@@ -194,8 +194,8 @@ struct policy_selector
194194 __make_scaled_segmented_radix_sort_policy (
195195 256 , 23 , BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, radix_bits),
196196 sub_warp_merge_sort_policy{
197- 256 , large_items ? 8 : 2 , small_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_TRANSPOSE , LOAD_LDG},
198- sub_warp_merge_sort_policy{256 , 16 , medium_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_TRANSPOSE , LOAD_LDG},
197+ 256 , large_items ? 8 : 2 , small_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_DIRECT , LOAD_LDG},
198+ sub_warp_merge_sort_policy{256 , 16 , medium_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_DIRECT , LOAD_LDG},
199199 500 };
200200 }
201201
@@ -208,8 +208,8 @@ struct policy_selector
208208 __make_scaled_segmented_radix_sort_policy (
209209 256 , 23 , BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, radix_bits),
210210 sub_warp_merge_sort_policy{
211- 256 , keys_only ? 4 : 2 , small_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_TRANSPOSE , LOAD_DEFAULT},
212- sub_warp_merge_sort_policy{256 , 32 , medium_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_TRANSPOSE , LOAD_DEFAULT},
211+ 256 , keys_only ? 4 : 2 , small_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_DIRECT , LOAD_DEFAULT},
212+ sub_warp_merge_sort_policy{256 , 32 , medium_itp, WARP_LOAD_TRANSPOSE, WARP_STORE_DIRECT , LOAD_DEFAULT},
213213 500 };
214214 }
215215
0 commit comments