Title: Fast evaluation of elementary functions with medium precision

Topics: reliable numerical algorithms, high performance computing

Address

Research team: MAX, Algebraic modeling and symbolic computation

Contacts

Context

The MAX team is searching for PhD candidates on the themes of the “ODELIX” ERC Advanced Grant. The present M2 internship proposal allows applicants to familiarize themselves with these themes. Upon successful completion of the internship, there will be an opportunity to pursue with a PhD.

Applying

Applications can be sent by email to the above contact persons; they should include a CV, a transcript of records, a letter of motivation, and optional recommendation letters. Note that we have no hard deadlines; new applications will be considered at any time of the year, as long as the Odelix project runs.

The class of elementary functions consists of the exponential, the trigonometric functions , , , the hyperbolic functions , , , and their functional inverses , , , , , , and . In C and C++, these functions are in the standard libm library. This implementation is highly optimized, but limited to single (24 bits) and double (53 bits) precisions. On the other side of the spectrum, dedicated libraries such as Mpfr [1] can be used to evaluate elementary functions efficiently with very large precisions. In the context of modern safety-critical applications such as aerospace engineering and autonomous vehicles design, computing with precision higher than the standard ones is required but the full power of Mpfr may be an overkill.

The aim of this internship is to produce a highly optimized implementation for “medium precisions” until bits. This raises several interesting challenges:

1. How to represent medium precision floating point numbers?

2. How to exploit SIMD (single instruction multiple data) instructions in modern processors to speed-up evaluations by an appreciable constant factor.

3. How to exploit special properties of elementary functions for the design of efficient implementations for each of the target precisions?

4. How to guarantee the correctness of the algorithms modulo rounding errors?

For the first two items, one may use and further extend techniques that were developed in [3, 2]. The goal is to produce a C or C++ library that can reliably evaluate elementary functions for any medium precision; see also [4].

Optionally, one may examine the generalization of the methods to other special functions. Of particular interest in this respect is the class of holonomic functions, i.e. functions that satisfy a non-trivial linear differential equation for polynomials . In [5], a library is described that allows to efficiently and reliably evaluate such functions with high precision. It would be interesting to have a similar library for medium precisions.