clMagma spotrf weird behavior

noob · Post by **noob** » Sun Jul 14, 2013 4:03 pm

Hi,

I was running some of the clMagma test cases on my amd64 linux, and I observed that spotrf in very inaccurate for matrices with sizes between 4096 and 8192. Interestingly, it works great for smaller or larger matrices. Here's the test output

Code: Select all

$ export LD_LIBRARY_PATH=/opt/acml5.3.1/gfortran64_mp/lib:/opt/clAmdBlas-1.10.321/lib64 
$ export MAGMA_CL_DIR=../interface_opencl 
$ ./testing_spotrf_gpu

Usage: 
  testing_spotrf_gpu -N 1024

Initializing clMAGMA runtime ...


  N    CPU GFlop/s (sec)    GPU GFlop/s (sec)    ||R_magma-R_lapack||_F / ||R_lapack||_F
========================================================================================
 1024      15.90 (  0.02)      21.75 (  0.02)         6.047437e-08
 2048      39.45 (  0.07)      41.75 (  0.07)         7.084088e-08
 3072      51.80 (  0.19)     240.15 (  0.04)         5.906473e-08
 4032      61.75 (  0.35)     372.92 (  0.06)         5.172135e-08
magma_spotrf had error 1089.
 5184      64.87 (  0.72)     2411.94 (  0.02)         6.033692e+01
magma_spotrf had error 1953.
 6048      69.36 (  1.06)     1473.26 (  0.05)         6.025791e+01
magma_spotrf had error 3105.
 7200      78.79 (  1.58)     1312.31 (  0.09)         6.249601e+01
magma_spotrf had error 3972.
 8064      82.13 (  2.13)     1092.68 (  0.16)         6.248589e+01
 8928      85.94 (  2.76)     596.50 (  0.40)         1.831093e-07
10560      87.78 (  4.47)     643.52 (  0.61)         2.243686e-07

I have a Radeon HD 7950, and have also observed that for a dimension of 4096 my computer completely hangs when I run the above test.

Judging from the running times, the CPU computation seems to always succeed (running times are increasing), which the GPU computation seems to fail (running times are suspiciously small). Is this known and is there a workaround?

Thanks!