Use simple masking for AVX2 version of CopyAlpha so it can be implemented using…

Use simple masking for AVX2 version of CopyAlpha so it can be implemented using a more generic bit mask function in future, and use more broadly known and optimized opcodes that will always be fast.  Same performance as vblend.
BUG=none
TEST=CopyAlpha*
R=johannkoenig@google.com

Review URL: https://webrtc-codereview.appspot.com/2393005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@813 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 812
+Version: 813
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -168,7 +168,6 @@ extern "C" {
 // TODO(fbarchard): Optimize and enable
 // #define HAS_ARGBLUMACOLORTABLEROW_SSSE3
 #define HAS_ARGBCOPYALPHAROW_SSE2
-#define HAS_ARGBCOPYALPHAROW_SSE41
 
 // Caveat: Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
@@ -702,7 +701,6 @@ void CopyRow_C(const uint8* src, uint8* dst, int count);
 
 void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width);
 void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width);
-void ARGBCopyAlphaRow_SSE41(const uint8* src, uint8* dst, int width);
 void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width);
 
 void SetRow_X86(uint8* dst, uint32 v32, int count);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 812
+#define LIBYUV_VERSION 813
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -2195,11 +2195,6 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
     ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
   }
 #endif
-#if defined(HAS_ARGBCOPYALPHAROW_SSE41)
-  if (TestCpuFlag(kCpuHasSSE41) && IS_ALIGNED(width, 8)) {
-    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE41;
-  }
-#endif
 #if defined(HAS_ARGBCOPYALPHAROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
     ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;

source/row_win.cc

@@ -3640,35 +3640,6 @@ void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_ARGBCOPYALPHAROW_SSE2
 
-#ifdef HAS_ARGBCOPYALPHAROW_SSE41
-// width in pixels
-__declspec(naked) __declspec(align(16))
-void ARGBCopyAlphaRow_SSE41(const uint8* src, uint8* dst, int width) {
-  __asm {
-    mov        eax, [esp + 4]   // src
-    mov        edx, [esp + 8]   // dst
-    mov        ecx, [esp + 12]  // count
-    pcmpeqb    xmm0, xmm0       // generate mask 0x00ffffff
-    psrld      xmm0, 8
-
-    align      4
-  convertloop:
-    movdqu     xmm1, [eax]
-    movdqu     xmm2, [eax + 16]
-    lea        eax, [eax + 32]
-    pblendvb   xmm1, [edx], xmm0
-    pblendvb   xmm2, [edx + 16], xmm0
-    movdqu     [edx], xmm1
-    movdqu     [edx + 16], xmm2
-    lea        edx, [edx + 32]
-    sub        ecx, 8
-    jg         convertloop
-
-    ret
-  }
-}
-#endif  // HAS_ARGBCOPYALPHAROW_SSE41
-
 #ifdef HAS_ARGBCOPYALPHAROW_AVX2
 // width in pixels
 __declspec(naked) __declspec(align(16))
@@ -3677,18 +3648,21 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
     mov        eax, [esp + 4]   // src
     mov        edx, [esp + 8]   // dst
     mov        ecx, [esp + 12]  // count
-    vpcmpeqb   ymm0, ymm0, ymm0  // generate mask 0x00ffffff
-    vpsrld     ymm0, ymm0, 8
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm1, ymm0, 8    // generate mask 0x00ffffff
+    vpslld     ymm0, ymm0, 24   // generate mask 0xff000000
 
     align      4
   convertloop:
-    vmovdqu    ymm1, [eax]
-    vmovdqu    ymm2, [eax + 32]
+    vpand      ymm2, ymm0, [eax]
+    vpand      ymm3, ymm0, [eax + 32]
     lea        eax, [eax + 64]
-    vpblendvb  ymm1, ymm1, [edx], ymm0
-    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
-    vmovdqu    [edx], ymm1
-    vmovdqu    [edx + 32], ymm2
+    vpand      ymm4, ymm1, [edx]
+    vpand      ymm5, ymm1, [edx + 32]
+    vpor       ymm2, ymm2, ymm4
+    vpor       ymm3, ymm3, ymm5
+    vmovdqu    [edx], ymm2
+    vmovdqu    [edx + 32], ymm3
     lea        edx, [edx + 64]
     sub        ecx, 16
     jg         convertloop
@@ -6958,7 +6932,8 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     // 2 pixel loop.
     align      16
  convertloop:
-// (slow)   pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
     movq       xmm0, qword ptr [eax]  // BGRABGRA
     lea        eax, [eax + 8]
     punpcklbw  xmm0, xmm3