twinvq.c
Go to the documentation of this file.
1 /*
2  * TwinVQ decoder
3  * Copyright (c) 2009 Vitor Sessak
4  *
5  * This file is part of Libav.
6  *
7  * Libav is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * Libav is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 #include "avcodec.h"
23 #include "get_bits.h"
24 #include "dsputil.h"
25 #include "fft.h"
26 #include "lsp.h"
27 #include "sinewin.h"
28 
29 #include <math.h>
30 #include <stdint.h>
31 
32 #include "twinvq_data.h"
33 
34 enum FrameType {
35  FT_SHORT = 0,
39 };
40 
44 struct FrameMode {
45  uint8_t sub;
46  const uint16_t *bark_tab;
47 
49  uint8_t bark_env_size;
50 
51  const int16_t *bark_cb;
52  uint8_t bark_n_coef;
53  uint8_t bark_n_bit;
54 
56 
57  const int16_t *cb0;
58  const int16_t *cb1;
60 
61  uint8_t cb_len_read;
62 };
63 
68 typedef struct {
69  struct FrameMode fmode[3];
70 
71  uint16_t size;
72  uint8_t n_lsp;
73  const float *lspcodebook;
74 
75  /* number of bits of the different LSP CB coefficients */
76  uint8_t lsp_bit0;
77  uint8_t lsp_bit1;
78  uint8_t lsp_bit2;
79 
80  uint8_t lsp_split;
81  const int16_t *ppc_shape_cb;
82 
84  uint8_t ppc_period_bit;
85 
86  uint8_t ppc_shape_bit;
87  uint8_t ppc_shape_len;
88  uint8_t pgain_bit;
89 
91  uint16_t peak_per2wid;
92 } ModeTab;
93 
94 static const ModeTab mode_08_08 = {
95  {
96  { 8, bark_tab_s08_64, 10, tab.fcb08s , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
97  { 2, bark_tab_m08_256, 20, tab.fcb08m , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
98  { 1, bark_tab_l08_512, 30, tab.fcb08l , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
99  },
100  512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40
101 };
102 
103 static const ModeTab mode_11_08 = {
104  {
105  { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
106  { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
107  { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
108  },
109  512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
110 };
111 
112 static const ModeTab mode_11_10 = {
113  {
114  { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
115  { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
116  { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
117  },
118  512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
119 };
120 
121 static const ModeTab mode_16_16 = {
122  {
123  { 8, bark_tab_s16_128, 10, tab.fcb16s , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
124  { 2, bark_tab_m16_512, 20, tab.fcb16m , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
125  { 1, bark_tab_l16_1024,30, tab.fcb16l , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
126  },
127  1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180
128 };
129 
130 static const ModeTab mode_22_20 = {
131  {
132  { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
133  { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
134  { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
135  },
136  1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
137 };
138 
139 static const ModeTab mode_22_24 = {
140  {
141  { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
142  { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
143  { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
144  },
145  1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
146 };
147 
148 static const ModeTab mode_22_32 = {
149  {
150  { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
151  { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
152  { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
153  },
154  512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
155 };
156 
157 static const ModeTab mode_44_40 = {
158  {
159  {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
160  { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
161  { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
162  },
163  2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
164 };
165 
166 static const ModeTab mode_44_48 = {
167  {
168  {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
169  { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
170  { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
171  },
172  2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
173 };
174 
175 typedef struct TwinContext {
180 
181  const ModeTab *mtab;
182 
183  // history
184  float lsp_hist[2][20];
185  float bark_hist[3][2][40];
186 
187  // bitstream parameters
188  int16_t permut[4][4096];
189  uint8_t length[4][2];
190  uint8_t length_change[4];
191  uint8_t bits_main_spec[2][4][2];
193  int n_div[4];
194 
195  float *spectrum;
196  float *curr_frame;
197  float *prev_frame;
200 
201  float *cos_tabs[3];
202 
203  // scratch buffers
204  float *tmp_buf;
205 } TwinContext;
206 
207 #define PPC_SHAPE_CB_SIZE 64
208 #define PPC_SHAPE_LEN_MAX 60
209 #define SUB_AMP_MAX 4500.0
210 #define MULAW_MU 100.0
211 #define GAIN_BITS 8
212 #define AMP_MAX 13000.0
213 #define SUB_GAIN_BITS 5
214 #define WINDOW_TYPE_BITS 4
215 #define PGAIN_MU 200
216 #define LSP_COEFS_MAX 20
217 #define LSP_SPLIT_MAX 4
218 #define CHANNELS_MAX 2
219 #define SUBBLOCKS_MAX 16
220 #define BARK_N_COEF_MAX 4
221 
223 static void memset_float(float *buf, float val, int size)
224 {
225  while (size--)
226  *buf++ = val;
227 }
228 
241 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
242 {
243  int j;
244  float p = 0.5f;
245  float q = 0.5f;
246  float two_cos_w = 2.0f*cos_val;
247 
248  for (j = 0; j + 1 < order; j += 2*2) {
249  // Unroll the loop once since order is a multiple of four
250  q *= lsp[j ] - two_cos_w;
251  p *= lsp[j+1] - two_cos_w;
252 
253  q *= lsp[j+2] - two_cos_w;
254  p *= lsp[j+3] - two_cos_w;
255  }
256 
257  p *= p * (2.0f - two_cos_w);
258  q *= q * (2.0f + two_cos_w);
259 
260  return 0.5 / (p + q);
261 }
262 
266 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
267 {
268  int i;
269  const ModeTab *mtab = tctx->mtab;
270  int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
271 
272  for (i = 0; i < size_s/2; i++) {
273  float cos_i = tctx->cos_tabs[0][i];
274  lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
275  lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
276  }
277 }
278 
279 static void interpolate(float *out, float v1, float v2, int size)
280 {
281  int i;
282  float step = (v1 - v2)/(size + 1);
283 
284  for (i = 0; i < size; i++) {
285  v2 += step;
286  out[i] = v2;
287  }
288 }
289 
290 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
291 {
292  return part ? -cos_tab[size - idx - 1] :
293  cos_tab[ idx ];
294 }
295 
310 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
311  enum FrameType ftype,
312  float *out, const float *in,
313  int size, int step, int part)
314 {
315  int i;
316  const ModeTab *mtab = tctx->mtab;
317  const float *cos_tab = tctx->cos_tabs[ftype];
318 
319  // Fill the 's'
320  for (i = 0; i < size; i += step)
321  out[i] =
323  get_cos(i, part, cos_tab, size),
324  mtab->n_lsp);
325 
326  // Fill the 'iiiibiiii'
327  for (i = step; i <= size - 2*step; i += step) {
328  if (out[i + step] + out[i - step] > 1.95*out[i] ||
329  out[i + step] >= out[i - step]) {
330  interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
331  } else {
332  out[i - step/2] =
334  get_cos(i-step/2, part, cos_tab, size),
335  mtab->n_lsp);
336  interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1);
337  interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1);
338  }
339  }
340 
341  interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
342 }
343 
344 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
345  const float *buf, float *lpc,
346  int size, int step)
347 {
348  eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0);
349  eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
350 
351  interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
352 
353  memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
354 }
355 
361 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
362  enum FrameType ftype,
363  const int16_t *cb0, const int16_t *cb1, int cb_len)
364 {
365  int pos = 0;
366  int i, j;
367 
368  for (i = 0; i < tctx->n_div[ftype]; i++) {
369  int tmp0, tmp1;
370  int sign0 = 1;
371  int sign1 = 1;
372  const int16_t *tab0, *tab1;
373  int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
374  int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
375 
376  int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
377  if (bits == 7) {
378  if (get_bits1(gb))
379  sign0 = -1;
380  bits = 6;
381  }
382  tmp0 = get_bits(gb, bits);
383 
384  bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
385 
386  if (bits == 7) {
387  if (get_bits1(gb))
388  sign1 = -1;
389 
390  bits = 6;
391  }
392  tmp1 = get_bits(gb, bits);
393 
394  tab0 = cb0 + tmp0*cb_len;
395  tab1 = cb1 + tmp1*cb_len;
396 
397  for (j = 0; j < length; j++)
398  out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
399 
400  pos += length;
401  }
402 
403 }
404 
405 static inline float mulawinv(float y, float clip, float mu)
406 {
407  y = av_clipf(y/clip, -1, 1);
408  return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
409 }
410 
431 static int very_broken_op(int a, int b)
432 {
433  int x = a*b + 200;
434  int size;
435  const uint8_t *rtab;
436 
437  if (x%400 || b%5)
438  return x/400;
439 
440  x /= 400;
441 
442  size = tabs[b/5].size;
443  rtab = tabs[b/5].tab;
444  return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
445 }
446 
452 static void add_peak(int period, int width, const float *shape,
453  float ppc_gain, float *speech, int len)
454 {
455  int i, j;
456 
457  const float *shape_end = shape + len;
458  int center;
459 
460  // First peak centered around zero
461  for (i = 0; i < width/2; i++)
462  speech[i] += ppc_gain * *shape++;
463 
464  for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
465  center = very_broken_op(period, i);
466  for (j = -width/2; j < (width+1)/2; j++)
467  speech[j+center] += ppc_gain * *shape++;
468  }
469 
470  // For the last block, be careful not to go beyond the end of the buffer
471  center = very_broken_op(period, i);
472  for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
473  speech[j+center] += ppc_gain * *shape++;
474 }
475 
476 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
477  float ppc_gain, float *speech)
478 {
479  const ModeTab *mtab = tctx->mtab;
480  int isampf = tctx->avctx->sample_rate/1000;
481  int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
482  int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf);
483  int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
484  int period_range = max_period - min_period;
485 
486  // This is actually the period multiplied by 400. It is just linearly coded
487  // between its maximum and minimum value.
488  int period = min_period +
489  ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
490  int width;
491 
492  if (isampf == 22 && ibps == 32) {
493  // For some unknown reason, NTT decided to code this case differently...
494  width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
495  } else
496  width = (period )* mtab->peak_per2wid/(400*mtab->size);
497 
498  add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
499 }
500 
501 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
502  float *out)
503 {
504  const ModeTab *mtab = tctx->mtab;
505  int i, j;
506  int sub = mtab->fmode[ftype].sub;
507  float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
508  float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
509 
510  if (ftype == FT_LONG) {
511  for (i = 0; i < tctx->avctx->channels; i++)
512  out[i] = (1./(1<<13)) *
513  mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
514  AMP_MAX, MULAW_MU);
515  } else {
516  for (i = 0; i < tctx->avctx->channels; i++) {
517  float val = (1./(1<<23)) *
518  mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
519  AMP_MAX, MULAW_MU);
520 
521  for (j = 0; j < sub; j++) {
522  out[i*sub + j] =
523  val*mulawinv(sub_step* 0.5 +
524  sub_step* get_bits(gb, SUB_GAIN_BITS),
526  }
527  }
528  }
529 }
530 
537 static void rearrange_lsp(int order, float *lsp, float min_dist)
538 {
539  int i;
540  float min_dist2 = min_dist * 0.5;
541  for (i = 1; i < order; i++)
542  if (lsp[i] - lsp[i-1] < min_dist) {
543  float avg = (lsp[i] + lsp[i-1]) * 0.5;
544 
545  lsp[i-1] = avg - min_dist2;
546  lsp[i ] = avg + min_dist2;
547  }
548 }
549 
550 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
551  int lpc_hist_idx, float *lsp, float *hist)
552 {
553  const ModeTab *mtab = tctx->mtab;
554  int i, j;
555 
556  const float *cb = mtab->lspcodebook;
557  const float *cb2 = cb + (1 << mtab->lsp_bit1)*mtab->n_lsp;
558  const float *cb3 = cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
559 
560  const int8_t funny_rounding[4] = {
561  -2,
562  mtab->lsp_split == 4 ? -2 : 1,
563  mtab->lsp_split == 4 ? -2 : 1,
564  0
565  };
566 
567  j = 0;
568  for (i = 0; i < mtab->lsp_split; i++) {
569  int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
570  for (; j < chunk_end; j++)
571  lsp[j] = cb [lpc_idx1 * mtab->n_lsp + j] +
572  cb2[lpc_idx2[i] * mtab->n_lsp + j];
573  }
574 
575  rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
576 
577  for (i = 0; i < mtab->n_lsp; i++) {
578  float tmp1 = 1. - cb3[lpc_hist_idx*mtab->n_lsp + i];
579  float tmp2 = hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
580  hist[i] = lsp[i];
581  lsp[i] = lsp[i] * tmp1 + tmp2;
582  }
583 
584  rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
585  rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
587 }
588 
589 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
590  enum FrameType ftype, float *lpc)
591 {
592  int i;
593  int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
594 
595  for (i = 0; i < tctx->mtab->n_lsp; i++)
596  lsp[i] = 2*cos(lsp[i]);
597 
598  switch (ftype) {
599  case FT_LONG:
600  eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
601  break;
602  case FT_MEDIUM:
603  eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
604  break;
605  case FT_SHORT:
606  eval_lpcenv(tctx, lsp, lpc);
607  break;
608  }
609 }
610 
611 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
612  float *in, float *prev, int ch)
613 {
614  FFTContext *mdct = &tctx->mdct_ctx[ftype];
615  const ModeTab *mtab = tctx->mtab;
616  int bsize = mtab->size / mtab->fmode[ftype].sub;
617  int size = mtab->size;
618  float *buf1 = tctx->tmp_buf;
619  int j;
620  int wsize; // Window size
621  float *out = tctx->curr_frame + 2*ch*mtab->size;
622  float *out2 = out;
623  float *prev_buf;
624  int first_wsize;
625 
626  static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1};
627  int types_sizes[] = {
628  mtab->size / mtab->fmode[FT_LONG ].sub,
629  mtab->size / mtab->fmode[FT_MEDIUM].sub,
630  mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
631  };
632 
633  wsize = types_sizes[wtype_to_wsize[wtype]];
634  first_wsize = wsize;
635  prev_buf = prev + (size - bsize)/2;
636 
637  for (j = 0; j < mtab->fmode[ftype].sub; j++) {
638  int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
639 
640  if (!j && wtype == 4)
641  sub_wtype = 4;
642  else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
643  sub_wtype = 7;
644 
645  wsize = types_sizes[wtype_to_wsize[sub_wtype]];
646 
647  mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
648 
649  tctx->dsp.vector_fmul_window(out2,
650  prev_buf + (bsize-wsize)/2,
651  buf1 + bsize*j,
652  ff_sine_windows[av_log2(wsize)],
653  wsize/2);
654  out2 += wsize;
655 
656  memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
657 
658  out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
659 
660  prev_buf = buf1 + bsize*j + bsize/2;
661  }
662 
663  tctx->last_block_pos[ch] = (size + first_wsize)/2;
664 }
665 
666 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
667  float *out)
668 {
669  const ModeTab *mtab = tctx->mtab;
670  int size1, size2;
671  float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
672  int i;
673 
674  for (i = 0; i < tctx->avctx->channels; i++) {
675  imdct_and_window(tctx, ftype, wtype,
676  tctx->spectrum + i*mtab->size,
677  prev_buf + 2*i*mtab->size,
678  i);
679  }
680 
681  if (!out)
682  return;
683 
684  size2 = tctx->last_block_pos[0];
685  size1 = mtab->size - size2;
686  if (tctx->avctx->channels == 2) {
687  tctx->dsp.butterflies_float_interleave(out, prev_buf,
688  &prev_buf[2*mtab->size],
689  size1);
690 
691  out += 2 * size1;
692 
694  &tctx->curr_frame[2*mtab->size],
695  size2);
696  } else {
697  memcpy(out, prev_buf, size1 * sizeof(*out));
698 
699  out += size1;
700 
701  memcpy(out, tctx->curr_frame, size2 * sizeof(*out));
702  }
703 
704 }
705 
706 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
707  int ch, float *out, float gain, enum FrameType ftype)
708 {
709  const ModeTab *mtab = tctx->mtab;
710  int i,j;
711  float *hist = tctx->bark_hist[ftype][ch];
712  float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
713  int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
714  int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
715  int idx = 0;
716 
717  for (i = 0; i < fw_cb_len; i++)
718  for (j = 0; j < bark_n_coef; j++, idx++) {
719  float tmp2 =
720  mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
721  float st = use_hist ?
722  (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
723 
724  hist[idx] = tmp2;
725  if (st < -1.) st = 1.;
726 
727  memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
728  out += mtab->fmode[ftype].bark_tab[idx];
729  }
730 
731 }
732 
734  float *out, enum FrameType ftype)
735 {
736  const ModeTab *mtab = tctx->mtab;
737  int channels = tctx->avctx->channels;
738  int sub = mtab->fmode[ftype].sub;
739  int block_size = mtab->size / sub;
740  float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
741  float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
742  uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
743  uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
744 
745  uint8_t lpc_idx1[CHANNELS_MAX];
746  uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
747  uint8_t lpc_hist_idx[CHANNELS_MAX];
748 
749  int i, j, k;
750 
751  dequant(tctx, gb, out, ftype,
752  mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
753  mtab->fmode[ftype].cb_len_read);
754 
755  for (i = 0; i < channels; i++)
756  for (j = 0; j < sub; j++)
757  for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
758  bark1[i][j][k] =
759  get_bits(gb, mtab->fmode[ftype].bark_n_bit);
760 
761  for (i = 0; i < channels; i++)
762  for (j = 0; j < sub; j++)
763  bark_use_hist[i][j] = get_bits1(gb);
764 
765  dec_gain(tctx, gb, ftype, gain);
766 
767  for (i = 0; i < channels; i++) {
768  lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
769  lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1);
770 
771  for (j = 0; j < tctx->mtab->lsp_split; j++)
772  lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
773  }
774 
775  if (ftype == FT_LONG) {
776  int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
777  tctx->n_div[3];
778  dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
779  mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
780  }
781 
782  for (i = 0; i < channels; i++) {
783  float *chunk = out + mtab->size * i;
784  float lsp[LSP_COEFS_MAX];
785 
786  for (j = 0; j < sub; j++) {
787  dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
788  tctx->tmp_buf, gain[sub*i+j], ftype);
789 
790  tctx->dsp.vector_fmul(chunk + block_size*j, chunk + block_size*j, tctx->tmp_buf,
791  block_size);
792 
793  }
794 
795  if (ftype == FT_LONG) {
796  float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
797  int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
798  int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
799  float v = 1./8192*
800  mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
801 
802  decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
803  chunk);
804  }
805 
806  decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
807  tctx->lsp_hist[i]);
808 
809  dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
810 
811  for (j = 0; j < mtab->fmode[ftype].sub; j++) {
812  tctx->dsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
813  chunk += block_size;
814  }
815  }
816 }
817 
818 static int twin_decode_frame(AVCodecContext * avctx, void *data,
819  int *got_frame_ptr, AVPacket *avpkt)
820 {
821  const uint8_t *buf = avpkt->data;
822  int buf_size = avpkt->size;
823  TwinContext *tctx = avctx->priv_data;
824  GetBitContext gb;
825  const ModeTab *mtab = tctx->mtab;
826  float *out = NULL;
827  enum FrameType ftype;
828  int window_type, ret;
829  static const enum FrameType wtype_to_ftype_table[] = {
831  FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM
832  };
833 
834  if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
835  av_log(avctx, AV_LOG_ERROR,
836  "Frame too small (%d bytes). Truncated file?\n", buf_size);
837  return AVERROR(EINVAL);
838  }
839 
840  /* get output buffer */
841  if (tctx->discarded_packets >= 2) {
842  tctx->frame.nb_samples = mtab->size;
843  if ((ret = avctx->get_buffer(avctx, &tctx->frame)) < 0) {
844  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
845  return ret;
846  }
847  out = (float *)tctx->frame.data[0];
848  }
849 
850  init_get_bits(&gb, buf, buf_size * 8);
851  skip_bits(&gb, get_bits(&gb, 8));
852  window_type = get_bits(&gb, WINDOW_TYPE_BITS);
853 
854  if (window_type > 8) {
855  av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
856  return -1;
857  }
858 
859  ftype = wtype_to_ftype_table[window_type];
860 
861  read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
862 
863  imdct_output(tctx, ftype, window_type, out);
864 
865  FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
866 
867  if (tctx->discarded_packets < 2) {
868  tctx->discarded_packets++;
869  *got_frame_ptr = 0;
870  return buf_size;
871  }
872 
873  *got_frame_ptr = 1;
874  *(AVFrame *)data = tctx->frame;;
875 
876  return buf_size;
877 }
878 
883 {
884  int i, j, ret;
885  const ModeTab *mtab = tctx->mtab;
886  int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
887  int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
888  int channels = tctx->avctx->channels;
889  float norm = channels == 1 ? 2. : 1.;
890 
891  for (i = 0; i < 3; i++) {
892  int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
893  if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
894  -sqrt(norm/bsize) / (1<<15))))
895  return ret;
896  }
897 
898  FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
899  mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
900 
901  FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
902  2 * mtab->size * channels * sizeof(*tctx->spectrum),
903  alloc_fail);
904  FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
905  2 * mtab->size * channels * sizeof(*tctx->curr_frame),
906  alloc_fail);
907  FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
908  2 * mtab->size * channels * sizeof(*tctx->prev_frame),
909  alloc_fail);
910 
911  for (i = 0; i < 3; i++) {
912  int m = 4*mtab->size/mtab->fmode[i].sub;
913  double freq = 2*M_PI/m;
914  FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
915  (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
916 
917  for (j = 0; j <= m/8; j++)
918  tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
919  for (j = 1; j < m/8; j++)
920  tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
921  }
922 
923 
924  ff_init_ff_sine_windows(av_log2(size_m));
925  ff_init_ff_sine_windows(av_log2(size_s/2));
926  ff_init_ff_sine_windows(av_log2(mtab->size));
927 
928  return 0;
929 alloc_fail:
930  return AVERROR(ENOMEM);
931 }
932 
939 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
940  int block_size,
941  const uint8_t line_len[2], int length_div,
942  enum FrameType ftype)
943 
944 {
945  int i,j;
946 
947  for (i = 0; i < line_len[0]; i++) {
948  int shift;
949 
950  if (num_blocks == 1 ||
951  (ftype == FT_LONG && num_vect % num_blocks) ||
952  (ftype != FT_LONG && num_vect & 1 ) ||
953  i == line_len[1]) {
954  shift = 0;
955  } else if (ftype == FT_LONG) {
956  shift = i;
957  } else
958  shift = i*i;
959 
960  for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
961  tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
962  }
963 }
964 
980 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
981  const uint8_t line_len[2], int length_div)
982 {
983  int i,j;
984  int cont= 0;
985  for (i = 0; i < num_vect; i++)
986  for (j = 0; j < line_len[i >= length_div]; j++)
987  out[cont++] = in[j*num_vect + i];
988 }
989 
990 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
991 {
992  int block_size = size/n_blocks;
993  int i;
994 
995  for (i = 0; i < size; i++)
996  out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
997 }
998 
999 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
1000 {
1001  int block_size;
1002  const ModeTab *mtab = tctx->mtab;
1003  int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
1004  int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
1005 
1006  if (ftype == FT_PPC) {
1007  size = tctx->avctx->channels;
1008  block_size = mtab->ppc_shape_len;
1009  } else
1010  block_size = mtab->size / mtab->fmode[ftype].sub;
1011 
1012  permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
1013  block_size, tctx->length[ftype],
1014  tctx->length_change[ftype], ftype);
1015 
1016  transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
1017  tctx->length[ftype], tctx->length_change[ftype]);
1018 
1019  linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
1020  size*block_size);
1021 }
1022 
1024 {
1025  const ModeTab *mtab = tctx->mtab;
1026  int n_ch = tctx->avctx->channels;
1027  int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
1028  tctx->avctx->sample_rate;
1029 
1030  int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
1031  mtab->lsp_split*mtab->lsp_bit2);
1032 
1033  int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
1034  mtab->ppc_period_bit);
1035 
1036  int bsize_no_main_cb[3];
1037  int bse_bits[3];
1038  int i;
1039  enum FrameType frametype;
1040 
1041  for (i = 0; i < 3; i++)
1042  // +1 for history usage switch
1043  bse_bits[i] = n_ch *
1044  (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
1045 
1046  bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1047  WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
1048 
1049  for (i = 0; i < 2; i++)
1050  bsize_no_main_cb[i] =
1051  lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
1052  mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
1053 
1054  // The remaining bits are all used for the main spectrum coefficients
1055  for (i = 0; i < 4; i++) {
1056  int bit_size;
1057  int vect_size;
1058  int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1059  if (i == 3) {
1060  bit_size = n_ch * mtab->ppc_shape_bit;
1061  vect_size = n_ch * mtab->ppc_shape_len;
1062  } else {
1063  bit_size = total_fr_bits - bsize_no_main_cb[i];
1064  vect_size = n_ch * mtab->size;
1065  }
1066 
1067  tctx->n_div[i] = (bit_size + 13) / 14;
1068 
1069  rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1070  rounded_down = (bit_size )/tctx->n_div[i];
1071  num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1072  num_rounded_up = tctx->n_div[i] - num_rounded_down;
1073  tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2;
1074  tctx->bits_main_spec[1][i][0] = (rounded_up )/2;
1075  tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
1076  tctx->bits_main_spec[1][i][1] = (rounded_down )/2;
1077  tctx->bits_main_spec_change[i] = num_rounded_up;
1078 
1079  rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1080  rounded_down = (vect_size )/tctx->n_div[i];
1081  num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1082  num_rounded_up = tctx->n_div[i] - num_rounded_down;
1083  tctx->length[i][0] = rounded_up;
1084  tctx->length[i][1] = rounded_down;
1085  tctx->length_change[i] = num_rounded_up;
1086  }
1087 
1088  for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
1089  construct_perm_table(tctx, frametype);
1090 }
1091 
1093 {
1094  TwinContext *tctx = avctx->priv_data;
1095  int i;
1096 
1097  for (i = 0; i < 3; i++) {
1098  ff_mdct_end(&tctx->mdct_ctx[i]);
1099  av_free(tctx->cos_tabs[i]);
1100  }
1101 
1102 
1103  av_free(tctx->curr_frame);
1104  av_free(tctx->spectrum);
1105  av_free(tctx->prev_frame);
1106  av_free(tctx->tmp_buf);
1107 
1108  return 0;
1109 }
1110 
1112 {
1113  int ret;
1114  TwinContext *tctx = avctx->priv_data;
1115  int isampf, ibps;
1116 
1117  tctx->avctx = avctx;
1118  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
1119 
1120  if (!avctx->extradata || avctx->extradata_size < 12) {
1121  av_log(avctx, AV_LOG_ERROR, "Missing or incomplete extradata\n");
1122  return AVERROR_INVALIDDATA;
1123  }
1124  avctx->channels = AV_RB32(avctx->extradata ) + 1;
1125  avctx->bit_rate = AV_RB32(avctx->extradata + 4) * 1000;
1126  isampf = AV_RB32(avctx->extradata + 8);
1127  switch (isampf) {
1128  case 44: avctx->sample_rate = 44100; break;
1129  case 22: avctx->sample_rate = 22050; break;
1130  case 11: avctx->sample_rate = 11025; break;
1131  default: avctx->sample_rate = isampf * 1000; break;
1132  }
1133 
1134  if (avctx->channels > CHANNELS_MAX) {
1135  av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1136  avctx->channels);
1137  return -1;
1138  }
1139  ibps = avctx->bit_rate / (1000 * avctx->channels);
1140 
1141  switch ((isampf << 8) + ibps) {
1142  case (8 <<8) + 8: tctx->mtab = &mode_08_08; break;
1143  case (11<<8) + 8: tctx->mtab = &mode_11_08; break;
1144  case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
1145  case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
1146  case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
1147  case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
1148  case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
1149  case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
1150  case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
1151  default:
1152  av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
1153  return -1;
1154  }
1155 
1156  dsputil_init(&tctx->dsp, avctx);
1157  if ((ret = init_mdct_win(tctx))) {
1158  av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
1159  twin_decode_close(avctx);
1160  return ret;
1161  }
1162  init_bitstream_params(tctx);
1163 
1164  memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1165 
1167  avctx->coded_frame = &tctx->frame;
1168 
1169  return 0;
1170 }
1171 
1173  .name = "twinvq",
1174  .type = AVMEDIA_TYPE_AUDIO,
1175  .id = CODEC_ID_TWINVQ,
1176  .priv_data_size = sizeof(TwinContext),
1180  .capabilities = CODEC_CAP_DR1,
1181  .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
1182 };