[Mlir-commits] [mlir] [MLIR][XeGPU] Introduce `xegpu::uArch` usage in target-sensitive passes (PR #163801)

[Charitha Saumya]

================
@@ -456,7 +482,22 @@ void LayoutInfoPropagation::visitPrefetchNdOp(
   // Here we assign the default layout to the tensor descriptor operand of
   // prefetch.
   auto tdescTy = prefetch.getTensorDescType();
-  auto prefetchLayout = getDefaultSIMTLayoutInfo(tdescTy);
+
+  auto uArch = getUArch(getChipStr(prefetch).value_or(""));
+  int subgroupSize = uArch->getSubgroupSize();
+  auto uArchInstruction =
+      std::static_pointer_cast<xegpu::uArch::PrefetchNdInstruction>(
+          uArch->getInstruction(xegpu::uArch::InstructionKind::STORE_ND));
+  int maxVecLength =
+      uArchInstruction
+          ->getSortedLaneVectorLengths(tdescTy.getElementTypeBitWidth())
+          .back();
+  SmallVector<int> instData;
+  if (tdescTy.getRank() == 1)
+    instData = {subgroupSize};
+  else
+    instData = {maxVecLength, subgroupSize};
----------------
charithaintc wrote:

here shouldn't we take the min of tile size and uArch max size?

https://github.com/llvm/llvm-project/pull/163801