[Mlir-commits] [mlir] [MLIR][XeGPU] Optimize create_mask sg-to-lane lowering (PR #221198)

[Artem Kroviakov]

================
@@ -864,28 +864,8 @@ struct SgToLaneVectorBitcast : public OpConversionPattern<vector::BitCastOp> {
 };
 
 /// Distributes a subgroup-level vector.create_mask or vector.constant_mask op
-/// to lane-level. Uses `computeDistributedCoords()` to obtain the
-/// coordinates each lane owns, then compares each coordinate against the
-/// original mask bounds using `arith.cmpi slt`. The per-element boolean
-/// results are assembled into the distributed mask vector.
-///
-/// For multi-dimensional masks, the element is in-bounds when ALL dimensions
-/// satisfy `coord[i] < bound[i]`.
-///
-/// Example (1D):
-///   layout = #xegpu.layout<lane_layout = [16], lane_data = [1]>
-///   %mask = vector.create_mask %m0 : vector<16xi1>
-/// For lane k, computeDistributedCoords gives coord = [k], so:
-///   %in_bounds = arith.cmpi slt, %coord, %m0  →  i1
-///   %mask = vector.broadcast %in_bounds : i1 to vector<1xi1>
-///
-/// Example (2D):
-///   layout = #xegpu.layout<lane_layout = [8, 2], lane_data = [1, 1]>
-///   %mask = vector.create_mask %m0, %m1 : vector<8x4xi1>
-/// Each WI owns a 1x2 slice. computeDistributedCoords returns 2 coords:
-///   [[r0, c0], [r0, c1]]
-/// For each coord: in_bounds = (r < m0) && (c < m1)
-///   %mask = vector.from_elements %bit0, %bit1 : vector<1x2xi1>
+/// to lane-level. An element is in-bounds when every dim satisfies
+/// `coord[d] < bound[d]`.
----------------
akroviakov wrote:

Added

https://github.com/llvm/llvm-project/pull/221198