Sunita Chandrasekaran is assistant professor in the Computer and Information Sciences Department at the University of Delaware. Her research interests include exploring the suitability of high-level programming models and runtime systems for HPC and embedded platforms, and migrating scientific applications to heterogeneous computing systems. Dr. Chandrasekaran was a post-doctoral fellow at the University of Houston and holds a Ph.D. from Nanyang Technological University, Singapore. She is a member of OpenACC, OpenMP, MCA and SPEC HPG. She has served on the program committees of various conferences and workshops including SC, ISC, ICPP, CCGrid, Cluster, and PACT, and has co-chaired parallel programming workshops co-located with SC, ISC, IPDPS, and SIAM.

Guido Juckeland is head of the Computational Science Group, Department for Information Services and Computing, Helmholtz-Zentrum Dresden-Rossendorf, and coordinates the work of the GPU Center of Excellence at Dresden. He and also represents HZDR at the SPEC High Performance Group and OpenACC committee. He received his Ph.D. from Technische Universität Dresden for his work on performance analysis for hardware accelerators. He was a Gordon Bell Award Finalist in 2013. Previously he worked as the IT-architect and post-doctoral researcher for the Center for Information Services and High Performance Computing (ZIH) at TU Dresden, Germany. He has served on the program committees of various conferences and workshops, including ISC, EuroPar, CCGrid, ASHES, P^3MA, PMBS, WACCPD, and PACT, and has co-chaired parallel programming workshops co-located with SC.

Foreword xv

Preface xxi

Acknowledgments xxiii

About the Contributors xxv

Chapter 1: OpenACC in a Nutshell 1

1.1 OpenACC Syntax 3

1.2 Compute Constructs 6

1.3 The Data Environment 11

1.4 Summary 15

1.5 Exercises 15

Chapter 2: Loop-Level Parallelism 17

2.1 Kernels Versus Parallel Loops 18

2.2 Three Levels of Parallelism 21

2.3 Other Loop Constructs 24

2.4 Summary 30

2.5 Exercises 31

Chapter 3: Programming Tools for OpenACC 33

3.1 Common Characteristics of Architectures 34

3.2 Compiling OpenACC Code 35

3.3 Performance Analysis of OpenACC Applications 36

3.4 Identifying Bugs in OpenACC Programs 51

3.5 Summary 53

3.6 Exercises 54

Chapter 4: Using OpenACC for Your First Program 59

4.1 Case Study 59

4.2 Creating a Naive Parallel Version 68

4.3 Performance of OpenACC Programs 71

4.4 An Optimized Parallel Version 73

4.5 Summary 78

4.6 Exercises 79

Chapter 5: Compiling OpenACC 81

5.1 The Challenges of Parallelism 82

5.2 Restructuring Compilers 88

5.3 Compiling OpenACC 92

5.4 Summary 97

5.5 Exercises 97

Chapter 6: Best Programming Practices 101

6.1 General Guidelines 102

6.2 Maximize On-Device Compute 105

6.3 Optimize Data Locality 108

6.4 A Representative Example 112

6.5 Summary 118

6.6 Exercises 119

Chapter 7: OpenACC and Performance Portability 121

7.1 Challenges 121

7.2 Target Architectures 123

7.3 OpenACC for Performance Portability 124

7.4 Code Refactoring for Performance Portability126

7.5 Summary 132

7.6 Exercises133

Chapter 8: Additional Approaches to Parallel Programming 135

8.1 Programming Models135

8.2 Programming Model Components142

8.3 A Case Study 155

8.4 Summary170

8.5 Exercises170

Chapter 9: OpenACC and Interoperability 173

9.1 Calling Native Device Code from OpenACC 174

9.2 Calling OpenACC from Native Device Code 181

9.3 Advanced Interoperability Topics 182

9.4 Summary185

9.5 Exercises185

Chapter 10: Advanced OpenACC 187

10.1 Asynchronous Operations 187

10.2 Multidevice Programming 204

10.3 Summary 213

10.4 Exercises 213

Chapter 11: Innovative Research Ideas Using OpenACC, Part I 215

11.1 Sunway OpenACC 215

11.2 Compiler Transformation of Nested Loops for Accelerators 224

Chapter 12: Innovative Research Ideas Using OpenACC, Part II 237

12.1 A Framework for Directive-Based High-Performance Reconfigurable Computing 237

12.2 Programming Accelerated Clusters Using XcalableACC 253

Index 269