port I444 to ARGB to matrix. Add I444 to ABGR.

R=harryjin@google.com
BUG=libyuv:488,libyuv:490

Review URL: https://codereview.chromium.org/1348763005 .

include/libyuv/row.h

@@ -117,7 +117,6 @@ extern "C" {
 #define HAS_I422TORGBAROW_SSSE3
 #define HAS_I422TOUYVYROW_SSE2
 #define HAS_I422TOYUY2ROW_SSE2
-#define HAS_I444TOARGBROW_SSSE3
 #define HAS_J400TOARGBROW_SSE2
 #define HAS_J422TOARGBROW_SSSE3
 #define HAS_J422TOABGRROW_SSSE3
@@ -150,6 +149,10 @@ extern "C" {
 #define HAS_YUY2TOUV422ROW_SSE2
 #define HAS_YUY2TOUVROW_SSE2
 #define HAS_YUY2TOYROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_I444TOABGRROW_SSSE3
+#define HAS_I444TOARGBMATRIXROW_SSSE3
+#define HAS_I444TOABGRMATRIXROW_SSSE3
 
 // Effects:
 #define HAS_ARGBADDROW_SSE2
@@ -192,15 +195,6 @@ extern "C" {
 #define HAS_I422TOABGRMATRIXROW_SSSE3
 #endif
 
-// The following are available for Visual C and clangcl 32 bit:
-// TODO(fbarchard): Port to gcc.
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
-    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
-#define HAS_I444TOABGRROW_SSSE3
-#define HAS_I444TOARGBMATRIXROW_SSSE3
-#define HAS_I444TOABGRMATRIXROW_SSSE3
-#endif
-
 // The following are available for AVX2 Visual C and clangcl 32 bit:
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \

source/row_gcc.cc

@@ -1584,18 +1584,19 @@ YuvConstants SIMD_ALIGNED(kYuvHConstants) = {
     "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_rgba]) "           \n"           \
     "lea       " MEMLEA(0x20, [dst_rgba]) ",%[dst_rgba]          \n"
 
-void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
-                                const uint8* u_buf,
-                                const uint8* v_buf,
-                                uint8* dst_argb,
-                                int width) {
+void OMITFP I444ToARGBMatrixRow_SSSE3(const uint8* y_buf,
+                                      const uint8* u_buf,
+                                      const uint8* v_buf,
+                                      uint8* dst_argb,
+                                      struct YuvConstants* YuvConstants,
+                                      int width) {
   asm volatile (
     "sub       %[u_buf],%[v_buf]               \n"
     "pcmpeqb   %%xmm5,%%xmm5                   \n"
     LABELALIGN
   "1:                                          \n"
     READYUV444
-    YUVTORGB(kYuvConstants)
+    YUVTORGB(YuvConstants)
     STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
@@ -1604,7 +1605,34 @@ void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : [YuvConstants]"r"(YuvConstants)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I444ToABGRMatrixRow_SSSE3(const uint8* y_buf,
+                                      const uint8* u_buf,
+                                      const uint8* v_buf,
+                                      uint8* dst_abgr,
+                                      struct YuvConstants* YuvConstants,
+                                      int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB(YuvConstants)
+    STOREABGR
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [YuvConstants]"r"(YuvConstants)  // %[kYuvConstants]
   : "memory", "cc", NACL_R14
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
@@ -1711,7 +1739,7 @@ void OMITFP I422ToARGBMatrixRow_SSSE3(const uint8* y_buf,
     LABELALIGN
   "1:                                          \n"
     READYUV422
-    YUVTORGB(kYuvConstants)
+    YUVTORGB(YuvConstants)
     STOREARGB
     "sub       $0x8,%[width]                   \n"
     "jg        1b                              \n"
@@ -1720,7 +1748,7 @@ void OMITFP I422ToARGBMatrixRow_SSSE3(const uint8* y_buf,
     [v_buf]"+r"(v_buf),    // %[v_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
-  : [kYuvConstants]"r"(YuvConstants)  // %[kYuvConstants]
+  : [YuvConstants]"r"(YuvConstants)  // %[kYuvConstants]
   : "memory", "cc", NACL_R14
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );