<phlebas>
qunaibit has been rendering some graphs using matplotlib vs matplotlib hpy on graalpython, with promosing results (~10x faster), even though the interface to numpy means that we're still running in "mixed mode". would anyone from the pypy side have time to look at qunaibit's PRs and/or update pypy's hpy with those changes for comparative measurements?
mattip has quit [Ping timeout: 240 seconds]
mattip has joined #hpy
<antocuni>
wow, that's impressive! It's worth a blog post
<antocuni>
I can't promise to do anything in the next few days because I'm busy with all the paperwork/preparation for my relocation, but hopefully I'll have some time next week
<mattip>
I see three PRs in hpy by quna1bit, or are we talking about some other PR?
<phlebas>
yes, those PRs. there's also the migration for matplotlib that we still have to push out (but mohaned just went to bed, so it'll have to wait :)
<Hodgestar>
phlebas: How does the C++ support fit into the matplotlib picture?
<Hodgestar>
Also, woot!
<phlebas>
Hodgestar: mohaned knows the details, but iirc the main issue was just fixing to allow compilation when used in C++ code
<phlebas>
i just saw i forgot to mention: cpython regresses ~5% on that same benchmark when run with the hpy port. that's something we still need to look at also
<phlebas>
(one iteration on cpython: ~19s with hpy, just under 18s with legacy c api)
<antocuni>
5% of slowdown when targeting the CPython ABI? That's indeed a bit surprising and we should look at it
<antocuni>
what is the performance if you target the hpy universal ABI in CPython?
<ronan>
phlebas: Nice!
<ronan>
I'll have a look at the PRs
krono has quit [Read error: Connection reset by peer]