ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Design of the scheduler of accelerator device driver supporting OpenCL application for heterogeneous accelerator in Supreme-K system
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
Youngho Kim, Eunji Lim, Shinyoung Ahn, Yoomi Park
Issue Date
2022-12
Citation
International Conference on Internet (ICONI) 2022, pp.297-299
Language
English
Type
Conference Paper
Abstract
In this paper, we present a design on the development of an accelerator device driver for the ultra-high performance processor SoC, which is composed of general-purpose ARM cores and the heterogeneous accelerator, named eXtreme Processing Unit (XPU). We describe our ongoing work on scheduler of accelerator device driver, a software-based OpenCL work group scheduling technique that minimize the scheduling overhead through the concurrent execution of the generation of scheduling metadata and the processing of work group assignment to XPU computing unit and parallel computational task execution in XPU. In addition, we provide a polling-based completion detection mechanism for the OpenCL work group execution, which reduces the scheduling overhead through dynamic update of the polling frequency.
KSP Keywords
Concurrent Execution, Device driver, Processing unit, Scheduling overhead, Scheduling technique, Ultra-high performance, completion detection, detection mechanism, dynamic update, group scheduling, task execution