IRC channel logs


back to list of logs

<sofi>Hello! :)
<sofi>Hi there! Is there any good advice for how to tune a *BLAS implementation in a reproductible enviroment/way without having to manually tweak it? Our current implementation is utilizing ATLAS quite heavily, however it self-tunes itself at build time so i don't think its a good fit for a Guix enviroment.
<sofi>To specify a bit, this is for a quite unique `cortex-a76.cortex-a55` rk3588 aarch64 chipset.
<efraim>I assume openblas supports that, but you can make the shared library smaller by compiling a variant to only support that chipset
<efraim>* supports that without needing to recompile openblas
<efraim> something like this might work
<efraim>sofi ^
<sofi>Oh that is an interesting link, I will definetly try to change that up to fit this core.
<sofi>Also, maybe a bit of a dumb question, but would running on something like Debian cause issues for Guix HPC performance?
<efraim>Not specifically, everything in Guix is supposed to be linked to only other packages in Guix, with anything dlopen'ed getting the full path to the necessary library.
<efraim>it might if there's a lot of background processes, but that's true for anything
<sofi>Yeah, I will also try to figure out what black magic ATLAS is doing to beat out our last *BLAS benchmarking. As its doing something strange to get like a 4 fold improvement to the previous method we used.
<sofi>Annoyingly, I do not remember what the previous method was.