RAWToJ400 and RGBToJ400 use 2 step row function for Intel. RAWToJ400 Was 3996…

RAWToJ400 and RGBToJ400 use 2 step row function for Intel. RAWToJ400 Was 3996 ms, now 3309.  20.7% faster.

Call a row function for each row, based on ARGBToI400 code.
But implement row functions as 2 step conversion.  Adds the
row functions:
RAWToYJ, RGBToYJ, SSSE3 and AVX2 versions, and Any versions.
The smaller row buffer is more cache friendly on large images.

The max cache size can be configured, and is currently:
// Maximum temporary width for wrappers to process at a time, in pixels.
And the row buffer is
  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
So 8192 bytes are used for the row buffer, leaving the rest for source
and destination buffers.

blaze-bin/third_party/libyuv/libyuv_test '--gunit_filter=*R*To?400_Opt' --libyuv_width=3600 --libyuv_height=2500 --libyuv_repeat=1000 --libyuv_flags=-1 --libyuv_cpu_info=-1 | sortms

Was
RAWToJ400_Opt (3996 ms)
ARGBToI400_Opt (3964 ms)
RGB24ToJ400_Opt (3960 ms)
ARGBToJ400_Opt (3909 ms)
RGBAToJ400_Opt (3885 ms)

Now
ARGBToJ400_Opt (4091 ms)
ARGBToI400_Opt (3936 ms)
RGBAToJ400_Opt (3428 ms)
RGB24ToJ400_Opt (3324 ms)
RAWToJ400_Opt (3309 ms)

Bug: libyuv:854, b/147753855
Change-Id: Ieb65fbda94e812c737f4c3c74107354b73c4bcd2
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/2016203Reviewed-by: richard winterton <rrwinterton@gmail.com>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1743
+Version: 1744
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -123,6 +123,8 @@ extern "C" {
 #define HAS_RAWTOYROW_SSSE3
 #define HAS_RGB24TOARGBROW_SSSE3
 #define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB24TOYJROW_SSSE3
+#define HAS_RAWTOYJROW_SSSE3
 #define HAS_RGB565TOARGBROW_SSE2
 #define HAS_RGBATOUVROW_SSSE3
 #define HAS_RGBATOYROW_SSSE3
@@ -194,6 +196,8 @@ extern "C" {
 #define HAS_ARGBTOUVROW_AVX2
 #define HAS_ARGBTOYJROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
+#define HAS_RGB24TOYJROW_AVX2
+#define HAS_RAWTOYJROW_AVX2
 #define HAS_COPYROW_AVX
 #define HAS_H422TOARGBROW_AVX2
 #define HAS_HALFFLOATROW_AVX2
@@ -973,7 +977,11 @@ void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width);
 void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width);
 void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width);
 void RGB24ToYRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
 void RAWToYRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_y, int width);
 void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width);
@@ -1187,8 +1195,12 @@ void RGBAToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void BGRAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ABGRToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGBAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RGB24ToYRow_Any_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
-void RAWToYRow_Any_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RGB24ToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGBAToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1743
+#define LIBYUV_VERSION 1744
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/convert.cc

@@ -2380,27 +2380,38 @@ int RGB24ToJ400(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
   void (*RGB24ToYJRow)(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) =
       RGB24ToYJRow_C;
-#else
-  void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
-      RGB24ToARGBRow_C;
-  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
-      ARGBToYJRow_C;
-#endif
   if (!src_rgb24 || !dst_yj || width <= 0 || height == 0) {
     return -1;
   }
-  // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
     src_stride_rgb24 = -src_stride_rgb24;
   }
-
-// Neon version does direct RGB24 to YJ.
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_yj = 0;
+  }
+#if defined(HAS_RGB24TOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYJRow = RGB24ToYJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_RGB24TOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
@@ -2408,83 +2419,28 @@ int RGB24ToJ400(const uint8_t* src_rgb24,
       RGB24ToYJRow = RGB24ToYJRow_NEON;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MSA)
+#endif
+#if defined(HAS_RGB24TOYJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGB24ToYJRow = RGB24ToYJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       RGB24ToYJRow = RGB24ToYJRow_MSA;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MMI)
+#endif
+#if defined(HAS_RGB24TOYJROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RGB24ToYJRow = RGB24ToYJRow_Any_MMI;
     if (IS_ALIGNED(width, 8)) {
       RGB24ToYJRow = RGB24ToYJRow_MMI;
     }
   }
-// Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
-#if defined(HAS_RGB24TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      RGB24ToARGBRow = RGB24ToARGBRow_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
-#endif
-
-  {
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
-    // 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_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
-      RGB24ToYJRow(src_rgb24, dst_yj, width);
-      RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_yj + dst_stride_yj, width);
-#else
-      RGB24ToARGBRow(src_rgb24, row, width);
-      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
-      ARGBToYJRow(row, dst_yj, width);
-      ARGBToYJRow(row + kRowSize, dst_yj + dst_stride_yj, width);
-#endif
-      src_rgb24 += src_stride_rgb24 * 2;
-      dst_yj += dst_stride_yj * 2;
-    }
-    if (height & 1) {
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
-      RGB24ToYJRow(src_rgb24, dst_yj, width);
-#else
-      RGB24ToARGBRow(src_rgb24, row, width);
-      ARGBToYJRow(row, dst_yj, width);
-#endif
-    }
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
-    free_aligned_buffer_64(row);
-#endif
+  for (y = 0; y < height; ++y) {
+    RGB24ToYJRow(src_rgb24, dst_yj, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_yj += dst_stride_yj;
   }
   return 0;
 }
@@ -2498,27 +2454,38 @@ int RAWToJ400(const uint8_t* src_raw,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
-     defined(HAS_RAWTOYJROW_MMI))
   void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_yj, int width) =
       RAWToYJRow_C;
-#else
-  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
-      RAWToARGBRow_C;
-  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
-      ARGBToYJRow_C;
-#endif
   if (!src_raw || !dst_yj || 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;
   }
-
-// Neon version does direct RAW to YJ.
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_yj = 0;
+  }
+#if defined(HAS_RAWTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToYJRow = RAWToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RAWToYJRow = RAWToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_RAWTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RAWToYJRow = RAWToYJRow_Any_NEON;
@@ -2526,83 +2493,28 @@ int RAWToJ400(const uint8_t* src_raw,
       RAWToYJRow = RAWToYJRow_NEON;
     }
   }
-#elif defined(HAS_RAWTOYJROW_MSA)
+#endif
+#if defined(HAS_RAWTOYJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RAWToYJRow = RAWToYJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       RAWToYJRow = RAWToYJRow_MSA;
     }
   }
-#elif defined(HAS_RAWTOYJROW_MMI)
+#endif
+#if defined(HAS_RAWTOYJROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RAWToYJRow = RAWToYJRow_Any_MMI;
     if (IS_ALIGNED(width, 8)) {
       RAWToYJRow = RAWToYJRow_MMI;
     }
   }
-// Other platforms do intermediate conversion from RAW to ARGB.
-#else
-#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
-#endif
-
-  {
-#if !(defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
-      defined(HAS_RAWTOYJROW_MMI))
-    // 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_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
-     defined(HAS_RAWTOYJROW_MMI))
-      RAWToYJRow(src_raw, dst_yj, width);
-      RAWToYJRow(src_raw + src_stride_raw, dst_yj + dst_stride_yj, width);
-#else
-      RAWToARGBRow(src_raw, row, width);
-      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
-      ARGBToYJRow(row, dst_yj, width);
-      ARGBToYJRow(row + kRowSize, dst_yj + dst_stride_yj, width);
-#endif
-      src_raw += src_stride_raw * 2;
-      dst_yj += dst_stride_yj * 2;
-    }
-    if (height & 1) {
-#if (defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
-     defined(HAS_RAWTOYJROW_MMI))
-      RAWToYJRow(src_raw, dst_yj, width);
-#else
-      RAWToARGBRow(src_raw, row, width);
-      ARGBToYJRow(row, dst_yj, width);
-#endif
-    }
-#if !(defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
-      defined(HAS_RAWTOYJROW_MMI))
-    free_aligned_buffer_64(row);
-#endif
+  for (y = 0; y < height; ++y) {
+    RAWToYJRow(src_raw, dst_yj, width);
+    src_raw += src_stride_raw;
+    dst_yj += dst_stride_yj;
   }
   return 0;
 }

source/row_any.cc

@@ -695,6 +695,12 @@ ANY11(RGBAToYRow_Any_MMI, RGBAToYRow_MMI, 0, 4, 1, 7)
 #ifdef HAS_RGB24TOYROW_NEON
 ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
 #endif
+#ifdef HAS_RGB24TOYJROW_AVX2
+ANY11(RGB24ToYJRow_Any_AVX2, RGB24ToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RGB24TOYJROW_SSSE3
+ANY11(RGB24ToYJRow_Any_SSSE3, RGB24ToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
 #ifdef HAS_RGB24TOYJROW_NEON
 ANY11(RGB24ToYJRow_Any_NEON, RGB24ToYJRow_NEON, 0, 3, 1, 7)
 #endif
@@ -707,6 +713,12 @@ ANY11(RGB24ToYRow_Any_MMI, RGB24ToYRow_MMI, 0, 3, 1, 7)
 #ifdef HAS_RAWTOYROW_NEON
 ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
 #endif
+#ifdef HAS_RAWTOYJROW_AVX2
+ANY11(RAWToYJRow_Any_AVX2, RAWToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RAWTOYJROW_SSSE3
+ANY11(RAWToYJRow_Any_SSSE3, RAWToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
 #ifdef HAS_RAWTOYJROW_NEON
 ANY11(RAWToYJRow_Any_NEON, RAWToYJRow_NEON, 0, 3, 1, 7)
 #endif

source/row_common.cc

@@ -3307,6 +3307,70 @@ void NV12ToRGB565Row_AVX2(const uint8_t* src_y,
 }
 #endif
 
+#ifdef HAS_RGB24TOYJROW_AVX2
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  //HAS_RGB24TOYJROW_AVX2
+
+#ifdef HAS_RAWTOYJROW_AVX2
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  //HAS_RAWTOYJROW_AVX2
+
+#ifdef HAS_RGB24TOYJROW_SSSE3
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  //HAS_RGB24TOYJROW_SSSE3
+
+#ifdef HAS_RAWTOYJROW_SSSE3
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  //HAS_RAWTOYJROW_SSSE3
+
 float ScaleSumSamples_C(const float* src, float* dst, float scale, int width) {
   float fsum = 0.f;
   int i;