2 * Copyright (c) 2012 Pavel Koshevoy <pkoshevoy at gmail dot com>
4 * This file is part of FFmpeg.
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.
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.
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
23 * tempo scaling audio filter -- an implementation of WSOLA algorithm
25 * Based on MIT licensed yaeAudioTempoFilter.h and yaeAudioFragment.h
26 * from Apprentice Video player by Pavel Koshevoy.
27 * https://sourceforge.net/projects/apprenticevideo/
29 * An explanation of SOLA algorithm is available at
30 * http://www.surina.net/article/time-and-pitch-scaling.html
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.
36 * The advantage of WSOLA algorithm is that the overlap region size is
37 * always the same, therefore the blending function is constant and
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"
54 * A fragment of audio waveform
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
62 // original packed multi-channel samples:
65 // number of samples in this fragment:
68 // rDFT transform of the down-mixed mono fragment, used for
69 // fast waveform alignment via correlation in frequency domain:
74 * Filter state machine states
80 YAE_OUTPUT_OVERLAP_ADD,
85 * Filter state machine
87 typedef struct ATempoContext {
90 // ring-buffer of input samples, necessary because some times
91 // input fragment position may be adjusted backwards:
94 // ring-buffer maximum capacity, expressed in sample rate time base:
97 // ring-buffer house keeping:
102 // 0: input sample position corresponding to the ring buffer tail
103 // 1: output sample position
106 // first input timestamp, all other timestamps are offset by this one
110 enum AVSampleFormat format;
112 // number of channels:
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
119 // fragment window size, power-of-two integer:
122 // Hann window coefficients, for feathering
123 // (blending) the overlapping fragment region:
126 // tempo scaling factor:
129 // a snapshot of previous fragment input and output position values
130 // captured when the tempo scale factor was set most recently:
133 // current/previous fragment ring-buffer:
134 AudioFragment frag[2];
136 // current fragment index:
142 // for fast correlation calculation in frequency domain:
143 RDFTContext *real_to_complex;
144 RDFTContext *complex_to_real;
145 FFTSample *correlation;
147 // for managing AVFilterPad.request_frame and AVFilterPad.filter_frame
151 uint64_t nsamples_in;
152 uint64_t nsamples_out;
155 #define YAE_ATEMPO_MIN 0.5
156 #define YAE_ATEMPO_MAX 100.0
158 #define OFFSET(x) offsetof(ATempoContext, x)
160 static const AVOption atempo_options[] = {
161 { "tempo", "set tempo scale factor",
162 OFFSET(tempo), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 },
165 AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_RUNTIME_PARAM },
169 AVFILTER_DEFINE_CLASS(atempo);
171 inline static AudioFragment *yae_curr_frag(ATempoContext *atempo)
173 return &atempo->frag[atempo->nfrag % 2];
176 inline static AudioFragment *yae_prev_frag(ATempoContext *atempo)
178 return &atempo->frag[(atempo->nfrag + 1) % 2];
182 * Reset filter to initial state, do not deallocate existing local buffers.
184 static void yae_clear(ATempoContext *atempo)
191 atempo->state = YAE_LOAD_FRAGMENT;
192 atempo->start_pts = AV_NOPTS_VALUE;
194 atempo->position[0] = 0;
195 atempo->position[1] = 0;
197 atempo->origin[0] = 0;
198 atempo->origin[1] = 0;
200 atempo->frag[0].position[0] = 0;
201 atempo->frag[0].position[1] = 0;
202 atempo->frag[0].nsamples = 0;
204 atempo->frag[1].position[0] = 0;
205 atempo->frag[1].position[1] = 0;
206 atempo->frag[1].nsamples = 0;
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);
214 av_frame_free(&atempo->dst_buffer);
216 atempo->dst_end = NULL;
218 atempo->nsamples_in = 0;
219 atempo->nsamples_out = 0;
223 * Reset filter to initial state and deallocate all buffers.
225 static void yae_release_buffers(ATempoContext *atempo)
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);
234 av_freep(&atempo->buffer);
235 av_freep(&atempo->hann);
236 av_freep(&atempo->correlation);
238 av_rdft_end(atempo->real_to_complex);
239 atempo->real_to_complex = NULL;
241 av_rdft_end(atempo->complex_to_real);
242 atempo->complex_to_real = NULL;
245 /* av_realloc is not aligned enough; fortunately, the data does not need to
247 #define RE_MALLOC_OR_FAIL(field, field_size) \
250 field = av_malloc(field_size); \
252 yae_release_buffers(atempo); \
253 return AVERROR(ENOMEM); \
258 * Prepare filter for processing audio data of given format,
259 * sample rate and number of channels.
261 static int yae_reset(ATempoContext *atempo,
262 enum AVSampleFormat format,
266 const int sample_size = av_get_bytes_per_sample(format);
267 uint32_t nlevels = 0;
271 atempo->format = format;
272 atempo->channels = channels;
273 atempo->stride = sample_size * channels;
275 // pick a segment window size:
276 atempo->window = sample_rate / 24;
278 // adjust window size to be a power-of-two integer:
279 nlevels = av_log2(atempo->window);
281 av_assert0(pot <= atempo->window);
283 if (pot < atempo->window) {
284 atempo->window = pot * 2;
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));
294 // initialize rDFT contexts:
295 av_rdft_end(atempo->real_to_complex);
296 atempo->real_to_complex = NULL;
298 av_rdft_end(atempo->complex_to_real);
299 atempo->complex_to_real = NULL;
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);
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);
313 RE_MALLOC_OR_FAIL(atempo->correlation, atempo->window * sizeof(FFTComplex));
315 atempo->ring = atempo->window * 3;
316 RE_MALLOC_OR_FAIL(atempo->buffer, atempo->ring * atempo->stride);
318 // initialize the Hann window function:
319 RE_MALLOC_OR_FAIL(atempo->hann, atempo->window * sizeof(float));
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;
331 static int yae_update(AVFilterContext *ctx)
333 const AudioFragment *prev;
334 ATempoContext *atempo = ctx->priv;
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;
343 * A helper macro for initializing complex data buffer with scalar data
346 #define yae_init_xdat(scalar_type, scalar_max) \
348 const uint8_t *src_end = src + \
349 frag->nsamples * atempo->channels * sizeof(scalar_type); \
351 FFTSample *xdat = frag->xdat; \
354 if (atempo->channels == 1) { \
355 for (; src < src_end; xdat++) { \
356 tmp = *(const scalar_type *)src; \
357 src += sizeof(scalar_type); \
359 *xdat = (FFTSample)tmp; \
362 FFTSample s, max, ti, si; \
365 for (; src < src_end; xdat++) { \
366 tmp = *(const scalar_type *)src; \
367 src += sizeof(scalar_type); \
369 max = (FFTSample)tmp; \
370 s = FFMIN((FFTSample)scalar_max, \
371 (FFTSample)fabsf(max)); \
373 for (i = 1; i < atempo->channels; i++) { \
374 tmp = *(const scalar_type *)src; \
375 src += sizeof(scalar_type); \
377 ti = (FFTSample)tmp; \
378 si = FFMIN((FFTSample)scalar_max, \
379 (FFTSample)fabsf(ti)); \
393 * Initialize complex data buffer of a given audio fragment
394 * with down-mixed mono data of appropriate scalar type.
396 static void yae_downmix(ATempoContext *atempo, AudioFragment *frag)
399 const uint8_t *src = frag->data;
401 // init complex data buffer used for FFT and Correlation:
402 memset(frag->xdat, 0, sizeof(FFTComplex) * atempo->window);
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);
418 * Populate the internal data buffer on as-needed basis.
421 * 0 if requested data was already available or was successfully loaded,
422 * AVERROR(EAGAIN) if more input data is required.
424 static int yae_load_data(ATempoContext *atempo,
425 const uint8_t **src_ref,
426 const uint8_t *src_end,
430 const uint8_t *src = *src_ref;
431 const int read_size = stop_here - atempo->position[0];
433 if (stop_here <= atempo->position[0]) {
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);
440 while (atempo->position[0] < stop_here && src < src_end) {
441 int src_samples = (src_end - src) / atempo->stride;
443 // load data piece-wise, in order to avoid complicating the logic:
444 int nsamples = FFMIN(read_size, src_samples);
448 nsamples = FFMIN(nsamples, atempo->ring);
449 na = FFMIN(nsamples, atempo->ring - atempo->tail);
450 nb = FFMIN(nsamples - na, atempo->ring);
453 uint8_t *a = atempo->buffer + atempo->tail * atempo->stride;
454 memcpy(a, src, na * atempo->stride);
456 src += na * atempo->stride;
457 atempo->position[0] += na;
459 atempo->size = FFMIN(atempo->size + na, atempo->ring);
460 atempo->tail = (atempo->tail + na) % atempo->ring;
462 atempo->size < atempo->ring ?
463 atempo->tail - atempo->size :
468 uint8_t *b = atempo->buffer;
469 memcpy(b, src, nb * atempo->stride);
471 src += nb * atempo->stride;
472 atempo->position[0] += nb;
474 atempo->size = FFMIN(atempo->size + nb, atempo->ring);
475 atempo->tail = (atempo->tail + nb) % atempo->ring;
477 atempo->size < atempo->ring ?
478 atempo->tail - atempo->size :
483 // pass back the updated source buffer pointer:
487 av_assert0(atempo->position[0] <= stop_here);
489 return atempo->position[0] == stop_here ? 0 : AVERROR(EAGAIN);
493 * Populate current audio fragment data buffer.
496 * 0 when the fragment is ready,
497 * AVERROR(EAGAIN) if more input data is required.
499 static int yae_load_frag(ATempoContext *atempo,
500 const uint8_t **src_ref,
501 const uint8_t *src_end)
504 AudioFragment *frag = yae_curr_frag(atempo);
506 int64_t missing, start, zeros;
508 const uint8_t *a, *b;
509 int i0, i1, n0, n1, na, nb;
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);
516 // calculate the number of samples we don't have:
518 stop_here > atempo->position[0] ?
519 stop_here - atempo->position[0] : 0;
522 missing < (int64_t)atempo->window ?
523 (uint32_t)(atempo->window - missing) : 0;
525 // setup the output buffer:
526 frag->nsamples = nsamples;
529 start = atempo->position[0] - atempo->size;
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);
537 memset(dst, 0, zeros * atempo->stride);
538 dst += zeros * atempo->stride;
541 if (zeros == nsamples) {
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);
550 nb = atempo->head < atempo->tail ? 0 : atempo->tail;
553 av_assert0(nsamples <= zeros + na + nb);
555 a = atempo->buffer + atempo->head * atempo->stride;
558 i0 = frag->position[0] + zeros - start;
559 i1 = i0 < na ? 0 : i0 - na;
561 n0 = i0 < na ? FFMIN(na - i0, (int)(nsamples - zeros)) : 0;
562 n1 = nsamples - zeros - n0;
565 memcpy(dst, a + i0 * atempo->stride, n0 * atempo->stride);
566 dst += n0 * atempo->stride;
570 memcpy(dst, b + i1 * atempo->stride, n1 * atempo->stride);
577 * Prepare for loading next audio fragment.
579 static void yae_advance_to_next_frag(ATempoContext *atempo)
581 const double fragment_step = atempo->tempo * (double)(atempo->window / 2);
583 const AudioFragment *prev;
587 prev = yae_prev_frag(atempo);
588 frag = yae_curr_frag(atempo);
590 frag->position[0] = prev->position[0] + (int64_t)fragment_step;
591 frag->position[1] = prev->position[1] + atempo->window / 2;
596 * Calculate cross-correlation via rDFT.
598 * Multiply two vectors of complex numbers (result of real_to_complex rDFT)
599 * and transform back via complex_to_real rDFT.
601 static void yae_xcorr_via_rdft(FFTSample *xcorr,
602 RDFTContext *complex_to_real,
603 const FFTComplex *xa,
604 const FFTComplex *xb,
607 FFTComplex *xc = (FFTComplex *)xcorr;
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]).
614 xc->re = xa->re * xb->re;
615 xc->im = xa->im * xb->im;
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);
625 // apply inverse rDFT:
626 av_rdft_calc(complex_to_real, xcorr);
630 * Calculate alignment offset for given fragment
631 * relative to the previous fragment.
633 * @return alignment offset of current fragment relative to previous.
635 static int yae_align(AudioFragment *frag,
636 const AudioFragment *prev,
640 FFTSample *correlation,
641 RDFTContext *complex_to_real)
643 int best_offset = -drift;
644 FFTSample best_metric = -FLT_MAX;
651 yae_xcorr_via_rdft(correlation,
653 (const FFTComplex *)prev->xdat,
654 (const FFTComplex *)frag->xdat,
657 // identify search window boundaries:
658 i0 = FFMAX(window / 2 - delta_max - drift, 0);
659 i0 = FFMIN(i0, window);
661 i1 = FFMIN(window / 2 + delta_max - drift, window - window / 16);
664 // identify cross-correlation peaks within search window:
665 xcorr = correlation + i0;
667 for (i = i0; i < i1; i++, xcorr++) {
668 FFTSample metric = *xcorr;
671 FFTSample drifti = (FFTSample)(drift + i);
672 metric *= drifti * (FFTSample)(i - i0) * (FFTSample)(i1 - i);
674 if (metric > best_metric) {
675 best_metric = metric;
676 best_offset = i - window / 2;
684 * Adjust current fragment position for better alignment
685 * with previous fragment.
687 * @return alignment correction.
689 static int yae_adjust_position(ATempoContext *atempo)
691 const AudioFragment *prev = yae_prev_frag(atempo);
692 AudioFragment *frag = yae_curr_frag(atempo);
694 const double prev_output_position =
695 (double)(prev->position[1] - atempo->origin[1] + atempo->window / 2) *
698 const double ideal_output_position =
699 (double)(prev->position[0] - atempo->origin[0] + atempo->window / 2);
701 const int drift = (int)(prev_output_position - ideal_output_position);
703 const int delta_max = atempo->window / 2;
704 const int correction = yae_align(frag,
710 atempo->complex_to_real);
713 // adjust fragment position:
714 frag->position[0] -= correction;
716 // clear so that the fragment can be reloaded:
724 * A helper macro for blending the overlap region of previous
725 * and current audio fragment.
727 #define yae_blend(scalar_type) \
729 const scalar_type *aaa = (const scalar_type *)a; \
730 const scalar_type *bbb = (const scalar_type *)b; \
732 scalar_type *out = (scalar_type *)dst; \
733 scalar_type *out_end = (scalar_type *)dst_end; \
736 for (i = 0; i < overlap && out < out_end; \
737 i++, atempo->position[1]++, wa++, wb++) { \
742 for (j = 0; j < atempo->channels; \
743 j++, aaa++, bbb++, out++) { \
744 float t0 = (float)*aaa; \
745 float t1 = (float)*bbb; \
748 frag->position[0] + i < 0 ? \
750 (scalar_type)(t0 * w0 + t1 * w1); \
753 dst = (uint8_t *)out; \
757 * Blend the overlap region of previous and current audio fragment
758 * and output the results to the given destination buffer.
761 * 0 if the overlap region was completely stored in the dst buffer,
762 * AVERROR(EAGAIN) if more destination buffer space is required.
764 static int yae_overlap_add(ATempoContext *atempo,
769 const AudioFragment *prev = yae_prev_frag(atempo);
770 const AudioFragment *frag = yae_curr_frag(atempo);
772 const int64_t start_here = FFMAX(atempo->position[1],
775 const int64_t stop_here = FFMIN(prev->position[1] + prev->nsamples,
776 frag->position[1] + frag->nsamples);
778 const int64_t overlap = stop_here - start_here;
780 const int64_t ia = start_here - prev->position[1];
781 const int64_t ib = start_here - frag->position[1];
783 const float *wa = atempo->hann + ia;
784 const float *wb = atempo->hann + ib;
786 const uint8_t *a = prev->data + ia * atempo->stride;
787 const uint8_t *b = frag->data + ib * atempo->stride;
789 uint8_t *dst = *dst_ref;
791 av_assert0(start_here <= stop_here &&
792 frag->position[1] <= start_here &&
793 overlap <= frag->nsamples);
795 if (atempo->format == AV_SAMPLE_FMT_U8) {
797 } else if (atempo->format == AV_SAMPLE_FMT_S16) {
799 } else if (atempo->format == AV_SAMPLE_FMT_S32) {
801 } else if (atempo->format == AV_SAMPLE_FMT_FLT) {
803 } else if (atempo->format == AV_SAMPLE_FMT_DBL) {
807 // pass-back the updated destination buffer pointer:
810 return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
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.
819 yae_apply(ATempoContext *atempo,
820 const uint8_t **src_ref,
821 const uint8_t *src_end,
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) {
833 yae_downmix(atempo, yae_curr_frag(atempo));
836 av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
838 // must load the second fragment before alignment can start:
839 if (!atempo->nfrag) {
840 yae_advance_to_next_frag(atempo);
844 atempo->state = YAE_ADJUST_POSITION;
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;
854 atempo->state = YAE_OUTPUT_OVERLAP_ADD;
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) {
865 yae_downmix(atempo, yae_curr_frag(atempo));
868 av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
870 atempo->state = YAE_OUTPUT_OVERLAP_ADD;
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) {
879 // advance to the next fragment, repeat:
880 yae_advance_to_next_frag(atempo);
881 atempo->state = YAE_LOAD_FRAGMENT;
887 * Flush any buffered data from the filter.
890 * 0 if all data was completely stored in the dst buffer,
891 * AVERROR(EAGAIN) if more destination buffer space is required.
893 static int yae_flush(ATempoContext *atempo,
897 AudioFragment *frag = yae_curr_frag(atempo);
910 atempo->state = YAE_FLUSH_OUTPUT;
912 if (!atempo->nfrag) {
913 // there is nothing to flush:
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:
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);
929 yae_downmix(atempo, frag);
932 av_rdft_calc(atempo->real_to_complex, frag->xdat);
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);
942 // flush the overlap region:
943 overlap_end = frag->position[1] + FFMIN(atempo->window / 2,
946 while (atempo->position[1] < overlap_end) {
947 if (yae_overlap_add(atempo, dst_ref, dst_end) != 0) {
948 return AVERROR(EAGAIN);
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);
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);
964 src = frag->data + offset * atempo->stride;
965 dst = (uint8_t *)*dst_ref;
967 src_size = (int)(stop_here - start_here) * atempo->stride;
968 dst_size = dst_end - dst;
969 nbytes = FFMIN(src_size, dst_size);
971 memcpy(dst, src, nbytes);
974 atempo->position[1] += (nbytes / atempo->stride);
976 // pass-back the updated destination buffer pointer:
977 *dst_ref = (uint8_t *)dst;
979 return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
982 static av_cold int init(AVFilterContext *ctx)
984 ATempoContext *atempo = ctx->priv;
985 atempo->format = AV_SAMPLE_FMT_NONE;
986 atempo->state = YAE_LOAD_FRAGMENT;
990 static av_cold void uninit(AVFilterContext *ctx)
992 ATempoContext *atempo = ctx->priv;
993 yae_release_buffers(atempo);
996 static int query_formats(AVFilterContext *ctx)
998 AVFilterChannelLayouts *layouts = NULL;
999 AVFilterFormats *formats = NULL;
1001 // WSOLA necessitates an internal sliding window ring buffer
1002 // for incoming audio stream.
1004 // Planar sample formats are too cumbersome to store in a ring buffer,
1005 // therefore planar sample formats are not supported.
1007 static const enum AVSampleFormat sample_fmts[] = {
1017 layouts = ff_all_channel_counts();
1019 return AVERROR(ENOMEM);
1021 ret = ff_set_common_channel_layouts(ctx, layouts);
1025 formats = ff_make_format_list(sample_fmts);
1027 return AVERROR(ENOMEM);
1029 ret = ff_set_common_formats(ctx, formats);
1033 formats = ff_all_samplerates();
1035 return AVERROR(ENOMEM);
1037 return ff_set_common_samplerates(ctx, formats);
1040 static int config_props(AVFilterLink *inlink)
1042 AVFilterContext *ctx = inlink->dst;
1043 ATempoContext *atempo = ctx->priv;
1045 enum AVSampleFormat format = inlink->format;
1046 int sample_rate = (int)inlink->sample_rate;
1048 return yae_reset(atempo, format, sample_rate, inlink->channels);
1051 static int push_samples(ATempoContext *atempo,
1052 AVFilterLink *outlink,
1057 atempo->dst_buffer->sample_rate = outlink->sample_rate;
1058 atempo->dst_buffer->nb_samples = n_out;
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);
1066 ret = ff_filter_frame(outlink, atempo->dst_buffer);
1067 atempo->dst_buffer = NULL;
1069 atempo->dst_end = NULL;
1073 atempo->nsamples_out += n_out;
1077 static int filter_frame(AVFilterLink *inlink, AVFrame *src_buffer)
1079 AVFilterContext *ctx = inlink->dst;
1080 ATempoContext *atempo = ctx->priv;
1081 AVFilterLink *outlink = ctx->outputs[0];
1084 int n_in = src_buffer->nb_samples;
1085 int n_out = (int)(0.5 + ((double)n_in) / atempo->tempo);
1087 const uint8_t *src = src_buffer->data[0];
1088 const uint8_t *src_end = src + n_in * atempo->stride;
1090 if (atempo->start_pts == AV_NOPTS_VALUE)
1091 atempo->start_pts = av_rescale_q(src_buffer->pts,
1093 outlink->time_base);
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);
1102 av_frame_copy_props(atempo->dst_buffer, src_buffer);
1104 atempo->dst = atempo->dst_buffer->data[0];
1105 atempo->dst_end = atempo->dst + n_out * atempo->stride;
1108 yae_apply(atempo, &src, src_end, &atempo->dst, atempo->dst_end);
1110 if (atempo->dst == atempo->dst_end) {
1111 int n_samples = ((atempo->dst - atempo->dst_buffer->data[0]) /
1113 ret = push_samples(atempo, outlink, n_samples);
1119 atempo->nsamples_in += n_in;
1121 av_frame_free(&src_buffer);
1125 static int request_frame(AVFilterLink *outlink)
1127 AVFilterContext *ctx = outlink->src;
1128 ATempoContext *atempo = ctx->priv;
1131 ret = ff_request_frame(ctx->inputs[0]);
1133 if (ret == AVERROR_EOF) {
1134 // flush the filter:
1135 int n_max = atempo->ring;
1137 int err = AVERROR(EAGAIN);
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);
1145 atempo->dst = atempo->dst_buffer->data[0];
1146 atempo->dst_end = atempo->dst + n_max * atempo->stride;
1149 err = yae_flush(atempo, &atempo->dst, atempo->dst_end);
1151 n_out = ((atempo->dst - atempo->dst_buffer->data[0]) /
1155 ret = push_samples(atempo, outlink, n_out);
1161 av_frame_free(&atempo->dst_buffer);
1163 atempo->dst_end = NULL;
1171 static int process_command(AVFilterContext *ctx,
1178 int ret = ff_filter_process_command(ctx, cmd, arg, res, res_len, flags);
1183 return yae_update(ctx);
1186 static const AVFilterPad atempo_inputs[] = {
1189 .type = AVMEDIA_TYPE_AUDIO,
1190 .filter_frame = filter_frame,
1191 .config_props = config_props,
1196 static const AVFilterPad atempo_outputs[] = {
1199 .request_frame = request_frame,
1200 .type = AVMEDIA_TYPE_AUDIO,
1205 const AVFilter ff_af_atempo = {
1207 .description = NULL_IF_CONFIG_SMALL("Adjust audio tempo."),
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,