YUY2 and UYVY to ARGB AVX2 versions via wrappers.

BUG=403
TESTED=UNTESTED
R=harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/34339004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1291 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1286
+Version: 1291
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -181,6 +181,15 @@ extern "C" {
 #define VISUALC_HAS_AVX2 1
 #endif  // VisualStudio >= 2012
 
+// The following are available require VS2012.  Port to GCC.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+// TODO(fbarchard): fix AVX2 versions of YUV conversion.  bug=393
+#define HAS_I422TOABGRROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_I422TOBGRAROW_AVX2
+#define HAS_I422TORGBAROW_AVX2
+#endif
+
 // The following are available on all x86 platforms, but
 // require VS2012, clang 3.4 or gcc 4.7.
 // The code supports NaCL but requires a new compiler and validator.
@@ -207,6 +216,12 @@ extern "C" {
 #define HAS_YUY2TOUVROW_AVX2
 #define HAS_YUY2TOYROW_AVX2
 
+// The following require HAS_I422TOARGBROW_AVX2
+#if defined(HAS_I422TOARGBROW_AVX2)
+#define HAS_YUY2TOARGBROW_AVX2
+#define HAS_UYVYTOARGBROW_AVX2
+#endif
+
 // Effects:
 #define HAS_ARGBADDROW_AVX2
 #define HAS_ARGBATTENUATEROW_AVX2
@@ -215,14 +230,6 @@ extern "C" {
 #define HAS_ARGBUNATTENUATEROW_AVX2
 #endif
 
-// The following are available require VS2012.  Port to GCC.
-#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
-// TODO(fbarchard): fix AVX2 versions of YUV conversion.  bug=393
-#define HAS_I422TOABGRROW_AVX2
-#define HAS_I422TOARGBROW_AVX2
-#define HAS_I422TOBGRAROW_AVX2
-#define HAS_I422TORGBAROW_AVX2
-#endif
 
 // The following are Yasm x86 only:
 // TODO(fbarchard): Port AVX2 to inline.
@@ -358,6 +365,7 @@ extern "C" {
 
 #if defined(_MSC_VER) && !defined(__CLR_VER)
 #define SIMD_ALIGNED(var) __declspec(align(16)) var
+#define SIMD_ALIGNED32(var) __declspec(align(64)) var
 typedef __declspec(align(16)) int16 vec16[8];
 typedef __declspec(align(16)) int32 vec32[4];
 typedef __declspec(align(16)) int8 vec8[16];
@@ -374,6 +382,7 @@ typedef __declspec(align(32)) uint8 ulvec8[32];
 #elif defined(__GNUC__)
 // Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
 #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+#define SIMD_ALIGNED32(var) var __attribute__((aligned(64)))
 typedef int16 __attribute__((vector_size(16))) vec16;
 typedef int32 __attribute__((vector_size(16))) vec32;
 typedef int8 __attribute__((vector_size(16))) vec8;
@@ -388,6 +397,7 @@ typedef uint32 __attribute__((vector_size(32))) ulvec32;
 typedef uint8 __attribute__((vector_size(32))) ulvec8;
 #else
 #define SIMD_ALIGNED(var) var
+#define SIMD_ALIGNED32(var) var
 typedef int16 vec16[8];
 typedef int32 vec32[4];
 typedef int8 vec8[16];
@@ -1043,6 +1053,12 @@ void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
 void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
                          uint8* dst_argb,
                          int width);
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
 void J422ToARGBRow_SSSE3(const uint8* src_y,
                          const uint8* src_u,
                          const uint8* src_v,
@@ -1145,6 +1161,12 @@ void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
 void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
                              uint8* dst_argb,
                              int width);
+void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
 void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
                              const uint8* src_u,
                              const uint8* src_v,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1286
+#define LIBYUV_VERSION 1291
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/convert_argb.cc

@@ -843,7 +843,6 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
     src_stride_yuy2 = dst_stride_argb = 0;
   }
 #if defined(HAS_YUY2TOARGBROW_SSSE3)
-  // Posix is 16, Windows is 8.
   if (TestCpuFlag(kCpuHasSSSE3)) {
     YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
@@ -851,6 +850,14 @@ int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
     }
   }
 #endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_YUY2TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
@@ -893,7 +900,6 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
     src_stride_uyvy = dst_stride_argb = 0;
   }
 #if defined(HAS_UYVYTOARGBROW_SSSE3)
-  // Posix is 16, Windows is 8.
   if (TestCpuFlag(kCpuHasSSSE3)) {
     UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
@@ -901,6 +907,14 @@ int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
     }
   }
 #endif
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToARGBRow = UYVYToARGBRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_UYVYTOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     UYVYToARGBRow = UYVYToARGBRow_Any_NEON;

source/row_any.cc

@@ -162,27 +162,20 @@ RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
        3, 4, 2)
 #endif
 #if defined(HAS_I400TOARGBROW_SSE2)
-RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C,
+RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C, 7, 1, 4)
-       7, 1, 4)
 #endif
 #if defined(HAS_YTOARGBROW_SSE2)
-RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
+RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C, 7, 1, 4)
-       7, 1, 4)
 #endif
 #if defined(HAS_YTOARGBROW_AVX2)
-RGBANY(YToARGBRow_Any_AVX2, YToARGBRow_AVX2, YToARGBRow_C,
+RGBANY(YToARGBRow_Any_AVX2, YToARGBRow_AVX2, YToARGBRow_C, 15, 1, 4)
-       15, 1, 4)
 #endif
 #if defined(HAS_YUY2TOARGBROW_SSSE3)
-RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C,
+RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C, 15, 2, 4)
-       15, 2, 4)
+RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, UYVYToARGBRow_C, 15, 2, 4)
-RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, UYVYToARGBRow_C,
-       15, 2, 4)
-// These require alignment on ARGB, so C is used for remainder.
 RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
        15, 3, 4)
-RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
+RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C, 15, 3, 4)
-       15, 3, 4)
 RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
        7, 2, 4)
 RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
@@ -190,6 +183,10 @@ RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
 RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
        7, 2, 4)
 #endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+RGBANY(YUY2ToARGBRow_Any_AVX2, YUY2ToARGBRow_AVX2, YUY2ToARGBRow_C, 31, 2, 4)
+RGBANY(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, UYVYToARGBRow_C, 31, 2, 4)
+#endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
 RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
 RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
@@ -199,14 +196,10 @@ RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
        7, 4, 2)
 RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
        7, 4, 2)
-RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
+RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C, 7, 1, 4)
-       7, 1, 4)
+RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C, 7, 1, 4)
-RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
+RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C, 7, 2, 4)
-       7, 1, 4)
+RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C, 7, 2, 4)
-RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
-       7, 2, 4)
-RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
-       7, 2, 4)
 #endif
 #undef RGBANY
 

source/row_common.cc

@@ -2123,7 +2123,7 @@ void I422ToUYVYRow_C(const uint8* src_y,
 }
 
 // Maximum temporary width for wrappers to process at a time, in pixels.
-#define MAXTWIDTH 4096
+#define MAXTWIDTH 2048
 
 #if !defined(LIBYUV_DISABLE_X86) && defined(HAS_I422TOARGBROW_SSSE3)
 // row_win.cc has asm version, but GCC uses 2 step wrapper.
@@ -2250,6 +2250,40 @@ void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, uint8* dst_argb, int width) {
 #endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
 #endif  // !defined(LIBYUV_DISABLE_X86)
 
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV conversion.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_AVX2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_AVX2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV conversion.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_AVX2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_AVX2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
 void ARGBPolynomialRow_C(const uint8* src_argb,
                          uint8* dst_argb, const float* poly,
                          int width) {