Section: Overall Objectives
Keywords : computer arithmetic, integer computation, approximate computation, floating-point representation, elementary function, reliability of numerical software, multiple-precision arithmetic, interval arithmetic, computer algebra, finite field, linear algebra, lattice basis reduction, VLSI circuit, FPGA circuit, low-power operator, embedded chips.
The Arénaire project aims at elaborating and consolidating knowledge in the field of Computer Arithmetic. Reliability, accuracy, and performance are the major goals that drive our studies. Our goals address various domains such as floating-point numbers, intervals, rational numbers, or finite fields. We study basic arithmetic operators such as adders, dividers, etc. We work on new operators for the evaluation of elementary and special functions (log , cos , erf , etc.), and also consider the composition of previous operators. In addition to these studies on the arithmetic operators themselves, our research focuses on specific application domains (cryptography, signal processing, linear algebra, lattice basis reduction, etc.) for a better understanding of the impact of the arithmetic choices on solving methods in scientific computing.
We contribute to the improvement of the available arithmetic on computers, processors, dedicated or embedded chips, etc., both at the hardware level and at the software level. Improving computing does not necessarily mean getting more accurate results or getting them faster: we also take into account other constraints such as power consumption, code size, or the reliability of numerical software. All branches of the project focus on algorithmic research and on the development and the diffusion of corresponding libraries, either in hardware or in software. Some distinctive features of our libraries are numerical quality, reliability, and performance.
The study of the number systems and, more generally, of data representations is a first topic of uttermost importance in the project. Typical examples are: the redundant number systems used inside multipliers and dividers; alternatives to floating-point representation for special purpose systems; finite field representations with a strong impact on cryptographic hardware circuits; the performance of an interval arithmetic that heavily depends on the underlying real arithmetic.
Another general objective of the project is to improve the validation of computed data, we mean to provide more guarantees on the quality of the results. For a few years we have been handling those validation aspects in the following three complementary ways: through better qualitative properties and specifications (correct rounding, error bound representation, and portability in floating-point arithmetic); by proposing a development methodology based on proof assistants; by studying and allowing the cooperation of various kinds of arithmetics such as constant precision, intervals, arbitrary precision and exact numbers.
These goals may be organized in four directions: hardware arithmetic , software arithmetic for algebraic and elementary functions , formal proof and validation , and arithmetics and algorithms for scientific computing . These directions are not independent and have strong interactions. For example, elementary functions are also studied for hardware targets, and scientific computing aspects concern most of the components of Arénaire.
Hardware arithmetic. From the mobile phone to the supercomputer, every computing system relies on a small set of computing primitives implemented in hardware. Our goal is to study the design of such arithmetic primitives, from basic operations such as the addition and the multiplication to more complex ones such as the division, the square root, cryptographic primitives, and even elementary functions. Arithmetic operators are relatively small hardware blocks at the scale of an integrated circuit, and are best described in a structural manner: a large operator is assembled from smaller ones, down to the granularity of the bit. This study requires knowledge of the hardware targets (ASICs, FPGAs), their metrics (area, delay, power), their constraints, and their specific language and tools. The input and output number systems are typically given (integer, fixed-point of floating-point), but internally, non-standard internal number systems may be successfully used.
Algebraic and elementary functions. Computer designers still have to implement the basic arithmetic functions for a medium-size precision. Addition and multiplication have been much studied but their performance may remain critical (silicon area or speed). Division and square root are less critical, however there is still room for improvement (e.g., the dedicated case where one of the inputs is constant). Research on new algorithms and architectures for elementary functions is also very active. Arénaire has a strong reputation in these domains and will keep contributing to their expansion. Thanks to past and recent efforts, the semantics of floating-point arithmetic has much improved. The adoption of the IEEE-754 standard for floating-point arithmetic has represented a key point for improving numerical reliability. Standardization is also related to properties of floating-point arithmetic (invariants that operators or sequences of operators may satisfy). Our goal is to establish and handle new properties in our developments (correct rounding, error bounds, etc.) and then to have those results integrated into the future computer arithmetic standards.
Formal proof and validation. For certifying the properties we identify, and the compliance of the numerical programs we develop with their specifications, we rely on formal proving. Proofs are checked using proof assistants such as Coq and PVS. In particular, this is made possible by a careful specification of the arithmetic operators that are involved. Further, an increasingly growing demand exists for certified numerical results, that is, for computing with known or controlled errors. We answer with the conception of modern and efficient error-measurement tools (roundoff errors, polynomial and power series approximation errors, etc.).
Arithmetics and algorithms. When conventional floating-point arithmetic does not suffice, we use other kinds of arithmetics. Especially in the matter of error bounds, we work on interval arithmetic libraries, including arbitrary precision intervals. Here a main domain of application is global optimization. Original algorithms dedicated to this type of arithmetic must be designed in order to get accurate solutions, or sometimes simply to avoid divergence (e.g., infinite intervals). We also investigate exact arithmetics for computing in algebraic domains such as finite fields, unlimited precision integers, and polynomials. A main objective is a better understanding of the influence of the output specification (approximate within a fixed interval, correctly rounded, exact, etc.) on the complexity estimates for the problems (e.g., linear algebra in mathematical computing).
Our work in Arénaire since its creation in 1998, and especially since 2002, provides us a strong expertise in computer arithmetic. This knowledge, together with the technology progress both in software and hardware, draws the evolution of our objectives towards the synthesis of validated algorithms .