使用DirectX11像素着色器在GPU上将DXGI_FORMAT_B8G8R8A8_UNORM转换为NV12颜色格式

Question

我正在编写一个代码，使用桌面复制来捕获桌面，并使用Intel硬件MFT将其编码为h264。编码器仅接受NV12格式作为输入。我已经得到了一个DXGI_FORMAT_B8G8R8A8_UNORM到NV12转换器(https://github.com/NVIDIA/video-sdk-samples/blob/master/nvEncDXGIOutputDuplicationSample/Preproc.cpp)，它基于DirectX VideoProcessor，可以正常工作。

问题在于，在某些英特尔图形硬件上，VideoProcessor仅支持从DXGI_FORMAT_B8G8R8A8_UNORM转换为YUY2而不是NV12，我通过GetVideoProcessorOutputFormats确认了这一点。尽管VideoProcessor Blt没有任何错误，我可以看到输出视频中的帧有点像素化，但只有仔细观察才能注意到。

我猜VideoProcessor已经简单地转向下一个支持的输出格式(YUY2)，而我无意中将其提供给了编码器，编码器认为输入已经配置成NV12。由于NV12和YUY2之间几乎没有差异，例如字节顺序和亚采样，因此没有失败或主要帧损坏。此外，在支持NV12转换的硬件上，我没有像素化问题。

因此，我决定使用基于像素着色器的颜色转换，该着色器基于此代码(https://github.com/bavulapati/DXGICaptureDXColorSpaceConversionIntelEncode/blob/master/DXGICaptureDXColorSpaceConversionIntelEncode/DuplicationManager.cpp)。我能够使像素着色器工作，我已经上传了我的代码(https://codeshare.io/5PJjxP)供参考(尽可能简化)。

现在，我剩下两个通道，分别是色度和亮度(ID3D11Texture2D纹理)。我真的很困惑如何有效地将这两个单独的通道打包成一个ID3D11Texture2D纹理，以便可以将其提供给编码器。有没有一种在GPU中有效地将Y和UV通道打包到单个ID3D11Texture2D中的方法？由于CPU成本高昂，并且不能提供最佳帧速率，事实上，我对基于CPU的方法甚至不愿意复制纹理到CPU。我正在考虑在GPU中进行处理，而无需在CPU和GPU之间进行任何来回拷贝的方法。

我已经研究了相当长的一段时间，但没有取得任何进展，希望有所帮助。

/**
* This method is incomplete. It's just a template of what I want to achieve.
*/

HRESULT CreateNV12TextureFromLumaAndChromaSurface(ID3D11Texture2D** pOutputTexture)
{
    HRESULT hr = S_OK;

    try
    {
        //Copying from GPU to CPU. Bad :(
        m_pD3D11DeviceContext->CopyResource(m_CPUAccessibleLuminanceSurf, m_LuminanceSurf);

        D3D11_MAPPED_SUBRESOURCE resource;
        UINT subresource = D3D11CalcSubresource(0, 0, 0);

        HRESULT hr = m_pD3D11DeviceContext->Map(m_CPUAccessibleLuminanceSurf, subresource, D3D11_MAP_READ, 0, &resource);

        BYTE* sptr = reinterpret_cast<BYTE*>(resource.pData);
        BYTE* dptrY = nullptr; // point to the address of Y channel in output surface

        //Store Image Pitch
        int m_ImagePitch = resource.RowPitch;

        int height = GetImageHeight();
        int width = GetImageWidth();

        for (int i = 0; i < height; i++)
        {
            memcpy_s(dptrY, m_ImagePitch, sptr, m_ImagePitch);

            sptr += m_ImagePitch;
            dptrY += m_ImagePitch;
        }

        m_pD3D11DeviceContext->Unmap(m_CPUAccessibleLuminanceSurf, subresource);

        //Copying from GPU to CPU. Bad :(
        m_pD3D11DeviceContext->CopyResource(m_CPUAccessibleChrominanceSurf, m_ChrominanceSurf);
        hr = m_pD3D11DeviceContext->Map(m_CPUAccessibleChrominanceSurf, subresource, D3D11_MAP_READ, 0, &resource);

        sptr = reinterpret_cast<BYTE*>(resource.pData);
        BYTE* dptrUV = nullptr; // point to the address of UV channel in output surface

        m_ImagePitch = resource.RowPitch;
        height /= 2;
        width /= 2;

        for (int i = 0; i < height; i++)
        {
            memcpy_s(dptrUV, m_ImagePitch, sptr, m_ImagePitch);

            sptr += m_ImagePitch;
            dptrUV += m_ImagePitch;
        }

        m_pD3D11DeviceContext->Unmap(m_CPUAccessibleChrominanceSurf, subresource);
    }
    catch(HRESULT){}

    return hr;
}

绘制 NV12：

 //
// Draw frame for NV12 texture
//
HRESULT DrawNV12Frame(ID3D11Texture2D* inputTexture)
{
    HRESULT hr;

    // If window was resized, resize swapchain
    if (!m_bIntialized)
    {
        HRESULT Ret = InitializeNV12Surfaces(inputTexture);
        if (!SUCCEEDED(Ret))
        {
            return Ret;
        }

        m_bIntialized = true;
    }

    m_pD3D11DeviceContext->CopyResource(m_ShaderResourceSurf, inputTexture);

    D3D11_TEXTURE2D_DESC FrameDesc;
    m_ShaderResourceSurf->GetDesc(&FrameDesc);

    D3D11_SHADER_RESOURCE_VIEW_DESC ShaderDesc;
    ShaderDesc.Format = FrameDesc.Format;
    ShaderDesc.ViewDimension = D3D11_SRV_DIMENSION_TEXTURE2D;
    ShaderDesc.Texture2D.MostDetailedMip = FrameDesc.MipLevels - 1;
    ShaderDesc.Texture2D.MipLevels = FrameDesc.MipLevels;

    // Create new shader resource view
    ID3D11ShaderResourceView* ShaderResource = nullptr;
    hr = m_pD3D11Device->CreateShaderResourceView(m_ShaderResourceSurf, &ShaderDesc, &ShaderResource);

    IF_FAILED_THROW(hr);

    m_pD3D11DeviceContext->PSSetShaderResources(0, 1, &ShaderResource);

    // Set resources
    m_pD3D11DeviceContext->OMSetRenderTargets(1, &m_pLumaRT, nullptr);
    m_pD3D11DeviceContext->PSSetShader(m_pPixelShaderLuma, nullptr, 0);
    m_pD3D11DeviceContext->RSSetViewports(1, &m_VPLuminance);

    // Draw textured quad onto render target
    m_pD3D11DeviceContext->Draw(NUMVERTICES, 0);

    m_pD3D11DeviceContext->OMSetRenderTargets(1, &m_pChromaRT, nullptr);
    m_pD3D11DeviceContext->PSSetShader(m_pPixelShaderChroma, nullptr, 0);
    m_pD3D11DeviceContext->RSSetViewports(1, &m_VPChrominance);

    // Draw textured quad onto render target
    m_pD3D11DeviceContext->Draw(NUMVERTICES, 0);

    // Release shader resource
    ShaderResource->Release();
    ShaderResource = nullptr;

    return S_OK;
}

初始化着色器：

void SetViewPort(D3D11_VIEWPORT* VP, UINT Width, UINT Height)
{
    VP->Width = static_cast<FLOAT>(Width);
    VP->Height = static_cast<FLOAT>(Height);
    VP->MinDepth = 0.0f;
    VP->MaxDepth = 1.0f;
    VP->TopLeftX = 0;
    VP->TopLeftY = 0;
}

HRESULT MakeRTV(ID3D11RenderTargetView** pRTV, ID3D11Texture2D* pSurf)
{
    if (*pRTV)
    {
        (*pRTV)->Release();
        *pRTV = nullptr;
    }
    // Create a render target view
    HRESULT hr = m_pD3D11Device->CreateRenderTargetView(pSurf, nullptr, pRTV);

    IF_FAILED_THROW(hr);

    return S_OK;
}

HRESULT InitializeNV12Surfaces(ID3D11Texture2D* inputTexture)
{
    ReleaseSurfaces();

    D3D11_TEXTURE2D_DESC lOutputDuplDesc;
    inputTexture->GetDesc(&lOutputDuplDesc);


    // Create shared texture for all duplication threads to draw into
    D3D11_TEXTURE2D_DESC DeskTexD;
    RtlZeroMemory(&DeskTexD, sizeof(D3D11_TEXTURE2D_DESC));
    DeskTexD.Width = lOutputDuplDesc.Width;
    DeskTexD.Height = lOutputDuplDesc.Height;
    DeskTexD.MipLevels = 1;
    DeskTexD.ArraySize = 1;
    DeskTexD.Format = lOutputDuplDesc.Format;
    DeskTexD.SampleDesc.Count = 1;
    DeskTexD.Usage = D3D11_USAGE_DEFAULT;
    DeskTexD.BindFlags = D3D11_BIND_SHADER_RESOURCE;

    HRESULT hr = m_pD3D11Device->CreateTexture2D(&DeskTexD, nullptr, &m_ShaderResourceSurf);
    IF_FAILED_THROW(hr);

    DeskTexD.Format = DXGI_FORMAT_R8_UNORM;
    DeskTexD.BindFlags = D3D11_BIND_RENDER_TARGET;

    hr = m_pD3D11Device->CreateTexture2D(&DeskTexD, nullptr, &m_LuminanceSurf);
    IF_FAILED_THROW(hr);

    DeskTexD.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
    DeskTexD.Usage = D3D11_USAGE_STAGING;
    DeskTexD.BindFlags = 0;

    hr = m_pD3D11Device->CreateTexture2D(&DeskTexD, NULL, &m_CPUAccessibleLuminanceSurf);
    IF_FAILED_THROW(hr);

    SetViewPort(&m_VPLuminance, DeskTexD.Width, DeskTexD.Height);

    HRESULT Ret = MakeRTV(&m_pLumaRT, m_LuminanceSurf);
    if (!SUCCEEDED(Ret))
        return Ret;

    DeskTexD.Width = lOutputDuplDesc.Width / 2;
    DeskTexD.Height = lOutputDuplDesc.Height / 2;
    DeskTexD.Format = DXGI_FORMAT_R8G8_UNORM;

    DeskTexD.Usage = D3D11_USAGE_DEFAULT;
    DeskTexD.CPUAccessFlags = 0;
    DeskTexD.BindFlags = D3D11_BIND_RENDER_TARGET;

    hr = m_pD3D11Device->CreateTexture2D(&DeskTexD, nullptr, &m_ChrominanceSurf);
    IF_FAILED_THROW(hr);

    DeskTexD.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
    DeskTexD.Usage = D3D11_USAGE_STAGING;
    DeskTexD.BindFlags = 0;

    hr = m_pD3D11Device->CreateTexture2D(&DeskTexD, NULL, &m_CPUAccessibleChrominanceSurf);
    IF_FAILED_THROW(hr);

    SetViewPort(&m_VPChrominance, DeskTexD.Width, DeskTexD.Height);
    return MakeRTV(&m_pChromaRT, m_ChrominanceSurf);
}

HRESULT InitVertexShader(ID3D11VertexShader** ppID3D11VertexShader)
{
    HRESULT hr = S_OK;
    UINT Size = ARRAYSIZE(g_VS);

    try
    {
        IF_FAILED_THROW(m_pD3D11Device->CreateVertexShader(g_VS, Size, NULL, ppID3D11VertexShader));;

        m_pD3D11DeviceContext->VSSetShader(m_pVertexShader, nullptr, 0);

        // Vertices for drawing whole texture
        VERTEX Vertices[NUMVERTICES] =
        {
            { XMFLOAT3(-1.0f, -1.0f, 0), XMFLOAT2(0.0f, 1.0f) },
            { XMFLOAT3(-1.0f, 1.0f, 0), XMFLOAT2(0.0f, 0.0f) },
            { XMFLOAT3(1.0f, -1.0f, 0), XMFLOAT2(1.0f, 1.0f) },
            { XMFLOAT3(1.0f, -1.0f, 0), XMFLOAT2(1.0f, 1.0f) },
            { XMFLOAT3(-1.0f, 1.0f, 0), XMFLOAT2(0.0f, 0.0f) },
            { XMFLOAT3(1.0f, 1.0f, 0), XMFLOAT2(1.0f, 0.0f) },
        };

        UINT Stride = sizeof(VERTEX);
        UINT Offset = 0;

        D3D11_BUFFER_DESC BufferDesc;
        RtlZeroMemory(&BufferDesc, sizeof(BufferDesc));
        BufferDesc.Usage = D3D11_USAGE_DEFAULT;
        BufferDesc.ByteWidth = sizeof(VERTEX) * NUMVERTICES;
        BufferDesc.BindFlags = D3D11_BIND_VERTEX_BUFFER;
        BufferDesc.CPUAccessFlags = 0;
        D3D11_SUBRESOURCE_DATA InitData;
        RtlZeroMemory(&InitData, sizeof(InitData));
        InitData.pSysMem = Vertices;

        // Create vertex buffer
        IF_FAILED_THROW(m_pD3D11Device->CreateBuffer(&BufferDesc, &InitData, &m_VertexBuffer));

        m_pD3D11DeviceContext->IASetVertexBuffers(0, 1, &m_VertexBuffer, &Stride, &Offset);
        m_pD3D11DeviceContext->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);

        D3D11_INPUT_ELEMENT_DESC Layout[] =
        {
            { "POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0 },
            { "TEXCOORD", 0, DXGI_FORMAT_R32G32_FLOAT, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0 }
        };

        UINT NumElements = ARRAYSIZE(Layout);
        hr = m_pD3D11Device->CreateInputLayout(Layout, NumElements, g_VS, Size, &m_pVertexLayout);

        m_pD3D11DeviceContext->IASetInputLayout(m_pVertexLayout);
    }
    catch (HRESULT) {}

    return hr;
}

HRESULT InitPixelShaders()
{
    HRESULT hr = S_OK;
    // Refer https://codeshare.io/5PJjxP for g_PS_Y & g_PS_UV blobs
    try
    {
        UINT Size = ARRAYSIZE(g_PS_Y);
        hr = m_pD3D11Device->CreatePixelShader(g_PS_Y, Size, nullptr, &m_pPixelShaderChroma);

        IF_FAILED_THROW(hr);

        Size = ARRAYSIZE(g_PS_UV);
        hr = m_pD3D11Device->CreatePixelShader(g_PS_UV, Size, nullptr, &m_pPixelShaderLuma);

        IF_FAILED_THROW(hr);
    }
    catch (HRESULT) {}

    return hr;
}

- iamrameshkumar

---

这个需要进行验证，但我认为在只能输出YUY2的硬件上，硬件编码器也会接受YUY2。因此，在这种情况下，您可以检查并直接将VideoProcessor的输出馈送到编码器中。 - VuVirt

@VuVirt，我也是这么想的，但当我尝试使用YUY2作为输入类型枚举硬件编码器时，没有编码器返回。 - iamrameshkumar

你有没有尝试过在双GPU的电脑上运行？ - VuVirt

我确定我没有在一台装有多个显卡的机器上运行它。我仍然在想这种不兼容性是如何发生的。我会在这个帖子中尝试更新更多细节。 - iamrameshkumar

Answer 1

我正在尝试使用DirectX11在GPU中将RGBA转换为NV12。

这是一个不错的挑战，因为我对Directx11不熟悉，所以这是我的第一次实验。

请查看此项目以获取更新：D3D11ShaderNV12

在我的当前实现（可能不是最后的实现）中，这是我做的：

• 步骤1：将DXGI_FORMAT_B8G8R8A8_UNORM用作输入纹理
• 步骤2：制作第一个通道着色器以获取3个纹理（Y:Luma、U:ChromaCb和V:ChromaCr）：请参见YCbCrPS2.hlsl
• 步骤3：Y为DXGI_FORMAT_R8_UNORM格式，并准备好用于最终的NV12纹理
• 步骤4：需要通过第二个通道着色器对UV进行下采样处理：请参见ScreenPS2.hlsl（使用线性过滤）
• 步骤5：第三个通道着色器对Y纹理进行采样
• 步骤6：第四个通道着色器使用移位纹理对UV纹理进行采样（我认为还可以使用其他技术）

我的最终纹理不是DXGI_FORMAT_NV12，而是类似的DXGI_FORMAT_R8_UNORM纹理。我的计算机是Windows7，因此不支持DXGI_FORMAT_NV12。我稍后将在另一台计算机上尝试。

带有图片的流程：

- mofo77

---

太好了。这正是我在寻找的。谢谢。 - iamrameshkumar

你可以尝试使用ID3D11DeviceContext :: GenerateMips调用替换第二个渲染通道。它在GPU驱动程序内部实现，可能比代码中的额外渲染通道更快。 - Soonts

我不知道它是否更快，但我添加了一个变量来使用GenerateMips，而不是着色器。这是一种有趣的技术。感谢您的建议。 - mofo77

@mofo77，感谢您的提示。我已经检查了编写的代码，它很好地展示了如何从RGB纹理中分离Y/U/V平面并对U/V平面进行下采样，但最终只将Y/U/V平面合并为DXGI_FORMAT_R8_UNORM平面，而不是真正的DXGI_FORMAT_NV12。我通过fork您的源代码库进行了一些更改以实现它 https://github.com/wangf1978/DirectXVideoScreen - ravin.wang