Re: GSOC 2020 Idea Discussion

[Benson Muite]

On Tue, Feb 25, 2020, at 12:25 PM, Carlo De Falco wrote:

Il giorno 25 feb 2020, alle ore 08:56, Benson Muite <address@hidden> ha scritto:

Backends may change and are architecture specific, some backends are:

1) MKLDNN https://github.com/intel/mkl-dnn

2) MLOpen https://rocm.github.io/miopen.html

3) CuDNN https://developer.nvidia.com/cudnn

4) Arm NN https://www.arm.com/products/silicon-ip-cpu/machine-learning/arm-nn

Maybe challenging to support all of these.

Might it be possible to import and export models in ONNX (https://onnx.ai/) format? In cases where one needs very high performance in the training phase, one can use an optimized library. Are developers of the libraries you are considering to integrate supportive?

When considering building an interface between Octave and an external library please also take into account possible license incompatibility issues.

I am for example not sure whether it is legally possible to distribute software that links together Octave and CUDA.

When preparing a project proposal a student should carefully look into this sort of issues, we have had in the past GSoC students that produced good 

looking code that we were never able to actually use because of licensing and copyright issues.

c.

Good point. However, there are some workarounds, for example one can use Cublas in Octave by linking an installed library:

https://devblogs.nvidia.com/drop-in-acceleration-gnu-octave/

There should of course be a default possibly less performant and portable option that does not need a separate installation due to licensing issues.