Line: Add support for interleaved geometries in raycast().

[manthrax]

Correctly compute stride and offset for reading vertices from InterleavedBufferAttribute.

#20365

[gkjohnson]

Does it make sense to use Vector3.fromBufferAttribute here instead? Then it's agnostic to the underlying attribute type.

[manthrax]

Yup that makes great sense.

[manthrax]

@gkjohnson I looked into the implementation of fromBufferAttribute, and it looks like it involves more function calls and index multiplies than computing the indexing inline.. and this is a pretty performance critical piece of code, and am leery of changing the runtime characteristics. I already agonized over the 2 tests of attributes.position.isInterleavedBufferAttribute but I figure the jit can optimize it trivially. I'm not as confident that it will inline the operations invoked by fromBufferAttribute..

fromBufferAttribute calls:
attribute.getX( index ) <- returns this.array[ index * this.itemSize ]
attribute.getY( index ); <-returns this.array[ index * this.itemSize + 1 ]
attribute.getZ( index ); <-returns this.array[ index * this.itemSize + 2 ]

whereas the inline version only has to do index * this.itemSize once, and then fetches consecutive vertices.

on a 1million point/line cloud that could be 2m more multiplies and 3m function calls on a raycast...

Of course I could be wrong and the jit inlines all of it no problem, but on less capable browsers it may still be relevant.

Thoughts?

[usefulthink]

Here's the alternate version using fromBufferAttribute(): dev...usefulthink:patch-1

Now, I'm not sure if we will ever see any more types of BufferAttributes (only I could think of would be float64-emulation like deckgl does it), in which case the version using fromBufferAttribute would be preferrable. I think something could be done to improve the perfomance in that implementation (maybe adding BufferAttribute.toVector3(index, targetVector) or something like that.

Also the js raycaster already isn't very high-performing to begin with and I think most people dealing with larger geometries will go for alternatives like GPU picking anyway.

[manthrax]

Sounds good to me. :)

[Mugen87]

Please use Vector3.fromBufferAttribute(). It's not acceptable that interleaved buffer logic leaks out to the raycasting code.

[manthrax]

@Mugen87 Can you take @usefulthink s patch as is? If it's more helpful to you, i can update this PR...

[Mugen87]

I guess it's easier if you just update the PR 👍

[manthrax]

Updated PR.. thanks @Mugen87, @usefulthink. :)

[mrdoob]

Thanks!