Автор курса - известный специалист по компьютерной
арифметике Dr. Florent de Dinechin, университет Лиона, Франция.
Dr. Florent de Dinechin – сотрудник института, в котором был разработан стандарт
представления чисел с плавающий запятой IEEE-754, один из авторов книги по
вещественной арифметике «Handbook of Floating-Point Arithmetic» http://www.amazon.com/gp/product/081764704X.
Цель
курса: введение в проблематику решения задач вычислительного
характера в условиях ограничений представления чисел на
ЭВМ.
Предварительная программа курса:
Lecture 1: Introduction
Number
representation
Introduction to floating-point
Lecture 2: The floating-point
environment
The dark ages of anarchy
Floating-point as it should be (The IEEE754-1985 standard)
Floating-point as it is in current processors
Floating-point as it is in current languages
Floating-point as it is in current operating systems
Lecture 3: Science around binary
floating-point
Representation tips and
tricks
Rounding to an integer
Sterbenz
Lemma
Exact addition
Exact
multiplication
Application: evaluating a polynomial to 60
bits of accuracy
Compensated sums
Compensated Horner evaluation
Lecture 4: The 2008 revision of the IEEE-754
standard
The new binary formats
The
FMA
The decimal fiasco
Elementary
functions
Sums and sums of products
Lecture 5: Elementary functions and their
evaluation
Examples of argument reduction
Polynomial
approximation
Evaluation of polynomials
Overall error
analysis
Ziv's technique for correct rounding
Lecture 6: Variations on the exponential
Baseline double-precision version
Using double-extended
Single-precision versions
Vector versions
Correctly rounded
versions
Correctly rounded vector versions
GPU
version
FPGA version
Lecture 7: Automating error analysis
Formal definitions of floating-point numbers
Interval
arithmetic
Taylor and Chebychev models
Computing infinite
norms
Managing rounding errors with high-level formal proof tools
Lecture 8: Automating code generation for function
evaluation
Machine-efficient polynomial approximation
Constrained polynomial approximation
Optimizing polynomial coefficient
sizes
Overlapping double-double and triple-double arithmetic
Lecture 9: Variations around the trigonometric
functions
Radian versus degrees
Cody and
Waite argument reduction
Paine and Hainek argument
reduction
Inverse trigonometric functions
Lecture 10: Solving the table maker's
dilemma
Lefèvre algorithms
Analysis of some surprising
results
SLZ algorithm
Brute-force approaches
Презентации к лекциям
Видеозаписи
лекций