Presentation
Floating-Point Autotuner for CPU-Based Mixed-Precision Applications
SessionResearch Posters
Event Type
Poster
TP
EX
TimeThursday, November 15th8:30am - 5pm
LocationC2/3/4 Ballroom
DescriptionIn this poster, we present the design and development of an autotuning tool for floating-point code. The goal is to balance accuracy and performance in order to produce an efficient and accurate mixed-precision program. The tuner starts by maximizing accuracy through the use of a high-precision library called CAMPARY and then achieves performance gains under a given error bound by tuning down groups of variables and operations from the higher precision down to double precision. We tested our tuning strategy on a computational fluid dynamics benchmark where we show a 4x speedup relative to the fully high-precision version during the iterative tuning process and achieve an average absolute error of 2.8E-16 compared with the reference solution computed using the 256-bit GNU MPFR extended precision library.
Archive