[Mlir-commits] [mlir] [mlir][vector] Fix n-d transfer write distribution (PR #83215)

Lei Zhang llvmlistbot at llvm.org
Tue Feb 27 19:32:46 PST 2024


================
@@ -526,7 +535,25 @@ struct WarpOpTransferWrite : public OpRewritePattern<WarpExecuteOnLane0Op> {
     // 4. Reindex the write using the distribution map.
     auto newWarpOp =
         newWriteOp.getVector().getDefiningOp<WarpExecuteOnLane0Op>();
+
----------------
antiagainst wrote:

It would be nice to have a comment here to say what we are doing in a high level for the following code. Something like if the distributed vector is more than 1-D, we need to delinearize  the 1-D lane ID into the matching dimension. 

https://github.com/llvm/llvm-project/pull/83215


More information about the Mlir-commits mailing list