Introduction to GPU Programming Training Course

Introduction

• Defining GPU programming
• The rationale behind using GPU acceleration
• Key challenges and trade-offs in GPU development
• Overview of available frameworks and tools
• Strategies for selecting the right framework and tool for specific applications

OpenCL

• An overview of OpenCL
• Advantages and limitations of the OpenCL ecosystem
• Configuring the OpenCL development environment
• Developing a basic OpenCL program for vector addition
• Using the OpenCL API to query device info, manage memory, handle host-device data transfers, launch kernels, and synchronize threads
• Writing device-side kernels using the OpenCL C language
• Leveraging OpenCL built-in functions, variables, and libraries for common operations
• Optimizing performance by utilizing OpenCL memory spaces: global, local, constant, and private
• Controlling parallelism via the OpenCL execution model: work-items, work-groups, and ND-ranges
• Debugging and testing OpenCL programs with tools like CodeXL
• Optimizing OpenCL code through coalescing, caching, prefetching, and profiling

CUDA

• An overview of CUDA
• Strengths and weaknesses of the CUDA platform
• Setting up the CUDA development environment
• Creating a basic CUDA program for vector addition
• Utilizing the CUDA API for device queries, memory management, data copying, kernel launches, and thread synchronization
• Writing kernels with CUDA C/C++ for device execution and data manipulation
• Employing CUDA built-in functions, variables, and libraries for standard tasks
• Optimizing data flow using CUDA memory spaces: global, shared, constant, and local
• Managing parallelism via the CUDA execution model: threads, blocks, and grids
• Debugging and testing CUDA applications using CUDA-GDB, CUDA-MEMCHECK, and NVIDIA Nsight
• Enhancing CUDA performance with coalescing, caching, prefetching, and profiling

ROCm

• An overview of ROCm
• Benefits and drawbacks of the ROCm stack
• Preparing the ROCm development environment
• Building a basic ROCm program for vector addition
• Using the ROCm API to manage device info, memory, data transfers, kernel launches, and synchronization
• Writing kernels with ROCm C/C++ for device-side data manipulation
• Utilizing ROCm built-in functions, variables, and libraries for common operations
• Optimizing performance via ROCm memory spaces: global, local, constant, and private
• Controlling parallelism using the ROCm execution model: threads, blocks, and grids
• Debugging and testing ROCm programs with the ROCm Debugger and ROCm Profiler
• Optimizing ROCm code through coalescing, caching, prefetching, and profiling

HIP

• An overview of HIP
• Advantages and disadvantages of using HIP
• Setting up the HIP development environment
• Creating a basic HIP program for vector addition
• Writing kernels using the HIP language for device execution and data manipulation
• Utilizing HIP built-in functions, variables, and libraries for common tasks
• Optimizing data access using HIP memory spaces: global, shared, constant, and local
• Controlling parallelism via the HIP execution model: threads, blocks, and grids
• Debugging and testing HIP programs using the ROCm Debugger and ROCm Profiler
• Optimizing HIP code through coalescing, caching, prefetching, and profiling

Comparison

• Comparing the features, performance, and compatibility of OpenCL, CUDA, ROCm, and HIP
• Evaluating GPU applications using benchmarks and performance metrics
• Adopting best practices and expert tips for effective GPU programming
• Exploring current trends and future challenges in the GPU programming landscape

Summary and Next Steps