TwinVQ decoder
[libav.git] / libavcodec / twinvq.c
1 /*
2 * TwinVQ decoder
3 * Copyright (c) 2009 Vitor Sessak
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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
26 #include <math.h>
27 #include <stdint.h>
28
29 #include "twinvq_data.h"
30
31 enum FrameType {
32 FT_SHORT = 0, ///< Short frame (divided in n sub-blocks)
33 FT_MEDIUM, ///< Medium frame (divided in m<n sub-blocks)
34 FT_LONG, ///< Long frame (single sub-block + PPC)
35 FT_PPC, ///< Periodic Peak Component (part of the long frame)
36 };
37
38 /**
39 * Parameters and tables that are different for each frame type
40 */
41 struct FrameMode {
42 uint8_t sub; ///< Number subblocks in each frame
43 const uint16_t *bark_tab;
44
45 /** number of distinct bark scale envelope values */
46 uint8_t bark_env_size;
47
48 const int16_t *bark_cb; ///< codebook for the bark scale envelope (BSE)
49 uint8_t bark_n_coef;///< number of BSE CB coefficients to read
50 uint8_t bark_n_bit; ///< number of bits of the BSE coefs
51
52 //@{
53 /** main codebooks for spectrum data */
54 const int16_t *cb0;
55 const int16_t *cb1;
56 //@}
57
58 uint8_t cb_len_read; ///< number of spectrum coefficients to read
59 };
60
61 /**
62 * Parameters and tables that are different for every combination of
63 * bitrate/sample rate
64 */
65 typedef struct {
66 struct FrameMode fmode[3]; ///< frame type-dependant parameters
67
68 uint16_t size; ///< frame size in samples
69 uint8_t n_lsp; ///< number of lsp coefficients
70 const float *lspcodebook;
71
72 /* number of bits of the different LSP CB coefficients */
73 uint8_t lsp_bit0;
74 uint8_t lsp_bit1;
75 uint8_t lsp_bit2;
76
77 uint8_t lsp_split; ///< number of CB entries for the LSP decoding
78 const int16_t *ppc_shape_cb; ///< PPC shape CB
79
80 /** number of the bits for the PPC period value */
81 uint8_t ppc_period_bit;
82
83 uint8_t ppc_shape_bit; ///< number of bits of the PPC shape CB coeffs
84 uint8_t ppc_shape_len; ///< size of PPC shape CB
85 uint8_t pgain_bit; ///< bits for PPC gain
86
87 /** constant for peak period to peak width conversion */
88 uint16_t peak_per2wid;
89 } ModeTab;
90
91 static const ModeTab mode_08_08 = {
92 {
93 { 8, bark_tab_s08_64, 10, tab.fcb08s , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
94 { 2, bark_tab_m08_256, 20, tab.fcb08m , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
95 { 1, bark_tab_l08_512, 30, tab.fcb08l , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
96 },
97 512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40
98 };
99
100 static const ModeTab mode_11_08 = {
101 {
102 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
103 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
104 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
105 },
106 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
107 };
108
109 static const ModeTab mode_11_10 = {
110 {
111 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
112 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
113 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
114 },
115 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
116 };
117
118 static const ModeTab mode_16_16 = {
119 {
120 { 8, bark_tab_s16_128, 10, tab.fcb16s , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
121 { 2, bark_tab_m16_512, 20, tab.fcb16m , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
122 { 1, bark_tab_l16_1024,30, tab.fcb16l , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
123 },
124 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180
125 };
126
127 static const ModeTab mode_22_20 = {
128 {
129 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
130 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
131 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
132 },
133 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
134 };
135
136 static const ModeTab mode_22_24 = {
137 {
138 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
139 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
140 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
141 },
142 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
143 };
144
145 static const ModeTab mode_22_32 = {
146 {
147 { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
148 { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
149 { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
150 },
151 512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
152 };
153
154 static const ModeTab mode_44_40 = {
155 {
156 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
157 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
158 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
159 },
160 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
161 };
162
163 static const ModeTab mode_44_48 = {
164 {
165 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
166 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
167 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
168 },
169 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
170 };
171
172 typedef struct TwinContext {
173 AVCodecContext *avctx;
174 DSPContext dsp;
175 MDCTContext mdct_ctx[3];
176
177 const ModeTab *mtab;
178
179 // history
180 float lsp_hist[2][20]; ///< LSP coefficients of the last frame
181 float bark_hist[3][2][40]; ///< BSE coefficients of last frame
182
183 // bitstream parameters
184 int16_t permut[4][4096];
185 uint8_t length[4][2]; ///< main codebook stride
186 uint8_t length_change[4];
187 uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
188 int bits_main_spec_change[4];
189 int n_div[4];
190
191 float *spectrum;
192 float *curr_frame; ///< non-interleaved output
193 float *prev_frame; ///< non-interleaved previous frame
194 int last_block_pos[2];
195
196 float *cos_tabs[3];
197
198 // scratch buffers
199 float *tmp_buf;
200 } TwinContext;
201
202 #define PPC_SHAPE_CB_SIZE 64
203 #define SUB_AMP_MAX 4500.0
204 #define MULAW_MU 100.0
205 #define GAIN_BITS 8
206 #define AMP_MAX 13000.0
207 #define SUB_GAIN_BITS 5
208 #define WINDOW_TYPE_BITS 4
209 #define PGAIN_MU 200
210
211 /** @note not speed critical, hence not optimized */
212 static void memset_float(float *buf, float val, int size)
213 {
214 while (size--)
215 *buf++ = val;
216 }
217
218 /**
219 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
220 * spectrum pairs.
221 *
222 * @param lsp a vector of the cosinus of the LSP values
223 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
224 * @param order the order of the LSP (and the size of the *lsp buffer). Must
225 * be a multiple of four.
226 * @return the LPC value
227 *
228 * @todo reuse code from vorbis_dec.c: vorbis_floor0_decode
229 */
230 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
231 {
232 int j;
233 float p = 0.5f;
234 float q = 0.5f;
235 float two_cos_w = 2.0f*cos_val;
236
237 for (j=0; j+1 < order; j += 2*2) {
238 // Unroll the loop once since order is a multiple of four
239 q *= lsp[j ] - two_cos_w;
240 p *= lsp[j+1] - two_cos_w;
241
242 q *= lsp[j+2] - two_cos_w;
243 p *= lsp[j+3] - two_cos_w;
244 }
245
246 p *= p * (2.0f - two_cos_w);
247 q *= q * (2.0f + two_cos_w);
248
249 return 0.5 / (p + q);
250 }
251
252 /**
253 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
254 */
255 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
256 {
257 int i;
258 const ModeTab *mtab = tctx->mtab;
259 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
260
261 for (i=0; i < size_s/2; i++) {
262 float cos_i = tctx->cos_tabs[0][i];
263 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
264 lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
265 }
266 }
267
268 static void interpolate(float *out, float v1, float v2, int size)
269 {
270 int i;
271 float step = (v1 - v2)/(size + 1);
272
273 for (i=0; i < size; i++) {
274 v2 += step;
275 out[i] = v2;
276 }
277 }
278
279 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
280 {
281 return part ? -cos_tab[size - idx - 1] :
282 cos_tab[ idx ];
283 }
284
285 /**
286 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
287 * Probably for speed reasons, the coefficients are evaluated as
288 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
289 * where s is an evaluated value, i is a value interpolated from the others
290 * and b might be either calculated or interpolated, depending on an
291 * unexplained condition.
292 *
293 * @param step the size of a block "siiiibiiii"
294 * @param in the cosinus of the LSP data
295 * @param part is 0 for 0...PI (positive cossinus values) and 1 for PI...2PI
296 (negative cossinus values)
297 * @param size the size of the whole output
298 */
299 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
300 enum FrameType ftype,
301 float *out, const float *in,
302 int size, int step, int part)
303 {
304 int i;
305 const ModeTab *mtab = tctx->mtab;
306 const float *cos_tab = tctx->cos_tabs[ftype];
307
308 // Fill the 's'
309 for (i=0; i < size; i += step)
310 out[i] =
311 eval_lpc_spectrum(in,
312 get_cos(i, part, cos_tab, size),
313 mtab->n_lsp);
314
315 // Fill the 'iiiibiiii'
316 for (i=step; i <= size - 2*step; i += step) {
317 if (out[i + step] + out[i - step] > 1.95*out[i] ||
318 out[i + step] >= out[i - step]) {
319 interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
320 } else {
321 out[i - step/2] =
322 eval_lpc_spectrum(in,
323 get_cos(i-step/2, part, cos_tab, size),
324 mtab->n_lsp);
325 interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1);
326 interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1);
327 }
328 }
329
330 interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
331 }
332
333 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
334 const float *buf, float *lpc,
335 int size, int step)
336 {
337 eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0);
338 eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
339
340 interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
341
342 memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
343 }
344
345 /**
346 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
347 * bitstream, sum the corresponding vectors and write the result to *out
348 * after permutation.
349 */
350 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
351 enum FrameType ftype,
352 const int16_t *cb0, const int16_t *cb1, int cb_len)
353 {
354 int pos = 0;
355 int i, j;
356
357 for (i=0; i < tctx->n_div[ftype]; i++) {
358 int tmp0, tmp1;
359 int sign0 = 1;
360 int sign1 = 1;
361 const int16_t *tab0, *tab1;
362 int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
363 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
364
365 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
366 if (bits == 7) {
367 if (get_bits1(gb))
368 sign0 = -1;
369 bits = 6;
370 }
371 tmp0 = get_bits(gb, bits);
372
373 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
374
375 if (bits == 7) {
376 if (get_bits1(gb))
377 sign1 = -1;
378
379 bits = 6;
380 }
381 tmp1 = get_bits(gb, bits);
382
383 tab0 = cb0 + tmp0*cb_len;
384 tab1 = cb1 + tmp1*cb_len;
385
386 for (j=0; j < length; j++)
387 out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
388
389 pos += length;
390 }
391
392 }
393
394 static inline float mulawinv(float y, float clip, float mu)
395 {
396 y = av_clipf(y/clip, -1, 1);
397 return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
398 }
399
400 /**
401 * Evaluate a*b/400 rounded to the nearest integer. When, for example,
402 * a*b == 200 and the nearest integer is ill-defined, use a table to emulate
403 * the following broken float-based implementation used by the binary decoder:
404 *
405 * \code
406 * static int very_broken_op(int a, int b)
407 * {
408 * static float test; // Ugh, force gcc to do the division first...
409 *
410 * test = a/400.;
411 * return b * test + 0.5;
412 * }
413 * \endcode
414 *
415 * @note if this function is replaced by just ROUNDED_DIV(a*b,400.), the stddev
416 * between the original file (before encoding with Yamaha encoder) and the
417 * decoded output increases, which leads one to believe that the encoder expects
418 * exactly this broken calculation.
419 */
420 static int very_broken_op(int a, int b)
421 {
422 int x = a*b + 200;
423 int size;
424 const uint8_t *rtab;
425
426 if (x%400 || b%5)
427 return x/400;
428
429 x /= 400;
430
431 size = tabs[b/5].size;
432 rtab = tabs[b/5].tab;
433 return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
434 }
435
436 /**
437 * Sum to data a periodic peak of a given period, width and shape.
438 *
439 * @param period the period of the peak divised by 400.0
440 */
441 static void add_peak(int period, int width, const float *shape,
442 float ppc_gain, float *speech, int len)
443 {
444 int i, j;
445
446 const float *shape_end = shape + len;
447 int center;
448
449 // First peak centered around zero
450 for (i=0; i < width/2; i++)
451 speech[i] += ppc_gain * *shape++;
452
453 for (i=1; i < ROUNDED_DIV(len,width) ; i++) {
454 center = very_broken_op(period, i);
455 for (j=-width/2; j < (width+1)/2; j++)
456 speech[j+center] += ppc_gain * *shape++;
457 }
458
459 // For the last block, be careful not to go beyond the end of the buffer
460 center = very_broken_op(period, i);
461 for (j=-width/2; j < (width+1)/2 && shape < shape_end; j++)
462 speech[j+center] += ppc_gain * *shape++;
463 }
464
465 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
466 float ppc_gain, float *speech)
467 {
468 const ModeTab *mtab = tctx->mtab;
469 int isampf = tctx->avctx->sample_rate/1000;
470 int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
471 int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf);
472 int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
473 int period_range = max_period - min_period;
474
475 // This is actually the period multiplied by 400. It is just linearly coded
476 // between its maximum and minimum value.
477 int period = min_period +
478 ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
479 int width;
480
481 if (isampf == 22 && ibps == 32) {
482 // For some unknown reason, NTT decided to code this case differently...
483 width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
484 } else
485 width = (period )* mtab->peak_per2wid/(400*mtab->size);
486
487 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
488 }
489
490 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
491 float *out)
492 {
493 const ModeTab *mtab = tctx->mtab;
494 int i, j;
495 int sub = mtab->fmode[ftype].sub;
496 float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
497 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
498
499 if (ftype == FT_LONG) {
500 for (i=0; i < tctx->avctx->channels; i++)
501 out[i] = (1./(1<<13)) *
502 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
503 AMP_MAX, MULAW_MU);
504 } else {
505 for (i=0; i < tctx->avctx->channels; i++) {
506 float val = (1./(1<<23)) *
507 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
508 AMP_MAX, MULAW_MU);
509
510 for (j=0; j < sub; j++) {
511 out[i*sub + j] =
512 val*mulawinv(sub_step* 0.5 +
513 sub_step* get_bits(gb, SUB_GAIN_BITS),
514 SUB_AMP_MAX, MULAW_MU);
515 }
516 }
517 }
518 }
519
520 /**
521 * Rearrange the LSP coefficients so that they have a minimum distance of
522 * min_dist. This function does it exactly as described in section of 3.2.4
523 * of the G.729 specification (but interestingly is different from what the
524 * reference decoder actually does).
525 */
526 static void rearrange_lsp(int order, float *lsp, float min_dist)
527 {
528 int i;
529 float min_dist2 = min_dist * 0.5;
530 for (i=1; i < order; i++)
531 if (lsp[i] - lsp[i-1] < min_dist) {
532 float avg = (lsp[i] + lsp[i-1]) * 0.5;
533
534 lsp[i-1] = avg - min_dist2;
535 lsp[i ] = avg + min_dist2;
536 }
537 }
538
539 static void bubblesort(float *lsp, int lp_order)
540 {
541 int i,j;
542
543 /* sort lsp in ascending order. float bubble agorithm,
544 O(n) if data already sorted, O(n^2) - otherwise */
545 for (i=0; i < lp_order - 1; i++)
546 for (j=i; j >= 0 && lsp[j] > lsp[j+1]; j--)
547 FFSWAP(float, lsp[j], lsp[j+1]);
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);
586 bubblesort(lsp, mtab->n_lsp);
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 const ModeTab *mtab = tctx->mtab;
615 int bsize = mtab->size / mtab->fmode[ftype].sub;
616 int size = mtab->size;
617 float *buf1 = tctx->tmp_buf;
618 int j;
619 int wsize; // Window size
620 float *out = tctx->curr_frame + 2*ch*mtab->size;
621 float *out2 = out;
622 float *prev_buf;
623 int first_wsize;
624
625 static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1};
626 int types_sizes[] = {
627 mtab->size / mtab->fmode[FT_LONG ].sub,
628 mtab->size / mtab->fmode[FT_MEDIUM].sub,
629 mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
630 };
631
632 wsize = types_sizes[wtype_to_wsize[wtype]];
633 first_wsize = wsize;
634 prev_buf = prev + (size - bsize)/2;
635
636 for (j=0; j < mtab->fmode[ftype].sub; j++) {
637 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
638
639 if (!j && wtype == 4)
640 sub_wtype = 4;
641 else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
642 sub_wtype = 7;
643
644 wsize = types_sizes[wtype_to_wsize[sub_wtype]];
645
646 ff_imdct_half(&tctx->mdct_ctx[ftype], buf1 + bsize*j, in + bsize*j);
647
648 tctx->dsp.vector_fmul_window(out2,
649 prev_buf + (bsize-wsize)/2,
650 buf1 + bsize*j,
651 ff_sine_windows[av_log2(wsize) - 7],
652 0.0,
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 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
671 int i, j;
672
673 for (i=0; i < tctx->avctx->channels; i++) {
674 imdct_and_window(tctx, ftype, wtype,
675 tctx->spectrum + i*mtab->size,
676 prev_buf + 2*i*mtab->size,
677 i);
678 }
679
680 if (tctx->avctx->channels == 2) {
681 for (i=0; i < mtab->size - tctx->last_block_pos[0]; i++) {
682 float f1 = prev_buf[ i];
683 float f2 = prev_buf[2*mtab->size + i];
684 out[2*i ] = f1 + f2;
685 out[2*i + 1] = f1 - f2;
686 }
687 for (j=0; i < mtab->size; j++,i++) {
688 float f1 = tctx->curr_frame[ j];
689 float f2 = tctx->curr_frame[2*mtab->size + j];
690 out[2*i ] = f1 + f2;
691 out[2*i + 1] = f1 - f2;
692 }
693 } else {
694 memcpy(out, prev_buf,
695 (mtab->size - tctx->last_block_pos[0]) * sizeof(*out));
696
697 out += mtab->size - tctx->last_block_pos[0];
698
699 memcpy(out, tctx->curr_frame,
700 (tctx->last_block_pos[0]) * sizeof(*out));
701 }
702
703 }
704
705 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
706 int ch, float *out, float gain, enum FrameType ftype)
707 {
708 const ModeTab *mtab = tctx->mtab;
709 int i,j;
710 float *hist = tctx->bark_hist[ftype][ch];
711 float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
712 int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
713 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
714 int idx = 0;
715
716 for (i=0; i < fw_cb_len; i++)
717 for (j=0; j < bark_n_coef; j++, idx++) {
718 float tmp2 =
719 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
720 float st = use_hist ?
721 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
722
723 hist[idx] = tmp2;
724 if (st < -1.) st = 1.;
725
726 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
727 out += mtab->fmode[ftype].bark_tab[idx];
728 }
729
730 }
731
732 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
733 float *out, enum FrameType ftype)
734 {
735 const ModeTab *mtab = tctx->mtab;
736 int channels = tctx->avctx->channels;
737 int sub = mtab->fmode[ftype].sub;
738 int block_size = mtab->size / sub;
739 float gain[channels*sub];
740 float ppc_shape[mtab->ppc_shape_len * channels * 4];
741 uint8_t bark1[channels][sub][mtab->fmode[ftype].bark_n_coef];
742 uint8_t bark_use_hist[channels][sub];
743
744 uint8_t lpc_idx1[channels];
745 uint8_t lpc_idx2[channels][tctx->mtab->lsp_split];
746 uint8_t lpc_hist_idx[channels];
747
748 int i, j, k;
749
750 dequant(tctx, gb, out, ftype,
751 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
752 mtab->fmode[ftype].cb_len_read);
753
754 for (i=0; i < channels; i++)
755 for (j=0; j < sub; j++)
756 for (k=0; k < mtab->fmode[ftype].bark_n_coef; k++)
757 bark1[i][j][k] =
758 get_bits(gb, mtab->fmode[ftype].bark_n_bit);
759
760 for (i=0; i < channels; i++)
761 for (j=0; j < sub; j++)
762 bark_use_hist[i][j] = get_bits1(gb);
763
764 dec_gain(tctx, gb, ftype, gain);
765
766 for (i=0; i < channels; i++) {
767 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
768 lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1);
769
770 for (j=0; j < tctx->mtab->lsp_split; j++)
771 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
772 }
773
774 if (ftype == FT_LONG) {
775 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
776 tctx->n_div[3];
777 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
778 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
779 }
780
781 for (i=0; i < channels; i++) {
782 float *chunk = out + mtab->size * i;
783 float lsp[tctx->mtab->n_lsp];
784
785 for (j=0; j < sub; j++) {
786 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
787 tctx->tmp_buf, gain[sub*i+j], ftype);
788
789 tctx->dsp.vector_fmul(chunk + block_size*j, tctx->tmp_buf,
790 block_size);
791
792 }
793
794 if (ftype == FT_LONG) {
795 float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
796 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
797 int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
798 float v = 1./8192*
799 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
800
801 decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
802 chunk);
803 }
804
805 decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
806 tctx->lsp_hist[i]);
807
808 dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
809
810 for (j=0; j < mtab->fmode[ftype].sub; j++) {
811 tctx->dsp.vector_fmul(chunk, tctx->tmp_buf, block_size);
812 chunk += block_size;
813 }
814 }
815 }
816
817 static int twin_decode_frame(AVCodecContext * avctx, void *data,
818 int *data_size, AVPacket *avpkt)
819 {
820 const uint8_t *buf = avpkt->data;
821 int buf_size = avpkt->size;
822 TwinContext *tctx = avctx->priv_data;
823 GetBitContext gb;
824 const ModeTab *mtab = tctx->mtab;
825 float *out = data;
826 enum FrameType ftype;
827 int window_type;
828 int i;
829 static const enum FrameType wtype_to_ftype_table[] = {
830 FT_LONG, FT_LONG, FT_SHORT, FT_LONG,
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 *data_size = 0;
838 return buf_size;
839 }
840
841 init_get_bits(&gb, buf, buf_size * 8);
842 skip_bits(&gb, get_bits(&gb, 8));
843 window_type = get_bits(&gb, WINDOW_TYPE_BITS);
844
845 if (window_type > 8) {
846 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
847 return -1;
848 }
849
850 ftype = wtype_to_ftype_table[window_type];
851
852 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
853
854 imdct_output(tctx, ftype, window_type, out);
855
856 FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
857
858 if (tctx->avctx->frame_number < 2) {
859 *data_size=0;
860 return buf_size;
861 }
862
863 for (i=0; i < avctx->channels * mtab->size; i++)
864 out[i] = av_clipf(out[i], -32700./(1<<15), 32700./(1<<15));
865
866 *data_size = mtab->size*avctx->channels*4;
867
868 return buf_size;
869 }
870
871 /**
872 * Init IMDCT and windowing tables
873 */
874 static av_cold void init_mdct_win(TwinContext *tctx)
875 {
876 int i,j;
877 const ModeTab *mtab = tctx->mtab;
878 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
879 int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
880 int channels = tctx->avctx->channels;
881 float norm = channels == 1 ? 2. : 1.;
882
883 for (i=0; i < 3; i++) {
884 int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
885 ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
886 -sqrt(norm/bsize) / (1<<15));
887 }
888
889 tctx->tmp_buf = av_malloc(mtab->size * sizeof(*tctx->tmp_buf));
890
891 tctx->spectrum = av_malloc(2*mtab->size*channels*sizeof(float));
892 tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float));
893 tctx->prev_frame = av_malloc(2*mtab->size*channels*sizeof(float));
894
895 for(i=0; i < 3; i++) {
896 int m = 4*mtab->size/mtab->fmode[i].sub;
897 double freq = 2*M_PI/m;
898 tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs));
899
900 for (j=0; j <= m/8; j++)
901 tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
902 for (j=1; j < m/8; j++)
903 tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
904 }
905
906
907 ff_sine_window_init(ff_sine_windows[av_log2(size_m) - 7], size_m );
908 ff_sine_window_init(ff_sine_windows[av_log2(size_s/2) - 7], size_s/2);
909 ff_sine_window_init(ff_sine_windows[av_log2(mtab->size) - 7], mtab->size);
910 }
911
912 /**
913 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
914 * each line do a cyclic permutation, i.e.
915 * abcdefghijklm -> defghijklmabc
916 * where the amount to be shifted is evaluated depending on the column.
917 */
918 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
919 int block_size,
920 const uint8_t line_len[2], int length_div,
921 enum FrameType ftype)
922
923 {
924 int i,j;
925
926 for (i=0; i < line_len[0]; i++) {
927 int shift;
928
929 if (num_blocks == 1 ||
930 (ftype == FT_LONG && num_vect % num_blocks) ||
931 (ftype != FT_LONG && num_vect & 1 ) ||
932 i == line_len[1]) {
933 shift = 0;
934 } else if (ftype == FT_LONG) {
935 shift = i;
936 } else
937 shift = i*i;
938
939 for (j=0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
940 tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
941 }
942 }
943
944 /**
945 * Interpret the input data as in the following table:
946 *
947 * \verbatim
948 *
949 * abcdefgh
950 * ijklmnop
951 * qrstuvw
952 * x123456
953 *
954 * \endverbatim
955 *
956 * and transpose it, giving the output
957 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
958 */
959 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
960 const uint8_t line_len[2], int length_div)
961 {
962 int i,j;
963 int cont= 0;
964 for (i=0; i < num_vect; i++)
965 for (j=0; j < line_len[i >= length_div]; j++)
966 out[cont++] = in[j*num_vect + i];
967 }
968
969 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
970 {
971 int block_size = size/n_blocks;
972 int i;
973
974 for (i=0; i < size; i++)
975 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
976 }
977
978 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
979 {
980 int block_size;
981 const ModeTab *mtab = tctx->mtab;
982 int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
983 int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
984
985 if (ftype == FT_PPC) {
986 size = tctx->avctx->channels;
987 block_size = mtab->ppc_shape_len;
988 } else
989 block_size = mtab->size / mtab->fmode[ftype].sub;
990
991 permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
992 block_size, tctx->length[ftype],
993 tctx->length_change[ftype], ftype);
994
995 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
996 tctx->length[ftype], tctx->length_change[ftype]);
997
998 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
999 size*block_size);
1000 }
1001
1002 static av_cold void init_bitstream_params(TwinContext *tctx)
1003 {
1004 const ModeTab *mtab = tctx->mtab;
1005 int n_ch = tctx->avctx->channels;
1006 int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
1007 tctx->avctx->sample_rate;
1008
1009 int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
1010 mtab->lsp_split*mtab->lsp_bit2);
1011
1012 int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
1013 mtab->ppc_period_bit);
1014
1015 int bsize_no_main_cb[3];
1016 int bse_bits[3];
1017 int i;
1018
1019 for (i=0; i < 3; i++)
1020 // +1 for history usage switch
1021 bse_bits[i] = n_ch *
1022 (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
1023
1024 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1025 WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
1026
1027 for (i=0; i < 2; i++)
1028 bsize_no_main_cb[i] =
1029 lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
1030 mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
1031
1032 // The remaining bits are all used for the main spectrum coefficients
1033 for (i=0; i < 4; i++) {
1034 int bit_size;
1035 int vect_size;
1036 int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1037 if (i == 3) {
1038 bit_size = n_ch * mtab->ppc_shape_bit;
1039 vect_size = n_ch * mtab->ppc_shape_len;
1040 } else {
1041 bit_size = total_fr_bits - bsize_no_main_cb[i];
1042 vect_size = n_ch * mtab->size;
1043 }
1044
1045 tctx->n_div[i] = (bit_size + 13) / 14;
1046
1047 rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1048 rounded_down = (bit_size )/tctx->n_div[i];
1049 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1050 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1051 tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2;
1052 tctx->bits_main_spec[1][i][0] = (rounded_up )/2;
1053 tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
1054 tctx->bits_main_spec[1][i][1] = (rounded_down )/2;
1055 tctx->bits_main_spec_change[i] = num_rounded_up;
1056
1057 rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1058 rounded_down = (vect_size )/tctx->n_div[i];
1059 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1060 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1061 tctx->length[i][0] = rounded_up;
1062 tctx->length[i][1] = rounded_down;
1063 tctx->length_change[i] = num_rounded_up;
1064 }
1065
1066 for (i=0; i < 4; i++)
1067 construct_perm_table(tctx, i);
1068 }
1069
1070 static av_cold int twin_decode_init(AVCodecContext *avctx)
1071 {
1072 TwinContext *tctx = avctx->priv_data;
1073 int isampf = avctx->sample_rate/1000;
1074 int ibps = avctx->bit_rate/(1000 * avctx->channels);
1075
1076 tctx->avctx = avctx;
1077 avctx->sample_fmt = SAMPLE_FMT_FLT;
1078
1079 if (avctx->channels > 2) {
1080 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1081 avctx->channels);
1082 return -1;
1083 }
1084
1085 switch ((isampf << 8) + ibps) {
1086 case (8 <<8) + 8: tctx->mtab = &mode_08_08; break;
1087 case (11<<8) + 8: tctx->mtab = &mode_11_08; break;
1088 case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
1089 case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
1090 case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
1091 case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
1092 case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
1093 case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
1094 case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
1095 default:
1096 av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
1097 return -1;
1098 }
1099
1100 dsputil_init(&tctx->dsp, avctx);
1101 init_mdct_win(tctx);
1102 init_bitstream_params(tctx);
1103
1104 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1105
1106 return 0;
1107 }
1108
1109 static av_cold int twin_decode_close(AVCodecContext *avctx)
1110 {
1111 TwinContext *tctx = avctx->priv_data;
1112 int i;
1113
1114 for (i=0; i < 3; i++) {
1115 ff_mdct_end(&tctx->mdct_ctx[i]);
1116 av_free(tctx->cos_tabs[i]);
1117 }
1118
1119
1120 av_free(tctx->curr_frame);
1121 av_free(tctx->spectrum);
1122 av_free(tctx->prev_frame);
1123 av_free(tctx->tmp_buf);
1124
1125 return 0;
1126 }
1127
1128 AVCodec twinvq_decoder =
1129 {
1130 "twinvq",
1131 CODEC_TYPE_AUDIO,
1132 CODEC_ID_TWINVQ,
1133 sizeof(TwinContext),
1134 twin_decode_init,
1135 NULL,
1136 twin_decode_close,
1137 twin_decode_frame,
1138 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
1139 };