I currently have a C/CUDA program which uses multiple CPUs to generate values in parallel to be passed to kernel functions which run on a GPU.

I see two general ways of handling the workflow:

1. Divide the work among CPUs, then within each single CPU context, run the GPU kernel function
2. Divide the work among CPUs, then, once all CPUs have completed their work, consolidate all values into a single array, and pass to GPU.

pretheoretically, is one of these options clearly more efficient than the other? Or is the answer case-specific? If case-specific, what are the relevant parameters most determinant of efficiency?

thank you for your time considering this question