Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

source/row_gcc.cc

Index: source/row_gcc.cc
diff --git a/source/row_gcc.cc b/source/row_gcc.cc
index 12c7dd884de344f8cbe533b5bb430e353223fcaf..112a7208532281a5ebee5d5439ed5a719397fe92 100644
--- a/source/row_gcc.cc
+++ b/source/row_gcc.cc
@@ -3531,27 +3531,30 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
     "sub        %2,%1                          \n"
     "sub        %2,%3                          \n"
 
-    // 16 pixel loop.
+    // 32 pixel loop.
     LABELALIGN
   "1:                                          \n"
-    "vmovdqu    (%2),%%xmm0                    \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vmovdqu    (%2),%%ymm0                    \n"
+    "vpunpckhbw %%ymm0,%%ymm0,%%ymm3           \n"
     "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpxor      %%ymm5,%%ymm3,%%ymm3           \n"
     "vpxor      %%ymm5,%%ymm0,%%ymm0           \n"
-    "vmovdqu    (%0,%2,1),%%xmm1               \n"
-    "vmovdqu    (%1,%2,1),%%xmm2               \n"
-    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
-    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vmovdqu    (%0,%2,1),%%ymm1               \n"
+    "vmovdqu    (%1,%2,1),%%ymm2               \n"
+    "vpunpckhbw %%ymm2,%%ymm1,%%ymm4           \n"
     "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
+    "vpsubb     %%ymm6,%%ymm1,%%ymm4           \n"
     "vpsubb     %%ymm6,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
     "vpmaddubsw %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpaddw     %%ymm7,%%ymm3,%%ymm3           \n"
     "vpaddw     %%ymm7,%%ymm0,%%ymm0           \n"
+    "vpsrlw     $0x8,%%ymm3,%%ymm3             \n"
     "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-    "vmovdqu    %%xmm0,(%3,%2,1)               \n"
-    "lea        0x10(%2),%2                    \n"
-    "sub        $0x10,%4                       \n"
+    "vpackuswb  %%ymm3,%%ymm0,%%ymm0           \n"
+    "vmovdqu    %%ymm0,(%3,%2,1)               \n"
+    "lea        0x20(%2),%2                    \n"
+    "sub        $0x20,%4                       \n"
     "jg        1b                              \n"
     "vzeroupper                                \n"
   : "+r"(src0),       // %0
@@ -3559,7 +3562,8 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
     "+r"(alpha),      // %2
     "+r"(dst),        // %3
     "+r"(width)       // %4
-  :: "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7"
+  :: "memory", "cc", "eax",
+     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 #endif  // HAS_BLENDPLANEROW_AVX2