performance variation when measuring the perf with "do_bench()" in triton.testing

[stephen-youn]

Hi All,
I have a question on what would the reason that we have a variation in the measured performance when we run do_bench and measure the perf through torch.cuda.event (link).
and i wonder whether we need extra line of "torch.cuda.synchronize()" for every end_event record in the code (link)
the other euqestion is how the "5" comes as the initial number of runs to estimate n_warmpup and n_repeat. and what's the reason for formula behind this.
thanks