[water] Fix elements_per_thread propagation to ignore memory operands #623
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
Implements elements per thread propagation for MMA operations. Fixes iree-org#608. Signed-off-by: tyb0807 <[email protected]>
Changes: - ReadOp: Only propagate attribute to result (register), ignore memory - WriteOp: Only validate/propagate with register operand, ignore memory This fixes false positives where memory resharding was incorrectly flagged as propagation errors. Fixes iree-org#622. Signed-off-by: tyb0807 <[email protected]>
ftynse
approved these changes
Dec 23, 2025
Contributor
ftynse
left a comment
ftynse
left a comment
There was a problem hiding this comment.
stacked commit LGTM modulo nits
|
|
||
| def ReadOp : WaveOp<"read", [ | ||
| WaveInferTypeOpInterface, IdentityTypeInferenceOpTrait, | ||
| WaveElementsPerThreadOpInterface, AttrBasedElementsPerThreadOpTrait, |
Contributor
There was a problem hiding this comment.
This trait becomes unused, remove it entirely.
Comment on lines
+1353
to
+1354
| // ReadOp doesn't propagate backward to memory operand | ||
| // Memory is decoupled from register dataflow for elements_per_thread |
Contributor
There was a problem hiding this comment.
Ultra-nit: use full sentences in comments and terminate them with a full stop. This will make it more readable when the comment is inevitably reflowed by tooling.
| llvm::MutableArrayRef<wave::ElementsPerThreadLatticeValue>, | ||
| llvm::raw_ostream &errs) { | ||
| // WriteOp only validates that elements_per_thread attribute matches register | ||
| // operand Memory operand is ignored for propagation - you can write to memory |
Contributor
There was a problem hiding this comment.
Like here, if you had a full stop, it would have been more readable.
| // Validate register operand (value_to_store) matches attribute | ||
| wave::ElementsPerThreadLatticeValue expectedValue(*elementsPerThread); | ||
| llvm::ArrayRef<wave::ElementsPerThreadLatticeValue> valueOnly = | ||
| operandElements.slice(0, 1); // Only first operand (value_to_store) |
Contributor
There was a problem hiding this comment.
Nit: use the getValueToStoreMutable().getOperandNumber() trick to avoid hardcoding positions.
| func.func @read_write_conflict(%mem: !wave.tensor<[@M] of f16, <global>>) attributes {wave.hyperparameters = #wave.hyperparameters<{M = 128}>} { | ||
| %reg = wave.read %mem {elements_per_thread = 4} : (!wave.tensor<[@M] of f16, <global>>) -> !wave.tensor<[@M] of f16, <register>> | ||
| // expected-error @below {{failed to propagate elements per thread backward: mismatch between elements_per_thread attribute (8) and operand #0 (4)}} | ||
| // expected-error @below {{failed to propagate elements per thread backward: mismatch between elements_per_thread attribute (8) and register operand #0 (4)}} |
Contributor
There was a problem hiding this comment.
Nit: "register operand #0" reads like there are multiple register operands. I'd consider just keeping as is.
| // CHECK-LABEL: @write_backward_propagation | ||
| func.func @write_backward_propagation(%mem: !wave.tensor<[@M] of f16, <shared>>) attributes {wave.hyperparameters = #wave.hyperparameters<{M = 128}>} { | ||
| %cst = arith.constant 0.0 : f16 | ||
| // RegisterOp without explicit elements_per_thread - should get it from backward propagation |
Contributor
There was a problem hiding this comment.
I think we punted on having explicit elements_per_thread on registers
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Stacked PRs, do not merge.
Changes:
This fixes false positives where memory resharding was incorrectly
flagged as propagation errors.
Fixes #622.