examples

Imaging examples

Three complete SAR imaging algorithms, each following the same layout so they can be compared side by side:

examples/
├── common/            shared infrastructure
│   ├── params.py      RadarParams, ALOS_PARAMS, synthetic_params
│   ├── simulate.py    point-target echo simulator
│   ├── plot.py        dB-scale image saving
│   └── alos.py        ALOS-1 raw loading and scene post-processing
├── data/              shared dataset: CEOS extraction tooling + ALOS product
├── wka/               omega-K (wavenumber domain)
├── rda/               Range-Doppler
└── csa/               Chirp Scaling (interpolation-free)

Every algorithm directory contains:

File	Purpose
algorithm.py	the imaging chain in the DSL: build_kernel(n, params), make_inputs(n, params)
reference.py	NumPy reference implementation (numerical ground truth)
run_cpu.py	full flow on the cpu backend: simulate, compile, focus, save a PNG
run_scalehls.py	full flow on the scalehls backend: trace and emit a Vitis HLS C++ design

Each run_*.py is self-contained -- one file, one backend, whole flow:

# from the repository root, after `make build` (and `make scalehls` for HLS)
python examples/wka/run_cpu.py --n 512
python examples/wka/run_scalehls.py --n 256
python examples/rda/run_cpu.py
python examples/csa/run_scalehls.py

All three algorithms also process the real ALOS-1 San Francisco dataset (16384 x 16384 raw echoes, extracted once via data/extract_alos.py): each directory has a run_alos.py. Urban-area image contrast at full size: omega-K 131.6, Range-Doppler 130.3, Chirp Scaling 126.1 -- all focus the scene in ~3.5 s on a large multi-core CPU.

All three algorithms are validated in test/python/: DSL vs reference equivalence, point-target focusing quality, and cross-algorithm agreement on target positions.