Re: GPUs and CUDA

[Gene <gene.ressler@gmail.com>]

On Jul 12, 6:26 am, "ldries46" <bertus.dr...@planet.nl> wrote:
> |"Rego, P."  schreef in berichtnews:783efc18-a581-4ae2-bde4-04afa3568ff8@glegroupsg2000goo.googlegroups.com...
> |
> |>This would allow something like
> |> Pragma CUDA( Ralph );
> |>to indicate that the subprogram "Ralph" is to be run on the GPUs; and
> |>it would also allow non-CUDA enabled compilers to compile the source
> |>regardless of whether or not their compiler actually knows about CUDA.
> |
> |That's the idea. It would be very good.
>
> At this moment I don't know if this is a good idea, because CUDA is a
> language developped only for NVidia GPU's. One of the purposes of Ada I
> think is that a program can always be used on another computer with only the
> a simple recompilation. I would like a Pragma GPU( Ralph ) that can check
> which GPU is used and if that GPU has a CUDA like environment available.
>
> L. Dries

Indeed, GPUs are highly specialized processors that excel mostly at
certain kinds of floating point computations with limited branching
behaviors.  Compiler tricks to find good mappings for general code
onto GPUs are very much a current topic of research.  Nvidia's CUDA
implementation relies on language extensions to C and Fortran buit
into GCC.  I don't think it would be useful or even possible to map
GPU-style threads into Ada tasks.  So getting access to CUDA
primitives from Ada looks like a tough problem, certainly more than
adding a Pragma and tweaking the GNAT code generator.

On the other hand, there are many numerical libraries already
rewritten and tuned to exploit CUDA.  Ada bindings for these would be
very useful and straingthforwrd to build.