Entry 20

The rigid body integrated and kept orthonormal

Files7 in 4 directories
crates/hwtr-game/src/body.rsmath.rscar.rscrates/hwtr-game/tests/body.rsmath.rscrates/hwtr-hle/src/port.rsdocs/engine/rigid-body.md

The motion half of the car's step is ported. integrate (0x8006c504) turns the force and torque sums from 19 into motion, orthonormalize (0x80025be4) repairs the rotation afterwards, and wheel_spin (0x80044fc4) sets how fast each wheel is drawn turning. In the running game, 9151 calls of the ported functions matched over 2400 frames of scripted driving from the race start (55 skipped because an interrupt landed inside). The layout and the steps are on the rigid body page.

The car carries a rigid body at +0x30. car_update calls 0x8006c504(car + 0x30, dt), and the body's fields line up with the car's known ones: position at body +0xdc is car +0x10c, velocity, rotation and the two sums likewise. The port defines those car fields through the body (POS = BODY + body::POS). The integrator has four callers, so it is a general rigid body, not just a car part.

The car's "up" vector was gravity. 18 named car +0xec "up" from its use in update_wheels. The integrator multiplies body +0xbc (car +0xec) by strength × mass and adds it to the force. Its value is (0, 0, -1) and the strength 386 in/s², so it is gravity's direction. The +0x54c count is then of grounded wheels on a floor (normal within 60° of straight up). Both names are corrected.

The 64-bit pieces. Angular momentum, torque and both inverse inertias are 64-bit. Two 730-instruction helpers take their arguments by value on the stack: an 88-byte block (a 72-byte s64[3][3] plus 16 bytes) and a 32-byte MATRIX, with the output pointer after them. They multiply A · M (0x80026884) and M · A (0x800273ec), so R I⁻¹ Rᵀ is two calls. Speed and spin rate both go through the 64-bit square root that vec_length uses, so the port's Tables::length serves both.

A wrong save state. The "desert1-drive" state saved at frame 330 has the player nearly stopped, with the body asleep: the start countdown ends after frame 300. Holding X does accelerate (about 830 in/s by frame 700). A third state, desert1-speed (frame 700, in a turn), is now in the tests. The shadow runs were not affected; they drive for hundreds of frames past the countdown.

Mutation testing again separated real gaps from equivalent mutants. The clamps compared with > against >= cannot be told apart, because at exactly the limit the scale is div_fx(x, x) = 4096. A 16-bit clamp in place of truncation in orthonormalize is equivalent too, since normalised components never reach 16 bits. Everything else, from the shift counts and the skew matrix's signs to the transpose, is caught.

Was unknown

what 0x8006b754 and 0x8007c894 integrate; where mass, inertia and gravity are set up; car +0x770; the airborne control 0x8003d71c and the rest of car_update

Closed by No. 75 — Closing the questions later work answered.