Implement closest_point function

Background:

Training SDF + texture model was not available because closest_point function wasn't implemented. (see #162) To traing the SDF + texture model, I implement closest_point function.

Changes:

Change all float variables in wisp/csrc/external/mesh2sdf_kernel.cu into double variables. Previously, when using float variables, there were accumulated numerical errors on SDF and closest points' coordinates. (the numerical SDF errors even reaches 1.xxx ...)
Add mesh2sdf_triangle_gpu_fast_nopre function in wisp/csrc/external/mesh2sdf_kernel.cu which can return both SDF and closest triangle.
Implement closest_point function in wisp/ops/mesh/closest_point.py. closest points are calculated from closest triangles returned from mesh2sdf_triangle_gpu_fast_nopre function.

Tests:

Unfortunately, I'm not good at pytest code. I just tested the code by printing results on terminal.

Test process:

Add print() in closest_tex() function (wisp/ops/mesh/closest_tex.py)


import torch
import numpy as np
from .barycentric_coordinates import barycentric_coordinates
from .closest_point import closest_point
from .sample_tex import sample_tex

def closest_tex( V : torch.Tensor, F : torch.Tensor, TV : torch.Tensor, TF : torch.Tensor, materials, points : torch.Tensor): """Returns the closest texture for a set of points.

    V (torch.FloatTensor): mesh vertices of shape [V, 3] 
    F (torch.LongTensor): mesh face indices of shape [F, 3]
    TV (torch.FloatTensor): 
    TF (torch.FloatTensor):
    materials:
    points (torch.FloatTensor): sample locations of shape [N, 3]

Returns:
    (torch.FloatTensor): texture samples of shape [N, 3]
"""

TV = TV.to(V.device)
TF = TF.to(V.device)
points = points.to(V.device)

dist, hit_pts, hit_tidx = closest_point(V, F, points)

##### !!!test by printing!!! #####
######################### 
print(dist)  # should be same with below result
print(((hit_pts - points) ** 2).sum(dim=1).sqrt())  # distance between hit_pts and points should be same with dist
######################### 
######################### 

hit_F = F[hit_tidx]
hit_V = V[hit_F]
BC = barycentric_coordinates(hit_pts.cuda(), hit_V[:,0], hit_V[:,1], hit_V[:,2])

hit_TF = TF[hit_tidx]
hit_TM = hit_TF[...,3]
hit_TF = hit_TF[...,:3]

if TV.shape[0] > 0:
    hit_TV = TV[hit_TF]
    hit_Tp = (hit_TV * BC.unsqueeze(-1)).sum(1)
else:
    hit_Tp = BC

rgb = sample_tex(hit_Tp, hit_TM, materials)

return rgb, hit_pts, dist

2. Run following code

from wisp.datasets.formats import MeshSampledSDFDataset

dataset = MeshSampledSDFDataset(mesh_path="data/obj/mesh.obj", split="train", sample_mode=["trace"], num_samples=5000, sample_tex=True, mode_norm="sphere")



Results:

I also trained the SDF + texture model on some .obj files after implementing closest_point function. (However, I don't include training code on this PR)

First image is .obj model in blender and second image is the SDF + texture model result. (wrong eye coloring is due to wrong .obj texture data, so you can ignore it)

![Untitled (9)](https://github.com/NVIDIAGameWorks/kaolin-wisp/assets/98357201/f58e8554-6bf5-4a18-89e7-4ce71b2aa8ed)
![Untitled (10)](https://github.com/NVIDIAGameWorks/kaolin-wisp/assets/98357201/2bccf84e-19a6-470a-94d9-2f6ab24870f2)

NVIDIAGameWorks / kaolin-wisp

Implement closest_point function #173