[mlir][linalg] Migrate Detensorize pass to new dialect conversion driver #152912
+537
−109
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
matthias-springer
requested review from
dcaballe,
nicolasvasilache and
rengolin
as code owners
|
@llvm/pr-subscribers-mlir @llvm/pr-subscribers-mlir-linalg Author: Matthias Springer (matthias-springer) ChangesThe pass used to access erased operations and block arguments in the type converter. That is no longer supported in the new conversion driver. Full diff: https://github.com/llvm/llvm-project/pull/152912.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Detensorize.cpp b/mlir/lib/Dialect/Linalg/Transforms/Detensorize.cpp
index 830905495e759..221f95a8d8f33 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Detensorize.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Detensorize.cpp
@@ -458,6 +458,22 @@ struct LinalgDetensorize
}
};
+ /// A listener that forwards notifyBlockErased and notifyOperationErased to
+ /// the given callbacks.
+ struct CallbackListener : public RewriterBase::Listener {
+ CallbackListener(std::function<void(Operation *op)> onOperationErased,
+ std::function<void(Block *block)> onBlockErased)
+ : onOperationErased(onOperationErased), onBlockErased(onBlockErased) {}
+
+ void notifyBlockErased(Block *block) override { onBlockErased(block); }
+ void notifyOperationErased(Operation *op) override {
+ onOperationErased(op);
+ }
+
+ std::function<void(Operation *op)> onOperationErased;
+ std::function<void(Block *block)> onBlockErased;
+ };
+
void runOnOperation() override {
MLIRContext *context = &getContext();
DetensorizeTypeConverter typeConverter;
@@ -551,8 +567,22 @@ struct LinalgDetensorize
populateBranchOpInterfaceTypeConversionPattern(patterns, typeConverter,
shouldConvertBranchOperand);
- if (failed(
- applyFullConversion(getOperation(), target, std::move(patterns))))
+ ConversionConfig config;
+ auto onOperationErased = [&](Operation *op) {
+ opsToDetensor.erase(op);
+ detensorableBranchOps.erase(op);
+ };
+ auto onBlockErased = [&](Block *block) {
+ for (BlockArgument arg : block->getArguments()) {
+ blockArgsToDetensor.erase(arg);
+ }
+ };
+ CallbackListener listener(onOperationErased, onBlockErased);
+
+ config.listener = &listener;
+ config.allowPatternRollback = false;
+ if (failed(applyFullConversion(getOperation(), target, std::move(patterns),
+ config)))
signalPassFailure();
RewritePatternSet canonPatterns(context);
diff --git a/mlir/test/Dialect/Linalg/detensorize_0d.mlir b/mlir/test/Dialect/Linalg/detensorize_0d.mlir
index 74931cb0830bc..5c29b04630cad 100644
--- a/mlir/test/Dialect/Linalg/detensorize_0d.mlir
+++ b/mlir/test/Dialect/Linalg/detensorize_0d.mlir
@@ -53,10 +53,11 @@ func.func @detensor_op_sequence(%arg1: tensor<f32>, %arg2: tensor<f32>) -> tenso
}
// CHECK-LABEL: func @detensor_op_sequence
// CHECK-SAME: (%[[arg1:.*]]: tensor<f32>, %[[arg2:.*]]: tensor<f32>)
-// CHECK-DAG: %[[arg1_val:.*]] = tensor.extract %[[arg1]]
+// CHECK-DAG: %[[arg1_val_1:.*]] = tensor.extract %[[arg1]]
// CHECK-DAG: %[[arg2_val:.*]] = tensor.extract %[[arg2]]
-// CHECK: %[[detensored_res:.*]] = arith.addf %[[arg1_val]], %[[arg2_val]]
-// CHECK: %[[detensored_res2:.*]] = arith.mulf %[[arg1_val]], %[[detensored_res]]
+// CHECK-DAG: %[[arg1_val_2:.*]] = tensor.extract %[[arg1]]
+// CHECK: %[[detensored_res:.*]] = arith.addf %[[arg1_val_2]], %[[arg2_val]]
+// CHECK: %[[detensored_res2:.*]] = arith.mulf %[[arg1_val_1]], %[[detensored_res]]
// CHECK: %[[detensored_res3:.*]] = arith.divf %[[detensored_res]], %[[detensored_res2]]
// CHECK: %[[new_tensor_res:.*]] = tensor.from_elements %[[detensored_res3]]
// CHECK: return %[[new_tensor_res]]
|
rengolin
reviewed
Aug 10, 2025
| // CHECK: %[[detensored_res:.*]] = arith.addf %[[arg1_val]], %[[arg2_val]] | ||
| // CHECK: %[[detensored_res2:.*]] = arith.mulf %[[arg1_val]], %[[detensored_res]] | ||
| // CHECK-DAG: %[[arg1_val_2:.*]] = tensor.extract %[[arg1]] | ||
| // CHECK: %[[detensored_res:.*]] = arith.addf %[[arg1_val_2]], %[[arg2_val]] |
There was a problem hiding this comment.
CHECK-DAG can come in any order, but the op here specifies them in a particular order, and the two arg1 vals have the same pattern. This may randomly fail.
|
|
||
| config.listener = &listener; | ||
| config.allowPatternRollback = false; | ||
| if (failed(applyFullConversion(getOperation(), target, std::move(patterns), |
There was a problem hiding this comment.
Since the callback doesn't return anything, it won't change the success/failure result, so this should be fine.
matthias-springer
force-pushed
the
users/matthias-springer/prototype_one_shot
branch
from
09d8ee9 to
2422ce2
Compare
matthias-springer
requested review from
Mogball,
grypp,
aartbik,
PeimingLiu,
yinying-lisa-li and
banach-space
as code owners
matthias-springer
force-pushed
the
users/matthias-springer/prototype_one_shot
branch
from
2422ce2 to
5ade1f5
Compare
Base automatically changed from
users/matthias-springer/prototype_one_shot
to
main
August 13, 2025 15:40
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.
The pass used to access erased operations and block arguments in the type converter. That is no longer supported in the new conversion driver.
Depends on #151865.