Add Blas3 to the sources

[amueller]

@fabianp could you please give me a hand in that?

[fabianp]

I've been trying to avoid this since it's a LOT of code with lots of special cases (much much more than level1-2) and because level3 is O(n^3) operations probably inlining in cython is not so important and the python functions returned by get_blas_funcs would do the job.

What's the particular usercase? Is the function missing from get_blas_funcs (in that case I would gladly help you with f2py) ?

[amueller]

There are two use-cases: the speed-up of euclidean distances by @vene #1006 and neural networks.
I profiled the neural network and with the current batch-size, there is significant overhead.
Even without batches, there is a vector x vector (i.e. rank 1) matrix multiplication. That is also blas3, right?

in case of Vlad's PR, the compilation worked, but there was an linker error on my box.

[fabianp]

If you are speaking about ger that's level 2.

[fabianp]

Anyway sure I could help with particular problems but adding the full level3 specification is IMHO a bit overkill.

[fabianp]

Also, if you want to team up with the scipy guys (I talked with Ralph Gommers at EuroScipy) they where open-minded to expose a C API for BLAS functions (via Cython) , which frankly would be of great benefit for us.

[amueller]

Yeah, exposing blas via Cython is really what we want. We would have to maintain a backport but that would be better than doing it ourselves.

Yes, I wanted ger. Having the same for batches would be gemm, right? So this would be what I want....

[amueller]

Not sure what @vene wants, though.

[fabianp]

Adding gemm seems reasonable to me. I see in @vene's PR references to dgemm so I suppose the goal is to add that to the cblas directory, right ?

[amueller]

yes, that would be great :)

[fabianp]

Done, see #1313