[Mlir-commits] [mlir] [mlir][sparse] end-to-end matmul between Dense and BSR tensors (PR #71448)

[Aart Bik]

================
@@ -139,13 +139,13 @@ class SparsificationAndBufferizationPass
     // of `bufferization.alloc_tensor` ops.
     {
       OpPassManager pm("builtin.module");
-      pm.addPass(
-          createSparseReinterpretMapPass(ReinterpretMapScope::kGenericOnly));
+      // We need to reinterpret maps on GenericOp, EmptyOp, and AllocTensorOp.
+      // empty BSR
+      pm.addPass(createSparseReinterpretMapPass(ReinterpretMapScope::kAll));
       pm.addPass(createSparsificationPass(sparsificationOptions));
       pm.addNestedPass<func::FuncOp>(createStageSparseOperationsPass());
----------------
aartbik wrote:

I suspect that our GPU libgen path will break after this, since we recognize the original matmul linalg patterns. We should probably run just the GPU sparsification even earlier

https://github.com/llvm/llvm-project/pull/71448