rotate nv12 any width

BUG=libyuv:464
R=harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/55709004.

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1461
+Version: 1462
 License: BSD
 License File: LICENSE
 

include/libyuv/rotate_row.h

@@ -59,7 +59,7 @@ extern "C" {
     defined(__mips__) && \
     defined(__mips_dsp) && (__mips_dsp_rev >= 2)
 #define HAS_TRANSPOSEWX8_MIPS_DSPR2
-#define HAS_TRANSPOSEUVWx8_MIPS_DSPR2
+#define HAS_TRANSPOSEUVWX8_MIPS_DSPR2
 #endif  // defined(__mips__)
 
 void TransposeWxH_C(const uint8* src, int src_stride,
@@ -103,6 +103,16 @@ void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
                                uint8* dst_a, int dst_stride_a,
                                uint8* dst_b, int dst_stride_b, int width);
 
+void TransposeUVWx8_Any_SSE2(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_Any_NEON(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_Any_MIPS_DSPR2(const uint8* src, int src_stride,
+                                   uint8* dst_a, int dst_stride_a,
+                                   uint8* dst_b, int dst_stride_b, int width);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1461
+#define LIBYUV_VERSION 1462
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/rotate.cc

@@ -204,11 +204,14 @@ void TransposeUV(const uint8* src, int src_stride,
   }
 #endif
 #if defined(HAS_TRANSPOSEUVWX8_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
-    TransposeUVWx8 = TransposeUVWx8_SSE2;
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    TransposeUVWx8 = TransposeUVWx8_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx8 = TransposeUVWx8_SSE2;
+    }
   }
 #endif
-#if defined(HAS_TRANSPOSEUVWx8_MIPS_DSPR2)
+#if defined(HAS_TRANSPOSEUVWX8_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
       IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
     TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;

source/rotate_any.cc

@@ -18,7 +18,7 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#define TANY(NAMEANY, TPOS_SIMD, TPOS_C, MASK)                                 \
+#define TANY(NAMEANY, TPOS_SIMD, MASK)                                         \
     void NAMEANY(const uint8* src, int src_stride,                             \
                  uint8* dst, int dst_stride, int width) {                      \
       int r = width & MASK;                                                    \
@@ -26,24 +26,49 @@ extern "C" {
       if (n > 0) {                                                             \
         TPOS_SIMD(src, src_stride, dst, dst_stride, n);                        \
       }                                                                        \
-      TPOS_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r);        \
+      TransposeWx8_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r);\
     }
 
 #ifdef HAS_TRANSPOSEWX8_NEON
-TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, TransposeWx8_C, 7)
+TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, 7)
 #endif
 #ifdef HAS_TRANSPOSEWX8_SSSE3
-TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, TransposeWx8_C, 7)
+TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, 7)
 #endif
 #ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
-TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, TransposeWx8_C, 15)
+TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15)
 #endif
 #ifdef HAS_TRANSPOSEWX8_MIPS_DSPR2
-TANY(TransposeWx8_Any_MIPS_DSPR2, TransposeWx8_MIPS_DSPR2, TransposeWx8_C, 7)
+TANY(TransposeWx8_Any_MIPS_DSPR2, TransposeWx8_MIPS_DSPR2, 7)
 #endif
-
 #undef TANY
 
+#define TUVANY(NAMEANY, TPOS_SIMD, MASK)                                       \
+    void NAMEANY(const uint8* src, int src_stride,                             \
+                uint8* dst_a, int dst_stride_a,                                \
+                uint8* dst_b, int dst_stride_b, int width) {                   \
+      int r = width & MASK;                                                    \
+      int n = width - r;                                                       \
+      if (n > 0) {                                                             \
+        TPOS_SIMD(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,   \
+                  n);                                                          \
+      }                                                                        \
+      TransposeUVWx8_C(src + n * 2, src_stride,                                \
+                       dst_a + n * dst_stride_a, dst_stride_a,                 \
+                       dst_b + n * dst_stride_b, dst_stride_b, r);             \
+    }
+
+#ifdef HAS_TRANSPOSEUVWX8_NEON
+TUVANY(TransposeUVWx8_Any_NEON, TransposeUVWx8_NEON, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX8_SSE2
+TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX8_MIPS_DSPR2
+TUVANY(TransposeUVWx8_Any_MIPS_DSPR2, TransposeUVWx8_MIPS_DSPR2, 7)
+#endif
+#undef TUVANY
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv