See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/47735229

Handbook of Floating-Point Arithmetic

Book · January 2010

DOI: 10.1007/978-0-8176-4705-6 · Source: OAI

CITATIONS READS

421 8,292

9 authors, including:

Jean-Michel Muller Florent De Dinechin

Ecole normale supérieure de Lyon Institut National des Sciences Appliquées de Lyon
307 PUBLICATIONS   4,390 CITATIONS    136 PUBLICATIONS   2,973 CITATIONS   

SEE PROFILE SEE PROFILE

Claude-Pierre Jeannerod Vincent Lefèvre

National Institute for Research in Computer Science and Control National Institute for Research in Computer Science and Control
90 PUBLICATIONS   1,354 CITATIONS    78 PUBLICATIONS   2,023 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GNU MPFR View project

All content following this page was uploaded by Florent De Dinechin on 29 May 2014.

The user has requested enhancement of the downloaded file.

Jean-Michel Muller
Nicolas Brisebarre
Florent de Dinechin
Claude-Pierre Jeannerod
Vincent Lerevre
Guillaume Melquiond
Nathalie Revol
Damien Stehle
Serge Tones

Handbook of
Floating-Point
Arithmetic

Birkhäuser
Boston • Basel • Berlin
Contents

Preface xv
List of Figures xvii
List of Tables xxi

I Introduction, Basic Definitions, and Standards 1

1 Introduction 3
1.1 Some History 3
1.2 Desirable Properties 6
1.3 Some Strange Behaviors 7
1.3.1 Some famous bugs 7
1.3.2 Difficult problems 8

2 Definitions and Basic Notions 13

2.1 Floating-Point Numbers 13
2.2 Rounding 20
2.2.1 Rounding modes 20
2.2.2 Useful properties 22
2.2.3 Relative error due to rounding 23
2.3 Exceptions 25
2.4 Lost or Preserved Properties of the Arithmetic on the Real
Numbers 27
2.5 Note on the Choice of the Radix 29
2.5.1 Representation errors 29
2.5.2 A case for radix 10 30
2.6 Tools for Manipulating Floating-Point Errors 32
2.6.1 The ulp function 32
2.6.2 Errors in ulps and relative errors 37
2.6.3 An example: iterated products 37
2.6.4 Unit roundoff 39
2.7 Note on Radix Conversion 40

vi Contents

2.7.1 Conditions on the formats 40

2.7.2 Conversion algorithms 43
2.8 The Fused Multiply-Add (FMA) Instruction 51
2.9 Interval Arithmetic 51
2.9.1 Intervalls with floating-point bounds 52
2.9.2 Optimized rounding 52

3 Floating-Point Formats and Environment 55

3.1
The IEEE 754-1985 Standard 56
3.1.1 Formats specified by IEEE 754-1985 56
3.1.2 Little-endian, big-endian 60
3.1.3 Rounding modes specified by IEEE 754-1985 61
3.1.4 Operations specified by IEEE 754-1985 62
3.1.5 Exceptions specified by IEEE 754-1985 66
3.1.6 Special values 69
3.2 The IEEE 854-1987 Standard 70
3.2.1 Constraints internal to a format 70
3.2.2 Various formats and the constraints between them . . . 71
3.2.3 Conversions between floating-point numbers and
decimal strings 72
3.2.4 Rounding 73
3.2.5 Operations 73
3.2.6 Comparisons 74
3.2.7 Exceptions 74
3.3 The Need for a Revision 74
3.3.1 A typical problem: "double rounding" 75
3.3.2 Various ambiguities 77
3.4 The New IEEE 754-2008 Standard 79
3.4.1 Formats specified by the revised standard 80
3.4.2 Binary interchange format encodings 81
3.4.3 Decimal interchange format encodings 82
3.4.4 Larger formats 92
3.4.5 Extended and extendable precisions 92
3.4.6 Attributes 93
3.4.7 Operations specified by the standard 97
3.4.8 Comparisons 99
3.4.9 Conversions 99
3.4.10 Default exception handling 100
3.4.11 Recommended transcendental functions 103
3.5 Floating-Point Hardware in Current Processors 104
3.5.1 The common hardware denominator 104
3.5.2 Fused multiply-add 104
3.5.3 Extended precision 104
3.5.4 Rounding and precision control 105


Contents vii

3.5.5 SIMD instructions 106

3.5.6 Floating-point on x86 processors: SSE2 versus x87 . . 106
3.5.7 Decimal arithmetic 107
3.6 Floating-Point Hardware in Recent Graphics Processing Units 108
3.7 Relations with Programming Languages 109
3.7.1 The Language Independent Arithmetic (LIA) standard 109
3.7.2 Programming languages 110
3.8 Checking the Environment 110
3.8.1 MACHAR 111
3.8.2 Paranoia 111
3.8.3 UCBTest 115
3.8.4 TestFloat 116
3.8.5 IeeeCC754 116
3.8.6 Miscellaneous 116

II Cleverly Using Floating Point Arithmetic

- 117

4 Basic Properties and Algorithms 119

4.1 Testing the Computational Environment 119
4.1.1 Computing the radix 119
4.1.2 Computing the precision 121
4.2 Exact Operations 122
4.2.1 Exact addition 122
4.2.2 Exact multiplications and divisions 124
4.3 Accurate Computations of Sums of Two Numbers 125
4.3.1 The Fast2Sum algorithm 126
4.3.2 The 2Sum algorithm 129
4.3.3 If we do not use rounding to nearest 131
4.4 Computation of Products 132
4.4.1 Veltkamp splitting 132
4.4.2 Dekker's multiplication algorithm 135
4.5 Complex numbers 139
4.5.1 Various error bounds 140
4.5.2 Error bound for complex multiplication 141
4.5.3 Complex division 144
4.5.4 Complex square root 149

5 The Fused Multiply-Add Instruction 151

5.1 The 2Mu1tFMA Algorithm 152
5.2 Computation of Residuals of Division and Square Root 153
5.3 Newton—Raphson-Based Division with an FMA 155
5.3.1 Variants of the Newton—Raphson iteration 155
viii Contents

5.3.2 Using the Newton-Raphson iteration for correctly

rounded division 160
5.4 Newton-Raphson-Based Square Root with an FMA 167
5.4.1 The basic iterations 167
5.4.2 Using the Newton-Raphson iteration for correctly
rounded square roots 168
5.5 Multiplication by an Arbitrary-Precision Constant 171
5.5.1 Checking for a given constant C if Algorithm 5.2 will
always work 172
5.6 Evaluation of the Error of an FMA 175
5.7 Evaluation of Integer Powers 177

6 Enhanced Floating-Point Sums, Dot Products, and Polynomial

Values 181
6.1 Preliminaries 182
6.1.1 Floating-point arithmetic models 183
6.1.2 Notation for error analysis and classical error estimates 184
6.1.3 Properties for deriving running error bounds 187
6.2 Computing Validated Running Error Bounds 188
6.3 Computing Sums More Accurately 190
6.3.1 Reordering the operands, and a bit more 190
6.3.2 Compensated sums 192
6.3.3 Implementing a "long accumulator" 199
6.3.4 On the sum of three floating-point numbers 199
6.4 Compensated Dot Products 201
6.5 Compensated Polynomial Evaluation 203

7 Languages and Compilers 205

7.1 A Play with Many Actors 205
7.1.1 Floating-point evaluation in programming languages . 206
7.1.2 Processors, compilers, and operating systems 208
7.1.3 In the hands of the programmer 209
7.2 Floating Point in the C Language 209
7.2.1 Standard C99 headers and IEEE 754-1985 support . 209
7.2.2 Types 210
7.2.3 Expression evaluation 213
7.2.4 Code transformations 216
7.2.5 Enabling unsafe optimizations 217
7.2.6 Summary: a few horror stories 218
7.3 Floating-Point Arithmetic in the C++ Language 220
7.3.1 Semantics 220
7.3.2 Numeric limits 221
7.3.3 Overloaded functions 222
7.4 FORTRAN Floating Point in a Nutshell 223
Contents ix

7.4.1 Philosophy 223

7.4.2 IEEE 754 support in FORTRAN 226
7.5 Java Floating Point in a Nutshell 227
7.5.1 Philosophy 227
7.5.2 Types and classes 228
7.5.3 Infinities, NaNs, and signed zeros 230
7.5.4 Missing features 231
7.5.5 Reproducibility 232
7.5.6 The BigDecimal package 233
7.6 Conclusion 234

III Implementing Floating Point Operators

- 237
8 Algorithms for the Five Basic Operations 239
8.1 Overview of Basic Operation Implementation 239
8.2 Implementing IEEE 754-2008 Rounding 241
8.2.1 Rounding a nonzero finite value with unbounded
exponent range 241
8.2.2 Overflow 243
8.2.3 Underflow and subnormal results 244
8.2.4 The inexact excep tion 245
8.2.5 Rounding for actual operations 245
8.3 Floating-Point Addition and Subtraction 246
8.3.1 Decimal addition 249
8.3.2 Decimal addition using binary encoding 250
8.3.3 Subnormal inputs and outputs in binary addition . . 251
8.4 Floating-Point Multiplication 251
8.4.1 Normal case 252
8.4.2 Handling subnormal numbers in binary multiplication 252
8.4.3 Decimal specifics 253
8.5 Floating-Point Fused Multiply-Add 254
8.5.1 Case analysis for normal inputs 254
8.5.2 Handling subnormal inputs 258
8.5.3 Handling decimal cohorts 259
8.5.4 Overview of a binary FMA implementation 259
8.6 Floating-Point Division 262
8.6.1 Overview and special cases 262
8.6.2 Computing the significand quotient 263
8.6.3 Managing subnormal numbers 264
8.6.4 The inexact exception 265
8.6.5 Decimal specifics 265
8.7 Floating-Point Square Root 265
8.7.1 Overview and special cases 265
 Contents

8.7.2 Computing the significand square root 266

8.7.3 Managing subnormal numbers 267
8.7.4 The inexact exception 267
8.7.5 Decimal specifics 267

9 Hardware Implementation of Floating-Point Arithmetic 269

9.1 Introduction and Context 269
9.1.1 Processor internal formats 269
9.1.2 Hardware handling of subnormal numbers 270
9.1.3 Full-custom VLSI versus reconfigurable circuits 271
9.1.4 Hardware decimal arithmetic 272
9.1.5 Pipelining 273
9.2 The Primitives and Their Cost 274
9.2.1 Integer adders 274
9.2.2 Digit-by-integer multiplication in hardware 280
9.2.3 Using nonstandard representations of numbers 280
9.2.4 Binary integer multiplication 281
9.2.5 Decimal integer multiplication 283
9.2.6 Shifters 284
9.2.7 Leading-zero courtters 284
9.2.8 Tables and table-based methods for fixed-point
function approximation 286
9.3 Binary Floating-Point Addition 288
9.3.1 Overview 288
9.3.2 A first dual-path architecture 289
9.3.3 Leading-zero anticipation 291
9.3.4 Probing further on floating-point adders 295
9.4 Binary Floating-Point Multiplication 296
9.4.1 Basic architecture 296
9.4.2 FPGA implementation 296
9.4.3 VLSI implementation optimized for delay 298
9.4.4 Managing subnormals 301
9.5 Binary Fused Multiply-Add 302
9.5.1 Classic architecture 303
9.5.2 To probe further 305
9.6 Division 305
9.6.1 Digit-recurrence division 306
9.6.2 Decimal division 309
9.7 Conclusion: Beyond the FPU 309
9.7.1 Optimization in context of standard operators 310
9.7.2 Operation with a constant operand 311
9.7.3 Block floating point 313
9.7.4 Specific architectures for accumulation 313
9.7.5 Coarser-grain operators 317
Contents xi

9.8 Probing Further 320

10 Software Implementation of Floating-Point Arithmetic 321

10.1 Implementation Context 322
10.1.1 Standard encoding of binary floating-point data 322
10.1.2 Available integer operators 323
10.1.3 First examples 326
10.1.4 Design choices and optimizations 328
10.2 Binary Floating-Point Addition 329
10.2.1 Handling special values 330
10.2.2 Computing the sign of the result 332
10.2.3 Swapping the operands and computing the alignment
shift 333
10.2.4 Getting the correctly rounded result 335
10.3 Binary Floating-Point Multiplication 341
10.3.1 Handling special values 341
10.3.2 Sign and exponent computation 343
10.3.3 Overflow detection 345
10.3.4 Getting the correctly rounded result 346
10.4 Binary Floating-Point Division 349
10.4.1 Handling special values 350
10.4.2 Sign and exponent computation 351
10.4.3 Overflow detection 354
10.4.4 Getting the correctly rounded result 355
10.5 Binary Floating-Point Square Root 361
10.5.1 Handling special values 362
10.5.2 Exponent computation 364
10.5.3 Getting the correctly rounded result 365

IV Elementary Functions 373

11 Evaluating Floating-Point Elementary Functions 375
11.1 Basic Range Reduction Algorithms 379
11.1.1 Cody and Waite's reduction algorithm 379
11.1.2 Payne and Hanek's algorithm 381
11.2 Bounding the Relative Error of Range Reduction 382
11.3 More Sophisticated Range Reduction Algorithms 384
11.3.1 An example of range reduction for the exponential
function 386
11.3.2 An example of range reduction for the logarithm 387
11.4 Polynomial or Rational Approximations 388
11.4.1 L2 case 389
11.4.2 L" , or minimax case 390
xii Contents

11.4.3 "Truncated" approximations 392

11.5 Evaluating Polynomials 393
11.6 Correct Rounding of Elementary Functions to binary64 . . . 394
11.6.1 The Table Maker's Dilemma and Ziv's onion peeling
strategy 394
11.6.2 When the TMD is solved 395
11.6.3 Rounding test 396
11.6.4 Accurate second step 400
11.6.5 Error analysis and the accuracy/performance tradeoff 401
11.7 Computing Error Bounds 402
11.7.1 The point with efficient code 402
11.7.2 Example: a "double-double" polynomial evaluation . 403
12 Solving the Table Maker's Dilemma 405
12.1 Introduction 405
12.1.1 The Table Maker's Dilemma 406
12.1.2 Brief history of the TMD 410
12.1.3 Organization of the chapter 411
12.2 Preliminary Remarks on the Table Maker's Dilemma 412
12.2.1 Statistical arguments: what can be expected in practice 412
12.2.2 In some domains, there is no need to find worst cases . 416
12.2.3 Deducing the worst cases from other functions or
domains 419
12.3 The Table Maker's Dilemma for Algebraic Functions 420
12.3.1 Algebraic and transcendental numbers and functions 420
12.3.2 The elementary case of quotients 422
12.3.3 Around Liouville's theorem 424
12.3.4 Generating bad rounding cases for the square root
using Hensel 2-adic lifting 425
12.4 Solving the Table Maker's Dilemma for Arbitrary Functions 429
12.4.1 Lindemann's theorem: application to some
transcendental functions 429
12.4.2 A theorem of Nesterenko and Waldschmidt 430
12.4.3 A first method: tabulated differences 432
12.4.4 From the TMD to the distance between a grid and a
segment 434
12.4.5 Linear approximation: Lefevre's algorithm 436
12.4.6 The SLZ algorithm 443
12.4.7 Periodic functions on large arguments 448
12.5 Some Results 449
12.5.1 Worst cases for the exponential, logarithmic,
trigonometric, and hyperbolic functions 449
12.5.2 A special case: integer powers 458
12.6 Current Limits and Perspectives 458
 Contents xiii

V Extensions 461
13 Formalisms for Certifying Floating-Point Algorithms 463
13.1 Formalizing Floating-Point Arithmetic 463
13.1.1 Defining floating-point numbers 464
13.1.2 Simplifying the definition 466
13.1.3 Defining rounding operators 467
13.1.4 Extending the set of numbers 470
13.2 Formalisms for Certifying Algorithms by Hand 471
13.2.1 Hardware units 471
13.2.2 Low-level algorithms 472
13.2.3 Advanced algorithms 473
13.3 Automating Proofs 474
13.3.1 Computing on bounds 475
13.3.2 Counting digits 477
13.3.3 Manipulating expressions 479
13.3.4 Handling the relative error 483
13.4 Using Gappa 484
13.4.1 Toy implementation of sine 484
13.4.2 Integer division on Itanium 488
14 Extending the Precision 493
14.1 Double-Words, Triple-Words. 494
14.1.1 Double-word arithmetic 495
14.1.2 Static triple-word arithmetic 498
14.1.3 Quad-word arithmetic 500
14.2 Floating-Point Expansions 503
14.3 Floating-Point Numbers with Batched Additional Exponent 509
14.4 Large Precision Relying on Processor Integers 510
14.4.1 Using arbitrary-precision integer arithmetic for
arbitrary-precision floating-point arithmetic 512
14.4.2 A brief introduction to arbitrary-precision integer
arithmetic 513

VI Perspectives and Appendix 517

15 Conclusion and Perspectives 519

16 Appendix: Number Theory Tools for Floating-Point Arithmetic 521

16.1 Continued Fractions 521
16.2 The LLL Algorithm 524

Bibliography 529

Index 567

View publication stats