[Mlir-commits] [mlir] [MLIR][XeGPU] Allow load/store/prefetch uses [memref+offset] instead of tdesc (PR #150576)

Adam Siemieniuk llvmlistbot at llvm.org
Wed Jul 30 05:40:46 PDT 2025


================
@@ -628,35 +628,65 @@ def XeGPU_PrefetchOp : XeGPU_Op<"prefetch", []> {
     As compared to prefetch_nd, which works on non-scattered TensorDesc,
     it works on scattered TensorDesc instead.
 
-    Example:
+    Example 1:
     ```mlir
       xegpu.prefetch %tdesc {l1_hint = #xegpu.cache_hint<cached>,
                              l2_hint = #xegpu.cache_hint<cached>,
                              l3_hint = #xegpu.cache_hint<cached>}
         : !xegpu.tensor_desc<16xf16>
     ```
+    
+    Example 2:
+    A variant accepts memref as base pointer and an offset instead of scattered TensorTdesc. It combines "create scattered TensorTdesc" and "prefetch with scattered TensorTdesc". The source operand could be a raw pointer (uint64_t). Please refer to create_tdesc for the restriction of memref. 
----------------
adam-smnk wrote:

nit: line breaks for readability

https://github.com/llvm/llvm-project/pull/150576


More information about the Mlir-commits mailing list