[Mlir-commits] [mlir] [mlir][xegpu] Block a convert_layout at a tile both layouts accept (PR #227892)

[Jianhui Li]

================
@@ -112,11 +112,12 @@ XeGPUBlockingPass::getTileShape(Operation *op) const {
         convertLayoutOp.getEffectiveInputLayout().getEffectiveInstDataAsInt();
     auto targetInstData =
         convertLayoutOp.getTargetLayout().getEffectiveInstDataAsInt();
-    // return the one with larger size
-    if (computeProduct(inputInstData) >= computeProduct(targetInstData))
-      return inputInstData;
-    else
-      return targetInstData;
+    assert(inputInstData.size() == targetInstData.size() &&
+           "convert_layout layouts must both carry inst_data of the same rank");
+    SmallVector<int64_t> tile(inputInstData.size());
+    for (size_t i = 0; i < tile.size(); ++i)
+      tile[i] = std::max(inputInstData[i], targetInstData[i]);
----------------
Jianhui-Li wrote:

Note that the inst_data picked is only feed to convert_layout. For convert_layout op, it has to conduct on a new data tile, even if it is bigger than other uArch sizes,or else the layout lost the meaning.  

https://github.com/llvm/llvm-project/pull/227892