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
6 * This file is part of FFmpeg.
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.
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.
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
25 * Calculate VIF between two input videos.
30 #include "libavutil/avstring.h"
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
34 #include "framesync.h"
35 #include "drawutils.h"
41 typedef struct VIFContext {
44 const AVPixFmtDescriptor *desc;
58 #define OFFSET(x) offsetof(VIFContext, x)
60 static const AVOption vif_options[] = {
64 AVFILTER_DEFINE_CLASS(vif);
66 static const uint8_t vif_filter1d_width1[4] = { 17, 9, 5, 3 };
68 static const float vif_filter1d_table[4][17] =
71 0.00745626912, 0.0142655009, 0.0250313189, 0.0402820669, 0.0594526194,
72 0.0804751068, 0.0999041125, 0.113746084, 0.118773937, 0.113746084,
73 0.0999041125, 0.0804751068, 0.0594526194, 0.0402820669, 0.0250313189,
74 0.0142655009, 0.00745626912
77 0.0189780835, 0.0558981746, 0.120920904, 0.192116052, 0.224173605,
78 0.192116052, 0.120920904, 0.0558981746, 0.0189780835
81 0.054488685, 0.244201347, 0.402619958, 0.244201347, 0.054488685
84 0.166378498, 0.667243004, 0.166378498
88 typedef struct ThreadData {
99 static void vif_dec2(const float *src, float *dst, int w, int h,
100 int src_stride, int dst_stride)
102 const int dst_px_stride = dst_stride / 2;
104 for (int i = 0; i < h / 2; i++) {
105 for (int j = 0; j < w / 2; j++)
106 dst[i * dst_px_stride + j] = src[(i * 2) * src_stride + (j * 2)];
110 static void vif_statistic(const float *mu1_sq, const float *mu2_sq,
111 const float *mu1_mu2, const float *xx_filt,
112 const float *yy_filt, const float *xy_filt,
113 float *num, float *den, int w, int h)
115 static const float sigma_nsq = 2;
116 static const float sigma_max_inv = 4.0/(255.0*255.0);
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;
125 for (int i = 0; i < h; i++) {
126 float accum_inner_num = 0.f;
127 float accum_inner_den = 0.f;
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];
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;
141 sigma1_sq = FFMAX(sigma1_sq, 0.0f);
142 sigma2_sq = FFMAX(sigma2_sq, 0.0f);
143 sigma12 = FFMAX(sigma12, 0.0f);
145 g = sigma12 / (sigma1_sq + eps);
146 sv_sq = sigma2_sq - g * sigma12;
148 if (sigma1_sq < eps) {
154 if (sigma2_sq < eps) {
163 sv_sq = FFMAX(sv_sq, eps);
165 g = FFMIN(g, gain_limit);
167 num_val = log2f(1.0f + g * g * sigma1_sq / (sv_sq + sigma_nsq));
168 den_val = log2f(1.0f + sigma1_sq / sigma_nsq);
173 if (sigma1_sq < sigma_nsq) {
174 num_val = 1.0f - sigma2_sq * sigma_max_inv;
178 accum_inner_num += num_val;
179 accum_inner_den += den_val;
182 accum_num += accum_inner_num;
183 accum_den += accum_inner_den;
190 static void vif_xx_yy_xy(const float *x, const float *y, float *xx, float *yy,
191 float *xy, int w, int h)
193 for (int i = 0; i < h; i++) {
194 for (int j = 0; j < w; j++) {
197 float xxval = xval * xval;
198 float yyval = yval * yval;
199 float xyval = xval * yval;
214 static int vif_filter1d(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
216 ThreadData *td = arg;
217 const float *filter = td->filter;
218 const float *src = td->src;
219 float *dst = td->dst;
222 int src_stride = td->src_stride;
223 int dst_stride = td->dst_stride;
224 int filt_w = td->filter_width;
225 float *temp = td->temp[jobnr];
226 const int slice_start = (h * jobnr) / nb_jobs;
227 const int slice_end = (h * (jobnr+1)) / nb_jobs;
229 for (int i = slice_start; i < slice_end; i++) {
230 /** Vertical pass. */
231 for (int j = 0; j < w; j++) {
234 if (i >= filt_w / 2 && i < h - filt_w / 2 - 1) {
235 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
236 const float filt_coeff = filter[filt_i];
238 int ii = i - filt_w / 2 + filt_i;
240 img_coeff = src[ii * src_stride + j];
241 sum += filt_coeff * img_coeff;
244 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
245 const float filt_coeff = filter[filt_i];
246 int ii = i - filt_w / 2 + filt_i;
249 ii = ii < 0 ? -ii : (ii >= h ? 2 * h - ii - 1 : ii);
251 img_coeff = src[ii * src_stride + j];
252 sum += filt_coeff * img_coeff;
259 /** Horizontal pass. */
260 for (int j = 0; j < w; j++) {
263 if (j >= filt_w / 2 && j < w - filt_w / 2 - 1) {
264 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
265 const float filt_coeff = filter[filt_j];
266 int jj = j - filt_w / 2 + filt_j;
269 img_coeff = temp[jj];
270 sum += filt_coeff * img_coeff;
273 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
274 const float filt_coeff = filter[filt_j];
275 int jj = j - filt_w / 2 + filt_j;
278 jj = jj < 0 ? -jj : (jj >= w ? 2 * w - jj - 1 : jj);
280 img_coeff = temp[jj];
281 sum += filt_coeff * img_coeff;
285 dst[i * dst_stride + j] = sum;
292 int ff_compute_vif2(AVFilterContext *ctx,
293 const float *ref, const float *main, int w, int h,
294 int ref_stride, int main_stride, float *score,
295 float *data_buf[14], float **temp,
299 float *ref_scale = data_buf[0];
300 float *main_scale = data_buf[1];
301 float *ref_sq = data_buf[2];
302 float *main_sq = data_buf[3];
303 float *ref_main = data_buf[4];
304 float *mu1 = data_buf[5];
305 float *mu2 = data_buf[6];
306 float *mu1_sq = data_buf[7];
307 float *mu2_sq = data_buf[8];
308 float *mu1_mu2 = data_buf[9];
309 float *ref_sq_filt = data_buf[10];
310 float *main_sq_filt = data_buf[11];
311 float *ref_main_filt = data_buf[12];
313 float *curr_ref_scale = (float *)ref;
314 float *curr_main_scale = (float *)main;
315 int curr_ref_stride = ref_stride;
316 int curr_main_stride = main_stride;
321 for (int scale = 0; scale < 4; scale++) {
322 const float *filter = vif_filter1d_table[scale];
323 int filter_width = vif_filter1d_width1[scale];
324 const int nb_threads = FFMIN(h, gnb_threads);
329 td.filter_width = filter_width;
332 td.src = curr_ref_scale;
336 td.src_stride = curr_ref_stride;
339 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
341 td.src = curr_main_scale;
343 td.src_stride = curr_main_stride;
344 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
346 vif_dec2(mu1, ref_scale, buf_valid_w, buf_valid_h, w, w);
347 vif_dec2(mu2, main_scale, buf_valid_w, buf_valid_h, w, w);
355 curr_ref_scale = ref_scale;
356 curr_main_scale = main_scale;
359 curr_main_stride = w;
362 td.src = curr_ref_scale;
366 td.src_stride = curr_ref_stride;
369 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
371 td.src = curr_main_scale;
373 td.src_stride = curr_main_stride;
374 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
376 vif_xx_yy_xy(mu1, mu2, mu1_sq, mu2_sq, mu1_mu2, w, h);
378 vif_xx_yy_xy(curr_ref_scale, curr_main_scale, ref_sq, main_sq, ref_main, w, h);
381 td.dst = ref_sq_filt;
383 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
386 td.dst = main_sq_filt;
388 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
391 td.dst = ref_main_filt;
392 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
394 vif_statistic(mu1_sq, mu2_sq, mu1_mu2, ref_sq_filt, main_sq_filt,
395 ref_main_filt, &num, &den, w, h);
397 score[scale] = den <= FLT_EPSILON ? 1.f : num / den;
403 #define offset_fn(type, bits) \
404 static void offset_##bits##bit(VIFContext *s, \
405 const AVFrame *ref, \
406 AVFrame *main, int stride)\
411 int ref_stride = ref->linesize[0]; \
412 int main_stride = main->linesize[0]; \
414 const type *ref_ptr = (const type *) ref->data[0]; \
415 const type *main_ptr = (const type *) main->data[0]; \
417 float *ref_ptr_data = s->ref_data; \
418 float *main_ptr_data = s->main_data; \
420 for (int i = 0; i < h; i++) { \
421 for (int j = 0; j < w; j++) { \
422 ref_ptr_data[j] = ref_ptr[j] - 128.f; \
423 main_ptr_data[j] = main_ptr[j] - 128.f; \
425 ref_ptr += ref_stride / sizeof(type); \
427 main_ptr += main_stride / sizeof(type); \
428 main_ptr_data += w; \
432 offset_fn(uint8_t, 8)
433 offset_fn(uint16_t, 10)
435 static void set_meta(AVDictionary **metadata, const char *key, float d)
438 snprintf(value, sizeof(value), "%f", d);
439 av_dict_set(metadata, key, value, 0);
442 static AVFrame *do_vif(AVFilterContext *ctx, AVFrame *main, const AVFrame *ref)
444 VIFContext *s = ctx->priv;
445 AVDictionary **metadata = &main->metadata;
448 if (s->desc->comp[0].depth <= 8) {
449 offset_8bit(s, ref, main, s->width);
451 offset_10bit(s, ref, main, s->width);
455 s->ref_data, s->main_data, s->width,
456 s->height, s->width, s->width,
457 score, s->data_buf, s->temp,
460 set_meta(metadata, "lavfi.vif.scale.0", score[0]);
461 set_meta(metadata, "lavfi.vif.scale.1", score[1]);
462 set_meta(metadata, "lavfi.vif.scale.2", score[2]);
463 set_meta(metadata, "lavfi.vif.scale.3", score[3]);
465 for (int i = 0; i < 4; i++) {
466 s->vif_min[i] = FFMIN(s->vif_min[i], score[i]);
467 s->vif_max[i] = FFMAX(s->vif_max[i], score[i]);
468 s->vif_sum[i] += score[i];
476 static int query_formats(AVFilterContext *ctx)
478 static const enum AVPixelFormat pix_fmts[] = {
479 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV420P,
480 AV_PIX_FMT_YUV444P10LE, AV_PIX_FMT_YUV422P10LE, AV_PIX_FMT_YUV420P10LE,
484 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
486 return AVERROR(ENOMEM);
487 return ff_set_common_formats(ctx, fmts_list);
490 static int config_input_ref(AVFilterLink *inlink)
492 AVFilterContext *ctx = inlink->dst;
493 VIFContext *s = ctx->priv;
495 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
496 ctx->inputs[0]->h != ctx->inputs[1]->h) {
497 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
498 return AVERROR(EINVAL);
500 if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
501 av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
502 return AVERROR(EINVAL);
505 s->desc = av_pix_fmt_desc_get(inlink->format);
506 s->width = ctx->inputs[0]->w;
507 s->height = ctx->inputs[0]->h;
508 s->nb_threads = ff_filter_get_nb_threads(ctx);
510 for (int i = 0; i < 4; i++) {
511 s->vif_min[i] = DBL_MAX;
512 s->vif_max[i] = -DBL_MAX;
515 for (int i = 0; i < 13; i++) {
516 if (!(s->data_buf[i] = av_calloc(s->width, s->height * sizeof(float))))
517 return AVERROR(ENOMEM);
520 if (!(s->ref_data = av_calloc(s->width, s->height * sizeof(float))))
521 return AVERROR(ENOMEM);
523 if (!(s->main_data = av_calloc(s->width, s->height * sizeof(float))))
524 return AVERROR(ENOMEM);
526 if (!(s->temp = av_calloc(s->nb_threads, sizeof(s->temp[0]))))
527 return AVERROR(ENOMEM);
529 for (int i = 0; i < s->nb_threads; i++) {
530 if (!(s->temp[i] = av_calloc(s->width, sizeof(float))))
531 return AVERROR(ENOMEM);
537 static int process_frame(FFFrameSync *fs)
539 AVFilterContext *ctx = fs->parent;
540 VIFContext *s = fs->opaque;
541 AVFilterLink *outlink = ctx->outputs[0];
542 AVFrame *out, *main = NULL, *ref = NULL;
545 ret = ff_framesync_dualinput_get(fs, &main, &ref);
549 if (ctx->is_disabled || !ref) {
552 out = do_vif(ctx, main, ref);
555 out->pts = av_rescale_q(s->fs.pts, s->fs.time_base, outlink->time_base);
557 return ff_filter_frame(outlink, out);
561 static int config_output(AVFilterLink *outlink)
563 AVFilterContext *ctx = outlink->src;
564 VIFContext *s = ctx->priv;
565 AVFilterLink *mainlink = ctx->inputs[0];
569 outlink->w = mainlink->w;
570 outlink->h = mainlink->h;
571 outlink->time_base = mainlink->time_base;
572 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
573 outlink->frame_rate = mainlink->frame_rate;
574 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
578 in[0].time_base = mainlink->time_base;
579 in[1].time_base = ctx->inputs[1]->time_base;
581 in[0].before = EXT_STOP;
582 in[0].after = EXT_STOP;
584 in[1].before = EXT_STOP;
585 in[1].after = EXT_STOP;
587 s->fs.on_event = process_frame;
589 return ff_framesync_configure(&s->fs);
592 static int activate(AVFilterContext *ctx)
594 VIFContext *s = ctx->priv;
595 return ff_framesync_activate(&s->fs);
598 static av_cold void uninit(AVFilterContext *ctx)
600 VIFContext *s = ctx->priv;
602 if (s->nb_frames > 0) {
603 for (int i = 0; i < 4; i++)
604 av_log(ctx, AV_LOG_INFO, "VIF scale=%d average:%f min:%f: max:%f\n",
605 i, s->vif_sum[i] / s->nb_frames, s->vif_min[i], s->vif_max[i]);
608 for (int i = 0; i < 13; i++)
609 av_freep(&s->data_buf[i]);
611 av_freep(&s->ref_data);
612 av_freep(&s->main_data);
614 for (int i = 0; i < s->nb_threads && s->temp; i++)
615 av_freep(&s->temp[i]);
619 ff_framesync_uninit(&s->fs);
622 static const AVFilterPad vif_inputs[] = {
625 .type = AVMEDIA_TYPE_VIDEO,
628 .type = AVMEDIA_TYPE_VIDEO,
629 .config_props = config_input_ref,
634 static const AVFilterPad vif_outputs[] = {
637 .type = AVMEDIA_TYPE_VIDEO,
638 .config_props = config_output,
643 AVFilter ff_vf_vif = {
645 .description = NULL_IF_CONFIG_SMALL("Calculate the VIF between two video streams."),
647 .query_formats = query_formats,
648 .priv_size = sizeof(VIFContext),
649 .priv_class = &vif_class,
650 .activate = activate,
651 .inputs = vif_inputs,
652 .outputs = vif_outputs,
653 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,