DOCTORAL SEMINAR

Optimizing number representation of approximable programs

Speaker
Mr Ho Nhut-Minh
Advisor
Dr Wong Weng Fai, Associate Professor, School of Computing


01 Nov 2018 Thursday, 09:00 AM to 10:00 AM

Executive Classroom, COM2-04-02

Abstract:

Recent years, along with the emergence of approximable programs such as deep neural networks, big data, media/video processing and simulation applications, there were many opportunities opened on hardware's ability to support approximate data types: limited precision fixed point in FPGAs, half precision and 8 bits integer in Graphics Processing Units and Tensor Cores. Despite extensive research in approximate computing, the problem of how to efficiently map user's required quality of service to number representations on the target hardware is still an open question. It is partly because of the complexity associated with the search space in the number of bits required to represent each variable in the program. This thesis proposal will tackle the problem of allocating the number of bits that will be used to represent each variable in a given program on a target hardware architecture using two methods: a heuristic search algorithm and an analytical model of rounding error. With the resultant optimized representation for each variable, a set of methods to rewrite and deploy programs and to take advantage of the approximate data types in various hardware configurations will also be presented in this proposal.