git.sesse.net Git - vlc/blob - src/ac3_decoder/ac3_imdct.c

   1 /*****************************************************************************
   2  * ac3_imdct.c: ac3 DCT
   3  *****************************************************************************
   4  * Copyright (C) 1999, 2000 VideoLAN
   5  * $Id: ac3_imdct.c,v 1.19 2001/05/14 15:58:04 reno Exp $
   6  *
   7  * Authors: Michel Kaempf <maxx@via.ecp.fr>
   8  *          Aaron Holtzman <aholtzma@engr.uvic.ca>
   9  *          Renaud Dartus <reno@videolan.org>
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2 of the License, or
  14  * (at your option) any later version.
  15  *
  16  * This program is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  * GNU General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU General Public License
  22  * along with this program; if not, write to the Free Software
  23  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  24  *****************************************************************************/
  25
  26 #include "defs.h"
  27
  28 #include <string.h>                                              /* memcpy() */
  29
  30 #include <math.h>
  31 #include <stdio.h>
  32
  33 #include "config.h"
  34 #include "common.h"
  35 #include "threads.h"
  36 #include "mtime.h"
  37
  38 #include "stream_control.h"
  39 #include "input_ext-dec.h"
  40
  41 #include "ac3_decoder.h"
  42
  43 #include "ac3_imdct_c.h"                                     /* imdct_init_c */
  44 #include "ac3_imdct_sse.h"                                 /* imdct_init_sse */
  45
  46 #include "tests.h"                                                /* TestCPU */
  47
  48 #ifndef M_PI
  49 #   define M_PI 3.14159265358979323846
  50 #endif
  51
  52
  53 void imdct_init(imdct_t * p_imdct)
  54 {
  55         int i;
  56         float scale = 181.019;
  57 #if 0
  58         if ( TestCPU (CPU_CAPABILITY_SSE) )
  59     {
  60         imdct_init_sse (p_imdct);
  61     }
  62     else
  63 #endif
  64     {
  65         imdct_init_c (p_imdct);
  66     }
  67
  68         /* More twiddle factors to turn IFFT into IMDCT */
  69         for (i=0; i < 64; i++) {
  70                 p_imdct->xcos2[i] = cos(2.0f * M_PI * (8*i+1)/(4*N)) * scale;
  71                 p_imdct->xsin2[i] = sin(2.0f * M_PI * (8*i+1)/(4*N)) * scale;
  72         }
  73 }
  74
  75 void imdct (ac3dec_t * p_ac3dec, s16 * buffer)
  76 {
  77         int   i;
  78         int   doable = 0;
  79         float *center=NULL, *left, *right, *left_sur, *right_sur;
  80         float *delay_left, *delay_right;
  81         float *delay1_left, *delay1_right, *delay1_center, *delay1_sr, *delay1_sl;
  82         float right_tmp, left_tmp;
  83         void (*do_imdct)(imdct_t * p_imdct, float data[], float delay[]);
  84
  85         /* test if dm in frequency is doable */
  86         if (!(doable = p_ac3dec->audblk.blksw[0]))
  87     {
  88                 do_imdct = p_ac3dec->imdct.imdct_do_512;
  89     }
  90         else
  91     {
  92                 do_imdct = imdct_do_256; /* There is only a C function */
  93     }
  94
  95         /* downmix in the frequency domain if all the channels
  96          * use the same imdct */
  97         for (i=0; i < p_ac3dec->bsi.nfchans; i++)
  98     {
  99                 if (doable != p_ac3dec->audblk.blksw[i])
 100         {
 101                         do_imdct = NULL;
 102                         break;
 103                 }
 104         }
 105
 106     if (do_imdct)
 107     {
 108                 /* dowmix first and imdct */
 109         switch(p_ac3dec->bsi.acmod)
 110         {
 111                 case 7:         /* 3/2 */
 112                         p_ac3dec->downmix.downmix_3f_2r_to_2ch (p_ac3dec->samples[0], &p_ac3dec->dm_par);
 113                         break;
 114                 case 6:         /* 2/2 */
 115                         p_ac3dec->downmix.downmix_2f_2r_to_2ch (p_ac3dec->samples[0], &p_ac3dec->dm_par);
 116                         break;
 117                 case 5:         /* 3/1 */
 118                         p_ac3dec->downmix.downmix_3f_1r_to_2ch (p_ac3dec->samples[0], &p_ac3dec->dm_par);
 119                         break;
 120                 case 4:         /* 2/1 */
 121                         p_ac3dec->downmix.downmix_2f_1r_to_2ch (p_ac3dec->samples[0], &p_ac3dec->dm_par);
 122                         break;
 123                 case 3:         /* 3/0 */
 124                         p_ac3dec->downmix.downmix_3f_0r_to_2ch (p_ac3dec->samples[0], &p_ac3dec->dm_par);
 125                         break;
 126                 case 2:
 127                         break;
 128                 default:        /* 1/0 */
 129 //                      if (p_ac3dec->bsi.acmod == 1)
 130                                 center = p_ac3dec->samples[0];
 131 //                      else if (p_ac3dec->bsi.acmod == 0)
 132 //                  center = samples[ac3_config.dual_mono_ch_sel];
 133                 do_imdct(&p_ac3dec->imdct, center, p_ac3dec->imdct.delay[0]); /* no downmix*/
 134
 135                         p_ac3dec->downmix.stream_sample_1ch_to_s16 (buffer, center);
 136
 137                     return;
 138                 break;
 139         }
 140
 141                 do_imdct (&p_ac3dec->imdct, p_ac3dec->samples[0], p_ac3dec->imdct.delay[0]);
 142                 do_imdct (&p_ac3dec->imdct, p_ac3dec->samples[1], p_ac3dec->imdct.delay[1]);
 143                 p_ac3dec->downmix.stream_sample_2ch_to_s16(buffer, p_ac3dec->samples[0], p_ac3dec->samples[1]);
 144
 145         } else {
 146         /* imdct and then downmix
 147                  * delay and samples should be saved and mixed
 148                  * fprintf(stderr, "time domain downmix\n"); */
 149                 for (i=0; i<p_ac3dec->bsi.nfchans; i++)
 150         {
 151                         if (p_ac3dec->audblk.blksw[i])
 152                 /* There is only a C function */
 153                                 imdct_do_256_nol (&p_ac3dec->imdct, p_ac3dec->samples[i], p_ac3dec->imdct.delay1[i]);
 154                         else
 155                                 p_ac3dec->imdct.imdct_do_512_nol (&p_ac3dec->imdct, p_ac3dec->samples[i], p_ac3dec->imdct.delay1[i]);
 156                 }
 157
 158                 /* mix the sample, overlap */
 159                 switch(p_ac3dec->bsi.acmod)
 160         {
 161                 case 7:         /* 3/2 */
 162                         left = p_ac3dec->samples[0];
 163                         center = p_ac3dec->samples[1];
 164                         right = p_ac3dec->samples[2];
 165                         left_sur = p_ac3dec->samples[3];
 166                         right_sur = p_ac3dec->samples[4];
 167                         delay_left = p_ac3dec->imdct.delay[0];
 168                         delay_right = p_ac3dec->imdct.delay[1];
 169                         delay1_left = p_ac3dec->imdct.delay1[0];
 170                         delay1_center = p_ac3dec->imdct.delay1[1];
 171                         delay1_right = p_ac3dec->imdct.delay1[2];
 172                 delay1_sl = p_ac3dec->imdct.delay1[3];
 173                         delay1_sr = p_ac3dec->imdct.delay1[4];
 174
 175                         for (i = 0; i < 256; i++) {
 176                                 left_tmp = p_ac3dec->dm_par.unit * *left++  + p_ac3dec->dm_par.clev * *center  + p_ac3dec->dm_par.slev * *left_sur++;
 177                                 right_tmp= p_ac3dec->dm_par.unit * *right++ + p_ac3dec->dm_par.clev * *center++ + p_ac3dec->dm_par.slev * *right_sur++;
 178                                 *buffer++ = (s16)(left_tmp + *delay_left);
 179                                 *buffer++ = (s16)(right_tmp + *delay_right);
 180                                 *delay_left++ = p_ac3dec->dm_par.unit * *delay1_left++  + p_ac3dec->dm_par.clev * *delay1_center  + p_ac3dec->dm_par.slev * *delay1_sl++;
 181                                 *delay_right++ = p_ac3dec->dm_par.unit * *delay1_right++ + p_ac3dec->dm_par.clev * *center++ + p_ac3dec->dm_par.slev * *delay1_sr++;
 182                         }
 183                         break;
 184                 case 6:         /* 2/2 */
 185                         left = p_ac3dec->samples[0];
 186                         right = p_ac3dec->samples[1];
 187                         left_sur = p_ac3dec->samples[2];
 188                         right_sur = p_ac3dec->samples[3];
 189                         delay_left = p_ac3dec->imdct.delay[0];
 190                         delay_right = p_ac3dec->imdct.delay[1];
 191                         delay1_left = p_ac3dec->imdct.delay1[0];
 192                         delay1_right = p_ac3dec->imdct.delay1[1];
 193                         delay1_sl = p_ac3dec->imdct.delay1[2];
 194                         delay1_sr = p_ac3dec->imdct.delay1[3];
 195
 196                         for (i = 0; i < 256; i++) {
 197                                 left_tmp = p_ac3dec->dm_par.unit * *left++  + p_ac3dec->dm_par.slev * *left_sur++;
 198                                 right_tmp= p_ac3dec->dm_par.unit * *right++ + p_ac3dec->dm_par.slev * *right_sur++;
 199                                 *buffer++ = (s16)(left_tmp + *delay_left);
 200                                 *buffer++ = (s16)(right_tmp + *delay_right);
 201                                 *delay_left++ = p_ac3dec->dm_par.unit * *delay1_left++  + p_ac3dec->dm_par.slev * *delay1_sl++;
 202                                 *delay_right++ = p_ac3dec->dm_par.unit * *delay1_right++ + p_ac3dec->dm_par.slev * *delay1_sr++;
 203                         }
 204                         break;
 205                 case 5:         /* 3/1 */
 206                         left = p_ac3dec->samples[0];
 207                         center = p_ac3dec->samples[1];
 208                         right = p_ac3dec->samples[2];
 209                         right_sur = p_ac3dec->samples[3];
 210                         delay_left = p_ac3dec->imdct.delay[0];
 211                         delay_right = p_ac3dec->imdct.delay[1];
 212                         delay1_left = p_ac3dec->imdct.delay1[0];
 213                         delay1_center = p_ac3dec->imdct.delay1[1];
 214                         delay1_right = p_ac3dec->imdct.delay1[2];
 215                         delay1_sl = p_ac3dec->imdct.delay1[3];
 216
 217                         for (i = 0; i < 256; i++) {
 218                                 left_tmp = p_ac3dec->dm_par.unit * *left++  + p_ac3dec->dm_par.clev * *center  - p_ac3dec->dm_par.slev * *right_sur;
 219                                 right_tmp= p_ac3dec->dm_par.unit * *right++ + p_ac3dec->dm_par.clev * *center++ + p_ac3dec->dm_par.slev * *right_sur++;
 220                                 *buffer++ = (s16)(left_tmp + *delay_left);
 221                                 *buffer++ = (s16)(right_tmp + *delay_right);
 222                                 *delay_left++ = p_ac3dec->dm_par.unit * *delay1_left++  + p_ac3dec->dm_par.clev * *delay1_center  + p_ac3dec->dm_par.slev * *delay1_sl;
 223                                 *delay_right++ = p_ac3dec->dm_par.unit * *delay1_right++ + p_ac3dec->dm_par.clev * *center++ + p_ac3dec->dm_par.slev * *delay1_sl++;
 224                         }
 225                         break;
 226                 case 4:         /* 2/1 */
 227                         left = p_ac3dec->samples[0];
 228                         right = p_ac3dec->samples[1];
 229                         right_sur = p_ac3dec->samples[2];
 230                         delay_left = p_ac3dec->imdct.delay[0];
 231                         delay_right = p_ac3dec->imdct.delay[1];
 232                         delay1_left = p_ac3dec->imdct.delay1[0];
 233                         delay1_right = p_ac3dec->imdct.delay1[1];
 234                         delay1_sl = p_ac3dec->imdct.delay1[2];
 235
 236                         for (i = 0; i < 256; i++) {
 237                                 left_tmp = p_ac3dec->dm_par.unit * *left++ - p_ac3dec->dm_par.slev * *right_sur;
 238                                 right_tmp= p_ac3dec->dm_par.unit * *right++ + p_ac3dec->dm_par.slev * *right_sur++;
 239                                 *buffer++ = (s16)(left_tmp + *delay_left);
 240                                 *buffer++ = (s16)(right_tmp + *delay_right);
 241                                 *delay_left++ = p_ac3dec->dm_par.unit * *delay1_left++ + p_ac3dec->dm_par.slev * *delay1_sl;
 242                                 *delay_right++ = p_ac3dec->dm_par.unit * *delay1_right++ + p_ac3dec->dm_par.slev * *delay1_sl++;
 243                         }
 244                         break;
 245                 case 3:         /* 3/0 */
 246                         left = p_ac3dec->samples[0];
 247                         center = p_ac3dec->samples[1];
 248                         right = p_ac3dec->samples[2];
 249                         delay_left = p_ac3dec->imdct.delay[0];
 250                         delay_right = p_ac3dec->imdct.delay[1];
 251                         delay1_left = p_ac3dec->imdct.delay1[0];
 252                         delay1_center = p_ac3dec->imdct.delay1[1];
 253                         delay1_right = p_ac3dec->imdct.delay1[2];
 254
 255                         for (i = 0; i < 256; i++) {
 256                                 left_tmp = p_ac3dec->dm_par.unit * *left++  + p_ac3dec->dm_par.clev * *center;
 257                                 right_tmp= p_ac3dec->dm_par.unit * *right++ + p_ac3dec->dm_par.clev * *center++;
 258                                 *buffer++ = (s16)(left_tmp + *delay_left);
 259                                 *buffer++ = (s16)(right_tmp + *delay_right);
 260                                 *delay_left++ = p_ac3dec->dm_par.unit * *delay1_left++  + p_ac3dec->dm_par.clev * *delay1_center;
 261                                 *delay_right++ = p_ac3dec->dm_par.unit * *delay1_right++ + p_ac3dec->dm_par.clev * *center++;
 262                         }
 263                         break;
 264                 case 2:         /* copy to output */
 265                         for (i = 0; i < 256; i++) {
 266                                 *buffer++ = (s16)p_ac3dec->samples[0][i];
 267                                 *buffer++ = (s16)p_ac3dec->samples[1][i];
 268                         }
 269                         break;
 270                 }
 271         }
 272 }