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
|
// 5x5 convolution with 25 samples
// Applies mat4 weights per sample
fn cnn_conv5x5(
tex: texture_2d<f32>,
samp: sampler,
uv: vec2<f32>,
resolution: vec2<f32>,
weights: array<mat4x4<f32>, 25>,
bias: vec4<f32>
) -> vec4<f32> {
let step = 1.0 / resolution;
var sum = bias;
var idx = 0;
for (var dy = -2; dy <= 2; dy++) {
for (var dx = -2; dx <= 2; dx++) {
let offset = vec2<f32>(f32(dx), f32(dy)) * step;
let sample = textureSample(tex, samp, uv + offset);
sum += weights[idx] * sample;
idx++;
}
}
return sum;
}
fn cnn_conv5x5_with_coord(
tex: texture_2d<f32>,
samp: sampler,
uv: vec2<f32>,
resolution: vec2<f32>,
rgba_weights: array<mat4x4<f32>, 25>,
coord_weights: mat2x4<f32>,
bias: vec4<f32>
) -> vec4<f32> {
let step = 1.0 / resolution;
var sum = bias;
sum += coord_weights * uv;
var idx = 0;
for (var dy = -2; dy <= 2; dy++) {
for (var dx = -2; dx <= 2; dx++) {
let offset = vec2<f32>(f32(dx), f32(dy)) * step;
let rgba = textureSample(tex, samp, uv + offset);
sum += rgba_weights[idx] * rgba;
idx++;
}
}
return sum;
}
// 5×5 variant for 7→4 channels (RGBD output)
// weights: array<array<f32, 8>, 100> (25 positions × 4 channels, each with 7 weights + bias)
fn cnn_conv5x5_7to4(
tex: texture_2d<f32>,
samp: sampler,
uv: vec2<f32>,
resolution: vec2<f32>,
original: vec4<f32>,
weights: array<array<f32, 8>, 100>
) -> vec4<f32> {
let step = 1.0 / resolution;
let gray_01 = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;
let gray = (gray_01 - 0.5) * 2.0;
let uv_norm = (uv - 0.5) * 2.0;
var sum = vec4<f32>(0.0);
var pos = 0;
for (var dy = -2; dy <= 2; dy++) {
for (var dx = -2; dx <= 2; dx++) {
let offset = vec2<f32>(f32(dx), f32(dy)) * step;
let rgbd_01 = textureSample(tex, samp, uv + offset);
let rgbd = (rgbd_01 - 0.5) * 2.0;
let inputs = array<f32, 7>(
rgbd.r, rgbd.g, rgbd.b, rgbd.a,
uv_norm.x, uv_norm.y, gray
);
for (var out_c = 0; out_c < 4; out_c++) {
let idx = pos * 4 + out_c;
var channel_sum = weights[idx][7];
for (var in_c = 0; in_c < 7; in_c++) {
channel_sum += weights[idx][in_c] * inputs[in_c];
}
sum[out_c] += channel_sum;
}
pos++;
}
}
return sum;
}
// 5×5 variant for 7→1 channel (scalar output)
// weights: array<array<f32, 8>, 25> (25 positions, each with 7 weights + bias)
fn cnn_conv5x5_7to1(
tex: texture_2d<f32>,
samp: sampler,
uv: vec2<f32>,
resolution: vec2<f32>,
original: vec4<f32>,
weights: array<array<f32, 8>, 25>
) -> f32 {
let step = 1.0 / resolution;
let gray_01 = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;
let gray = (gray_01 - 0.5) * 2.0;
let uv_norm = (uv - 0.5) * 2.0;
var sum = 0.0;
var pos = 0;
for (var dy = -2; dy <= 2; dy++) {
for (var dx = -2; dx <= 2; dx++) {
let offset = vec2<f32>(f32(dx), f32(dy)) * step;
let rgbd_01 = textureSample(tex, samp, uv + offset);
let rgbd = (rgbd_01 - 0.5) * 2.0;
sum += weights[pos][0] * rgbd.r;
sum += weights[pos][1] * rgbd.g;
sum += weights[pos][2] * rgbd.b;
sum += weights[pos][3] * rgbd.a;
sum += weights[pos][4] * uv_norm.x;
sum += weights[pos][5] * uv_norm.y;
sum += weights[pos][6] * gray;
sum += weights[pos][7]; // Bias
pos++;
}
}
return sum;
}
|
