K-ballo changed the topic of #ste||ar to: STE||AR: Systems Technology, Emergent Parallelism, and Algorithm Research | stellar.cct.lsu.edu | HPX: A cure for performance impaired parallel applications | github.com/STEllAR-GROUP/hpx | Buildbot: http://rostam.cct.lsu.edu/ | Log: http://irclog.cct.lsu.edu/
nanmiao has quit [Quit: Connection closed]
hkaiser has quit [Quit: bye]
<bering[m]> The bridge error might be due to recent events at freenode.
heller1 has quit [Quit: Idle for 30+ days]
hkaiser has joined #ste||ar
<ms[m]> gnikunj: would you mind trying this for your kokkos segfaults: https://github.com/msimberg/kokkos/commit/baef375deb5b2856294bbb982486809239c9ef8c?
<ms[m]> I couldn't reproduce a segfault happening where you said it was happening but I could reproduce a segfault quite regularly
<gnikunj[m]> ms thanks! Will do
<ms[m]> in principle I could say that you were using the backend wrongly, but it's subtle enough that it deserves a change
<ms[m]> it might also explain the performance differences you see
<ms[m]> I think I told you at some point that the instances are not threadsafe
<ms[m]> what that means though is that kokkos view construction is not threadsafe, because at the moment they all use the default instance for initialization
<ms[m]> I think there's been some talk among the kokkos developers that one could initialize a view using a specific instance but it's not there yet
<ms[m]> if you don't need to zero-initialize whatever views you're using you might also want to do Kokkos::View<...> v(Kokkos::view_alloc(Kokkos::WithoutInitializing, "v"), ...) which skips the initialization (so the backend isn't involved, it only does allocation)
parsa has quit [Quit: Free ZNC ~ Powered by LunarBNC: https://LunarBNC.net]
parsa has joined #ste||ar
diehlpk_work has quit [Quit: Leaving]
diehlpk_work has joined #ste||ar
nanmiao has joined #ste||ar
nanmiao has quit [Quit: Connection closed]
nanmiao has joined #ste||ar
Vir is now known as Vir_migrated_to_
Vir_migrated_to_ is now known as Vir_libera_chat
<pedro_barbosa[m]> ty
<gnikunj[m]> ms: how is view's default instance construction leading to segmentation fault?
<ms[m]> gnikunj: not the view's default construction
<ms[m]> all views are zero initialized by default in parallel using the corresponding backend
<ms[m]> and since the view constructor doesn't take an execution space instance, it uses the default instance
<ms[m]> and since the instances are not threadsafe, constructing those views inside tasks is dangerous
<gnikunj[m]> aha
<gnikunj[m]> got it
<gonidelis[m]> hkaiser: this guy get's this simple bench slower by parallelizing, simply because.... it's simple
<gonidelis[m]> and at the end he says: "Well maybe HPX async could do your job"
<gonidelis[m]> :DDDDD
<hkaiser> nod
parsa has quit [Quit: Free ZNC ~ Powered by LunarBNC: https://LunarBNC.net]
parsa has joined #ste||ar