Fast Walsh-Hadamard Transform CUDA bindings for JAX
Credit to Tri Dao for the CUDA kernel implementation from HazyResearch/structured-nets
pip install .TODO
- make a simple implementation (for non-GPU)
- benchmark fused vs simple implementaiton
- vmap rules
- only supports float32, can se also support bfloat16??
- async stream for better performance?