|
cuDDL
|
How a replicated_database search splits and delivers its tiles. More...
Public Attributes | |
| all_to_all_shard | shard {} |
| This process's part of the pairs, when several processes share the search. | |
| bool | in_query_order = false |
| Deliver tiles one at a time in increasing query order, for example to write output that does not depend on the device count. | |
| std::function< uint64_t()> | next_tile {} |
Claim tiles from a counter that several processes share instead of keeping the tiles shard assigns, so faster processes do more of the search; see all_to_all_tile_queue. | |
How a replicated_database search splits and delivers its tiles.
Definition at line 40 of file replicated_database.cuh.
| bool cuddl::replica_search::in_query_order = false |
Deliver tiles one at a time in increasing query order, for example to write output that does not depend on the device count.
Each replica then waits for its turn before its callback, so callbacks no longer overlap. Not available with next_tile.
Definition at line 46 of file replicated_database.cuh.
| std::function<uint64_t()> cuddl::replica_search::next_tile {} |
Claim tiles from a counter that several processes share instead of keeping the tiles shard assigns, so faster processes do more of the search; see all_to_all_tile_queue.
shard.count must still give the number of processes.
Definition at line 50 of file replicated_database.cuh.
| all_to_all_shard cuddl::replica_search::shard {} |
This process's part of the pairs, when several processes share the search.
Definition at line 42 of file replicated_database.cuh.