Global Optimization of Operand Transfer Fusion in Heterogeneous Computing
2019 (English)In: SCOPES '19: Proceedings of the 22nd International Workshop on Software and Compilers for Embedded Systems, Association for Computing Machinery (ACM), 2019, p. 49-58Conference paper, Published paper (Refereed)
Abstract [en]
We consider the problem of minimizing, for a dataflow graph of kernel calls, the overall number of operand data transfers, and thus, the accumulated transfer startup overhead, in heterogeneous systems with non-shared memory. Our approach analyzes the kernel-operand dependence graph and reorders the operand arrays in memory such that transfers and memory allocations of multiple operands adjacent in memory can be merged, saving transfer startup costs and memory allocation overheads.
Place, publisher, year, edition, pages
Association for Computing Machinery (ACM), 2019. p. 49-58
Keywords [en]
heterogeneous computing, program optimization, GPU, memory mapping, kernel dataflow graph, Hamiltonian path, message fusion, allocation fusion, distributed memory, Heterogeneous computing, data transfer fusion, CUDA
National Category
Computer Sciences
Identifiers
URN: urn:nbn:se:liu:diva-168603DOI: 10.1145/3323439.3323981ISBN: 9781450367622 (print)OAI: oai:DiVA.org:liu-168603DiVA, id: diva2:1461468
Conference
22nd International Workshop on Software and Compilers for Embedded Systems (SCOPES-2019), St. Goar, Germany, May 2019
Funder
EU, Horizon 2020, 8010152020-08-262020-08-262020-08-26