Optimize BlendPlane SIMD support for odd width.

source/planar_functions.cc

Index: source/planar_functions.cc
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 459df6fe0792de0ad4b1513e2c376dc5e4fb5dd2..e2bc1a5db82d2f0646d251b25e96ab0e7542a339 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -610,8 +610,7 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
 
 #if defined(HAS_BLENDPLANEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-// TODO(fbarchard): Implement any versions for odd width.
-//  BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
+  BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
       BlendPlaneRow = BlendPlaneRow_SSSE3;
     }
@@ -619,8 +618,8 @@ int BlendPlane(const uint8* src_y0, int src_stride_y0,
 #endif
 #if defined(HAS_BLENDPLANEROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-//  BlendPlaneRow = BlendPlaneRow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
+  BlendPlaneRow = BlendPlaneRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
       BlendPlaneRow = BlendPlaneRow_AVX2;
     }
   }
@@ -678,17 +677,16 @@ int I420Blend(const uint8* src_y0, int src_stride_y0,
 
 #if defined(HAS_BLENDPLANEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-// TODO(fbarchard): Implement any versions for odd width.
-//  BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
-    if (IS_ALIGNED(halfwidth, 8)) {
+  BlendPlaneRow = BlendPlaneRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
       BlendPlaneRow = BlendPlaneRow_SSSE3;
     }
   }
 #endif
 #if defined(HAS_BLENDPLANEROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-//  BlendPlaneRow = BlendPlaneRow_Any_AVX2;
-    if (IS_ALIGNED(halfwidth, 16)) {
+  BlendPlaneRow = BlendPlaneRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {

[Diony Rosa, 2015/12/08 20:06:53]

You're also using BlendPlaneRow on the chroma channels, so don't you need to
verify that halfwidth is a multiple of 32?

[fbarchard, 2015/12/08 20:09:47]

Done.

       BlendPlaneRow = BlendPlaneRow_AVX2;
     }
   }