Add SplitUVPlanes and MergeUVPlanes

source/planar_functions.cc

Index: source/planar_functions.cc
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 237ab68315be82d0f40d2b8b4ea6305277ae7e6e..565174547083f85126a87383719a8d16201c3f01 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -31,6 +31,12 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
                int width, int height) {
   int y;
   void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;

[fbarchard1, 2016/08/24 17:42:24]

copy plane is really intended as a low level with minimal overhead.  i'm
thinking the planar functions shouldnt support negative height, the convert
higher level should?

+  }
   // Coalesce rows.
   if (src_stride_y == width &&
       dst_stride_y == width) {
@@ -2476,8 +2482,6 @@ int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
                int width, int height) {
   int y;
   int halfwidth = (width + 1) >> 1;
-  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) = SplitUVRow_C;
   void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
@@ -2492,30 +2496,6 @@ int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
     src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
     src_stride_yuy2 = -src_stride_yuy2;
   }
-#if defined(HAS_SPLITUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    SplitUVRow = SplitUVRow_Any_SSE2;
-    if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_SSE2;
-    }
-  }
-#endif
-#if defined(HAS_SPLITUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    SplitUVRow = SplitUVRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      SplitUVRow = SplitUVRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_SPLITUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    SplitUVRow = SplitUVRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_NEON;
-    }
-  }
-#endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
@@ -2546,6 +2526,12 @@ int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
     // row of y and 2 rows of uv
     align_buffer_64(rows, awidth * 3);
 
+    SplitUVRowFunction SplitUVRow = GetOptimizedSplitUVRowFunction(
+        src_yuy2, src_stride_yuy2,
+        rows, awidth,
+        rows, awidth,
+        awidth);
+
     for (y = 0; y < height - 1; y += 2) {
       // Split Y from UV.
       SplitUVRow(src_yuy2, rows, rows + awidth, awidth);
@@ -2559,8 +2545,9 @@ int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
     }
     if (height & 1) {
       // Split Y from UV.
-      SplitUVRow(src_yuy2, rows, dst_uv, awidth);
+      SplitUVRow(src_yuy2, rows, rows + awidth, awidth);
       memcpy(dst_y, rows, width);
+      memcpy(dst_uv, rows + awidth, awidth);

[fbarchard1, 2016/08/24 17:42:24]

can you explain this change?  is there a bug?  The original code seems right to
me.

     }
     free_aligned_buffer_64(rows);
   }
@@ -2574,8 +2561,6 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
                int width, int height) {
   int y;
   int halfwidth = (width + 1) >> 1;
-  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
-                     int width) = SplitUVRow_C;
   void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
@@ -2590,30 +2575,6 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
     src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
     src_stride_uyvy = -src_stride_uyvy;
   }
-#if defined(HAS_SPLITUVROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    SplitUVRow = SplitUVRow_Any_SSE2;
-    if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_SSE2;
-    }
-  }
-#endif
-#if defined(HAS_SPLITUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    SplitUVRow = SplitUVRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      SplitUVRow = SplitUVRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_SPLITUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    SplitUVRow = SplitUVRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
-      SplitUVRow = SplitUVRow_NEON;
-    }
-  }
-#endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
@@ -2644,6 +2605,12 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
     // row of y and 2 rows of uv
     align_buffer_64(rows, awidth * 3);
 
+    SplitUVRowFunction SplitUVRow = GetOptimizedSplitUVRowFunction(
+        src_uyvy, src_stride_uyvy,
+        rows, awidth,
+        rows, awidth,
+        awidth);
+
     for (y = 0; y < height - 1; y += 2) {
       // Split Y from UV.
       SplitUVRow(src_uyvy, rows + awidth, rows, awidth);
@@ -2657,8 +2624,9 @@ int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
     }
     if (height & 1) {
       // Split Y from UV.
-      SplitUVRow(src_uyvy, dst_uv, rows, awidth);
+      SplitUVRow(src_uyvy, rows + awidth, rows, awidth);
       memcpy(dst_y, rows, width);
+      memcpy(dst_uv, rows + awidth, awidth);

[fbarchard1, 2016/08/24 17:42:24]

ditto.  this seems like it would cause a UV buffer overflow on odd height.
You can set libyuv width/height via options and run valgrind or msan to check.

     }
     free_aligned_buffer_64(rows);
   }