Confirming the match

The Python golden and the Vitis kernels are compared bit-for-bit by a small build DAG — the same gen → csim → compare machinery the rigorous examples/schemas/fixedpoint conformance uses.

The build DAG

def build_basic_vec_dag() -> BuildDag:
    dag = BuildDag()
    dag.add(SourceStep(artifact="basic_vec_source", path=...))   # the sources
    dag.add(SourceStep(artifact="kernels_source",   path=...))
    dag.add(SourceStep(artifact="run_tcl",          path=...))
    dag.add(GenStep(name="gen"))                                 # write kernels + vectors + golden
    dag.add(RunStep(name="run"))                                 # Vitis csim + compare bits
    return dag

• gen — for each kind, computes the Python golden (the operators), renders the matching kernel, and writes kernel.cpp, in_{a,b,c}.txt, and expected.json (the golden bits). No Vitis needed.
• run — runs each kernel in Vitis C-sim and asserts the emitted bits equal expected.json exactly; the first mismatch stops the build:

failed = [r for r in results if not r["exact"]]
if failed:
    raise RuntimeError(f"STOP — Vitis disagreed with the Python golden: {failed[0]}")

Running it

python examples/basic_vec/basic_vec_build.py --through gen   # generate (no Vitis)
python examples/basic_vec/basic_vec_build.py --through run   # the bit-exact conformance (Vitis)

When run passes, the Python operator model and the Vitis kernel produced identical bits for int, float, and fixed — the whole point of basic_vec: a vectorized Python golden that predicts the hardware exactly. The same contract runs as a test in tests/examples/test_basic_vec.py (under pytest -m vitis).

See also

• Vectorization guide — the concepts this example embodies (the two paths, growth rules, why vectorized sim is fast and bit-exact).
• examples/schemas/fixedpoint — the rigorous all-modes/all-widths conformance sweep.