I think calling thrust::count for 1000 times adds more latency than computation itself. Thrust supports sort, upper_bound and adjacent_difference methods to do this in only 3-5 operations.

I think calling thrust::count for 1000 times adds more latency than computation itself. Thrust supports sort, upper_bound and adjacent_difference methods to do this in only 3-5 operations like this:

• sort 1M inputs

• adjacent_difference

• segmented-reduction or reduce_by_key