-
Notifications
You must be signed in to change notification settings - Fork 41
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Double buffering improvements #1511
base: develop
Are you sure you want to change the base?
Conversation
- Split the LDS reads and MFMA/WMMA into two independent loops - Have them into two separate stages (so that they can be executed in parallel) This is to make our pipeline similar to what CK is doing in: - https://github.com/ROCm/composable_kernel/blob/6d073d31bbc7d39d8b170d549f2af61970378150/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp
6c301ff
to
c9930bc
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Doesn't this make BlockwiseGemm kinda obselete?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If you're comfortable leaving dead code running around, I don't see any deep reason not to land this.
Since it doesn't look like it makes sense to e2e test this, can we get a lit test for this new structure? Maybe some sort of internal forceDoubleBufferForTesting
attribute
Re: BlockwiseGemm: I don't think so, because in BlockwiseGemm the intuition is that we do everything at the same time |
Re BlockwiseGemm, yeah, I misread the comment Re testing ... yeah, I'd like a |
This is fine with me for now as I had this long standing opinion, the interface for BlockwiseGemm is too limiting -- the fact that it requires LDS buffers to be passed in as inputs. Long-term : What would be better (as most other compilers IREE, Triton) is we do BlockwiseGemm take register input with the layout (i.e. tid, iter --> blockwise_tensor). Then let the lowering decides whether it needs to swizzle stuff with whatever way (DPP, LDS, etc) is suitable. We could even define a op : blockwise_copy to encapsulate layout changes. In the case of user-defined buffering (i.e. what we currently do in gridwise lowering of gemm), user is free to also load directly into the appropiate layout a BlockwiseGemm would prefer. So this would strictly limit the BlockwiseGemm to be just "compute" part of the split you are doing here. |
@giuseros should we close this ? I dont think this is happening. correct? |
This is to make our pipeline similar to what CK is doing in: