Edit

Share via

DXGKARG_CREATEHWQUEUE structure (d3dkmddi.h)

  • Article

The DXGKARG_CREATEHWQUEUE structure is used with DxgkDdiCreateHwQueue to create a hardware queue.

Syntax

typedef struct _DXGKARG_CREATEHWQUEUE {
  HANDLE                    hHwQueue;
  D3DDDI_CREATEHWQUEUEFLAGS Flags;
  UINT                      PrivateDriverDataSize;
  VOID                      *pPrivateDriverData;
  D3DKMT_HANDLE             hHwQueueProgressFence;
  VOID                      *HwQueueProgressFenceCPUVirtualAddress;
  D3DGPU_VIRTUAL_ADDRESS    HwQueueProgressFenceGPUVirtualAddress;
} DXGKARG_CREATEHWQUEUE;

Members

hHwQueue

[in/out] On input, Dxgkrnl's handle to the hardware queue. On output, the driver's handle to the hardware queue.

Flags

[in] A D3DDDI_CREATEHWQUEUEFLAGS structure containing the queue creation flags.

PrivateDriverDataSize

[in] Size in bytes of the private driver data that pPrivateDriverData points to.

pPrivateDriverData

[in/out] Pointer to the private driver data that was passed by the user-mode driver's pfnCreateHwContextCb function. Since pPrivateDriverData is an in-out parameter, KMD can also return private data back to the UMD.

hHwQueueProgressFence

[in] Handle to the GPU synchronization object created by the OS for this hardware queue. The progress fence is used to synchronize DMA buffer completion on this hardware queue. See Remarks.

HwQueueProgressFenceCPUVirtualAddress

[in] Kernel-mode CPU virtual address that can be used to read the currently signaled value of the hardware queue progress fence.

HwQueueProgressFenceGPUVirtualAddress

[in] GPU virtual address that can be used to read or write the value of the hardware queue progress fence. The queue progress fence value must monotonically increase with each new submission to the hardware queue. GPU instructions to update the queue progress fence must be inserted by the user-mode driver at the end of the DMA buffer being tracked.

Remarks

For kernel submission HWQueues, the OS will allocate the progress fence as a native fence type on GPUs that support native fences. When the NativeProgressFence flag is set, KMD's DxgkDdiCreateHwQueue must read the corresponding hHwQueueNativeProgressFence field that specifies the driver handle to the native fence object previously created in DxgkDdiCreateNativeFence. If the NativeProgressFence flag is FALSE then this HWQueue's progress fence is a regular monitored fence object, and KMD must read the legacy hHwQueueProgressFence field like before.

Upon creating a native fence for a HWQueue, the OS will also set D3DKMT_CREATEHWQUEUE::Flags.NativeProgressFence to TRUE, so that the runtime is also aware that the HWQueue's progress fence was allocated as a native fence.

For user submission HWQueues, OS will allocate its progress fence as a native fence in DxgkDdiCreateHwQueueForUserModeSubmission() by definition.

Note on driver compatibility: Dxgkrnl will only write to the hHwQueueNativeProgressFence field on drivers that support the NativeGpuFence cap. This ensures that OS will only reference this field on in-development and future drivers that are compiled with this updated header. KMD will only read this field if DXGKARG_CREATEHWQUEUE.Flags.NativeProgressFence is TRUE. This ensures that KMD only references this field on the latest OS builds that support native progress fence and this updated header.

Requirements

Requirement Value
Minimum supported client Windows 10, version 1703 (WDDM 2.2)
Header d3dkmddi.h

See also

DxgkDdiCreateHwQueue