Skip to content

[mlir][sparse] clone a empty sparse tensor when fuse convert into pro… #92158

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
May 14, 2024
Merged

[mlir][sparse] clone a empty sparse tensor when fuse convert into pro… #92158

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0ba059e
[mlir][sparse] clone a empty sparse tensor when fuse convert into pro…
May 14, 2024
73ac45f
add test
May 14, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 7 additions & 7 deletions mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorRewriting.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -302,17 +302,17 @@ struct FoldConvertIntoProducer : public OpRewritePattern<ConvertOp> {
!producer.getResult(0).hasOneUse()) {
return failure();
}
// Clone the materialization operation, but update the result to sparse.
rewriter.setInsertionPoint(producer);
Operation *init = producer.getDpsInitOperand(0)->get().getDefiningOp();
Operation *cloned = rewriter.clone(*init);
cloned->getResult(0).setType(op.getResult().getType());

rewriter.modifyOpInPlace(producer, [&]() {
producer.getDpsInitsMutable().assign(cloned->getResults());
producer.getResult(0).setType(op.getResult().getType());
});

Operation *materializeOp =
producer.getDpsInitOperand(0)->get().getDefiningOp();

rewriter.modifyOpInPlace(materializeOp, [&]() {
materializeOp->getResult(0).setType(op.getResult().getType());
});

rewriter.replaceAllOpUsesWith(op, producer);
op->erase();

Expand Down
44 changes: 44 additions & 0 deletions mlir/test/Dialect/SparseTensor/fuse_sparse_convert_into_producer.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,50 @@ func.func @fold_convert(%arg0: tensor<128x32x32x1xf32>, %arg1: tensor<128x32x32x
return %2 : tensor<128x32x32x1xf32, #CCCD>
}

#trait_bin = {
indexing_maps = [
affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>,
affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>,
affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
}

// CHECK-FOLD-LABEL: func.func @fold_convert_multi_use(
// CHECK-FOLD: tensor.empty() : tensor<128x32x32x1xf32>
// CHECK-FOLD: linalg.generic
// CHECK-FOLD: tensor.empty() : tensor<128x32x32x1xf32, #sparse>
// CHECK-FOLD: linalg.generic
// CHECK-FOLD-NOT: sparse_tensor.convert
func.func @fold_convert_multi_use(%arg0: tensor<128x32x32x1xf32>, %arg1: tensor<128x32x32x1xf32>,
%arg2: tensor<128x32x32x1xf32>, %arg3: tensor<128x32x32x1xf32>) -> (tensor<128x32x32x1xf32>, tensor<128x32x32x1xf32, #CCCD>) {
%cst = arith.constant 0.000000e+00 : f32
%cst_0 = arith.constant 1.000000e+00 : f32
%cst_1 = arith.constant 1.000000e+00 : f32

%0 = tensor.empty() : tensor<128x32x32x1xf32>
%1 = linalg.generic #trait_bin
ins(%arg0, %arg1 : tensor<128x32x32x1xf32>, tensor<128x32x32x1xf32>)
outs(%0 : tensor<128x32x32x1xf32>) {
^bb0(%in: f32, %in_1: f32, %out: f32):
%3 = arith.mulf %in, %in_1 : f32
linalg.yield %3 : f32
} -> tensor<128x32x32x1xf32>

// A second kernel that uses %0 as the init operand.
%3 = linalg.generic #trait_bin
ins(%arg2, %arg3 : tensor<128x32x32x1xf32>, tensor<128x32x32x1xf32>)
outs(%0 : tensor<128x32x32x1xf32>) {
^bb0(%in: f32, %in_1: f32, %out: f32):
%3 = arith.mulf %in, %in_1 : f32
linalg.yield %3 : f32
} -> tensor<128x32x32x1xf32>
%4 = sparse_tensor.convert %3 : tensor<128x32x32x1xf32> to tensor<128x32x32x1xf32, #CCCD>

return %1, %4 : tensor<128x32x32x1xf32>, tensor<128x32x32x1xf32, #CCCD>
}



// FIXME: The following kernel is not sparsifiable because `arith.select`
// operations is not handled by the sparse compiler at the moment.
Expand Down
Loading