| |
| |
| Preface | |
| |
| |
| Acknowledgments | |
| |
| |
| Dedication | |
| |
| |
| |
| Introduction | |
| |
| |
| |
| GPUs as Parallel Computers | |
| |
| |
| |
| Architecture of a Modern GPU | |
| |
| |
| |
| Why More Speed or Parallelism? | |
| |
| |
| |
| Parallel Programming Languages and Models | |
| |
| |
| |
| Overarching Goals | |
| |
| |
| |
| Organization of the Book | |
| |
| |
| |
| History Of GPU Computing | |
| |
| |
| |
| Evolution of Graphics Pipelines | |
| |
| |
| |
| The Era of Fixed-Function Graphics Pipelines | |
| |
| |
| |
| Evolution of Programmable Real-Time Graphics | |
| |
| |
| |
| Unified Graphics and Computing Processors | |
| |
| |
| |
| GPGPU: An Intermediate Step | |
| |
| |
| |
| GPU Computing | |
| |
| |
| |
| Scalable GPUs | |
| |
| |
| |
| Recent Developments | |
| |
| |
| |
| Future Trends | |
| |
| |
| |
| Introduction To Cuda | |
| |
| |
| |
| Data Parallelism | |
| |
| |
| |
| Cuda Program Structure | |
| |
| |
| |
| A Matrix-Matrix Multiplication Example | |
| |
| |
| |
| Device Memories and Data Transfer | |
| |
| |
| |
| Kernel Functions and Threading | |
| |
| |
| |
| Summary | |
| |
| |
| |
| Function declarations | |
| |
| |
| |
| Kernel launch | |
| |
| |
| |
| Predefined variables | |
| |
| |
| |
| Runtime API | |
| |
| |
| |
| Cuda Threads | |
| |
| |
| |
| Cuda Thread Organization | |
| |
| |
| |
| Using blockIdx and threadIdx | |
| |
| |
| |
| Synchronization and Transparent Scalability | |
| |
| |
| |
| Thread Assignment | |
| |
| |
| |
| Thread Scheduling and Latency Tolerance | |
| |
| |
| |
| Summary | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Cuda� Memories | |
| |
| |
| |
| Importance of Memory Access Efficiency | |
| |
| |
| |
| CUDA Device Memory Types | |
| |
| |
| |
| A Strategy for Reducing Global Memory Traffic | |
| |
| |
| |
| Memory as a Limiting Factor to Parallelism | |
| |
| |
| |
| Summary | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Performance On Siderations | |
| |
| |
| |
| More on Thread Execution | |
| |
| |
| |
| Global Memory Bandwidth | |
| |
| |
| |
| Dynamic Partitioning of SM Resources | |
| |
| |
| |
| Data Prefetching | |
| |
| |
| |
| Instruction Mix | |
| |
| |
| |
| Thread Granularity | |
| |
| |
| |
| Measured Performance and Summary | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Floating Point Considerations | |
| |
| |
| |
| Floating-Point Format | |
| |
| |
| |
| Normalized Representation of M | |
| |
| |
| |
| Excess Encoding of E | |
| |
| |
| |
| Representable Numbers | |
| |
| |
| |
| Special Bit Patterns and Precision | |
| |
| |
| |
| Arithmetic Accuracy and Rounding | |
| |
| |
| |
| Algorithm Considerations | |
| |
| |
| |
| Summary | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Application Case Study: Advanced MRI Reconstruction | |
| |
| |
| |
| Application Background | |
| |
| |
| |
| Iterative Reconstruction | |
| |
| |
| |
| Computing F<sup>H</sup>d | |
| |
| |
| |
| Determine the Kernel Parallelism Structure | |
| |
| |
| |
| Getting Around the Memory Bandwidth Limitation | |
| |
| |
| |
| Using Hardware Trigonometry Functions | |
| |
| |
| |
| Experimental Performance Tuning | |
| |
| |
| |
| Final Evaluation | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Application Case Study: Molecular Visualization and Analysis | |
| |
| |
| |
| Application Background | |
| |
| |
| |
| A Simple Kernel Implementation | |
| |
| |
| |
| Instruction Execution Efficiency | |
| |
| |
| |
| Memory Coalescing | |
| |
| |
| |
| Additional Performance Comparisons | |
| |
| |
| |
| Using Multiple GPUs | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Parallel Programming and Computational Thinking | |
| |
| |
| |
| Goals of Parallel Programming | |
| |
| |
| |
| Problem Decomposition | |
| |
| |
| |
| Algorithm Selection | |
| |
| |
| |
| Computational Thinking | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| A Brief Introduction To Opencl� | |
| |
| |
| |
| Background | |
| |
| |
| |
| Data Parallelism Model | |
| |
| |
| |
| Device Architecture | |
| |
| |
| |
| Kernel Functions | |
| |
| |
| |
| Device Management and Kernel Launch | |
| |
| |
| |
| Electrostatic Potential Map in OpenCL | |
| |
| |
| |
| Summary | |
| |
| |
| |
| Exercises | |
| |
| |
| |
| Conclusion And Future Outlook | |
| |
| |
| |
| Goals Revisited | |
| |
| |
| |
| Memory Architecture Evolution | |
| |
| |
| |
| Large Virtual and Physical Address Spaces | |
| |
| |
| |
| Unified Device Memory Space | |
| |
| |
| |
| Configurable Caching and Scratch Pad | |
| |
| |
| |
| Enhanced Atomic Operations | |
| |
| |
| |
| Enhanced Global Memory Access | |
| |
| |
| |
| Kernel Execution Control Evolution | |
| |
| |
| |
| Function Calls within Kernel Functions | |
| |
| |
| |
| Exception Handling in Kernel Functions | |
| |
| |
| |
| Simultaneous Execution of Multiple Kernels | |
| |
| |
| |
| Interruptible Kernels | |
| |
| |
| |
| Core Performance | |
| |
| |
| |
| Double-Precision Speed | |
| |
| |
| |
| Better Control Flow Efficiency | |
| |
| |
| |
| Programming Environment | |
| |
| |
| |
| A Bright Outlook | |
| |
| |
| |
| Matrix Multiplication Host-Only Version Source Code | |
| |
| |
| |
| matrixmul . cu | |
| |
| |
| |
| matrixmul_gold.cpp | |
| |
| |
| |
| matrixmul . h | |
| |
| |
| |
| assist.h | |
| |
| |
| |
| Expected Output | |
| |
| |
| |
| GPU Compute Capabilities | |
| |
| |
| |
| GPU Compute Capability Tables | |
| |
| |
| |
| Memory Coalescing Variations | |
| |
| |
| Index | |