RadixSortStoreAlgorithm#

enum cub::RadixSortStoreAlgorithm#

cub::RadixSortStoreAlgorithm enumerates different algorithms to write partitioned elements (keys, values) stored in shared memory into global memory.

Currently applies only to writing 4B keys in full tiles; in all other cases, RADIX_SORT_STORE_DIRECT is used.

Values:

enumerator RADIX_SORT_STORE_DIRECT#

Elements are statically distributed among block threads, which write them into the appropriate partition in global memory.

This results in fewer instructions and more writes in flight at a given moment, but may generate more transactions.

enumerator RADIX_SORT_STORE_ALIGNED#

Elements are distributed among warps in a block distribution.

Each warp goes through its elements and tries to write them while minimizing the number of memory transactions. This results in fewer memory transactions, but more instructions and less writes in flight at a given moment.