Skip to content

Personal · 2026

leetGPU

CUDA kernels written from scratch, from matrix multiplication to an online softmax.

CUDA · C++

GitHub

The situation

one value per threadmax and sum, in one passlog₂ n steps

LeetGPU is a set of GPU programming challenges. I solve them on my own time: element-wise kernels (ReLU, SiLU, SwiGLU), matrix addition, multiplication and transpose, a 1D convolution, a parallel reduction and an online softmax.

Results

kernels
~20

github.com/anassee15/leetGPU

About 20 kernels so far, all on GitHub.

Next

HES-SO · bachelor · 2023Connect Four, read by a webcam