YToARGB AVX2 port from SSE2

BUG=393
TESTED=YToARGB unittest
R=brucedawson@google.com, harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/41679004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1258 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/row.h

@@ -220,6 +220,11 @@ extern "C" {
 #define HAS_ARGBUNATTENUATEROW_AVX2
 #endif
 
+// The following are available require VS2012
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_YTOARGBROW_AVX2
+#endif
+
 // The following are Yasm x86 only:
 // TODO(fbarchard): Port AVX2 to inline.
 #if !defined(LIBYUV_DISABLE_X86) && defined(HAVE_YASM)
@@ -980,9 +985,6 @@ void I422ToRGB565Row_C(const uint8* src_y,
                        const uint8* src_v,
                        uint8* dst_rgb565,
                        int width);
-void YToARGBRow_C(const uint8* src_y,
-                  uint8* dst_argb,
-                  int width);
 void I422ToARGBRow_AVX2(const uint8* src_y,
                         const uint8* src_u,
                         const uint8* src_v,
@@ -1182,15 +1184,25 @@ void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
                             const uint8* src_v,
                             uint8* dst_argb,
                             int width);
+
+void YToARGBRow_C(const uint8* src_y,
+                  uint8* dst_argb,
+                  int width);
 void YToARGBRow_SSE2(const uint8* src_y,
                      uint8* dst_argb,
                      int width);
+void YToARGBRow_AVX2(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
 void YToARGBRow_NEON(const uint8* src_y,
                      uint8* dst_argb,
                      int width);
 void YToARGBRow_Any_SSE2(const uint8* src_y,
                          uint8* dst_argb,
                          int width);
+void YToARGBRow_Any_AVX2(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
 void YToARGBRow_Any_NEON(const uint8* src_y,
                          uint8* dst_argb,
                          int width);

source/convert_argb.cc

@@ -276,6 +276,14 @@ int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_YTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YToARGBRow = YToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      YToARGBRow = YToARGBRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_YTOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     YToARGBRow = YToARGBRow_Any_NEON;

source/row_any.cc

@@ -169,6 +169,10 @@ RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C,
 RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
        7, 1, 4)
 #endif
+#if defined(HAS_YTOARGBROW_AVX2)
+RGBANY(YToARGBRow_Any_AVX2, YToARGBRow_AVX2, YToARGBRow_C,
+       15, 1, 4)
+#endif
 #if defined(HAS_YUY2TOARGBROW_SSSE3)
 RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C,
        15, 2, 4)

source/row_win.cc

@@ -2299,8 +2299,8 @@ void I422ToRGBARow_SSSE3(const uint8* y_buf,
 
 #endif  // HAS_I422TOARGBROW_SSSE3
 
-// TODO(fbarchard): Remove shift by 6.
 #ifdef HAS_YTOARGBROW_SSE2
+// 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void YToARGBRow_SSE2(const uint8* y_buf,
                      uint8* rgb_buf,
@@ -2341,12 +2341,62 @@ void YToARGBRow_SSE2(const uint8* y_buf,
     lea        edx,  [edx + 32]
     sub        ecx, 8
     jg         convertloop
-
     ret
   }
 }
 #endif  // HAS_YTOARGBROW_SSE2
 
+#ifdef HAS_YTOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void YToARGBRow_AVX2(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) {
+  __asm {
+    vpcmpeqb   ymm4, ymm4, ymm4     // generate mask 0xff000000
+    vpslld     ymm4, ymm4, 24
+    mov        eax, 0x04ad04ad      // 04ad = 1197 = round(1.164 * 64 * 16)
+    vmovd      xmm3, eax
+    vbroadcastss ymm3, xmm3
+    mov        eax, 0x4a7f4a7f      // 4a7f = 19071 = round(1.164 * 64 * 256)
+    vmovd      xmm2, eax
+    vbroadcastss ymm2, xmm2
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+    vmovdqu    xmm0, [eax]
+    lea        eax, [eax + 16]
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklbw ymm0, ymm0, ymm0           // Y.Y
+    vpmulhuw   ymm0, ymm0, ymm2
+    vpsubusw   ymm0, ymm0, ymm3
+    vpsrlw     ymm0, ymm0, 6
+    vpackuswb  ymm0, ymm0, ymm0           // G.  still mutated: 3120
+
+    // TODO(fbarchard): Weave alpha with unpack.
+    // Step 2: Weave into ARGB
+    vpunpcklbw ymm1, ymm0, ymm0           // GG - mutates
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm1           // GGGG first 4 pixels
+    vpunpckhwd ymm1, ymm1, ymm1           // GGGG next 4 pixels
+    vpor       ymm0, ymm0, ymm4
+    vpor       ymm1, ymm1, ymm4
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YTOARGBROW_AVX2
+
+
 #ifdef HAS_MIRRORROW_SSSE3
 // Shuffle table for reversing the bytes.
 static const uvec8 kShuffleMirror = {

unit_test/convert_test.cc

@@ -1255,11 +1255,13 @@ TEST_F(libyuvTest, TestYToARGB) {
   YToARGB(y, 0, argb, 0, 32, 1);
 
   for (int i = 0; i < 32; ++i) {
-    printf("%d: %d <-> %d,%d,%d,%d\n", y[i], expectedg[i],
+    printf("%2d %d: %d <-> %d,%d,%d,%d\n", i, y[i], expectedg[i],
            argb[i * 4 + 0],
            argb[i * 4 + 1],
            argb[i * 4 + 2],
            argb[i * 4 + 3]);
+  }
+  for (int i = 0; i < 32; ++i) {
     EXPECT_NEAR(expectedg[i], argb[i * 4 + 0], 1);
   }
 }