[Mlir-commits] [mlir] [MLIR][XeGPU] Add non-splat constant distribution in SgToLane pass (PR #205575)

[Jianhui Li]

================
@@ -180,6 +180,27 @@ gpu.func @arith_constant_non_splat() {
   gpu.return
 }
 
+// With lane_data > 1 each lane owns a lane_data-sized block, so multiple
+// elements are extracted and assembled with vector.from_elements.
+// CHECK-LABEL: gpu.func @arith_constant_non_splat_lane_data
+// CHECK: %[[CST:.*]] = arith.constant dense<[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]> : vector<32xindex>
+// CHECK: %[[LANE:.*]] = gpu.lane_id
+// CHECK: %[[ELEM0:.*]] = vector.extract %[[CST]][%{{.*}}] : index from vector<32xindex>
+// CHECK: %[[ELEM1:.*]] = vector.extract %[[CST]][%{{.*}}] : index from vector<32xindex>
+// CHECK: %[[RES:.*]] = vector.from_elements %[[ELEM0]], %[[ELEM1]] : vector<2xindex>
----------------
Jianhui-Li wrote:

The form_elements always form a 1d vector.  What if the constant data is nd? I think it should keep original rank by doing extract_slice/insert_slice. 

https://github.com/llvm/llvm-project/pull/205575