See this Medium post for more information.
You can use Pannotia
to make sure you built GPGPU-Sim correctly.
Below are the instructions to test pagerank
in Pannotia:
hg clone http://gem5-gpu.cs.wisc.edu/repo/benchmarks/pannotia
cd pannotia/pagerank
make
If you want to make the test take a shorter amount of time, change #define ITER
in pagerank.cu
to 1.
Next, you need to make symlinks to the GPU configuration files.
Pick the appropriate GPU, then:
ln -s gpgpu-sim_distribution/configs/<your gpu>/config_fermi_islip.icnt
ln -s gpgpu-sim_distribution/configs/<your gpu>/gpgpusim.config
ln -s gpgpu-sim_distribution/configs/<your gpu>/gpuwattch_<your gpu>.xml
Make sure you source gpgpu-sim_distribution/setup_environment
before attempting to run the workload.
Using coAuthorsDBLP.graph as an example, use the following to run pagerank:
./pagerank coAuthorsDBLP.graph 1
You can find this container on Docker Hub, under the name jlperona/gpgpu-sim-build. This Dockerfile and container will build and run GPGPU-Sim as is, but has a very outdated version of CUDA installed (3.2.14).
You can find this container on Docker Hub, under the name jlperona/gpgpu-sim-build-update. This Dockerfile will build GPGPU-Sim, but programs will not use GPGPU-Sim. Instead, they attempt to use the system CUDA installation. I have not been able to figure out why.