FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavfilter/vf_nnedi.c
Date: 2024-10-17 22:24:08
Exec Total Coverage
Lines: 0 549 0.0%
Functions: 0 37 0.0%
Branches: 0 252 0.0%
Line Branch Exec Source
1 /*
2 * Copyright (C) 2010-2011 Kevin Stone
3 * Copyright (C) 2016 Paul B Mahol
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License along
18 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
20 */
21
22 #include <float.h>
23
24 #include "libavutil/common.h"
25 #include "libavutil/file_open.h"
26 #include "libavutil/float_dsp.h"
27 #include "libavutil/imgutils.h"
28 #include "libavutil/mem.h"
29 #include "libavutil/mem_internal.h"
30 #include "libavutil/opt.h"
31 #include "libavutil/pixdesc.h"
32 #include "avfilter.h"
33 #include "filters.h"
34 #include "video.h"
35
36 static const size_t NNEDI_WEIGHTS_SIZE = 13574928;
37 static const uint8_t NNEDI_XDIM[] = { 8, 16, 32, 48, 8, 16, 32 };
38 static const uint8_t NNEDI_YDIM[] = { 6, 6, 6, 6, 4, 4, 4 };
39 static const uint16_t NNEDI_NNS[] = { 16, 32, 64, 128, 256 };
40
41 typedef struct PrescreenerCoefficients {
42 DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4];
43 DECLARE_ALIGNED(32, float, bias_l0)[4];
44
45 DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
46 DECLARE_ALIGNED(32, float, bias_l1)[4];
47
48 DECLARE_ALIGNED(32, float, kernel_l2)[4][8];
49 DECLARE_ALIGNED(32, float, bias_l2)[4];
50 } PrescreenerCoefficients;
51
52 typedef struct PredictorCoefficients {
53 int xdim, ydim, nns, nsize;
54 float *data;
55 float *softmax_q1;
56 float *elliott_q1;
57 float *softmax_bias_q1;
58 float *elliott_bias_q1;
59 float *softmax_q2;
60 float *elliott_q2;
61 float *softmax_bias_q2;
62 float *elliott_bias_q2;
63 } PredictorCoefficients;
64
65 typedef struct NNEDIContext {
66 const AVClass *class;
67
68 char *weights_file;
69
70 AVFrame *prev;
71 int eof;
72 int64_t pts;
73
74 AVFloatDSPContext *fdsp;
75 int depth;
76 int nb_planes;
77 int nb_threads;
78 int linesize[4];
79 int planewidth[4];
80 int planeheight[4];
81 int field_n;
82
83 PrescreenerCoefficients prescreener[4];
84 PredictorCoefficients coeffs[2][5][7];
85
86 float half;
87 float in_scale;
88 float out_scale;
89
90 // Parameters
91 int deint;
92 int field;
93 int process_plane;
94 int nsize;
95 int nnsparam;
96 int qual;
97 int etype;
98 int pscrn;
99
100 int input_size;
101 uint8_t **prescreen_buf;
102 float **input_buf;
103 float **output_buf;
104
105 void (*read)(const uint8_t *src, float *dst,
106 int src_stride, int dst_stride,
107 int width, int height, float scale);
108 void (*write)(const float *src, uint8_t *dst,
109 int src_stride, int dst_stride,
110 int width, int height, int depth, float scale);
111 void (*prescreen[2])(AVFilterContext *ctx,
112 const void *src, ptrdiff_t src_stride,
113 uint8_t *prescreen, int N,
114 const PrescreenerCoefficients *const coeffs);
115 } NNEDIContext;
116
117 #define OFFSET(x) offsetof(NNEDIContext, x)
118 #define RFLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
119 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
120
121 static const AVOption nnedi_options[] = {
122 {"weights", "set weights file", OFFSET(weights_file), AV_OPT_TYPE_STRING, {.str="nnedi3_weights.bin"}, 0, 0, FLAGS },
123 {"deint", "set which frames to deinterlace", OFFSET(deint), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, .unit = "deint" },
124 {"all", "deinterlace all frames", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "deint" },
125 {"interlaced", "only deinterlace frames marked as interlaced", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "deint" },
126 {"field", "set mode of operation", OFFSET(field), AV_OPT_TYPE_INT, {.i64=-1}, -2, 3, RFLAGS, .unit = "field" },
127 {"af", "use frame flags, both fields", 0, AV_OPT_TYPE_CONST, {.i64=-2}, 0, 0, RFLAGS, .unit = "field" },
128 {"a", "use frame flags, single field", 0, AV_OPT_TYPE_CONST, {.i64=-1}, 0, 0, RFLAGS, .unit = "field" },
129 {"t", "use top field only", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "field" },
130 {"b", "use bottom field only", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "field" },
131 {"tf", "use both fields, top first", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, .unit = "field" },
132 {"bf", "use both fields, bottom first", 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, .unit = "field" },
133 {"planes", "set which planes to process", OFFSET(process_plane), AV_OPT_TYPE_INT, {.i64=7}, 0, 15, RFLAGS },
134 {"nsize", "set size of local neighborhood around each pixel, used by the predictor neural network", OFFSET(nsize), AV_OPT_TYPE_INT, {.i64=6}, 0, 6, RFLAGS, .unit = "nsize" },
135 {"s8x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "nsize" },
136 {"s16x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "nsize" },
137 {"s32x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, .unit = "nsize" },
138 {"s48x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, .unit = "nsize" },
139 {"s8x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, .unit = "nsize" },
140 {"s16x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=5}, 0, 0, RFLAGS, .unit = "nsize" },
141 {"s32x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=6}, 0, 0, RFLAGS, .unit = "nsize" },
142 {"nns", "set number of neurons in predictor neural network", OFFSET(nnsparam), AV_OPT_TYPE_INT, {.i64=1}, 0, 4, RFLAGS, .unit = "nns" },
143 {"n16", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "nns" },
144 {"n32", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "nns" },
145 {"n64", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, .unit = "nns" },
146 {"n128", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, .unit = "nns" },
147 {"n256", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, .unit = "nns" },
148 {"qual", "set quality", OFFSET(qual), AV_OPT_TYPE_INT, {.i64=1}, 1, 2, RFLAGS, .unit = "qual" },
149 {"fast", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "qual" },
150 {"slow", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, .unit = "qual" },
151 {"etype", "set which set of weights to use in the predictor", OFFSET(etype), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, .unit = "etype" },
152 {"a", "weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "etype" },
153 {"abs","weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "etype" },
154 {"s", "weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "etype" },
155 {"mse","weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "etype" },
156 {"pscrn", "set prescreening", OFFSET(pscrn), AV_OPT_TYPE_INT, {.i64=2}, 0, 4, RFLAGS, .unit = "pscrn" },
157 {"none", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, .unit = "pscrn" },
158 {"original", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, .unit = "pscrn" },
159 {"new", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, .unit = "pscrn" },
160 {"new2", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, .unit = "pscrn" },
161 {"new3", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, .unit = "pscrn" },
162 { NULL }
163 };
164
165 AVFILTER_DEFINE_CLASS(nnedi);
166
167 ✗ static int config_output(AVFilterLink *outlink)
168 {
169 ✗ AVFilterContext *ctx = outlink->src;
170 ✗ const NNEDIContext *const s = ctx->priv;
171
172 ✗ outlink->time_base = av_mul_q(ctx->inputs[0]->time_base, (AVRational){1, 2});
173 ✗ outlink->w = ctx->inputs[0]->w;
174 ✗ outlink->h = ctx->inputs[0]->h;
175
176 ✗ if (s->field == -2 || s->field > 1) {
177 ✗ FilterLink *il = ff_filter_link(ctx->inputs[0]);
178 ✗ FilterLink *ol = ff_filter_link(outlink);
179 ✗ ol->frame_rate = av_mul_q(il->frame_rate, (AVRational){2, 1});
180 }
181
182 ✗ return 0;
183 }
184
185 static const enum AVPixelFormat pix_fmts[] = {
186 AV_PIX_FMT_GRAY8,
187 AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10, AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
188 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
189 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
190 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
191 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
192 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
193 AV_PIX_FMT_YUVJ411P,
194 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
195 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRAP,
196 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
197 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
198 AV_PIX_FMT_YUV440P10,
199 AV_PIX_FMT_YUV420P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV444P12,
200 AV_PIX_FMT_YUV440P12,
201 AV_PIX_FMT_YUV420P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV444P14,
202 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
203 AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
204 AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
205 AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
206 AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
207 AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
208 AV_PIX_FMT_NONE
209 };
210
211 ✗ static float dot_dsp(const NNEDIContext *const s, const float *kernel, const float *input,
212 int n, float scale, float bias)
213 {
214 float sum, y;
215
216 ✗ sum = s->fdsp->scalarproduct_float(kernel, input, n);
217
218 ✗ y = sum * scale + bias + 1e-20f;
219
220 ✗ return y;
221 }
222
223 ✗ static float elliott(float x)
224 {
225 ✗ return x / (1.0f + fabsf(x));
226 }
227
228 ✗ static void transform_elliott(float *input, int size)
229 {
230 ✗ for (int i = 0; i < size; i++)
231 ✗ input[i] = elliott(input[i]);
232 ✗ }
233
234 ✗ static void process_old(AVFilterContext *ctx,
235 const void *src, ptrdiff_t src_stride,
236 uint8_t *prescreen, int N,
237 const PrescreenerCoefficients *const m_data)
238 {
239 ✗ NNEDIContext *s = ctx->priv;
240 ✗ const float *src_p = src;
241
242 // Adjust source pointer to point to top-left of filter window.
243 ✗ const float *window = src_p - 2 * src_stride - 5;
244
245 ✗ for (int j = 0; j < N; j++) {
246 ✗ LOCAL_ALIGNED_32(float, input, [48]);
247 float state[12];
248
249 ✗ for (int i = 0; i < 4; i++)
250 ✗ memcpy(input + i * 12, window + i * src_stride + j, 12 * sizeof(float));
251
252 // Layer 0.
253 ✗ for (int n = 0; n < 4; n++)
254 ✗ state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 48, 1.0f, m_data->bias_l0[n]);
255 ✗ transform_elliott(state + 1, 3);
256
257 // Layer 1.
258 ✗ for (int n = 0; n < 4; n++)
259 ✗ state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
260 ✗ transform_elliott(state + 4, 3);
261
262 // Layer 2.
263 ✗ for (int n = 0; n < 4; n++)
264 ✗ state[n + 8] = dot_dsp(s, m_data->kernel_l2[n], state, 8, 1.0f, m_data->bias_l2[n]);
265
266 ✗ prescreen[j] = FFMAX(state[10], state[11]) <= FFMAX(state[8], state[9]) ? 255 : 0;
267 }
268 ✗ }
269
270 ✗ static void process_new(AVFilterContext *ctx,
271 const void *src, ptrdiff_t src_stride,
272 uint8_t *prescreen, int N,
273 const PrescreenerCoefficients *const m_data)
274 {
275 ✗ NNEDIContext *s = ctx->priv;
276 ✗ const float *src_p = src;
277
278 // Adjust source pointer to point to top-left of filter window.
279 ✗ const float *window = src_p - 2 * src_stride - 6;
280
281 ✗ for (int j = 0; j < N; j += 4) {
282 ✗ LOCAL_ALIGNED_32(float, input, [64]);
283 float state[8];
284
285 ✗ for (int i = 0; i < 4; i++)
286 ✗ memcpy(input + i * 16, window + i * src_stride + j, 16 * sizeof(float));
287
288 ✗ for (int n = 0; n < 4; n++)
289 ✗ state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 64, 1.0f, m_data->bias_l0[n]);
290 ✗ transform_elliott(state, 4);
291
292 ✗ for (int n = 0; n < 4; n++)
293 ✗ state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
294
295 ✗ for (int n = 0; n < 4; n++)
296 ✗ prescreen[j + n] = state[n + 4] > 0.f;
297 }
298 ✗ }
299
300 ✗ static int filter_offset(int nn, const PredictorCoefficients *const model)
301 {
302 ✗ return nn * model->nsize;
303 }
304
305 ✗ static const float *softmax_q1_filter(int nn,
306 const PredictorCoefficients *const model)
307 {
308 ✗ return model->softmax_q1 + filter_offset(nn, model);
309 }
310
311 ✗ static const float *elliott_q1_filter(int nn,
312 const PredictorCoefficients *const model)
313 {
314 ✗ return model->elliott_q1 + filter_offset(nn, model);
315 }
316
317 ✗ static const float *softmax_q2_filter(int nn,
318 const PredictorCoefficients *const model)
319 {
320 ✗ return model->softmax_q2 + filter_offset(nn, model);
321 }
322
323 ✗ static const float *elliott_q2_filter(int nn,
324 const PredictorCoefficients *const model)
325 {
326 ✗ return model->elliott_q2 + filter_offset(nn, model);
327 }
328
329 ✗ static void gather_input(const float *src, ptrdiff_t src_stride,
330 float *buf, float mstd[4],
331 const PredictorCoefficients *const model)
332 {
333 ✗ const float scale = 1.f / model->nsize;
334 ✗ float sum = 0.f;
335 ✗ float sum_sq = 0.f;
336 float tmp;
337
338 ✗ for (int i = 0; i < model->ydim; i++) {
339 ✗ memcpy(buf, src, model->xdim * sizeof(float));
340
341 ✗ for (int j = 0; j < model->xdim; j++) {
342 ✗ const float val = src[j];
343
344 ✗ sum += val;
345 ✗ sum_sq += val * val;
346 }
347
348 ✗ src += src_stride;
349 ✗ buf += model->xdim;
350 }
351
352 ✗ mstd[0] = sum * scale;
353 ✗ mstd[3] = 0.f;
354
355 ✗ tmp = sum_sq * scale - mstd[0] * mstd[0];
356 ✗ if (tmp < FLT_EPSILON) {
357 ✗ mstd[1] = 0.0f;
358 ✗ mstd[2] = 0.0f;
359 } else {
360 ✗ mstd[1] = sqrtf(tmp);
361 ✗ mstd[2] = 1.0f / mstd[1];
362 }
363 ✗ }
364
365 ✗ static float softmax_exp(float x)
366 {
367 ✗ return expf(av_clipf(x, -80.f, 80.f));
368 }
369
370 ✗ static void transform_softmax_exp(float *input, int size)
371 {
372 ✗ for (int i = 0; i < size; i++)
373 ✗ input[i] = softmax_exp(input[i]);
374 ✗ }
375
376 ✗ static void wae5(const float *softmax, const float *el,
377 int n, float mstd[4])
378 {
379 ✗ float vsum = 0.0f, wsum = 0.0f;
380
381 ✗ for (int i = 0; i < n; i++) {
382 ✗ vsum += softmax[i] * elliott(el[i]);
383 ✗ wsum += softmax[i];
384 }
385
386 ✗ if (wsum > 1e-10f)
387 ✗ mstd[3] += (5.0f * vsum) / wsum * mstd[1] + mstd[0];
388 else
389 ✗ mstd[3] += mstd[0];
390 ✗ }
391
392 ✗ static void predictor(AVFilterContext *ctx,
393 const void *src, ptrdiff_t src_stride, void *dst,
394 const uint8_t *prescreen, int N,
395 const PredictorCoefficients *const model, int use_q2)
396 {
397 ✗ const NNEDIContext *const s = ctx->priv;
398 ✗ const float *src_p = src;
399 ✗ float *dst_p = dst;
400
401 // Adjust source pointer to point to top-left of filter window.
402 ✗ const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1);
403 ✗ const int filter_size = model->nsize;
404 ✗ const int nns = model->nns;
405
406 ✗ for (int i = 0; i < N; i++) {
407 ✗ LOCAL_ALIGNED_32(float, input, [48 * 6]);
408 float activation[256 * 2];
409 float mstd[4];
410 float scale;
411
412 ✗ if (prescreen[i])
413 ✗ continue;
414
415 ✗ gather_input(window + i, src_stride, input, mstd, model);
416 ✗ scale = mstd[2];
417
418 ✗ for (int nn = 0; nn < nns; nn++)
419 ✗ activation[nn] = dot_dsp(s, softmax_q1_filter(nn, model), input, filter_size, scale, model->softmax_bias_q1[nn]);
420
421 ✗ for (int nn = 0; nn < nns; nn++)
422 ✗ activation[nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]);
423
424 ✗ transform_softmax_exp(activation, nns);
425 ✗ wae5(activation, activation + nns, nns, mstd);
426
427 ✗ if (use_q2) {
428 ✗ for (int nn = 0; nn < nns; nn++)
429 ✗ activation[nn] = dot_dsp(s, softmax_q2_filter(nn, model), input, filter_size, scale, model->softmax_bias_q2[nn]);
430
431 ✗ for (int nn = 0; nn < nns; nn++)
432 ✗ activation[nns + nn] = dot_dsp(s, elliott_q2_filter(nn, model), input, filter_size, scale, model->elliott_bias_q2[nn]);
433
434 ✗ transform_softmax_exp(activation, nns);
435 ✗ wae5(activation, activation + nns, nns, mstd);
436 }
437
438 ✗ dst_p[i] = mstd[3] * (use_q2 ? 0.5f : 1.f);
439 }
440 ✗ }
441
442 ✗ static void read_bytes(const uint8_t *src, float *dst,
443 int src_stride, int dst_stride,
444 int width, int height, float scale)
445 {
446 ✗ for (int y = 0; y < height; y++) {
447 ✗ for (int x = 0; x < 32; x++)
448 ✗ dst[-x - 1] = src[x];
449
450 ✗ for (int x = 0; x < width; x++)
451 ✗ dst[x] = src[x];
452
453 ✗ for (int x = 0; x < 32; x++)
454 ✗ dst[width + x] = src[width - x - 1];
455
456 ✗ dst += dst_stride;
457 ✗ src += src_stride;
458 }
459 ✗ }
460
461 ✗ static void read_words(const uint8_t *srcp, float *dst,
462 int src_stride, int dst_stride,
463 int width, int height, float scale)
464 {
465 ✗ const uint16_t *src = (const uint16_t *)srcp;
466
467 ✗ src_stride /= 2;
468
469 ✗ for (int y = 0; y < height; y++) {
470 ✗ for (int x = 0; x < 32; x++)
471 ✗ dst[-x - 1] = src[x] * scale;
472
473 ✗ for (int x = 0; x < width; x++)
474 ✗ dst[x] = src[x] * scale;
475
476 ✗ for (int x = 0; x < 32; x++)
477 ✗ dst[width + x] = src[width - x - 1] * scale;
478
479 ✗ dst += dst_stride;
480 ✗ src += src_stride;
481 }
482 ✗ }
483
484 ✗ static void write_bytes(const float *src, uint8_t *dst,
485 int src_stride, int dst_stride,
486 int width, int height, int depth,
487 float scale)
488 {
489 ✗ for (int y = 0; y < height; y++) {
490 ✗ for (int x = 0; x < width; x++)
491 ✗ dst[x] = av_clip_uint8(src[x]);
492
493 ✗ dst += dst_stride;
494 ✗ src += src_stride;
495 }
496 ✗ }
497
498 ✗ static void write_words(const float *src, uint8_t *dstp,
499 int src_stride, int dst_stride,
500 int width, int height, int depth,
501 float scale)
502 {
503 ✗ uint16_t *dst = (uint16_t *)dstp;
504
505 ✗ dst_stride /= 2;
506
507 ✗ for (int y = 0; y < height; y++) {
508 ✗ for (int x = 0; x < width; x++)
509 ✗ dst[x] = av_clip_uintp2_c(src[x] * scale, depth);
510
511 ✗ dst += dst_stride;
512 ✗ src += src_stride;
513 }
514 ✗ }
515
516 ✗ static void interpolation(const void *src, ptrdiff_t src_stride,
517 void *dst, const uint8_t *prescreen, int n)
518 {
519 ✗ const float *src_p = src;
520 ✗ float *dst_p = dst;
521 ✗ const float *window = src_p - 2 * src_stride;
522
523 ✗ for (int i = 0; i < n; i++) {
524 ✗ float accum = 0.0f;
525
526 ✗ if (!prescreen[i])
527 ✗ continue;
528
529 ✗ accum += (-3.0f / 32.0f) * window[0 * src_stride + i];
530 ✗ accum += (19.0f / 32.0f) * window[1 * src_stride + i];
531 ✗ accum += (19.0f / 32.0f) * window[2 * src_stride + i];
532 ✗ accum += (-3.0f / 32.0f) * window[3 * src_stride + i];
533
534 ✗ dst_p[i] = accum;
535 }
536 ✗ }
537
538 ✗ static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
539 {
540 ✗ const NNEDIContext *const s = ctx->priv;
541 ✗ AVFrame *out = arg;
542 ✗ AVFrame *in = s->prev;
543 ✗ const float in_scale = s->in_scale;
544 ✗ const float out_scale = s->out_scale;
545 ✗ const int depth = s->depth;
546 ✗ const int interlaced = !!(in->flags & AV_FRAME_FLAG_INTERLACED);
547 ✗ const int tff = s->field_n == (s->field < 0 ? interlaced ? (in->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST) : 1 :
548 ✗ (s->field & 1) ^ 1);
549
550
551 ✗ for (int p = 0; p < s->nb_planes; p++) {
552 ✗ const int height = s->planeheight[p];
553 ✗ const int width = s->planewidth[p];
554 ✗ const int slice_start = 2 * ((height / 2 * jobnr) / nb_jobs);
555 ✗ const int slice_end = 2 * ((height / 2 * (jobnr+1)) / nb_jobs);
556 ✗ const uint8_t *src_data = in->data[p];
557 ✗ uint8_t *dst_data = out->data[p];
558 ✗ uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
559 ✗ const int src_linesize = in->linesize[p];
560 ✗ const int dst_linesize = out->linesize[p];
561 ✗ uint8_t *prescreen_buf = s->prescreen_buf[jobnr];
562 ✗ float *srcbuf = s->input_buf[jobnr];
563 ✗ const int srcbuf_stride = width + 64;
564 ✗ float *dstbuf = s->output_buf[jobnr];
565 ✗ const int dstbuf_stride = width;
566 ✗ const int slice_height = (slice_end - slice_start) / 2;
567 ✗ const int last_slice = slice_end == height;
568 const uint8_t *in_line;
569 uint8_t *out_line;
570 int y_out;
571
572 ✗ if (!(s->process_plane & (1 << p))) {
573 ✗ av_image_copy_plane(dst, out->linesize[p],
574 ✗ in->data[p] + slice_start * in->linesize[p],
575 in->linesize[p],
576 ✗ s->linesize[p], slice_end - slice_start);
577 ✗ continue;
578 }
579
580 ✗ y_out = slice_start + (tff ^ (slice_start & 1));
581 ✗ in_line = src_data + (y_out * src_linesize);
582 ✗ out_line = dst_data + (y_out * dst_linesize);
583
584 ✗ while (y_out < slice_end) {
585 ✗ memcpy(out_line, in_line, s->linesize[p]);
586 ✗ y_out += 2;
587 ✗ in_line += src_linesize * 2;
588 ✗ out_line += dst_linesize * 2;
589 }
590
591 ✗ y_out = slice_start + ((!tff) ^ (slice_start & 1));
592
593 ✗ s->read(src_data + FFMAX(y_out - 5, tff) * src_linesize,
594 srcbuf + 32,
595 src_linesize * 2, srcbuf_stride,
596 width, 1, in_scale);
597 ✗ srcbuf += srcbuf_stride;
598
599 ✗ s->read(src_data + FFMAX(y_out - 3, tff) * src_linesize,
600 srcbuf + 32,
601 src_linesize * 2, srcbuf_stride,
602 width, 1, in_scale);
603 ✗ srcbuf += srcbuf_stride;
604
605 ✗ s->read(src_data + FFMAX(y_out - 1, tff) * src_linesize,
606 srcbuf + 32,
607 src_linesize * 2, srcbuf_stride,
608 width, 1, in_scale);
609 ✗ srcbuf += srcbuf_stride;
610
611 ✗ in_line = src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize;
612 ✗ out_line = dst_data + (y_out * dst_linesize);
613
614 ✗ s->read(in_line, srcbuf + 32, src_linesize * 2, srcbuf_stride,
615 width, slice_height - last_slice, in_scale);
616
617 ✗ y_out += (slice_height - last_slice) * 2;
618
619 ✗ s->read(src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize,
620 ✗ srcbuf + 32 + srcbuf_stride * (slice_height - last_slice),
621 src_linesize * 2, srcbuf_stride,
622 width, 1, in_scale);
623
624 ✗ s->read(src_data + FFMIN(y_out + 3, height - 1 - !tff) * src_linesize,
625 ✗ srcbuf + 32 + srcbuf_stride * (slice_height + 1 - last_slice),
626 src_linesize * 2, srcbuf_stride,
627 width, 1, in_scale);
628
629 ✗ s->read(src_data + FFMIN(y_out + 5, height - 1 - !tff) * src_linesize,
630 ✗ srcbuf + 32 + srcbuf_stride * (slice_height + 2 - last_slice),
631 src_linesize * 2, srcbuf_stride,
632 width, 1, in_scale);
633
634 ✗ for (int y = 0; y < slice_end - slice_start; y += 2) {
635 ✗ if (s->pscrn > 0)
636 ✗ s->prescreen[s->pscrn > 1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
637 srcbuf_stride, prescreen_buf, width,
638 ✗ &s->prescreener[s->pscrn - 1]);
639
640 ✗ predictor(ctx,
641 ✗ srcbuf + (y / 2) * srcbuf_stride + 32,
642 srcbuf_stride,
643 ✗ dstbuf + (y / 2) * dstbuf_stride,
644 prescreen_buf, width,
645 ✗ &s->coeffs[s->etype][s->nnsparam][s->nsize], s->qual == 2);
646
647 ✗ if (s->pscrn > 0)
648 ✗ interpolation(srcbuf + (y / 2) * srcbuf_stride + 32,
649 srcbuf_stride,
650 ✗ dstbuf + (y / 2) * dstbuf_stride,
651 prescreen_buf, width);
652 }
653
654 ✗ s->write(dstbuf, out_line, dstbuf_stride, dst_linesize * 2,
655 width, slice_height, depth, out_scale);
656 }
657
658 ✗ return 0;
659 }
660
661 ✗ static int get_frame(AVFilterContext *ctx, int is_second)
662 {
663 ✗ NNEDIContext *s = ctx->priv;
664 ✗ AVFilterLink *outlink = ctx->outputs[0];
665 AVFrame *dst;
666
667 ✗ dst = ff_get_video_buffer(outlink, outlink->w, outlink->h);
668 ✗ if (!dst)
669 ✗ return AVERROR(ENOMEM);
670 ✗ av_frame_copy_props(dst, s->prev);
671 #if FF_API_INTERLACED_FRAME
672 FF_DISABLE_DEPRECATION_WARNINGS
673 ✗ dst->interlaced_frame = 0;
674 FF_ENABLE_DEPRECATION_WARNINGS
675 #endif
676 ✗ dst->flags &= ~AV_FRAME_FLAG_INTERLACED;
677 ✗ dst->pts = s->pts;
678
679 ✗ ff_filter_execute(ctx, filter_slice, dst, NULL,
680 ✗ FFMIN(s->planeheight[1] / 2, s->nb_threads));
681
682 ✗ if (s->field == -2 || s->field > 1)
683 ✗ s->field_n = !s->field_n;
684
685 ✗ return ff_filter_frame(outlink, dst);
686 }
687
688 ✗ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
689 {
690 ✗ AVFilterContext *ctx = inlink->dst;
691 ✗ NNEDIContext *s = ctx->priv;
692 int ret;
693
694 ✗ if (!s->prev) {
695 ✗ s->prev = in;
696 ✗ return 0;
697 }
698
699 ✗ if ((s->deint && !(s->prev->flags & AV_FRAME_FLAG_INTERLACED)) || ctx->is_disabled) {
700 ✗ s->prev->pts *= 2;
701 ✗ ret = ff_filter_frame(ctx->outputs[0], s->prev);
702 ✗ s->prev = in;
703 ✗ return ret;
704 }
705
706 ✗ s->pts = s->prev->pts * 2;
707 ✗ ret = get_frame(ctx, 0);
708 ✗ if (ret < 0 || (s->field > -2 && s->field < 2)) {
709 ✗ av_frame_free(&s->prev);
710 ✗ s->prev = in;
711 ✗ return ret;
712 }
713
714 ✗ s->pts = s->prev->pts + in->pts;
715 ✗ ret = get_frame(ctx, 1);
716 ✗ av_frame_free(&s->prev);
717 ✗ s->prev = in;
718 ✗ return ret;
719 }
720
721 ✗ static int request_frame(AVFilterLink *link)
722 {
723 ✗ AVFilterContext *ctx = link->src;
724 ✗ NNEDIContext *s = ctx->priv;
725 int ret;
726
727 ✗ if (s->eof)
728 ✗ return AVERROR_EOF;
729
730 ✗ ret = ff_request_frame(ctx->inputs[0]);
731
732 ✗ if (ret == AVERROR_EOF && s->prev) {
733 ✗ AVFrame *next = av_frame_clone(s->prev);
734 ✗ FilterLink *l = ff_filter_link(ctx->outputs[0]);
735
736 ✗ if (!next)
737 ✗ return AVERROR(ENOMEM);
738
739 ✗ next->pts = s->prev->pts + av_rescale_q(1, av_inv_q(l->frame_rate),
740 ✗ ctx->outputs[0]->time_base);
741 ✗ s->eof = 1;
742
743 ✗ ret = filter_frame(ctx->inputs[0], next);
744 ✗ } else if (ret < 0) {
745 ✗ return ret;
746 }
747
748 ✗ return ret;
749 }
750
751 ✗ static void copy_weights(float *dst, int n, const float **data)
752 {
753 ✗ memcpy(dst, *data, n * sizeof(float));
754 ✗ *data += n;
755 ✗ }
756
757 ✗ static float *allocate(float **ptr, int size)
758 {
759 ✗ float *ret = *ptr;
760
761 ✗ *ptr += size;
762
763 ✗ return ret;
764 }
765
766 ✗ static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int nns)
767 {
768 ✗ int filter_size = nns * xdim * ydim;
769 ✗ int bias_size = nns;
770 float *data;
771
772 ✗ data = av_calloc(filter_size + bias_size, 4 * sizeof(float));
773 ✗ if (!data)
774 ✗ return AVERROR(ENOMEM);
775
776 ✗ coeffs->data = data;
777 ✗ coeffs->xdim = xdim;
778 ✗ coeffs->ydim = ydim;
779 ✗ coeffs->nsize = xdim * ydim;
780 ✗ coeffs->nns = nns;
781
782 ✗ coeffs->softmax_q1 = allocate(&data, filter_size);
783 ✗ coeffs->elliott_q1 = allocate(&data, filter_size);
784 ✗ coeffs->softmax_bias_q1 = allocate(&data, bias_size);
785 ✗ coeffs->elliott_bias_q1 = allocate(&data, bias_size);
786
787 ✗ coeffs->softmax_q2 = allocate(&data, filter_size);
788 ✗ coeffs->elliott_q2 = allocate(&data, filter_size);
789 ✗ coeffs->softmax_bias_q2 = allocate(&data, bias_size);
790 ✗ coeffs->elliott_bias_q2 = allocate(&data, bias_size);
791
792 ✗ return 0;
793 }
794
795 ✗ static int read_weights(AVFilterContext *ctx, const float *bdata)
796 {
797 ✗ NNEDIContext *s = ctx->priv;
798 int ret;
799
800 ✗ copy_weights(&s->prescreener[0].kernel_l0[0][0], 4 * 48, &bdata);
801 ✗ copy_weights(s->prescreener[0].bias_l0, 4, &bdata);
802
803 ✗ copy_weights(&s->prescreener[0].kernel_l1[0][0], 4 * 4, &bdata);
804 ✗ copy_weights(s->prescreener[0].bias_l1, 4, &bdata);
805
806 ✗ copy_weights(&s->prescreener[0].kernel_l2[0][0], 4 * 8, &bdata);
807 ✗ copy_weights(s->prescreener[0].bias_l2, 4, &bdata);
808
809 ✗ for (int i = 0; i < 3; i++) {
810 ✗ PrescreenerCoefficients *data = &s->prescreener[i + 1];
811 float kernel_l0_shuffled[4 * 64];
812 float kernel_l1_shuffled[4 * 4];
813
814 ✗ copy_weights(kernel_l0_shuffled, 4 * 64, &bdata);
815 ✗ copy_weights(data->bias_l0, 4, &bdata);
816
817 ✗ copy_weights(kernel_l1_shuffled, 4 * 4, &bdata);
818 ✗ copy_weights(data->bias_l1, 4, &bdata);
819
820 ✗ for (int n = 0; n < 4; n++) {
821 ✗ for (int k = 0; k < 64; k++)
822 ✗ data->kernel_l0[n][k] = kernel_l0_shuffled[(k / 8) * 32 + n * 8 + k % 8];
823 ✗ for (int k = 0; k < 4; k++)
824 ✗ data->kernel_l1[n][k] = kernel_l1_shuffled[k * 4 + n];
825 }
826 }
827
828 ✗ for (int m = 0; m < 2; m++) {
829 // Grouping by neuron count.
830 ✗ for (int i = 0; i < 5; i++) {
831 ✗ const int nns = NNEDI_NNS[i];
832
833 // Grouping by window size.
834 ✗ for (int j = 0; j < 7; j++) {
835 ✗ PredictorCoefficients *model = &s->coeffs[m][i][j];
836 ✗ const int xdim = NNEDI_XDIM[j];
837 ✗ const int ydim = NNEDI_YDIM[j];
838 ✗ const int filter_size = xdim * ydim;
839
840 ✗ ret = allocate_model(model, xdim, ydim, nns);
841 ✗ if (ret < 0)
842 ✗ return ret;
843
844 // Quality 1 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
845 ✗ copy_weights(model->softmax_q1, nns * filter_size, &bdata);
846 ✗ copy_weights(model->elliott_q1, nns * filter_size, &bdata);
847
848 // Quality 1 model bias. NNS[i] * 2 coefficients.
849 ✗ copy_weights(model->softmax_bias_q1, nns, &bdata);
850 ✗ copy_weights(model->elliott_bias_q1, nns, &bdata);
851
852 // Quality 2 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
853 ✗ copy_weights(model->softmax_q2, nns * filter_size, &bdata);
854 ✗ copy_weights(model->elliott_q2, nns * filter_size, &bdata);
855
856 // Quality 2 model bias. NNS[i] * 2 coefficients.
857 ✗ copy_weights(model->softmax_bias_q2, nns, &bdata);
858 ✗ copy_weights(model->elliott_bias_q2, nns, &bdata);
859 }
860 }
861 }
862
863 ✗ return 0;
864 }
865
866 ✗ static float mean(const float *input, int size)
867 {
868 ✗ float sum = 0.f;
869
870 ✗ for (int i = 0; i < size; i++)
871 ✗ sum += input[i];
872
873 ✗ return sum / size;
874 }
875
876 ✗ static void transform(float *input, int size, float mean, float half)
877 {
878 ✗ for (int i = 0; i < size; i++)
879 ✗ input[i] = (input[i] - mean) / half;
880 ✗ }
881
882 ✗ static void subtract_mean_old(PrescreenerCoefficients *coeffs, float half)
883 {
884 ✗ for (int n = 0; n < 4; n++) {
885 ✗ float m = mean(coeffs->kernel_l0[n], 48);
886
887 ✗ transform(coeffs->kernel_l0[n], 48, m, half);
888 }
889 ✗ }
890
891 ✗ static void subtract_mean_new(PrescreenerCoefficients *coeffs, float half)
892 {
893 ✗ for (int n = 0; n < 4; n++) {
894 ✗ float m = mean(coeffs->kernel_l0[n], 64);
895
896 ✗ transform(coeffs->kernel_l0[n], 64, m, half);
897 }
898 ✗ }
899
900 ✗ static void subtract_mean_predictor(PredictorCoefficients *model)
901 {
902 ✗ const int filter_size = model->nsize;
903 ✗ const int nns = model->nns;
904 ✗ const float scale = 1.f / nns;
905
906 double softmax_means[256]; // Average of individual softmax filters.
907 double elliott_means[256]; // Average of individual elliott filters.
908 ✗ double mean_filter[48 * 6] = { 0 }; // Pointwise average of all softmax filters.
909 double mean_bias;
910
911 // Quality 1.
912 ✗ for (int nn = 0; nn < nns; nn++) {
913 ✗ softmax_means[nn] = mean(model->softmax_q1 + nn * filter_size, filter_size);
914 ✗ elliott_means[nn] = mean(model->elliott_q1 + nn * filter_size, filter_size);
915
916 ✗ for (int k = 0; k < filter_size; k++)
917 ✗ mean_filter[k] += model->softmax_q1[nn * filter_size + k] - softmax_means[nn];
918 }
919
920 ✗ for (int k = 0; k < filter_size; k++)
921 ✗ mean_filter[k] *= scale;
922
923 ✗ mean_bias = mean(model->softmax_bias_q1, nns);
924
925 ✗ for (int nn = 0; nn < nns; nn++) {
926 ✗ for (int k = 0; k < filter_size; k++) {
927 ✗ model->softmax_q1[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
928 ✗ model->elliott_q1[nn * filter_size + k] -= elliott_means[nn];
929 }
930 ✗ model->softmax_bias_q1[nn] -= mean_bias;
931 }
932
933 // Quality 2.
934 ✗ memset(mean_filter, 0, sizeof(mean_filter));
935
936 ✗ for (int nn = 0; nn < nns; nn++) {
937 ✗ softmax_means[nn] = mean(model->softmax_q2 + nn * filter_size, filter_size);
938 ✗ elliott_means[nn] = mean(model->elliott_q2 + nn * filter_size, filter_size);
939
940 ✗ for (int k = 0; k < filter_size; k++) {
941 ✗ mean_filter[k] += model->softmax_q2[nn * filter_size + k] - softmax_means[nn];
942 }
943 }
944
945 ✗ for (int k = 0; k < filter_size; k++)
946 ✗ mean_filter[k] *= scale;
947
948 ✗ mean_bias = mean(model->softmax_bias_q2, nns);
949
950 ✗ for (int nn = 0; nn < nns; nn++) {
951 ✗ for (int k = 0; k < filter_size; k++) {
952 ✗ model->softmax_q2[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
953 ✗ model->elliott_q2[nn * filter_size + k] -= elliott_means[nn];
954 }
955
956 ✗ model->softmax_bias_q2[nn] -= mean_bias;
957 }
958 ✗ }
959
960 ✗ static av_cold int init(AVFilterContext *ctx)
961 {
962 ✗ NNEDIContext *s = ctx->priv;
963 ✗ FILE *weights_file = NULL;
964 int64_t weights_size;
965 float *bdata;
966 size_t bytes_read;
967 ✗ int ret = 0;
968
969 ✗ weights_file = avpriv_fopen_utf8(s->weights_file, "rb");
970 ✗ if (!weights_file) {
971 ✗ av_log(ctx, AV_LOG_ERROR, "No weights file provided, aborting!\n");
972 ✗ return AVERROR(EINVAL);
973 }
974
975 ✗ if (fseek(weights_file, 0, SEEK_END)) {
976 ✗ av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the end of weights file.\n");
977 ✗ fclose(weights_file);
978 ✗ return AVERROR(EINVAL);
979 }
980
981 ✗ weights_size = ftell(weights_file);
982
983 ✗ if (weights_size == -1) {
984 ✗ fclose(weights_file);
985 ✗ av_log(ctx, AV_LOG_ERROR, "Couldn't get size of weights file.\n");
986 ✗ return AVERROR(EINVAL);
987 ✗ } else if (weights_size != NNEDI_WEIGHTS_SIZE) {
988 ✗ fclose(weights_file);
989 ✗ av_log(ctx, AV_LOG_ERROR, "Unexpected weights file size.\n");
990 ✗ return AVERROR(EINVAL);
991 }
992
993 ✗ if (fseek(weights_file, 0, SEEK_SET)) {
994 ✗ fclose(weights_file);
995 ✗ av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the start of weights file.\n");
996 ✗ return AVERROR(EINVAL);
997 }
998
999 ✗ bdata = av_malloc(NNEDI_WEIGHTS_SIZE);
1000 ✗ if (!bdata) {
1001 ✗ fclose(weights_file);
1002 ✗ return AVERROR(ENOMEM);
1003 }
1004
1005 ✗ bytes_read = fread(bdata, 1, NNEDI_WEIGHTS_SIZE, weights_file);
1006 ✗ if (bytes_read != NNEDI_WEIGHTS_SIZE) {
1007 ✗ fclose(weights_file);
1008 ✗ ret = AVERROR_INVALIDDATA;
1009 ✗ av_log(ctx, AV_LOG_ERROR, "Couldn't read weights file.\n");
1010 ✗ goto fail;
1011 }
1012
1013 ✗ fclose(weights_file);
1014
1015 ✗ s->fdsp = avpriv_float_dsp_alloc(0);
1016 ✗ if (!s->fdsp) {
1017 ✗ ret = AVERROR(ENOMEM);
1018 ✗ goto fail;
1019 }
1020
1021 ✗ ret = read_weights(ctx, bdata);
1022 ✗ if (ret < 0)
1023 ✗ goto fail;
1024
1025 ✗ fail:
1026 ✗ av_free(bdata);
1027 ✗ return ret;
1028 }
1029
1030 ✗ static int config_input(AVFilterLink *inlink)
1031 {
1032 ✗ AVFilterContext *ctx = inlink->dst;
1033 ✗ NNEDIContext *s = ctx->priv;
1034 ✗ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
1035 int ret;
1036
1037 ✗ s->depth = desc->comp[0].depth;
1038 ✗ s->nb_threads = ff_filter_get_nb_threads(ctx);
1039 ✗ s->nb_planes = av_pix_fmt_count_planes(inlink->format);
1040 ✗ if ((ret = av_image_fill_linesizes(s->linesize, inlink->format, inlink->w)) < 0)
1041 ✗ return ret;
1042
1043 ✗ s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
1044 ✗ s->planewidth[0] = s->planewidth[3] = inlink->w;
1045 ✗ s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
1046 ✗ s->planeheight[0] = s->planeheight[3] = inlink->h;
1047
1048 ✗ s->half = ((1 << 8) - 1) / 2.f;
1049 ✗ s->out_scale = 1 << (s->depth - 8);
1050 ✗ s->in_scale = 1.f / s->out_scale;
1051
1052 ✗ switch (s->depth) {
1053 ✗ case 8:
1054 ✗ s->read = read_bytes;
1055 ✗ s->write = write_bytes;
1056 ✗ break;
1057 ✗ default:
1058 ✗ s->read = read_words;
1059 ✗ s->write = write_words;
1060 ✗ break;
1061 }
1062
1063 ✗ subtract_mean_old(&s->prescreener[0], s->half);
1064 ✗ subtract_mean_new(&s->prescreener[1], s->half);
1065 ✗ subtract_mean_new(&s->prescreener[2], s->half);
1066 ✗ subtract_mean_new(&s->prescreener[3], s->half);
1067
1068 ✗ s->prescreen[0] = process_old;
1069 ✗ s->prescreen[1] = process_new;
1070
1071 ✗ for (int i = 0; i < 2; i++) {
1072 ✗ for (int j = 0; j < 5; j++) {
1073 ✗ for (int k = 0; k < 7; k++)
1074 ✗ subtract_mean_predictor(&s->coeffs[i][j][k]);
1075 }
1076 }
1077
1078 ✗ s->input_size = (s->planewidth[0] + 64) * (s->planeheight[0] + 6);
1079 ✗ s->input_buf = av_calloc(s->nb_threads, sizeof(*s->input_buf));
1080 ✗ if (!s->input_buf)
1081 ✗ return AVERROR(ENOMEM);
1082
1083 ✗ for (int i = 0; i < s->nb_threads; i++) {
1084 ✗ s->input_buf[i] = av_calloc(s->input_size, sizeof(**s->input_buf));
1085 ✗ if (!s->input_buf[i])
1086 ✗ return AVERROR(ENOMEM);
1087 }
1088
1089 ✗ s->output_buf = av_calloc(s->nb_threads, sizeof(*s->output_buf));
1090 ✗ if (!s->output_buf)
1091 ✗ return AVERROR(ENOMEM);
1092
1093 ✗ for (int i = 0; i < s->nb_threads; i++) {
1094 ✗ s->output_buf[i] = av_calloc(s->input_size, sizeof(**s->output_buf));
1095 ✗ if (!s->output_buf[i])
1096 ✗ return AVERROR(ENOMEM);
1097 }
1098
1099 ✗ s->prescreen_buf = av_calloc(s->nb_threads, sizeof(*s->prescreen_buf));
1100 ✗ if (!s->prescreen_buf)
1101 ✗ return AVERROR(ENOMEM);
1102
1103 ✗ for (int i = 0; i < s->nb_threads; i++) {
1104 ✗ s->prescreen_buf[i] = av_calloc(s->planewidth[0], sizeof(**s->prescreen_buf));
1105 ✗ if (!s->prescreen_buf[i])
1106 ✗ return AVERROR(ENOMEM);
1107 }
1108
1109 ✗ return 0;
1110 }
1111
1112 ✗ static av_cold void uninit(AVFilterContext *ctx)
1113 {
1114 ✗ NNEDIContext *s = ctx->priv;
1115
1116 ✗ for (int i = 0; i < s->nb_threads && s->prescreen_buf; i++)
1117 ✗ av_freep(&s->prescreen_buf[i]);
1118
1119 ✗ av_freep(&s->prescreen_buf);
1120
1121 ✗ for (int i = 0; i < s->nb_threads && s->input_buf; i++)
1122 ✗ av_freep(&s->input_buf[i]);
1123
1124 ✗ av_freep(&s->input_buf);
1125
1126 ✗ for (int i = 0; i < s->nb_threads && s->output_buf; i++)
1127 ✗ av_freep(&s->output_buf[i]);
1128
1129 ✗ av_freep(&s->output_buf);
1130 ✗ av_freep(&s->fdsp);
1131
1132 ✗ for (int i = 0; i < 2; i++) {
1133 ✗ for (int j = 0; j < 5; j++) {
1134 ✗ for (int k = 0; k < 7; k++) {
1135 ✗ av_freep(&s->coeffs[i][j][k].data);
1136 }
1137 }
1138 }
1139
1140 ✗ av_frame_free(&s->prev);
1141 ✗ }
1142
1143 static const AVFilterPad inputs[] = {
1144 {
1145 .name = "default",
1146 .type = AVMEDIA_TYPE_VIDEO,
1147 .filter_frame = filter_frame,
1148 .config_props = config_input,
1149 },
1150 };
1151
1152 static const AVFilterPad outputs[] = {
1153 {
1154 .name = "default",
1155 .type = AVMEDIA_TYPE_VIDEO,
1156 .config_props = config_output,
1157 .request_frame = request_frame,
1158 },
1159 };
1160
1161 const AVFilter ff_vf_nnedi = {
1162 .name = "nnedi",
1163 .description = NULL_IF_CONFIG_SMALL("Apply neural network edge directed interpolation intra-only deinterlacer."),
1164 .priv_size = sizeof(NNEDIContext),
1165 .priv_class = &nnedi_class,
1166 .init = init,
1167 .uninit = uninit,
1168 FILTER_INPUTS(inputs),
1169 FILTER_OUTPUTS(outputs),
1170 FILTER_PIXFMTS_ARRAY(pix_fmts),
1171 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
1172 .process_command = ff_filter_process_command,
1173 };
1174