CNN-Core

A CNN Core Generator and Code on ARIANNA Trigger System

Dependencies

pip install -r requirements.txt

The conversion and build flow requires a working Vitis HLS installation. The current project targets:

xcku5p-ffvb676-2-e

Model Conversion

The conversion entry point is:

python scripts/convert_homogeneous.py

This script loads the HGQ/Keras model from models/hgq_config_beta7_gamma6_p1_cl_best_v3.keras, extracts the homogeneous quantization settings, rebuilds the network with standard Keras layers, and converts it to an hls4ml Vitis project under cnn_core_project/.

The generated design uses io_stream and applies per-layer fixed-point precision to the convolution and dense layers. The test data in data/ is also used to compare the Keras output against the compiled HLS model.

Main outputs:

cnn_core_project/keras_model.keras
cnn_core_project/hls4ml_config.yml
cnn_core_project/firmware/
cnn_core_project/tb_data/

Bender

bender update

Vitis HLS Build

Edit cnn_core_project/build_opt.tcl

array set opt {
    reset      0
    csim       1
    synth      1
    cosim      1
    validation 1
    export     1
    vsynth     1
    fifo_opt   1
}

Recommended full local check:

csim       1
synth      1
cosim      1
validation 1
export     1
vsynth     1
fifo_opt   1

Then run:

cd cnn_core_project
vitis_hls -f build_prj.tcl

The vsynth step runs vivado_synth.tcl, which performs Vivado synthesis and optimization on the generated RTL. It does not run full place-and-route implementation.

Reports

Check the HLS synthesis report for latency, initiation interval, and HLS resource estimates:

cnn_core_project/cnn_core_prj/solution1/syn/report/cnn_core_csynth.rpt

Check the Vivado synthesis utilization report for post-synthesis resource usage:

cnn_core_project/vivado_synth.rpt

Typical metrics to record:

Latency cycles
Initiation interval
Estimated clock period
LUT / FF / DSP / BRAM usage

For post-implementation timing and utilization, run a separate Vivado implementation flow with place_design and route_design, then collect report_timing_summary and report_utilization.

Always check the large files before pushing.

git rev-list --objects --all \
| git cat-file --batch-check='%(objecttype) %(objectname) %(objectsize) %(rest)' \
| sed -n 's/^blob //p' \
| sort --numeric-sort --key=2 --reverse \
| head -n 10 \
| cut -c 1-12,41- \
| numfmt --field=2 --to=iec-i --suffix=B --padding=7

Release

Use it to update Bender.yml before release.

bash ./scripts/generate_bender_sources.sh

License

This project is licensed under the MIT License - see the LICENSE file for details