[Mlir-commits] [mlir] [mlir][scf] Add getPartialResultTilePosition to PartialReductionOpInterface (PR #120465)

[llvmlistbot at llvm.org]

================
@@ -476,29 +484,74 @@ struct LinalgOpPartialReductionInterface
                                          Location loc, ValueRange partialReduce,
                                          ArrayRef<int> reductionDims) const {
     auto linalgOp = cast<LinalgOp>(op);
-    SmallVector<int64_t> reductionDimsInt64(reductionDims);
-    auto reduction = b.create<linalg::ReduceOp>(
-        loc, partialReduce, linalgOp.getDpsInits(), reductionDimsInt64,
-        [&linalgOp](OpBuilder &b, Location loc, ValueRange inputs) {
-          int64_t numInits = linalgOp.getNumDpsInits();
-          SmallVector<Value> yieldedValues;
-          for (int idx : llvm::seq<int>(0, numInits)) {
+
+    // Permute the reduction dims as permuted by the partial result map.
+
+    int64_t numInits = linalgOp.getNumDpsInits();
+    SmallVector<Operation *> mergeOperations;
+    SmallVector<Value> replacements;
+    for (int idx : llvm::seq(numInits)) {
+      // linalg.reduce's iteration space is the result's iteration space (and
----------------
MaheshRavishankar wrote:

Is this correct? The iteration space of the `linalg.reduce` should be the operands iteration space cause the result does not have the reduction dimension.

https://github.com/llvm/llvm-project/pull/120465