Start DX12 backend #80
Conversation
Very early so far, but cool to have a branch for it.
Chipping away at the dx12 backend. This should more or less do the signalling to the CPU that the command buffer is done (ie wire up the fence). It also creates buffer objects.
This brings the signature current so it compiles, but the implementations are just stubs for now.
Create compute pipelines from shader source and descriptor sets. This gets it to the point where it can run the collatz example. Still WIP and with rough edges, of course.
|
Is there a reason you're using D3D12 and not D3D11? Since last we spoke, I figured out why the shaders wouldn't compile for the D3D11 cs_5_0 profile (I had introduced non-uniform barriers) and now they compile fine. Note that I haven't actually run the shaders in D3D11 yet. |
|
It's a good question, ultimately I want to support both. I go back and forth about whether I want to target "advanced features" including subgroups and descriptor indexing, and those are DX12 only. I think the current plan is to get SM5 working first, then explore those advanced features later (with runtime query). I had this code lying around for a while, based on Brian's piet-dx12 prototype, so it was handy. |
These function, but can use some work. First, the buffer situation is worse than it should be. It should be possible to create a single readback buffer rather then copy from gpu-local to host-coherent. Second, the command buffer `finish_timestamps` call doesn't correlate to anything in Vulkan, so needs plumbing up through the hub in one form or other when that happens. I'm inclined to make it ergonomic by doing a bit of resource tracking that will trigger the appropriate call (and subsequent host barrier) in the `finish` method on the command buffer.
Rework the entire mechanism for specifying memory for creating resources, inferring the correct options from the new usage flags.
Adds image data types and operations. At this point, lightly tested.
Update the backend to reflect trait changes.
|
Merging this now even though it's in rough state, as I'm working concurrently on a number of subproblems (I've had to change the signatures of a few Device methods, mostly because outside Vulkan it's not valid to assume Fence and Semaphore are Copy). I'll continue to refine this as part of the #95 work. |
Very early so far, but uploading the progress I made.
Much of the code is adapted from Brian Merchant's https://github.com/bzm3r/piet-dx12 codebase.