J400ToARGB switch from SSE2 to AVX2

- port for row_win - remove unused HAS_ macros Was C/SSE2 MSVC J400ToARGB_Opt (1967 ms) Clang J400ToARGB_Opt (568 ms) Now AVX2 MSVC J400ToARGB_Opt (411 ms) Clang J400ToARGB_Opt (418 ms) Test: libyuv_unittest --gtest_filter=*J400ToARGB* Bug: libyuv:508639302 Change-Id: Ifdfb026832b708b61f55477250cc5ee52449f421 TAG=agy CONV=186608fc-966a-4ea7-bf57-9fe07cc1383c Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/7877368 Commit-Queue: Frank Barchard <fbarchard@google.com> Reviewed-by: Justin Green <greenjustin@google.com>
2026-06-15 00:16:08 +08:00 · 2026-05-28 21:06:47 -07:00 · 2026-05-28 21:06:47 -07:00 · e449eb2172
commit e449eb2172
parent 904f562d86
7 changed files with 115 additions and 56 deletions
--- a/README.chromium
+++ b/README.chromium
@ -1,6 +1,6 @@
 Name: libyuv
 URL: https://chromium.googlesource.com/libyuv/libyuv/
-Version: 1941
+Version: 1942
 Revision: DEPS
 License: BSD-3-Clause
 License File: LICENSE
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@ -49,7 +49,6 @@ extern "C" {
 #define HAS_ARGBTORGB565ROW_SSE2
 #define HAS_COPYROW_ERMS
 #define HAS_COPYROW_SSE2
-#define HAS_H422TOARGBROW_SSSE3
 #define HAS_I422TOARGB1555ROW_SSSE3
 #define HAS_I422TOARGB4444ROW_SSSE3
 #define HAS_I422TOARGBROW_SSSE3
@ -60,11 +59,8 @@ extern "C" {
 #define HAS_I422TOYUY2ROW_SSE2
 #define HAS_I444TOARGBROW_SSSE3
 #define HAS_I444TORGB24ROW_SSSE3
-#define HAS_J400TOARGBROW_SSE2
-#define HAS_J422TOARGBROW_SSSE3
 #define HAS_MERGEUVROW_SSE2
 #define HAS_MIRRORROW_SSSE3
-#define HAS_MIRRORSPLITUVROW_SSSE3
 #define HAS_NV12TOARGBROW_SSSE3
 #define HAS_NV12TORGB24ROW_SSSE3
 #define HAS_NV12TORGB565ROW_SSSE3
@ -78,8 +74,6 @@ extern "C" {
 #define HAS_SETROW_X86
 #define HAS_SPLITUVROW_SSE2
 #define HAS_UYVYTOARGBROW_SSSE3
-#define HAS_UYVYTOUV422ROW_SSE2
-#define HAS_UYVYTOUVROW_SSE2
 #define HAS_UYVYTOYROW_SSE2
 #define HAS_YUY2TOARGBROW_SSSE3
 #define HAS_YUY2TOUV422ROW_SSE2
@ -163,7 +157,6 @@ extern "C" {
 #define HAS_ARGBSHUFFLEROW_AVX2
 #define HAS_ARGBTORGB565DITHERROW_AVX2
 #define HAS_COPYROW_AVX
-#define HAS_H422TOARGBROW_AVX2
 #define HAS_HALFFLOATROW_AVX2
 #define HAS_I422TOARGB1555ROW_AVX2
 #define HAS_I422TOARGB4444ROW_AVX2
@ -173,7 +166,7 @@ extern "C" {
 #define HAS_I422TORGBAROW_AVX2
 #define HAS_I444TOARGBROW_AVX2
 #define HAS_I444TORGB24ROW_AVX2
-#define HAS_J422TOARGBROW_AVX2
+#define HAS_J400TOARGBROW_AVX2
 #define HAS_MIRRORROW_AVX2
 #define HAS_MIRRORSPLITUVROW_AVX2
 #define HAS_NV12TOARGBROW_AVX2
@ -183,8 +176,6 @@ extern "C" {
 #define HAS_NV21TORGB24ROW_AVX2
 #define HAS_SPLITUVROW_AVX2
 #define HAS_UYVYTOARGBROW_AVX2
-#define HAS_UYVYTOUV422ROW_AVX2
-#define HAS_UYVYTOUVROW_AVX2
 #define HAS_UYVYTOYROW_AVX2
 #define HAS_YUY2TOARGBROW_AVX2
 #define HAS_YUY2TOUV422ROW_AVX2
@ -256,16 +247,11 @@ extern "C" {
 #define HAS_ARGBTOYROW_SSSE3
 #define HAS_ARGBTOYMATRIXROW_SSSE3
 #define HAS_BGRATOYROW_SSSE3
-#define HAS_ABGRTOYROW_SSSE3
-#define HAS_RGBATOYROW_SSSE3

 // TODO: adjust row_win to use 8 bit negative coefficients.
 #define HAS_ABGRTOUVJROW_SSSE3
 #define HAS_ARGBTOUVJROW_SSSE3
-#define HAS_ABGRTOUVROW_SSSE3
 #define HAS_ARGBTOUVROW_SSSE3
-#define HAS_BGRATOUVROW_SSSE3
-#define HAS_RGBATOUVROW_SSSE3
 #define HAS_ARGBTOUVMATRIXROW_SSSE3
 #define HAS_ARGBTOUV444MATRIXROW_SSSE3

@ -372,8 +358,7 @@ extern "C" {
 #define HAS_ARGBTOYJROW_AVX2
 #define HAS_ABGRTOYJROW_AVX2
 #define HAS_RGBATOYJROW_AVX2
-#define HAS_RGBATOYROW_AVX2
-#define HAS_BGRATOYROW_AVX2
+#define HAS_J400TOARGBROW_AVX2
 #endif

 // The following are available for AVX512 clang x86 platforms:
@ -397,8 +382,6 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && defined(__x86_64__) && \
    (defined(CLANG_HAS_AVX512))
 #define HAS_I422TOARGBROW_AVX512BW
-#define HAS_I444TOARGBROW_AVX512BW
-#define HAS_I444TORGB24ROW_AVX512BW
 #define HAS_ARGBTOUV444ROW_AVX512BW
 #define HAS_ARGBTOUV444MATRIXROW_AVX512BW
 #define HAS_ARGBTOYROW_AVX512BW
@ -407,6 +390,7 @@ extern "C" {
 #define HAS_ARGBTOUVROW_AVX512BW
 #define HAS_ARGBTOUVJROW_AVX512BW
 #define HAS_ARGBTOUVMATRIXROW_AVX512BW
+#define HAS_J400TOARGBROW_AVX512BW
 #endif

 // The following are available on Neon platforms:
@ -449,7 +433,6 @@ extern "C" {
 #define HAS_ARGBTOYMATRIXROW_NEON
 #endif
 #define HAS_ARGBTOYROW_NEON
-#define HAS_AYUVTOUVROW_NEON
 #define HAS_AYUVTOVUROW_NEON
 #define HAS_AYUVTOYROW_NEON
 #define HAS_BGRATOUVROW_NEON
@ -523,7 +506,6 @@ extern "C" {
 #define HAS_SWAPUVROW_NEON
 #define HAS_UNPACKMT2T_NEON
 #define HAS_UYVYTOARGBROW_NEON
-#define HAS_UYVYTOUV422ROW_NEON
 #define HAS_UYVYTOUVROW_NEON
 #define HAS_UYVYTOYROW_NEON
 #define HAS_YUY2TOARGBROW_NEON
@ -4386,15 +4368,15 @@ void AB64ToARGBRow_Any_NEON(const uint16_t* src_ptr,
                            uint8_t* dst_ptr,
                            int width);

-void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_AVX512BW(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_RVV(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width);
-void J400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
-                            uint8_t* dst_ptr,
-                            int width);
+void J400ToARGBRow_Any_AVX512BW(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
 void J400ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
                            uint8_t* dst_ptr,
                            int width);
--- 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 1941
+#define LIBYUV_VERSION 1942

 #endif  // INCLUDE_LIBYUV_VERSION_H_
--- a/source/convert_argb.cc
+++ b/source/convert_argb.cc
@ -3280,14 +3280,7 @@ int J400ToARGB(const uint8_t* src_y,
    height = 1;
    src_stride_y = dst_stride_argb = 0;
  }
-#if defined(HAS_J400TOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    J400ToARGBRow = J400ToARGBRow_Any_SSE2;
-    if (IS_ALIGNED(width, 8)) {
-      J400ToARGBRow = J400ToARGBRow_SSE2;
-    }
-  }
-#endif
+
 #if defined(HAS_J400TOARGBROW_AVX2)
  if (TestCpuFlag(kCpuHasAVX2)) {
    J400ToARGBRow = J400ToARGBRow_Any_AVX2;
@ -3296,6 +3289,14 @@ int J400ToARGB(const uint8_t* src_y,
    }
  }
 #endif
+#if defined(HAS_J400TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW)) {
+    J400ToARGBRow = J400ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      J400ToARGBRow = J400ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_J400TOARGBROW_NEON)
  if (TestCpuFlag(kCpuHasNEON)) {
    J400ToARGBRow = J400ToARGBRow_Any_NEON;
--- a/source/row_any.cc
+++ b/source/row_any.cc
@ -983,8 +983,9 @@ ANY11(ABGRToAR30Row_Any_AVX2, ABGRToAR30Row_AVX2, 0, 4, 4, 7)
 #if defined(HAS_ARGBTOAR30ROW_AVX2)
 ANY11(ARGBToAR30Row_Any_AVX2, ARGBToAR30Row_AVX2, 0, 4, 4, 7)
 #endif
-#if defined(HAS_J400TOARGBROW_SSE2)
-ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7)
+
+#if defined(HAS_J400TOARGBROW_AVX512BW)
+ANY11(J400ToARGBRow_Any_AVX512BW, J400ToARGBRow_AVX512BW, 0, 1, 4, 31)
 #endif
 #if defined(HAS_J400TOARGBROW_AVX2)
 ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15)
--- a/source/row_gcc.cc
+++ b/source/row_gcc.cc
@ -117,34 +117,74 @@ static const lvec8 kShuffleNV21 = {
 };
 #endif  // HAS_RGB24TOARGBROW_SSSE3

-#ifdef HAS_J400TOARGBROW_SSE2
-void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+#if defined(HAS_J400TOARGBROW_AVX2) || defined(HAS_J400TOARGBROW_AVX512BW)
+alignas(64) static const uint8_t kShuffleMaskJ400ToARGB[64] = {
+    0u, 0u, 0u, 128u, 1u, 1u, 1u, 128u, 2u, 2u, 2u, 128u, 3u, 3u, 3u, 128u,
+    4u, 4u, 4u, 128u, 5u, 5u, 5u, 128u, 6u, 6u, 6u, 128u, 7u, 7u, 7u, 128u,
+    8u, 8u, 8u, 128u, 9u, 9u, 9u, 128u, 10u, 10u, 10u, 128u, 11u, 11u, 11u, 128u,
+    12u, 12u, 12u, 128u, 13u, 13u, 13u, 128u, 14u, 14u, 14u, 128u, 15u, 15u, 15u, 128u
+};
+#endif
+
+#ifdef HAS_J400TOARGBROW_AVX2
+void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width) {
  asm volatile(
-      "pcmpeqb     %%xmm5,%%xmm5                 \n"
-      "pslld       $0x18,%%xmm5                  \n"
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"
+      "vpslld      $0x18,%%ymm7,%%ymm7           \n"
+      "vmovdqa     (%3),%%ymm5                     \n"
+      "vmovdqa     0x20(%3),%%ymm6                  \n"

      LABELALIGN
      "1:          \n"
-      "movq        (%0),%%xmm0                   \n"
-      "lea         0x8(%0),%0                    \n"
-      "punpcklbw   %%xmm0,%%xmm0                 \n"
-      "movdqa      %%xmm0,%%xmm1                 \n"
-      "punpcklwd   %%xmm0,%%xmm0                 \n"
-      "punpckhwd   %%xmm1,%%xmm1                 \n"
-      "por         %%xmm5,%%xmm0                 \n"
-      "por         %%xmm5,%%xmm1                 \n"
-      "movdqu      %%xmm0,(%1)                   \n"
-      "movdqu      %%xmm1,0x10(%1)               \n"
-      "lea         0x20(%1),%1                   \n"
-      "sub         $0x8,%2                       \n"
+      "vbroadcasti128 (%0),%%ymm0                \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm0,%%ymm2          \n"
+      "vpor        %%ymm7,%%ymm1,%%ymm1          \n"
+      "vpor        %%ymm7,%%ymm2,%%ymm2          \n"
+      "vmovdqu     %%ymm1,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x20(%1)               \n"
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
      "jg          1b                            \n"
+      "vzeroupper  \n"
      : "+r"(src_y),     // %0
        "+r"(dst_argb),  // %1
        "+r"(width)      // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1", "xmm5");
+      : "r"(kShuffleMaskJ400ToARGB) // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7");
 }
-#endif  // HAS_J400TOARGBROW_SSE2
+#endif  // HAS_J400TOARGBROW_AVX2
+
+#ifdef HAS_J400TOARGBROW_AVX512BW
+void J400ToARGBRow_AVX512BW(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "vpternlogd  $0xff,%%zmm7,%%zmm7,%%zmm7    \n"  // 0xffffffff
+      "vpslld      $0x18,%%zmm7,%%zmm7           \n"  // 0xff000000
+      "vmovdqa64   %3,%%zmm5                     \n"
+
+      LABELALIGN
+      "1:          \n"
+      "vbroadcasti32x4 (%0),%%zmm0               \n"
+      "vbroadcasti32x4 0x10(%0),%%zmm1          \n"
+      "vpshufb     %%zmm5,%%zmm0,%%zmm0          \n"
+      "vpshufb     %%zmm5,%%zmm1,%%zmm1          \n"
+      "vpord       %%zmm7,%%zmm0,%%zmm0          \n"
+      "vpord       %%zmm7,%%zmm1,%%zmm1          \n"
+      "vmovdqu64   %%zmm0,(%1)                   \n"
+      "vmovdqu64   %%zmm1,0x40(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x80(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper  \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "m"(kShuffleMaskJ400ToARGB) // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5", "xmm7");
+}
+#endif  // HAS_J400TOARGBROW_AVX512BW

 #ifdef HAS_RGB24TOARGBROW_SSSE3
 void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
--- a/source/row_win.cc
+++ b/source/row_win.cc
@ -669,6 +669,41 @@ void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 }
 #endif

+#ifdef HAS_J400TOARGBROW_AVX2
+alignas(32) static const uint8_t kShuffleMaskJ400ToARGB_0[32] = {
+    0u, 0u, 0u, 128u, 1u, 1u, 1u, 128u, 2u, 2u, 2u, 128u, 3u, 3u, 3u, 128u,
+    4u, 4u, 4u, 128u, 5u, 5u, 5u, 128u, 6u, 6u, 6u, 128u, 7u, 7u, 7u, 128u
+};
+alignas(32) static const uint8_t kShuffleMaskJ400ToARGB_1[32] = {
+    8u, 8u, 8u, 128u, 9u, 9u, 9u, 128u, 10u, 10u, 10u, 128u, 11u, 11u, 11u, 128u,
+    12u, 12u, 12u, 128u, 13u, 13u, 13u, 128u, 14u, 14u, 14u, 128u, 15u, 15u, 15u, 128u
+};
+
+LIBYUV_TARGET_AVX2
+void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  __m256i ymm_mask0 = _mm256_load_si256((const __m256i*)kShuffleMaskJ400ToARGB_0);
+  __m256i ymm_mask1 = _mm256_load_si256((const __m256i*)kShuffleMaskJ400ToARGB_1);
+  __m256i ymm_alpha = _mm256_set1_epi32((int)0xff000000u);
+
+  while (width > 0) {
+    __m256i ymm0 = _mm256_broadcastsi128_si256(_mm_loadu_si128((const __m128i*)src_y));
+
+    __m256i ymm1 = _mm256_shuffle_epi8(ymm0, ymm_mask0);
+    __m256i ymm2 = _mm256_shuffle_epi8(ymm0, ymm_mask1);
+
+    ymm1 = _mm256_or_si256(ymm1, ymm_alpha);
+    ymm2 = _mm256_or_si256(ymm2, ymm_alpha);
+
+    _mm256_storeu_si256((__m256i*)dst_argb, ymm1);
+    _mm256_storeu_si256((__m256i*)(dst_argb + 32), ymm2);
+
+    src_y += 16;
+    dst_argb += 64;
+    width -= 16;
+  }
+}
+#endif  // HAS_J400TOARGBROW_AVX2
+
 #endif

 #ifdef __cplusplus