[llvm] [offload] Allow replay repetitions and report basic timing (PR #193388)

[Kevin Sala Penades via llvm-commits]

kevinsala wrote:

@koparasy
> Most of the optimizations that I can think off affect the device execution time and not the host execution time. I have observed that host measuring can be extremely unreliable especially on fast kernels.

I agree. I believe both timings are interesting to report. We have plans to extend and provide more precise kernel time through stream events. But it will require a few more changes in the plugins since it should be a more generic mechanism (not only for record replay). For now, users can leverage profiling tools to extract kernel time.

https://github.com/llvm/llvm-project/pull/193388