Add SplitUVPlanes and MergeUVPlanes

source/row_common.cc

Index: source/row_common.cc
diff --git a/source/row_common.cc b/source/row_common.cc
index 229da97527c440b5786f4560828e033aa90c3a6f..b5e42b3aec26b028eb10ff0bf52d0f3cd71e0dd1 100644
--- a/source/row_common.cc
+++ b/source/row_common.cc
@@ -13,6 +13,7 @@
 #include <string.h>  // For memcpy and memset.
 
 #include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -1800,6 +1801,37 @@ void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
   }
 }
 
+SplitUVRowFunction GetOptimizedSplitUVRowFunction(

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

suggest removing.  common is not meant for this sort of function.
For each row function, there should be a _NEON, _SSE2, and _C version, which
goes here.
If we do getters, I think this belongs in its own file, or row_any.c

+    const uint8 *src_uv, int src_stride_uv,
+    const uint8 *dst_u, int dst_stride_u,
+    const uint8 *dst_v, int dst_stride_v,
+    int width) {
+#if defined(HAS_SPLITUVROW_DSPR2)
+  if (TestCpuFlag(kCpuHasDSPR2) &&
+      IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+      IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+      IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
+    return IS_ALIGNED(width, 16) ? SplitUVRow_DSPR2 : SplitUVRow_Any_DSPR2;
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    return IS_ALIGNED(width, 16) ? SplitUVRow_NEON : SplitUVRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    return IS_ALIGNED(width, 32) ? SplitUVRow_AVX2 : SplitUVRow_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    return IS_ALIGNED(width, 16) ? SplitUVRow_SSE2 : SplitUVRow_Any_SSE2;
+  }
+#endif
+  return SplitUVRow_C;
+}
+
 void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
                   int width) {
   int x;
@@ -1816,6 +1848,25 @@ void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
   }
 }
 
+MergeUVRowFunction GetOptimizedMergeUVRowFunction(int width) {
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    return IS_ALIGNED(width, 16) ? MergeUVRow_NEON : MergeUVRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    return IS_ALIGNED(width, 32) ? MergeUVRow_AVX2 : MergeUVRow_Any_AVX2;
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    return IS_ALIGNED(width, 16) ? MergeUVRow_SSE2 : MergeUVRow_Any_SSE2;
+  }
+#endif
+  return MergeUVRow_C;
+}
+
 void CopyRow_C(const uint8* src, uint8* dst, int count) {
   memcpy(dst, src, count);
 }