Groupwise GEMM Full Kernel Tuning #4521
Open
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
Summary: This diff expands the templatization of FBGEMM's cutlass groupwise GEMM to include options to limit the swizzling of loaded tiles and implicitly transpose tiles. We introduce a bit of additional tuning that shows this allows us to get decent performance in both compute and memory bound domains. In a followup diff, we will dramatically expand these heuristics for better performance. Differential Revision: D78537461
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D78537466 |
Summary: Pull Request resolved: pytorch#4521 This diff introduces extensive kernel tuning and heuristics for Cutlass version of DeepGemm. With these heuristics, performance of the cutlass kernel is generally as good or better than DeepGemm for an equivalent workload and numerics are identical. This makes cutlass kernels in FBGEMM quite a bit easier to use than DeepGEMM as they are AOT compiled and do not require runtime tuning or JIT. We also expect substantially better performance for memory bound workloads like decode. The key tricks we used to get this performance were limiting tile swizzling for compute bound shapes, which reduces register pressure and dramatically improves performance and doing an implicit input transpose for memory bound shapes that allows more efficient WGMMA tiling. Benchmarking across a wide range of shapes shows the general performance benefits of these kernels in memory bound domains and their competitiveness with deepgemm in compute bound domains. {F1980376140} Differential Revision: D78537466
|
This pull request was exported from Phabricator. Differential Revision: D78537466 |
jwfromm
force-pushed
the
export-D78537466
branch
from
6f874ab to
11d27ab
Compare
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.
Summary:
This diff introduces extensive kernel tuning and heuristics for Cutlass version of DeepGemm. With these heuristics, performance of the cutlass kernel is generally as good or better than DeepGemm for an equivalent workload and numerics are identical. This makes cutlass kernels in FBGEMM quite a bit easier to use than DeepGEMM as they are AOT compiled and do not require runtime tuning or JIT. We also expect substantially better performance for memory bound workloads like decode.
The key tricks we used to get this performance were limiting tile swizzling for compute bound shapes, which reduces register pressure and dramatically improves performance and doing an implicit input transpose for memory bound shapes that allows more efficient WGMMA tiling.
Benchmarking across a wide range of shapes shows the general performance benefits of these kernels in memory bound domains and their competitiveness with deepgemm in compute bound domains.
{F1980376140}
Differential Revision: D78537466