Hi Aanchan, > Just to give you an intro: > I work in speech recognition, with C/C++ toolkits like HTK and Kaldi. We > have added some of our own libraries to these codebases maintained > internally at McGill. I am also familiar with Python, Perl, Bash and > CUDA. I use a Python wrapper called Gnumpy (wrapped around a CUDA matrix > library called CUDAMat) to train neural networks on GPU boards > routinely. I use Linear Algebra in my work routinely.
Ah, I see, thanks for the informations, all good stuff :-) > I had a very nice conversation with Philippe on IRC who mentioned that > there are some issues with GEMV in PyViennaCL. Since I am familiar with > Gnumpy I could look into that. Hmm, that might be pretty hard to debug for a start, since it connects the most high-level-layer with the most low-level functionality. Also it seems like Philippe found a fix for this just shortly after your chat. > Philippe also mentioned that the current > implementation of the SVD is a bit slow. I could profile that, and start > from there. If necessary start with a newer implementation and see if I > get any better. This is actually a very good starting point. We get a few user requests on a faster SVD, and I think there is quite a bit of optimization potential for the existing OpenCL implementation. A bachelor student will also port this to CUDA and OpenMP in the next months, so if you want to look at the OpenCL part now, it would certainly help. > I was also recently working with the Generalized SVD. As > of the 70's the GSVD was developed for N=2 matrices. Recently a group at > Utah has developed it for three or more matrices > :http://www.alterlab.org/HO_GSVD/. N=2 matrices case is based off the QR > decomposition, for N>2 a QR decomposition based method is not that > straightforward. But given their current (non-QR) solution I could try > and implement that. This also sounds like a great project to start with. Although I only looked at the paper briefly, the algorithm looks sufficiently compact. More importantly, you also seem to have direct use for the implementation, which is a good thing maintenance-wise :-) > Do these sound like good things to start with? Generalized SVD and 'standard' SVD are certainly good things to look at if you're familiar with the linear algebra behind it. I can also send you slides about a randomized SVD algorithm, which I know is fairly easy to implement for all three compute backends with the existing functionality in ViennaCL. Are you fine with issuing pull requests through GitHub for a start? If you want to start off with a stable state, I suggest you start (i.e. fork) early next week when the 1.6.0 release is out. :-) Best regards, Karli ------------------------------------------------------------------------------ _______________________________________________ ViennaCL-devel mailing list ViennaCL-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/viennacl-devel
