Port I411ToARGBRow to AVX2.

source/row_gcc.cc

 // VERSION 2
 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
@@ skipping 1413 matching lines @@
     "m"(kRGBAToV),  // %5
     "m"(kRGBAToU),  // %6
     "m"(kAddUV128)  // %7
   : "memory", "cc", NACL_R14
     "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
   );
 }
 
 #if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2)
 
-// Read 8 UV from 411
+// Read 8 UV from 444
 #define READYUV444                                                             \
     "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
     MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
     "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
     "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
     "movq       " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
     "punpcklbw  %%xmm4,%%xmm4                                   \n"            \
     "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"
 
 // Read 4 UV from 422, upsample to 8 UV
@@ skipping 500 matching lines @@
     "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
     "vpunpcklwd %%ymm0,%%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"        \
     "vmovdqu    " MEMACCESS([a_buf]) ",%%xmm5                       \n"        \
     "vpermq     $0xd8,%%ymm5,%%ymm5                                 \n"        \
     "lea        " MEMLEA(0x10, [a_buf]) ",%[a_buf]                  \n"
 
+// Read 4 UV from 411, upsample to 16 UV.
+#define READYUV411_AVX2                                                        \
+    "vmovd      " MEMACCESS([u_buf]) ",%%xmm0                       \n"        \
+    MEMOPREG(vmovd, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]                   \n"        \
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
+    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"        \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
+    "vpunpckldq %%ymm0,%%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 NV12, upsample to 16 UV.
 #define READNV12_AVX2                                                          \
     "vmovdqu    " MEMACCESS([uv_buf]) ",%%xmm0                      \n"        \
     "lea        " MEMLEA(0x10, [uv_buf]) ",%[uv_buf]                \n"        \
     "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
     "vpunpcklwd %%ymm0,%%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"
@@ skipping 95 matching lines @@
 // 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"
+    "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
 
+#ifdef HAS_I411TOARGBROW_AVX2
+// 16 pixels
+// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I411ToARGBRow_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"
+    READYUV411_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_I411TOARGBROW_AVX2
+
 #if defined(HAS_I422TOARGBROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
 void OMITFP I422ToARGBRow_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"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
     LABELALIGN
   "1:                                          \n"
     READYUV422_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]
@@ skipping 3276 matching lines @@
   );
 }
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif