Allow putting NV12 software GMBs into overlays on windows

gpu/ipc/service/direct_composition_surface_win.cc

 // Copyright 2017 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 "gpu/ipc/service/direct_composition_surface_win.h"
 
 #include <d3d11_1.h>
 #include <dcomptypes.h>
 
 #include <deque>
@@ skipping 12 matching lines @@
 #include "ui/display/display_switches.h"
 #include "ui/gfx/color_space_win.h"
 #include "ui/gfx/geometry/size_conversions.h"
 #include "ui/gfx/native_widget_types.h"
 #include "ui/gfx/transform.h"
 #include "ui/gl/dc_renderer_layer_params.h"
 #include "ui/gl/egl_util.h"
 #include "ui/gl/gl_angle_util_win.h"
 #include "ui/gl/gl_context.h"
 #include "ui/gl/gl_image_dxgi.h"
+#include "ui/gl/gl_image_memory.h"
 #include "ui/gl/gl_surface_egl.h"
 #include "ui/gl/scoped_make_current.h"
 
 #ifndef EGL_ANGLE_flexible_surface_compatibility
 #define EGL_ANGLE_flexible_surface_compatibility 1
 #define EGL_FLEXIBLE_SURFACE_COMPATIBILITY_SUPPORTED_ANGLE 0x33A6
 #endif /* EGL_ANGLE_flexible_surface_compatibility */
 
 #ifndef EGL_ANGLE_d3d_texture_client_buffer
 #define EGL_ANGLE_d3d_texture_client_buffer 1
@@ skipping 158 matching lines @@
   float swap_chain_scale_x() const { return swap_chain_scale_x_; }
   float swap_chain_scale_y() const { return swap_chain_scale_y_; }
   const base::win::ScopedComPtr<IDXGISwapChain1>& swap_chain() const {
     return swap_chain_;
   }
 
  private:
   using PFN_DCOMPOSITION_CREATE_SURFACE_HANDLE =
       HRESULT(WINAPI*)(DWORD, SECURITY_ATTRIBUTES*, HANDLE*);
 
+  bool UploadMemoryImages(gl::GLImageMemory* y_image_memory,

[sunnyps1, 2017/05/04 22:26:56]

nit: UploadVideoImages

+                          gl::GLImageMemory* uv_image_memory);
   // Returns true if the video processor changed.
   bool InitializeVideoProcessor(const gfx::Size& in_size,
                                 const gfx::Size& out_size);
   void ReallocateSwapChain(bool yuy2);
   bool ShouldBeYUY2();
 
   DCLayerTree* surface_;
   PFN_DCOMPOSITION_CREATE_SURFACE_HANDLE create_surface_handle_function_;
 
   gfx::Size swap_chain_size_;
   gfx::Size processor_input_size_;
   gfx::Size processor_output_size_;
   bool is_yuy2_swapchain_ = false;
 
   // This is the scale from the swapchain size to the size of the contents
   // onscreen.
   float swap_chain_scale_x_ = 0.0f;
   float swap_chain_scale_y_ = 0.0f;
 
   PresentationHistory presentation_history_;
   bool failed_to_create_yuy2_swapchain_ = false;
   int frames_since_color_space_change_ = 0;
 
   // This is the GLImage that was presented in the last frame.
-  scoped_refptr<gl::GLImageDXGI> last_gl_image_;
+  scoped_refptr<gl::GLImage> last_gl_image_;
+
+  base::win::ScopedComPtr<ID3D11Texture2D> staging_texture_;
+  gfx::Size staging_texture_size_;
 
   base::win::ScopedComPtr<ID3D11Device> d3d11_device_;
   base::win::ScopedComPtr<IDXGISwapChain1> swap_chain_;
   base::win::ScopedComPtr<ID3D11VideoProcessorOutputView> out_view_;
   base::win::ScopedComPtr<ID3D11VideoProcessor> video_processor_;
   base::win::ScopedComPtr<ID3D11VideoProcessorEnumerator>
       video_processor_enumerator_;
   base::win::ScopedComPtr<ID3D11VideoDevice> video_device_;
   base::win::ScopedComPtr<ID3D11VideoContext> video_context_;
 
@@ skipping 91 matching lines @@
 
   if (is_yuy2_swapchain_) {
     // Switch to BGRA once 3/4 of presents are composed.
     return composition_count < (PresentationHistory::kPresentsToStore * 3 / 4);
   } else {
     // Switch to YUY2 once 3/4 are using overlays (or unknown).
     return composition_count < (PresentationHistory::kPresentsToStore / 4);
   }
 }
 
+bool DCLayerTree::SwapChainPresenter::UploadMemoryImages(
+    gl::GLImageMemory* y_image_memory,
+    gl::GLImageMemory* uv_image_memory) {
+  gfx::Size texture_size = y_image_memory->GetSize();
+  gfx::Size uv_image_size = uv_image_memory->GetSize();
+  if (uv_image_size.height() != texture_size.height() / 2 ||
+      uv_image_size.width() != texture_size.width() / 2 ||
+      y_image_memory->format() != gfx::BufferFormat::R_8 ||
+      uv_image_memory->format() != gfx::BufferFormat::RG_88) {
+    DVLOG(ERROR) << "Invalid NV12 GLImageMemory properties.";
+    return false;
+  }
+
+  if (!staging_texture_ || (staging_texture_size_ != texture_size)) {
+    staging_texture_.Reset();
+    D3D11_TEXTURE2D_DESC desc = {};
+    desc.Width = texture_size.width();
+    desc.Height = texture_size.height();
+    desc.Format = DXGI_FORMAT_NV12;
+    desc.MipLevels = 1;
+    desc.ArraySize = 1;
+    desc.Usage = D3D11_USAGE_DYNAMIC;
+
+    // This isn't actually bound to a decoder, but dynamic textures need
+    // BindFlags to be nonzero and D3D11_BIND_DECODER also works when creating
+    // a VideoProcessorInputView.
+    desc.BindFlags = D3D11_BIND_DECODER;
+    desc.CPUAccessFlags = D3D11_CPU_ACCESS_WRITE;
+    desc.MiscFlags = 0;
+    desc.SampleDesc.Count = 1;
+    base::win::ScopedComPtr<ID3D11Texture2D> texture;
+    HRESULT hr = d3d11_device_->CreateTexture2D(&desc, nullptr,
+                                                staging_texture_.Receive());
+    CHECK(SUCCEEDED(hr)) << std::hex << hr;

[sunnyps1, 2017/05/04 22:26:56]

nit: can you write a more descriptive error message? or omit it completely?

+    staging_texture_size_ = texture_size;
+  }
+  base::win::ScopedComPtr<ID3D11DeviceContext> context;
+  d3d11_device_->GetImmediateContext(context.Receive());
+  D3D11_MAPPED_SUBRESOURCE mapped_resource;
+  HRESULT hr = context->Map(staging_texture_.Get(), 0, D3D11_MAP_WRITE_DISCARD,
+                            0, &mapped_resource);
+  CHECK(SUCCEEDED(hr)) << std::hex << hr;

[sunnyps1, 2017/05/04 22:26:56]

nit: same here

+
+  size_t dest_stride = mapped_resource.RowPitch;
+  for (int y = 0; y < texture_size.height(); y++) {
+    const uint8_t* y_source =
+        y_image_memory->memory() + y * y_image_memory->stride();
+    uint8_t* dest =
+        reinterpret_cast<uint8_t*>(mapped_resource.pData) + dest_stride * y;
+    memcpy(dest, y_source, texture_size.width());
+  }
+
+  uint8_t* uv_dest_plane_start =
+      reinterpret_cast<uint8_t*>(mapped_resource.pData) +
+      dest_stride * texture_size.height();
+  for (int y = 0; y < texture_size.height() / 2; y++) {

[sunnyps1, 2017/05/04 22:26:56]

nit: uv_image_size.height() instead of texture_size.height() / 2

+    const uint8_t* uv_source =
+        uv_image_memory->memory() + y * uv_image_memory->stride();
+    uint8_t* dest = uv_dest_plane_start + dest_stride * y;
+    memcpy(dest, uv_source, texture_size.width());
+  }
+  context->Unmap(staging_texture_.Get(), 0);
+  return true;
+}
+
 void DCLayerTree::SwapChainPresenter::PresentToSwapChain(
     const ui::DCRendererLayerParams& params) {
   gl::GLImageDXGI* image_dxgi =
       gl::GLImageDXGI::FromGLImage(params.image[0].get());
-  DCHECK(image_dxgi);
+  gl::GLImageMemory* y_image_memory = nullptr;
+  gl::GLImageMemory* uv_image_memory = nullptr;
+  if (params.image.size() >= 2) {
+    y_image_memory = gl::GLImageMemory::FromGLImage(params.image[0].get());
+    uv_image_memory = gl::GLImageMemory::FromGLImage(params.image[1].get());
+  }
+
+  if (!image_dxgi && (!y_image_memory || !uv_image_memory)) {
+    DLOG(ERROR) << "Video GLImages are missing";
+    last_gl_image_ = nullptr;
+    return;
+  }
 
   // Swap chain size is the minimum of the on-screen size and the source
   // size so the video processor can do the minimal amount of work and
   // the overlay has to read the minimal amount of data.
   // DWM is also less likely to promote a surface to an overlay if it's
   // much larger than its area on-screen.
   gfx::Rect bounds_rect = params.rect;
   gfx::Size ceiled_input_size = gfx::ToCeiledSize(params.contents_rect.size());
   gfx::Size swap_chain_size = bounds_rect.size();
   swap_chain_size.SetToMin(ceiled_input_size);
 
   // YUY2 surfaces must have an even width.
   if (swap_chain_size.width() % 2 == 1)
     swap_chain_size.set_width(swap_chain_size.width() + 1);
 
   InitializeVideoProcessor(ceiled_input_size, swap_chain_size);
 
   bool yuy2_swapchain = ShouldBeYUY2();
   bool first_present = false;
   if (!swap_chain_ || swap_chain_size_ != swap_chain_size ||
       ((yuy2_swapchain != is_yuy2_swapchain_) &&
        !failed_to_create_yuy2_swapchain_)) {
     first_present = true;
     swap_chain_size_ = swap_chain_size;
     swap_chain_.Reset();
     ReallocateSwapChain(yuy2_swapchain);
-  } else if (last_gl_image_ == image_dxgi) {
+  } else if (last_gl_image_ == params.image[0]) {

[sunnyps1, 2017/05/04 22:26:56]

Does this check also work for YUV? Can't UV change but not Y?

     // The swap chain is presenting the same image as last swap, which means
     // that the image was never returned to the video decoder and should have
     // the same contents as last time. It shouldn't need to be redrawn.
     return;
   }
 
-  last_gl_image_ = image_dxgi;
+  last_gl_image_ = params.image[0];
+
+  base::win::ScopedComPtr<ID3D11Texture2D> input_texture;
+  UINT input_level;
+  if (image_dxgi) {
+    input_texture = image_dxgi->texture();
+    input_level = (UINT)image_dxgi->level();
+    staging_texture_.Reset();
+  } else {
+    DCHECK(!!y_image_memory);

[sunnyps1, 2017/05/04 22:26:56]

nit: DCHECK(y_image_memory)

+    DCHECK(!!uv_image_memory);
+    if (!UploadMemoryImages(y_image_memory, uv_image_memory))
+      return;
+    DCHECK(staging_texture_);
+    input_texture = staging_texture_;
+    input_level = 0;
+  }
 
   if (!out_view_) {
     base::win::ScopedComPtr<ID3D11Texture2D> texture;
     swap_chain_->GetBuffer(0, IID_PPV_ARGS(texture.Receive()));
     D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC out_desc = {};
     out_desc.ViewDimension = D3D11_VPOV_DIMENSION_TEXTURE2D;
     out_desc.Texture2D.MipSlice = 0;
     HRESULT hr = video_device_->CreateVideoProcessorOutputView(
         texture.Get(), video_processor_enumerator_.Get(), &out_desc,
         out_view_.Receive());
@@ skipping 36 matching lines @@
       D3D11_VIDEO_PROCESSOR_COLOR_SPACE d3d11_color_space =
           gfx::ColorSpaceWin::GetD3D11ColorSpace(output_color_space);
       video_context_->VideoProcessorSetOutputColorSpace(video_processor_.Get(),
                                                         &d3d11_color_space);
     }
   }
 
   {
     D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC in_desc = {};
     in_desc.ViewDimension = D3D11_VPIV_DIMENSION_TEXTURE2D;
-    in_desc.Texture2D.ArraySlice = (UINT)image_dxgi->level();
+    in_desc.Texture2D.ArraySlice = input_level;
     base::win::ScopedComPtr<ID3D11VideoProcessorInputView> in_view;
     HRESULT hr = video_device_->CreateVideoProcessorInputView(
-        image_dxgi->texture().Get(), video_processor_enumerator_.Get(),
-        &in_desc, in_view.Receive());
+        input_texture.Get(), video_processor_enumerator_.Get(), &in_desc,
+        in_view.Receive());
     CHECK(SUCCEEDED(hr));
 
     D3D11_VIDEO_PROCESSOR_STREAM stream = {};
     stream.Enable = true;
     stream.OutputIndex = 0;
     stream.InputFrameOrField = 0;
     stream.PastFrames = 0;
     stream.FutureFrames = 0;
     stream.pInputSurface = in_view.Get();
     RECT dest_rect = gfx::Rect(swap_chain_size).ToRECT();
@@ skipping 297 matching lines @@
   // The overall visual tree has one clip visual for every overlay (including
   // the main plane). The clip visuals are in z_order and are all children of
   // a root visual. Each clip visual has a child visual that has the actual
   // plane content.
 
   for (size_t i = 0; i < pending_overlays_.size(); i++) {
     ui::DCRendererLayerParams& params = *pending_overlays_[i];
     VisualInfo* visual_info = &visual_info_[i];
 
     InitVisual(i);
-    if (params.image.size() > 0 && params.image[0] &&
-        params.image[0]->GetType() == gl::GLImage::Type::DXGI_IMAGE) {
+    if (params.image.size() >= 1 && params.image[0]) {
       UpdateVisualForVideo(visual_info, params);
     } else if (params.image.empty()) {
       UpdateVisualForBackbuffer(visual_info, params);
     } else {
       CHECK(false);
     }
     UpdateVisualClip(visual_info, params);
   }
 
   HRESULT hr = dcomp_device_->Commit();
@@ skipping 382 matching lines @@
 DirectCompositionSurfaceWin::GetWindowTaskRunnerForTesting() {
   return child_window_.GetTaskRunnerForTesting();
 }
 
 base::win::ScopedComPtr<IDXGISwapChain1>
 DirectCompositionSurfaceWin::GetLayerSwapChainForTesting(size_t index) const {
   return layer_tree_->GetLayerSwapChainForTesting(index);
 }
 
 }  // namespace gpu