Re: [Starpu-devel] About StarPU

[Jeff DiMarco <jdimarco218@gmail.com>]

Thanks for the help, I can reference A from within the cuda kernel.  Why am I able to reference it from the cpu_func as opposed to cuda_func that calls the kernel?  Does GET_PTR just work that way then?

Thanks so much,
Jeff

On Mon, Oct 22, 2012 at 7:25 PM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Sylvain HENRY, le Tue 23 Oct 2012 01:22:02 +0200, a écrit :

> I think you should use STARPU_VECTOR_GET_DEV_HANDLE instead of
> STARPU_MATRIX_GET_PTR.

Well, these are the same for CUDA, actually. GET_PTR is actually
preferred in this case, because it's really supposed to be a (CUDA)
pointer.

Jeff DiMarco wrote:
>     float * A = (float *)STARPU_MATRIX_GET_PTR(buffers[0]);
>
>     Doing this then referencing A[0] produces a Segmentation fault.  Am I doing
>     something blatantly wrong that I cannot see?

When executing the CUDA codelet, the pointer returned by GET_PTR is a
*CUDA* pointer, so you can not dereference it from the host.

Samuel