I400ToARGB_AVX2 port from SSE2 to AVX2.

BUG=403
TESTED=libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter=*I400ToARGB*
R=brucedawson@google.com

Review URL: https://webrtc-codereview.appspot.com/46569004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1322 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/row.h

@@ -200,6 +200,7 @@ extern "C" {
 #define HAS_I422TOARGB4444ROW_AVX2
 #define HAS_I444TOARGBROW_AVX2
 #define HAS_I411TOARGBROW_AVX2
+#define HAS_I400TOARGBROW_AVX2
 // TODO(fbarchard): Port to Neon
 #define HAS_ARGBTORGB565DITHERROW_SSE2
 #define HAS_ARGBTORGB565DITHERROW_AVX2
@@ -935,9 +936,11 @@ void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
 void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
 
 void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
 void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
 void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
 void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
 void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
 
 void I444ToARGBRow_C(const uint8* src_y,

source/convert_argb.cc

@@ -349,6 +349,14 @@ int I400ToARGB(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I400ToARGBRow = I400ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I400TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I400ToARGBRow = I400ToARGBRow_Any_NEON;

source/row_any.cc

@@ -192,6 +192,9 @@ RGBANY(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, ARGBToARGB4444Row_C,
 #if defined(HAS_I400TOARGBROW_SSE2)
 RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C, 1, 4, 7)
 #endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+RGBANY(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, I400ToARGBRow_C, 1, 4, 15)
+#endif
 #if defined(HAS_YTOARGBROW_SSE2)
 RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C, 1, 4, 7)
 #endif

source/row_win.cc

@@ -284,6 +284,38 @@ void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
   }
 }
 
+#ifdef HAS_I400TOARGBROW_AVX2
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov         eax, [esp + 4]        // src_y
+    mov         edx, [esp + 8]        // dst_argb
+    mov         ecx, [esp + 12]       // pix
+    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0xff000000
+    vpslld      ymm5, ymm5, 24
+
+  convertloop:
+    vmovdqu     xmm0, [eax]
+    lea         eax,  [eax + 16]
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpcklbw  ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpckhwd  ymm1, ymm0, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm0
+    vpor        ymm0, ymm0, ymm5
+    vpor        ymm1, ymm1, ymm5
+    vmovdqu     [edx], ymm0
+    vmovdqu     [edx + 32], ymm1
+    lea         edx, [edx + 64]
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I400TOARGBROW_AVX2
+
 __declspec(naked) __declspec(align(16))
 void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
   __asm {