[Mlir-commits] [mlir] [MLIR][XeGPU] Extend SGMapAttr and Add ConvertLayoutOp (PR #132425)

[Adam Siemieniuk]

================
@@ -154,33 +154,111 @@ def XeGPU_FenceScopeAttr:
     let assemblyFormat = "$value";
 }
 
-def XeGPU_SGMapAttr : XeGPUAttr<"SGMap", "sg_map"> {
+def XeGPU_LayoutAttr : XeGPUAttr<"Layout", "layout"> {
   let summary = [{
-    Describes the mapping between work item (WI) and the 2D tensor specified by the tensor descriptor.
+    Describes the data distribution to subgroups and work-items for a tensor
+    specified by the tensor descriptor.
   }];
   let description = [{
-    To distribute the XeGPU operation to work items, the tensor_desc must be specified with the sg_map
-    attribute at the tensor description creation time.
-    Within the `sg_map`, `wi_layout` specifies the layout of work items,
-    describing the mapping of work items to the tensor.
-    wi_layout[0] x wi_layout[1] must be equal to the total number of work items within a subgroup.
-    `wi_data` specifies the minimum number of data elements assigned to each work item for a single distribution.
-
-    E.g., #xegpu.sg_map<wi_layout = [1, 16], wi_data = [1, 1]>
-    In this example, the subgroup has 16 work items in wi_layout=[1, 16],
-    each accessing 1 element as specified by wi_data=[1, 1].
-
-    `wi_data[0] * wi_data[1]` can be greater than 1, meaning that each work item operates on multiple elements,
-    which is eventually lowered to "SIMT-flavor" vector, like SPIR-V vector or llvm vector, or packed to a storage data type.
-    The multiple elements indicated by `wi_data` can only be from one dimension and must be contiguous in the memory along either dimension.
+    XeGPU operations use `LayoutAttr` to define how data is distributed across subgroups and work-items.
+    This attribute is specified in tensor descriptors during tensor description creation. `LayoutAttr`
+    includes the following parameters:
+
+    * `sg_layout`: Specifies the total number of subgroups and their layout within a workgroup.
+      It is mandatory for workgroup-level programming and optional for subgroup programming. Its
+      presence implies workgroup-level code.
+    * `sg_data`: Defines the data size accessed per subgroup. It is optionally used with `sg_layout`
+      for workgroup-level programming. When it is left empty, the size accessed per subgroup can be
+      derived from the tensor shape and `sg_layout` using the formula:
+      `sg_data[i] = tensor_shape[i] / sg_layout[i]`.
+    * `inst_data`: Specifies the data size that is processed by an instruction. It is optionally
----------------
adam-smnk wrote:

+1

https://github.com/llvm/llvm-project/pull/132425