Parallella Community

I am trying to get basic OpenCL programs to run on the Epiphany processor, but when I run them as a the regular linaro user, I get mmap failure during the compilation. If I instead run them as root user, the Epiphany accelerator is not found.

Any answer can help solve this is highly appreciated.

Thanks for the fast reply. I've sent an email to the given adress.

guen:

I have seen this behavior when using R to run openCL code on the Epiphany. My advice

1) Always make sure you run your code as root
2) Somewhere in your code, at the beginning when it tries to determine what devices are available, your code will most likely query for CL_DEVICE_TYPE_GPU type devices. What I have found is that the epiphany does NOT identify itself as a GPU, so it is not detected by the query. However, the Epiphany will be detected if your code does instead query for CL_DEVICE_TYPE_ALL devices.

If you query for CL_DEVICE_TYPE_ALL you should find two devices, the ARM CPU and Epiphany. If you just want Epiphany, you should query CL_DEVICE_TYPE_ACCELERATOR. That is how its supposed to work since we classify Epiphany as an accelerator.

Ok. Thanks. I can see the epiphany device support is not being loaded. Let me try to recreate this locally to figure out why. On the platform I just checked where this build was created there is no issue.

This has nothing to do with ocl.conf, I was just suggesting to create one. For our software we have deprecated the ICD enumeration system because its not a very good design, causes problems, and is very limited in features. You may not have a platform where this is relevant, but if you were to have several parallella boards networked together, you can export the the Epiphany coprocessors so they could all be use from one host program, as an example. The ocl.conf file allows you to just add CLRPC servers to support this. There are other features in future that will use ocl.conf as the primary means of configuring your system.