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

source/row_win.cc

 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 4123 matching lines @@
     vbroadcastss ymm7, xmm7
     mov         eax, [esp + 8 + 4]   // src0
     mov         edx, [esp + 8 + 8]   // src1
     mov         esi, [esp + 8 + 12]  // alpha
     mov         edi, [esp + 8 + 16]  // dst
     mov         ecx, [esp + 8 + 20]  // width
     sub         eax, esi
     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
     pop         esi
     vzeroupper
     ret
   }
 }
 #endif  // HAS_BLENDPLANEROW_AVX2
 
@@ skipping 2184 matching lines @@
 }
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
 #endif  // defined(_M_X64)
 #endif  // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64))
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif