Room for More

Occupancy on a GPU streaming multiprocessor (SM) is the ratio of active warps to the maximum warps the SM can host concurrently. Which of the following is most likely to reduce occupancy for a kernel?

Reducing the number of arithmetic instructions per warp

High per-thread register usage, since more registers per thread means fewer resident warps fit within the SM's fixed register file

Increasing the total number of thread blocks launched in the grid

Using coalesced memory accesses instead of strided ones