Taichi: Productive, portable, and performant GPU programming in Python
github.com
github.com
If you have an Nvidia CUDA GPU, consider https://github.com/NVIDIA/warp instead.
https://www.youtube.com/watch?v=V3Q9IjsgXvA
NVidia not only made CUDA great for C, C++, Fortran, and any language with a PTX backend, they are now adding a couple of Python frameworks as DSL for GPU JITs into the mix.
Meanwhile competition is too busy trying to replicate CUDA C without the ecosystem.
This isn't on-going work - this is my work from last year. I'm not Guray but I'm one of the people that built out the python bindings in MLIR (and I helped him with his demo). I'm not doing it anymore except for ~once every 3 months (and that's mostly some small fix). Also note the PTX emission in LLVM is far behind what NVIDIA has internally.