cramsay / conifer

A playground for parallel and multiplierless FIR filters with Clash

Conifer

A playground for parallel and multiplierless FIR filters with Clash.

This work, presented first at the 2021 Asilomar conference, uses Clash to compose two interesting filter techniques --- the Fast FIR Algorithm (FFA) for smaller parallel structures, and a set of different Multiple Constant Multiplication (MCM) blocks for the subfilter arithmetic. This is presented as a solution for front-end digital filtering in RFSoC devices, but is actually quite widely applicable.

Both of these techniques are well studied in the literature but very rarely implemented. We argue that this is only an undiscovered pearl because of our (unfortunately uncommon) choice of language and tooling. Traditionally, designers would use ad-hoc circuit generators written in a software language. These are used essentially as a black-box, prove difficult to test, and are even more painful to compose. Clash gives us the tools we need to describe and compose these algorithms (with quite a lot of compile-time complexity) in a native hardware description language.

For more info, see the prepublication versions of our presentation, slides, and poster.

Usage

We have packaged our circuit generator as cabal and nix packages in clash/.

If you want to have a quick play around, the top-level shell.nix file will source everything you need. You could start by looking at one example filter (and our scripts to implement it out-of-context with Vivado) in the synth/ folder. Make sure you have Vivado in your PATH and you have installed the nix package manager.

me@computer:~/conifer $ nix-shell

[nix-shell:~/conifer]$ clashi synth/Filter.hs 
Clashi, version 1.4.3 (using clash-lib, version 1.4.3):
https://clash-lang.org/  :? for help
[1 of 1] Compiling Main             ( synth/Filter.hs, interpreted )
Ok, one module loaded.
*Main> ...

Have a look at the examples in clash/src/ExampleFilters.hs for guidance, or just have a browse in clash/ if you're curious.

Reproducible Research

We present a lot of utilisation/timing results in our publication. We want this to be as reproducible as possible. If you want to recreate our results, or do something similar with your own extensions, take a look at the analysis folder. There's another shell.nix environment that should include everything you need. We offer our result-generating scripts as two Jupyter Notebooks:

1. ImplementationResults.ipynb --- Scripts to implement our designs (and the LogiCore FIR Compiler's as a reference) with Vivado's out-of-context flow, plot the results, and export them as CSVs. See coeffs.py and filters.py for most of the automation and mechanics.
2. CoefficientSymbolicTests.ipynb --- Presents an experimental verification (using symbolic programming with sympy) of the equations we have derived for the number of multiplications our filter structures really need to implement when presented with various common styles of symmetry present in real-world coefficient sets.

View and interact with them with:

me@computer:~/conifer/analysis $ nix-shell --option sandbox false --command "jupyter lab"

This should open a browser window with the Jupyter Lab interface in a new tab.

License

Conifer is licensed under GPLv2. Please feel free to use it accordingly for educational and academic uses, including extending it for your own publications. For commercial uses, this license gets much more restrictive. Contact us if you want to negotiate a new license for commercial use.