CopyRow_AVX which supports unaligned pointers for Sandy Bridge CPU.

BUG=363
TESTED=out\release\libyuv_unittest --gtest_filter=*ARGBToARGB_*
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/31489004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1097 16f28f9a-4ce2-e073-06de-1de4eb20be90

9 files changed, 81 insertions, 2 deletions

diff --git a/README.chromium b/README.chromium
index 895e95a..a2b74e7 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1096
+Version: 1097
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index 05a7888..e6e2c98 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -122,6 +122,7 @@ extern "C" {
 #define HAS_BGRATOUVROW_SSSE3
 #define HAS_BGRATOYROW_SSSE3
 #define HAS_COPYROW_ERMS
+#define HAS_COPYROW_AVX
 #define HAS_COPYROW_SSE2
 #define HAS_COPYROW_X86
 #define HAS_HALFROW_SSE2
@@ -891,6 +892,7 @@ void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
                          int width);
 
 void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_AVX(const uint8* src, uint8* dst, int count);
 void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
 void CopyRow_X86(const uint8* src, uint8* dst, int count);
 void CopyRow_NEON(const uint8* src, uint8* dst, int count);
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index ab370ff..ca5be9b 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1096
+#define LIBYUV_VERSION 1097
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/convert.cc b/source/convert.cc
index c31ecf2..f205143 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -201,6 +201,11 @@ static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
     CopyRow = CopyRow_SSE2;
   }
 #endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX) && IS_ALIGNED(width, 64)) {
+    CopyRow = CopyRow_AVX;
+  }
+#endif
 #if defined(HAS_COPYROW_ERMS)
   if (TestCpuFlag(kCpuHasERMS)) {
     CopyRow = CopyRow_ERMS;
@@ -441,6 +446,11 @@ int Q420ToI420(const uint8* src_y, int src_stride_y,
     CopyRow = CopyRow_SSE2;
   }
 #endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX) && IS_ALIGNED(width, 64)) {
+    CopyRow = CopyRow_AVX;
+  }
+#endif
 #if defined(HAS_COPYROW_ERMS)
   if (TestCpuFlag(kCpuHasERMS)) {
     CopyRow = CopyRow_ERMS;
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 3857008..3cfa6ce 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -53,6 +53,11 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
     CopyRow = CopyRow_SSE2;
   }
 #endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX) && IS_ALIGNED(width, 64)) {
+    CopyRow = CopyRow_AVX;
+  }
+#endif
 #if defined(HAS_COPYROW_ERMS)
   if (TestCpuFlag(kCpuHasERMS)) {
     CopyRow = CopyRow_ERMS;
diff --git a/source/rotate.cc b/source/rotate.cc
index 7f9b13d..890c4b5 100644
--- a/source/rotate.cc
+++ b/source/rotate.cc
@@ -946,6 +946,11 @@ void RotatePlane180(const uint8* src, int src_stride,
     CopyRow = CopyRow_SSE2;
   }
 #endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX) && IS_ALIGNED(width, 64)) {
+    CopyRow = CopyRow_AVX;
+  }
+#endif
 #if defined(HAS_COPYROW_ERMS)
   if (TestCpuFlag(kCpuHasERMS)) {
     CopyRow = CopyRow_ERMS;
diff --git a/source/rotate_argb.cc b/source/rotate_argb.cc
index ab0f9ce..a8d7fc2 100644
--- a/source/rotate_argb.cc
+++ b/source/rotate_argb.cc
@@ -136,6 +136,11 @@ void ARGBRotate180(const uint8* src, int src_stride,
     CopyRow = CopyRow_SSE2;
   }
 #endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX) && IS_ALIGNED(width, 64)) {
+    CopyRow = CopyRow_AVX;
+  }
+#endif
 #if defined(HAS_COPYROW_ERMS)
   if (TestCpuFlag(kCpuHasERMS)) {
     CopyRow = CopyRow_ERMS;
diff --git a/source/row_posix.cc b/source/row_posix.cc
index 106fda5..b21002a 100644
--- a/source/row_posix.cc
+++ b/source/row_posix.cc
@@ -3266,6 +3266,31 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_SSE2
 
+#ifdef HAS_COPYROW_AVX
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%ymm0         \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%ymm0," MEMACCESS(1) "         \n"
+    "movdqa    %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x40,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "ymm0", "ymm1"
+#endif
+  );
+}
+#endif  // HAS_COPYROW_AVX
+
 #ifdef HAS_COPYROW_X86
 void CopyRow_X86(const uint8* src, uint8* dst, int width) {
   size_t width_tmp = (size_t)(width);
diff --git a/source/row_win.cc b/source/row_win.cc
index d79c353..f507718 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -3687,6 +3687,32 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_SSE2
 
+#ifdef HAS_COPYROW_AVX
+// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time.
+__declspec(naked) __declspec(align(16))
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 64
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_COPYROW_AVX
+
 // Unaligned Multiple of 1.
 __declspec(naked) __declspec(align(16))
 void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
@@ -3704,6 +3730,7 @@ void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
 }
 
 #ifdef HAS_COPYROW_X86
+// Unaligned Multiple of 4.
 __declspec(naked) __declspec(align(16))
 void CopyRow_X86(const uint8* src, uint8* dst, int count) {
   __asm {