future multi-threading for-loops
nagle at animats.com
Tue Feb 5 19:11:53 CET 2008
Christian Heimes wrote:
> castironpi at gmail.com wrote:
>> Multi-threaded control flow is a worthwhile priority.
> It is? That's totally new to me. Given the fact that threads don't scale
> I highly doubt your claim, too.
There's plenty that can be done to automatically extract parallelism
from programs, but given the architecture of CPython, with the "global
interpreter lock" and the inability to detect dynamism at compile time,
it's probably not going to be feasible with that implementation.
I've been to a recent talk at Stanford (in EE380) in which
someone was describing an optimizing compiler for Matlab intended
to generate production DSP code. I've heard two other talks
on how to automatically parallelize loops for execution in a GPU.
It's quite possible to do extreme optimizations like that.
But those are things you do after solving the problems of being
10x-30x slower than C.
The real optimization trick for Python is figuring out at compile time what
might change at run time, and what won't be. Then all the things that
can't change can be hard-bound during compilation. Shed Skin does
some of that.
More information about the Python-list