ARGBMirror AVX2

BUG=none
TEST=out\release\libyuv_unittest --gtest_filter=*ARGBMirror*
Review URL: https://webrtc-codereview.appspot.com/1159005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@594 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 590
+Version: 594
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -139,6 +139,7 @@ extern "C" {
 #define HAS_UYVYTOYROW_AVX2
 #define HAS_HALFROW_AVX2
 #define HAS_MIRRORROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
 
 // Effects
 #define HAS_ARGBATTENUATEROW_AVX2
@@ -574,6 +575,7 @@ void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
 void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
                    int width);
 
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
 void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
 void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
 void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 590
+#define LIBYUV_VERSION 594
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/convert_from.cc

@@ -50,22 +50,28 @@ int I420ToI422(const uint8* src_y, int src_stride_y,
   }
   int halfwidth = (width + 1) >> 1;
   void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(halfwidth, 32)) {
-    CopyRow = CopyRow_NEON;
-  }
-#elif defined(HAS_COPYROW_X86)
+#if defined(HAS_COPYROW_X86)
   if (IS_ALIGNED(halfwidth, 4)) {
     CopyRow = CopyRow_X86;
+  }
+#endif
 #if defined(HAS_COPYROW_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(halfwidth, 32) &&
-        IS_ALIGNED(src_u, 16) && IS_ALIGNED(src_stride_u, 16) &&
-        IS_ALIGNED(src_v, 16) && IS_ALIGNED(src_stride_v, 16) &&
-        IS_ALIGNED(dst_u, 16) && IS_ALIGNED(dst_stride_u, 16) &&
-        IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
-      CopyRow = CopyRow_SSE2;
-    }
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(halfwidth, 32) &&
+      IS_ALIGNED(src_u, 16) && IS_ALIGNED(src_stride_u, 16) &&
+      IS_ALIGNED(src_v, 16) && IS_ALIGNED(src_stride_v, 16) &&
+      IS_ALIGNED(dst_u, 16) && IS_ALIGNED(dst_stride_u, 16) &&
+      IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    CopyRow = CopyRow_AVX2;
+  }
 #endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(halfwidth, 32)) {
+    CopyRow = CopyRow_NEON;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)

source/planar_functions.cc

@@ -448,7 +448,15 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
     ARGBMirrorRow = ARGBMirrorRow_SSSE3;
   }
-#elif defined(HAS_ARGBMIRRORROW_NEON)
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  bool clear = false;
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    clear = true;
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
     ARGBMirrorRow = ARGBMirrorRow_NEON;
   }
@@ -460,6 +468,12 @@ int ARGBMirror(const uint8* src_argb, int src_stride_argb,
     src_argb += src_stride_argb;
     dst_argb += dst_stride_argb;
   }
+
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (clear) {
+    __asm vzeroupper;
+  }
+#endif
   return 0;
 }
 

source/rotate_argb.cc

@@ -99,7 +99,15 @@ void ARGBRotate180(const uint8* src, int src_stride,
       IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
     ARGBMirrorRow = ARGBMirrorRow_SSSE3;
   }
-#elif defined(HAS_ARGBMIRRORROW_NEON)
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  bool clear = false;
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    clear = true;
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
     ARGBMirrorRow = ARGBMirrorRow_NEON;
   }
@@ -151,6 +159,11 @@ void ARGBRotate180(const uint8* src, int src_stride,
     src_bot -= src_stride;
     dst_bot -= dst_stride;
   }
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (clear) {
+    __asm vzeroupper;
+  }
+#endif
 }
 
 LIBYUV_API

source/row_win.cc

@@ -3056,6 +3056,33 @@ void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_ARGBMIRRORROW_SSSE3
 
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 32]
+    vmovdqa   ymm5, kARGBShuffleMirror_AVX2
+
+    align      16
+ convertloop:
+    vpermd    ymm0, ymm5, [eax + ecx * 4]  // permute dword order
+    sub       ecx, 8
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
 #ifdef HAS_SPLITUVROW_SSE2
 __declspec(naked) __declspec(align(16))
 void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {