[mlir][sparse] add source materizalization callback for sparse tensor codegen type converter.

[Peiming]

Summary

Required by scf.for to achieve 1:N type conversion (See D136314).

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorPasses.cpp

169

This looks okay, but shall we do this in the constructor in the SparseTensorCodegen.cpp file?
Just so that all logic is in one place?

mlir::SparseTensorTypeToBufferConverter::SparseTensorTypeToBufferConverter() {

addConversion([](Type type) { return type; });
addConversion(convertSparseTensorType);

}

[Peiming]

address comments from Aart.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorCodegen.cpp

754 ↗

(On Diff #469293)

Last nit: Not a sparse tensor.
;=)

(capital N)

757 ↗

(On Diff #469293)

Ah, this is a nice side-effect of moving it ;)
genTuple reuse

[Peiming]

fix comment.

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorPasses.cpp

struct SparseTensorCodegenPass
 
   SparseTensorCodegenPass() = default;
   SparseTensorCodegenPass(const SparseTensorCodegenPass &pass) = default;
 
   void runOnOperation() override {
     auto *ctx = &getContext();
     RewritePatternSet patterns(ctx);
     SparseTensorTypeToBufferConverter converter;
+    // Required by scf.for 1:N type conversion.
+    converter.addSourceMaterialization(
+        [](OpBuilder &builder, RankedTensorType tp, ValueRange inputs,
+           Location loc) -> Optional<Value> {
+          if (!getSparseTensorEncoding(tp))
+            // not a sparse tensor.
+            return llvm::None;
+          // Sparse compiler knows how to cancel out these casts.
+          return builder
+              .create<UnrealizedConversionCastOp>(loc, TypeRange(tp), inputs)
+              .getResult(0);
+        });
+
     ConversionTarget target(*ctx);
     // Most ops in the sparse dialect must go!
     target.addIllegalDialect<SparseTensorDialect>();
     target.addLegalOp<SortOp>();
     target.addLegalOp<PushBackOp>();
     // All dynamic rules below accept new function, call, return, and various
     // tensor and bufferization operations as legal output of the rewriting
     // provided that all sparse tensor types have been fully rewritten.