-
Notifications
You must be signed in to change notification settings - Fork 0
/
rt-gpu.py
191 lines (144 loc) · 6.3 KB
/
rt-gpu.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
#!/usr/local/bin/python3
# -*- coding: UTF-8 -*-
import torch
import argparse
import numpy as np
from itertools import repeat
from utils import XMLReader
try:
from scipy.misc import imsave
except:
from matplotlib.pyplot.plt import imsave
def parse_args():
parser = argparse.ArgumentParser('Ray tracer implemented in python.')
parser.add_argument('--input_file', type=str,
required=True, help='input XML file')
parser.add_argument('--output_file', type=str,
default='fig.png', help='output figure')
parser.add_argument('--max_depth', type=int,
default=3, help='max depth of recursion')
parser.add_argument('--img_size', nargs=2, type=int,
default=(100, 100), help='resolution of output figure')
return parser.parse_args()
def inner1d(x, y):
n, m = x.size()
return torch.bmm(x.view(n, 1, m), y.view(n, m, 1)).squeeze(2)
def normalize(x):
return x / np.linalg.norm(x)
def rotate(mat, rad, axis):
rot_mats = np.array([
[1., 0., 0.],
[0., np.cos(rad), -np.sin(rad)],
[0., np.sin(rad), np.cos(rad)],
[np.cos(rad), 0., np.sin(rad)],
[0., 1., 0.],
[-np.sin(rad), 0., np.cos(rad)],
[np.cos(rad), -np.sin(rad), 0.],
[np.sin(rad), np.cos(rad), 0.],
[0., 0., 1.]
])
rot_mat = {
'x': rot_mats[:3, :],
'y': rot_mats[3:6, :],
'z': rot_mats[6:, :]
}[axis]
return np.dot(mat, rot_mat)
class RayTracer(object):
def __init__(self, mat_c, mat_p, mat_n, mat_e, mat_spec, mat_refl, mat_refr):
self.num_tris = mat_p.shape[0]
self.mat_c = np.clip(mat_c, 0., 1.)
self.mat_p = torch.cuda.FloatTensor(mat_p)
self.mat_n = torch.cuda.FloatTensor(mat_n)
self.mat_e = torch.cuda.FloatTensor(mat_e)
self.mat_spec = torch.cuda.FloatTensor(mat_spec)
self.mat_refl = mat_refl
self.mat_refr = torch.cuda.FloatTensor(mat_refr)
# vertex color = averaged color of neighboring triangles
vtx_dict = {}
mat_p_rnd = np.around(mat_p, 4)
for idx, vtx in enumerate(mat_p_rnd.reshape(-1, 3)):
vtx = tuple(vtx)
if vtx not in vtx_dict:
vtx_dict[vtx] = []
vtx_dict[vtx].append(np.unravel_index(idx, mat_p_rnd.shape[:-1]))
self.mat_vtx_c = np.empty(mat_p_rnd.shape, dtype=np.float32)
for vtx, idx in vtx_dict.items():
idx = np.array(idx)
self.mat_vtx_c[idx[:, 0], idx[:, 1], :] = self.mat_c[idx[:, 0]].mean(0)
# speed up intersection test
self.v0 = self.mat_p[:, 2] - self.mat_p[:, 0]
self.v1 = self.mat_p[:, 1] - self.mat_p[:, 0]
self.d00 = inner1d(self.v0, self.v0)
self.d01 = inner1d(self.v0, self.v1)
self.d11 = inner1d(self.v1, self.v1)
self.invDenom = 1. / (self.d00 * self.d11 - self.d01 * self.d01)
def trace(self, img_size, ori, dst, scene, max_depth=3):
img = np.zeros(img_size + (3,))
x_coord = dst[0]
ray_ori, ray_drt = [], []
for row, z in enumerate(np.linspace(scene[1], scene[3], img_size[1])):
for col, y in enumerate(np.linspace(scene[0], scene[2], img_size[0])):
dst = np.array([x_coord, y, z])
drt = normalize(dst - ori)
ray_ori.append(torch.cuda.FloatTensor(dst))
ray_drt.append(torch.cuda.FloatTensor(drt))
img = [self._trace_ray(ori, drt, refl, depth) for ori, drt, refl, depth in zip(ray_ori, ray_drt, repeat(1.), repeat(max_depth))]
img = np.array(img)
img = np.clip(img, 0., 1.)
img = img.reshape(img_size + (-1,))
img = np.flipud(img)
return img
def _trace_ray(self, ray_ori, ray_drt, refl, depth):
ret = self._intersect(ray_ori, ray_drt)
if ret is None:
return np.array([0., 0., 0.], dtype=np.float32)
idx, pnt_int, (u, v) = ret
# vertice color interpolation
tri_int = self.mat_vtx_c[idx].squeeze()
color = (1. - u - v) * tri_int[0] + v * tri_int[1] + u * tri_int[2]
refl_int = self.mat_refl[idx]
if depth > 1 and refl * refl_int > 0.01:
new_ray_drt = ray_drt - 2 * torch.dot(ray_drt, self.mat_n[idx, :].squeeze()) * self.mat_n[idx, :]
new_ray_drt = (new_ray_drt / new_ray_drt.norm(p=2)).squeeze()
new_ray_ori = pnt_int + 1e-3 * new_ray_drt
ret_color = self._trace_ray(new_ray_ori, new_ray_drt, refl * refl_int, depth - 1)
color = color + ret_color
return refl * color
def _intersect(self, ray_ori, ray_drt):
ray_drt = torch.unsqueeze(ray_drt, 1)
denom = torch.mm(self.mat_n, ray_drt) + 1e-12
dist = inner1d(self.mat_p[:, 0, :].squeeze() -
ray_ori, self.mat_n) / denom
pnt_int = dist * ray_drt.t() + ray_ori
# Barycentric Technique
v2 = pnt_int - self.mat_p[:, 0]
d02 = inner1d(self.v0, v2)
d12 = inner1d(self.v1, v2)
u = (self.d11 * d02 - self.d01 * d12) * self.invDenom
v = (self.d00 * d12 - self.d01 * d02) * self.invDenom
# inside triangle
dist[~((u >= 0.) & (v >= 0.) & (u + v < 1.)) | (dist <= 0.)] = np.inf
if (dist == np.inf).all():
return None
_, idx_min = torch.min(dist, 0)
return idx_min, pnt_int[idx_min, :], (u[idx_min].cpu().numpy().squeeze(), v[idx_min].cpu().numpy().squeeze())
if __name__ == '__main__':
args = parse_args()
mat_c, mat_p, mat_e, mat_spec, mat_refl, mat_refr = XMLReader.read_tri(
args.input_file)
mat_p = mat_p.reshape(-1, 3)
mat_p = rotate(mat_p, np.pi * -.42, 'x')
mat_p = rotate(mat_p, np.pi * -.55, 'z')
mat_p = rotate(mat_p, np.pi * -.05, 'y')
mat_p = mat_p.reshape(-1, 3, 3)
# normal vector
mat_n = np.cross(mat_p[:, 0] - mat_p[:, 1], mat_p[:, 1] - mat_p[:, 2])
mat_n = mat_n / np.expand_dims(np.linalg.norm(mat_n, axis=1), axis=1)
tracer = RayTracer(mat_c, mat_p, mat_n, mat_e,
mat_spec, mat_refl, mat_refr)
img_size = tuple(args.img_size)
ori = np.array([40., 0., 0.], dtype=np.float32)
dst = np.array([20., 0., 0.], dtype=np.float32)
scene = (-15, -15, 10, 10)
img = tracer.trace(img_size, ori, dst, scene, max_depth=args.max_depth)
imsave(args.output_file, img)