What would be really nice would be for the matrix multiplication phrase in
J (+/ . *), the matrix inverse phrase (%.), and other J matrix operations,
to detect the presence of a GPU and use "special code" to accelerate these
matrix operations when a GPU is present. J code running with a GPU
shouldn't look any different that J code running on a scalar machine. It
should just run a LOT faster.

Skip

Skip Cave
Cave Consulting LLC


On Sat, Aug 30, 2014 at 12:14 AM, Raul Miller <[email protected]> wrote:

> This isn't really worth a pull request, but double->float is 1&(3!:5)
> and float->double is _1&(3!:5)
>
> This is documented at
> http://www.jsoftware.com/help/dictionary/dx003.htm. (Note that floats
> are represented as a sequence of literals, because J can't work with
> them.)
>
>    $1&(3!:5) o.1 2
> 4
>    $1&(3!:5) o.1 2
> 8
>    a.i.1&(3!:5) o.1
> 219 15 73 64
>
> Thanks,
>
> --
> Raul
>
>
> On Sat, Aug 30, 2014 at 1:04 AM, Scott Locklin <[email protected]> wrote:
> > So, it took me all of 20 minutes to pull dgemm into J for a matrix
> > multiplication speedup.
> > I stuck it here, along with an org-emacs TODO list for making this
> actually
> > happen.
> > It's all "busy work" as far as I can tell, though it would be my first
> time
> > writing code that links to CUDA.
> >
> > Either way, the dgemm wrapper should eventually make its way into the API
> > stuff, as it's a pretty good speedup over +/ .* for bigger array problems
> >
> > https://github.com/locklin/jCUDA
> >
> > Feel free to pitch in on the busy work if anyone has problems that would
> > benefit from this.
> >
> > -Scott
> > ----------------------------------------------------------------------
> > For information about J forums see http://www.jsoftware.com/forums.htm
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
>
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to