Intel SWBC2; 0-9764832-1-1; Production Proof; 02-07-06 #1
The Software
Optimization Cookbook
High-Performance Recipes for IA-32
Platforms
Second Edition
Richard Gerber
Aart J.C. Bik
Kevin B. Smith
Xinmin Tian
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Copyright 2006 Intel Corporation. All rights reserved.
ISBN 0-9764832-1-1
No part of this publication may be reproduced, stored in a retrieval system or transmitted in any
form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise,
except as permitted under Sections 107 or 108 of the 1976 United States Copyright Act, without
either the prior written permission of the Publisher, or authorization through payment of the
appropriate per-copy fee to the Copyright Clearance Center, 222 Rosewood Drive, Danvers,
MA 01923, (978) 750-8400, fax (978) 750-4744. Requests to the Publisher for permission
should be addressed to the Publisher, Intel Press, Intel Corporation, 2111 NE 25 Avenue, JF3330, Hillsboro, OR 97124-5961. E-Mail: intelpress@[Link].
This publication is designed to provide accurate and authoritative information in regard to the
subject matter covered. It is sold with the understanding that the publisher is not engaged in
professional services. If professional advice or other expert assistance is required, the services
of a competent professional person should be sought.
Intel Corporation may have patents or pending patent applications, trademarks, copyrights, or
other intellectual property rights that relate to the presented subject matter. The furnishing of
documents and other materials and information does not provide any license, express or
implied, by estoppel or otherwise, to any such patents, trademarks, copyrights, or other
intellectual property rights.
Intel may make changes to specifications, product descriptions, and plans at any time, without
notice.
Intel products are not intended for use in medical, life saving, life sustaining, critical control or
safety systems, or in nuclear facility applications.
Intel, Intel logo, Intel386, Intel486, Intel NetBurst, Intel XScale, Itanium, MMX, Pentium, VTune, and
Xeon are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United
States and other countries.
Other names and brands may be claimed as the property of others.
This book is printed on acid-free paper.
Publisher: Richard Bowles
Editor: David B. Spencer
Content Architect: Stuart Goldstein
Text Design & Composition: IMS Pubs
Graphic Art: IMS Pubs (illustrations), Ted Cyrek (cover)
Library of Congress Cataloging in Publication Data:
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1
First printing, March 2006
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Contents
Preface
xi
Part I
Performance Tools and Concepts
Chapter 1
Introduction
Software Optimization 4
Software Optimization Pitfalls 4
The Software Optimization Process 6
Chapter 2
The Benchmark
The Attributes of the Benchmark 10
Repeatable (Required) 10
Representative (Required) 11
Easy to Run (Required) 11
Verifiable (Required) 11
Measure Elapsed Time (Optional) 12
Complete Coverage (Situation-dependent) 12
Precision (Situation-dependent) 12
Benchmark Examples 13
iii
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
iv
The Software Optimization Cookbook
Chapter 3
Performance Tools
19
Timing Mechanisms 19
Optimizing Compilers 21
Using the Intel C++ and Fortran Compilers 22
Optimizing for Specific Processors 23
Writing Functions Specific to One Processor 25
Other Compiler Optimizations 26
Types of Software Profilers 27
Performance Monitor 28
VTune Performance Analyzer 29
Sampling 29
Call Graph Profiling 31
Intel Compiler Codecov Profiler 33
Microsoft Visual C++ Profiler 34
Sampling Versus Instrumentation 36
Trial and Error, Common Sense, and Patience 37
Chapter 4
The Hotspot
41
What Causes Hotspots and Cold-spots? 43
More Than Just Time 44
Uniform Execution and No Hotspots 46
Chapter 5
Processor Architecture
51
Functional Blocks 52
Two Cheeseburgers Please! 54
Instruction Fetch and Decode 57
Instruction Execution 59
Retirement 62
Registers and Memory 63
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Contents
Part II
Chapter 6
Performance Issues
Algorithms
67
69
Computational Complexity 69
Choice of Instructions 70
Data Dependencies and Instruction Parallelism 75
Memory Requirements 79
Parallel Algorithms 79
Generality of Algorithms 80
Detecting Algorithm Issues 81
Chapter 7
Branching
89
Finding the Critical Mis-predicted Branches 92
Step 1: Find the Mis-predicted Branches 92
Step 2: Find the Time-consuming Hotspots 93
Step 3: Determine the Percentage of Mis-predicted Branches 94
Final Sanity Check 95
The Different Types of Branches 96
Making Branches More Predictable 99
Removing Branches with CMOV 100
Removing Branches with Masks 101
Removing Branches with Min/Max Instructions 103
Removing Branches by Doing Extra Work 103
Chapter 8
Memory
107
Memory Overview 108
Main Memory and Virtual Memory 109
Processor Caches 109
Cache Details 112
Hardware Prefetch 114
Software Prefetch 114
Writing Data Without the Cache: Non-temporal Writes 115
Issues Affecting Memory Performance 117
Cache Compulsory Loads 117
Cache Capacity Loads 118
Cache Conflict Loads 119
Cache Efficiency 120
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
vi
The Software Optimization Cookbook
Store Forwarding 120
Data Alignment 122
Compilers and Data Alignment 123
Software Prefetch 124
Detecting Memory Issues 125
Finding Page Misses 126
Finding Store Forwarding Problems 130
Finding L1 Cache Misses 131
Understanding Potential Improvement 133
Fixing Memory Problems 135
Chapter 9
Loops
143
Data Dependences 144
Loop Distribution and Fusion 146
Loop Peeling 149
Loop Unrolling and Re-rolling 149
Loop Interchanging 153
Loop Invariant Computations 155
Loop Invariant Branches 156
Loop Invariant Results 157
Chapter 10 Slow Operations
159
Slow Instructions 159
Lookup Tables 161
System Calls 164
System Idle Process 168
Chapter 11 Floating Point
175
Numeric Exceptions 176
Flush-to-Zero and Denormals are Zero 179
Precision 180
Packed and Scalar Mode 184
Float-to-Integer Conversions, Rounding 185
Floor and Ceil Functions 186
Floating-point Manipulation Tricks 186
FP-to-Integer Conversion 186
Square Root 187
Reciprocal Square Root 187
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Contents
Chapter 12 SIMD Technology
191
An Introduction to SIMD Technology 192
MMX Technology 192
Streaming SIMD Extensions 193
Using SIMD Technology 194
Automatic Vectorization 195
C++ Class Libraries 196
Intrinsics 197
Inline Assembly Language 198
Advantages and Disadvantages of the Four Methods 199
SIMD Technology Considerations 200
Determining Where to Use SIMD Technology 201
Memory Alignment 201
Data Layout 203
Selecting an Appropriate Packed Data Type 205
Compatibility of SIMD and x87 FPU Calculations 207
Chapter 13 Automatic Vectorization
211
Compiler Switches for Vectorization 211
Commonly Used Compiler Switches 211
Compiler Switches Example 214
Compiler Hints for Vectorization 215
Commonly Used Compiler Hints 215
Compiler Hints Examples 220
Vectorization Guidelines 222
Design and Implementation Considerations 222
Usage of Vectorization Diagnostics 224
Minimize Potential Aliasing and Side Effects 228
Programming Style 232
Target Architectures 233
Chapter 14 Processor-Specific Optimizations
32-bit Intel Architectures 239
The Pentium M Processor 242
L1 Instruction Cache 243
Instruction Decoding 243
Instruction Latencies 244
239
vii
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
viii
The Software Optimization Cookbook
Instruction Set 245
Floating-Point Control Register 246
MXCSR Status Register 246
L1 Data Cache 247
Memory Prefetch 247
Processor Events 248
Partial Register Stalls 248
Partial Flag Stall 249
Pause Instruction 250
Chapter 15 Introduction to Multiprocessing
253
Parallel Programming 254
Thread Management 256
High-level Threading with OpenMP 256
Low-level Threading 259
Threading Goals 260
Threading Issues 261
Intel Compilers and Threading Tools 265
Chapter 16 Multithreading with OpenMP
269
OpenMP Key Elements 269
Multithreading Execution Model 274
OpenMP Memory Model 276
Limitations of OpenMP 280
Compiling OpenMP Programs 281
Automatic Parallelization 283
Multithreading Guidelines 286
Chapter 17 Taskqueuing and Advanced Topics
291
TaskqueuingIntel Extension to OpenMP 291
Taskqueuing Execution Model 291
Taskq and Task Constructs 294
Threading the N-Queens Program: A Case Study 297
Thread-Level Pipeline Parallelism 302
Exploiting Nested Parallelism 306
Multi-level Parallelism 311
Thread Affinity Insight 313
Understanding Loop Scheduling 315
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Contents
Part III
Design and Application Optimization
321
Chapter 18 Case Study: Threading a Video Codec
323
Initial Performance of the H.264 Encoder 324
Parallelization of the H.264 Encoder 325
Task and Data Domain Decomposition 325
Slice-Level Parallelism 327
Frame-Level Parallelism 328
Implementation Using Two Slice Queues 329
Implementation Using Task Queuing Model 331
Performance 333
Tradeoff Between Increased Speed and Effective Compression
333
Performance on Multiprocessor with HT Technology 335
Understanding the Performance 336
Multithreading Overhead 340
Further Performance Tuning 341
Summary of Threading 342
Chapter 19 Designing for Performance
345
Data Movement 346
Memory and Parallelism 347
Performance Experiments for Design 348
Algorithms 349
Chapter 20 Putting it Together: Basic Optimizations
Picking the Low-Hanging Fruit 353
The Application 354
Follow Along 356
The Benchmark 356
Interpret the Benchmark Results 357
Improving FloattoLong Conversions 359
Parallelizing the Algorithm 360
Automatic Vectorization to the Rescue 360
Instruction Level Parallelism with the Intrinsics 362
Summary of Optimizations 363
353
ix
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
The Software Optimization Cookbook
Chapter 21 Putting It Together: The Last Ten Percent
Speed of Light 365
Greater SIMD Efficiency 367
One Final Optimization 370
Summary of Optimizations 371
References
Index
379
373
365
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Preface
The first edition of The Software Optimization Cookbook continues to
be one of the most popular books offered by Intel Press. Feedback
received from readers indicates that the book fills a gap between
introductory textbooks that deal with program optimizations in general
and advanced manuals that deal with all aspects of the Intel architecture
in particular. The introduction of the Intel Extended Memory 64
Technology (Intel EM64T) and multi-core processing together with the
growing popularity of the Hyper-Threading Technology, OpenMP , and
multimedia extensions have outdated the first edition, however. The
continuing demand for an intermediate level introduction to these topics
has prompted Intel Press to ask three additional Intel experts to team up
with the original author to provide an expanded and updated second
edition of the book.
The Software Optimization Cookbook, Second Edition, provides
updated recipes for high-performance applications on Intel platforms.
Through simple explanations and examples, the authors show you how
to address performance issues with algorithms, memory access, branch
prediction, automatic vectorization, SIMD instructions, multiple threads,
and floating-point calculations. Software developers learn how to take
advantage of Intel EM64T, multi-core processing, Hyper-Threading
Technology, OpenMP, and multimedia extensions. This book guides you
through the growing collection of software tools, compiler switches, and
coding optimizations, showing you efficient ways to improve the
performance of software applications for Intel platforms. Software
developers who want to understand the latest techniques for delivering
more performance and to fine-tune their coding skills will benefit from
this book.
xi
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
xii
The Software Optimization Cookbook
Acknowledgements
The authors would like to thank everyone who has contributed in one
way or the other to this book or to the Intel compilers and performance
analysis tools. In particular, the authors are grateful for the efforts of the
following people: Zia Ansari, Pete Baker, Mitch Bodart, Christopher Bord,
Mark Buxton, Ryan Carlson, Yen-Kuang Chen, Joshua Chia, Martyn
Corden, Robert Cox, William E. Damon III, Max J. Domeika, Milind
Girkar, Koby Gottlieb, Grant Haab, Mohammad Haghighat, Jay
Hoeflinger, John Holm, Alexander Isaev, Michael Julier, Wei Li,
Christopher Lishka, Diana King, Knud Kirkegaard, David Kreitzer, Tim
Mattson, Eric Moore, Dan Macri, Andrey Naraikin, Kannan Narayanan,
Clark Nelson, John Ng, Paul Peterson, Michael Ross, Hideki Saito, Sanjiv
Shah, Ernesto Su, Sara Sarmiento, William Savage, Dale Schouten, David
Sehr, Ronak Singhal, Kevin J. Smith, Stacey Smith, Craig Stoller, Bob
Valentine, and Ronen Zohar.
A special thanks to our evaluators, Tim Carver formerly of Intel, Walt
Dixon of GE Global Research, Lars Petter Endresen at Scandpower
Petroleum Technology AS, James H. Hill of the United States Postal
Service, Brian Kennedy at Jeppesen/Boeing, Jitendra Maheshwari at
Zeesoft Inc., Kevin Ruland at the University of Kansas Natural History
Museum, and Robert van Engelen at Florida State University, for
providing valuable feedback on preliminary drafts of the book.
The people of Intel Press have been very helpful during the writing
process. In particular, we would like to acknowledge the help of our
editor David Spencer and content architect Stuart Goldstein for their kind
and professional guidance during all stages of the publishing process.
Intel SWBC2; 0-9764832-1-1; Production Proof; 01-30-06 #1
Preface
xiii
Readers Can Help, Too
To prepare this second edition, the authoring team carefully reviewed
the material in the first edition and updated most of it. In particular,
Richard Gerber updated the putting-it-together material. Aart Bik updated
Chapters 2, 9, and 12, and he added Chapter 13 on automatic
vectorization. Kevin Smith updated Chapter 3 through 8, 10, 11, and 14.
Xinmin Tian updated Chapter 15 and added Chapters 16, 17, and 18 on
OpenMP parallelization.
Having done so, we are anxious to keep the book up-to-date. We
would like to hear your questions and requests for clarification. We post
contact information, errata, and additional materials on this books Web
site.
Richard Gerber
Aart J. C. Bik
Kevin B. Smith
Xinmin Tian
