Add ARGBExtractAlpha_AVX2 function

source/row_gcc.cc

Index: source/row_gcc.cc
diff --git a/source/row_gcc.cc b/source/row_gcc.cc
index bf9ddde42c3a4c0e759d831f90c06996b49e8419..80fbbb8717890957dab76bbd3f90d5e865111f7a 100644
--- a/source/row_gcc.cc
+++ b/source/row_gcc.cc
@@ -2860,6 +2860,36 @@ void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
 }
 #endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+void ARGBExtractAlphaRow_AVX2(const uint8* src_argb, uint8* dst_a, int width) {
+ asm volatile (
+    "vmovdqa    %3,%%ymm6                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ", %%ymm0        \n"
+    "vmovdqu   " MEMACCESS2(0x20, 0) ", %%ymm1 \n"
+    "lea       " MEMLEA(0x40, 0) ", %0         \n"
+    "vpsrld     $0x18, %%ymm0, %%ymm0          \n"
+    "vpsrld     $0x18, %%ymm1, %%ymm1          \n"
+    "vpackssdw  %%ymm1, %%ymm0, %%ymm0         \n"  // mutates
+    "vpackuswb  %%ymm0, %%ymm0, %%ymm0         \n"  // mutates
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vmovdqu    %%xmm0," MEMACCESS(1) "        \n"
+    "lea        " MEMLEA(0x10, 1) ", %1        \n"
+    "sub        $0x10, %2                      \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_a),     // %1
+    "+rm"(width)     // %2

[wangcheng, 2016/10/13 19:52:18]

Do you need to check "width" is multiply of 16?

[fbarchard1, 2016/10/13 21:46:27]

planar_functions.cc does that check:

#if defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
  if (TestCpuFlag(kCpuHasAVX2)) {
    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_AVX2
                                                : ARGBExtractAlphaRow_Any_AVX2;
  }
#endif

+  : "m"(kPermdARGBToY_AVX)  // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm6"
+  );
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2
+
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
 void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {