Enhanced Rep Mov String version of CopyRow for posix and use cpu detect for ERMS

BUG=213
TEST=none
Review URL: https://webrtc-codereview.appspot.com/1306008

git-svn-id: http://libyuv.googlecode.com/svn/trunk@658 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/row.h

@@ -61,6 +61,7 @@ extern "C" {
 #define HAS_BGRATOYROW_SSSE3
 #define HAS_COPYROW_SSE2
 #define HAS_COPYROW_X86
+#define HAS_COPYROW_ERMS
 #define HAS_HALFROW_SSE2
 #define HAS_I400TOARGBROW_SSE2
 #define HAS_I411TOARGBROW_SSSE3
@@ -130,7 +131,6 @@ extern "C" {
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
 #define HAS_ARGBCOLORTABLEROW_X86
-#define HAS_COPYROW_AVX2
 // Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
 // TODO(fbarchard): Hook these up to all functions. e.g. format conversion.
@@ -649,7 +649,7 @@ 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_AVX2(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);
 void CopyRow_MIPS(const uint8* src, uint8* dst, int count);

source/convert.cc

@@ -308,10 +308,10 @@ static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
+#if defined(HAS_COPYROW_ERMS)
   // TODO(fbarchard): Detect Fast String support.
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_NEON)
@@ -539,9 +539,9 @@ int Q420ToI420(const uint8* src_y, int src_stride_y,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)

source/convert_from.cc

@@ -64,9 +64,9 @@ int I420ToI422(const uint8* src_y, int src_stride_y,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_NEON)

source/planar_functions.cc

@@ -49,10 +49,9 @@ void CopyPlane(const uint8* src_y, int src_stride_y,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
-  // TODO(fbarchard): Detect Fast String support.
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_NEON)

source/rotate.cc

@@ -911,10 +911,9 @@ void RotatePlane180(const uint8* src, int src_stride,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
-  // TODO(fbarchard): Detect Fast String support.
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)

source/rotate_argb.cc

@@ -128,10 +128,9 @@ void ARGBRotate180(const uint8* src, int src_stride,
     CopyRow = CopyRow_SSE2;
   }
 #endif
-#if defined(HAS_COPYROW_AVX2)
-  // TODO(fbarchard): Detect Fast String support.
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    CopyRow = CopyRow_AVX2;
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
   }
 #endif
 #if defined(HAS_COPYROW_MIPS)

source/row_posix.cc

@@ -3027,6 +3027,19 @@ void CopyRow_X86(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_COPYROW_X86
 
+// Unaligned Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = static_cast<size_t>(width);
+  asm volatile (
+    "rep movsb                                 \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+
 #ifdef HAS_SETROW_X86
 void SetRow_X86(uint8* dst, uint32 v32, int width) {
   size_t width_tmp = static_cast<size_t>(width);

source/row_win.cc

@@ -3497,10 +3497,9 @@ void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_SSE2
 
-#ifdef HAS_COPYROW_AVX2
 // Unaligned Multiple of 1.
 __declspec(naked) __declspec(align(16))
-void CopyRow_AVX2(const uint8* src, uint8* dst, int count) {
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
   __asm {
     mov        eax, esi
     mov        edx, edi
@@ -3513,7 +3512,6 @@ void CopyRow_AVX2(const uint8* src, uint8* dst, int count) {
     ret
   }
 }
-#endif  // HAS_COPYROW_AVX2
 
 #ifdef HAS_COPYROW_X86
 __declspec(naked) __declspec(align(16))