Image Use memcpy() in fill(), fill_rect(), _put_pixelb(), _get_pixelb()

Co-authored-by: Lightning_A <[email protected]>

core/io/image.cpp

@@ -86,20 +86,14 @@ SaveEXRFunc Image::save_exr_func = nullptr;
 
 SavePNGBufferFunc Image::save_png_buffer_func = nullptr;
 
-void Image::_put_pixelb(int p_x, int p_y, uint32_t p_pixelsize, uint8_t *p_data, const uint8_t *p_pixel) {
-	uint32_t ofs = (p_y * width + p_x) * p_pixelsize;
-
-	for (uint32_t i = 0; i < p_pixelsize; i++) {
-		p_data[ofs + i] = p_pixel[i];
-	}
+void Image::_put_pixelb(int p_x, int p_y, uint32_t p_pixel_size, uint8_t *p_data, const uint8_t *p_pixel) {
+	uint32_t ofs = (p_y * width + p_x) * p_pixel_size;
+	memcpy(p_data + ofs, p_pixel, p_pixel_size);
 }
 
-void Image::_get_pixelb(int p_x, int p_y, uint32_t p_pixelsize, const uint8_t *p_data, uint8_t *p_pixel) {
-	uint32_t ofs = (p_y * width + p_x) * p_pixelsize;
-
-	for (uint32_t i = 0; i < p_pixelsize; i++) {
-		p_pixel[i] = p_data[ofs + i];
-	}
+void Image::_get_pixelb(int p_x, int p_y, uint32_t p_pixel_size, const uint8_t *p_data, uint8_t *p_pixel) {
+	uint32_t ofs = (p_y * width + p_x) * p_pixel_size;
+	memcpy(p_pixel, p_data + ofs, p_pixel_size);
 }
 
 int Image::get_format_pixel_size(Format p_format) {
@@ -2697,6 +2691,19 @@ void Image::blend_rect_mask(const Ref<Image> &p_src, const Ref<Image> &p_mask, c
 	}
 }
 
+// Repeats `p_pixel` `p_count` times in consecutive memory.
+// Results in the original pixel and `p_count - 1` subsequent copies of it.
+void Image::_repeat_pixel_over_subsequent_memory(uint8_t *p_pixel, int p_pixel_size, int p_count) {
+	int offset = 1;
+	for (int stride = 1; offset + stride <= p_count; stride *= 2) {
+		memcpy(p_pixel + offset * p_pixel_size, p_pixel, stride * p_pixel_size);
+		offset += stride;
+	}
+	if (offset < p_count) {
+		memcpy(p_pixel + offset * p_pixel_size, p_pixel, (p_count - offset) * p_pixel_size);
+	}
+}
+
 void Image::fill(const Color &p_color) {
 	ERR_FAIL_COND_MSG(!_can_modify(format), "Cannot fill in compressed or custom image formats.");
 
@@ -2707,15 +2714,7 @@ void Image::fill(const Color &p_color) {
 	// Put first pixel with the format-aware API.
 	_set_color_at_ofs(dst_data_ptr, 0, p_color);
 
-	for (int y = 0; y < height; y++) {
-		for (int x = 0; x < width; x++) {
-			uint8_t *dst = &dst_data_ptr[(y * width + x) * pixel_size];
-
-			for (int k = 0; k < pixel_size; k++) {
-				dst[k] = dst_data_ptr[k];
-			}
-		}
-	}
+	_repeat_pixel_over_subsequent_memory(dst_data_ptr, pixel_size, width * height);
 }
 
 void Image::fill_rect(const Rect2 &p_rect, const Color &p_color) {
@@ -2734,13 +2733,13 @@ void Image::fill_rect(const Rect2 &p_rect, const Color &p_color) {
 	uint8_t *rect_first_pixel_ptr = &dst_data_ptr[(r.position.y * width + r.position.x) * pixel_size];
 	_set_color_at_ofs(rect_first_pixel_ptr, 0, p_color);
 
-	for (int y = r.position.y; y < r.position.y + r.size.y; y++) {
-		for (int x = r.position.x; x < r.position.x + r.size.x; x++) {
-			uint8_t *dst = &dst_data_ptr[(y * width + x) * pixel_size];
-
-			for (int k = 0; k < pixel_size; k++) {
-				dst[k] = rect_first_pixel_ptr[k];
-			}
+	if (r.size.x == width) {
+		// No need to fill rows separately.
+		_repeat_pixel_over_subsequent_memory(rect_first_pixel_ptr, pixel_size, width * r.size.y);
+	} else {
+		_repeat_pixel_over_subsequent_memory(rect_first_pixel_ptr, pixel_size, r.size.x);
+		for (int y = 1; y < r.size.y; y++) {
+			memcpy(rect_first_pixel_ptr + y * width * pixel_size, rect_first_pixel_ptr, r.size.x * pixel_size);
 		}
 	}
 }

core/io/image.h

@@ -190,8 +190,10 @@ private:
 	static int _get_dst_image_size(int p_width, int p_height, Format p_format, int &r_mipmaps, int p_mipmaps = -1, int *r_mm_width = nullptr, int *r_mm_height = nullptr);
 	bool _can_modify(Format p_format) const;
 
-	_FORCE_INLINE_ void _put_pixelb(int p_x, int p_y, uint32_t p_pixelsize, uint8_t *p_data, const uint8_t *p_pixel);
-	_FORCE_INLINE_ void _get_pixelb(int p_x, int p_y, uint32_t p_pixelsize, const uint8_t *p_data, uint8_t *p_pixel);
+	_FORCE_INLINE_ void _put_pixelb(int p_x, int p_y, uint32_t p_pixel_size, uint8_t *p_data, const uint8_t *p_pixel);
+	_FORCE_INLINE_ void _get_pixelb(int p_x, int p_y, uint32_t p_pixel_size, const uint8_t *p_data, uint8_t *p_pixel);
+
+	_FORCE_INLINE_ void _repeat_pixel_over_subsequent_memory(uint8_t *p_pixel, int p_pixel_size, int p_count);
 
 	void _set_data(const Dictionary &p_data);
 	Dictionary _get_data() const;