Description

• Provide an interface similar to cudaOccupancyMaxActiveBlocksPerMultiprocessor to facilitate max_blocks estimation
• Add an option to disable the grid-stride loop

Context

I noticed that the grid-stride loop is used by default in cuda_kernel_template_forward. I have a basic understanding of its benefits, but I've encountered some issues:

• The grid-stride loop is enabled by default, but in cases where I'm certain it's unnecessary, disabling it could provide a slight performance improvement (possibly due to better register usage).
• It's not trivial to estimate an appropriate max_blocks value — I typically need to use tools like Nsight Compute to determine how many blocks can run on an SM before arriving at the final max_blocks setting.
• The grid-stride loop uses size_t (64-bit integer), which may introduce unnecessary 64-bit arithmetic in some cases.