Learn CUDA

A small, CMake-driven collection of CUDA examples. Each .cu file under src/ builds into its own executable, preserving the src/ folder structure under build-*/bin/.

What’s inside

• src/001-basic: launch basics (streams/events, graph launch, dynamic parallelism, etc.)
• src/002-device: querying device properties
• src/003-memory: allocation and host↔device copies, shared memory
• src/004-vector: vector add/copy/dot/sum
• src/005-matrix: matrix multiply/transpose/mat-vec, WMMA example
• src/006-ptx: PTX-related examples

Build outputs land here (example):

• build-release/bin/001-basic/hello-world
• build-debug/bin/004-vector/vector-add

Requirements

For both Dev Container and local

• NVIDIA GPU + compatible NVIDIA driver
• CUDA Toolkit (provides nvcc)
• CMake >= 3.25
• Ninja
• A C++20-capable compiler toolchain

Extra for VS Code Dev Container

• VS Code + “Dev Containers” extension
• Docker
• NVIDIA Container Toolkit (so the container can access your GPU)

Setup

Clone Dependency

This repo expects the dependency folder to exist at cpp-bench-utils/ (repo root).

git clone https://github.com/KindRoach/cpp-bench-utils.git

a) VS Code Dev Container

1. Open this folder in VS Code.
2. Run: Dev Containers: Reopen in Container.
3. Ensure GPU access works inside the container (you should be able to run nvidia-smi).

b) Local

1. Install CUDA Toolkit, CMake (>= 3.25), Ninja, and a C++ toolchain.
2. Clone this repo.

How to build, run and debug

Cmake Configure

This repo uses CMake configure presets from CMakePresets.json:

cmake --preset release

Other useful presets:

• cmake --preset debug
• cmake --preset relwithdebinfo

Build

cmake --build build-release

For debug builds:

cmake --build build-debug

Run

Executables are written under build-*/bin/<category>/.

Examples:

./build-release/bin/001-basic/hello-world
./build-release/bin/004-vector/vector-add

Debug

VS Code (recommended)

Install the “CMake Tools” extension, then:

1. Select a configure preset (e.g. debug or relwithdebinfo).
2. Configure + build.
3. Set the target executable as the debug program (from build-*/bin/...).

If you use relwithdebinfo, you get a good balance of debug symbols and runtime speed.