[Mlir-commits] [mlir] [MLIR][XeGPU] Validate DPAS operand types against uArch in XeGPUToXeVM conversion (PR #185081)

[Arjun Bhamra]

================
@@ -902,6 +903,43 @@ class DpasToXeVMPattern : public OpConversionPattern<xegpu::DpasOp> {
     auto bTy = cast<VectorType>(op.getRhs().getType());
     auto resultType = cast<VectorType>(op.getResultType());
 
+    // get the correct dpasInst by getting info from chip
+    auto chipStr = xegpu::getChipStr(op);
+    if (!chipStr)
+      return rewriter.notifyMatchFailure(op, "cannot determine target chip");
+
+    const auto *uArch = mlir::xegpu::uArch::getUArch(*chipStr);
+    if (!uArch)
+      return rewriter.notifyMatchFailure(op, "unsupported target uArch");
+
+    auto *dpasInst = const_cast<xegpu::uArch::SubgroupMatrixMultiplyAcc *>(
----------------
abhamra wrote:

I believe so, because currently `getSupportedTypes` is not a const function right now, and we need that. However, that seems like a reasonable change. What do you think?

https://github.com/llvm/llvm-project/pull/185081