Use cuda events API to make profiler better on GPU

[abadams]

Currently the sampling profiler doesn't measure anything really meaningful for GPU schedules. Using the CUDA events API we could do better and at least get an accurate per-kernel runtime like nvprof does.

[mcourteaux]

I'm not 100% sure what you suggest here, but I'd like to point out that I managed to run nvprof and visualize the result with nvvp of GPU-enabled Halide applications. Very insightful!

I'm not aware of a sampling profiler for Halide. Are you talking about just the Linux perf tool?

[abadams]

Halide has a built-in sampling profiler that turns on when you add the -profile target flag. It's not particularly useful for the GPU though, which is what this issue is about. I also use nvvp currently for GPU schedules.