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

source/row_win.cc

Index: source/row_win.cc
diff --git a/source/row_win.cc b/source/row_win.cc
index 13076ce604edf2532ab96571a4396fa91816b084..0a28430081f0df89d4b2778b99e8880c86b25f62 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -4141,26 +4141,29 @@ void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1,
     sub         edx, esi
     sub         edi, esi
 
-    // 16 pixel loop.
+    // 32 pixel loop.
   convertloop16:
-    vmovdqu     xmm0, [esi]        // alpha
-    vpermq      ymm0, ymm0, 0xd8
-    vpunpcklbw  ymm0, ymm0, ymm0
+    vmovdqu     ymm0, [esi]        // alpha
+    vpunpckhbw  ymm3, ymm0, ymm0   // 8..15, 24..31
+    vpunpcklbw  ymm0, ymm0, ymm0   // 0..7, 16..23
+    vpxor       ymm3, ymm3, ymm5   // a, 255-a
     vpxor       ymm0, ymm0, ymm5   // a, 255-a
-    vmovdqu     xmm1, [eax + esi]  // src0
-    vmovdqu     xmm2, [edx + esi]  // src1
-    vpermq      ymm1, ymm1, 0xd8
-    vpermq      ymm2, ymm2, 0xd8
+    vmovdqu     ymm1, [eax + esi]  // src0
+    vmovdqu     ymm2, [edx + esi]  // src1
+    vpunpckhbw  ymm4, ymm1, ymm2
     vpunpcklbw  ymm1, ymm1, ymm2
+    vpsubb      ymm4, ymm4, ymm6   // bias src0/1 - 128
     vpsubb      ymm1, ymm1, ymm6   // bias src0/1 - 128
+    vpmaddubsw  ymm3, ymm3, ymm4
     vpmaddubsw  ymm0, ymm0, ymm1
+    vpaddw      ymm3, ymm3, ymm7   // unbias result - 32768 and round.
     vpaddw      ymm0, ymm0, ymm7   // unbias result - 32768 and round.
+    vpsrlw      ymm3, ymm3, 8
     vpsrlw      ymm0, ymm0, 8
-    vpackuswb   ymm0, ymm0, ymm0
-    vpermq      ymm0, ymm0, 0xd8
-    vmovdqu     [edi + esi], xmm0
-    lea         esi, [esi + 16]
-    sub         ecx, 16
+    vpackuswb   ymm0, ymm0, ymm3
+    vmovdqu     [edi + esi], ymm0
+    lea         esi, [esi + 32]
+    sub         ecx, 32
     jg          convertloop16
 
     pop         edi