_0az3 has quit [*.net *.split]
_0az3 has joined #hpy
ronan has joined #hpy
<mattip> the author of piconumpy has some new benchmark results for HPy in universal and cpy abi modes on CPython and PyPy (and others)
<mattip> the timings for PyPy and CPython in the universal mode are similar (cpython is a bit faster), which is a little disappointing
<mattip> since it would be nice if PyPy were faster than CPython
<antocuni> mattip: actually, being as fast as CPython on such a benchmark is not too bad
<antocuni> ideally, I would like to arrive at a point in which PyPy is "never slower, often faster" than CPython
<antocuni> for the specific case of piconumpy, I and arigato investigated a bit the performance. I summarized our findings here: https://foss.heptapod.net/pypy/pypy/-/issues/3542
<antocuni> I think there is a lot to win here, but it's not an easy fix
<antocuni> I even tried to trick arigato to do it but without much success :)
<Hodgestar> Nice write up!
<Hodgestar> I guess the real comparison for PyPy is "how much faster is HPy piconumpy than cpyext piconumpy on PyPy".
<Hodgestar> And it looks like PyPy + HPy is *much* (i.e. 4x) faster than PyPy + cpyext, which looks pretty good.
<antocuni> yes, but the final goal is to be competitive with CPython, so it makes a lot of sense to benchmark it
<Hodgestar> Agreed, both are valuable. We should just remember to be happy that alternative implementations can be 4x faster. :)