When the render mode is Direct2D ThebesLayerD3D9 should use D3D9/D3D10 interop in order to allow cairo to use the Direct2D backend to draw directly to the ThebesLayer's texture.

This is needed to be able to enable D3D9 layers for D2D enabled machines. This in turn is needed for fast webgl, hardware accelerated YUV->RGB conversion and some other things.

This patch allows creating a Direct2D surface from a SharedHandle which can come from another device. It also allows the D2D device to be flushed in such a way that the surface is ready to be used by another device.

This does the job. These patches depend on the cairo_d2d_device_t and D3D9Ex patches. We may want to spend some time making CanvasLayer use this as well, so we get super-fast canvas too. But that will require some restructuring. It's probably best combined with the work to make ANGLE interop with us.

Comment on attachment 464302 [details] [diff] [review] Part 1: Allow creating a D2D surface from a shared handle >+void >+cairo_d2d_flush_device(cairo_device_t *device) >+{ >+ cairo_d2d_device_t *d2d_device = reinterpret_cast<cairo_d2d_device_t*>(device); >+ // Here it becomes interesting, this flush method is generally called when >+ // interop is going on between our device and another device. The >+ // synchronisation between these devices is not always that great. The >+ // device flush method may flush the device's command queue, but it gives >+ // no guarantee that the device will actually be done with those commands, >+ // and so the surface may still not be complete when the external device >+ // chooses to use it. The EVENT query will actually tell us when the GPU >+ // is completely done with our commands. >+ D3D10_QUERY_DESC queryDesc; >+ queryDesc.MiscFlags = 0; >+ queryDesc.Query = D3D10_QUERY_EVENT; >+ RefPtr<ID3D10Query> query; >+ >+ d2d_device->mD3D10Device->CreateQuery(&queryDesc, &query); >+ Probably worth adding a comment that says "Begin() is disabled with D3D10_QUERY_EVENT" >+ query->End(); Should we flush() the device here? >+ >+ BOOL done = FALSE; >+ while (!done) { >+ // This will return S_OK and done = FALSE when the GPU is not done, and >+ // S_OK and done = TRUE when the GPU is done. Any other return value >+ // means we need to break out or risk an infinite loop. >+ if (FAILED(query->GetData(&done, sizeof(BOOL), 0))) { >+ break; >+ } Will this spin waiting for GetData to give done = TRUE; >+cairo_surface_t * >+cairo_d2d_surface_create_for_handle(cairo_device_t *device, HANDLE handle, cairo_content_t content) >+{ [snip] >+ hr = d2d_device->mD3D10Device->OpenSharedResource(handle, >+ __uuidof(ID3D10Resource), >+ (void**)&newSurf->surface); >+ >+ RefPtr<ID3D10Texture2D> texture; >+ RefPtr<IDXGISurface> dxgiSurface; >+ D2D1_BITMAP_PROPERTIES bitProps; >+ D2D1_RENDER_TARGET_PROPERTIES props; >+ DXGI_FORMAT format; >+ DXGI_SURFACE_DESC desc; >+ >+ if (FAILED(hr)) { >+ goto FAIL_CREATEHANDLE; >+ } Move this closer to the call that you're testing the return value of. >diff --git a/gfx/cairo/cairo/src/cairo-win32.h b/gfx/cairo/cairo/src/cairo-win32.h >--- a/gfx/cairo/cairo/src/cairo-win32.h >+++ b/gfx/cairo/cairo/src/cairo-win32.h > /** >+ * Flushes a D3D device. In most cases the surface backend will do this >+ * internally, but when using a surfaces created from a shared handle this >+ * should be executed manually when a different device is going to be accessing >+ * the same surface data >+ */ >+void >+cairo_d2d_flush_device(cairo_device_t *device); finish may be a better name to use here. finish matches the semantics of glFlush()/glFinish()

(In reply to comment #5) > Comment on attachment 464302 [details] [diff] [review] > Should we flush() the device here? GetData does this unless you specify some 'NOFLUSH' flag. > > >+ > >+ BOOL done = FALSE; > >+ while (!done) { > >+ // This will return S_OK and done = FALSE when the GPU is not done, and > >+ // S_OK and done = TRUE when the GPU is done. Any other return value > >+ // means we need to break out or risk an infinite loop. > >+ if (FAILED(query->GetData(&done, sizeof(BOOL), 0))) { > >+ break; > >+ } > > Will this spin waiting for GetData to give done = TRUE; It will, this is when the GPU is completely done executing commands. > > >+ RefPtr<ID3D10Texture2D> texture; > >+ RefPtr<IDXGISurface> dxgiSurface; > >+ D2D1_BITMAP_PROPERTIES bitProps; > >+ D2D1_RENDER_TARGET_PROPERTIES props; > >+ DXGI_FORMAT format; > >+ DXGI_SURFACE_DESC desc; > >+ > >+ if (FAILED(hr)) { > >+ goto FAIL_CREATEHANDLE; > >+ } > > Move this closer to the call that you're testing the return value of. This mainly means moving the declarations to above the call. I don't think we can put declarations after a forward goto, can we?

Comment on attachment 464310 [details] [diff] [review] Part 3: Use D2D interop in D3D9 layers when D2D is enabled Looks fine, but: >+ if (gfxWindowsPlatform::GetPlatform()->GetRenderMode() == >+ gfxWindowsPlatform::RENDER_DIRECT2D) { >+ if (mD3DManager->deviceManager()->IsD3D9Ex()) { >+ // Strange? D2D but no D3D9Ex? This comment is misplaced/misstated -- you maybe want something like "// if we have d2d we should have d3d9ex"..

Pushed http://hg.mozilla.org/mozilla-central/rev/2cd5d96fcc70. Pushed http://hg.mozilla.org/mozilla-central/rev/91bf043c5365. Pushed http://hg.mozilla.org/mozilla-central/rev/72ec4df3de76.