video-streaming - 输入少量输入样本后，英特尔图形硬件 H264 MFT ProcessInput 调用失败，同样适用于 Nvidia 硬件 MFT

Question

我正在使用 DesktopDuplication API 捕获桌面并将样本从 RGBA 转换为 GPU 中的 NV12，并将其提供给 MediaFoundation 硬件 H264 MFT。这适用于 Nvidia 图形，也适用于软件编码器，但当只有英特尔图形硬件 MFT 可用时失败。如果我回退到软件 MFT，该代码在同一台英特尔图形机上运行良好。我还确保编码实际上是在 Nvidia 图形机器上的硬件中完成的。

在 Intel 显卡上，MFT 返回 MEError ( "Unspecified error" )，这仅在输入第一个样本之后发生，随后对 ProcessInput 的调用(当事件生成器触发 METransformNeedInput 时)返回 “被调用者当前不接受进一步的输入” . MFT 在返回这些错误之前很少消耗更多样本。这种行为令人困惑，我仅在事件生成器通过 IMFAsyncCallback 异步触发 METransformNeedInput 时才提供样本，并且还正确检查是否在提供样本后立即触发 METransformHaveOutput。当相同的异步逻辑与 Nvidia 硬件 MFT 和 Microsoft 软件编码器一起正常工作时，这真的让我感到困惑。

还有一个类似的unresolved question在英特尔论坛本身。我的代码类似于 intel 线程中提到的代码，除了我还将 d3d 设备管理器设置为编码器，如下所示。

而且，还有其他三个堆栈溢出线程报告了类似的问题，但没有给出解决方案 (MFTransform encoder->ProcessInput returns E_FAIL &
How to create IMFSample from D11 texture for Intel MFT encoder & Asynchronous MFT is not sending MFTransformHaveOutput Event(Intel Hardware MJPEG Decoder MFT) )。我已经尝试了所有可能的选项，但没有任何改进。

Color converter code取自英特尔媒体 sdk 样本。我还上传了我的完整代码 here .

设置d3d管理器的方法:

void SetD3dManager() {

    HRESULT hr = S_OK;

    if (!deviceManager) {

        // Create device manager
        hr = MFCreateDXGIDeviceManager(&resetToken, &deviceManager);
    }

    if (SUCCEEDED(hr)) 
    {
        if (!pD3dDevice) {

            pD3dDevice = GetDeviceDirect3D(0);
        }
    }

    if (pD3dDevice) {

        // NOTE: Getting ready for multi-threaded operation
        const CComQIPtr<ID3D10Multithread> pMultithread = pD3dDevice;
        pMultithread->SetMultithreadProtected(TRUE);

        hr = deviceManager->ResetDevice(pD3dDevice, resetToken);
        CHECK_HR(_pTransform->ProcessMessage(MFT_MESSAGE_SET_D3D_MANAGER, reinterpret_cast<ULONG_PTR>(deviceManager.p)), "Failed to set device manager.");
    }
    else {
        cout << "Failed to get d3d device";
    }
}

Getd3d设备:

CComPtr<ID3D11Device> GetDeviceDirect3D(UINT idxVideoAdapter)
{
    // Create DXGI factory:
    CComPtr<IDXGIFactory1> dxgiFactory;
    DXGI_ADAPTER_DESC1 dxgiAdapterDesc;

    // Direct3D feature level codes and names:

    struct KeyValPair { int code; const char* name; };

    const KeyValPair d3dFLevelNames[] =
    {
        KeyValPair{ D3D_FEATURE_LEVEL_9_1, "Direct3D 9.1" },
        KeyValPair{ D3D_FEATURE_LEVEL_9_2, "Direct3D 9.2" },
        KeyValPair{ D3D_FEATURE_LEVEL_9_3, "Direct3D 9.3" },
        KeyValPair{ D3D_FEATURE_LEVEL_10_0, "Direct3D 10.0" },
        KeyValPair{ D3D_FEATURE_LEVEL_10_1, "Direct3D 10.1" },
        KeyValPair{ D3D_FEATURE_LEVEL_11_0, "Direct3D 11.0" },
        KeyValPair{ D3D_FEATURE_LEVEL_11_1, "Direct3D 11.1" },
    };

    // Feature levels for Direct3D support
    const D3D_FEATURE_LEVEL d3dFeatureLevels[] =
    {
        D3D_FEATURE_LEVEL_11_1,
        D3D_FEATURE_LEVEL_11_0,
        D3D_FEATURE_LEVEL_10_1,
        D3D_FEATURE_LEVEL_10_0,
        D3D_FEATURE_LEVEL_9_3,
        D3D_FEATURE_LEVEL_9_2,
        D3D_FEATURE_LEVEL_9_1,
    };

    constexpr auto nFeatLevels = static_cast<UINT> ((sizeof d3dFeatureLevels) / sizeof(D3D_FEATURE_LEVEL));

    CComPtr<IDXGIAdapter1> dxgiAdapter;
    D3D_FEATURE_LEVEL featLevelCodeSuccess;
    CComPtr<ID3D11Device> d3dDx11Device;

    std::wstring_convert<std::codecvt_utf8<wchar_t>> transcoder;

    HRESULT hr = CreateDXGIFactory1(IID_PPV_ARGS(&dxgiFactory));
    CHECK_HR(hr, "Failed to create DXGI factory");

    // Get a video adapter:
    dxgiFactory->EnumAdapters1(idxVideoAdapter, &dxgiAdapter);

    // Get video adapter description:
    dxgiAdapter->GetDesc1(&dxgiAdapterDesc);

    CHECK_HR(hr, "Failed to retrieve DXGI video adapter description");

    std::cout << "Selected DXGI video adapter is \'"
        << transcoder.to_bytes(dxgiAdapterDesc.Description) << '\'' << std::endl;

    // Create Direct3D device:
    hr = D3D11CreateDevice(
        dxgiAdapter,
        D3D_DRIVER_TYPE_UNKNOWN,
        nullptr,
        (0 * D3D11_CREATE_DEVICE_SINGLETHREADED) | D3D11_CREATE_DEVICE_VIDEO_SUPPORT,
        d3dFeatureLevels,
        nFeatLevels,
        D3D11_SDK_VERSION,
        &d3dDx11Device,
        &featLevelCodeSuccess,
        nullptr
    );

    // Might have failed for lack of Direct3D 11.1 runtime:
    if (hr == E_INVALIDARG)
    {
        // Try again without Direct3D 11.1:
        hr = D3D11CreateDevice(
            dxgiAdapter,
            D3D_DRIVER_TYPE_UNKNOWN,
            nullptr,
            (0 * D3D11_CREATE_DEVICE_SINGLETHREADED) | D3D11_CREATE_DEVICE_VIDEO_SUPPORT,
            d3dFeatureLevels + 1,
            nFeatLevels - 1,
            D3D11_SDK_VERSION,
            &d3dDx11Device,
            &featLevelCodeSuccess,
            nullptr
        );
    }

    // Get name of Direct3D feature level that succeeded upon device creation:
    std::cout << "Hardware device supports " << std::find_if(
        d3dFLevelNames,
        d3dFLevelNames + nFeatLevels,
        [featLevelCodeSuccess](const KeyValPair& entry)
        {
            return entry.code == featLevelCodeSuccess;
        }
    )->name << std::endl;

done:

    return d3dDx11Device;
}

异步回调实现:

struct EncoderCallbacks : IMFAsyncCallback
{
    EncoderCallbacks(IMFTransform* encoder)
    {
        TickEvent = CreateEvent(0, FALSE, FALSE, 0);
        _pEncoder = encoder;
    }

    ~EncoderCallbacks()
    {
        eventGen = nullptr;
        CloseHandle(TickEvent);
    }

    bool Initialize() {

        _pEncoder->QueryInterface(IID_PPV_ARGS(&eventGen));

        if (eventGen) {

            eventGen->BeginGetEvent(this, 0);
            return true;
        }

        return false;
    }

    // dummy IUnknown impl
    virtual HRESULT STDMETHODCALLTYPE QueryInterface(REFIID riid, void** ppvObject) override { return E_NOTIMPL; }
    virtual ULONG STDMETHODCALLTYPE AddRef(void) override { return 1; }
    virtual ULONG STDMETHODCALLTYPE Release(void) override { return 1; }

    virtual HRESULT STDMETHODCALLTYPE GetParameters(DWORD* pdwFlags, DWORD* pdwQueue) override
    {
        // we return immediately and don't do anything except signaling another thread
        *pdwFlags = MFASYNC_SIGNAL_CALLBACK;
        *pdwQueue = MFASYNC_CALLBACK_QUEUE_IO;
        return S_OK;
    }

    virtual HRESULT STDMETHODCALLTYPE Invoke(IMFAsyncResult* pAsyncResult) override
    {
        IMFMediaEvent* event = 0;
        eventGen->EndGetEvent(pAsyncResult, &event);
        if (event)
        {
            MediaEventType type;
            event->GetType(&type);
            switch (type)
            {
            case METransformNeedInput: InterlockedIncrement(&NeedsInput); break;
            case METransformHaveOutput: InterlockedIncrement(&HasOutput); break;
            }
            event->Release();
            SetEvent(TickEvent);
        }

        eventGen->BeginGetEvent(this, 0);
        return S_OK;
    }

    CComQIPtr<IMFMediaEventGenerator> eventGen = nullptr;
    HANDLE TickEvent;
    IMFTransform* _pEncoder = nullptr;

    unsigned int NeedsInput = 0;
    unsigned int HasOutput = 0;
};

生成样本方法:

bool GenerateSampleAsync() {

    DWORD processOutputStatus = 0;
    HRESULT mftProcessOutput = S_OK;
    bool frameSent = false;

    // Create sample
    CComPtr<IMFSample> currentVideoSample = nullptr;

    MFT_OUTPUT_STREAM_INFO StreamInfo;

    // wait for any callback to come in
    WaitForSingleObject(_pEventCallback->TickEvent, INFINITE);

    while (_pEventCallback->NeedsInput) {

        if (!currentVideoSample) {

            (pDesktopDuplication)->releaseBuffer();
            (pDesktopDuplication)->cleanUpCurrentFrameObjects();

            bool bTimeout = false;

            if (pDesktopDuplication->GetCurrentFrameAsVideoSample((void**)& currentVideoSample, waitTime, bTimeout, deviceRect, deviceRect.Width(), deviceRect.Height())) {

                prevVideoSample = currentVideoSample;
            }
            // Feed the previous sample to the encoder in case of no update in display
            else {
                currentVideoSample = prevVideoSample;
            }
        }

        if (currentVideoSample)
        {
            InterlockedDecrement(&_pEventCallback->NeedsInput);
            _frameCount++;

            CHECK_HR(currentVideoSample->SetSampleTime(mTimeStamp), "Error setting the video sample time.");
            CHECK_HR(currentVideoSample->SetSampleDuration(VIDEO_FRAME_DURATION), "Error getting video sample duration.");

            CHECK_HR(_pTransform->ProcessInput(inputStreamID, currentVideoSample, 0), "The resampler H264 ProcessInput call failed.");

            mTimeStamp += VIDEO_FRAME_DURATION;
        }
    }

    while (_pEventCallback->HasOutput) {

        CComPtr<IMFSample> mftOutSample = nullptr;
        CComPtr<IMFMediaBuffer> pOutMediaBuffer = nullptr;

        InterlockedDecrement(&_pEventCallback->HasOutput);

        CHECK_HR(_pTransform->GetOutputStreamInfo(outputStreamID, &StreamInfo), "Failed to get output stream info from H264 MFT.");

        CHECK_HR(MFCreateSample(&mftOutSample), "Failed to create MF sample.");
        CHECK_HR(MFCreateMemoryBuffer(StreamInfo.cbSize, &pOutMediaBuffer), "Failed to create memory buffer.");
        CHECK_HR(mftOutSample->AddBuffer(pOutMediaBuffer), "Failed to add sample to buffer.");

        MFT_OUTPUT_DATA_BUFFER _outputDataBuffer;
        memset(&_outputDataBuffer, 0, sizeof _outputDataBuffer);
        _outputDataBuffer.dwStreamID = outputStreamID;
        _outputDataBuffer.dwStatus = 0;
        _outputDataBuffer.pEvents = nullptr;
        _outputDataBuffer.pSample = mftOutSample;

        mftProcessOutput = _pTransform->ProcessOutput(0, 1, &_outputDataBuffer, &processOutputStatus);

        if (mftProcessOutput != MF_E_TRANSFORM_NEED_MORE_INPUT)
        {
            if (_outputDataBuffer.pSample) {

                CComPtr<IMFMediaBuffer> buf = NULL;
                DWORD bufLength;
                CHECK_HR(_outputDataBuffer.pSample->ConvertToContiguousBuffer(&buf), "ConvertToContiguousBuffer failed.");

                if (buf) {

                    CHECK_HR(buf->GetCurrentLength(&bufLength), "Get buffer length failed.");
                    BYTE* rawBuffer = NULL;

                    fFrameSize = bufLength;
                    fDurationInMicroseconds = 0;
                    gettimeofday(&fPresentationTime, NULL);

                    buf->Lock(&rawBuffer, NULL, NULL);
                    memmove(fTo, rawBuffer, fFrameSize > fMaxSize ? fMaxSize : fFrameSize);

                    bytesTransfered += bufLength;

                    FramedSource::afterGetting(this);

                    buf->Unlock();

                    frameSent = true;
                }
            }

            if (_outputDataBuffer.pEvents)
                _outputDataBuffer.pEvents->Release();
        }
        else if (MF_E_TRANSFORM_STREAM_CHANGE == mftProcessOutput) {

            // some encoders want to renegotiate the output format. 
            if (_outputDataBuffer.dwStatus & MFT_OUTPUT_DATA_BUFFER_FORMAT_CHANGE)
            {
                CComPtr<IMFMediaType> pNewOutputMediaType = nullptr;
                HRESULT res = _pTransform->GetOutputAvailableType(outputStreamID, 1, &pNewOutputMediaType);

                res = _pTransform->SetOutputType(0, pNewOutputMediaType, 0);//setting the type again
                CHECK_HR(res, "Failed to set output type during stream change");
            }
        }
        else {
            HandleFailure();
        }
    }

    return frameSent;
}

创建视频样本和颜色转换:

bool GetCurrentFrameAsVideoSample(void **videoSample, int waitTime, bool &isTimeout, CRect &deviceRect, int surfaceWidth, int surfaceHeight)
{

FRAME_DATA currentFrameData;

m_LastErrorCode = m_DuplicationManager.GetFrame(&currentFrameData, waitTime, &isTimeout);

if (!isTimeout && SUCCEEDED(m_LastErrorCode)) {

    m_CurrentFrameTexture = currentFrameData.Frame;

    if (!pDstTexture) {

        D3D11_TEXTURE2D_DESC desc;
        ZeroMemory(&desc, sizeof(D3D11_TEXTURE2D_DESC));

        desc.Format = DXGI_FORMAT_NV12;
        desc.Width = surfaceWidth;
        desc.Height = surfaceHeight;
        desc.MipLevels = 1;
        desc.ArraySize = 1;
        desc.SampleDesc.Count = 1;
        desc.CPUAccessFlags = 0;
        desc.Usage = D3D11_USAGE_DEFAULT;
        desc.BindFlags = D3D11_BIND_RENDER_TARGET;

        m_LastErrorCode = m_Id3d11Device->CreateTexture2D(&desc, NULL, &pDstTexture);
    }

    if (m_CurrentFrameTexture && pDstTexture) {

        // Copy diff area texels to new temp texture
        //m_Id3d11DeviceContext->CopySubresourceRegion(pNewTexture, D3D11CalcSubresource(0, 0, 1), 0, 0, 0, m_CurrentFrameTexture, 0, NULL);

        HRESULT hr = pColorConv->Convert(m_CurrentFrameTexture, pDstTexture);

        if (SUCCEEDED(hr)) { 

            CComPtr<IMFMediaBuffer> pMediaBuffer = nullptr;

            MFCreateDXGISurfaceBuffer(__uuidof(ID3D11Texture2D), pDstTexture, 0, FALSE, (IMFMediaBuffer**)&pMediaBuffer);

            if (pMediaBuffer) {

                CComPtr<IMF2DBuffer> p2DBuffer = NULL;
                DWORD length = 0;
                (((IMFMediaBuffer*)pMediaBuffer))->QueryInterface(__uuidof(IMF2DBuffer), reinterpret_cast<void**>(&p2DBuffer));
                p2DBuffer->GetContiguousLength(&length);
                (((IMFMediaBuffer*)pMediaBuffer))->SetCurrentLength(length);

                //MFCreateVideoSampleFromSurface(NULL, (IMFSample**)videoSample);
                MFCreateSample((IMFSample * *)videoSample);

                if (videoSample) {

                    (*((IMFSample **)videoSample))->AddBuffer((((IMFMediaBuffer*)pMediaBuffer)));
                }

                return true;
            }
        }
    }
}

return false;
}

机器中的英特尔显卡驱动程序已经是最新的。





只有 TransformNeedInput 事件一直被触发，但编码器提示它不能接受更多的输入。 TransformHaveOutput 事件从未被触发。



intel 和 msdn 论坛上报告了类似的问题:
1) https://software.intel.com/en-us/forums/intel-media-sdk/topic/607189
2) https://social.msdn.microsoft.com/Forums/SECURITY/en-US/fe051dd5-b522-4e4b-9cbb-2c06a5450e40/imfsinkwriter-merit-validation-failed-for-mft-intel-quick-sync-video-h264-encoder-mft?forum=mediafoundationdevelopment

更新:
我试图只模拟输入源(通过以编程方式创建动画矩形 NV12 样本)，而其他所有内容都保持不变。这一次，英特尔编码器没有任何提示，我什至得到了输出样本。除了英特尔编码器的输出视频失真，而 Nvidia 编码器工作得非常好。

此外，我仍然收到带有英特尔编码器的原始 NV12 源的 ProcessInput 错误。我对 Nvidia MFT 和软件编码器没有任何问题。

Intel硬件MFT输出:(请看Nvidia编码器输出)


Nvidia 硬件 MFT 输出:


Nvidia图形使用统计:


英特尔图形使用统计(我不明白为什么 GPU 引擎显示为视频解码):

Answer 1

我看了你的代码。

根据您的帖子，我怀疑是英特尔视频处理器问题。

我的操作系统是 Win7，所以我决定在我的 Nvidia 卡上使用 D3D9Device 测试视频处理器的行为，然后在 Intel HD Graphics 4000 上测试。

我想视频处理器功能对于 D3D9Device 的行为方式与对于 D3D11Device 的行为方式相同。当然，这将是必要的检查。

所以我做了这个程序来检查:https://github.com/mofo7777/DirectXVideoScreen (参见 D3D9VideoProcessor 子项目)

看来您没有检查有关视频处理器功能的足够内容。

使用 IDXVAHD_Device::GetVideoProcessorDeviceCaps，这是我检查的内容:

DXVAHD_VPDEVCAPS.MaxInputStreams > 0

DXVAHD_VPDEVCAPS.VideoProcessorCount > 0

DXVAHD_VPDEVCAPS.OutputFormatCount > 0

DXVAHD_VPDEVCAPS.InputFormatCount > 0

DXVAHD_VPDEVCAPS.InputPool == D3DPOOL_DEFAULT

我还检查了 IDXVAHD_Device::GetVideoProcessorOutputFormats 和 IDXVAHD_Device::GetVideoProcessorInputFormats 支持的输入和输出格式。

这就是我发现 Nvidia GPU 和 Intel GPU 之间存在差异的地方。

NVIDIA : 4 输出格式

D3DFMT_A8R8G8B8

D3DFMT_X8R8G8B8

D3DFMT_YUY2

D3DFMT_NV12

INTEL : 3 输出格式

D3DFMT_A8R8G8B8

D3DFMT_X8R8G8B8

D3DFMT_YUY2

在 Intel HD Graphics 4000 上，不支持 NV12 输出格式。

此外，为了让程序正常工作，我需要在使用 VideoProcessBltHD 之前设置流状态:

DXVAHD_STREAM_STATE_D3DFORMAT

DXVAHD_STREAM_STATE_FRAME_FORMAT

DXVAHD_STREAM_STATE_INPUT_COLOR_SPACE

DXVAHD_STREAM_STATE_SOURCE_RECT

DXVAHD_STREAM_STATE_DESTINATION_RECT

对于 D3D11:

ID3D11VideoProcessorEnumerator::GetVideoProcessorCaps == IDXVAHD_Device::GetVideoProcessorDeviceCaps

(D3D11_VIDEO_PROCESSOR_FORMAT_SUPPORT_OUTPUT)
ID3D11VideoProcessorEnumerator::CheckVideoProcessorFormat == IDXVAHD_Device::GetVideoProcessorOutputFormats

(D3D11_VIDEO_PROCESSOR_FORMAT_SUPPORT_INPUT)
ID3D11VideoProcessorEnumerator::CheckVideoProcessorFormat == IDXVAHD_Device::GetVideoProcessorInputFormats

ID3D11VideoContext::(...) == IDXVAHD_VideoProcessor::SetVideoProcessStreamState

您能否首先验证您的 GPU 的视频处理器功能。你看到和我看到的一样的区别吗？

这是我们需要知道的第一件事，从我在您的 github 项目中看到的内容来看，您的程序似乎没有检查这一点。