FFmpeg coverage

Line	Branch	Exec	Source
1			/*
2			* Copyright (c) 2015-2016 mawen1250
3			* Copyright (c) 2018 Paul B Mahol
4			*
5			* This file is part of FFmpeg.
6			*
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:
13			*
14			* The above copyright notice and this permission notice shall be included in all
15			* copies or substantial portions of the Software.
16			*
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
23			* SOFTWARE.
24			*/
25
26			/**
27			* @todo
28			* - opponent color space
29			* - temporal support
30			*/
31
32			#include <float.h>
33
34			#include "libavutil/cpu.h"
35			#include "libavutil/imgutils.h"
36			#include "libavutil/mem.h"
37			#include "libavutil/opt.h"
38			#include "libavutil/pixdesc.h"
39			#include "libavutil/tx.h"
40			#include "avfilter.h"
41			#include "filters.h"
42			#include "framesync.h"
43			#include "internal.h"
44			#include "video.h"
45
46			#define MAX_NB_THREADS 32
47
48			enum FilterModes {
49			BASIC,
50			FINAL,
51			NB_MODES,
52			};
53
54			typedef struct ThreadData {
55			const uint8_t *src;
56			int src_linesize;
57			const uint8_t *ref;
58			int ref_linesize;
59			int plane;
60			} ThreadData;
61
62			typedef struct PosCode {
63			int x, y;
64			} PosCode;
65
66			typedef struct PosPairCode {
67			double score;
68			int x, y;
69			} PosPairCode;
70
71			typedef struct SliceContext {
72			AVTXContext *gdctf, *gdcti;
73			av_tx_fn tx_fn_g, itx_fn_g;
74			AVTXContext *dctf, *dcti;
75			av_tx_fn tx_fn, itx_fn;
76			float *bufferh;
77			float *buffert;
78			float *bufferv;
79			float *bufferz;
80			float *buffer;
81			float *rbufferh;
82			float *rbufferv;
83			float *rbufferz;
84			float *rbuffer;
85			float *num, *den;
86			PosPairCode match_blocks[256];
87			int nb_match_blocks;
88			PosCode *search_positions;
89			} SliceContext;
90
91			typedef struct BM3DContext {
92			const AVClass *class;
93
94			float sigma;
95			int block_size;
96			int block_step;
97			int group_size;
98			int bm_range;
99			int bm_step;
100			float th_mse;
101			float hard_threshold;
102			int mode;
103			int ref;
104			int planes;
105
106			int depth;
107			int max;
108			int nb_planes;
109			int planewidth[4];
110			int planeheight[4];
111			int pblock_size;
112			int pgroup_size;
113
114			SliceContext slices[MAX_NB_THREADS];
115
116			FFFrameSync fs;
117			int nb_threads;
118
119			void (*get_block_row)(const uint8_t *srcp, int src_linesize,
120			int y, int x, int block_size, float *dst);
121			double (*do_block_ssd)(struct BM3DContext *s, PosCode *pos,
122			const uint8_t *src, int src_stride,
123			int r_y, int r_x);
124			void (*do_output)(struct BM3DContext *s, uint8_t *dst, int dst_linesize,
125			int plane, int nb_jobs);
126			void (*block_filtering)(struct BM3DContext *s,
127			const uint8_t *src, int src_linesize,
128			const uint8_t *ref, int ref_linesize,
129			int y, int x, int plane, int jobnr);
130			} BM3DContext;
131
132			#define OFFSET(x) offsetof(BM3DContext, x)
133			#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
134
135			static const AVOption bm3d_options[] = {
136			{ "sigma", "set denoising strength",
137			OFFSET(sigma), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 99999.9, FLAGS },
138			{ "block", "set size of local patch",
139			OFFSET(block_size), AV_OPT_TYPE_INT, {.i64=16}, 8, 64, FLAGS },
140			{ "bstep", "set sliding step for processing blocks",
141			OFFSET(block_step), AV_OPT_TYPE_INT, {.i64=4}, 1, 64, FLAGS },
142			{ "group", "set maximal number of similar blocks",
143			OFFSET(group_size), AV_OPT_TYPE_INT, {.i64=1}, 1, 256, FLAGS },
144			{ "range", "set block matching range",
145			OFFSET(bm_range), AV_OPT_TYPE_INT, {.i64=9}, 1, INT32_MAX, FLAGS },
146			{ "mstep", "set step for block matching",
147			OFFSET(bm_step), AV_OPT_TYPE_INT, {.i64=1}, 1, 64, FLAGS },
148			{ "thmse", "set threshold of mean square error for block matching",
149			OFFSET(th_mse), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, INT32_MAX, FLAGS },
150			{ "hdthr", "set hard threshold for 3D transfer domain",
151			OFFSET(hard_threshold), AV_OPT_TYPE_FLOAT, {.dbl=2.7}, 0, INT32_MAX, FLAGS },
152			{ "estim", "set filtering estimation mode",
153			OFFSET(mode), AV_OPT_TYPE_INT, {.i64=BASIC}, 0, NB_MODES-1, FLAGS, .unit = "mode" },
154			{ "basic", "basic estimate",
155			0, AV_OPT_TYPE_CONST, {.i64=BASIC}, 0, 0, FLAGS, .unit = "mode" },
156			{ "final", "final estimate",
157			0, AV_OPT_TYPE_CONST, {.i64=FINAL}, 0, 0, FLAGS, .unit = "mode" },
158			{ "ref", "have reference stream",
159			OFFSET(ref), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
160			{ "planes", "set planes to filter",
161			OFFSET(planes), AV_OPT_TYPE_INT, {.i64=7}, 0, 15, FLAGS },
162			{ NULL }
163			};
164
165			AVFILTER_DEFINE_CLASS(bm3d);
166
167			static const enum AVPixelFormat pix_fmts[] = {
168			AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
169			AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, 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,
175			AV_PIX_FMT_YUVJ411P,
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,
185			AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
186			AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
187			AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
188			AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
189			AV_PIX_FMT_GBRAP, AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
190			AV_PIX_FMT_NONE
191			};
192
193		✗	static int do_search_boundary(int pos, int plane_boundary, int search_range, int search_step)
194			{
195			int search_boundary;
196
197		✗	search_range = search_range / search_step * search_step;
198
199		✗	if (pos == plane_boundary) {
200		✗	search_boundary = plane_boundary;
201		✗	} else if (pos > plane_boundary) {
202		✗	search_boundary = pos - search_range;
203
204		✗	while (search_boundary < plane_boundary) {
205		✗	search_boundary += search_step;
206			}
207			} else {
208		✗	search_boundary = pos + search_range;
209
210		✗	while (search_boundary > plane_boundary) {
211		✗	search_boundary -= search_step;
212			}
213			}
214
215		✗	return search_boundary;
216			}
217
218		✗	static int search_boundary(int plane_boundary, int search_range, int search_step, int vertical, int y, int x)
219			{
220		✗	return do_search_boundary(vertical ? y : x, plane_boundary, search_range, search_step);
221			}
222
223		✗	static int cmp_scores(const void *a, const void *b)
224			{
225		✗	const struct PosPairCode *pair1 = a;
226		✗	const struct PosPairCode *pair2 = b;
227		✗	return FFDIFFSIGN(pair1->score, pair2->score);
228			}
229
230		✗	static double do_block_ssd(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
231			{
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;
235		✗	double dist = 0.;
236			int x, y;
237
238		✗	for (y = 0; y < block_size; y++) {
239		✗	for (x = 0; x < block_size; x++) {
240		✗	double temp = refp[x] - srcp[x];
241		✗	dist += temp * temp;
242			}
243
244		✗	srcp += src_stride;
245		✗	refp += src_stride;
246			}
247
248		✗	return dist;
249			}
250
251		✗	static double do_block_ssd16(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
252			{
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;
256		✗	double dist = 0.;
257			int x, y;
258
259		✗	for (y = 0; y < block_size; y++) {
260		✗	for (x = 0; x < block_size; x++) {
261		✗	double temp = refp[x] - srcp[x];
262		✗	dist += temp * temp;
263			}
264
265		✗	srcp += src_stride / 2;
266		✗	refp += src_stride / 2;
267			}
268
269		✗	return dist;
270			}
271
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)
275			{
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 index = sc->nb_match_blocks;
281
282		✗	for (int i = 0; i < search_size; i++) {
283		✗	PosCode pos = search_pos[i];
284			double dist;
285
286		✗	dist = s->do_block_ssd(s, &pos, src, src_stride, r_y, r_x);
287
288			// Only match similar blocks but not identical blocks
289		✗	if (dist <= th_sse && dist != 0) {
290		✗	const double score = dist * distMul;
291
292		✗	if (index >= s->group_size && score >= sc->match_blocks[index - 1].score) {
293		✗	continue;
294			}
295
296		✗	if (index >= s->group_size)
297		✗	index = s->group_size - 1;
298
299		✗	sc->match_blocks[index].score = score;
300		✗	sc->match_blocks[index].y = pos.y;
301		✗	sc->match_blocks[index].x = pos.x;
302		✗	index++;
303		✗	qsort(sc->match_blocks, index, sizeof(PosPairCode), cmp_scores);
304			}
305			}
306
307		✗	sc->nb_match_blocks = index;
308		✗	}
309
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)
312			{
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);
323		✗	int index = 0;
324
325		✗	for (int j = t; j <= b; j += step) {
326		✗	for (int i = l; i <= r; i += step) {
327			PosCode pos;
328
329		✗	if (exclude_cur_pos > 0 && j == y && i == x) {
330		✗	continue;
331			}
332
333		✗	pos.y = j;
334		✗	pos.x = i;
335		✗	sc->search_positions[index++] = pos;
336			}
337			}
338
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;
344			}
345
346		✗	do_block_matching_multi(s, ref, ref_linesize, s->bm_range,
347		✗	sc->search_positions, index, s->th_mse, y, x, plane, jobnr);
348		✗	}
349
350		✗	static void block_matching(BM3DContext *s, const uint8_t *ref, int ref_linesize,
351			int j, int i, int plane, int jobnr)
352			{
353		✗	SliceContext *sc = &s->slices[jobnr];
354
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;
360		✗	return;
361			}
362
363		✗	sc->nb_match_blocks = 0;
364		✗	block_matching_multi(s, ref, ref_linesize, j, i, 1, plane, jobnr);
365			}
366
367		✗	static void get_block_row(const uint8_t *srcp, int src_linesize,
368			int y, int x, int block_size, float *dst)
369			{
370		✗	const uint8_t *src = srcp + y * src_linesize + x;
371
372		✗	for (int j = 0; j < block_size; j++)
373		✗	dst[j] = src[j];
374		✗	}
375
376		✗	static void get_block_row16(const uint8_t *srcp, int src_linesize,
377			int y, int x, int block_size, float *dst)
378			{
379		✗	const uint16_t *src = (uint16_t *)srcp + y * src_linesize / 2 + x;
380
381		✗	for (int j = 0; j < block_size; j++)
382		✗	dst[j] = src[j];
383		✗	}
384
385		✗	static void basic_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
386			const uint8_t *ref, int ref_linesize,
387			int y, int x, int plane, int jobnr)
388			{
389		✗	SliceContext *sc = &s->slices[jobnr];
390		✗	const int pblock_size = s->pblock_size;
391		✗	const int buffer_linesize = s->pblock_size * s->pblock_size;
392		✗	const int nb_match_blocks = sc->nb_match_blocks;
393		✗	const int block_size = s->block_size;
394		✗	const int width = s->planewidth[plane];
395		✗	const int pgroup_size = s->pgroup_size;
396		✗	const int group_size = s->group_size;
397		✗	float *buffer = sc->buffer;
398		✗	float *bufferh = sc->bufferh;
399		✗	float *buffert = sc->buffert;
400		✗	float *bufferv = sc->bufferv;
401		✗	float *bufferz = sc->bufferz;
402			float threshold[4];
403			float den_weight, num_weight;
404		✗	int retained = 0;
405
406		✗	for (int k = 0; k < nb_match_blocks; k++) {
407		✗	const int y = sc->match_blocks[k].y;
408		✗	const int x = sc->match_blocks[k].x;
409
410		✗	for (int i = 0; i < block_size; i++) {
411		✗	s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + pblock_size * i);
412		✗	sc->tx_fn(sc->dctf, buffert, bufferh + pblock_size * i, sizeof(float));
413		✗	for (int j = 0; j < block_size; j++)
414		✗	bufferv[j * pblock_size + i] = buffert[j];
415			}
416
417		✗	for (int i = 0; i < block_size; i++) {
418		✗	sc->tx_fn(sc->dctf, buffert, bufferv + i * pblock_size, sizeof(float));
419		✗	memcpy(buffer + k * buffer_linesize + i * pblock_size,
420			buffert, block_size * sizeof(float));
421			}
422			}
423
424		✗	for (int i = 0; i < block_size; i++) {
425		✗	for (int j = 0; j < block_size; j++) {
426		✗	for (int k = 0; k < nb_match_blocks; k++)
427		✗	bufferz[k] = buffer[buffer_linesize * k + i * pblock_size + j];
428		✗	if (group_size > 1)
429		✗	sc->tx_fn_g(sc->gdctf, bufferz, bufferz, sizeof(float));
430		✗	bufferz += pgroup_size;
431			}
432			}
433
434		✗	threshold[0] = s->hard_threshold * s->sigma * M_SQRT2 * 4.f * block_size * block_size * (1 << (s->depth - 8)) / 255.f;
435		✗	threshold[1] = threshold[0] * sqrtf(2.f);
436		✗	threshold[2] = threshold[0] * 2.f;
437		✗	threshold[3] = threshold[0] * sqrtf(8.f);
438		✗	bufferz = sc->bufferz;
439
440		✗	for (int i = 0; i < block_size; i++) {
441		✗	for (int j = 0; j < block_size; j++) {
442		✗	for (int k = 0; k < nb_match_blocks; k++) {
443		✗	const float thresh = threshold[(j == 0) + (i == 0) + (k == 0)];
444
445		✗	if (bufferz[k] > thresh || bufferz[k] < -thresh) {
446		✗	retained++;
447			} else {
448		✗	bufferz[k] = 0;
449			}
450			}
451		✗	bufferz += pgroup_size;
452			}
453			}
454
455		✗	bufferz = sc->bufferz;
456		✗	buffer = sc->buffer;
457		✗	for (int i = 0; i < block_size; i++) {
458		✗	for (int j = 0; j < block_size; j++) {
459		✗	if (group_size > 1)
460		✗	sc->itx_fn_g(sc->gdcti, bufferz, bufferz, sizeof(float));
461		✗	for (int k = 0; k < nb_match_blocks; k++)
462		✗	buffer[buffer_linesize * k + i * pblock_size + j] = bufferz[k];
463		✗	bufferz += pgroup_size;
464			}
465			}
466
467		✗	den_weight = retained < 1 ? 1.f : 1.f / retained;
468		✗	num_weight = den_weight;
469
470		✗	buffer = sc->buffer;
471		✗	for (int k = 0; k < nb_match_blocks; k++) {
472		✗	float *num = sc->num + y * width + x;
473		✗	float *den = sc->den + y * width + x;
474
475		✗	for (int i = 0; i < block_size; i++) {
476		✗	memcpy(bufferv + i * pblock_size,
477		✗	buffer + k * buffer_linesize + i * pblock_size,
478			block_size * sizeof(float));
479			}
480
481		✗	for (int i = 0; i < block_size; i++) {
482		✗	sc->itx_fn(sc->dcti, buffert, bufferv + i * pblock_size, sizeof(float));
483		✗	for (int j = 0; j < block_size; j++)
484		✗	bufferh[j * pblock_size + i] = buffert[j];
485			}
486
487		✗	for (int i = 0; i < block_size; i++) {
488		✗	sc->itx_fn(sc->dcti, buffert, bufferh + pblock_size * i, sizeof(float));
489		✗	for (int j = 0; j < block_size; j++) {
490		✗	num[j] += buffert[j] * num_weight;
491		✗	den[j] += den_weight;
492			}
493		✗	num += width;
494		✗	den += width;
495			}
496			}
497		✗	}
498
499		✗	static void final_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
500			const uint8_t *ref, int ref_linesize,
501			int y, int x, int plane, int jobnr)
502			{
503		✗	SliceContext *sc = &s->slices[jobnr];
504		✗	const int pblock_size = s->pblock_size;
505		✗	const int buffer_linesize = s->pblock_size * s->pblock_size;
506		✗	const int nb_match_blocks = sc->nb_match_blocks;
507		✗	const int block_size = s->block_size;
508		✗	const int width = s->planewidth[plane];
509		✗	const int pgroup_size = s->pgroup_size;
510		✗	const int group_size = s->group_size;
511		✗	const float sigma_sqr = s->sigma * s->sigma;
512		✗	float *buffer = sc->buffer;
513		✗	float *bufferh = sc->bufferh;
514		✗	float *bufferv = sc->bufferv;
515		✗	float *bufferz = sc->bufferz;
516		✗	float *rbuffer = sc->rbuffer;
517		✗	float *rbufferh = sc->rbufferh;
518		✗	float *rbufferv = sc->rbufferv;
519		✗	float *rbufferz = sc->rbufferz;
520			float den_weight, num_weight;
521		✗	float l2_wiener = 0;
522
523		✗	for (int k = 0; k < nb_match_blocks; k++) {
524		✗	const int y = sc->match_blocks[k].y;
525		✗	const int x = sc->match_blocks[k].x;
526
527		✗	for (int i = 0; i < block_size; i++) {
528		✗	s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + pblock_size * i);
529		✗	s->get_block_row(ref, ref_linesize, y + i, x, block_size, rbufferh + pblock_size * i);
530		✗	sc->tx_fn(sc->dctf, bufferh + pblock_size * i, bufferh + pblock_size * i, sizeof(float));
531		✗	sc->tx_fn(sc->dctf, rbufferh + pblock_size * i, rbufferh + pblock_size * i, sizeof(float));
532			}
533
534		✗	for (int i = 0; i < block_size; i++) {
535		✗	for (int j = 0; j < block_size; j++) {
536		✗	bufferv[i * pblock_size + j] = bufferh[j * pblock_size + i];
537		✗	rbufferv[i * pblock_size + j] = rbufferh[j * pblock_size + i];
538			}
539		✗	sc->tx_fn(sc->dctf, bufferv + i * pblock_size, bufferv + i * pblock_size, sizeof(float));
540		✗	sc->tx_fn(sc->dctf, rbufferv + i * pblock_size, rbufferv + i * pblock_size, sizeof(float));
541			}
542
543		✗	for (int i = 0; i < block_size; i++) {
544		✗	memcpy(buffer + k * buffer_linesize + i * pblock_size,
545		✗	bufferv + i * pblock_size, block_size * sizeof(float));
546		✗	memcpy(rbuffer + k * buffer_linesize + i * pblock_size,
547		✗	rbufferv + i * pblock_size, block_size * sizeof(float));
548			}
549			}
550
551		✗	for (int i = 0; i < block_size; i++) {
552		✗	for (int j = 0; j < block_size; j++) {
553		✗	for (int k = 0; k < nb_match_blocks; k++) {
554		✗	bufferz[k] = buffer[buffer_linesize * k + i * pblock_size + j];
555		✗	rbufferz[k] = rbuffer[buffer_linesize * k + i * pblock_size + j];
556			}
557		✗	if (group_size > 1) {
558		✗	sc->tx_fn_g(sc->gdctf, bufferz, bufferz, sizeof(float));
559		✗	sc->tx_fn_g(sc->gdctf, rbufferz, rbufferz, sizeof(float));
560			}
561		✗	bufferz += pgroup_size;
562		✗	rbufferz += pgroup_size;
563			}
564			}
565
566		✗	bufferz = sc->bufferz;
567		✗	rbufferz = sc->rbufferz;
568
569		✗	for (int i = 0; i < block_size; i++) {
570		✗	for (int j = 0; j < block_size; j++) {
571		✗	for (int k = 0; k < nb_match_blocks; k++) {
572		✗	const float ref_sqr = rbufferz[k] * rbufferz[k];
573		✗	float wiener_coef = ref_sqr / (ref_sqr + sigma_sqr);
574
575		✗	if (isnan(wiener_coef))
576		✗	wiener_coef = 1;
577		✗	bufferz[k] *= wiener_coef;
578		✗	l2_wiener += wiener_coef * wiener_coef;
579			}
580		✗	bufferz += pgroup_size;
581		✗	rbufferz += pgroup_size;
582			}
583			}
584
585		✗	bufferz = sc->bufferz;
586		✗	buffer = sc->buffer;
587		✗	for (int i = 0; i < block_size; i++) {
588		✗	for (int j = 0; j < block_size; j++) {
589		✗	if (group_size > 1)
590		✗	sc->itx_fn_g(sc->gdcti, bufferz, bufferz, sizeof(float));
591		✗	for (int k = 0; k < nb_match_blocks; k++) {
592		✗	buffer[buffer_linesize * k + i * pblock_size + j] = bufferz[k];
593			}
594		✗	bufferz += pgroup_size;
595			}
596			}
597
598		✗	l2_wiener = FFMAX(l2_wiener, 1e-15f);
599		✗	den_weight = 1.f / l2_wiener;
600		✗	num_weight = den_weight;
601
602		✗	for (int k = 0; k < nb_match_blocks; k++) {
603		✗	float *num = sc->num + y * width + x;
604		✗	float *den = sc->den + y * width + x;
605
606		✗	for (int i = 0; i < block_size; i++) {
607		✗	memcpy(bufferv + i * pblock_size,
608		✗	buffer + k * buffer_linesize + i * pblock_size,
609			block_size * sizeof(float));
610			}
611
612		✗	for (int i = 0; i < block_size; i++) {
613		✗	sc->itx_fn(sc->dcti, bufferv + pblock_size * i, bufferv + pblock_size * i, sizeof(float));
614		✗	for (int j = 0; j < block_size; j++) {
615		✗	bufferh[j * pblock_size + i] = bufferv[i * pblock_size + j];
616			}
617			}
618
619		✗	for (int i = 0; i < block_size; i++) {
620		✗	sc->itx_fn(sc->dcti, bufferh + pblock_size * i, bufferh + pblock_size * i, sizeof(float));
621		✗	for (int j = 0; j < block_size; j++) {
622		✗	num[j] += bufferh[i * pblock_size + j] * num_weight;
623		✗	den[j] += den_weight;
624			}
625		✗	num += width;
626		✗	den += width;
627			}
628			}
629		✗	}
630
631		✗	static void do_output(BM3DContext *s, uint8_t *dst, int dst_linesize,
632			int plane, int nb_jobs)
633			{
634		✗	const int height = s->planeheight[plane];
635		✗	const int width = s->planewidth[plane];
636
637		✗	for (int i = 0; i < height; i++) {
638		✗	for (int j = 0; j < width; j++) {
639		✗	uint8_t *dstp = dst + i * dst_linesize;
640		✗	float sum_den = 0.f;
641		✗	float sum_num = 0.f;
642
643		✗	for (int k = 0; k < nb_jobs; k++) {
644		✗	SliceContext *sc = &s->slices[k];
645		✗	float num = sc->num[i * width + j];
646		✗	float den = sc->den[i * width + j];
647
648		✗	sum_num += num;
649		✗	sum_den += den;
650			}
651
652		✗	dstp[j] = av_clip_uint8(lrintf(sum_num / sum_den));
653			}
654			}
655		✗	}
656
657		✗	static void do_output16(BM3DContext *s, uint8_t *dst, int dst_linesize,
658			int plane, int nb_jobs)
659			{
660		✗	const int height = s->planeheight[plane];
661		✗	const int width = s->planewidth[plane];
662		✗	const int depth = s->depth;
663
664		✗	for (int i = 0; i < height; i++) {
665		✗	for (int j = 0; j < width; j++) {
666		✗	uint16_t *dstp = (uint16_t *)dst + i * dst_linesize / 2;
667		✗	float sum_den = 0.f;
668		✗	float sum_num = 0.f;
669
670		✗	for (int k = 0; k < nb_jobs; k++) {
671		✗	SliceContext *sc = &s->slices[k];
672		✗	float num = sc->num[i * width + j];
673		✗	float den = sc->den[i * width + j];
674
675		✗	sum_num += num;
676		✗	sum_den += den;
677			}
678
679		✗	dstp[j] = av_clip_uintp2_c(lrintf(sum_num / sum_den), depth);
680			}
681			}
682		✗	}
683
684		✗	static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
685			{
686		✗	BM3DContext *s = ctx->priv;
687		✗	SliceContext *sc = &s->slices[jobnr];
688		✗	const int block_step = s->block_step;
689		✗	ThreadData *td = arg;
690		✗	const uint8_t *src = td->src;
691		✗	const uint8_t *ref = td->ref;
692		✗	const int src_linesize = td->src_linesize;
693		✗	const int ref_linesize = td->ref_linesize;
694		✗	const int plane = td->plane;
695		✗	const int width = s->planewidth[plane];
696		✗	const int height = s->planeheight[plane];
697		✗	const int block_pos_bottom = FFMAX(0, height - s->block_size);
698		✗	const int block_pos_right = FFMAX(0, width - s->block_size);
699		✗	const int slice_start = (((height + block_step - 1) / block_step) * jobnr / nb_jobs) * block_step;
700		✗	const int slice_end = (jobnr == nb_jobs - 1) ? block_pos_bottom + block_step :
701		✗	(((height + block_step - 1) / block_step) * (jobnr + 1) / nb_jobs) * block_step;
702
703		✗	memset(sc->num, 0, width * height * sizeof(float));
704		✗	memset(sc->den, 0, width * height * sizeof(float));
705
706		✗	for (int j = slice_start; j < slice_end; j += block_step) {
707		✗	if (j > block_pos_bottom) {
708		✗	j = block_pos_bottom;
709			}
710
711		✗	for (int i = 0; i < block_pos_right + block_step; i += block_step) {
712		✗	if (i > block_pos_right) {
713		✗	i = block_pos_right;
714			}
715
716		✗	block_matching(s, ref, ref_linesize, j, i, plane, jobnr);
717
718		✗	s->block_filtering(s, src, src_linesize,
719			ref, ref_linesize, j, i, plane, jobnr);
720			}
721			}
722
723		✗	return 0;
724			}
725
726		✗	static int filter_frame(AVFilterContext *ctx, AVFrame **out, AVFrame *in, AVFrame *ref)
727			{
728		✗	BM3DContext *s = ctx->priv;
729		✗	AVFilterLink *outlink = ctx->outputs[0];
730			int p;
731
732		✗	*out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
733		✗	if (!*out)
734		✗	return AVERROR(ENOMEM);
735		✗	av_frame_copy_props(*out, in);
736
737		✗	for (p = 0; p < s->nb_planes; p++) {
738		✗	const int nb_jobs = FFMAX(1, FFMIN(s->nb_threads, s->planeheight[p] / s->block_size));
739			ThreadData td;
740
741		✗	if (!((1 << p) & s->planes) || ctx->is_disabled) {
742		✗	av_image_copy_plane((*out)->data[p], (*out)->linesize[p],
743		✗	in->data[p], in->linesize[p],
744		✗	s->planewidth[p] * (1 + (s->depth > 8)), s->planeheight[p]);
745		✗	continue;
746			}
747
748		✗	td.src = in->data[p];
749		✗	td.src_linesize = in->linesize[p];
750		✗	td.ref = ref->data[p];
751		✗	td.ref_linesize = ref->linesize[p];
752		✗	td.plane = p;
753		✗	ff_filter_execute(ctx, filter_slice, &td, NULL, nb_jobs);
754
755		✗	s->do_output(s, (*out)->data[p], (*out)->linesize[p], p, nb_jobs);
756			}
757
758		✗	return 0;
759			}
760
761			#define SQR(x) ((x) * (x))
762
763		✗	static int config_input(AVFilterLink *inlink)
764			{
765		✗	const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
766		✗	AVFilterContext *ctx = inlink->dst;
767		✗	BM3DContext *s = ctx->priv;
768
769		✗	s->nb_threads = FFMIN(ff_filter_get_nb_threads(ctx), MAX_NB_THREADS);
770		✗	s->nb_planes = av_pix_fmt_count_planes(inlink->format);
771		✗	s->depth = desc->comp[0].depth;
772		✗	s->max = (1 << s->depth) - 1;
773		✗	s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
774		✗	s->planeheight[0] = s->planeheight[3] = inlink->h;
775		✗	s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
776		✗	s->planewidth[0] = s->planewidth[3] = inlink->w;
777		✗	s->pblock_size = FFALIGN(s->block_size * 2, av_cpu_max_align());
778		✗	s->pgroup_size = FFALIGN(s->group_size * 2, av_cpu_max_align());
779
780		✗	for (int i = 0; i < s->nb_threads; i++) {
781		✗	SliceContext *sc = &s->slices[i];
782		✗	float iscale = 0.5f / s->block_size;
783		✗	float scale = 1.f;
784			int ret;
785
786		✗	sc->num = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(float));
787		✗	sc->den = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(float));
788		✗	if (!sc->num || !sc->den)
789		✗	return AVERROR(ENOMEM);
790
791		✗	ret = av_tx_init(&sc->dctf, &sc->tx_fn, AV_TX_FLOAT_DCT, 0, s->block_size >> 0, &scale, 0);
792		✗	if (ret < 0)
793		✗	return ret;
794
795		✗	ret = av_tx_init(&sc->dcti, &sc->itx_fn, AV_TX_FLOAT_DCT, 1, s->block_size >> 1, &iscale, 0);
796		✗	if (ret < 0)
797		✗	return ret;
798
799		✗	if (s->group_size > 1) {
800		✗	float iscale = 0.5f / s->group_size;
801
802		✗	ret = av_tx_init(&sc->gdctf, &sc->tx_fn_g, AV_TX_FLOAT_DCT, 0, s->group_size >> 0, &scale, 0);
803		✗	if (ret < 0)
804		✗	return ret;
805
806		✗	ret = av_tx_init(&sc->gdcti, &sc->itx_fn_g, AV_TX_FLOAT_DCT, 1, s->group_size >> 1, &iscale, 0);
807		✗	if (ret < 0)
808		✗	return ret;
809			}
810
811		✗	sc->buffer = av_calloc(s->pblock_size * s->pblock_size * s->pgroup_size, sizeof(*sc->buffer));
812		✗	sc->bufferz = av_calloc(s->pblock_size * s->pblock_size * s->pgroup_size, sizeof(*sc->bufferz));
813		✗	sc->bufferh = av_calloc(s->pblock_size * s->pblock_size, sizeof(*sc->bufferh));
814		✗	sc->bufferv = av_calloc(s->pblock_size * s->pblock_size, sizeof(*sc->bufferv));
815		✗	sc->buffert = av_calloc(s->pblock_size, sizeof(*sc->buffert));
816		✗	if (!sc->bufferh || !sc->bufferv || !sc->buffer || !sc->bufferz || !sc->buffert)
817		✗	return AVERROR(ENOMEM);
818
819		✗	if (s->mode == FINAL) {
820		✗	sc->rbuffer = av_calloc(s->pblock_size * s->pblock_size * s->pgroup_size, sizeof(*sc->rbuffer));
821		✗	sc->rbufferz = av_calloc(s->pblock_size * s->pblock_size * s->pgroup_size, sizeof(*sc->rbufferz));
822		✗	sc->rbufferh = av_calloc(s->pblock_size * s->pblock_size, sizeof(*sc->rbufferh));
823		✗	sc->rbufferv = av_calloc(s->pblock_size * s->pblock_size, sizeof(*sc->rbufferv));
824		✗	if (!sc->rbufferh || !sc->rbufferv || !sc->rbuffer || !sc->rbufferz)
825		✗	return AVERROR(ENOMEM);
826			}
827
828		✗	sc->search_positions = av_calloc(SQR(2 * s->bm_range / s->bm_step + 1), sizeof(*sc->search_positions));
829		✗	if (!sc->search_positions)
830		✗	return AVERROR(ENOMEM);
831			}
832
833		✗	s->do_output = do_output;
834		✗	s->do_block_ssd = do_block_ssd;
835		✗	s->get_block_row = get_block_row;
836
837		✗	if (s->depth > 8) {
838		✗	s->do_output = do_output16;
839		✗	s->do_block_ssd = do_block_ssd16;
840		✗	s->get_block_row = get_block_row16;
841			}
842
843		✗	return 0;
844			}
845
846		✗	static int activate(AVFilterContext *ctx)
847			{
848		✗	BM3DContext *s = ctx->priv;
849
850		✗	if (!s->ref) {
851		✗	AVFrame *frame = NULL;
852		✗	AVFrame *out = NULL;
853			int ret, status;
854			int64_t pts;
855
856		✗	FF_FILTER_FORWARD_STATUS_BACK(ctx->outputs[0], ctx->inputs[0]);
857
858		✗	if ((ret = ff_inlink_consume_frame(ctx->inputs[0], &frame)) > 0) {
859		✗	ret = filter_frame(ctx, &out, frame, frame);
860		✗	av_frame_free(&frame);
861		✗	if (ret < 0)
862		✗	return ret;
863		✗	ret = ff_filter_frame(ctx->outputs[0], out);
864			}
865		✗	if (ret < 0) {
866		✗	return ret;
867		✗	} else if (ff_inlink_acknowledge_status(ctx->inputs[0], &status, &pts)) {
868		✗	ff_outlink_set_status(ctx->outputs[0], status, pts);
869		✗	return 0;
870			} else {
871		✗	if (ff_outlink_frame_wanted(ctx->outputs[0]))
872		✗	ff_inlink_request_frame(ctx->inputs[0]);
873		✗	return 0;
874			}
875			} else {
876		✗	return ff_framesync_activate(&s->fs);
877			}
878			}
879
880		✗	static int process_frame(FFFrameSync *fs)
881			{
882		✗	AVFilterContext *ctx = fs->parent;
883		✗	BM3DContext *s = fs->opaque;
884		✗	AVFilterLink *outlink = ctx->outputs[0];
885		✗	AVFrame *out = NULL, *src, *ref;
886			int ret;
887
888		✗	if ((ret = ff_framesync_get_frame(&s->fs, 0, &src, 0)) < 0 ||
889		✗	(ret = ff_framesync_get_frame(&s->fs, 1, &ref, 0)) < 0)
890		✗	return ret;
891
892		✗	if ((ret = filter_frame(ctx, &out, src, ref)) < 0)
893		✗	return ret;
894
895		✗	out->pts = av_rescale_q(src->pts, s->fs.time_base, outlink->time_base);
896
897		✗	return ff_filter_frame(outlink, out);
898			}
899
900		✗	static av_cold int init(AVFilterContext *ctx)
901			{
902		✗	BM3DContext *s = ctx->priv;
903		✗	AVFilterPad pad = { 0 };
904			int ret;
905
906		✗	if (s->mode == BASIC) {
907		✗	if (s->th_mse == 0.f)
908		✗	s->th_mse = 400.f + s->sigma * 80.f;
909		✗	s->block_filtering = basic_block_filtering;
910		✗	} else if (s->mode == FINAL) {
911		✗	if (!s->ref) {
912		✗	av_log(ctx, AV_LOG_WARNING, "Reference stream is mandatory in final estimation mode.\n");
913		✗	s->ref = 1;
914			}
915		✗	if (s->th_mse == 0.f)
916		✗	s->th_mse = 200.f + s->sigma * 10.f;
917
918		✗	s->block_filtering = final_block_filtering;
919			} else {
920		✗	return AVERROR_BUG;
921			}
922
923		✗	if (s->block_step > s->block_size) {
924		✗	av_log(ctx, AV_LOG_WARNING, "bstep: %d can't be bigger than block size. Changing to %d.\n",
925			s->block_step, s->block_size);
926		✗	s->block_step = s->block_size;
927			}
928
929		✗	if (s->bm_step > s->bm_range) {
930		✗	av_log(ctx, AV_LOG_WARNING, "mstep: %d can't be bigger than block matching range. Changing to %d.\n",
931			s->bm_step, s->bm_range);
932		✗	s->bm_step = s->bm_range;
933			}
934
935		✗	pad.type = AVMEDIA_TYPE_VIDEO;
936		✗	pad.name = "source";
937		✗	pad.config_props = config_input;
938
939		✗	if ((ret = ff_append_inpad(ctx, &pad)) < 0)
940		✗	return ret;
941
942		✗	if (s->ref) {
943		✗	pad.type = AVMEDIA_TYPE_VIDEO;
944		✗	pad.name = "reference";
945		✗	pad.config_props = NULL;
946
947		✗	if ((ret = ff_append_inpad(ctx, &pad)) < 0)
948		✗	return ret;
949			}
950
951		✗	return 0;
952			}
953
954		✗	static int config_output(AVFilterLink *outlink)
955			{
956		✗	AVFilterContext *ctx = outlink->src;
957		✗	BM3DContext *s = ctx->priv;
958		✗	AVFilterLink *src = ctx->inputs[0];
959			AVFilterLink *ref;
960			FFFrameSyncIn *in;
961			int ret;
962
963		✗	if (s->ref) {
964		✗	ref = ctx->inputs[1];
965
966		✗	if (src->w != ref->w ||
967		✗	src->h != ref->h) {
968		✗	av_log(ctx, AV_LOG_ERROR, "First input link %s parameters "
969			"(size %dx%d) do not match the corresponding "
970			"second input link %s parameters (%dx%d) ",
971		✗	ctx->input_pads[0].name, src->w, src->h,
972		✗	ctx->input_pads[1].name, ref->w, ref->h);
973		✗	return AVERROR(EINVAL);
974			}
975			}
976
977		✗	outlink->w = src->w;
978		✗	outlink->h = src->h;
979		✗	outlink->time_base = src->time_base;
980		✗	outlink->sample_aspect_ratio = src->sample_aspect_ratio;
981		✗	outlink->frame_rate = src->frame_rate;
982
983		✗	if (!s->ref)
984		✗	return 0;
985
986		✗	if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
987		✗	return ret;
988
989		✗	in = s->fs.in;
990		✗	in[0].time_base = src->time_base;
991		✗	in[1].time_base = ref->time_base;
992		✗	in[0].sync = 1;
993		✗	in[0].before = EXT_STOP;
994		✗	in[0].after = EXT_STOP;
995		✗	in[1].sync = 1;
996		✗	in[1].before = EXT_STOP;
997		✗	in[1].after = EXT_STOP;
998		✗	s->fs.opaque = s;
999		✗	s->fs.on_event = process_frame;
1000
1001		✗	return ff_framesync_configure(&s->fs);
1002			}
1003
1004		✗	static av_cold void uninit(AVFilterContext *ctx)
1005			{
1006		✗	BM3DContext *s = ctx->priv;
1007
1008		✗	if (s->ref)
1009		✗	ff_framesync_uninit(&s->fs);
1010
1011		✗	for (int i = 0; i < s->nb_threads; i++) {
1012		✗	SliceContext *sc = &s->slices[i];
1013
1014		✗	av_freep(&sc->num);
1015		✗	av_freep(&sc->den);
1016
1017		✗	av_tx_uninit(&sc->gdctf);
1018		✗	av_tx_uninit(&sc->gdcti);
1019		✗	av_tx_uninit(&sc->dctf);
1020		✗	av_tx_uninit(&sc->dcti);
1021
1022		✗	av_freep(&sc->buffer);
1023		✗	av_freep(&sc->bufferh);
1024		✗	av_freep(&sc->buffert);
1025		✗	av_freep(&sc->bufferv);
1026		✗	av_freep(&sc->bufferz);
1027		✗	av_freep(&sc->rbuffer);
1028		✗	av_freep(&sc->rbufferh);
1029		✗	av_freep(&sc->rbufferv);
1030		✗	av_freep(&sc->rbufferz);
1031
1032		✗	av_freep(&sc->search_positions);
1033			}
1034		✗	}
1035
1036			static const AVFilterPad bm3d_outputs[] = {
1037			{
1038			.name = "default",
1039			.type = AVMEDIA_TYPE_VIDEO,
1040			.config_props = config_output,
1041			},
1042			};
1043
1044			const AVFilter ff_vf_bm3d = {
1045			.name = "bm3d",
1046			.description = NULL_IF_CONFIG_SMALL("Block-Matching 3D denoiser."),
1047			.priv_size = sizeof(BM3DContext),
1048			.init = init,
1049			.uninit = uninit,
1050			.activate = activate,
1051			.inputs = NULL,
1052			FILTER_OUTPUTS(bm3d_outputs),
1053			FILTER_PIXFMTS_ARRAY(pix_fmts),
1054			.priv_class = &bm3d_class,
1055			.flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
1056			AVFILTER_FLAG_DYNAMIC_INPUTS |
1057			AVFILTER_FLAG_SLICE_THREADS,
1058			};
1059