ARGBPolynomial ported to gcc

BUG=265
TESTED=try bots
R=changjun.yang@intel.com

Review URL: https://webrtc-codereview.appspot.com/2191007

git-svn-id: http://libyuv.googlecode.com/svn/trunk@786 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 785
+Version: 786
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -98,6 +98,7 @@ extern "C" {
 // Conversions:
 #define HAS_ARGB1555TOARGBROW_SSE2
 #define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
 #define HAS_ARGBTOARGB1555ROW_SSE2
 #define HAS_ARGBTOARGB4444ROW_SSE2
 #define HAS_ARGBTORAWROW_SSSE3
@@ -142,7 +143,6 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 // Effects:
 #define HAS_ARGBCOLORTABLEROW_X86
-#define HAS_ARGBPOLYNOMIALROW_SSE2
 #define HAS_RGBCOLORTABLEROW_X86
 #define HAS_ARGBLUMACOLORTABLEROW_SSE2
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 785
+#define LIBYUV_VERSION 786
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_common.cc

@@ -2075,13 +2075,22 @@ void ARGBLumaColorTableRow_C(const uint8* src_argb,
     dst_argb[3] = src_argb[3];
     const uint8* luma1 = RGBToYJx256(src_argb[6], src_argb[5], src_argb[4]) +
       luma;
-    dst_argb[4] = luma0[src_argb[4]];
-    dst_argb[5] = luma0[src_argb[5]];
-    dst_argb[6] = luma0[src_argb[6]];
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
     dst_argb[7] = src_argb[7];
     src_argb += 8;
     dst_argb += 8;
   }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = RGBToYJx256(src_argb[2], src_argb[1], src_argb[0]) +
+      luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
 }
 
 #undef clamp0

source/row_posix.cc

@@ -5816,6 +5816,65 @@ int FixedDiv_X86(int num, int div) {
   return num;
 }
 #endif  // HAS_FIXEDDIV_X86
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    ".p2align  4                               \n"
+  "1:                                          \n"
+    "movq      (%0),%%xmm0                     \n"
+    "lea       0x8(%0),%0                      \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     0x10(%3),%%xmm0                 \n"
+    "mulps     0x10(%3),%%xmm4                 \n"
+    "addps     (%3),%%xmm0                     \n"
+    "addps     (%3),%%xmm4                     \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     0x20(%3),%%xmm2                 \n"
+    "mulps     0x20(%3),%%xmm6                 \n"
+    "mulps     0x30(%3),%%xmm1                 \n"
+    "mulps     0x30(%3),%%xmm5                 \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x2,%2                         \n"
+    "movq      %%xmm0,(%1)                     \n"
+    "lea       0x8(%1),%1                      \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus

source/row_win.cc

@@ -6781,6 +6781,7 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     mov        ecx, [esp + 4 + 16]  /* width */
     pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
 
+    // 2 pixel loop.
     align      16
  convertloop:
 // (slow)   pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
@@ -6846,6 +6847,7 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
     mov        edx, [esp + 8]   /* dst_argb */
     mov        ecx, [esp + 16]  /* width */
 
+    // 2 pixel loop.
     align      16
  convertloop:
     vpmovzxbd  ymm0, qword ptr [eax]  // 2 BGRA pixels