Here we provide help for setting up general-purpose computing on graphics processing units (GPGPU) using CUDA. Performing (part of) the calculations on a graphics card can result in a significant speed-up. Currently two codes in AMUSE support or require GPGPU: phi-GRAPE (using Sapporo) and Octgrav.
In the AMUSE root directory a self-help script can be found. If building or testing any of the codes mentioned above fails and you wonder why, it will hopefully provide you with helpful suggestions. From a command line run the bash script cuda_self_help:
First determine the model of your GPU.
> nvidia-settings -q gpus
- Click on “Apple Menu”
- Click on “About this Mac”
- Click on “More Info”
- Select “Graphics/Displays” under Contents list
Check whether your GPU model is listed among Nvidia’s CUDA-enabled GPUs.
If not, download and install it from CUDA Development Tools.
After installing the CUDA TK and SDK, make sure the environment variables CUDA_TK and CUDA_SDK are set correctly. For shell (bash) you need to do:
export CUDA_TK=/path/to/cuda_tk export CUDA_SDK=/path/to/cuda_sdk
‘/path/to/cuda_tk’ should hold directories named ‘include’, ‘lib’, and ‘lib64’ (where libcudart.so is located)
‘/path/to/cuda_sdk’ should hold a directory named ‘common/inc’ (where various header files are located)
We recommend you add these lines to your ‘.bashrc’ file so that the variables are set correctly for all sessions. If you have a C shell you need to do a setenv and edit the ‘.cshrc file.
AMUSE needs to be configured with the option --enable-cuda. See GPU.
Now try building for example Octgrav and run the nosetests (from AMUSE root directory), but first re-initialize mpd (or it will remember its original environment):
mpdallexit mpd & make octgrav.code nosetests ./test/codes_tests/test_octgrav.py
If this fails, please contact us through the ‘amusecode’ google group, or on IRC at the #amuse channel on irc.freenode.net.