Arrayfire on Multiple GPUs and Backends

[Logan Johnson]

Hi,

I'm interested in using Arrayfire for some of our applications. I'm very impressed with how quickly I've been able to prototype some algorithms within the framework. Two important questions:

1. Multi-GPU support - looks like this is an iffy subject for Arrayfire (https://github.com/arrayfire/arrayfire/issues/907). Just double checking that this is the latest on multi-gpu support.

2. Multiple backends - Is it possible to use both CUDA and OpenCL platforms within the same application? Documentation leads me to believe that AF will throw an exception if I try to do that, but I want to ask the experts if it's possible. Our existing code base uses both due to uncertainty around NVidia's commitment to OpenCL.

[Pavan Yalamanchili]

Hi logan,

1) You can use multiple GPUs using arrayfire. However you will need to use it in a single threaded mode. Considering that most operations in ArrayFire are asynchronous this should still provide a way to run GPUs concurrently.

2) You can use multiple platforms in the same application for task parallelism. You can however not combine the arrays from one backend with the other. This will give exceptions.

We plan to address both issues in a better way in the future by making our backends thread safe and having some way interop arrays between various backends.

--

Pavan

--
You received this message because you are subscribed to the Google Groups "ArrayFire Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to arrayfire-use...@googlegroups.com.
To post to this group, send email to arrayfi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/arrayfire-users/a020b001-ed21-4c74-8405-247ca66ec8d3%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.