Future features

[yoyolicoris]

• Time-varying biquad (time-varying FIRs + sample_wise_lpc) #7
• Parallel scan for first-order filter #8
• Parallel scan for time-invariant, high-order filter with state-space form #9
• migrate to cpp API torch::autograd::Function
  • [cm] Initial CPP forward implementation #10
  • feat: LPC CUDA kernel #24
  • swap out python custom function
• Metal kernel for MacOS
• torch.compile compatible

[yoyolicoris]

We plan to have c++ kernels cover all the computation in the next release v0.7.
We won't distribute pre-built wheels, but source distribution, so the users have to compile them themselves. There is not much difference since the previous numba kernel use JIT and relies on user's compiler.

The numba kernels will still be there in case the compilation failed. They will be removed entirely in v1.0.
We can have pre-built binaries once I figure out how to set up a CICD pipeline with GPU machines.

[yoyolicoris]

PyTorch 2 compliant guide
https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit?usp=sharing