3 * Copyright (c) 2009 Vitor Sessak
5 * This file is part of FFmpeg.
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.
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.
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
32 #include "twinvq_data.h"
35 FT_SHORT = 0, ///< Short frame (divided in n sub-blocks)
36 FT_MEDIUM, ///< Medium frame (divided in m<n sub-blocks)
37 FT_LONG, ///< Long frame (single sub-block + PPC)
38 FT_PPC, ///< Periodic Peak Component (part of the long frame)
42 * Parameters and tables that are different for each frame type
45 uint8_t sub; ///< Number subblocks in each frame
46 const uint16_t *bark_tab;
48 /** number of distinct bark scale envelope values */
49 uint8_t bark_env_size;
51 const int16_t *bark_cb; ///< codebook for the bark scale envelope (BSE)
52 uint8_t bark_n_coef;///< number of BSE CB coefficients to read
53 uint8_t bark_n_bit; ///< number of bits of the BSE coefs
56 /** main codebooks for spectrum data */
61 uint8_t cb_len_read; ///< number of spectrum coefficients to read
65 * Parameters and tables that are different for every combination of
69 struct FrameMode fmode[3]; ///< frame type-dependant parameters
71 uint16_t size; ///< frame size in samples
72 uint8_t n_lsp; ///< number of lsp coefficients
73 const float *lspcodebook;
75 /* number of bits of the different LSP CB coefficients */
80 uint8_t lsp_split; ///< number of CB entries for the LSP decoding
81 const int16_t *ppc_shape_cb; ///< PPC shape CB
83 /** number of the bits for the PPC period value */
84 uint8_t ppc_period_bit;
86 uint8_t ppc_shape_bit; ///< number of bits of the PPC shape CB coeffs
87 uint8_t ppc_shape_len; ///< size of PPC shape CB
88 uint8_t pgain_bit; ///< bits for PPC gain
90 /** constant for peak period to peak width conversion */
91 uint16_t peak_per2wid;
94 static const ModeTab mode_08_08 = {
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}
100 512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40
103 static const ModeTab mode_11_08 = {
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}
109 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
112 static const ModeTab mode_11_10 = {
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}
118 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
121 static const ModeTab mode_16_16 = {
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}
127 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180
130 static const ModeTab mode_22_20 = {
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}
136 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
139 static const ModeTab mode_22_24 = {
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}
145 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
148 static const ModeTab mode_22_32 = {
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}
154 512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
157 static const ModeTab mode_44_40 = {
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}
163 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
166 static const ModeTab mode_44_48 = {
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}
172 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
175 typedef struct TwinContext {
176 AVCodecContext *avctx;
178 FFTContext mdct_ctx[3];
183 float lsp_hist[2][20]; ///< LSP coefficients of the last frame
184 float bark_hist[3][2][40]; ///< BSE coefficients of last frame
186 // bitstream parameters
187 int16_t permut[4][4096];
188 uint8_t length[4][2]; ///< main codebook stride
189 uint8_t length_change[4];
190 uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
191 int bits_main_spec_change[4];
195 float *curr_frame; ///< non-interleaved output
196 float *prev_frame; ///< non-interleaved previous frame
197 int last_block_pos[2];
205 #define PPC_SHAPE_CB_SIZE 64
206 #define PPC_SHAPE_LEN_MAX 60
207 #define SUB_AMP_MAX 4500.0
208 #define MULAW_MU 100.0
210 #define AMP_MAX 13000.0
211 #define SUB_GAIN_BITS 5
212 #define WINDOW_TYPE_BITS 4
214 #define LSP_COEFS_MAX 20
215 #define LSP_SPLIT_MAX 4
216 #define CHANNELS_MAX 2
217 #define SUBBLOCKS_MAX 16
218 #define BARK_N_COEF_MAX 4
220 /** @note not speed critical, hence not optimized */
221 static void memset_float(float *buf, float val, int size)
228 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
231 * @param lsp a vector of the cosinus of the LSP values
232 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
233 * @param order the order of the LSP (and the size of the *lsp buffer). Must
234 * be a multiple of four.
235 * @return the LPC value
237 * @todo reuse code from vorbis_dec.c: vorbis_floor0_decode
239 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
244 float two_cos_w = 2.0f*cos_val;
246 for (j = 0; j + 1 < order; j += 2*2) {
247 // Unroll the loop once since order is a multiple of four
248 q *= lsp[j ] - two_cos_w;
249 p *= lsp[j+1] - two_cos_w;
251 q *= lsp[j+2] - two_cos_w;
252 p *= lsp[j+3] - two_cos_w;
255 p *= p * (2.0f - two_cos_w);
256 q *= q * (2.0f + two_cos_w);
258 return 0.5 / (p + q);
262 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
264 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
267 const ModeTab *mtab = tctx->mtab;
268 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
270 for (i = 0; i < size_s/2; i++) {
271 float cos_i = tctx->cos_tabs[0][i];
272 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
273 lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
277 static void interpolate(float *out, float v1, float v2, int size)
280 float step = (v1 - v2)/(size + 1);
282 for (i = 0; i < size; i++) {
288 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
290 return part ? -cos_tab[size - idx - 1] :
295 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
296 * Probably for speed reasons, the coefficients are evaluated as
297 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
298 * where s is an evaluated value, i is a value interpolated from the others
299 * and b might be either calculated or interpolated, depending on an
300 * unexplained condition.
302 * @param step the size of a block "siiiibiiii"
303 * @param in the cosinus of the LSP data
304 * @param part is 0 for 0...PI (positive cossinus values) and 1 for PI...2PI
305 (negative cossinus values)
306 * @param size the size of the whole output
308 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
309 enum FrameType ftype,
310 float *out, const float *in,
311 int size, int step, int part)
314 const ModeTab *mtab = tctx->mtab;
315 const float *cos_tab = tctx->cos_tabs[ftype];
318 for (i = 0; i < size; i += step)
320 eval_lpc_spectrum(in,
321 get_cos(i, part, cos_tab, size),
324 // Fill the 'iiiibiiii'
325 for (i = step; i <= size - 2*step; i += step) {
326 if (out[i + step] + out[i - step] > 1.95*out[i] ||
327 out[i + step] >= out[i - step]) {
328 interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
331 eval_lpc_spectrum(in,
332 get_cos(i-step/2, part, cos_tab, size),
334 interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1);
335 interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1);
339 interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
342 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
343 const float *buf, float *lpc,
346 eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0);
347 eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
349 interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
351 memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
355 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
356 * bitstream, sum the corresponding vectors and write the result to *out
359 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
360 enum FrameType ftype,
361 const int16_t *cb0, const int16_t *cb1, int cb_len)
366 for (i = 0; i < tctx->n_div[ftype]; i++) {
370 const int16_t *tab0, *tab1;
371 int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
372 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
374 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
380 tmp0 = get_bits(gb, bits);
382 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
390 tmp1 = get_bits(gb, bits);
392 tab0 = cb0 + tmp0*cb_len;
393 tab1 = cb1 + tmp1*cb_len;
395 for (j = 0; j < length; j++)
396 out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
403 static inline float mulawinv(float y, float clip, float mu)
405 y = av_clipf(y/clip, -1, 1);
406 return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
410 * Evaluate a*b/400 rounded to the nearest integer. When, for example,
411 * a*b == 200 and the nearest integer is ill-defined, use a table to emulate
412 * the following broken float-based implementation used by the binary decoder:
415 * static int very_broken_op(int a, int b)
417 * static float test; // Ugh, force gcc to do the division first...
420 * return b * test + 0.5;
424 * @note if this function is replaced by just ROUNDED_DIV(a*b,400.), the stddev
425 * between the original file (before encoding with Yamaha encoder) and the
426 * decoded output increases, which leads one to believe that the encoder expects
427 * exactly this broken calculation.
429 static int very_broken_op(int a, int b)
440 size = tabs[b/5].size;
441 rtab = tabs[b/5].tab;
442 return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
446 * Sum to data a periodic peak of a given period, width and shape.
448 * @param period the period of the peak divised by 400.0
450 static void add_peak(int period, int width, const float *shape,
451 float ppc_gain, float *speech, int len)
455 const float *shape_end = shape + len;
458 // First peak centered around zero
459 for (i = 0; i < width/2; i++)
460 speech[i] += ppc_gain * *shape++;
462 for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
463 center = very_broken_op(period, i);
464 for (j = -width/2; j < (width+1)/2; j++)
465 speech[j+center] += ppc_gain * *shape++;
468 // For the last block, be careful not to go beyond the end of the buffer
469 center = very_broken_op(period, i);
470 for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
471 speech[j+center] += ppc_gain * *shape++;
474 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
475 float ppc_gain, float *speech)
477 const ModeTab *mtab = tctx->mtab;
478 int isampf = tctx->avctx->sample_rate/1000;
479 int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
480 int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf);
481 int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
482 int period_range = max_period - min_period;
484 // This is actually the period multiplied by 400. It is just linearly coded
485 // between its maximum and minimum value.
486 int period = min_period +
487 ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
490 if (isampf == 22 && ibps == 32) {
491 // For some unknown reason, NTT decided to code this case differently...
492 width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
494 width = (period )* mtab->peak_per2wid/(400*mtab->size);
496 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
499 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
502 const ModeTab *mtab = tctx->mtab;
504 int sub = mtab->fmode[ftype].sub;
505 float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
506 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
508 if (ftype == FT_LONG) {
509 for (i = 0; i < tctx->avctx->channels; i++)
510 out[i] = (1./(1<<13)) *
511 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
514 for (i = 0; i < tctx->avctx->channels; i++) {
515 float val = (1./(1<<23)) *
516 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
519 for (j = 0; j < sub; j++) {
521 val*mulawinv(sub_step* 0.5 +
522 sub_step* get_bits(gb, SUB_GAIN_BITS),
523 SUB_AMP_MAX, MULAW_MU);
530 * Rearrange the LSP coefficients so that they have a minimum distance of
531 * min_dist. This function does it exactly as described in section of 3.2.4
532 * of the G.729 specification (but interestingly is different from what the
533 * reference decoder actually does).
535 static void rearrange_lsp(int order, float *lsp, float min_dist)
538 float min_dist2 = min_dist * 0.5;
539 for (i = 1; i < order; i++)
540 if (lsp[i] - lsp[i-1] < min_dist) {
541 float avg = (lsp[i] + lsp[i-1]) * 0.5;
543 lsp[i-1] = avg - min_dist2;
544 lsp[i ] = avg + min_dist2;
548 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
549 int lpc_hist_idx, float *lsp, float *hist)
551 const ModeTab *mtab = tctx->mtab;
554 const float *cb = mtab->lspcodebook;
555 const float *cb2 = cb + (1 << mtab->lsp_bit1)*mtab->n_lsp;
556 const float *cb3 = cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
558 const int8_t funny_rounding[4] = {
560 mtab->lsp_split == 4 ? -2 : 1,
561 mtab->lsp_split == 4 ? -2 : 1,
566 for (i = 0; i < mtab->lsp_split; i++) {
567 int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
568 for (; j < chunk_end; j++)
569 lsp[j] = cb [lpc_idx1 * mtab->n_lsp + j] +
570 cb2[lpc_idx2[i] * mtab->n_lsp + j];
573 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
575 for (i = 0; i < mtab->n_lsp; i++) {
576 float tmp1 = 1. - cb3[lpc_hist_idx*mtab->n_lsp + i];
577 float tmp2 = hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
579 lsp[i] = lsp[i] * tmp1 + tmp2;
582 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
583 rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
584 ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
587 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
588 enum FrameType ftype, float *lpc)
591 int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
593 for (i = 0; i < tctx->mtab->n_lsp; i++)
594 lsp[i] = 2*cos(lsp[i]);
598 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
601 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
604 eval_lpcenv(tctx, lsp, lpc);
609 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
610 float *in, float *prev, int ch)
612 FFTContext *mdct = &tctx->mdct_ctx[ftype];
613 const ModeTab *mtab = tctx->mtab;
614 int bsize = mtab->size / mtab->fmode[ftype].sub;
615 int size = mtab->size;
616 float *buf1 = tctx->tmp_buf;
618 int wsize; // Window size
619 float *out = tctx->curr_frame + 2*ch*mtab->size;
624 static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1};
625 int types_sizes[] = {
626 mtab->size / mtab->fmode[FT_LONG ].sub,
627 mtab->size / mtab->fmode[FT_MEDIUM].sub,
628 mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
631 wsize = types_sizes[wtype_to_wsize[wtype]];
633 prev_buf = prev + (size - bsize)/2;
635 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
636 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
638 if (!j && wtype == 4)
640 else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
643 wsize = types_sizes[wtype_to_wsize[sub_wtype]];
645 mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
647 tctx->dsp.vector_fmul_window(out2,
648 prev_buf + (bsize-wsize)/2,
650 ff_sine_windows[av_log2(wsize)],
654 memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
656 out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
658 prev_buf = buf1 + bsize*j + bsize/2;
661 tctx->last_block_pos[ch] = (size + first_wsize)/2;
664 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
667 const ModeTab *mtab = tctx->mtab;
668 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
671 for (i = 0; i < tctx->avctx->channels; i++) {
672 imdct_and_window(tctx, ftype, wtype,
673 tctx->spectrum + i*mtab->size,
674 prev_buf + 2*i*mtab->size,
678 if (tctx->avctx->channels == 2) {
679 for (i = 0; i < mtab->size - tctx->last_block_pos[0]; i++) {
680 float f1 = prev_buf[ i];
681 float f2 = prev_buf[2*mtab->size + i];
683 out[2*i + 1] = f1 - f2;
685 for (j = 0; i < mtab->size; j++,i++) {
686 float f1 = tctx->curr_frame[ j];
687 float f2 = tctx->curr_frame[2*mtab->size + j];
689 out[2*i + 1] = f1 - f2;
692 memcpy(out, prev_buf,
693 (mtab->size - tctx->last_block_pos[0]) * sizeof(*out));
695 out += mtab->size - tctx->last_block_pos[0];
697 memcpy(out, tctx->curr_frame,
698 (tctx->last_block_pos[0]) * sizeof(*out));
703 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
704 int ch, float *out, float gain, enum FrameType ftype)
706 const ModeTab *mtab = tctx->mtab;
708 float *hist = tctx->bark_hist[ftype][ch];
709 float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
710 int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
711 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
714 for (i = 0; i < fw_cb_len; i++)
715 for (j = 0; j < bark_n_coef; j++, idx++) {
717 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
718 float st = use_hist ?
719 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
722 if (st < -1.) st = 1.;
724 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
725 out += mtab->fmode[ftype].bark_tab[idx];
730 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
731 float *out, enum FrameType ftype)
733 const ModeTab *mtab = tctx->mtab;
734 int channels = tctx->avctx->channels;
735 int sub = mtab->fmode[ftype].sub;
736 int block_size = mtab->size / sub;
737 float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
738 float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
739 uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
740 uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
742 uint8_t lpc_idx1[CHANNELS_MAX];
743 uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
744 uint8_t lpc_hist_idx[CHANNELS_MAX];
748 dequant(tctx, gb, out, ftype,
749 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
750 mtab->fmode[ftype].cb_len_read);
752 for (i = 0; i < channels; i++)
753 for (j = 0; j < sub; j++)
754 for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
756 get_bits(gb, mtab->fmode[ftype].bark_n_bit);
758 for (i = 0; i < channels; i++)
759 for (j = 0; j < sub; j++)
760 bark_use_hist[i][j] = get_bits1(gb);
762 dec_gain(tctx, gb, ftype, gain);
764 for (i = 0; i < channels; i++) {
765 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
766 lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1);
768 for (j = 0; j < tctx->mtab->lsp_split; j++)
769 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
772 if (ftype == FT_LONG) {
773 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
775 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
776 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
779 for (i = 0; i < channels; i++) {
780 float *chunk = out + mtab->size * i;
781 float lsp[LSP_COEFS_MAX];
783 for (j = 0; j < sub; j++) {
784 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
785 tctx->tmp_buf, gain[sub*i+j], ftype);
787 tctx->dsp.vector_fmul(chunk + block_size*j, chunk + block_size*j, tctx->tmp_buf,
792 if (ftype == FT_LONG) {
793 float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
794 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
795 int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
797 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
799 decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
803 decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
806 dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
808 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
809 tctx->dsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
815 static int twin_decode_frame(AVCodecContext * avctx, void *data,
816 int *data_size, AVPacket *avpkt)
818 const uint8_t *buf = avpkt->data;
819 int buf_size = avpkt->size;
820 TwinContext *tctx = avctx->priv_data;
822 const ModeTab *mtab = tctx->mtab;
824 enum FrameType ftype;
826 static const enum FrameType wtype_to_ftype_table[] = {
827 FT_LONG, FT_LONG, FT_SHORT, FT_LONG,
828 FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM
831 if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
832 av_log(avctx, AV_LOG_ERROR,
833 "Frame too small (%d bytes). Truncated file?\n", buf_size);
838 init_get_bits(&gb, buf, buf_size * 8);
839 skip_bits(&gb, get_bits(&gb, 8));
840 window_type = get_bits(&gb, WINDOW_TYPE_BITS);
842 if (window_type > 8) {
843 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
847 ftype = wtype_to_ftype_table[window_type];
849 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
851 imdct_output(tctx, ftype, window_type, out);
853 FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
855 if (tctx->avctx->frame_number < 2) {
860 *data_size = mtab->size*avctx->channels*4;
866 * Init IMDCT and windowing tables
868 static av_cold void init_mdct_win(TwinContext *tctx)
871 const ModeTab *mtab = tctx->mtab;
872 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
873 int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
874 int channels = tctx->avctx->channels;
875 float norm = channels == 1 ? 2. : 1.;
877 for (i = 0; i < 3; i++) {
878 int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
879 ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
880 -sqrt(norm/bsize) / (1<<15));
883 tctx->tmp_buf = av_malloc(mtab->size * sizeof(*tctx->tmp_buf));
885 tctx->spectrum = av_malloc(2*mtab->size*channels*sizeof(float));
886 tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float));
887 tctx->prev_frame = av_malloc(2*mtab->size*channels*sizeof(float));
889 for (i = 0; i < 3; i++) {
890 int m = 4*mtab->size/mtab->fmode[i].sub;
891 double freq = 2*M_PI/m;
892 tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs));
894 for (j = 0; j <= m/8; j++)
895 tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
896 for (j = 1; j < m/8; j++)
897 tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
901 ff_init_ff_sine_windows(av_log2(size_m));
902 ff_init_ff_sine_windows(av_log2(size_s/2));
903 ff_init_ff_sine_windows(av_log2(mtab->size));
907 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
908 * each line do a cyclic permutation, i.e.
909 * abcdefghijklm -> defghijklmabc
910 * where the amount to be shifted is evaluated depending on the column.
912 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
914 const uint8_t line_len[2], int length_div,
915 enum FrameType ftype)
920 for (i = 0; i < line_len[0]; i++) {
923 if (num_blocks == 1 ||
924 (ftype == FT_LONG && num_vect % num_blocks) ||
925 (ftype != FT_LONG && num_vect & 1 ) ||
928 } else if (ftype == FT_LONG) {
933 for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
934 tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
939 * Interpret the input data as in the following table:
950 * and transpose it, giving the output
951 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
953 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
954 const uint8_t line_len[2], int length_div)
958 for (i = 0; i < num_vect; i++)
959 for (j = 0; j < line_len[i >= length_div]; j++)
960 out[cont++] = in[j*num_vect + i];
963 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
965 int block_size = size/n_blocks;
968 for (i = 0; i < size; i++)
969 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
972 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
975 const ModeTab *mtab = tctx->mtab;
976 int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
977 int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
979 if (ftype == FT_PPC) {
980 size = tctx->avctx->channels;
981 block_size = mtab->ppc_shape_len;
983 block_size = mtab->size / mtab->fmode[ftype].sub;
985 permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
986 block_size, tctx->length[ftype],
987 tctx->length_change[ftype], ftype);
989 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
990 tctx->length[ftype], tctx->length_change[ftype]);
992 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
996 static av_cold void init_bitstream_params(TwinContext *tctx)
998 const ModeTab *mtab = tctx->mtab;
999 int n_ch = tctx->avctx->channels;
1000 int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
1001 tctx->avctx->sample_rate;
1003 int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
1004 mtab->lsp_split*mtab->lsp_bit2);
1006 int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
1007 mtab->ppc_period_bit);
1009 int bsize_no_main_cb[3];
1012 enum FrameType frametype;
1014 for (i = 0; i < 3; i++)
1015 // +1 for history usage switch
1016 bse_bits[i] = n_ch *
1017 (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
1019 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1020 WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
1022 for (i = 0; i < 2; i++)
1023 bsize_no_main_cb[i] =
1024 lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
1025 mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
1027 // The remaining bits are all used for the main spectrum coefficients
1028 for (i = 0; i < 4; i++) {
1031 int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1033 bit_size = n_ch * mtab->ppc_shape_bit;
1034 vect_size = n_ch * mtab->ppc_shape_len;
1036 bit_size = total_fr_bits - bsize_no_main_cb[i];
1037 vect_size = n_ch * mtab->size;
1040 tctx->n_div[i] = (bit_size + 13) / 14;
1042 rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1043 rounded_down = (bit_size )/tctx->n_div[i];
1044 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1045 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1046 tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2;
1047 tctx->bits_main_spec[1][i][0] = (rounded_up )/2;
1048 tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
1049 tctx->bits_main_spec[1][i][1] = (rounded_down )/2;
1050 tctx->bits_main_spec_change[i] = num_rounded_up;
1052 rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1053 rounded_down = (vect_size )/tctx->n_div[i];
1054 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1055 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1056 tctx->length[i][0] = rounded_up;
1057 tctx->length[i][1] = rounded_down;
1058 tctx->length_change[i] = num_rounded_up;
1061 for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
1062 construct_perm_table(tctx, frametype);
1065 static av_cold int twin_decode_init(AVCodecContext *avctx)
1067 TwinContext *tctx = avctx->priv_data;
1068 int isampf = avctx->sample_rate/1000;
1069 int ibps = avctx->bit_rate/(1000 * avctx->channels);
1071 tctx->avctx = avctx;
1072 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
1074 if (avctx->channels > CHANNELS_MAX) {
1075 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1080 switch ((isampf << 8) + ibps) {
1081 case (8 <<8) + 8: tctx->mtab = &mode_08_08; break;
1082 case (11<<8) + 8: tctx->mtab = &mode_11_08; break;
1083 case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
1084 case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
1085 case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
1086 case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
1087 case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
1088 case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
1089 case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
1091 av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
1095 dsputil_init(&tctx->dsp, avctx);
1096 init_mdct_win(tctx);
1097 init_bitstream_params(tctx);
1099 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1104 static av_cold int twin_decode_close(AVCodecContext *avctx)
1106 TwinContext *tctx = avctx->priv_data;
1109 for (i = 0; i < 3; i++) {
1110 ff_mdct_end(&tctx->mdct_ctx[i]);
1111 av_free(tctx->cos_tabs[i]);
1115 av_free(tctx->curr_frame);
1116 av_free(tctx->spectrum);
1117 av_free(tctx->prev_frame);
1118 av_free(tctx->tmp_buf);
1123 AVCodec ff_twinvq_decoder =
1128 sizeof(TwinContext),
1133 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),