openmp

Using OpenMP in Owl

Owl provides an option of using OpenMP in compiling. The script here aims to evaluate the impact of using OpenMP on several core N-dimensional array operations. In this doc we briefly introduce how to reproduce the results. Note: The current script selects only a few operations, and is not well-automated. It will be updated.

1. Compile and install Owl without the OpenMP option, and then run the script crosspoint.ml. It will generate a file eval_omp_cross.csv. Then run cp eval_omp_cross.csv openmp_cross.csv.

2. Compile and install Owl with the OpenMP option, setting the number of threads using omp_set_num_threads in source code owl/core/openmp/owl_ndarray_maths_map_omp.h (or change the environment variable) to two. And then run the script again. Manually copy the generated data (two columns) from eval_omp_cross.csv and append them to openmp_cross.csv.

3. Repeat step 2 with number of threads set to four.

4. Now we have an openmp_cross.csv file with seven columns. Run the script draw_fig.py to visualise these data.