occt-booltest

OpenCASCADE (OCCT 8) boolean intersection stress-test harness and kernel-independent benchmark corpus.

Two separated modes (two binaries)

The native-modeling mode and the STEP mode are fully separate programs sharing only common.cpp (bottle model, geometry queries, STEP I/O, validated boolean):

binary	source	input	per-iteration pipeline
booltest	main.cpp	models the bottle in-process (MakeBottle)	copy → rotate → boolean, pure in-memory handoff (body1 ← body4 handle assignment, lossless); STEP files are written as archives only
boolstep	step_main.cpp	a STEP model file (--in bottle.step)	read STEP → copy → rotate → boolean → write STEP (or report failure); the written body4.step is re-read as the next iteration's input

Prepare the STEP input once with the modeling tool:

./build/mkbottle bottle.step     # native modeling, writes the STEP input

Native modeling mode

Per test case (axis × theta):

1. body1 = the OCCT tutorial bottle (MakeBottle 50×70×30, V0 = 6960.775 mm³)
2. body3 = body1 rotated by theta about axis through its bbox center
3. body4 = BRepAlgoAPI_Common(body1 solids, body3 solids)
4. on success body1 ← body4 (in memory), repeat until boolean failure / volume < V0/1000 / max iterations / face-explosion guard (>4000 faces) / wall-clock cap

Results are validated for physical plausibility: 0 < V ≤ min(Vin1, Vin3). OCCT silently returns garbage (empty, negative, oversized results) in near-degenerate configurations — those are recorded as failures with the offending shapes archived (BREP + STEP + screenshots).

STEP mode

Every case starts from the input STEP file and chains strictly file-to-file: read STEP → copy (BRepBuilderAPI_Copy) → rotate → intersect → write STEP. Each iteration directory holds the kernel-independent input pair (body1.step + body3.step) plus the result (body4.step); on failure the pair is kept with info.txt and the run reports it. boolstep links no Qt — it is a short, readable, file-in/file-out tester.

Build & run

cmake -B build && cmake --build build

# native modeling mode (screenshots via QOpenGLWidget)
./build/booltest --out output \
    --axes x,y,z,diag --thetas 1,2,5,15,45,90,180 \
    --max-iters 500 --case-seconds 300

# STEP mode (prepare the input once, then)
./build/mkbottle bottle.step
./build/boolstep --in bottle.step --out output_step \
    --axes x,y,z,diag --thetas 1,2,5,15,45,90,180 \
    --max-iters 500 --case-seconds 300

Dependencies: OCCT (homebrew opencascade), Qt6 (only for booltest offscreen screenshots via QOpenGLWidget grabFramebuffer).

Key findings (see corpus summary.csv for raw data)

• Exact-coincidence geometry kills the boolean: rotating about z (near-symmetry axis of the bottle) fails 7/7 in memory mode. A STEP round-trip (volume perturbation ~4e-8%) revives 4 of those 7 — "mathematically identical faces" are harder than "nearly identical".
• Silent invalid results: intersections larger than both inputs (up to 9.5× input volume), negative volumes, and near-zero slivers are returned with status OK. The plausibility check catches them.
• Cross-run nondeterminism: identical binaries + identical case can end differently in different processes (archived evidence in corpus).
• STEP round-trip is a double-edged sword: merges coplanar faces (4363 → 34 faces on axisx_theta15) and shrinks some results by >99%, but adds noise that breaks thin slivers that succeeded in memory.

Benchmark corpus (STEP mode)

Published as GitHub Release assets (not in git):

asset	contents
booltest-corpus-step-v1-core.zip	bottle.step, summaries, all failure archives (brep + info + png), first iteration of every case
booltest-corpus-step-v1-full.zip	everything: 2389 input pairs (body1+body3 per iteration), 2378 results (body4)

To benchmark your kernel: read each iteration's body1.step + body3.step, intersect, compare your volume against body4.step / all_iterations.csv. summary.csv rows carry io_mode so memory-mode and step-mode runs are never confused.

License: code MIT; corpus data CC-BY-4.0.