David Boehme

Results 48 comments of David Boehme

Oh good, you already implemented this, I thought I'd have to deal with the Kokkos fencing stuff now 😄 Do I interpret this correctly that you tell Kokkos "I don't...

Yeah, `CALI_CONFIG="cuda-activity-report(profile.kokkos)"` should work. Here's some more information about these reports: https://software.llnl.gov/Caliper/CUDA.html

Might be easier to just give the kokkos service a config flag that turns on fencing. Is there actually another situation besides the UVM event mapping where we really need...

> > First: hi! Hope you're enjoying AMD > > Had my first day today. I haven't sent out an email yet (only bc of procrastination) so I suspect this...

Is this ready to merge now @DavidPoliakoff ?

Oh, sorry for the delay. I took a closer look at this again - it seems the only thing that disables the fencing now is if `cuda-activity-report` is active? Seems...

> Why turn it off, though? If I'm not doing cuda-activity-report, the answers I get will be garbage, no? There are other configs (cuda-activity-profile) which also work for asynchronous tasks....

There's the cuda-activity-profile (in addition to -report), whatever we'll do for HIP, etc. Plus maybe you want to see the asynchronous times. I'm just saying it would be nice to...

> An aggressive proposal might be for configs to say what resources they require fences of (this is independently useful information). Then Kokkos could just query "ah, does the requested...

Hi @mrzv , Thanks for the report. Currently that's not directly possible, but I'll try to come up with something. You can create flat profiles with `select event.end#annotation,sum(time.inclusive.duration) group by...