ARGBSobel

BUG=201
TEST=Sobel*
Review URL: https://webrtc-codereview.appspot.com/1221005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@609 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 608
+Version: 609
 License: BSD
 License File: LICENSE
 

include/libyuv/planar_functions.h

@@ -347,6 +347,12 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
                 uint8* dst_argb, int dst_stride_argb,
                 const uint8* shuffler, int width, int height);
 
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
 #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 608
+#define LIBYUV_VERSION 609
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

libyuv_test.gyp

@@ -71,7 +71,6 @@
         }],
       ], # conditions
     },
-
   ], # targets
 }
 

source/planar_functions.cc

@@ -1822,6 +1822,159 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
   return 0;
 }
 
+// Sobel ARGB effect.
+// TODO(fbarchard): Enable AVX2.  Mixing SSSE3 and AVX2 requires zeroupper.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  if (!src_argb  || !dst_argb ||
+      width <= 0 || height == 0 || width > kMaxStride) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      // Assumed row buffer aligned.
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2_DISABLED)
+  bool clear = false;
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    clear = true;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#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;
+#if defined(HAS_SOBELYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SobelYRow = SobelYRow_SSSE3;
+  }
+#endif
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+#if defined(HAS_SOBELXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SobelXRow = SobelXRow_SSSE3;
+  }
+#endif
+
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+#if defined(HAS_ARGBADDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_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
+
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  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;
+  uint8* row_y1 = row_y0 + kMaxStride;
+  uint8* row_y2 = row_y1 + kMaxStride;
+  ARGBToYRow(src_argb, row_y0, width);
+  row_y0[-1] = row_y0[0];
+  row_y0[width] = row_y0[width - 1];
+  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.
+    if (y < (height - 1)) {
+      src_argb += src_stride_argb;
+    }
+    ARGBToYRow(src_argb, row_y2, width);
+    row_y2[-1] = row_y2[0];
+    row_y2[width] = row_y2[width - 1];
+
+    SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+    SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, 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;
+    row_y0 = row_y1;
+    row_y1 = row_y2;
+    row_y2 = row_yt;
+
+    dst_argb += dst_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_AVX2_DISABLED)
+  if (clear) {
+    __asm vzeroupper;
+  }
+#endif
+  return 0;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

unit_test/planar_test.cc

@@ -1232,4 +1232,67 @@ TEST_F(libyuvTest, ARGBSubtract_Opt) {
   EXPECT_LE(max_diff, 1);
 }
 
+static int TestSobel(int width, int height, int benchmark_iterations,
+                     int invert, int off) {
+  const int kBpp = 4;
+  const int kStride = (width * kBpp + 15) & ~15;
+  align_buffer_64(src_argb_a, kStride * height + off);
+  align_buffer_64(dst_argb_c, kStride * height);
+  align_buffer_64(dst_argb_opt, kStride * height);
+  srandom(time(NULL));
+  for (int i = 0; i < kStride * height; ++i) {
+    src_argb_a[i + off] = (random() & 0xff);
+  }
+  memset(dst_argb_c, 0, kStride * height);
+  memset(dst_argb_opt, 0, kStride * height);
+
+  MaskCpuFlags(0);
+  ARGBSobel(src_argb_a + off, kStride,
+            dst_argb_c, kStride,
+            width, invert * height);
+  MaskCpuFlags(-1);
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    ARGBSobel(src_argb_a + off, kStride,
+              dst_argb_opt, kStride,
+              width, invert * height);
+  }
+  int max_diff = 0;
+  for (int i = 0; i < kStride * height; ++i) {
+    int abs_diff =
+        abs(static_cast<int>(dst_argb_c[i]) -
+            static_cast<int>(dst_argb_opt[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  free_aligned_buffer_64(src_argb_a)
+  free_aligned_buffer_64(dst_argb_c)
+  free_aligned_buffer_64(dst_argb_opt)
+  return max_diff;
+}
+
+TEST_F(libyuvTest, ARGBSobel_Any) {
+  int max_diff = TestSobel(benchmark_width_ - 1, benchmark_height_,
+                           benchmark_iterations_, +1, 0);
+  EXPECT_LE(max_diff, 14);
+}
+
+TEST_F(libyuvTest, ARGBSobel_Unaligned) {
+  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
+                           benchmark_iterations_, +1, 1);
+  EXPECT_LE(max_diff, 14);
+}
+
+TEST_F(libyuvTest, ARGBSobel_Invert) {
+  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
+                           benchmark_iterations_, -1, 0);
+  EXPECT_LE(max_diff, 14);
+}
+
+TEST_F(libyuvTest, ARGBSobel_Opt) {
+  int max_diff = TestSobel(benchmark_width_, benchmark_height_,
+                           benchmark_iterations_, +1, 0);
+  EXPECT_LE(max_diff, 14);
+}
+
 }  // namespace libyuv

unit_test/unit_test.cc

@@ -39,8 +39,9 @@ libyuvTest::libyuvTest() : rotate_max_w_(128), rotate_max_h_(128),
     if (height) {
       benchmark_height_ = atoi(height);  // NOLINT
     }
-    benchmark_pixels_div256_ = (benchmark_iterations_ * benchmark_width_ *
-        benchmark_height_ + 255) / 256;
+    benchmark_pixels_div256_ = static_cast<int>(
+        (static_cast<double>(benchmark_width_ *
+        benchmark_height_) * benchmark_iterations_  + 255.0) / 256.0);
 }
 
 int main(int argc, char** argv) {