Resources Numerical accuracy Calculation policy
Vertex skinning by bones on the GPU for the real-time view and the path tracer.
The WGSL kernel and its bone packing work in 32-bit and feed the renderers. The CPU path in rig.js is the reference the exports use.
2 calculations in 1 source file. The code is read from the released app when this page loads. It is the calculation people run, not a copy kept for this page.
Tests None found by the review
Packs the bone matrices (or dual quaternions) for the kernel.
Float32Array.from rounds every 64-bit matrix entry to 32-bit.
export function packBones(M, mode) {
const nb = M.length / 16;
if (mode === 'dqs') return Float32Array.from(dualQuats(M, nb));
return Float32Array.from(M);
}Runs one skin on the card and reads the points back.
The points come back from the GPU buffer as a Float32Array. The code comment says this call is for tests.
async skin({ rest, joints, weights, n, max = 4, M, mode = 'lbs' }) {
const bonesData = packBones(M, mode);
const b = buffersFor({ rest, joints, weights, n, max, bones: M.length / 16 });
device.queue.writeBuffer(b.params, 0, new Uint32Array([n, mode === 'dqs' ? 1 : 0, max, 0]));
device.queue.writeBuffer(b.bones, 0, bonesData);
const enc = device.createCommandEncoder();
encode(enc, b, n);
enc.copyBufferToBuffer(b.out, 0, b.read, 0, n * 12);
device.queue.submit([enc.finish()]);
await b.read.mapAsync(GPUMapMode.READ);
const out = new Float32Array(b.read.getMappedRange().slice(0));
b.read.unmap();
return out;
},Computed to draw the screen, often in 32-bit on the graphics processor. No number a person reads, saves or exports comes from it.