ARGBToUV SSE use average of 4 pixels

- Was using avgb twice for non-exact and C for exact.

On Skylake Xeon:

Now SSE3
ARGBToJ420_Opt (326 ms)

Was
Exact C
ARGBToJ420_Opt (871 ms)
Not exact AVX2
ARGBToJ420_Opt (237 ms)
Not exact SSSE3
ARGBToJ420_Opt (312 ms)

Bug: 381138208
Change-Id: I6d1081bb52e36f06736c0c6575fa82bb2268629b
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/6629605
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Ben Weiss <bweiss@google.com>

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: https://chromium.googlesource.com/libyuv/libyuv/
-Version: 1912
+Version: 1913
 License: BSD-3-Clause
 License File: LICENSE
 Shipped: yes

include/libyuv/row.h

@@ -79,12 +79,6 @@ extern "C" {
 #define HAS_YUY2TOUV422ROW_SSE2
 #define HAS_YUY2TOUVROW_SSE2
 #define HAS_YUY2TOYROW_SSE2
-#if !defined(LIBYUV_BIT_EXACT)
-#define HAS_ABGRTOUVROW_SSSE3
-#define HAS_ARGBTOUVROW_SSSE3
-#define HAS_BGRATOUVROW_SSSE3
-#define HAS_RGBATOUVROW_SSSE3
-#endif
 
 // Effects:
 #define HAS_ARGBADDROW_SSE2
@@ -272,11 +266,13 @@ extern "C" {
 #define HAS_RGB24TOYROW_SSSE3
 #define HAS_RGBATOYROW_SSSE3
 
-#if !defined(LIBYUV_BIT_EXACT)
 // TODO: adjust row_win to use 8 bit negative coefficients.
 #define HAS_ABGRTOUVJROW_SSSE3
 #define HAS_ARGBTOUVJROW_SSSE3
-#endif
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_RGBATOUVROW_SSSE3
 
 #if defined(__x86_64__) || !defined(__pic__)
 // TODO(fbarchard): fix build error on android_full_debug=1
@@ -350,12 +346,11 @@ extern "C" {
 #define HAS_SPLITXRGBROW_AVX2
 #define HAS_SWAPUVROW_AVX2
 #define HAS_YUY2TONVUVROW_AVX2
-#if !defined(LIBYUV_BIT_EXACT)
+// TODO: Port SSSE3 to AVX2
-#define HAS_ABGRTOUVJROW_AVX2
+// #define HAS_ABGRTOUVJROW_AVX2
-#define HAS_ABGRTOUVROW_AVX2
+// #define HAS_ABGRTOUVROW_AVX2
-#define HAS_ARGBTOUVJROW_AVX2
+// #define HAS_ARGBTOUVJROW_AVX2
-#define HAS_ARGBTOUVROW_AVX2
+// #define HAS_ARGBTOUVROW_AVX2
-#endif
 
 #if defined(__x86_64__) || !defined(__pic__)
 // TODO(fbarchard): fix build error on android_full_debug=1

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1912
+#define LIBYUV_VERSION 1913
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/row_common.cc

@@ -36,10 +36,6 @@ extern "C" {
 // LIBYUV_UNLIMITED_BT709
 // LIBYUV_UNLIMITED_BT2020
 
-#if !defined(LIBYUV_BIT_EXACT) && (defined(__x86_64__) || defined(_M_X64) || \
-                                   defined(__i386__) || defined(_M_IX86))
-#define LIBYUV_ARGBTOUV_PAVGB 1
-#endif
 #if defined(LIBYUV_BIT_EXACT)
 #define LIBYUV_UNATTENUATE_DUP 1
 #endif
@@ -627,7 +623,6 @@ void AR64ShuffleRow_C(const uint8_t* src_ar64,
 // b -0.1406 * 128 = −17.9968 = -18
 // g -0.7344 * 128 = −94.0032 = -94
 // r  0.875  * 128 = 112.0    = 112
-
 static __inline uint8_t RGBToY(uint8_t r, uint8_t g, uint8_t b) {
   return STATIC_CAST(uint8_t, (66 * r + 129 * g + 25 * b + 0x1080) >> 8);
 }
@@ -639,46 +634,6 @@ static __inline uint8_t RGBToV(uint8_t r, uint8_t g, uint8_t b) {
 }
 #define AVGB(a, b) (((a) + (b) + 1) >> 1)
 
-// ARGBToY_C and ARGBToUV_C
-// Intel version of UV mimic SSE/AVX which does 2 pavgb
-#if defined(LIBYUV_ARGBTOUV_PAVGB)
-#define MAKEROWY(NAME, R, G, B, BPP)                                       \
-  void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
-    int x;                                                                 \
-    for (x = 0; x < width; ++x) {                                          \
-      dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]);               \
-      src_rgb += BPP;                                                      \
-      dst_y += 1;                                                          \
-    }                                                                      \
-  }                                                                        \
-  void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
-                       uint8_t* dst_u, uint8_t* dst_v, int width) {        \
-    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                    \
-    int x;                                                                 \
-    for (x = 0; x < width - 1; x += 2) {                                   \
-      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                     \
-                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));        \
-      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                     \
-                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));        \
-      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                     \
-                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));        \
-      dst_u[0] = RGBToU(ar, ag, ab);                                       \
-      dst_v[0] = RGBToV(ar, ag, ab);                                       \
-      src_rgb += BPP * 2;                                                  \
-      src_rgb1 += BPP * 2;                                                 \
-      dst_u += 1;                                                          \
-      dst_v += 1;                                                          \
-    }                                                                      \
-    if (width & 1) {                                                       \
-      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                          \
-      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                          \
-      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                          \
-      dst_u[0] = RGBToU(ar, ag, ab);                                       \
-      dst_v[0] = RGBToV(ar, ag, ab);                                       \
-    }                                                                      \
-  }
-#else
-// ARM version does average of 4 pixels with rounding
 #define MAKEROWY(NAME, R, G, B, BPP)                                       \
   void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
     int x;                                                                 \
@@ -717,7 +672,6 @@ static __inline uint8_t RGBToV(uint8_t r, uint8_t g, uint8_t b) {
       dst_v[0] = RGBToV(ar, ag, ab);                                       \
     }                                                                      \
   }
-#endif
 
 MAKEROWY(ARGB, 2, 1, 0, 4)
 MAKEROWY(BGRA, 1, 2, 3, 4)
@@ -756,45 +710,6 @@ static __inline uint8_t RGBToVJ(uint8_t r, uint8_t g, uint8_t b) {
 }
 
 // ARGBToYJ_C and ARGBToUVJ_C
-// Intel version mimic SSE/AVX which does 2 pavgb
-#if defined(LIBYUV_ARGBTOUV_PAVGB)
-#define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
-  void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
-    int x;                                                                  \
-    for (x = 0; x < width; ++x) {                                           \
-      dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]);               \
-      src_rgb += BPP;                                                       \
-      dst_y += 1;                                                           \
-    }                                                                       \
-  }                                                                         \
-  void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
-                        uint8_t* dst_u, uint8_t* dst_v, int width) {        \
-    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                     \
-    int x;                                                                  \
-    for (x = 0; x < width - 1; x += 2) {                                    \
-      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                      \
-                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));         \
-      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                      \
-                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));         \
-      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                      \
-                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));         \
-      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
-      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
-      src_rgb += BPP * 2;                                                   \
-      src_rgb1 += BPP * 2;                                                  \
-      dst_u += 1;                                                           \
-      dst_v += 1;                                                           \
-    }                                                                       \
-    if (width & 1) {                                                        \
-      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                           \
-      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                           \
-      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                           \
-      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
-      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
-    }                                                                       \
-  }
-#else
-// ARM version does average of 4 pixels with rounding
 #define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
   void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
     int x;                                                                  \
@@ -834,8 +749,6 @@ static __inline uint8_t RGBToVJ(uint8_t r, uint8_t g, uint8_t b) {
     }                                                                       \
   }
 
-#endif
-
 MAKEROWYJ(ARGB, 2, 1, 0, 4)
 MAKEROWYJ(ABGR, 0, 1, 2, 4)
 MAKEROWYJ(RGBA, 3, 2, 1, 4)
@@ -928,19 +841,11 @@ void RGB565ToUVRow_C(const uint8_t* src_rgb565,
     g3 = STATIC_CAST(uint8_t, (g3 << 2) | (g3 >> 4));
     r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2));
 
-#if defined(LIBYUV_ARGBTOUV_PAVGB)
-    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
-    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
-    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-#else
     uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2;
     uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2;
     uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2;
     dst_u[0] = RGBToU(r, g, b);
     dst_v[0] = RGBToV(r, g, b);
-#endif
 
     src_rgb565 += 4;
     next_rgb565 += 4;
@@ -1009,19 +914,11 @@ void ARGB1555ToUVRow_C(const uint8_t* src_argb1555,
     g3 = STATIC_CAST(uint8_t, (g3 << 3) | (g3 >> 2));
     r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2));
 
-#if defined(LIBYUV_ARGBTOUV_PAVGB)
-    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
-    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
-    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-#else
     uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2;
     uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2;
     uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2;
     dst_u[0] = RGBToU(r, g, b);
     dst_v[0] = RGBToV(r, g, b);
-#endif
 
     src_argb1555 += 4;
     next_argb1555 += 4;
@@ -1087,19 +984,11 @@ void ARGB4444ToUVRow_C(const uint8_t* src_argb4444,
     g3 = STATIC_CAST(uint8_t, (g3 << 4) | g3);
     r3 = STATIC_CAST(uint8_t, (r3 << 4) | r3);
 
-#if defined(LIBYUV_ARGBTOUV_PAVGB)
-    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
-    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
-    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
-    dst_u[0] = RGBToU(ar, ag, ab);
-    dst_v[0] = RGBToV(ar, ag, ab);
-#else
     uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2;
     uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2;
     uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2;
     dst_u[0] = RGBToU(r, g, b);
     dst_v[0] = RGBToV(r, g, b);
-#endif
 
     src_argb4444 += 4;
     next_argb4444 += 4;

source/row_gcc.cc

@@ -1650,7 +1650,7 @@ void ARGBToUV444MatrixRow_SSSE3(const uint8_t* src_argb,
 #if defined(__i386__)
         "+m"(width)  // %3
 #else
-        "+rm"(width)           // %3
+        "+rm"(width)  // %3
 #endif
       : "m"(rgbuvconstants->kRGBToU),  // %4
         "m"(rgbuvconstants->kRGBToV),  // %5
@@ -1721,7 +1721,7 @@ void ARGBToUV444MatrixRow_AVX2(const uint8_t* src_argb,
 #if defined(__i386__)
         "+m"(width)  // %3
 #else
-        "+rm"(width)           // %3
+        "+rm"(width)  // %3
 #endif
       : "m"(rgbuvconstants->kRGBToU),  // %4
         "m"(rgbuvconstants->kRGBToV),  // %5
@@ -1733,161 +1733,71 @@ void ARGBToUV444MatrixRow_AVX2(const uint8_t* src_argb,
 #endif  // HAS_ARGBTOUV444ROW_AVX2
 
 #ifdef HAS_ARGBTOUVROW_SSSE3
-
+// 8x2 -> 4x1 ARGB pixels converted to 4 U and 4 V
-void OMITFP
+// ARGBToUV does rounding average of 4 ARGB pixels
-ARGBToUVMatrixRow_SSSE3(const uint8_t* src_argb,
+void ARGBToUVMatrixRow_SSSE3(const uint8_t* src_argb,
-                        int src_stride_argb,
+                             int src_stride_argb,
-                        uint8_t* dst_u,
+                             uint8_t* dst_u,
-                        uint8_t* dst_v,
+                             uint8_t* dst_v,
-                        int width,
+                             int width,
-                        const struct RgbUVConstants* rgbuvconstants) {
+                             const struct RgbUVConstants* rgbuvconstants) {
   asm volatile(
-      "movdqa      %0,%%xmm3                     \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"  // 0x0101
-      "movdqa      %1,%%xmm4                     \n"
+      "pabsb       %%xmm4,%%xmm4                 \n"
-      "movdqa      %2,%%xmm5                     \n"
+      "movdqa      %5,%%xmm6                     \n"  // ARGB to U
-      :
+      "movdqa      %6,%%xmm7                     \n"  // ARGB to V
-      : "m"(rgbuvconstants->kRGBToU),  // %0
-        "m"(rgbuvconstants->kRGBToV),  // %1
-        "m"(kAddUV128)                 // %2
-      : "xmm3", "xmm4", "xmm5");
-
-      asm volatile("sub         %1,%2                         \n"
-
-               LABELALIGN
-      "1:          \n"
-      "movdqu      (%0),%%xmm0                   \n"
-      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
-      "pavgb       %%xmm7,%%xmm0                 \n"
-      "movdqu      0x10(%0),%%xmm1               \n"
-      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
-      "pavgb       %%xmm7,%%xmm1                 \n"
-      "movdqu      0x20(%0),%%xmm2               \n"
-      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
-      "pavgb       %%xmm7,%%xmm2                 \n"
-      "movdqu      0x30(%0),%%xmm6               \n"
-      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
-      "pavgb       %%xmm7,%%xmm6                 \n"
-      "lea         0x40(%0),%0                   \n"
-      "movdqa      %%xmm0,%%xmm7                 \n"
-      "shufps      $0x88,%%xmm1,%%xmm0           \n"
-      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
-      "pavgb       %%xmm7,%%xmm0                 \n"
-      "movdqa      %%xmm2,%%xmm7                 \n"
-      "shufps      $0x88,%%xmm6,%%xmm2           \n"
-      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
-      "pavgb       %%xmm7,%%xmm2                 \n"
-
-      "movdqa      %%xmm0,%%xmm1                 \n"
-      "movdqa      %%xmm2,%%xmm6                 \n"
-      "pmaddubsw   %%xmm3,%%xmm0                 \n"
-      "pmaddubsw   %%xmm3,%%xmm2                 \n"
-      "pmaddubsw   %%xmm4,%%xmm1                 \n"
-      "pmaddubsw   %%xmm4,%%xmm6                 \n"
-      "phaddw      %%xmm2,%%xmm0                 \n"
-      "phaddw      %%xmm6,%%xmm1                 \n"
-      "movdqa      %%xmm5,%%xmm2                 \n"
-      "movdqa      %%xmm5,%%xmm6                 \n"
-      "psubw       %%xmm0,%%xmm2                 \n"
-      "psubw       %%xmm1,%%xmm6                 \n"
-      "psrlw       $0x8,%%xmm2                   \n"
-      "psrlw       $0x8,%%xmm6                   \n"
-      "packuswb    %%xmm6,%%xmm2                 \n"
-      "movlps      %%xmm2,(%1)                   \n"
-      "movhps      %%xmm2,0x00(%1,%2,1)          \n"
-      "lea         0x8(%1),%1                    \n"
-      "subl        $0x10,%3                      \n"
-      "jg          1b                            \n"
-               : "+r"(src_argb),  // %0
-                 "+r"(dst_u),     // %1
-                 "+r"(dst_v),     // %2
-#if defined(__i386__)
-                 "+m"(width)  // %3
-#else
-                 "+rm"(width)  // %3
-#endif
-               : "r"((intptr_t)(src_stride_argb))  // %4
-               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
-                 "xmm6", "xmm7");
-}
-
-#endif  // HAS_ARGBTOUVROW_SSSE3
-
-// vpshufb for vphaddw + vpackuswb packed to shorts.
-// Coefficients expressed as negatives to allow 128
-struct UVMatrixConstants {
-  lvec8 kShufARGBToUV;
-  ulvec8 kAddUV128;
-};
-
-static const UVMatrixConstants kShufARGBToUV_AVX = {
-    0, 1,   8, 9,   2, 3,   10, 11,  4, 5,   12, 13,  6, 7,   14, 15,
-    0, 1,   8, 9,   2, 3,   10, 11,  4, 5,   12, 13,  6, 7,   14, 15,
-    0, 128, 0, 128, 0, 128, 0,  128, 0, 128, 0,  128, 0, 128, 0,  128,
-    0, 128, 0, 128, 0, 128, 0,  128, 0, 128, 0,  128, 0, 128, 0,  128};
-
-void OMITFP
-ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb,
-                       int src_stride_argb,
-                       uint8_t* dst_u,
-                       uint8_t* dst_v,
-                       int width,
-                       const struct RgbUVConstants* rgbuvconstants) {
-  asm volatile(
-      "vbroadcastf128 %0,%%ymm6                  \n"
-      "vbroadcastf128 %1,%%ymm7                  \n"
-      :
-      : "m"(rgbuvconstants->kRGBToU),  // %0
-        "m"(rgbuvconstants->kRGBToV)   // %1
-      :);
-
-  asm volatile(
-#if !defined(__i386__)
-      "vmovdqa     0(%5),%%ymm8                  \n"
-#endif
-      "vmovdqa     32(%5),%%ymm5                 \n"
       "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:          \n"
-      "vmovdqu     (%0),%%ymm0                   \n"
+      "movdqu      (%0),%%xmm0                   \n"  // Read 8 ARGB Pixels
-      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "movdqu      0x10(%0),%%xmm5               \n"
-      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
-      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "shufps      $0x88,%%xmm5,%%xmm0           \n"  // Even pixels
-      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "shufps      $0xdd,%%xmm5,%%xmm1           \n"  // Odd pixels
-      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "movdqa      %%xmm0,%%xmm5                 \n"
-      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // aarrgbb
-      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "punpckhbw   %%xmm5,%%xmm1                 \n"
-      "lea         0x80(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"  // paired add argb
-      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
-      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
-      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
-      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
-      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
-      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
 
-      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm1          \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"  // Read 2nd row
-      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm3          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm5          \n"
-      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm0          \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
-      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm2          \n"
+      "shufps      $0x88,%%xmm5,%%xmm2           \n"  // Even
-      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "shufps      $0xdd,%%xmm5,%%xmm3           \n"  // Odd pixels
-      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
-      "vpsubw      %%ymm0,%%ymm5,%%ymm0          \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"  // aarrgbb
-      "vpsubw      %%ymm1,%%ymm5,%%ymm1          \n"
+      "punpckhbw   %%xmm5,%%xmm3                 \n"
-      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"  // argb
-      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
-      "vpackuswb   %%ymm0,%%ymm1,%%ymm0          \n"
+
-      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "pxor        %%xmm5,%%xmm5                 \n"  // constant 0 for pavgw
-#if defined(__i386__)
+      "paddw       %%xmm2,%%xmm0                 \n"
-      "vpshufb     (%5),%%ymm0,%%ymm0            \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
-#else
+      "psrlw       $1,%%xmm0                     \n"  // round
-      "vpshufb     %%ymm8,%%ymm0,%%ymm0            \n"
+      "psrlw       $1,%%xmm1                     \n"
-#endif
+      "pavgw       %%xmm5,%%xmm0                 \n"
-      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
-      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"  // 4 ARGB pixels
-      "lea         0x10(%1),%1                   \n"
+
-      "subl        $0x20,%3                      \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pmaddubsw   %%xmm6,%%xmm0                 \n"  // u
+      "pmaddubsw   %%xmm7,%%xmm1                 \n"  // v
+      "phaddw      %%xmm1,%%xmm0                 \n"  // uuuuvvvv
+
+      "movdqa      %7,%%xmm2                     \n"  // 0x8000
+      "psubw       %%xmm0,%%xmm2                 \n"  // unsigned 0 to 0xffff
+      "psrlw       $0x8,%%xmm2                   \n"
+      "packuswb    %%xmm2,%%xmm2                 \n"
+      "movd        %%xmm2,(%1)                   \n"  // Write 4 U's
+      "shufps      $0xdd,%%xmm2,%%xmm2           \n"
+      "movd        %%xmm2,0x00(%1,%2,1)          \n"  // Write 4 V's
+
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x4(%1),%1                    \n"
+      "subl        $0x8,%3                       \n"
       "jg          1b                            \n"
-      "vzeroupper  \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1897,10 +1807,14 @@ ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb,
         "+rm"(width)  // %3
 #endif
       : "r"((intptr_t)(src_stride_argb)),  // %4
-        "r"(&kShufARGBToUV_AVX)            // %5
+        "m"(rgbuvconstants->kRGBToU),      // %5
+        "m"(rgbuvconstants->kRGBToV),      // %6
+        "m"(kAddUV128)                     // %7
+
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
         "xmm7");
 }
+#endif  // HAS_ARGBTOUVROW_SSSE3
 
 #ifdef HAS_ARGBTOUV444ROW_SSSE3
 
@@ -1944,14 +1858,6 @@ void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
   ARGBToUVMatrixRow_SSSE3(src_argb, src_stride_argb, dst_u, dst_v, width,
                           &kARGBI601UVConstants);
 }
-void ARGBToUVRow_AVX2(const uint8_t* src_argb,
-                      int src_stride_argb,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width) {
-  ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width,
-                         &kARGBI601UVConstants);
-}
 
 static const struct RgbUVConstants kABGRI601UVConstants = {
     {38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0},
@@ -1966,15 +1872,6 @@ void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
                           &kABGRI601UVConstants);
 }
 
-void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
-                      int src_stride_abgr,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width) {
-  ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width,
-                         &kABGRI601UVConstants);
-}
-
 static const struct RgbUVConstants kBGRAI601UVConstants = {
     {0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112},
     {0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18}};
@@ -2035,28 +1932,10 @@ void ARGBToUVJ444Row_AVX2(const uint8_t* src_argb,
 }
 #endif  // HAS_ARGBTOUVJ444ROW_AVX2
 
-void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
-                       int src_stride_argb,
-                       uint8_t* dst_u,
-                       uint8_t* dst_v,
-                       int width) {
-  ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width,
-                         &kARGBJPEGUVConstants);
-}
-
 static const struct RgbUVConstants kABGRJPEGUVConstants = {
     {43, 85, -128, 0, 43, 85, -128, 0, 43, 85, -128, 0, 43, 85, -128, 0},
     {-128, 107, 21, 0, -128, 107, 21, 0, -128, 107, 21, 0, -128, 107, 21, 0}};
 
-void ABGRToUVJRow_AVX2(const uint8_t* src_abgr,
-                       int src_stride_abgr,
-                       uint8_t* dst_u,
-                       uint8_t* dst_v,
-                       int width) {
-  ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width,
-                         &kABGRJPEGUVConstants);
-}
-
 #ifdef HAS_ARGBTOUVJROW_SSSE3
 void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
                         int src_stride_argb,

source/row_win.cc

@@ -184,7 +184,7 @@ void I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
 // 32 bit
 #else  // defined(_M_X64)
 
-#ifdef HAS_ARGBTOUVROW_SSSE3
+// if HAS_ARGBTOUVROW_SSSE3
 
 // 8 bit fixed point 0.5, for bias of UV.
 static const ulvec8 kBiasUV128 = {
@@ -221,7 +221,7 @@ static const lvec8 kShuffleUYVYUV = {0,  2,  0,  2,  4,  6,  4,  6,  8,  10, 8,
 // JPeg full range.
 static const vec8 kARGBToYJ = {15, 75, 38, 0, 15, 75, 38, 0,
                                15, 75, 38, 0, 15, 75, 38, 0};
-#endif
+// endif
 
 // vpermd for vphaddw + vpackuswb vpermd.
 static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7};
@@ -1208,6 +1208,8 @@ __declspec(naked) void ARGBToYRow_SSSE3(const uint8_t* src_argb,
   }
 }
 
+#ifdef HAS_ARGBTOUVROW_SSSE3
+
 // Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
 // Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
 __declspec(naked) void ARGBToYJRow_SSSE3(const uint8_t* src_argb,
@@ -1244,6 +1246,7 @@ __declspec(naked) void ARGBToYJRow_SSSE3(const uint8_t* src_argb,
     ret
   }
 }
+#endif
 
 #ifdef HAS_ARGBTOYROW_AVX2
 
@@ -1430,6 +1433,8 @@ __declspec(naked) void RGBAToYRow_SSSE3(const uint8_t* src_argb,
   }
 }
 
+#ifdef HAS_ARGBTOUVROW_SSSE3
+
 __declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
@@ -1578,6 +1583,7 @@ __declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
     ret
   }
 }
+#endif
 
 #ifdef HAS_ARGBTOUVROW_AVX2
 __declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb,

unit_test/convert_argb_test.cc

@@ -2721,6 +2721,86 @@ TEST_F(LibYUVConvertTest, TestUYVYToARGB) {
   EXPECT_EQ(3486643515u, checksum);
 }
 
+#ifdef ENABLE_ROW_TESTS
+TEST_F(LibYUVConvertTest, TestARGBToUVRow) {
+  SIMD_ALIGNED(uint8_t orig_argb_pixels[256]);
+  SIMD_ALIGNED(uint8_t dest_u[32]);
+  SIMD_ALIGNED(uint8_t dest_v[32]);
+
+  for (int i = 0; i < 256; ++i) {
+    orig_argb_pixels[i] = i * 43;
+  }
+
+  orig_argb_pixels[0] = 0xff;  // blue
+  orig_argb_pixels[1] = 0x0;
+  orig_argb_pixels[2] = 0x0;
+  orig_argb_pixels[3] = 0xff;
+  orig_argb_pixels[4] = 0xff;  // blue
+  orig_argb_pixels[5] = 0x0;
+  orig_argb_pixels[6] = 0x0;
+  orig_argb_pixels[7] = 0xff;
+
+  orig_argb_pixels[8] = 0x0;
+  orig_argb_pixels[9] = 0xff;  // green
+  orig_argb_pixels[10] = 0x0;
+  orig_argb_pixels[11] = 0xff;
+  orig_argb_pixels[12] = 0x0;
+  orig_argb_pixels[13] = 0xff;  // green
+  orig_argb_pixels[14] = 0x0;
+  orig_argb_pixels[15] = 0xff;
+
+  orig_argb_pixels[16] = 0x0;
+  orig_argb_pixels[17] = 0x0;
+  orig_argb_pixels[18] = 0xff;  // red
+  orig_argb_pixels[19] = 0xff;
+  orig_argb_pixels[20] = 0x0;
+  orig_argb_pixels[21] = 0x0;
+  orig_argb_pixels[22] = 0xff;  // red
+  orig_argb_pixels[23] = 0xff;
+
+  orig_argb_pixels[24] = 0xff;
+  orig_argb_pixels[25] = 0xff;
+  orig_argb_pixels[26] = 0xff;  // white
+  orig_argb_pixels[27] = 0xff;
+  orig_argb_pixels[28] = 0xff;
+  orig_argb_pixels[29] = 0xff;
+  orig_argb_pixels[30] = 0xff;  // white
+  orig_argb_pixels[31] = 0xff;
+
+  int benchmark_iterations =
+      benchmark_width_ * benchmark_height_ * benchmark_iterations_ / 32;
+
+  for (int i = 0; i < benchmark_iterations; ++i) {
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+    int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+    if (has_ssse3) {
+      ARGBToUVRow_SSSE3(&orig_argb_pixels[0], 0, &dest_u[0], &dest_v[0], 64);
+    } else {
+      ARGBToUVRow_C(&orig_argb_pixels[0], 0, &dest_u[0], &dest_v[0], 64);
+    }
+#elif defined(HAS_ARGBTOUVROW_NEON)
+    ARGBToUVRow_NEON(&orig_argb_pixels[0], 0, &dest_u[0], &dest_v[0], 64);
+#else
+    ARGBToUVRow_C(&orig_argb_pixels[0], 0, &dest_u[0], &dest_v[0], 64);
+#endif
+  }
+  printf("u: ");
+  for (int i = 0; i < 32; ++i) {
+    printf("%3d ", (int)dest_u[i]);
+  }
+  printf("\nv: ");
+  for (int i = 0; i < 32; ++i) {
+    printf("%3d ", (int)dest_v[i]);
+  }
+  printf("\n");
+
+  uint32_t checksum_u = HashDjb2(&dest_u[0], sizeof(dest_u), 5381);
+  EXPECT_EQ(192508756u, checksum_u);
+  uint32_t checksum_v = HashDjb2(&dest_v[0], sizeof(dest_v), 5381);
+  EXPECT_EQ(2590663990u, checksum_v);
+}
+#endif
+
 #if defined(__x86_64__) || defined(_M_X64) || defined(__aarch64__)
 TEST_F(LibYUVConvertTest, TestI400LargeSize) {
   // The width and height are chosen as follows: