BufferGeometry.merge or BufferGeometryUtils.mergeBufferGeometries: Add support for transformation matrix

Summary of feature request

.addScaledVector but instead of adding a scaled vector, merging a transformed BufferGeometry

Description of feature request

For optimization purposes, I want to render a single buffer geometry instead of rendering each object in my scene separately

const geometry = THREE.BufferGeometryUtils.mergeBufferGeometries(this.children.map(child =>
    child.geometry.clone().applyMatrix4(child.matrix) // The issue
));
this.add(new THREE.Mesh(geometry, this.children[0].material));
for (const child of this.children) {
     child.visible = false;
}

(simplified code, not tested but should convey my point)

Unfortunately, the existing buffer geometry merge methods don’t support transforming each merged vertex by a custom matrix. That means I have to clone each child’s geometry and apply the matrix to the entire clone at once.

This is a waste of memory, and could be avoided with

• BufferGeometry.merge(bufferGeometry, optionalMatrix)
• BufferGeometryUtils.mergeBufferGeometries(bufferGeometries, optionalMatrices)

The 2nd option is slightly more awkward but is preferred because BufferGeometryUtils is lossless (doesn’t require manual index argument)

If a matrix is passed in, each vertex being merged in should have it applied.

Note: I know I could get around this by pre-applyMatrix4-ing all my objects’ geometries, but I don’t want to do that…I want them to share a single, applyMatrix4’d buffer geometry instance.

Note 2: For my use case, each child’s BufferGeometry is the same. A solution specific to my case is :

• BufferGeometryUtils.stamp(bufferGeometry, matrices)

Which would take a single geometry and stamp it at multiple locations/rotations/scales defined by each matrix.

Three.js version

• Dev
• r114

[donmccurdy]

@finnbear both your results and your code look reasonable to me.

I expect that a carefully designed function could write arbitrary objects, with arbitrary transforms, into a large geometry buffer efficiently enough to make the 50-100ms latency go away. For an imaginary BufferGeometryBuilder example:

var merged = new BufferGeometry();
merged.setAttribute( 'position', new Float32Array( 1e6 ), 3 );
merged.setAttribute( 'normal', new Float32Array( 1e6 ), 3 );
merged.setAttribute( 'uv', new Float32Array( 1e6 ), 2 );
merged.setIndex( new Uint8Array( 1e6 ), 3 );

var builder = new THREE.BufferGeometryBuilder( merged );
var offset1 = builder.append( geometry, matrix1 );
var offset2 = builder.append( geometry, matrix2 );
var offset3 = builder.append( geometry2, matrix1 );

// later
builder.update( offset2, geometry, matrix3 );

However, this sort of API seems better as a new thing, rather than a change to mergeBufferGeometries.

[donmccurdy]

What do you mean by this?

bufferGeometry1.merge(bufferGeometry2) overwrites data rather than doing a union. This is rarely what users want.

That means I have to clone each child’s geometry and apply the matrix to the entire clone at once. … This is a waste of memory …

I’m not sure how we’d implement it in BufferGeometryUtils any other way, without duplicating a fair bit of code from:

• BufferGeometry.prototype.applyMatrix4
• BufferAttribute.prototype.applyMatrix4
• BufferAttribute.prototype.applyNormalMatrix
• BufferAttribute.prototype.transformDirection

I agree with @gkjohnson that InstancedMesh may be a better fit for this use case.