From 6398e1d4ebc50935a29c3532abcc1a73568fa4fa Mon Sep 17 00:00:00 2001
From: "fbarchard@google.com"
 <fbarchard@google.com@16f28f9a-4ce2-e073-06de-1de4eb20be90>
Date: Wed, 11 Jul 2012 19:12:32 +0000
Subject: [PATCH] Port ARGBShade to Mac BUG=none TEST=none Review URL:
 https://webrtc-codereview.appspot.com/679009

git-svn-id: http://libyuv.googlecode.com/svn/trunk@299 16f28f9a-4ce2-e073-06de-1de4eb20be90
---
 README.chromium          |  2 +-
 include/libyuv/version.h |  2 +-
 source/row.h             |  2 +-
 source/row_posix.cc      | 37 +++++++++++++++++++++++++++++++++++++
 4 files changed, 40 insertions(+), 3 deletions(-)

diff --git a/README.chromium b/README.chromium
index 5928823f8..57582730e 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 298
+Version: 299
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index 3f3f947bd..8aabbb247 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 298
+#define LIBYUV_VERSION 299
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/row.h b/source/row.h
index bc59694a6..9b720b6da 100644
--- a/source/row.h
+++ b/source/row.h
@@ -80,12 +80,12 @@ extern "C" {
 #define HAS_ARGBQUANTIZEROW_SSE2
 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
 #define HAS_CUMULATIVESUMTOAVERAGE_SSE2
+#define HAS_ARGBSHADE_SSE2
 #endif
 
 // The following are Windows only:
 #if !defined(YUV_DISABLE_ASM) && defined(_M_IX86)
 #define HAS_ARGBCOLORTABLEROW_X86
-#define HAS_ARGBSHADE_SSE2
 #endif
 
 // The following are disabled when SSSE3 is available:
diff --git a/source/row_posix.cc b/source/row_posix.cc
index aa24149ca..0041dcece 100644
--- a/source/row_posix.cc
+++ b/source/row_posix.cc
@@ -3178,7 +3178,44 @@ void CumulativeSumToAverage_SSE2(const int32* topleft, const int32* botleft,
   );
 }
 #endif  // HAS_CUMULATIVESUMTOAVERAGE_SSE2
+#ifdef HAS_ARGBSHADE_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "sub       %0,%1                           \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
 
+    // 4 pixel loop.
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movdqa    (%0),%%xmm0                     \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0,(%0,%1,1)                \n"
+    "lea       0x10(%0),%0                     \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHADE_SSE2
 
 #endif  // defined(__x86_64__) || defined(__i386__)