Handle scale down by factor of 2 efficiently by calling SIMD for multiple of 16…

Handle scale down by factor of 2 efficiently by calling SIMD for multiple of 16 destination pixels, and C for remainder.
BUG=314
TESTED=out\release\libyuv_unittest.exe --gtest_catch_exceptions=0 --gtest_filter=*.ScaleDownBy2*
R=bcornell@google.com

Review URL: https://webrtc-codereview.appspot.com/48689004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1344 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/scale_row.h

@@ -222,6 +222,12 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
 void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
                                 ptrdiff_t src_stride,
                                 uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
 void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
                        uint16* dst_ptr, int src_width,
                        int src_height);
@@ -271,10 +277,8 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
 // Note - not static due to reuse in convert for 444 to 420.
 void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst, int dst_width);
-
 void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                               uint8* dst, int dst_width);
-
 void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst, int dst_width);
 
@@ -309,6 +313,13 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
                                ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width);
 
+void ScaleRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleRowDown2Linear_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst, int dst_width);
+void ScaleRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+
 void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                        uint16* dst_ptr, int src_width, int src_height);
 

source/scale.cc

@@ -54,17 +54,27 @@ static void ScalePlaneDown2(int src_width, int src_height,
   }
 
 #if defined(HAS_SCALEROWDOWN2_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
-        ScaleRowDown2Box_NEON);
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_NEON :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON :
+        ScaleRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
+          ScaleRowDown2Box_NEON);
+    }
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
-        (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
-        ScaleRowDown2Box_SSE2);
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSE2 :
+        ScaleRowDown2Box_Any_SSE2);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)

source/scale_any.cc

@@ -35,6 +35,36 @@ CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
 #endif
 #undef CANY
 
+// Fixed scale down.
+#define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int n = dst_width & ~MASK;                                               \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                     dst_ptr + n * BPP, dst_width & MASK);                     \
+    }
+
+
+#ifdef HAS_SCALEROWDOWN2_SSE2
+SDANY(ScaleRowDown2_Any_SSE2, ScaleRowDown2_SSE2, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_SSE2, ScaleRowDown2Linear_SSE2,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_SSE2, ScaleRowDown2Box_SSE2,
+      ScaleRowDown2Box_C, 2, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN2_NEON
+SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_C, 2, 1, 15)
+#endif
+
+#undef SDANY
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

source/scale_win.cc

@@ -666,7 +666,6 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 }
 
 // Bilinear column filtering. SSSE3 version.
-// TODO(fbarchard): Port to Neon
 // TODO(fbarchard): Switch the following:
 //    xor        ebx, ebx
 //    mov        bx, word ptr [esi + eax]  // 2 source x0 pixels