[Mlir-commits] [mlir] [MLIR][SCF] Add an API to fuse consumer to a producer within scf loop (PR #88712)

Abhishek Varma llvmlistbot at llvm.org
Mon May 6 21:45:13 PDT 2024


================
@@ -1100,6 +1102,459 @@ mlir::scf::tileConsumerAndFuseProducersUsingSCF(
                                    replacements};
 }
 
+//===----------------------------------------------------------------------===//
+// tileAndFuseConsumerUsingSCF implementation.
+//===----------------------------------------------------------------------===//
+
+/// A utility function that checks whether the passed value has only one user.
+/// In case the defining operation is a tensor.insert_slice, it checks if the
+/// user is scf.yield.
+static LogicalResult checkAssumptionForFusingConsumer(Value result) {
+  Value::use_range uses = result.getUses();
+  if (!llvm::hasSingleElement(uses)) {
+    LLVM_DEBUG(llvm::dbgs() << "Too many uses of the candidate slice op\n");
+    return failure();
+  }
+  OpOperand &operandUse = (*uses.begin());
+  Operation *userOp = operandUse.getOwner();
+  if (!isa<scf::YieldOp>(userOp)) {
+    LLVM_DEBUG(llvm::dbgs()
+               << "Expected scf.yield to be the only user, but got -> "
+               << (*userOp));
+    return failure();
+  }
+  return success();
+}
+
+/// Fetch the first untiled consumer of a scf.for's result which is yielded by
+/// a tensor.insert_slice. This function makes the following assumptions :-
+/// 1.  tensor.insert_slice has scf.yield as its only user.
+/// 2.  scf.for's corresponding result has only one use.
+static FailureOr<OpOperand *>
+getUntiledConsumerFromSlice(tensor::InsertSliceOp candidateSliceOp) {
----------------
Abhishek-Varma wrote:

The API is pivoting on a `tensor.insert_slice` that has only one `scf.yield` as its user and only one `use`.

And the implementation is already enforcing the check [here](https://github.com/llvm/llvm-project/blob/53e607e9467f38dc97656000439ea115d57621e3/mlir/lib/Dialect/SCF/Transforms/TileUsingInterface.cpp#L1109-L1127).

https://github.com/llvm/llvm-project/pull/88712


More information about the Mlir-commits mailing list