[nicolaihaeni]

With more than one camera per call, rasterization_2dgs expands the opacities over the cameras without a copy, and the 2DGS rasterizer rejects the non-contiguous result:

RuntimeError: opacities must be contiguous

This makes the expanded opacities contiguous, as the 3DGS path does before its kernels. Checked: a 2-camera call now gives exactly the same outputs as two 1-camera calls, and gradients within ~1e-7 relative L2.

Repro (fails on main in the forward):

import torch
from gsplat import rasterization_2dgs
n, C = 20_000, 2
means = (torch.randn(n, 3) * 0.7 + torch.tensor([0, 0, 4.0])).cuda().requires_grad_()
quats = torch.nn.functional.normalize(torch.randn(n, 4), dim=-1).cuda().requires_grad_()
scales = (torch.rand(n, 3) * 0.03).cuda().requires_grad_()
opacities = torch.rand(n).cuda().requires_grad_()
colors = (torch.randn(n, 16, 3) * 0.3).cuda().requires_grad_()
viewmats = torch.eye(4).cuda()[None].repeat(C, 1, 1)
Ks = torch.tensor([[500.0, 0, 320], [0, 500.0, 240], [0, 0, 1]]).cuda()[None].repeat(C, 1, 1)
out = rasterization_2dgs(means, quats, scales, opacities, colors, viewmats, Ks, 640, 480,
                         sh_degree=3, render_mode="RGB+ED", distloss=True)