AMD GPU Programming Training Course

Introduction

• Overview of ROCm
• Introduction to HIP
• Comparing ROCm, CUDA, and OpenCL
• Features and architectural overview of ROCm and HIP
• Preparing the development environment

Getting Started

• Initiating a new ROCm project in Visual Studio Code
• Examining project structure and file organization
• Compiling and executing the program
• Displaying output using printf and fprintf functions

ROCm API

• The role of the ROCm API within the host program
• Querying device information and capabilities via the ROCm API
• Managing device memory allocation and deallocation using the ROCm API
• Transferring data between host and device with the ROCm API
• Launching kernels and synchronizing threads using the ROCm API
• Handling errors and exceptions through the ROCm API

HIP Language

• The role of the HIP language in device-side programming
• Writing kernels for GPU execution and data manipulation using HIP
• Utilizing HIP data types, qualifiers, operators, and expressions
• Applying HIP built-in functions, variables, and libraries for common operations

ROCm and HIP Memory Model

• Distinguishing between host and device memory models
• Managing memory spaces such as global, shared, constant, and local
• Working with memory objects like pointers, arrays, textures, and surfaces
• Applying memory access modes including read-only, write-only, and read-write
• Understanding the memory consistency model and synchronization mechanisms

ROCm and HIP Execution Model

• Differences between host and device execution models
• Defining parallelism using threads, blocks, and grids
• Utilizing thread functions such as hipThreadIdx_x, hipBlockIdx_x, and hipBlockDim_x
• Employing block functions like __syncthreads and __threadfence_block
• Using grid functions including hipGridDim_x, hipGridSync, and cooperative groups

Debugging

• Identifying common errors and bugs in ROCm and HIP programs
• Inspecting variables, breakpoints, and call stacks using the Visual Studio Code debugger
• Debugging ROCm and HIP programs on AMD devices with the ROCm Debugger
• Analyzing program performance on AMD devices using the ROCm Profiler

Optimization

• Factors influencing the performance of ROCm and HIP programs
• Improving memory throughput via coalescing techniques
• Reducing memory latency through caching and prefetching strategies
• Optimizing memory access and bandwidth using shared and local memory techniques
• Measuring and enhancing execution time and resource utilization through profiling and dedicated tools

Summary and Next Steps