Use separate render and decode D3D11Devices

See #1424
2026-02-16 10:40:59 +00:00 · 2026-01-19 14:25:08 -06:00
parent 218ffc2d55
commit f6e08f8a43
2 changed files with 339 additions and 143 deletions
--- a/app/streaming/video/ffmpeg-renderers/d3d11va.cpp
+++ b/app/streaming/video/ffmpeg-renderers/d3d11va.cpp
@@ -108,24 +108,144 @@ D3D11VARenderer::~D3D11VARenderer()
    m_VideoBlendState.Reset();

    m_PreviousFrameRenderedFence.Reset();
-    m_DecoderShaderBindFence.Reset();
+    m_DecodeD2RFence.Reset();
+    m_DecodeR2DFence.Reset();
+    m_RenderD2RFence.Reset();
+    m_RenderR2DFence.Reset();

    m_RenderTargetView.Reset();
    m_SwapChain.Reset();

+    m_RenderSharedTextureArray.Reset();
+
    av_buffer_unref(&m_HwDeviceContext);
+    m_DecodeDevice.Reset();
+    m_DecodeDeviceContext.Reset();

    // Force destruction of the swapchain immediately
-    if (m_DeviceContext != nullptr) {
-        m_DeviceContext->ClearState();
-        m_DeviceContext->Flush();
+    if (m_RenderDeviceContext != nullptr) {
+        m_RenderDeviceContext->ClearState();
+        m_RenderDeviceContext->Flush();
    }

-    m_Device.Reset();
-    m_DeviceContext.Reset();
+    m_RenderDevice.Reset();
+    m_RenderDeviceContext.Reset();
    m_Factory.Reset();
 }

+bool D3D11VARenderer::createSharedFencePair(UINT64 initialValue, ID3D11Device5* dev1, ID3D11Device5* dev2, ComPtr<ID3D11Fence>& dev1Fence, ComPtr<ID3D11Fence>& dev2Fence)
+{
+    HRESULT hr;
+
+    hr = dev1->CreateFence(initialValue, D3D11_FENCE_FLAG_SHARED, IID_PPV_ARGS(&dev1Fence));
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11Device5::CreateFence() failed: %x",
+                     hr);
+        return false;
+    }
+
+    HANDLE fenceHandle;
+    hr = dev1Fence->CreateSharedHandle(nullptr, GENERIC_ALL, nullptr, &fenceHandle);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11Fence::CreateSharedHandle() failed: %x",
+                     hr);
+        dev1Fence.Reset();
+        return false;
+    }
+
+    hr = dev2->OpenSharedFence(fenceHandle, IID_PPV_ARGS(&dev2Fence));
+    CloseHandle(fenceHandle);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11Device5::OpenSharedFence() failed: %x",
+                     hr);
+        dev1Fence.Reset();
+        return false;
+    }
+
+    return true;
+}
+
+bool D3D11VARenderer::setupSharedDevice(IDXGIAdapter1* adapter)
+{
+    const D3D_FEATURE_LEVEL supportedFeatureLevels[] = { D3D_FEATURE_LEVEL_11_1, D3D_FEATURE_LEVEL_11_0 };
+    D3D_FEATURE_LEVEL featureLevel;
+    HRESULT hr;
+    ComPtr<ID3D11Device> device;
+    ComPtr<ID3D11DeviceContext> deviceContext;
+    bool success = false;
+
+    // We don't support cross-device sharing without fences
+    if (m_FenceType == SupportedFenceType::None) {
+        return false;
+    }
+
+    // If we're going to use separate devices for decoding and rendering, create the decoding device
+    hr = D3D11CreateDevice(adapter,
+                           D3D_DRIVER_TYPE_UNKNOWN,
+                           nullptr,
+                           D3D11_CREATE_DEVICE_VIDEO_SUPPORT
+#ifdef QT_DEBUG
+                               | D3D11_CREATE_DEVICE_DEBUG
+#endif
+                           ,
+                           supportedFeatureLevels,
+                           ARRAYSIZE(supportedFeatureLevels),
+                           D3D11_SDK_VERSION,
+                           &device,
+                           &featureLevel,
+                           &deviceContext);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "D3D11CreateDevice() failed: %x",
+                     hr);
+        return false;
+    }
+
+    hr = device.As(&m_DecodeDevice);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11Device::QueryInterface(ID3D11Device1) failed: %x",
+                     hr);
+        goto Exit;
+    }
+
+    hr = deviceContext.As(&m_DecodeDeviceContext);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11DeviceContext::QueryInterface(ID3D11DeviceContext1) failed: %x",
+                     hr);
+        goto Exit;
+    }
+
+    // Create our decode->render fence
+    m_DecodeRenderSyncFenceValue = 0;
+    if (!createSharedFencePair(1, m_DecodeDevice.Get(), m_RenderDevice.Get(), m_DecodeD2RFence, m_RenderD2RFence)) {
+        goto Exit;
+    }
+
+    // Create our render->decode fence
+    if (!createSharedFencePair(0, m_DecodeDevice.Get(), m_RenderDevice.Get(), m_DecodeR2DFence, m_RenderR2DFence)) {
+        goto Exit;
+    }
+
+    success = true;
+Exit:
+    if (!success) {
+        m_DecodeD2RFence.Reset();
+        m_RenderD2RFence.Reset();
+        m_DecodeR2DFence.Reset();
+        m_RenderR2DFence.Reset();
+
+        m_DecodeDevice.Reset();
+        m_RenderDevice.Reset();
+    }
+
+    return success;
+}
+
 bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapterNotFound)
 {
    const D3D_FEATURE_LEVEL supportedFeatureLevels[] = { D3D_FEATURE_LEVEL_11_1, D3D_FEATURE_LEVEL_11_0 };
@@ -134,10 +254,13 @@ bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapter
    DXGI_ADAPTER_DESC1 adapterDesc;
    D3D_FEATURE_LEVEL featureLevel;
    HRESULT hr;
+    ComPtr<ID3D11Device> device;
    ComPtr<ID3D11DeviceContext> deviceContext;

-    SDL_assert(!m_Device);
-    SDL_assert(!m_DeviceContext);
+    SDL_assert(!m_RenderDevice);
+    SDL_assert(!m_RenderDeviceContext);
+    SDL_assert(!m_DecodeDevice);
+    SDL_assert(!m_DecodeDeviceContext);

    hr = m_Factory->EnumAdapters1(adapterIndex, &adapter);
    if (hr == DXGI_ERROR_NOT_FOUND) {
@@ -182,7 +305,7 @@ bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapter
                           supportedFeatureLevels,
                           ARRAYSIZE(supportedFeatureLevels),
                           D3D11_SDK_VERSION,
-                           &m_Device,
+                           &device,
                           &featureLevel,
                           &deviceContext);
    if (FAILED(hr)) {
@@ -194,8 +317,6 @@ bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapter
    else if (adapterDesc.VendorId == 0x8086 && featureLevel <= D3D_FEATURE_LEVEL_11_0 && !qEnvironmentVariableIntValue("D3D11VA_ENABLED")) {
        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
                    "Avoiding D3D11VA on old pre-FL11.1 Intel GPU. Set D3D11VA_ENABLED=1 to override.");
-        m_DeviceContext.Reset();
-        m_Device.Reset();
        goto Exit;
    }
    else if (featureLevel >= D3D_FEATURE_LEVEL_11_0) {
@@ -204,27 +325,22 @@ bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapter
        m_DevicesWithFL11Support++;
    }

-    hr = deviceContext.As(&m_DeviceContext);
+    hr = device.As(&m_RenderDevice);
+    if (FAILED(hr)) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                     "ID3D11Device::QueryInterface(ID3D11Device1) failed: %x",
+                     hr);
+        goto Exit;
+    }
+
+    hr = deviceContext.As(&m_RenderDeviceContext);
    if (FAILED(hr)) {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                     "ID3D11DeviceContext::QueryInterface(ID3D11DeviceContext1) failed: %x",
                     hr);
-        m_DeviceContext.Reset();
-        m_Device.Reset();
        goto Exit;
    }

-    if (Utils::getEnvironmentVariableOverride("D3D11VA_FORCE_BIND", &m_BindDecoderOutputTextures)) {
-        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
-                    "Using D3D11VA_FORCE_BIND to override default bind/copy logic");
-    }
-    else {
-        // Skip copying to our own internal texture on Intel GPUs due to
-        // significant performance impact of the extra copy. See:
-        // https://github.com/moonlight-stream/moonlight-qt/issues/1304
-        m_BindDecoderOutputTextures = adapterDesc.VendorId == 0x8086;
-    }
-
    // Check which fence types are supported by this GPU
    {
        m_FenceType = SupportedFenceType::None;
@@ -245,62 +361,72 @@ bool D3D11VARenderer::createDeviceByAdapterIndex(int adapterIndex, bool* adapter
        }

        if (m_FenceType != SupportedFenceType::None) {
-            ComPtr<ID3D11Device5> device5;
-            ComPtr<ID3D11DeviceContext4> deviceContext4;
-            if (SUCCEEDED(m_Device.As(&device5)) && SUCCEEDED(m_DeviceContext.As(&deviceContext4))) {
-                // If this GPU supports monitored fences, use one to wait until the previous frame
-                // has finished rendering before starting on the next one. This reduces latency by
-                // avoiding stalling during rendering after we've already grabbed the next frame
-                // to render, and also avoids stalling the decoder by releasing a surface back to
-                // the pool before we've finished reading from it (causing a stall if the decoder
-                // tries to write again).
-                if (m_FenceType == SupportedFenceType::Monitored) {
-                    m_PreviousFrameRenderedFenceValue = 0;
-                    hr = device5->CreateFence(m_PreviousFrameRenderedFenceValue,
-                                              m_FenceType == SupportedFenceType::Monitored ?
-                                                  D3D11_FENCE_FLAG_NONE : D3D11_FENCE_FLAG_NON_MONITORED,
-                                              IID_PPV_ARGS(&m_PreviousFrameRenderedFence));
-                    if (FAILED(hr)) {
-                        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
-                                    "ID3D11Device5::CreateFence() failed: %x",
-                                    hr);
-                        // Non-fatal
-                    }
-
-                    // Create an auto-reset event for our fence to signal
-                    m_PreviousFrameRenderedEvent.Attach(CreateEvent(NULL, FALSE, TRUE, NULL));
+            // If this GPU supports monitored fences, use one to wait until the previous frame
+            // has finished rendering before starting on the next one. This reduces latency by
+            // avoiding stalling during rendering after we've already grabbed the next frame
+            // to render, and also avoids stalling the decoder by releasing a surface back to
+            // the pool before we've finished reading from it (causing a stall if the decoder
+            // tries to write again).
+            if (m_FenceType == SupportedFenceType::Monitored) {
+                m_PreviousFrameRenderedFenceValue = 0;
+                hr = m_RenderDevice->CreateFence(m_PreviousFrameRenderedFenceValue,
+                                                 D3D11_FENCE_FLAG_NONE,
+                                                 IID_PPV_ARGS(&m_PreviousFrameRenderedFence));
+                if (FAILED(hr)) {
+                    SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
+                                "ID3D11Device5::CreateFence() failed: %x",
+                                hr);
+                    // Non-fatal
                }

-                // If we're binding output textures, we signal and wait for a fence prior to rendering
-                // as a workaround for some driver bugs that can cause the GPU driver to fail to insert
-                // a dependency between the decoder engine and the 3D engine. This seems to be a much
-                // less well-tested path in most drivers than the video->copy or video->video path.
-                if (m_BindDecoderOutputTextures) {
-                    m_DecoderShaderBindFenceValue = 0;
-                    hr = device5->CreateFence(m_DecoderShaderBindFenceValue,
-                                              m_FenceType == SupportedFenceType::Monitored ?
-                                                  D3D11_FENCE_FLAG_NONE : D3D11_FENCE_FLAG_NON_MONITORED,
-                                              IID_PPV_ARGS(&m_DecoderShaderBindFence));
-                    if (FAILED(hr)) {
-                        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
-                                    "ID3D11Device5::CreateFence() failed: %x",
-                                    hr);
-                        // Non-fatal
-                    }
-                }
+                // Create an auto-reset event for our fence to signal
+                m_PreviousFrameRenderedEvent.Attach(CreateEvent(NULL, FALSE, TRUE, NULL));
            }
        }
    }

+    if (Utils::getEnvironmentVariableOverride("D3D11VA_FORCE_BIND", &m_BindDecoderOutputTextures)) {
+        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
+                    "Using D3D11VA_FORCE_BIND to override default bind/copy logic");
+    }
+    else {
+        // Skip copying to our own internal texture on Intel GPUs due to
+        // significant performance impact of the extra copy. See:
+        // https://github.com/moonlight-stream/moonlight-qt/issues/1304
+        m_BindDecoderOutputTextures = adapterDesc.VendorId == 0x8086;
+    }
+
+    bool separateDevices;
+    if (Utils::getEnvironmentVariableOverride("D3D11VA_FORCE_SEPARATE_DEVICES", &separateDevices)) {
+        SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
+                    "Using D3D11VA_FORCE_SEPARATE_DEVICES to override default logic");
+    }
+    else {
+        D3D11_FEATURE_DATA_D3D11_OPTIONS d3d11Options;
+
+        // Check if cross-device sharing works for YUV textures and fences are supported
+        hr = m_RenderDevice->CheckFeatureSupport(D3D11_FEATURE_D3D11_OPTIONS, &d3d11Options, sizeof(d3d11Options));
+        separateDevices = SUCCEEDED(hr) && d3d11Options.ExtendedResourceSharing && m_FenceType != SupportedFenceType::None;
+    }
+
+    // If we're going to use separate devices for decoding and rendering, create the decoding device
+    if (!separateDevices || !setupSharedDevice(adapter.Get())) {
+        m_DecodeDevice = m_RenderDevice;
+        m_DecodeDeviceContext = m_RenderDeviceContext;
+        separateDevices = false;
+    }
+
    SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
                "Decoder texture access: %s (fence: %s)",
                m_BindDecoderOutputTextures ? "bind" : "copy",
                 m_FenceType == SupportedFenceType::Monitored ? "monitored" :
                    (m_FenceType == SupportedFenceType::NonMonitored ? "non-monitored" : "unsupported"));

+    SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
+                "Using %s device for decoding and rendering",
+                separateDevices ? "separate" : "shared");
+
    if (!checkDecoderSupport(adapter.Get())) {
-        m_DeviceContext.Reset();
-        m_Device.Reset();
        goto Exit;
    }
    else {
@@ -314,6 +440,12 @@ Exit:
    if (adapterNotFound != nullptr) {
        *adapterNotFound = !adapter;
    }
+    if (!success) {
+        m_RenderDeviceContext.Reset();
+        m_RenderDevice.Reset();
+        m_DecodeDeviceContext.Reset();
+        m_DecodeDevice.Reset();
+    }
    return success;
 }

@@ -379,8 +511,8 @@ bool D3D11VARenderer::initialize(PDECODER_PARAMETERS params)
        }

        if (adapterNotFound) {
-            SDL_assert(!m_Device);
-            SDL_assert(!m_DeviceContext);
+            SDL_assert(!m_RenderDevice);
+            SDL_assert(!m_RenderDeviceContext);
            return false;
        }
    }
@@ -469,7 +601,7 @@ bool D3D11VARenderer::initialize(PDECODER_PARAMETERS params)
    // Always use windowed or borderless windowed mode.. SDL does mode-setting for us in
    // full-screen exclusive mode (SDL_WINDOW_FULLSCREEN), so this actually works out okay.
    ComPtr<IDXGISwapChain1> swapChain;
-    hr = m_Factory->CreateSwapChainForHwnd(m_Device.Get(),
+    hr = m_Factory->CreateSwapChainForHwnd(m_RenderDevice.Get(),
                                           info.info.win.window,
                                           &swapChainDesc,
                                           nullptr,
@@ -514,8 +646,8 @@ bool D3D11VARenderer::initialize(PDECODER_PARAMETERS params)
        AVD3D11VADeviceContext* d3d11vaDeviceContext = (AVD3D11VADeviceContext*)deviceContext->hwctx;

        // FFmpeg will take ownership of these pointers, so we use CopyTo() to bump the ref count
-        m_Device.CopyTo(&d3d11vaDeviceContext->device);
-        m_DeviceContext.CopyTo(&d3d11vaDeviceContext->device_context);
+        m_DecodeDevice.CopyTo(&d3d11vaDeviceContext->device);
+        m_DecodeDeviceContext.CopyTo(&d3d11vaDeviceContext->device_context);

        // Set lock functions that we will use to synchronize with FFmpeg's usage of our device context
        d3d11vaDeviceContext->lock = lockContext;
@@ -568,6 +700,11 @@ bool D3D11VARenderer::prepareDecoderContextInGetFormat(AVCodecContext *context,
        d3d11vaFramesContext->BindFlags |= D3D11_BIND_SHADER_RESOURCE;
    }

+    // If we're using separate decode and render devices, we need to create shared textures
+    if (m_DecodeDevice != m_RenderDevice) {
+        d3d11vaFramesContext->MiscFlags |= D3D11_RESOURCE_MISC_SHARED | D3D11_RESOURCE_MISC_SHARED_NTHANDLE;
+    }
+
    // Mimic the logic in ff_decode_get_hw_frames_ctx() which adds an extra 3 frames
    if (framesContext->initial_pool_size) {
        framesContext->initial_pool_size += 3;
@@ -594,15 +731,17 @@ void D3D11VARenderer::renderFrame(AVFrame* frame)
 {
    // Acquire the context lock for rendering to prevent concurrent
    // access from inside FFmpeg's decoding code
-    lockContext(this);
+    if (m_DecodeDevice == m_RenderDevice) {
+        lockContext(this);
+    }

    // Clear the back buffer
    const float clearColor[4] = {0.0f, 0.0f, 0.0f, 1.0f};
-    m_DeviceContext->ClearRenderTargetView(m_RenderTargetView.Get(), clearColor);
+    m_RenderDeviceContext->ClearRenderTargetView(m_RenderTargetView.Get(), clearColor);

    // Bind the back buffer. This needs to be done each time,
    // because the render target view will be unbound by Present().
-    m_DeviceContext->OMSetRenderTargets(1, m_RenderTargetView.GetAddressOf(), nullptr);
+    m_RenderDeviceContext->OMSetRenderTargets(1, m_RenderTargetView.GetAddressOf(), nullptr);

    // Render our video frame with the aspect-ratio adjusted viewport
    renderVideo(frame);
@@ -656,8 +795,10 @@ void D3D11VARenderer::renderFrame(AVFrame* frame)
    // Present according to the decoder parameters
    hr = m_SwapChain->Present(0, flags);

-    // Release the context lock
-    unlockContext(this);
+    if (m_DecodeDevice == m_RenderDevice) {
+        // Release the context lock
+        unlockContext(this);
+    }

    if (FAILED(hr)) {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -701,16 +842,16 @@ void D3D11VARenderer::renderOverlay(Overlay::OverlayType type)
    // Bind vertex buffer
    UINT stride = sizeof(VERTEX);
    UINT offset = 0;
-    m_DeviceContext->IASetVertexBuffers(0, 1, overlayVertexBuffer.GetAddressOf(), &stride, &offset);
+    m_RenderDeviceContext->IASetVertexBuffers(0, 1, overlayVertexBuffer.GetAddressOf(), &stride, &offset);

    // Bind pixel shader and resources
-    m_DeviceContext->PSSetShader(m_OverlayPixelShader.Get(), nullptr, 0);
-    m_DeviceContext->PSSetShaderResources(0, 1, overlayTextureResourceView.GetAddressOf());
+    m_RenderDeviceContext->PSSetShader(m_OverlayPixelShader.Get(), nullptr, 0);
+    m_RenderDeviceContext->PSSetShaderResources(0, 1, overlayTextureResourceView.GetAddressOf());

    // Draw the overlay with alpha blending
-    m_DeviceContext->OMSetBlendState(m_OverlayBlendState.Get(), nullptr, 0xffffffff);
-    m_DeviceContext->DrawIndexed(6, 0, 0);
-    m_DeviceContext->OMSetBlendState(m_VideoBlendState.Get(), nullptr, 0xffffffff);
+    m_RenderDeviceContext->OMSetBlendState(m_OverlayBlendState.Get(), nullptr, 0xffffffff);
+    m_RenderDeviceContext->DrawIndexed(6, 0, 0);
+    m_RenderDeviceContext->OMSetBlendState(m_VideoBlendState.Get(), nullptr, 0xffffffff);
 }

 void D3D11VARenderer::bindVideoVertexBuffer(bool frameChanged, AVFrame* frame)
@@ -754,7 +895,7 @@ void D3D11VARenderer::bindVideoVertexBuffer(bool frameChanged, AVFrame* frame)
        D3D11_SUBRESOURCE_DATA vbData = {};
        vbData.pSysMem = verts;

-        HRESULT hr = m_Device->CreateBuffer(&vbDesc, &vbData, &m_VideoVertexBuffer);
+        HRESULT hr = m_RenderDevice->CreateBuffer(&vbDesc, &vbData, &m_VideoVertexBuffer);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreateBuffer() failed: %x",
@@ -766,7 +907,7 @@ void D3D11VARenderer::bindVideoVertexBuffer(bool frameChanged, AVFrame* frame)
    // Bind video rendering vertex buffer
    UINT stride = sizeof(VERTEX);
    UINT offset = 0;
-    m_DeviceContext->IASetVertexBuffers(0, 1, m_VideoVertexBuffer.GetAddressOf(), &stride, &offset);
+    m_RenderDeviceContext->IASetVertexBuffers(0, 1, m_VideoVertexBuffer.GetAddressOf(), &stride, &offset);
 }

 void D3D11VARenderer::bindColorConversion(bool frameChanged, AVFrame* frame)
@@ -779,10 +920,10 @@ void D3D11VARenderer::bindColorConversion(bool frameChanged, AVFrame* frame)
        switch (m_TextureFormat)
        {
        case DXGI_FORMAT_AYUV:
-            m_DeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_AYUV].Get(), nullptr, 0);
+            m_RenderDeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_AYUV].Get(), nullptr, 0);
            break;
        case DXGI_FORMAT_Y410:
-            m_DeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_Y410].Get(), nullptr, 0);
+            m_RenderDeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_Y410].Get(), nullptr, 0);
            break;
        default:
            SDL_assert(false);
@@ -790,7 +931,7 @@ void D3D11VARenderer::bindColorConversion(bool frameChanged, AVFrame* frame)
    }
    else {
        // We'll need to use the generic 4:2:0 shader for this colorspace and color range combo
-        m_DeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_YUV_420].Get(), nullptr, 0);
+        m_RenderDeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::GENERIC_YUV_420].Get(), nullptr, 0);
    }

    // If nothing has changed since last frame, we're done
@@ -835,9 +976,9 @@ void D3D11VARenderer::bindColorConversion(bool frameChanged, AVFrame* frame)
    constData.pSysMem = &constBuf;

    ComPtr<ID3D11Buffer> constantBuffer;
-    HRESULT hr = m_Device->CreateBuffer(&constDesc, &constData, &constantBuffer);
+    HRESULT hr = m_RenderDevice->CreateBuffer(&constDesc, &constData, &constantBuffer);
    if (SUCCEEDED(hr)) {
-        m_DeviceContext->PSSetConstantBuffers(0, 1, constantBuffer.GetAddressOf());
+        m_RenderDeviceContext->PSSetConstantBuffers(0, 1, constantBuffer.GetAddressOf());
    }
    else {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -849,6 +990,18 @@ void D3D11VARenderer::bindColorConversion(bool frameChanged, AVFrame* frame)

 void D3D11VARenderer::renderVideo(AVFrame* frame)
 {
+    // Insert a fence to force the render context to wait for the decode context to finish writing
+    if (m_DecodeDevice != m_RenderDevice) {
+        SDL_assert(m_DecodeD2RFence);
+        SDL_assert(m_RenderD2RFence);
+
+        lockContext(this);
+        if (SUCCEEDED(m_DecodeDeviceContext->Signal(m_DecodeD2RFence.Get(), m_DecodeRenderSyncFenceValue))) {
+            m_RenderDeviceContext->Wait(m_RenderD2RFence.Get(), m_DecodeRenderSyncFenceValue++);
+        }
+        unlockContext(this);
+    }
+
    UINT srvIndex;
    if (m_BindDecoderOutputTextures) {
        // Our indexing logic depends on a direct mapping into m_VideoTextureResourceViews
@@ -861,25 +1014,13 @@ void D3D11VARenderer::renderVideo(AVFrame* frame)
                         srvIndex);
            return;
        }
-
-        // Insert a fence to force proper synchronization between the video engine and
-        // 3D engine. Some GPU drivers (HD 4000, MTT S70) have bugs that prevent this
-        // data dependency from being handled properly on its own, which can lead to
-        // rendering artifacts and video lag (rendering old frames).
-        if (m_DecoderShaderBindFence) {
-            ComPtr<ID3D11DeviceContext4> deviceContext4;
-            if (SUCCEEDED(m_DeviceContext.As(&deviceContext4))) {
-                if (SUCCEEDED(deviceContext4->Signal(m_DecoderShaderBindFence.Get(), ++m_DecoderShaderBindFenceValue))) {
-                    deviceContext4->Wait(m_DecoderShaderBindFence.Get(), m_DecoderShaderBindFenceValue);
-                }
-            }
-        }
    }
    else {
        // Copy this frame into our video texture
-        m_DeviceContext->CopySubresourceRegion1(m_VideoTexture.Get(), 0, 0, 0, 0,
-                                                (ID3D11Resource*)frame->data[0], (int)(intptr_t)frame->data[1],
-                                                nullptr, D3D11_COPY_DISCARD);
+        m_RenderDeviceContext->CopySubresourceRegion1(m_VideoTexture.Get(), 0, 0, 0, 0,
+                                                      m_RenderSharedTextureArray.Get(),
+                                                      (int)(intptr_t)frame->data[1],
+                                                      nullptr, D3D11_COPY_DISCARD);

        // SRV 0 is always mapped to the video texture
        srvIndex = 0;
@@ -895,19 +1036,31 @@ void D3D11VARenderer::renderVideo(AVFrame* frame)

    // Bind SRVs for this frame
    ID3D11ShaderResourceView* frameSrvs[] = { m_VideoTextureResourceViews[srvIndex][0].Get(), m_VideoTextureResourceViews[srvIndex][1].Get() };
-    m_DeviceContext->PSSetShaderResources(0, 2, frameSrvs);
+    m_RenderDeviceContext->PSSetShaderResources(0, 2, frameSrvs);

    // Draw the video
-    m_DeviceContext->DrawIndexed(6, 0, 0);
+    m_RenderDeviceContext->DrawIndexed(6, 0, 0);

    // Unbind SRVs for this frame
    ID3D11ShaderResourceView* nullSrvs[2] = {};
-    m_DeviceContext->PSSetShaderResources(0, 2, nullSrvs);
+    m_RenderDeviceContext->PSSetShaderResources(0, 2, nullSrvs);
+
+    // Insert a fence to force the decode context to wait for the render context to finish reading
+    if (m_DecodeDevice != m_RenderDevice) {
+        SDL_assert(m_DecodeR2DFence);
+        SDL_assert(m_RenderR2DFence);
+
+        if (SUCCEEDED(m_RenderDeviceContext->Signal(m_RenderR2DFence.Get(), m_DecodeRenderSyncFenceValue))) {
+            lockContext(this);
+            m_DecodeDeviceContext->Wait(m_DecodeR2DFence.Get(), m_DecodeRenderSyncFenceValue++);
+            unlockContext(this);
+        }
+    }

    // Trigger our fence to signal after this video frame has been rendered
    if (m_PreviousFrameRenderedFence) {
        ComPtr<ID3D11DeviceContext4> deviceContext4;
-        if (SUCCEEDED(m_DeviceContext.As(&deviceContext4))) {
+        if (SUCCEEDED(m_RenderDeviceContext.As(&deviceContext4))) {
            if (SUCCEEDED(deviceContext4->Signal(m_PreviousFrameRenderedFence.Get(), m_PreviousFrameRenderedFenceValue + 1))) {
                m_PreviousFrameRenderedFenceValue++;
            }
@@ -962,7 +1115,7 @@ void D3D11VARenderer::notifyOverlayUpdated(Overlay::OverlayType type)
    texData.SysMemPitch = newSurface->pitch;

    ComPtr<ID3D11Texture2D> newTexture;
-    hr = m_Device->CreateTexture2D(&texDesc, &texData, &newTexture);
+    hr = m_RenderDevice->CreateTexture2D(&texDesc, &texData, &newTexture);
    if (FAILED(hr)) {
        SDL_FreeSurface(newSurface);
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -972,7 +1125,7 @@ void D3D11VARenderer::notifyOverlayUpdated(Overlay::OverlayType type)
    }

    ComPtr<ID3D11ShaderResourceView> newTextureResourceView;
-    hr = m_Device->CreateShaderResourceView((ID3D11Resource*)newTexture.Get(), nullptr, &newTextureResourceView);
+    hr = m_RenderDevice->CreateShaderResourceView((ID3D11Resource*)newTexture.Get(), nullptr, &newTextureResourceView);
    if (FAILED(hr)) {
        SDL_FreeSurface(newSurface);
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1038,7 +1191,7 @@ bool D3D11VARenderer::createOverlayVertexBuffer(Overlay::OverlayType type, int w
    D3D11_SUBRESOURCE_DATA vbData = {};
    vbData.pSysMem = verts;

-    HRESULT hr = m_Device->CreateBuffer(&vbDesc, &vbData, &newVertexBuffer);
+    HRESULT hr = m_RenderDevice->CreateBuffer(&vbDesc, &vbData, &newVertexBuffer);
    if (FAILED(hr)) {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                     "ID3D11Device::CreateBuffer() failed: %x",
@@ -1107,7 +1260,7 @@ bool D3D11VARenderer::notifyWindowChanged(PWINDOW_STATE_CHANGE_INFO stateInfo)

        // We must release all references to the back buffer
        m_RenderTargetView.Reset();
-        m_DeviceContext->Flush();
+        m_RenderDeviceContext->Flush();

        HRESULT hr = m_SwapChain->ResizeBuffers(0, stateInfo->width, stateInfo->height, DXGI_FORMAT_UNKNOWN, swapchainDesc.Flags);
        if (FAILED(hr)) {
@@ -1175,7 +1328,7 @@ bool D3D11VARenderer::checkDecoderSupport(IDXGIAdapter* adapter)
    }

    // Derive a ID3D11VideoDevice from our ID3D11Device.
-    hr = m_Device.As(&videoDevice);
+    hr = m_RenderDevice.As(&videoDevice);
    if (FAILED(hr)) {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                     "ID3D11Device::QueryInterface(ID3D11VideoDevice) failed: %x",
@@ -1394,16 +1547,16 @@ bool D3D11VARenderer::setupRenderingResources()
 {
    HRESULT hr;

-    m_DeviceContext->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+    m_RenderDeviceContext->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);

    // We use a common vertex shader for all pixel shaders
    {
        QByteArray vertexShaderBytecode = Path::readDataFile("d3d11_vertex.fxc");

        ComPtr<ID3D11VertexShader> vertexShader;
-        hr = m_Device->CreateVertexShader(vertexShaderBytecode.constData(), vertexShaderBytecode.length(), nullptr, &vertexShader);
+        hr = m_RenderDevice->CreateVertexShader(vertexShaderBytecode.constData(), vertexShaderBytecode.length(), nullptr, &vertexShader);
        if (SUCCEEDED(hr)) {
-            m_DeviceContext->VSSetShader(vertexShader.Get(), nullptr, 0);
+            m_RenderDeviceContext->VSSetShader(vertexShader.Get(), nullptr, 0);
        }
        else {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1418,9 +1571,9 @@ bool D3D11VARenderer::setupRenderingResources()
            { "TEXCOORD", 0, DXGI_FORMAT_R32G32_FLOAT, 0, 8, D3D11_INPUT_PER_VERTEX_DATA, 0 },
        };
        ComPtr<ID3D11InputLayout> inputLayout;
-        hr = m_Device->CreateInputLayout(vertexDesc, ARRAYSIZE(vertexDesc), vertexShaderBytecode.constData(), vertexShaderBytecode.length(), &inputLayout);
+        hr = m_RenderDevice->CreateInputLayout(vertexDesc, ARRAYSIZE(vertexDesc), vertexShaderBytecode.constData(), vertexShaderBytecode.length(), &inputLayout);
        if (SUCCEEDED(hr)) {
-            m_DeviceContext->IASetInputLayout(inputLayout.Get());
+            m_RenderDeviceContext->IASetInputLayout(inputLayout.Get());
        }
        else {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1433,7 +1586,7 @@ bool D3D11VARenderer::setupRenderingResources()
    {
        QByteArray overlayPixelShaderBytecode = Path::readDataFile("d3d11_overlay_pixel.fxc");

-        hr = m_Device->CreatePixelShader(overlayPixelShaderBytecode.constData(), overlayPixelShaderBytecode.length(), nullptr, &m_OverlayPixelShader);
+        hr = m_RenderDevice->CreatePixelShader(overlayPixelShaderBytecode.constData(), overlayPixelShaderBytecode.length(), nullptr, &m_OverlayPixelShader);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreatePixelShader() failed: %x",
@@ -1446,7 +1599,7 @@ bool D3D11VARenderer::setupRenderingResources()
    {
        QByteArray videoPixelShaderBytecode = Path::readDataFile(k_VideoShaderNames[i]);

-        hr = m_Device->CreatePixelShader(videoPixelShaderBytecode.constData(), videoPixelShaderBytecode.length(), nullptr, &m_VideoPixelShaders[i]);
+        hr = m_RenderDevice->CreatePixelShader(videoPixelShaderBytecode.constData(), videoPixelShaderBytecode.length(), nullptr, &m_VideoPixelShaders[i]);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreatePixelShader() failed: %x",
@@ -1469,9 +1622,9 @@ bool D3D11VARenderer::setupRenderingResources()
        samplerDesc.MaxLOD = D3D11_FLOAT32_MAX;

        ComPtr<ID3D11SamplerState> sampler;
-        hr = m_Device->CreateSamplerState(&samplerDesc,  &sampler);
+        hr = m_RenderDevice->CreateSamplerState(&samplerDesc,  &sampler);
        if (SUCCEEDED(hr)) {
-            m_DeviceContext->PSSetSamplers(0, 1, sampler.GetAddressOf());
+            m_RenderDeviceContext->PSSetSamplers(0, 1, sampler.GetAddressOf());
        }
        else {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1497,9 +1650,9 @@ bool D3D11VARenderer::setupRenderingResources()
        indexBufferData.SysMemPitch = sizeof(int);

        ComPtr<ID3D11Buffer> indexBuffer;
-        hr = m_Device->CreateBuffer(&indexBufferDesc, &indexBufferData, &indexBuffer);
+        hr = m_RenderDevice->CreateBuffer(&indexBufferDesc, &indexBufferData, &indexBuffer);
        if (SUCCEEDED(hr)) {
-            m_DeviceContext->IASetIndexBuffer(indexBuffer.Get(), DXGI_FORMAT_R32_UINT, 0);
+            m_RenderDeviceContext->IASetIndexBuffer(indexBuffer.Get(), DXGI_FORMAT_R32_UINT, 0);
        }
        else {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1523,7 +1676,7 @@ bool D3D11VARenderer::setupRenderingResources()
        blendDesc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
        blendDesc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL;

-        hr = m_Device->CreateBlendState(&blendDesc, &m_OverlayBlendState);
+        hr = m_RenderDevice->CreateBlendState(&blendDesc, &m_OverlayBlendState);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreateBlendState() failed: %x",
@@ -1540,9 +1693,9 @@ bool D3D11VARenderer::setupRenderingResources()
        blendDesc.RenderTarget[0].BlendEnable = FALSE;
        blendDesc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL;

-        hr = m_Device->CreateBlendState(&blendDesc, &m_VideoBlendState);
+        hr = m_RenderDevice->CreateBlendState(&blendDesc, &m_VideoBlendState);
        if (SUCCEEDED(hr)) {
-            m_DeviceContext->OMSetBlendState(m_VideoBlendState.Get(), nullptr, 0xffffffff);
+            m_RenderDeviceContext->OMSetBlendState(m_VideoBlendState.Get(), nullptr, 0xffffffff);
        }
        else {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
@@ -1566,7 +1719,7 @@ bool D3D11VARenderer::setupSwapchainDependentResources()
    // Create our render target view
    {
        ComPtr<ID3D11Resource> backBufferResource;
-        hr = m_SwapChain->GetBuffer(0, __uuidof(ID3D11Resource), (void**)&backBufferResource);
+        hr = m_SwapChain->GetBuffer(0, IID_PPV_ARGS(&backBufferResource));
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "IDXGISwapChain::GetBuffer() failed: %x",
@@ -1574,7 +1727,7 @@ bool D3D11VARenderer::setupSwapchainDependentResources()
            return false;
        }

-        hr = m_Device->CreateRenderTargetView(backBufferResource.Get(), nullptr, &m_RenderTargetView);
+        hr = m_RenderDevice->CreateRenderTargetView(backBufferResource.Get(), nullptr, &m_RenderTargetView);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreateRenderTargetView() failed: %x",
@@ -1594,7 +1747,7 @@ bool D3D11VARenderer::setupSwapchainDependentResources()
        viewport.MinDepth = 0;
        viewport.MaxDepth = 1;

-        m_DeviceContext->RSSetViewports(1, &viewport);
+        m_RenderDeviceContext->RSSetViewports(1, &viewport);
    }

    return true;
@@ -1605,9 +1758,44 @@ bool D3D11VARenderer::setupFrameRenderingResources(AVHWFramesContext* framesCont
 {
    auto d3d11vaFramesContext = (AVD3D11VAFramesContext*)framesContext->hwctx;

+    // Open the decoder texture array on the renderer device if we're using separate devices
+    if (m_DecodeDevice != m_RenderDevice) {
+        ComPtr<IDXGIResource1> dxgiDecoderResource;
+
+        HRESULT hr = d3d11vaFramesContext->texture_infos->texture->QueryInterface(IID_PPV_ARGS(&dxgiDecoderResource));
+        if (FAILED(hr)) {
+            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                         "ID3D11Texture2D::QueryInterface(IDXGIResource1) failed: %x",
+                         hr);
+            return false;
+        }
+
+        HANDLE sharedHandle;
+        hr = dxgiDecoderResource->CreateSharedHandle(nullptr, DXGI_SHARED_RESOURCE_READ, nullptr, &sharedHandle);
+        if (FAILED(hr)) {
+            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                         "IDXGIResource1::CreateSharedHandle() failed: %x",
+                         hr);
+            return false;
+        }
+
+        hr = m_RenderDevice->OpenSharedResource1(sharedHandle, IID_PPV_ARGS(&m_RenderSharedTextureArray));
+        CloseHandle(sharedHandle);
+        if (FAILED(hr)) {
+            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
+                         "ID3D11Device1::OpenSharedResource1() failed: %x",
+                         hr);
+            return false;
+        }
+    }
+    else {
+        d3d11vaFramesContext->texture_infos->texture->AddRef();
+        m_RenderSharedTextureArray.Attach(d3d11vaFramesContext->texture_infos->texture);
+    }
+
    // Query the format of the underlying texture array
    D3D11_TEXTURE2D_DESC textureDesc;
-    d3d11vaFramesContext->texture_infos->texture->GetDesc(&textureDesc);
+    m_RenderSharedTextureArray->GetDesc(&textureDesc);
    m_TextureFormat = textureDesc.Format;

    if (m_BindDecoderOutputTextures) {
@@ -1660,7 +1848,7 @@ bool D3D11VARenderer::setupVideoTexture(AVHWFramesContext* framesContext)
    texDesc.CPUAccessFlags = 0;
    texDesc.MiscFlags = 0;

-    hr = m_Device->CreateTexture2D(&texDesc, nullptr, &m_VideoTexture);
+    hr = m_RenderDevice->CreateTexture2D(&texDesc, nullptr, &m_VideoTexture);
    if (FAILED(hr)) {
        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                     "ID3D11Device::CreateTexture2D() failed: %x",
@@ -1681,7 +1869,7 @@ bool D3D11VARenderer::setupVideoTexture(AVHWFramesContext* framesContext)
        SDL_assert(srvIndex < m_VideoTextureResourceViews[0].size());

        srvDesc.Format = srvFormat;
-        hr = m_Device->CreateShaderResourceView(m_VideoTexture.Get(), &srvDesc, &m_VideoTextureResourceViews[0][srvIndex]);
+        hr = m_RenderDevice->CreateShaderResourceView(m_VideoTexture.Get(), &srvDesc, &m_VideoTextureResourceViews[0][srvIndex]);
        if (FAILED(hr)) {
            SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                         "ID3D11Device::CreateShaderResourceView() failed: %x",
@@ -1723,9 +1911,9 @@ bool D3D11VARenderer::setupTexturePoolViews(AVHWFramesContext* framesContext)
            SDL_assert(srvIndex < m_VideoTextureResourceViews[i].size());

            srvDesc.Format = srvFormat;
-            hr = m_Device->CreateShaderResourceView(d3d11vaFramesContext->texture_infos[i].texture,
-                                                    &srvDesc,
-                                                    &m_VideoTextureResourceViews[i][srvIndex]);
+            hr = m_RenderDevice->CreateShaderResourceView(m_RenderSharedTextureArray.Get(),
+                                                          &srvDesc,
+                                                          &m_VideoTextureResourceViews[i][srvIndex]);
            if (FAILED(hr)) {
                SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
                             "ID3D11Device::CreateShaderResourceView() failed: %x",