Topic 11. Computing on graphics processors: GPU architecture, the CUDA C++ language, and the ILGPU library for C#

Goal: become familiar with GPU architecture and the CUDA programming model; learn to write CUDA C++ kernels with one- and two-dimensional grids, manage host and device memory, use shared memory, block synchronization, and atomic operations, and measure time correctly with CUDA events; master the ILGPU library for GPU computing in C# and compare CPU and GPU performance, taking data transfers into account.

Lecture contents

1. GPU architecture and the CUDA ecosystem — GPU and CPU architecture · The NVIDIA ecosystem and CUDA in WSL2
2. The CUDA C++ programming model — The CUDA C++ programming model · Host and device memory
3. Building and basic parallel patterns — Building a project and handling errors · Basic parallel patterns
4. Shared memory, CUDA streams, and libraries — Shared memory and block synchronization · Asynchronous execution: CUDA streams and events · CUDA libraries and profiling
5. ILGPU, comparisons, and common mistakes — The ILGPU library for C# · Comparing CPU and GPU · Common mistakes

Practice and review

• Practice — 3 worked examples with complete code and output
• Tasks — 30 variants, each with tasks at three difficulty levels
• Summary — conclusions, self-check questions, and useful links