Re: [Starpu-devel] Problem running CUDA task

[Miguel Palhas <mpalhas@gmail.com>]

This is an old question so i'll have to check if i'm thinking this right, but i believe the problem was due to the machine i use having 2 GPU's, one of which only supports an older computational capability. half the times, the kernel was being sent to the secondary GPU which didn't support the compiled kernel correctly, and probably issued an error.

setting the STARPU_NCUDA=1 environment variable solved the problem

Anyway, apologies for not answering here when i found the solution (i forgot about this mail in the meantime) and thanks for the answer anyway

Best regards

Miguel Palhas

On Wed, Jun 5, 2013 at 3:09 PM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Hello,

It seems you never got any answer. Here is some feedback.

Miguel Palhas, le Mon 06 May 2013 00:43:39 +0200, a écrit :

> If required, i can provide the full source code i used

It would probably help, since otherwise we can't really try to reproduce
your issue. The few bits you have pasted seem completely alright.

> Also, another question: I tried, in order to fix this, to place a
> starpu_data_acquire(vector_handle, STARPU_R) right before the
> starpu_data_unregister call, but then the program hangs completely at that
> point. What is going on here?

Again, without seeing the rest of the program it's hard to say.

Miguel Palhas, le Mon 06 May 2013 13:47:44 +0200, a écrit :

> An update. It seems the problem is with the kernel itself not being called.
>

>   printf("i'm in the proxy function %d %d\n", nblocks, threads_per_block);
>   cudaPrintfInit();
>   cuda_kernel_impl<<<nblocks, threads_per_block, 0,
> starpu_cuda_get_local_stream()>>>(val, n, *factor);
>   cudaStreamSynchronize(starpu_cuda_get_local_stream());

You should probably also test the value returned by
cudaStreamSynchronize, to make sure CUDA didn't report an error.

> function, like shown in my first email. They match the output of the cuPrintf
> function during the kernel
> Sometimes, however, i get this:
>
> i'm in the proxy function 1 64
> 0.000000

Which indeed looks like something does not get called. You could put
another printf after the cuda_kernel_impl() call, to make sure that the
piece of host code at least has been executed.

Samuel

--
Cumprimentos

Miguel Palhas