- author = "Bobby R. Bruce and Justyna Petke",
- title = "Towards automatic generation and insertion of {OpenACC} directives",
- institution = "University College, London",
- year = "2018",
- number = "RN/18/04",
- address = "London, UK",
- month = "12 " # apr,
- keywords = "genetic algorithms, genetic programming, genetic improvement, GPU, parallel computing",
-
URL = "
http://www.cs.ucl.ac.uk/fileadmin/UCL-CS/research/Research_Notes/RN_18_04.pdf",
- size = "35 pages",
- abstract = "While the use of hardware accelerators, like GPUs, can significantly improve software performance, developers often lack the expertise or time to properly translate source code to do so. In this paper we highlight two approaches to automatically offload computationally intensive tasks to a system's GPU by generating and inserting OpenACC directives; one using grammar-based genetic programming, and another using a bespoke four stage process. We find that the grammar-based genetic programming approach reduces execution time by 2.60percent on average, across the applications studied, while the bespoke four-stage approach reduces execution time by 2.44percent. Despite this, our investigation shows a handwritten OpenACC implementation is capable of reducing execution time by 65.68percent. Comparing the differences, we identified a promising avenue for future research: combining genetic improvement with better handling of data to and from the GPU.",
-
notes = "Seoul National University NAS Parallel Benchmark
suite
GB-GP-Parallelisation and Four-Stage-Parallelisation 'both under-perform due to their poor handling of how program variables are transferred to and from the GPU' grammar-based GP. Greedy algorithms and evolutionary strategies",
