How to get regular float32 geometry of a mesh compressed using gltfpack - three.js

Question

I used mesh decimation using gltfpack with the command
gltfpack -i input.glb -o output.glb -si 0.01
this reduces my mesh geometry triangles by 99 percent.
Now my output.glb has geometry.position as an interleaved Buffer Attribute of data type unsigned Int 16. I am using ammo.js to make its physics body, which requires geometry as a regular float32 array. But my attempts to convert it by dividing by 2^32 - 1, 2^31 - 1 have failed.
I get the body but positioned with different size and position, and it doesn't align with the threejs rendered model.
Is there a way to a regular float32 array so that I can pass it to the createConvexHullPhysicsShape function of Ammo.js
I've added a sandBox that shows the issue. issue

Answer 1

Steps for quantizing and de-quantizing a normalized array can be found in the KHR_mesh_quantization specification, equivalent to what hardware APIs will do. For int16 that would be:
float = max(int / 32767.0, -1.0)
This does assume that the values are normalized. In gltfpack I think that is an option. You can check attribute.normalized in three.js. If it is false there is no need for conversion, you can just use the values as-is.
Finally, note that if the buffer is interleaved, you will have to de-interleave, as the buffer might contain vertex attributes other than position.
UPDATE: September 9, 2022
The positions this GLB do not actually appear to be normalized, in the WebGL sense. They are large integers, and there's an inverse scaling on the parent nodes that brings them back down into a smaller floating point range. You'll need to account for that when extracting positions from the Mesh to create a convex hull. For example —
import { deinterleaveAttribute } from "three/examples/jsm/utils/BufferGeometryUtils.js";
function getVertexPositions(root) {
const positions = [];
root.updateMatrixWorld(true);
root.traverse((obj) => {
if (obj.type !== "Mesh") return;
let position = obj.geometry.attributes.position;
// de-interleave the array
position = deinterleaveAttribute(position);
// cast from Int16 to Float32
position = new THREE.BufferAttribute(
new Float32Array(position.array),
position.itemSize,
false
);
// apply mesh scaling to the positions array.
position.applyMatrix4(obj.matrixWorld);
positions.push(position);
});
return positions;
};
Result: