Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
TEST=/usr/local/google/home/fbarchard/intelsde/sde -skx -- out/Release/libyuv_unittest --gtest_filter=*Extract*
R=wangcheng@google.com, magjed@chromium.org

Review URL: https://codereview.chromium.org/2420553002 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1625
+Version: 1626
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -239,6 +239,15 @@ extern "C" {
 #define HAS_BLENDPLANEROW_AVX2
 #endif
 
+
+// The following are available clang 3.4 or gcc 4.7.
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2)) && \
+    !(defined(__clang__) && defined(_M_IX86) )
+#define HAS_ARGBEXTRACTALPHAROW_AVX2
+#endif
+
 // The following are available for AVX2 Visual C and clangcl 32 bit:
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
@@ -880,9 +889,12 @@ void ARGBCopyAlphaRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
 
 void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width);
 void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width);
+void ARGBExtractAlphaRow_AVX2(const uint8* src_argb, uint8* dst_a, int width);
 void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width);
 void ARGBExtractAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_a,
                                   int width);
+void ARGBExtractAlphaRow_Any_AVX2(const uint8* src_argb, uint8* dst_a,
+                                  int width);
 void ARGBExtractAlphaRow_Any_NEON(const uint8* src_argb, uint8* dst_a,
                                   int width);
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1625
+#define LIBYUV_VERSION 1626
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/planar_functions.cc

@@ -2702,6 +2702,12 @@ int ARGBExtractAlpha(const uint8* src_argb, int src_stride,
                                                : ARGBExtractAlphaRow_Any_SSE2;
   }
 #endif
+#if defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 32) ? ARGBExtractAlphaRow_AVX2
+                                                : ARGBExtractAlphaRow_Any_AVX2;
+  }
+#endif
 #if defined(HAS_ARGBEXTRACTALPHAROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_NEON

source/row_any.cc

@@ -474,6 +474,9 @@ ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
 #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
 ANY11(ARGBExtractAlphaRow_Any_SSE2, ARGBExtractAlphaRow_SSE2, 0, 4, 1, 7)
 #endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+ANY11(ARGBExtractAlphaRow_Any_AVX2, ARGBExtractAlphaRow_AVX2, 0, 4, 1, 32)
+#endif
 #ifdef HAS_ARGBEXTRACTALPHAROW_NEON
 ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 15)
 #endif

source/row_gcc.cc

@@ -2860,6 +2860,47 @@ void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
 }
 #endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
+static const uvec8 kShuffleAlphaShort_AVX2 = {
+  3u, 128u, 128u, 128u, 7u, 128u, 128u, 128u,
+  11u, 128u, 128u, 128u, 15u, 128u, 128u, 128u
+};
+
+void ARGBExtractAlphaRow_AVX2(const uint8* src_argb, uint8* dst_a, int width) {
+ asm volatile (
+    "vmovdqa    %3,%%ymm4                      \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ", %%ymm0        \n"
+    "vmovdqu   " MEMACCESS2(0x20, 0) ", %%ymm1 \n"
+    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n" // vpsrld $0x18, %%ymm0
+    "vpshufb    %%ymm5,%%ymm1,%%ymm1           \n"
+    "vmovdqu   " MEMACCESS2(0x40, 0) ", %%ymm2 \n"
+    "vmovdqu   " MEMACCESS2(0x60, 0) ", %%ymm3 \n"
+    "lea       " MEMLEA(0x80, 0) ", %0         \n"
+    "vpackssdw  %%ymm1, %%ymm0, %%ymm0         \n"  // mutates
+    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpackssdw  %%ymm3, %%ymm2, %%ymm2         \n"  // mutates
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm4,%%ymm0           \n"  // unmutate.
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub        $0x20, %2                      \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_a),     // %1
+    "+rm"(width)     // %2
+  : "m"(kPermdARGBToY_AVX),  // %3
+    "m"(kShuffleAlphaShort_AVX2)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
 void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {