Synthesizing Efficient Low-Precision Kernels

Anastasiia Izycheva, Eva Darulova, Helmut Seidl

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

4 Scopus citations


In this paper, we present a fully automated approach for synthesizing fast numerical kernels with guaranteed error bounds. The kernels we target contain elementary functions such as sine and logarithm, which are widely used in scientific computing, embedded as well as machine-learning programs. However, standard library implementations of these functions are often overly accurate and therefore unnecessarily expensive. Our approach trades superfluous accuracy against performance by approximating elementary function calls by polynomials and by implementing arithmetic operations in low-precision fixed-point arithmetic. Our algorithm soundly distributes and guarantees an overall error budget specified by the user. The evaluation on benchmarks from different domains shows significant performance improvements of 2.23 $$\times $$ on average compared to state-of-the-art implementations of such kernel functions.

Original languageEnglish
Title of host publicationAutomated Technology for Verification and Analysis- 17th International Symposium, AVTA 2019, Proceedings
EditorsYu-Fang Chen, Chih-Hong Cheng, Javier Esparza
Number of pages20
ISBN (Print)9783030317836
StatePublished - 2019
Event17th International Symposium on Automated Technology for Verification and Analysis, ATVA 2019 - Taipei, Taiwan, Province of China
Duration: 28 Oct 201931 Oct 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11781 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference17th International Symposium on Automated Technology for Verification and Analysis, ATVA 2019
Country/TerritoryTaiwan, Province of China


  • Approximation
  • Elementary functions
  • Finite precision
  • Synthesis


Dive into the research topics of 'Synthesizing Efficient Low-Precision Kernels'. Together they form a unique fingerprint.

Cite this