diff options
Diffstat (limited to 'source/row_any.cc')
-rw-r--r-- | source/row_any.cc | 140 |
1 files changed, 102 insertions, 38 deletions
diff --git a/source/row_any.cc b/source/row_any.cc index 97ef844..b1ede4e 100644 --- a/source/row_any.cc +++ b/source/row_any.cc @@ -35,19 +35,19 @@ extern "C" { } #ifdef HAS_I422TOARGBROW_SSSE3 -YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C, +YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, I422ToARGBRow_C, 1, 4, 7) #endif // HAS_I422TOARGBROW_SSSE3 #ifdef HAS_I444TOARGBROW_SSSE3 -YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C, +YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, I444ToARGBRow_C, 0, 4, 7) -YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C, +YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, I411ToARGBRow_C, 2, 4, 7) -YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C, +YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, I422ToBGRARow_C, 1, 4, 7) -YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C, +YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, I422ToABGRRow_C, 1, 4, 7) -YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C, +YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, I422ToRGBARow_C, 1, 4, 7) // I422ToRGB565Row_SSSE3 is unaligned. YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C, @@ -65,6 +65,9 @@ YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15) #ifdef HAS_I422TOARGBROW_AVX2 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15) #endif // HAS_I422TOARGBROW_AVX2 +#ifdef HAS_I422TOBGRAROW_AVX2 +YANY(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, I422ToBGRARow_C, 1, 4, 15) +#endif // HAS_I422TOBGRAROW_AVX2 #ifdef HAS_I422TOARGBROW_NEON YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7) YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7) @@ -79,9 +82,13 @@ YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C, YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C, 1, 2, 7) YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7) +#endif // HAS_I422TOARGBROW_NEON +#ifdef HAS_I422TOYUY2ROW_NEON YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15) +#endif // HAS_I422TOYUY2ROW_NEON +#ifdef HAS_I422TOUYVYROW_NEON YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15) -#endif // HAS_I422TOARGBROW_NEON +#endif // HAS_I422TOUYVYROW_NEON #undef YANY // Wrappers to handle odd width @@ -98,9 +105,9 @@ YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15) } #ifdef HAS_NV12TOARGBROW_SSSE3 -NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C, +NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, NV12ToARGBRow_C, 0, 4) -NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C, +NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, NV21ToARGBRow_C, 0, 4) #endif // HAS_NV12TOARGBROW_SSSE3 #ifdef HAS_NV12TOARGBROW_NEON @@ -141,15 +148,15 @@ RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C, 3, 4, 2) #endif #if defined(HAS_I400TOARGBROW_SSE2) -RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C, +RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C, 7, 1, 4) #endif #if defined(HAS_YTOARGBROW_SSE2) RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C, 7, 1, 4) -RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C, +RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C, 15, 2, 4) -RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C, +RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, UYVYToARGBRow_C, 15, 2, 4) // These require alignment on ARGB, so C is used for remainder. RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C, @@ -227,35 +234,67 @@ YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32) YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32) #endif #ifdef HAS_ARGBTOYROW_SSSE3 -YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16) +YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 4, 1, 16) #endif #ifdef HAS_BGRATOYROW_SSSE3 -YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16) -YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16) -YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16) -YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16) -YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16) +YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 4, 1, 16) +YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 4, 1, 16) +YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 4, 1, 16) +YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 2, 1, 16) +YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 2, 1, 16) #endif #ifdef HAS_ARGBTOYJROW_SSSE3 -YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16) +YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 4, 1, 16) #endif #ifdef HAS_ARGBTOYROW_NEON YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8) +#endif +#ifdef HAS_ARGBTOYJROW_NEON YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8) +#endif +#ifdef HAS_BGRATOYROW_NEON YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8) +#endif +#ifdef HAS_ABGRTOYROW_NEON YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8) +#endif +#ifdef HAS_RGBATOYROW_NEON YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8) +#endif +#ifdef HAS_RGB24TOYROW_NEON YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8) +#endif +#ifdef HAS_RAWTOYROW_NEON YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8) +#endif +#ifdef HAS_RGB565TOYROW_NEON YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8) +#endif +#ifdef HAS_ARGB1555TOYROW_NEON YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8) +#endif +#ifdef HAS_ARGB4444TOYROW_NEON YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8) +#endif +#ifdef HAS_YUY2TOYROW_NEON YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16) +#endif +#ifdef HAS_UYVYTOYROW_NEON YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16) +#endif +#ifdef HAS_RGB24TOARGBROW_NEON YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8) +#endif +#ifdef HAS_RAWTOARGBROW_NEON YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8) +#endif +#ifdef HAS_RGB565TOARGBROW_NEON YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8) +#endif +#ifdef HAS_ARGB1555TOARGBROW_NEON YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8) +#endif +#ifdef HAS_ARGB4444TOARGBROW_NEON YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8) #endif #undef YANY @@ -313,27 +352,50 @@ UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31) UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31) #endif #ifdef HAS_ARGBTOUVROW_SSSE3 -UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15) -UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C, - 4, 15) -UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15) -UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15) -UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15) -UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15) -UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15) +UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, ARGBToUVRow_C, 4, 15) +UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, ARGBToUVJRow_C, 4, 15) +UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, BGRAToUVRow_C, 4, 15) +UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, ABGRToUVRow_C, 4, 15) +UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, RGBAToUVRow_C, 4, 15) +#endif +#ifdef HAS_YUY2TOUVROW_SSE2 +UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, YUY2ToUVRow_C, 2, 15) +UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, UYVYToUVRow_C, 2, 15) #endif #ifdef HAS_ARGBTOUVROW_NEON UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15) +#endif +#ifdef HAS_ARGBTOUVJROW_NEON UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15) +#endif +#ifdef HAS_BGRATOUVROW_NEON UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15) +#endif +#ifdef HAS_ABGRTOUVROW_NEON UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15) +#endif +#ifdef HAS_RGBATOUVROW_NEON UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15) +#endif +#ifdef HAS_RGB24TOUVROW_NEON UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15) +#endif +#ifdef HAS_RAWTOUVROW_NEON UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15) +#endif +#ifdef HAS_RGB565TOUVROW_NEON UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15) +#endif +#ifdef HAS_ARGB1555TOUVROW_NEON UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15) +#endif +#ifdef HAS_ARGB4444TOUVROW_NEON UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15) +#endif +#ifdef HAS_YUY2TOUVROW_NEON UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15) +#endif +#ifdef HAS_UYVYTOUVROW_NEON UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15) #endif #undef UVANY @@ -350,7 +412,7 @@ UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15) } #ifdef HAS_ARGBTOUV444ROW_SSSE3 -UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3, +UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, ARGBToUV444Row_C, 4, 15, 0) #endif #ifdef HAS_YUY2TOUV422ROW_AVX2 @@ -359,12 +421,14 @@ UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2, UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2, UYVYToUV422Row_C, 2, 31, 1) #endif -#ifdef HAS_ARGBTOUVROW_SSSE3 -UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3, +#ifdef HAS_ARGBTOUV422ROW_SSSE3 +UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3, ARGBToUV422Row_C, 4, 15, 1) -UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2, +#endif +#ifdef HAS_YUY2TOUV422ROW_SSE2 +UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2, YUY2ToUV422Row_C, 2, 15, 1) -UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2, +UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2, UYVYToUV422Row_C, 2, 15, 1) #endif #ifdef HAS_YUY2TOUV422ROW_NEON @@ -393,7 +457,7 @@ UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, } #ifdef HAS_SPLITUVROW_SSE2 -SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15) +SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, SplitUVRow_C, 15) #endif #ifdef HAS_SPLITUVROW_AVX2 SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31) @@ -402,7 +466,7 @@ SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31) SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15) #endif #ifdef HAS_SPLITUVROW_MIPS_DSPR2 -SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2, +SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2, SplitUVRow_C, 15) #endif #undef SPLITUVROWANY @@ -419,7 +483,7 @@ SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2, } #ifdef HAS_MERGEUVROW_SSE2 -MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15) +MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, MergeUVRow_C, 15) #endif #ifdef HAS_MERGEUVROW_AVX2 MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31) @@ -490,7 +554,7 @@ YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2, ARGBShuffleRow_C, 4, 4, 3) #endif #ifdef HAS_ARGBSHUFFLEROW_SSSE3 -YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3, +YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, ARGBShuffleRow_C, 4, 4, 7) #endif #ifdef HAS_ARGBSHUFFLEROW_AVX2 @@ -521,11 +585,11 @@ NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, InterpolateRow_C, 1, 1, 32) #endif #ifdef HAS_INTERPOLATEROW_SSSE3 -NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3, +NANY(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, InterpolateRow_C, 1, 1, 15) #endif #ifdef HAS_INTERPOLATEROW_SSE2 -NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2, +NANY(InterpolateRow_Any_SSE2, InterpolateRow_SSE2, InterpolateRow_C, 1, 1, 15) #endif #ifdef HAS_INTERPOLATEROW_NEON |