perhaps integration should be even at JCublas level. http://www.jcuda.org/jcuda/jcublas/JCublas.html
David Hall's code in Breeze's `gust` is probably a big source for idea rip-offs :) for such integration. On Fri, Jul 18, 2014 at 11:42 AM, Dmitriy Lyubimov <[email protected]> wrote: > For committers who still have time/ are working on Mahout: > > Since we know that current performance is function of in-core performance, > how about throwing in Lapack-based and GPU based backs under in-core Matrix > api? > > Breeze has already moved to do both (it supports jBlas/Lapack for dense > matrices, and has a separate experimental GPU module as an add-on). > > on Lapack side, it could be directly integrated. > > for GPU side, there is this thing to look at > https://github.com/BIDData/BIDMat which could be integrated directly, or > merge-ported (similar to Colt effort). It even has a bit quirky matrix dsl > for scala. perhaps some additional ideas how it may all work together, > could also be had from looking at a sister project BidMach -- i haven't > studied it extensively, it looks like an attempt to standardize learning > process. > > Before long, both techniques are to be a new norm for distributed > computation systems. Last chance not to fall behind. > > Any takers? >
