Problem with Magma sgetrf on non-GTX 280

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
evanlezar
Posts: 33
Joined: Tue Aug 25, 2009 7:20 pm
Location: Stellenbosch, South Africa
Contact:

Problem with Magma sgetrf on non-GTX 280

Post by evanlezar » Tue Aug 25, 2009 7:32 pm

Hi

I have downloaded and built MAGMA, and would just like to say how EXCITED I am about it.

However, I run into some problems when I run the sgetrf tests on both a 8600 GT and a 9500 GT.

When running the tests for matrix sizes greater than the block size (which I have changed set in the magma_get_sgetrf_nb function in get_np.c).

For example if the block size returns 128 if m is less than 2048, then the following output is obtained when running for -N 129:

./testing_sgetrf_gpu -N 129
device 0: GeForce 9500 GT, 1350.0 MHz clock, 1023.3 MB memory


N CPU GFlop/s GPU GFlop/s ||PA-LU|| / (||A||*N)
==========================================================
129 0.62 0.49 1.718544e-01


as soon as the matrix size is 128 or less, it seems to work well:

./testing_sgetrf_gpu -N 128
device 0: GeForce 9500 GT, 1350.0 MHz clock, 1023.3 MB memory


N CPU GFlop/s GPU GFlop/s ||PA-LU|| / (||A||*N)
==========================================================
128 0.73 0.70 1.657047e-09
45 0.42 0.40 2.046684e-09


Does anyone have any ideas as to how I can go about attempting to debug and fix this?

I have also tested the code on another machine running a GTX 280, and it appears to be running well.

Thanks
Evan

admin
Site Admin
Posts: 20
Joined: Tue Aug 04, 2009 12:23 pm

Re: Problem with Magma sgetrf on non-GTX 280

Post by admin » Wed Aug 26, 2009 4:06 pm

Evan,

Thanks for trying out MAGMA and the feedback!
Thanks also for giving us a temporary account to fix the problem.

It turned out a MAGMA library (libmagmablas.a) had to be recompiled from the sources. The magmablas sources will be included with the distribution of MAGMA to avoid similar problems.

The correct results now look like

Code: Select all

stomov@genugtig:~/magma_0.1/testing$  ./testing_sgetrf
device 0: GeForce 9500 GT, 1350.0 MHz clock, 1023.3 MB memory

Usage: 
  testing_sgetrf -N 1024

  N    CPU GFlop/s    GPU GFlop/s    ||PA-LU|| / (||A||*N)
==========================================================
 1024      1.16          14.63         1.933232e-09
 2048      1.94          22.49         1.788219e-09
...

stomov@lefauve:~/magma_0.1/testing$ ./testing_sgetrf
device 0: GeForce 8600 GT, 1404.0 MHz clock, 511.3 MB memory

Usage: 
  testing_sgetrf -N 1024

  N    CPU GFlop/s    GPU GFlop/s    ||PA-LU|| / (||A||*N)
==========================================================
 1024      1.70           7.96         2.289750e-09
 2048      3.87          19.19         2.111563e-09
...
Regards,
Stan Tomov

evanlezar
Posts: 33
Joined: Tue Aug 25, 2009 7:20 pm
Location: Stellenbosch, South Africa
Contact:

Re: Problem with Magma sgetrf on non-GTX 280

Post by evanlezar » Wed Aug 26, 2009 5:24 pm

Stan,

Thank you for looking at this so quickly. The use of the machine is the least that I could do.

I wanted to ask, do you have some kind of testing/benchmarking program that we could perhaps help out with? This would in theory pick up bugs like this a lot earlier and also give other users an idea of the kind of performance that they can expect.

Thanks again for you effort.
Evan

Post Reply