From 7892ea1fe14533d197f11a4f720d877eebaeb0b8 Mon Sep 17 00:00:00 2001
From: "fbarchard@google.com" <fbarchard@google.com>
Date: Mon, 15 Dec 2014 18:59:23 +0000
Subject: [PATCH] Fix for ARGBToUV on AVX2 BUG=269 TESTED=local testing
 R=harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/33669004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1202 16f28f9a-4ce2-e073-06de-1de4eb20be90
---
 README.chromium          |  2 +-
 include/libyuv/version.h |  2 +-
 source/row_posix.cc      |  2 ++
 source/row_win.cc        | 12 +++++++++---
 4 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/README.chromium b/README.chromium
index edfeedc11..8680ab393 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1201
+Version: 1203
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index 9ceeff62c..8cc1b5ed0 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 1201
+#define LIBYUV_VERSION 1203
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/row_posix.cc b/source/row_posix.cc
index 78997edc2..38886c15d 100644
--- a/source/row_posix.cc
+++ b/source/row_posix.cc
@@ -983,6 +983,7 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
     "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
     "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
+
     "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
     "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
     "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
@@ -995,6 +996,7 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
     "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
     "vpshufb    %8,%%ymm0,%%ymm0               \n"
     "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
+
     "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
     VEXTOPMEM(vextractf128,ymm0,0x1,1,2,1) // vextractf128 $0x1,%%ymm0,(%1,%2,1)
     "lea       " MEMLEA(0x10,1) ",%1           \n"
diff --git a/source/row_win.cc b/source/row_win.cc
index 4996de00f..31af591c9 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -1473,7 +1473,6 @@ void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
 }
 #endif  // HAS_ARGBTOYROW_SSSE3
 
-
 #if defined(HAS_I422TOARGBROW_AVX2) || defined(HAS_I422TOBGRAROW_AVX2)
 static const lvec8 kUVToB_AVX = {
   UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
@@ -1502,6 +1501,7 @@ static const lvec16 kUVBiasG_AVX = {
 static const lvec16 kUVBiasR_AVX = {
   BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR
 };
+#endif  // defined(HAS_I422TOARGBROW_AVX2) || defined(HAS_I422TOBGRAROW_AVX2)
 
 // Read 8 UV from 422, upsample to 16 UV.
 #define READYUV422_AVX2 __asm {                                                \
@@ -1540,7 +1540,7 @@ static const lvec16 kUVBiasR_AVX = {
     __asm vpackuswb  ymm2, ymm2, ymm2           /* R */                        \
   }
 
-#if defined(HAS_I422TOARGBROW_AVX2)
+#ifdef HAS_I422TOARGBROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
 __declspec(naked) __declspec(align(16))
@@ -1584,7 +1584,9 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
     ret
   }
 }
+#endif  // HAS_I422TOARGBROW_AVX2
 
+#ifdef HAS_I422TOBGRAROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
 // TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
@@ -1629,7 +1631,9 @@ void I422ToBGRARow_AVX2(const uint8* y_buf,
     ret
   }
 }
+#endif  // HAS_I422TOBGRAROW_AVX2
 
+#ifdef HAS_I422TORGBAROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
 // TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
@@ -1674,7 +1678,9 @@ void I422ToRGBARow_AVX2(const uint8* y_buf,
     ret
   }
 }
+#endif  // HAS_I422TORGBAROW_AVX2
 
+#ifdef HAS_I422TOABGRROW_AVX2
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
 // TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
@@ -1719,7 +1725,7 @@ void I422ToABGRRow_AVX2(const uint8* y_buf,
     ret
   }
 }
-#endif  // HAS_I422TOARGBROW_AVX2
+#endif  // HAS_I422TOABGRROW_AVX2
 
 #ifdef HAS_I422TOARGBROW_SSSE3
 // TODO(fbarchard): Read that does half size on Y and treats 420 as 444.