Neon version of new SetRow functions for rectangles.

BUG=387
TESTED=untested
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/39449004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1220 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1219
+Version: 1220
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -302,7 +302,7 @@ extern "C" {
 #define HAS_RGB565TOYROW_NEON
 #define HAS_RGBATOUVROW_NEON
 #define HAS_RGBATOYROW_NEON
-#define HAS_SETROW_NEON
+// #define HAS_SETROW_NEON
 #define HAS_SPLITUVROW_NEON
 #define HAS_UYVYTOARGBROW_NEON
 #define HAS_UYVYTOUV422ROW_NEON
@@ -332,7 +332,7 @@ extern "C" {
 #define HAS_SOBELXYROW_NEON
 #define HAS_SOBELYROW_NEON
 #define HAS_ARGBCOLORMATRIXROW_NEON
-#define HAS_ARGBSETROWS_NEON
+// #define HAS_ARGBSETROWS_NEON
 #define HAS_ARGBSHUFFLEROW_NEON
 #endif
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1219
+#define LIBYUV_VERSION 1220
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_neon.cc

@@ -845,10 +845,10 @@ void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
   );
 }
 
-// SetRow8 writes 'count' bytes using a 32 bit value repeated.
-void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint32 v8, int count) {
   asm volatile (
-    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+    "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
   "1:                                          \n"
     "subs      %1, %1, #16                     \n"  // 16 bytes per loop
     MEMACCESS(0)
@@ -856,19 +856,25 @@ void SetRow_NEON(uint8* dst, uint32 v32, int count) {
     "bgt       1b                              \n"
   : "+r"(dst),   // %0
     "+r"(count)  // %1
-  : "r"(v32)     // %2
+  : "r"(v8)      // %2
   : "cc", "memory", "q0"
   );
 }
 
-// TODO(fbarchard): Make fully assembler
-// SetRow32 writes 'count' words using a 32 bit value repeated.
-void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
-                      int dst_stride, int height) {
-  for (int y = 0; y < height; ++y) {
-    SetRow_NEON(dst, v32, width << 2);
-    dst += dst_stride;
-  }
+// ARGBSetRow writes 'count' pixels using an 32 bit value repeated.
+void ARGBSetRow_NEON(uint8* dst, uint32 v8, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 pixels per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
 }
 
 void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {

source/row_neon64.cc

@@ -735,11 +735,10 @@ void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_NEON
 
-// SetRow8 writes 'count' bytes using a 32 bit value repeated.
-#ifdef HAS_SETROW_NEON
-void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint32 v8, int count) {
   asm volatile (
-    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+    "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
   "1:                                          \n"
     "subs      %1, %1, #16                     \n"  // 16 bytes per loop
     MEMACCESS(0)
@@ -747,23 +746,25 @@ void SetRow_NEON(uint8* dst, uint32 v32, int count) {
     "b.gt      1b                              \n"
   : "+r"(dst),   // %0
     "+r"(count)  // %1
-  : "r"(v32)     // %2
+  : "r"(v8)      // %2
   : "cc", "memory", "v0"
   );
 }
-#endif  // HAS_SETROW_NEON
 
-// TODO(fbarchard): Make fully assembler
-// SetRow32 writes 'count' words using a 32 bit value repeated.
-#ifdef HAS_ARGBSETROWS_NEON
-void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
-                      int dst_stride, int height) {
-  for (int y = 0; y < height; ++y) {
-    SetRow_NEON(dst, v32, width << 2);
-    dst += dst_stride;
-  }
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 ints per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "v0"
+  );
 }
-#endif  // HAS_ARGBSETROWS_NEON
 
 #ifdef HAS_MIRRORROW_NEON
 void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {