LibreCUDA aims to replace the CUDA driver API, enabling running CUDA code on Nvidia GPUs without relying on the proprietary CUDA runtime. By communicating directly with the hardware, it can upload CUDA ELF binaries onto the GPU and launch them using the command queue. Currently, it can allocate and free GPU memory, map memory for CPU access, and perform basic CUDA functions. The project is in early stages and not ready for production. Contributors are encouraged to submit issues and pull requests to enhance LibreCUDA, which is licensed under the MIT License.
https://github.com/mikex86/LibreCuda