diff options
Diffstat (limited to 'files/source/convert_from_argb.cc')
| -rw-r--r-- | files/source/convert_from_argb.cc | 1125 |
1 files changed, 833 insertions, 292 deletions
diff --git a/files/source/convert_from_argb.cc b/files/source/convert_from_argb.cc index fbcd039d..e50c2af3 100644 --- a/files/source/convert_from_argb.cc +++ b/files/source/convert_from_argb.cc @@ -76,11 +76,11 @@ int ARGBToI444(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOUV444ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUV444Row = ARGBToUV444Row_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToUV444Row = ARGBToUV444Row_MMI; +#if defined(HAS_ARGBTOUV444ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToUV444Row = ARGBToUV444Row_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToUV444Row = ARGBToUV444Row_LASX; } } #endif @@ -103,7 +103,7 @@ int ARGBToI444(const uint8_t* src_argb, #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -116,11 +116,11 @@ int ARGBToI444(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; } } #endif @@ -170,30 +170,42 @@ int ARGBToI422(const uint8_t* src_argb, height = 1; src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0; } -#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) +#if defined(HAS_ARGBTOYROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVRow = ARGBToUVRow_Any_SSSE3; ARGBToYRow = ARGBToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_SSSE3; ARGBToYRow = ARGBToYRow_SSSE3; } } #endif -#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) +#if defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { - ARGBToUVRow = ARGBToUVRow_Any_AVX2; ARGBToYRow = ARGBToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_AVX2; ARGBToYRow = ARGBToYRow_AVX2; } } #endif +#if defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + } + } +#endif #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -206,37 +218,26 @@ int ARGBToI422(const uint8_t* src_argb, } } #endif - -#if defined(HAS_ARGBTOYROW_MSA) +#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYRow = ARGBToYRow_Any_MSA; + ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOUVROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVRow = ARGBToUVRow_MSA; } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVRow = ARGBToUVRow_Any_MMI; - if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + ARGBToUVRow = ARGBToUVRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; + ARGBToUVRow = ARGBToUVRow_LASX; } } #endif @@ -279,30 +280,10 @@ int ARGBToNV12(const uint8_t* src_argb, src_argb = src_argb + (height - 1) * src_stride_argb; src_stride_argb = -src_stride_argb; } -#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) - if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVRow = ARGBToUVRow_Any_SSSE3; - ARGBToYRow = ARGBToYRow_Any_SSSE3; - if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_SSSE3; - ARGBToYRow = ARGBToYRow_SSSE3; - } - } -#endif -#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) - if (TestCpuFlag(kCpuHasAVX2)) { - ARGBToUVRow = ARGBToUVRow_Any_AVX2; - ARGBToYRow = ARGBToYRow_Any_AVX2; - if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_AVX2; - ARGBToYRow = ARGBToYRow_AVX2; - } - } -#endif #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -315,35 +296,57 @@ int ARGBToNV12(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToYRow = ARGBToYRow_Any_MSA; +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToYRow = ARGBToYRow_MSA; + ARGBToYRow = ARGBToYRow_SSSE3; } } #endif -#if defined(HAS_ARGBTOUVROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVRow = ARGBToUVRow_Any_MSA; +#if defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_MSA; + ARGBToYRow = ARGBToYRow_AVX2; } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; +#if defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; } } #endif -#if defined(HAS_ARGBTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVRow = ARGBToUVRow_Any_MMI; +#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + ARGBToYRow = ARGBToYRow_Any_MSA; + ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_MMI; + ARGBToYRow = ARGBToYRow_MSA; + } + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_MSA; + } + } +#endif +#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + ARGBToUVRow = ARGBToUVRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; + ARGBToUVRow = ARGBToUVRow_LASX; } } #endif @@ -379,11 +382,11 @@ int ARGBToNV12(const uint8_t* src_argb, } } #endif -#if defined(HAS_MERGEUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - MergeUVRow_ = MergeUVRow_Any_MMI; - if (IS_ALIGNED(halfwidth, 8)) { - MergeUVRow_ = MergeUVRow_MMI; +#if defined(HAS_MERGEUVROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + MergeUVRow_ = MergeUVRow_Any_LSX; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_LSX; } } #endif @@ -439,30 +442,42 @@ int ARGBToNV21(const uint8_t* src_argb, src_argb = src_argb + (height - 1) * src_stride_argb; src_stride_argb = -src_stride_argb; } -#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) +#if defined(HAS_ARGBTOYROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVRow = ARGBToUVRow_Any_SSSE3; ARGBToYRow = ARGBToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_SSSE3; ARGBToYRow = ARGBToYRow_SSSE3; } } #endif -#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) +#if defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { - ARGBToUVRow = ARGBToUVRow_Any_AVX2; ARGBToYRow = ARGBToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_AVX2; ARGBToYRow = ARGBToYRow_AVX2; } } #endif +#if defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + } + } +#endif #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -475,39 +490,28 @@ int ARGBToNV21(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MSA) +#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYRow = ARGBToYRow_Any_MSA; + ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOUVROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVRow = ARGBToUVRow_MSA; } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVRow = ARGBToUVRow_Any_MMI; - if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + ARGBToUVRow = ARGBToUVRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; + ARGBToUVRow = ARGBToUVRow_LASX; } } #endif - #if defined(HAS_MERGEUVROW_SSE2) if (TestCpuFlag(kCpuHasSSE2)) { MergeUVRow_ = MergeUVRow_Any_SSE2; @@ -540,11 +544,11 @@ int ARGBToNV21(const uint8_t* src_argb, } } #endif -#if defined(HAS_MERGEUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - MergeUVRow_ = MergeUVRow_Any_MMI; - if (IS_ALIGNED(halfwidth, 8)) { - MergeUVRow_ = MergeUVRow_MMI; +#if defined(HAS_MERGEUVROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + MergeUVRow_ = MergeUVRow_Any_LSX; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_LSX; } } #endif @@ -599,30 +603,42 @@ int ABGRToNV12(const uint8_t* src_abgr, src_abgr = src_abgr + (height - 1) * src_stride_abgr; src_stride_abgr = -src_stride_abgr; } -#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3) +#if defined(HAS_ABGRTOYROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ABGRToUVRow = ABGRToUVRow_Any_SSSE3; ABGRToYRow = ABGRToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ABGRToUVRow = ABGRToUVRow_SSSE3; ABGRToYRow = ABGRToYRow_SSSE3; } } #endif -#if defined(HAS_ABGRTOYROW_AVX2) && defined(HAS_ABGRTOUVROW_AVX2) +#if defined(HAS_ABGRTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ABGRToUVRow = ABGRToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ABGRToUVRow = ABGRToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ABGRTOYROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { - ABGRToUVRow = ABGRToUVRow_Any_AVX2; ABGRToYRow = ABGRToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ABGRToUVRow = ABGRToUVRow_AVX2; ABGRToYRow = ABGRToYRow_AVX2; } } #endif +#if defined(HAS_ABGRTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ABGRToUVRow = ABGRToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ABGRToUVRow = ABGRToUVRow_AVX2; + } + } +#endif #if defined(HAS_ABGRTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ABGRToYRow = ABGRToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ABGRToYRow = ABGRToYRow_NEON; } } @@ -635,35 +651,167 @@ int ABGRToNV12(const uint8_t* src_abgr, } } #endif -#if defined(HAS_ABGRTOYROW_MSA) +#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ABGRToYRow = ABGRToYRow_Any_MSA; + ABGRToUVRow = ABGRToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ABGRToYRow = ABGRToYRow_MSA; } + if (IS_ALIGNED(width, 32)) { + ABGRToUVRow = ABGRToUVRow_MSA; + } + } +#endif +#if defined(HAS_MERGEUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MergeUVRow_ = MergeUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_SSE2; + } + } +#endif +#if defined(HAS_MERGEUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MergeUVRow_ = MergeUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + MergeUVRow_ = MergeUVRow_AVX2; + } } #endif -#if defined(HAS_ABGRTOUVROW_MSA) +#if defined(HAS_MERGEUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MergeUVRow_ = MergeUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_NEON; + } + } +#endif +#if defined(HAS_MERGEUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { - ABGRToUVRow = ABGRToUVRow_Any_MSA; + MergeUVRow_ = MergeUVRow_Any_MSA; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_MSA; + } + } +#endif +#if defined(HAS_MERGEUVROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + MergeUVRow_ = MergeUVRow_Any_LSX; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_LSX; + } + } +#endif + { + // Allocate a rows of uv. + align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2); + uint8_t* row_v = row_u + ((halfwidth + 31) & ~31); + + for (y = 0; y < height - 1; y += 2) { + ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width); + MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + ABGRToYRow(src_abgr, dst_y, width); + ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width); + src_abgr += src_stride_abgr * 2; + dst_y += dst_stride_y * 2; + dst_uv += dst_stride_uv; + } + if (height & 1) { + ABGRToUVRow(src_abgr, 0, row_u, row_v, width); + MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + ABGRToYRow(src_abgr, dst_y, width); + } + free_aligned_buffer_64(row_u); + } + return 0; +} + +// Same as NV12 but U and V swapped. +LIBYUV_API +int ABGRToNV21(const uint8_t* src_abgr, + int src_stride_abgr, + uint8_t* dst_y, + int dst_stride_y, + uint8_t* dst_vu, + int dst_stride_vu, + int width, + int height) { + int y; + int halfwidth = (width + 1) >> 1; + void (*ABGRToUVRow)(const uint8_t* src_abgr0, int src_stride_abgr, + uint8_t* dst_u, uint8_t* dst_v, int width) = + ABGRToUVRow_C; + void (*ABGRToYRow)(const uint8_t* src_abgr, uint8_t* dst_y, int width) = + ABGRToYRow_C; + void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v, + uint8_t* dst_vu, int width) = MergeUVRow_C; + if (!src_abgr || !dst_y || !dst_vu || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_abgr = src_abgr + (height - 1) * src_stride_abgr; + src_stride_abgr = -src_stride_abgr; + } +#if defined(HAS_ABGRTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ABGRToYRow = ABGRToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ABGRToYRow = ABGRToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ABGRTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ABGRToUVRow = ABGRToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ABGRToUVRow = ABGRToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ABGRTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ABGRToYRow = ABGRToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ABGRToUVRow = ABGRToUVRow_MSA; + ABGRToYRow = ABGRToYRow_AVX2; } } #endif -#if defined(HAS_ABGRTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ABGRToYRow = ABGRToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ABGRToYRow = ABGRToYRow_MMI; +#if defined(HAS_ABGRTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ABGRToUVRow = ABGRToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ABGRToUVRow = ABGRToUVRow_AVX2; + } + } +#endif +#if defined(HAS_ABGRTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ABGRToYRow = ABGRToYRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ABGRToYRow = ABGRToYRow_NEON; + } + } +#endif +#if defined(HAS_ABGRTOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ABGRToUVRow = ABGRToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ABGRToUVRow = ABGRToUVRow_NEON; } } #endif -#if defined(HAS_ABGRTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ABGRToUVRow = ABGRToUVRow_Any_MMI; +#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + ABGRToYRow = ABGRToYRow_Any_MSA; + ABGRToUVRow = ABGRToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { - ABGRToUVRow = ABGRToUVRow_MMI; + ABGRToYRow = ABGRToYRow_MSA; + } + if (IS_ALIGNED(width, 32)) { + ABGRToUVRow = ABGRToUVRow_MSA; } } #endif @@ -699,11 +847,11 @@ int ABGRToNV12(const uint8_t* src_abgr, } } #endif -#if defined(HAS_MERGEUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - MergeUVRow_ = MergeUVRow_Any_MMI; - if (IS_ALIGNED(halfwidth, 8)) { - MergeUVRow_ = MergeUVRow_MMI; +#if defined(HAS_MERGEUVROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + MergeUVRow_ = MergeUVRow_Any_LSX; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_LSX; } } #endif @@ -714,16 +862,16 @@ int ABGRToNV12(const uint8_t* src_abgr, for (y = 0; y < height - 1; y += 2) { ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width); - MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); ABGRToYRow(src_abgr, dst_y, width); ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width); src_abgr += src_stride_abgr * 2; dst_y += dst_stride_y * 2; - dst_uv += dst_stride_uv; + dst_vu += dst_stride_vu; } if (height & 1) { ABGRToUVRow(src_abgr, 0, row_u, row_v, width); - MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); ABGRToYRow(src_abgr, dst_y, width); } free_aligned_buffer_64(row_u); @@ -764,30 +912,42 @@ int ARGBToYUY2(const uint8_t* src_argb, height = 1; src_stride_argb = dst_stride_yuy2 = 0; } -#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) +#if defined(HAS_ARGBTOYROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVRow = ARGBToUVRow_Any_SSSE3; ARGBToYRow = ARGBToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_SSSE3; ARGBToYRow = ARGBToYRow_SSSE3; } } #endif -#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) +#if defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { - ARGBToUVRow = ARGBToUVRow_Any_AVX2; ARGBToYRow = ARGBToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_AVX2; ARGBToYRow = ARGBToYRow_AVX2; } } #endif +#if defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + } + } +#endif #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -800,35 +960,25 @@ int ARGBToYUY2(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MSA) +#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYRow = ARGBToYRow_Any_MSA; + ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOUVROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVRow = ARGBToUVRow_MSA; } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVRow = ARGBToUVRow_Any_MMI; - if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + ARGBToUVRow = ARGBToUVRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; + ARGBToUVRow = ARGBToUVRow_LASX; } } #endif @@ -864,11 +1014,11 @@ int ARGBToYUY2(const uint8_t* src_argb, } } #endif -#if defined(HAS_I422TOYUY2ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - I422ToYUY2Row = I422ToYUY2Row_Any_MMI; - if (IS_ALIGNED(width, 8)) { - I422ToYUY2Row = I422ToYUY2Row_MMI; +#if defined(HAS_I422TOYUY2ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + I422ToYUY2Row = I422ToYUY2Row_Any_LASX; + if (IS_ALIGNED(width, 32)) { + I422ToYUY2Row = I422ToYUY2Row_LASX; } } #endif @@ -925,30 +1075,42 @@ int ARGBToUYVY(const uint8_t* src_argb, height = 1; src_stride_argb = dst_stride_uyvy = 0; } -#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) +#if defined(HAS_ARGBTOYROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVRow = ARGBToUVRow_Any_SSSE3; ARGBToYRow = ARGBToYRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_SSSE3; ARGBToYRow = ARGBToYRow_SSSE3; } } #endif -#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) +#if defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { - ARGBToUVRow = ARGBToUVRow_Any_AVX2; ARGBToYRow = ARGBToYRow_Any_AVX2; if (IS_ALIGNED(width, 32)) { - ARGBToUVRow = ARGBToUVRow_AVX2; ARGBToYRow = ARGBToYRow_AVX2; } } #endif +#if defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + } + } +#endif #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -961,35 +1123,25 @@ int ARGBToUYVY(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MSA) +#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYRow = ARGBToYRow_Any_MSA; + ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOUVROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVRow = ARGBToUVRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVRow = ARGBToUVRow_MSA; } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVRow = ARGBToUVRow_Any_MMI; - if (IS_ALIGNED(width, 16)) { - ARGBToUVRow = ARGBToUVRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + ARGBToUVRow = ARGBToUVRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; + ARGBToUVRow = ARGBToUVRow_LASX; } } #endif @@ -1025,11 +1177,11 @@ int ARGBToUYVY(const uint8_t* src_argb, } } #endif -#if defined(HAS_I422TOUYVYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - I422ToUYVYRow = I422ToUYVYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - I422ToUYVYRow = I422ToUYVYRow_MMI; +#if defined(HAS_I422TOUYVYROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + I422ToUYVYRow = I422ToUYVYRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + I422ToUYVYRow = I422ToUYVYRow_LASX; } } #endif @@ -1097,7 +1249,7 @@ int ARGBToI400(const uint8_t* src_argb, #if defined(HAS_ARGBTOYROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYRow = ARGBToYRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYRow = ARGBToYRow_NEON; } } @@ -1110,11 +1262,11 @@ int ARGBToI400(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYRow = ARGBToYRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYRow = ARGBToYRow_MMI; +#if defined(HAS_ARGBTOYROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYRow = ARGBToYRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_LASX; } } #endif @@ -1195,7 +1347,7 @@ int ARGBToRGB24(const uint8_t* src_argb, #if defined(HAS_ARGBTORGB24ROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToRGB24Row = ARGBToRGB24Row_NEON; } } @@ -1208,11 +1360,11 @@ int ARGBToRGB24(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTORGB24ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToRGB24Row = ARGBToRGB24Row_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToRGB24Row = ARGBToRGB24Row_MMI; +#if defined(HAS_ARGBTORGB24ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToRGB24Row = ARGBToRGB24Row_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToRGB24Row = ARGBToRGB24Row_LASX; } } #endif @@ -1282,11 +1434,11 @@ int ARGBToRAW(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTORAWROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToRAWRow = ARGBToRAWRow_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToRAWRow = ARGBToRAWRow_MMI; +#if defined(HAS_ARGBTORAWROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToRAWRow = ARGBToRAWRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToRAWRow = ARGBToRAWRow_LASX; } } #endif @@ -1360,11 +1512,11 @@ int ARGBToRGB565Dither(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTORGB565DITHERROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MMI; +#if defined(HAS_ARGBTORGB565DITHERROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_LASX; + if (IS_ALIGNED(width, 16)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_LASX; } } #endif @@ -1437,11 +1589,11 @@ int ARGBToRGB565(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTORGB565ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToRGB565Row = ARGBToRGB565Row_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToRGB565Row = ARGBToRGB565Row_MMI; +#if defined(HAS_ARGBTORGB565ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToRGB565Row = ARGBToRGB565Row_Any_LASX; + if (IS_ALIGNED(width, 16)) { + ARGBToRGB565Row = ARGBToRGB565Row_LASX; } } #endif @@ -1511,11 +1663,11 @@ int ARGBToARGB1555(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOARGB1555ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToARGB1555Row = ARGBToARGB1555Row_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToARGB1555Row = ARGBToARGB1555Row_MMI; +#if defined(HAS_ARGBTOARGB1555ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_Any_LASX; + if (IS_ALIGNED(width, 16)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_LASX; } } #endif @@ -1585,11 +1737,11 @@ int ARGBToARGB4444(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOARGB4444ROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToARGB4444Row = ARGBToARGB4444Row_Any_MMI; - if (IS_ALIGNED(width, 4)) { - ARGBToARGB4444Row = ARGBToARGB4444Row_MMI; +#if defined(HAS_ARGBTOARGB4444ROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_Any_LASX; + if (IS_ALIGNED(width, 16)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_LASX; } } #endif @@ -1727,16 +1879,22 @@ int ARGBToJ420(const uint8_t* src_argb, src_argb = src_argb + (height - 1) * src_stride_argb; src_stride_argb = -src_stride_argb; } -#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3) +#if defined(HAS_ARGBTOYJROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; ARGBToYJRow = ARGBToYJRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVJRow = ARGBToUVJRow_SSSE3; ARGBToYJRow = ARGBToYJRow_SSSE3; } } #endif +#if defined(HAS_ARGBTOUVJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJRow = ARGBToUVJRow_SSSE3; + } + } +#endif #if defined(HAS_ARGBTOYJROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { ARGBToYJRow = ARGBToYJRow_Any_AVX2; @@ -1748,7 +1906,7 @@ int ARGBToJ420(const uint8_t* src_argb, #if defined(HAS_ARGBTOYJROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYJRow = ARGBToYJRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYJRow = ARGBToYJRow_NEON; } } @@ -1761,35 +1919,35 @@ int ARGBToJ420(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYJROW_MSA) +#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYJRow = ARGBToYJRow_Any_MSA; + ARGBToUVJRow = ARGBToUVJRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYJRow = ARGBToYJRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOYJROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYJRow = ARGBToYJRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYJRow = ARGBToYJRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVJROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVJRow = ARGBToUVJRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVJRow = ARGBToUVJRow_MSA; } } #endif -#if defined(HAS_ARGBTOUVJROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVJRow = ARGBToUVJRow_Any_MMI; +#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + ARGBToYJRow = ARGBToYJRow_Any_LSX; + ARGBToUVJRow = ARGBToUVJRow_Any_LSX; if (IS_ALIGNED(width, 16)) { - ARGBToUVJRow = ARGBToUVJRow_MMI; + ARGBToYJRow = ARGBToYJRow_LSX; + ARGBToUVJRow = ARGBToUVJRow_LSX; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYJRow = ARGBToYJRow_Any_LASX; + ARGBToUVJRow = ARGBToUVJRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_LASX; + ARGBToUVJRow = ARGBToUVJRow_LASX; } } #endif @@ -1844,16 +2002,22 @@ int ARGBToJ422(const uint8_t* src_argb, height = 1; src_stride_argb = dst_stride_yj = dst_stride_u = dst_stride_v = 0; } -#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3) +#if defined(HAS_ARGBTOYJROW_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { - ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; ARGBToYJRow = ARGBToYJRow_Any_SSSE3; if (IS_ALIGNED(width, 16)) { - ARGBToUVJRow = ARGBToUVJRow_SSSE3; ARGBToYJRow = ARGBToYJRow_SSSE3; } } #endif +#if defined(HAS_ARGBTOUVJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJRow = ARGBToUVJRow_SSSE3; + } + } +#endif #if defined(HAS_ARGBTOYJROW_AVX2) if (TestCpuFlag(kCpuHasAVX2)) { ARGBToYJRow = ARGBToYJRow_Any_AVX2; @@ -1865,7 +2029,7 @@ int ARGBToJ422(const uint8_t* src_argb, #if defined(HAS_ARGBTOYJROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYJRow = ARGBToYJRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYJRow = ARGBToYJRow_NEON; } } @@ -1878,35 +2042,35 @@ int ARGBToJ422(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYJROW_MSA) +#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA) if (TestCpuFlag(kCpuHasMSA)) { ARGBToYJRow = ARGBToYJRow_Any_MSA; + ARGBToUVJRow = ARGBToUVJRow_Any_MSA; if (IS_ALIGNED(width, 16)) { ARGBToYJRow = ARGBToYJRow_MSA; } - } -#endif -#if defined(HAS_ARGBTOYJROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYJRow = ARGBToYJRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYJRow = ARGBToYJRow_MMI; - } - } -#endif -#if defined(HAS_ARGBTOUVJROW_MSA) - if (TestCpuFlag(kCpuHasMSA)) { - ARGBToUVJRow = ARGBToUVJRow_Any_MSA; if (IS_ALIGNED(width, 32)) { ARGBToUVJRow = ARGBToUVJRow_MSA; } } #endif -#if defined(HAS_ARGBTOUVJROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToUVJRow = ARGBToUVJRow_Any_MMI; +#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + ARGBToYJRow = ARGBToYJRow_Any_LSX; + ARGBToUVJRow = ARGBToUVJRow_Any_LSX; if (IS_ALIGNED(width, 16)) { - ARGBToUVJRow = ARGBToUVJRow_MMI; + ARGBToYJRow = ARGBToYJRow_LSX; + ARGBToUVJRow = ARGBToUVJRow_LSX; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX) + if (TestCpuFlag(kCpuHasLASX)) { + ARGBToYJRow = ARGBToYJRow_Any_LASX; + ARGBToUVJRow = ARGBToUVJRow_Any_LASX; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_LASX; + ARGBToUVJRow = ARGBToUVJRow_LASX; } } #endif @@ -1922,6 +2086,124 @@ int ARGBToJ422(const uint8_t* src_argb, return 0; } +// Convert ARGB to AR64. +LIBYUV_API +int ARGBToAR64(const uint8_t* src_argb, + int src_stride_argb, + uint16_t* dst_ar64, + int dst_stride_ar64, + int width, + int height) { + int y; + void (*ARGBToAR64Row)(const uint8_t* src_argb, uint16_t* dst_ar64, + int width) = ARGBToAR64Row_C; + if (!src_argb || !dst_ar64 || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && dst_stride_ar64 == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_ar64 = 0; + } +#if defined(HAS_ARGBTOAR64ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToAR64Row = ARGBToAR64Row_Any_SSSE3; + if (IS_ALIGNED(width, 4)) { + ARGBToAR64Row = ARGBToAR64Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOAR64ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToAR64Row = ARGBToAR64Row_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToAR64Row = ARGBToAR64Row_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOAR64ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToAR64Row = ARGBToAR64Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToAR64Row = ARGBToAR64Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToAR64Row(src_argb, dst_ar64, width); + src_argb += src_stride_argb; + dst_ar64 += dst_stride_ar64; + } + return 0; +} + +// Convert ARGB to AB64. +LIBYUV_API +int ARGBToAB64(const uint8_t* src_argb, + int src_stride_argb, + uint16_t* dst_ab64, + int dst_stride_ab64, + int width, + int height) { + int y; + void (*ARGBToAB64Row)(const uint8_t* src_argb, uint16_t* dst_ar64, + int width) = ARGBToAB64Row_C; + if (!src_argb || !dst_ab64 || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && dst_stride_ab64 == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_ab64 = 0; + } +#if defined(HAS_ARGBTOAB64ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToAB64Row = ARGBToAB64Row_Any_SSSE3; + if (IS_ALIGNED(width, 4)) { + ARGBToAB64Row = ARGBToAB64Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOAB64ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToAB64Row = ARGBToAB64Row_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToAB64Row = ARGBToAB64Row_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOAB64ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToAB64Row = ARGBToAB64Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToAB64Row = ARGBToAB64Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToAB64Row(src_argb, dst_ab64, width); + src_argb += src_stride_argb; + dst_ab64 += dst_stride_ab64; + } + return 0; +} + // Convert ARGB to J400. LIBYUV_API int ARGBToJ400(const uint8_t* src_argb, @@ -1966,7 +2248,7 @@ int ARGBToJ400(const uint8_t* src_argb, #if defined(HAS_ARGBTOYJROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { ARGBToYJRow = ARGBToYJRow_Any_NEON; - if (IS_ALIGNED(width, 8)) { + if (IS_ALIGNED(width, 16)) { ARGBToYJRow = ARGBToYJRow_NEON; } } @@ -1979,23 +2261,282 @@ int ARGBToJ400(const uint8_t* src_argb, } } #endif -#if defined(HAS_ARGBTOYJROW_MMI) - if (TestCpuFlag(kCpuHasMMI)) { - ARGBToYJRow = ARGBToYJRow_Any_MMI; - if (IS_ALIGNED(width, 8)) { - ARGBToYJRow = ARGBToYJRow_MMI; + + for (y = 0; y < height; ++y) { + ARGBToYJRow(src_argb, dst_yj, width); + src_argb += src_stride_argb; + dst_yj += dst_stride_yj; + } + return 0; +} + +// Convert RGBA to J400. +LIBYUV_API +int RGBAToJ400(const uint8_t* src_rgba, + int src_stride_rgba, + uint8_t* dst_yj, + int dst_stride_yj, + int width, + int height) { + int y; + void (*RGBAToYJRow)(const uint8_t* src_rgba, uint8_t* dst_yj, int width) = + RGBAToYJRow_C; + if (!src_rgba || !dst_yj || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_rgba = src_rgba + (height - 1) * src_stride_rgba; + src_stride_rgba = -src_stride_rgba; + } + // Coalesce rows. + if (src_stride_rgba == width * 4 && dst_stride_yj == width) { + width *= height; + height = 1; + src_stride_rgba = dst_stride_yj = 0; + } +#if defined(HAS_RGBATOYJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RGBAToYJRow = RGBAToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RGBAToYJRow = RGBAToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_RGBATOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + RGBAToYJRow = RGBAToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + RGBAToYJRow = RGBAToYJRow_AVX2; + } + } +#endif +#if defined(HAS_RGBATOYJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGBAToYJRow = RGBAToYJRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + RGBAToYJRow = RGBAToYJRow_NEON; + } + } +#endif +#if defined(HAS_RGBATOYJROW_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + RGBAToYJRow = RGBAToYJRow_Any_MSA; + if (IS_ALIGNED(width, 16)) { + RGBAToYJRow = RGBAToYJRow_MSA; } } #endif for (y = 0; y < height; ++y) { - ARGBToYJRow(src_argb, dst_yj, width); - src_argb += src_stride_argb; + RGBAToYJRow(src_rgba, dst_yj, width); + src_rgba += src_stride_rgba; dst_yj += dst_stride_yj; } return 0; } +// Enabled if 1 pass is available +#if defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) +#define HAS_RAWTOYJROW +#endif + +// RAW to JNV21 full range NV21 +LIBYUV_API +int RAWToJNV21(const uint8_t* src_raw, + int src_stride_raw, + uint8_t* dst_y, + int dst_stride_y, + uint8_t* dst_vu, + int dst_stride_vu, + int width, + int height) { + int y; + int halfwidth = (width + 1) >> 1; +#if defined(HAS_RAWTOYJROW) + void (*RAWToUVJRow)(const uint8_t* src_raw, int src_stride_raw, + uint8_t* dst_u, uint8_t* dst_v, int width) = + RAWToUVJRow_C; + void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) = + RAWToYJRow_C; +#else + void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) = + RAWToARGBRow_C; + void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb, + uint8_t* dst_u, uint8_t* dst_v, int width) = + ARGBToUVJRow_C; + void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) = + ARGBToYJRow_C; +#endif + void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v, + uint8_t* dst_vu, int width) = MergeUVRow_C; + if (!src_raw || !dst_y || !dst_vu || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_raw = src_raw + (height - 1) * src_stride_raw; + src_stride_raw = -src_stride_raw; + } + +#if defined(HAS_RAWTOYJROW) + +// Neon version does direct RAW to YUV. +#if defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RAWToUVJRow = RAWToUVJRow_Any_NEON; + RAWToYJRow = RAWToYJRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + RAWToYJRow = RAWToYJRow_NEON; + RAWToUVJRow = RAWToUVJRow_NEON; + } + } +#endif +#if defined(HAS_RAWTOYJROW_MSA) && defined(HAS_RAWTOUVJROW_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + RAWToUVJRow = RAWToUVJRow_Any_MSA; + RAWToYJRow = RAWToYJRow_Any_MSA; + if (IS_ALIGNED(width, 16)) { + RAWToYJRow = RAWToYJRow_MSA; + RAWToUVJRow = RAWToUVJRow_MSA; + } + } +#endif + +// Other platforms do intermediate conversion from RAW to ARGB. +#else // HAS_RAWTOYJROW + +#if defined(HAS_RAWTOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RAWToARGBRow = RAWToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RAWToARGBRow = RAWToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYJRow = ARGBToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYJRow = ARGBToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYJRow = ARGBToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOUVJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJRow = ARGBToUVJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUVJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVJRow = ARGBToUVJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVJRow = ARGBToUVJRow_AVX2; + } + } +#endif +#endif // HAS_RAWTOYJROW +#if defined(HAS_MERGEUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MergeUVRow_ = MergeUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_SSE2; + } + } +#endif +#if defined(HAS_MERGEUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MergeUVRow_ = MergeUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + MergeUVRow_ = MergeUVRow_AVX2; + } + } +#endif +#if defined(HAS_MERGEUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MergeUVRow_ = MergeUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_NEON; + } + } +#endif +#if defined(HAS_MERGEUVROW_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + MergeUVRow_ = MergeUVRow_Any_MSA; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_MSA; + } + } +#endif +#if defined(HAS_MERGEUVROW_LSX) + if (TestCpuFlag(kCpuHasLSX)) { + MergeUVRow_ = MergeUVRow_Any_LSX; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_LSX; + } + } +#endif + { + // Allocate a row of uv. + align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2); + uint8_t* row_v = row_u + ((halfwidth + 31) & ~31); +#if !defined(HAS_RAWTOYJROW) + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_RAWTOYJROW) + RAWToUVJRow(src_raw, src_stride_raw, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); + RAWToYJRow(src_raw, dst_y, width); + RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width); +#else + RAWToARGBRow(src_raw, row, width); + RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width); + ARGBToUVJRow(row, kRowSize, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); + ARGBToYJRow(row, dst_y, width); + ARGBToYJRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_raw += src_stride_raw * 2; + dst_y += dst_stride_y * 2; + dst_vu += dst_stride_vu; + } + if (height & 1) { +#if defined(HAS_RAWTOYJROW) + RAWToUVJRow(src_raw, 0, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); + RAWToYJRow(src_raw, dst_y, width); +#else + RAWToARGBRow(src_raw, row, width); + ARGBToUVJRow(row, 0, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_vu, halfwidth); + ARGBToYJRow(row, dst_y, width); +#endif + } +#if !defined(HAS_RAWTOYJROW) + free_aligned_buffer_64(row); +#endif + free_aligned_buffer_64(row_u); + } + return 0; +} +#undef HAS_RAWTOYJROW + #ifdef __cplusplus } // extern "C" } // namespace libyuv |