Re: [Starpu-devel] data_unregister_submit() does not flush data from GPU to CPU

[Aleksandr Mikhalev <aleksandr.mikhalev@kaust.edu.sa>]

Thank you very much for fast answer!

So, I misunderstood meaning of "*_submit()” family of functions. At first, I 
thought it adds asynchronous task, where the handle is flushed to its origin 
asynchronously: nobody knows when it will happen, but it will happen for 
sure. I input data on CPU, create task to update this data on GPU and do not 
do anything else, just wait for all tasks to finish. Resulting data on GPU is 
what I need to display. If I use starpu_data_unregister(), data is brought 
back from GPU to CPU and handle is destroyed, but if I use 
starpu_data_unregister_submit(), handle is destroyed without copying at all, 
even if I wait for all tasks to finish. From my point of view, it is 
counterintuitive, that synchronous call copies data and asynchronous call 
does not copy data. Can you confirm, that this difference is intended?

Proposed solution with data_acquire is much more advanced, adding explicit 
empty task to synchronise data is easier.

Best regards,
Aleksandr.

> On 24 Jul 2018, at 18:39, Samuel Thibault <samuel.thibault@inria.fr> wrote:
> 
> Hello,
> 
> Aleksandr Mikhalev, le mar. 24 juil. 2018 16:18:36 +0300, a ecrit:
>> If I understood correctly, idea of “submit” functions is to hint StarPU to 
>> flush data from GPU to CPU,
> 
> No, "submit" means that it is asynchronous, i.e. it retuns immediately
> without waiting for the result.
> 
>> StarPU does not copy output data from GPU to CPU if corresponding handle 
>> is destroyed by starpu_data_unregister_submit().
> 
> Well, yes, and that is what is documented.
> 
> I don't actually really see how it would be useful to copy output
> back from GPU to CPU: since it is asynchronous, you have no way of
> knowing when that will be effective. If you want to bring back data
> to the CPU, either use starpu_data_acquire(STARPU_R) (synchronous) or
> starpu_data_acquire_cb(STARPU_R) (asynchronous but with a callback
> called when the data up to date on the CPU) and starpu_data_release()
> 
>> The same holds true for starpu_invalidate_data_submit().
> 
> Well, yes, invalidate_data_submit() is explicitly exactly meant to
> discard the value of a data.
> 
> Samuel


-- 

This message and its contents, including attachments are intended solely 
for the original recipient. If you are not the intended recipient or have 
received this message in error, please notify me immediately and delete 
this message from your computer system. Any unauthorized use or 
distribution is prohibited. Please consider the environment before printing 
this email.