It looks like there's a reason for that, and it relates to the different way cuBLAS handles memory. See more info here: http://stackoverflow.com/a/18820210/931700
You can try linking with the thunking library and see if there is still any improvement. On Wednesday, 12 February 2014 13:02:41 UTC-6, Elliot Saba wrote: > > As far as I know, nobody has tried this before. It looks like cuBLAS does > not > conform<http://docs.nvidia.com/cuda/cublas/#cublas-level-2-function-reference>to > the standard BLAS API, (e.g. their functions, instead of being named > something like dgemv() are named cublasDgemv(), and take two extra > parameters at the beginning) so you won't be able to just plug the cuBLAS > shared library into Julia have have it "work". You will need to write some > kind of wrapper library around cuBLAS that does any special > initialization/cleanup that cuBLAS needs, as well as mapping "standard" > BLAS library names to cuBLAS library names. > > A project similar in spirit was the libgfortblas > library<https://github.com/JuliaLang/julia/blob/master/deps/gfortblas.c>that > Jameson wrote, to map our C calls to the cblas versions of Apple's > BLAS libraries. Looking at that could provide some inspiration for this > project, as it does pretty much everything I outlined above, except that > cuBLAS will (of course) require more work, as you'll need to wrap basically > the entire library instead of just a few functions. > > > On Wed, Feb 12, 2014 at 9:53 AM, Pedro Rafael > <[email protected]<javascript:> > > wrote: > >> Many thanks Iain Dunning. You have been very enlightening. Can we link >> the library cublas Julia? >> >> cuBLAS: https://developer.nvidia.com/cuBLAS >> >> >> Hugs, >> Pedro Rafael. >> >> >> 2014-02-12 14:28 GMT-03:00 Iain Dunning <[email protected] <javascript:> >> >: >> >> 1. I think the default is to use OpenBLAS on all platforms, but this can >>> be swapped out if you'd like to use a different BLAS implementation. >>> 2. I believe it already is. The core language certainly is, the rest is >>> just a package ecosystem issue. We have database packages, graphics >>> packages, web/internet packages, etc. - a good sign that it is >>> "general-purpose". >>> 3. I imagine use of the GPU in core Julia is a long way off or ever - >>> but I'm sure there will be more and more packages that do clever things in >>> this area. There certainly is a lot of interest in this area. We have CUDA >>> and OpenCL packages for now, at least. >>> >>> >>> On Wednesday, February 12, 2014 12:16:05 PM UTC-5, Pedro Rafael wrote: >>>> >>>> [1] The installation of Julia from the compilation of the source code >>>> makes use of OpenBlas library? >>>> [2] Julia will also be a general-purpose language like Python? >>>> [3] The use of GPU in the future will be native? >>>> >>>> Regards, >>>> Pedro Rafael Diniz Marinho. >>>> >>> >> >
