[NVIDIA][Backend] Add CoalesceAsyncCopy Pass for in-DotOpEnc Upcasting #5222
+234
−34
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
ggengnv
force-pushed
the
coalesce-upcast
branch
from
9a9fcb0
to
f1af158
Compare
ggengnv
changed the title
ggengnv
changed the title
ThomasRaoux
reviewed
Nov 22, 2024
| Value mask = copyOp.getMask(); | ||
| Value other = copyOp.getOther(); | ||
| auto srcTy = cast<RankedTensorType>(src.getType()); | ||
| auto blockEnc = cast<BlockedEncodingAttr>(srcTy.getEncoding()); |
There was a problem hiding this comment.
you can't assume the copy will use blocked layout
Comment on lines
98
to
102
| // replace the asyncCopy | ||
| auto newCopyOp = rewriter.create<AsyncCopyGlobalToLocalOp>( | ||
| copyOp.getLoc(), src, copyOp.getResult(), mask, other, | ||
| copyOp.getCache(), copyOp.getEvict(), copyOp.getIsVolatile()); | ||
| rewriter.replaceOp(copyOp, newCopyOp); |
There was a problem hiding this comment.
nit, you could do in place update
| #include "mlir/Support/LLVM.h" | ||
| #include "mlir/Transforms/Passes.h" | ||
| #include "triton/Analysis/Utility.h" | ||
| #include "triton/Conversion/TritonGPUToLLVM/Utility.h" |
There was a problem hiding this comment.
nit: this is a bit of a layering violation, getRegToSharedLayout probably belongs to triton gpu dialect utils.
|
Addressed comments - moved util to |
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Outdated suggestions cannot be applied.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 follow-up to the dotOp hoisting optimization for WGMMA (MMAv3). See #5003 (comment)
In short, when upcasting operand A in registers prior to WGMMA and when pipelining is enabled,
AsyncCopyGLobalToLocal's src gmem blocked encoding will havesizePerThread> smem view'svec(along the contiguous dimension). This will resulting in multiplecp.asyncinstructions being generated for a contiguous global data segment, resulting in uncoalesced loads. This was previously confirmed in ncu. See above comment for an example.I've added a generalized fix in a new pass after the pipeliner. I've reused the logic in the LLVM lowering for
AsyncCopyGlobalToLocalto calculate the max contiguous copy size. I compare that to the blockEnc'ssizePerThreadalong the inner (contiguous) dimension. If the former is less than latter, I set the latter to former.When A is k-major, can verify a small perf improvement and that ncu no longer reports uncoalesced loads.
When A is m-major, this pass is a no-op because
copy size == sizePerThread == 16ptal, thanks @ThomasRaoux