GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavfilter/af_dynaudnorm.c Lines: 0 396 0.0 %
Date: 2020-10-23 17:01:47 Branches: 0 264 0.0 %

Line Branch Exec Source
1
/*
2
 * Dynamic Audio Normalizer
3
 * Copyright (c) 2015 LoRd_MuldeR <mulder2@gmx.de>. Some rights reserved.
4
 *
5
 * This file is part of FFmpeg.
6
 *
7
 * FFmpeg is free software; you can redistribute it and/or
8
 * modify it under the terms of the GNU Lesser General Public
9
 * License as published by the Free Software Foundation; either
10
 * version 2.1 of the License, or (at your option) any later version.
11
 *
12
 * FFmpeg is distributed in the hope that it will be useful,
13
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15
 * Lesser General Public License for more details.
16
 *
17
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with FFmpeg; if not, write to the Free Software
19
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20
 */
21
22
/**
23
 * @file
24
 * Dynamic Audio Normalizer
25
 */
26
27
#include <float.h>
28
29
#include "libavutil/avassert.h"
30
#include "libavutil/opt.h"
31
32
#define MIN_FILTER_SIZE 3
33
#define MAX_FILTER_SIZE 301
34
35
#define FF_BUFQUEUE_SIZE (MAX_FILTER_SIZE + 1)
36
#include "libavfilter/bufferqueue.h"
37
38
#include "audio.h"
39
#include "avfilter.h"
40
#include "filters.h"
41
#include "internal.h"
42
43
typedef struct local_gain {
44
    double max_gain;
45
    double threshold;
46
} local_gain;
47
48
typedef struct cqueue {
49
    double *elements;
50
    int size;
51
    int max_size;
52
    int nb_elements;
53
} cqueue;
54
55
typedef struct DynamicAudioNormalizerContext {
56
    const AVClass *class;
57
58
    struct FFBufQueue queue;
59
60
    int frame_len;
61
    int frame_len_msec;
62
    int filter_size;
63
    int dc_correction;
64
    int channels_coupled;
65
    int alt_boundary_mode;
66
67
    double peak_value;
68
    double max_amplification;
69
    double target_rms;
70
    double compress_factor;
71
    double threshold;
72
    double *prev_amplification_factor;
73
    double *dc_correction_value;
74
    double *compress_threshold;
75
    double *weights;
76
77
    int channels;
78
    int eof;
79
    int64_t pts;
80
81
    cqueue **gain_history_original;
82
    cqueue **gain_history_minimum;
83
    cqueue **gain_history_smoothed;
84
    cqueue **threshold_history;
85
86
    cqueue *is_enabled;
87
} DynamicAudioNormalizerContext;
88
89
#define OFFSET(x) offsetof(DynamicAudioNormalizerContext, x)
90
#define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
91
92
static const AVOption dynaudnorm_options[] = {
93
    { "framelen",    "set the frame length in msec",     OFFSET(frame_len_msec),    AV_OPT_TYPE_INT,    {.i64 = 500},   10,  8000, FLAGS },
94
    { "f",           "set the frame length in msec",     OFFSET(frame_len_msec),    AV_OPT_TYPE_INT,    {.i64 = 500},   10,  8000, FLAGS },
95
    { "gausssize",   "set the filter size",              OFFSET(filter_size),       AV_OPT_TYPE_INT,    {.i64 = 31},     3,   301, FLAGS },
96
    { "g",           "set the filter size",              OFFSET(filter_size),       AV_OPT_TYPE_INT,    {.i64 = 31},     3,   301, FLAGS },
97
    { "peak",        "set the peak value",               OFFSET(peak_value),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.95}, 0.0,   1.0, FLAGS },
98
    { "p",           "set the peak value",               OFFSET(peak_value),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.95}, 0.0,   1.0, FLAGS },
99
    { "maxgain",     "set the max amplification",        OFFSET(max_amplification), AV_OPT_TYPE_DOUBLE, {.dbl = 10.0}, 1.0, 100.0, FLAGS },
100
    { "m",           "set the max amplification",        OFFSET(max_amplification), AV_OPT_TYPE_DOUBLE, {.dbl = 10.0}, 1.0, 100.0, FLAGS },
101
    { "targetrms",   "set the target RMS",               OFFSET(target_rms),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,   1.0, FLAGS },
102
    { "r",           "set the target RMS",               OFFSET(target_rms),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,   1.0, FLAGS },
103
    { "coupling",    "set channel coupling",             OFFSET(channels_coupled),  AV_OPT_TYPE_BOOL,   {.i64 = 1},      0,     1, FLAGS },
104
    { "n",           "set channel coupling",             OFFSET(channels_coupled),  AV_OPT_TYPE_BOOL,   {.i64 = 1},      0,     1, FLAGS },
105
    { "correctdc",   "set DC correction",                OFFSET(dc_correction),     AV_OPT_TYPE_BOOL,   {.i64 = 0},      0,     1, FLAGS },
106
    { "c",           "set DC correction",                OFFSET(dc_correction),     AV_OPT_TYPE_BOOL,   {.i64 = 0},      0,     1, FLAGS },
107
    { "altboundary", "set alternative boundary mode",    OFFSET(alt_boundary_mode), AV_OPT_TYPE_BOOL,   {.i64 = 0},      0,     1, FLAGS },
108
    { "b",           "set alternative boundary mode",    OFFSET(alt_boundary_mode), AV_OPT_TYPE_BOOL,   {.i64 = 0},      0,     1, FLAGS },
109
    { "compress",    "set the compress factor",          OFFSET(compress_factor),   AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,  30.0, FLAGS },
110
    { "s",           "set the compress factor",          OFFSET(compress_factor),   AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,  30.0, FLAGS },
111
    { "threshold",   "set the threshold value",          OFFSET(threshold),         AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,   1.0, FLAGS },
112
    { "t",           "set the threshold value",          OFFSET(threshold),         AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,   1.0, FLAGS },
113
    { NULL }
114
};
115
116
AVFILTER_DEFINE_CLASS(dynaudnorm);
117
118
static av_cold int init(AVFilterContext *ctx)
119
{
120
    DynamicAudioNormalizerContext *s = ctx->priv;
121
122
    if (!(s->filter_size & 1)) {
123
        av_log(ctx, AV_LOG_WARNING, "filter size %d is invalid. Changing to an odd value.\n", s->filter_size);
124
        s->filter_size |= 1;
125
    }
126
127
    return 0;
128
}
129
130
static int query_formats(AVFilterContext *ctx)
131
{
132
    AVFilterFormats *formats;
133
    AVFilterChannelLayouts *layouts;
134
    static const enum AVSampleFormat sample_fmts[] = {
135
        AV_SAMPLE_FMT_DBLP,
136
        AV_SAMPLE_FMT_NONE
137
    };
138
    int ret;
139
140
    layouts = ff_all_channel_counts();
141
    if (!layouts)
142
        return AVERROR(ENOMEM);
143
    ret = ff_set_common_channel_layouts(ctx, layouts);
144
    if (ret < 0)
145
        return ret;
146
147
    formats = ff_make_format_list(sample_fmts);
148
    if (!formats)
149
        return AVERROR(ENOMEM);
150
    ret = ff_set_common_formats(ctx, formats);
151
    if (ret < 0)
152
        return ret;
153
154
    formats = ff_all_samplerates();
155
    if (!formats)
156
        return AVERROR(ENOMEM);
157
    return ff_set_common_samplerates(ctx, formats);
158
}
159
160
static inline int frame_size(int sample_rate, int frame_len_msec)
161
{
162
    const int frame_size = lrint((double)sample_rate * (frame_len_msec / 1000.0));
163
    return frame_size + (frame_size % 2);
164
}
165
166
static cqueue *cqueue_create(int size, int max_size)
167
{
168
    cqueue *q;
169
170
    if (max_size < size)
171
        return NULL;
172
173
    q = av_malloc(sizeof(cqueue));
174
    if (!q)
175
        return NULL;
176
177
    q->max_size = max_size;
178
    q->size = size;
179
    q->nb_elements = 0;
180
181
    q->elements = av_malloc_array(max_size, sizeof(double));
182
    if (!q->elements) {
183
        av_free(q);
184
        return NULL;
185
    }
186
187
    return q;
188
}
189
190
static void cqueue_free(cqueue *q)
191
{
192
    if (q)
193
        av_free(q->elements);
194
    av_free(q);
195
}
196
197
static int cqueue_size(cqueue *q)
198
{
199
    return q->nb_elements;
200
}
201
202
static int cqueue_empty(cqueue *q)
203
{
204
    return q->nb_elements <= 0;
205
}
206
207
static int cqueue_enqueue(cqueue *q, double element)
208
{
209
    av_assert2(q->nb_elements < q->max_size);
210
211
    q->elements[q->nb_elements] = element;
212
    q->nb_elements++;
213
214
    return 0;
215
}
216
217
static double cqueue_peek(cqueue *q, int index)
218
{
219
    av_assert2(index < q->nb_elements);
220
    return q->elements[index];
221
}
222
223
static int cqueue_dequeue(cqueue *q, double *element)
224
{
225
    av_assert2(!cqueue_empty(q));
226
227
    *element = q->elements[0];
228
    memmove(&q->elements[0], &q->elements[1], (q->nb_elements - 1) * sizeof(double));
229
    q->nb_elements--;
230
231
    return 0;
232
}
233
234
static int cqueue_pop(cqueue *q)
235
{
236
    av_assert2(!cqueue_empty(q));
237
238
    memmove(&q->elements[0], &q->elements[1], (q->nb_elements - 1) * sizeof(double));
239
    q->nb_elements--;
240
241
    return 0;
242
}
243
244
static void cqueue_resize(cqueue *q, int new_size)
245
{
246
    av_assert2(q->max_size >= new_size);
247
    av_assert2(MIN_FILTER_SIZE <= new_size);
248
249
    if (new_size > q->nb_elements) {
250
        const int side = (new_size - q->nb_elements) / 2;
251
252
        memmove(q->elements + side, q->elements, sizeof(double) * q->nb_elements);
253
        for (int i = 0; i < side; i++)
254
            q->elements[i] = q->elements[side];
255
        q->nb_elements = new_size - 1 - side;
256
    } else {
257
        int count = (q->size - new_size + 1) / 2;
258
259
        while (count-- > 0)
260
            cqueue_pop(q);
261
    }
262
263
    q->size = new_size;
264
}
265
266
static void init_gaussian_filter(DynamicAudioNormalizerContext *s)
267
{
268
    double total_weight = 0.0;
269
    const double sigma = (((s->filter_size / 2.0) - 1.0) / 3.0) + (1.0 / 3.0);
270
    double adjust;
271
    int i;
272
273
    // Pre-compute constants
274
    const int offset = s->filter_size / 2;
275
    const double c1 = 1.0 / (sigma * sqrt(2.0 * M_PI));
276
    const double c2 = 2.0 * sigma * sigma;
277
278
    // Compute weights
279
    for (i = 0; i < s->filter_size; i++) {
280
        const int x = i - offset;
281
282
        s->weights[i] = c1 * exp(-x * x / c2);
283
        total_weight += s->weights[i];
284
    }
285
286
    // Adjust weights
287
    adjust = 1.0 / total_weight;
288
    for (i = 0; i < s->filter_size; i++) {
289
        s->weights[i] *= adjust;
290
    }
291
}
292
293
static av_cold void uninit(AVFilterContext *ctx)
294
{
295
    DynamicAudioNormalizerContext *s = ctx->priv;
296
    int c;
297
298
    av_freep(&s->prev_amplification_factor);
299
    av_freep(&s->dc_correction_value);
300
    av_freep(&s->compress_threshold);
301
302
    for (c = 0; c < s->channels; c++) {
303
        if (s->gain_history_original)
304
            cqueue_free(s->gain_history_original[c]);
305
        if (s->gain_history_minimum)
306
            cqueue_free(s->gain_history_minimum[c]);
307
        if (s->gain_history_smoothed)
308
            cqueue_free(s->gain_history_smoothed[c]);
309
        if (s->threshold_history)
310
            cqueue_free(s->threshold_history[c]);
311
    }
312
313
    av_freep(&s->gain_history_original);
314
    av_freep(&s->gain_history_minimum);
315
    av_freep(&s->gain_history_smoothed);
316
    av_freep(&s->threshold_history);
317
318
    cqueue_free(s->is_enabled);
319
    s->is_enabled = NULL;
320
321
    av_freep(&s->weights);
322
323
    ff_bufqueue_discard_all(&s->queue);
324
}
325
326
static int config_input(AVFilterLink *inlink)
327
{
328
    AVFilterContext *ctx = inlink->dst;
329
    DynamicAudioNormalizerContext *s = ctx->priv;
330
    int c;
331
332
    uninit(ctx);
333
334
    s->channels = inlink->channels;
335
    s->frame_len = frame_size(inlink->sample_rate, s->frame_len_msec);
336
    av_log(ctx, AV_LOG_DEBUG, "frame len %d\n", s->frame_len);
337
338
    s->prev_amplification_factor = av_malloc_array(inlink->channels, sizeof(*s->prev_amplification_factor));
339
    s->dc_correction_value = av_calloc(inlink->channels, sizeof(*s->dc_correction_value));
340
    s->compress_threshold = av_calloc(inlink->channels, sizeof(*s->compress_threshold));
341
    s->gain_history_original = av_calloc(inlink->channels, sizeof(*s->gain_history_original));
342
    s->gain_history_minimum = av_calloc(inlink->channels, sizeof(*s->gain_history_minimum));
343
    s->gain_history_smoothed = av_calloc(inlink->channels, sizeof(*s->gain_history_smoothed));
344
    s->threshold_history = av_calloc(inlink->channels, sizeof(*s->threshold_history));
345
    s->weights = av_malloc_array(MAX_FILTER_SIZE, sizeof(*s->weights));
346
    s->is_enabled = cqueue_create(s->filter_size, MAX_FILTER_SIZE);
347
    if (!s->prev_amplification_factor || !s->dc_correction_value ||
348
        !s->compress_threshold ||
349
        !s->gain_history_original || !s->gain_history_minimum ||
350
        !s->gain_history_smoothed || !s->threshold_history ||
351
        !s->is_enabled || !s->weights)
352
        return AVERROR(ENOMEM);
353
354
    for (c = 0; c < inlink->channels; c++) {
355
        s->prev_amplification_factor[c] = 1.0;
356
357
        s->gain_history_original[c] = cqueue_create(s->filter_size, MAX_FILTER_SIZE);
358
        s->gain_history_minimum[c]  = cqueue_create(s->filter_size, MAX_FILTER_SIZE);
359
        s->gain_history_smoothed[c] = cqueue_create(s->filter_size, MAX_FILTER_SIZE);
360
        s->threshold_history[c]     = cqueue_create(s->filter_size, MAX_FILTER_SIZE);
361
362
        if (!s->gain_history_original[c] || !s->gain_history_minimum[c] ||
363
            !s->gain_history_smoothed[c] || !s->threshold_history[c])
364
            return AVERROR(ENOMEM);
365
    }
366
367
    init_gaussian_filter(s);
368
369
    return 0;
370
}
371
372
static inline double fade(double prev, double next, int pos, int length)
373
{
374
    const double step_size = 1.0 / length;
375
    const double f0 = 1.0 - (step_size * (pos + 1.0));
376
    const double f1 = 1.0 - f0;
377
    return f0 * prev + f1 * next;
378
}
379
380
static inline double pow_2(const double value)
381
{
382
    return value * value;
383
}
384
385
static inline double bound(const double threshold, const double val)
386
{
387
    const double CONST = 0.8862269254527580136490837416705725913987747280611935; //sqrt(PI) / 2.0
388
    return erf(CONST * (val / threshold)) * threshold;
389
}
390
391
static double find_peak_magnitude(AVFrame *frame, int channel)
392
{
393
    double max = DBL_EPSILON;
394
    int c, i;
395
396
    if (channel == -1) {
397
        for (c = 0; c < frame->channels; c++) {
398
            double *data_ptr = (double *)frame->extended_data[c];
399
400
            for (i = 0; i < frame->nb_samples; i++)
401
                max = FFMAX(max, fabs(data_ptr[i]));
402
        }
403
    } else {
404
        double *data_ptr = (double *)frame->extended_data[channel];
405
406
        for (i = 0; i < frame->nb_samples; i++)
407
            max = FFMAX(max, fabs(data_ptr[i]));
408
    }
409
410
    return max;
411
}
412
413
static double compute_frame_rms(AVFrame *frame, int channel)
414
{
415
    double rms_value = 0.0;
416
    int c, i;
417
418
    if (channel == -1) {
419
        for (c = 0; c < frame->channels; c++) {
420
            const double *data_ptr = (double *)frame->extended_data[c];
421
422
            for (i = 0; i < frame->nb_samples; i++) {
423
                rms_value += pow_2(data_ptr[i]);
424
            }
425
        }
426
427
        rms_value /= frame->nb_samples * frame->channels;
428
    } else {
429
        const double *data_ptr = (double *)frame->extended_data[channel];
430
        for (i = 0; i < frame->nb_samples; i++) {
431
            rms_value += pow_2(data_ptr[i]);
432
        }
433
434
        rms_value /= frame->nb_samples;
435
    }
436
437
    return FFMAX(sqrt(rms_value), DBL_EPSILON);
438
}
439
440
static local_gain get_max_local_gain(DynamicAudioNormalizerContext *s, AVFrame *frame,
441
                                     int channel)
442
{
443
    const double peak_magnitude = find_peak_magnitude(frame, channel);
444
    const double maximum_gain = s->peak_value / peak_magnitude;
445
    const double rms_gain = s->target_rms > DBL_EPSILON ? (s->target_rms / compute_frame_rms(frame, channel)) : DBL_MAX;
446
    local_gain gain;
447
448
    gain.threshold = peak_magnitude > s->threshold;
449
    gain.max_gain  = bound(s->max_amplification, FFMIN(maximum_gain, rms_gain));
450
451
    return gain;
452
}
453
454
static double minimum_filter(cqueue *q)
455
{
456
    double min = DBL_MAX;
457
    int i;
458
459
    for (i = 0; i < cqueue_size(q); i++) {
460
        min = FFMIN(min, cqueue_peek(q, i));
461
    }
462
463
    return min;
464
}
465
466
static double gaussian_filter(DynamicAudioNormalizerContext *s, cqueue *q, cqueue *tq)
467
{
468
    double result = 0.0, tsum = 0.0;
469
    int i;
470
471
    for (i = 0; i < cqueue_size(q); i++) {
472
        tsum += cqueue_peek(tq, i) * s->weights[i];
473
        result += cqueue_peek(q, i) * s->weights[i] * cqueue_peek(tq, i);
474
    }
475
476
    if (tsum == 0.0)
477
        result = 1.0;
478
479
    return result;
480
}
481
482
static void update_gain_history(DynamicAudioNormalizerContext *s, int channel,
483
                                local_gain gain)
484
{
485
    if (cqueue_empty(s->gain_history_original[channel])) {
486
        const int pre_fill_size = s->filter_size / 2;
487
        const double initial_value = s->alt_boundary_mode ? gain.max_gain : s->peak_value;
488
489
        s->prev_amplification_factor[channel] = initial_value;
490
491
        while (cqueue_size(s->gain_history_original[channel]) < pre_fill_size) {
492
            cqueue_enqueue(s->gain_history_original[channel], initial_value);
493
            cqueue_enqueue(s->threshold_history[channel], gain.threshold);
494
        }
495
    }
496
497
    cqueue_enqueue(s->gain_history_original[channel], gain.max_gain);
498
499
    while (cqueue_size(s->gain_history_original[channel]) >= s->filter_size) {
500
        double minimum;
501
502
        if (cqueue_empty(s->gain_history_minimum[channel])) {
503
            const int pre_fill_size = s->filter_size / 2;
504
            double initial_value = s->alt_boundary_mode ? cqueue_peek(s->gain_history_original[channel], 0) : 1.0;
505
            int input = pre_fill_size;
506
507
            while (cqueue_size(s->gain_history_minimum[channel]) < pre_fill_size) {
508
                input++;
509
                initial_value = FFMIN(initial_value, cqueue_peek(s->gain_history_original[channel], input));
510
                cqueue_enqueue(s->gain_history_minimum[channel], initial_value);
511
            }
512
        }
513
514
        minimum = minimum_filter(s->gain_history_original[channel]);
515
516
        cqueue_enqueue(s->gain_history_minimum[channel], minimum);
517
518
        cqueue_enqueue(s->threshold_history[channel], gain.threshold);
519
520
        cqueue_pop(s->gain_history_original[channel]);
521
    }
522
523
    while (cqueue_size(s->gain_history_minimum[channel]) >= s->filter_size) {
524
        double smoothed, limit;
525
526
        smoothed = gaussian_filter(s, s->gain_history_minimum[channel], s->threshold_history[channel]);
527
        limit    = cqueue_peek(s->gain_history_original[channel], 0);
528
        smoothed = FFMIN(smoothed, limit);
529
530
        cqueue_enqueue(s->gain_history_smoothed[channel], smoothed);
531
532
        cqueue_pop(s->gain_history_minimum[channel]);
533
        cqueue_pop(s->threshold_history[channel]);
534
    }
535
}
536
537
static inline double update_value(double new, double old, double aggressiveness)
538
{
539
    av_assert0((aggressiveness >= 0.0) && (aggressiveness <= 1.0));
540
    return aggressiveness * new + (1.0 - aggressiveness) * old;
541
}
542
543
static void perform_dc_correction(DynamicAudioNormalizerContext *s, AVFrame *frame)
544
{
545
    const double diff = 1.0 / frame->nb_samples;
546
    int is_first_frame = cqueue_empty(s->gain_history_original[0]);
547
    int c, i;
548
549
    for (c = 0; c < s->channels; c++) {
550
        double *dst_ptr = (double *)frame->extended_data[c];
551
        double current_average_value = 0.0;
552
        double prev_value;
553
554
        for (i = 0; i < frame->nb_samples; i++)
555
            current_average_value += dst_ptr[i] * diff;
556
557
        prev_value = is_first_frame ? current_average_value : s->dc_correction_value[c];
558
        s->dc_correction_value[c] = is_first_frame ? current_average_value : update_value(current_average_value, s->dc_correction_value[c], 0.1);
559
560
        for (i = 0; i < frame->nb_samples; i++) {
561
            dst_ptr[i] -= fade(prev_value, s->dc_correction_value[c], i, frame->nb_samples);
562
        }
563
    }
564
}
565
566
static double setup_compress_thresh(double threshold)
567
{
568
    if ((threshold > DBL_EPSILON) && (threshold < (1.0 - DBL_EPSILON))) {
569
        double current_threshold = threshold;
570
        double step_size = 1.0;
571
572
        while (step_size > DBL_EPSILON) {
573
            while ((llrint((current_threshold + step_size) * (UINT64_C(1) << 63)) >
574
                    llrint(current_threshold * (UINT64_C(1) << 63))) &&
575
                   (bound(current_threshold + step_size, 1.0) <= threshold)) {
576
                current_threshold += step_size;
577
            }
578
579
            step_size /= 2.0;
580
        }
581
582
        return current_threshold;
583
    } else {
584
        return threshold;
585
    }
586
}
587
588
static double compute_frame_std_dev(DynamicAudioNormalizerContext *s,
589
                                    AVFrame *frame, int channel)
590
{
591
    double variance = 0.0;
592
    int i, c;
593
594
    if (channel == -1) {
595
        for (c = 0; c < s->channels; c++) {
596
            const double *data_ptr = (double *)frame->extended_data[c];
597
598
            for (i = 0; i < frame->nb_samples; i++) {
599
                variance += pow_2(data_ptr[i]);  // Assume that MEAN is *zero*
600
            }
601
        }
602
        variance /= (s->channels * frame->nb_samples) - 1;
603
    } else {
604
        const double *data_ptr = (double *)frame->extended_data[channel];
605
606
        for (i = 0; i < frame->nb_samples; i++) {
607
            variance += pow_2(data_ptr[i]);      // Assume that MEAN is *zero*
608
        }
609
        variance /= frame->nb_samples - 1;
610
    }
611
612
    return FFMAX(sqrt(variance), DBL_EPSILON);
613
}
614
615
static void perform_compression(DynamicAudioNormalizerContext *s, AVFrame *frame)
616
{
617
    int is_first_frame = cqueue_empty(s->gain_history_original[0]);
618
    int c, i;
619
620
    if (s->channels_coupled) {
621
        const double standard_deviation = compute_frame_std_dev(s, frame, -1);
622
        const double current_threshold  = FFMIN(1.0, s->compress_factor * standard_deviation);
623
624
        const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[0];
625
        double prev_actual_thresh, curr_actual_thresh;
626
        s->compress_threshold[0] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[0], (1.0/3.0));
627
628
        prev_actual_thresh = setup_compress_thresh(prev_value);
629
        curr_actual_thresh = setup_compress_thresh(s->compress_threshold[0]);
630
631
        for (c = 0; c < s->channels; c++) {
632
            double *const dst_ptr = (double *)frame->extended_data[c];
633
            for (i = 0; i < frame->nb_samples; i++) {
634
                const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, frame->nb_samples);
635
                dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
636
            }
637
        }
638
    } else {
639
        for (c = 0; c < s->channels; c++) {
640
            const double standard_deviation = compute_frame_std_dev(s, frame, c);
641
            const double current_threshold  = setup_compress_thresh(FFMIN(1.0, s->compress_factor * standard_deviation));
642
643
            const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[c];
644
            double prev_actual_thresh, curr_actual_thresh;
645
            double *dst_ptr;
646
            s->compress_threshold[c] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[c], 1.0/3.0);
647
648
            prev_actual_thresh = setup_compress_thresh(prev_value);
649
            curr_actual_thresh = setup_compress_thresh(s->compress_threshold[c]);
650
651
            dst_ptr = (double *)frame->extended_data[c];
652
            for (i = 0; i < frame->nb_samples; i++) {
653
                const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, frame->nb_samples);
654
                dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
655
            }
656
        }
657
    }
658
}
659
660
static void analyze_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
661
{
662
    if (s->dc_correction) {
663
        perform_dc_correction(s, frame);
664
    }
665
666
    if (s->compress_factor > DBL_EPSILON) {
667
        perform_compression(s, frame);
668
    }
669
670
    if (s->channels_coupled) {
671
        const local_gain gain = get_max_local_gain(s, frame, -1);
672
        int c;
673
674
        for (c = 0; c < s->channels; c++)
675
            update_gain_history(s, c, gain);
676
    } else {
677
        int c;
678
679
        for (c = 0; c < s->channels; c++)
680
            update_gain_history(s, c, get_max_local_gain(s, frame, c));
681
    }
682
}
683
684
static void amplify_frame(DynamicAudioNormalizerContext *s, AVFrame *frame, int enabled)
685
{
686
    int c, i;
687
688
    for (c = 0; c < s->channels; c++) {
689
        double *dst_ptr = (double *)frame->extended_data[c];
690
        double current_amplification_factor;
691
692
        cqueue_dequeue(s->gain_history_smoothed[c], &current_amplification_factor);
693
694
        for (i = 0; i < frame->nb_samples && enabled; i++) {
695
            const double amplification_factor = fade(s->prev_amplification_factor[c],
696
                                                     current_amplification_factor, i,
697
                                                     frame->nb_samples);
698
699
            dst_ptr[i] *= amplification_factor;
700
        }
701
702
        s->prev_amplification_factor[c] = current_amplification_factor;
703
    }
704
}
705
706
static int filter_frame(AVFilterLink *inlink, AVFrame *in)
707
{
708
    AVFilterContext *ctx = inlink->dst;
709
    DynamicAudioNormalizerContext *s = ctx->priv;
710
    AVFilterLink *outlink = ctx->outputs[0];
711
    int ret = 1;
712
713
    while (((s->queue.available >= s->filter_size) ||
714
            (s->eof && s->queue.available)) &&
715
           !cqueue_empty(s->gain_history_smoothed[0])) {
716
        AVFrame *out = ff_bufqueue_get(&s->queue);
717
        double is_enabled;
718
719
        cqueue_dequeue(s->is_enabled, &is_enabled);
720
721
        amplify_frame(s, out, is_enabled > 0.);
722
        ret = ff_filter_frame(outlink, out);
723
    }
724
725
    av_frame_make_writable(in);
726
    analyze_frame(s, in);
727
    if (!s->eof) {
728
        ff_bufqueue_add(ctx, &s->queue, in);
729
        cqueue_enqueue(s->is_enabled, !ctx->is_disabled);
730
    } else {
731
        av_frame_free(&in);
732
    }
733
734
    return ret;
735
}
736
737
static int flush_buffer(DynamicAudioNormalizerContext *s, AVFilterLink *inlink,
738
                        AVFilterLink *outlink)
739
{
740
    AVFrame *out = ff_get_audio_buffer(outlink, s->frame_len);
741
    int c, i;
742
743
    if (!out)
744
        return AVERROR(ENOMEM);
745
746
    for (c = 0; c < s->channels; c++) {
747
        double *dst_ptr = (double *)out->extended_data[c];
748
749
        for (i = 0; i < out->nb_samples; i++) {
750
            dst_ptr[i] = s->alt_boundary_mode ? DBL_EPSILON : ((s->target_rms > DBL_EPSILON) ? FFMIN(s->peak_value, s->target_rms) : s->peak_value);
751
            if (s->dc_correction) {
752
                dst_ptr[i] *= ((i % 2) == 1) ? -1 : 1;
753
                dst_ptr[i] += s->dc_correction_value[c];
754
            }
755
        }
756
    }
757
758
    return filter_frame(inlink, out);
759
}
760
761
static int flush(AVFilterLink *outlink)
762
{
763
    AVFilterContext *ctx = outlink->src;
764
    DynamicAudioNormalizerContext *s = ctx->priv;
765
    int ret = 0;
766
767
    if (!cqueue_empty(s->gain_history_smoothed[0])) {
768
        ret = flush_buffer(s, ctx->inputs[0], outlink);
769
    } else if (s->queue.available) {
770
        AVFrame *out = ff_bufqueue_get(&s->queue);
771
772
        s->pts = out->pts;
773
        ret = ff_filter_frame(outlink, out);
774
    }
775
776
    return ret;
777
}
778
779
static int activate(AVFilterContext *ctx)
780
{
781
    AVFilterLink *inlink = ctx->inputs[0];
782
    AVFilterLink *outlink = ctx->outputs[0];
783
    DynamicAudioNormalizerContext *s = ctx->priv;
784
    AVFrame *in = NULL;
785
    int ret = 0, status;
786
    int64_t pts;
787
788
    FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
789
790
    if (!s->eof) {
791
        ret = ff_inlink_consume_samples(inlink, s->frame_len, s->frame_len, &in);
792
        if (ret < 0)
793
            return ret;
794
        if (ret > 0) {
795
            ret = filter_frame(inlink, in);
796
            if (ret <= 0)
797
                return ret;
798
        }
799
800
        if (ff_inlink_queued_samples(inlink) >= s->frame_len) {
801
            ff_filter_set_ready(ctx, 10);
802
            return 0;
803
        }
804
    }
805
806
    if (!s->eof && ff_inlink_acknowledge_status(inlink, &status, &pts)) {
807
        if (status == AVERROR_EOF)
808
            s->eof = 1;
809
    }
810
811
    if (s->eof && s->queue.available)
812
        return flush(outlink);
813
814
    if (s->eof && !s->queue.available) {
815
        ff_outlink_set_status(outlink, AVERROR_EOF, s->pts);
816
        return 0;
817
    }
818
819
    if (!s->eof)
820
        FF_FILTER_FORWARD_WANTED(outlink, inlink);
821
822
    return FFERROR_NOT_READY;
823
}
824
825
static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
826
                           char *res, int res_len, int flags)
827
{
828
    DynamicAudioNormalizerContext *s = ctx->priv;
829
    AVFilterLink *inlink = ctx->inputs[0];
830
    int prev_filter_size = s->filter_size;
831
    int ret;
832
833
    ret = ff_filter_process_command(ctx, cmd, args, res, res_len, flags);
834
    if (ret < 0)
835
        return ret;
836
837
    s->filter_size |= 1;
838
    if (prev_filter_size != s->filter_size) {
839
        init_gaussian_filter(s);
840
841
        for (int c = 0; c < s->channels; c++) {
842
            cqueue_resize(s->gain_history_original[c], s->filter_size);
843
            cqueue_resize(s->gain_history_minimum[c], s->filter_size);
844
            cqueue_resize(s->threshold_history[c], s->filter_size);
845
        }
846
    }
847
848
    s->frame_len = frame_size(inlink->sample_rate, s->frame_len_msec);
849
850
    return 0;
851
}
852
853
static const AVFilterPad avfilter_af_dynaudnorm_inputs[] = {
854
    {
855
        .name           = "default",
856
        .type           = AVMEDIA_TYPE_AUDIO,
857
        .config_props   = config_input,
858
    },
859
    { NULL }
860
};
861
862
static const AVFilterPad avfilter_af_dynaudnorm_outputs[] = {
863
    {
864
        .name          = "default",
865
        .type          = AVMEDIA_TYPE_AUDIO,
866
    },
867
    { NULL }
868
};
869
870
AVFilter ff_af_dynaudnorm = {
871
    .name          = "dynaudnorm",
872
    .description   = NULL_IF_CONFIG_SMALL("Dynamic Audio Normalizer."),
873
    .query_formats = query_formats,
874
    .priv_size     = sizeof(DynamicAudioNormalizerContext),
875
    .init          = init,
876
    .uninit        = uninit,
877
    .activate      = activate,
878
    .inputs        = avfilter_af_dynaudnorm_inputs,
879
    .outputs       = avfilter_af_dynaudnorm_outputs,
880
    .priv_class    = &dynaudnorm_class,
881
    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,
882
    .process_command = process_command,
883
};