point sample 64 bit column filter

BUG=302
TESTED=ARGBScaleClipTo320x240_None etc
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/6539004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@946 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 942
+Version: 946
 License: BSD
 License File: LICENSE
 

include/libyuv/scale_row.h

@@ -140,6 +140,8 @@ void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
                                uint8* dst_argb, int dst_width);
 void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
                      int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
 void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
                         int dst_width, int, int);
 void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 942
+#define LIBYUV_VERSION 946
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/scale_argb.cc

@@ -322,8 +322,9 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
   void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
       int dst_width, int x, int dx) =
       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
-  if (filtering && src_width >= 32768) {
-    ScaleARGBFilterCols = ScaleARGBFilterCols64_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
   }
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -508,8 +509,9 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
   void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
       int dst_width, int x, int dx) =
       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
-  if (filtering && src_width >= 32768) {
-    ScaleARGBFilterCols = ScaleARGBFilterCols64_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
   }
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
@@ -622,9 +624,10 @@ static void ScaleARGBSimple(int src_width, int src_height,
                             const uint8* src_argb, uint8* dst_argb,
                             int x, int dx, int y, int dy) {
   void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
-      int dst_width, int x, int dx) = ScaleARGBCols_C;
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
 #if defined(HAS_SCALEARGBCOLS_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
+  if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBCols = ScaleARGBCols_SSE2;
   }
 #endif

source/scale_common.cc

@@ -427,6 +427,23 @@ void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
   }
 }
 
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = static_cast<int64>(x32);
+  const uint32* src = reinterpret_cast<const uint32*>(src_argb);
+  uint32* dst = reinterpret_cast<uint32*>(dst_argb);
+  for (int j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
 // Scales a single row of pixels up by 2x using point sampling.
 void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
                         int dst_width, int, int) {