git.sesse.net Git - ffmpeg/blob - libavfilter/af_dynaudnorm.c

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