Skip to content

Commit ecdd20c

Browse files
committed
kernel_bench: rms norm: cross-subgroup reduction is not supported
1 parent ec1a426 commit ecdd20c

1 file changed

Lines changed: 3 additions & 3 deletions

File tree

‎examples/xegpu/kernel_bench.py‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -208,9 +208,9 @@ def infer_parameters(
208208
"layer_kind": "reduction",
209209
"sizes": shape,
210210
"wg_tile": [1, 0, 128, 128],
211-
"wg_subtile": [0, 0, 16, 32],
212-
"sg_tile": [0, 2, 16, 32],
213-
"reduction_tile": [0, 32, 0, 0],
211+
"wg_subtile": [0, 0, 64, 128],
212+
"sg_tile": [0, 0, 8, 32],
213+
"reduction_tile": [0, 8, 0, 0],
214214
"subgroup_size": 16,
215215
}
216216
else:

0 commit comments

Comments
 (0)