Is there a good implementation of reduction algorithm callable from kernel with dynamic parallelism?

Question

I see reductions algorithms in CUDA (such as summation and maximization over a range of elements) discussed in previous posts, but with dynamic parallelism, they could potentially be implemented in a different way. Is there a more efficient implementation which is callable from inside the kernels?

Answer 1

Is there a more efficient implementation which is callable from inside the kernels?

CUB provides a CUDA reduction primitive compatible with dynamic parallelism, namely, that can be called within kernels.