Replace Add+YToARGB with SobelRow

BUG=201
TEST=Sobel*
Review URL: https://webrtc-codereview.appspot.com/1217005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@611 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 610
+Version: 611
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -122,6 +122,7 @@ extern "C" {
 #define HAS_COPYROW_AVX2
 #define HAS_SOBELXROW_SSSE3
 #define HAS_SOBELYROW_SSSE3
+#define HAS_SOBELROW_SSE2
 #define HAS_SOBELXYROW_SSE2
 // Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
@@ -1431,6 +1432,10 @@ void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
                  uint8* dst_sobely, int width);
 void SobelYRow_SSSE3(const uint8* src_y0, const uint8* src_y1,
                      uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
 void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
                   uint8* dst_argb, int width);
 void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 610
+#define LIBYUV_VERSION 611
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -1870,25 +1870,6 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
     }
   }
 #endif
-  void (*YToARGBRow)(const uint8* y_buf,
-                     uint8* rgb_buf,
-                     int width) = YToARGBRow_C;
-#if defined(HAS_YTOARGBROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
-      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-    YToARGBRow = YToARGBRow_Any_SSE2;
-    if (IS_ALIGNED(width, 8)) {
-      YToARGBRow = YToARGBRow_SSE2;
-    }
-  }
-#elif defined(HAS_YTOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    YToARGBRow = YToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      YToARGBRow = YToARGBRow_NEON;
-    }
-  }
-#endif
 
   void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
                     uint8* dst_sobely, int width) = SobelYRow_C;
@@ -1905,23 +1886,12 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
     SobelXRow = SobelXRow_SSSE3;
   }
 #endif
-
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
-  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                   uint8* dst_argb, int width) = SobelRow_C;
-                     int width) = ARGBAddRow_C;
+#if defined(HAS_SOBELROW_SSE2)
-#if defined(HAS_ARGBADDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
-  if (TestCpuFlag(kCpuHasSSE2)) {
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-    ARGBAddRow = ARGBAddRow_SSE2;
+    SobelRow = SobelRow_SSE2;
-  }
-#endif
-#if defined(HAS_ARGBADDROW_AVX2_DISABLED)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    clear = true;
-    ARGBAddRow = ARGBAddRow_AVX2;
-  }
-#endif
-#if defined(HAS_ARGBADDROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBAddRow = ARGBAddRow_NEON;
   }
 #endif
 
@@ -1929,7 +1899,6 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
   SIMD_ALIGNED(uint8 row_y[(kMaxStride + kEdge) * 3 + kEdge]);
   SIMD_ALIGNED(uint8 row_sobelx[kMaxStride]);
   SIMD_ALIGNED(uint8 row_sobely[kMaxStride]);
-  SIMD_ALIGNED(uint8 row_sobel[kMaxStride]);
 
   // Convert first row.
   uint8* row_y0 = row_y + kEdge;
@@ -1941,7 +1910,6 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
   ARGBToYRow(src_argb, row_y1, width);
   row_y1[-1] = row_y1[0];
   row_y1[width] = row_y1[width - 1];
-  int awidth = (width + 3) >> 2;
 
   for (int y = 0; y < height; ++y) {
     // Convert next row of ARGB to Y.
@@ -1954,10 +1922,7 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
 
     SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
     SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
-
+    SobelRow(row_sobelx, row_sobely, dst_argb, width);
-    ARGBAddRow(row_sobelx, row_sobely, row_sobel, awidth);
-
-    YToARGBRow(row_sobel, dst_argb, width);
 
     // Cycle thru circular queue of 3 row_y buffers.
     uint8* row_yt = row_y0;
@@ -1975,7 +1940,6 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
-
 // SobelXY ARGB effect.
 // Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
 // TODO(fbarchard): Enable AVX2.  Mixing SSSE3 and AVX2 requires zeroupper.
@@ -2042,7 +2006,7 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
 #endif
 
   void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
-                    uint8* dst_argb, int width) = SobelXYRow_C;
+                     uint8* dst_argb, int width) = SobelXYRow_C;
 #if defined(HAS_SOBELXYROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
@@ -2065,7 +2029,6 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
   ARGBToYRow(src_argb, row_y1, width);
   row_y1[-1] = row_y1[0];
   row_y1[width] = row_y1[width - 1];
-  int awidth = (width + 3) >> 2;
 
   for (int y = 0; y < height; ++y) {
     // Convert next row of ARGB to Y.

source/row_common.cc

@@ -759,6 +759,23 @@ void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
   }
 }
 
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  for (int i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = r + b;
+    if (s > 255) {
+      s = 255;
+    }
+    dst_argb[0] = static_cast<uint8>(s);
+    dst_argb[1] = static_cast<uint8>(s);
+    dst_argb[2] = static_cast<uint8>(s);
+    dst_argb[3] = static_cast<uint8>(255u);
+    dst_argb += 4;
+  }
+}
+
 void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
                   uint8* dst_argb, int width) {
   for (int i = 0; i < width; ++i) {

source/row_win.cc

@@ -5133,6 +5133,58 @@ void SobelYRow_SSSE3(const uint8* src_y0, const uint8* src_y1,
 }
 #endif  // HAS_SOBELYROW_SSSE3
 
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked) __declspec(align(16))
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+    align      16
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm1
+    movdqa     [edx + 16], xmm2
+    movdqa     [edx + 32], xmm3
+    movdqa     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
 #ifdef HAS_SOBELXYROW_SSE2
 // Mixes Sobel X, Sobel Y and Sobel into ARGB.
 // A = 255
@@ -5182,7 +5234,7 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
     ret
   }
 }
-#endif // HAS_SOBELXYROW_SSE2
+#endif  // HAS_SOBELXYROW_SSE2
 
 #ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
 // Consider float CumulativeSum.

unit_test/planar_test.cc

@@ -962,6 +962,47 @@ TEST_F(libyuvTest, TestSobelY) {
 #endif
 }
 
+TEST_F(libyuvTest, TestSobel) {
+  SIMD_ALIGNED(uint8 orig_sobelx[256]);
+  SIMD_ALIGNED(uint8 orig_sobely[256]);
+  SIMD_ALIGNED(uint8 sobel_pixels_c[256 * 4]);
+
+  for (int i = 0; i < 256; ++i) {
+    orig_sobelx[i] = i;
+    orig_sobely[i] = i * 2;
+  }
+
+  SobelRow_C(orig_sobelx, orig_sobely, sobel_pixels_c, 256);
+
+  EXPECT_EQ(0u, sobel_pixels_c[0]);
+  EXPECT_EQ(3u, sobel_pixels_c[4]);
+  EXPECT_EQ(3u, sobel_pixels_c[5]);
+  EXPECT_EQ(3u, sobel_pixels_c[6]);
+  EXPECT_EQ(255u, sobel_pixels_c[7]);
+  EXPECT_EQ(6u, sobel_pixels_c[8]);
+  EXPECT_EQ(6u, sobel_pixels_c[9]);
+  EXPECT_EQ(6u, sobel_pixels_c[10]);
+  EXPECT_EQ(255u, sobel_pixels_c[7]);
+  EXPECT_EQ(255u, sobel_pixels_c[100 * 4 + 1]);
+  EXPECT_EQ(255u, sobel_pixels_c[255 * 4 + 1]);
+#if defined(HAS_SOBELROW_SSE2)
+  SIMD_ALIGNED(uint8 sobel_pixels_opt[256 * 4]);
+  int has_sse2 = TestCpuFlag(kCpuHasSSE2);
+  if (has_sse2) {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelRow_SSE2(orig_sobelx, orig_sobely, sobel_pixels_opt, 256);
+    }
+  } else {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelRow_C(orig_sobelx, orig_sobely, sobel_pixels_opt, 256);
+    }
+  }
+  for (int i = 0; i < 16; ++i) {
+    EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
+  }
+#endif
+}
+
 TEST_F(libyuvTest, TestSobelXY) {
   SIMD_ALIGNED(uint8 orig_sobelx[256]);
   SIMD_ALIGNED(uint8 orig_sobely[256]);