[Mlir-commits] [mlir] [mlir][ArmNeon] Updates LowerContractionToSMMLAPattern with vecmat unroll patterns (PR #86005)

[Diego Caballero]

================
@@ -150,16 +150,30 @@ class LowerContractionToSMMLAPattern
       Value tiledAcc =
           extractOperand(op.getAcc(), accPermutationMap, accOffsets);
 
+      // With vecmat, tiled LHS and ACC will contain only one of 2 necessary
+      // rows along dimM. Broadcast both to the full width
+      if (isVecmat) {
+        auto lhsBroadcastType = VectorType::get(
+            {2, 8}, tiledLhs.getType().cast<ShapedType>().getElementType());
+        tiledLhs = rewriter.create<vector::BroadcastOp>(loc, lhsBroadcastType,
+                                                        tiledLhs);
+        auto accBroadcastType = VectorType::get(
+            {2, 2}, tiledAcc.getType().cast<ShapedType>().getElementType());
+        tiledAcc = rewriter.create<vector::BroadcastOp>(loc, accBroadcastType,
+                                                        tiledAcc);
----------------
dcaballe wrote:

Yes, that would work! I'm not sure if we would have to introduce the LLVM dialect dependence, though. I think in general we've been doing the same but using `arith.constant 0`  instead of the undef.

https://github.com/llvm/llvm-project/pull/86005