Resource lifetime and synchronization
Just as with Direct3D 12, your DirectML application must (in order to avoid undefined behavior) correctly manage object lifetimes and synchronization between the CPU and the GPU. DirectML follows an identical resource lifetime model to that of Direct3D 12.
- Lifetime dependencies between two CPU objects are maintained by DirectML using strong reference counts. Your application needn't manually manage CPU lifetime dependencies. For example, every device child holds a strong reference to its parent device.
- Lifetime dependencies between GPU objects—or dependencies that span across CPU and GPU—aren't automatically managed. It's your application's responsibility to ensure that GPU resources live at least until all work using that resource has completed execution on the GPU.
DirectML devices
The DirectML device is a thread-safe stateless factory object. Every device child (see IDMLDeviceChild) holds a strong reference to its parent DirectML device (see IDMLDevice). That means that you can always retrieve the parent device interface from any device child interface.
A DirectML device in turn holds a strong reference to the Direct3D 12 device that was used to create it (see ID3D12Device, and derived interfaces).
Because the DirectML device is stateless, it's implicitly thread-safe. You may call methods on the DirectML device from multiple threads simultaneously without the need for external synchronization.
However, unlike the Direct3D 12 device, the DirectML device is not a singleton object. You're free to create as many DirectML devices as you wish. However, you may not mix and match device children that belong to different devices. For example, IDMLBindingTable and IDMLCompiledOperator are two kinds of device children (both interfaces derive directly or indirectly from IDMLDeviceChild). And you may not use a binding table (IDMLBindingTable) to bind for an operator (IDMLCompiledOperator) if the operator and the binding table belong to different DirectML device instances.
Because the DirectML device is not a singleton, device removal occurs on a per-device basis, rather than being a process-wide event as it is for a Direct3D 12 device. For more information, see Handling errors and device removal in DirectML.
Lifetime requirements of GPU resources
Like Direct3D 12, DirectML doesn't automatically synchronize between the CPU and GPU; nor does it automatically keep resources alive while they're in use by the GPU. Instead, these are responsibilities of your application.
When executing a command list that contains DirectML dispatches, your application must ensure that GPU resources are kept alive until all work using those resources has completed execution on the GPU.
In the case of IDMLCommandRecorder::RecordDispatch for a DirectML operator, that includes the following objects.
- The IDMLCompiledOperator being executed (or IDMLOperatorInitializer instead, if performing operator initialization).
- The IDMLCompiledOperator backing the binding table being used to bind the operator.
- The ID3D12Resource objects bound as the inputs/outputs of the operator.
- The ID3D12Resource objects bound as persistent and temporary resources, if applicable.
- The ID3D12CommandAllocator backing the command list itself.
Not all DirectML interfaces represent GPU resources. For example, a binding table does not need to be kept alive until all dispatches using it have completed execution on the GPU. That's because the binding table itself doesn't own any GPU resources. Rather, the descriptor heap does. Therefore, the underlying descriptor heap is the object that must be kept alive until execution completes, and not the binding table itself.
A similar concept exists in Direct3D 12. A command allocator must be kept alive until all executions using it have completed on the GPU; since it owns GPU memory. But, a command list doesn't own GPU memory, so it can therefore be reset or released as soon as it's been submitted for execution.
In DirectML, compiled operators (IDMLCompiledOperator) and operator initializers (IDMLOperatorInitializer) both own GPU resources directly, so they must be kept alive until all dispatches using them have completed execution on the GPU. In addition, any Direct3D 12 resource being used (command allocators, descriptor heaps, buffers, as examples) must similarly be kept alive by your application.
If you prematurely release an object while it's still in use by the GPU, the result is undefined behavior, which has the potential to cause device removal or other errors.
CPU and GPU synchronization
DirectML doesn't itself submit any work for execution on the GPU. Instead, the IDMLCommandRecorder::RecordDispatch method records the dispatch of that work into a command list for later execution. Your application must then close and submit its command list for execution by calling ID3D12CommandQueue::ExecuteCommandLists, as with any Direct3D 12 command list.
Because DirectML doesn't itself submit any work for execution on the GPU, it also doesn't create any fences, nor perform any form of CPU/GPU synchronization on your behalf. It is the responsibility of your application to use the appropriate Direct3D 12 primitives to wait for submitted work to complete execution on the GPU, if necessary. For more info, see ID3D12Fence and ID3D12CommandQueue::Signal.