Quantize for Neon

BUG=167
TEST=./libyuv_unittest --gtest_filter=*Quantize
Review URL: https://webrtc-codereview.appspot.com/972010

git-svn-id: http://libyuv.googlecode.com/svn/trunk@508 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 507
+Version: 508
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -222,6 +222,7 @@ extern "C" {
 #define HAS_ARGBINTERPOLATEROW_NEON
 #define HAS_ARGBBLENDROW_NEON
 #define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
 #endif
 
 // The following are available on Mips platforms
@@ -1231,6 +1232,8 @@ void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
                        int interval_offset, int width);
 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
 
 // Used for blur.
 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 507
+#define LIBYUV_VERSION 508
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -984,6 +984,10 @@ int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
     ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
   }
+#elif defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
 #endif
   uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
   for (int y = 0; y < height; ++y) {

source/row_neon.cc

@@ -2473,6 +2473,47 @@ void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
 }
 #endif
 
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
 #endif  // __ARM_NEON__
 
 #ifdef __cplusplus