Kernel launch parameters for thrust::for_each (and other algorithms)

[mz-zoox]

Is there a way to customize the kernel launch parameter for thrust algorithms? thrust::for_each always launches 512 CUDA threads per block. I am wondering if it something user can customize for performance tuning?

Also related to launch parameters, but possible a new topic entirely. Is it possible to use shared memory in functor passed to thrust::for_each and if so, is dynamic shared memory possible and how to specify the size?

[alliepiper]

Thrust does not expose block size or shmem allocations to users, these are always hardcoded for each algorithm/architecture. You'll need to write a custom CUDA kernel outside of Thrust if you want this level of control.