AVX2 YUV alpha blender and improved unittests

include/libyuv/row.h

Index: include/libyuv/row.h
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index deed8a422df5f17e507de684b3fbc4e5b8d11a31..cf96c0516aa675c26f1f9096fc1395370d23faa4 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -233,6 +233,7 @@ extern "C" {
 #define HAS_ARGBMULTIPLYROW_AVX2
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBUNATTENUATEROW_AVX2
+#define HAS_BLENDPLANEROW_AVX2
 #endif
 
 // The following are available for AVX2 Visual C and clangcl 32 bit:
@@ -253,12 +254,6 @@ extern "C" {
 #define HAS_RGB565TOARGBROW_AVX2
 #endif
 
-// The following are available for 32 bit Visual C and clangcl 32 bit:
-// TODO(fbarchard): Port to gcc.
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
-#define HAS_BLENDPLANEROW_SSSE3
-#endif
-
 // The following are also available on x64 Visual C.
 #if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
     (!defined(__clang__) || defined(__SSSE3__))
@@ -1464,6 +1459,12 @@ void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
 // Unattenuated planar alpha blend.
 void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1,
                          const uint8* alpha, uint8* dst, int width);
+void BlendPlaneRow_Any_SSSE3(const uint8* src0, const uint8* src1,
+                             const uint8* alpha, uint8* dst, int width);
+void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
+                        const uint8* alpha, uint8* dst, int width);
+void BlendPlaneRow_Any_AVX2(const uint8* src0, const uint8* src1,
+                            const uint8* alpha, uint8* dst, int width);
 void BlendPlaneRow_C(const uint8* src0, const uint8* src1,
                      const uint8* alpha, uint8* dst, int width);