Hello everyone,
I am using the 28Aug12 LAMMPS distribution and I am trying to get the GPU acceleration working on Ubuntu 11.04- 64 bit. I have 2 gtx 580’s and a core i-7 3960x. I have successfully compiled the gpu lib and compiled lammps with openmpi making sure to use make yes-gpu and make yes-kspace. I can also run the examples in the CUDA-SDK and nvc_get_devices registers that both gtx 580’s are good to go. I am able to run the examples for the cpu code but when I use the gpu input files with the benchmark examples I get:
running: in.lj.gpu
Cuda driver error 101 in call at file ‘geryon/nvd_device.h’ in line 232.
or
Cuda driver error 100 in call at file ‘geryon/nvd_device.h’ in line 232.
To be clear, I use the commands in the README file located in bench/GPU/ for GPU acceleration not USER-CUDA.
After looking through this list serve I found that someone had this problem and solved it by replacing
int my_gpu=node_rank/_procs_per_gpu;
in lib/gpu/pair_gpu_device.cpp
with
int my_gpu=node_rank/_procs_per_gpu+first_gpu;
shown here: http://lammps.sandia.gov/threads/msg19380.html
Unfortunately I can’t seem to find pair_gpu_devices.cpp to make this change. Considering that the suggestion was made in 2011 I am guessing that the file name has changed or the correction was put in place.
I was wondering if it is because I am also using the first GTX 580 as video output which could cause the timing issue so I went into the in.lj.gpu file and changed
if “$g == 1” then “package gpu force/neigh 0 0 1”
to
if “$g == 1” then “package gpu force/neigh 1 1 1”
so that I could push the work load to the second card. This was unsuccessful.
Any suggestions would be helpful.
Thank you for your time,
David W. Caldwell II