Page 1 of 1

clMagma: zhetrd gives NaN or loops infinitely for N>=2048

Posted: Tue Apr 08, 2014 8:12 am
by Jan Kucera
I believe the title says it. The result depends on the choice of OpenCL.

If I use fglrx-driver 13.12 plus those AMD libraries which come with Debian Jessie, it gives just NaN values on output.

If I use any of 13.12 or 14.3 fglrx drivers plus the APPSDK Version 2.9 from AMD, it loops infinitely.
(In either case, for N<2047 zhetrd seems to work perfectly.)

I have compiled clmagma-1.1.0 with mkl and clAmdBlas-1.10.321.

Any hint would be welcome.

JK

Re: clMagma: zhetrd gives NaN or loops infinitely for N>=204

Posted: Wed Apr 09, 2014 10:10 am
by Jan Kucera
I have experimented a bit playing with the version of OpenCL libraries..., unsetting the
GPU_FORCE_64BIT_PTR to avoid possible problems due to 64bit addressing and setting
GPU_MAX_ALLOC_PERCENT=40 to avoid problem with accessible memory exceeding
the addressable space (my Tahiti card has 6GB RAM).

Just to illustrate the problem in one particular case when the test runs for a finite, I obtained this

Code: Select all

root@machine:/usr/local/src/clmagma-1.1.0/testing# ./testing_zhetrd -N 2047
  testing_zhetrd -L|U -N 2047

Initializing clMAGMA runtime ...
Device: Tahiti (memory  3.2 GB, max allocation  2.5 GB, driver  1348.5 (VM))
using default MAGMA_CL_DIR = /usr/local/magma/cl
  N    CPU GFlop/s    GPU GFlop/s   |A-QHQ'|/N|A|  |I-QQ'|/N 
=============================================================
 2047    19.66         20.18       2.011490e-17 7.243096e-17
Just by changing N from 2047 to 2048 I obtained this

Code: Select all

root@machine:/usr/local/src/clmagma-1.1.0/testing# ./testing_zhetrd -N 2048
  testing_zhetrd -L|U -N 2048

Initializing clMAGMA runtime ...
Device: Tahiti (memory  3.2 GB, max allocation  2.5 GB, driver  1348.5 (VM))
using default MAGMA_CL_DIR = /usr/local/magma/cl
  N    CPU GFlop/s    GPU GFlop/s   |A-QHQ'|/N|A|  |I-QQ'|/N 
=============================================================
 2048    17.16         32.61       5.000000e-01 5.000000e-01
JK