AVX2 version of half rows.

BUG=none
TEST=out\release\libyuv_unittest.exe --gtest_filter=*I422ToI420*
Review URL: https://webrtc-codereview.appspot.com/1157004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@590 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 589
+Version: 590
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -137,6 +137,7 @@ extern "C" {
 #define HAS_UYVYTOUV422ROW_AVX2
 #define HAS_UYVYTOUVROW_AVX2
 #define HAS_UYVYTOYROW_AVX2
+#define HAS_HALFROW_AVX2
 
 // Effects
 #define HAS_ARGBATTENUATEROW_AVX2
@@ -1269,6 +1270,8 @@ void HalfRow_C(const uint8* src_uv, int src_uv_stride,
                uint8* dst_uv, int pix);
 void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
                   uint8* dst_uv, int pix);
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
 void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
                   uint8* dst_uv, int pix);
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 589
+#define LIBYUV_VERSION 590
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/convert.cc

@@ -97,7 +97,15 @@ int I422ToI420(const uint8* src_y, int src_stride_y,
       IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
     HalfRow = HalfRow_SSE2;
   }
-#elif defined(HAS_HALFROW_NEON)
+#endif
+#if defined(HAS_HALFROW_AVX2)
+  bool clear = false;
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(halfwidth, 32)) {
+    clear = true;
+    HalfRow = HalfRow_AVX2;
+  }
+#endif
+#if defined(HAS_HALFROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(halfwidth, 16)) {
     HalfRow = HalfRow_NEON;
   }
@@ -128,6 +136,11 @@ int I422ToI420(const uint8* src_y, int src_stride_y,
   if (height & 1) {
     HalfRow(src_v, 0, dst_v, halfwidth);
   }
+#if defined(HAS_HALFROW_AVX2)
+  if (clear) {
+    __asm vzeroupper;
+  }
+#endif
   return 0;
 }
 

source/row_win.cc

@@ -5531,6 +5531,32 @@ void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
   }
 }
 
+#ifdef HAS_HALFROW_AVX2
+__declspec(naked) __declspec(align(16))
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      16
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vpavgb     ymm0, ymm0, [eax + edx]
+    sub        ecx, 32
+    vmovdqu    [eax + edi], ymm0
+    lea        eax,  [eax + 32]
+    jg         convertloop
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_HALFROW_AVX2
+
 __declspec(naked) __declspec(align(16))
 void ARGBToBayerRow_SSSE3(const uint8* src_argb,
                           uint8* dst_bayer, uint32 selector, int pix) {