git.sesse.net Git - vlc/blob - modules/audio_filter/resampler/bandlimited.c

   1 /*****************************************************************************
   2  * bandlimited.c : band-limited interpolation resampler
   3  *****************************************************************************
   4  * Copyright (C) 2002, 2006 the VideoLAN team
   5  * $Id$
   6  *
   7  * Authors: Gildas Bazin <gbazin@netcourrier.com>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  22  *****************************************************************************/
  23
  24 /*****************************************************************************
  25  * Preamble:
  26  *
  27  * This implementation of the band-limited interpolationis based on the
  28  * following paper:
  29  * http://ccrma-www.stanford.edu/~jos/resample/resample.html
  30  *
  31  * It uses a Kaiser-windowed sinc-function low-pass filter and the width of the
  32  * filter is 13 samples.
  33  *
  34  *****************************************************************************/
  35
  36 #ifdef HAVE_CONFIG_H
  37 # include "config.h"
  38 #endif
  39
  40 #include <vlc/vlc.h>
  41 #include <vlc_aout.h>
  42
  43 #include "bandlimited.h"
  44
  45 /*****************************************************************************
  46  * Local prototypes
  47  *****************************************************************************/
  48 static int  Create    ( vlc_object_t * );
  49 static void Close     ( vlc_object_t * );
  50 static void DoWork    ( aout_instance_t *, aout_filter_t *, aout_buffer_t *,
  51                         aout_buffer_t * );
  52
  53 static void FilterFloatUP( float Imp[], float ImpD[], uint16_t Nwing,
  54                            float *f_in, float *f_out, uint32_t ui_remainder,
  55                            uint32_t ui_output_rate, int16_t Inc,
  56                            int i_nb_channels );
  57
  58 static void FilterFloatUD( float Imp[], float ImpD[], uint16_t Nwing,
  59                            float *f_in, float *f_out, uint32_t ui_remainder,
  60                            uint32_t ui_output_rate, uint32_t ui_input_rate,
  61                            int16_t Inc, int i_nb_channels );
  62
  63 /*****************************************************************************
  64  * Local structures
  65  *****************************************************************************/
  66 struct aout_filter_sys_t
  67 {
  68     int32_t *p_buf;                        /* this filter introduces a delay */
  69     int i_buf_size;
  70
  71     int i_old_rate;
  72     double d_old_factor;
  73     int i_old_wing;
  74
  75     unsigned int i_remainder;                /* remainder of previous sample */
  76
  77     audio_date_t end_date;
  78 };
  79
  80 /*****************************************************************************
  81  * Module descriptor
  82  *****************************************************************************/
  83 vlc_module_begin();
  84     set_category( CAT_AUDIO );
  85     set_subcategory( SUBCAT_AUDIO_MISC );
  86     set_description( _("Audio filter for band-limited interpolation resampling") );
  87     set_capability( "audio filter", 20 );
  88     set_callbacks( Create, Close );
  89 vlc_module_end();
  90
  91 /*****************************************************************************
  92  * Create: allocate linear resampler
  93  *****************************************************************************/
  94 static int Create( vlc_object_t *p_this )
  95 {
  96     aout_filter_t * p_filter = (aout_filter_t *)p_this;
  97     double d_factor;
  98     int i_filter_wing;
  99
 100     if ( p_filter->input.i_rate == p_filter->output.i_rate
 101           || p_filter->input.i_format != p_filter->output.i_format
 102           || p_filter->input.i_physical_channels
 103               != p_filter->output.i_physical_channels
 104           || p_filter->input.i_original_channels
 105               != p_filter->output.i_original_channels
 106           || p_filter->input.i_format != VLC_FOURCC('f','l','3','2') )
 107     {
 108         return VLC_EGENERIC;
 109     }
 110
 111 #if !defined( __APPLE__ )
 112     if( !config_GetInt( p_this, "hq-resampling" ) )
 113     {
 114         return VLC_EGENERIC;
 115     }
 116 #endif
 117
 118     /* Allocate the memory needed to store the module's structure */
 119     p_filter->p_sys = malloc( sizeof(struct aout_filter_sys_t) );
 120     if( p_filter->p_sys == NULL )
 121     {
 122         msg_Err( p_filter, "out of memory" );
 123         return VLC_ENOMEM;
 124     }
 125
 126     /* Calculate worst case for the length of the filter wing */
 127     d_factor = (double)p_filter->output.i_rate
 128                         / p_filter->input.i_rate / AOUT_MAX_INPUT_RATE;
 129     i_filter_wing = ((SMALL_FILTER_NMULT + 1)/2.0)
 130                       * __MAX(1.0, 1.0/d_factor) + 10;
 131     p_filter->p_sys->i_buf_size = aout_FormatNbChannels( &p_filter->input ) *
 132         sizeof(int32_t) * 2 * i_filter_wing;
 133
 134     /* Allocate enough memory to buffer previous samples */
 135     p_filter->p_sys->p_buf = malloc( p_filter->p_sys->i_buf_size );
 136     if( p_filter->p_sys->p_buf == NULL )
 137     {
 138         msg_Err( p_filter, "out of memory" );
 139         return VLC_ENOMEM;
 140     }
 141
 142     p_filter->p_sys->i_old_wing = 0;
 143     p_filter->pf_do_work = DoWork;
 144
 145     /* We don't want a new buffer to be created because we're not sure we'll
 146      * actually need to resample anything. */
 147     p_filter->b_in_place = true;
 148
 149     return VLC_SUCCESS;
 150 }
 151
 152 /*****************************************************************************
 153  * Close: free our resources
 154  *****************************************************************************/
 155 static void Close( vlc_object_t * p_this )
 156 {
 157     aout_filter_t * p_filter = (aout_filter_t *)p_this;
 158     free( p_filter->p_sys->p_buf );
 159     free( p_filter->p_sys );
 160 }
 161
 162 /*****************************************************************************
 163  * DoWork: convert a buffer
 164  *****************************************************************************/
 165 static void DoWork( aout_instance_t * p_aout, aout_filter_t * p_filter,
 166                     aout_buffer_t * p_in_buf, aout_buffer_t * p_out_buf )
 167 {
 168     float *p_in, *p_in_orig, *p_out = (float *)p_out_buf->p_buffer;
 169
 170     int i_nb_channels = aout_FormatNbChannels( &p_filter->input );
 171     int i_in_nb = p_in_buf->i_nb_samples;
 172     int i_in, i_out = 0;
 173     double d_factor, d_scale_factor, d_old_scale_factor;
 174     int i_filter_wing;
 175 #if 0
 176     int i;
 177 #endif
 178
 179     /* Check if we really need to run the resampler */
 180     if( p_aout->mixer.mixer.i_rate == p_filter->input.i_rate )
 181     {
 182         if( /*p_filter->b_continuity && /--* What difference does it make ? :) */
 183             p_filter->p_sys->i_old_wing &&
 184             p_in_buf->i_size >=
 185               p_in_buf->i_nb_bytes + p_filter->p_sys->i_old_wing *
 186               p_filter->input.i_bytes_per_frame )
 187         {
 188             /* output the whole thing with the samples from last time */
 189             memmove( ((float *)(p_in_buf->p_buffer)) +
 190                      i_nb_channels * p_filter->p_sys->i_old_wing,
 191                      p_in_buf->p_buffer, p_in_buf->i_nb_bytes );
 192             memcpy( p_in_buf->p_buffer, p_filter->p_sys->p_buf +
 193                     i_nb_channels * p_filter->p_sys->i_old_wing,
 194                     p_filter->p_sys->i_old_wing *
 195                     p_filter->input.i_bytes_per_frame );
 196
 197             p_out_buf->i_nb_samples = p_in_buf->i_nb_samples +
 198                 p_filter->p_sys->i_old_wing;
 199
 200             p_out_buf->start_date = aout_DateGet( &p_filter->p_sys->end_date );
 201             p_out_buf->end_date =
 202                 aout_DateIncrement( &p_filter->p_sys->end_date,
 203                                     p_out_buf->i_nb_samples );
 204
 205             p_out_buf->i_nb_bytes = p_out_buf->i_nb_samples *
 206                 p_filter->input.i_bytes_per_frame;
 207         }
 208         p_filter->b_continuity = false;
 209         p_filter->p_sys->i_old_wing = 0;
 210         return;
 211     }
 212
 213     if( !p_filter->b_continuity )
 214     {
 215         /* Continuity in sound samples has been broken, we'd better reset
 216          * everything. */
 217         p_filter->b_continuity = true;
 218         p_filter->p_sys->i_remainder = 0;
 219         aout_DateInit( &p_filter->p_sys->end_date, p_filter->output.i_rate );
 220         aout_DateSet( &p_filter->p_sys->end_date, p_in_buf->start_date );
 221         p_filter->p_sys->i_old_rate   = p_filter->input.i_rate;
 222         p_filter->p_sys->d_old_factor = 1;
 223         p_filter->p_sys->i_old_wing   = 0;
 224     }
 225
 226 #if 0
 227     msg_Err( p_filter, "old rate: %i, old factor: %f, old wing: %i, i_in: %i",
 228              p_filter->p_sys->i_old_rate, p_filter->p_sys->d_old_factor,
 229              p_filter->p_sys->i_old_wing, i_in_nb );
 230 #endif
 231
 232     /* Prepare the source buffer */
 233     i_in_nb += (p_filter->p_sys->i_old_wing * 2);
 234 #ifdef HAVE_ALLOCA
 235     p_in = p_in_orig = (float *)alloca( i_in_nb *
 236                                         p_filter->input.i_bytes_per_frame );
 237 #else
 238     p_in = p_in_orig = (float *)malloc( i_in_nb *
 239                                         p_filter->input.i_bytes_per_frame );
 240 #endif
 241     if( p_in == NULL )
 242     {
 243         return;
 244     }
 245
 246     /* Copy all our samples in p_in */
 247     if( p_filter->p_sys->i_old_wing )
 248     {
 249         p_aout->p_libvlc->pf_memcpy( p_in, p_filter->p_sys->p_buf,
 250                                   p_filter->p_sys->i_old_wing * 2 *
 251                                   p_filter->input.i_bytes_per_frame );
 252     }
 253     p_aout->p_libvlc->pf_memcpy( p_in + p_filter->p_sys->i_old_wing * 2 *
 254                               i_nb_channels, p_in_buf->p_buffer,
 255                               p_in_buf->i_nb_samples *
 256                               p_filter->input.i_bytes_per_frame );
 257
 258     /* Make sure the output buffer is reset */
 259     memset( p_out, 0, p_out_buf->i_size );
 260
 261     /* Calculate the new length of the filter wing */
 262     d_factor = (double)p_aout->mixer.mixer.i_rate / p_filter->input.i_rate;
 263     i_filter_wing = ((SMALL_FILTER_NMULT+1)/2.0) * __MAX(1.0,1.0/d_factor) + 1;
 264
 265     /* Account for increased filter gain when using factors less than 1 */
 266     d_old_scale_factor = SMALL_FILTER_SCALE *
 267         p_filter->p_sys->d_old_factor + 0.5;
 268     d_scale_factor = SMALL_FILTER_SCALE * d_factor + 0.5;
 269
 270     /* Apply the old rate until we have enough samples for the new one */
 271     i_in = p_filter->p_sys->i_old_wing;
 272     p_in += p_filter->p_sys->i_old_wing * i_nb_channels;
 273     for( ; i_in < i_filter_wing &&
 274            (i_in + p_filter->p_sys->i_old_wing) < i_in_nb; i_in++ )
 275     {
 276         if( p_filter->p_sys->d_old_factor == 1 )
 277         {
 278             /* Just copy the samples */
 279             memcpy( p_out, p_in,
 280                     p_filter->input.i_bytes_per_frame );
 281             p_in += i_nb_channels;
 282             p_out += i_nb_channels;
 283             i_out++;
 284             continue;
 285         }
 286
 287         while( p_filter->p_sys->i_remainder < p_filter->output.i_rate )
 288         {
 289
 290             if( p_filter->p_sys->d_old_factor >= 1 )
 291             {
 292                 /* FilterFloatUP() is faster if we can use it */
 293
 294                 /* Perform left-wing inner product */
 295                 FilterFloatUP( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 296                                SMALL_FILTER_NWING, p_in, p_out,
 297                                p_filter->p_sys->i_remainder,
 298                                p_filter->output.i_rate,
 299                                -1, i_nb_channels );
 300                 /* Perform right-wing inner product */
 301                 FilterFloatUP( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 302                                SMALL_FILTER_NWING, p_in + i_nb_channels, p_out,
 303                                p_filter->output.i_rate -
 304                                p_filter->p_sys->i_remainder,
 305                                p_filter->output.i_rate,
 306                                1, i_nb_channels );
 307
 308 #if 0
 309                 /* Normalize for unity filter gain */
 310                 for( i = 0; i < i_nb_channels; i++ )
 311                 {
 312                     *(p_out+i) *= d_old_scale_factor;
 313                 }
 314 #endif
 315
 316                 /* Sanity check */
 317                 if( p_out_buf->i_size/p_filter->input.i_bytes_per_frame
 318                     <= (unsigned int)i_out+1 )
 319                 {
 320                     p_out += i_nb_channels;
 321                     i_out++;
 322                     p_filter->p_sys->i_remainder += p_filter->input.i_rate;
 323                     break;
 324                 }
 325             }
 326             else
 327             {
 328                 /* Perform left-wing inner product */
 329                 FilterFloatUD( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 330                                SMALL_FILTER_NWING, p_in, p_out,
 331                                p_filter->p_sys->i_remainder,
 332                                p_filter->output.i_rate, p_filter->input.i_rate,
 333                                -1, i_nb_channels );
 334                 /* Perform right-wing inner product */
 335                 FilterFloatUD( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 336                                SMALL_FILTER_NWING, p_in + i_nb_channels, p_out,
 337                                p_filter->output.i_rate -
 338                                p_filter->p_sys->i_remainder,
 339                                p_filter->output.i_rate, p_filter->input.i_rate,
 340                                1, i_nb_channels );
 341             }
 342
 343             p_out += i_nb_channels;
 344             i_out++;
 345
 346             p_filter->p_sys->i_remainder += p_filter->input.i_rate;
 347         }
 348
 349         p_in += i_nb_channels;
 350         p_filter->p_sys->i_remainder -= p_filter->output.i_rate;
 351     }
 352
 353     /* Apply the new rate for the rest of the samples */
 354     if( i_in < i_in_nb - i_filter_wing )
 355     {
 356         p_filter->p_sys->i_old_rate   = p_filter->input.i_rate;
 357         p_filter->p_sys->d_old_factor = d_factor;
 358         p_filter->p_sys->i_old_wing   = i_filter_wing;
 359     }
 360     for( ; i_in < i_in_nb - i_filter_wing; i_in++ )
 361     {
 362         while( p_filter->p_sys->i_remainder < p_filter->output.i_rate )
 363         {
 364
 365             if( d_factor >= 1 )
 366             {
 367                 /* FilterFloatUP() is faster if we can use it */
 368
 369                 /* Perform left-wing inner product */
 370                 FilterFloatUP( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 371                                SMALL_FILTER_NWING, p_in, p_out,
 372                                p_filter->p_sys->i_remainder,
 373                                p_filter->output.i_rate,
 374                                -1, i_nb_channels );
 375
 376                 /* Perform right-wing inner product */
 377                 FilterFloatUP( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 378                                SMALL_FILTER_NWING, p_in + i_nb_channels, p_out,
 379                                p_filter->output.i_rate -
 380                                p_filter->p_sys->i_remainder,
 381                                p_filter->output.i_rate,
 382                                1, i_nb_channels );
 383
 384 #if 0
 385                 /* Normalize for unity filter gain */
 386                 for( i = 0; i < i_nb_channels; i++ )
 387                 {
 388                     *(p_out+i) *= d_old_scale_factor;
 389                 }
 390 #endif
 391                 /* Sanity check */
 392                 if( p_out_buf->i_size/p_filter->input.i_bytes_per_frame
 393                     <= (unsigned int)i_out+1 )
 394                 {
 395                     p_out += i_nb_channels;
 396                     i_out++;
 397                     p_filter->p_sys->i_remainder += p_filter->input.i_rate;
 398                     break;
 399                 }
 400             }
 401             else
 402             {
 403                 /* Perform left-wing inner product */
 404                 FilterFloatUD( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 405                                SMALL_FILTER_NWING, p_in, p_out,
 406                                p_filter->p_sys->i_remainder,
 407                                p_filter->output.i_rate, p_filter->input.i_rate,
 408                                -1, i_nb_channels );
 409                 /* Perform right-wing inner product */
 410                 FilterFloatUD( SMALL_FILTER_FLOAT_IMP, SMALL_FILTER_FLOAT_IMPD,
 411                                SMALL_FILTER_NWING, p_in + i_nb_channels, p_out,
 412                                p_filter->output.i_rate -
 413                                p_filter->p_sys->i_remainder,
 414                                p_filter->output.i_rate, p_filter->input.i_rate,
 415                                1, i_nb_channels );
 416             }
 417
 418             p_out += i_nb_channels;
 419             i_out++;
 420
 421             p_filter->p_sys->i_remainder += p_filter->input.i_rate;
 422         }
 423
 424         p_in += i_nb_channels;
 425         p_filter->p_sys->i_remainder -= p_filter->output.i_rate;
 426     }
 427
 428     /* Buffer i_filter_wing * 2 samples for next time */
 429     if( p_filter->p_sys->i_old_wing )
 430     {
 431         memcpy( p_filter->p_sys->p_buf,
 432                 p_in_orig + (i_in_nb - 2 * p_filter->p_sys->i_old_wing) *
 433                 i_nb_channels, (2 * p_filter->p_sys->i_old_wing) *
 434                 p_filter->input.i_bytes_per_frame );
 435     }
 436
 437 #if 0
 438     msg_Err( p_filter, "p_out size: %i, nb bytes out: %i", p_out_buf->i_size,
 439              i_out * p_filter->input.i_bytes_per_frame );
 440 #endif
 441
 442     /* Free the temp buffer */
 443 #ifndef HAVE_ALLOCA
 444     free( p_in_orig );
 445 #endif
 446
 447     /* Finalize aout buffer */
 448     p_out_buf->i_nb_samples = i_out;
 449     p_out_buf->start_date = aout_DateGet( &p_filter->p_sys->end_date );
 450     p_out_buf->end_date = aout_DateIncrement( &p_filter->p_sys->end_date,
 451                                               p_out_buf->i_nb_samples );
 452
 453     p_out_buf->i_nb_bytes = p_out_buf->i_nb_samples *
 454         i_nb_channels * sizeof(int32_t);
 455
 456 }
 457
 458 void FilterFloatUP( float Imp[], float ImpD[], uint16_t Nwing, float *p_in,
 459                     float *p_out, uint32_t ui_remainder,
 460                     uint32_t ui_output_rate, int16_t Inc, int i_nb_channels )
 461 {
 462     float *Hp, *Hdp, *End;
 463     float t, temp;
 464     uint32_t ui_linear_remainder;
 465     int i;
 466
 467     Hp = &Imp[(ui_remainder<<Nhc)/ui_output_rate];
 468     Hdp = &ImpD[(ui_remainder<<Nhc)/ui_output_rate];
 469
 470     End = &Imp[Nwing];
 471
 472     ui_linear_remainder = (ui_remainder<<Nhc) -
 473                             (ui_remainder<<Nhc)/ui_output_rate*ui_output_rate;
 474
 475     if (Inc == 1)               /* If doing right wing...              */
 476     {                           /* ...drop extra coeff, so when Ph is  */
 477         End--;                  /*    0.5, we don't do too many mult's */
 478         if (ui_remainder == 0)  /* If the phase is zero...           */
 479         {                       /* ...then we've already skipped the */
 480             Hp += Npc;          /*    first sample, so we must also  */
 481             Hdp += Npc;         /*    skip ahead in Imp[] and ImpD[] */
 482         }
 483     }
 484
 485     while (Hp < End) {
 486         t = *Hp;                /* Get filter coeff */
 487                                 /* t is now interp'd filter coeff */
 488         t += *Hdp * ui_linear_remainder / ui_output_rate / Npc;
 489         for( i = 0; i < i_nb_channels; i++ )
 490         {
 491             temp = t;
 492             temp *= *(p_in+i);  /* Mult coeff by input sample */
 493             *(p_out+i) += temp; /* The filter output */
 494         }
 495         Hdp += Npc;             /* Filter coeff differences step */
 496         Hp += Npc;              /* Filter coeff step */
 497         p_in += (Inc * i_nb_channels); /* Input signal step */
 498     }
 499 }
 500
 501 void FilterFloatUD( float Imp[], float ImpD[], uint16_t Nwing, float *p_in,
 502                     float *p_out, uint32_t ui_remainder,
 503                     uint32_t ui_output_rate, uint32_t ui_input_rate,
 504                     int16_t Inc, int i_nb_channels )
 505 {
 506     float *Hp, *Hdp, *End;
 507     float t, temp;
 508     uint32_t ui_linear_remainder;
 509     int i, ui_counter = 0;
 510
 511     Hp = Imp + (ui_remainder<<Nhc) / ui_input_rate;
 512     Hdp = ImpD  + (ui_remainder<<Nhc) / ui_input_rate;
 513
 514     End = &Imp[Nwing];
 515
 516     if (Inc == 1)               /* If doing right wing...              */
 517     {                           /* ...drop extra coeff, so when Ph is  */
 518         End--;                  /*    0.5, we don't do too many mult's */
 519         if (ui_remainder == 0)  /* If the phase is zero...           */
 520         {                       /* ...then we've already skipped the */
 521             Hp = Imp +          /* first sample, so we must also  */
 522                   (ui_output_rate << Nhc) / ui_input_rate;
 523             Hdp = ImpD +        /* skip ahead in Imp[] and ImpD[] */
 524                   (ui_output_rate << Nhc) / ui_input_rate;
 525             ui_counter++;
 526         }
 527     }
 528
 529     while (Hp < End) {
 530         t = *Hp;                /* Get filter coeff */
 531                                 /* t is now interp'd filter coeff */
 532         ui_linear_remainder =
 533           ((ui_output_rate * ui_counter + ui_remainder)<< Nhc) -
 534           ((ui_output_rate * ui_counter + ui_remainder)<< Nhc) /
 535           ui_input_rate * ui_input_rate;
 536         t += *Hdp * ui_linear_remainder / ui_input_rate / Npc;
 537         for( i = 0; i < i_nb_channels; i++ )
 538         {
 539             temp = t;
 540             temp *= *(p_in+i);  /* Mult coeff by input sample */
 541             *(p_out+i) += temp; /* The filter output */
 542         }
 543
 544         ui_counter++;
 545
 546         /* Filter coeff step */
 547         Hp = Imp + ((ui_output_rate * ui_counter + ui_remainder)<< Nhc)
 548                     / ui_input_rate;
 549         /* Filter coeff differences step */
 550         Hdp = ImpD + ((ui_output_rate * ui_counter + ui_remainder)<< Nhc)
 551                      / ui_input_rate;
 552
 553         p_in += (Inc * i_nb_channels); /* Input signal step */
 554     }
 555 }