Self Tests

🌐 English · Русский

← GPU Renderer · Project Structure →

Self-Tests

Nova ships a suite of headless self-tests that run as argument branches of the sample app — no interactive input, no rendering to a live terminal — so they can run on CI and guard against regressions. Each prints a PASS / FAIL line.

Running them

Build once, then run a test by name:

dotnet build -c Release
dotnet run --project SampleGame -c Release --no-build -- <testname>

For example:

dotnet run --project SampleGame -c Release --no-build -- gputest

Contributors typically build with dotnet build -c Release --no-incremental (a plain incremental build can report a false "0 warnings").

The tests

Name	What it guards
impulsetest	The impulse-based physics solver — a sphere and box dropping, resting, bouncing, sliding, tumbling, stacking, scattering and sleeping across a dt sweep (Stages 1–6).
jointtest	The joints — a ball-socket pendulum, a hinge with an angle limit + motor, and a distance joint as a rigid rod and a soft spring, all solved in the same constraint solver as contacts.
hulltest	The convex hull — quickhull, the support function, and closed-polyhedron mass properties (volume / centre of mass / inertia vs the analytic cube).
gjktest	GJK / EPA — closest distance + witness points between separated convex shapes, and the penetration normal + depth for overlapping ones.
hullphystest	The dynamic convex-hull collider — a hull rests flat on a box / another hull / a mesh floor (at the same height), stacks, rests and slides on a real sloped mesh (not its bounding box), and tumbles to rest.
cutovertest	Stale-engine-key world loads, and the physics net-sync roundtrip (authority steps → serialize → client dead-reckons + eases → converges).
bvhtest	The BVH build/traversal.
worldtest	The world system.
editortest	The in-scene editor end-to-end: lights, camera objects (Fixed/Follow, F7/F8 cycles, clip-avoidance, split-screen aim), HUD layout, pure editor math, joint authoring + net sync, the dynamic mesh→hull bridge, and the player capsule.
picktest	Aim-select / object picking.
worldsynctest	A full world round-trip through WorldSync (every object/light/camera/physics/texture field + compare).
collisiontest	The pure camera-bubble collision resolvers and 3D SAT.
physicstest	The shared pure physics math (quaternion orientation, angle lerp, dead-reckon, restitution combine, closest-point-on-triangle).
gputest	CPU↔GPU parity — renders fixed scenes (incl. textured) on both paths and compares pixel-by-pixel.
texturetest	The PNG decode / sampling / UV / params / mip pipeline.
splashtest	The startup-splash block-letter helpers.
uitest	The setup-wizard console-UI toolkit (layout, focus nav, widgets, mouse hit-test, wizard flow).
colortest	A visual-only color diagnostic (no PASS/FAIL line).
uidemo	A visual-only interactive demo of the console-UI toolkit (font-zoom check).

What "gputest Δ=0" means

gputest renders the same scene with the CPU raytracer and the GPU kernel and compares them pixel-by-pixel:

• With shadows off, every feature is deterministic, so the two images must be exactly equal — the pixel difference Δ = 0.
• With shadows on, only a thin shadow-boundary band is allowed to differ.
• For textures, mesh primitives and imported meshes stay Δ = 0; the sphere's analytic UV allows a thin seam/pole band, and a mipmapped object allows a thin band.

A Δ=0 result is the proof that a change to shading / intersection / compositing / BVH / texture math kept the CPU and GPU renderers in lockstep. See GPU Renderer.