ARGBCopyAlpha for effects

BUG=none
TEST=none
R=johannkoenig@google.com

Review URL: https://webrtc-codereview.appspot.com/2385004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@810 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 809
+Version: 810
 License: BSD
 License File: LICENSE
 

include/libyuv/planar_functions.h

@@ -209,18 +209,6 @@ int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
                    const int8* matrix_rgb,
                    int x, int y, int width, int height);
 
-#ifdef __cplusplus
-}  // extern "C"
-// Deprecated. Temporary API mapper.
-int inline ARGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
-                           const int8* matrix_rgb,
-                           int x, int y, int width, int height) {
-  return RGBColorMatrix(dst_argb, dst_stride_argb, matrix_rgb,
-                        x, y, width, height);
-}
-extern "C" {
-#endif
-
 // Apply a color table each ARGB pixel.
 // Table contains 256 ARGB values.
 LIBYUV_API
@@ -273,6 +261,12 @@ int ARGBCopy(const uint8* src_argb, int src_stride_argb,
              uint8* dst_argb, int dst_stride_argb,
              int width, int height);
 
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
 typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
                              uint8* dst_argb, int width);
 

include/libyuv/row.h

@@ -167,6 +167,8 @@ extern "C" {
 // Effects:
 // TODO(fbarchard): Optimize and enable
 // #define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// TODO(fbarchard): Optimize and enable
+// #define HAS_ARGBCOPYALPHAROW_SSE2
 
 // Caveat: Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
@@ -697,6 +699,9 @@ void CopyRow_NEON(const uint8* src, uint8* dst, int count);
 void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
 void CopyRow_C(const uint8* src, uint8* dst, int count);
 
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width);
+
 void SetRow_X86(uint8* dst, uint32 v32, int count);
 void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
                      int dst_stride, int height);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 809
+#define LIBYUV_VERSION 810
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -2162,6 +2162,44 @@ int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Copy ARGB with optional flipping
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  // TODO(fbarchard): Consider macro for boiler plate checks, invert and/or
+  // row coalesce.
+  if (!src_argb || !dst_argb ||
+      width <= 0 || height == 0) {
+    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;
+  }
+  // Coalesce contiguous rows.
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
+    return ARGBCopyAlpha(src_argb, 0,
+                         dst_argb, 0,
+                         width * height, 1);
+  }
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+  for (int y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

source/row_common.cc

@@ -2099,6 +2099,18 @@ void ARGBLumaColorTableRow_C(const uint8* src_argb,
   }
 }
 
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  for (int i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
 #undef clamp0
 #undef clamp255
 

source/row_win.cc

@@ -3603,6 +3603,37 @@ void CopyRow_X86(const uint8* src, uint8* dst, int count) {
 }
 #endif  // HAS_COPYROW_X86
 
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        edx, edi
+    mov        eax, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm5, xmm5  // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      16
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    maskmovdqu xmm0, xmm5
+    lea        edi, [edi + 16]
+    maskmovdqu xmm1, xmm5
+    lea        edi, [edi + 16]
+    sub        ecx, 8
+    jg         convertloop
+
+    mov        edi, edx
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
 #ifdef HAS_SETROW_X86
 // SetRow8 writes 'count' bytes using a 32 bit value repeated.
 __declspec(naked) __declspec(align(16))

unit_test/planar_test.cc

@@ -1906,5 +1906,34 @@ TEST_F(libyuvTest, TestARGBLumaColorTable) {
   }
 }
 
+TEST_F(libyuvTest, TestARGBCopyAlpha) {
+  const int kSize = benchmark_width_ * benchmark_height_ * 4;
+  align_buffer_64(orig_pixels, kSize);
+  align_buffer_64(dst_pixels_opt, kSize);
+  align_buffer_64(dst_pixels_c, kSize);
+
+  MemRandomize(orig_pixels, kSize);
+  MemRandomize(dst_pixels_opt, kSize);
+  memcpy(dst_pixels_c, dst_pixels_opt, kSize);
+
+  MaskCpuFlags(0);
+  ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4,
+                dst_pixels_c, benchmark_width_ * 4,
+                benchmark_width_, benchmark_height_);
+  MaskCpuFlags(-1);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ARGBCopyAlpha(orig_pixels, benchmark_width_ * 4,
+                  dst_pixels_opt, benchmark_width_ * 4,
+                  benchmark_width_, benchmark_height_);
+  }
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_64(dst_pixels_c)
+  free_aligned_buffer_64(dst_pixels_opt)
+  free_aligned_buffer_64(orig_pixels)
+}
 
 }  // namespace libyuv