GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavfilter/af_atempo.c Lines: 0 422 0.0 %
Date: 2020-08-14 10:39:37 Branches: 0 300 0.0 %

Line Branch Exec Source
1
/*
2
 * Copyright (c) 2012 Pavel Koshevoy <pkoshevoy at gmail dot com>
3
 *
4
 * This file is part of FFmpeg.
5
 *
6
 * FFmpeg is free software; you can redistribute it and/or
7
 * modify it under the terms of the GNU Lesser General Public
8
 * License as published by the Free Software Foundation; either
9
 * version 2.1 of the License, or (at your option) any later version.
10
 *
11
 * FFmpeg is distributed in the hope that it will be useful,
12
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14
 * Lesser General Public License for more details.
15
 *
16
 * You should have received a copy of the GNU Lesser General Public
17
 * License along with FFmpeg; if not, write to the Free Software
18
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19
 */
20
21
/**
22
 * @file
23
 * tempo scaling audio filter -- an implementation of WSOLA algorithm
24
 *
25
 * Based on MIT licensed yaeAudioTempoFilter.h and yaeAudioFragment.h
26
 * from Apprentice Video player by Pavel Koshevoy.
27
 * https://sourceforge.net/projects/apprenticevideo/
28
 *
29
 * An explanation of SOLA algorithm is available at
30
 * http://www.surina.net/article/time-and-pitch-scaling.html
31
 *
32
 * WSOLA is very similar to SOLA, only one major difference exists between
33
 * these algorithms.  SOLA shifts audio fragments along the output stream,
34
 * where as WSOLA shifts audio fragments along the input stream.
35
 *
36
 * The advantage of WSOLA algorithm is that the overlap region size is
37
 * always the same, therefore the blending function is constant and
38
 * can be precomputed.
39
 */
40
41
#include <float.h>
42
#include "libavcodec/avfft.h"
43
#include "libavutil/avassert.h"
44
#include "libavutil/avstring.h"
45
#include "libavutil/channel_layout.h"
46
#include "libavutil/eval.h"
47
#include "libavutil/opt.h"
48
#include "libavutil/samplefmt.h"
49
#include "avfilter.h"
50
#include "audio.h"
51
#include "internal.h"
52
53
/**
54
 * A fragment of audio waveform
55
 */
56
typedef struct AudioFragment {
57
    // index of the first sample of this fragment in the overall waveform;
58
    // 0: input sample position
59
    // 1: output sample position
60
    int64_t position[2];
61
62
    // original packed multi-channel samples:
63
    uint8_t *data;
64
65
    // number of samples in this fragment:
66
    int nsamples;
67
68
    // rDFT transform of the down-mixed mono fragment, used for
69
    // fast waveform alignment via correlation in frequency domain:
70
    FFTSample *xdat;
71
} AudioFragment;
72
73
/**
74
 * Filter state machine states
75
 */
76
typedef enum {
77
    YAE_LOAD_FRAGMENT,
78
    YAE_ADJUST_POSITION,
79
    YAE_RELOAD_FRAGMENT,
80
    YAE_OUTPUT_OVERLAP_ADD,
81
    YAE_FLUSH_OUTPUT,
82
} FilterState;
83
84
/**
85
 * Filter state machine
86
 */
87
typedef struct ATempoContext {
88
    const AVClass *class;
89
90
    // ring-buffer of input samples, necessary because some times
91
    // input fragment position may be adjusted backwards:
92
    uint8_t *buffer;
93
94
    // ring-buffer maximum capacity, expressed in sample rate time base:
95
    int ring;
96
97
    // ring-buffer house keeping:
98
    int size;
99
    int head;
100
    int tail;
101
102
    // 0: input sample position corresponding to the ring buffer tail
103
    // 1: output sample position
104
    int64_t position[2];
105
106
    // first input timestamp, all other timestamps are offset by this one
107
    int64_t start_pts;
108
109
    // sample format:
110
    enum AVSampleFormat format;
111
112
    // number of channels:
113
    int channels;
114
115
    // row of bytes to skip from one sample to next, across multple channels;
116
    // stride = (number-of-channels * bits-per-sample-per-channel) / 8
117
    int stride;
118
119
    // fragment window size, power-of-two integer:
120
    int window;
121
122
    // Hann window coefficients, for feathering
123
    // (blending) the overlapping fragment region:
124
    float *hann;
125
126
    // tempo scaling factor:
127
    double tempo;
128
129
    // a snapshot of previous fragment input and output position values
130
    // captured when the tempo scale factor was set most recently:
131
    int64_t origin[2];
132
133
    // current/previous fragment ring-buffer:
134
    AudioFragment frag[2];
135
136
    // current fragment index:
137
    uint64_t nfrag;
138
139
    // current state:
140
    FilterState state;
141
142
    // for fast correlation calculation in frequency domain:
143
    RDFTContext *real_to_complex;
144
    RDFTContext *complex_to_real;
145
    FFTSample *correlation;
146
147
    // for managing AVFilterPad.request_frame and AVFilterPad.filter_frame
148
    AVFrame *dst_buffer;
149
    uint8_t *dst;
150
    uint8_t *dst_end;
151
    uint64_t nsamples_in;
152
    uint64_t nsamples_out;
153
} ATempoContext;
154
155
#define YAE_ATEMPO_MIN 0.5
156
#define YAE_ATEMPO_MAX 100.0
157
158
#define OFFSET(x) offsetof(ATempoContext, x)
159
160
static const AVOption atempo_options[] = {
161
    { "tempo", "set tempo scale factor",
162
      OFFSET(tempo), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 },
163
      YAE_ATEMPO_MIN,
164
      YAE_ATEMPO_MAX,
165
      AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_RUNTIME_PARAM },
166
    { NULL }
167
};
168
169
AVFILTER_DEFINE_CLASS(atempo);
170
171
inline static AudioFragment *yae_curr_frag(ATempoContext *atempo)
172
{
173
    return &atempo->frag[atempo->nfrag % 2];
174
}
175
176
inline static AudioFragment *yae_prev_frag(ATempoContext *atempo)
177
{
178
    return &atempo->frag[(atempo->nfrag + 1) % 2];
179
}
180
181
/**
182
 * Reset filter to initial state, do not deallocate existing local buffers.
183
 */
184
static void yae_clear(ATempoContext *atempo)
185
{
186
    atempo->size = 0;
187
    atempo->head = 0;
188
    atempo->tail = 0;
189
190
    atempo->nfrag = 0;
191
    atempo->state = YAE_LOAD_FRAGMENT;
192
    atempo->start_pts = AV_NOPTS_VALUE;
193
194
    atempo->position[0] = 0;
195
    atempo->position[1] = 0;
196
197
    atempo->origin[0] = 0;
198
    atempo->origin[1] = 0;
199
200
    atempo->frag[0].position[0] = 0;
201
    atempo->frag[0].position[1] = 0;
202
    atempo->frag[0].nsamples    = 0;
203
204
    atempo->frag[1].position[0] = 0;
205
    atempo->frag[1].position[1] = 0;
206
    atempo->frag[1].nsamples    = 0;
207
208
    // shift left position of 1st fragment by half a window
209
    // so that no re-normalization would be required for
210
    // the left half of the 1st fragment:
211
    atempo->frag[0].position[0] = -(int64_t)(atempo->window / 2);
212
    atempo->frag[0].position[1] = -(int64_t)(atempo->window / 2);
213
214
    av_frame_free(&atempo->dst_buffer);
215
    atempo->dst     = NULL;
216
    atempo->dst_end = NULL;
217
218
    atempo->nsamples_in       = 0;
219
    atempo->nsamples_out      = 0;
220
}
221
222
/**
223
 * Reset filter to initial state and deallocate all buffers.
224
 */
225
static void yae_release_buffers(ATempoContext *atempo)
226
{
227
    yae_clear(atempo);
228
229
    av_freep(&atempo->frag[0].data);
230
    av_freep(&atempo->frag[1].data);
231
    av_freep(&atempo->frag[0].xdat);
232
    av_freep(&atempo->frag[1].xdat);
233
234
    av_freep(&atempo->buffer);
235
    av_freep(&atempo->hann);
236
    av_freep(&atempo->correlation);
237
238
    av_rdft_end(atempo->real_to_complex);
239
    atempo->real_to_complex = NULL;
240
241
    av_rdft_end(atempo->complex_to_real);
242
    atempo->complex_to_real = NULL;
243
}
244
245
/* av_realloc is not aligned enough; fortunately, the data does not need to
246
 * be preserved */
247
#define RE_MALLOC_OR_FAIL(field, field_size)                    \
248
    do {                                                        \
249
        av_freep(&field);                                       \
250
        field = av_malloc(field_size);                          \
251
        if (!field) {                                           \
252
            yae_release_buffers(atempo);                        \
253
            return AVERROR(ENOMEM);                             \
254
        }                                                       \
255
    } while (0)
256
257
/**
258
 * Prepare filter for processing audio data of given format,
259
 * sample rate and number of channels.
260
 */
261
static int yae_reset(ATempoContext *atempo,
262
                     enum AVSampleFormat format,
263
                     int sample_rate,
264
                     int channels)
265
{
266
    const int sample_size = av_get_bytes_per_sample(format);
267
    uint32_t nlevels  = 0;
268
    uint32_t pot;
269
    int i;
270
271
    atempo->format   = format;
272
    atempo->channels = channels;
273
    atempo->stride   = sample_size * channels;
274
275
    // pick a segment window size:
276
    atempo->window = sample_rate / 24;
277
278
    // adjust window size to be a power-of-two integer:
279
    nlevels = av_log2(atempo->window);
280
    pot = 1 << nlevels;
281
    av_assert0(pot <= atempo->window);
282
283
    if (pot < atempo->window) {
284
        atempo->window = pot * 2;
285
        nlevels++;
286
    }
287
288
    // initialize audio fragment buffers:
289
    RE_MALLOC_OR_FAIL(atempo->frag[0].data, atempo->window * atempo->stride);
290
    RE_MALLOC_OR_FAIL(atempo->frag[1].data, atempo->window * atempo->stride);
291
    RE_MALLOC_OR_FAIL(atempo->frag[0].xdat, atempo->window * sizeof(FFTComplex));
292
    RE_MALLOC_OR_FAIL(atempo->frag[1].xdat, atempo->window * sizeof(FFTComplex));
293
294
    // initialize rDFT contexts:
295
    av_rdft_end(atempo->real_to_complex);
296
    atempo->real_to_complex = NULL;
297
298
    av_rdft_end(atempo->complex_to_real);
299
    atempo->complex_to_real = NULL;
300
301
    atempo->real_to_complex = av_rdft_init(nlevels + 1, DFT_R2C);
302
    if (!atempo->real_to_complex) {
303
        yae_release_buffers(atempo);
304
        return AVERROR(ENOMEM);
305
    }
306
307
    atempo->complex_to_real = av_rdft_init(nlevels + 1, IDFT_C2R);
308
    if (!atempo->complex_to_real) {
309
        yae_release_buffers(atempo);
310
        return AVERROR(ENOMEM);
311
    }
312
313
    RE_MALLOC_OR_FAIL(atempo->correlation, atempo->window * sizeof(FFTComplex));
314
315
    atempo->ring = atempo->window * 3;
316
    RE_MALLOC_OR_FAIL(atempo->buffer, atempo->ring * atempo->stride);
317
318
    // initialize the Hann window function:
319
    RE_MALLOC_OR_FAIL(atempo->hann, atempo->window * sizeof(float));
320
321
    for (i = 0; i < atempo->window; i++) {
322
        double t = (double)i / (double)(atempo->window - 1);
323
        double h = 0.5 * (1.0 - cos(2.0 * M_PI * t));
324
        atempo->hann[i] = (float)h;
325
    }
326
327
    yae_clear(atempo);
328
    return 0;
329
}
330
331
static int yae_update(AVFilterContext *ctx)
332
{
333
    const AudioFragment *prev;
334
    ATempoContext *atempo = ctx->priv;
335
336
    prev = yae_prev_frag(atempo);
337
    atempo->origin[0] = prev->position[0] + atempo->window / 2;
338
    atempo->origin[1] = prev->position[1] + atempo->window / 2;
339
    return 0;
340
}
341
342
/**
343
 * A helper macro for initializing complex data buffer with scalar data
344
 * of a given type.
345
 */
346
#define yae_init_xdat(scalar_type, scalar_max)                          \
347
    do {                                                                \
348
        const uint8_t *src_end = src +                                  \
349
            frag->nsamples * atempo->channels * sizeof(scalar_type);    \
350
                                                                        \
351
        FFTSample *xdat = frag->xdat;                                   \
352
        scalar_type tmp;                                                \
353
                                                                        \
354
        if (atempo->channels == 1) {                                    \
355
            for (; src < src_end; xdat++) {                             \
356
                tmp = *(const scalar_type *)src;                        \
357
                src += sizeof(scalar_type);                             \
358
                                                                        \
359
                *xdat = (FFTSample)tmp;                                 \
360
            }                                                           \
361
        } else {                                                        \
362
            FFTSample s, max, ti, si;                                   \
363
            int i;                                                      \
364
                                                                        \
365
            for (; src < src_end; xdat++) {                             \
366
                tmp = *(const scalar_type *)src;                        \
367
                src += sizeof(scalar_type);                             \
368
                                                                        \
369
                max = (FFTSample)tmp;                                   \
370
                s = FFMIN((FFTSample)scalar_max,                        \
371
                          (FFTSample)fabsf(max));                       \
372
                                                                        \
373
                for (i = 1; i < atempo->channels; i++) {                \
374
                    tmp = *(const scalar_type *)src;                    \
375
                    src += sizeof(scalar_type);                         \
376
                                                                        \
377
                    ti = (FFTSample)tmp;                                \
378
                    si = FFMIN((FFTSample)scalar_max,                   \
379
                               (FFTSample)fabsf(ti));                   \
380
                                                                        \
381
                    if (s < si) {                                       \
382
                        s   = si;                                       \
383
                        max = ti;                                       \
384
                    }                                                   \
385
                }                                                       \
386
                                                                        \
387
                *xdat = max;                                            \
388
            }                                                           \
389
        }                                                               \
390
    } while (0)
391
392
/**
393
 * Initialize complex data buffer of a given audio fragment
394
 * with down-mixed mono data of appropriate scalar type.
395
 */
396
static void yae_downmix(ATempoContext *atempo, AudioFragment *frag)
397
{
398
    // shortcuts:
399
    const uint8_t *src = frag->data;
400
401
    // init complex data buffer used for FFT and Correlation:
402
    memset(frag->xdat, 0, sizeof(FFTComplex) * atempo->window);
403
404
    if (atempo->format == AV_SAMPLE_FMT_U8) {
405
        yae_init_xdat(uint8_t, 127);
406
    } else if (atempo->format == AV_SAMPLE_FMT_S16) {
407
        yae_init_xdat(int16_t, 32767);
408
    } else if (atempo->format == AV_SAMPLE_FMT_S32) {
409
        yae_init_xdat(int, 2147483647);
410
    } else if (atempo->format == AV_SAMPLE_FMT_FLT) {
411
        yae_init_xdat(float, 1);
412
    } else if (atempo->format == AV_SAMPLE_FMT_DBL) {
413
        yae_init_xdat(double, 1);
414
    }
415
}
416
417
/**
418
 * Populate the internal data buffer on as-needed basis.
419
 *
420
 * @return
421
 *   0 if requested data was already available or was successfully loaded,
422
 *   AVERROR(EAGAIN) if more input data is required.
423
 */
424
static int yae_load_data(ATempoContext *atempo,
425
                         const uint8_t **src_ref,
426
                         const uint8_t *src_end,
427
                         int64_t stop_here)
428
{
429
    // shortcut:
430
    const uint8_t *src = *src_ref;
431
    const int read_size = stop_here - atempo->position[0];
432
433
    if (stop_here <= atempo->position[0]) {
434
        return 0;
435
    }
436
437
    // samples are not expected to be skipped, unless tempo is greater than 2:
438
    av_assert0(read_size <= atempo->ring || atempo->tempo > 2.0);
439
440
    while (atempo->position[0] < stop_here && src < src_end) {
441
        int src_samples = (src_end - src) / atempo->stride;
442
443
        // load data piece-wise, in order to avoid complicating the logic:
444
        int nsamples = FFMIN(read_size, src_samples);
445
        int na;
446
        int nb;
447
448
        nsamples = FFMIN(nsamples, atempo->ring);
449
        na = FFMIN(nsamples, atempo->ring - atempo->tail);
450
        nb = FFMIN(nsamples - na, atempo->ring);
451
452
        if (na) {
453
            uint8_t *a = atempo->buffer + atempo->tail * atempo->stride;
454
            memcpy(a, src, na * atempo->stride);
455
456
            src += na * atempo->stride;
457
            atempo->position[0] += na;
458
459
            atempo->size = FFMIN(atempo->size + na, atempo->ring);
460
            atempo->tail = (atempo->tail + na) % atempo->ring;
461
            atempo->head =
462
                atempo->size < atempo->ring ?
463
                atempo->tail - atempo->size :
464
                atempo->tail;
465
        }
466
467
        if (nb) {
468
            uint8_t *b = atempo->buffer;
469
            memcpy(b, src, nb * atempo->stride);
470
471
            src += nb * atempo->stride;
472
            atempo->position[0] += nb;
473
474
            atempo->size = FFMIN(atempo->size + nb, atempo->ring);
475
            atempo->tail = (atempo->tail + nb) % atempo->ring;
476
            atempo->head =
477
                atempo->size < atempo->ring ?
478
                atempo->tail - atempo->size :
479
                atempo->tail;
480
        }
481
    }
482
483
    // pass back the updated source buffer pointer:
484
    *src_ref = src;
485
486
    // sanity check:
487
    av_assert0(atempo->position[0] <= stop_here);
488
489
    return atempo->position[0] == stop_here ? 0 : AVERROR(EAGAIN);
490
}
491
492
/**
493
 * Populate current audio fragment data buffer.
494
 *
495
 * @return
496
 *   0 when the fragment is ready,
497
 *   AVERROR(EAGAIN) if more input data is required.
498
 */
499
static int yae_load_frag(ATempoContext *atempo,
500
                         const uint8_t **src_ref,
501
                         const uint8_t *src_end)
502
{
503
    // shortcuts:
504
    AudioFragment *frag = yae_curr_frag(atempo);
505
    uint8_t *dst;
506
    int64_t missing, start, zeros;
507
    uint32_t nsamples;
508
    const uint8_t *a, *b;
509
    int i0, i1, n0, n1, na, nb;
510
511
    int64_t stop_here = frag->position[0] + atempo->window;
512
    if (src_ref && yae_load_data(atempo, src_ref, src_end, stop_here) != 0) {
513
        return AVERROR(EAGAIN);
514
    }
515
516
    // calculate the number of samples we don't have:
517
    missing =
518
        stop_here > atempo->position[0] ?
519
        stop_here - atempo->position[0] : 0;
520
521
    nsamples =
522
        missing < (int64_t)atempo->window ?
523
        (uint32_t)(atempo->window - missing) : 0;
524
525
    // setup the output buffer:
526
    frag->nsamples = nsamples;
527
    dst = frag->data;
528
529
    start = atempo->position[0] - atempo->size;
530
    zeros = 0;
531
532
    if (frag->position[0] < start) {
533
        // what we don't have we substitute with zeros:
534
        zeros = FFMIN(start - frag->position[0], (int64_t)nsamples);
535
        av_assert0(zeros != nsamples);
536
537
        memset(dst, 0, zeros * atempo->stride);
538
        dst += zeros * atempo->stride;
539
    }
540
541
    if (zeros == nsamples) {
542
        return 0;
543
    }
544
545
    // get the remaining data from the ring buffer:
546
    na = (atempo->head < atempo->tail ?
547
          atempo->tail - atempo->head :
548
          atempo->ring - atempo->head);
549
550
    nb = atempo->head < atempo->tail ? 0 : atempo->tail;
551
552
    // sanity check:
553
    av_assert0(nsamples <= zeros + na + nb);
554
555
    a = atempo->buffer + atempo->head * atempo->stride;
556
    b = atempo->buffer;
557
558
    i0 = frag->position[0] + zeros - start;
559
    i1 = i0 < na ? 0 : i0 - na;
560
561
    n0 = i0 < na ? FFMIN(na - i0, (int)(nsamples - zeros)) : 0;
562
    n1 = nsamples - zeros - n0;
563
564
    if (n0) {
565
        memcpy(dst, a + i0 * atempo->stride, n0 * atempo->stride);
566
        dst += n0 * atempo->stride;
567
    }
568
569
    if (n1) {
570
        memcpy(dst, b + i1 * atempo->stride, n1 * atempo->stride);
571
    }
572
573
    return 0;
574
}
575
576
/**
577
 * Prepare for loading next audio fragment.
578
 */
579
static void yae_advance_to_next_frag(ATempoContext *atempo)
580
{
581
    const double fragment_step = atempo->tempo * (double)(atempo->window / 2);
582
583
    const AudioFragment *prev;
584
    AudioFragment       *frag;
585
586
    atempo->nfrag++;
587
    prev = yae_prev_frag(atempo);
588
    frag = yae_curr_frag(atempo);
589
590
    frag->position[0] = prev->position[0] + (int64_t)fragment_step;
591
    frag->position[1] = prev->position[1] + atempo->window / 2;
592
    frag->nsamples    = 0;
593
}
594
595
/**
596
 * Calculate cross-correlation via rDFT.
597
 *
598
 * Multiply two vectors of complex numbers (result of real_to_complex rDFT)
599
 * and transform back via complex_to_real rDFT.
600
 */
601
static void yae_xcorr_via_rdft(FFTSample *xcorr,
602
                               RDFTContext *complex_to_real,
603
                               const FFTComplex *xa,
604
                               const FFTComplex *xb,
605
                               const int window)
606
{
607
    FFTComplex *xc = (FFTComplex *)xcorr;
608
    int i;
609
610
    // NOTE: first element requires special care -- Given Y = rDFT(X),
611
    // Im(Y[0]) and Im(Y[N/2]) are always zero, therefore av_rdft_calc
612
    // stores Re(Y[N/2]) in place of Im(Y[0]).
613
614
    xc->re = xa->re * xb->re;
615
    xc->im = xa->im * xb->im;
616
    xa++;
617
    xb++;
618
    xc++;
619
620
    for (i = 1; i < window; i++, xa++, xb++, xc++) {
621
        xc->re = (xa->re * xb->re + xa->im * xb->im);
622
        xc->im = (xa->im * xb->re - xa->re * xb->im);
623
    }
624
625
    // apply inverse rDFT:
626
    av_rdft_calc(complex_to_real, xcorr);
627
}
628
629
/**
630
 * Calculate alignment offset for given fragment
631
 * relative to the previous fragment.
632
 *
633
 * @return alignment offset of current fragment relative to previous.
634
 */
635
static int yae_align(AudioFragment *frag,
636
                     const AudioFragment *prev,
637
                     const int window,
638
                     const int delta_max,
639
                     const int drift,
640
                     FFTSample *correlation,
641
                     RDFTContext *complex_to_real)
642
{
643
    int       best_offset = -drift;
644
    FFTSample best_metric = -FLT_MAX;
645
    FFTSample *xcorr;
646
647
    int i0;
648
    int i1;
649
    int i;
650
651
    yae_xcorr_via_rdft(correlation,
652
                       complex_to_real,
653
                       (const FFTComplex *)prev->xdat,
654
                       (const FFTComplex *)frag->xdat,
655
                       window);
656
657
    // identify search window boundaries:
658
    i0 = FFMAX(window / 2 - delta_max - drift, 0);
659
    i0 = FFMIN(i0, window);
660
661
    i1 = FFMIN(window / 2 + delta_max - drift, window - window / 16);
662
    i1 = FFMAX(i1, 0);
663
664
    // identify cross-correlation peaks within search window:
665
    xcorr = correlation + i0;
666
667
    for (i = i0; i < i1; i++, xcorr++) {
668
        FFTSample metric = *xcorr;
669
670
        // normalize:
671
        FFTSample drifti = (FFTSample)(drift + i);
672
        metric *= drifti * (FFTSample)(i - i0) * (FFTSample)(i1 - i);
673
674
        if (metric > best_metric) {
675
            best_metric = metric;
676
            best_offset = i - window / 2;
677
        }
678
    }
679
680
    return best_offset;
681
}
682
683
/**
684
 * Adjust current fragment position for better alignment
685
 * with previous fragment.
686
 *
687
 * @return alignment correction.
688
 */
689
static int yae_adjust_position(ATempoContext *atempo)
690
{
691
    const AudioFragment *prev = yae_prev_frag(atempo);
692
    AudioFragment       *frag = yae_curr_frag(atempo);
693
694
    const double prev_output_position =
695
        (double)(prev->position[1] - atempo->origin[1] + atempo->window / 2) *
696
        atempo->tempo;
697
698
    const double ideal_output_position =
699
        (double)(prev->position[0] - atempo->origin[0] + atempo->window / 2);
700
701
    const int drift = (int)(prev_output_position - ideal_output_position);
702
703
    const int delta_max  = atempo->window / 2;
704
    const int correction = yae_align(frag,
705
                                     prev,
706
                                     atempo->window,
707
                                     delta_max,
708
                                     drift,
709
                                     atempo->correlation,
710
                                     atempo->complex_to_real);
711
712
    if (correction) {
713
        // adjust fragment position:
714
        frag->position[0] -= correction;
715
716
        // clear so that the fragment can be reloaded:
717
        frag->nsamples = 0;
718
    }
719
720
    return correction;
721
}
722
723
/**
724
 * A helper macro for blending the overlap region of previous
725
 * and current audio fragment.
726
 */
727
#define yae_blend(scalar_type)                                          \
728
    do {                                                                \
729
        const scalar_type *aaa = (const scalar_type *)a;                \
730
        const scalar_type *bbb = (const scalar_type *)b;                \
731
                                                                        \
732
        scalar_type *out     = (scalar_type *)dst;                      \
733
        scalar_type *out_end = (scalar_type *)dst_end;                  \
734
        int64_t i;                                                      \
735
                                                                        \
736
        for (i = 0; i < overlap && out < out_end;                       \
737
             i++, atempo->position[1]++, wa++, wb++) {                  \
738
            float w0 = *wa;                                             \
739
            float w1 = *wb;                                             \
740
            int j;                                                      \
741
                                                                        \
742
            for (j = 0; j < atempo->channels;                           \
743
                 j++, aaa++, bbb++, out++) {                            \
744
                float t0 = (float)*aaa;                                 \
745
                float t1 = (float)*bbb;                                 \
746
                                                                        \
747
                *out =                                                  \
748
                    frag->position[0] + i < 0 ?                         \
749
                    *aaa :                                              \
750
                    (scalar_type)(t0 * w0 + t1 * w1);                   \
751
            }                                                           \
752
        }                                                               \
753
        dst = (uint8_t *)out;                                           \
754
    } while (0)
755
756
/**
757
 * Blend the overlap region of previous and current audio fragment
758
 * and output the results to the given destination buffer.
759
 *
760
 * @return
761
 *   0 if the overlap region was completely stored in the dst buffer,
762
 *   AVERROR(EAGAIN) if more destination buffer space is required.
763
 */
764
static int yae_overlap_add(ATempoContext *atempo,
765
                           uint8_t **dst_ref,
766
                           uint8_t *dst_end)
767
{
768
    // shortcuts:
769
    const AudioFragment *prev = yae_prev_frag(atempo);
770
    const AudioFragment *frag = yae_curr_frag(atempo);
771
772
    const int64_t start_here = FFMAX(atempo->position[1],
773
                                     frag->position[1]);
774
775
    const int64_t stop_here = FFMIN(prev->position[1] + prev->nsamples,
776
                                    frag->position[1] + frag->nsamples);
777
778
    const int64_t overlap = stop_here - start_here;
779
780
    const int64_t ia = start_here - prev->position[1];
781
    const int64_t ib = start_here - frag->position[1];
782
783
    const float *wa = atempo->hann + ia;
784
    const float *wb = atempo->hann + ib;
785
786
    const uint8_t *a = prev->data + ia * atempo->stride;
787
    const uint8_t *b = frag->data + ib * atempo->stride;
788
789
    uint8_t *dst = *dst_ref;
790
791
    av_assert0(start_here <= stop_here &&
792
               frag->position[1] <= start_here &&
793
               overlap <= frag->nsamples);
794
795
    if (atempo->format == AV_SAMPLE_FMT_U8) {
796
        yae_blend(uint8_t);
797
    } else if (atempo->format == AV_SAMPLE_FMT_S16) {
798
        yae_blend(int16_t);
799
    } else if (atempo->format == AV_SAMPLE_FMT_S32) {
800
        yae_blend(int);
801
    } else if (atempo->format == AV_SAMPLE_FMT_FLT) {
802
        yae_blend(float);
803
    } else if (atempo->format == AV_SAMPLE_FMT_DBL) {
804
        yae_blend(double);
805
    }
806
807
    // pass-back the updated destination buffer pointer:
808
    *dst_ref = dst;
809
810
    return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
811
}
812
813
/**
814
 * Feed as much data to the filter as it is able to consume
815
 * and receive as much processed data in the destination buffer
816
 * as it is able to produce or store.
817
 */
818
static void
819
yae_apply(ATempoContext *atempo,
820
          const uint8_t **src_ref,
821
          const uint8_t *src_end,
822
          uint8_t **dst_ref,
823
          uint8_t *dst_end)
824
{
825
    while (1) {
826
        if (atempo->state == YAE_LOAD_FRAGMENT) {
827
            // load additional data for the current fragment:
828
            if (yae_load_frag(atempo, src_ref, src_end) != 0) {
829
                break;
830
            }
831
832
            // down-mix to mono:
833
            yae_downmix(atempo, yae_curr_frag(atempo));
834
835
            // apply rDFT:
836
            av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
837
838
            // must load the second fragment before alignment can start:
839
            if (!atempo->nfrag) {
840
                yae_advance_to_next_frag(atempo);
841
                continue;
842
            }
843
844
            atempo->state = YAE_ADJUST_POSITION;
845
        }
846
847
        if (atempo->state == YAE_ADJUST_POSITION) {
848
            // adjust position for better alignment:
849
            if (yae_adjust_position(atempo)) {
850
                // reload the fragment at the corrected position, so that the
851
                // Hann window blending would not require normalization:
852
                atempo->state = YAE_RELOAD_FRAGMENT;
853
            } else {
854
                atempo->state = YAE_OUTPUT_OVERLAP_ADD;
855
            }
856
        }
857
858
        if (atempo->state == YAE_RELOAD_FRAGMENT) {
859
            // load additional data if necessary due to position adjustment:
860
            if (yae_load_frag(atempo, src_ref, src_end) != 0) {
861
                break;
862
            }
863
864
            // down-mix to mono:
865
            yae_downmix(atempo, yae_curr_frag(atempo));
866
867
            // apply rDFT:
868
            av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
869
870
            atempo->state = YAE_OUTPUT_OVERLAP_ADD;
871
        }
872
873
        if (atempo->state == YAE_OUTPUT_OVERLAP_ADD) {
874
            // overlap-add and output the result:
875
            if (yae_overlap_add(atempo, dst_ref, dst_end) != 0) {
876
                break;
877
            }
878
879
            // advance to the next fragment, repeat:
880
            yae_advance_to_next_frag(atempo);
881
            atempo->state = YAE_LOAD_FRAGMENT;
882
        }
883
    }
884
}
885
886
/**
887
 * Flush any buffered data from the filter.
888
 *
889
 * @return
890
 *   0 if all data was completely stored in the dst buffer,
891
 *   AVERROR(EAGAIN) if more destination buffer space is required.
892
 */
893
static int yae_flush(ATempoContext *atempo,
894
                     uint8_t **dst_ref,
895
                     uint8_t *dst_end)
896
{
897
    AudioFragment *frag = yae_curr_frag(atempo);
898
    int64_t overlap_end;
899
    int64_t start_here;
900
    int64_t stop_here;
901
    int64_t offset;
902
903
    const uint8_t *src;
904
    uint8_t *dst;
905
906
    int src_size;
907
    int dst_size;
908
    int nbytes;
909
910
    atempo->state = YAE_FLUSH_OUTPUT;
911
912
    if (!atempo->nfrag) {
913
        // there is nothing to flush:
914
        return 0;
915
    }
916
917
    if (atempo->position[0] == frag->position[0] + frag->nsamples &&
918
        atempo->position[1] == frag->position[1] + frag->nsamples) {
919
        // the current fragment is already flushed:
920
        return 0;
921
    }
922
923
    if (frag->position[0] + frag->nsamples < atempo->position[0]) {
924
        // finish loading the current (possibly partial) fragment:
925
        yae_load_frag(atempo, NULL, NULL);
926
927
        if (atempo->nfrag) {
928
            // down-mix to mono:
929
            yae_downmix(atempo, frag);
930
931
            // apply rDFT:
932
            av_rdft_calc(atempo->real_to_complex, frag->xdat);
933
934
            // align current fragment to previous fragment:
935
            if (yae_adjust_position(atempo)) {
936
                // reload the current fragment due to adjusted position:
937
                yae_load_frag(atempo, NULL, NULL);
938
            }
939
        }
940
    }
941
942
    // flush the overlap region:
943
    overlap_end = frag->position[1] + FFMIN(atempo->window / 2,
944
                                            frag->nsamples);
945
946
    while (atempo->position[1] < overlap_end) {
947
        if (yae_overlap_add(atempo, dst_ref, dst_end) != 0) {
948
            return AVERROR(EAGAIN);
949
        }
950
    }
951
952
    // check whether all of the input samples have been consumed:
953
    if (frag->position[0] + frag->nsamples < atempo->position[0]) {
954
        yae_advance_to_next_frag(atempo);
955
        return AVERROR(EAGAIN);
956
    }
957
958
    // flush the remainder of the current fragment:
959
    start_here = FFMAX(atempo->position[1], overlap_end);
960
    stop_here  = frag->position[1] + frag->nsamples;
961
    offset     = start_here - frag->position[1];
962
    av_assert0(start_here <= stop_here && frag->position[1] <= start_here);
963
964
    src = frag->data + offset * atempo->stride;
965
    dst = (uint8_t *)*dst_ref;
966
967
    src_size = (int)(stop_here - start_here) * atempo->stride;
968
    dst_size = dst_end - dst;
969
    nbytes = FFMIN(src_size, dst_size);
970
971
    memcpy(dst, src, nbytes);
972
    dst += nbytes;
973
974
    atempo->position[1] += (nbytes / atempo->stride);
975
976
    // pass-back the updated destination buffer pointer:
977
    *dst_ref = (uint8_t *)dst;
978
979
    return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
980
}
981
982
static av_cold int init(AVFilterContext *ctx)
983
{
984
    ATempoContext *atempo = ctx->priv;
985
    atempo->format = AV_SAMPLE_FMT_NONE;
986
    atempo->state  = YAE_LOAD_FRAGMENT;
987
    return 0;
988
}
989
990
static av_cold void uninit(AVFilterContext *ctx)
991
{
992
    ATempoContext *atempo = ctx->priv;
993
    yae_release_buffers(atempo);
994
}
995
996
static int query_formats(AVFilterContext *ctx)
997
{
998
    AVFilterChannelLayouts *layouts = NULL;
999
    AVFilterFormats        *formats = NULL;
1000
1001
    // WSOLA necessitates an internal sliding window ring buffer
1002
    // for incoming audio stream.
1003
    //
1004
    // Planar sample formats are too cumbersome to store in a ring buffer,
1005
    // therefore planar sample formats are not supported.
1006
    //
1007
    static const enum AVSampleFormat sample_fmts[] = {
1008
        AV_SAMPLE_FMT_U8,
1009
        AV_SAMPLE_FMT_S16,
1010
        AV_SAMPLE_FMT_S32,
1011
        AV_SAMPLE_FMT_FLT,
1012
        AV_SAMPLE_FMT_DBL,
1013
        AV_SAMPLE_FMT_NONE
1014
    };
1015
    int ret;
1016
1017
    layouts = ff_all_channel_counts();
1018
    if (!layouts) {
1019
        return AVERROR(ENOMEM);
1020
    }
1021
    ret = ff_set_common_channel_layouts(ctx, layouts);
1022
    if (ret < 0)
1023
        return ret;
1024
1025
    formats = ff_make_format_list(sample_fmts);
1026
    if (!formats) {
1027
        return AVERROR(ENOMEM);
1028
    }
1029
    ret = ff_set_common_formats(ctx, formats);
1030
    if (ret < 0)
1031
        return ret;
1032
1033
    formats = ff_all_samplerates();
1034
    if (!formats) {
1035
        return AVERROR(ENOMEM);
1036
    }
1037
    return ff_set_common_samplerates(ctx, formats);
1038
}
1039
1040
static int config_props(AVFilterLink *inlink)
1041
{
1042
    AVFilterContext  *ctx = inlink->dst;
1043
    ATempoContext *atempo = ctx->priv;
1044
1045
    enum AVSampleFormat format = inlink->format;
1046
    int sample_rate = (int)inlink->sample_rate;
1047
1048
    return yae_reset(atempo, format, sample_rate, inlink->channels);
1049
}
1050
1051
static int push_samples(ATempoContext *atempo,
1052
                        AVFilterLink *outlink,
1053
                        int n_out)
1054
{
1055
    int ret;
1056
1057
    atempo->dst_buffer->sample_rate = outlink->sample_rate;
1058
    atempo->dst_buffer->nb_samples  = n_out;
1059
1060
    // adjust the PTS:
1061
    atempo->dst_buffer->pts = atempo->start_pts +
1062
        av_rescale_q(atempo->nsamples_out,
1063
                     (AVRational){ 1, outlink->sample_rate },
1064
                     outlink->time_base);
1065
1066
    ret = ff_filter_frame(outlink, atempo->dst_buffer);
1067
    atempo->dst_buffer = NULL;
1068
    atempo->dst        = NULL;
1069
    atempo->dst_end    = NULL;
1070
    if (ret < 0)
1071
        return ret;
1072
1073
    atempo->nsamples_out += n_out;
1074
    return 0;
1075
}
1076
1077
static int filter_frame(AVFilterLink *inlink, AVFrame *src_buffer)
1078
{
1079
    AVFilterContext  *ctx = inlink->dst;
1080
    ATempoContext *atempo = ctx->priv;
1081
    AVFilterLink *outlink = ctx->outputs[0];
1082
1083
    int ret = 0;
1084
    int n_in = src_buffer->nb_samples;
1085
    int n_out = (int)(0.5 + ((double)n_in) / atempo->tempo);
1086
1087
    const uint8_t *src = src_buffer->data[0];
1088
    const uint8_t *src_end = src + n_in * atempo->stride;
1089
1090
    if (atempo->start_pts == AV_NOPTS_VALUE)
1091
        atempo->start_pts = av_rescale_q(src_buffer->pts,
1092
                                         inlink->time_base,
1093
                                         outlink->time_base);
1094
1095
    while (src < src_end) {
1096
        if (!atempo->dst_buffer) {
1097
            atempo->dst_buffer = ff_get_audio_buffer(outlink, n_out);
1098
            if (!atempo->dst_buffer) {
1099
                av_frame_free(&src_buffer);
1100
                return AVERROR(ENOMEM);
1101
            }
1102
            av_frame_copy_props(atempo->dst_buffer, src_buffer);
1103
1104
            atempo->dst = atempo->dst_buffer->data[0];
1105
            atempo->dst_end = atempo->dst + n_out * atempo->stride;
1106
        }
1107
1108
        yae_apply(atempo, &src, src_end, &atempo->dst, atempo->dst_end);
1109
1110
        if (atempo->dst == atempo->dst_end) {
1111
            int n_samples = ((atempo->dst - atempo->dst_buffer->data[0]) /
1112
                             atempo->stride);
1113
            ret = push_samples(atempo, outlink, n_samples);
1114
            if (ret < 0)
1115
                goto end;
1116
        }
1117
    }
1118
1119
    atempo->nsamples_in += n_in;
1120
end:
1121
    av_frame_free(&src_buffer);
1122
    return ret;
1123
}
1124
1125
static int request_frame(AVFilterLink *outlink)
1126
{
1127
    AVFilterContext  *ctx = outlink->src;
1128
    ATempoContext *atempo = ctx->priv;
1129
    int ret;
1130
1131
    ret = ff_request_frame(ctx->inputs[0]);
1132
1133
    if (ret == AVERROR_EOF) {
1134
        // flush the filter:
1135
        int n_max = atempo->ring;
1136
        int n_out;
1137
        int err = AVERROR(EAGAIN);
1138
1139
        while (err == AVERROR(EAGAIN)) {
1140
            if (!atempo->dst_buffer) {
1141
                atempo->dst_buffer = ff_get_audio_buffer(outlink, n_max);
1142
                if (!atempo->dst_buffer)
1143
                    return AVERROR(ENOMEM);
1144
1145
                atempo->dst = atempo->dst_buffer->data[0];
1146
                atempo->dst_end = atempo->dst + n_max * atempo->stride;
1147
            }
1148
1149
            err = yae_flush(atempo, &atempo->dst, atempo->dst_end);
1150
1151
            n_out = ((atempo->dst - atempo->dst_buffer->data[0]) /
1152
                     atempo->stride);
1153
1154
            if (n_out) {
1155
                ret = push_samples(atempo, outlink, n_out);
1156
                if (ret < 0)
1157
                    return ret;
1158
            }
1159
        }
1160
1161
        av_frame_free(&atempo->dst_buffer);
1162
        atempo->dst     = NULL;
1163
        atempo->dst_end = NULL;
1164
1165
        return AVERROR_EOF;
1166
    }
1167
1168
    return ret;
1169
}
1170
1171
static int process_command(AVFilterContext *ctx,
1172
                           const char *cmd,
1173
                           const char *arg,
1174
                           char *res,
1175
                           int res_len,
1176
                           int flags)
1177
{
1178
    int ret = ff_filter_process_command(ctx, cmd, arg, res, res_len, flags);
1179
1180
    if (ret < 0)
1181
        return ret;
1182
1183
    return yae_update(ctx);
1184
}
1185
1186
static const AVFilterPad atempo_inputs[] = {
1187
    {
1188
        .name         = "default",
1189
        .type         = AVMEDIA_TYPE_AUDIO,
1190
        .filter_frame = filter_frame,
1191
        .config_props = config_props,
1192
    },
1193
    { NULL }
1194
};
1195
1196
static const AVFilterPad atempo_outputs[] = {
1197
    {
1198
        .name          = "default",
1199
        .request_frame = request_frame,
1200
        .type          = AVMEDIA_TYPE_AUDIO,
1201
    },
1202
    { NULL }
1203
};
1204
1205
AVFilter ff_af_atempo = {
1206
    .name            = "atempo",
1207
    .description     = NULL_IF_CONFIG_SMALL("Adjust audio tempo."),
1208
    .init            = init,
1209
    .uninit          = uninit,
1210
    .query_formats   = query_formats,
1211
    .process_command = process_command,
1212
    .priv_size       = sizeof(ATempoContext),
1213
    .priv_class      = &atempo_class,
1214
    .inputs          = atempo_inputs,
1215
    .outputs         = atempo_outputs,
1216
};