[Mlir-commits] [mlir] [Linalg] Add basic infra to add matchers for linalg.*conv*/*pool* ops (PR #163724)

[Andrzej Warzyński]

================
@@ -240,6 +240,472 @@ bool isReductionIterator(utils::IteratorType iteratorType) {
   return iteratorType == utils::IteratorType::reduction;
 }
 
+//===----------------------------------------------------------------------===//
+// Convolution matcher utilities
+//===----------------------------------------------------------------------===//
+
+/// Utility to match block body for linalg.pool* ops.
+template <typename... OpTypes>
+static bool bodyMatcherForPoolOps(Value yieldVal, Block *body) {
+  Operation *defOp = yieldVal.getDefiningOp();
+  if (!(isa_and_present<OpTypes>(defOp) || ...))
+    return false;
+
+  BlockArgument lhsArg = dyn_cast<BlockArgument>(defOp->getOperand(0));
+  BlockArgument rhsArg = dyn_cast<BlockArgument>(defOp->getOperand(1));
+  if (!lhsArg || !rhsArg)
+    return false;
+  return true;
+}
+
+static bool bodyMatcherForMaxSignedPoolOps(Value yieldVal, Block *body) {
+  return bodyMatcherForPoolOps<arith::MaximumFOp, arith::MaxSIOp>(yieldVal,
+                                                                  body);
+}
+
+// max_unsigned ops should not allow float data type.
+// TODO: Retire OPDSL logic. Refer to :
+// https://github.com/llvm/llvm-project/pull/163724#discussion_r2438940337
----------------
banach-space wrote:

IMO, a link to the GitHub issue as the central point for the discussion would be better: https://github.com/llvm/llvm-project/issues/164800#issuecomment-3436374358

Similar comment below.

https://github.com/llvm/llvm-project/pull/163724