GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavfilter/vf_bm3d.c Lines: 0 536 0.0 %
Date: 2020-09-25 23:16:12 Branches: 0 278 0.0 %

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
 * - non-power of 2 DCT
29
 * - opponent color space
30
 * - temporal support
31
 */
32
33
#include <float.h>
34
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"
40
#include "avfilter.h"
41
#include "filters.h"
42
#include "formats.h"
43
#include "framesync.h"
44
#include "internal.h"
45
#include "video.h"
46
47
#define MAX_NB_THREADS 32
48
49
enum FilterModes {
50
    BASIC,
51
    FINAL,
52
    NB_MODES,
53
};
54
55
typedef struct ThreadData {
56
    const uint8_t *src;
57
    int src_linesize;
58
    const uint8_t *ref;
59
    int ref_linesize;
60
    int plane;
61
} ThreadData;
62
63
typedef struct PosCode {
64
    int x, y;
65
} PosCode;
66
67
typedef struct PosPairCode {
68
    double score;
69
    int x, y;
70
} PosPairCode;
71
72
typedef struct SliceContext {
73
    DCTContext *gdctf, *gdcti;
74
    DCTContext *dctf, *dcti;
75
    FFTSample *bufferh;
76
    FFTSample *bufferv;
77
    FFTSample *bufferz;
78
    FFTSample *buffer;
79
    FFTSample *rbufferh;
80
    FFTSample *rbufferv;
81
    FFTSample *rbufferz;
82
    FFTSample *rbuffer;
83
    float *num, *den;
84
    PosPairCode match_blocks[256];
85
    int nb_match_blocks;
86
    PosCode *search_positions;
87
} SliceContext;
88
89
typedef struct BM3DContext {
90
    const AVClass *class;
91
92
    float sigma;
93
    int block_size;
94
    int block_step;
95
    int group_size;
96
    int bm_range;
97
    int bm_step;
98
    float th_mse;
99
    float hard_threshold;
100
    int mode;
101
    int ref;
102
    int planes;
103
104
    int depth;
105
    int max;
106
    int nb_planes;
107
    int planewidth[4];
108
    int planeheight[4];
109
    int group_bits;
110
    int pgroup_size;
111
112
    SliceContext slices[MAX_NB_THREADS];
113
114
    FFFrameSync fs;
115
    int nb_threads;
116
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,
121
                           int r_y, int r_x);
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);
128
} BM3DContext;
129
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_BOOL,  {.i64=0},     0,            1, FLAGS },
157
    { "planes", "set planes to filter",
158
        OFFSET(planes),         AV_OPT_TYPE_INT,   {.i64=7},     0,           15, FLAGS },
159
    { NULL }
160
};
161
162
AVFILTER_DEFINE_CLASS(bm3d);
163
164
static int query_formats(AVFilterContext *ctx)
165
{
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,
174
        AV_PIX_FMT_YUVJ411P,
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,
184
        AV_PIX_FMT_YUVA420P,  AV_PIX_FMT_YUVA422P,   AV_PIX_FMT_YUVA444P,
185
        AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
186
        AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
187
        AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
188
        AV_PIX_FMT_GBRAP,     AV_PIX_FMT_GBRAP10,    AV_PIX_FMT_GBRAP12,    AV_PIX_FMT_GBRAP16,
189
        AV_PIX_FMT_NONE
190
    };
191
192
    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
193
    if (!fmts_list)
194
        return AVERROR(ENOMEM);
195
    return ff_set_common_formats(ctx, fmts_list);
196
}
197
198
static int do_search_boundary(int pos, int plane_boundary, int search_range, int search_step)
199
{
200
    int search_boundary;
201
202
    search_range = search_range / search_step * search_step;
203
204
    if (pos == plane_boundary) {
205
        search_boundary = plane_boundary;
206
    } else if (pos > plane_boundary) {
207
        search_boundary = pos - search_range;
208
209
        while (search_boundary < plane_boundary) {
210
            search_boundary += search_step;
211
        }
212
    } else {
213
        search_boundary = pos + search_range;
214
215
        while (search_boundary > plane_boundary) {
216
            search_boundary -= search_step;
217
        }
218
    }
219
220
    return search_boundary;
221
}
222
223
static int search_boundary(int plane_boundary, int search_range, int search_step, int vertical, int y, int x)
224
{
225
    return do_search_boundary(vertical ? y : x, plane_boundary, search_range, search_step);
226
}
227
228
static int cmp_scores(const void *a, const void *b)
229
{
230
    const struct PosPairCode *pair1 = a;
231
    const struct PosPairCode *pair2 = b;
232
    return FFDIFFSIGN(pair1->score, pair2->score);
233
}
234
235
static double do_block_ssd(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
236
{
237
    const uint8_t *srcp = src + pos->y * src_stride + pos->x;
238
    const uint8_t *refp = src + r_y * src_stride + r_x;
239
    const int block_size = s->block_size;
240
    double dist = 0.;
241
    int x, y;
242
243
    for (y = 0; y < block_size; y++) {
244
        for (x = 0; x < block_size; x++) {
245
            double temp = refp[x] - srcp[x];
246
            dist += temp * temp;
247
        }
248
249
        srcp += src_stride;
250
        refp += src_stride;
251
    }
252
253
    return dist;
254
}
255
256
static double do_block_ssd16(BM3DContext *s, PosCode *pos, const uint8_t *src, int src_stride, int r_y, int r_x)
257
{
258
    const uint16_t *srcp = (uint16_t *)src + pos->y * src_stride / 2 + pos->x;
259
    const uint16_t *refp = (uint16_t *)src + r_y * src_stride / 2 + r_x;
260
    const int block_size = s->block_size;
261
    double dist = 0.;
262
    int x, y;
263
264
    for (y = 0; y < block_size; y++) {
265
        for (x = 0; x < block_size; x++) {
266
            double temp = refp[x] - srcp[x];
267
            dist += temp * temp;
268
        }
269
270
        srcp += src_stride / 2;
271
        refp += src_stride / 2;
272
    }
273
274
    return dist;
275
}
276
277
static void do_block_matching_multi(BM3DContext *s, const uint8_t *src, int src_stride, int src_range,
278
                                    const PosCode *search_pos, int search_size, float th_mse,
279
                                    int r_y, int r_x, int plane, int jobnr)
280
{
281
    SliceContext *sc = &s->slices[jobnr];
282
    double MSE2SSE = s->group_size * s->block_size * s->block_size * src_range * src_range / (s->max * s->max);
283
    double distMul = 1. / MSE2SSE;
284
    double th_sse = th_mse * MSE2SSE;
285
    int i, index = sc->nb_match_blocks;
286
287
    for (i = 0; i < search_size; i++) {
288
        PosCode pos = search_pos[i];
289
        double dist;
290
291
        dist = s->do_block_ssd(s, &pos, src, src_stride, r_y, r_x);
292
293
        // Only match similar blocks but not identical blocks
294
        if (dist <= th_sse && dist != 0) {
295
            const double score = dist * distMul;
296
297
            if (index >= s->group_size && score >= sc->match_blocks[index - 1].score) {
298
                continue;
299
            }
300
301
            if (index >= s->group_size)
302
                index = s->group_size - 1;
303
304
            sc->match_blocks[index].score = score;
305
            sc->match_blocks[index].y = pos.y;
306
            sc->match_blocks[index].x = pos.x;
307
            index++;
308
            qsort(sc->match_blocks, index, sizeof(PosPairCode), cmp_scores);
309
        }
310
    }
311
312
    sc->nb_match_blocks = index;
313
}
314
315
static void block_matching_multi(BM3DContext *s, const uint8_t *ref, int ref_linesize, int y, int x,
316
                                 int exclude_cur_pos, int plane, int jobnr)
317
{
318
    SliceContext *sc = &s->slices[jobnr];
319
    const int width = s->planewidth[plane];
320
    const int height = s->planeheight[plane];
321
    const int block_size = s->block_size;
322
    const int step = s->bm_step;
323
    const int range = s->bm_range / step * step;
324
    int l = search_boundary(0, range, step, 0, y, x);
325
    int r = search_boundary(width - block_size, range, step, 0, y, x);
326
    int t = search_boundary(0, range, step, 1, y, x);
327
    int b = search_boundary(height - block_size, range, step, 1, y, x);
328
    int j, i, index = 0;
329
330
    for (j = t; j <= b; j += step) {
331
        for (i = l; i <= r; i += step) {
332
            PosCode pos;
333
334
            if (exclude_cur_pos > 0 && j == y && i == x) {
335
                continue;
336
            }
337
338
            pos.y = j;
339
            pos.x = i;
340
            sc->search_positions[index++] = pos;
341
        }
342
    }
343
344
    if (exclude_cur_pos == 1) {
345
        sc->match_blocks[0].score = 0;
346
        sc->match_blocks[0].y = y;
347
        sc->match_blocks[0].x = x;
348
        sc->nb_match_blocks = 1;
349
    }
350
351
    do_block_matching_multi(s, ref, ref_linesize, s->bm_range,
352
                            sc->search_positions, index, s->th_mse, y, x, plane, jobnr);
353
}
354
355
static void block_matching(BM3DContext *s, const uint8_t *ref, int ref_linesize,
356
                           int j, int i, int plane, int jobnr)
357
{
358
    SliceContext *sc = &s->slices[jobnr];
359
360
    if (s->group_size == 1 || s->th_mse <= 0.f) {
361
        sc->match_blocks[0].score = 1;
362
        sc->match_blocks[0].x = i;
363
        sc->match_blocks[0].y = j;
364
        sc->nb_match_blocks = 1;
365
        return;
366
    }
367
368
    sc->nb_match_blocks = 0;
369
    block_matching_multi(s, ref, ref_linesize, j, i, 1, plane, jobnr);
370
}
371
372
static void get_block_row(const uint8_t *srcp, int src_linesize,
373
                          int y, int x, int block_size, float *dst)
374
{
375
    const uint8_t *src = srcp + y * src_linesize + x;
376
    int j;
377
378
    for (j = 0; j < block_size; j++) {
379
        dst[j] = src[j];
380
    }
381
}
382
383
static void get_block_row16(const uint8_t *srcp, int src_linesize,
384
                            int y, int x, int block_size, float *dst)
385
{
386
    const uint16_t *src = (uint16_t *)srcp + y * src_linesize / 2 + x;
387
    int j;
388
389
    for (j = 0; j < block_size; j++) {
390
        dst[j] = src[j];
391
    }
392
}
393
394
static void basic_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
395
                                  const uint8_t *ref, int ref_linesize,
396
                                  int y, int x, int plane, int jobnr)
397
{
398
    SliceContext *sc = &s->slices[jobnr];
399
    const int buffer_linesize = s->block_size * s->block_size;
400
    const int nb_match_blocks = sc->nb_match_blocks;
401
    const int block_size = s->block_size;
402
    const int width = s->planewidth[plane];
403
    const int pgroup_size = s->pgroup_size;
404
    const int group_size = s->group_size;
405
    float *buffer = sc->buffer;
406
    float *bufferh = sc->bufferh;
407
    float *bufferv = sc->bufferv;
408
    float *bufferz = sc->bufferz;
409
    float threshold[4];
410
    float den_weight, num_weight;
411
    int retained = 0;
412
    int i, j, k;
413
414
    for (k = 0; k < nb_match_blocks; k++) {
415
        const int y = sc->match_blocks[k].y;
416
        const int x = sc->match_blocks[k].x;
417
418
        for (i = 0; i < block_size; i++) {
419
            s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
420
            av_dct_calc(sc->dctf, bufferh + block_size * i);
421
        }
422
423
        for (i = 0; i < block_size; i++) {
424
            for (j = 0; j < block_size; j++) {
425
                bufferv[i * block_size + j] = bufferh[j * block_size + i];
426
            }
427
            av_dct_calc(sc->dctf, bufferv + i * block_size);
428
        }
429
430
        for (i = 0; i < block_size; i++) {
431
            memcpy(buffer + k * buffer_linesize + i * block_size,
432
                   bufferv + i * block_size, block_size * 4);
433
        }
434
    }
435
436
    for (i = 0; i < block_size; i++) {
437
        for (j = 0; j < block_size; j++) {
438
            for (k = 0; k < nb_match_blocks; k++)
439
                bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
440
            if (group_size > 1)
441
                av_dct_calc(sc->gdctf, bufferz);
442
            bufferz += pgroup_size;
443
        }
444
    }
445
446
    threshold[0] = s->hard_threshold * s->sigma * M_SQRT2 * block_size * block_size * (1 << (s->depth - 8)) / 255.f;
447
    threshold[1] = threshold[0] * sqrtf(2.f);
448
    threshold[2] = threshold[0] * 2.f;
449
    threshold[3] = threshold[0] * sqrtf(8.f);
450
    bufferz = sc->bufferz;
451
452
    for (i = 0; i < block_size; i++) {
453
        for (j = 0; j < block_size; j++) {
454
            for (k = 0; k < nb_match_blocks; k++) {
455
                const float thresh = threshold[(j == 0) + (i == 0) + (k == 0)];
456
457
                if (bufferz[k] > thresh || bufferz[k] < -thresh) {
458
                    retained++;
459
                } else {
460
                    bufferz[k] = 0;
461
                }
462
            }
463
            bufferz += pgroup_size;
464
        }
465
    }
466
467
    bufferz = sc->bufferz;
468
    buffer = sc->buffer;
469
    for (i = 0; i < block_size; i++) {
470
        for (j = 0; j < block_size; j++) {
471
            if (group_size > 1)
472
                av_dct_calc(sc->gdcti, bufferz);
473
            for (k = 0; k < nb_match_blocks; k++) {
474
                buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
475
            }
476
            bufferz += pgroup_size;
477
        }
478
    }
479
480
    den_weight = retained < 1 ? 1.f : 1.f / retained;
481
    num_weight = den_weight;
482
483
    buffer = sc->buffer;
484
    for (k = 0; k < nb_match_blocks; k++) {
485
        float *num = sc->num + y * width + x;
486
        float *den = sc->den + y * width + x;
487
488
        for (i = 0; i < block_size; i++) {
489
            memcpy(bufferv + i * block_size,
490
                   buffer + k * buffer_linesize + i * block_size,
491
                   block_size * 4);
492
        }
493
494
        for (i = 0; i < block_size; i++) {
495
            av_dct_calc(sc->dcti, bufferv + block_size * i);
496
            for (j = 0; j < block_size; j++) {
497
                bufferh[j * block_size + i] = bufferv[i * block_size + j];
498
            }
499
        }
500
501
        for (i = 0; i < block_size; i++) {
502
            av_dct_calc(sc->dcti, bufferh + block_size * i);
503
            for (j = 0; j < block_size; j++) {
504
                num[j] += bufferh[i * block_size + j] * num_weight;
505
                den[j] += den_weight;
506
            }
507
            num += width;
508
            den += width;
509
        }
510
    }
511
}
512
513
static void final_block_filtering(BM3DContext *s, const uint8_t *src, int src_linesize,
514
                                  const uint8_t *ref, int ref_linesize,
515
                                  int y, int x, int plane, int jobnr)
516
{
517
    SliceContext *sc = &s->slices[jobnr];
518
    const int buffer_linesize = s->block_size * s->block_size;
519
    const int nb_match_blocks = sc->nb_match_blocks;
520
    const int block_size = s->block_size;
521
    const int width = s->planewidth[plane];
522
    const int pgroup_size = s->pgroup_size;
523
    const int group_size = s->group_size;
524
    const float sigma_sqr = s->sigma * s->sigma;
525
    float *buffer = sc->buffer;
526
    float *bufferh = sc->bufferh;
527
    float *bufferv = sc->bufferv;
528
    float *bufferz = sc->bufferz;
529
    float *rbuffer = sc->rbuffer;
530
    float *rbufferh = sc->rbufferh;
531
    float *rbufferv = sc->rbufferv;
532
    float *rbufferz = sc->rbufferz;
533
    float den_weight, num_weight;
534
    float l2_wiener = 0;
535
    int i, j, k;
536
537
    for (k = 0; k < nb_match_blocks; k++) {
538
        const int y = sc->match_blocks[k].y;
539
        const int x = sc->match_blocks[k].x;
540
541
        for (i = 0; i < block_size; i++) {
542
            s->get_block_row(src, src_linesize, y + i, x, block_size, bufferh + block_size * i);
543
            s->get_block_row(ref, ref_linesize, y + i, x, block_size, rbufferh + block_size * i);
544
            av_dct_calc(sc->dctf, bufferh + block_size * i);
545
            av_dct_calc(sc->dctf, rbufferh + block_size * i);
546
        }
547
548
        for (i = 0; i < block_size; i++) {
549
            for (j = 0; j < block_size; j++) {
550
                bufferv[i * block_size + j] = bufferh[j * block_size + i];
551
                rbufferv[i * block_size + j] = rbufferh[j * block_size + i];
552
            }
553
            av_dct_calc(sc->dctf, bufferv + i * block_size);
554
            av_dct_calc(sc->dctf, rbufferv + i * block_size);
555
        }
556
557
        for (i = 0; i < block_size; i++) {
558
            memcpy(buffer + k * buffer_linesize + i * block_size,
559
                   bufferv + i * block_size, block_size * 4);
560
            memcpy(rbuffer + k * buffer_linesize + i * block_size,
561
                   rbufferv + i * block_size, block_size * 4);
562
        }
563
    }
564
565
    for (i = 0; i < block_size; i++) {
566
        for (j = 0; j < block_size; j++) {
567
            for (k = 0; k < nb_match_blocks; k++) {
568
                bufferz[k] = buffer[buffer_linesize * k + i * block_size + j];
569
                rbufferz[k] = rbuffer[buffer_linesize * k + i * block_size + j];
570
            }
571
            if (group_size > 1) {
572
                av_dct_calc(sc->gdctf, bufferz);
573
                av_dct_calc(sc->gdctf, rbufferz);
574
            }
575
            bufferz += pgroup_size;
576
            rbufferz += pgroup_size;
577
        }
578
    }
579
580
    bufferz = sc->bufferz;
581
    rbufferz = sc->rbufferz;
582
583
    for (i = 0; i < block_size; i++) {
584
        for (j = 0; j < block_size; j++) {
585
            for (k = 0; k < nb_match_blocks; k++) {
586
                const float ref_sqr = rbufferz[k] * rbufferz[k];
587
                float wiener_coef = ref_sqr / (ref_sqr + sigma_sqr);
588
589
                if (isnan(wiener_coef))
590
                   wiener_coef = 1;
591
                bufferz[k] *= wiener_coef;
592
                l2_wiener += wiener_coef * wiener_coef;
593
            }
594
            bufferz += pgroup_size;
595
            rbufferz += pgroup_size;
596
        }
597
    }
598
599
    bufferz = sc->bufferz;
600
    buffer = sc->buffer;
601
    for (i = 0; i < block_size; i++) {
602
        for (j = 0; j < block_size; j++) {
603
            if (group_size > 1)
604
                av_dct_calc(sc->gdcti, bufferz);
605
            for (k = 0; k < nb_match_blocks; k++) {
606
                buffer[buffer_linesize * k + i * block_size + j] = bufferz[k];
607
            }
608
            bufferz += pgroup_size;
609
        }
610
    }
611
612
    l2_wiener = FFMAX(l2_wiener, 1e-15f);
613
    den_weight = 1.f / l2_wiener;
614
    num_weight = den_weight;
615
616
    for (k = 0; k < nb_match_blocks; k++) {
617
        float *num = sc->num + y * width + x;
618
        float *den = sc->den + y * width + x;
619
620
        for (i = 0; i < block_size; i++) {
621
            memcpy(bufferv + i * block_size,
622
                   buffer + k * buffer_linesize + i * block_size,
623
                   block_size * 4);
624
        }
625
626
        for (i = 0; i < block_size; i++) {
627
            av_dct_calc(sc->dcti, bufferv + block_size * i);
628
            for (j = 0; j < block_size; j++) {
629
                bufferh[j * block_size + i] = bufferv[i * block_size + j];
630
            }
631
        }
632
633
        for (i = 0; i < block_size; i++) {
634
            av_dct_calc(sc->dcti, bufferh + block_size * i);
635
            for (j = 0; j < block_size; j++) {
636
                num[j] += bufferh[i * block_size + j] * num_weight;
637
                den[j] += den_weight;
638
            }
639
            num += width;
640
            den += width;
641
        }
642
    }
643
}
644
645
static void do_output(BM3DContext *s, uint8_t *dst, int dst_linesize,
646
                      int plane, int nb_jobs)
647
{
648
    const int height = s->planeheight[plane];
649
    const int width = s->planewidth[plane];
650
    int i, j, k;
651
652
    for (i = 0; i < height; i++) {
653
        for (j = 0; j < width; j++) {
654
            uint8_t *dstp = dst + i * dst_linesize;
655
            float sum_den = 0.f;
656
            float sum_num = 0.f;
657
658
            for (k = 0; k < nb_jobs; k++) {
659
                SliceContext *sc = &s->slices[k];
660
                float num = sc->num[i * width + j];
661
                float den = sc->den[i * width + j];
662
663
                sum_num += num;
664
                sum_den += den;
665
            }
666
667
            dstp[j] = av_clip_uint8(lrintf(sum_num / sum_den));
668
        }
669
    }
670
}
671
672
static void do_output16(BM3DContext *s, uint8_t *dst, int dst_linesize,
673
                        int plane, int nb_jobs)
674
{
675
    const int height = s->planeheight[plane];
676
    const int width = s->planewidth[plane];
677
    const int depth = s->depth;
678
    int i, j, k;
679
680
    for (i = 0; i < height; i++) {
681
        for (j = 0; j < width; j++) {
682
            uint16_t *dstp = (uint16_t *)dst + i * dst_linesize / 2;
683
            float sum_den = 0.f;
684
            float sum_num = 0.f;
685
686
            for (k = 0; k < nb_jobs; k++) {
687
                SliceContext *sc = &s->slices[k];
688
                float num = sc->num[i * width + j];
689
                float den = sc->den[i * width + j];
690
691
                sum_num += num;
692
                sum_den += den;
693
            }
694
695
            dstp[j] = av_clip_uintp2_c(lrintf(sum_num / sum_den), depth);
696
        }
697
    }
698
}
699
700
static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
701
{
702
    BM3DContext *s = ctx->priv;
703
    SliceContext *sc = &s->slices[jobnr];
704
    const int block_step = s->block_step;
705
    ThreadData *td = arg;
706
    const uint8_t *src = td->src;
707
    const uint8_t *ref = td->ref;
708
    const int src_linesize = td->src_linesize;
709
    const int ref_linesize = td->ref_linesize;
710
    const int plane = td->plane;
711
    const int width = s->planewidth[plane];
712
    const int height = s->planeheight[plane];
713
    const int block_pos_bottom = FFMAX(0, height - s->block_size);
714
    const int block_pos_right  = FFMAX(0, width - s->block_size);
715
    const int slice_start = (((height + block_step - 1) / block_step) * jobnr / nb_jobs) * block_step;
716
    const int slice_end = (jobnr == nb_jobs - 1) ? block_pos_bottom + block_step :
717
                          (((height + block_step - 1) / block_step) * (jobnr + 1) / nb_jobs) * block_step;
718
    int i, j;
719
720
    memset(sc->num, 0, width * height * sizeof(FFTSample));
721
    memset(sc->den, 0, width * height * sizeof(FFTSample));
722
723
    for (j = slice_start; j < slice_end; j += block_step) {
724
        if (j > block_pos_bottom) {
725
            j = block_pos_bottom;
726
        }
727
728
        for (i = 0; i < block_pos_right + block_step; i += block_step) {
729
            if (i > block_pos_right) {
730
                i = block_pos_right;
731
            }
732
733
            block_matching(s, ref, ref_linesize, j, i, plane, jobnr);
734
735
            s->block_filtering(s, src, src_linesize,
736
                               ref, ref_linesize, j, i, plane, jobnr);
737
        }
738
    }
739
740
    return 0;
741
}
742
743
static int filter_frame(AVFilterContext *ctx, AVFrame **out, AVFrame *in, AVFrame *ref)
744
{
745
    BM3DContext *s = ctx->priv;
746
    AVFilterLink *outlink = ctx->outputs[0];
747
    int p;
748
749
    *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
750
    if (!*out)
751
        return AVERROR(ENOMEM);
752
    av_frame_copy_props(*out, in);
753
754
    for (p = 0; p < s->nb_planes; p++) {
755
        const int nb_jobs = FFMAX(1, FFMIN(s->nb_threads, s->planeheight[p] / s->block_size));
756
        ThreadData td;
757
758
        if (!((1 << p) & s->planes) || ctx->is_disabled) {
759
            av_image_copy_plane((*out)->data[p], (*out)->linesize[p],
760
                                in->data[p], in->linesize[p],
761
                                s->planewidth[p], s->planeheight[p]);
762
            continue;
763
        }
764
765
        td.src = in->data[p];
766
        td.src_linesize = in->linesize[p];
767
        td.ref = ref->data[p];
768
        td.ref_linesize = ref->linesize[p];
769
        td.plane = p;
770
        ctx->internal->execute(ctx, filter_slice, &td, NULL, nb_jobs);
771
772
        s->do_output(s, (*out)->data[p], (*out)->linesize[p], p, nb_jobs);
773
    }
774
775
    return 0;
776
}
777
778
#define SQR(x) ((x) * (x))
779
780
static int config_input(AVFilterLink *inlink)
781
{
782
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
783
    AVFilterContext *ctx = inlink->dst;
784
    BM3DContext *s = ctx->priv;
785
    int i, group_bits;
786
787
    s->nb_threads = FFMIN(ff_filter_get_nb_threads(ctx), MAX_NB_THREADS);
788
    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
789
    s->depth = desc->comp[0].depth;
790
    s->max = (1 << s->depth) - 1;
791
    s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
792
    s->planeheight[0] = s->planeheight[3] = inlink->h;
793
    s->planewidth[1]  = s->planewidth[2]  = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
794
    s->planewidth[0]  = s->planewidth[3]  = inlink->w;
795
796
    for (group_bits = 4; 1 << group_bits < s->group_size; group_bits++);
797
    s->group_bits = group_bits;
798
    s->pgroup_size = 1 << group_bits;
799
800
    for (i = 0; i < s->nb_threads; i++) {
801
        SliceContext *sc = &s->slices[i];
802
803
        sc->num = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(FFTSample));
804
        sc->den = av_calloc(FFALIGN(s->planewidth[0], s->block_size) * FFALIGN(s->planeheight[0], s->block_size), sizeof(FFTSample));
805
        if (!sc->num || !sc->den)
806
            return AVERROR(ENOMEM);
807
808
        sc->dctf = av_dct_init(av_log2(s->block_size), DCT_II);
809
        sc->dcti = av_dct_init(av_log2(s->block_size), DCT_III);
810
        if (!sc->dctf || !sc->dcti)
811
            return AVERROR(ENOMEM);
812
813
        if (s->group_bits > 1) {
814
            sc->gdctf = av_dct_init(s->group_bits, DCT_II);
815
            sc->gdcti = av_dct_init(s->group_bits, DCT_III);
816
            if (!sc->gdctf || !sc->gdcti)
817
                return AVERROR(ENOMEM);
818
        }
819
820
        sc->buffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->buffer));
821
        sc->bufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->bufferz));
822
        sc->bufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferh));
823
        sc->bufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->bufferv));
824
        if (!sc->bufferh || !sc->bufferv || !sc->buffer || !sc->bufferz)
825
            return AVERROR(ENOMEM);
826
827
        if (s->mode == FINAL) {
828
            sc->rbuffer = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbuffer));
829
            sc->rbufferz = av_calloc(s->block_size * s->block_size * s->pgroup_size, sizeof(*sc->rbufferz));
830
            sc->rbufferh = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferh));
831
            sc->rbufferv = av_calloc(s->block_size * s->block_size, sizeof(*sc->rbufferv));
832
            if (!sc->rbufferh || !sc->rbufferv || !sc->rbuffer || !sc->rbufferz)
833
                return AVERROR(ENOMEM);
834
        }
835
836
        sc->search_positions = av_calloc(SQR(2 * s->bm_range / s->bm_step + 1), sizeof(*sc->search_positions));
837
        if (!sc->search_positions)
838
            return AVERROR(ENOMEM);
839
    }
840
841
    s->do_output = do_output;
842
    s->do_block_ssd = do_block_ssd;
843
    s->get_block_row = get_block_row;
844
845
    if (s->depth > 8) {
846
        s->do_output = do_output16;
847
        s->do_block_ssd = do_block_ssd16;
848
        s->get_block_row = get_block_row16;
849
    }
850
851
    return 0;
852
}
853
854
static int activate(AVFilterContext *ctx)
855
{
856
    BM3DContext *s = ctx->priv;
857
858
    if (!s->ref) {
859
        AVFrame *frame = NULL;
860
        AVFrame *out = NULL;
861
        int ret, status;
862
        int64_t pts;
863
864
        FF_FILTER_FORWARD_STATUS_BACK(ctx->outputs[0], ctx->inputs[0]);
865
866
        if ((ret = ff_inlink_consume_frame(ctx->inputs[0], &frame)) > 0) {
867
            ret = filter_frame(ctx, &out, frame, frame);
868
            av_frame_free(&frame);
869
            if (ret < 0)
870
                return ret;
871
            ret = ff_filter_frame(ctx->outputs[0], out);
872
        }
873
        if (ret < 0) {
874
            return ret;
875
        } else if (ff_inlink_acknowledge_status(ctx->inputs[0], &status, &pts)) {
876
            ff_outlink_set_status(ctx->outputs[0], status, pts);
877
            return 0;
878
        } else {
879
            if (ff_outlink_frame_wanted(ctx->outputs[0]))
880
                ff_inlink_request_frame(ctx->inputs[0]);
881
            return 0;
882
        }
883
    } else {
884
        return ff_framesync_activate(&s->fs);
885
    }
886
}
887
888
static int process_frame(FFFrameSync *fs)
889
{
890
    AVFilterContext *ctx = fs->parent;
891
    BM3DContext *s = fs->opaque;
892
    AVFilterLink *outlink = ctx->outputs[0];
893
    AVFrame *out = NULL, *src, *ref;
894
    int ret;
895
896
    if ((ret = ff_framesync_get_frame(&s->fs, 0, &src, 0)) < 0 ||
897
        (ret = ff_framesync_get_frame(&s->fs, 1, &ref, 0)) < 0)
898
        return ret;
899
900
    if ((ret = filter_frame(ctx, &out, src, ref)) < 0)
901
        return ret;
902
903
    out->pts = av_rescale_q(src->pts, s->fs.time_base, outlink->time_base);
904
905
    return ff_filter_frame(outlink, out);
906
}
907
908
static av_cold int init(AVFilterContext *ctx)
909
{
910
    BM3DContext *s = ctx->priv;
911
    AVFilterPad pad = { 0 };
912
    int ret;
913
914
    if (s->mode == BASIC) {
915
        if (s->th_mse == 0.f)
916
            s->th_mse = 400.f + s->sigma * 80.f;
917
        s->block_filtering = basic_block_filtering;
918
    } else if (s->mode == FINAL) {
919
        if (!s->ref) {
920
            av_log(ctx, AV_LOG_WARNING, "Reference stream is mandatory in final estimation mode.\n");
921
            s->ref = 1;
922
        }
923
        if (s->th_mse == 0.f)
924
            s->th_mse = 200.f + s->sigma * 10.f;
925
926
        s->block_filtering = final_block_filtering;
927
    } else {
928
        return AVERROR_BUG;
929
    }
930
931
    s->block_size = 1 << s->block_size;
932
933
    if (s->block_step > s->block_size) {
934
        av_log(ctx, AV_LOG_WARNING, "bstep: %d can't be bigger than block size. Changing to %d.\n",
935
               s->block_step, s->block_size);
936
        s->block_step = s->block_size;
937
    }
938
    if (s->bm_step > s->bm_range) {
939
        av_log(ctx, AV_LOG_WARNING, "mstep: %d can't be bigger than block matching range. Changing to %d.\n",
940
               s->bm_step, s->bm_range);
941
        s->bm_step = s->bm_range;
942
    }
943
944
    pad.type         = AVMEDIA_TYPE_VIDEO;
945
    pad.name         = "source";
946
    pad.config_props = config_input;
947
948
    if ((ret = ff_insert_inpad(ctx, 0, &pad)) < 0)
949
        return ret;
950
951
    if (s->ref) {
952
        pad.type         = AVMEDIA_TYPE_VIDEO;
953
        pad.name         = "reference";
954
        pad.config_props = NULL;
955
956
        if ((ret = ff_insert_inpad(ctx, 1, &pad)) < 0)
957
            return ret;
958
    }
959
960
    return 0;
961
}
962
963
static int config_output(AVFilterLink *outlink)
964
{
965
    AVFilterContext *ctx = outlink->src;
966
    BM3DContext *s = ctx->priv;
967
    AVFilterLink *src = ctx->inputs[0];
968
    AVFilterLink *ref;
969
    FFFrameSyncIn *in;
970
    int ret;
971
972
    if (s->ref) {
973
        ref = ctx->inputs[1];
974
975
        if (src->format != ref->format) {
976
            av_log(ctx, AV_LOG_ERROR, "inputs must be of same pixel format\n");
977
            return AVERROR(EINVAL);
978
        }
979
        if (src->w                       != ref->w ||
980
            src->h                       != ref->h) {
981
            av_log(ctx, AV_LOG_ERROR, "First input link %s parameters "
982
                   "(size %dx%d) do not match the corresponding "
983
                   "second input link %s parameters (%dx%d) ",
984
                   ctx->input_pads[0].name, src->w, src->h,
985
                   ctx->input_pads[1].name, ref->w, ref->h);
986
            return AVERROR(EINVAL);
987
        }
988
    }
989
990
    outlink->w = src->w;
991
    outlink->h = src->h;
992
    outlink->time_base = src->time_base;
993
    outlink->sample_aspect_ratio = src->sample_aspect_ratio;
994
    outlink->frame_rate = src->frame_rate;
995
996
    if (!s->ref)
997
        return 0;
998
999
    if ((ret = ff_framesync_init(&s->fs, ctx, 2)) < 0)
1000
        return ret;
1001
1002
    in = s->fs.in;
1003
    in[0].time_base = src->time_base;
1004
    in[1].time_base = ref->time_base;
1005
    in[0].sync   = 1;
1006
    in[0].before = EXT_STOP;
1007
    in[0].after  = EXT_STOP;
1008
    in[1].sync   = 1;
1009
    in[1].before = EXT_STOP;
1010
    in[1].after  = EXT_STOP;
1011
    s->fs.opaque   = s;
1012
    s->fs.on_event = process_frame;
1013
1014
    return ff_framesync_configure(&s->fs);
1015
}
1016
1017
static av_cold void uninit(AVFilterContext *ctx)
1018
{
1019
    BM3DContext *s = ctx->priv;
1020
    int i;
1021
1022
    if (s->ref)
1023
        ff_framesync_uninit(&s->fs);
1024
1025
    for (i = 0; i < s->nb_threads; i++) {
1026
        SliceContext *sc = &s->slices[i];
1027
1028
        av_freep(&sc->num);
1029
        av_freep(&sc->den);
1030
1031
        av_dct_end(sc->gdctf);
1032
        av_dct_end(sc->gdcti);
1033
        av_dct_end(sc->dctf);
1034
        av_dct_end(sc->dcti);
1035
1036
        av_freep(&sc->buffer);
1037
        av_freep(&sc->bufferh);
1038
        av_freep(&sc->bufferv);
1039
        av_freep(&sc->bufferz);
1040
        av_freep(&sc->rbuffer);
1041
        av_freep(&sc->rbufferh);
1042
        av_freep(&sc->rbufferv);
1043
        av_freep(&sc->rbufferz);
1044
1045
        av_freep(&sc->search_positions);
1046
    }
1047
}
1048
1049
static const AVFilterPad bm3d_outputs[] = {
1050
    {
1051
        .name         = "default",
1052
        .type         = AVMEDIA_TYPE_VIDEO,
1053
        .config_props = config_output,
1054
    },
1055
    { NULL }
1056
};
1057
1058
AVFilter ff_vf_bm3d = {
1059
    .name          = "bm3d",
1060
    .description   = NULL_IF_CONFIG_SMALL("Block-Matching 3D denoiser."),
1061
    .priv_size     = sizeof(BM3DContext),
1062
    .init          = init,
1063
    .uninit        = uninit,
1064
    .activate      = activate,
1065
    .query_formats = query_formats,
1066
    .inputs        = NULL,
1067
    .outputs       = bm3d_outputs,
1068
    .priv_class    = &bm3d_class,
1069
    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
1070
                     AVFILTER_FLAG_DYNAMIC_INPUTS |
1071
                     AVFILTER_FLAG_SLICE_THREADS,
1072
};