matlisp-users

[Matlisp-users] m.*! not native?

[Jefferson P. <jp...@cs...>]

Hi,

I was profiling some of my code yesterday, and noticed that a single m.*!
call was taking more than 90% of the run time for my program.  After
poking around a little in mtimes.lisp, it seems that m.*! isn't native.
Is this true?

I don't know BLAS very well.  Is there no way to do element-wise
multiplication in BLAS?  The non-native routine is a couple of orders of
magnitude slower than native routines of the same complexity: (e.g. m.+!
or outer-product of vectors).

If BLAS can't do this, it seems like it would be an easy Fortran routine
to write.  I could do it, though I'm not familiar with how to do foreign
functions in Allegro or cmucl.

Jeff

Re: [Matlisp-users] m.*! not native?

[Tunc S. <si...@ee...>]

what do you mean by native.  both allegro and cmucl produce native code.
looking at mtimes.lisp; it looks like all the declarations are in place
for m.*! so the only reason that it may be slow is if the optimization
flags were not set during compilation.  I personally never checked that
they are.  It is worth checking this.

Otherwise, a single loop in lisp with full declarations should not be
orders of magnitude slower than fortran.

 -- Tunc

Jefferson Provost wrote:
> 
> Hi,
> 
> I was profiling some of my code yesterday, and noticed that a single m.*!
> call was taking more than 90% of the run time for my program.  After
> poking around a little in mtimes.lisp, it seems that m.*! isn't native.
> Is this true?
> 
> I don't know BLAS very well.  Is there no way to do element-wise
> multiplication in BLAS?  The non-native routine is a couple of orders of
> magnitude slower than native routines of the same complexity: (e.g. m.+!
> or outer-product of vectors).
> 
> If BLAS can't do this, it seems like it would be an easy Fortran routine
> to write.  I could do it, though I'm not familiar with how to do foreign
> functions in Allegro or cmucl.
> 
> Jeff
> 
> _______________________________________________
> Matlisp-users mailing list
> Mat...@li...
> https://lists.sourceforge.net/lists/listinfo/matlisp-users