Re: [Starpu-devel] Data interfaces in StarPU

[Cyril Roelandt <cyril.roelandt@inria.fr>]

On 08/28/2012 04:10 PM, Albert Sidelnik wrote:
> Hi Cyril,
>
> I was under the impression that the copy/allocate registered routines (e.g.
> copy_ram_to_ram, allocate_variable_buffer_on_node) would also be invoked
> even for CPU->CPU tasks. If not, where/how does that occur?

If you run a task on a CPU, and then another task on another CPU, then 
the copy_ram_to_ram function will not be called, because both CPUs use 
the same memory node (the RAM).

You could take a look at starpu_data_cpy (src/util/starpu_data_cpy.c) to 
see how to trigger a RAM to RAM copy.

> What I meant by passing pointers around is the situation where I would like
> a codelet to propagate (or forward) data through a chain of codelets
> without having to allocate and copy the data when each codelet starts.
> Instead I would like to just pass on an address to the original data among
> the chain of codelets. I hope that makes sense, and if not, I can give you
> a better example what I am trying to do.

The thing is, if you need to use a given piece of data on different 
memory nodes (say, the RAM and the memory of your GPU) then you need to 
make a copy. How would you handle this case ?

StarPU does not allocate/copy memory every time a codelet starts. It 
just registers a piece of data (when the application calls 
starpu_variable_data_register(), for instance) and moves it between 
memory nodes when it is needed. So if you run the following tasks 
sequentially:

T1: on CPU0
T2: on CPU1
T3: on GPU0
T4: on CPU2

StarPU will only run 2 copies: one from the RAM to the GPU, and one from 
the GPU to the RAM.

WBR,
Cyril Roelandt.