Analysis and optimization of an implementation of hierarchical Tucker tensors

Abstract: The hierarchical Tucker tensor format is a format for approximated tensors which has been decomposed into smaller parts. This structure can be useful in high-dimensional computer simulations. One implementation of this structure is in a library called htlib. The goal of this project was to parallelize htlib with OpenMP and to explore other ways of improving the performance of htlib with regards to the execution time. One loop was parallelized with OpenMP and for some uses it resulted in a speedup over 1.8 for that part of the code but it had a small effect on the library as a whole. This project also compared different implementations of matrix multiplication. This showed that good speedup can be gained from using an optimized third party implementation and also that it might be possible to decrease the execution time even more when implemented by oneself. Although this project managed to reduce the execution time of htlib, there is still more work that can be done to the library to decrease the execution time further.