[Mlir-commits] [mlir] [MLIR][XeGPU] Add sg_map attribute to support Work Item level semantics (PR #108864)

[Petr Kurapov]

================
@@ -77,12 +77,16 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
 
     // A TensorDesc with 8x16 f32 elements for a memory region in shared memory space.
     xegpu.tensor_desc<8x16xf32, #xegpu.tdesc_attr<memory_scope = slm>>
+
+    // A TensorDesc with a sg_map
+    xegpu.tensor_desc<8x16xf32, #xegpu.sg_map<wi_layout = [1, 16], wi_data = [1, 1]>>
     ```
   }];
 
   let parameters = (ins ArrayRefParameter<"int64_t">: $shape,
                         "mlir::Type": $elementType,
-                        OptionalParameter<"mlir::Attribute">: $encoding);
+                        OptionalParameter<"mlir::Attribute">: $encoding,
+                        OptionalParameter<"mlir::Attribute">: $sg_map);
----------------
kurapov-peter wrote:

> Is it also available to scattered TensorDesc (created by create_tdesc)

Yes, it produces the same type, so the map can (and should) be attached to the scattered variation. It works now but I'll add a test to cover it.

> If so, is there any restrictions on it?

There are no additional restrictions on the layout; it only has to match the sub-group size. The data WI field represents the minimum requirement for the distribution (e.g., to cover packing). We may want to bind it to the chunk size somehow but I don't think it is necessary at the moment.

https://github.com/llvm/llvm-project/pull/108864