@@ -141,8 +141,7 @@ struct copy_op<dtype, load, store, M, N, false> {
141
141
//
142
142
launch<
143
143
copy_kernel_vectorized<decltype (S), decltype (D), decltype (tiled_load),
144
- decltype (tiled_store), load>, CopyKernelVectorizedName<decltype (S), decltype (D), decltype (tiled_load),
145
- decltype (tiled_store), load>>(
144
+ decltype (tiled_store), load>, CopyKernelVectorizedName<load, Int<M>>>(
146
145
launch_policy{
147
146
cutlasscompat::dim3 (1 ), blockDim,
148
147
kernel_properties{sycl_exp::sub_group_size<SUBGROUP_SIZE>}},
@@ -194,8 +193,7 @@ struct copy_op<char, load, XE_2D_U8x2x32_ST_N, M, N, false> {
194
193
//
195
194
launch<
196
195
copy_kernel_vectorized<decltype (S), decltype (D), decltype (tiled_load),
197
- decltype (tiled_store), load>, CopyKernelVectorizedName<decltype (S), decltype (D), decltype (tiled_load),
198
- decltype (tiled_store), load>>(
196
+ decltype (tiled_store), load>, CopyKernelVectorizedName<load>>(
199
197
launch_policy{
200
198
cutlasscompat::dim3 (1 ), blockDim,
201
199
kernel_properties{sycl_exp::sub_group_size<SUBGROUP_SIZE>}},
@@ -247,8 +245,7 @@ struct copy_op<uint16_t, load, XE_2D_U16x2x16_ST_N, M, N, false> {
247
245
//
248
246
launch<
249
247
copy_kernel_vectorized<decltype (S), decltype (D), decltype (tiled_load),
250
- decltype (tiled_store), load>, CopyKernelVectorizedName<decltype (S), decltype (D), decltype (tiled_load),
251
- decltype (tiled_store), load>>(
248
+ decltype (tiled_store), load>, CopyKernelVectorizedName<load>>(
252
249
launch_policy{
253
250
cutlasscompat::dim3 (1 ), blockDim,
254
251
kernel_properties{sycl_exp::sub_group_size<SUBGROUP_SIZE>}},
@@ -310,8 +307,7 @@ struct copy_op<uint32_t, load, store, M_, N_, true> {
310
307
//
311
308
launch<
312
309
copy_kernel_vectorized<decltype (S), decltype (D), decltype (tiled_load),
313
- decltype (tiled_store), load>, CopyKernelVectorizedName<decltype (S), decltype (D), decltype (tiled_load),
314
- decltype (tiled_store), load>>(
310
+ decltype (tiled_store), load>, CopyKernelVectorizedName<load, Int<M>>>(
315
311
launch_policy{
316
312
cutlasscompat::dim3 (1 ), blockDim,
317
313
kernel_properties{sycl_exp::sub_group_size<SUBGROUP_SIZE>}},
0 commit comments