[Mlir-commits] [mlir] [MLIR][XeGPU] Add lowering from transfer_read/transfer_write to load_gather/store_scatter (PR #152429)

[Jianhui Li]

================
@@ -165,6 +496,20 @@ struct TransferReadLowering : public OpRewritePattern<vector::TransferReadOp> {
     if (failed(transferPreconditions(rewriter, readOp)))
       return failure();
 
+    auto chip = xegpu::getXeGPUChipStr(readOp);
+    if ( chip != "pvc" && chip != "bmg") {
----------------
Jianhui-Li wrote:

If the target is "PVC" or "BMG", lower the transfer_read to 2d block load/store. For other targets, lower it to scattered load/store.  

https://github.com/llvm/llvm-project/pull/152429