Prefix sum implementation WIP #14
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sorta works but deadlocks on larger inputs.
Still doesn't fix deadlocks tho :/
Still WIP
Fastest results on AMD at workgroup = 1024. Note, this has atomicOr workaround for correctness. Also note, not all targets will support a workgroup of this size; on shipping, we'd need to query and select at runtime.
Do a small sequential scan at the leaf of the hierarchy. That amortizes both the workgroup-scope tree reduction and the (still sequential) decoupled look-back to a larger number of inputs. Note: this falls short of a real performance evaluation because there's no attempt to warm up the GPU clock. But it's valid as a very rough swag.
Better for performance analaysis
Naga will accept ordinary loads and stores to atomic types, but tint will not.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a draft, I'm still working on it. I'll likely create another subdirectory for tests and add this to that, as overwriting the main hello example is not very good form. But I'm doing it for expedience.
The version at tip of tree as I write this (87e5b20) works well on AMD 5700 XT. In fact, it works very well - I'm seeing 36.4 billion elements/s, which is excellent. It's within a sliver of a compute shader that just copies input to output, and looking at GPU counters suggests that memory bandwidth is pretty well saturated.
This version also makes some progress on each spin, so does not depend on strong forward progress guarantees from the GPU.
That said, I am employing the atomicOr workaround for the atomic bugs I'm seeing, otherwise I get both incorrect results and hangs (try N_DATA = 1 << 17 for a nice mix of the two). I will probably work on a simplified version of the test to exercise the atomic problems without bringing in all of the complexity of full prefix sum.