[Mlir-commits] [mlir] [mlir][x86] Decouple Accumulator from ADD-Based Vector Contractions (over loops) (PR #204327)

[Julian Oppermann]

================
@@ -415,25 +415,21 @@ struct VectorContractBF16ToFMA
       Operation *accReadOp1 =
           traceToVectorReadLikeParentOperation(pairContractOp.getAcc());
 
-      // Iterate down to find the users of contact operations until it is store
-      // or transfer_write.
-      Operation *resultWriteOp0 =
-          traceToVectorWriteLikeUserOperation(contractOp.getResult());
-      Operation *resultWriteOp1 =
-          traceToVectorWriteLikeUserOperation(pairContractOp.getResult());
-
-      // Shuffle the accumulators of the contract operations.
-      LogicalResult readShuffle =
-          shuffleAfterReadLikeOp(rewriter, accReadOp0, accReadOp1, contractOp,
-                                 pairContractOp, nonUnitDim, accTy);
-
-      if (failed(readShuffle))
-        return rewriter.notifyMatchFailure(
-            contractOp, "Accumulator read is not by transfer_read or load");
+      if (!(dyn_cast<arith::ConstantOp>(accReadOp0))) {
----------------
jopperm wrote:

Again, what if the accumulator was a non-zero constant?

https://github.com/llvm/llvm-project/pull/204327