Further NaN problem with dsgesv_gpu

[fletchjp]

I have done some further tests on the 7520 problem reported at the top of this thread. I am working now with RC5.

I can report that the test as supplied fails for the 7520 case on two different hardwares, both the Fermi I used and also a C1xxx.

Also, I tried running just the single line at 7520 and sizes around it. After some other sizes e.g. 7540 all running O.K. I reran 7520 and that also ran correctly. The implication is that this is another issue around memory allocation and initialisation.

I hope that helps.

John

[fletchjp]

Hello

I have just started to test with MAGMA 1.3.0 and I am still getting this bug with DSGESV which I reported in early 2011.

Note the line starting 7520.

John

Code: Select all

fletcher@fletcher-desktop:~/magma-1.3.0/testing$ ./testing_dsgesv_gpu
MAGMA 1.3.0
device 0: GeForce GTX 460, 1400.0 MHz clock, 2047.2 MB memory, capability 2.1

Usage:
  testing_dsgesv_gpu -nrhs 1 -N 1024

Epsilon(double): 1.110223e-16
Epsilon(single): 5.960464e-08

  N   DP-Factor  DP-Solve  SP-Factor  SP-Solve  MP-Solve  ||b-Ax||/||A||  NumIter
==================================================================================
 1024   27.53     24.56     62.96      48.50      22.60     2.119594e-16      3
 2048   47.45     43.78    140.13     120.93     101.13     2.047622e-15      3
 3072   59.18     56.60    180.76     162.64     147.95     2.986802e-16      3
 4032   62.78     60.95    244.82     221.62     194.84     9.296049e-14      4
 5184   65.56     64.38    276.29     258.37     238.53     4.860976e-16      3
 6016   66.97     65.99    289.30     274.61     247.69     1.762242e-14      4
 7040   68.03     67.34    299.59     287.09     269.98     1.343772e-15      3
 7520   69.19     68.46    302.53     291.20     289.44     0.000000e+00      0
 8064   70.23     69.59    311.62     301.81     278.30     3.651095e-15      4
 8192   70.37     69.76    307.58     297.77     279.63     2.647270e-15      3