Neon version of Gray

BUG=176
TEST=./libyuv_unittest --gtest_filter=**Gray*
Review URL: https://webrtc-codereview.appspot.com/929039

git-svn-id: http://libyuv.googlecode.com/svn/trunk@510 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 509
+Version: 510
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -224,6 +224,8 @@ extern "C" {
 #define HAS_ARGBATTENUATEROW_NEON
 #define HAS_ARGBQUANTIZEROW_NEON
 #define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
 #endif
 
 // The following are available on Mips platforms
@@ -1218,9 +1220,11 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
 
 void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
 
 void ARGBSepiaRow_C(uint8* dst_argb, int width);
 void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
 
 void ARGBColorMatrixRow_C(uint8* dst_argb, const int8* matrix_argb, int width);
 void ARGBColorMatrixRow_SSSE3(uint8* dst_argb, const int8* matrix_argb,
@@ -1236,6 +1240,13 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width);
 
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
 // Used for blur.
 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
                                     int width, int area, uint8* dst, int count);
@@ -1247,12 +1258,6 @@ void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
 void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
                                const int32* previous_cumsum, int width);
 
-void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
-                    uint32 value);
-void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value);
-void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
-                       uint32 value);
 
 LIBYUV_API
 void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 509
+#define LIBYUV_VERSION 510
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -856,6 +856,10 @@ int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
     ARGBGrayRow = ARGBGrayRow_SSSE3;
   }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
 #endif
 
   for (int y = 0; y < height; ++y) {
@@ -881,6 +885,10 @@ int ARGBGray(uint8* dst_argb, int dst_stride_argb,
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
     ARGBGrayRow = ARGBGrayRow_SSSE3;
   }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
 #endif
   uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
   for (int y = 0; y < height; ++y) {

source/row_common.cc

@@ -563,7 +563,7 @@ void ARGBToUV411Row_C(const uint8* src_argb,
 // 0.11 * B + 0.59 * G + 0.30 * R
 // Coefficients rounded to multiple of 2 for consistency with SSSE3 version.
 static __inline int RGBToGray(uint8 r, uint8 g, uint8 b) {
-  return (( 76 * r + 152 * g +  28 * b) >> 8);
+  return (28 * b + 152 * g + 76 * r) >> 8;
 }
 
 void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {

source/row_neon.cc

@@ -2550,6 +2550,34 @@ void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
   );
 }
 
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToY but different constants, no round and stores ARGB.
+// C code is (28 * b + 152 * g + 76 * r) >> 8;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #14                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #76                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.2578 coefficient
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqshrun.s16 d0, q2, #7                    \n"  // 16 bit to 8 bit Y
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
 #endif  // __ARM_NEON__
 
 #ifdef __cplusplus