Where should board support files for the VivadoAccelerator backend be shared?

[boyesm]

Hello,

Although hls4ml is phasing out the use of VivadoAccelerator in favour of backends that use modern AMD tools, several of the current tutorials and documentation require VivadoAccelerator to complete [1]. In my opinion, this makes the VivadoAccelerator backend still relevant to working with this project when using AMD FPGAs.

Additionally, VivadoAccelerator is the only backend for AMD FPGAs that allows for the generation of bitstream files directly from Python. This feature significantly streamlines the development process, however, as of the current version only 6 types of board are supported. This led me to create my own board support files for the board I was working with (RealDigital AUP-ZU3.) If someone had distributed the files that I eventually produced, it would've saved me significant amounts of time, which is why I would like to contribute those files.

However, other PRs that attempt to add additional board support files are currently marked as 'wont-merge' [2] [3] [4]. So what would be the best way for myself and others to share the board support files that they've created? I could put them in my own repository/fork but people that need them might not find them there.

If there is any work surrounding this issue that you would like me to complete, please request it from me and I will work to complete it. Thank you for maintaining such an incredible tool.

Regards,
Malcolm Boyes

[bo3z]

Thank you for raising this question.

In general, I would recommend basing your contribution on top of the new VitisAccelerator backend branches (#1376 or #991) as these will be maintained longer term and follow a largely similar syntax to the prior VivadoAccelerator backend.

If you would like to contribute your board support packages to hls4ml, there are two options. First, like the other PRs, opening a PR with the target board as a reference for other users, though keeping in mind that the PR is unlikely to be merged, as VivadoAccelerator backend is being phased out, and, we are unable to support boards to which we don't have access to.

Second, as an alternative, we could think of something like the Extension API and contrib folder for Accelerator backends, for which, board support files could exist in some external repos (e.g., as submodule) and "injected" into hls4ml via an API similar to the Extension API. Currently, the Extension API allows users to define custom HLS kernels and layers, and, register them with hls4ml, while still relying on the rest of the infrastructure for other layers. If you would like to contribute to hls4ml, you could work on an extension-like API for the accelerator backends, and, if working correctly could include submodules to external board support file repos, or, allow users to define their own.