Is it possible to pre-program certain data to a specific cache level in the CUDA core?

Question

I understand that Fermi GPUs support prefetching to L1 or L2 cache. However, I cannot find anything in the CUDA Reference Guide.

Does CUDA Dues allow my kernel code to pre-select specific data at a specific cache level?

Answer 1

Well, not at the instruction level, but detailed information on prefetching in GPUs is here:

Multi-threaded prefetching mechanisms for GPGPU applications
(paper at the 2010 ACM Microarchitecture Symposium)

You can find instruction instructions in the nVIDIA PTX ISA document ; The corresponding prefetch and prefetchu instructions.