Cleanup.

include/bx/simd128_langext.inl

@@ -3,8 +3,8 @@
  * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
  */
 
-#ifndef BX_SIMD_LANGEXT_H_HEADER_GUARD
-#define BX_SIMD_LANGEXT_H_HEADER_GUARD
+#ifndef BX_SIMD128_LANGEXT_H_HEADER_GUARD
+#define BX_SIMD128_LANGEXT_H_HEADER_GUARD
 
 #define float4_rcp           float4_rcp_ni
 #define float4_orx           float4_orx_ni
@@ -512,4 +512,4 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw , 0xf);
 
 } // namespace bx
 
-#endif // BX_SIMD_LANGEXT_H_HEADER_GUARD
+#endif // BX_SIMD128_LANGEXT_H_HEADER_GUARD

include/bx/simd128_neon.inl

@@ -3,8 +3,8 @@
  * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
  */
 
-#ifndef BX_SIMD_NEON_H_HEADER_GUARD
-#define BX_SIMD_NEON_H_HEADER_GUARD
+#ifndef BX_SIMD128_NEON_H_HEADER_GUARD
+#define BX_SIMD128_NEON_H_HEADER_GUARD
 
 #define simd_rcp           simd_rcp_ni
 #define simd_orx           simd_orx_ni
@@ -50,7 +50,7 @@ namespace bx
 				return __builtin_shuffle(_a, (uint32x4_t){ ELEM##_x, ELEM##_y, ELEM##_z, ELEM##_w }); \
 			}
 
-#include "simd_swizzle.inl"
+#include "simd128_swizzle.inl"
 
 #undef BX_SIMD128_IMPLEMENT_SWIZZLE
 #undef ELEMw
@@ -559,4 +559,4 @@ BX_SIMD128_IMPLEMENT_TEST(yzw, yzww);
 
 } // namespace bx
 
-#endif // BX_SIMD_NEON_H_HEADER_GUARD
+#endif // BX_SIMD128_NEON_H_HEADER_GUARD

include/bx/simd128_ref.inl

@@ -3,8 +3,8 @@
  * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
  */
 
-#ifndef BX_SIMD_REF_H_HEADER_GUARD
-#define BX_SIMD_REF_H_HEADER_GUARD
+#ifndef BX_SIMD128_REF_H_HEADER_GUARD
+#define BX_SIMD128_REF_H_HEADER_GUARD
 
 #include <math.h> // sqrtf
 
@@ -57,7 +57,7 @@ namespace bx
 				return result; \
 			}
 
-#include "simd_swizzle.inl"
+#include "simd128_swizzle.inl"
 
 #undef BX_SIMD128_IMPLEMENT_SWIZZLE
 #undef ELEMw
@@ -647,4 +647,4 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw , 0xf);
 
 } // namespace bx
 
-#endif // BX_SIMD_REF_H_HEADER_GUARD
+#endif // BX_SIMD128_REF_H_HEADER_GUARD

include/bx/simd128_sse.inl

@@ -3,8 +3,8 @@
  * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
  */
 
-#ifndef BX_SIMD_SSE_H_HEADER_GUARD
-#define BX_SIMD_SSE_H_HEADER_GUARD
+#ifndef BX_SIMD128_SSE_H_HEADER_GUARD
+#define BX_SIMD128_SSE_H_HEADER_GUARD
 
 #include "simd_ni.inl"
 
@@ -21,7 +21,7 @@ namespace bx
 				return _mm_shuffle_ps( _a, _a, _MM_SHUFFLE(ELEM##_w, ELEM##_z, ELEM##_y, ELEM##_x ) ); \
 			}
 
-#include "simd_swizzle.inl"
+#include "simd128_swizzle.inl"
 
 #undef BX_SIMD128_IMPLEMENT_SWIZZLE
 #undef ELEMw
@@ -644,4 +644,4 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw , 0xf);
 
 } // namespace bx
 
-#endif // BX_SIMD_SSE_H_HEADER_GUARD
+#endif // BX_SIMD128_SSE_H_HEADER_GUARD

include/bx/simd256_avx.inl

@@ -0,0 +1,9 @@
+/*
+ * Copyright 2010-2016 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
+ */
+
+#ifndef BX_SIMD256_AVX_H_HEADER_GUARD
+#define BX_SIMD256_AVX_H_HEADER_GUARD
+
+#endif // BX_SIMD256_AVX_H_HEADER_GUARD

include/bx/simd256_ref.inl

@@ -0,0 +1,9 @@
+/*
+ * Copyright 2010-2016 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
+ */
+
+#ifndef BX_SIMD256_REF_H_HEADER_GUARD
+#define BX_SIMD256_REF_H_HEADER_GUARD
+
+#endif // BX_SIMD256_REF_H_HEADER_GUARD

include/bx/simd_t.h

@@ -11,10 +11,16 @@
 #define BX_SIMD_FORCE_INLINE BX_FORCE_INLINE
 #define BX_SIMD_INLINE inline
 
-#define BX_SIMD_SSE     0
 #define BX_SIMD_AVX     0
-#define BX_SIMD_NEON    0
 #define BX_SIMD_LANGEXT 0
+#define BX_SIMD_NEON    0
+#define BX_SIMD_SSE     0
+
+#if defined(__AVX__) || defined(__AVX2__)
+#	include <immintrin.h>
+#	undef  BX_SIMD_AVX
+#	define BX_SIMD_AVX 1
+#endif //
 
 #if defined(__SSE2__) || (BX_COMPILER_MSVC && (BX_ARCH_64BIT || _M_IX86_FP >= 2) )
 #	include <emmintrin.h> // __m128i
@@ -46,7 +52,7 @@ namespace bx
 #define BX_SIMD128_IMPLEMENT_SWIZZLE(_x, _y, _z, _w) \
 			template<typename Ty> \
 			BX_SIMD_FORCE_INLINE Ty simd_swiz_##_x##_y##_z##_w(Ty _a);
-#include "simd_swizzle.inl"
+#include "simd128_swizzle.inl"
 
 #undef BX_SIMD128_IMPLEMENT_SWIZZLE
 #undef ELEMw
@@ -321,14 +327,10 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw);
 	template<typename Ty>
 	BX_SIMD_INLINE Ty simd_floor(Ty _a);
 
-#if BX_SIMD_SSE
-	typedef __m128 simd128_sse_t;
+#if BX_SIMD_AVX
+	typedef __m256 simd256_avx_t;
 #endif // BX_SIMD_SSE
 
-#if BX_SIMD_NEON
-	typedef float32x4_t simd128_neon_t;
-#endif // BX_SIMD_NEON
-
 #if BX_SIMD_LANGEXT
 	union simd128_langext_t
 	{
@@ -342,6 +344,14 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw);
 	};
 #endif // BX_SIMD_LANGEXT
 
+#if BX_SIMD_NEON
+	typedef float32x4_t simd128_neon_t;
+#endif // BX_SIMD_NEON
+
+#if BX_SIMD_SSE
+	typedef __m128 simd128_sse_t;
+#endif // BX_SIMD_SSE
+
 	union simd128_ref_t
 	{
 		float    fxyzw[4];
@@ -352,26 +362,31 @@ BX_SIMD128_IMPLEMENT_TEST(xyzw);
 
 } // namespace bx
 
-#if BX_SIMD_SSE
-#	include "simd128_sse.inl"
-#endif // BX_SIMD_SSE
+#if BX_SIMD_AVX
+#	include "simd256_avx.inl"
+#endif // BX_SIMD_AVX
+
+#if BX_SIMD_LANGEXT
+#	include "simd128_langext.inl"
+#endif // BX_SIMD_LANGEXT
 
 #if BX_SIMD_NEON
 #	include "simd128_neon.inl"
 #endif // BX_SIMD_NEON
 
-#if BX_SIMD_LANGEXT
-#	include "simd128_langext.inl"
-#endif // BX_SIMD_LANGEXT
+#if BX_SIMD_SSE
+#	include "simd128_sse.inl"
+#endif // BX_SIMD_SSE
 
 #include "simd128_ref.inl"
+#include "simd256_ref.inl"
 
 namespace bx
 {
-#if !( BX_SIMD_SSE \
-	|| BX_SIMD_AVX \
-	|| BX_SIMD_NEON \
+#if !( BX_SIMD_AVX \
 	|| BX_SIMD_LANGEXT \
+	|| BX_SIMD_NEON \
+	|| BX_SIMD_SSE \
 	 )
 #	ifndef BX_SIMD_WARN_REFERENCE_IMPL
 #		define BX_SIMD_WARN_REFERENCE_IMPL 0
@@ -418,6 +433,6 @@ namespace bx
 	{
 		return simd_isplat<simd128_t>(_a);
 	}
-}
+} // namespace bx
 
 #endif // BX_SIMD_T_H_HEADER_GUARD

tests/simd_t.cpp

@@ -98,7 +98,7 @@ TEST(simd_swizzle)
 #define BX_SIMD128_IMPLEMENT_SWIZZLE(_x, _y, _z, _w) \
 			simd_check_string("" #_x #_y #_z #_w "", simd_swiz_##_x##_y##_z##_w(xyzw) ); \
 
-#include <bx/simd_swizzle.inl>
+#include <bx/simd128_swizzle.inl>
 
 #undef BX_SIMD128_IMPLEMENT_SWIZZLE
 #undef ELEMw