High Performance Stencil Code Generation with LIFT
- Submitting institution
-
University of Glasgow
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 11-04919
- Type
- E - Conference contribution
- DOI
-
10.1145/3168824
- Title of conference / published proceedings
- International Symposium on Code Generation and Optimization (CGO'18)
- First page
- 100
- Volume
- -
- Issue
- -
- ISSN
- -
- Open access status
- -
- Month of publication
- -
- Year of publication
- 2018
- URL
-
http://eprints.gla.ac.uk/153441/
- Supplementary information
-
-
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- No
- Number of additional authors
-
4
- Research group(s)
-
-
- Citation count
- 24
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- ORIGINALITY: The paper presents a novel formulation of stencil computations as composable parallel patterns allowing, for the first time, to reuse commonly known parallel patterns and optimisations defined on them for stencil computations. SIGNIFICANCE: the work has won the best paper award in the premier conference on compilation. RIGOUR: The paper introduces and integrates new parallel patterns into the existing formal Lift language. We present the formal description of the overlap tiling optimisation as a semantic preserving rewrite rule, and extensive experimental results demonstrate the performance of stencil applications against the state-of-the-art polyhedral compiler on three GPUs (desktop and mobile).
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -