v0.2.4 — Two pads, four-fighter results and picture baselines
Test tooling only. Nothing in how the game plays changes from v0.2.3.
New
- A two-pad match. The
versusscenario plugs in a second emulated
controller: both pads press START, pick their fighters and fight, and the check
passes only if both fighters are pad-controlled, both walked and both landed a
hit that did damage. It is part of the full regression. - A four-fighter match played to its result.
ffa-resultasserts that every
fighter but the winner was recorded as put out, each once, at zero health. - Two matches in one launch.
two-matchesplays a One on One to its result,
goes back through the summary and the menus, and plays a Free For All to its
result. Each match is judged on its own events. - Picture baselines for still screens. The crib menu, the Learn Moves list
and the move preview are compared against images reviewed on the machine, with
measured tolerances and masks for what legitimately varies. Baselines are the
game's artwork, so they stay in the data folder and are not in this repository;
the check reportsblockedwhere none exist. - Captures at a fight step.
--step-shotstakes the frame at a given
simulation step instead of a host-time offset.
See Gameplay test harness.
What was learned
- In a repeatable fight the fighters, health bars, clock, arena and camera come
out pixel-identical between runs. The crowd does not. A new probe that counts
every random draw (RECOMP_TEST_DRAWS) showed why: the background crowd and
the 3D spectators both start animating while the match loads and leave that
window in a slightly different state each run, although every seed is the same. - The spectators push fighters, so that difference can reach the fight: pinned
pairs parted at step 871 with the CPU fighter's position the first thing to
differ. This is the likely cause of the divergence noted for v0.2.3.
Details:docs/research/crowd-nondeterminism.md. - Pad scripts can wait for the Nth occurrence of a screen event, which is what
lets one launch go round the menus twice.
Build this version
git clone --recursive --branch v0.2.4 https://github.com/vyanhursky/DefJamFFNY-recomp.git
cd DefJamFFNY-recompUpdating from v0.2.3: rerun the lift and the build (.\scripts\recomp.ps1, then
.\scripts\build.ps1 -Preset win-x64-release). This is a source release: no
game data, generated code, captures or executable.
Validation
On one machine (Windows 11, 2560x1440 at 120 Hz):
- Release: the fourteen-check regression passed in 58 minutes, including the
two-pad match (26 of 26 assertions), the combat check (27 of 27) and the replay
check against the golden hash. The opt-in checks passed: the four-fighter result
(28 of 28), the repeat check and the picture check (4 of 4 comparisons).
two-matchespassed once (27 of 27 in 731 seconds). - Debug: fight, two-pad, combat and replay passed. The full Debug regression
was not rerun. - The random-draw probe was added after those runs. With it in, the unit tests,
the three reference frames, fight, four-fighter fight, combat, two-pad, replay
and repeat passed on Release, and fight and replay on Debug. The Story routes,
the Terrordome match and the boot soak were not rerun with it. - Unit tests: 170 pass with the compiler available.
Limits
- The two-pad check uses emulated pads; real controllers, and three or four pads,
have not been checked. - Fight pictures are not compared (the crowd). Sounds are not tied to events.
- A pinned fight is not guaranteed to repeat even for thirty seconds (see above).
Thereplaycheck has passed every time it has run and can fail by chance. - A human win, time-up, save and load, cutscene skipping and long sessions are
not covered. - The rare sound-library crash guarded in v0.2.3 still has its cause in the
emulated audio chip; fixing that is a to-do.