[mlir][sparse] add linalg elt-wise fusion to sparsifier pipeline

[aartbik]

yields better IR in general, and all end-to-end tests pass!

[llvmbot]

@llvm/pr-subscribers-mlir-sparse

@llvm/pr-subscribers-mlir

Author: Aart Bik (aartbik)

Changes

yields better IR in general, and all end-to-end tests pass!

---

Full diff: https://github.com/llvm/llvm-project/pull/90924.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp (+2-1)

diff --git a/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp b/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
index e58503d508ced4..ab4d90ec745d4a 100644
--- a/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
+++ b/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
@@ -31,8 +31,9 @@
 
 void mlir::sparse_tensor::buildSparsifier(OpPassManager &pm,
                                           const SparsifierOptions &options) {
-  // Rewrite named linalg ops into generic ops.
+  // Rewrite named linalg ops into generic ops and apply fusion.
   pm.addNestedPass<func::FuncOp>(createLinalgGeneralizeNamedOpsPass());
+  pm.addNestedPass<func::FuncOp>(createLinalgElementwiseOpFusionPass());
 
   // Sparsification and bufferization mini-pipeline.
   pm.addPass(createSparsificationAndBufferizationPass(