Evaluation Artifact for HPAC-Offload (https://github.com/LLNL/HPAC/tree/hpac_offload)

Before running any experiments, the HPAC-Offload compiler and runtime system must be built, and the input data used for the analsysis downloaded. We have provided the script setup.sh to perform both actions. In the root direcotry of this repository, simply run ./setup.sh $NTHREADS, where NTHREADS will be used to build the compiler. On an AMD Epyc system using 64 threads, the build takes about 1 hour. Build times will vary with the sytem architecture and the number of threads used.

**All following instructions assume that the build is complete, and that the directory HPAC exists in the root directory of the repository.

The following lists the expected time needed to perform the accuracy/performance trade-off experiments for two different configurations: small and large. The small configuration runs only those data points shown in the paper, where we take the fastest and slowest 10% of all experiments for 10 different x-axis intervals. The large configuration runs experiments for the entire Cartesian space.

To run the NVIDIA performance/accuracy trade-off experiments, first build the HPAC-Offload compiler using the above steps. Then, enter the directory experiments/nvidia/tradeoff_space. In this directory, you'll see two YAML configuration files: config_all.yaml and config_small.yaml. These files correspond to the large and small configurations, respectively. Files containing the individual experimental configuration for all benchmarks and each approximation technique are in the experiment_config directory. For more information on running these experiments, see the README in the NVIDIA directory.