[Mlir-commits] [mlir] [mlir][xegpu] Relax rank restriction of TensorDescType (PR #145916)

[Adam Siemieniuk]

================
@@ -67,8 +67,11 @@ def XeGPU_ScatterTensorDescAttr: XeGPU_TensorDescAttr<"ScatterTensorDesc", "scat
         TensorDesc is located, `Global` device memory or `Shared` local memory.
         It is default to `Global`.
 
-    2.  `chunk_size`: indicates number of contiguous elements accessed for each
-        offset, default is 1. It is used with `scattered` attr only.
+    2. `chunk_size`: Specifies the number of contiguous elements accessed per offset.
+      The default value is 1. While XeGPU supports a range of chunk sizes, hardware
+      may only allow specific values (e.g., 1, 2, 3, 4, 8, 16, 32, 64, 128, 256).
+      Therefore, XeGPU will legalize the chunk size as needed prior to lowering to
+      hardware instructions.
----------------
adam-smnk wrote:

nit: I agree with the design direction but it's a bit too detailed info especially for a "random" attribute parameter doc

This kind of general note (about validation in general) would be suitable for the top-level dialect description.

https://github.com/llvm/llvm-project/pull/145916