GPU concurrency: weak behaviours and programming assumptions
- Submitting institution
-
Imperial College of Science, Technology and Medicine
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 2228
- Type
- E - Conference contribution
- DOI
-
10.1145/2775054.2694391
- Title of conference / published proceedings
- ACM Sigplan Notices
- First page
- 577
- Volume
- 50
- Issue
- 4
- ISSN
- 1523-2867
- Open access status
- Out of scope for open access requirements
- Month of publication
- March
- Year of publication
- 2015
- URL
-
-
- Supplementary information
-
10.1145/2775054.2694391
- 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
-
7
- Research group(s)
-
-
- Citation count
- 25
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- The first large-scale empirical study of NVIDIA and AMD GPU memory consistency models. Led to discovery of defects in CUDA code examples in published memory model documentation from NVIDIA; engagement with NVIDIA led to them fixing this documentation (https://developer.nvidia.com/cuda-example-errata-page). A follow-on collaboration with AMD enabled early defect discovery in next-generation AMD GPU design (OOPSLA'15; https://dl.acm.org/citation.cfm?id=2814283; AMD, contact: FoEREF@ic.ac.uk). GPU memory model research building on this work (PLDI'16; https://dl.acm.org/citation.cfm?id=2908114 and POPL'17; https://dl.acm.org/citation.cfm?id=3009857) underpins the industry-standard Vulkan memory model (NVIDIA, contact: FoEREF@ic.ac.uk), and formed the basis for work in Donaldson's EPSRC Early Career Fellowship (EP/N026314/1; £1M). ASPLOS'15 acceptance rate: 16.7%./287.
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -
