[Mlir-commits] [mlir] [MLIR][XeGPU] Matrix load/store subgroup distribution (PR #165008)

[Jianhui Li]

================
@@ -38,47 +38,52 @@ void XeGPUDialect::initialize() {
       >();
 }
 
-/// Generates instructions to compute offsets for a subgroup identified by
-/// its multidimensional indices (sgId), using the specified subgroup layout
-/// (sgLayout), subgroup data dimensions (sizePerSg), and the overall data
-/// dimensions (sizePerWg).
+// A `srcShape` consists of N distribution units, each being `subShapesLayout` x
+// `subShape`. A `delinearizedId` is used to identify a particular `subShape`
+// within each distribution unit.
+// Example:
+// WG data is 128x256. SG data is 16x32, in 4x2 layout, this gives a
+// distribution unit of shape 64x64, we have 2x4 such distribution units.
+// `delinearizedId` is used to identify a 16x32 of a subgroup in each
+// distribution unit.
 static SmallVector<SmallVector<Value>>
-genOffsetsComputingInsts(OpBuilder &builder, Location loc,
-                         SmallVector<Value> sgId, ArrayRef<int64_t> sgLayout,
-                         ArrayRef<int64_t> sizePerSg,
-                         ArrayRef<int64_t> sizePerWg) {
-
+genOffsets(OpBuilder &builder, Location loc, SmallVector<Value> delinearizedId,
----------------
Jianhui-Li wrote:

I think the function should be genCoords(). Inside the function, when we compute each individual value, it is fine to still use offset. 

A coordinate (coord) in n-d tensor is a vector of logical offsets.  

https://github.com/llvm/llvm-project/pull/165008