On Mon, 23 Apr 2007 15:13:36 -0400
"Terry Reedy" <[EMAIL PROTECTED]> wrote:

> 
> | Naturally PyPy wouldn't be able to use its magical optimization abilities 
> on those bits
> 
> Many to most of 'those bits' have been heavily optimized already, which is 
> why to use them.
> 
> Terry Jan Reedy

Yes, but how about this idea: that we would like to take vector expressions,
such as the sum of three arrays, X+Y+Z, and instead of creating a temporary 
array
to house X+Y (and then summing with Z), we sum the elements from all three 
arrays at once.

This kind of calculation is very typical in numpy/R, where temporary arrays are 
created, transversed
then destroyed alot, and the CPU cache overheats, and people become generally 
unhappy.

Last year I had a look at teaching the rtyper to understand numpy expressions 
(see pypy.rpython.numpy),
with the idea of heading in this more intelligent direction. This year i 
discovered the pypy JIT and
had a brain orgasm; it would be very interesting to use (+develop) the JIT 
towards these cache friendly
ideas also.

It really makes no sense to reimplement R, i think, UNLESS you can interface 
with the present R libraries.
The big strength of R is it's many many libraries for statistics etc. I don't 
think it's much of an
interesting language in and of itself.

Simon.
_______________________________________________
[email protected]
http://codespeak.net/mailman/listinfo/pypy-dev

Reply via email to