通过缓冲区读回数据

若要从 GPU 回读数据(例如,捕获屏幕截图),需使用回读堆。 这种技术与通过缓冲区上传纹理数据有关,但有一些不同。

  • 若要读回数据,请创建一个堆, 其中D3D12_HEAP_TYPE 设置为 D3D12_HEAP_TYPE_READBACK,而不是 D3D12_HEAP_TYPE_UPLOAD
  • 回读堆上的资源必须始终是 D3D12_RESOURCE_DIMENSION_BUFFER
  • 使用栅栏来检测 GPU 完成帧处理的时间(当它完成将数据写入输出缓冲区时)。 这很重要,因为 ID3D12Resource::Map 方法不与 GPU 同步(相反,Direct3D 11 的对应方法会与之同步)。 Direct3D 12 映射调用的运行状况类似于使用 NO_OVERWRITE 标志调用 Direct3D 11 对应映射
  • 数据准备就绪后(包括任何必要的资源屏障),调用 ID3D12Resource::Map,使回读数据对 CPU 可见

代码示例

下方的代码示例显示通过缓冲区将数据从 GPU 回读到 CPU 的过程的概要。


// The output buffer (created below) is on a default heap, so only the GPU can access it.

D3D12_HEAP_PROPERTIES defaultHeapProperties{ CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT) };
D3D12_RESOURCE_DESC outputBufferDesc{ CD3DX12_RESOURCE_DESC::Buffer(outputBufferSize, D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS) };
winrt::com_ptr<::ID3D12Resource> outputBuffer;
winrt::check_hresult(d3d12Device->CreateCommittedResource(
    &defaultHeapProperties,
    D3D12_HEAP_FLAG_NONE,
    &outputBufferDesc,
    D3D12_RESOURCE_STATE_COPY_DEST,
    nullptr,
    __uuidof(outputBuffer),
    outputBuffer.put_void()));

// The readback buffer (created below) is on a readback heap, so that the CPU can access it.

D3D12_HEAP_PROPERTIES readbackHeapProperties{ CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_READBACK) };
D3D12_RESOURCE_DESC readbackBufferDesc{ CD3DX12_RESOURCE_DESC::Buffer(outputBufferSize) };
winrt::com_ptr<::ID3D12Resource> readbackBuffer;
winrt::check_hresult(d3d12Device->CreateCommittedResource(
    &readbackHeapProperties,
    D3D12_HEAP_FLAG_NONE,
    &readbackBufferDesc,
    D3D12_RESOURCE_STATE_COPY_DEST,
    nullptr,
    __uuidof(readbackBuffer),
    readbackBuffer.put_void()));

{
    D3D12_RESOURCE_BARRIER outputBufferResourceBarrier
    {
        CD3DX12_RESOURCE_BARRIER::Transition(
            outputBuffer.get(),
            D3D12_RESOURCE_STATE_COPY_DEST,
            D3D12_RESOURCE_STATE_COPY_SOURCE)
    };
    commandList->ResourceBarrier(1, &outputBufferResourceBarrier);
}

commandList->CopyResource(readbackBuffer.get(), outputBuffer.get());

// Code goes here to close, execute (and optionally reset) the command list, and also
// to use a fence to wait for the command queue.

// The code below assumes that the GPU wrote FLOATs to the buffer.

D3D12_RANGE readbackBufferRange{ 0, outputBufferSize };
FLOAT * pReadbackBufferData{};
winrt::check_hresult(
    readbackBuffer->Map
    (
        0,
        &readbackBufferRange,
        reinterpret_cast<void**>(&pReadbackBufferData)
    )
);

// Code goes here to access the data via pReadbackBufferData.

D3D12_RANGE emptyRange{ 0, 0 };
readbackBuffer->Unmap
(
    0,
    &emptyRange
);

有关读取呈现目标纹理并将其作为文件写入磁盘的屏幕截图例程的完整实现,请参阅 适用于 DX12ScreenGrab 的 DirectX 工具包。