2 * scaletempo audio filter
4 * scale tempo while maintaining pitch
5 * (WSOLA technique with cross correlation)
6 * inspired by SoundTouch library by Olli Parviainen
9 * - produce 'stride' output samples per loop
10 * - consume stride*scale input samples per loop
12 * to produce smoother transitions between strides, blend next overlap
13 * samples from last stride with correlated samples of current input
15 * Copyright (c) 2007 Robert Juliano
17 * This file is part of MPlayer.
19 * MPlayer is free software; you can redistribute it and/or modify
20 * it under the terms of the GNU General Public License as published by
21 * the Free Software Foundation; either version 2 of the License, or
22 * (at your option) any later version.
24 * MPlayer is distributed in the hope that it will be useful,
25 * but WITHOUT ANY WARRANTY; without even the implied warranty of
26 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
27 * GNU General Public License for more details.
29 * You should have received a copy of the GNU General Public License along
30 * with MPlayer; if not, write to the Free Software Foundation, Inc.,
31 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
39 #include "libavutil/common.h"
40 #include "subopt-helper.h"
42 // Data for specific instances of this filter
43 typedef struct af_scaletempo_s
48 float frames_stride_scaled
;
49 float frames_stride_error
;
52 float bytes_stride_scaled
;
64 void (*output_overlap
)(struct af_scaletempo_s
* s
, void* out_buf
, int bytes_off
);
70 int (*best_overlap_offset
)(struct af_scaletempo_s
* s
);
74 float percent_overlap
;
80 static int fill_queue(struct af_instance_s
* af
, af_data_t
* data
, int offset
)
82 af_scaletempo_t
* s
= af
->setup
;
83 int bytes_in
= data
->len
- offset
;
84 int offset_unchanged
= offset
;
86 if (s
->bytes_to_slide
> 0) {
87 if (s
->bytes_to_slide
< s
->bytes_queued
) {
88 int bytes_move
= s
->bytes_queued
- s
->bytes_to_slide
;
90 s
->buf_queue
+ s
->bytes_to_slide
,
92 s
->bytes_to_slide
= 0;
93 s
->bytes_queued
= bytes_move
;
96 s
->bytes_to_slide
-= s
->bytes_queued
;
97 bytes_skip
= FFMIN(s
->bytes_to_slide
, bytes_in
);
99 s
->bytes_to_slide
-= bytes_skip
;
100 offset
+= bytes_skip
;
101 bytes_in
-= bytes_skip
;
106 int bytes_copy
= FFMIN(s
->bytes_queue
- s
->bytes_queued
, bytes_in
);
107 memcpy(s
->buf_queue
+ s
->bytes_queued
,
108 (int8_t*)data
->audio
+ offset
,
110 s
->bytes_queued
+= bytes_copy
;
111 offset
+= bytes_copy
;
114 return offset
- offset_unchanged
;
117 #define UNROLL_PADDING (4*4)
119 static int best_overlap_offset_float(af_scaletempo_t
* s
)
121 float *pw
, *po
, *ppc
, *search_start
;
122 float best_corr
= INT_MIN
;
126 pw
= s
->table_window
;
128 po
+= s
->num_channels
;
129 ppc
= s
->buf_pre_corr
;
130 for (i
=s
->num_channels
; i
<s
->samples_overlap
; i
++) {
131 *ppc
++ = *pw
++ * *po
++;
134 search_start
= (float*)s
->buf_queue
+ s
->num_channels
;
135 for (off
=0; off
<s
->frames_search
; off
++) {
137 float* ps
= search_start
;
138 ppc
= s
->buf_pre_corr
;
139 for (i
=s
->num_channels
; i
<s
->samples_overlap
; i
++) {
140 corr
+= *ppc
++ * *ps
++;
142 if (corr
> best_corr
) {
146 search_start
+= s
->num_channels
;
149 return best_off
* 4 * s
->num_channels
;
152 static int best_overlap_offset_s16(af_scaletempo_t
* s
)
155 int16_t *po
, *search_start
;
156 int64_t best_corr
= INT64_MIN
;
161 pw
= s
->table_window
;
163 po
+= s
->num_channels
;
164 ppc
= s
->buf_pre_corr
;
165 for (i
=s
->num_channels
; i
<s
->samples_overlap
; i
++) {
166 *ppc
++ = ( *pw
++ * *po
++ ) >> 15;
169 search_start
= (int16_t*)s
->buf_queue
+ s
->num_channels
;
170 for (off
=0; off
<s
->frames_search
; off
++) {
172 int16_t* ps
= search_start
;
173 ppc
= s
->buf_pre_corr
;
174 ppc
+= s
->samples_overlap
- s
->num_channels
;
175 ps
+= s
->samples_overlap
- s
->num_channels
;
176 i
= -(s
->samples_overlap
- s
->num_channels
);
178 corr
+= ppc
[i
+0] * ps
[i
+0];
179 corr
+= ppc
[i
+1] * ps
[i
+1];
180 corr
+= ppc
[i
+2] * ps
[i
+2];
181 corr
+= ppc
[i
+3] * ps
[i
+3];
184 if (corr
> best_corr
) {
188 search_start
+= s
->num_channels
;
191 return best_off
* 2 * s
->num_channels
;
194 static void output_overlap_float(af_scaletempo_t
* s
, void* buf_out
,
197 float* pout
= buf_out
;
198 float* pb
= s
->table_blend
;
199 float* po
= s
->buf_overlap
;
200 float* pin
= (float*)(s
->buf_queue
+ bytes_off
);
202 for (i
=0; i
<s
->samples_overlap
; i
++) {
203 *pout
++ = *po
- *pb
++ * ( *po
- *pin
++ ); po
++;
206 static void output_overlap_s16(af_scaletempo_t
* s
, void* buf_out
,
209 int16_t* pout
= buf_out
;
210 int32_t* pb
= s
->table_blend
;
211 int16_t* po
= s
->buf_overlap
;
212 int16_t* pin
= (int16_t*)(s
->buf_queue
+ bytes_off
);
214 for (i
=0; i
<s
->samples_overlap
; i
++) {
215 *pout
++ = *po
- ( ( *pb
++ * ( *po
- *pin
++ ) ) >> 16 ); po
++;
219 // Filter data through filter
220 static af_data_t
* play(struct af_instance_s
* af
, af_data_t
* data
)
222 af_scaletempo_t
* s
= af
->setup
;
227 if (s
->scale
== 1.0) {
232 // RESIZE_LOCAL_BUFFER - can't use macro
233 max_bytes_out
= ((int)(data
->len
/ s
->bytes_stride_scaled
) + 1) * s
->bytes_stride
;
234 if (max_bytes_out
> af
->data
->len
) {
235 mp_msg(MSGT_AFILTER
, MSGL_V
, "[libaf] Reallocating memory in module %s, "
236 "old len = %i, new len = %i\n",af
->info
->name
,af
->data
->len
,max_bytes_out
);
237 af
->data
->audio
= realloc(af
->data
->audio
, max_bytes_out
);
238 if (!af
->data
->audio
) {
239 mp_msg(MSGT_AFILTER
, MSGL_FATAL
, "[libaf] Could not allocate memory\n");
242 af
->data
->len
= max_bytes_out
;
245 offset_in
= fill_queue(af
, data
, 0);
246 pout
= af
->data
->audio
;
247 while (s
->bytes_queued
>= s
->bytes_queue
) {
253 if (s
->output_overlap
) {
254 if (s
->best_overlap_offset
)
255 bytes_off
= s
->best_overlap_offset(s
);
256 s
->output_overlap(s
, pout
, bytes_off
);
258 memcpy(pout
+ s
->bytes_overlap
,
259 s
->buf_queue
+ bytes_off
+ s
->bytes_overlap
,
261 pout
+= s
->bytes_stride
;
264 memcpy(s
->buf_overlap
,
265 s
->buf_queue
+ bytes_off
+ s
->bytes_stride
,
267 tf
= s
->frames_stride_scaled
+ s
->frames_stride_error
;
269 s
->frames_stride_error
= tf
- ti
;
270 s
->bytes_to_slide
= ti
* s
->bytes_per_frame
;
272 offset_in
+= fill_queue(af
, data
, offset_in
);
275 // This filter can have a negative delay when scale > 1:
276 // output corresponding to some length of input can be decided and written
277 // after receiving only a part of that input.
278 af
->delay
= s
->bytes_queued
- s
->bytes_to_slide
;
280 data
->audio
= af
->data
->audio
;
281 data
->len
= pout
- (int8_t *)af
->data
->audio
;
285 // Initialization and runtime control
286 static int control(struct af_instance_s
* af
, int cmd
, void* arg
)
288 af_scaletempo_t
* s
= af
->setup
;
290 case AF_CONTROL_REINIT
:{
291 af_data_t
* data
= (af_data_t
*)arg
;
292 float srate
= data
->rate
/ 1000;
296 int frames_stride
, frames_overlap
;
299 mp_msg(MSGT_AFILTER
, MSGL_V
,
300 "[scaletempo] %.3f speed * %.3f scale_nominal = %.3f\n",
301 s
->speed
, s
->scale_nominal
, s
->scale
);
303 if (s
->scale
== 1.0) {
304 if (s
->speed_tempo
&& s
->speed_pitch
)
306 memcpy(af
->data
, data
, sizeof(af_data_t
));
307 return af_test_output(af
, data
);
310 af
->data
->rate
= data
->rate
;
311 af
->data
->nch
= data
->nch
;
312 if ( data
->format
== AF_FORMAT_S16_LE
313 || data
->format
== AF_FORMAT_S16_BE
) {
315 af
->data
->format
= AF_FORMAT_S16_NE
;
316 af
->data
->bps
= bps
= 2;
318 af
->data
->format
= AF_FORMAT_FLOAT_NE
;
319 af
->data
->bps
= bps
= 4;
322 frames_stride
= srate
* s
->ms_stride
;
323 s
->bytes_stride
= frames_stride
* bps
* nch
;
324 s
->bytes_stride_scaled
= s
->scale
* s
->bytes_stride
;
325 s
->frames_stride_scaled
= s
->scale
* frames_stride
;
326 s
->frames_stride_error
= 0;
327 af
->mul
= (double)s
->bytes_stride
/ s
->bytes_stride_scaled
;
329 frames_overlap
= frames_stride
* s
->percent_overlap
;
330 if (frames_overlap
<= 0) {
331 s
->bytes_standing
= s
->bytes_stride
;
332 s
->samples_standing
= s
->bytes_standing
/ bps
;
333 s
->output_overlap
= NULL
;
335 s
->samples_overlap
= frames_overlap
* nch
;
336 s
->bytes_overlap
= frames_overlap
* nch
* bps
;
337 s
->bytes_standing
= s
->bytes_stride
- s
->bytes_overlap
;
338 s
->samples_standing
= s
->bytes_standing
/ bps
;
339 s
->buf_overlap
= realloc(s
->buf_overlap
, s
->bytes_overlap
);
340 s
->table_blend
= realloc(s
->table_blend
, s
->bytes_overlap
* 4);
341 if(!s
->buf_overlap
|| !s
->table_blend
) {
342 mp_msg(MSGT_AFILTER
, MSGL_FATAL
, "[scaletempo] Out of memory\n");
345 memset(s
->buf_overlap
, 0, s
->bytes_overlap
);
347 int32_t* pb
= s
->table_blend
;
349 for (i
=0; i
<frames_overlap
; i
++) {
350 int32_t v
= blend
/ frames_overlap
;
351 for (j
=0; j
<nch
; j
++) {
354 blend
+= 65536; // 2^16
356 s
->output_overlap
= output_overlap_s16
;
358 float* pb
= s
->table_blend
;
359 for (i
=0; i
<frames_overlap
; i
++) {
360 float v
= i
/ (float)frames_overlap
;
361 for (j
=0; j
<nch
; j
++) {
365 s
->output_overlap
= output_overlap_float
;
369 s
->frames_search
= (frames_overlap
> 1) ? srate
* s
->ms_search
: 0;
370 if (s
->frames_search
<= 0) {
371 s
->best_overlap_offset
= NULL
;
374 int64_t t
= frames_overlap
;
375 int32_t n
= 8589934588LL / (t
* t
); // 4 * (2^31 - 1) / t^2
377 s
->buf_pre_corr
= realloc(s
->buf_pre_corr
, s
->bytes_overlap
* 2 + UNROLL_PADDING
);
378 s
->table_window
= realloc(s
->table_window
, s
->bytes_overlap
* 2 - nch
* bps
* 2);
379 if(!s
->buf_pre_corr
|| !s
->table_window
) {
380 mp_msg(MSGT_AFILTER
, MSGL_FATAL
, "[scaletempo] Out of memory\n");
383 memset((char *)s
->buf_pre_corr
+ s
->bytes_overlap
* 2, 0, UNROLL_PADDING
);
384 pw
= s
->table_window
;
385 for (i
=1; i
<frames_overlap
; i
++) {
386 int32_t v
= ( i
* (t
- i
) * n
) >> 15;
387 for (j
=0; j
<nch
; j
++) {
391 s
->best_overlap_offset
= best_overlap_offset_s16
;
394 s
->buf_pre_corr
= realloc(s
->buf_pre_corr
, s
->bytes_overlap
);
395 s
->table_window
= realloc(s
->table_window
, s
->bytes_overlap
- nch
* bps
);
396 if(!s
->buf_pre_corr
|| !s
->table_window
) {
397 mp_msg(MSGT_AFILTER
, MSGL_FATAL
, "[scaletempo] Out of memory\n");
400 pw
= s
->table_window
;
401 for (i
=1; i
<frames_overlap
; i
++) {
402 float v
= i
* (frames_overlap
- i
);
403 for (j
=0; j
<nch
; j
++) {
407 s
->best_overlap_offset
= best_overlap_offset_float
;
411 s
->bytes_per_frame
= bps
* nch
;
412 s
->num_channels
= nch
;
415 = (s
->frames_search
+ frames_stride
+ frames_overlap
) * bps
* nch
;
416 s
->buf_queue
= realloc(s
->buf_queue
, s
->bytes_queue
+ UNROLL_PADDING
);
418 mp_msg(MSGT_AFILTER
, MSGL_FATAL
, "[scaletempo] Out of memory\n");
422 mp_msg (MSGT_AFILTER
, MSGL_DBG2
, "[scaletempo] "
423 "%.2f stride_in, %i stride_out, %i standing, "
424 "%i overlap, %i search, %i queue, %s mode\n",
425 s
->frames_stride_scaled
,
426 (int)(s
->bytes_stride
/ nch
/ bps
),
427 (int)(s
->bytes_standing
/ nch
/ bps
),
428 (int)(s
->bytes_overlap
/ nch
/ bps
),
430 (int)(s
->bytes_queue
/ nch
/ bps
),
431 (use_int
?"s16":"float"));
433 return af_test_output(af
, (af_data_t
*)arg
);
435 case AF_CONTROL_PLAYBACK_SPEED
| AF_CONTROL_SET
:{
436 if (s
->speed_tempo
) {
437 if (s
->speed_pitch
) {
440 s
->speed
= *(float*)arg
;
441 s
->scale
= s
->speed
* s
->scale_nominal
;
443 if (s
->speed_pitch
) {
444 s
->speed
= 1 / *(float*)arg
;
445 s
->scale
= s
->speed
* s
->scale_nominal
;
451 case AF_CONTROL_SCALETEMPO_AMOUNT
| AF_CONTROL_SET
:{
452 s
->scale
= *(float*)arg
;
453 s
->scale
= s
->speed
* s
->scale_nominal
;
456 case AF_CONTROL_SCALETEMPO_AMOUNT
| AF_CONTROL_GET
:
457 *(float*)arg
= s
->scale
;
459 case AF_CONTROL_COMMAND_LINE
:{
462 {"scale", OPT_ARG_FLOAT
, &s
->scale_nominal
, NULL
},
463 {"stride", OPT_ARG_FLOAT
, &s
->ms_stride
, NULL
},
464 {"overlap", OPT_ARG_FLOAT
, &s
->percent_overlap
, NULL
},
465 {"search", OPT_ARG_FLOAT
, &s
->ms_search
, NULL
},
466 {"speed", OPT_ARG_STR
, &speed
, NULL
},
469 if (subopt_parse(arg
, subopts
) != 0) {
472 if (s
->scale_nominal
<= 0) {
473 mp_msg(MSGT_AFILTER
, MSGL_ERR
, "[scaletempo] %s: %s: scale > 0\n",
474 mp_gtext("error parsing command line"),
475 mp_gtext("value out of range"));
478 if (s
->ms_stride
<= 0) {
479 mp_msg(MSGT_AFILTER
, MSGL_ERR
, "[scaletempo] %s: %s: stride > 0\n",
480 mp_gtext("error parsing command line"),
481 mp_gtext("value out of range"));
484 if (s
->percent_overlap
< 0 || s
->percent_overlap
> 1) {
485 mp_msg(MSGT_AFILTER
, MSGL_ERR
,
486 "[scaletempo] %s: %s: 0 <= overlap <= 1\n",
487 mp_gtext("error parsing command line"),
488 mp_gtext("value out of range"));
491 if (s
->ms_search
< 0) {
492 mp_msg(MSGT_AFILTER
, MSGL_ERR
, "[scaletempo] %s: %s: search >= 0\n",
493 mp_gtext("error parsing command line"),
494 mp_gtext("value out of range"));
498 if (strcmp(speed
.str
, "pitch") == 0) {
501 } else if (strcmp(speed
.str
, "tempo") == 0) {
504 } else if (strcmp(speed
.str
, "none") == 0) {
507 } else if (strcmp(speed
.str
, "both") == 0) {
511 mp_msg(MSGT_AFILTER
, MSGL_ERR
,
512 "[scaletempo] %s: %s: speed=[pitch|tempo|none|both]\n",
513 mp_gtext("error parsing command line"),
514 mp_gtext("value out of range"));
518 s
->scale
= s
->speed
* s
->scale_nominal
;
519 mp_msg(MSGT_AFILTER
, MSGL_DBG2
, "[scaletempo] %6.3f scale, %6.2f stride, %6.2f overlap, %6.2f search, speed = %s\n", s
->scale_nominal
, s
->ms_stride
, s
->percent_overlap
, s
->ms_search
, (s
->speed_tempo
?(s
->speed_pitch
?"tempo and speed":"tempo"):(s
->speed_pitch
?"pitch":"none")));
527 static void uninit(struct af_instance_s
* af
)
529 af_scaletempo_t
* s
= af
->setup
;
530 free(af
->data
->audio
);
533 free(s
->buf_overlap
);
534 free(s
->buf_pre_corr
);
535 free(s
->table_blend
);
536 free(s
->table_window
);
540 // Allocate memory and set function pointers
541 static int af_open(af_instance_t
* af
){
544 af
->control
= control
;
548 af
->data
= calloc(1,sizeof(af_data_t
));
549 af
->setup
= calloc(1,sizeof(af_scaletempo_t
));
550 if(af
->data
== NULL
|| af
->setup
== NULL
)
554 s
->scale
= s
->speed
= s
->scale_nominal
= 1.0;
558 s
->percent_overlap
= .20;
564 // Description of this filter
565 af_info_t af_info_scaletempo
= {
566 "Scale audio tempo while maintaining pitch",