Compute Queues

Basics

A compute queue is represented by the ID3D12CommandQueue interface. Each queue is associated with a specific D3D12_COMMAND_LIST_TYPE (typically D3D12_COMMAND_LIST_TYPE_COMPUTE for compute work).

// Creating a compute command queue
D3D12_COMMAND_QUEUE_DESC desc = {};
desc.Type = D3D12_COMMAND_LIST_TYPE_COMPUTE;
desc.Priority = D3D12_COMMAND_QUEUE_PRIORITY_NORMAL;
desc.Flags = D3D12_COMMAND_QUEUE_FLAG_NONE;
desc.NodeMask = 0;

ID3D12CommandQueue* pComputeQueue = nullptr;
device->CreateCommandQueue(&desc, IID_PPV_ARGS(&pComputeQueue));

Submitting Work

Work is submitted via command lists. Record your compute commands into an ID3D12GraphicsCommandList and then close and execute them on the queue.

// Record a compute dispatch
ID3D12GraphicsCommandList* cmdList = nullptr;
device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_COMPUTE, pAllocator, nullptr, IID_PPV_ARGS(&cmdList));

cmdList->SetPipelineState(pComputePSO);
cmdList->SetComputeRootSignature(pRootSig);
cmdList->Dispatch(threadGroupX, threadGroupY, threadGroupZ);
cmdList->Close();

ID3D12CommandList* ppLists[] = { cmdList };
pComputeQueue->ExecuteCommandLists(_countof(ppLists), ppLists);

Synchronization

Use fences to synchronize CPU and GPU or to coordinate multiple queues.

// Create a fence
ID3D12Fence* pFence = nullptr;
device->CreateFence(0, D3D12_FENCE_FLAG_NONE, IID_PPV_ARGS(&pFence));
UINT64 fenceValue = 1;
HANDLE fenceEvent = CreateEvent(nullptr, FALSE, FALSE, nullptr);

// Signal and wait
pComputeQueue->Signal(pFence, fenceValue);
if(pFence->GetCompletedValue() < fenceValue){
    pFence->SetEventOnCompletion(fenceValue, fenceEvent);
    WaitForSingleObject(fenceEvent, INFINITE);
}
fenceValue++;

Advanced Topics