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);
171 num_val = den_val = 1.f;
173 accum_inner_num += num_val;
174 accum_inner_den += den_val;
177 accum_num += accum_inner_num;
178 accum_den += accum_inner_den;
185 static void vif_xx_yy_xy(const float *x, const float *y, float *xx, float *yy,
186 float *xy, int w, int h)
188 for (int i = 0; i < h; i++) {
189 for (int j = 0; j < w; j++) {
192 float xxval = xval * xval;
193 float yyval = yval * yval;
194 float xyval = xval * yval;
209 static int vif_filter1d(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
211 ThreadData *td = arg;
212 const float *filter = td->filter;
213 const float *src = td->src;
214 float *dst = td->dst;
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;
224 for (int i = slice_start; i < slice_end; i++) {
225 /** Vertical pass. */
226 for (int j = 0; j < w; j++) {
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];
233 int ii = i - filt_w / 2 + filt_i;
235 img_coeff = src[ii * src_stride + j];
236 sum += filt_coeff * img_coeff;
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;
244 ii = ii < 0 ? -ii : (ii >= h ? 2 * h - ii - 1 : ii);
246 img_coeff = src[ii * src_stride + j];
247 sum += filt_coeff * img_coeff;
254 /** Horizontal pass. */
255 for (int j = 0; j < w; j++) {
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;
264 img_coeff = temp[jj];
265 sum += filt_coeff * img_coeff;
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;
273 jj = jj < 0 ? -jj : (jj >= w ? 2 * w - jj - 1 : jj);
275 img_coeff = temp[jj];
276 sum += filt_coeff * img_coeff;
280 dst[i * dst_stride + j] = sum;
287 int ff_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 *data_buf[14], float **temp,
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];
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;
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);
324 td.filter_width = filter_width;
327 td.src = curr_ref_scale;
331 td.src_stride = curr_ref_stride;
334 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
336 td.src = curr_main_scale;
338 td.src_stride = curr_main_stride;
339 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
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);
350 curr_ref_scale = ref_scale;
351 curr_main_scale = main_scale;
354 curr_main_stride = w;
357 td.src = curr_ref_scale;
361 td.src_stride = curr_ref_stride;
364 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
366 td.src = curr_main_scale;
368 td.src_stride = curr_main_stride;
369 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
371 vif_xx_yy_xy(mu1, mu2, mu1_sq, mu2_sq, mu1_mu2, w, h);
373 vif_xx_yy_xy(curr_ref_scale, curr_main_scale, ref_sq, main_sq, ref_main, w, h);
376 td.dst = ref_sq_filt;
378 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
381 td.dst = main_sq_filt;
383 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
386 td.dst = ref_main_filt;
387 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
389 vif_statistic(mu1_sq, mu2_sq, mu1_mu2, ref_sq_filt, main_sq_filt,
390 ref_main_filt, &num, &den, w, h);
392 score[scale] = den <= FLT_EPSILON ? 1.f : num / den;
398 #define offset_fn(type, bits) \
399 static void offset_##bits##bit(VIFContext *s, \
400 const AVFrame *ref, \
401 AVFrame *main, int stride)\
406 int ref_stride = ref->linesize[0]; \
407 int main_stride = main->linesize[0]; \
409 const type *ref_ptr = (const type *) ref->data[0]; \
410 const type *main_ptr = (const type *) main->data[0]; \
412 float *ref_ptr_data = s->ref_data; \
413 float *main_ptr_data = s->main_data; \
415 for (int i = 0; i < h; i++) { \
416 for (int j = 0; j < w; j++) { \
417 ref_ptr_data[j] = ref_ptr[j] - 128.f; \
418 main_ptr_data[j] = main_ptr[j] - 128.f; \
420 ref_ptr += ref_stride / sizeof(type); \
422 main_ptr += main_stride / sizeof(type); \
423 main_ptr_data += w; \
427 offset_fn(uint8_t, 8)
428 offset_fn(uint16_t, 10)
430 static void set_meta(AVDictionary **metadata, const char *key, float d)
433 snprintf(value, sizeof(value), "%f", d);
434 av_dict_set(metadata, key, value, 0);
437 static AVFrame *do_vif(AVFilterContext *ctx, AVFrame *main, const AVFrame *ref)
439 VIFContext *s = ctx->priv;
440 AVDictionary **metadata = &main->metadata;
443 if (s->desc->comp[0].depth <= 8) {
444 offset_8bit(s, ref, main, s->width);
446 offset_10bit(s, ref, main, s->width);
450 s->ref_data, s->main_data, s->width,
451 s->height, s->width, s->width,
452 score, s->data_buf, s->temp,
455 set_meta(metadata, "lavfi.vif.scale.0", score[0]);
456 set_meta(metadata, "lavfi.vif.scale.1", score[1]);
457 set_meta(metadata, "lavfi.vif.scale.2", score[2]);
458 set_meta(metadata, "lavfi.vif.scale.3", score[3]);
460 for (int i = 0; i < 4; i++) {
461 s->vif_min[i] = FFMIN(s->vif_min[i], score[i]);
462 s->vif_max[i] = FFMAX(s->vif_max[i], score[i]);
463 s->vif_sum[i] += score[i];
471 static int query_formats(AVFilterContext *ctx)
473 static const enum AVPixelFormat pix_fmts[] = {
474 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV420P,
475 AV_PIX_FMT_YUV444P10LE, AV_PIX_FMT_YUV422P10LE, AV_PIX_FMT_YUV420P10LE,
479 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
481 return AVERROR(ENOMEM);
482 return ff_set_common_formats(ctx, fmts_list);
485 static int config_input_ref(AVFilterLink *inlink)
487 AVFilterContext *ctx = inlink->dst;
488 VIFContext *s = ctx->priv;
490 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
491 ctx->inputs[0]->h != ctx->inputs[1]->h) {
492 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
493 return AVERROR(EINVAL);
495 if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
496 av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
497 return AVERROR(EINVAL);
500 s->desc = av_pix_fmt_desc_get(inlink->format);
501 s->width = ctx->inputs[0]->w;
502 s->height = ctx->inputs[0]->h;
503 s->nb_threads = ff_filter_get_nb_threads(ctx);
505 for (int i = 0; i < 4; i++) {
506 s->vif_min[i] = DBL_MAX;
507 s->vif_max[i] = -DBL_MAX;
510 for (int i = 0; i < 13; i++) {
511 if (!(s->data_buf[i] = av_calloc(s->width, s->height * sizeof(float))))
512 return AVERROR(ENOMEM);
515 if (!(s->ref_data = av_calloc(s->width, s->height * sizeof(float))))
516 return AVERROR(ENOMEM);
518 if (!(s->main_data = av_calloc(s->width, s->height * sizeof(float))))
519 return AVERROR(ENOMEM);
521 if (!(s->temp = av_calloc(s->nb_threads, sizeof(s->temp[0]))))
522 return AVERROR(ENOMEM);
524 for (int i = 0; i < s->nb_threads; i++) {
525 if (!(s->temp[i] = av_calloc(s->width, sizeof(float))))
526 return AVERROR(ENOMEM);
532 static int process_frame(FFFrameSync *fs)
534 AVFilterContext *ctx = fs->parent;
535 VIFContext *s = fs->opaque;
536 AVFilterLink *outlink = ctx->outputs[0];
537 AVFrame *out_frame, *main_frame = NULL, *ref_frame = NULL;
540 ret = ff_framesync_dualinput_get(fs, &main_frame, &ref_frame);
544 if (ctx->is_disabled || !ref_frame) {
545 out_frame = main_frame;
547 out_frame = do_vif(ctx, main_frame, ref_frame);
550 out_frame->pts = av_rescale_q(s->fs.pts, s->fs.time_base, outlink->time_base);
552 return ff_filter_frame(outlink, out_frame);
556 static int config_output(AVFilterLink *outlink)
558 AVFilterContext *ctx = outlink->src;
559 VIFContext *s = ctx->priv;
560 AVFilterLink *mainlink = ctx->inputs[0];
564 outlink->w = mainlink->w;
565 outlink->h = mainlink->h;
566 outlink->time_base = mainlink->time_base;
567 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
568 outlink->frame_rate = mainlink->frame_rate;
569 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
573 in[0].time_base = mainlink->time_base;
574 in[1].time_base = ctx->inputs[1]->time_base;
576 in[0].before = EXT_STOP;
577 in[0].after = EXT_STOP;
579 in[1].before = EXT_STOP;
580 in[1].after = EXT_STOP;
582 s->fs.on_event = process_frame;
584 return ff_framesync_configure(&s->fs);
587 static int activate(AVFilterContext *ctx)
589 VIFContext *s = ctx->priv;
590 return ff_framesync_activate(&s->fs);
593 static av_cold void uninit(AVFilterContext *ctx)
595 VIFContext *s = ctx->priv;
597 if (s->nb_frames > 0) {
598 for (int i = 0; i < 4; i++)
599 av_log(ctx, AV_LOG_INFO, "VIF scale=%d average:%f min:%f: max:%f\n",
600 i, s->vif_sum[i] / s->nb_frames, s->vif_min[i], s->vif_max[i]);
603 for (int i = 0; i < 13; i++)
604 av_freep(&s->data_buf[i]);
606 av_freep(&s->ref_data);
607 av_freep(&s->main_data);
609 for (int i = 0; i < s->nb_threads && s->temp; i++)
610 av_freep(&s->temp[i]);
614 ff_framesync_uninit(&s->fs);
617 static const AVFilterPad vif_inputs[] = {
620 .type = AVMEDIA_TYPE_VIDEO,
623 .type = AVMEDIA_TYPE_VIDEO,
624 .config_props = config_input_ref,
629 static const AVFilterPad vif_outputs[] = {
632 .type = AVMEDIA_TYPE_VIDEO,
633 .config_props = config_output,
638 AVFilter ff_vf_vif = {
640 .description = NULL_IF_CONFIG_SMALL("Calculate the VIF between two video streams."),
642 .query_formats = query_formats,
643 .priv_size = sizeof(VIFContext),
644 .priv_class = &vif_class,
645 .activate = activate,
646 .inputs = vif_inputs,
647 .outputs = vif_outputs,
648 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,