NVIDIA/cccl

Move CUB tunings to dedicated headers

Closed this issue · 0 comments

All CUB algorithms have tunings they select via the policy hub, but they are sometimes placed in the dispatcher implementation header (e.g. cub/device/dispatch/dispatch_radix_sort.cuh) and sometimes in dedicated tuning headers (e.g. cub/device/dispatch/tuning/tuning_radix_sort.cuh).

We should finish the effort @gevtushenko started a while back and move all policy hubs and tunings to dedicated headers. This makes the files a bit more maintainable, especially during tuning when we have to extend the policy hub with new tunings.

Applies to the folloing CUB algos:

  • adjacent_difference
  • batch_memcpy
  • merge
  • merge_sort
  • reduce
  • radix_sort
  • segmented_sort
  • transform