CAS in f2py
Ondrej Certik
quoting from issues:
> Code generation has not high priority at the moment but
> this is certainly something that I would need in projects
> that will use python based CA, e.g. in g3f2py.
What do you need CAS for in f2py? For simplifying fortran expressions?
Or evaluating some numerical things (mpmath)?
Ondrej
Pearu Peterson
1) to hold expressions of array dimensions.
2) to do simple manipulations with array dimensions for dealing with
the fact that in C array indices start from 0 but in Fortran array
indices start from 1 by default (or from any integer if so specified).
3) to convert Fortran expressions to C expressions.
The current f2py in numpy has very trivial support for the
above but it is not always sufficient.
Pearu
Ondrej Certik
>
>
>
>
> On Dec 8, 9:33 pm, "Ondrej Certik" <ond...@certik.cz> wrote:
>
> > > Code generation has not high priority at the moment but
> > > this is certainly something that I would need in projects
> > > that will use python based CA, e.g. in g3f2py.
> >
> > What do you need CAS for in f2py? For simplifying fortran expressions?
> > Or evaluating some numerical things (mpmath)?
>
> CAS in f2py is mainly needed for the following:
>
> 1) to hold expressions of array dimensions.
> 2) to do simple manipulations with array dimensions for dealing with
> the fact that in C array indices start from 0 but in Fortran array
> indices start from 1 by default (or from any integer if so specified).
I see - like n**2, or (n+1)**2. Those are simple things, possible with
the current sympycore/sympy.
I think real application is the code Kent sent - and that is still
kind of slow in sympy. I am curious how it performs in sympycore.
Ondrej
Pearu Peterson
> I think real application is the code Kent sent
ok ok, you probably meant that f2py would use only a
very small subset of sympy/sympycore and that is true.
> - and that is still kind of slow in sympy. I am curious how it performs in sympycore.
recent sympy in hg (with sympy the test runs 6.04secs,
with sympycore it takes 1.58secs in my laptop).
Btw, sympycore does not use caching almost at all,
which is an open possibility to improve sympycore speed even more
and is left to future when sympycore has enough unittests.
Pearu
Ondrej Certik
>
> On Dec 8, 10:10 pm, "Ondrej Certik" <ond...@certik.cz> wrote:
>
> > I think real application is the code Kent sent
>
> :) do you think that f2py is an academic application;)
f2py is not an academic application at all. But using sympy(core) in
f2py for matrix dimensions is an academic application of sympy(core).
:)
> ok ok, you probably meant that f2py would use only a
> very small subset of sympy/sympycore and that is true.
yep.
> > - and that is still kind of slow in sympy. I am curious how it performs in sympycore.
>
> The Kent test is about 3.85 times faster with sympycore compared to
> recent sympy in hg (with sympy the test runs 6.04secs,
> with sympycore it takes 1.58secs in my laptop).
>
> Btw, sympycore does not use caching almost at all,
> which is an open possibility to improve sympycore speed even more
> and is left to future when sympycore has enough unittests.
That is interesting. I think the biggest speedup you made in the new
core (now in sympy) was to use __new__ instead of __init__. The
caching improves things too.
I'll try now to port the dict based add and mul to sympy, this should
provide the major speedup, right? The next thing is the "is_Mul", I
know that they are fast, but I don't like the fact, that it pollutes
the namespace when tab-completing in ipython. Maybe it could be cured
somehow.
Ondrej