[Mlir-commits] [mlir] [mlir][sparse] simplify ConvertOp rewriting rules (PR #68350)

[Yinying Li]

================
@@ -1066,6 +1066,44 @@ OpFoldResult ConvertOp::fold(FoldAdaptor adaptor) {
   return {};
 }
 
+bool ConvertOp::directConvertable() {
+  if (isSortCOOConvert())
+    return false;
+
+  SparseTensorType srcStt = getSparseTensorType(getSource());
+  SparseTensorType dstStt = getSparseTensorType(getDest());
+
+  // We can always directly convert to unordered sparse tensor or dense tensor
+  // since dense tensor support random access.
+  if (dstStt.isAllDense() || !dstStt.isAllOrdered())
+    return true;
+
+  if (srcStt.isAllOrdered() && dstStt.isAllOrdered() &&
+      srcStt.hasSameDimToLvl(dstStt)) {
+    return true;
+  }
+
+  // Source and dest tensors are ordered in different ways. We only do direct
+  // dense to sparse conversion when the dense input is defined by a sparse
+  // constant. Note that we can theoritically always directly convert from dense
----------------
yinying-lisa-li wrote:

typo: theoretically?

https://github.com/llvm/llvm-project/pull/68350