Compiling CP2K with GPU support doesn't utilize GPU

[Ray Schireman]

Hi all,

I compiled cp2k on a Slurm compute node with a V100 with the following command:

./install_cp2k_toolchain.sh --enable-cuda=yes --gpu-ver=V100 --with-openmpi=install

No errors were reported, but when I go to run a benchmark (on the same node), the GPU is not being utilized. Here's the command I used:

mpirun -np 1 ../../exe/local/cp2k.popt -i H2O-2048.inp

In fact, the calculation results in an error after a few seconds:

Extrapolation method: initial_guess
[dg-gpunode04:20950:0:20950] Caught signal 11 (Segmentation fault: address not mapped to object at address 0x2b551b96b000)
==== backtrace (tid:  20950) ====
=================================

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x2b46c0edf62f in ???
#1  0x2b46c08a40a7 in ???

I'm not really sure what's going on, my CPU version works just fine (besides poor performance on multiple nodes). I would appreciate any feedback.

Thank you for your time,

Ray

[Ray Schireman]

Also utilized the binary in local_cuda, still have the same issue