ARGBToYJ_AVX2 port to AVX2.

BUG=none
TEST=none
Review URL: https://webrtc-codereview.appspot.com/1272008

git-svn-id: http://libyuv.googlecode.com/svn/trunk@640 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/row.h

@@ -137,6 +137,7 @@ extern "C" {
 #define HAS_ARGBSHUFFLEROW_AVX2
 #define HAS_ARGBTOUVROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
 #define HAS_HALFROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
@@ -403,6 +404,8 @@ void UYVYToARGBRow_NEON(const uint8* src_uyvy,
 void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
 void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
 void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
 void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
 void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
 void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);

source/convert_from_argb.cc

@@ -1012,6 +1012,14 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
@@ -1077,6 +1085,14 @@ int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;

source/row_any.cc

@@ -204,6 +204,7 @@ BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
 
 #ifdef HAS_ARGBTOYROW_AVX2
 YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
+YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
 YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
 YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
 #endif

source/row_win.cc

@@ -35,6 +35,11 @@ static const lvec8 kARGBToY_AVX = {
   13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
 };
 
+static const lvec8 kARGBToYJ_AVX = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0,
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
 static const vec8 kARGBToU = {
   112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
 };
@@ -119,6 +124,9 @@ static const uvec8 kAddY16 = {
 static const vec16 kAddYJ64 = {
   64, 64, 64, 64, 64, 64, 64, 64
 };
+static const lvec16 kAddYJ64_AVX = {
+  64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64
+};
 
 static const ulvec8 kAddY16_AVX = {
   16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
@@ -760,6 +768,48 @@ void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
 }
 #endif  //  HAS_ARGBTOYROW_AVX2
 
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vmovdqa    ymm4, kARGBToYJ_AVX
+    vmovdqa    ymm5, kAddYJ64_AVX
+    vmovdqa    ymm6, kShufARGBToY_AVX
+
+    align      16
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
 __declspec(naked) __declspec(align(16))
 void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
   __asm {