A multi-precision function interpolator, which is fitted in with the IEEE-754 single precision floating point standard, is proposed in this paper. It provides logarithms, exponentials, reciprocal and square root reciprocal operations. Each operation is able to dynamically select four different precision modes in demand. The hardware architecture is designed with fully pipeline in order to comply with hardware architectures of general digital signal processors (DSPs) and graphics processors (GPUs). When considering the usefulness of each precision mode, it is designed to minimize the error among various modes as far as possible in the beginning. According to the precision from high to low, function interpolator can provide 23, 18, 13 and 8-bit accuracy respectively in spite of the rounding effect. This function interpolator is designed based on the look-up table method. It can get the approximation value of target function through the calculation of quadratic polynomial. The coefficient of quadratic polynomial is obtained by piecewise minimax approximation. Before implementing the hardware, we use the Maple algebra software to generate the quadratic polynomial coefficients of aforementioned four operations, and estimate whether these coefficients can meet IEEE-754 single precision floating point standard. In addition, we take the exhaustive search to check the results generated by our implementation to make sure that it can meet the requirements for various operations and precision modes. When performing one of the above four operations, only the tables of the operation are used to obtain the quadratic polynomial coefficient. Therefore, we can take the advantage of the tri-state buffer as a switch to reduce dynamic power consumption of tables for the other three operations. In addition, when performing lower precision modes, we can turn off a part of hardwares, which are used to calculate the quadratic polynomial, to save the power consumption more effectively. By providing multi-precision hardware, we hope users or developers, those who use the battery device, can choose a lower precision mode within the permissible error range to extend the battery life.

<p>This work presents the design and implementation of a decimation filter for a three bits sigma delta analog to digital converter. The input is audio with a oversampling ratio of 32. Filter optimization and tradeoffs concerning the design is described. The filter is a multistage filter consisting of two cascaded FIR filters. The arithmetic unit is a multi-precision unit that can handle three or 24 bits MAC operations. The designed decimation filter is synthesized on standard cells of a 0.13 μm CMOS library.</p>

This work presents the design and implementation of a decimation filter for a three bits sigma delta analog to digital converter. The input is audio with a oversampling ratio of 32. Filter optimization and tradeoffs concerning the design is described. The filter is a multistage filter consisting of two cascaded FIR filters. The arithmetic unit is a multi-precision unit that can handle three or 24 bits MAC operations. The designed decimation filter is synthesized on standard cells of a 0.13 μm CMOS library.

Context. The cryptographically secure pseudo-random number generator Blum Blum Shub (BBS) is a simple algorithm with a strong security proof, however it requires very large numbers to be secure, which makes it computationally heavy. The Graphics Processing Unit (GPU) is a common vector processor originally dedicated to computer-game graphics, but has since been adapted to perform general-purpose computing. The GPU has a large potential for fast general-purpose parallel computing but due to its architecture it is difficult to adapt certain algorithms to utilise the full computational power of the GPU. Objectives. The objective of this thesis was to investigate if an implementation of the BBS pseudo-random number generator algorithm on the GPU would be faster than a CPU implementation. Methods. In this thesis, we modelled the performance of a multi-precision number system with different data types; to decide which data type should be used for a multi-precision number system implementation on the GPU. The multi-precision number system design was based on a positional number system. Because multi-precision numbers were used, conventional methods for arithmetic were not efficient or practical. Therefore, addition was performed by using Lazy Addition that allows larger carry values in order to limit the amount of carry propagation required to perform addition. Carry propagation was done by using a technique derived from a Kogge-Stone carry look-ahead adder. Single-precision multiplication was done using Dekker splits and multi-precision modular multiplication used Montgomery multiplication. Results. Our results showed that using the floating-point data type would yield greater performance for a multi-precision number system on the GPU compared to using the integer data type. The performance results from our GPU bound BBS implementation was about 4 times slower than a CPU version implemented with the GNU Multiple Precision Arithmetic Library (GMP). Conclusions. The conclusion made from this thesis, is that our GPU bound BBS implementation, is not a suitable alternative or replacement for the CPU bound implementation.

Nowadays high-precision computations are still more desired. Either for simulation on a level of atoms where every digit is important and inaccurary in computation can cause invalid result or numerical approximations in partial differential equations solving where a small deviation causes a result to be useless. The computations are carried over data types with precision of order hundred to thousand digits, or even more. This creates pressure on time complexity of problem solving and so it is essential to find very efficient methods for computation. Every complex physical problem is usually described by a system of equations frequently containing elementary functions like sinus, cosines or exponentials. The aim of the work is to design and implement methods that for a given precision, arbitrary elementary function and a point compute its value in the most efficent way. The core of the work is an application of methods based on AGM (arithmetic-geometric mean) with a time complexity of order $O(M(n)\log_2{n})$ 9(expresed for multiplication $M(n)$). The complexity can not be improved. There are many libraries supporting multi-precision atithmetic, one of which is GMP and is about to be used for efficent method implementation. In the end all implemented methods are compared with existing ones.

De nombreux problèmes de calcul numérique demandent parfois à effectuer des calculs très précis. L'étude desystèmes dynamiques chaotiques fournit des exemples très connus: la stabilité du système solaire ou l’itération à longterme de l'attracteur de Lorenz qui constitue un des premiers modèles de prédiction de l'évolution météorologique. Ons'intéresse aussi aux problèmes d'optimisation semi-définie positive mal-posés qui apparaissent dans la chimie oul'informatique quantique.Pour tenter de résoudre ces problèmes avec des ordinateurs, chaque opération arithmétique de base (addition,multiplication, division, racine carrée) demande une plus grande précision que celle offerte par les systèmes usuels(binary32 and binary64). Il existe des logiciels «multi-précision» qui permettent de manipuler des nombres avec unetrès grande précision, mais leur généralité (ils sont capables de manipuler des nombres de millions de chiffres) empêched’atteindre de hautes performances. L’objectif majeur de cette thèse a été de développer un nouveau logiciel à la foissuffisamment précis, rapide et sûr : on calcule avec quelques dizaines de chiffres (quelques centaines de bits) deprécision, sur des architectures hautement parallèles comme les processeurs graphiques et on démontre des bornesd'erreur afin d'être capables d’obtenir des résultats certains. / Many numerical problems require some very accurate computations. Examples can be found in the field ofdynamical systems, like the long-term stability of the solar system or the long-term iteration of the Lorenz attractor thatis one of the first models used for meteorological predictions. We are also interested in ill-posed semi-definite positiveoptimization problems that appear in quantum chemistry or quantum information.In order to tackle these problems using computers, every basic arithmetic operation (addition, multiplication,division, square root) requires more precision than the ones offered by common processors (binary32 and binary64).There exist multiple-precision libraries that allow the manipulation of very high precision numbers, but their generality(they are able to handle numbers with millions of digits) is quite a heavy alternative when high performance is needed.The major objective of this thesis was to design and develop a new arithmetic library that offers sufficient precision, isfast and also certified. We offer accuracy up to a few tens of digits (a few hundred bits) on both common CPU processorsand on highly parallel architectures, such as graphical cards (GPUs). We ensure the results obtained by providing thealgorithms with correctness and error bound proofs.

This thesis considers new methods that exploit recent developments in computer technology to address three extant problems in the area of Finance and Econometrics. The problem of Asian option pricing has endured for the last two decades in spite of many attempts to find a robust solution across all parameter values. All recently proposed methods are shown to fail when computations are conducted using standard machine precision because as more and more accuracy is forced upon the problem, round-off error begins to propagate. Using recent methods from numerical analysis based on multi-precision arithmetic, we show using the Mathematica platform that all extant methods have efficacy when computations use sufficient arithmetic precision. This creates the proper framework to compare and contrast the methods based on criteria such as computational speed for a given accuracy. Numerical methods based on a deformation of the Bromwich contour in the Geman-Yor Laplace transform are found to perform best provided the normalized strike price is above a given threshold; otherwise methods based on Euler approximation are preferred. The same methods are applied in two other contexts: the simulation of stable distributions and the computation of unit root densities in Econometrics. The stable densities are all nested in a general function called a Fox H function. The same computational difficulties as above apply when using only double-precision arithmetic but are again solved using higher arithmetic precision. We also consider simulating the densities of infinitely divisible distributions associated with hyperbolic functions. Finally, our methods are applied to unit root densities. Focusing on the two fundamental densities, we show our methods perform favorably against the extant methods of Monte Carlo simulation, the Imhof algorithm and some analytical expressions derived principally by Abadir. Using Mathematica, the main two-dimensional Laplace transform in this context is reduced to a one-dimensional problem.