README.md

softmax-kernels

GPU kernel optimization: Softmax

Usage

The following code was tested using the docker image: nvidia/cuda:12.4.0-devel-ubuntu22.04 on a Geforce RTX 2070

• Build Python library with CUDA bindings

cd cuda
pip install .

• Test both implementations against Pytorch baseline

python3 assertions.py

• Run a benchmark

python3 benchmark.py

• Profile both implementations

ncu --set full [-o output_path] python3 -O assertions.py

Articles

• Part 1: First CUDA implementation

• Part 2: Beating naive Triton implementation

• Part 3: Online Softmax