forked from leejet/stable-diffusion.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathpreprocessing.hpp
More file actions
278 lines (247 loc) · 11.1 KB
/
Copy pathpreprocessing.hpp
File metadata and controls
278 lines (247 loc) · 11.1 KB
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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
#ifndef __PREPROCESSING_HPP__
#define __PREPROCESSING_HPP__
#include <cmath>
#include <limits>
#include "ggml_extend.hpp"
#define M_PI_ 3.14159265358979323846f
static inline int64_t preprocessing_offset_4d(const sd::Tensor<float>& tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
const auto& shape = tensor.shape();
int64_t n0 = shape.size() > 0 ? shape[0] : 1;
int64_t n1 = shape.size() > 1 ? shape[1] : 1;
int64_t n2 = shape.size() > 2 ? shape[2] : 1;
return ((i3 * n2 + i2) * n1 + i1) * n0 + i0;
}
static inline float preprocessing_get_4d(const sd::Tensor<float>& tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
return tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))];
}
static inline void preprocessing_set_4d(sd::Tensor<float>& tensor, float value, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))] = value;
}
static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t image) {
sd::Tensor<float> tensor({static_cast<int64_t>(image.width), static_cast<int64_t>(image.height), static_cast<int64_t>(image.channel), 1});
for (uint32_t y = 0; y < image.height; ++y) {
for (uint32_t x = 0; x < image.width; ++x) {
for (uint32_t c = 0; c < image.channel; ++c) {
preprocessing_set_4d(tensor, sd_image_get_f32(image, x, y, c), x, y, c, 0);
}
}
}
return tensor;
}
static inline void preprocessing_tensor_to_sd_image(const sd::Tensor<float>& tensor, uint8_t* image_data) {
GGML_ASSERT(tensor.dim() == 4);
GGML_ASSERT(tensor.shape()[3] == 1);
GGML_ASSERT(image_data != nullptr);
int width = static_cast<int>(tensor.shape()[0]);
int height = static_cast<int>(tensor.shape()[1]);
int channel = static_cast<int>(tensor.shape()[2]);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
for (int c = 0; c < channel; ++c) {
float value = preprocessing_get_4d(tensor, x, y, c, 0);
value = std::min(1.0f, std::max(0.0f, value));
image_data[(y * width + x) * channel + c] = static_cast<uint8_t>(std::round(value * 255.0f));
}
}
}
}
static inline sd::Tensor<float> gaussian_kernel_tensor(int kernel_size) {
sd::Tensor<float> kernel({kernel_size, kernel_size, 1, 1});
int ks_mid = kernel_size / 2;
float sigma = 1.4f;
float normal = 1.f / (2.0f * M_PI_ * std::pow(sigma, 2.0f));
for (int y = 0; y < kernel_size; ++y) {
float gx = static_cast<float>(-ks_mid + y);
for (int x = 0; x < kernel_size; ++x) {
float gy = static_cast<float>(-ks_mid + x);
float k = std::exp(-((gx * gx + gy * gy) / (2.0f * std::pow(sigma, 2.0f)))) * normal;
preprocessing_set_4d(kernel, k, x, y, 0, 0);
}
}
return kernel;
}
static inline sd::Tensor<float> convolve_tensor(const sd::Tensor<float>& input, const sd::Tensor<float>& kernel, int padding) {
GGML_ASSERT(input.dim() == 4);
GGML_ASSERT(kernel.dim() == 4);
GGML_ASSERT(input.shape()[3] == 1);
GGML_ASSERT(kernel.shape()[2] == 1);
GGML_ASSERT(kernel.shape()[3] == 1);
sd::Tensor<float> output(input.shape());
int64_t width = input.shape()[0];
int64_t height = input.shape()[1];
int64_t channels = input.shape()[2];
int64_t kernel_w = kernel.shape()[0];
int64_t kernel_h = kernel.shape()[1];
for (int64_t c = 0; c < channels; ++c) {
for (int64_t y = 0; y < height; ++y) {
for (int64_t x = 0; x < width; ++x) {
float sum = 0.0f;
for (int64_t ky = 0; ky < kernel_h; ++ky) {
int64_t iy = y + ky - padding;
if (iy < 0 || iy >= height) {
continue;
}
for (int64_t kx = 0; kx < kernel_w; ++kx) {
int64_t ix = x + kx - padding;
if (ix < 0 || ix >= width) {
continue;
}
sum += preprocessing_get_4d(input, ix, iy, c, 0) * preprocessing_get_4d(kernel, kx, ky, 0, 0);
}
}
preprocessing_set_4d(output, sum, x, y, c, 0);
}
}
}
return output;
}
static inline sd::Tensor<float> grayscale_tensor(const sd::Tensor<float>& rgb_img) {
GGML_ASSERT(rgb_img.dim() == 4);
GGML_ASSERT(rgb_img.shape()[2] >= 3);
sd::Tensor<float> grayscale({rgb_img.shape()[0], rgb_img.shape()[1], 1, rgb_img.shape()[3]});
for (int64_t iy = 0; iy < rgb_img.shape()[1]; ++iy) {
for (int64_t ix = 0; ix < rgb_img.shape()[0]; ++ix) {
float r = preprocessing_get_4d(rgb_img, ix, iy, 0, 0);
float g = preprocessing_get_4d(rgb_img, ix, iy, 1, 0);
float b = preprocessing_get_4d(rgb_img, ix, iy, 2, 0);
float gray = 0.2989f * r + 0.5870f * g + 0.1140f * b;
preprocessing_set_4d(grayscale, gray, ix, iy, 0, 0);
}
}
return grayscale;
}
static inline sd::Tensor<float> tensor_hypot(const sd::Tensor<float>& x, const sd::Tensor<float>& y) {
sd::tensor_check_same_shape(x, y);
sd::Tensor<float> out(x.shape());
for (int64_t i = 0; i < out.numel(); ++i) {
out[i] = std::sqrt(x[i] * x[i] + y[i] * y[i]);
}
return out;
}
static inline sd::Tensor<float> tensor_arctan2(const sd::Tensor<float>& x, const sd::Tensor<float>& y) {
sd::tensor_check_same_shape(x, y);
sd::Tensor<float> out(x.shape());
for (int64_t i = 0; i < out.numel(); ++i) {
out[i] = std::atan2(y[i], x[i]);
}
return out;
}
static inline void normalize_tensor(sd::Tensor<float>* g) {
GGML_ASSERT(g != nullptr);
if (g->empty()) {
return;
}
float max_value = -std::numeric_limits<float>::infinity();
for (int64_t i = 0; i < g->numel(); ++i) {
max_value = std::max(max_value, (*g)[i]);
}
if (max_value == 0.0f || !std::isfinite(max_value)) {
return;
}
*g *= (1.0f / max_value);
}
static inline sd::Tensor<float> non_max_supression(const sd::Tensor<float>& G, const sd::Tensor<float>& D) {
GGML_ASSERT(G.shape() == D.shape());
sd::Tensor<float> result = sd::Tensor<float>::zeros(G.shape());
for (int64_t iy = 1; iy < result.shape()[1] - 1; ++iy) {
for (int64_t ix = 1; ix < result.shape()[0] - 1; ++ix) {
float angle = preprocessing_get_4d(D, ix, iy, 0, 0) * 180.0f / M_PI_;
angle = angle < 0.0f ? angle + 180.0f : angle;
float q = 1.0f;
float r = 1.0f;
if ((0 >= angle && angle < 22.5f) || (157.5f >= angle && angle <= 180.0f)) {
q = preprocessing_get_4d(G, ix, iy + 1, 0, 0);
r = preprocessing_get_4d(G, ix, iy - 1, 0, 0);
} else if (22.5f >= angle && angle < 67.5f) {
q = preprocessing_get_4d(G, ix + 1, iy - 1, 0, 0);
r = preprocessing_get_4d(G, ix - 1, iy + 1, 0, 0);
} else if (67.5f >= angle && angle < 112.5f) {
q = preprocessing_get_4d(G, ix + 1, iy, 0, 0);
r = preprocessing_get_4d(G, ix - 1, iy, 0, 0);
} else if (112.5f >= angle && angle < 157.5f) {
q = preprocessing_get_4d(G, ix - 1, iy - 1, 0, 0);
r = preprocessing_get_4d(G, ix + 1, iy + 1, 0, 0);
}
float cur = preprocessing_get_4d(G, ix, iy, 0, 0);
preprocessing_set_4d(result, (cur >= q && cur >= r) ? cur : 0.0f, ix, iy, 0, 0);
}
}
return result;
}
static inline void threshold_hystersis(sd::Tensor<float>* img, float high_threshold, float low_threshold, float weak, float strong) {
GGML_ASSERT(img != nullptr);
if (img->empty()) {
return;
}
float max_value = -std::numeric_limits<float>::infinity();
for (int64_t i = 0; i < img->numel(); ++i) {
max_value = std::max(max_value, (*img)[i]);
}
float ht = max_value * high_threshold;
float lt = ht * low_threshold;
for (int64_t i = 0; i < img->numel(); ++i) {
float img_v = (*img)[i];
if (img_v >= ht) {
(*img)[i] = strong;
} else if (img_v <= ht && img_v >= lt) {
(*img)[i] = weak;
}
}
for (int64_t iy = 0; iy < img->shape()[1]; ++iy) {
for (int64_t ix = 0; ix < img->shape()[0]; ++ix) {
if (!(ix >= 3 && ix <= img->shape()[0] - 3 && iy >= 3 && iy <= img->shape()[1] - 3)) {
preprocessing_set_4d(*img, 0.0f, ix, iy, 0, 0);
}
}
}
for (int64_t iy = 1; iy < img->shape()[1] - 1; ++iy) {
for (int64_t ix = 1; ix < img->shape()[0] - 1; ++ix) {
float imd_v = preprocessing_get_4d(*img, ix, iy, 0, 0);
if (imd_v == weak) {
bool has_strong_neighbor =
preprocessing_get_4d(*img, ix + 1, iy - 1, 0, 0) == strong ||
preprocessing_get_4d(*img, ix + 1, iy, 0, 0) == strong ||
preprocessing_get_4d(*img, ix, iy - 1, 0, 0) == strong ||
preprocessing_get_4d(*img, ix, iy + 1, 0, 0) == strong ||
preprocessing_get_4d(*img, ix - 1, iy - 1, 0, 0) == strong ||
preprocessing_get_4d(*img, ix - 1, iy, 0, 0) == strong;
preprocessing_set_4d(*img, has_strong_neighbor ? strong : 0.0f, ix, iy, 0, 0);
}
}
}
}
bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
float kX[9] = {
-1, 0, 1,
-2, 0, 2,
-1, 0, 1};
float kY[9] = {
1, 2, 1,
0, 0, 0,
-1, -2, -1};
sd::Tensor<float> gkernel = gaussian_kernel_tensor(5);
sd::Tensor<float> sf_kx({3, 3, 1, 1}, std::vector<float>(kX, kX + 9));
sd::Tensor<float> sf_ky({3, 3, 1, 1}, std::vector<float>(kY, kY + 9));
sd::Tensor<float> image = sd_image_to_preprocessing_tensor(img);
sd::Tensor<float> image_gray = grayscale_tensor(image);
image_gray = convolve_tensor(image_gray, gkernel, 2);
sd::Tensor<float> iX = convolve_tensor(image_gray, sf_kx, 1);
sd::Tensor<float> iY = convolve_tensor(image_gray, sf_ky, 1);
sd::Tensor<float> G = tensor_hypot(iX, iY);
normalize_tensor(&G);
sd::Tensor<float> theta = tensor_arctan2(iX, iY);
image_gray = non_max_supression(G, theta);
threshold_hystersis(&image_gray, high_threshold, low_threshold, weak, strong);
for (uint32_t iy = 0; iy < img.height; ++iy) {
for (uint32_t ix = 0; ix < img.width; ++ix) {
float gray = preprocessing_get_4d(image_gray, ix, iy, 0, 0);
gray = inverse ? 1.0f - gray : gray;
for (uint32_t c = 0; c < img.channel; ++c) {
preprocessing_set_4d(image, gray, ix, iy, c, 0);
}
}
}
preprocessing_tensor_to_sd_image(image, img.data);
return true;
}
#endif // __PREPROCESSING_HPP__