Port Polynomial AVX2 code to GCC/NaCL

BUG=269
TESTED=untested
R=johannkoenig@google.com

Review URL: https://webrtc-codereview.appspot.com/2262004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@795 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 794
+Version: 795
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -90,7 +90,6 @@ extern "C" {
 #define HAS_YUY2TOYROW_SSE2
 #endif
 
-
 // The following are available on all x86 platforms except NaCL x64:
 #if !defined(LIBYUV_DISABLE_X86) && \
     (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
@@ -142,6 +141,16 @@ extern "C" {
 #define HAS_RGBCOLORTABLEROW_X86
 #endif
 
+// The following are available on all x86 platforms, including NaCL, but
+// require VS2012, llvm or NaCL.
+// Caveat: llvm 3.1 required, but does not provide a version.
+#if !defined(LIBYUV_DISABLE_X86) && \
+  ((defined(_M_IX86) && defined(_MSC_VER) && _MSC_VER >= 1700) || \
+  defined(__native_client__) || defined(__llvm__))
+// Effects:
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#endif
+
 // The following are Windows only:
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
@@ -173,7 +182,6 @@ extern "C" {
 #define HAS_ARGBATTENUATEROW_AVX2
 #define HAS_ARGBMIRRORROW_AVX2
 #define HAS_ARGBMULTIPLYROW_AVX2
-#define HAS_ARGBPOLYNOMIALROW_AVX2
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBUNATTENUATEROW_AVX2
 #endif  // _MSC_VER >= 1700

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 794
+#define LIBYUV_VERSION 795
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_posix.cc

@@ -5875,6 +5875,52 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_SSE2
 
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vmovdqu   "MEMACCESS(3)",%%xmm4           \n"
+    "vmovdqu   "MEMACCESS2(0x10,3)",%%xmm5     \n"
+    "vmovdqu   "MEMACCESS2(0x20,3)",%%xmm6     \n"
+    "vmovdqu   "MEMACCESS2(0x30,3)",%%xmm7     \n"
+    "vpermq    $0x44,%%ymm4,%%ymm4             \n"
+    "vpermq    $0x44,%%ymm5,%%ymm5             \n"
+    "vpermq    $0x44,%%ymm6,%%ymm6             \n"
+    "vpermq    $0x44,%%ymm7,%%ymm7             \n"
+
+    // 2 pixel loop.
+    ".p2align  4                               \n"
+  "1:                                          \n"
+    "vpmovzxbd   "MEMACCESS(0)",%%ymm0         \n"  // 2 ARGB pixels
+    "lea         "MEMLEA(0x8,0)",%0            \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "sub         $0x2,%2                       \n"
+    "movq        %%xmm0,"MEMACCESS(1)"         \n"
+    "lea         "MEMLEA(0x8,1)",%1            \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.
 void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,