@@ -128,6 +128,9 @@ def bundle_xegpu_reduction_schedule(
128128 assert len (wg_tile ) == len (sg_tile ) == len (reduction_tile ), (
129129 "wg_tile, sg_tile, and reduction_tile must have the same number of dimensions"
130130 )
131+ assert all (red == 0 or sg == 0 for red , sg in zip (reduction_tile , sg_tile )), (
132+ "Reduction dimensions cannot be distributed across subgroups"
133+ )
131134 if wg_subtile :
132135 assert len (wg_subtile ) == len (wg_tile ), (
133136 "wg_subtile must have the same number of dimensions as wg_tile"
@@ -299,11 +302,9 @@ def tile_and_fuse_reduction(reduction_op, tile_sizes):
299302 launch_op = match_and_split (func , ops = {"gpu.launch" })
300303 num_subgroups = 1
301304 _wg_tile = wg_subtile if apply_wg_subtile else wg_tile
302- for i , ( wg , red , sg ) in enumerate ( zip (_wg_tile , reduction_tile , sg_tile ) ):
305+ for wg , sg in zip (_wg_tile , sg_tile ):
303306 if wg > 0 and sg > 0 :
304307 num_subgroups *= wg // sg
305- if red > 0 and sg > 0 :
306- num_subgroups *= red // sg
307308 num_threads = num_subgroups * subgroup_size
308309 xegpu .set_gpu_launch_threads (launch_op [0 ], threads = [num_threads , 1 , 1 ])
309310
@@ -329,9 +330,7 @@ def tile_and_fuse_reduction(reduction_op, tile_sizes):
329330 store_nd_ops = match (gpu_func , ops = {"xegpu.store_nd" })
330331 store_matrix_ops = match (gpu_func , ops = {"xegpu.store_matrix" })
331332 sg_layout = [1 ] * ndims
332- for i , (wg , red , sg ) in enumerate (zip (_wg_tile , reduction_tile , sg_tile )):
333- if red > 0 and sg > 0 :
334- sg_layout [i ] = int (red // sg )
333+ for i , (wg , sg ) in enumerate (zip (_wg_tile , sg_tile )):
335334 if wg > 0 and sg > 0 :
336335 sg_layout [i ] = int (wg // sg )
337336 sg_data = [1 ] * ndims
0 commit comments