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;
59 #define OFFSET(x) offsetof(VIFContext, x)
61 static const AVOption vif_options[] = {
65 AVFILTER_DEFINE_CLASS(vif);
67 static const uint8_t vif_filter1d_width1[4] = { 17, 9, 5, 3 };
69 static const float vif_filter1d_table[4][17] =
72 0.00745626912, 0.0142655009, 0.0250313189, 0.0402820669, 0.0594526194,
73 0.0804751068, 0.0999041125, 0.113746084, 0.118773937, 0.113746084,
74 0.0999041125, 0.0804751068, 0.0594526194, 0.0402820669, 0.0250313189,
75 0.0142655009, 0.00745626912
78 0.0189780835, 0.0558981746, 0.120920904, 0.192116052, 0.224173605,
79 0.192116052, 0.120920904, 0.0558981746, 0.0189780835
82 0.054488685, 0.244201347, 0.402619958, 0.244201347, 0.054488685
85 0.166378498, 0.667243004, 0.166378498
89 typedef struct ThreadData {
100 static void vif_dec2(const float *src, float *dst, int w, int h,
101 int src_stride, int dst_stride)
103 const int dst_px_stride = dst_stride / 2;
105 for (int i = 0; i < h / 2; i++) {
106 for (int j = 0; j < w / 2; j++)
107 dst[i * dst_px_stride + j] = src[(i * 2) * src_stride + (j * 2)];
111 static void vif_statistic(const float *mu1_sq, const float *mu2_sq,
112 const float *mu1_mu2, const float *xx_filt,
113 const float *yy_filt, const float *xy_filt,
114 float *num, float *den, int w, int h)
116 static const float sigma_nsq = 2;
117 float mu1_sq_val, mu2_sq_val, mu1_mu2_val, xx_filt_val, yy_filt_val, xy_filt_val;
118 float sigma1_sq, sigma2_sq, sigma12, g, sv_sq, eps = 1.0e-10f;
119 float gain_limit = 100.f;
120 float num_val, den_val;
121 float accum_num = 0.0f;
122 float accum_den = 0.0f;
124 for (int i = 0; i < h; i++) {
125 float accum_inner_num = 0.f;
126 float accum_inner_den = 0.f;
128 for (int j = 0; j < w; j++) {
129 mu1_sq_val = mu1_sq[i * w + j];
130 mu2_sq_val = mu2_sq[i * w + j];
131 mu1_mu2_val = mu1_mu2[i * w + j];
132 xx_filt_val = xx_filt[i * w + j];
133 yy_filt_val = yy_filt[i * w + j];
134 xy_filt_val = xy_filt[i * w + j];
136 sigma1_sq = xx_filt_val - mu1_sq_val;
137 sigma2_sq = yy_filt_val - mu2_sq_val;
138 sigma12 = xy_filt_val - mu1_mu2_val;
140 sigma1_sq = FFMAX(sigma1_sq, 0.0f);
141 sigma2_sq = FFMAX(sigma2_sq, 0.0f);
142 sigma12 = FFMAX(sigma12, 0.0f);
144 g = sigma12 / (sigma1_sq + eps);
145 sv_sq = sigma2_sq - g * sigma12;
147 if (sigma1_sq < eps) {
153 if (sigma2_sq < eps) {
162 sv_sq = FFMAX(sv_sq, eps);
164 g = FFMIN(g, gain_limit);
166 num_val = log2f(1.0f + g * g * sigma1_sq / (sv_sq + sigma_nsq));
167 den_val = log2f(1.0f + sigma1_sq / sigma_nsq);
170 num_val = den_val = 1.f;
172 accum_inner_num += num_val;
173 accum_inner_den += den_val;
176 accum_num += accum_inner_num;
177 accum_den += accum_inner_den;
184 static void vif_xx_yy_xy(const float *x, const float *y, float *xx, float *yy,
185 float *xy, int w, int h)
187 for (int i = 0; i < h; i++) {
188 for (int j = 0; j < w; j++) {
191 float xxval = xval * xval;
192 float yyval = yval * yval;
193 float xyval = xval * yval;
208 static int vif_filter1d(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
210 ThreadData *td = arg;
211 const float *filter = td->filter;
212 const float *src = td->src;
213 float *dst = td->dst;
216 int src_stride = td->src_stride;
217 int dst_stride = td->dst_stride;
218 int filt_w = td->filter_width;
219 float *temp = td->temp[jobnr];
220 const int slice_start = (h * jobnr) / nb_jobs;
221 const int slice_end = (h * (jobnr+1)) / nb_jobs;
223 for (int i = slice_start; i < slice_end; i++) {
224 /** Vertical pass. */
225 for (int j = 0; j < w; j++) {
228 if (i >= filt_w / 2 && i < h - filt_w / 2 - 1) {
229 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
230 const float filt_coeff = filter[filt_i];
232 int ii = i - filt_w / 2 + filt_i;
234 img_coeff = src[ii * src_stride + j];
235 sum += filt_coeff * img_coeff;
238 for (int filt_i = 0; filt_i < filt_w; filt_i++) {
239 const float filt_coeff = filter[filt_i];
240 int ii = i - filt_w / 2 + filt_i;
243 ii = ii < 0 ? -ii : (ii >= h ? 2 * h - ii - 1 : ii);
245 img_coeff = src[ii * src_stride + j];
246 sum += filt_coeff * img_coeff;
253 /** Horizontal pass. */
254 for (int j = 0; j < w; j++) {
257 if (j >= filt_w / 2 && j < w - filt_w / 2 - 1) {
258 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
259 const float filt_coeff = filter[filt_j];
260 int jj = j - filt_w / 2 + filt_j;
263 img_coeff = temp[jj];
264 sum += filt_coeff * img_coeff;
267 for (int filt_j = 0; filt_j < filt_w; filt_j++) {
268 const float filt_coeff = filter[filt_j];
269 int jj = j - filt_w / 2 + filt_j;
272 jj = jj < 0 ? -jj : (jj >= w ? 2 * w - jj - 1 : jj);
274 img_coeff = temp[jj];
275 sum += filt_coeff * img_coeff;
279 dst[i * dst_stride + j] = sum;
286 int ff_compute_vif2(AVFilterContext *ctx,
287 const float *ref, const float *main, int w, int h,
288 int ref_stride, int main_stride, float *score,
289 float *data_buf[14], float **temp,
293 float *ref_scale = data_buf[0];
294 float *main_scale = data_buf[1];
295 float *ref_sq = data_buf[2];
296 float *main_sq = data_buf[3];
297 float *ref_main = data_buf[4];
298 float *mu1 = data_buf[5];
299 float *mu2 = data_buf[6];
300 float *mu1_sq = data_buf[7];
301 float *mu2_sq = data_buf[8];
302 float *mu1_mu2 = data_buf[9];
303 float *ref_sq_filt = data_buf[10];
304 float *main_sq_filt = data_buf[11];
305 float *ref_main_filt = data_buf[12];
307 float *curr_ref_scale = (float *)ref;
308 float *curr_main_scale = (float *)main;
309 int curr_ref_stride = ref_stride;
310 int curr_main_stride = main_stride;
315 for (int scale = 0; scale < 4; scale++) {
316 const float *filter = vif_filter1d_table[scale];
317 int filter_width = vif_filter1d_width1[scale];
318 const int nb_threads = FFMIN(h, gnb_threads);
323 td.filter_width = filter_width;
326 td.src = curr_ref_scale;
330 td.src_stride = curr_ref_stride;
333 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
335 td.src = curr_main_scale;
337 td.src_stride = curr_main_stride;
338 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
340 vif_dec2(mu1, ref_scale, buf_valid_w, buf_valid_h, w, w);
341 vif_dec2(mu2, main_scale, buf_valid_w, buf_valid_h, w, w);
349 curr_ref_scale = ref_scale;
350 curr_main_scale = main_scale;
353 curr_main_stride = w;
356 td.src = curr_ref_scale;
360 td.src_stride = curr_ref_stride;
363 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
365 td.src = curr_main_scale;
367 td.src_stride = curr_main_stride;
368 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
370 vif_xx_yy_xy(mu1, mu2, mu1_sq, mu2_sq, mu1_mu2, w, h);
372 vif_xx_yy_xy(curr_ref_scale, curr_main_scale, ref_sq, main_sq, ref_main, w, h);
375 td.dst = ref_sq_filt;
377 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
380 td.dst = main_sq_filt;
382 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
385 td.dst = ref_main_filt;
386 ctx->internal->execute(ctx, vif_filter1d, &td, NULL, nb_threads);
388 vif_statistic(mu1_sq, mu2_sq, mu1_mu2, ref_sq_filt, main_sq_filt,
389 ref_main_filt, &num, &den, w, h);
391 score[scale] = den <= FLT_EPSILON ? 1.f : num / den;
397 #define offset_fn(type, bits) \
398 static void offset_##bits##bit(VIFContext *s, \
399 const AVFrame *ref, \
400 AVFrame *main, int stride)\
405 int ref_stride = ref->linesize[0]; \
406 int main_stride = main->linesize[0]; \
408 const type *ref_ptr = (const type *) ref->data[0]; \
409 const type *main_ptr = (const type *) main->data[0]; \
411 const float factor = s->factor; \
413 float *ref_ptr_data = s->ref_data; \
414 float *main_ptr_data = s->main_data; \
416 for (int i = 0; i < h; i++) { \
417 for (int j = 0; j < w; j++) { \
418 ref_ptr_data[j] = ref_ptr[j] * factor - 128.f; \
419 main_ptr_data[j] = main_ptr[j] * factor - 128.f; \
421 ref_ptr += ref_stride / sizeof(type); \
423 main_ptr += main_stride / sizeof(type); \
424 main_ptr_data += w; \
428 offset_fn(uint8_t, 8)
429 offset_fn(uint16_t, 16)
431 static void set_meta(AVDictionary **metadata, const char *key, float d)
434 snprintf(value, sizeof(value), "%f", d);
435 av_dict_set(metadata, key, value, 0);
438 static AVFrame *do_vif(AVFilterContext *ctx, AVFrame *main, const AVFrame *ref)
440 VIFContext *s = ctx->priv;
441 AVDictionary **metadata = &main->metadata;
444 s->factor = 1.f / (1 << (s->desc->comp[0].depth - 8));
445 if (s->desc->comp[0].depth <= 8) {
446 offset_8bit(s, ref, main, s->width);
448 offset_16bit(s, ref, main, s->width);
452 s->ref_data, s->main_data, s->width,
453 s->height, s->width, s->width,
454 score, s->data_buf, s->temp,
457 set_meta(metadata, "lavfi.vif.scale.0", score[0]);
458 set_meta(metadata, "lavfi.vif.scale.1", score[1]);
459 set_meta(metadata, "lavfi.vif.scale.2", score[2]);
460 set_meta(metadata, "lavfi.vif.scale.3", score[3]);
462 for (int i = 0; i < 4; i++) {
463 s->vif_min[i] = FFMIN(s->vif_min[i], score[i]);
464 s->vif_max[i] = FFMAX(s->vif_max[i], score[i]);
465 s->vif_sum[i] += score[i];
473 static int query_formats(AVFilterContext *ctx)
475 static const enum AVPixelFormat pix_fmts[] = {
476 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
477 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
478 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
479 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV410P,
480 AV_PIX_FMT_YUVJ411P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
481 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
482 #define PF(suf) AV_PIX_FMT_YUV420##suf, AV_PIX_FMT_YUV422##suf, AV_PIX_FMT_YUV444##suf
483 PF(P9), PF(P10), PF(P12), PF(P14), PF(P16),
487 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
489 return AVERROR(ENOMEM);
490 return ff_set_common_formats(ctx, fmts_list);
493 static int config_input_ref(AVFilterLink *inlink)
495 AVFilterContext *ctx = inlink->dst;
496 VIFContext *s = ctx->priv;
498 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
499 ctx->inputs[0]->h != ctx->inputs[1]->h) {
500 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
501 return AVERROR(EINVAL);
503 if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
504 av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
505 return AVERROR(EINVAL);
508 s->desc = av_pix_fmt_desc_get(inlink->format);
509 s->width = ctx->inputs[0]->w;
510 s->height = ctx->inputs[0]->h;
511 s->nb_threads = ff_filter_get_nb_threads(ctx);
513 for (int i = 0; i < 4; i++) {
514 s->vif_min[i] = DBL_MAX;
515 s->vif_max[i] = -DBL_MAX;
518 for (int i = 0; i < 13; i++) {
519 if (!(s->data_buf[i] = av_calloc(s->width, s->height * sizeof(float))))
520 return AVERROR(ENOMEM);
523 if (!(s->ref_data = av_calloc(s->width, s->height * sizeof(float))))
524 return AVERROR(ENOMEM);
526 if (!(s->main_data = av_calloc(s->width, s->height * sizeof(float))))
527 return AVERROR(ENOMEM);
529 if (!(s->temp = av_calloc(s->nb_threads, sizeof(s->temp[0]))))
530 return AVERROR(ENOMEM);
532 for (int i = 0; i < s->nb_threads; i++) {
533 if (!(s->temp[i] = av_calloc(s->width, sizeof(float))))
534 return AVERROR(ENOMEM);
540 static int process_frame(FFFrameSync *fs)
542 AVFilterContext *ctx = fs->parent;
543 VIFContext *s = fs->opaque;
544 AVFilterLink *outlink = ctx->outputs[0];
545 AVFrame *out_frame, *main_frame = NULL, *ref_frame = NULL;
548 ret = ff_framesync_dualinput_get(fs, &main_frame, &ref_frame);
552 if (ctx->is_disabled || !ref_frame) {
553 out_frame = main_frame;
555 out_frame = do_vif(ctx, main_frame, ref_frame);
558 out_frame->pts = av_rescale_q(s->fs.pts, s->fs.time_base, outlink->time_base);
560 return ff_filter_frame(outlink, out_frame);
564 static int config_output(AVFilterLink *outlink)
566 AVFilterContext *ctx = outlink->src;
567 VIFContext *s = ctx->priv;
568 AVFilterLink *mainlink = ctx->inputs[0];
572 outlink->w = mainlink->w;
573 outlink->h = mainlink->h;
574 outlink->time_base = mainlink->time_base;
575 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
576 outlink->frame_rate = mainlink->frame_rate;
577 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
581 in[0].time_base = mainlink->time_base;
582 in[1].time_base = ctx->inputs[1]->time_base;
584 in[0].before = EXT_STOP;
585 in[0].after = EXT_STOP;
587 in[1].before = EXT_STOP;
588 in[1].after = EXT_STOP;
590 s->fs.on_event = process_frame;
592 return ff_framesync_configure(&s->fs);
595 static int activate(AVFilterContext *ctx)
597 VIFContext *s = ctx->priv;
598 return ff_framesync_activate(&s->fs);
601 static av_cold void uninit(AVFilterContext *ctx)
603 VIFContext *s = ctx->priv;
605 if (s->nb_frames > 0) {
606 for (int i = 0; i < 4; i++)
607 av_log(ctx, AV_LOG_INFO, "VIF scale=%d average:%f min:%f: max:%f\n",
608 i, s->vif_sum[i] / s->nb_frames, s->vif_min[i], s->vif_max[i]);
611 for (int i = 0; i < 13; i++)
612 av_freep(&s->data_buf[i]);
614 av_freep(&s->ref_data);
615 av_freep(&s->main_data);
617 for (int i = 0; i < s->nb_threads && s->temp; i++)
618 av_freep(&s->temp[i]);
622 ff_framesync_uninit(&s->fs);
625 static const AVFilterPad vif_inputs[] = {
628 .type = AVMEDIA_TYPE_VIDEO,
631 .type = AVMEDIA_TYPE_VIDEO,
632 .config_props = config_input_ref,
637 static const AVFilterPad vif_outputs[] = {
640 .type = AVMEDIA_TYPE_VIDEO,
641 .config_props = config_output,
646 AVFilter ff_vf_vif = {
648 .description = NULL_IF_CONFIG_SMALL("Calculate the VIF between two video streams."),
650 .query_formats = query_formats,
651 .priv_size = sizeof(VIFContext),
652 .priv_class = &vif_class,
653 .activate = activate,
654 .inputs = vif_inputs,
655 .outputs = vif_outputs,
656 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,