Enable SIMD for exact RGB to Y conversions

Bug: libyuv:908, b/202888439
Change-Id: Icc5470b85d91b441ded9958ee04b4f32246646f0
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3230489
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Mirko Bonadei <mbonadei@chromium.org>

1 files changed, 275 insertions, 216 deletions

diff --git a/source/convert.cc b/source/convert.cc
index 69f7fb6e..b5e241e1 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -1368,38 +1368,54 @@ int ARGBToI420(const uint8_t* src_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_NEON) && defined(HAS_ARGBTOUVROW_NEON)
+#if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVRow = ARGBToUVRow_NEON;
-      }
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#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_MMI) && defined(HAS_ARGBTOUVROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     ARGBToYRow = ARGBToYRow_Any_MMI;
@@ -1560,26 +1576,38 @@ int ABGRToI420(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;
@@ -1662,16 +1690,22 @@ int RGBAToI420(const uint8_t* src_rgba,
     src_rgba = src_rgba + (height - 1) * src_stride_rgba;
     src_stride_rgba = -src_stride_rgba;
   }
-#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+#if defined(HAS_RGBATOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
     RGBAToYRow = RGBAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      RGBAToUVRow = RGBAToUVRow_SSSE3;
       RGBAToYRow = RGBAToYRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_RGBATOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGBAToYRow = RGBAToYRow_Any_NEON;
@@ -1727,6 +1761,12 @@ int RGBAToI420(const uint8_t* src_rgba,
   return 0;
 }
 
+// Enabled if 1 pass is available
+#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
+     defined(HAS_RGB24TOYROW_MMI))
+#define HAS_RGB24TOYROW
+#endif
+
 // Convert RGB24 to I420.
 LIBYUV_API
 int RGB24ToI420(const uint8_t* src_rgb24,
@@ -1740,8 +1780,7 @@ int RGB24ToI420(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
   void (*RGB24ToUVRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
                        uint8_t* dst_u, uint8_t* dst_v, int width) =
       RGB24ToUVRow_C;
@@ -1766,6 +1805,8 @@ int RGB24ToI420(const uint8_t* src_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
 
+#if defined(HAS_RGB24TOYROW)
+
 // Neon version does direct RGB24 to YUV.
 #if defined(HAS_RGB24TOYROW_NEON) && defined(HAS_RGB24TOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
@@ -1778,8 +1819,7 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       }
     }
   }
-// MMI and MSA version does direct RGB24 to YUV.
-#elif (defined(HAS_RGB24TOYROW_MMI) || defined(HAS_RGB24TOYROW_MSA))
+#endif
 #if defined(HAS_RGB24TOYROW_MMI) && defined(HAS_RGB24TOUVROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RGB24ToUVRow = RGB24ToUVRow_Any_MMI;
@@ -1802,16 +1842,10 @@ int RGB24ToI420(const uint8_t* src_rgb24,
     }
   }
 #endif
+
 // Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
-#if defined(HAS_RGB24TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
-    }
-  }
-#endif
+#else  // HAS_RGB24TOYROW
+
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
@@ -1820,51 +1854,49 @@ int RGB24ToI420(const uint8_t* src_rgb24,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_NEON) && defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVRow = ARGBToUVRow_NEON;
-      }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if 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 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
-#endif
+#endif  // HAS_RGB24TOYROW
 
   {
-#if !(defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-      defined(HAS_RGB24TOYROW_MMI))
+#if !defined(HAS_RGB24TOYROW)
     // 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_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
       RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
@@ -1881,8 +1913,7 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
 #else
@@ -1891,15 +1922,20 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-      defined(HAS_RGB24TOYROW_MMI))
+#if !defined(HAS_RGB24TOYROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RGB24TOYROW
+
+// Enabled if 1 pass is available
+#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
+     defined(HAS_RGB24TOYJROW_MMI))
+#define HAS_RGB24TOYJROW
+#endif
 
-// TODO(fbarchard): Use Matrix version to implement I420 and J420.
 // Convert RGB24 to J420.
 LIBYUV_API
 int RGB24ToJ420(const uint8_t* src_rgb24,
@@ -1913,10 +1949,9 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)) || \
-    defined(HAS_RGB24TOYJROW_MSA) || defined(HAS_RGB24TOYJROW_MMI)
+#if defined(HAS_RGB24TOYJROW)
   void (*RGB24ToUVJRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
-                        uint8_t* dst_u, uint8_t* dst_v, int width) =
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
       RGB24ToUVJRow_C;
   void (*RGB24ToYJRow)(const uint8_t* src_rgb24, uint8_t* dst_y, int width) =
       RGB24ToYJRow_C;
@@ -1924,7 +1959,7 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
   void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
       RGB24ToARGBRow_C;
   void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
-                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+                      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;
@@ -1939,6 +1974,8 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
 
+#if defined(HAS_RGB24TOYJROW)
+
 // Neon version does direct RGB24 to YUV.
 #if defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
@@ -1951,8 +1988,7 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       }
     }
   }
-// MMI and MSA version does direct RGB24 to YUV.
-#elif (defined(HAS_RGB24TOYJROW_MMI) || defined(HAS_RGB24TOYJROW_MSA))
+#endif
 #if defined(HAS_RGB24TOYJROW_MMI) && defined(HAS_RGB24TOUVJROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RGB24ToUVJRow = RGB24ToUVJRow_Any_MMI;
@@ -1975,15 +2011,10 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
     }
   }
 #endif
-#else
-#if defined(HAS_RGB24TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
-    }
-  }
-#endif
+
+// Other platforms do intermediate conversion from RGB24 to ARGB.
+#else  // HAS_RGB24TOYJROW
+
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
@@ -1992,51 +2023,49 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_NEON) && defined(HAS_ARGBTOUVJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
-    ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVJRow = ARGBToUVJRow_NEON;
-      }
+#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_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#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;
-    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVJRow = ARGBToUVJRow_SSSE3;
-      ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
+#if defined(HAS_ARGBTOUVJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
-    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVJRow = ARGBToUVJRow_AVX2;
-      ARGBToYJRow = ARGBToYJRow_AVX2;
     }
   }
 #endif
-#endif
+#endif  // HAS_RGB24TOYJROW
 
   {
-#if !((defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)) || \
-      defined(HAS_RGB24TOYJROW_MSA) || defined(HAS_RGB24TOYJROW_MMI))
+#if !defined(HAS_RGB24TOYJROW)
     // 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_RGB24TOUVJROW_NEON)) || \
-     defined(HAS_RGB24TOYJROW_MSA) || defined(HAS_RGB24TOYJROW_MMI))
+#if defined(HAS_RGB24TOYJROW)
       RGB24ToUVJRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
       RGB24ToYJRow(src_rgb24, dst_y, width);
       RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
@@ -2053,8 +2082,7 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if ((defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)) || \
-     defined(HAS_RGB24TOYJROW_MSA) || defined(HAS_RGB24TOYJROW_MMI))
+#if defined(HAS_RGB24TOYJROW)
       RGB24ToUVJRow(src_rgb24, 0, dst_u, dst_v, width);
       RGB24ToYJRow(src_rgb24, dst_y, width);
 #else
@@ -2063,31 +2091,37 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       ARGBToYJRow(row, dst_y, width);
 #endif
     }
-#if !((defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)) || \
-      defined(HAS_RGB24TOYJROW_MSA) || defined(HAS_RGB24TOYJROW_MMI))
+#if !defined(HAS_RGB24TOYJROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RGB24TOYJROW
+
+// Enabled if 1 pass is available
+#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
+     defined(HAS_RAWTOYROW_MMI))
+#define HAS_RAWTOYROW
+#endif
 
 // Convert RAW to I420.
 LIBYUV_API
 int RAWToI420(const uint8_t* src_raw,
-              int src_stride_raw,
-              uint8_t* dst_y,
-              int dst_stride_y,
-              uint8_t* dst_u,
-              int dst_stride_u,
-              uint8_t* dst_v,
-              int dst_stride_v,
-              int width,
-              int height) {
+                int src_stride_raw,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height) {
   int y;
-#if (defined(HAS_RAWTOYROW_NEON) && defined(HAS_RAWTOUVROW_NEON)) || \
-    defined(HAS_RAWTOYROW_MSA) || defined(HAS_RAWTOYROW_MMI)
-  void (*RAWToUVRow)(const uint8_t* src_raw, int src_stride_raw, uint8_t* dst_u,
-                     uint8_t* dst_v, int width) = RAWToUVRow_C;
+#if defined(HAS_RAWTOYROW)
+  void (*RAWToUVRow)(const uint8_t* src_raw, int src_stride_raw,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RAWToUVRow_C;
   void (*RAWToYRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
       RAWToYRow_C;
 #else
@@ -2109,6 +2143,8 @@ int RAWToI420(const uint8_t* src_raw,
     src_stride_raw = -src_stride_raw;
   }
 
+#if defined(HAS_RAWTOYROW)
+
 // Neon version does direct RAW to YUV.
 #if defined(HAS_RAWTOYROW_NEON) && defined(HAS_RAWTOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
@@ -2121,8 +2157,7 @@ int RAWToI420(const uint8_t* src_raw,
       }
     }
   }
-// MMI and MSA version does direct RAW to YUV.
-#elif (defined(HAS_RAWTOYROW_MMI) || defined(HAS_RAWTOYROW_MSA))
+#endif
 #if defined(HAS_RAWTOYROW_MMI) && defined(HAS_RAWTOUVROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RAWToUVRow = RAWToUVRow_Any_MMI;
@@ -2145,28 +2180,10 @@ int RAWToI420(const uint8_t* src_raw,
     }
   }
 #endif
+
 // Other platforms do intermediate conversion from RAW to ARGB.
-#else
-#if defined(HAS_RAWTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    RAWToARGBRow = RAWToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      RAWToARGBRow = RAWToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOYROW_NEON) && defined(HAS_ARGBTOUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVRow = ARGBToUVRow_Any_NEON;
-    ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVRow = ARGBToUVRow_NEON;
-      }
-    }
-  }
-#endif
+#else  // HAS_RAWTOYROW
+
 #if defined(HAS_RAWTOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
@@ -2175,39 +2192,49 @@ int RAWToI420(const uint8_t* src_raw,
     }
   }
 #endif
-#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_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_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
 #endif
+#endif  // HAS_RAWTOYROW
 
   {
-#if !(defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-      defined(HAS_RAWTOYROW_MMI))
+#if !defined(HAS_RAWTOYROW)
     // 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_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-     defined(HAS_RAWTOYROW_MMI))
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
       RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
@@ -2224,8 +2251,7 @@ int RAWToI420(const uint8_t* src_raw,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-     defined(HAS_RAWTOYROW_MMI))
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
 #else
@@ -2234,32 +2260,36 @@ int RAWToI420(const uint8_t* src_raw,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-      defined(HAS_RAWTOYROW_MMI))
+#if !defined(HAS_RAWTOYROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RAWTOYROW
+
+// Enabled if 1 pass is available
+#if (defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA) || \
+     defined(HAS_RAWTOYJROW_MMI))
+#define HAS_RAWTOYJROW
+#endif
 
-// TODO(fbarchard): Use Matrix version to implement I420 and J420.
 // Convert RAW to J420.
 LIBYUV_API
 int RAWToJ420(const uint8_t* src_raw,
-              int src_stride_raw,
-              uint8_t* dst_y,
-              int dst_stride_y,
-              uint8_t* dst_u,
-              int dst_stride_u,
-              uint8_t* dst_v,
-              int dst_stride_v,
-              int width,
-              int height) {
+                int src_stride_raw,
+                uint8_t* dst_y,
+                int dst_stride_y,
+                uint8_t* dst_u,
+                int dst_stride_u,
+                uint8_t* dst_v,
+                int dst_stride_v,
+                int width,
+                int height) {
   int y;
-#if (defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)) || \
-    defined(HAS_RAWTOYJROW_MSA) || defined(HAS_RAWTOYJROW_MMI)
+#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) =
+                       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;
@@ -2267,7 +2297,7 @@ int RAWToJ420(const uint8_t* src_raw,
   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) =
+                      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;
@@ -2282,6 +2312,8 @@ int RAWToJ420(const uint8_t* src_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)) {
@@ -2294,8 +2326,7 @@ int RAWToJ420(const uint8_t* src_raw,
       }
     }
   }
-// MMI and MSA version does direct RAW to YUV.
-#elif (defined(HAS_RAWTOYJROW_MMI) || defined(HAS_RAWTOYJROW_MSA))
+#endif
 #if defined(HAS_RAWTOYJROW_MMI) && defined(HAS_RAWTOUVJROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     RAWToUVJRow = RAWToUVJRow_Any_MMI;
@@ -2318,27 +2349,10 @@ int RAWToJ420(const uint8_t* src_raw,
     }
   }
 #endif
-#else
-#if defined(HAS_RAWTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    RAWToARGBRow = RAWToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      RAWToARGBRow = RAWToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOYJROW_NEON) && defined(HAS_ARGBTOUVJROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
-    ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVJRow = ARGBToUVJRow_NEON;
-      }
-    }
-  }
-#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;
@@ -2347,39 +2361,49 @@ int RAWToJ420(const uint8_t* src_raw,
     }
   }
 #endif
-#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_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
+#if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVJRow = ARGBToUVJRow_AVX2;
       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_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)) || \
-      defined(HAS_RAWTOYJROW_MSA) || defined(HAS_RAWTOYJROW_MMI))
+#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_NEON) && defined(HAS_RAWTOUVJROW_NEON)) || \
-     defined(HAS_RAWTOYJROW_MSA) || defined(HAS_RAWTOYJROW_MMI))
+#if defined(HAS_RAWTOYJROW)
       RAWToUVJRow(src_raw, src_stride_raw, dst_u, dst_v, width);
       RAWToYJRow(src_raw, dst_y, width);
       RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
@@ -2396,8 +2420,7 @@ int RAWToJ420(const uint8_t* src_raw,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if ((defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)) || \
-     defined(HAS_RAWTOYJROW_MSA) || defined(HAS_RAWTOYJROW_MMI))
+#if defined(HAS_RAWTOYJROW)
       RAWToUVJRow(src_raw, 0, dst_u, dst_v, width);
       RAWToYJRow(src_raw, dst_y, width);
 #else
@@ -2406,13 +2429,13 @@ int RAWToJ420(const uint8_t* src_raw,
       ARGBToYJRow(row, dst_y, width);
 #endif
     }
-#if !((defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)) || \
-      defined(HAS_RAWTOYJROW_MSA) || defined(HAS_RAWTOYJROW_MMI))
+#if !defined(HAS_RAWTOYJROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RAWTOYJROW
 
 // Convert RGB565 to I420.
 LIBYUV_API
@@ -2507,26 +2530,38 @@ int RGB565ToI420(const uint8_t* src_rgb565,
     }
   }
 #endif
-#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
 #endif
   {
 #if !(defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
@@ -2666,26 +2701,38 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
     }
   }
 #endif
-#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
 #endif
   {
 #if !(defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
@@ -2822,26 +2869,38 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
     }
   }
 #endif
-#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_MMI) && defined(HAS_ARGBTOUVROW_MMI)
   if (TestCpuFlag(kCpuHasMMI)) {
     ARGBToUVRow = ARGBToUVRow_Any_MMI;