diff --git a/README.chromium b/README.chromium index 9642ba633..8b6708f23 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: https://chromium.googlesource.com/libyuv/libyuv/ -Version: 1957 +Version: 1958 Revision: DEPS License: BSD-3-Clause License File: LICENSE diff --git a/include/libyuv/row.h b/include/libyuv/row.h index 25c146d63..22b53a219 100644 --- a/include/libyuv/row.h +++ b/include/libyuv/row.h @@ -393,6 +393,7 @@ extern "C" { #if !defined(LIBYUV_DISABLE_X86) && \ (defined(__x86_64__) || defined(_M_X64)) && defined(CLANG_HAS_AVX512) #define HAS_I422TOARGBROW_AVX512BW +#define HAS_I422TOAR30ROW_AVX512BW #define HAS_ARGBSHUFFLEROW_AVX512BW #define HAS_ARGBTOUV444MATRIXROW_AVX512BW #define HAS_RGBTOUV444MATRIXROW_AVX512BW @@ -4465,6 +4466,12 @@ void I422ToARGBRow_AVX512BW(const uint8_t* y_buf, uint8_t* dst_argb, const struct YuvConstants* yuvconstants, int width); +void I422ToAR30Row_AVX512BW(const uint8_t* y_buf, + const uint8_t* u_buf, + const uint8_t* v_buf, + uint8_t* dst_ar30, + const struct YuvConstants* yuvconstants, + int width); void I422ToRGBARow_AVX2(const uint8_t* y_buf, const uint8_t* u_buf, const uint8_t* v_buf, @@ -4793,6 +4800,12 @@ void I422ToARGBRow_Any_AVX512BW(const uint8_t* y_buf, uint8_t* dst_ptr, const struct YuvConstants* yuvconstants, int width); +void I422ToAR30Row_Any_AVX512BW(const uint8_t* y_buf, + const uint8_t* u_buf, + const uint8_t* v_buf, + uint8_t* dst_ptr, + const struct YuvConstants* yuvconstants, + int width); void I422ToRGBARow_Any_AVX2(const uint8_t* y_buf, const uint8_t* u_buf, const uint8_t* v_buf, diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 91ef0c164..cd1d85c71 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1957 +#define LIBYUV_VERSION 1958 #endif // INCLUDE_LIBYUV_VERSION_H_ diff --git a/source/convert_argb.cc b/source/convert_argb.cc index 98c7effb9..f20083257 100644 --- a/source/convert_argb.cc +++ b/source/convert_argb.cc @@ -6512,6 +6512,15 @@ int I420ToAR30Matrix(const uint8_t* src_y, } } #endif +#if defined(HAS_I422TOAR30ROW_AVX512BW) + if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) == + (kCpuHasAVX512BW | kCpuHasAVX512VL)) { + I422ToAR30Row = I422ToAR30Row_Any_AVX512BW; + if (IS_ALIGNED(width, 32)) { + I422ToAR30Row = I422ToAR30Row_AVX512BW; + } + } +#endif #if defined(HAS_I422TOAR30ROW_NEON) if (TestCpuFlag(kCpuHasNEON)) { I422ToAR30Row = I422ToAR30Row_Any_NEON; diff --git a/source/row_any.cc b/source/row_any.cc index b09e8137e..38784e144 100644 --- a/source/row_any.cc +++ b/source/row_any.cc @@ -398,6 +398,9 @@ ANY31C(I422ToAR30Row_Any_SSSE3, I422ToAR30Row_SSSE3, 1, 0, 4, 7) #ifdef HAS_I422TOAR30ROW_AVX2 ANY31C(I422ToAR30Row_Any_AVX2, I422ToAR30Row_AVX2, 1, 0, 4, 15) #endif +#ifdef HAS_I422TOAR30ROW_AVX512BW +ANY31C(I422ToAR30Row_Any_AVX512BW, I422ToAR30Row_AVX512BW, 1, 0, 4, 31) +#endif #ifdef HAS_I444TOARGBROW_SSSE3 ANY31C(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7) #endif diff --git a/source/row_gcc.cc b/source/row_gcc.cc index 46bdd02e8..0cb829584 100644 --- a/source/row_gcc.cc +++ b/source/row_gcc.cc @@ -3436,6 +3436,33 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf, "vmovups %%zmm0,0x40(%[dst_argb]) \n" \ "lea 0x80(%[dst_argb]), %[dst_argb] \n" +// Store 32 AR30 values. +#define STOREAR30_AVX512BW \ + "vpsraw $0x4,%%zmm0,%%zmm0 \n" \ + "vpsraw $0x4,%%zmm1,%%zmm1 \n" \ + "vpsraw $0x4,%%zmm2,%%zmm2 \n" \ + "vpminsw %%zmm7,%%zmm0,%%zmm0 \n" \ + "vpminsw %%zmm7,%%zmm1,%%zmm1 \n" \ + "vpminsw %%zmm7,%%zmm2,%%zmm2 \n" \ + "vpmaxsw %%zmm6,%%zmm0,%%zmm0 \n" \ + "vpmaxsw %%zmm6,%%zmm1,%%zmm1 \n" \ + "vpmaxsw %%zmm6,%%zmm2,%%zmm2 \n" \ + "vpsllw $0x4,%%zmm2,%%zmm2 \n" \ + "vpermq %%zmm0,%%zmm18,%%zmm0 \n" \ + "vpermq %%zmm1,%%zmm18,%%zmm1 \n" \ + "vpermq %%zmm2,%%zmm18,%%zmm2 \n" \ + "vpunpckhwd %%zmm2,%%zmm0,%%zmm3 \n" \ + "vpunpcklwd %%zmm2,%%zmm0,%%zmm0 \n" \ + "vpunpckhwd %%zmm5,%%zmm1,%%zmm2 \n" \ + "vpunpcklwd %%zmm5,%%zmm1,%%zmm1 \n" \ + "vpslld $0xa,%%zmm1,%%zmm1 \n" \ + "vpslld $0xa,%%zmm2,%%zmm2 \n" \ + "vpord %%zmm1,%%zmm0,%%zmm0 \n" \ + "vpord %%zmm2,%%zmm3,%%zmm3 \n" \ + "vmovdqu32 %%zmm0,(%[dst_ar30]) \n" \ + "vmovdqu32 %%zmm3,0x40(%[dst_ar30]) \n" \ + "lea 0x80(%[dst_ar30]), %[dst_ar30] \n" + // Store 16 AR30 values. #define STOREAR30_AVX2 \ "vpsraw $0x4,%%ymm0,%%ymm0 \n" \ @@ -3705,6 +3732,51 @@ void OMITFP I422ToRGB24Row_AVX512VBMI(const uint8_t* y_buf, } #endif // HAS_I422TORGB24ROW_AVX512VBMI +#if defined(HAS_I422TOAR30ROW_AVX512BW) +// 32 pixels +// 16 UV values upsampled to 32 UV, mixed with 32 Y producing 32 AR30 (128 +// bytes). +void OMITFP I422ToAR30Row_AVX512BW(const uint8_t* y_buf, + const uint8_t* u_buf, + const uint8_t* v_buf, + uint8_t* dst_ar30, + const struct YuvConstants* yuvconstants, + int width) { + asm volatile ( + YUVTORGB_SETUP_AVX512BW(yuvconstants) + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants + "vpsrlw $14,%%ymm5,%%ymm5 \n" + "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits + "vpbroadcastq %%xmm5,%%zmm5 \n" + "vpxord %%zmm6,%%zmm6,%%zmm6 \n" // 0 for min + "vpternlogd $0xff,%%zmm7,%%zmm7,%%zmm7 \n" // 1023 for max + "vpsrlw $6,%%zmm7,%%zmm7 \n" + + LABELALIGN + "1: \n" + READYUV422_AVX512BW + YUVTORGB16_AVX512BW(yuvconstants) + STOREAR30_AVX512BW + "sub $0x20,%[width] \n" + "jg 1b \n" + + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] + [width]"+rm"(width) // %[width] + : [yuvconstants]"r"(yuvconstants), // %[yuvconstants] + [quadsplitperm]"r"(kSplitQuadWords), // %[quadsplitperm] + [dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm] + [unperm]"r"(kUnpermuteAVX512) // %[unperm] + : "memory", "cc", YUVTORGB_REGS_AVX512BW + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_I422TOAR30ROW_AVX512BW + #endif // HAS_I422TOARGBROW_AVX512BW #if defined(HAS_I422TOAR30ROW_AVX2)