I420ToYUY2_AVX2 port

I420 and I422 To YUY2 and UYVY ported from SSE2 to AVX2.

Was SSE2
I420ToYUY2_Opt (135 ms)
I420ToUYVY_Opt (148 ms)
I422ToYUY2_Opt (145 ms)
I422ToUYVY_Opt (142 ms)

Now AVX2
I420ToYUY2_Opt (133 ms)
I420ToUYVY_Opt (130 ms)
I422ToYUY2_Opt (127 ms)
I422ToUYVY_Opt (137 ms)

Bug: libyuv:556
Test: out/Release/libyuv_unittest --sandbox_unittests --gtest_filter=*I42?To*UY*Opt
Change-Id: Ic35f97cee02dc009fd98785589ba17c7cf50bb35
Reviewed-on: https://chromium-review.googlesource.com/892493
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: richard winterton <rrwinterton@gmail.com>

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1694
+Version: 1695
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -276,6 +276,8 @@ extern "C" {
 #define HAS_I210TOARGBROW_AVX2
 #define HAS_I210TOAR30ROW_AVX2
 #define HAS_I422TOAR30ROW_AVX2
+#define HAS_I422TOUYVYROW_AVX2
+#define HAS_I422TOYUY2ROW_AVX2
 #define HAS_MERGEUVROW_16_AVX2
 #define HAS_MULTIPLYROW_16_AVX2
 #endif
@@ -2412,8 +2414,12 @@ void ARGBToARGB1555Row_Any_SSE2(const uint8_t* src_argb,
 void ARGBToARGB4444Row_Any_SSE2(const uint8_t* src_argb,
                                 uint8_t* dst_rgb,
                                 int width);
-void ABGRToAR30Row_Any_SSSE3(const uint8_t* src_abgr, uint8_t* dst_ar30, int width);
-void ARGBToAR30Row_Any_SSSE3(const uint8_t* src_argb, uint8_t* dst_ar30, int width);
+void ABGRToAR30Row_Any_SSSE3(const uint8_t* src_abgr,
+                             uint8_t* dst_ar30,
+                             int width);
+void ARGBToAR30Row_Any_SSSE3(const uint8_t* src_argb,
+                             uint8_t* dst_ar30,
+                             int width);
 
 void ARGBToRGB565DitherRow_Any_SSE2(const uint8_t* src_argb,
                                     uint8_t* dst_rgb,
@@ -2433,8 +2439,12 @@ void ARGBToARGB1555Row_Any_AVX2(const uint8_t* src_argb,
 void ARGBToARGB4444Row_Any_AVX2(const uint8_t* src_argb,
                                 uint8_t* dst_rgb,
                                 int width);
-void ABGRToAR30Row_Any_AVX2(const uint8_t* src_abgr, uint8_t* dst_ar30, int width);
-void ARGBToAR30Row_Any_AVX2(const uint8_t* src_argb, uint8_t* dst_ar30, int width);
+void ABGRToAR30Row_Any_AVX2(const uint8_t* src_abgr,
+                            uint8_t* dst_ar30,
+                            int width);
+void ARGBToAR30Row_Any_AVX2(const uint8_t* src_argb,
+                            uint8_t* dst_ar30,
+                            int width);
 
 void ARGBToRGB24Row_Any_NEON(const uint8_t* src_argb,
                              uint8_t* dst_rgb,
@@ -2840,6 +2850,26 @@ void I422ToUYVYRow_Any_SSE2(const uint8_t* src_y,
                             const uint8_t* src_v,
                             uint8_t* dst_uyvy,
                             int width);
+void I422ToYUY2Row_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
+void I422ToUYVYRow_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
+void I422ToYUY2Row_Any_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_yuy2,
+                            int width);
+void I422ToUYVYRow_Any_AVX2(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_uyvy,
+                            int width);
 void I422ToYUY2Row_NEON(const uint8_t* src_y,
                         const uint8_t* src_u,
                         const uint8_t* src_v,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1694
+#define LIBYUV_VERSION 1695
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/convert_from.cc

@@ -219,6 +219,14 @@ int I422ToYUY2(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -270,6 +278,14 @@ int I420ToYUY2(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -341,6 +357,14 @@ int I422ToUYVY(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
@@ -400,6 +424,14 @@ int I420ToUYVY(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;

source/convert_from_argb.cc

@@ -583,6 +583,14 @@ int ARGBToYUY2(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOYUY2ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOYUY2ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
@@ -712,6 +720,14 @@ int ARGBToUYVY(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_I422TOUYVYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOUYVYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToUYVYRow = I422ToUYVYRow_Any_NEON;

source/row_any.cc

@@ -96,6 +96,10 @@ ANY31(MergeRGBRow_Any_NEON, MergeRGBRow_NEON, 0, 0, 3, 15)
 ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15)
 ANY31(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, 1, 1, 4, 15)
 #endif
+#ifdef HAS_I422TOYUY2ROW_AVX2
+ANY31(I422ToYUY2Row_Any_AVX2, I422ToYUY2Row_AVX2, 1, 1, 4, 31)
+ANY31(I422ToUYVYRow_Any_AVX2, I422ToUYVYRow_AVX2, 1, 1, 4, 31)
+#endif
 #ifdef HAS_I422TOYUY2ROW_NEON
 ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15)
 #endif

source/row_gcc.cc

@@ -6041,6 +6041,88 @@ void I422ToUYVYRow_SSE2(const uint8_t* src_y,
 }
 #endif  // HAS_I422TOUYVYROW_SSE2
 
+#ifdef HAS_I422TOYUY2ROW_AVX2
+// TODO(fbarchard): Consider vmovhps to avoid vpermq
+
+void I422ToYUY2Row_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_frame,
+                        int width) {
+  asm volatile(
+
+      "sub       %1,%2                             \n"
+
+      LABELALIGN
+      "1:                                          \n"
+      "vmovdqu    (%1),%%xmm2                      \n"
+      "vmovdqu    0x00(%1,%2,1),%%xmm3             \n"
+      "lea        0x10(%1),%1                      \n"
+      "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+      "vpermq     $0xd8,%%ymm3,%%ymm3              \n"
+      "vpunpcklbw %%ymm3,%%ymm2,%%ymm2             \n"
+      "vmovdqu    (%0),%%ymm0                      \n"
+      "lea        0x20(%0),%0                      \n"
+      "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+      "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
+      "vpunpckhbw %%ymm2,%%ymm0,%%ymm1             \n"
+      "vpunpcklbw %%ymm2,%%ymm0,%%ymm0             \n"
+      "vmovdqu    %%ymm0,(%3)                      \n"
+      "vmovdqu    %%ymm1,0x20(%3)                  \n"
+      "lea        0x40(%3),%3                      \n"
+      "sub        $0x20,%4                         \n"
+      "jg         1b                               \n"
+      "vzeroupper                                  \n"
+      : "+r"(src_y),      // %0
+        "+r"(src_u),      // %1
+        "+r"(src_v),      // %2
+        "+r"(dst_frame),  // %3
+        "+rm"(width)      // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif  // HAS_I422TOYUY2ROW_AVX2
+
+#ifdef HAS_I422TOUYVYROW_AVX2
+void I422ToUYVYRow_AVX2(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_frame,
+                        int width) {
+  asm volatile(
+
+      "sub        %1,%2                            \n"
+
+      LABELALIGN
+      "1:                                          \n"
+      "vmovdqu    (%1),%%xmm2                      \n"
+      "vmovdqu    0x00(%1,%2,1),%%xmm3             \n"
+      "lea        0x10(%1),%1                      \n"
+      "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+      "vpermq     $0xd8,%%ymm3,%%ymm3              \n"
+      "vpunpcklbw %%ymm3,%%ymm2,%%ymm2             \n"
+      "vmovdqu    (%0),%%ymm0                      \n"
+      "lea        0x20(%0),%0                      \n"
+      "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+      "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
+      "vpunpcklbw %%ymm0,%%ymm2,%%ymm1             \n"
+      "vpunpckhbw %%ymm0,%%ymm2,%%ymm2             \n"
+      "vmovdqu    %%ymm1,(%3)                      \n"
+      "vmovdqu    %%ymm2,0x20(%3)                  \n"
+      "lea        0x40(%3),%3                      \n"
+      "sub        $0x20,%4                         \n"
+      "jg         1b                               \n"
+      "vzeroupper                                  \n"
+      : "+r"(src_y),      // %0
+        "+r"(src_u),      // %1
+        "+r"(src_v),      // %2
+        "+r"(dst_frame),  // %3
+        "+rm"(width)      // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif  // HAS_I422TOUYVYROW_AVX2
+
 #ifdef HAS_ARGBPOLYNOMIALROW_SSE2
 void ARGBPolynomialRow_SSE2(const uint8_t* src_argb,
                             uint8_t* dst_argb,

source/scale_gcc.cc

@@ -1324,6 +1324,7 @@ void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
       "movd      %%xmm0,(%0)                     \n"
 
       LABELALIGN "99:                            \n"  // clang-format error.
+
       : "+r"(dst_argb),    // %0
         "+r"(src_argb),    // %1
         "+rm"(dst_width),  // %2