output_data, outrow + 1,
+                         v_expand - 1, cinfo->output_width);
+    }
+    inrow++;
+    outrow += v_expand;
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v1_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+              _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  _JSAMPARRAY output_data = *output_data_ptr;
+  register _JSAMPROW inptr, outptr;
+  register _JSAMPLE invalue;
+  _JSAMPROW outend;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+              _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  _JSAMPARRAY output_data = *output_data_ptr;
+  register _JSAMPROW inptr, outptr;
+  register _JSAMPLE invalue;
+  _JSAMPROW outend;
+  int inrow, outrow;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+    _jcopy_sample_rows(output_data, outrow, output_data, outrow + 1, 1,
+                       cinfo->output_width);
+    inrow++;
+    outrow += 2;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+ *
+ * The upsampling algorithm is linear interpolation between pixel centers,
+ * also known as a "triangle filter".  This is a good compromise between
+ * speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+ * of the way between input pixel centers.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  _JSAMPARRAY output_data = *output_data_ptr;
+  register _JSAMPROW inptr, outptr;
+  register int invalue;
+  register JDIMENSION colctr;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    /* Special case for first column */
+    invalue = *inptr++;
+    *outptr++ = (_JSAMPLE)invalue;
+    *outptr++ = (_JSAMPLE)((invalue * 3 + inptr[0] + 2) >> 2);
+
+    for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+      /* General case: 3/4 * nearer pixel + 1/4 * further pixel */
+      invalue = (*inptr++) * 3;
+      *outptr++ = (_JSAMPLE)((invalue + inptr[-2] + 1) >> 2);
+      *outptr++ = (_JSAMPLE)((invalue + inptr[0] + 2) >> 2);
+    }
+
+    /* Special case for last column */
+    invalue = *inptr;
+    *outptr++ = (_JSAMPLE)((invalue * 3 + inptr[-1] + 1) >> 2);
+    *outptr++ = (_JSAMPLE)invalue;
+  }
+}
+
+
+/*
+ * Fancy processing for 1:1 horizontal and 2:1 vertical (4:4:0 subsampling).
+ *
+ * This is a less common case, but it can be encountered when losslessly
+ * rotating/transposing a JPEG file that uses 4:2:2 chroma subsampling.
+ */
+
+METHODDEF(void)
+h1v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  _JSAMPARRAY output_data = *output_data_ptr;
+  _JSAMPROW inptr0, inptr1, outptr;
+#if BITS_IN_JSAMPLE == 8
+  int thiscolsum, bias;
+#else
+  JLONG thiscolsum, bias;
+#endif
+  JDIMENSION colctr;
+  int inrow, outrow, v;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    for (v = 0; v < 2; v++) {
+      /* inptr0 points to nearest input row, inptr1 points to next nearest */
+      inptr0 = input_data[inrow];
+      if (v == 0) {             /* next nearest is row above */
+        inptr1 = input_data[inrow - 1];
+        bias = 1;
+      } else {                  /* next nearest is row below */
+        inptr1 = input_data[inrow + 1];
+        bias = 2;
+      }
+      outptr = output_data[outrow++];
+
+      for (colctr = 0; colctr < compptr->downsampled_width; colctr++) {
+        thiscolsum = (*inptr0++) * 3 + (*inptr1++);
+        *outptr++ = (_JSAMPLE)((thiscolsum + bias) >> 2);
+      }
+    }
+    inrow++;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * Again a triangle filter; see comments for h2v1 case, above.
+ *
+ * It is OK for us to reference the adjacent input rows because we demanded
+ * context from the main buffer controller (see initialization code).
+ */
+
+METHODDEF(void)
+h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    _JSAMPARRAY input_data, _JSAMPARRAY *output_data_ptr)
+{
+  _JSAMPARRAY output_data = *output_data_ptr;
+  register _JSAMPROW inptr0, inptr1, outptr;
+#if BITS_IN_JSAMPLE == 8
+  register int thiscolsum, lastcolsum, nextcolsum;
+#else
+  register JLONG thiscolsum, lastcolsum, nextcolsum;
+#endif
+  register JDIMENSION colctr;
+  int inrow, outrow, v;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    for (v = 0; v < 2; v++) {
+      /* inptr0 points to nearest input row, inptr1 points to next nearest */
+      inptr0 = input_data[inrow];
+      if (v == 0)               /* next nearest is row above */
+        inptr1 = input_data[inrow - 1];
+      else                      /* next nearest is row below */
+        inptr1 = input_data[inrow + 1];
+      outptr = output_data[outrow++];
+
+      /* Special case for first column */
+      thiscolsum = (*inptr0++) * 3 + (*inptr1++);
+      nextcolsum = (*inptr0++) * 3 + (*inptr1++);
+      *outptr++ = (_JSAMPLE)((thiscolsum * 4 + 8) >> 4);
+      *outptr++ = (_JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4);
+      lastcolsum = thiscolsum;  thiscolsum = nextcolsum;
+
+      for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+        /* General case: 3/4 * nearer pixel + 1/4 * further pixel in each */
+        /* dimension, thus 9/16, 3/16, 3/16, 1/16 overall */
+        nextcolsum = (*inptr0++) * 3 + (*inptr1++);
+        *outptr++ = (_JSAMPLE)((thiscolsum * 3 + lastcolsum + 8) >> 4);
+        *outptr++ = (_JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4);
+        lastcolsum = thiscolsum;  thiscolsum = nextcolsum;
+      }
+
+      /* Special case for last column */
+      *outptr++ = (_JSAMPLE)((thiscolsum * 3 + lastcolsum + 8) >> 4);
+      *outptr++ = (_JSAMPLE)((thiscolsum * 4 + 7) >> 4);
+    }
+    inrow++;
+  }
+}
+
+
+/*
+ * Module initialization routine for upsampling.
+ */
+
+GLOBAL(void)
+_jinit_upsampler(j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+  int ci;
+  jpeg_component_info *compptr;
+  boolean need_buffer, do_fancy;
+  int h_in_group, v_in_group, h_out_group, v_out_group;
+
+#ifdef D_LOSSLESS_SUPPORTED
+  if (cinfo->master->lossless) {
+#if BITS_IN_JSAMPLE == 8
+    if (cinfo->data_precision > BITS_IN_JSAMPLE || cinfo->data_precision < 2)
+#else
+    if (cinfo->data_precision > BITS_IN_JSAMPLE ||
+        cinfo->data_precision < BITS_IN_JSAMPLE - 3)
+#endif
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  } else
+#endif
+  {
+    if (cinfo->data_precision != BITS_IN_JSAMPLE)
+      ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+  }
+
+  if (!cinfo->master->jinit_upsampler_no_alloc) {
+    upsample = (my_upsample_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+                                  sizeof(my_upsampler));
+    cinfo->upsample = (struct jpeg_upsampler *)upsample;
+    upsample->pub.start_pass = start_pass_upsample;
+    upsample->pub._upsample = sep_upsample;
+    upsample->pub.need_context_rows = FALSE; /* until we find out differently */
+  } else
+    upsample = (my_upsample_ptr)cinfo->upsample;
+
+  if (cinfo->CCIR601_sampling)  /* this isn't supported */
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* jdmainct.c doesn't support context rows when min_DCT_scaled_size = 1,
+   * so don't ask for it.
+   */
+  do_fancy = cinfo->do_fancy_upsampling && cinfo->_min_DCT_scaled_size > 1;
+
+  /* Verify we can handle the sampling factors, select per-component methods,
+   * and create storage as needed.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Compute size of an "input group" after IDCT scaling.  This many samples
+     * are to be converted to max_h_samp_factor * max_v_samp_factor pixels.
+     */
+    h_in_group = (compptr->h_samp_factor * compptr->_DCT_scaled_size) /
+                 cinfo->_min_DCT_scaled_size;
+    v_in_group = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+                 cinfo->_min_DCT_scaled_size;
+    h_out_group = cinfo->max_h_samp_factor;
+    v_out_group = cinfo->max_v_samp_factor;
+    upsample->rowgroup_height[ci] = v_in_group; /* save for use later */
+    need_buffer = TRUE;
+    if (!compptr->component_needed) {
+      /* Don't bother to upsample an uninteresting component. */
+      upsample->methods[ci] = noop_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group == h_out_group && v_in_group == v_out_group) {
+      /* Fullsize components can be processed without any work. */
+      upsample->methods[ci] = fullsize_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group * 2 == h_out_group && v_in_group == v_out_group) {
+      /* Special cases for 2h1v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2) {
+#ifdef WITH_SIMD
+        if (jsimd_can_h2v1_fancy_upsample())
+          upsample->methods[ci] = jsimd_h2v1_fancy_upsample;
+        else
+#endif
+          upsample->methods[ci] = h2v1_fancy_upsample;
+      } else {
+#ifdef WITH_SIMD
+        if (jsimd_can_h2v1_upsample())
+          upsample->methods[ci] = jsimd_h2v1_upsample;
+        else
+#endif
+          upsample->methods[ci] = h2v1_upsample;
+      }
+    } else if (h_in_group == h_out_group &&
+               v_in_group * 2 == v_out_group && do_fancy) {
+      /* Non-fancy upsampling is handled by the generic method */
+#if defined(WITH_SIMD) && (defined(__arm__) || defined(__aarch64__) || \
+                           defined(_M_ARM) || defined(_M_ARM64))
+      if (jsimd_can_h1v2_fancy_upsample())
+        upsample->methods[ci] = jsimd_h1v2_fancy_upsample;
+      else
+#endif
+        upsample->methods[ci] = h1v2_fancy_upsample;
+      upsample->pub.need_context_rows = TRUE;
+    } else if (h_in_group * 2 == h_out_group &&
+               v_in_group * 2 == v_out_group) {
+      /* Special cases for 2h2v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2) {
+#ifdef WITH_SIMD
+        if (jsimd_can_h2v2_fancy_upsample())
+          upsample->methods[ci] = jsimd_h2v2_fancy_upsample;
+        else
+#endif
+          upsample->methods[ci] = h2v2_fancy_upsample;
+        upsample->pub.need_context_rows = TRUE;
+      } else {
+#ifdef WITH_SIMD
+        if (jsimd_can_h2v2_upsample())
+          upsample->methods[ci] = jsimd_h2v2_upsample;
+        else
+#endif
+          upsample->methods[ci] = h2v2_upsample;
+      }
+    } else if ((h_out_group % h_in_group) == 0 &&
+               (v_out_group % v_in_group) == 0) {
+      /* Generic integral-factors upsampling method */
+#if defined(WITH_SIMD) && defined(__mips__)
+      if (jsimd_can_int_upsample())
+        upsample->methods[ci] = jsimd_int_upsample;
+      else
+#endif
+        upsample->methods[ci] = int_upsample;
+      upsample->h_expand[ci] = (UINT8)(h_out_group / h_in_group);
+      upsample->v_expand[ci] = (UINT8)(v_out_group / v_in_group);
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+    if (need_buffer && !cinfo->master->jinit_upsampler_no_alloc) {
+      upsample->color_buf[ci] = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
+        ((j_common_ptr)cinfo, JPOOL_IMAGE,
+         (JDIMENSION)jround_up((long)cinfo->output_width,
+                               (long)cinfo->max_h_samp_factor),
+         (JDIMENSION)cinfo->max_v_samp_factor);
+    }
+  }
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */

thirdparty/libjpeg-turbo/src/jdsample.h

@@ -0,0 +1,53 @@
+/*
+ * jdsample.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+#include "jsamplecomp.h"
+
+
+/* Pointer to routine to upsample a single component */
+typedef void (*upsample1_ptr) (j_decompress_ptr cinfo,
+                               jpeg_component_info *compptr,
+                               _JSAMPARRAY input_data,
+                               _JSAMPARRAY *output_data_ptr);
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;    /* public fields */
+
+  /* Color conversion buffer.  When using separate upsampling and color
+   * conversion steps, this buffer holds one upsampled row group until it
+   * has been color converted and output.
+   * Note: we do not allocate any storage for component(s) which are full-size,
+   * ie do not need rescaling.  The corresponding entry of color_buf[] is
+   * simply set to point to the input data array, thereby avoiding copying.
+   */
+  _JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  /* Per-component upsampling method pointers */
+  upsample1_ptr methods[MAX_COMPONENTS];
+
+  int next_row_out;             /* counts rows emitted from color_buf */
+  JDIMENSION rows_to_go;        /* counts rows remaining in image */
+
+  /* Height of an input row group for each component. */
+  int rowgroup_height[MAX_COMPONENTS];
+
+  /* These arrays save pixel expansion factors so that int_expand need not
+   * recompute them each time.  They are unused for other upsampling methods.
+   */
+  UINT8 h_expand[MAX_COMPONENTS];
+  UINT8 v_expand[MAX_COMPONENTS];
+} my_upsampler;
+
+typedef my_upsampler *my_upsample_ptr;

thirdparty/libjpeg-turbo/src/jdtrans.c

@@ -0,0 +1,162 @@
+/*
+ * jdtrans.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains library routines for transcoding decompression,
+ * that is, reading raw DCT coefficient arrays from an input JPEG file.
+ * The routines in jdapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegapicomp.h"
+
+
+/* Forward declarations */
+LOCAL(void) transdecode_master_selection(j_decompress_ptr cinfo);
+
+
+/*
+ * Read the coefficient arrays from a JPEG file.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * The entire image is read into a set of virtual coefficient-block arrays,
+ * one per component.  The return value is a pointer to the array of
+ * virtual-array descriptors.  These can be manipulated directly via the
+ * JPEG memory manager, or handed off to jpeg_write_coefficients().
+ * To release the memory occupied by the virtual arrays, call
+ * jpeg_finish_decompress() when done with the data.
+ *
+ * An alternative usage is to simply obtain access to the coefficient arrays
+ * during a buffered-image-mode decompression operation.  This is allowed
+ * after any jpeg_finish_output() call.  The arrays can be accessed until
+ * jpeg_finish_decompress() is called.  (Note that any call to the library
+ * may reposition the arrays, so don't rely on access_virt_barray() results
+ * to stay valid across library calls.)
+ *
+ * Returns NULL if suspended.  This case need be checked only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(jvirt_barray_ptr *)
+jpeg_read_coefficients(j_decompress_ptr cinfo)
+{
+  if (cinfo->master->lossless)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize active modules */
+    transdecode_master_selection(cinfo);
+    cinfo->global_state = DSTATE_RDCOEFS;
+  }
+  if (cinfo->global_state == DSTATE_RDCOEFS) {
+    /* Absorb whole file into the coef buffer */
+    for (;;) {
+      int retcode;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL)
+        (*cinfo->progress->progress_monitor) ((j_common_ptr)cinfo);
+      /* Absorb some more input */
+      retcode = (*cinfo->inputctl->consume_input) (cinfo);
+      if (retcode == JPEG_SUSPENDED)
+        return NULL;
+      if (retcode == JPEG_REACHED_EOI)
+        break;
+      /* Advance progress counter if appropriate */
+      if (cinfo->progress != NULL &&
+          (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+        if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+          /* startup underestimated number of scans; ratchet up one scan */
+          cinfo->progress->pass_limit += (long)cinfo->total_iMCU_rows;
+        }
+      }
+    }
+    /* Set state so that jpeg_finish_decompress does the right thing */
+    cinfo->global_state = DSTATE_STOPPING;
+  }
+  /* At this point we should be in state DSTATE_STOPPING if being used
+   * standalone, or in state DSTATE_BUFIMAGE if being invoked to get access
+   * to the coefficients during a full buffered-image-mode decompression.
+   */
+  if ((cinfo->global_state == DSTATE_STOPPING ||
+       cinfo->global_state == DSTATE_BUFIMAGE) && cinfo->buffered_image) {
+    return cinfo->coef->coef_arrays;
+  }
+  /* Oops, improper usage */
+  ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return NULL;                  /* keep compiler happy */
+}
+
+
+/*
+ * Master selection of decompression modules for transcoding.
+ * This substitutes for jdmaster.c's initialization of the full decompressor.
+ */
+
+LOCAL(void)
+transdecode_master_selection(j_decompress_ptr cinfo)
+{
+  /* This is effectively a buffered-image operation. */
+  cinfo->buffered_image = TRUE;
+
+#if JPEG_LIB_VERSION >= 80
+  /* Compute output image dimensions and related values. */
+  jpeg_core_output_dimensions(cinfo);
+#endif
+
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+#ifdef D_ARITH_CODING_SUPPORTED
+    jinit_arith_decoder(cinfo);
+#else
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+#endif
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef D_PROGRESSIVE_SUPPORTED
+      jinit_phuff_decoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_decoder(cinfo);
+  }
+
+  /* Always get a full-image coefficient buffer. */
+  if (cinfo->data_precision == 12)
+    j12init_d_coef_controller(cinfo, TRUE);
+  else
+    jinit_d_coef_controller(cinfo, TRUE);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr)cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+  /* Initialize progress monitoring. */
+  if (cinfo->progress != NULL) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else if (cinfo->inputctl->has_multiple_scans) {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    } else {
+      nscans = 1;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long)cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = 1;
+  }
+}

thirdparty/libjpeg-turbo/src/jerror.c

@@ -0,0 +1,243 @@
+/*
+ * jerror.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, 2024, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains simple error-reporting and trace-message routines.
+ * These are suitable for Unix-like systems and others where writing to
+ * stderr is the right thing to do.  Many applications will want to replace
+ * some or all of these routines.
+ *
+ * If you define USE_WINDOWS_MESSAGEBOX in jconfig.h or in the makefile,
+ * you get a Windows-specific hack to display error messages in a dialog box.
+ * It ain't much, but it beats dropping error messages into the bit bucket,
+ * which is what happens to output to stderr under most Windows C compilers.
+ *
+ * These routines are used by both the compression and decompression code.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jversion.h"
+#include "jerror.h"
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+#include <windows.h>
+#endif
+
+#ifndef EXIT_FAILURE            /* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+
+
+/*
+ * Create the message string table.
+ * We do this from the master message list in jerror.h by re-reading
+ * jerror.h with a suitable definition for macro JMESSAGE.
+ */
+
+#define JMESSAGE(code, string)  string,
+
+static const char * const jpeg_std_message_table[] = {
+#include "jerror.h"
+  NULL
+};
+
+
+/*
+ * Error exit handler: must not return to caller.
+ *
+ * Applications may override this if they want to get control back after
+ * an error.  Typically one would longjmp somewhere instead of exiting.
+ * The setjmp buffer can be made a private field within an expanded error
+ * handler object.  Note that the info needed to generate an error message
+ * is stored in the error object, so you can generate the message now or
+ * later, at your convenience.
+ * You should make sure that the JPEG object is cleaned up (with jpeg_abort
+ * or jpeg_destroy) at some point.
+ */
+
+METHODDEF(void)
+error_exit(j_common_ptr cinfo)
+{
+  /* Always display the message */
+  (*cinfo->err->output_message) (cinfo);
+
+  /* Let the memory manager delete any temp files before we die */
+  jpeg_destroy(cinfo);
+
+  exit(EXIT_FAILURE);
+}
+
+
+/*
+ * Actual output of an error or trace message.
+ * Applications may override this method to send JPEG messages somewhere
+ * other than stderr.
+ *
+ * On Windows, printing to stderr is generally completely useless,
+ * so we provide optional code to produce an error-dialog popup.
+ * Most Windows applications will still prefer to override this routine,
+ * but if they don't, it'll do something at least marginally useful.
+ *
+ * NOTE: to use the library in an environment that doesn't support the
+ * C stdio library, you may have to delete the call to fprintf() entirely,
+ * not just not use this routine.
+ */
+
+METHODDEF(void)
+output_message(j_common_ptr cinfo)
+{
+  char buffer[JMSG_LENGTH_MAX];
+
+  /* Create the message */
+  (*cinfo->err->format_message) (cinfo, buffer);
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+  /* Display it in a message dialog box */
+  MessageBox(GetActiveWindow(), buffer, "JPEG Library Error",
+             MB_OK | MB_ICONERROR);
+#else
+  /* Send it to stderr, adding a newline */
+  fprintf(stderr, "%s\n", buffer);
+#endif
+}
+
+
+/*
+ * Decide whether to emit a trace or warning message.
+ * msg_level is one of:
+ *   -1: recoverable corrupt-data warning, may want to abort.
+ *    0: important advisory messages (always display to user).
+ *    1: first level of tracing detail.
+ *    2,3,...: successively more detailed tracing messages.
+ * An application might override this method if it wanted to abort on warnings
+ * or change the policy about which messages to display.
+ */
+
+METHODDEF(void)
+emit_message(j_common_ptr cinfo, int msg_level)
+{
+  struct jpeg_error_mgr *err = cinfo->err;
+
+  if (msg_level < 0) {
+    /* It's a warning message.  Since corrupt files may generate many warnings,
+     * the policy implemented here is to show only the first warning,
+     * unless trace_level >= 3.
+     */
+    if (err->num_warnings == 0 || err->trace_level >= 3)
+      (*err->output_message) (cinfo);
+    /* Always count warnings in num_warnings. */
+    err->num_warnings++;
+  } else {
+    /* It's a trace message.  Show it if trace_level >= msg_level. */
+    if (err->trace_level >= msg_level)
+      (*err->output_message) (cinfo);
+  }
+}
+
+
+/*
+ * Format a message string for the most recent JPEG error or message.
+ * The message is stored into buffer, which should be at least JMSG_LENGTH_MAX
+ * characters.  Note that no '\n' character is added to the string.
+ * Few applications should need to override this method.
+ */
+
+METHODDEF(void)
+format_message(j_common_ptr cinfo, char *buffer)
+{
+  struct jpeg_error_mgr *err = cinfo->err;
+  int msg_code = err->msg_code;
+  const char *msgtext = NULL;
+  const char *msgptr;
+  char ch;
+  boolean isstring;
+
+  /* Look up message string in proper table */
+  if (msg_code > 0 && msg_code <= err->last_jpeg_message) {
+    msgtext = err->jpeg_message_table[msg_code];
+  } else if (err->addon_message_table != NULL &&
+             msg_code >= err->first_addon_message &&
+             msg_code <= err->last_addon_message) {
+    msgtext = err->addon_message_table[msg_code - err->first_addon_message];
+  }
+
+  /* Defend against bogus message number */
+  if (msgtext == NULL) {
+    err->msg_parm.i[0] = msg_code;
+    msgtext = err->jpeg_message_table[0];
+  }
+
+  /* Check for string parameter, as indicated by %s in the message text */
+  isstring = FALSE;
+  msgptr = msgtext;
+  while ((ch = *msgptr++) != '\0') {
+    if (ch == '%') {
+      if (*msgptr == 's') isstring = TRUE;
+      break;
+    }
+  }
+
+  /* Format the message into the passed buffer */
+  if (isstring)
+    SNPRINTF(buffer, JMSG_LENGTH_MAX, msgtext, err->msg_parm.s);
+  else
+    SNPRINTF(buffer, JMSG_LENGTH_MAX, msgtext,
+             err->msg_parm.i[0], err->msg_parm.i[1],
+             err->msg_parm.i[2], err->msg_parm.i[3],
+             err->msg_parm.i[4], err->msg_parm.i[5],
+             err->msg_parm.i[6], err->msg_parm.i[7]);
+}
+
+
+/*
+ * Reset error state variables at start of a new image.
+ * This is called during compression startup to reset trace/error
+ * processing to default state, without losing any application-specific
+ * method pointers.  An application might possibly want to override
+ * this method if it has additional error processing state.
+ */
+
+METHODDEF(void)
+reset_error_mgr(j_common_ptr cinfo)
+{
+  cinfo->err->num_warnings = 0;
+  /* trace_level is not reset since it is an application-supplied parameter */
+  cinfo->err->msg_code = 0;     /* may be useful as a flag for "no error" */
+}
+
+
+/*
+ * Fill in the standard error-handling methods in a jpeg_error_mgr object.
+ * Typical call is:
+ *      struct jpeg_compress_struct cinfo;
+ *      struct jpeg_error_mgr err;
+ *
+ *      cinfo.err = jpeg_std_error(&err);
+ * after which the application may override some of the methods.
+ */
+
+GLOBAL(struct jpeg_error_mgr *)
+jpeg_std_error(struct jpeg_error_mgr *err)
+{
+  memset(err, 0, sizeof(struct jpeg_error_mgr));
+
+  err->error_exit = error_exit;
+  err->emit_message = emit_message;
+  err->output_message = output_message;
+  err->format_message = format_message;
+  err->reset_error_mgr = reset_error_mgr;
+
+  /* Initialize message table pointers */
+  err->jpeg_message_table = jpeg_std_message_table;
+  err->last_jpeg_message = (int)JMSG_LASTMSGCODE - 1;
+
+  return err;
+}

thirdparty/libjpeg-turbo/src/jerror.h

@@ -0,0 +1,336 @@
+/*
+ * jerror.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 1997-2009 by Guido Vollbeding.
+ * Lossless JPEG Modifications:
+ * Copyright (C) 1999, Ken Murchison.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2014, 2017, 2021-2023, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file defines the error and message codes for the JPEG library.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ * A set of error-reporting macros are defined too.  Some applications using
+ * the JPEG library may wish to include this file to get the error codes
+ * and/or the macros.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef JERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code, string)
+#endif /* JERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code, string)  code,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_NOMESSAGE, "Bogus message code %d") /* Must be first entry! */
+
+/* For maintenance convenience, list is alphabetical by message code name */
+#if JPEG_LIB_VERSION < 70
+JMESSAGE(JERR_ARITH_NOTIMPL, "Sorry, arithmetic coding is not implemented")
+#endif
+JMESSAGE(JERR_BAD_ALIGN_TYPE, "ALIGN_TYPE is wrong, please fix")
+JMESSAGE(JERR_BAD_ALLOC_CHUNK, "MAX_ALLOC_CHUNK is wrong, please fix")
+JMESSAGE(JERR_BAD_BUFFER_MODE, "Bogus buffer control mode")
+JMESSAGE(JERR_BAD_COMPONENT_ID, "Invalid component ID %d in SOS")
+#if JPEG_LIB_VERSION >= 70
+JMESSAGE(JERR_BAD_CROP_SPEC, "Invalid crop request")
+#endif
+JMESSAGE(JERR_BAD_DCT_COEF,
+         "DCT coefficient (lossy) or spatial difference (lossless) out of range")
+JMESSAGE(JERR_BAD_DCTSIZE, "IDCT output block size %d not supported")
+#if JPEG_LIB_VERSION >= 70
+JMESSAGE(JERR_BAD_DROP_SAMPLING,
+         "Component index %d: mismatching sampling ratio %d:%d, %d:%d, %c")
+#endif
+JMESSAGE(JERR_BAD_HUFF_TABLE, "Bogus Huffman table definition")
+JMESSAGE(JERR_BAD_IN_COLORSPACE, "Bogus input colorspace")
+JMESSAGE(JERR_BAD_J_COLORSPACE, "Bogus JPEG colorspace")
+JMESSAGE(JERR_BAD_LENGTH, "Bogus marker length")
+JMESSAGE(JERR_BAD_LIB_VERSION,
+         "Wrong JPEG library version: library is %d, caller expects %d")
+JMESSAGE(JERR_BAD_MCU_SIZE, "Sampling factors too large for interleaved scan")
+JMESSAGE(JERR_BAD_POOL_ID, "Invalid memory pool code %d")
+JMESSAGE(JERR_BAD_PRECISION, "Unsupported JPEG data precision %d")
+JMESSAGE(JERR_BAD_PROGRESSION,
+         "Invalid progressive/lossless parameters Ss=%d Se=%d Ah=%d Al=%d")
+JMESSAGE(JERR_BAD_PROG_SCRIPT,
+         "Invalid progressive/lossless parameters at scan script entry %d")
+JMESSAGE(JERR_BAD_SAMPLING, "Bogus sampling factors")
+JMESSAGE(JERR_BAD_SCAN_SCRIPT, "Invalid scan script at entry %d")
+JMESSAGE(JERR_BAD_STATE, "Improper call to JPEG library in state %d")
+JMESSAGE(JERR_BAD_STRUCT_SIZE,
+         "JPEG parameter struct mismatch: library thinks size is %u, caller expects %u")
+JMESSAGE(JERR_BAD_VIRTUAL_ACCESS, "Bogus virtual array access")
+JMESSAGE(JERR_BUFFER_SIZE, "Buffer passed to JPEG library is too small")
+JMESSAGE(JERR_CANT_SUSPEND, "Suspension not allowed here")
+JMESSAGE(JERR_CCIR601_NOTIMPL, "CCIR601 sampling not implemented yet")
+JMESSAGE(JERR_COMPONENT_COUNT, "Too many color components: %d, max %d")
+JMESSAGE(JERR_CONVERSION_NOTIMPL, "Unsupported color conversion request")
+JMESSAGE(JERR_DAC_INDEX, "Bogus DAC index %d")
+JMESSAGE(JERR_DAC_VALUE, "Bogus DAC value 0x%x")
+JMESSAGE(JERR_DHT_INDEX, "Bogus DHT index %d")
+JMESSAGE(JERR_DQT_INDEX, "Bogus DQT index %d")
+JMESSAGE(JERR_EMPTY_IMAGE, "Empty JPEG image (DNL not supported)")
+JMESSAGE(JERR_EMS_READ, "Read from EMS failed")
+JMESSAGE(JERR_EMS_WRITE, "Write to EMS failed")
+JMESSAGE(JERR_EOI_EXPECTED, "Didn't expect more than one scan")
+JMESSAGE(JERR_FILE_READ, "Input file read error")
+JMESSAGE(JERR_FILE_WRITE, "Output file write error --- out of disk space?")
+JMESSAGE(JERR_FRACT_SAMPLE_NOTIMPL, "Fractional sampling not implemented yet")
+JMESSAGE(JERR_HUFF_CLEN_OVERFLOW, "Huffman code size table overflow")
+JMESSAGE(JERR_HUFF_MISSING_CODE, "Missing Huffman code table entry")
+JMESSAGE(JERR_IMAGE_TOO_BIG, "Maximum supported image dimension is %u pixels")
+JMESSAGE(JERR_INPUT_EMPTY, "Empty input file")
+JMESSAGE(JERR_INPUT_EOF, "Premature end of input file")
+JMESSAGE(JERR_MISMATCHED_QUANT_TABLE,
+         "Cannot transcode due to multiple use of quantization table %d")
+JMESSAGE(JERR_MISSING_DATA, "Scan script does not transmit all data")
+JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
+JMESSAGE(JERR_NOTIMPL, "Requested features are incompatible")
+JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
+#if JPEG_LIB_VERSION >= 70
+JMESSAGE(JERR_NO_ARITH_TABLE, "Arithmetic table 0x%02x was not defined")
+#endif
+JMESSAGE(JERR_NO_BACKING_STORE, "Memory limit exceeded")
+JMESSAGE(JERR_NO_HUFF_TABLE, "Huffman table 0x%02x was not defined")
+JMESSAGE(JERR_NO_IMAGE, "JPEG datastream contains no image")
+JMESSAGE(JERR_NO_QUANT_TABLE, "Quantization table 0x%02x was not defined")
+JMESSAGE(JERR_NO_SOI, "Not a JPEG file: starts with 0x%02x 0x%02x")
+JMESSAGE(JERR_OUT_OF_MEMORY, "Insufficient memory (case %d)")
+JMESSAGE(JERR_QUANT_COMPONENTS,
+         "Cannot quantize more than %d color components")
+JMESSAGE(JERR_QUANT_FEW_COLORS, "Cannot quantize to fewer than %d colors")
+JMESSAGE(JERR_QUANT_MANY_COLORS, "Cannot quantize to more than %d colors")
+JMESSAGE(JERR_SOF_DUPLICATE, "Invalid JPEG file structure: two SOF markers")
+JMESSAGE(JERR_SOF_NO_SOS, "Invalid JPEG file structure: missing SOS marker")
+JMESSAGE(JERR_SOF_UNSUPPORTED, "Unsupported JPEG process: SOF type 0x%02x")
+JMESSAGE(JERR_SOI_DUPLICATE, "Invalid JPEG file structure: two SOI markers")
+JMESSAGE(JERR_SOS_NO_SOF, "Invalid JPEG file structure: SOS before SOF")
+JMESSAGE(JERR_TFILE_CREATE, "Failed to create temporary file %s")
+JMESSAGE(JERR_TFILE_READ, "Read failed on temporary file")
+JMESSAGE(JERR_TFILE_SEEK, "Seek failed on temporary file")
+JMESSAGE(JERR_TFILE_WRITE,
+         "Write failed on temporary file --- out of disk space?")
+JMESSAGE(JERR_TOO_LITTLE_DATA, "Application transferred too few scanlines")
+JMESSAGE(JERR_UNKNOWN_MARKER, "Unsupported marker type 0x%02x")
+JMESSAGE(JERR_VIRTUAL_BUG, "Virtual array controller messed up")
+JMESSAGE(JERR_WIDTH_OVERFLOW, "Image too wide for this implementation")
+JMESSAGE(JERR_XMS_READ, "Read from XMS failed")
+JMESSAGE(JERR_XMS_WRITE, "Write to XMS failed")
+JMESSAGE(JMSG_COPYRIGHT, JCOPYRIGHT_SHORT)
+JMESSAGE(JMSG_VERSION, JVERSION)
+JMESSAGE(JTRC_16BIT_TABLES,
+         "Caution: quantization tables are too coarse for baseline JPEG")
+JMESSAGE(JTRC_ADOBE,
+         "Adobe APP14 marker: version %d, flags 0x%04x 0x%04x, transform %d")
+JMESSAGE(JTRC_APP0, "Unknown APP0 marker (not JFIF), length %u")
+JMESSAGE(JTRC_APP14, "Unknown APP14 marker (not Adobe), length %u")
+JMESSAGE(JTRC_DAC, "Define Arithmetic Table 0x%02x: 0x%02x")
+JMESSAGE(JTRC_DHT, "Define Huffman Table 0x%02x")
+JMESSAGE(JTRC_DQT, "Define Quantization Table %d  precision %d")
+JMESSAGE(JTRC_DRI, "Define Restart Interval %u")
+JMESSAGE(JTRC_EMS_CLOSE, "Freed EMS handle %u")
+JMESSAGE(JTRC_EMS_OPEN, "Obtained EMS handle %u")
+JMESSAGE(JTRC_EOI, "End Of Image")
+JMESSAGE(JTRC_HUFFBITS, "        %3d %3d %3d %3d %3d %3d %3d %3d")
+JMESSAGE(JTRC_JFIF, "JFIF APP0 marker: version %d.%02d, density %dx%d  %d")
+JMESSAGE(JTRC_JFIF_BADTHUMBNAILSIZE,
+         "Warning: thumbnail image size does not match data length %u")
+JMESSAGE(JTRC_JFIF_EXTENSION, "JFIF extension marker: type 0x%02x, length %u")
+JMESSAGE(JTRC_JFIF_THUMBNAIL, "    with %d x %d thumbnail image")
+JMESSAGE(JTRC_MISC_MARKER, "Miscellaneous marker 0x%02x, length %u")
+JMESSAGE(JTRC_PARMLESS_MARKER, "Unexpected marker 0x%02x")
+JMESSAGE(JTRC_QUANTVALS, "        %4u %4u %4u %4u %4u %4u %4u %4u")
+JMESSAGE(JTRC_QUANT_3_NCOLORS, "Quantizing to %d = %d*%d*%d colors")
+JMESSAGE(JTRC_QUANT_NCOLORS, "Quantizing to %d colors")
+JMESSAGE(JTRC_QUANT_SELECTED, "Selected %d colors for quantization")
+JMESSAGE(JTRC_RECOVERY_ACTION, "At marker 0x%02x, recovery action %d")
+JMESSAGE(JTRC_RST, "RST%d")
+JMESSAGE(JTRC_SMOOTH_NOTIMPL,
+         "Smoothing not supported with nonstandard sampling ratios")
+JMESSAGE(JTRC_SOF, "Start Of Frame 0x%02x: width=%u, height=%u, components=%d")
+JMESSAGE(JTRC_SOF_COMPONENT, "    Component %d: %dhx%dv q=%d")
+JMESSAGE(JTRC_SOI, "Start of Image")
+JMESSAGE(JTRC_SOS, "Start Of Scan: %d components")
+JMESSAGE(JTRC_SOS_COMPONENT, "    Component %d: dc=%d ac=%d")
+JMESSAGE(JTRC_SOS_PARAMS, "  Ss=%d, Se=%d, Ah=%d, Al=%d")
+JMESSAGE(JTRC_TFILE_CLOSE, "Closed temporary file %s")
+JMESSAGE(JTRC_TFILE_OPEN, "Opened temporary file %s")
+JMESSAGE(JTRC_THUMB_JPEG,
+         "JFIF extension marker: JPEG-compressed thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_PALETTE,
+         "JFIF extension marker: palette thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_RGB,
+         "JFIF extension marker: RGB thumbnail image, length %u")
+JMESSAGE(JTRC_UNKNOWN_IDS,
+         "Unrecognized component IDs %d %d %d, assuming YCbCr (lossy) or RGB (lossless)")
+JMESSAGE(JTRC_XMS_CLOSE, "Freed XMS handle %u")
+JMESSAGE(JTRC_XMS_OPEN, "Obtained XMS handle %u")
+JMESSAGE(JWRN_ADOBE_XFORM, "Unknown Adobe color transform code %d")
+#if JPEG_LIB_VERSION >= 70
+JMESSAGE(JWRN_ARITH_BAD_CODE, "Corrupt JPEG data: bad arithmetic code")
+#endif
+JMESSAGE(JWRN_BOGUS_PROGRESSION,
+         "Inconsistent progression sequence for component %d coefficient %d")
+JMESSAGE(JWRN_EXTRANEOUS_DATA,
+         "Corrupt JPEG data: %u extraneous bytes before marker 0x%02x")
+JMESSAGE(JWRN_HIT_MARKER, "Corrupt JPEG data: premature end of data segment")
+JMESSAGE(JWRN_HUFF_BAD_CODE, "Corrupt JPEG data: bad Huffman code")
+JMESSAGE(JWRN_JFIF_MAJOR, "Warning: unknown JFIF revision number %d.%02d")
+JMESSAGE(JWRN_JPEG_EOF, "Premature end of JPEG file")
+JMESSAGE(JWRN_MUST_RESYNC,
+         "Corrupt JPEG data: found marker 0x%02x instead of RST%d")
+JMESSAGE(JWRN_NOT_SEQUENTIAL, "Invalid SOS parameters for sequential JPEG")
+JMESSAGE(JWRN_TOO_MUCH_DATA, "Application transferred too many scanlines")
+#if JPEG_LIB_VERSION < 70
+JMESSAGE(JERR_BAD_CROP_SPEC, "Invalid crop request")
+#if defined(C_ARITH_CODING_SUPPORTED) || defined(D_ARITH_CODING_SUPPORTED)
+JMESSAGE(JERR_NO_ARITH_TABLE, "Arithmetic table 0x%02x was not defined")
+JMESSAGE(JWRN_ARITH_BAD_CODE, "Corrupt JPEG data: bad arithmetic code")
+#endif
+#endif
+JMESSAGE(JWRN_BOGUS_ICC, "Corrupt JPEG data: bad ICC marker")
+#if JPEG_LIB_VERSION < 70
+JMESSAGE(JERR_BAD_DROP_SAMPLING,
+         "Component index %d: mismatching sampling ratio %d:%d, %d:%d, %c")
+#endif
+JMESSAGE(JERR_BAD_RESTART,
+         "Invalid restart interval %d; must be an integer multiple of the number of MCUs in an MCU row (%d)")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTMSGCODE
+} J_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE
+
+
+#ifndef JERROR_H
+#define JERROR_H
+
+/* Macros to simplify using the error and trace message stuff */
+/* The first parameter is either type of cinfo pointer */
+
+/* Fatal errors (print message and exit) */
+#define ERREXIT(cinfo, code) \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT1(cinfo, code, p1) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT2(cinfo, code, p1, p2) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT3(cinfo, code, p1, p2, p3) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT4(cinfo, code, p1, p2, p3, p4) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT6(cinfo, code, p1, p2, p3, p4, p5, p6) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (cinfo)->err->msg_parm.i[4] = (p5), \
+   (cinfo)->err->msg_parm.i[5] = (p6), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXITS(cinfo, code, str) \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (cinfo)->err->msg_parm.s[JMSG_STR_PARM_MAX - 1] = '\0', \
+   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+
+#define MAKESTMT(stuff)         do { stuff } while (0)
+
+/* Nonfatal errors (we can keep going, but the data is probably corrupt) */
+#define WARNMS(cinfo, code) \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), -1))
+#define WARNMS1(cinfo, code, p1) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), -1))
+#define WARNMS2(cinfo, code, p1, p2) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), -1))
+
+/* Informational/debugging messages */
+#define TRACEMS(cinfo, lvl, code) \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)))
+#define TRACEMS1(cinfo, lvl, code, p1) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)))
+#define TRACEMS2(cinfo, lvl, code, p1, p2) \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)))
+#define TRACEMS3(cinfo, lvl, code, p1, p2, p3) \
+  MAKESTMT(int *_mp = (cinfo)->err->msg_parm.i; \
+           _mp[0] = (p1);  _mp[1] = (p2);  _mp[2] = (p3); \
+           (cinfo)->err->msg_code = (code); \
+           (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)); )
+#define TRACEMS4(cinfo, lvl, code, p1, p2, p3, p4) \
+  MAKESTMT(int *_mp = (cinfo)->err->msg_parm.i; \
+           _mp[0] = (p1);  _mp[1] = (p2);  _mp[2] = (p3);  _mp[3] = (p4); \
+           (cinfo)->err->msg_code = (code); \
+           (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)); )
+#define TRACEMS5(cinfo, lvl, code, p1, p2, p3, p4, p5) \
+  MAKESTMT(int *_mp = (cinfo)->err->msg_parm.i; \
+           _mp[0] = (p1);  _mp[1] = (p2);  _mp[2] = (p3);  _mp[3] = (p4); \
+           _mp[4] = (p5); \
+           (cinfo)->err->msg_code = (code); \
+           (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)); )
+#define TRACEMS8(cinfo, lvl, code, p1, p2, p3, p4, p5, p6, p7, p8) \
+  MAKESTMT(int *_mp = (cinfo)->err->msg_parm.i; \
+           _mp[0] = (p1);  _mp[1] = (p2);  _mp[2] = (p3);  _mp[3] = (p4); \
+           _mp[4] = (p5);  _mp[5] = (p6);  _mp[6] = (p7);  _mp[7] = (p8); \
+           (cinfo)->err->msg_code = (code); \
+           (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)); )
+#define TRACEMSS(cinfo, lvl, code, str) \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (cinfo)->err->msg_parm.s[JMSG_STR_PARM_MAX - 1] = '\0', \
+   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)))
+
+#endif /* JERROR_H */

thirdparty/libjpeg-turbo/src/jfdctflt.c

@@ -0,0 +1,169 @@
+/*
+ * jfdctflt.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a floating-point implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * This implementation should be more accurate than either of the integer
+ * DCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README.ijg).  The following
+ * code is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_float(FAST_FLOAT *data)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
+  FAST_FLOAT *dataptr;
+  int ctr;
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;        /* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT)0.707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;    /* phase 5 */
+    dataptr[6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;        /* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT)0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT)0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT)1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT)0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;            /* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;      /* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;         /* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
+    tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
+    tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
+    tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
+    tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
+    tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
+    tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
+    tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;        /* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[DCTSIZE * 0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE * 4] = tmp10 - tmp11;
+
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT)0.707106781); /* c4 */
+    dataptr[DCTSIZE * 2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE * 6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;        /* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT)0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT)0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT)1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT)0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;            /* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE * 5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE * 3] = z13 - z2;
+    dataptr[DCTSIZE * 1] = z11 + z4;
+    dataptr[DCTSIZE * 7] = z11 - z4;
+
+    dataptr++;                  /* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */

thirdparty/libjpeg-turbo/src/jfdctfst.c

@@ -0,0 +1,227 @@
+/*
+ * jfdctfst.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README.ijg).  The following
+ * code is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jfdctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * Again to save a few shifts, the intermediate results between pass 1 and
+ * pass 2 are not upscaled, but are represented only to integral precision.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#define CONST_BITS  8
+
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_0_382683433  ((JLONG)98)            /* FIX(0.382683433) */
+#define FIX_0_541196100  ((JLONG)139)           /* FIX(0.541196100) */
+#define FIX_0_707106781  ((JLONG)181)           /* FIX(0.707106781) */
+#define FIX_1_306562965  ((JLONG)334)           /* FIX(1.306562965) */
+#else
+#define FIX_0_382683433  FIX(0.382683433)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_707106781  FIX(0.707106781)
+#define FIX_1_306562965  FIX(1.306562965)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x, n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an JLONG constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var, const)  ((DCTELEM)DESCALE((var) * (const), CONST_BITS))
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+_jpeg_fdct_ifast(DCTELEM *data)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z1, z2, z3, z4, z5, z11, z13;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;        /* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;    /* phase 5 */
+    dataptr[6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;        /* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;            /* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;      /* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;         /* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
+    tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
+    tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
+    tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
+    tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
+    tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
+    tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
+    tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;        /* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[DCTSIZE * 0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE * 4] = tmp10 - tmp11;
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[DCTSIZE * 2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE * 6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;        /* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;            /* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE * 5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE * 3] = z13 - z2;
+    dataptr[DCTSIZE * 1] = z11 + z4;
+    dataptr[DCTSIZE * 7] = z11 - z4;
+
+    dataptr++;                  /* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */

thirdparty/libjpeg-turbo/src/jfdctint.c

@@ -0,0 +1,288 @@
+/*
+ * jfdctint.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a slower but more accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true DCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D DCT,
+ * because the y0 and y4 outputs need not be divided by sqrt(N).
+ * In the IJG code, this factor of 8 is removed by the quantization step
+ * (in jcdctmgr.c), NOT in this module.
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (For 12-bit sample data, the intermediate
+ * array is JLONG anyway.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1           /* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((JLONG)2446)          /* FIX(0.298631336) */
+#define FIX_0_390180644  ((JLONG)3196)          /* FIX(0.390180644) */
+#define FIX_0_541196100  ((JLONG)4433)          /* FIX(0.541196100) */
+#define FIX_0_765366865  ((JLONG)6270)          /* FIX(0.765366865) */
+#define FIX_0_899976223  ((JLONG)7373)          /* FIX(0.899976223) */
+#define FIX_1_175875602  ((JLONG)9633)          /* FIX(1.175875602) */
+#define FIX_1_501321110  ((JLONG)12299)         /* FIX(1.501321110) */
+#define FIX_1_847759065  ((JLONG)15137)         /* FIX(1.847759065) */
+#define FIX_1_961570560  ((JLONG)16069)         /* FIX(1.961570560) */
+#define FIX_2_053119869  ((JLONG)16819)         /* FIX(2.053119869) */
+#define FIX_2_562915447  ((JLONG)20995)         /* FIX(2.562915447) */
+#define FIX_3_072711026  ((JLONG)25172)         /* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an JLONG variable by an JLONG constant to yield an JLONG result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var, const)  MULTIPLY16C16(var, const)
+#else
+#define MULTIPLY(var, const)  ((var) * (const))
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+_jpeg_fdct_islow(DCTELEM *data)
+{
+  JLONG tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  JLONG tmp10, tmp11, tmp12, tmp13;
+  JLONG z1, z2, z3, z4, z5;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[0] = (DCTELEM)LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS);
+    dataptr[4] = (DCTELEM)LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+                                  CONST_BITS - PASS1_BITS);
+    dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, -FIX_1_847759065),
+                                  CONST_BITS - PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, -FIX_0_899976223); /* sqrt(2) * ( c7-c3) */
+    z2 = MULTIPLY(z2, -FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, -FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, -FIX_0_390180644); /* sqrt(2) * ( c5-c3) */
+
+    z3 += z5;
+    z4 += z5;
+
+    dataptr[7] = (DCTELEM)DESCALE(tmp4 + z1 + z3, CONST_BITS - PASS1_BITS);
+    dataptr[5] = (DCTELEM)DESCALE(tmp5 + z2 + z4, CONST_BITS - PASS1_BITS);
+    dataptr[3] = (DCTELEM)DESCALE(tmp6 + z2 + z3, CONST_BITS - PASS1_BITS);
+    dataptr[1] = (DCTELEM)DESCALE(tmp7 + z1 + z4, CONST_BITS - PASS1_BITS);
+
+    dataptr += DCTSIZE;         /* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
+    tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
+    tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
+    tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
+    tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
+    tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
+    tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
+    tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
+
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[DCTSIZE * 0] = (DCTELEM)DESCALE(tmp10 + tmp11, PASS1_BITS);
+    dataptr[DCTSIZE * 4] = (DCTELEM)DESCALE(tmp10 - tmp11, PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[DCTSIZE * 2] =
+      (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+                       CONST_BITS + PASS1_BITS);
+    dataptr[DCTSIZE * 6] =
+      (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, -FIX_1_847759065),
+                       CONST_BITS + PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, -FIX_0_899976223); /* sqrt(2) * ( c7-c3) */
+    z2 = MULTIPLY(z2, -FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, -FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, -FIX_0_390180644); /* sqrt(2) * ( c5-c3) */
+
+    z3 += z5;
+    z4 += z5;
+
+    dataptr[DCTSIZE * 7] = (DCTELEM)DESCALE(tmp4 + z1 + z3,
+                                            CONST_BITS + PASS1_BITS);
+    dataptr[DCTSIZE * 5] = (DCTELEM)DESCALE(tmp5 + z2 + z4,
+                                            CONST_BITS + PASS1_BITS);
+    dataptr[DCTSIZE * 3] = (DCTELEM)DESCALE(tmp6 + z2 + z3,
+                                            CONST_BITS + PASS1_BITS);
+    dataptr[DCTSIZE * 1] = (DCTELEM)DESCALE(tmp7 + z1 + z4,
+                                            CONST_BITS + PASS1_BITS);
+
+    dataptr++;                  /* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_ISLOW_SUPPORTED */

thirdparty/libjpeg-turbo/src/jidctflt.c

@@ -0,0 +1,240 @@
+/*
+ * jidctflt.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2010 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2014, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a floating-point implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * This implementation should be more accurate than either of the integer
+ * IDCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README.ijg).  The following
+ * code is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a float result.
+ */
+
+#define DEQUANTIZE(coef, quantval)  (((FAST_FLOAT)(coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+_jpeg_idct_float(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  FLOAT_MULT_TYPE *quantptr;
+  FAST_FLOAT *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = (_JSAMPLE *)cinfo->sample_range_limit;
+  int ctr;
+  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
+#define _0_125  ((FLOAT_MULT_TYPE)0.125)
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (FLOAT_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+
+    if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
+        inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
+        inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
+        inptr[DCTSIZE * 7] == 0) {
+      /* AC terms all zero */
+      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE * 0],
+                                    quantptr[DCTSIZE * 0] * _0_125);
+
+      wsptr[DCTSIZE * 0] = dcval;
+      wsptr[DCTSIZE * 1] = dcval;
+      wsptr[DCTSIZE * 2] = dcval;
+      wsptr[DCTSIZE * 3] = dcval;
+      wsptr[DCTSIZE * 4] = dcval;
+      wsptr[DCTSIZE * 5] = dcval;
+      wsptr[DCTSIZE * 6] = dcval;
+      wsptr[DCTSIZE * 7] = dcval;
+
+      inptr++;                  /* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0] * _0_125);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2] * _0_125);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4] * _0_125);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6] * _0_125);
+
+    tmp10 = tmp0 + tmp2;        /* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;        /* phases 5-3 */
+    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT)1.414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;       /* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1] * _0_125);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3] * _0_125);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5] * _0_125);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7] * _0_125);
+
+    z13 = tmp6 + tmp5;          /* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;           /* phase 5 */
+    tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562); /* 2*c4 */
+
+    z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
+    tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
+    tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;        /* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 - tmp5;
+
+    wsptr[DCTSIZE * 0] = tmp0 + tmp7;
+    wsptr[DCTSIZE * 7] = tmp0 - tmp7;
+    wsptr[DCTSIZE * 1] = tmp1 + tmp6;
+    wsptr[DCTSIZE * 6] = tmp1 - tmp6;
+    wsptr[DCTSIZE * 2] = tmp2 + tmp5;
+    wsptr[DCTSIZE * 5] = tmp2 - tmp5;
+    wsptr[DCTSIZE * 3] = tmp3 + tmp4;
+    wsptr[DCTSIZE * 4] = tmp3 - tmp4;
+
+    inptr++;                    /* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+
+  /* Pass 2: process rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * And testing floats for zero is relatively expensive, so we don't bother.
+     */
+
+    /* Even part */
+
+    /* Apply signed->unsigned and prepare float->int conversion */
+    z5 = wsptr[0] + ((FAST_FLOAT)_CENTERJSAMPLE + (FAST_FLOAT)0.5);
+    tmp10 = z5 + wsptr[4];
+    tmp11 = z5 - wsptr[4];
+
+    tmp13 = wsptr[2] + wsptr[6];
+    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT)1.414213562) - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = wsptr[5] + wsptr[3];
+    z10 = wsptr[5] - wsptr[3];
+    z11 = wsptr[1] + wsptr[7];
+    z12 = wsptr[1] - wsptr[7];
+
+    tmp7 = z11 + z13;
+    tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562);
+
+    z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
+    tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
+    tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 - tmp5;
+
+    /* Final output stage: float->int conversion and range-limit */
+
+    outptr[0] = range_limit[((int)(tmp0 + tmp7)) & RANGE_MASK];
+    outptr[7] = range_limit[((int)(tmp0 - tmp7)) & RANGE_MASK];
+    outptr[1] = range_limit[((int)(tmp1 + tmp6)) & RANGE_MASK];
+    outptr[6] = range_limit[((int)(tmp1 - tmp6)) & RANGE_MASK];
+    outptr[2] = range_limit[((int)(tmp2 + tmp5)) & RANGE_MASK];
+    outptr[5] = range_limit[((int)(tmp2 - tmp5)) & RANGE_MASK];
+    outptr[3] = range_limit[((int)(tmp3 + tmp4)) & RANGE_MASK];
+    outptr[4] = range_limit[((int)(tmp3 - tmp4)) & RANGE_MASK];
+
+    wsptr += DCTSIZE;           /* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */

thirdparty/libjpeg-turbo/src/jidctfst.c

@@ -0,0 +1,371 @@
+/*
+ * jidctfst.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README.ijg).  The following
+ * code is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jidctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * The dequantized coefficients are not integers because the AA&N scaling
+ * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
+ * so that the first and second IDCT rounds have the same input scaling.
+ * For 8-bit samples, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
+ * avoid a descaling shift; this compromises accuracy rather drastically
+ * for small quantization table entries, but it saves a lot of shifts.
+ * For 12-bit samples, there's no hope of using 16x16 multiplies anyway,
+ * so we use a much larger scaling factor to preserve accuracy.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  8
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  8
+#define PASS1_BITS  1           /* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_1_082392200  ((JLONG)277)           /* FIX(1.082392200) */
+#define FIX_1_414213562  ((JLONG)362)           /* FIX(1.414213562) */
+#define FIX_1_847759065  ((JLONG)473)           /* FIX(1.847759065) */
+#define FIX_2_613125930  ((JLONG)669)           /* FIX(2.613125930) */
+#else
+#define FIX_1_082392200  FIX(1.082392200)
+#define FIX_1_414213562  FIX(1.414213562)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_613125930  FIX(2.613125930)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x, n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an JLONG constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var, const)  ((DCTELEM)DESCALE((var) * (const), CONST_BITS))
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
+ * multiplication will do.  For 12-bit data, the multiplier table is
+ * declared JLONG, so a 32-bit multiply will be used.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define DEQUANTIZE(coef, quantval)  (((IFAST_MULT_TYPE)(coef)) * (quantval))
+#else
+#define DEQUANTIZE(coef, quantval) \
+  DESCALE((coef) * (quantval), IFAST_SCALE_BITS - PASS1_BITS)
+#endif
+
+
+/* Like DESCALE, but applies to a DCTELEM and produces an int.
+ * We assume that int right shift is unsigned if JLONG right shift is.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS    DCTELEM ishift_temp;
+#if BITS_IN_JSAMPLE == 8
+#define DCTELEMBITS  16         /* DCTELEM may be 16 or 32 bits */
+#else
+#define DCTELEMBITS  32         /* DCTELEM must be 32 bits */
+#endif
+#define IRIGHT_SHIFT(x, shft) \
+  ((ishift_temp = (x)) < 0 ? \
+   (ishift_temp >> (shft)) | ((~((DCTELEM)0)) << (DCTELEMBITS - (shft))) : \
+   (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x, shft)   ((x) >> (shft))
+#endif
+
+#ifdef USE_ACCURATE_ROUNDING
+#define IDESCALE(x, n)  ((int)IRIGHT_SHIFT((x) + (1 << ((n) - 1)), n))
+#else
+#define IDESCALE(x, n)  ((int)IRIGHT_SHIFT(x, n))
+#endif
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+_jpeg_idct_ifast(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  IFAST_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];      /* buffers data between passes */
+  SHIFT_TEMPS                   /* for DESCALE */
+  ISHIFT_TEMPS                  /* for IDESCALE */
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (IFAST_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+
+    if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
+        inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
+        inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
+        inptr[DCTSIZE * 7] == 0) {
+      /* AC terms all zero */
+      int dcval = (int)DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+
+      wsptr[DCTSIZE * 0] = dcval;
+      wsptr[DCTSIZE * 1] = dcval;
+      wsptr[DCTSIZE * 2] = dcval;
+      wsptr[DCTSIZE * 3] = dcval;
+      wsptr[DCTSIZE * 4] = dcval;
+      wsptr[DCTSIZE * 5] = dcval;
+      wsptr[DCTSIZE * 6] = dcval;
+      wsptr[DCTSIZE * 7] = dcval;
+
+      inptr++;                  /* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp10 = tmp0 + tmp2;        /* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;        /* phases 5-3 */
+    tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;       /* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    z13 = tmp6 + tmp5;          /* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;           /* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, -FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;        /* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE * 0] = (int)(tmp0 + tmp7);
+    wsptr[DCTSIZE * 7] = (int)(tmp0 - tmp7);
+    wsptr[DCTSIZE * 1] = (int)(tmp1 + tmp6);
+    wsptr[DCTSIZE * 6] = (int)(tmp1 - tmp6);
+    wsptr[DCTSIZE * 2] = (int)(tmp2 + tmp5);
+    wsptr[DCTSIZE * 5] = (int)(tmp2 - tmp5);
+    wsptr[DCTSIZE * 4] = (int)(tmp3 + tmp4);
+    wsptr[DCTSIZE * 3] = (int)(tmp3 - tmp4);
+
+    inptr++;                    /* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+        wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      _JSAMPLE dcval =
+        range_limit[IDESCALE(wsptr[0], PASS1_BITS + 3) & RANGE_MASK];
+
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;         /* advance pointer to next row */
+      continue;
+    }
+#endif
+
+    /* Even part */
+
+    tmp10 = ((DCTELEM)wsptr[0] + (DCTELEM)wsptr[4]);
+    tmp11 = ((DCTELEM)wsptr[0] - (DCTELEM)wsptr[4]);
+
+    tmp13 = ((DCTELEM)wsptr[2] + (DCTELEM)wsptr[6]);
+    tmp12 =
+      MULTIPLY((DCTELEM)wsptr[2] - (DCTELEM)wsptr[6], FIX_1_414213562) - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = (DCTELEM)wsptr[5] + (DCTELEM)wsptr[3];
+    z10 = (DCTELEM)wsptr[5] - (DCTELEM)wsptr[3];
+    z11 = (DCTELEM)wsptr[1] + (DCTELEM)wsptr[7];
+    z12 = (DCTELEM)wsptr[1] - (DCTELEM)wsptr[7];
+
+    tmp7 = z11 + z13;           /* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, -FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;        /* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] =
+      range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[7] =
+      range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[1] =
+      range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[6] =
+      range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[2] =
+      range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[5] =
+      range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[4] =
+      range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS + 3) & RANGE_MASK];
+    outptr[3] =
+      range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS + 3) & RANGE_MASK];
+
+    wsptr += DCTSIZE;           /* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */

thirdparty/libjpeg-turbo/src/jidctint.c

@@ -0,0 +1,2627 @@
+/*
+ * jidctint.c
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modification developed 2002-2018 by Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2015, 2020, 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains a slower but more accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ *
+ * We also provide IDCT routines with various output sample block sizes for
+ * direct resolution reduction or enlargement without additional resampling:
+ * NxN (N=1...16) pixels for one 8x8 input DCT block.
+ *
+ * For N<8 we simply take the corresponding low-frequency coefficients of
+ * the 8x8 input DCT block and apply an NxN point IDCT on the sub-block
+ * to yield the downscaled outputs.
+ * This can be seen as direct low-pass downsampling from the DCT domain
+ * point of view rather than the usual spatial domain point of view,
+ * yielding significant computational savings and results at least
+ * as good as common bilinear (averaging) spatial downsampling.
+ *
+ * For N>8 we apply a partial NxN IDCT on the 8 input coefficients as
+ * lower frequencies and higher frequencies assumed to be zero.
+ * It turns out that the computational effort is similar to the 8x8 IDCT
+ * regarding the output size.
+ * Furthermore, the scaling and descaling is the same for all IDCT sizes.
+ *
+ * CAUTION: We rely on the FIX() macro except for the N=1,2,4,8 cases
+ * since there would be too many additional constants to pre-calculate.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"               /* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true IDCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D IDCT,
+ * because the y0 and y4 inputs need not be divided by sqrt(N).
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (To scale up 12-bit sample data further, an
+ * intermediate JLONG array would be needed.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1           /* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((JLONG)2446)          /* FIX(0.298631336) */
+#define FIX_0_390180644  ((JLONG)3196)          /* FIX(0.390180644) */
+#define FIX_0_541196100  ((JLONG)4433)          /* FIX(0.541196100) */
+#define FIX_0_765366865  ((JLONG)6270)          /* FIX(0.765366865) */
+#define FIX_0_899976223  ((JLONG)7373)          /* FIX(0.899976223) */
+#define FIX_1_175875602  ((JLONG)9633)          /* FIX(1.175875602) */
+#define FIX_1_501321110  ((JLONG)12299)         /* FIX(1.501321110) */
+#define FIX_1_847759065  ((JLONG)15137)         /* FIX(1.847759065) */
+#define FIX_1_961570560  ((JLONG)16069)         /* FIX(1.961570560) */
+#define FIX_2_053119869  ((JLONG)16819)         /* FIX(2.053119869) */
+#define FIX_2_562915447  ((JLONG)20995)         /* FIX(2.562915447) */
+#define FIX_3_072711026  ((JLONG)25172)         /* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an JLONG variable by an JLONG constant to yield an JLONG result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var, const)  MULTIPLY16C16(var, const)
+#else
+#define MULTIPLY(var, const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef, quantval)  (((ISLOW_MULT_TYPE)(coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+_jpeg_idct_islow(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp2, tmp3;
+  JLONG tmp10, tmp11, tmp12, tmp13;
+  JLONG z1, z2, z3, z4, z5;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];      /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+  /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+
+    if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
+        inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
+        inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
+        inptr[DCTSIZE * 7] == 0) {
+      /* AC terms all zero */
+      int dcval = LEFT_SHIFT(DEQUANTIZE(inptr[DCTSIZE * 0],
+                             quantptr[DCTSIZE * 0]), PASS1_BITS);
+
+      wsptr[DCTSIZE * 0] = dcval;
+      wsptr[DCTSIZE * 1] = dcval;
+      wsptr[DCTSIZE * 2] = dcval;
+      wsptr[DCTSIZE * 3] = dcval;
+      wsptr[DCTSIZE * 4] = dcval;
+      wsptr[DCTSIZE * 5] = dcval;
+      wsptr[DCTSIZE * 6] = dcval;
+      wsptr[DCTSIZE * 7] = dcval;
+
+      inptr++;                  /* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+
+    tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS);
+    tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS);
+
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, -FIX_0_899976223); /* sqrt(2) * ( c7-c3) */
+    z2 = MULTIPLY(z2, -FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, -FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, -FIX_0_390180644); /* sqrt(2) * ( c5-c3) */
+
+    z3 += z5;
+    z4 += z5;
+
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+
+    wsptr[DCTSIZE * 0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[DCTSIZE * 4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS - PASS1_BITS);
+
+    inptr++;                    /* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+        wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      _JSAMPLE dcval = range_limit[(int)DESCALE((JLONG)wsptr[0],
+                                                PASS1_BITS + 3) & RANGE_MASK];
+
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;         /* advance pointer to next row */
+      continue;
+    }
+#endif
+
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+
+    z2 = (JLONG)wsptr[2];
+    z3 = (JLONG)wsptr[6];
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+
+    tmp0 = LEFT_SHIFT((JLONG)wsptr[0] + (JLONG)wsptr[4], CONST_BITS);
+    tmp1 = LEFT_SHIFT((JLONG)wsptr[0] - (JLONG)wsptr[4], CONST_BITS);
+
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+
+    tmp0 = (JLONG)wsptr[7];
+    tmp1 = (JLONG)wsptr[5];
+    tmp2 = (JLONG)wsptr[3];
+    tmp3 = (JLONG)wsptr[1];
+
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, -FIX_0_899976223); /* sqrt(2) * ( c7-c3) */
+    z2 = MULTIPLY(z2, -FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, -FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, -FIX_0_390180644); /* sqrt(2) * ( c5-c3) */
+
+    z3 += z5;
+    z4 += z5;
+
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+
+    outptr[0] = range_limit[(int)DESCALE(tmp10 + tmp3,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[7] = range_limit[(int)DESCALE(tmp10 - tmp3,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)DESCALE(tmp11 + tmp2,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[6] = range_limit[(int)DESCALE(tmp11 - tmp2,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)DESCALE(tmp12 + tmp1,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[5] = range_limit[(int)DESCALE(tmp12 - tmp1,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)DESCALE(tmp13 + tmp0,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)DESCALE(tmp13 - tmp0,
+                                         CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += DCTSIZE;           /* advance pointer to next row */
+  }
+}
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 7x7 output block.
+ *
+ * Optimized algorithm with 12 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/14).
+ */
+
+GLOBAL(void)
+_jpeg_idct_7x7(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp2, tmp10, tmp11, tmp12, tmp13;
+  JLONG z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[7 * 7];         /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp13 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp13 = LEFT_SHIFT(tmp13, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp13 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp10 = MULTIPLY(z2 - z3, FIX(0.881747734));     /* c4 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.314692123));     /* c6 */
+    tmp11 = tmp10 + tmp12 + tmp13 - MULTIPLY(z2, FIX(1.841218003)); /* c2+c4-c6 */
+    tmp0 = z1 + z3;
+    z2 -= tmp0;
+    tmp0 = MULTIPLY(tmp0, FIX(1.274162392)) + tmp13; /* c2 */
+    tmp10 += tmp0 - MULTIPLY(z3, FIX(0.077722536));  /* c2-c4-c6 */
+    tmp12 += tmp0 - MULTIPLY(z1, FIX(2.470602249));  /* c2+c4+c6 */
+    tmp13 += MULTIPLY(z2, FIX(1.414213562));         /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+
+    tmp1 = MULTIPLY(z1 + z2, FIX(0.935414347));      /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(z1 - z2, FIX(0.170262339));      /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(z2 + z3, -FIX(1.378756276));     /* -c1 */
+    tmp1 += tmp2;
+    z2 = MULTIPLY(z1 + z3, FIX(0.613604268));        /* c5 */
+    tmp0 += z2;
+    tmp2 += z2 + MULTIPLY(z3, FIX(1.870828693));     /* c3+c1-c5 */
+
+    /* Final output stage */
+
+    wsptr[7 * 0] = (int)RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 6] = (int)RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 1] = (int)RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 5] = (int)RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 2] = (int)RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 4] = (int)RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[7 * 3] = (int)RIGHT_SHIFT(tmp13, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 7 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp13 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp13 = LEFT_SHIFT(tmp13, CONST_BITS);
+
+    z1 = (JLONG)wsptr[2];
+    z2 = (JLONG)wsptr[4];
+    z3 = (JLONG)wsptr[6];
+
+    tmp10 = MULTIPLY(z2 - z3, FIX(0.881747734));     /* c4 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.314692123));     /* c6 */
+    tmp11 = tmp10 + tmp12 + tmp13 - MULTIPLY(z2, FIX(1.841218003)); /* c2+c4-c6 */
+    tmp0 = z1 + z3;
+    z2 -= tmp0;
+    tmp0 = MULTIPLY(tmp0, FIX(1.274162392)) + tmp13; /* c2 */
+    tmp10 += tmp0 - MULTIPLY(z3, FIX(0.077722536));  /* c2-c4-c6 */
+    tmp12 += tmp0 - MULTIPLY(z1, FIX(2.470602249));  /* c2+c4+c6 */
+    tmp13 += MULTIPLY(z2, FIX(1.414213562));         /* c0 */
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+
+    tmp1 = MULTIPLY(z1 + z2, FIX(0.935414347));      /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(z1 - z2, FIX(0.170262339));      /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(z2 + z3, -FIX(1.378756276));     /* -c1 */
+    tmp1 += tmp2;
+    z2 = MULTIPLY(z1 + z3, FIX(0.613604268));        /* c5 */
+    tmp0 += z2;
+    tmp2 += z2 + MULTIPLY(z3, FIX(1.870828693));     /* c3+c1-c5 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp10 + tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[6] = range_limit[(int)RIGHT_SHIFT(tmp10 - tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp11 + tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[5] = range_limit[(int)RIGHT_SHIFT(tmp11 - tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp12 + tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)RIGHT_SHIFT(tmp12 - tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)RIGHT_SHIFT(tmp13,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 7;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 6x6 output block.
+ *
+ * Optimized algorithm with 3 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/12).
+ */
+
+GLOBAL(void)
+_jpeg_idct_6x6(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp2, tmp10, tmp11, tmp12;
+  JLONG z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[6 * 6];         /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS - PASS1_BITS - 1);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    tmp10 = MULTIPLY(tmp2, FIX(0.707106781));   /* c4 */
+    tmp1 = tmp0 + tmp10;
+    tmp11 = RIGHT_SHIFT(tmp0 - tmp10 - tmp10, CONST_BITS - PASS1_BITS);
+    tmp10 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    tmp0 = MULTIPLY(tmp10, FIX(1.224744871));   /* c2 */
+    tmp10 = tmp1 + tmp0;
+    tmp12 = tmp1 - tmp0;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    tmp1 = MULTIPLY(z1 + z3, FIX(0.366025404)); /* c5 */
+    tmp0 = tmp1 + LEFT_SHIFT(z1 + z2, CONST_BITS);
+    tmp2 = tmp1 + LEFT_SHIFT(z3 - z2, CONST_BITS);
+    tmp1 = LEFT_SHIFT(z1 - z2 - z3, PASS1_BITS);
+
+    /* Final output stage */
+
+    wsptr[6 * 0] = (int)RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[6 * 5] = (int)RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[6 * 1] = (int)(tmp11 + tmp1);
+    wsptr[6 * 4] = (int)(tmp11 - tmp1);
+    wsptr[6 * 2] = (int)RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[6 * 3] = (int)RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 6 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    tmp2 = (JLONG)wsptr[4];
+    tmp10 = MULTIPLY(tmp2, FIX(0.707106781));   /* c4 */
+    tmp1 = tmp0 + tmp10;
+    tmp11 = tmp0 - tmp10 - tmp10;
+    tmp10 = (JLONG)wsptr[2];
+    tmp0 = MULTIPLY(tmp10, FIX(1.224744871));   /* c2 */
+    tmp10 = tmp1 + tmp0;
+    tmp12 = tmp1 - tmp0;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    tmp1 = MULTIPLY(z1 + z3, FIX(0.366025404)); /* c5 */
+    tmp0 = tmp1 + LEFT_SHIFT(z1 + z2, CONST_BITS);
+    tmp2 = tmp1 + LEFT_SHIFT(z3 - z2, CONST_BITS);
+    tmp1 = LEFT_SHIFT(z1 - z2 - z3, CONST_BITS);
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp10 + tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[5] = range_limit[(int)RIGHT_SHIFT(tmp10 - tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp11 + tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)RIGHT_SHIFT(tmp11 - tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp12 + tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)RIGHT_SHIFT(tmp12 - tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 6;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 5x5 output block.
+ *
+ * Optimized algorithm with 5 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/10).
+ */
+
+GLOBAL(void)
+_jpeg_idct_5x5(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp10, tmp11, tmp12;
+  JLONG z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[5 * 5];         /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp12 = LEFT_SHIFT(tmp12, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp12 += ONE << (CONST_BITS - PASS1_BITS - 1);
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z1 = MULTIPLY(tmp0 + tmp1, FIX(0.790569415)); /* (c2+c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.353553391)); /* (c2-c4)/2 */
+    z3 = tmp12 + z2;
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z1;
+    tmp12 -= LEFT_SHIFT(z2, 2);
+
+    /* Odd part */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));     /* c3 */
+    tmp0 = z1 + MULTIPLY(z2, FIX(0.513743148));   /* c1-c3 */
+    tmp1 = z1 - MULTIPLY(z3, FIX(2.176250899));   /* c1+c3 */
+
+    /* Final output stage */
+
+    wsptr[5 * 0] = (int)RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[5 * 4] = (int)RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[5 * 1] = (int)RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[5 * 3] = (int)RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[5 * 2] = (int)RIGHT_SHIFT(tmp12, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 5 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp12 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp12 = LEFT_SHIFT(tmp12, CONST_BITS);
+    tmp0 = (JLONG)wsptr[2];
+    tmp1 = (JLONG)wsptr[4];
+    z1 = MULTIPLY(tmp0 + tmp1, FIX(0.790569415)); /* (c2+c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.353553391)); /* (c2-c4)/2 */
+    z3 = tmp12 + z2;
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z1;
+    tmp12 -= LEFT_SHIFT(z2, 2);
+
+    /* Odd part */
+
+    z2 = (JLONG)wsptr[1];
+    z3 = (JLONG)wsptr[3];
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));     /* c3 */
+    tmp0 = z1 + MULTIPLY(z2, FIX(0.513743148));   /* c1-c3 */
+    tmp1 = z1 - MULTIPLY(z3, FIX(2.176250899));   /* c1+c3 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp10 + tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)RIGHT_SHIFT(tmp10 - tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp11 + tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)RIGHT_SHIFT(tmp11 - tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp12,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 5;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 3x3 output block.
+ *
+ * Optimized algorithm with 2 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/6).
+ */
+
+GLOBAL(void)
+_jpeg_idct_3x3(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  JLONG tmp0, tmp2, tmp10, tmp12;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[3 * 3];         /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 3; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS - PASS1_BITS - 1);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    tmp12 = MULTIPLY(tmp2, FIX(0.707106781)); /* c2 */
+    tmp10 = tmp0 + tmp12;
+    tmp2 = tmp0 - tmp12 - tmp12;
+
+    /* Odd part */
+
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    tmp0 = MULTIPLY(tmp12, FIX(1.224744871)); /* c1 */
+
+    /* Final output stage */
+
+    wsptr[3 * 0] = (int)RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[3 * 2] = (int)RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[3 * 1] = (int)RIGHT_SHIFT(tmp2, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 3 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 3; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    tmp2 = (JLONG)wsptr[2];
+    tmp12 = MULTIPLY(tmp2, FIX(0.707106781)); /* c2 */
+    tmp10 = tmp0 + tmp12;
+    tmp2 = tmp0 - tmp12 - tmp12;
+
+    /* Odd part */
+
+    tmp12 = (JLONG)wsptr[1];
+    tmp0 = MULTIPLY(tmp12, FIX(1.224744871)); /* c1 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp10 + tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp10 - tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 3;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 9x9 output block.
+ *
+ * Optimized algorithm with 10 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/18).
+ */
+
+GLOBAL(void)
+_jpeg_idct_9x9(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp2, tmp3, tmp10, tmp11, tmp12, tmp13, tmp14;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 9];         /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp3 = MULTIPLY(z3, FIX(0.707106781));      /* c6 */
+    tmp1 = tmp0 + tmp3;
+    tmp2 = tmp0 - tmp3 - tmp3;
+
+    tmp0 = MULTIPLY(z1 - z2, FIX(0.707106781)); /* c6 */
+    tmp11 = tmp2 + tmp0;
+    tmp14 = tmp2 - tmp0 - tmp0;
+
+    tmp0 = MULTIPLY(z1 + z2, FIX(1.328926049)); /* c2 */
+    tmp2 = MULTIPLY(z1, FIX(1.083350441));      /* c4 */
+    tmp3 = MULTIPLY(z2, FIX(0.245575608));      /* c8 */
+
+    tmp10 = tmp1 + tmp0 - tmp3;
+    tmp12 = tmp1 - tmp0 + tmp2;
+    tmp13 = tmp1 - tmp2 + tmp3;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    z2 = MULTIPLY(z2, -FIX(1.224744871));            /* -c3 */
+
+    tmp2 = MULTIPLY(z1 + z3, FIX(0.909038955));      /* c5 */
+    tmp3 = MULTIPLY(z1 + z4, FIX(0.483689525));      /* c7 */
+    tmp0 = tmp2 + tmp3 - z2;
+    tmp1 = MULTIPLY(z3 - z4, FIX(1.392728481));      /* c1 */
+    tmp2 += z2 - tmp1;
+    tmp3 += z2 + tmp1;
+    tmp1 = MULTIPLY(z1 - z3 - z4, FIX(1.224744871)); /* c3 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0] = (int)RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8] = (int)RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1] = (int)RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7] = (int)RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2] = (int)RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6] = (int)RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3] = (int)RIGHT_SHIFT(tmp13 + tmp3, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5] = (int)RIGHT_SHIFT(tmp13 - tmp3, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4] = (int)RIGHT_SHIFT(tmp14, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 9 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 9; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+
+    z1 = (JLONG)wsptr[2];
+    z2 = (JLONG)wsptr[4];
+    z3 = (JLONG)wsptr[6];
+
+    tmp3 = MULTIPLY(z3, FIX(0.707106781));      /* c6 */
+    tmp1 = tmp0 + tmp3;
+    tmp2 = tmp0 - tmp3 - tmp3;
+
+    tmp0 = MULTIPLY(z1 - z2, FIX(0.707106781)); /* c6 */
+    tmp11 = tmp2 + tmp0;
+    tmp14 = tmp2 - tmp0 - tmp0;
+
+    tmp0 = MULTIPLY(z1 + z2, FIX(1.328926049)); /* c2 */
+    tmp2 = MULTIPLY(z1, FIX(1.083350441));      /* c4 */
+    tmp3 = MULTIPLY(z2, FIX(0.245575608));      /* c8 */
+
+    tmp10 = tmp1 + tmp0 - tmp3;
+    tmp12 = tmp1 - tmp0 + tmp2;
+    tmp13 = tmp1 - tmp2 + tmp3;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+
+    z2 = MULTIPLY(z2, -FIX(1.224744871));            /* -c3 */
+
+    tmp2 = MULTIPLY(z1 + z3, FIX(0.909038955));      /* c5 */
+    tmp3 = MULTIPLY(z1 + z4, FIX(0.483689525));      /* c7 */
+    tmp0 = tmp2 + tmp3 - z2;
+    tmp1 = MULTIPLY(z3 - z4, FIX(1.392728481));      /* c1 */
+    tmp2 += z2 - tmp1;
+    tmp3 += z2 + tmp1;
+    tmp1 = MULTIPLY(z1 - z3 - z4, FIX(1.224744871)); /* c3 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp10 + tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[8] = range_limit[(int)RIGHT_SHIFT(tmp10 - tmp0,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp11 + tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[7] = range_limit[(int)RIGHT_SHIFT(tmp11 - tmp1,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp12 + tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[6] = range_limit[(int)RIGHT_SHIFT(tmp12 - tmp2,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)RIGHT_SHIFT(tmp13 + tmp3,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[5] = range_limit[(int)RIGHT_SHIFT(tmp13 - tmp3,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)RIGHT_SHIFT(tmp14,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 10x10 output block.
+ *
+ * Optimized algorithm with 12 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/20).
+ */
+
+GLOBAL(void)
+_jpeg_idct_10x10(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24;
+  JLONG z1, z2, z3, z4, z5;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 10];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z3 = LEFT_SHIFT(z3, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    z3 += ONE << (CONST_BITS - PASS1_BITS - 1);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z1 = MULTIPLY(z4, FIX(1.144122806));         /* c4 */
+    z2 = MULTIPLY(z4, FIX(0.437016024));         /* c8 */
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z2;
+
+    tmp22 = RIGHT_SHIFT(z3 - LEFT_SHIFT(z1 - z2, 1),
+                        CONST_BITS - PASS1_BITS); /* c0 = (c4-c8)*2 */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));    /* c6 */
+    tmp12 = z1 + MULTIPLY(z2, FIX(0.513743148)); /* c2-c6 */
+    tmp13 = z1 - MULTIPLY(z3, FIX(2.176250899)); /* c2+c6 */
+
+    tmp20 = tmp10 + tmp12;
+    tmp24 = tmp10 - tmp12;
+    tmp21 = tmp11 + tmp13;
+    tmp23 = tmp11 - tmp13;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp11 = z2 + z4;
+    tmp13 = z2 - z4;
+
+    tmp12 = MULTIPLY(tmp13, FIX(0.309016994));        /* (c3-c7)/2 */
+    z5 = LEFT_SHIFT(z3, CONST_BITS);
+
+    z2 = MULTIPLY(tmp11, FIX(0.951056516));           /* (c3+c7)/2 */
+    z4 = z5 + tmp12;
+
+    tmp10 = MULTIPLY(z1, FIX(1.396802247)) + z2 + z4; /* c1 */
+    tmp14 = MULTIPLY(z1, FIX(0.221231742)) - z2 + z4; /* c9 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.587785252));           /* (c1-c9)/2 */
+    z4 = z5 - tmp12 - LEFT_SHIFT(tmp13, CONST_BITS - 1);
+
+    tmp12 = LEFT_SHIFT(z1 - tmp13 - z3, PASS1_BITS);
+
+    tmp11 = MULTIPLY(z1, FIX(1.260073511)) - z2 - z4; /* c3 */
+    tmp13 = MULTIPLY(z1, FIX(0.642039522)) - z2 + z4; /* c7 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0] = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1] = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8] = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2] = (int)(tmp22 + tmp12);
+    wsptr[8 * 7] = (int)(tmp22 - tmp12);
+    wsptr[8 * 3] = (int)RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6] = (int)RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4] = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5] = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 10 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 10; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    z3 = LEFT_SHIFT(z3, CONST_BITS);
+    z4 = (JLONG)wsptr[4];
+    z1 = MULTIPLY(z4, FIX(1.144122806));         /* c4 */
+    z2 = MULTIPLY(z4, FIX(0.437016024));         /* c8 */
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z2;
+
+    tmp22 = z3 - LEFT_SHIFT(z1 - z2, 1);         /* c0 = (c4-c8)*2 */
+
+    z2 = (JLONG)wsptr[2];
+    z3 = (JLONG)wsptr[6];
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));    /* c6 */
+    tmp12 = z1 + MULTIPLY(z2, FIX(0.513743148)); /* c2-c6 */
+    tmp13 = z1 - MULTIPLY(z3, FIX(2.176250899)); /* c2+c6 */
+
+    tmp20 = tmp10 + tmp12;
+    tmp24 = tmp10 - tmp12;
+    tmp21 = tmp11 + tmp13;
+    tmp23 = tmp11 - tmp13;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z3 = LEFT_SHIFT(z3, CONST_BITS);
+    z4 = (JLONG)wsptr[7];
+
+    tmp11 = z2 + z4;
+    tmp13 = z2 - z4;
+
+    tmp12 = MULTIPLY(tmp13, FIX(0.309016994));        /* (c3-c7)/2 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.951056516));           /* (c3+c7)/2 */
+    z4 = z3 + tmp12;
+
+    tmp10 = MULTIPLY(z1, FIX(1.396802247)) + z2 + z4; /* c1 */
+    tmp14 = MULTIPLY(z1, FIX(0.221231742)) - z2 + z4; /* c9 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.587785252));           /* (c1-c9)/2 */
+    z4 = z3 - tmp12 - LEFT_SHIFT(tmp13, CONST_BITS - 1);
+
+    tmp12 = LEFT_SHIFT(z1 - tmp13, CONST_BITS) - z3;
+
+    tmp11 = MULTIPLY(z1, FIX(1.260073511)) - z2 - z4; /* c3 */
+    tmp13 = MULTIPLY(z1, FIX(0.642039522)) - z2 + z4; /* c7 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[9] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[1] = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[8] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[2] = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[7] = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[3] = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[6] = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[4] = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+    outptr[5] = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                             CONST_BITS + PASS1_BITS + 3) &
+                            RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing an 11x11 output block.
+ *
+ * Optimized algorithm with 24 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/22).
+ */
+
+GLOBAL(void)
+_jpeg_idct_11x11(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 11];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp10 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp10 = LEFT_SHIFT(tmp10, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp10 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp20 = MULTIPLY(z2 - z3, FIX(2.546640132));     /* c2+c4 */
+    tmp23 = MULTIPLY(z2 - z1, FIX(0.430815045));     /* c2-c6 */
+    z4 = z1 + z3;
+    tmp24 = MULTIPLY(z4, -FIX(1.155664402));         /* -(c2-c10) */
+    z4 -= z2;
+    tmp25 = tmp10 + MULTIPLY(z4, FIX(1.356927976));  /* c2 */
+    tmp21 = tmp20 + tmp23 + tmp25 -
+            MULTIPLY(z2, FIX(1.821790775));          /* c2+c4+c10-c6 */
+    tmp20 += tmp25 + MULTIPLY(z3, FIX(2.115825087)); /* c4+c6 */
+    tmp23 += tmp25 - MULTIPLY(z1, FIX(1.513598477)); /* c6+c8 */
+    tmp24 += tmp25;
+    tmp22 = tmp24 - MULTIPLY(z3, FIX(0.788749120));  /* c8+c10 */
+    tmp24 += MULTIPLY(z2, FIX(1.944413522)) -        /* c2+c8 */
+             MULTIPLY(z1, FIX(1.390975730));         /* c4+c10 */
+    tmp25 = tmp10 - MULTIPLY(z4, FIX(1.414213562));  /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp11 = z1 + z2;
+    tmp14 = MULTIPLY(tmp11 + z3 + z4, FIX(0.398430003)); /* c9 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.887983902));           /* c3-c9 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(0.670361295));         /* c5-c9 */
+    tmp13 = tmp14 + MULTIPLY(z1 + z4, FIX(0.366151574)); /* c7-c9 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+            MULTIPLY(z1, FIX(0.923107866));              /* c7+c5+c3-c1-2*c9 */
+    z1    = tmp14 - MULTIPLY(z2 + z3, FIX(1.163011579)); /* c7+c9 */
+    tmp11 += z1 + MULTIPLY(z2, FIX(2.073276588));        /* c1+c7+3*c9-c3 */
+    tmp12 += z1 - MULTIPLY(z3, FIX(1.192193623));        /* c3+c5-c7-c9 */
+    z1    = MULTIPLY(z2 + z4, -FIX(1.798248910));        /* -(c1+c9) */
+    tmp11 += z1;
+    tmp13 += z1 + MULTIPLY(z4, FIX(2.102458632));        /* c1+c5+c9-c7 */
+    tmp14 += MULTIPLY(z2, -FIX(1.467221301)) +           /* -(c5+c9) */
+             MULTIPLY(z3, FIX(1.001388905)) -            /* c1-c9 */
+             MULTIPLY(z4, FIX(1.684843907));             /* c3+c9 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 10] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 11 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 11; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp10 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp10 = LEFT_SHIFT(tmp10, CONST_BITS);
+
+    z1 = (JLONG)wsptr[2];
+    z2 = (JLONG)wsptr[4];
+    z3 = (JLONG)wsptr[6];
+
+    tmp20 = MULTIPLY(z2 - z3, FIX(2.546640132));     /* c2+c4 */
+    tmp23 = MULTIPLY(z2 - z1, FIX(0.430815045));     /* c2-c6 */
+    z4 = z1 + z3;
+    tmp24 = MULTIPLY(z4, -FIX(1.155664402));         /* -(c2-c10) */
+    z4 -= z2;
+    tmp25 = tmp10 + MULTIPLY(z4, FIX(1.356927976));  /* c2 */
+    tmp21 = tmp20 + tmp23 + tmp25 -
+            MULTIPLY(z2, FIX(1.821790775));          /* c2+c4+c10-c6 */
+    tmp20 += tmp25 + MULTIPLY(z3, FIX(2.115825087)); /* c4+c6 */
+    tmp23 += tmp25 - MULTIPLY(z1, FIX(1.513598477)); /* c6+c8 */
+    tmp24 += tmp25;
+    tmp22 = tmp24 - MULTIPLY(z3, FIX(0.788749120));  /* c8+c10 */
+    tmp24 += MULTIPLY(z2, FIX(1.944413522)) -        /* c2+c8 */
+             MULTIPLY(z1, FIX(1.390975730));         /* c4+c10 */
+    tmp25 = tmp10 - MULTIPLY(z4, FIX(1.414213562));  /* c0 */
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+
+    tmp11 = z1 + z2;
+    tmp14 = MULTIPLY(tmp11 + z3 + z4, FIX(0.398430003)); /* c9 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.887983902));           /* c3-c9 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(0.670361295));         /* c5-c9 */
+    tmp13 = tmp14 + MULTIPLY(z1 + z4, FIX(0.366151574)); /* c7-c9 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+            MULTIPLY(z1, FIX(0.923107866));              /* c7+c5+c3-c1-2*c9 */
+    z1    = tmp14 - MULTIPLY(z2 + z3, FIX(1.163011579)); /* c7+c9 */
+    tmp11 += z1 + MULTIPLY(z2, FIX(2.073276588));        /* c1+c7+3*c9-c3 */
+    tmp12 += z1 - MULTIPLY(z3, FIX(1.192193623));        /* c3+c5-c7-c9 */
+    z1    = MULTIPLY(z2 + z4, -FIX(1.798248910));        /* -(c1+c9) */
+    tmp11 += z1;
+    tmp13 += z1 + MULTIPLY(z4, FIX(2.102458632));        /* c1+c5+c9-c7 */
+    tmp14 += MULTIPLY(z2, -FIX(1.467221301)) +           /* -(c5+c9) */
+             MULTIPLY(z3, FIX(1.001388905)) -            /* c1-c9 */
+             MULTIPLY(z4, FIX(1.684843907));             /* c3+c9 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 12x12 output block.
+ *
+ * Optimized algorithm with 15 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/24).
+ */
+
+GLOBAL(void)
+_jpeg_idct_12x12(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 12];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z3 = LEFT_SHIFT(z3, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    z3 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z4 = MULTIPLY(z4, FIX(1.224744871)); /* c4 */
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z4 = MULTIPLY(z1, FIX(1.366025404)); /* c2 */
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+    z2 = LEFT_SHIFT(z2, CONST_BITS);
+
+    tmp12 = z1 - z2;
+
+    tmp21 = z3 + tmp12;
+    tmp24 = z3 - tmp12;
+
+    tmp12 = z4 + z2;
+
+    tmp20 = tmp10 + tmp12;
+    tmp25 = tmp10 - tmp12;
+
+    tmp12 = z4 - z1 - z2;
+
+    tmp22 = tmp11 + tmp12;
+    tmp23 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp11 = MULTIPLY(z2, FIX(1.306562965));                  /* c3 */
+    tmp14 = MULTIPLY(z2, -FIX_0_541196100);                  /* -c9 */
+
+    tmp10 = z1 + z3;
+    tmp15 = MULTIPLY(tmp10 + z4, FIX(0.860918669));          /* c7 */
+    tmp12 = tmp15 + MULTIPLY(tmp10, FIX(0.261052384));       /* c5-c7 */
+    tmp10 = tmp12 + tmp11 + MULTIPLY(z1, FIX(0.280143716));  /* c1-c5 */
+    tmp13 = MULTIPLY(z3 + z4, -FIX(1.045510580));            /* -(c7+c11) */
+    tmp12 += tmp13 + tmp14 - MULTIPLY(z3, FIX(1.478575242)); /* c1+c5-c7-c11 */
+    tmp13 += tmp15 - tmp11 + MULTIPLY(z4, FIX(1.586706681)); /* c1+c11 */
+    tmp15 += tmp14 - MULTIPLY(z1, FIX(0.676326758)) -        /* c7-c11 */
+             MULTIPLY(z4, FIX(1.982889723));                 /* c5+c7 */
+
+    z1 -= z4;
+    z2 -= z3;
+    z3 = MULTIPLY(z1 + z2, FIX_0_541196100);                 /* c9 */
+    tmp11 = z3 + MULTIPLY(z1, FIX_0_765366865);              /* c3-c9 */
+    tmp14 = z3 - MULTIPLY(z2, FIX_1_847759065);              /* c3+c9 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 11] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 10] = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 12 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 12; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    z3 = LEFT_SHIFT(z3, CONST_BITS);
+
+    z4 = (JLONG)wsptr[4];
+    z4 = MULTIPLY(z4, FIX(1.224744871)); /* c4 */
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    z1 = (JLONG)wsptr[2];
+    z4 = MULTIPLY(z1, FIX(1.366025404)); /* c2 */
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    z2 = (JLONG)wsptr[6];
+    z2 = LEFT_SHIFT(z2, CONST_BITS);
+
+    tmp12 = z1 - z2;
+
+    tmp21 = z3 + tmp12;
+    tmp24 = z3 - tmp12;
+
+    tmp12 = z4 + z2;
+
+    tmp20 = tmp10 + tmp12;
+    tmp25 = tmp10 - tmp12;
+
+    tmp12 = z4 - z1 - z2;
+
+    tmp22 = tmp11 + tmp12;
+    tmp23 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+
+    tmp11 = MULTIPLY(z2, FIX(1.306562965));                  /* c3 */
+    tmp14 = MULTIPLY(z2, -FIX_0_541196100);                  /* -c9 */
+
+    tmp10 = z1 + z3;
+    tmp15 = MULTIPLY(tmp10 + z4, FIX(0.860918669));          /* c7 */
+    tmp12 = tmp15 + MULTIPLY(tmp10, FIX(0.261052384));       /* c5-c7 */
+    tmp10 = tmp12 + tmp11 + MULTIPLY(z1, FIX(0.280143716));  /* c1-c5 */
+    tmp13 = MULTIPLY(z3 + z4, -FIX(1.045510580));            /* -(c7+c11) */
+    tmp12 += tmp13 + tmp14 - MULTIPLY(z3, FIX(1.478575242)); /* c1+c5-c7-c11 */
+    tmp13 += tmp15 - tmp11 + MULTIPLY(z4, FIX(1.586706681)); /* c1+c11 */
+    tmp15 += tmp14 - MULTIPLY(z1, FIX(0.676326758)) -        /* c7-c11 */
+             MULTIPLY(z4, FIX(1.982889723));                 /* c5+c7 */
+
+    z1 -= z4;
+    z2 -= z3;
+    z3 = MULTIPLY(z1 + z2, FIX_0_541196100);                 /* c9 */
+    tmp11 = z3 + MULTIPLY(z1, FIX_0_765366865);              /* c3-c9 */
+    tmp14 = z3 - MULTIPLY(z2, FIX_1_847759065);              /* c3+c9 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[11] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25 + tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp25 - tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 13x13 output block.
+ *
+ * Optimized algorithm with 29 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/26).
+ */
+
+GLOBAL(void)
+_jpeg_idct_13x13(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 13];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    tmp12 = MULTIPLY(tmp10, FIX(1.155388986));                /* (c4+c6)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.096834934)) + z1;           /* (c4-c6)/2 */
+
+    tmp20 = MULTIPLY(z2, FIX(1.373119086)) + tmp12 + tmp13;   /* c2 */
+    tmp22 = MULTIPLY(z2, FIX(0.501487041)) - tmp12 + tmp13;   /* c10 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.316450131));                /* (c8-c12)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.486914739)) + z1;           /* (c8+c12)/2 */
+
+    tmp21 = MULTIPLY(z2, FIX(1.058554052)) - tmp12 + tmp13;   /* c6 */
+    tmp25 = MULTIPLY(z2, -FIX(1.252223920)) + tmp12 + tmp13;  /* c4 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.435816023));                /* (c2-c10)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.937303064)) - z1;           /* (c2+c10)/2 */
+
+    tmp23 = MULTIPLY(z2, -FIX(0.170464608)) - tmp12 - tmp13;  /* c12 */
+    tmp24 = MULTIPLY(z2, -FIX(0.803364869)) + tmp12 - tmp13;  /* c8 */
+
+    tmp26 = MULTIPLY(tmp11 - z2, FIX(1.414213562)) + z1;      /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.322312651));     /* c3 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(1.163874945));     /* c5 */
+    tmp15 = z1 + z4;
+    tmp13 = MULTIPLY(tmp15, FIX(0.937797057));       /* c7 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+            MULTIPLY(z1, FIX(2.020082300));          /* c7+c5+c3-c1 */
+    tmp14 = MULTIPLY(z2 + z3, -FIX(0.338443458));    /* -c11 */
+    tmp11 += tmp14 + MULTIPLY(z2, FIX(0.837223564)); /* c5+c9+c11-c3 */
+    tmp12 += tmp14 - MULTIPLY(z3, FIX(1.572116027)); /* c1+c5-c9-c11 */
+    tmp14 = MULTIPLY(z2 + z4, -FIX(1.163874945));    /* -c5 */
+    tmp11 += tmp14;
+    tmp13 += tmp14 + MULTIPLY(z4, FIX(2.205608352)); /* c3+c5+c9-c7 */
+    tmp14 = MULTIPLY(z3 + z4, -FIX(0.657217813));    /* -c9 */
+    tmp12 += tmp14;
+    tmp13 += tmp14;
+    tmp15 = MULTIPLY(tmp15, FIX(0.338443458));       /* c11 */
+    tmp14 = tmp15 + MULTIPLY(z1, FIX(0.318774355)) - /* c9-c11 */
+            MULTIPLY(z2, FIX(0.466105296));          /* c1-c7 */
+    z1    = MULTIPLY(z3 - z2, FIX(0.937797057));     /* c7 */
+    tmp14 += z1;
+    tmp15 += z1 + MULTIPLY(z3, FIX(0.384515595)) -   /* c3-c7 */
+             MULTIPLY(z4, FIX(1.742345811));         /* c1+c11 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 12] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 11] = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 10] = (int)RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp26, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 13 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 13; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+
+    z2 = (JLONG)wsptr[2];
+    z3 = (JLONG)wsptr[4];
+    z4 = (JLONG)wsptr[6];
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    tmp12 = MULTIPLY(tmp10, FIX(1.155388986));                /* (c4+c6)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.096834934)) + z1;           /* (c4-c6)/2 */
+
+    tmp20 = MULTIPLY(z2, FIX(1.373119086)) + tmp12 + tmp13;   /* c2 */
+    tmp22 = MULTIPLY(z2, FIX(0.501487041)) - tmp12 + tmp13;   /* c10 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.316450131));                /* (c8-c12)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.486914739)) + z1;           /* (c8+c12)/2 */
+
+    tmp21 = MULTIPLY(z2, FIX(1.058554052)) - tmp12 + tmp13;   /* c6 */
+    tmp25 = MULTIPLY(z2, -FIX(1.252223920)) + tmp12 + tmp13;  /* c4 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.435816023));                /* (c2-c10)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.937303064)) - z1;           /* (c2+c10)/2 */
+
+    tmp23 = MULTIPLY(z2, -FIX(0.170464608)) - tmp12 - tmp13;  /* c12 */
+    tmp24 = MULTIPLY(z2, -FIX(0.803364869)) + tmp12 - tmp13;  /* c8 */
+
+    tmp26 = MULTIPLY(tmp11 - z2, FIX(1.414213562)) + z1;      /* c0 */
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.322312651));     /* c3 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(1.163874945));     /* c5 */
+    tmp15 = z1 + z4;
+    tmp13 = MULTIPLY(tmp15, FIX(0.937797057));       /* c7 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+            MULTIPLY(z1, FIX(2.020082300));          /* c7+c5+c3-c1 */
+    tmp14 = MULTIPLY(z2 + z3, -FIX(0.338443458));    /* -c11 */
+    tmp11 += tmp14 + MULTIPLY(z2, FIX(0.837223564)); /* c5+c9+c11-c3 */
+    tmp12 += tmp14 - MULTIPLY(z3, FIX(1.572116027)); /* c1+c5-c9-c11 */
+    tmp14 = MULTIPLY(z2 + z4, -FIX(1.163874945));    /* -c5 */
+    tmp11 += tmp14;
+    tmp13 += tmp14 + MULTIPLY(z4, FIX(2.205608352)); /* c3+c5+c9-c7 */
+    tmp14 = MULTIPLY(z3 + z4, -FIX(0.657217813));    /* -c9 */
+    tmp12 += tmp14;
+    tmp13 += tmp14;
+    tmp15 = MULTIPLY(tmp15, FIX(0.338443458));       /* c11 */
+    tmp14 = tmp15 + MULTIPLY(z1, FIX(0.318774355)) - /* c9-c11 */
+            MULTIPLY(z2, FIX(0.466105296));          /* c1-c7 */
+    z1    = MULTIPLY(z3 - z2, FIX(0.937797057));     /* c7 */
+    tmp14 += z1;
+    tmp15 += z1 + MULTIPLY(z3, FIX(0.384515595)) -   /* c3-c7 */
+             MULTIPLY(z4, FIX(1.742345811));         /* c1+c11 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[12] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[11] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25 + tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp25 - tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp26,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 14x14 output block.
+ *
+ * Optimized algorithm with 20 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/28).
+ */
+
+GLOBAL(void)
+_jpeg_idct_14x14(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 14];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS - PASS1_BITS - 1);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z2 = MULTIPLY(z4, FIX(1.274162392));         /* c4 */
+    z3 = MULTIPLY(z4, FIX(0.314692123));         /* c12 */
+    z4 = MULTIPLY(z4, FIX(0.881747734));         /* c8 */
+
+    tmp10 = z1 + z2;
+    tmp11 = z1 + z3;
+    tmp12 = z1 - z4;
+
+    tmp23 = RIGHT_SHIFT(z1 - LEFT_SHIFT(z2 + z3 - z4, 1),
+                        CONST_BITS - PASS1_BITS); /* c0 = (c4+c12-c8)*2 */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    z3 = MULTIPLY(z1 + z2, FIX(1.105676686));    /* c6 */
+
+    tmp13 = z3 + MULTIPLY(z1, FIX(0.273079590)); /* c2-c6 */
+    tmp14 = z3 - MULTIPLY(z2, FIX(1.719280954)); /* c6+c10 */
+    tmp15 = MULTIPLY(z1, FIX(0.613604268)) -     /* c10 */
+            MULTIPLY(z2, FIX(1.378756276));      /* c2 */
+
+    tmp20 = tmp10 + tmp13;
+    tmp26 = tmp10 - tmp13;
+    tmp21 = tmp11 + tmp14;
+    tmp25 = tmp11 - tmp14;
+    tmp22 = tmp12 + tmp15;
+    tmp24 = tmp12 - tmp15;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+    tmp13 = LEFT_SHIFT(z4, CONST_BITS);
+
+    tmp14 = z1 + z3;
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.334852607));           /* c3 */
+    tmp12 = MULTIPLY(tmp14, FIX(1.197448846));             /* c5 */
+    tmp10 = tmp11 + tmp12 + tmp13 - MULTIPLY(z1, FIX(1.126980169)); /* c3+c5-c1 */
+    tmp14 = MULTIPLY(tmp14, FIX(0.752406978));             /* c9 */
+    tmp16 = tmp14 - MULTIPLY(z1, FIX(1.061150426));        /* c9+c11-c13 */
+    z1    -= z2;
+    tmp15 = MULTIPLY(z1, FIX(0.467085129)) - tmp13;        /* c11 */
+    tmp16 += tmp15;
+    z1    += z4;
+    z4    = MULTIPLY(z2 + z3, -FIX(0.158341681)) - tmp13;  /* -c13 */
+    tmp11 += z4 - MULTIPLY(z2, FIX(0.424103948));          /* c3-c9-c13 */
+    tmp12 += z4 - MULTIPLY(z3, FIX(2.373959773));          /* c3+c5-c13 */
+    z4    = MULTIPLY(z3 - z2, FIX(1.405321284));           /* c1 */
+    tmp14 += z4 + tmp13 - MULTIPLY(z3, FIX(1.6906431334)); /* c1+c9-c11 */
+    tmp15 += z4 + MULTIPLY(z2, FIX(0.674957567));          /* c1+c11-c5 */
+
+    tmp13 = LEFT_SHIFT(z1 - z3, PASS1_BITS);
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 13] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 12] = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 11] = (int)RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)(tmp23 + tmp13);
+    wsptr[8 * 10] = (int)(tmp23 - tmp13);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp26 + tmp16, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp26 - tmp16, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 14 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 14; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    z4 = (JLONG)wsptr[4];
+    z2 = MULTIPLY(z4, FIX(1.274162392));         /* c4 */
+    z3 = MULTIPLY(z4, FIX(0.314692123));         /* c12 */
+    z4 = MULTIPLY(z4, FIX(0.881747734));         /* c8 */
+
+    tmp10 = z1 + z2;
+    tmp11 = z1 + z3;
+    tmp12 = z1 - z4;
+
+    tmp23 = z1 - LEFT_SHIFT(z2 + z3 - z4, 1);    /* c0 = (c4+c12-c8)*2 */
+
+    z1 = (JLONG)wsptr[2];
+    z2 = (JLONG)wsptr[6];
+
+    z3 = MULTIPLY(z1 + z2, FIX(1.105676686));    /* c6 */
+
+    tmp13 = z3 + MULTIPLY(z1, FIX(0.273079590)); /* c2-c6 */
+    tmp14 = z3 - MULTIPLY(z2, FIX(1.719280954)); /* c6+c10 */
+    tmp15 = MULTIPLY(z1, FIX(0.613604268)) -     /* c10 */
+            MULTIPLY(z2, FIX(1.378756276));      /* c2 */
+
+    tmp20 = tmp10 + tmp13;
+    tmp26 = tmp10 - tmp13;
+    tmp21 = tmp11 + tmp14;
+    tmp25 = tmp11 - tmp14;
+    tmp22 = tmp12 + tmp15;
+    tmp24 = tmp12 - tmp15;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+    z4 = LEFT_SHIFT(z4, CONST_BITS);
+
+    tmp14 = z1 + z3;
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.334852607));           /* c3 */
+    tmp12 = MULTIPLY(tmp14, FIX(1.197448846));             /* c5 */
+    tmp10 = tmp11 + tmp12 + z4 - MULTIPLY(z1, FIX(1.126980169)); /* c3+c5-c1 */
+    tmp14 = MULTIPLY(tmp14, FIX(0.752406978));             /* c9 */
+    tmp16 = tmp14 - MULTIPLY(z1, FIX(1.061150426));        /* c9+c11-c13 */
+    z1    -= z2;
+    tmp15 = MULTIPLY(z1, FIX(0.467085129)) - z4;           /* c11 */
+    tmp16 += tmp15;
+    tmp13 = MULTIPLY(z2 + z3, -FIX(0.158341681)) - z4;     /* -c13 */
+    tmp11 += tmp13 - MULTIPLY(z2, FIX(0.424103948));       /* c3-c9-c13 */
+    tmp12 += tmp13 - MULTIPLY(z3, FIX(2.373959773));       /* c3+c5-c13 */
+    tmp13 = MULTIPLY(z3 - z2, FIX(1.405321284));           /* c1 */
+    tmp14 += tmp13 + z4 - MULTIPLY(z3, FIX(1.6906431334)); /* c1+c9-c11 */
+    tmp15 += tmp13 + MULTIPLY(z2, FIX(0.674957567));       /* c1+c11-c5 */
+
+    tmp13 = LEFT_SHIFT(z1 - z3, CONST_BITS) + z4;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[13] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[12] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[11] = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25 + tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp25 - tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp26 + tmp16,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp26 - tmp16,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 15x15 output block.
+ *
+ * Optimized algorithm with 22 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/30).
+ */
+
+GLOBAL(void)
+_jpeg_idct_15x15(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26, tmp27;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 15];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+
+    tmp10 = MULTIPLY(z4, FIX(0.437016024)); /* c12 */
+    tmp11 = MULTIPLY(z4, FIX(1.144122806)); /* c6 */
+
+    tmp12 = z1 - tmp10;
+    tmp13 = z1 + tmp11;
+    z1 -= LEFT_SHIFT(tmp11 - tmp10, 1);     /* c0 = (c6-c12)*2 */
+
+    z4 = z2 - z3;
+    z3 += z2;
+    tmp10 = MULTIPLY(z3, FIX(1.337628990)); /* (c2+c4)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.045680613)); /* (c2-c4)/2 */
+    z2 = MULTIPLY(z2, FIX(1.439773946));    /* c4+c14 */
+
+    tmp20 = tmp13 + tmp10 + tmp11;
+    tmp23 = tmp12 - tmp10 + tmp11 + z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.547059574)); /* (c8+c14)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.399234004)); /* (c8-c14)/2 */
+
+    tmp25 = tmp13 - tmp10 - tmp11;
+    tmp26 = tmp12 + tmp10 - tmp11 - z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.790569415)); /* (c6+c12)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.353553391)); /* (c6-c12)/2 */
+
+    tmp21 = tmp12 + tmp10 + tmp11;
+    tmp24 = tmp13 - tmp10 + tmp11;
+    tmp11 += tmp11;
+    tmp22 = z1 + tmp11;                     /* c10 = c6-c12 */
+    tmp27 = z1 - tmp11 - tmp11;             /* c0 = (c6-c12)*2 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z3 = MULTIPLY(z4, FIX(1.224744871));                    /* c5 */
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp13 = z2 - z4;
+    tmp15 = MULTIPLY(z1 + tmp13, FIX(0.831253876));         /* c9 */
+    tmp11 = tmp15 + MULTIPLY(z1, FIX(0.513743148));         /* c3-c9 */
+    tmp14 = tmp15 - MULTIPLY(tmp13, FIX(2.176250899));      /* c3+c9 */
+
+    tmp13 = MULTIPLY(z2, -FIX(0.831253876));                /* -c9 */
+    tmp15 = MULTIPLY(z2, -FIX(1.344997024));                /* -c3 */
+    z2 = z1 - z4;
+    tmp12 = z3 + MULTIPLY(z2, FIX(1.406466353));            /* c1 */
+
+    tmp10 = tmp12 + MULTIPLY(z4, FIX(2.457431844)) - tmp15; /* c1+c7 */
+    tmp16 = tmp12 - MULTIPLY(z1, FIX(1.112434820)) + tmp13; /* c1-c13 */
+    tmp12 = MULTIPLY(z2, FIX(1.224744871)) - z3;            /* c5 */
+    z2 = MULTIPLY(z1 + z4, FIX(0.575212477));               /* c11 */
+    tmp13 += z2 + MULTIPLY(z1, FIX(0.475753014)) - z3;      /* c7-c11 */
+    tmp15 += z2 - MULTIPLY(z4, FIX(0.869244010)) + z3;      /* c11+c13 */
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 14] = (int)RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 13] = (int)RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 12] = (int)RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 11] = (int)RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 10] = (int)RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp26 + tmp16, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp26 - tmp16, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp27, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 15 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 15; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    z1 = LEFT_SHIFT(z1, CONST_BITS);
+
+    z2 = (JLONG)wsptr[2];
+    z3 = (JLONG)wsptr[4];
+    z4 = (JLONG)wsptr[6];
+
+    tmp10 = MULTIPLY(z4, FIX(0.437016024)); /* c12 */
+    tmp11 = MULTIPLY(z4, FIX(1.144122806)); /* c6 */
+
+    tmp12 = z1 - tmp10;
+    tmp13 = z1 + tmp11;
+    z1 -= LEFT_SHIFT(tmp11 - tmp10, 1);     /* c0 = (c6-c12)*2 */
+
+    z4 = z2 - z3;
+    z3 += z2;
+    tmp10 = MULTIPLY(z3, FIX(1.337628990)); /* (c2+c4)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.045680613)); /* (c2-c4)/2 */
+    z2 = MULTIPLY(z2, FIX(1.439773946));    /* c4+c14 */
+
+    tmp20 = tmp13 + tmp10 + tmp11;
+    tmp23 = tmp12 - tmp10 + tmp11 + z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.547059574)); /* (c8+c14)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.399234004)); /* (c8-c14)/2 */
+
+    tmp25 = tmp13 - tmp10 - tmp11;
+    tmp26 = tmp12 + tmp10 - tmp11 - z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.790569415)); /* (c6+c12)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.353553391)); /* (c6-c12)/2 */
+
+    tmp21 = tmp12 + tmp10 + tmp11;
+    tmp24 = tmp13 - tmp10 + tmp11;
+    tmp11 += tmp11;
+    tmp22 = z1 + tmp11;                     /* c10 = c6-c12 */
+    tmp27 = z1 - tmp11 - tmp11;             /* c0 = (c6-c12)*2 */
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z4 = (JLONG)wsptr[5];
+    z3 = MULTIPLY(z4, FIX(1.224744871));                    /* c5 */
+    z4 = (JLONG)wsptr[7];
+
+    tmp13 = z2 - z4;
+    tmp15 = MULTIPLY(z1 + tmp13, FIX(0.831253876));         /* c9 */
+    tmp11 = tmp15 + MULTIPLY(z1, FIX(0.513743148));         /* c3-c9 */
+    tmp14 = tmp15 - MULTIPLY(tmp13, FIX(2.176250899));      /* c3+c9 */
+
+    tmp13 = MULTIPLY(z2, -FIX(0.831253876));                /* -c9 */
+    tmp15 = MULTIPLY(z2, -FIX(1.344997024));                /* -c3 */
+    z2 = z1 - z4;
+    tmp12 = z3 + MULTIPLY(z2, FIX(1.406466353));            /* c1 */
+
+    tmp10 = tmp12 + MULTIPLY(z4, FIX(2.457431844)) - tmp15; /* c1+c7 */
+    tmp16 = tmp12 - MULTIPLY(z1, FIX(1.112434820)) + tmp13; /* c1-c13 */
+    tmp12 = MULTIPLY(z2, FIX(1.224744871)) - z3;            /* c5 */
+    z2 = MULTIPLY(z1 + z4, FIX(0.575212477));               /* c11 */
+    tmp13 += z2 + MULTIPLY(z1, FIX(0.475753014)) - z3;      /* c7-c11 */
+    tmp15 += z2 - MULTIPLY(z4, FIX(0.869244010)) + z3;      /* c11+c13 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[14] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[13] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[12] = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[11] = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp14,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25 + tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp25 - tmp15,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp26 + tmp16,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp26 - tmp16,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp27,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 16x16 output block.
+ *
+ * Optimized algorithm with 28 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/32).
+ */
+
+GLOBAL(void)
+_jpeg_idct_16x16(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, _JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  JLONG tmp0, tmp1, tmp2, tmp3, tmp10, tmp11, tmp12, tmp13;
+  JLONG tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26, tmp27;
+  JLONG z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE *quantptr;
+  int *wsptr;
+  _JSAMPROW outptr;
+  _JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8 * 16];        /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *)compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS - PASS1_BITS - 1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
+    tmp1 = MULTIPLY(z1, FIX(1.306562965));      /* c4[16] = c2[8] */
+    tmp2 = MULTIPLY(z1, FIX_0_541196100);       /* c12[16] = c6[8] */
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
+    z3 = z1 - z2;
+    z4 = MULTIPLY(z3, FIX(0.275899379));        /* c14[16] = c7[8] */
+    z3 = MULTIPLY(z3, FIX(1.387039845));        /* c2[16] = c1[8] */
+
+    tmp0 = z3 + MULTIPLY(z2, FIX_2_562915447);  /* (c6+c2)[16] = (c3+c1)[8] */
+    tmp1 = z4 + MULTIPLY(z1, FIX_0_899976223);  /* (c6-c14)[16] = (c3-c7)[8] */
+    tmp2 = z3 - MULTIPLY(z1, FIX(0.601344887)); /* (c2-c10)[16] = (c1-c5)[8] */
+    tmp3 = z4 - MULTIPLY(z2, FIX(0.509795579)); /* (c10-c14)[16] = (c5-c7)[8] */
+
+    tmp20 = tmp10 + tmp0;
+    tmp27 = tmp10 - tmp0;
+    tmp21 = tmp12 + tmp1;
+    tmp26 = tmp12 - tmp1;
+    tmp22 = tmp13 + tmp2;
+    tmp25 = tmp13 - tmp2;
+    tmp23 = tmp11 + tmp3;
+    tmp24 = tmp11 - tmp3;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
+
+    tmp11 = z1 + z3;
+
+    tmp1  = MULTIPLY(z1 + z2, FIX(1.353318001));   /* c3 */
+    tmp2  = MULTIPLY(tmp11,   FIX(1.247225013));   /* c5 */
+    tmp3  = MULTIPLY(z1 + z4, FIX(1.093201867));   /* c7 */
+    tmp10 = MULTIPLY(z1 - z4, FIX(0.897167586));   /* c9 */
+    tmp11 = MULTIPLY(tmp11,   FIX(0.666655658));   /* c11 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.410524528));   /* c13 */
+    tmp0  = tmp1 + tmp2 + tmp3 -
+            MULTIPLY(z1, FIX(2.286341144));        /* c7+c5+c3-c1 */
+    tmp13 = tmp10 + tmp11 + tmp12 -
+            MULTIPLY(z1, FIX(1.835730603));        /* c9+c11+c13-c15 */
+    z1    = MULTIPLY(z2 + z3, FIX(0.138617169));   /* c15 */
+    tmp1  += z1 + MULTIPLY(z2, FIX(0.071888074));  /* c9+c11-c3-c15 */
+    tmp2  += z1 - MULTIPLY(z3, FIX(1.125726048));  /* c5+c7+c15-c3 */
+    z1    = MULTIPLY(z3 - z2, FIX(1.407403738));   /* c1 */
+    tmp11 += z1 - MULTIPLY(z3, FIX(0.766367282));  /* c1+c11-c9-c13 */
+    tmp12 += z1 + MULTIPLY(z2, FIX(1.971951411));  /* c1+c5+c13-c7 */
+    z2    += z4;
+    z1    = MULTIPLY(z2, -FIX(0.666655658));       /* -c11 */
+    tmp1  += z1;
+    tmp3  += z1 + MULTIPLY(z4, FIX(1.065388962));  /* c3+c11+c15-c7 */
+    z2    = MULTIPLY(z2, -FIX(1.247225013));       /* -c5 */
+    tmp10 += z2 + MULTIPLY(z4, FIX(3.141271809));  /* c1+c5+c9-c13 */
+    tmp12 += z2;
+    z2    = MULTIPLY(z3 + z4, -FIX(1.353318001));  /* -c3 */
+    tmp2  += z2;
+    tmp3  += z2;
+    z2    = MULTIPLY(z4 - z3, FIX(0.410524528));   /* c13 */
+    tmp10 += z2;
+    tmp11 += z2;
+
+    /* Final output stage */
+
+    wsptr[8 * 0]  = (int)RIGHT_SHIFT(tmp20 + tmp0,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 15] = (int)RIGHT_SHIFT(tmp20 - tmp0,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 1]  = (int)RIGHT_SHIFT(tmp21 + tmp1,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 14] = (int)RIGHT_SHIFT(tmp21 - tmp1,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 2]  = (int)RIGHT_SHIFT(tmp22 + tmp2,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 13] = (int)RIGHT_SHIFT(tmp22 - tmp2,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 3]  = (int)RIGHT_SHIFT(tmp23 + tmp3,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 12] = (int)RIGHT_SHIFT(tmp23 - tmp3,  CONST_BITS - PASS1_BITS);
+    wsptr[8 * 4]  = (int)RIGHT_SHIFT(tmp24 + tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 11] = (int)RIGHT_SHIFT(tmp24 - tmp10, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 5]  = (int)RIGHT_SHIFT(tmp25 + tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 10] = (int)RIGHT_SHIFT(tmp25 - tmp11, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 6]  = (int)RIGHT_SHIFT(tmp26 + tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 9]  = (int)RIGHT_SHIFT(tmp26 - tmp12, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 7]  = (int)RIGHT_SHIFT(tmp27 + tmp13, CONST_BITS - PASS1_BITS);
+    wsptr[8 * 8]  = (int)RIGHT_SHIFT(tmp27 - tmp13, CONST_BITS - PASS1_BITS);
+  }
+
+  /* Pass 2: process 16 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 16; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (JLONG)wsptr[0] + (ONE << (PASS1_BITS + 2));
+    tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
+
+    z1 = (JLONG)wsptr[4];
+    tmp1 = MULTIPLY(z1, FIX(1.306562965));      /* c4[16] = c2[8] */
+    tmp2 = MULTIPLY(z1, FIX_0_541196100);       /* c12[16] = c6[8] */
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+
+    z1 = (JLONG)wsptr[2];
+    z2 = (JLONG)wsptr[6];
+    z3 = z1 - z2;
+    z4 = MULTIPLY(z3, FIX(0.275899379));        /* c14[16] = c7[8] */
+    z3 = MULTIPLY(z3, FIX(1.387039845));        /* c2[16] = c1[8] */
+
+    tmp0 = z3 + MULTIPLY(z2, FIX_2_562915447);  /* (c6+c2)[16] = (c3+c1)[8] */
+    tmp1 = z4 + MULTIPLY(z1, FIX_0_899976223);  /* (c6-c14)[16] = (c3-c7)[8] */
+    tmp2 = z3 - MULTIPLY(z1, FIX(0.601344887)); /* (c2-c10)[16] = (c1-c5)[8] */
+    tmp3 = z4 - MULTIPLY(z2, FIX(0.509795579)); /* (c10-c14)[16] = (c5-c7)[8] */
+
+    tmp20 = tmp10 + tmp0;
+    tmp27 = tmp10 - tmp0;
+    tmp21 = tmp12 + tmp1;
+    tmp26 = tmp12 - tmp1;
+    tmp22 = tmp13 + tmp2;
+    tmp25 = tmp13 - tmp2;
+    tmp23 = tmp11 + tmp3;
+    tmp24 = tmp11 - tmp3;
+
+    /* Odd part */
+
+    z1 = (JLONG)wsptr[1];
+    z2 = (JLONG)wsptr[3];
+    z3 = (JLONG)wsptr[5];
+    z4 = (JLONG)wsptr[7];
+
+    tmp11 = z1 + z3;
+
+    tmp1  = MULTIPLY(z1 + z2, FIX(1.353318001));   /* c3 */
+    tmp2  = MULTIPLY(tmp11,   FIX(1.247225013));   /* c5 */
+    tmp3  = MULTIPLY(z1 + z4, FIX(1.093201867));   /* c7 */
+    tmp10 = MULTIPLY(z1 - z4, FIX(0.897167586));   /* c9 */
+    tmp11 = MULTIPLY(tmp11,   FIX(0.666655658));   /* c11 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.410524528));   /* c13 */
+    tmp0  = tmp1 + tmp2 + tmp3 -
+            MULTIPLY(z1, FIX(2.286341144));        /* c7+c5+c3-c1 */
+    tmp13 = tmp10 + tmp11 + tmp12 -
+            MULTIPLY(z1, FIX(1.835730603));        /* c9+c11+c13-c15 */
+    z1    = MULTIPLY(z2 + z3, FIX(0.138617169));   /* c15 */
+    tmp1  += z1 + MULTIPLY(z2, FIX(0.071888074));  /* c9+c11-c3-c15 */
+    tmp2  += z1 - MULTIPLY(z3, FIX(1.125726048));  /* c5+c7+c15-c3 */
+    z1    = MULTIPLY(z3 - z2, FIX(1.407403738));   /* c1 */
+    tmp11 += z1 - MULTIPLY(z3, FIX(0.766367282));  /* c1+c11-c9-c13 */
+    tmp12 += z1 + MULTIPLY(z2, FIX(1.971951411));  /* c1+c5+c13-c7 */
+    z2    += z4;
+    z1    = MULTIPLY(z2, -FIX(0.666655658));       /* -c11 */
+    tmp1  += z1;
+    tmp3  += z1 + MULTIPLY(z4, FIX(1.065388962));  /* c3+c11+c15-c7 */
+    z2    = MULTIPLY(z2, -FIX(1.247225013));       /* -c5 */
+    tmp10 += z2 + MULTIPLY(z4, FIX(3.141271809));  /* c1+c5+c9-c13 */
+    tmp12 += z2;
+    z2    = MULTIPLY(z3 + z4, -FIX(1.353318001));  /* -c3 */
+    tmp2  += z2;
+    tmp3  += z2;
+    z2    = MULTIPLY(z4 - z3, FIX(0.410524528));   /* c13 */
+    tmp10 += z2;
+    tmp11 += z2;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int)RIGHT_SHIFT(tmp20 + tmp0,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[15] = range_limit[(int)RIGHT_SHIFT(tmp20 - tmp0,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[1]  = range_limit[(int)RIGHT_SHIFT(tmp21 + tmp1,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[14] = range_limit[(int)RIGHT_SHIFT(tmp21 - tmp1,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[2]  = range_limit[(int)RIGHT_SHIFT(tmp22 + tmp2,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[13] = range_limit[(int)RIGHT_SHIFT(tmp22 - tmp2,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[3]  = range_limit[(int)RIGHT_SHIFT(tmp23 + tmp3,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[12] = range_limit[(int)RIGHT_SHIFT(tmp23 - tmp3,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[4]  = range_limit[(int)RIGHT_SHIFT(tmp24 + tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[11] = range_limit[(int)RIGHT_SHIFT(tmp24 - tmp10,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[5]  = range_limit[(int)RIGHT_SHIFT(tmp25 + tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[10] = range_limit[(int)RIGHT_SHIFT(tmp25 - tmp11,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[6]  = range_limit[(int)RIGHT_SHIFT(tmp26 + tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[9]  = range_limit[(int)RIGHT_SHIFT(tmp26 - tmp12,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[7]  = range_limit[(int)RIGHT_SHIFT(tmp27 + tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+    outptr[8]  = range_limit[(int)RIGHT_SHIFT(tmp27 - tmp13,
+                                              CONST_BITS + PASS1_BITS + 3) &
+                             RANGE_MASK];
+
+    wsptr += 8;         /* advance pointer to next row */
+  }
+}
+
+#endif /* IDCT_SCALING_SUPPORTED */
+#endif /* DCT_ISLOW_SUPPORTED */