Copyrights to these papers may be held by the publishers. The download files are preprints. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.

Benjamin Hess (Master thesis, Computer Science, ETH Zurich, Switzerland, 2013)
Automatic Refactoring: Locality Friendly Interface Enhancements for Numerical Functions
Preprint (1.2 MB)
Bibtex

Recent improvements in processor architectures such as multiple cores and larger single-instruction multiple-data (SIMD) vector units increased the discrepancy between processing speed and memory bandwidth. Today, the memory bandwidth has become the biggest bottleneck in high performance domains. Numerical functions are often target to heavy optimizations to get as much performance as possible. These functions presume a specific data layout and have a fixed domain and range. Adjusting unsuitable data according to these requirements can take up a significant amount of the runtime due to heavy memory operations. By integrating layout, domain and range adjustments into numerical functions, memory bandwidth is saved as the adjustments happen inplace during execution of the function. Four transformations are provided by the developed tool to refactor the most common adjustments into numerical functions. Restrictions on the to transformed function ensure the correctness of the transformations. These transformations enable the function to directly work on previously incompatible data which makes the manual adjustments superfluous and saves memory bandwidth by not needing to adjust the data manually before calling the function. The runtime of the transformed function is highly dependent on the used function and transformation type, ranging from 50% slower to up to 5 times faster compared to applying the adjustments manually before or after the function. The developed tool shows that automatic refactoring with a subset of C is possible and allows a developer to enhance numerical functions with minimal additional effort providing more flexible and high-performance versions of the functions with only having to maintain the original function.

Keywords:
(No keyword)