YUV conversion on the GPU.

ui/surface/accelerated_surface_transformer_win_unittest.cc

-// Copyright (c) 2010 The Chromium Authors. All rights reserved.
+// Copyright (c) 2013 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include <d3d9.h>
 #include <random>
 
 #include "base/basictypes.h"
+#include "base/file_util.h"
 #include "base/hash.h"
 #include "base/scoped_native_library.h"
 #include "base/stringprintf.h"
+#include "base/time.h"
 #include "base/win/scoped_comptr.h"
 #include "base/win/windows_version.h"
+#include "media/base/simd/convert_rgb_to_yuv.h"
+#include "media/base/yuv_convert.h"
+#include "skia/ext/image_operations.h"
 #include "testing/gtest/include/gtest/gtest-param-test.h"
 #include "testing/gtest/include/gtest/gtest.h"
+#include "third_party/skia/include/core/SkBitmap.h"
+#include "third_party/skia/include/core/SkColor.h"
+#include "ui/gfx/codec/png_codec.h"
 #include "ui/gfx/rect.h"
 #include "ui/surface/accelerated_surface_transformer_win.h"
 #include "ui/surface/accelerated_surface_win.h"
 #include "ui/surface/d3d9_utils_win.h"
 
 namespace d3d_utils = ui_surface_d3d9_utils;
 
 using base::win::ScopedComPtr;
 using std::uniform_int_distribution;
 
-// Provides a reference rasterizer (all rendering done by software emulation)
-// Direct3D device, for use by unit tests.
+namespace {
+
+// Debug flag, useful when hacking on tests.
+const bool kDumpImagesOnFailure = false;
+
+SkBitmap ToSkBitmap(IDirect3DSurface9* surface, bool is_single_channel) {
+  D3DLOCKED_RECT locked_rect;
+  EXPECT_HRESULT_SUCCEEDED(
+      surface->LockRect(&locked_rect, NULL, D3DLOCK_READONLY));
+
+  SkBitmap result;
+  gfx::Size size = d3d_utils::GetSize(surface);
+  if (is_single_channel)
+    size = gfx::Size(size.width() * 4, size.height());
+  result.setConfig(SkBitmap::kARGB_8888_Config, size.width(), size.height());
+  result.setIsOpaque(true);
+  result.allocPixels();
+  result.lockPixels();
+  for (int y = 0; y < size.height(); ++y) {
+    uint8* row8 = reinterpret_cast<uint8*>(locked_rect.pBits) +
+        (y * locked_rect.Pitch);
+    if (is_single_channel) {
+      for (int x = 0; x < size.width(); ++x) {
+        *result.getAddr32(x, y) = SkColorSetRGB(row8[x], row8[x], row8[x]);
+      }
+    } else {
+      uint32* row32 = reinterpret_cast<uint32*>(row8);
+      for (int x = 0; x < size.width(); ++x) {
+        *result.getAddr32(x, y) = row32[x] | 0xFF000000;
+      }
+    }
+  }
+  result.unlockPixels();
+  result.setImmutable();
+  surface->UnlockRect();
+  return result;
+}
+
+bool WritePNGFile(const SkBitmap& bitmap, const FilePath& file_path) {
+  std::vector<unsigned char> png_data;
+  const bool discard_transparency = true;
+  if (gfx::PNGCodec::EncodeBGRASkBitmap(bitmap,
+                                        discard_transparency,
+                                        &png_data) &&
+      file_util::CreateDirectory(file_path.DirName())) {
+    char* data = reinterpret_cast<char*>(&png_data[0]);
+    int size = static_cast<int>(png_data.size());
+    return file_util::WriteFile(file_path, data, size) == size;
+  }
+  return false;
+}
+
+}  // namespace
+
+// Test fixture for AcceleratedSurfaceTransformer.
 //
 // This class is parameterized so that it runs only on Vista+. See
 // WindowsVersionIfVistaOrBetter() for details on this works.
 class AcceleratedSurfaceTransformerTest : public testing::TestWithParam<int> {
  public:
-  AcceleratedSurfaceTransformerTest() {};
+  AcceleratedSurfaceTransformerTest() : color_error_tolerance_(0) {};
 
   IDirect3DDevice9Ex* device() { return device_.get(); }
 
   virtual void SetUp() {
     if (!d3d_module_.is_valid()) {
       if (!d3d_utils::LoadD3D9(&d3d_module_)) {
         GTEST_FAIL() << "Could not load d3d9.dll";
         return;
       }
     }
@@ skipping 28 matching lines @@
   }
 
   // Driver workaround: on an Intel GPU (Mobile Intel 965 Express), it seems
   // necessary to flush between drawing and locking, for the synchronization
   // to behave properly.
   void BeforeLockWorkaround() {
     EXPECT_HRESULT_SUCCEEDED(
         device()->Present(0, 0, 0, 0));
   }
 
+  void WarnOnMissingFeatures(AcceleratedSurfaceTransformer* gpu_ops) {
+    // Prints a single warning line if some tests are feature-dependent
+    // and the feature is not supported by the current GPU.
+    if (!gpu_ops->device_supports_multiple_render_targets()) {
+      LOG(WARNING) << "MRT not supported, some tests will be skipped. "
+                   << GetAdapterInfo();
+    }
+  }
+
   // Locks and fills a surface with a checkerboard pattern where the colors
   // are random but the total image pattern is horizontally and vertically
   // symmetric.
   void FillSymmetricRandomCheckerboard(
       IDirect3DSurface9* lockable_surface,
       const gfx::Size& size,
       int checker_square_size) {
 
     D3DLOCKED_RECT locked_rect;
     ASSERT_HRESULT_SUCCEEDED(
         lockable_surface->LockRect(&locked_rect, NULL, D3DLOCK_DISCARD));
     DWORD* surface = reinterpret_cast<DWORD*>(locked_rect.pBits);
     ASSERT_EQ(0, locked_rect.Pitch % sizeof(DWORD));
     int pitch = locked_rect.Pitch / sizeof(DWORD);
 
-    for (int y = 0; y <= size.height() / 2; y += checker_square_size) {
-      for (int x = 0; x <= size.width() / 2; x += checker_square_size) {
+    for (int y = 0; y < (size.height() + 1) / 2; y += checker_square_size) {
+      for (int x = 0; x < (size.width() + 1) / 2; x += checker_square_size) {
         DWORD color = RandomColor();
         int y_limit = std::min(size.height() / 2, y + checker_square_size - 1);
         int x_limit = std::min(size.width() / 2, x + checker_square_size - 1);
         for (int y_lo = y; y_lo <= y_limit; y_lo++) {
           for (int x_lo = x; x_lo <= x_limit; x_lo++) {
             int y_hi = size.height() - 1 - y_lo;
             int x_hi = size.width() - 1 - x_lo;
             surface[x_lo + y_lo*pitch] = color;
             surface[x_lo + y_hi*pitch] = color;
             surface[x_hi + y_lo*pitch] = color;
@@ skipping 38 matching lines @@
   bool AssertSameColor(DWORD color_a, DWORD color_b) {
     if (color_a == color_b)
       return true;
     uint8* a = reinterpret_cast<uint8*>(&color_a);
     uint8* b = reinterpret_cast<uint8*>(&color_b);
     int max_error = 0;
     for (int i = 0; i < 4; i++)
       max_error = std::max(max_error,
           std::abs(static_cast<int>(a[i]) - b[i]));
 
-    if (max_error <= kAbsoluteColorErrorTolerance)
+    if (max_error <= color_error_tolerance())
       return true;
 
     std::string expected_color =
         StringPrintf("%3d, %3d, %3d, %3d", a[0], a[1], a[2], a[3]);
     std::string actual_color =
         StringPrintf("%3d, %3d, %3d, %3d", b[0], b[1], b[2], b[3]);
     EXPECT_EQ(expected_color, actual_color)
         << "Componentwise color difference was "
-        << max_error << "; max allowed is " << kAbsoluteColorErrorTolerance;
+        << max_error << "; max allowed is " << color_error_tolerance();
 
     return false;
   }
 
+bool AssertSameColor(uint8 color_a, uint8 color_b) {
+    if (color_a == color_b)
+      return true;
+    int max_error = std::abs((int) color_a - (int) color_b);
+    if (max_error <= color_error_tolerance())
+      return true;
+    ADD_FAILURE() << "Colors not equal: " << StringPrintf("0x%x", color_a)
+                  << " vs. " << StringPrintf("0x%x", color_b);
+    return false;
+  }
+
   // Asserts that an image is symmetric with respect to itself: both
   // horizontally and vertically, within the tolerance of AssertSameColor.
   void AssertSymmetry(IDirect3DSurface9* lockable_surface,
                       const gfx::Size& size) {
     BeforeLockWorkaround();
 
     D3DLOCKED_RECT locked_rect;
     ASSERT_HRESULT_SUCCEEDED(
         lockable_surface->LockRect(&locked_rect, NULL, D3DLOCK_READONLY));
     ASSERT_EQ(0, locked_rect.Pitch % sizeof(DWORD));
@@ skipping 51 matching lines @@
           actual->UnlockRect();
           GTEST_FAIL() << "Pixels (" << x << ", " << y << ") vs. "
                        << "(" << x << ", " << y_actual << ")";
         }
     }
     expected->UnlockRect();
     actual->UnlockRect();
   }
 
  protected:
-  static const int kAbsoluteColorErrorTolerance = 5;
-
   DWORD RandomColor() {
     return random_dword_(rng_);
   }
 
+  void set_color_error_tolerance(int value) {
+    color_error_tolerance_ = value;
+  }
+
+  int color_error_tolerance() {
+    return color_error_tolerance_;
+  }
+
   void DoResizeBilinearTest(AcceleratedSurfaceTransformer* gpu_ops,
                             const gfx::Size& src_size,
                             const gfx::Size& dst_size,
                             int checkerboard_size) {
 
     SCOPED_TRACE(
         StringPrintf("Resizing %dx%d -> %dx%d at checkerboard size of %d",
                      src_size.width(), src_size.height(),
                      dst_size.width(), dst_size.height(),
                      checkerboard_size));
 
+    set_color_error_tolerance(4);
+
     base::win::ScopedComPtr<IDirect3DSurface9> src, dst;
     ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(
         device(), src_size, src.Receive()))
             << "Could not create src render target";
     ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(
         device(), dst_size, dst.Receive()))
             << "Could not create dst render target";
 
     FillSymmetricRandomCheckerboard(src, src_size, checkerboard_size);
 
     ASSERT_TRUE(gpu_ops->ResizeBilinear(src, gfx::Rect(src_size), dst));
 
     AssertSymmetry(dst, dst_size);
   }
 
+  void CreateRandomCheckerboardTexture(
+      const gfx::Size& size,
+      int checkerboard_size,
+      IDirect3DSurface9** reference_surface,
+      IDirect3DTexture9** result) {
+    base::win::ScopedComPtr<IDirect3DSurface9> dst;
+    ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(device(), size,
+        reference_surface));
+    ASSERT_TRUE(d3d_utils::CreateTemporaryRenderTargetTexture(device(), size,
+        result, dst.Receive()));
+    FillRandomCheckerboard(*reference_surface, size, checkerboard_size);
+    ASSERT_HRESULT_SUCCEEDED(
+        device()->StretchRect(
+            *reference_surface, NULL, dst, NULL, D3DTEXF_NONE));
+  }
+
+  void AssertSame(int width_in_bytes, int height, uint8* reference,
+                  IDirect3DSurface9* lockable) {
+    BeforeLockWorkaround();
+
+    D3DLOCKED_RECT locked_rect;
+    ASSERT_HRESULT_SUCCEEDED(
+        lockable->LockRect(&locked_rect, NULL, D3DLOCK_READONLY));
+    uint8* actual = reinterpret_cast<uint8*>(locked_rect.pBits);
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width_in_bytes; ++x) {
+        if (!AssertSameColor(reference[y * width_in_bytes + x],
+                             actual[y * locked_rect.Pitch + x])) {
+          lockable->UnlockRect();
+          GTEST_FAIL() << "At pixel (" << x << ", " << y << ")";
+        }
+      }
+    }
+    lockable->UnlockRect();
+  }
+
   void DoCopyInvertedTest(AcceleratedSurfaceTransformer* gpu_ops,
                           const gfx::Size& size) {
 
     SCOPED_TRACE(
         StringPrintf("CopyInverted @ %dx%d", size.width(), size.height()));
 
-    base::win::ScopedComPtr<IDirect3DSurface9> checkerboard, src, dst;
-    base::win::ScopedComPtr<IDirect3DTexture9> src_texture;
-    ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(device(), size,
-        checkerboard.Receive())) << "Could not create src render target";;
-    ASSERT_TRUE(d3d_utils::CreateTemporaryRenderTargetTexture(device(), size,
-        src_texture.Receive(), src.Receive()))
-            << "Could not create src texture.";
-    ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(device(), size,
+    set_color_error_tolerance(0);
+
+    base::win::ScopedComPtr<IDirect3DSurface9> dst, reference_pattern;
+    base::win::ScopedComPtr<IDirect3DTexture9> src;
+
+    CreateRandomCheckerboardTexture(size, 1, reference_pattern.Receive(),
+                                    src.Receive());
+
+    // Alloc a slightly larger image 75% of the time, to test that the
+    // viewport is set properly.
+    const int kAlign = 4;
+    gfx::Size alloc_size((size.width() + kAlign - 1) / kAlign * kAlign,
+                         (size.height() + kAlign - 1) / kAlign * kAlign);
+
+    ASSERT_TRUE(d3d_utils::CreateTemporaryLockableSurface(device(), alloc_size,
         dst.Receive())) << "Could not create dst render target.";
 
-    FillRandomCheckerboard(checkerboard, size, 1);
-    ASSERT_HRESULT_SUCCEEDED(
-        device()->StretchRect(checkerboard, NULL, src, NULL, D3DTEXF_NONE));
-    ASSERT_TRUE(gpu_ops->CopyInverted(src_texture, dst, size));
-    AssertIsInvertedCopy(size, checkerboard, dst);
-  }
-
+    ASSERT_TRUE(gpu_ops->CopyInverted(src, dst, size));
+    AssertIsInvertedCopy(size, reference_pattern, dst);
+  }
+
+
+  void DoYUVConversionTest(AcceleratedSurfaceTransformer* gpu_ops,
+                           const gfx::Size& src_size,
+                           int checkerboard_size) {
+    // Test the non-MRT implementation, and the MRT implementation as well
+    // (if supported by the device).
+    ASSERT_NO_FATAL_FAILURE(
+        DoYUVConversionTest(gpu_ops, src_size, src_size,
+                            checkerboard_size, false));
+    if (gpu_ops->device_supports_multiple_render_targets()) {
+      ASSERT_NO_FATAL_FAILURE(
+          DoYUVConversionTest(gpu_ops, src_size, src_size,
+                              checkerboard_size, true));
+    }
+  }
+
+  void DoYUVConversionScaleTest(AcceleratedSurfaceTransformer* gpu_ops,
+                                const gfx::Size& src_size,
+                                const gfx::Size& dst_size) {
+    // Test the non-MRT implementation, and the MRT implementation as well
+    // (if supported by the device).
+    if (gpu_ops->device_supports_multiple_render_targets()) {
+      ASSERT_NO_FATAL_FAILURE(
+          DoYUVConversionTest(gpu_ops, src_size, dst_size, 4, true));
+    }
+    ASSERT_NO_FATAL_FAILURE(
+        DoYUVConversionTest(gpu_ops, src_size, dst_size, 4, false));
+  }
+
+  void DoYUVConversionTest(AcceleratedSurfaceTransformer* gpu_ops,
+                           const gfx::Size& src_size,
+                           const gfx::Size& dst_size,
+                           int checkerboard_size,
+                           boolean use_multi_render_targets) {
+    SCOPED_TRACE(
+        StringPrintf("YUV Converting %dx%d at checkerboard size of %d; MRT %s",
+                     src_size.width(), src_size.height(),
+                     checkerboard_size,
+                     use_multi_render_targets ? "enabled" : "disabled"));
+
+
+    base::win::ScopedComPtr<IDirect3DTexture9> src;
+    base::win::ScopedComPtr<IDirect3DSurface9> reference;
+    base::win::ScopedComPtr<IDirect3DSurface9> dst_y, dst_u, dst_v;
+
+    // TODO(ncarter): Use a better error metric that measures aggregate error
+    // rather than simply max error. There seems to be slightly more error at
+    // higher resolutions, maybe due to precision issues during rasterization
+    // (or maybe more pixels = more test trials). Results are usually to an
+    // error of 1, but we must use a tolerance of 3.
+    set_color_error_tolerance(3);
+    CreateRandomCheckerboardTexture(
+        src_size, checkerboard_size, reference.Receive(), src.Receive());
+
+    gfx::Size packed_y_size, packed_uv_size;
+
+    ASSERT_TRUE(gpu_ops->AllocYUVBuffers(dst_size,
+                                         &packed_y_size,
+                                         &packed_uv_size,
+                                         dst_y.Receive(),
+                                         dst_u.Receive(),
+                                         dst_v.Receive()));
+
+    // Actually do the conversion.
+    if (use_multi_render_targets) {
+      ASSERT_TRUE(gpu_ops->TransformRGBToYV12_MRT(src,
+                                                  dst_size,
+                                                  packed_y_size,
+                                                  packed_uv_size,
+                                                  dst_y,
+                                                  dst_u,
+                                                  dst_v));
+    } else {
+      ASSERT_TRUE(gpu_ops->TransformRGBToYV12_WithoutMRT(src,
+                                                         dst_size,
+                                                         packed_y_size,
+                                                         packed_uv_size,
+                                                         dst_y,
+                                                         dst_u,
+                                                         dst_v));
+    }
+
+    // UV size (in bytes/samples) is half, rounded up.
+    gfx::Size uv_size((dst_size.width() + 1) / 2,
+                      (dst_size.height() + 1) / 2);
+
+    // Generate a reference bitmap by calling a software implementation.
+    SkBitmap reference_rgb = ToSkBitmap(reference, false);
+    SkBitmap reference_rgb_scaled;
+    if (dst_size == src_size) {
+      reference_rgb_scaled = reference_rgb;
+    } else {
+      // We'll call Copy to do the bilinear scaling if needed.
+      base::win::ScopedComPtr<IDirect3DSurface9> reference_scaled;
+      ASSERT_TRUE(
+          d3d_utils::CreateTemporaryLockableSurface(
+              device(), dst_size, reference_scaled.Receive()));
+      ASSERT_TRUE(
+          gpu_ops->Copy(src, reference_scaled, dst_size));
+      BeforeLockWorkaround();
+      reference_rgb_scaled = ToSkBitmap(reference_scaled, false);
+    }
+
+    scoped_ptr<uint8[]> reference_y(new uint8[dst_size.GetArea()]);
+    scoped_ptr<uint8[]> reference_u(new uint8[uv_size.GetArea()]);
+    scoped_ptr<uint8[]> reference_v(new uint8[uv_size.GetArea()]);
+    reference_rgb_scaled.lockPixels();
+    media::ConvertRGB32ToYUV_SSE2_Reference(
+        reinterpret_cast<uint8*>(reference_rgb_scaled.getAddr32(0, 0)),
+        &reference_y[0],
+        &reference_u[0],
+        &reference_v[0],
+        dst_size.width(),
+        dst_size.height(),
+        reference_rgb_scaled.rowBytes(),
+        dst_size.width(),
+        uv_size.width());
+    reference_rgb_scaled.unlockPixels();
+
+    // Check for equality of the reference and the actual.
+    AssertSame(dst_size.width(), dst_size.height(), &reference_y[0], dst_y);
+    AssertSame(uv_size.width(), uv_size.height(), &reference_u[0], dst_u);
+    AssertSame(uv_size.width(), uv_size.height(), &reference_v[0], dst_v);
+
+    if (kDumpImagesOnFailure && HasFatalFailure()) {
+      // Note that this will dump the full u and v buffers, including
+      // extra columns added due to packing. That means up to 7 extra
+      // columns for uv, and up to 3 extra columns for y.
+      WritePNGFile(reference_rgb,
+                   FilePath(FILE_PATH_LITERAL("test_fail_src.png")));
+      WritePNGFile(reference_rgb_scaled,
+                   FilePath(FILE_PATH_LITERAL("test_fail_src_scaled.png")));
+      WritePNGFile(ToSkBitmap(dst_y, true),
+                   FilePath(FILE_PATH_LITERAL("test_fail_y.png")));
+      WritePNGFile(ToSkBitmap(dst_u, true),
+                   FilePath(FILE_PATH_LITERAL("test_fail_u.png")));
+      WritePNGFile(ToSkBitmap(dst_v, true),
+                   FilePath(FILE_PATH_LITERAL("test_fail_v.png")));
+    }
+  }
+
+  int color_error_tolerance_;
   uniform_int_distribution<DWORD> random_dword_;
   std::mt19937 rng_;
   base::ScopedNativeLibrary d3d_module_;
   base::win::ScopedComPtr<IDirect3DDevice9Ex> device_;
 };
 
 // Fails on some bots because Direct3D isn't allowed.
 TEST_P(AcceleratedSurfaceTransformerTest, Init) {
   SCOPED_TRACE(GetAdapterInfo());
   AcceleratedSurfaceTransformer gpu_ops;
   ASSERT_TRUE(gpu_ops.Init(device()));
+
+  WarnOnMissingFeatures(&gpu_ops);
 };
 
 // Fails on some bots because Direct3D isn't allowed.
 TEST_P(AcceleratedSurfaceTransformerTest, TestConsistentRandom) {
   // This behavior should be the same for every execution on every machine.
   // Otherwise tests might be flaky and impossible to debug.
   SeedRandom("AcceleratedSurfaceTransformerTest.TestConsistentRandom");
   ASSERT_EQ(2922058934, RandomColor());
 
   SeedRandom("AcceleratedSurfaceTransformerTest.TestConsistentRandom");
@@ skipping 16 matching lines @@
 
   uniform_int_distribution<int> size(1, 512);
 
   for (int i = 0; i < 100; ++i) {
     ASSERT_NO_FATAL_FAILURE(
         DoCopyInvertedTest(&t, gfx::Size(size(rng_), size(rng_))))
             << "At iteration " << i;
   }
 }
 
-
 // Fails on some bots because Direct3D isn't allowed.
 // Fails on other bots because of ResizeBilinear symmetry failures.
 // Should pass, at least, on NVIDIA Quadro 600.
 TEST_P(AcceleratedSurfaceTransformerTest, MixedOperations) {
   SCOPED_TRACE(GetAdapterInfo());
   SeedRandom("MixedOperations");
 
   AcceleratedSurfaceTransformer t;
   ASSERT_TRUE(t.Init(device()));
 
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(256, 256), gfx::Size(255, 255), 1));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(256, 256), gfx::Size(255, 255), 2));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(20, 107)));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(256, 256), gfx::Size(255, 255), 5));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(256, 256), gfx::Size(64, 64), 5));
   ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&t, gfx::Size(128, 128), 1));
+  ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(255, 255), gfx::Size(3, 3), 1));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(1412, 124)));
   ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&t, gfx::Size(100, 200), 1));
+  ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(255, 255), gfx::Size(257, 257), 1));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(255, 255), gfx::Size(257, 257), 2));
 
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(1512, 7)));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(255, 255), gfx::Size(257, 257), 5));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(150, 256), gfx::Size(126, 256), 8));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(1521, 3)));
   ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&t, gfx::Size(140, 181), 1));
+  ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(150, 256), gfx::Size(126, 256), 1));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(33, 712)));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(150, 256), gfx::Size(126, 8), 8));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&t, gfx::Size(33, 2)));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&t, gfx::Size(200, 256), gfx::Size(126, 8), 8));
 }
@@ skipping 25 matching lines @@
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&gpu_ops, gfx::Size(lo, h), gfx::Size(lo, lo), 1));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&gpu_ops, gfx::Size(lo, lo), gfx::Size(w, lo), lo));
   ASSERT_NO_FATAL_FAILURE(
       DoResizeBilinearTest(&gpu_ops, gfx::Size(lo, lo), gfx::Size(lo, h), lo));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&gpu_ops, gfx::Size(w, lo)));
   ASSERT_NO_FATAL_FAILURE(
       DoCopyInvertedTest(&gpu_ops, gfx::Size(lo, h)));
+
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(w, lo), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(lo, h), 1));
+
 }
 
 // Exercises ResizeBilinear with random minification cases where the
 // aspect ratio does not change.
 //
 // Fails on some bots because Direct3D isn't allowed.
 // Fails on other bots because of StretchRect symmetry failures.
 // Should pass, at least, on NVIDIA Quadro 600.
 TEST_P(AcceleratedSurfaceTransformerTest, MinifyUniform) {
   SCOPED_TRACE(GetAdapterInfo());
   SeedRandom("MinifyUniform");
 
   AcceleratedSurfaceTransformer gpu_ops;
   ASSERT_TRUE(gpu_ops.Init(device()));
 
-  int dims[] = { 21, 63, 64, 65, 99, 127, 128, 129, 192, 255, 256, 257};
-  int checkerboards[] = {1, 2, 3, 9};
+  const int dims[] = {21, 63, 64, 65, 99, 127, 128, 129, 192, 255, 256, 257};
+  const int checkerboards[] = {1, 2, 3, 9};
   uniform_int_distribution<int> dim(0, arraysize(dims) - 1);
   uniform_int_distribution<int> checkerboard(0, arraysize(checkerboards) - 1);
 
   for (int i = 0; i < 300; i++) {
     // Widths are picked so that dst is smaller than src.
     int dst_width = dims[dim(rng_)];
     int src_width = dims[dim(rng_)];
     if (src_width < dst_width)
       std::swap(dst_width, src_width);
 
@@ skipping 19 matching lines @@
 // image, but for the current implementation of ResizeBilinear, this does not
 // seem to be true (fails on NVIDIA Quadro 600; passes on
 // Intel Mobile 965 Express)
 TEST_P(AcceleratedSurfaceTransformerTest, DISABLED_MagnifyUniform) {
   SCOPED_TRACE(GetAdapterInfo());
   SeedRandom("MagnifyUniform");
 
   AcceleratedSurfaceTransformer gpu_ops;
   ASSERT_TRUE(gpu_ops.Init(device()));
 
-  int dims[] = {63, 64, 65, 99, 127, 128, 129, 192, 255, 256, 257};
-  int checkerboards[] = {1, 2, 3, 9};
+  const int dims[] = {63, 64, 65, 99, 127, 128, 129, 192, 255, 256, 257};
+  const int checkerboards[] = {1, 2, 3, 9};
   uniform_int_distribution<int> dim(0, arraysize(dims) - 1);
   uniform_int_distribution<int> checkerboard(0, arraysize(checkerboards) - 1);
 
   for (int i = 0; i < 50; i++) {
     // Widths are picked so that src is smaller than dst.
     int dst_width = dims[dim(rng_)];
     int src_width = dims[dim(rng_)];
     if (dst_width < src_width)
       std::swap(src_width, dst_width);
 
     int dst_height = dims[dim(rng_)];
     int src_height = static_cast<int>(
         static_cast<int64>(src_width) * dst_height / dst_width);
 
     int checkerboard_size = checkerboards[checkerboard(rng_)];
 
     ASSERT_NO_FATAL_FAILURE(
         DoResizeBilinearTest(&gpu_ops,
             gfx::Size(src_width, src_height),  // Src size (smaller)
             gfx::Size(dst_width, dst_height),  // Dst size (larger)
             checkerboard_size)) << "Failed on iteration " << i;
   }
 };
 
+TEST_P(AcceleratedSurfaceTransformerTest, RGBtoYUV) {
+  SeedRandom("RGBtoYUV");
+
+  AcceleratedSurfaceTransformer gpu_ops;
+  ASSERT_TRUE(gpu_ops.Init(device()));
+
+  // Start with some easy-to-debug cases. A checkerboard size of 1 is the
+  // best test, but larger checkerboard sizes give more insight into where
+  // a bug might be.
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(32, 32), 4));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(32, 32), 2));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(32, 32), 3));
+
+  // All cases of width (mod 8) and height (mod 8), using 1x1 checkerboard.
+  for (int w = 32; w < 40; ++w) {
+    for (int h = 32; h < 40; ++h) {
+      ASSERT_NO_FATAL_FAILURE(
+          DoYUVConversionTest(&gpu_ops, gfx::Size(w, h), 1));
+    }
+  }
+
+  // All the very small sizes which require the most shifting in the
+  // texture coordinates when doing alignment.
+  for (int w = 1; w <= 9; ++w) {
+    for (int h = 1; h <= 9; ++h) {
+      ASSERT_NO_FATAL_FAILURE(
+          DoYUVConversionTest(&gpu_ops, gfx::Size(w, h), 1));
+    }
+  }
+
+  // Random medium dimensions.
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(10, 142), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(124, 333), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(853, 225), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(231, 412), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(512, 128), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1024, 768), 1));
+
+  // Common video/monitor resolutions
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(800, 768), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1024, 768), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1280, 720), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1280, 720), 2));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1920, 1080), 1));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(1920, 1080), 2));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionTest(&gpu_ops, gfx::Size(2048, 1536), 1));
+}
+
+TEST_P(AcceleratedSurfaceTransformerTest, RGBtoYUVScaled) {
+  SeedRandom("RGBtoYUVScaled");
+
+  AcceleratedSurfaceTransformer gpu_ops;
+  ASSERT_TRUE(gpu_ops.Init(device()));
+
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionScaleTest(&gpu_ops, gfx::Size(32, 32), gfx::Size(64, 64)));
+
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionScaleTest(&gpu_ops, gfx::Size(32, 32), gfx::Size(16, 16)));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionScaleTest(&gpu_ops, gfx::Size(32, 32), gfx::Size(24, 24)));
+  ASSERT_NO_FATAL_FAILURE(
+      DoYUVConversionScaleTest(&gpu_ops, gfx::Size(32, 32), gfx::Size(48, 48)));
+}
+
 namespace {
 
 // Used to suppress test on Windows versions prior to Vista.
 std::vector<int> WindowsVersionIfVistaOrBetter() {
   std::vector<int> result;
   if (base::win::GetVersion() >= base::win::VERSION_VISTA) {
     result.push_back(base::win::GetVersion());
   }
   return result;
 }
 
 }  // namespace
 
 INSTANTIATE_TEST_CASE_P(VistaAndUp,
                         AcceleratedSurfaceTransformerTest,
                         ::testing::ValuesIn(WindowsVersionIfVistaOrBetter()));