[R-sig-hpc] A few questions on using R and CUDA

Charles Determan cdeterm@njr @end|ng |rom gm@||@com
Thu Mar 21 18:22:39 CET 2019


To my knowledge there is no CUDA LAPACK plugin.  NVBLAS is a CUDA
accelerated version of BLAS (https://developer.nvidia.com/cublas).  As
such, I wouldn't expect it to have any LAPACK.  If someone knows otherwise
or is aware of a GPU LAPACK plugin I would be very much interested.

Regards,
Charles

On Tue, Mar 19, 2019 at 7:40 PM Nafis Sadat <sadatnfs using uw.edu> wrote:

> Hey everyone,
>
> Thanks so much for all your help in setting this up. One thing I am seeing
> with the packages I'm interested in is that: even after LD_PRELOADing with
> the NVBLAS shared object, the LAPACK is still coming from the MKL library.
>
> Would anyone happen to know if there's a LAPACK library I can overload as
> well on top of the NVBLAS? I'm not sure yet why my packages are not picking
> up all the CUDA innards (I wonder if I need to re-compile with NVBLAS...)
> ________________________________
> From: Dirk Eddelbuettel <dirk.eddelbuettel using gmail.com> on behalf of Dirk
> Eddelbuettel <edd using debian.org>
> Sent: Thursday, March 7, 2019 3:15 PM
> To: Nafis Sadat
> Cc: Qiang Kou; Andrew Ekstrom; r-sig-hpc using r-project.org
> Subject: Re: [R-sig-hpc] A few questions on using R and CUDA
>
>
> On 7 March 2019 at 20:43, Nafis Sadat wrote:
> | Thanks everyone. Can confirm that it works!
>
> Excellent!  Maybe to close this out you can do a really small quick
> benchmark
> to see if you can find where on your system "N" might be such that costs of
> transport to the GPU does not outweight the gain from computing on the GPU?
> Nothing too fancy -- just a set of QR or SVD or dgemm (mat.mult.) for three
> or so different sample sizes?  If you have the time that is ...
>
> Dirk
>
> --
> http://dirk.eddelbuettel.com | @eddelbuettel | edd using debian.org
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-hpc mailing list
> R-sig-hpc using r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-hpc
>

	[[alternative HTML version deleted]]



More information about the R-sig-hpc mailing list