initial I444ToARGB avx2 code

source/row_gcc.cc

Index: source/row_gcc.cc
diff --git a/source/row_gcc.cc b/source/row_gcc.cc
index 9940cba125655c238f8bdc4b72e9a2070882b73d..348935405761d16bf12ed18278fe277a1559d3a6 100644
--- a/source/row_gcc.cc
+++ b/source/row_gcc.cc
@@ -1947,6 +1947,19 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
 
 #endif  // HAS_I422TOARGBROW_SSSE3
 
+// Read 16 UV from 444
+#define READYUV444_AVX2                                                        \
+    "vmovdqu    " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
+    MEMOPREG(vmovdqu, 0x00, [u_buf], [v_buf], 1, xmm1)                         \
+    "lea        " MEMLEA(0x10, [u_buf]) ",%[u_buf]                  \n"        \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
+    "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n"        \
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
+    "vmovdqu    " MEMACCESS([y_buf]) ",%%xmm4                       \n"        \
+    "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n"        \
+    "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
+    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
+
 // Read 8 UV from 422, upsample to 16 UV.
 #define READYUV422_AVX2                                                        \
     "vmovq      " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
@@ -2079,6 +2092,39 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
     "vmovdqu    %%ymm0," MEMACCESS2(0x20, [dst_argb]) "             \n"        \
     "lea       " MEMLEA(0x40, [dst_argb]) ", %[dst_argb]            \n"
 
+#ifdef HAS_I444TOARGBROW_AVX2
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I444ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", NACL_R14 YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I444TOARGBROW_AVX2
+
 #if defined(HAS_I422TOARGBROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
@@ -2091,7 +2137,7 @@ void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
     "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5           \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422_AVX2