2 * Copyright (c) 2015-2016 mawen1250
3 * Copyright (c) 2018 Paul B Mahol
5 * This file is part of FFmpeg.
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included in all
15 * copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
20 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
28 * - non-power of 2 DCT
29 * - opponent color space
35 #include "libavutil/avassert.h"
36 #include "libavutil/imgutils.h"
37 #include "libavutil/opt.h"
38 #include "libavutil/pixdesc.h"
39 #include "libavcodec/avfft.h"
43 #include "framesync.h"
47 #define MAX_NB_THREADS 32
55 typedef struct ThreadData {
63 typedef struct PosCode {
67 typedef struct PosPairCode {
72 typedef struct SliceContext {
73 DCTContext *gdctf, *gdcti;
74 DCTContext *dctf, *dcti;
84 PosPairCode match_blocks[256];
86 PosCode *search_positions;
89 typedef struct BM3DContext {
112 SliceContext slices[MAX_NB_THREADS];
117 void (*get_block_row)(const uint8_t *srcp, int src_linesize,
118 int y, int x, int block_size, float *dst);
119 double (*do_block_ssd)(struct BM3DContext *s, PosCode *pos,
120 const uint8_t *src, int src_stride,
122 void (*do_output)(struct BM3DContext *s, uint8_t *dst, int dst_linesize,
123 int plane, int nb_jobs);
124 void (*block_filtering)(struct BM3DContext *s,
125 const uint8_t *src, int src_linesize,
126 const uint8_t *ref, int ref_linesize,
127 int y, int x, int plane, int jobnr);
130 #define OFFSET(x) offsetof(BM3DContext, x)
131 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
132 static const AVOption bm3d_options[] = {
133 { "sigma", "set denoising strength",
134 OFFSET(sigma), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 99999.9, FLAGS },
135 { "block", "set log2(size) of local patch",
136 OFFSET(block_size), AV_OPT_TYPE_INT, {.i64=4}, 4, 6, FLAGS },
137 { "bstep", "set sliding step for processing blocks",
138 OFFSET(block_step), AV_OPT_TYPE_INT, {.i64=4}, 1, 64, FLAGS },
139 { "group", "set maximal number of similar blocks",
140 OFFSET(group_size), AV_OPT_TYPE_INT, {.i64=1}, 1, 256, FLAGS },
141 { "range", "set block matching range",
142 OFFSET(bm_range), AV_OPT_TYPE_INT, {.i64=9}, 1, INT32_MAX, FLAGS },
143 { "mstep", "set step for block matching",
144 OFFSET(bm_step), AV_OPT_TYPE_INT, {.i64=1}, 1, 64, FLAGS },
145 { "thmse", "set threshold of mean square error for block matching",
146 OFFSET(th_mse), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, INT32_MAX, FLAGS },
147 { "hdthr", "set hard threshold for 3D transfer domain",
148 OFFSET(hard_threshold), AV_OPT_TYPE_FLOAT, {.dbl=2.7}, 0, INT32_MAX, FLAGS },
149 { "estim", "set filtering estimation mode",
150 OFFSET(mode), AV_OPT_TYPE_INT, {.i64=BASIC}, 0, NB_MODES-1, FLAGS, "mode" },
151 { "basic", "basic estimate",
152 0, AV_OPT_TYPE_CONST, {.i64=BASIC}, 0, 0, FLAGS, "mode" },
153 { "final", "final estimate",
154 0, AV_OPT_TYPE_CONST, {.i64=FINAL}, 0, 0, FLAGS, "mode" },
155 { "ref", "have reference stream",
156 OFFSET(ref), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, FLAGS },
157 { "planes", "set planes to filter",
158 OFFSET(planes), AV_OPT_TYPE_INT, {.i64=7}, 0, 15, FLAGS },
162 AVFILTER_DEFINE_CLASS(bm3d);
164 static int query_formats(AVFilterContext *ctx)
166 static const enum AVPixelFormat pix_fmts[] = {
167 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
168 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
169 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
170 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
171 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
172 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
173 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
175 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
176 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
177 AV_PIX_FMT_YUV440P10,
178 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
179 AV_PIX_FMT_YUV440P12,
180 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
181 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
182 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
183 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
187 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
189 return AVERROR(ENOMEM);
190 return ff_set_common_formats(ctx, fmts_list);
193 static int do_search_boundary(int pos, int plane_boundary, int search_range, int search_step)
197 search_range = search_range / search_step * search_step;
199 if (pos == plane_boundary) {
200 search_boundary = plane_boundary;
201 } else if (pos > plane_boundary) {
202 search_boundary = pos - search_range;
204 while (search_boundary < plane_boundary) {
205 search_boundary += search_step;
208 search_boundary = pos + search_range;
210 while (search_boundary > plane_boundary) {
211 search_boundary -= search_step;
215 return search_boundary;
218 static int search_boundary(int plane_boundary, int search_range, int search_step, int vertical, int y, int x)
220 return do_search_boundary(vertical ? y : x, plane_boundary, search_range, search_step);
223 static int cmp_scores(const void *a, const void *b)
225 const struct PosPairCode *pair1 = a;
226 const struct PosPairCode *pair2 = b;
227 return FFDIFFSIGN(pair1->score, pair2->score);
230 static double do_block_ssd(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
232 const uint8_t *srcp = src + pos->y * src_stride + pos->x;
233 const uint8_t *refp = src + r_y * src_stride + r_x;
234 const int block_size = s->block_size;
238 for (y = 0; y < block_size; y++) {
239 for (x = 0; x < block_size; x++) {
240 double temp = refp[x] - srcp[x];
251 static double do_block_ssd16(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
253 const uint16_t *srcp = (uint16_t *)src + pos->y * src_stride / 2 + pos->x;
254 const uint16_t *refp = (uint16_t *)src + r_y * src_stride / 2 + r_x;
255 const int block_size = s->block_size;
259 for (y = 0; y < block_size; y++) {
260 for (x = 0; x < block_size; x++) {
261 double temp = refp[x] - srcp[x];
265 srcp += src_stride / 2;
266 refp += src_stride / 2;
272 static void do_block_matching_multi(BM3DContext *s, const uint8_t *src, int src_stride, int src_range,
273 const PosCode *search_pos, int search_size, float th_mse,
274 int r_y, int r_x, int plane, int jobnr)
276 SliceContext *sc = &s->slices[jobnr];
277 double MSE2SSE = s->group_size * s->block_size * s->block_size * src_range * src_range / (s->max * s->max);
278 double distMul = 1. / MSE2SSE;
279 double th_sse = th_mse * MSE2SSE;
280 int i, index = sc->nb_match_blocks;
282 for (i = 0; i < search_size; i++) {
283 PosCode pos = search_pos[i];
286 dist = s->do_block_ssd(s, &pos, src, src_stride, r_y, r_x);
288 // Only match similar blocks but not identical blocks
289 if (dist <= th_sse && dist != 0) {
290 const double score = dist * distMul;
292 if (index >= s->group_size && score >= sc->match_blocks[index - 1].score) {
296 if (index >= s->group_size)
297 index = s->group_size - 1;
299 sc->match_blocks[index].score = score;
300 sc->match_blocks[index].y = pos.y;
301 sc->match_blocks[index].x = pos.x;
303 qsort(sc->match_blocks, index, sizeof(PosPairCode), cmp_scores);
307 sc->nb_match_blocks = index;
310 static void block_matching_multi(BM3DContext *s, const uint8_t *ref, int ref_linesize, int y, int x,
311 int exclude_cur_pos, int plane, int jobnr)
313 SliceContext *sc = &s->slices[jobnr];
314 const int width = s->planewidth[plane];
315 const int height = s->planeheight[plane];
316 const int block_size = s->block_size;
317 const int step = s->bm_step;
318 const int range = s->bm_range / step * step;
319 int l = search_boundary(0, range, step, 0, y, x);
320 int r = search_boundary(width - block_size, range, step, 0, y, x);
321 int t = search_boundary(0, range, step, 1, y, x);
322 int b = search_boundary(height - block_size, range, step, 1, y, x);
325 for (j = t; j <= b; j += step) {
326 for (i = l; i <= r; i += step) {
329 if (exclude_cur_pos > 0 && j == y && i == x) {
335 sc->search_positions[index++] = pos;
339 if (exclude_cur_pos == 1) {
340 sc->match_blocks[0].score = 0;
341 sc->match_blocks[0].y = y;
342 sc->match_blocks[0].x = x;
343 sc->nb_match_blocks = 1;
346 do_block_matching_multi(s, ref, ref_linesize, s->bm_range,
347 sc->search_positions, index, s->th_mse, y, x, plane, jobnr);
350 static void block_matching(BM3DContext *s, const uint8_t *ref, int ref_linesize,
351 int j, int i, int plane, int jobnr)
353 SliceContext *sc = &s->slices[jobnr];
355 if (s->group_size == 1 || s->th_mse <= 0.f) {
356 sc->match_blocks[0].score = 1;
357 sc->match_blocks[0].x = i;
358 sc->match_blocks[0].y = j;
359 sc->nb_match_blocks = 1;
363 sc->nb_match_blocks = 0;
364 block_matching_multi(s, ref, ref_linesize, j, i, 1, plane, jobnr);
367 static void get_block_row(const uint8_t *srcp, int src_linesize,
368 int y, int x, int block_size, float *dst)
370 const uint8_t *src = srcp + y * src_linesize + x;
373 for (j = 0; j < block_size; j++) {
378 static void get_block_row16(const uint8_t *srcp, int src_linesize,
379 int y, int x, int block_size, float *dst)
381 const uint16_t *src = (uint16_t *)srcp + y * src_linesize / 2 + x;
384 for (j = 0; j < block_size; j++) {
389 static void basic_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
390 const uint8_t *ref, int ref_linesize,
391 int y, int x, int plane, int jobnr)
393 SliceContext *sc = &s->slices[jobnr];
394 const int buffer_linesize = s->block_size * s->block_size;
395 const int nb_match_blocks = sc->nb_match_blocks;
396 const int block_size = s->block_size;
397 const int width = s->planewidth[plane];
398 const int pgroup_size = s->pgroup_size;
399 const int group_size = s->group_size;
400 float *buffer = sc->buffer;
401 float *bufferh = sc->bufferh;
402 float *bufferv = sc->bufferv;
403 float *bufferz = sc->bufferz;
405 float den_weight, num_weight;
409 for (k = 0; k < nb_match_blocks; k++) {
410 const int y = sc->match_blocks[k].y;
411 const int x = sc->match_blocks[k].x;
413 for (i = 0; i < block_size; i++) {
414 s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
415 av_dct_calc(sc->dctf, bufferh + block_size * i);
418 for (i = 0; i < block_size; i++) {
419 for (j = 0; j < block_size; j++) {
420 bufferv[i * block_size + j] = bufferh[j * block_size + i];
422 av_dct_calc(sc->dctf, bufferv + i * block_size);
425 for (i = 0; i < block_size; i++) {
426 memcpy(buffer + k * buffer_linesize + i * block_size,
427 bufferv + i * block_size, block_size * 4);
431 for (i = 0; i < block_size; i++) {
432 for (j = 0; j < block_size; j++) {
433 for (k = 0; k < nb_match_blocks; k++)
434 bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
436 av_dct_calc(sc->gdctf, bufferz);
437 bufferz += pgroup_size;
441 threshold[0] = s->hard_threshold * s->sigma;
442 threshold[1] = threshold[0] * sqrtf(2.f);
443 threshold[2] = threshold[0] * 2.f;
444 threshold[3] = threshold[0] * sqrtf(8.f);
445 bufferz = sc->bufferz;
447 for (i = 0; i < block_size; i++) {
448 for (j = 0; j < block_size; j++) {
449 for (k = 0; k < nb_match_blocks; k++) {
450 const float thresh = threshold[(j == 0) + (i == 0) + (k == 0)];
452 if (bufferz[k] > thresh || bufferz[k] < -thresh) {
458 bufferz += pgroup_size;
462 bufferz = sc->bufferz;
464 for (i = 0; i < block_size; i++) {
465 for (j = 0; j < block_size; j++) {
467 av_dct_calc(sc->gdcti, bufferz);
468 for (k = 0; k < nb_match_blocks; k++) {
469 buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
471 bufferz += pgroup_size;
475 den_weight = retained < 1 ? 1.f : 1.f / retained;
476 num_weight = den_weight;
479 for (k = 0; k < nb_match_blocks; k++) {
480 float *num = sc->num + y * width + x;
481 float *den = sc->den + y * width + x;
483 for (i = 0; i < block_size; i++) {
484 memcpy(bufferv + i * block_size,
485 buffer + k * buffer_linesize + i * block_size,
489 for (i = 0; i < block_size; i++) {
490 av_dct_calc(sc->dcti, bufferv + block_size * i);
491 for (j = 0; j < block_size; j++) {
492 bufferh[j * block_size + i] = bufferv[i * block_size + j];
496 for (i = 0; i < block_size; i++) {
497 av_dct_calc(sc->dcti, bufferh + block_size * i);
498 for (j = 0; j < block_size; j++) {
499 num[j] += bufferh[i * block_size + j] * num_weight;
500 den[j] += den_weight;
508 static void final_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
509 const uint8_t *ref, int ref_linesize,
510 int y, int x, int plane, int jobnr)
512 SliceContext *sc = &s->slices[jobnr];
513 const int buffer_linesize = s->block_size * s->block_size;
514 const int nb_match_blocks = sc->nb_match_blocks;
515 const int block_size = s->block_size;
516 const int width = s->planewidth[plane];
517 const int pgroup_size = s->pgroup_size;
518 const int group_size = s->group_size;
519 const float sigma_sqr = s->sigma * s->sigma;
520 float *buffer = sc->buffer;
521 float *bufferh = sc->bufferh;
522 float *bufferv = sc->bufferv;
523 float *bufferz = sc->bufferz;
524 float *rbuffer = sc->rbuffer;
525 float *rbufferh = sc->rbufferh;
526 float *rbufferv = sc->rbufferv;
527 float *rbufferz = sc->rbufferz;
528 float den_weight, num_weight;
532 for (k = 0; k < nb_match_blocks; k++) {
533 const int y = sc->match_blocks[k].y;
534 const int x = sc->match_blocks[k].x;
536 for (i = 0; i < block_size; i++) {
537 s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
538 s->get_block_row(ref, ref_linesize, y + i, x, block_size, rbufferh + block_size * i);
539 av_dct_calc(sc->dctf, bufferh + block_size * i);
540 av_dct_calc(sc->dctf, rbufferh + block_size * i);
543 for (i = 0; i < block_size; i++) {
544 for (j = 0; j < block_size; j++) {
545 bufferv[i * block_size + j] = bufferh[j * block_size + i];
546 rbufferv[i * block_size + j] = rbufferh[j * block_size + i];
548 av_dct_calc(sc->dctf, bufferv + i * block_size);
549 av_dct_calc(sc->dctf, rbufferv + i * block_size);
552 for (i = 0; i < block_size; i++) {
553 memcpy(buffer + k * buffer_linesize + i * block_size,
554 bufferv + i * block_size, block_size * 4);
555 memcpy(rbuffer + k * buffer_linesize + i * block_size,
556 rbufferv + i * block_size, block_size * 4);
560 for (i = 0; i < block_size; i++) {
561 for (j = 0; j < block_size; j++) {
562 for (k = 0; k < nb_match_blocks; k++) {
563 bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
564 rbufferz[k] = rbuffer[buffer_linesize * k + i * block_size + j];
566 if (group_size > 1) {
567 av_dct_calc(sc->gdctf, bufferz);
568 av_dct_calc(sc->gdctf, rbufferz);
570 bufferz += pgroup_size;
571 rbufferz += pgroup_size;
575 bufferz = sc->bufferz;
576 rbufferz = sc->rbufferz;
578 for (i = 0; i < block_size; i++) {
579 for (j = 0; j < block_size; j++) {
580 for (k = 0; k < nb_match_blocks; k++) {
581 const float ref_sqr = rbufferz[k] * rbufferz[k];
582 float wiener_coef = ref_sqr / (ref_sqr + sigma_sqr);
584 if (isnan(wiener_coef))
586 bufferz[k] *= wiener_coef;
587 l2_wiener += wiener_coef * wiener_coef;
589 bufferz += pgroup_size;
590 rbufferz += pgroup_size;
594 bufferz = sc->bufferz;
596 for (i = 0; i < block_size; i++) {
597 for (j = 0; j < block_size; j++) {
599 av_dct_calc(sc->gdcti, bufferz);
600 for (k = 0; k < nb_match_blocks; k++) {
601 buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
603 bufferz += pgroup_size;
607 l2_wiener = FFMAX(l2_wiener, 1e-15f);
608 den_weight = 1.f / l2_wiener;
609 num_weight = den_weight;
611 for (k = 0; k < nb_match_blocks; k++) {
612 float *num = sc->num + y * width + x;
613 float *den = sc->den + y * width + x;
615 for (i = 0; i < block_size; i++) {
616 memcpy(bufferv + i * block_size,
617 buffer + k * buffer_linesize + i * block_size,
621 for (i = 0; i < block_size; i++) {
622 av_dct_calc(sc->dcti, bufferv + block_size * i);
623 for (j = 0; j < block_size; j++) {
624 bufferh[j * block_size + i] = bufferv[i * block_size + j];
628 for (i = 0; i < block_size; i++) {
629 av_dct_calc(sc->dcti, bufferh + block_size * i);
630 for (j = 0; j < block_size; j++) {
631 num[j] += bufferh[i * block_size + j] * num_weight;
632 den[j] += den_weight;
640 static void do_output(BM3DContext *s, uint8_t *dst, int dst_linesize,
641 int plane, int nb_jobs)
643 const int height = s->planeheight[plane];
644 const int width = s->planewidth[plane];
647 for (i = 0; i < height; i++) {
648 for (j = 0; j < width; j++) {
649 uint8_t *dstp = dst + i * dst_linesize;
653 for (k = 0; k < nb_jobs; k++) {
654 SliceContext *sc = &s->slices[k];
655 float num = sc->num[i * width + j];
656 float den = sc->den[i * width + j];
662 dstp[j] = av_clip_uint8(lrintf(sum_num / sum_den));
667 static void do_output16(BM3DContext *s, uint8_t *dst, int dst_linesize,
668 int plane, int nb_jobs)
670 const int height = s->planeheight[plane];
671 const int width = s->planewidth[plane];
672 const int depth = s->depth;
675 for (i = 0; i < height; i++) {
676 for (j = 0; j < width; j++) {
677 uint16_t *dstp = (uint16_t *)dst + i * dst_linesize / 2;
681 for (k = 0; k < nb_jobs; k++) {
682 SliceContext *sc = &s->slices[k];
683 float num = sc->num[i * width + j];
684 float den = sc->den[i * width + j];
690 dstp[j] = av_clip_uintp2_c(lrintf(sum_num / sum_den), depth);
695 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
697 BM3DContext *s = ctx->priv;
698 SliceContext *sc = &s->slices[jobnr];
699 const int block_step = s->block_step;
700 ThreadData *td = arg;
701 const uint8_t *src = td->src;
702 const uint8_t *ref = td->ref;
703 const int src_linesize = td->src_linesize;
704 const int ref_linesize = td->ref_linesize;
705 const int plane = td->plane;
706 const int width = s->planewidth[plane];
707 const int height = s->planeheight[plane];
708 const int block_pos_bottom = FFMAX(0, height - s->block_size);
709 const int block_pos_right = FFMAX(0, width - s->block_size);
710 const int slice_start = (((height + block_step - 1) / block_step) * jobnr / nb_jobs) * block_step;
711 const int slice_end = (jobnr == nb_jobs - 1) ? block_pos_bottom + block_step :
712 (((height + block_step - 1) / block_step) * (jobnr + 1) / nb_jobs) * block_step;
715 memset(sc->num, 0, width * height * sizeof(FFTSample));
716 memset(sc->den, 0, width * height * sizeof(FFTSample));
718 for (j = slice_start; j < slice_end; j += block_step) {
719 if (j > block_pos_bottom) {
720 j = block_pos_bottom;
723 for (i = 0; i < block_pos_right + block_step; i += block_step) {
724 if (i > block_pos_right) {
728 block_matching(s, ref, ref_linesize, j, i, plane, jobnr);
730 s->block_filtering(s, src, src_linesize,
731 ref, ref_linesize, j, i, plane, jobnr);
738 static int filter_frame(AVFilterContext *ctx, AVFrame **out, AVFrame *in, AVFrame *ref)
740 BM3DContext *s = ctx->priv;
741 AVFilterLink *outlink = ctx->outputs[0];
744 *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
746 return AVERROR(ENOMEM);
747 av_frame_copy_props(*out, in);
749 for (p = 0; p < s->nb_planes; p++) {
750 const int nb_jobs = FFMAX(1, FFMIN(s->nb_threads, s->planeheight[p] / s->block_size));
753 if (!((1 << p) & s->planes) || ctx->is_disabled) {
754 av_image_copy_plane((*out)->data[p], (*out)->linesize[p],
755 in->data[p], in->linesize[p],
756 s->planewidth[p], s->planeheight[p]);
760 td.src = in->data[p];
761 td.src_linesize = in->linesize[p];
762 td.ref = ref->data[p];
763 td.ref_linesize = ref->linesize[p];
765 ctx->internal->execute(ctx, filter_slice, &td, NULL, nb_jobs);
767 s->do_output(s, (*out)->data[p], (*out)->linesize[p], p, nb_jobs);
773 #define SQR(x) ((x) * (x))
775 static int config_input(AVFilterLink *inlink)
777 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
778 AVFilterContext *ctx = inlink->dst;
779 BM3DContext *s = ctx->priv;
782 s->nb_threads = FFMIN(ff_filter_get_nb_threads(ctx), MAX_NB_THREADS);
783 s->nb_planes = av_pix_fmt_count_planes(inlink->format);
784 s->depth = desc->comp[0].depth;
785 s->max = (1 << s->depth) - 1;
786 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
787 s->planeheight[0] = s->planeheight[3] = inlink->h;
788 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
789 s->planewidth[0] = s->planewidth[3] = inlink->w;
791 for (group_bits = 4; 1 << group_bits < s->group_size; group_bits++);
792 s->group_bits = group_bits;
793 s->pgroup_size = 1 << group_bits;
795 for (i = 0; i < s->nb_threads; i++) {
796 SliceContext *sc = &s->slices[i];
798 sc->num = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(FFTSample));
799 sc->den = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(FFTSample));
800 if (!sc->num || !sc->den)
801 return AVERROR(ENOMEM);
803 sc->dctf = av_dct_init(av_log2(s->block_size), DCT_II);
804 sc->dcti = av_dct_init(av_log2(s->block_size), DCT_III);
805 if (!sc->dctf || !sc->dcti)
806 return AVERROR(ENOMEM);
808 if (s->group_bits > 1) {
809 sc->gdctf = av_dct_init(s->group_bits, DCT_II);
810 sc->gdcti = av_dct_init(s->group_bits, DCT_III);
811 if (!sc->gdctf || !sc->gdcti)
812 return AVERROR(ENOMEM);
815 sc->buffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->buffer));
816 sc->bufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->bufferz));
817 sc->bufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferh));
818 sc->bufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferv));
819 if (!sc->bufferh || !sc->bufferv || !sc->buffer || !sc->bufferz)
820 return AVERROR(ENOMEM);
822 if (s->mode == FINAL) {
823 sc->rbuffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbuffer));
824 sc->rbufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbufferz));
825 sc->rbufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferh));
826 sc->rbufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferv));
827 if (!sc->rbufferh || !sc->rbufferv || !sc->rbuffer || !sc->rbufferz)
828 return AVERROR(ENOMEM);
831 sc->search_positions = av_calloc(SQR(2 * s->bm_range / s->bm_step + 1), sizeof(*sc->search_positions));
832 if (!sc->search_positions)
833 return AVERROR(ENOMEM);
836 s->do_output = do_output;
837 s->do_block_ssd = do_block_ssd;
838 s->get_block_row = get_block_row;
841 s->do_output = do_output16;
842 s->do_block_ssd = do_block_ssd16;
843 s->get_block_row = get_block_row16;
849 static int activate(AVFilterContext *ctx)
851 BM3DContext *s = ctx->priv;
854 AVFrame *frame = NULL;
859 FF_FILTER_FORWARD_STATUS_BACK(ctx->outputs[0], ctx->inputs[0]);
861 if ((ret = ff_inlink_consume_frame(ctx->inputs[0], &frame)) > 0) {
862 ret = filter_frame(ctx, &out, frame, frame);
863 av_frame_free(&frame);
866 ret = ff_filter_frame(ctx->outputs[0], out);
870 } else if (ff_inlink_acknowledge_status(ctx->inputs[0], &status, &pts)) {
871 ff_outlink_set_status(ctx->outputs[0], status, pts);
874 if (ff_outlink_frame_wanted(ctx->outputs[0]))
875 ff_inlink_request_frame(ctx->inputs[0]);
879 return ff_framesync_activate(&s->fs);
883 static int process_frame(FFFrameSync *fs)
885 AVFilterContext *ctx = fs->parent;
886 BM3DContext *s = fs->opaque;
887 AVFilterLink *outlink = ctx->outputs[0];
888 AVFrame *out = NULL, *src, *ref;
891 if ((ret = ff_framesync_get_frame(&s->fs, 0, &src, 0)) < 0 ||
892 (ret = ff_framesync_get_frame(&s->fs, 1, &ref, 0)) < 0)
895 if ((ret = filter_frame(ctx, &out, src, ref)) < 0)
898 out->pts = av_rescale_q(src->pts, s->fs.time_base, outlink->time_base);
900 return ff_filter_frame(outlink, out);
903 static av_cold int init(AVFilterContext *ctx)
905 BM3DContext *s = ctx->priv;
906 AVFilterPad pad = { 0 };
909 if (s->mode == BASIC) {
910 if (s->th_mse == 0.f)
911 s->th_mse = 400.f + s->sigma * 80.f;
912 s->block_filtering = basic_block_filtering;
913 } else if (s->mode == FINAL) {
915 av_log(ctx, AV_LOG_WARNING, "Reference stream is mandatory in final estimation mode.\n");
918 if (s->th_mse == 0.f)
919 s->th_mse = 200.f + s->sigma * 10.f;
921 s->block_filtering = final_block_filtering;
926 s->block_size = 1 << s->block_size;
928 if (s->block_step > s->block_size) {
929 av_log(ctx, AV_LOG_WARNING, "bstep: %d can't be bigger than block size. Changing to %d.\n",
930 s->block_step, s->block_size);
931 s->block_step = s->block_size;
933 if (s->bm_step > s->bm_range) {
934 av_log(ctx, AV_LOG_WARNING, "mstep: %d can't be bigger than block matching range. Changing to %d.\n",
935 s->bm_step, s->bm_range);
936 s->bm_step = s->bm_range;
939 pad.type = AVMEDIA_TYPE_VIDEO;
940 pad.name = av_strdup("source");
941 pad.config_props = config_input;
943 return AVERROR(ENOMEM);
945 if ((ret = ff_insert_inpad(ctx, 0, &pad)) < 0) {
951 pad.type = AVMEDIA_TYPE_VIDEO;
952 pad.name = av_strdup("reference");
953 pad.config_props = NULL;
955 return AVERROR(ENOMEM);
957 if ((ret = ff_insert_inpad(ctx, 1, &pad)) < 0) {
966 static int config_output(AVFilterLink *outlink)
968 AVFilterContext *ctx = outlink->src;
969 BM3DContext *s = ctx->priv;
970 AVFilterLink *src = ctx->inputs[0];
976 ref = ctx->inputs[1];
978 if (src->format != ref->format) {
979 av_log(ctx, AV_LOG_ERROR, "inputs must be of same pixel format\n");
980 return AVERROR(EINVAL);
982 if (src->w != ref->w ||
984 av_log(ctx, AV_LOG_ERROR, "First input link %s parameters "
985 "(size %dx%d) do not match the corresponding "
986 "second input link %s parameters (%dx%d) ",
987 ctx->input_pads[0].name, src->w, src->h,
988 ctx->input_pads[1].name, ref->w, ref->h);
989 return AVERROR(EINVAL);
995 outlink->time_base = src->time_base;
996 outlink->sample_aspect_ratio = src->sample_aspect_ratio;
997 outlink->frame_rate = src->frame_rate;
1002 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
1006 in[0].time_base = src->time_base;
1007 in[1].time_base = ref->time_base;
1009 in[0].before = EXT_STOP;
1010 in[0].after = EXT_STOP;
1012 in[1].before = EXT_STOP;
1013 in[1].after = EXT_STOP;
1015 s->fs.on_event = process_frame;
1017 return ff_framesync_configure(&s->fs);
1020 static av_cold void uninit(AVFilterContext *ctx)
1022 BM3DContext *s = ctx->priv;
1025 for (i = 0; i < ctx->nb_inputs; i++)
1026 av_freep(&ctx->input_pads[i].name);
1029 ff_framesync_uninit(&s->fs);
1031 for (i = 0; i < s->nb_threads; i++) {
1032 SliceContext *sc = &s->slices[i];
1037 av_dct_end(sc->gdctf);
1038 av_dct_end(sc->gdcti);
1039 av_dct_end(sc->dctf);
1040 av_dct_end(sc->dcti);
1042 av_freep(&sc->buffer);
1043 av_freep(&sc->bufferh);
1044 av_freep(&sc->bufferv);
1045 av_freep(&sc->bufferz);
1046 av_freep(&sc->rbuffer);
1047 av_freep(&sc->rbufferh);
1048 av_freep(&sc->rbufferv);
1049 av_freep(&sc->rbufferz);
1051 av_freep(&sc->search_positions);
1055 static const AVFilterPad bm3d_outputs[] = {
1058 .type = AVMEDIA_TYPE_VIDEO,
1059 .config_props = config_output,
1064 AVFilter ff_vf_bm3d = {
1066 .description = NULL_IF_CONFIG_SMALL("Block-Matching 3D denoiser."),
1067 .priv_size = sizeof(BM3DContext),
1070 .activate = activate,
1071 .query_formats = query_formats,
1073 .outputs = bm3d_outputs,
1074 .priv_class = &bm3d_class,
1075 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
1076 AVFILTER_FLAG_DYNAMIC_INPUTS |
1077 AVFILTER_FLAG_SLICE_THREADS,