Test routines for the sparse matrix methods in 1.6.2

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
fletchjp
Posts: 203
Joined: Mon Dec 27, 2010 7:29 pm

Test routines for the sparse matrix methods in 1.6.2

Post by fletchjp » Wed Jul 22, 2015 1:29 pm

Are there any test routines or examples available for the magma sparse routines in Magma 1.6.2?

I am particularly interesting in calling from FORTRAN and there seem to be few examples of anything for this.

Is there guidance for calling the C or C++ routines from FORTRAN?

Thanks

John

fletchjp
Posts: 203
Joined: Mon Dec 27, 2010 7:29 pm

Re: Test routines for the sparse matrix methods in 1.6.2

Post by fletchjp » Wed Jul 22, 2015 1:43 pm

I have now found that the sparse methods are in a folder sparse-iter and not built by default.

I can see from a failure that I need routines with names starting cusparse. What version of CUDA do I need for this?

I am currently using CUDA 4.1 which I guess is not up to date enough.

Thanks

John

hartwig anzt
Posts: 90
Joined: Tue Sep 02, 2014 5:44 pm

Re: Test routines for the sparse matrix methods in 1.6.2

Post by hartwig anzt » Wed Jul 22, 2015 3:25 pm

Dear John,

to generate the sparse library, just enter the sparse-iter folder and follow the instructions. The building process is very similar to the dense MAGMA.


MAGMA is in general backward-compatible. CUDA 4 is pretty old, if you can, I would recommend getting a newer version. The newer versions then also ship with cuSPARSE, which will be needed by the sparse-iter package. You can also install cuSPARSE separately if you want to keep CUDA 4.0, but I would really recommend upgrading. This may save you a lot work. Which kind of GPU do you want to run the code on?

Everything is however designed for C/C++, to use it with Fortran you may need to use wrappers.

Hope this helps.

Hartwig

fletchjp
Posts: 203
Joined: Mon Dec 27, 2010 7:29 pm

Re: Test routines for the sparse matrix methods in 1.6.2

Post by fletchjp » Wed Jul 22, 2015 5:44 pm

Hartwig

Thank you. My system is a GTX 460, which is a Fermi. My operating system is Ubuntu 12.04 although I do have more recent compilers than gcc 4.6 which is the default.

At present I am using CUDA 4.1 and compile using gcc 4.5. Rather to my surprise that will run both MAGMA 1.5.0 and 1.6.2 for most of the tests.

Code: Select all

MAGMA 1.6.2  compiled for CUDA capability >= 2.0
CUDA runtime 4010, driver 5000. MAGMA not compiled with OpenMP. 
ndevices 1
device 0: GeForce GTX 460, 1400.0 MHz clock, 2047.2 MB memory, capability 2.1
I am worried that if I upgrade too far either my hardware or Ubuntu 12.04 will not be supported. I think I am O.K. with CUDA 6.5 and possibly 7.0.

Can anyone advise me please?

Thank you.

John

hartwig anzt
Posts: 90
Joined: Tue Sep 02, 2014 5:44 pm

Re: Test routines for the sparse matrix methods in 1.6.2

Post by hartwig anzt » Wed Jul 22, 2015 9:33 pm

John,

I can completely understand your concern. I found a few reference for upgrading to 6.5 on ubuntu 12.04, so you may give it a try:

http://jasonjuang.blogspot.com/2013/09/ ... -1204.html
http://ivonneleonor.blogspot.com/2014/0 ... -from.html

According to NVIDIA, 7.0 is also compatible to ubuntu 12.04:
http://docs.nvidia.com/cuda/cuda-gettin ... z3gfpgqxC6

Let me know whether this works!

Thanks, Hartwig

fletchjp
Posts: 203
Joined: Mon Dec 27, 2010 7:29 pm

Re: Test routines for the sparse matrix methods in 1.6.2

Post by fletchjp » Thu Jul 23, 2015 5:49 am

Hartwig

Thank you for that. I am going to upgrade to CUDA 6.5.

I can also upgrade my Ubuntu to 14.04 but that is a big step which can disrupt other things I am doing. I think I do need to do it at some point.

Thanks again

John

Post Reply