使用DirectX11像素着色器将颜色从DXGI_FORMAT_B8G8R8A8_UNORM转换为GPU中的NV12


9

我正在编写一个代码,以使用桌面复制捕获桌面,并使用英特尔hardwareMFT将其编码为h264。编码器仅接受NV12格式作为输入。我有一个DXGI_FORMAT_B8G8R8A8_UNORM到NV12转换器(https://github.com/NVIDIA/video-sdk-samples/blob/master/nvEncDXGIOutputDuplicationSample/Preproc.cpp)可以正常工作,并且基于DirectX VideoProcessor。

问题是某些英特尔图形硬件上的VideoProcessor仅支持从DXGI_FORMAT_B8G8R8A8_UNORM到YUY2的转换,但不支持NV12的转换,我已经通过GetVideoProcessorOutputFormats枚举支持的格式来确认了相同的转换。尽管VideoProcessor Blt成功完成,没有任何错误,而且我可以看到输出视频中的帧有些像素化,但如果仔细观察,我会注意到它。

我猜想,VideoProcessor只是故障转移到了下一个受支持的输出格式(YUY2),而我在不知不觉中将其馈送到认为输入已按照配置在NV12中的编码器。NV12和YUY2之间几乎没有像字节顺序和子采样之类的差异,因此不会出现帧故障或严重损坏的情况。另外,在支持NV12转换的硬件上我也没有像素化问题。

所以我决定使用基于此代码的像素着色器进行颜色转换(https://github.com/bavulapati/DXGICaptureDXColorSpaceConversionIntelEncode/blob/master/DXGICaptureDXColorSpaceConversionIntelEncode/DuplicationManager.cpp)。我能够使像素着色器正常工作,我也已经在这里上传了我的代码(https://codeshare.io/5PJjxP)供参考(尽可能简化了它)。

现在,我剩下两个通道,分别是色度和亮度(ID3D11Texture2D纹理)。对于将两个独立的通道有效地打包到一个ID3D11Texture2D纹理中,以便将其馈送到编码器中,我确实感到困惑。有没有一种方法可以将Y和UV通道有效地打包到GPU中的单个ID3D11Texture2D中?我非常厌倦基于CPU的方法,因为它价格昂贵,并且无法提供最佳的帧速率。实际上,我什至都不愿将纹理复制到CPU。我正在考虑在GPU和GPU之间不做任何来回复制的方式。

我已经对此进行了相当长时间的研究,没有任何进展,我们将不胜感激。

/**
* 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;
}

抽奖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;
}

应该对此进行检查,但是我认为在VideoProcessor只能输出到YUY2的硬件上,硬件编码器也将接受YUY2。因此,在这种情况下,您可以检查它并将VideoProcessor输出直接输入到编码器。
VuVirt

@VuVirt,我也这么认为,但是当我尝试枚举YUY2作为输入类型的硬件编码器时,没有返回编码器。
拉姆

您是否可能在双GPU PC上尝试过?
VuVirt

我敢肯定,我不会在具有多个图形卡的计算机上运行它。我仍然想知道这种不兼容会如何发生。我将尝试在此线程中更新更多详细信息。
拉姆

Answers:


5

我正在尝试使用DirectX11将RGBA转换为仅在GPU中的NV12。

这是一个很好的挑战。我对Directx11不熟悉,所以这是我的第一个实验。

检查此项目以获取更新:D3D11ShaderNV12

在我当前的实现中(可能不是最后一个),这是我要做的:

  • 步骤1:使用DXGI_FORMAT_B8G8R8A8_UNORM作为输入纹理
  • 第2步:制作第1遍着色器以获取3种纹理(Y:Luma,U:ChromaCb和V:ChromaCr):请参见YCbCrPS2.hlsl
  • 步骤3:Y为DXGI_FORMAT_R8_UNORM,并准备好进行最终的NV12纹理
  • 步骤4:需要在第二遍着色器中对UV进行下采样:请参阅ScreenPS2.hlsl(使用线性过滤)
  • 步骤5:第三遍着色器以采样Y纹理
  • 第6步:第四遍着色器使用移位纹理对UV纹理进行采样(我认为可以使用其他技术)

着色器NV12

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,因此未处理DXGI_FORMAT_NV12。稍后我将在另一台计算机上尝试。

带图片的过程:

渲染目标


大。这正是我想要的。谢谢。
拉姆

您可以尝试用ID3D11DeviceContext :: GenerateMips调用替换第二个渲染过程。它是在GPU驱动程序内部深入实现的,可能比代码中额外的渲染过程更快。
即将

我不知道它是否更快,但是我添加了一个使用GenerateMips的变体,而不是着色器。这是一种有趣的技术。感谢您的提示。
mofo77,19年
By using our site, you acknowledge that you have read and understand our Cookie Policy and Privacy Policy.
Licensed under cc by-sa 3.0 with attribution required.