FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavfilter/vf_vif.c
Date: 2022-11-26 13:19:19
Exec Total Coverage
Lines: 0 294 0.0%
Branches: 0 120 0.0%

Line Branch Exec Source
1 /*
2 * Copyright (c) 2017 Ronald S. Bultje <rsbultje@gmail.com>
3 * Copyright (c) 2017 Ashish Pratap Singh <ashk43712@gmail.com>
4 * Copyright (c) 2021 Paul B Mahol
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * Calculate VIF between two input videos.
26 */
27
28 #include <float.h>
29
30 #include "libavutil/avstring.h"
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
33 #include "avfilter.h"
34 #include "framesync.h"
35 #include "drawutils.h"
36 #include "formats.h"
37 #include "internal.h"
38 #include "video.h"
39
40 #define NUM_DATA_BUFS 13
41
42 typedef struct VIFContext {
43 const AVClass *class;
44 FFFrameSync fs;
45 const AVPixFmtDescriptor *desc;
46 int width;
47 int height;
48 int nb_threads;
49 float factor;
50 float *data_buf[NUM_DATA_BUFS];
51 float **temp;
52 float *ref_data;
53 float *main_data;
54 double vif_sum[4];
55 double vif_min[4];
56 double vif_max[4];
57 uint64_t nb_frames;
58 } VIFContext;
59
60 #define OFFSET(x) offsetof(VIFContext, x)
61
62 static const AVOption vif_options[] = {
63 { NULL }
64 };
65
66 AVFILTER_DEFINE_CLASS(vif);
67
68 static const uint8_t vif_filter1d_width1[4] = { 17, 9, 5, 3 };
69
70 static const float vif_filter1d_table[4][17] =
71 {
72 {
73 0.00745626912, 0.0142655009, 0.0250313189, 0.0402820669, 0.0594526194,
74 0.0804751068, 0.0999041125, 0.113746084, 0.118773937, 0.113746084,
75 0.0999041125, 0.0804751068, 0.0594526194, 0.0402820669, 0.0250313189,
76 0.0142655009, 0.00745626912
77 },
78 {
79 0.0189780835, 0.0558981746, 0.120920904, 0.192116052, 0.224173605,
80 0.192116052, 0.120920904, 0.0558981746, 0.0189780835
81 },
82 {
83 0.054488685, 0.244201347, 0.402619958, 0.244201347, 0.054488685
84 },
85 {
86 0.166378498, 0.667243004, 0.166378498
87 }
88 };
89
90 typedef struct ThreadData {
91 const float *filter;
92 const float *src;
93 float *dst;
94 int w, h;
95 int src_stride;
96 int dst_stride;
97 int filter_width;
98 float **temp;
99 } ThreadData;
100
101 static void vif_dec2(const float *src, float *dst, int w, int h,
102 int src_stride, int dst_stride)
103 {
104 const int dst_px_stride = dst_stride / 2;
105
106 for (int i = 0; i < h / 2; i++) {
107 for (int j = 0; j < w / 2; j++)
108 dst[i * dst_px_stride + j] = src[(i * 2) * src_stride + (j * 2)];
109 }
110 }
111
112 static void vif_statistic(const float *mu1_sq, const float *mu2_sq,
113 const float *mu1_mu2, const float *xx_filt,
114 const float *yy_filt, const float *xy_filt,
115 float *num, float *den, int w, int h)
116 {
117 static const float sigma_nsq = 2;
118 float mu1_sq_val, mu2_sq_val, mu1_mu2_val, xx_filt_val, yy_filt_val, xy_filt_val;
119 float sigma1_sq, sigma2_sq, sigma12, g, sv_sq, eps = 1.0e-10f;
120 float gain_limit = 100.f;
121 float num_val, den_val;
122 float accum_num = 0.0f;
123 float accum_den = 0.0f;
124
125 for (int i = 0; i < h; i++) {
126 float accum_inner_num = 0.f;
127 float accum_inner_den = 0.f;
128
129 for (int j = 0; j < w; j++) {
130 mu1_sq_val = mu1_sq[i * w + j];
131 mu2_sq_val = mu2_sq[i * w + j];
132 mu1_mu2_val = mu1_mu2[i * w + j];
133 xx_filt_val = xx_filt[i * w + j];
134 yy_filt_val = yy_filt[i * w + j];
135 xy_filt_val = xy_filt[i * w + j];
136
137 sigma1_sq = xx_filt_val - mu1_sq_val;
138 sigma2_sq = yy_filt_val - mu2_sq_val;
139 sigma12 = xy_filt_val - mu1_mu2_val;
140
141 sigma1_sq = FFMAX(sigma1_sq, 0.0f);
142 sigma2_sq = FFMAX(sigma2_sq, 0.0f);
143 sigma12 = FFMAX(sigma12, 0.0f);
144
145 g = sigma12 / (sigma1_sq + eps);
146 sv_sq = sigma2_sq - g * sigma12;
147
148 if (sigma1_sq < eps) {
149 g = 0.0f;
150 sv_sq = sigma2_sq;
151 sigma1_sq = 0.0f;
152 }
153
154 if (sigma2_sq < eps) {
155 g = 0.0f;
156 sv_sq = 0.0f;
157 }
158
159 if (g < 0.0f) {
160 sv_sq = sigma2_sq;
161 g = 0.0f;
162 }
163 sv_sq = FFMAX(sv_sq, eps);
164
165 g = FFMIN(g, gain_limit);
166
167 num_val = log2f(1.0f + g * g * sigma1_sq / (sv_sq + sigma_nsq));
168 den_val = log2f(1.0f + sigma1_sq / sigma_nsq);
169
170 if (isnan(den_val))
171 num_val = den_val = 1.f;
172
173 accum_inner_num += num_val;
174 accum_inner_den += den_val;
175 }
176
177 accum_num += accum_inner_num;
178 accum_den += accum_inner_den;
179 }
180
181 num[0] = accum_num;
182 den[0] = accum_den;
183 }
184
185 static void vif_xx_yy_xy(const float *x, const float *y, float *xx, float *yy,
186 float *xy, int w, int h)
187 {
188 for (int i = 0; i < h; i++) {
189 for (int j = 0; j < w; j++) {
190 float xval = x[j];
191 float yval = y[j];
192 float xxval = xval * xval;
193 float yyval = yval * yval;
194 float xyval = xval * yval;
195
196 xx[j] = xxval;
197 yy[j] = yyval;
198 xy[j] = xyval;
199 }
200
201 xx += w;
202 yy += w;
203 xy += w;
204 x += w;
205 y += w;
206 }
207 }
208
209 static int vif_filter1d(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
210 {
211 ThreadData *td = arg;
212 const float *filter = td->filter;
213 const float *src = td->src;
214 float *dst = td->dst;
215 int w = td->w;
216 int h = td->h;
217 int src_stride = td->src_stride;
218 int dst_stride = td->dst_stride;
219 int filt_w = td->filter_width;
220 float *temp = td->temp[jobnr];
221 const int slice_start = (h * jobnr) / nb_jobs;
222 const int slice_end = (h * (jobnr+1)) / nb_jobs;
223
224 for (int i = slice_start; i < slice_end; i++) {
225 /** Vertical pass. */
226 for (int j = 0; j < w; j++) {
227 float sum = 0.f;
228
229 if (i >= filt_w / 2 && i < h - filt_w / 2 - 1) {
230 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
231 const float filt_coeff = filter[filt_i];
232 float img_coeff;
233 int ii = i - filt_w / 2 + filt_i;
234
235 img_coeff = src[ii * src_stride + j];
236 sum += filt_coeff * img_coeff;
237 }
238 } else {
239 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
240 const float filt_coeff = filter[filt_i];
241 int ii = i - filt_w / 2 + filt_i;
242 float img_coeff;
243
244 ii = ii < 0 ? -ii : (ii >= h ? 2 * h - ii - 1 : ii);
245
246 img_coeff = src[ii * src_stride + j];
247 sum += filt_coeff * img_coeff;
248 }
249 }
250
251 temp[j] = sum;
252 }
253
254 /** Horizontal pass. */
255 for (int j = 0; j < w; j++) {
256 float sum = 0.f;
257
258 if (j >= filt_w / 2 && j < w - filt_w / 2 - 1) {
259 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
260 const float filt_coeff = filter[filt_j];
261 int jj = j - filt_w / 2 + filt_j;
262 float img_coeff;
263
264 img_coeff = temp[jj];
265 sum += filt_coeff * img_coeff;
266 }
267 } else {
268 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
269 const float filt_coeff = filter[filt_j];
270 int jj = j - filt_w / 2 + filt_j;
271 float img_coeff;
272
273 jj = jj < 0 ? -jj : (jj >= w ? 2 * w - jj - 1 : jj);
274
275 img_coeff = temp[jj];
276 sum += filt_coeff * img_coeff;
277 }
278 }
279
280 dst[i * dst_stride + j] = sum;
281 }
282 }
283
284 return 0;
285 }
286
287 static int compute_vif2(AVFilterContext *ctx,
288 const float *ref, const float *main, int w, int h,
289 int ref_stride, int main_stride, float *score,
290 float *const data_buf[NUM_DATA_BUFS], float **temp,
291 int gnb_threads)
292 {
293 ThreadData td;
294 float *ref_scale = data_buf[0];
295 float *main_scale = data_buf[1];
296 float *ref_sq = data_buf[2];
297 float *main_sq = data_buf[3];
298 float *ref_main = data_buf[4];
299 float *mu1 = data_buf[5];
300 float *mu2 = data_buf[6];
301 float *mu1_sq = data_buf[7];
302 float *mu2_sq = data_buf[8];
303 float *mu1_mu2 = data_buf[9];
304 float *ref_sq_filt = data_buf[10];
305 float *main_sq_filt = data_buf[11];
306 float *ref_main_filt = data_buf[12];
307
308 float *curr_ref_scale = (float *)ref;
309 float *curr_main_scale = (float *)main;
310 int curr_ref_stride = ref_stride;
311 int curr_main_stride = main_stride;
312
313 float num = 0.f;
314 float den = 0.f;
315
316 for (int scale = 0; scale < 4; scale++) {
317 const float *filter = vif_filter1d_table[scale];
318 int filter_width = vif_filter1d_width1[scale];
319 const int nb_threads = FFMIN(h, gnb_threads);
320 int buf_valid_w = w;
321 int buf_valid_h = h;
322
323 td.filter = filter;
324 td.filter_width = filter_width;
325
326 if (scale > 0) {
327 td.src = curr_ref_scale;
328 td.dst = mu1;
329 td.w = w;
330 td.h = h;
331 td.src_stride = curr_ref_stride;
332 td.dst_stride = w;
333 td.temp = temp;
334 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
335
336 td.src = curr_main_scale;
337 td.dst = mu2;
338 td.src_stride = curr_main_stride;
339 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
340
341 vif_dec2(mu1, ref_scale, buf_valid_w, buf_valid_h, w, w);
342 vif_dec2(mu2, main_scale, buf_valid_w, buf_valid_h, w, w);
343
344 w = buf_valid_w / 2;
345 h = buf_valid_h / 2;
346
347 buf_valid_w = w;
348 buf_valid_h = h;
349
350 curr_ref_scale = ref_scale;
351 curr_main_scale = main_scale;
352
353 curr_ref_stride = w;
354 curr_main_stride = w;
355 }
356
357 td.src = curr_ref_scale;
358 td.dst = mu1;
359 td.w = w;
360 td.h = h;
361 td.src_stride = curr_ref_stride;
362 td.dst_stride = w;
363 td.temp = temp;
364 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
365
366 td.src = curr_main_scale;
367 td.dst = mu2;
368 td.src_stride = curr_main_stride;
369 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
370
371 vif_xx_yy_xy(mu1, mu2, mu1_sq, mu2_sq, mu1_mu2, w, h);
372
373 vif_xx_yy_xy(curr_ref_scale, curr_main_scale, ref_sq, main_sq, ref_main, w, h);
374
375 td.src = ref_sq;
376 td.dst = ref_sq_filt;
377 td.src_stride = w;
378 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
379
380 td.src = main_sq;
381 td.dst = main_sq_filt;
382 td.src_stride = w;
383 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
384
385 td.src = ref_main;
386 td.dst = ref_main_filt;
387 ff_filter_execute(ctx, vif_filter1d, &td, NULL, nb_threads);
388
389 vif_statistic(mu1_sq, mu2_sq, mu1_mu2, ref_sq_filt, main_sq_filt,
390 ref_main_filt, &num, &den, w, h);
391
392 score[scale] = den <= FLT_EPSILON ? 1.f : num / den;
393 }
394
395 return 0;
396 }
397
398 #define offset_fn(type, bits) \
399 static void offset_##bits##bit(VIFContext *s, \
400 const AVFrame *ref, \
401 AVFrame *main, int stride)\
402 { \
403 int w = s->width; \
404 int h = s->height; \
405 \
406 int ref_stride = ref->linesize[0]; \
407 int main_stride = main->linesize[0]; \
408 \
409 const type *ref_ptr = (const type *) ref->data[0]; \
410 const type *main_ptr = (const type *) main->data[0]; \
411 \
412 const float factor = s->factor; \
413 \
414 float *ref_ptr_data = s->ref_data; \
415 float *main_ptr_data = s->main_data; \
416 \
417 for (int i = 0; i < h; i++) { \
418 for (int j = 0; j < w; j++) { \
419 ref_ptr_data[j] = ref_ptr[j] * factor - 128.f; \
420 main_ptr_data[j] = main_ptr[j] * factor - 128.f; \
421 } \
422 ref_ptr += ref_stride / sizeof(type); \
423 ref_ptr_data += w; \
424 main_ptr += main_stride / sizeof(type); \
425 main_ptr_data += w; \
426 } \
427 }
428
429 offset_fn(uint8_t, 8)
430 offset_fn(uint16_t, 16)
431
432 static void set_meta(AVDictionary **metadata, const char *key, float d)
433 {
434 char value[257];
435 snprintf(value, sizeof(value), "%f", d);
436 av_dict_set(metadata, key, value, 0);
437 }
438
439 static AVFrame *do_vif(AVFilterContext *ctx, AVFrame *main, const AVFrame *ref)
440 {
441 VIFContext *s = ctx->priv;
442 AVDictionary **metadata = &main->metadata;
443 float score[4];
444
445 s->factor = 1.f / (1 << (s->desc->comp[0].depth - 8));
446 if (s->desc->comp[0].depth <= 8) {
447 offset_8bit(s, ref, main, s->width);
448 } else {
449 offset_16bit(s, ref, main, s->width);
450 }
451
452 compute_vif2(ctx, s->ref_data, s->main_data,
453 s->width, s->height, s->width, s->width,
454 score, s->data_buf, s->temp, s->nb_threads);
455
456 set_meta(metadata, "lavfi.vif.scale.0", score[0]);
457 set_meta(metadata, "lavfi.vif.scale.1", score[1]);
458 set_meta(metadata, "lavfi.vif.scale.2", score[2]);
459 set_meta(metadata, "lavfi.vif.scale.3", score[3]);
460
461 for (int i = 0; i < 4; i++) {
462 s->vif_min[i] = FFMIN(s->vif_min[i], score[i]);
463 s->vif_max[i] = FFMAX(s->vif_max[i], score[i]);
464 s->vif_sum[i] += score[i];
465 }
466
467 s->nb_frames++;
468
469 return main;
470 }
471
472 static const enum AVPixelFormat pix_fmts[] = {
473 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
474 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
475 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
476 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV410P,
477 AV_PIX_FMT_YUVJ411P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
478 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
479 #define PF(suf) AV_PIX_FMT_YUV420##suf, AV_PIX_FMT_YUV422##suf, AV_PIX_FMT_YUV444##suf
480 PF(P9), PF(P10), PF(P12), PF(P14), PF(P16),
481 AV_PIX_FMT_NONE
482 };
483
484 static int config_input_ref(AVFilterLink *inlink)
485 {
486 AVFilterContext *ctx = inlink->dst;
487 VIFContext *s = ctx->priv;
488
489 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
490 ctx->inputs[0]->h != ctx->inputs[1]->h) {
491 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
492 return AVERROR(EINVAL);
493 }
494
495 s->desc = av_pix_fmt_desc_get(inlink->format);
496 s->width = ctx->inputs[0]->w;
497 s->height = ctx->inputs[0]->h;
498 s->nb_threads = ff_filter_get_nb_threads(ctx);
499
500 for (int i = 0; i < 4; i++) {
501 s->vif_min[i] = DBL_MAX;
502 s->vif_max[i] = -DBL_MAX;
503 }
504
505 for (int i = 0; i < NUM_DATA_BUFS; i++) {
506 if (!(s->data_buf[i] = av_calloc(s->width, s->height * sizeof(float))))
507 return AVERROR(ENOMEM);
508 }
509
510 if (!(s->ref_data = av_calloc(s->width, s->height * sizeof(float))))
511 return AVERROR(ENOMEM);
512
513 if (!(s->main_data = av_calloc(s->width, s->height * sizeof(float))))
514 return AVERROR(ENOMEM);
515
516 if (!(s->temp = av_calloc(s->nb_threads, sizeof(s->temp[0]))))
517 return AVERROR(ENOMEM);
518
519 for (int i = 0; i < s->nb_threads; i++) {
520 if (!(s->temp[i] = av_calloc(s->width, sizeof(float))))
521 return AVERROR(ENOMEM);
522 }
523
524 return 0;
525 }
526
527 static int process_frame(FFFrameSync *fs)
528 {
529 AVFilterContext *ctx = fs->parent;
530 VIFContext *s = fs->opaque;
531 AVFilterLink *outlink = ctx->outputs[0];
532 AVFrame *out_frame, *main_frame = NULL, *ref_frame = NULL;
533 int ret;
534
535 ret = ff_framesync_dualinput_get(fs, &main_frame, &ref_frame);
536 if (ret < 0)
537 return ret;
538
539 if (ctx->is_disabled || !ref_frame) {
540 out_frame = main_frame;
541 } else {
542 out_frame = do_vif(ctx, main_frame, ref_frame);
543 }
544
545 out_frame->pts = av_rescale_q(s->fs.pts, s->fs.time_base, outlink->time_base);
546
547 return ff_filter_frame(outlink, out_frame);
548 }
549
550
551 static int config_output(AVFilterLink *outlink)
552 {
553 AVFilterContext *ctx = outlink->src;
554 VIFContext *s = ctx->priv;
555 AVFilterLink *mainlink = ctx->inputs[0];
556 FFFrameSyncIn *in;
557 int ret;
558
559 outlink->w = mainlink->w;
560 outlink->h = mainlink->h;
561 outlink->time_base = mainlink->time_base;
562 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
563 outlink->frame_rate = mainlink->frame_rate;
564 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
565 return ret;
566
567 in = s->fs.in;
568 in[0].time_base = mainlink->time_base;
569 in[1].time_base = ctx->inputs[1]->time_base;
570 in[0].sync = 2;
571 in[0].before = EXT_STOP;
572 in[0].after = EXT_STOP;
573 in[1].sync = 1;
574 in[1].before = EXT_STOP;
575 in[1].after = EXT_STOP;
576 s->fs.opaque = s;
577 s->fs.on_event = process_frame;
578
579 return ff_framesync_configure(&s->fs);
580 }
581
582 static int activate(AVFilterContext *ctx)
583 {
584 VIFContext *s = ctx->priv;
585 return ff_framesync_activate(&s->fs);
586 }
587
588 static av_cold void uninit(AVFilterContext *ctx)
589 {
590 VIFContext *s = ctx->priv;
591
592 if (s->nb_frames > 0) {
593 for (int i = 0; i < 4; i++)
594 av_log(ctx, AV_LOG_INFO, "VIF scale=%d average:%f min:%f: max:%f\n",
595 i, s->vif_sum[i] / s->nb_frames, s->vif_min[i], s->vif_max[i]);
596 }
597
598 for (int i = 0; i < NUM_DATA_BUFS; i++)
599 av_freep(&s->data_buf[i]);
600
601 av_freep(&s->ref_data);
602 av_freep(&s->main_data);
603
604 for (int i = 0; i < s->nb_threads && s->temp; i++)
605 av_freep(&s->temp[i]);
606
607 av_freep(&s->temp);
608
609 ff_framesync_uninit(&s->fs);
610 }
611
612 static const AVFilterPad vif_inputs[] = {
613 {
614 .name = "main",
615 .type = AVMEDIA_TYPE_VIDEO,
616 },{
617 .name = "reference",
618 .type = AVMEDIA_TYPE_VIDEO,
619 .config_props = config_input_ref,
620 },
621 };
622
623 static const AVFilterPad vif_outputs[] = {
624 {
625 .name = "default",
626 .type = AVMEDIA_TYPE_VIDEO,
627 .config_props = config_output,
628 },
629 };
630
631 const AVFilter ff_vf_vif = {
632 .name = "vif",
633 .description = NULL_IF_CONFIG_SMALL("Calculate the VIF between two video streams."),
634 .uninit = uninit,
635 .priv_size = sizeof(VIFContext),
636 .priv_class = &vif_class,
637 .activate = activate,
638 FILTER_INPUTS(vif_inputs),
639 FILTER_OUTPUTS(vif_outputs),
640 FILTER_PIXFMTS_ARRAY(pix_fmts),
641 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
642 AVFILTER_FLAG_SLICE_THREADS |
643 AVFILTER_FLAG_METADATA_ONLY,
644 };
645