[Arfrever] has quit [Killed (NickServ (GHOST command used by [Arfreve1]))]
[Arfrever] has joined #pypy
reneeontheweb has joined #pypy
Atque has joined #pypy
f4at has joined #pypy
f4at has quit [Quit: bye]
Atque has quit [Quit: ...]
reneeontheweb has quit [Quit: Client closed]
cfbolz_ has joined #pypy
idnar_ has joined #pypy
Techcable_ has joined #pypy
fijal_ has joined #pypy
EWDurbin_ has joined #pypy
alicetries has joined #pypy
alice has quit [Killed (tungsten.libera.chat (Nickname regained by services))]
alicetries is now known as alice
Lightsword_ has joined #pypy
Techcable has quit [*.net *.split]
cfbolz has quit [*.net *.split]
Lightsword has quit [*.net *.split]
EWDurbin has quit [*.net *.split]
fijal has quit [*.net *.split]
idnar has quit [*.net *.split]
cfbolz_ is now known as cfbolz
EWDurbin_ is now known as EWDurbin
Lightsword_ is now known as Lightsword
fijal_ is now known as fijal
Techcable_ is now known as Techcable
idnar_ is now known as idnar
greedom has joined #pypy
greedom has quit [Remote host closed the connection]
lritter has joined #pypy
greedom has joined #pypy
f4at has joined #pypy
greedom has quit [Remote host closed the connection]
<ronan>
I've been looking at the pyston benchmark (discussed at https://github.com/faster-cpython/ideas/discussions/375 ). The (somewhat) good news is that we're eventually faster than CPython and almost as fast as Pyston, the bad news is that it takes ~15000 iterations.
<cfbolz>
ronan: oh, wow
<cfbolz>
ronan: I had looked at it for like five minutes during the weekend, sqlalchemy is really quite tough for us in some places :-(
<cfbolz>
but it sounds like a good goal to work on maybe?
Corbin has quit [Remote host closed the connection]
Corbin has joined #pypy
<f4at>
I'm playing with PyPy. I know almost nothing :'(
<f4at>
Is It possible to run pypy with jit without recompiling?
<cfbolz>
f4at: hey, welcome!
<f4at>
hello :)
<cfbolz>
f4at: anything that interests you in particular?
<cfbolz>
the JIT can unfortunately not be tested without recompiling, only behaviour can
<f4at>
Well I want to play with the jit. But I don't think I'm qualified yet. So, I thought about optimising some features Like factorials and decimals because they are pretty slow. And the reason telco and bm_mdp and slower in pypy
<f4at>
are*
<ronan>
cfbolz: I haven't looked much at the server-side, I was surprised to see that the client side seems to take a significant amount of time and also warms up very slowly
greedom has joined #pypy
<cfbolz>
f4at: it's possible that the fractions thing is not jit related
<f4at>
Oh I think its pure python
<cfbolz>
f4at: yes, but it uses huge integers
<f4at>
cfbolz: PyPy is pretty fast (big fun) It loses to CPython only when the benchmark uses features that are not well optimised in PyPy. I was suprised when I found out that filter is x30 is slower than list comprehension.
<f4at>
cfboltz: any ideas why sort with key is still slower than Cpython?
<f4at>
cfboltz: I copied fractions source code from Cpython to PyPy. PyPy is now faster than CPython ¯\_( ͡❛ ͜ʖ ͡❛)_/¯
<cfbolz>
f4at: oh what!
<f4at>
:)
<cfbolz>
f4at: so diff it and see what the difference it
<f4at>
cfbolz: okay sure
<cfbolz>
f4at: is the python file from a newer python version
<cfbolz>
?
<f4at>
cfbolz: yes, Python 3.10
<f4at>
I have older versions installed I will test them
<f4at>
Well, didn't work with Python 3.7.
<f4at>
cfbolz: I guess, Fractions is not the reason the benchmark was slow after all
<cfbolz>
f4at: what do you mean "didn't work"?
<cfbolz>
to me it sounds like there maybe was an optimization in the fractions.py file in 3.10 that we could steal
<cfbolz>
f4at: so, sounds like a worthy backport to pypy I would say
<cfbolz>
I mean we will probably pick it up when we start working on 3.10 (hopefully later this year)
<cfbolz>
but in the meantime, it looks like it's purely an optimization so we can steal it ;-)
<cfbolz>
f4at: so, good find
<f4at>
cfboltz: Thank you :), Actually the older version is faster in Cpython but the newer version is faster PyPY
<f4at>
cfbolz: THe older version is x1.6 faster in CPython, the newer version is x1.4 faster in Python. Atleast in my test case
<f4at>
faster in PyPy*
<cfbolz>
ok
<cfbolz>
f4at: I just sent you a private message, did you see that?
<f4at>
ye :)
<ronan>
hmm, I guess we should be a bit more willing to modify the stdlib. CPython often write it with a mindset of "every bytecode costs the same", which is quite wrong on CPython and super-wrong on PyPy
Dejan has quit [Quit: Leaving]
f4at has quit [Quit: bye]
greedom has quit [Remote host closed the connection]
f4at has joined #pypy
<f4at>
cfbolz: I think I found a solution for Deciaml slow performance.
<f4at>
There is 2 versions of the decimal library. 1 is pure python the other is built-in switching the built-in one with the pure python versions.
<f4at>
Telco benchmark is now 13 times faster than Cpython, it was 1.65 slower than CPython. More than x20 speed up.