Re: GSoC 2017 project -- Speeding up tensor contraction in Octave

[John W. Eaton]

On 03/02/2017 09:52 AM, Susi Lehtola wrote:
> On 02/28/2017 12:58 AM, Somesh Singh wrote:
> > Hi All
> >
> > Tensor contraction is a widely used operation that comes up quite often
> > when performing operations on tensors. It is widely used in many domains
> > such as analysis of chemical compounds. There are a lot of ongoing
> > efforts to speed-up tensor contraction.
> > Many of the softwares for numerical computation have sound
> > implementations of tensor contraction. In this light, I believe it will
> > be a good idea to have a fast implementation of tensor contraction in
> > Octave for muti-core systems, since tensor contraction offers a lot of
> > scope for parallelism. I would like to propose this as a project for
> > GSoC 2017 and get involved with it.
> > Please share your thoughts on the suggestion.
>
> What would be the scope of this? Tensor contractions are typically
> handled using BLAS calls by permuting the contracted indices to either
> the start or the end, and then calling dgemv or dgemm to get the result.
> These would be AFAIK pretty straightforward to loop-parallellize.
>
> There are recent software developments that are able to do the tensor
> contraction without the permutations, but these basically involve novel
> implementations of BLAS...

I agree.  To me, it doesn't seem like something that should be specific 
to Octave.  It sounds like something that should be developed as a 
fairly low-level library (like the BLAS or LAPACK libraries) and then we 
would just use these functions in Octave.

jwe