jmitrevs/test-hls4ml-backend

Vivado AXI-master Backend and Baremetal Application

This is a workspace for testing the integration of the Vivado Accelerator AXI-master backend and baremetal applications. It is a work-in-progress repository, but it should help us to converge on a PR for the official hls4ml repository.

The hls4ml fork and branch that we use in this workspace is

• https://github.com/GiuseppeDiGuglielmo/hls4ml/tree/gdg/axi-m
• https://github.com/hls4ml-finn-mlperftiny/hls4ml/tree/fifo_depth_opt
• https://github.com/fnal-fastml/hls4ml/tree/external-weights-vivado-accelerator

conda env create -f environment.yml
conda activate hls4ml-vivado-accelerator
pip install qkeras==0.9.0
pip uninstall hls4ml
pip install git+https://github.com/fnal-fastml/hls4ml.git@external-weights-vivado-accelerator#egg=hls4ml[profiling]

A few notes

• Please notice the changes and additions:
  • convert_from_keras_model
  • write_header_file
• In particular, the function write_header_file in hls4ml is a draft. It should be general enough to support various models in the future.
• The call of the same function has to be embedded in the hls4ml backend generation flow (not at the top level).

Profile the Model

make run-profile

Run the Hardware/Software Flow

make run
cd sdk
make clean sdk gui

At this point if you have a board locally connected you can run the software and bitstream directly from Vivado SDK.