Re: [Starpu-devel] CUDA task and GCC plugin

[ludovic.courtes@inria.fr (Ludovic Courtès)]

Hi,

Mehdi AMINI <mehdi.amini@silkan.com> skribis:

> First I wonder if I can declare a task with only a CUDA implementation
> and no CPU implementation ?

In theory, it should be possible.

> Currently I always end up with a linker error.

Oups, that’s a genuine bug (I’ve reproduced it.)  I’ll let you know when
a fix is available.

> I also wonder if there is a more verbose mode for StarPU ? Using
> --enable-debug and --enable-verbose juste give some initialization
> informations but I would also expect an output each time a task or a
> data transfer is scheduled.

For such detailed information, you would probably need to generate
traces with FxT (info "(starpu) Off-line").

> I also have an issue related to implicit cpu implementations and -O3
> gcc flag which seems to be broken. I'm testing a simple task call with
> a CPU and a CUDA implementation. Using -O3 the CPU task is always
> executed while without the -O3 flag the CUDA task is correctly
> executed. Is it a known issue ?

I believe there’s nothing wrong here: the default scheduler (‘eager’)
makes non-deterministic scheduling decisions, and just happened to favor
CPU over CUDA for some reason.

Now, the CPU back-end can be disabled at run-time, forcing the scheduler
to use GPUs only:

  $ STARPU_NCPUS=0 ./compile-test.sh

  [...]

  Run with implicit CPU task, compiled and link at the same time
  GPU !
  Run with implicit CPU task, compiled and link separately
  GPU !
  Run with explicit CPU task, compiled and link at the same time
  GPU !
  Run with explicit CPU task, compiled and link separately
  GPU !

(Using ‘--disable-cpu’ should have the same effect; not sure why it
didn’t.)

The manual has more info about schedulers (info "(starpu) Task
scheduling policy").

Hope this helps,
Ludo’.