Optimize aarch64 GEMM kernel #32
Merged
Conversation
This file contains 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
Revise aarch64 kernel to use SIMD intrinsics. The structure is the same as the AVX 2 / FMA kernel, but the tile size is set to 8x8 as that performed best. On an M1 Mac performance for an M=N=K=1024 matmul increases from ~334 to ~418 GFLOPS.
robertknight
force-pushed
the
aarch64-gemm-v2
branch
from
bc57ee8 to
31622c1
Compare
Closed
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Outdated suggestions cannot be applied.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.
Revise aarch64 kernel to use SIMD intrinsics. The structure is the same as the AVX 2 / FMA kernel, but the tile size is set to 8x8 as that performed best. I did not port the prefetching logic as
std::arch::aarch64::_prefetchis currently nightly only :(On an M1 Mac (AWS mac2.metal) performance for an M=N=K=1024 matmul increases from ~334 to ~418 GFLOPS.
On a Graviton 2 (AWS c6g.xlarge) performance for an M=N=K=1024 matmul increases from ~115 to ~136 GFLOPS.
One observation from my tests is that ARM performance is much more sensitive to whether unrolling is used than the AVX kernel. Performance drops ~40 GFLOPS without it on the M1 in this test.
See #27