dependence on CUDA 6.5?

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
paciorek
Posts: 1
Joined: Tue Mar 17, 2015 8:02 pm

dependence on CUDA 6.5?

Post by paciorek » Wed Mar 18, 2015 1:19 pm

In trying to install MAGMA 1.6.1 with CUDA 6.0 on an Ubuntu 14.04 machine I ran into an issue with testing_zgeqrf_batched.cpp not compiling and had a few run-time failures when running the MAGMA tests. In the source code there is a comment about testing_zgeqrf_batched needing CUDA 6.5 but in the MAGMA documentation I'm not seeing any info on what versions of MAGMA depend on what versions of CUDA. Can anyone provide guidance on CUDA dependencies?

As a related note if there is a dependence on CUDA 6.5, I'll note that the Ubuntu 14.04 CUDA package is 6.0 so if recent versions of MAGMA require CUDA 6.5 that means Ubuntu users will need to install a newer version of CUDA from source rather than relying on the packaged version.

-Chris

----------------------------------------------------------------------------------------------
Chris Paciorek

Statistical Computing Consultant
Statistical Computing Facility, Econometrics Laboratory, Berkeley Research Computing

Office: 495 Evans Hall Email: paciorek@stat.berkeley.edu
Mailing Address: Voice: 510-842-6670
Department of Statistics Fax: 510-642-7892
367 Evans Hall Skype: cjpaciorek
University of California, Berkeley WWW: www.stat.berkeley.edu/~paciorek
Berkeley, CA 94720 USA Permanent forward: paciorek@alumni.cmu.edu

mgates3
Posts: 918
Joined: Fri Jan 06, 2012 2:13 pm

Re: dependence on CUDA 6.5?

Post by mgates3 » Wed Mar 25, 2015 4:47 pm

The MAGMA 1.6.1 library itself does not have any dependence on CUDA 6.5. For instance, you can 'make lib' with CUDA 6.0, at least. I haven't tested with older CUDA libraries.

The tester_zgeqrf_batched requires CUDA 6.5, only to call the equivalent cublas routine to compare performance. You can comment out cublasZgeqrfBatched and it should compile and run. Some of the error checks will be wrong. I've fixed this to check for the CUDA version.

-mark

Post Reply