aboutsummaryrefslogtreecommitdiffstats
path: root/libavfilter/vf_nlmeans.c
blob: 06233b0dd46a0cdbb9bc423bbc95467c728c68b9 (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
/*
 * Copyright (c) 2016 Clément Bœsch <u pkh me>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @todo
 * - better automatic defaults? see "Parameters" @ http://www.ipol.im/pub/art/2011/bcm_nlm/
 * - temporal support (probably doesn't need any displacement according to
 *   "Denoising image sequences does not require motion estimation")
 * - Bayer pixel format support for at least raw photos? (DNG support would be
 *   handy here)
 * - FATE test (probably needs visual threshold test mechanism due to the use
 *   of floats)
 */

#include "libavutil/avassert.h"
#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
#include "avfilter.h"
#include "formats.h"
#include "internal.h"
#include "vf_nlmeans.h"
#include "video.h"

struct weighted_avg {
    float total_weight;
    float sum;
};

typedef struct NLMeansContext {
    const AVClass *class;
    int nb_planes;
    int chroma_w, chroma_h;
    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
    double sigma;                               // denoising strength
    int patch_size,    patch_hsize;             // patch size and half size
    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
    int research_size,    research_hsize;       // research size and half size
    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
    uint32_t *ii_orig;                          // integral image
    uint32_t *ii;                               // integral image starting after the 0-line and 0-column
    int ii_w, ii_h;                             // width and height of the integral image
    ptrdiff_t ii_lz_32;                         // linesize in 32-bit units of the integral image
    struct weighted_avg *wa;                    // weighted average of every pixel
    ptrdiff_t wa_linesize;                      // linesize for wa in struct size unit
    float *weight_lut;                          // lookup table mapping (scaled) patch differences to their associated weights
    uint32_t max_meaningful_diff;               // maximum difference considered (if the patch difference is too high we ignore the pixel)
    NLMeansDSPContext dsp;
} NLMeansContext;

#define OFFSET(x) offsetof(NLMeansContext, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
static const AVOption nlmeans_options[] = {
    { "s",  "denoising strength", OFFSET(sigma), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 }, 1.0, 30.0, FLAGS },
    { "p",  "patch size",                   OFFSET(patch_size),    AV_OPT_TYPE_INT, { .i64 = 3*2+1 }, 0, 99, FLAGS },
    { "pc", "patch size for chroma planes", OFFSET(patch_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 },     0, 99, FLAGS },
    { "r",  "research window",                   OFFSET(research_size),    AV_OPT_TYPE_INT, { .i64 = 7*2+1 }, 0, 99, FLAGS },
    { "rc", "research window for chroma planes", OFFSET(research_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 },     0, 99, FLAGS },
    { NULL }
};

AVFILTER_DEFINE_CLASS(nlmeans);

static int query_formats(AVFilterContext *ctx)
{
    static const enum AVPixelFormat pix_fmts[] = {
        AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
        AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
        AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
        AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
        AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
        AV_PIX_FMT_YUVJ411P,
        AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
        AV_PIX_FMT_NONE
    };

    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
    if (!fmts_list)
        return AVERROR(ENOMEM);
    return ff_set_common_formats(ctx, fmts_list);
}

/**
 * Compute squared difference of the safe area (the zone where s1 and s2
 * overlap). It is likely the largest integral zone, so it is interesting to do
 * as little checks as possible; contrary to the unsafe version of this
 * function, we do not need any clipping here.
 *
 * The line above dst and the column to its left are always readable.
 */
static void compute_safe_ssd_integral_image_c(uint32_t *dst, ptrdiff_t dst_linesize_32,
                                              const uint8_t *s1, ptrdiff_t linesize1,
                                              const uint8_t *s2, ptrdiff_t linesize2,
                                              int w, int h)
{
    int x, y;
    const uint32_t *dst_top = dst - dst_linesize_32;

    /* SIMD-friendly assumptions allowed here */
    av_assert2(!(w & 0xf) && w >= 16 && h >= 1);

    for (y = 0; y < h; y++) {
        for (x = 0; x < w; x += 4) {
            const int d0 = s1[x    ] - s2[x    ];
            const int d1 = s1[x + 1] - s2[x + 1];
            const int d2 = s1[x + 2] - s2[x + 2];
            const int d3 = s1[x + 3] - s2[x + 3];

            dst[x    ] = dst_top[x    ] - dst_top[x - 1] + d0*d0;
            dst[x + 1] = dst_top[x + 1] - dst_top[x    ] + d1*d1;
            dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
            dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;

            dst[x    ] += dst[x - 1];
            dst[x + 1] += dst[x    ];
            dst[x + 2] += dst[x + 1];
            dst[x + 3] += dst[x + 2];
        }
        s1  += linesize1;
        s2  += linesize2;
        dst += dst_linesize_32;
        dst_top += dst_linesize_32;
    }
}

/**
 * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
 * be readable).
 *
 * On the other hand, the line above dst and the column to its left are always
 * readable.
 *
 * There is little point in having this function SIMDified as it is likely too
 * complex and only handle small portions of the image.
 *
 * @param dst               integral image
 * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
 * @param startx            integral starting x position
 * @param starty            integral starting y position
 * @param src               source plane buffer
 * @param linesize          source plane linesize
 * @param offx              source offsetting in x
 * @param offy              source offsetting in y
 * @paran r                 absolute maximum source offsetting
 * @param sw                source width
 * @param sh                source height
 * @param w                 width to compute
 * @param h                 height to compute
 */
static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t dst_linesize_32,
                                                     int startx, int starty,
                                                     const uint8_t *src, ptrdiff_t linesize,
                                                     int offx, int offy, int r, int sw, int sh,
                                                     int w, int h)
{
    int x, y;

    for (y = starty; y < starty + h; y++) {
        uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
        const int s1y = av_clip(y -  r,         0, sh - 1);
        const int s2y = av_clip(y - (r + offy), 0, sh - 1);

        for (x = startx; x < startx + w; x++) {
            const int s1x = av_clip(x -  r,         0, sw - 1);
            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
            const uint8_t v1 = src[s1y*linesize + s1x];
            const uint8_t v2 = src[s2y*linesize + s2x];
            const int d = v1 - v2;
            acc += d * d;
            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
        }
    }
}

/*
 * Compute the sum of squared difference integral image
 * http://www.ipol.im/pub/art/2014/57/
 * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
 *
 * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
 *                          an additional zeroed top line and column already
 *                          "applied" to the pointer value
 * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
 * @param src               source plane buffer
 * @param linesize          source plane linesize
 * @param offx              x-offsetting ranging in [-e;e]
 * @param offy              y-offsetting ranging in [-e;e]
 * @param w                 source width
 * @param h                 source height
 * @param e                 research padding edge
 */
static void compute_ssd_integral_image(const NLMeansDSPContext *dsp,
                                       uint32_t *ii, ptrdiff_t ii_linesize_32,
                                       const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
                                       int e, int w, int h)
{
    // ii has a surrounding padding of thickness "e"
    const int ii_w = w + e*2;
    const int ii_h = h + e*2;

    // we center the first source
    const int s1x = e;
    const int s1y = e;

    // 2nd source is the frame with offsetting
    const int s2x = e + offx;
    const int s2y = e + offy;

    // get the dimension of the overlapping rectangle where it is always safe
    // to compare the 2 sources pixels
    const int startx_safe = FFMAX(s1x, s2x);
    const int starty_safe = FFMAX(s1y, s2y);
    const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
    const int endy_safe   = FFMIN(s1y + h, s2y + h);

    // deduce the safe area width and height
    const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
    const int safe_ph = endy_safe - starty_safe;

    // adjusted end x position of the safe area after width of the safe area gets aligned
    const int endx_safe = startx_safe + safe_pw;

    // top part where only one of s1 and s2 is still readable, or none at all
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, 0,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w, starty_safe);

    // fill the left column integral required to compute the central
    // overlapping one
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, starty_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      startx_safe, safe_ph);

    // main and safe part of the integral
    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
    if (safe_pw && safe_ph)
        dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
                                             src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
                                             src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
                                             safe_pw, safe_ph);

    // right part of the integral
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      endx_safe, starty_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w - endx_safe, safe_ph);

    // bottom part where only one of s1 and s2 is still readable, or none at all
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, endy_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w, ii_h - endy_safe);
}

static int config_input(AVFilterLink *inlink)
{
    AVFilterContext *ctx = inlink->dst;
    NLMeansContext *s = ctx->priv;
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
    const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
                + FFMAX(s->patch_hsize,    s->patch_hsize_uv);

    s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
    s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
    s->nb_planes = av_pix_fmt_count_planes(inlink->format);

    /* Allocate the integral image with extra edges of thickness "e"
     *
     *   +_+-------------------------------+
     *   |0|0000000000000000000000000000000|
     *   +-x-------------------------------+
     *   |0|\    ^                         |
     *   |0| ii  | e                       |
     *   |0|     v                         |
     *   |0|   +-----------------------+   |
     *   |0|   |                       |   |
     *   |0|<->|                       |   |
     *   |0| e |                       |   |
     *   |0|   |                       |   |
     *   |0|   +-----------------------+   |
     *   |0|                               |
     *   |0|                               |
     *   |0|                               |
     *   +-+-------------------------------+
     */
    s->ii_w = inlink->w + e*2;
    s->ii_h = inlink->h + e*2;

    // align to 4 the linesize, "+1" is for the space of the left 0-column
    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);

    // "+1" is for the space of the top 0-line
    s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
    if (!s->ii_orig)
        return AVERROR(ENOMEM);

    // skip top 0-line and left 0-column
    s->ii = s->ii_orig + s->ii_lz_32 + 1;

    // allocate weighted average for every pixel
    s->wa_linesize = inlink->w;
    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
    if (!s->wa)
        return AVERROR(ENOMEM);

    return 0;
}

struct thread_data {
    const uint8_t *src;
    ptrdiff_t src_linesize;
    int startx, starty;
    int endx, endy;
    const uint32_t *ii_start;
    int p;
};

static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
{
    int x, y;
    NLMeansContext *s = ctx->priv;
    const struct thread_data *td = arg;
    const ptrdiff_t src_linesize = td->src_linesize;
    const int process_h = td->endy - td->starty;
    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
    const int starty = td->starty + slice_start;
    const int endy   = td->starty + slice_end;
    const int p = td->p;
    const uint32_t *ii = td->ii_start + (starty - p - 1) * s->ii_lz_32 - p - 1;
    const int dist_b = 2*p + 1;
    const int dist_d = dist_b * s->ii_lz_32;
    const int dist_e = dist_d + dist_b;

    for (y = starty; y < endy; y++) {
        const uint8_t *src = td->src + y*src_linesize;
        struct weighted_avg *wa = s->wa + y*s->wa_linesize;
        for (x = td->startx; x < td->endx; x++) {
            /*
             * M is a discrete map where every entry contains the sum of all the entries
             * in the rectangle from the top-left origin of M to its coordinate. In the
             * following schema, "i" contains the sum of the whole map:
             *
             * M = +----------+-----------------+----+
             *     |          |                 |    |
             *     |          |                 |    |
             *     |         a|                b|   c|
             *     +----------+-----------------+----+
             *     |          |                 |    |
             *     |          |                 |    |
             *     |          |        X        |    |
             *     |          |                 |    |
             *     |         d|                e|   f|
             *     +----------+-----------------+----+
             *     |          |                 |    |
             *     |         g|                h|   i|
             *     +----------+-----------------+----+
             *
             * The sum of the X box can be calculated with:
             *    X = e-d-b+a
             *
             * See https://en.wikipedia.org/wiki/Summed_area_table
             *
             * The compute*_ssd functions compute the integral image M where every entry
             * contains the sum of the squared difference of every corresponding pixels of
             * two input planes of the same size as M.
             */
            const uint32_t a = ii[x];
            const uint32_t b = ii[x + dist_b];
            const uint32_t d = ii[x + dist_d];
            const uint32_t e = ii[x + dist_e];
            const uint32_t patch_diff_sq = e - d - b + a;

            if (patch_diff_sq < s->max_meaningful_diff) {
                const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
                wa[x].total_weight += weight;
                wa[x].sum += weight * src[x];
            }
        }
        ii += s->ii_lz_32;
    }
    return 0;
}

static void weight_averages(uint8_t *dst, ptrdiff_t dst_linesize,
                            const uint8_t *src, ptrdiff_t src_linesize,
                            struct weighted_avg *wa, ptrdiff_t wa_linesize,
                            int w, int h)
{
    int x, y;

    for (y = 0; y < h; y++) {
        for (x = 0; x < w; x++) {
            // Also weight the centered pixel
            wa[x].total_weight += 1.f;
            wa[x].sum += 1.f * src[x];
            dst[x] = av_clip_uint8(wa[x].sum / wa[x].total_weight + 0.5f);
        }
        dst += dst_linesize;
        src += src_linesize;
        wa += wa_linesize;
    }
}

static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
                         uint8_t *dst, ptrdiff_t dst_linesize,
                         const uint8_t *src, ptrdiff_t src_linesize)
{
    int offx, offy;
    NLMeansContext *s = ctx->priv;
    /* patches center points cover the whole research window so the patches
     * themselves overflow the research window */
    const int e = r + p;
    /* focus an integral pointer on the centered image (s1) */
    const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;

    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));

    for (offy = -r; offy <= r; offy++) {
        for (offx = -r; offx <= r; offx++) {
            if (offx || offy) {
                struct thread_data td = {
                    .src          = src + offy*src_linesize + offx,
                    .src_linesize = src_linesize,
                    .startx       = FFMAX(0, -offx),
                    .starty       = FFMAX(0, -offy),
                    .endx         = FFMIN(w, w - offx),
                    .endy         = FFMIN(h, h - offy),
                    .ii_start     = centered_ii + offy*s->ii_lz_32 + offx,
                    .p            = p,
                };

                compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32,
                                           src, src_linesize,
                                           offx, offy, e, w, h);
                ctx->internal->execute(ctx, nlmeans_slice, &td, NULL,
                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
            }
        }
    }

    weight_averages(dst, dst_linesize, src, src_linesize,
                    s->wa, s->wa_linesize, w, h);

    return 0;
}

static int filter_frame(AVFilterLink *inlink, AVFrame *in)
{
    int i;
    AVFilterContext *ctx = inlink->dst;
    NLMeansContext *s = ctx->priv;
    AVFilterLink *outlink = ctx->outputs[0];

    AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
    if (!out) {
        av_frame_free(&in);
        return AVERROR(ENOMEM);
    }
    av_frame_copy_props(out, in);

    for (i = 0; i < s->nb_planes; i++) {
        const int w = i ? s->chroma_w          : inlink->w;
        const int h = i ? s->chroma_h          : inlink->h;
        const int p = i ? s->patch_hsize_uv    : s->patch_hsize;
        const int r = i ? s->research_hsize_uv : s->research_hsize;
        nlmeans_plane(ctx, w, h, p, r,
                      out->data[i], out->linesize[i],
                      in->data[i],  in->linesize[i]);
    }

    av_frame_free(&in);
    return ff_filter_frame(outlink, out);
}

#define CHECK_ODD_FIELD(field, name) do {                       \
    if (!(s->field & 1)) {                                      \
        s->field |= 1;                                          \
        av_log(ctx, AV_LOG_WARNING, name " size must be odd, "  \
               "setting it to %d\n", s->field);                 \
    }                                                           \
} while (0)

void ff_nlmeans_init(NLMeansDSPContext *dsp)
{
    dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c;

    if (ARCH_AARCH64)
        ff_nlmeans_init_aarch64(dsp);
}

static av_cold int init(AVFilterContext *ctx)
{
    int i;
    NLMeansContext *s = ctx->priv;
    const double h = s->sigma * 10.;

    s->pdiff_scale = 1. / (h * h);
    s->max_meaningful_diff = log(255.) / s->pdiff_scale;
    s->weight_lut = av_calloc(s->max_meaningful_diff, sizeof(*s->weight_lut));
    if (!s->weight_lut)
        return AVERROR(ENOMEM);
    for (i = 0; i < s->max_meaningful_diff; i++)
        s->weight_lut[i] = exp(-i * s->pdiff_scale);

    CHECK_ODD_FIELD(research_size,   "Luma research window");
    CHECK_ODD_FIELD(patch_size,      "Luma patch");

    if (!s->research_size_uv) s->research_size_uv = s->research_size;
    if (!s->patch_size_uv)    s->patch_size_uv    = s->patch_size;

    CHECK_ODD_FIELD(research_size_uv, "Chroma research window");
    CHECK_ODD_FIELD(patch_size_uv,    "Chroma patch");

    s->research_hsize    = s->research_size    / 2;
    s->research_hsize_uv = s->research_size_uv / 2;
    s->patch_hsize       = s->patch_size       / 2;
    s->patch_hsize_uv    = s->patch_size_uv    / 2;

    av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
           s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
           s->patch_size,    s->patch_size,    s->patch_size_uv,    s->patch_size_uv);

    ff_nlmeans_init(&s->dsp);

    return 0;
}

static av_cold void uninit(AVFilterContext *ctx)
{
    NLMeansContext *s = ctx->priv;
    av_freep(&s->weight_lut);
    av_freep(&s->ii_orig);
    av_freep(&s->wa);
}

static const AVFilterPad nlmeans_inputs[] = {
    {
        .name         = "default",
        .type         = AVMEDIA_TYPE_VIDEO,
        .config_props = config_input,
        .filter_frame = filter_frame,
    },
    { NULL }
};

static const AVFilterPad nlmeans_outputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO,
    },
    { NULL }
};

AVFilter ff_vf_nlmeans = {
    .name          = "nlmeans",
    .description   = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
    .priv_size     = sizeof(NLMeansContext),
    .init          = init,
    .uninit        = uninit,
    .query_formats = query_formats,
    .inputs        = nlmeans_inputs,
    .outputs       = nlmeans_outputs,
    .priv_class    = &nlmeans_class,
    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,
};