Parallella Community

Two updates to this thread -

First, a programming quick start using STDCL can be found here:

Second, OpenMP support is being investigated using libcoprthr as a co-processor threading library for Epiphany, proof-of-concept code looks good so far as a basis for a project.

Truly amazing! So now you have implemented a production OpenCL and prototypes of OpenMP and and MPI for the Epiphany, all using the same COPRTHR code base.(and with a team that is a fraction of what most large organizations have).

I know there are some important folks interested in OpenMP, so I am anxious to see how this unfolds.

Cheers,
Andreas

Awesome!!!

Hi everybody,

I am just curious to know where I could find the OpenMP SDK, because I was told that it is available.

On the other hand, how is or would be OpenMP implemented on the Epiphany? OpenMP is designed for SMP's and as far as I know the memory is splitted on two parts, one for the arm (host) and the other for the epiphany (device). Is the data copied between host and device implicitly using mailboxes?

If there is an SDK that already supports Epiphany I am not aware of it. We are working on something that would support OpenMP for the co-processor and allow explicit distributed memory management using support that already exists for malloc'ing device-shareable memory in the APIs provided in the COPRTHR SDK. It would be easy to implement automatic host-device data copy instrumentation but this will only work for toy use cases. Most applications will quickly require more control anyway. See for example the Intel pragma offload API for MIC, which quickly requires qualifying empty scopes to force data transfers for efficiency. There is an experimental feature of Parallella that would make automatic memory management more efficient, but its experimental at the moment. It relates to whether Parallella is an SMP platform or not - the issue is complicated, it does not have a clear answer at the moment.

Hi all,

How is the OpenMP implementation going? I've just received my parallella (finally arriving to the Old Continent) and I was hoping to do a matrix-matrix multiplication benchmark but going through the examples I thought there was not an OpenMP approach to use the Ephiphany coprocessor. Am I right?
My modest opinion is that it would be very important to access the coprocessor using OpenMP and MPI. That would definitely put parallella among other emergent low consumption options in HPC.

Hope is okay to share my ideas...

Regards,
Luiceur

Unfortunately that work has been pushed aside by other projects. I do agree that OpenMP and MPI both would provide excellent programming models for Epiphany, and the Parallella platform in particular since it effectively has shared memory between ARM and Epiphany. That is why we explored both of these through proof-of-concept efforts. This update only covers what I know from BDT. You might check of Adapteva is working on support for either or both internally.
-DAR

Hi, so can Open MP programs be run on the Epiphany cores or are you guys still working on it?

Thanks