-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrasterize_n_G_color.slang
191 lines (173 loc) · 9.65 KB
/
rasterize_n_G_color.slang
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
191
[Differentiable]
[PreferRecompute]
float3x3 getRotationMatrix(float4 r)
{
float3x3 R;
float4 r_n = r / length(r);
float q0 = r_n[0];
float q1 = r_n[1];
float q2 = r_n[2];
float q3 = r_n[3];
R[0] = float3(1 - 2 * q2 * q2 - 2 * q3 * q3, 2 * q1 * q2 - 2 * q0 * q3, 2 * q1 * q3 + 2 * q0 * q2);
R[1] = float3(2 * q1 * q2 + 2 * q0 * q3, 1 - 2 * q1 * q1 - 2 * q3 * q3, 2 * q2 * q3 - 2 * q0 * q1);
R[2] = float3(2 * q1 * q3 - 2 * q0 * q2, 2 * q2 * q3 + 2 * q0 * q1, 1 - 2 * q1 * q1 - 2 * q2 * q2);
return R;
}
[Differentiable]
[PreferRecompute]
float3x3 getScalingMatrix(float3 s)
{
float3x3 T;
T[0] = float3(s.x, 0, 0);
T[1] = float3(0, s.y, 0);
T[2] = float3(0, 0, s.z);
return T;
}
[Differentiable]
[PreferRecompute]
float2x2 compute_Conv2D(float4 mean, float4x4 viewM, float3x3 Sig, float4 view_angle)
{
float4 t = mul(viewM, mean);
// float limx = 1.3f * view_angle[0];
// float limy = 1.3f * view_angle[1];
// float txtz = t.x / t.z;
// float tytz = t.y / t.z;
// float focal_x = view_angle[2];
// float focal_y = view_angle[3];
// t.x = min(limx, max(-limx, txtz)) * t.z;
// t.y = min(limy, max(-limy, tytz)) * t.z;
float3x3 J = float3x3(1.0 / t.z, 0, -t.x / (t.z * t.z), 0, 1.0 / t.z, -t.y / (t.z * t.z), 0, 0, 0);
// float3x3 J = float3x3(focal_x / t.z, 0, -(focal_x * t.x) / (t.z * t.z), 0, focal_y / t.z, -(focal_y * t.y) / (t.z * t.z), 0, 0, 0);
float3x3 W = float3x3(viewM[0][0], viewM[0][1], viewM[0][2], viewM[1][0], viewM[1][1], viewM[1][2], viewM[2][0], viewM[2][1], viewM[2][2]);
float3x3 T = mul(J, W);
float3x3 T_t = transpose(T);
float3x3 Conv2D_ori = mul(T, mul(Sig, T_t));
float2x2 Conv2D = float2x2(Conv2D_ori[0][0] + 0.3, Conv2D_ori[0][1], Conv2D_ori[1][0], Conv2D_ori[1][1] + 0.3);
return Conv2D;
}
[CudaKernel]
[Differentiable]
[AutoPyBindCUDA]
void rasterize(
DiffTensorView mean,
DiffTensorView s,
DiffTensorView r,
TensorView<float> viewM,
TensorView<float> projM,
TensorView<float> view_angle,
int gaussian_range,
DiffTensorView color,
DiffTensorView output)
{
uint3 globalIdx = cudaBlockIdx() * cudaBlockDim() + cudaThreadIdx();
if (globalIdx.x > output.size(0) || globalIdx.y > output.size(1))
return;
// float T = output.load(uint3(globalIdx.x, globalIdx.y, 4));
// if (T < 0.0001)
// return;
float T = 1.0;
// in the order of tan_fovx, tan_fovy, focal_x, focal_y
float4 fov_info = float4(no_diff(view_angle[0]), no_diff(view_angle[1]), no_diff(view_angle[2]), no_diff(view_angle[3]));
// float4x4 _viewM = float4x4(no_diff(viewM[0]), no_diff(viewM[1]), no_diff(viewM[2]), no_diff(viewM[3]));
// _viewM[0][0] = no_diff(viewM[uint2(0, 0)]);
float4x4 _viewM;
// = float4x4(no_diff(viewM[0]), no_diff(viewM[1]), no_diff(viewM[2]), no_diff(viewM[3]));
_viewM[0] = float4(no_diff(viewM[uint2(0, 0)]), no_diff(viewM[uint2(0, 1)]), no_diff(viewM[uint2(0, 2)]), no_diff(viewM[uint2(0, 3)]));
_viewM[1] = float4(no_diff(viewM[uint2(1, 0)]), no_diff(viewM[uint2(1, 1)]), no_diff(viewM[uint2(1, 2)]), no_diff(viewM[uint2(1, 3)]));
_viewM[2] = float4(no_diff(viewM[uint2(2, 0)]), no_diff(viewM[uint2(2, 1)]), no_diff(viewM[uint2(2, 2)]), no_diff(viewM[uint2(2, 3)]));
_viewM[3] = float4(no_diff(viewM[uint2(3, 0)]), no_diff(viewM[uint2(3, 1)]), no_diff(viewM[uint2(3, 2)]), no_diff(viewM[uint2(3, 3)]));
float4x4 _projM;
// _projM[0] = float4(no_diff(projM[uint2(1, 0)]), no_diff(projM[uint2(1, 1)]), no_diff(projM[uint2(1, 2)]), no_diff(projM[uint2(1, 3)]));
_projM[0] = float4(no_diff(projM[uint2(0, 0)]), no_diff(projM[uint2(0, 1)]), no_diff(projM[uint2(0, 2)]), no_diff(projM[uint2(0, 3)]));
_projM[1] = float4(no_diff(projM[uint2(1, 0)]), no_diff(projM[uint2(1, 1)]), no_diff(projM[uint2(1, 2)]), no_diff(projM[uint2(1, 3)]));
_projM[2] = float4(no_diff(projM[uint2(2, 0)]), no_diff(projM[uint2(2, 1)]), no_diff(projM[uint2(2, 2)]), no_diff(projM[uint2(2, 3)]));
_projM[3] = float4(no_diff(projM[uint2(3, 0)]), no_diff(projM[uint2(3, 1)]), no_diff(projM[uint2(3, 2)]), no_diff(projM[uint2(3, 3)]));
float4 color_track = float4(0.0, 0.0, 0.0, 0.0);
for (int i = 0; i < 32; i++)
{
if (i == gaussian_range)
break;
float4 mean_0 = float4(mean[uint2(i, 0)], mean[uint2(i, 1)], mean[uint2(i, 2)], mean[uint2(i, 3)]);
float4 color_0 = float4(color[uint2(i, 0)], color[uint2(i, 1)], color[uint2(i, 2)], color[uint2(i, 3)]);
float4 _r = float4(r[uint2(i, 0)], r[uint2(i, 1)], r[uint2(i, 2)], r[uint2(i, 3)]);
float3 _s = float3(s[uint2(i, 0)], s[uint2(i, 1)], s[uint2(i, 2)]);
float3x3 R = getRotationMatrix(_r);
float3x3 S = getScalingMatrix(_s);
float3x3 R_t = transpose(R);
float3x3 S_t = transpose(S);
float3x3 Sig = mul(R, mul(S, mul(S_t, R_t)));
float2x2 Conv2D = compute_Conv2D(mean_0, _viewM, Sig, fov_info);
float det = Conv2D[0][0] * Conv2D[1][1] - Conv2D[0][1] * Conv2D[1][0];
float inv_det = 1.0 / det;
float2x2 conic = float2x2(Conv2D[1][1] * inv_det, -Conv2D[0][1] * inv_det, -Conv2D[1][0] * inv_det, Conv2D[0][0] * inv_det);
float mid = 0.5 * (conic[0][0] + conic[1][1]);
float lambda1 = mid + sqrt(max(0.1, mid * mid - det));
float lambda2 = mid - sqrt(max(0.1, mid * mid - det));
float radius = ceil(3.0 * sqrt(max(lambda1, lambda2)));
float4 mean_camera = mul(_viewM, mean_0);
float4 mean_proj = mul(_projM, mean_camera);
float2 mean_ndc = float2(mean_proj.x / mean_proj.w, mean_proj.y / mean_proj.w);
float2 mean_screen = float2((mean_ndc.x + 1.0) * 0.5 * output.size(0), (mean_ndc.y + 1) * 0.5 * output.size(1));
float x_dis = float(globalIdx.x) - mean_screen.x;
float y_dis = float(globalIdx.y) - mean_screen.y;
float power_con = -0.5 * (conic[0][0] * x_dis * x_dis + 2 * conic[0][1] * x_dis * y_dis + conic[1][1] * y_dis * y_dis);
if (power_con > 0)
continue;
float alpha = min(0.99, color_0.w * exp(power_con));
if (alpha < 1.0 / 255.0)
continue;
float test_t = (1.0 - alpha) * T;
if (test_t < 0.0001)
continue;
color_track += color_0 * alpha * T;
T = test_t;
}
output.store(uint3(globalIdx.x, globalIdx.y, 0), color_track[0]);
output.store(uint3(globalIdx.x, globalIdx.y, 1), color_track[1]);
output.store(uint3(globalIdx.x, globalIdx.y, 2), color_track[2]);
output.store(uint3(globalIdx.x, globalIdx.y, 3), color_track[3]);
return;
// if (globalIdx.x == 0) {
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), no_diff(viewM[uint2(0,0)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), no_diff(viewM[uint2(0,1)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), no_diff(viewM[uint2(0,2)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), no_diff(viewM[uint2(0,3)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), _viewM[0][0]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), _viewM[0][1]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), _viewM[0][2]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), _viewM[0][3]);
// output.store(uint3(globalIdx.x, globalIdx.y, 0), mean_camera[0]);
// output.store(uint3(globalIdx.x, globalIdx.y, 1), mean_camera[1]);
// output.store(uint3(globalIdx.x, globalIdx.y, 2), mean_camera[2]);
// output.store(uint3(globalIdx.x, globalIdx.y, 3), mean_camera[3]);
// } else {
// if (globalIdx.x == 1) {
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), no_diff(projM[uint2(0,0)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), no_diff(projM[uint2(0,1)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), no_diff(projM[uint2(0,2)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), no_diff(projM[uint2(0,3)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), mean_proj[0]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), mean_proj[1]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), mean_proj[2]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), mean_proj[3]);
// output.store(uint3(globalIdx.x, globalIdx.y, 0), _viewM[2][0]);
// output.store(uint3(globalIdx.x, globalIdx.y, 1), _viewM[2][1]);
// output.store(uint3(globalIdx.x, globalIdx.y, 2), _viewM[2][2]);
// output.store(uint3(globalIdx.x, globalIdx.y, 3), _viewM[2][3]);
// } else {
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), no_diff(viewM[uint2(2,0)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), no_diff(viewM[uint2(2,1)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), no_diff(viewM[uint2(2,2)]));
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), no_diff(viewM[uint2(2,3)]));
// output.store(uint3(globalIdx.x, globalIdx.y, 0), mean_ndc[0]);
// output.store(uint3(globalIdx.x, globalIdx.y, 1), mean_ndc[1]);
// output.store(uint3(globalIdx.x, globalIdx.y, 2), mean_screen[0]);
// output.store(uint3(globalIdx.x, globalIdx.y, 3), mean_screen[1]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 0), _viewM[2][0]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 1), _viewM[2][1]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 2), _viewM[2][2]);
// // output.store(uint3(globalIdx.x, globalIdx.y, 3), _viewM[2][3]);
// }
// }
// return;
}