scale to 3/4 or 3/8 with odd width destinations efficiently. previously if…

scale to 3/4 or 3/8 with odd width destinations efficiently.  previously if width was not multiple of what the simd loop would do (24), scaling would fall back on slower C code.  This change allows SIMD to be used for most of the scaling and C for the remainder, improving efficiency.
BUG=314
TESTED=set LIBYUV_WIDTH=1896 & ScaleDownBy3by4_*
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/48249004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1380 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1378
+Version: 1379
 License: BSD
 License File: LICENSE
 

include/libyuv/scale_row.h

@@ -261,6 +261,22 @@ void ScaleRowDown4_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
                             uint8* dst_ptr, int dst_width);
 void ScaleRowDown4Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_Any_SSSE3(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);
@@ -358,6 +374,21 @@ void ScaleRowDown4_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                             uint8* dst_ptr, int dst_width);
 void ScaleRowDown4Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+// 32 -> 12
+void ScaleRowDown38_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
 
 void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                        uint16* dst_ptr, int src_width, int src_height);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1378
+#define LIBYUV_VERSION 1379
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/scale.cc

@@ -277,7 +277,15 @@ static void ScalePlaneDown34(int src_width, int src_height,
     ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
   }
 #if defined(HAS_SCALEROWDOWN34_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_NEON;
+    }
+    if (dst_width % 24 == 0) {
       if (!filtering) {
         ScaleRowDown34_0 = ScaleRowDown34_NEON;
         ScaleRowDown34_1 = ScaleRowDown34_NEON;
@@ -286,9 +294,18 @@ static void ScalePlaneDown34(int src_width, int src_height,
         ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
       }
     }
+  }
 #endif
 #if defined(HAS_SCALEROWDOWN34_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_SSSE3;
+    }
+    if (dst_width % 24 == 0) {
       if (!filtering) {
         ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
         ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
@@ -297,6 +314,7 @@ static void ScalePlaneDown34(int src_width, int src_height,
         ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
       }
     }
+  }
 #endif
 #if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
@@ -450,8 +468,17 @@ static void ScalePlaneDown38(int src_width, int src_height,
     ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
     ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
   }
+
 #if defined(HAS_SCALEROWDOWN38_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_NEON;
+    }
+    if (dst_width % 12 == 0) {
       if (!filtering) {
         ScaleRowDown38_3 = ScaleRowDown38_NEON;
         ScaleRowDown38_2 = ScaleRowDown38_NEON;
@@ -460,13 +487,22 @@ static void ScalePlaneDown38(int src_width, int src_height,
         ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
       }
     }
+  }
 #endif
 #if defined(HAS_SCALEROWDOWN38_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+  if (TestCpuFlag(kCpuHasSSSE3)) {
     if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_SSSE3;
+    }
+    if (dst_width % 12 == 0 && !filtering) {
       ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
       ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
-    } else {
+    }
+    if (dst_width % 6 == 0 && filtering) {
       ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
       ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
     }

source/scale_any.cc

@@ -87,6 +87,40 @@ SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7)
 SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C,
       4, 1, 7)
 #endif
+
+
+#ifdef HAS_SCALEROWDOWN34_SSSE3
+SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3, ScaleRowDown34_C,
+      3 / 4, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3,
+      ScaleRowDown34_0_Box_C, 3 / 4, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3,
+      ScaleRowDown34_1_Box_C, 3 / 4, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_NEON
+SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON, ScaleRowDown34_C,
+      3 / 4, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON,
+      ScaleRowDown34_0_Box_C, 3 / 4, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON,
+      ScaleRowDown34_1_Box_C, 3 / 4, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN38_SSSE3
+SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3, ScaleRowDown38_C,
+      3 / 4, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3,
+      ScaleRowDown38_3_Box_C, 3 / 4, 1, 5)
+SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3,
+      ScaleRowDown38_2_Box_C, 3 / 4, 1, 5)
+#endif
+#ifdef HAS_SCALEROWDOWN38_NEON
+SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON, ScaleRowDown38_C,
+      3 / 4, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON,
+      ScaleRowDown38_3_Box_C, 3 / 4, 1, 11)
+SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON,
+      ScaleRowDown38_2_Box_C, 3 / 4, 1, 11)
+#endif
 #undef SDANY
 
 // Fixed scale down.

source/scale_win.cc

@@ -400,7 +400,6 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
 // Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
 // Then shuffled to do the scaling.
 
-// Note that movdqa+palign may be better than movdqu.
 __declspec(naked)
 void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                           uint8* dst_ptr, int dst_width) {