Improving Floating-Point Numbers: A Lazy Approach to Adaptive Accuracy Refinement for Numerical Computations

Kawabata, Hideyuki; Iwasaki, Hideya

doi:10.1007/978-3-662-49498-1_16

Cited by 2 publications

(1 citation statement)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We should note that other libraries such as iRRAM [23] and IFN [10] might be a few orders of magnitude faster than those examined in this paper. reuse rate were measured for solving Hilbert systems.…”

Section: Experiments 1: Solving Sensitive Linear Systems Ofmentioning

confidence: 79%

Speeding up Exact Real Arithmetic on Fast Binary Cauchy Sequences by Using Memoization Based on Quantized Precision

Kawabata

2017

Journal of Information Processing

Self Cite

View full text Add to dashboard Cite

Exact Real Arithmetic on Fast Binary Cauchy Sequences (FBCSs) provides us a simple and fairly fast way to obtain numerical results of arbitrary precision. The arithmetic on FBCSs can be implemented concisely in a lazy functional language with unlimited-length integer arithmetic, such that each FBCS is represented by a function that generates approximated values with respect to requested precisions. However, application of the arithmetic on FBCSs to programs such as matrix computations, that usually involve large amount of references to common subexpressions, requires care to avoid the blowup of the amount of computation caused by the fact that approximated values are not shared among multiple references. Although simple memoization might alleviate the situation, the effect would be limited since required precisions for subexpressions tend to be various. In this paper, we present an extended design of the arithmetic on FBCSs that enables the memoization based on quantized precision, that is expected to enlarge the reuse rate and reduce the amount of computation without sacrificing the properties of the arithmetic to be exact arithmetic. Numerical experiments by using our prototype libraries in Haskell demonstrated that our approach possesses the potential to outperform existing implementations by orders of magnitude in speed and memory consumption.

show abstract

Section: Experiments 1: Solving Sensitive Linear Systems Ofmentioning

confidence: 79%