GPU package with multiple platforms.

LAMMPS LAMMPS Mailing List Mirror
1

I recently built lammps with OpenCL support because we have a Xeon Phi to play around with. I have no technical issues to report there. However, we also have an nvidia M2050 that has CUDA/OpenCL support. When I try to run it on the GPU node, it always detects that I have OpenCL for my CPUs installed and uses those. I put the output of ocl_get_devices at the end of the email. It detects that I have 2 platforms but seemingly only 1 device. How can I get it to use the GPU? I also tried playing with the line:

package gpu force/neigh 0 0 1.0

but when I do anything besides 0’s it complains with:

ERROR: Could not find/initialize a specified accelerator device

I don’t think I’d need to recompile, would I?

Output of ocl_get_devices:

2

I seemed to have not quite finished the subscribe process before sending this; oops.

3

I seemed to have not quite finished the subscribe process before sending
this; oops.

I recently built lammps with OpenCL support because we have a Xeon Phi to
play around with. I have no technical issues to report there. However, we
also have an nvidia M2050 that has CUDA/OpenCL support. When I try to run it
on the GPU node, it always detects that I have OpenCL for my CPUs installed
and uses those. I put the output of ocl_get_devices at the end of the email.
It detects that I have 2 platforms but seemingly only 1 device. How can I
get it to use the GPU? I also tried playing with the line:

you should uninstall the ICD for the Intel OpenCL in
/etc/OpenCL/vendors from the node with the GPU. the GPU library in
LAMMPS is currently not set up to select between different platforms,
only between devices within a platform.

axel.

4

LAMMPS is not the only piece of software we use; uninstalling Intel OpenCL
is not an option. It seems the best thing to do in the meantime is to
compile a cuda lammps version and use that for when running strictly on the
GPU. I would like to only maintain one installation but things like this
happen. Thank you for the quick reply.

5

are you telling me, that you you have the xeon phi *and* the GPU on
the same physical node?
it doesn't look so from the output of ocl_get_devices.

also, i didn't tell you to uninstall Intel OpenCL, only to remove the
ICD on the machine with the GPU.
there is not much of a point have OpenCL favor the intel CPU driver
over the nvidia GPU driver on a node with a GPU.

axel.