Sobel functions

BUG=201
TEST=none
Review URL: https://webrtc-codereview.appspot.com/1200004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@608 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 607
+Version: 608
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -120,6 +120,8 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
 #define HAS_ARGBCOLORTABLEROW_X86
 #define HAS_COPYROW_AVX2
+#define HAS_SOBELXROW_SSSE3
+#define HAS_SOBELYROW_SSSE3
 // Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
 // TODO(fbarchard): Hook these up to all functions. e.g. format conversion.
@@ -1419,6 +1421,16 @@ void ARGBInterpolateRow_NEON(uint8* dst_argb, const uint8* src_argb,
                              ptrdiff_t src_stride_argb, int dst_width,
                              int source_y_fraction);
 
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSSE3(const uint8* src_y0, const uint8* src_y1,
+                     const uint8* src_y2, uint8* dst_sobelx, int width);
+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);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 607
+#define LIBYUV_VERSION 608
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_common.cc

@@ -712,6 +712,53 @@ void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
 }
 #undef SHADE
 
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  for (int i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = a_diff + b_diff * 2 + c_diff;
+    if (sobel < 0) {
+      sobel = -sobel;
+    }
+    if (sobel > 255) {
+      sobel = 255;
+    }
+    dst_sobelx[i] = static_cast<uint8>(sobel);
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  for (int i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = a_diff + b_diff * 2 + c_diff;
+    if (sobel < 0) {
+      sobel = -sobel;
+    }
+    if (sobel > 255) {
+      sobel = 255;
+    }
+    dst_sobely[i] = static_cast<uint8>(sobel);
+  }
+}
+
 void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
   // Copy a Y to RGB.
   for (int x = 0; x < width; ++x) {

source/row_win.cc

@@ -5027,6 +5027,112 @@ void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
 }
 #endif  // HAS_ARGBSUBTRACTROW_AVX2
 
+#ifdef HAS_SOBELXROW_SSSE3
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked) __declspec(align(16))
+void SobelXRow_SSSE3(const uint8* src_y0, const uint8* src_y1,
+                     const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      16
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pabsw      xmm0, xmm0   // SSSE3.  Could use SSE2 psubusw twice instead.
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSSE3
+
+#ifdef HAS_SOBELYROW_SSSE3
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked) __declspec(align(16))
+void SobelYRow_SSSE3(const uint8* src_y0, const uint8* src_y1,
+                     uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      16
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pabsw      xmm0, xmm0   // SSSE3.  Could use SSE2 psubusw twice instead.
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSSE3
+
 #ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
 // Consider float CumulativeSum.
 // Consider calling CumulativeSum one row at time as needed.

unit_test/planar_test.cc

@@ -20,6 +20,7 @@
 #include "libyuv/format_conversion.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
+#include "libyuv/row.h"  // For Sobel
 #include "../unit_test/unit_test.h"
 
 #if defined(_MSC_VER)
@@ -890,6 +891,77 @@ TEST_F(libyuvTest, TestAffine) {
 #endif
 }
 
+TEST_F(libyuvTest, TestSobelX) {
+  SIMD_ALIGNED(uint8 orig_pixels_0[256 + 2]);
+  SIMD_ALIGNED(uint8 orig_pixels_1[256 + 2]);
+  SIMD_ALIGNED(uint8 orig_pixels_2[256 + 2]);
+  SIMD_ALIGNED(uint8 sobel_pixels_c[256]);
+
+  for (int i = 0; i < 256 + 2; ++i) {
+    orig_pixels_0[i] = i;
+    orig_pixels_1[i] = i * 2;
+    orig_pixels_2[i] = i * 3;
+  }
+
+  SobelXRow_C(orig_pixels_0, orig_pixels_1, orig_pixels_2,
+              sobel_pixels_c, 256);
+
+  EXPECT_EQ(16u, sobel_pixels_c[0]);
+  EXPECT_EQ(16u, sobel_pixels_c[100]);
+  EXPECT_EQ(255u, sobel_pixels_c[255]);
+#if defined(HAS_SOBELXROW_SSSE3)
+  SIMD_ALIGNED(uint8 sobel_pixels_opt[256]);
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  if (has_ssse3) {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelXRow_SSSE3(orig_pixels_0, orig_pixels_1, orig_pixels_2,
+                      sobel_pixels_opt, 256);
+    }
+  } else {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelXRow_C(orig_pixels_0, orig_pixels_1, orig_pixels_2,
+                  sobel_pixels_opt, 256);
+    }
+  }
+  for (int i = 0; i < 256; ++i) {
+    EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
+  }
+#endif
+}
+
+TEST_F(libyuvTest, TestSobelY) {
+  SIMD_ALIGNED(uint8 orig_pixels_0[256 + 2]);
+  SIMD_ALIGNED(uint8 orig_pixels_1[256 + 2]);
+  SIMD_ALIGNED(uint8 sobel_pixels_c[256]);
+
+  for (int i = 0; i < 256 + 2; ++i) {
+    orig_pixels_0[i] = i;
+    orig_pixels_1[i] = i * 2;
+  }
+
+  SobelYRow_C(orig_pixels_0, orig_pixels_1, sobel_pixels_c, 256);
+
+  EXPECT_EQ(4u, sobel_pixels_c[0]);
+  EXPECT_EQ(255u, sobel_pixels_c[100]);
+  EXPECT_EQ(0u, sobel_pixels_c[255]);
+#if defined(HAS_SOBELYROW_SSSE3)
+  SIMD_ALIGNED(uint8 sobel_pixels_opt[256]);
+  int has_ssse3 = TestCpuFlag(kCpuHasSSSE3);
+  if (has_ssse3) {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelYRow_SSSE3(orig_pixels_0, orig_pixels_1, sobel_pixels_opt, 256);
+    }
+  } else {
+    for (int i = 0; i < benchmark_pixels_div256_; ++i) {
+      SobelYRow_C(orig_pixels_0, orig_pixels_1, sobel_pixels_opt, 256);
+    }
+  }
+  for (int i = 0; i < 256; ++i) {
+    EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
+  }
+#endif
+}
+
 TEST_F(libyuvTest, TestCopyPlane) {
   int err = 0;
   int yw = benchmark_width_;