On Fri, Feb 25, 2022 at 5:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:
>
> On Fri, Feb 25, 2022 at 05:38:45PM +0100, Magnus Hagander wrote:
> >
> > Per some off-list discussion with Julien, we have clearly been talking
> > in slightly different terms. So let's summarize the options into what
> > theÿ́d actually be:
> >
> > Option 0: what is int he patch now
> >
> > Option 1:
> > jit_count - number of executions using jit
> > total_jit_time - for sum of functions+inlining+optimization+emission time
> > min_jit_time - for sum of functions+inlining+optimization+emission time
> > max_jit_time - for sum of functions+inlining+optimization+emission time
> > mean_jit_time - for sum of functions+inlining+optimization+emission time
> > stddev_jit_time - for sum of functions+inlining+optimization+emission time
> > jit_functions - number of functions
> > jit_inlining_count - number of executions where inlining happened
> > jit_optimization_count - number of executions where optimization happened
> > jit_emission_count - number of executions where emission happened
> >
> > Option 2:
> > jit_count
> > jit_functions
> > jit_generation_time
> > jit_inlining_count
> > jit_inlining_time
> > jit_optimization_count
> > jit_optimization_time
> > jit_emission_count
> > jit_emission_time
>
> I'm for option 2, I think it's important to have the timing details for
> inlining and optimization and be able to compute correct stats.
I wonder if there might be an interesting middle ground, or if that is
making it too much. That is, we could have an
Option 3:
jit_count
total_jit_time - for sum of functions+inlining+optimization+emission time
min_jit_time - for sum of functions+inlining+optimization+emission time
max_jit_time - for sum of functions+inlining+optimization+emission time
mean_jit_time - for sum of functions+inlining+optimization+emission time
stddev_jit_time - for sum of functions+inlining+optimization+emission time
jit_functions
jit_generation_time
jit_inlining_count
jit_inlining_time
jit_optimization_count
jit_optimization_time
jit_emission_count
jit_emission_time
That is, we'd get the more detailed timings across the total time, but
not on the details. But that might be overkill.
But -- here's an updated patched based on Option 2.
--
Magnus Hagander
Me: https://www.hagander.net/
Work: https://www.redpill-linpro.com/
