coalesced reads/writes in CUDA

Question

Is there a way to check my kernel reads and writes in a coalesced way from/to global memory? I've been trying ways to make sure my kernel reads and writes to memory efficiently to get a better performance.

Thanks

Answer 1

Use a profiler such as nvprof

The gld_efficiency and gst_efficiency metrics will give you a direct measure of percentage of coalesced global loads and stores. For example on Linux:

nvprof --metrics gld_efficiency,gst_efficiency ./my_app