MergeUV AVX2 use vextractf128 to store results to avoid shuffling.

BUG=none
TESTED=intel sde on unittests
R=brucedawson@google.com

Review URL: https://webrtc-codereview.appspot.com/33369004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1178 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1177
+Version: 1178
 License: BSD
 License File: LICENSE
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1177
+#define LIBYUV_VERSION 1178
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_posix.cc

@@ -2446,8 +2446,6 @@ void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
 }
 #endif  // HAS_SPLITUVROW_SSE2
 
-// TODO(fbarchard): Consider vpunpcklbw, vpunpckhbw, store-low1, store-low2,
-// extract-high1, extract-high2.
 #ifdef HAS_MERGEUVROW_AVX2
 void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
                      int width) {
@@ -2458,13 +2456,12 @@ void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
     "vmovdqu   " MEMACCESS(0) ",%%ymm0           \n"
     MEMOPREG(vmovdqu,0x00,0,1,1,ymm1)             //  vmovdqu (%0,%1,1),%%ymm1
     "lea       " MEMLEA(0x20,0) ",%0             \n"
-
     "vpunpcklbw %%ymm1,%%ymm0,%%ymm2             \n"
     "vpunpckhbw %%ymm1,%%ymm0,%%ymm0             \n"
-    "vperm2i128 $0x20,%%ymm0,%%ymm2,%%ymm1       \n"
-    "vperm2i128 $0x31,%%ymm0,%%ymm2,%%ymm2       \n"
-    "vmovdqu    %%ymm1," MEMACCESS(2) "          \n"
-    "vmovdqu    %%ymm2," MEMACCESS2(0x20,2) "    \n"
+    "vextractf128 $0x0,%%ymm2," MEMACCESS(2) "   \n"
+    "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n"
+    "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n"
+    "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n"
     "lea       " MEMLEA(0x40,2) ",%2             \n"
     "sub       $0x20,%3                          \n"
     "jg        1b                                \n"

source/row_win.cc

@@ -2686,10 +2686,10 @@ void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
     lea        eax,  [eax + 32]
     vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
     vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
-    vperm2i128 ymm1, ymm2, ymm0, 0x20  // low 128 of ymm2 and low 128 of ymm0
-    vperm2i128 ymm2, ymm2, ymm0, 0x31  // high 128 of ymm2 and high 128 of ymm0
-    vmovdqu    [edi], ymm1
-    vmovdqu    [edi + 32], ymm2
+    vextractf128 [edi], ymm2, 0       // bytes 0..15
+    vextractf128 [edi + 16], ymm0, 0  // bytes 16..31
+    vextractf128 [edi + 32], ymm2, 1  // bytes 32..47
+    vextractf128 [edi + 48], ymm0, 1  // bytes 47..63
     lea        edi, [edi + 64]
     sub        ecx, 32
     jg         convertloop