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[] = {
168 AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
169 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY16,
170 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
171 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
172 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
173 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
174 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
176 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
177 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
178 AV_PIX_FMT_YUV440P10,
179 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
180 AV_PIX_FMT_YUV440P12,
181 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
182 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
183 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
184 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
188 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
190 return AVERROR(ENOMEM);
191 return ff_set_common_formats(ctx, fmts_list);
194 static int do_search_boundary(int pos, int plane_boundary, int search_range, int search_step)
198 search_range = search_range / search_step * search_step;
200 if (pos == plane_boundary) {
201 search_boundary = plane_boundary;
202 } else if (pos > plane_boundary) {
203 search_boundary = pos - search_range;
205 while (search_boundary < plane_boundary) {
206 search_boundary += search_step;
209 search_boundary = pos + search_range;
211 while (search_boundary > plane_boundary) {
212 search_boundary -= search_step;
216 return search_boundary;
219 static int search_boundary(int plane_boundary, int search_range, int search_step, int vertical, int y, int x)
221 return do_search_boundary(vertical ? y : x, plane_boundary, search_range, search_step);
224 static int cmp_scores(const void *a, const void *b)
226 const struct PosPairCode *pair1 = a;
227 const struct PosPairCode *pair2 = b;
228 return FFDIFFSIGN(pair1->score, pair2->score);
231 static double do_block_ssd(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
233 const uint8_t *srcp = src + pos->y * src_stride + pos->x;
234 const uint8_t *refp = src + r_y * src_stride + r_x;
235 const int block_size = s->block_size;
239 for (y = 0; y < block_size; y++) {
240 for (x = 0; x < block_size; x++) {
241 double temp = refp[x] - srcp[x];
252 static double do_block_ssd16(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
254 const uint16_t *srcp = (uint16_t *)src + pos->y * src_stride / 2 + pos->x;
255 const uint16_t *refp = (uint16_t *)src + r_y * src_stride / 2 + r_x;
256 const int block_size = s->block_size;
260 for (y = 0; y < block_size; y++) {
261 for (x = 0; x < block_size; x++) {
262 double temp = refp[x] - srcp[x];
266 srcp += src_stride / 2;
267 refp += src_stride / 2;
273 static void do_block_matching_multi(BM3DContext *s, const uint8_t *src, int src_stride, int src_range,
274 const PosCode *search_pos, int search_size, float th_mse,
275 int r_y, int r_x, int plane, int jobnr)
277 SliceContext *sc = &s->slices[jobnr];
278 double MSE2SSE = s->group_size * s->block_size * s->block_size * src_range * src_range / (s->max * s->max);
279 double distMul = 1. / MSE2SSE;
280 double th_sse = th_mse * MSE2SSE;
281 int i, index = sc->nb_match_blocks;
283 for (i = 0; i < search_size; i++) {
284 PosCode pos = search_pos[i];
287 dist = s->do_block_ssd(s, &pos, src, src_stride, r_y, r_x);
289 // Only match similar blocks but not identical blocks
290 if (dist <= th_sse && dist != 0) {
291 const double score = dist * distMul;
293 if (index >= s->group_size && score >= sc->match_blocks[index - 1].score) {
297 if (index >= s->group_size)
298 index = s->group_size - 1;
300 sc->match_blocks[index].score = score;
301 sc->match_blocks[index].y = pos.y;
302 sc->match_blocks[index].x = pos.x;
304 qsort(sc->match_blocks, index, sizeof(PosPairCode), cmp_scores);
308 sc->nb_match_blocks = index;
311 static void block_matching_multi(BM3DContext *s, const uint8_t *ref, int ref_linesize, int y, int x,
312 int exclude_cur_pos, int plane, int jobnr)
314 SliceContext *sc = &s->slices[jobnr];
315 const int width = s->planewidth[plane];
316 const int height = s->planeheight[plane];
317 const int block_size = s->block_size;
318 const int step = s->bm_step;
319 const int range = s->bm_range / step * step;
320 int l = search_boundary(0, range, step, 0, y, x);
321 int r = search_boundary(width - block_size, range, step, 0, y, x);
322 int t = search_boundary(0, range, step, 1, y, x);
323 int b = search_boundary(height - block_size, range, step, 1, y, x);
326 for (j = t; j <= b; j += step) {
327 for (i = l; i <= r; i += step) {
330 if (exclude_cur_pos > 0 && j == y && i == x) {
336 sc->search_positions[index++] = pos;
340 if (exclude_cur_pos == 1) {
341 sc->match_blocks[0].score = 0;
342 sc->match_blocks[0].y = y;
343 sc->match_blocks[0].x = x;
344 sc->nb_match_blocks = 1;
347 do_block_matching_multi(s, ref, ref_linesize, s->bm_range,
348 sc->search_positions, index, s->th_mse, y, x, plane, jobnr);
351 static void block_matching(BM3DContext *s, const uint8_t *ref, int ref_linesize,
352 int j, int i, int plane, int jobnr)
354 SliceContext *sc = &s->slices[jobnr];
356 if (s->group_size == 1 || s->th_mse <= 0.f) {
357 sc->match_blocks[0].score = 1;
358 sc->match_blocks[0].x = i;
359 sc->match_blocks[0].y = j;
360 sc->nb_match_blocks = 1;
364 sc->nb_match_blocks = 0;
365 block_matching_multi(s, ref, ref_linesize, j, i, 1, plane, jobnr);
368 static void get_block_row(const uint8_t *srcp, int src_linesize,
369 int y, int x, int block_size, float *dst)
371 const uint8_t *src = srcp + y * src_linesize + x;
374 for (j = 0; j < block_size; j++) {
379 static void get_block_row16(const uint8_t *srcp, int src_linesize,
380 int y, int x, int block_size, float *dst)
382 const uint16_t *src = (uint16_t *)srcp + y * src_linesize / 2 + x;
385 for (j = 0; j < block_size; j++) {
390 static void basic_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
391 const uint8_t *ref, int ref_linesize,
392 int y, int x, int plane, int jobnr)
394 SliceContext *sc = &s->slices[jobnr];
395 const int buffer_linesize = s->block_size * s->block_size;
396 const int nb_match_blocks = sc->nb_match_blocks;
397 const int block_size = s->block_size;
398 const int width = s->planewidth[plane];
399 const int pgroup_size = s->pgroup_size;
400 const int group_size = s->group_size;
401 float *buffer = sc->buffer;
402 float *bufferh = sc->bufferh;
403 float *bufferv = sc->bufferv;
404 float *bufferz = sc->bufferz;
406 float den_weight, num_weight;
410 for (k = 0; k < nb_match_blocks; k++) {
411 const int y = sc->match_blocks[k].y;
412 const int x = sc->match_blocks[k].x;
414 for (i = 0; i < block_size; i++) {
415 s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
416 av_dct_calc(sc->dctf, bufferh + block_size * i);
419 for (i = 0; i < block_size; i++) {
420 for (j = 0; j < block_size; j++) {
421 bufferv[i * block_size + j] = bufferh[j * block_size + i];
423 av_dct_calc(sc->dctf, bufferv + i * block_size);
426 for (i = 0; i < block_size; i++) {
427 memcpy(buffer + k * buffer_linesize + i * block_size,
428 bufferv + i * block_size, block_size * 4);
432 for (i = 0; i < block_size; i++) {
433 for (j = 0; j < block_size; j++) {
434 for (k = 0; k < nb_match_blocks; k++)
435 bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
437 av_dct_calc(sc->gdctf, bufferz);
438 bufferz += pgroup_size;
442 threshold[0] = s->hard_threshold * s->sigma;
443 threshold[1] = threshold[0] * sqrtf(2.f);
444 threshold[2] = threshold[0] * 2.f;
445 threshold[3] = threshold[0] * sqrtf(8.f);
446 bufferz = sc->bufferz;
448 for (i = 0; i < block_size; i++) {
449 for (j = 0; j < block_size; j++) {
450 for (k = 0; k < nb_match_blocks; k++) {
451 const float thresh = threshold[(j == 0) + (i == 0) + (k == 0)];
453 if (bufferz[k] > thresh || bufferz[k] < -thresh) {
459 bufferz += pgroup_size;
463 bufferz = sc->bufferz;
465 for (i = 0; i < block_size; i++) {
466 for (j = 0; j < block_size; j++) {
468 av_dct_calc(sc->gdcti, bufferz);
469 for (k = 0; k < nb_match_blocks; k++) {
470 buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
472 bufferz += pgroup_size;
476 den_weight = retained < 1 ? 1.f : 1.f / retained;
477 num_weight = den_weight;
480 for (k = 0; k < nb_match_blocks; k++) {
481 float *num = sc->num + y * width + x;
482 float *den = sc->den + y * width + x;
484 for (i = 0; i < block_size; i++) {
485 memcpy(bufferv + i * block_size,
486 buffer + k * buffer_linesize + i * block_size,
490 for (i = 0; i < block_size; i++) {
491 av_dct_calc(sc->dcti, bufferv + block_size * i);
492 for (j = 0; j < block_size; j++) {
493 bufferh[j * block_size + i] = bufferv[i * block_size + j];
497 for (i = 0; i < block_size; i++) {
498 av_dct_calc(sc->dcti, bufferh + block_size * i);
499 for (j = 0; j < block_size; j++) {
500 num[j] += bufferh[i * block_size + j] * num_weight;
501 den[j] += den_weight;
509 static void final_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
510 const uint8_t *ref, int ref_linesize,
511 int y, int x, int plane, int jobnr)
513 SliceContext *sc = &s->slices[jobnr];
514 const int buffer_linesize = s->block_size * s->block_size;
515 const int nb_match_blocks = sc->nb_match_blocks;
516 const int block_size = s->block_size;
517 const int width = s->planewidth[plane];
518 const int pgroup_size = s->pgroup_size;
519 const int group_size = s->group_size;
520 const float sigma_sqr = s->sigma * s->sigma;
521 float *buffer = sc->buffer;
522 float *bufferh = sc->bufferh;
523 float *bufferv = sc->bufferv;
524 float *bufferz = sc->bufferz;
525 float *rbuffer = sc->rbuffer;
526 float *rbufferh = sc->rbufferh;
527 float *rbufferv = sc->rbufferv;
528 float *rbufferz = sc->rbufferz;
529 float den_weight, num_weight;
533 for (k = 0; k < nb_match_blocks; k++) {
534 const int y = sc->match_blocks[k].y;
535 const int x = sc->match_blocks[k].x;
537 for (i = 0; i < block_size; i++) {
538 s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
539 s->get_block_row(ref, ref_linesize, y + i, x, block_size, rbufferh + block_size * i);
540 av_dct_calc(sc->dctf, bufferh + block_size * i);
541 av_dct_calc(sc->dctf, rbufferh + block_size * i);
544 for (i = 0; i < block_size; i++) {
545 for (j = 0; j < block_size; j++) {
546 bufferv[i * block_size + j] = bufferh[j * block_size + i];
547 rbufferv[i * block_size + j] = rbufferh[j * block_size + i];
549 av_dct_calc(sc->dctf, bufferv + i * block_size);
550 av_dct_calc(sc->dctf, rbufferv + i * block_size);
553 for (i = 0; i < block_size; i++) {
554 memcpy(buffer + k * buffer_linesize + i * block_size,
555 bufferv + i * block_size, block_size * 4);
556 memcpy(rbuffer + k * buffer_linesize + i * block_size,
557 rbufferv + i * block_size, block_size * 4);
561 for (i = 0; i < block_size; i++) {
562 for (j = 0; j < block_size; j++) {
563 for (k = 0; k < nb_match_blocks; k++) {
564 bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
565 rbufferz[k] = rbuffer[buffer_linesize * k + i * block_size + j];
567 if (group_size > 1) {
568 av_dct_calc(sc->gdctf, bufferz);
569 av_dct_calc(sc->gdctf, rbufferz);
571 bufferz += pgroup_size;
572 rbufferz += pgroup_size;
576 bufferz = sc->bufferz;
577 rbufferz = sc->rbufferz;
579 for (i = 0; i < block_size; i++) {
580 for (j = 0; j < block_size; j++) {
581 for (k = 0; k < nb_match_blocks; k++) {
582 const float ref_sqr = rbufferz[k] * rbufferz[k];
583 float wiener_coef = ref_sqr / (ref_sqr + sigma_sqr);
585 if (isnan(wiener_coef))
587 bufferz[k] *= wiener_coef;
588 l2_wiener += wiener_coef * wiener_coef;
590 bufferz += pgroup_size;
591 rbufferz += pgroup_size;
595 bufferz = sc->bufferz;
597 for (i = 0; i < block_size; i++) {
598 for (j = 0; j < block_size; j++) {
600 av_dct_calc(sc->gdcti, bufferz);
601 for (k = 0; k < nb_match_blocks; k++) {
602 buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
604 bufferz += pgroup_size;
608 l2_wiener = FFMAX(l2_wiener, 1e-15f);
609 den_weight = 1.f / l2_wiener;
610 num_weight = den_weight;
612 for (k = 0; k < nb_match_blocks; k++) {
613 float *num = sc->num + y * width + x;
614 float *den = sc->den + y * width + x;
616 for (i = 0; i < block_size; i++) {
617 memcpy(bufferv + i * block_size,
618 buffer + k * buffer_linesize + i * block_size,
622 for (i = 0; i < block_size; i++) {
623 av_dct_calc(sc->dcti, bufferv + block_size * i);
624 for (j = 0; j < block_size; j++) {
625 bufferh[j * block_size + i] = bufferv[i * block_size + j];
629 for (i = 0; i < block_size; i++) {
630 av_dct_calc(sc->dcti, bufferh + block_size * i);
631 for (j = 0; j < block_size; j++) {
632 num[j] += bufferh[i * block_size + j] * num_weight;
633 den[j] += den_weight;
641 static void do_output(BM3DContext *s, uint8_t *dst, int dst_linesize,
642 int plane, int nb_jobs)
644 const int height = s->planeheight[plane];
645 const int width = s->planewidth[plane];
648 for (i = 0; i < height; i++) {
649 for (j = 0; j < width; j++) {
650 uint8_t *dstp = dst + i * dst_linesize;
654 for (k = 0; k < nb_jobs; k++) {
655 SliceContext *sc = &s->slices[k];
656 float num = sc->num[i * width + j];
657 float den = sc->den[i * width + j];
663 dstp[j] = av_clip_uint8(sum_num / sum_den);
668 static void do_output16(BM3DContext *s, uint8_t *dst, int dst_linesize,
669 int plane, int nb_jobs)
671 const int height = s->planeheight[plane];
672 const int width = s->planewidth[plane];
673 const int depth = s->depth;
676 for (i = 0; i < height; i++) {
677 for (j = 0; j < width; j++) {
678 uint16_t *dstp = (uint16_t *)dst + i * dst_linesize / 2;
682 for (k = 0; k < nb_jobs; k++) {
683 SliceContext *sc = &s->slices[k];
684 float num = sc->num[i * width + j];
685 float den = sc->den[i * width + j];
691 dstp[j] = av_clip_uintp2_c(sum_num / sum_den, depth);
696 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
698 BM3DContext *s = ctx->priv;
699 SliceContext *sc = &s->slices[jobnr];
700 const int block_step = s->block_step;
701 ThreadData *td = arg;
702 const uint8_t *src = td->src;
703 const uint8_t *ref = td->ref;
704 const int src_linesize = td->src_linesize;
705 const int ref_linesize = td->ref_linesize;
706 const int plane = td->plane;
707 const int width = s->planewidth[plane];
708 const int height = s->planeheight[plane];
709 const int block_pos_bottom = height - s->block_size;
710 const int block_pos_right = width - s->block_size;
711 const int slice_start = (((height + block_step - 1) / block_step) * jobnr / nb_jobs) * block_step;
712 const int slice_end = (jobnr == nb_jobs - 1) ? block_pos_bottom + block_step :
713 (((height + block_step - 1) / block_step) * (jobnr + 1) / nb_jobs) * block_step;
716 memset(sc->num, 0, width * height * sizeof(FFTSample));
717 memset(sc->den, 0, width * height * sizeof(FFTSample));
719 for (j = slice_start; j < slice_end; j += block_step) {
720 if (j > block_pos_bottom) {
721 j = block_pos_bottom;
724 for (i = 0; i < block_pos_right + block_step; i += block_step) {
725 if (i > block_pos_right) {
729 block_matching(s, ref, ref_linesize, j, i, plane, jobnr);
731 s->block_filtering(s, src, src_linesize,
732 ref, ref_linesize, j, i, plane, jobnr);
739 static int filter_frame(AVFilterContext *ctx, AVFrame **out, AVFrame *in, AVFrame *ref)
741 BM3DContext *s = ctx->priv;
742 AVFilterLink *outlink = ctx->outputs[0];
745 *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
747 return AVERROR(ENOMEM);
748 av_frame_copy_props(*out, in);
750 for (p = 0; p < s->nb_planes; p++) {
751 const int nb_jobs = FFMIN(s->nb_threads, s->planeheight[p] / s->block_step);
754 if (!((1 << p) & s->planes) || ctx->is_disabled) {
755 av_image_copy_plane((*out)->data[p], (*out)->linesize[p],
756 in->data[p], in->linesize[p],
757 s->planewidth[p], s->planeheight[p]);
761 td.src = in->data[p];
762 td.src_linesize = in->linesize[p];
763 td.ref = ref->data[p];
764 td.ref_linesize = ref->linesize[p];
766 ctx->internal->execute(ctx, filter_slice, &td, NULL, nb_jobs);
768 s->do_output(s, (*out)->data[p], (*out)->linesize[p], p, nb_jobs);
774 #define SQR(x) ((x) * (x))
776 static int config_input(AVFilterLink *inlink)
778 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
779 AVFilterContext *ctx = inlink->dst;
780 BM3DContext *s = ctx->priv;
783 s->nb_threads = FFMIN(ff_filter_get_nb_threads(ctx), MAX_NB_THREADS);
784 s->nb_planes = av_pix_fmt_count_planes(inlink->format);
785 s->depth = desc->comp[0].depth;
786 s->max = (1 << s->depth) - 1;
787 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
788 s->planeheight[0] = s->planeheight[3] = inlink->h;
789 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
790 s->planewidth[0] = s->planewidth[3] = inlink->w;
792 for (group_bits = 4; 1 << group_bits < s->group_size; group_bits++);
793 s->group_bits = group_bits;
794 s->pgroup_size = 1 << group_bits;
796 for (i = 0; i < s->nb_threads; i++) {
797 SliceContext *sc = &s->slices[i];
799 sc->num = av_calloc(s->planewidth[0] * s->planeheight[0], sizeof(FFTSample));
800 sc->den = av_calloc(s->planewidth[0] * s->planeheight[0], sizeof(FFTSample));
801 if (!sc->num || !sc->den)
802 return AVERROR(ENOMEM);
804 sc->dctf = av_dct_init(av_log2(s->block_size), DCT_II);
805 sc->dcti = av_dct_init(av_log2(s->block_size), DCT_III);
806 if (!sc->dctf || !sc->dcti)
807 return AVERROR(ENOMEM);
809 if (s->group_bits > 1) {
810 sc->gdctf = av_dct_init(s->group_bits, DCT_II);
811 sc->gdcti = av_dct_init(s->group_bits, DCT_III);
812 if (!sc->gdctf || !sc->gdcti)
813 return AVERROR(ENOMEM);
816 sc->buffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->buffer));
817 sc->bufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->bufferz));
818 sc->bufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferh));
819 sc->bufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferv));
820 if (!sc->bufferh || !sc->bufferv || !sc->buffer || !sc->bufferz)
821 return AVERROR(ENOMEM);
823 if (s->mode == FINAL) {
824 sc->rbuffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbuffer));
825 sc->rbufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbufferz));
826 sc->rbufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferh));
827 sc->rbufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferv));
828 if (!sc->rbufferh || !sc->rbufferv || !sc->rbuffer || !sc->rbufferz)
829 return AVERROR(ENOMEM);
832 sc->search_positions = av_calloc(SQR(2 * s->bm_range / s->bm_step + 1), sizeof(*sc->search_positions));
833 if (!sc->search_positions)
834 return AVERROR(ENOMEM);
837 s->do_output = do_output;
838 s->do_block_ssd = do_block_ssd;
839 s->get_block_row = get_block_row;
842 s->do_output = do_output16;
843 s->do_block_ssd = do_block_ssd16;
844 s->get_block_row = get_block_row16;
850 static int activate(AVFilterContext *ctx)
852 BM3DContext *s = ctx->priv;
855 AVFrame *frame = NULL;
860 if ((ret = ff_inlink_consume_frame(ctx->inputs[0], &frame)) > 0) {
861 ret = filter_frame(ctx, &out, frame, frame);
862 av_frame_free(&frame);
865 ret = ff_filter_frame(ctx->outputs[0], out);
869 } else if (ff_inlink_acknowledge_status(ctx->inputs[0], &status, &pts)) {
870 ff_outlink_set_status(ctx->outputs[0], status, pts);
873 if (ff_outlink_frame_wanted(ctx->outputs[0]))
874 ff_inlink_request_frame(ctx->inputs[0]);
878 return ff_framesync_activate(&s->fs);
882 static int process_frame(FFFrameSync *fs)
884 AVFilterContext *ctx = fs->parent;
885 BM3DContext *s = fs->opaque;
886 AVFilterLink *outlink = ctx->outputs[0];
887 AVFrame *out = NULL, *src, *ref;
890 if ((ret = ff_framesync_get_frame(&s->fs, 0, &src, 0)) < 0 ||
891 (ret = ff_framesync_get_frame(&s->fs, 1, &ref, 0)) < 0)
894 if ((ret = filter_frame(ctx, &out, src, ref)) < 0)
897 out->pts = av_rescale_q(src->pts, s->fs.time_base, outlink->time_base);
899 return ff_filter_frame(outlink, out);
902 static av_cold int init(AVFilterContext *ctx)
904 BM3DContext *s = ctx->priv;
905 AVFilterPad pad = { 0 };
908 if (s->mode == BASIC) {
909 if (s->th_mse == 0.f)
910 s->th_mse = 400.f + s->sigma * 80.f;
911 s->block_filtering = basic_block_filtering;
912 } else if (s->mode == FINAL) {
914 av_log(ctx, AV_LOG_WARNING, "Reference stream is mandatory in final estimation mode.\n");
917 if (s->th_mse == 0.f)
918 s->th_mse = 200.f + s->sigma * 10.f;
920 s->block_filtering = final_block_filtering;
925 s->block_size = 1 << s->block_size;
927 if (s->block_step > s->block_size) {
928 av_log(ctx, AV_LOG_WARNING, "bstep: %d can't be bigger than block size. Changing to %d.\n",
929 s->block_step, s->block_size);
930 s->block_step = s->block_size;
932 if (s->bm_step > s->bm_range) {
933 av_log(ctx, AV_LOG_WARNING, "mstep: %d can't be bigger than block matching range. Changing to %d.\n",
934 s->bm_step, s->bm_range);
935 s->bm_step = s->bm_range;
938 pad.type = AVMEDIA_TYPE_VIDEO;
939 pad.name = av_strdup("source");
940 pad.config_props = config_input;
942 return AVERROR(ENOMEM);
944 if ((ret = ff_insert_inpad(ctx, 0, &pad)) < 0) {
950 pad.type = AVMEDIA_TYPE_VIDEO;
951 pad.name = av_strdup("reference");
952 pad.config_props = NULL;
954 return AVERROR(ENOMEM);
956 if ((ret = ff_insert_inpad(ctx, 1, &pad)) < 0) {
965 static int config_output(AVFilterLink *outlink)
967 AVFilterContext *ctx = outlink->src;
968 BM3DContext *s = ctx->priv;
969 AVFilterLink *src = ctx->inputs[0];
975 ref = ctx->inputs[1];
977 if (src->format != ref->format) {
978 av_log(ctx, AV_LOG_ERROR, "inputs must be of same pixel format\n");
979 return AVERROR(EINVAL);
981 if (src->w != ref->w ||
983 av_log(ctx, AV_LOG_ERROR, "First input link %s parameters "
984 "(size %dx%d) do not match the corresponding "
985 "second input link %s parameters (%dx%d) ",
986 ctx->input_pads[0].name, src->w, src->h,
987 ctx->input_pads[1].name, ref->w, ref->h);
988 return AVERROR(EINVAL);
994 outlink->time_base = src->time_base;
995 outlink->sample_aspect_ratio = src->sample_aspect_ratio;
996 outlink->frame_rate = src->frame_rate;
1001 if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
1005 in[0].time_base = src->time_base;
1006 in[1].time_base = ref->time_base;
1008 in[0].before = EXT_STOP;
1009 in[0].after = EXT_STOP;
1011 in[1].before = EXT_STOP;
1012 in[1].after = EXT_STOP;
1014 s->fs.on_event = process_frame;
1016 return ff_framesync_configure(&s->fs);
1019 static av_cold void uninit(AVFilterContext *ctx)
1021 BM3DContext *s = ctx->priv;
1024 for (i = 0; i < ctx->nb_inputs; i++)
1025 av_freep(&ctx->input_pads[i].name);
1028 ff_framesync_uninit(&s->fs);
1030 for (i = 0; i < s->nb_threads; i++) {
1031 SliceContext *sc = &s->slices[i];
1036 av_dct_end(sc->gdctf);
1037 av_dct_end(sc->gdcti);
1038 av_dct_end(sc->dctf);
1039 av_dct_end(sc->dcti);
1041 av_freep(&sc->buffer);
1042 av_freep(&sc->bufferh);
1043 av_freep(&sc->bufferv);
1044 av_freep(&sc->bufferz);
1045 av_freep(&sc->rbuffer);
1046 av_freep(&sc->rbufferh);
1047 av_freep(&sc->rbufferv);
1048 av_freep(&sc->rbufferz);
1050 av_freep(&sc->search_positions);
1054 static const AVFilterPad bm3d_outputs[] = {
1057 .type = AVMEDIA_TYPE_VIDEO,
1058 .config_props = config_output,
1063 AVFilter ff_vf_bm3d = {
1065 .description = NULL_IF_CONFIG_SMALL("Block-Matching 3D denoiser."),
1066 .priv_size = sizeof(BM3DContext),
1069 .activate = activate,
1070 .query_formats = query_formats,
1072 .outputs = bm3d_outputs,
1073 .priv_class = &bm3d_class,
1074 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
1075 AVFILTER_FLAG_DYNAMIC_INPUTS |
1076 AVFILTER_FLAG_SLICE_THREADS,