port ARGB to 565 dithering AVX2 code to GCC.

Previously the assembly code was only available to Windows.
This CL ports the AVX2 code to GCC syntax.

TBR=harryjin@google.com
BUG=libyuv:492

Review URL: https://codereview.chromium.org/1391273003 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1503
+Version: 1504
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -182,36 +182,6 @@ extern "C" {
 #define HAS_SOBELYROW_SSE2
 #endif
 
-// The following are also available on x64 Visual C.
-#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
-    (!defined(__clang__) || defined(__SSSE3__))
-#define HAS_I422ALPHATOABGRROW_SSSE3
-#define HAS_I422ALPHATOARGBROW_SSSE3
-#define HAS_I422TOABGRROW_SSSE3
-#define HAS_I422TOARGBROW_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) && \
-    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
-#define HAS_ARGB1555TOARGBROW_AVX2
-#define HAS_ARGB4444TOARGBROW_AVX2
-#define HAS_ARGBTOARGB1555ROW_AVX2
-#define HAS_ARGBTOARGB4444ROW_AVX2
-#define HAS_ARGBTORGB565DITHERROW_AVX2
-#define HAS_ARGBTORGB565ROW_AVX2
-#define HAS_I411TOARGBROW_AVX2
-#define HAS_I422TOARGB1555ROW_AVX2
-#define HAS_I422TOARGB4444ROW_AVX2
-#define HAS_I422TORGB565ROW_AVX2
-#define HAS_I444TOABGRROW_AVX2
-#define HAS_I444TOARGBROW_AVX2
-#define HAS_J400TOARGBROW_AVX2
-#define HAS_NV12TORGB565ROW_AVX2
-#define HAS_RGB565TOARGBROW_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.
@@ -222,6 +192,7 @@ extern "C" {
 #define HAS_ARGBMIRRORROW_AVX2
 #define HAS_ARGBPOLYNOMIALROW_AVX2
 #define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBTORGB565DITHERROW_AVX2
 #define HAS_ARGBTOUVROW_AVX2
 #define HAS_ARGBTOYJROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
@@ -262,6 +233,35 @@ extern "C" {
 #define HAS_ARGBUNATTENUATEROW_AVX2
 #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) && \
+    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
+#define HAS_ARGB1555TOARGBROW_AVX2
+#define HAS_ARGB4444TOARGBROW_AVX2
+#define HAS_ARGBTOARGB1555ROW_AVX2
+#define HAS_ARGBTOARGB4444ROW_AVX2
+#define HAS_ARGBTORGB565ROW_AVX2
+#define HAS_I411TOARGBROW_AVX2
+#define HAS_I422TOARGB1555ROW_AVX2
+#define HAS_I422TOARGB4444ROW_AVX2
+#define HAS_I422TORGB565ROW_AVX2
+#define HAS_I444TOABGRROW_AVX2
+#define HAS_I444TOARGBROW_AVX2
+#define HAS_J400TOARGBROW_AVX2
+#define HAS_NV12TORGB565ROW_AVX2
+#define HAS_RGB565TOARGBROW_AVX2
+#endif
+
+// The following are also available on x64 Visual C.
+#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
+    (!defined(__clang__) || defined(__SSSE3__))
+#define HAS_I422ALPHATOABGRROW_SSSE3
+#define HAS_I422ALPHATOARGBROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#endif
+
 // The following are available on Neon platforms:
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1503
+#define LIBYUV_VERSION 1504
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_gcc.cc

@@ -566,12 +566,58 @@ void ARGBToRGB565DitherRow_SSE2(const uint8* src, uint8* dst,
   : "+r"(src),  // %0
     "+r"(dst),  // %1
     "+r"(pix)   // %2
-  : "m"(dither4) // %3
+  : "rm"(dither4) // %3
   : "memory", "cc",
     "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
+#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
+void ARGBToRGB565DitherRow_AVX2(const uint8* src, uint8* dst,
+                                const uint32 dither4, int pix) {
+  asm volatile (
+    "vbroadcastss %3,%%xmm6                    \n"
+    "vpunpcklbw %%xmm6,%%xmm6,%%xmm6           \n"
+    "vpermq     $0xd8,%%ymm6,%%ymm6            \n"
+    "vpunpcklwd %%ymm6,%%ymm6,%%ymm6           \n"
+    "vpcmpeqb   %%ymm3,%%ymm3,%%ymm3           \n"
+    "vpsrld     $0x1b,%%ymm3,%%ymm3            \n"
+    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
+    "vpsrld     $0x1a,%%ymm4,%%ymm4            \n"
+    "vpslld     $0x5,%%ymm4,%%ymm4             \n"
+    "vpslld     $0xb,%%ymm3,%%ymm5             \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    (%0),%%ymm0                    \n"
+    "vpaddusb   %%ymm6,%%ymm0,%%ymm0           \n"
+    "vpsrld     $0x5,%%ymm0,%%ymm2             \n"
+    "vpsrld     $0x3,%%ymm0,%%ymm1             \n"
+    "vpsrld     $0x8,%%ymm0,%%ymm0             \n"
+    "vpand      %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpand      %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm2,%%ymm1,%%ymm1           \n"
+    "vpor       %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpackusdw  %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "lea        0x20(%0),%0                    \n"
+    "vmovdqu    %%xmm0,(%1)                    \n"
+    "lea        0x10(%1),%1                    \n"
+    "sub        $0x8,%2                        \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "rm"(dither4) // %3
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTORGB565DITHERROW_AVX2
+
+
 void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
   asm volatile (
     "pcmpeqb   %%xmm4,%%xmm4                   \n"