[gmx-developers] gromacs 5.1rc1 OpenCL problem with Parrinello-Rahman

Carlo Camilloni carlo.camilloni at gmail.com
Wed Jul 15 17:22:12 CEST 2015


Dear Mark and Szilard,

thanks for your answer. I filed a bug in redmine but in the meantime I was running more tests and I am a bit scared by what I found:

what I have done is the following I have performed a single step run with gmx51-rc1 compiled with cuda, again clang and so on
and compared the forces on the first step with -nb cpu or not (I am using -pforce 1), the forces are identical:

ie.:

cuda-gpu

step 0  atom      1  x    3.940    5.612    2.226  force  1.90839e+03
step 0  atom      2  x    3.852    5.659    2.211  force  4.24845e+02
step 0  atom      3  x    3.979    5.665    2.303  force  6.89472e+02
step 0  atom      4  x    3.992    5.610    2.139  force  7.42053e+02


cpu:

step 0  atom      1  x    3.940    5.612    2.226  force  1.90839e+03
step 0  atom      2  x    3.852    5.659    2.211  force  4.24845e+02
step 0  atom      3  x    3.979    5.665    2.303  force  6.89472e+02
step 0  atom      4  x    3.992    5.610    2.139  force  7.42053e+02

if I do the same test on the version compiled with OpenCL 

cpu: 

(the former are done on my macbook pro avx2_256  this latter on a MacPro avx_256, this should
explain the small differences in the forces)

step 0  atom      1  x    3.940    5.612    2.226  force  1.90838e+03
step 0  atom      2  x    3.852    5.659    2.211  force  4.24848e+02
step 0  atom      3  x    3.979    5.665    2.303  force  6.89470e+02
step 0  atom      4  x    3.992    5.610    2.139  force  7.42043e+02

opencl-gpu:
step 0  atom      1  x    3.940    5.612    2.226  force  1.48597e+03
step 0  atom      2  x    3.852    5.659    2.211  force  6.26942e+02
step 0  atom      3  x    3.979    5.665    2.303  force  8.44032e+02
step 0  atom      4  x    3.992    5.610    2.139  force  7.92786e+02

I am afraid there is something wrong  in OpenCL kernels.

I am using the topol-nvt-nogen.tpr I have uploaded on redmine.

Best,
Carlo





More information about the gromacs.org_gmx-developers mailing list