Export/Perf: use numpy in extract_primitives (2-8x speedup)

[scurest]

Work in progress. This is ready.

Rewrite extract_primitives to use numpy and foreach_get. This gives a pretty good perf boost; export times are now comparable to the other Python exporters (eg. FBX).

Some time comparisons.

Model	Before	After
CesiumMan, no anim	0.37s	0.09s
Ruby Rose, no anim (iss848)	6.0s	2.7s
01_vertex_count_32bit	2.6s	0.5s
Suzanne subdivided 4 times	9.6s	1.2s
AliciaSolid (iss967)	5.8s	1.7s
folded.blend (iss1112)	194s	35s

Dots

To replace the dict of verts, the idea is to use numpy.unique. We collect a "dot" for every loop in the mesh. Each dot stores the Blender vertex index (which fully determines all its per-vert data: position, vertex groups, etc), and all the per-loop data at that loop. To make a primitive, we slice out all the dots for the primitive's tris, use numpy.unique to dedupe them, and then read out the attribute arrays from the dots.

convert_swizzle_*

These aren't used by extract_primitives anymore, so I moved them to gltf2_blender_node where they're still called.

The transformations from eg. __convert_swizzle_normal are now part of the __get_normals function.

Changes

I think there are only two changes

• the bitangent vector is never stored. Instead, only the bitangent sign is tracked. This should only flip if we apply a transform that reverses orientation (determinant < 0). Since the Zup2Yup doesn't change orientation, the only reason it would flip would be the apply_matrix transform from skinning.
• there is no warning for 0-length tangents since convert_swizzle_tangent is gone. I can put one back in if you want though.

[fire]

Is there a good way to test on 2.83? Can you generate a zip for me?

[scurest]

@fire You'll need 2.90 for this. Actually the master branch just switched over to targeting 2.91, but this will still work on 2.90.

[scurest]

@julienduroure Think this is ready for review now, when you have time. I updated the OP.

[scurest]

Tested exporting original.glb from #1166 with default export options. The memory use also improves a lot.

• master
  • time: 118s
  • peak mem: 7.4G
• this branch
  • time: 14s
  • peak mem: 2.5G

[julienduroure]

Thanks !