osdir.com


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

How to make Python run as fast (or faster) than Julia


On Thu, Mar 08, 2018 at 08:44:16AM +1100, Chris Angelico wrote:
> On Thu, Mar 8, 2018 at 8:36 AM, Python <python at bladeshadow.org> wrote:
> > On Mon, Mar 05, 2018 at 04:09:48PM -0800, Dan Stromberg wrote:
> >> On Mon, Mar 5, 2018 at 3:53 PM, Python <python at bladeshadow.org> wrote:
> >> > On Sat, Mar 03, 2018 at 08:18:03AM +1100, Chris Angelico wrote:
> >> >> > Python is often a preferred solution because it is often fantastic for
> >> >> > rapid implementation and maintainability.  The GIL's interference
> >> >> > with threaded code performance has, for me at least, on several
> >> >> > occasions been...  disappointing (perf costs of removing it aside)
> >> >> > because it gets in the way of choosing Python for such solutions.
> >> >> > Jython and IronPython are simply not feasible options for me, for
> >> >> > multiple reasons that have zero to do with their technical
> >> >> > suitability.
> >> >>
> >> >> Have you actually tried it and run into problems,
> >> >
> >> > Yes.  It was years ago and I forget the details, but I even posted
> >> > some sample code here and was told (quite possibly by you) that it was
> >> > the GIL that was eating my lunch.  Someone suggested writing the bits
> >> > I wanted to thread as a C extension, which largely defeated the
> >> > purpose of using Python.  In at least one case I just used C++, and in
> >> > another I just ignored the problem until it went away.
> >>
> >> So how about a little Cython?
> >
> > Nope.  As a practical matter, the only Python option I have is
> > Python proper (i.e. CPython).
> 
> So you can use C++, but only if you write the entire program in it,
> instead of using Cython and only transforming the part that's actually
> slow?

That describes the situation closely enough to correctly that I don't
feel the need to elaborate, yes.

I will say that /in theory/ I can do whatever I want; in practice
there are significant barriers to integrate new tools into the build
system I'm required to use that make it prohibitive to do so.  There
are complexities there which I don't care to describe, but the point
remains, effectively, the only Python option I have available is
CPython.