Imperial College London

ProfessorAlastairDonaldson

Faculty of EngineeringDepartment of Computing

Professor of Programming Languages
 
 
 
//

Contact

 

+44 (0)20 7594 8266alastair.donaldson Website

 
 
//

Location

 

422Huxley BuildingSouth Kensington Campus

//

Summary

 

Publications

Citation

BibTex format

@inproceedings{Sorensen:2016:10.1145/3022671.2984032,
author = {Sorensen, T and Donaldson, AF and Batty, M and Gopalakrishnan, G and Rakamaric, Z},
doi = {10.1145/3022671.2984032},
pages = {39--58},
publisher = {ACM},
title = {Portable Inter-workgroup Barrier Synchronisation for GPUs},
url = {http://dx.doi.org/10.1145/3022671.2984032},
year = {2016}
}

RIS format (EndNote, RefMan)

TY  - CPAPER
AB - Despite the growing popularity of GPGPU programming,there is not yet a portable and formally-specified barrierthat one can use to synchronise across workgroups. Moreover,the occupancy-bound execution model of GPUs breaksassumptions inherent in traditional software execution barriers,exposing them to deadlock. We present an occupancydiscovery protocol that dynamically discovers a safe estimateof the occupancy for a given GPU and kernel, allowingfor a starvation-free (and hence, deadlock-free) interworkgroupbarrier by restricting the number of workgroupsaccording to this estimate. We implement this idea byadapting an existing, previously non-portable, GPU interworkgroupbarrier to use OpenCL 2.0 atomic operations,and prove that the barrier meets its natural specification interms of synchronisation.We assess the portability of our approach over eightGPUs spanning four vendors, comparing the performanceof our method against alternative methods. Our key findingsinclude: (1) the recall of our discovery protocol isnearly 100%; (2) runtime comparisons vary substantiallyacross GPUs and applications; and (3) our method providesportable and safe inter-workgroup synchronisation acrossthe applications we study.
AU - Sorensen,T
AU - Donaldson,AF
AU - Batty,M
AU - Gopalakrishnan,G
AU - Rakamaric,Z
DO - 10.1145/3022671.2984032
EP - 58
PB - ACM
PY - 2016///
SP - 39
TI - Portable Inter-workgroup Barrier Synchronisation for GPUs
UR - http://dx.doi.org/10.1145/3022671.2984032
UR - http://hdl.handle.net/10044/1/39597
ER -