Batchify all APIs

[JunchenLiu77]

This PR updated all the existing APIs (3DGS, 2DGS, 3DGUT) to support multiple batch dimensions eg. [A, B, C] -> [..., A, B, C], which

1. ensures backward compatibility.

2. is tested with various batch dimension configurations.

[liruilong940607]

Looks good from skimming through it! I'll take a closer look later

[JunchenLiu77]

benchmark batch feature with 10000 gaussians (profiling/batch.py)

Model	N Batches	Mem (GB)	Time [fwd]	Time [bwd]
3DGS	1	0.010	0.00037	0.00049
3DGS	4	0.040	0.00040	0.00079
3DGS	16	0.161	0.00093	0.00284
3DGS	64	0.642	0.00368	0.01124
3DGUT	1	0.010	0.00042	0.00070
3DGUT	4	0.040	0.00057	0.00128
3DGUT	16	0.162	0.00162	0.00513
3DGUT	64	0.641	0.00635	0.02031

[MrNeRF]

Hey cool pr. Does this also work with bilateral grid? Is distributed computing (Grendel) supported?

[liruilong940607]

Hey cool pr. Does this also work with bilateral grid? Is distributed computing (Grendel) supported?

Hey there! This is just to update the rasterization function to support batch over multiple scenes (on top of our previous support for batch over multiple viewpoints).

bilateral grid already supports arbitrary batching on the final rendered images -- so yes it works with bilateral grid.

For Grendel, we already supported it in gsplat -- you could training a single scene with multiple GPUs by distributing the GSs. This PR addresses batching over multiple scenes (for the use cases of those feed-forward models) which is not quite relevant / orthogonal to Grendel.

[MrNeRF]

Ok, got it! Thank you!

[JunchenLiu77]

0518.mov

[liruilong940607]

LGTM! Awesome job JC for such huge effort!

[JunchenLiu77]

0518.mov

Fix bug in this video.

gsplat-batch.mp4