Breeze 🌬️

Breeze is a header-only tensor (toy) library built with C++20 and CUDA. It provides a simple Tensor representation, smart RAII-based device memory management, common element-wise operations, and optimized matrix multiplication implementations. (for now)

Usage

The easiest way to build the project is using the Nix package manager, which guarantees consistent versions of the CUDA toolkit, compilers, and libraries.

Instructions to build and test demos

Once inside the development shell, run the following commands to configure and compile the project targets:

cmake -B build -S .
cmake --build build

Benchmark

A basic matrix multiplication benchmark. More stuff to be added soon.

Matrix Multiplication Benchmark

This was benchmarked on an RTX 3060 Mobile with 6GB VRAM. Currently, cold starts are unaccounted for.