Henry Rich <[email protected]> wrote:
> You would also have agonize over the fact that in J,
>    0 * _
> 0
> which is probably not what BLAS does.

In most cases it actually IS what BLAS does -- but I do agree that
agonizing is necessary. Careful unit testing is needed.

Here's an instructive example of agonizing:
http://www.r-bloggers.com/slowing-down-matrix-multiplication-in-r/

> And anyway, if you did modify cip.c, how would you distribute the fruits of
> your labor?

It's GPLed -- he'd probably post it on one of the public version
control repos and ask if jsoftware wanted to pull the changes into j8.

Raul said:
>  So if BLAS treats it differently, my idea of the correct solution
> would be to modify BLAS.

If you want the benefit of BLAS performance, you're stuck with BLAS'
optimizations.

> Henry Rich

-Wm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to