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
22 #include "libavutil/channel_layout.h"
23 #include "libavutil/float_dsp.h"
35 #include "twinvq_data.h"
38 FT_SHORT = 0, ///< Short frame (divided in n sub-blocks)
39 FT_MEDIUM, ///< Medium frame (divided in m<n sub-blocks)
40 FT_LONG, ///< Long frame (single sub-block + PPC)
41 FT_PPC, ///< Periodic Peak Component (part of the long frame)
45 * Parameters and tables that are different for each frame type
48 uint8_t sub; ///< Number subblocks in each frame
49 const uint16_t *bark_tab;
51 /** number of distinct bark scale envelope values */
52 uint8_t bark_env_size;
54 const int16_t *bark_cb; ///< codebook for the bark scale envelope (BSE)
55 uint8_t bark_n_coef;///< number of BSE CB coefficients to read
56 uint8_t bark_n_bit; ///< number of bits of the BSE coefs
59 /** main codebooks for spectrum data */
64 uint8_t cb_len_read; ///< number of spectrum coefficients to read
68 * Parameters and tables that are different for every combination of
72 struct FrameMode fmode[3]; ///< frame type-dependant parameters
74 uint16_t size; ///< frame size in samples
75 uint8_t n_lsp; ///< number of lsp coefficients
76 const float *lspcodebook;
78 /* number of bits of the different LSP CB coefficients */
83 uint8_t lsp_split; ///< number of CB entries for the LSP decoding
84 const int16_t *ppc_shape_cb; ///< PPC shape CB
86 /** number of the bits for the PPC period value */
87 uint8_t ppc_period_bit;
89 uint8_t ppc_shape_bit; ///< number of bits of the PPC shape CB coeffs
90 uint8_t ppc_shape_len; ///< size of PPC shape CB
91 uint8_t pgain_bit; ///< bits for PPC gain
93 /** constant for peak period to peak width conversion */
94 uint16_t peak_per2wid;
97 static const ModeTab mode_08_08 = {
99 { 8, bark_tab_s08_64, 10, tab.fcb08s , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
100 { 2, bark_tab_m08_256, 20, tab.fcb08m , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
101 { 1, bark_tab_l08_512, 30, tab.fcb08l , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
103 512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40
106 static const ModeTab mode_11_08 = {
108 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
109 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
110 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
112 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
115 static const ModeTab mode_11_10 = {
117 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
118 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
119 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
121 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
124 static const ModeTab mode_16_16 = {
126 { 8, bark_tab_s16_128, 10, tab.fcb16s , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
127 { 2, bark_tab_m16_512, 20, tab.fcb16m , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
128 { 1, bark_tab_l16_1024,30, tab.fcb16l , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
130 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180
133 static const ModeTab mode_22_20 = {
135 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
136 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
137 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
139 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
142 static const ModeTab mode_22_24 = {
144 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
145 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
146 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
148 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
151 static const ModeTab mode_22_32 = {
153 { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
154 { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
155 { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
157 512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
160 static const ModeTab mode_44_40 = {
162 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
163 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
164 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
166 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
169 static const ModeTab mode_44_48 = {
171 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
172 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
173 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
175 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
178 typedef struct TwinContext {
179 AVCodecContext *avctx;
180 AVFloatDSPContext fdsp;
181 FFTContext mdct_ctx[3];
186 float lsp_hist[2][20]; ///< LSP coefficients of the last frame
187 float bark_hist[3][2][40]; ///< BSE coefficients of last frame
189 // bitstream parameters
190 int16_t permut[4][4096];
191 uint8_t length[4][2]; ///< main codebook stride
192 uint8_t length_change[4];
193 uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
194 int bits_main_spec_change[4];
198 float *curr_frame; ///< non-interleaved output
199 float *prev_frame; ///< non-interleaved previous frame
200 int last_block_pos[2];
201 int discarded_packets;
209 #define PPC_SHAPE_CB_SIZE 64
210 #define PPC_SHAPE_LEN_MAX 60
211 #define SUB_AMP_MAX 4500.0
212 #define MULAW_MU 100.0
214 #define AMP_MAX 13000.0
215 #define SUB_GAIN_BITS 5
216 #define WINDOW_TYPE_BITS 4
218 #define LSP_COEFS_MAX 20
219 #define LSP_SPLIT_MAX 4
220 #define CHANNELS_MAX 2
221 #define SUBBLOCKS_MAX 16
222 #define BARK_N_COEF_MAX 4
224 /** @note not speed critical, hence not optimized */
225 static void memset_float(float *buf, float val, int size)
232 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
235 * @param lsp a vector of the cosinus of the LSP values
236 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
237 * @param order the order of the LSP (and the size of the *lsp buffer). Must
238 * be a multiple of four.
239 * @return the LPC value
241 * @todo reuse code from Vorbis decoder: vorbis_floor0_decode
243 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
248 float two_cos_w = 2.0f*cos_val;
250 for (j = 0; j + 1 < order; j += 2*2) {
251 // Unroll the loop once since order is a multiple of four
252 q *= lsp[j ] - two_cos_w;
253 p *= lsp[j+1] - two_cos_w;
255 q *= lsp[j+2] - two_cos_w;
256 p *= lsp[j+3] - two_cos_w;
259 p *= p * (2.0f - two_cos_w);
260 q *= q * (2.0f + two_cos_w);
262 return 0.5 / (p + q);
266 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
268 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
271 const ModeTab *mtab = tctx->mtab;
272 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
274 for (i = 0; i < size_s/2; i++) {
275 float cos_i = tctx->cos_tabs[0][i];
276 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
277 lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
281 static void interpolate(float *out, float v1, float v2, int size)
284 float step = (v1 - v2)/(size + 1);
286 for (i = 0; i < size; i++) {
292 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
294 return part ? -cos_tab[size - idx - 1] :
299 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
300 * Probably for speed reasons, the coefficients are evaluated as
301 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
302 * where s is an evaluated value, i is a value interpolated from the others
303 * and b might be either calculated or interpolated, depending on an
304 * unexplained condition.
306 * @param step the size of a block "siiiibiiii"
307 * @param in the cosinus of the LSP data
308 * @param part is 0 for 0...PI (positive cossinus values) and 1 for PI...2PI
309 (negative cossinus values)
310 * @param size the size of the whole output
312 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
313 enum FrameType ftype,
314 float *out, const float *in,
315 int size, int step, int part)
318 const ModeTab *mtab = tctx->mtab;
319 const float *cos_tab = tctx->cos_tabs[ftype];
322 for (i = 0; i < size; i += step)
324 eval_lpc_spectrum(in,
325 get_cos(i, part, cos_tab, size),
328 // Fill the 'iiiibiiii'
329 for (i = step; i <= size - 2*step; i += step) {
330 if (out[i + step] + out[i - step] > 1.95*out[i] ||
331 out[i + step] >= out[i - step]) {
332 interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
335 eval_lpc_spectrum(in,
336 get_cos(i-step/2, part, cos_tab, size),
338 interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1);
339 interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1);
343 interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
346 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
347 const float *buf, float *lpc,
350 eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0);
351 eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
353 interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
355 memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
359 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
360 * bitstream, sum the corresponding vectors and write the result to *out
363 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
364 enum FrameType ftype,
365 const int16_t *cb0, const int16_t *cb1, int cb_len)
370 for (i = 0; i < tctx->n_div[ftype]; i++) {
374 const int16_t *tab0, *tab1;
375 int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
376 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
378 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
384 tmp0 = get_bits(gb, bits);
386 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
394 tmp1 = get_bits(gb, bits);
396 tab0 = cb0 + tmp0*cb_len;
397 tab1 = cb1 + tmp1*cb_len;
399 for (j = 0; j < length; j++)
400 out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
407 static inline float mulawinv(float y, float clip, float mu)
409 y = av_clipf(y/clip, -1, 1);
410 return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
414 * Evaluate a*b/400 rounded to the nearest integer. When, for example,
415 * a*b == 200 and the nearest integer is ill-defined, use a table to emulate
416 * the following broken float-based implementation used by the binary decoder:
419 * static int very_broken_op(int a, int b)
421 * static float test; // Ugh, force gcc to do the division first...
424 * return b * test + 0.5;
428 * @note if this function is replaced by just ROUNDED_DIV(a*b,400.), the stddev
429 * between the original file (before encoding with Yamaha encoder) and the
430 * decoded output increases, which leads one to believe that the encoder expects
431 * exactly this broken calculation.
433 static int very_broken_op(int a, int b)
444 size = tabs[b/5].size;
445 rtab = tabs[b/5].tab;
446 return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
450 * Sum to data a periodic peak of a given period, width and shape.
452 * @param period the period of the peak divised by 400.0
454 static void add_peak(int period, int width, const float *shape,
455 float ppc_gain, float *speech, int len)
459 const float *shape_end = shape + len;
462 // First peak centered around zero
463 for (i = 0; i < width/2; i++)
464 speech[i] += ppc_gain * *shape++;
466 for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
467 center = very_broken_op(period, i);
468 for (j = -width/2; j < (width+1)/2; j++)
469 speech[j+center] += ppc_gain * *shape++;
472 // For the last block, be careful not to go beyond the end of the buffer
473 center = very_broken_op(period, i);
474 for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
475 speech[j+center] += ppc_gain * *shape++;
478 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
479 float ppc_gain, float *speech)
481 const ModeTab *mtab = tctx->mtab;
482 int isampf = tctx->avctx->sample_rate/1000;
483 int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
484 int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf);
485 int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
486 int period_range = max_period - min_period;
488 // This is actually the period multiplied by 400. It is just linearly coded
489 // between its maximum and minimum value.
490 int period = min_period +
491 ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
494 if (isampf == 22 && ibps == 32) {
495 // For some unknown reason, NTT decided to code this case differently...
496 width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
498 width = (period )* mtab->peak_per2wid/(400*mtab->size);
500 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
503 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
506 const ModeTab *mtab = tctx->mtab;
508 int sub = mtab->fmode[ftype].sub;
509 float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
510 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
512 if (ftype == FT_LONG) {
513 for (i = 0; i < tctx->avctx->channels; i++)
514 out[i] = (1./(1<<13)) *
515 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
518 for (i = 0; i < tctx->avctx->channels; i++) {
519 float val = (1./(1<<23)) *
520 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
523 for (j = 0; j < sub; j++) {
525 val*mulawinv(sub_step* 0.5 +
526 sub_step* get_bits(gb, SUB_GAIN_BITS),
527 SUB_AMP_MAX, MULAW_MU);
534 * Rearrange the LSP coefficients so that they have a minimum distance of
535 * min_dist. This function does it exactly as described in section of 3.2.4
536 * of the G.729 specification (but interestingly is different from what the
537 * reference decoder actually does).
539 static void rearrange_lsp(int order, float *lsp, float min_dist)
542 float min_dist2 = min_dist * 0.5;
543 for (i = 1; i < order; i++)
544 if (lsp[i] - lsp[i-1] < min_dist) {
545 float avg = (lsp[i] + lsp[i-1]) * 0.5;
547 lsp[i-1] = avg - min_dist2;
548 lsp[i ] = avg + min_dist2;
552 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
553 int lpc_hist_idx, float *lsp, float *hist)
555 const ModeTab *mtab = tctx->mtab;
558 const float *cb = mtab->lspcodebook;
559 const float *cb2 = cb + (1 << mtab->lsp_bit1)*mtab->n_lsp;
560 const float *cb3 = cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
562 const int8_t funny_rounding[4] = {
564 mtab->lsp_split == 4 ? -2 : 1,
565 mtab->lsp_split == 4 ? -2 : 1,
570 for (i = 0; i < mtab->lsp_split; i++) {
571 int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
572 for (; j < chunk_end; j++)
573 lsp[j] = cb [lpc_idx1 * mtab->n_lsp + j] +
574 cb2[lpc_idx2[i] * mtab->n_lsp + j];
577 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
579 for (i = 0; i < mtab->n_lsp; i++) {
580 float tmp1 = 1. - cb3[lpc_hist_idx*mtab->n_lsp + i];
581 float tmp2 = hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
583 lsp[i] = lsp[i] * tmp1 + tmp2;
586 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
587 rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
588 ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
591 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
592 enum FrameType ftype, float *lpc)
595 int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
597 for (i = 0; i < tctx->mtab->n_lsp; i++)
598 lsp[i] = 2*cos(lsp[i]);
602 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
605 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
608 eval_lpcenv(tctx, lsp, lpc);
613 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
614 float *in, float *prev, int ch)
616 FFTContext *mdct = &tctx->mdct_ctx[ftype];
617 const ModeTab *mtab = tctx->mtab;
618 int bsize = mtab->size / mtab->fmode[ftype].sub;
619 int size = mtab->size;
620 float *buf1 = tctx->tmp_buf;
622 int wsize; // Window size
623 float *out = tctx->curr_frame + 2*ch*mtab->size;
628 static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1};
629 int types_sizes[] = {
630 mtab->size / mtab->fmode[FT_LONG ].sub,
631 mtab->size / mtab->fmode[FT_MEDIUM].sub,
632 mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
635 wsize = types_sizes[wtype_to_wsize[wtype]];
637 prev_buf = prev + (size - bsize)/2;
639 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
640 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
642 if (!j && wtype == 4)
644 else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
647 wsize = types_sizes[wtype_to_wsize[sub_wtype]];
649 mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
651 tctx->fdsp.vector_fmul_window(out2, prev_buf + (bsize-wsize) / 2,
653 ff_sine_windows[av_log2(wsize)],
657 memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
659 out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
661 prev_buf = buf1 + bsize*j + bsize/2;
664 tctx->last_block_pos[ch] = (size + first_wsize)/2;
667 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
670 const ModeTab *mtab = tctx->mtab;
672 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
675 for (i = 0; i < tctx->avctx->channels; i++) {
676 imdct_and_window(tctx, ftype, wtype,
677 tctx->spectrum + i*mtab->size,
678 prev_buf + 2*i*mtab->size,
685 size2 = tctx->last_block_pos[0];
686 size1 = mtab->size - size2;
688 memcpy(&out[0][0 ], prev_buf, size1 * sizeof(out[0][0]));
689 memcpy(&out[0][size1], tctx->curr_frame, size2 * sizeof(out[0][0]));
691 if (tctx->avctx->channels == 2) {
692 memcpy(&out[1][0], &prev_buf[2*mtab->size], size1 * sizeof(out[1][0]));
693 memcpy(&out[1][size1], &tctx->curr_frame[2*mtab->size], size2 * sizeof(out[1][0]));
694 tctx->fdsp.butterflies_float(out[0], out[1], mtab->size);
698 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
699 int ch, float *out, float gain, enum FrameType ftype)
701 const ModeTab *mtab = tctx->mtab;
703 float *hist = tctx->bark_hist[ftype][ch];
704 float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
705 int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
706 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
709 for (i = 0; i < fw_cb_len; i++)
710 for (j = 0; j < bark_n_coef; j++, idx++) {
712 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
713 float st = use_hist ?
714 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
717 if (st < -1.) st = 1.;
719 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
720 out += mtab->fmode[ftype].bark_tab[idx];
725 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
726 float *out, enum FrameType ftype)
728 const ModeTab *mtab = tctx->mtab;
729 int channels = tctx->avctx->channels;
730 int sub = mtab->fmode[ftype].sub;
731 int block_size = mtab->size / sub;
732 float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
733 float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
734 uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
735 uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
737 uint8_t lpc_idx1[CHANNELS_MAX];
738 uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
739 uint8_t lpc_hist_idx[CHANNELS_MAX];
743 dequant(tctx, gb, out, ftype,
744 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
745 mtab->fmode[ftype].cb_len_read);
747 for (i = 0; i < channels; i++)
748 for (j = 0; j < sub; j++)
749 for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
751 get_bits(gb, mtab->fmode[ftype].bark_n_bit);
753 for (i = 0; i < channels; i++)
754 for (j = 0; j < sub; j++)
755 bark_use_hist[i][j] = get_bits1(gb);
757 dec_gain(tctx, gb, ftype, gain);
759 for (i = 0; i < channels; i++) {
760 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
761 lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1);
763 for (j = 0; j < tctx->mtab->lsp_split; j++)
764 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
767 if (ftype == FT_LONG) {
768 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
770 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
771 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
774 for (i = 0; i < channels; i++) {
775 float *chunk = out + mtab->size * i;
776 float lsp[LSP_COEFS_MAX];
778 for (j = 0; j < sub; j++) {
779 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
780 tctx->tmp_buf, gain[sub*i+j], ftype);
782 tctx->fdsp.vector_fmul(chunk + block_size*j, chunk + block_size*j,
783 tctx->tmp_buf, block_size);
787 if (ftype == FT_LONG) {
788 float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
789 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
790 int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
792 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
794 decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
798 decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
801 dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
803 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
804 tctx->fdsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
810 static int twin_decode_frame(AVCodecContext * avctx, void *data,
811 int *got_frame_ptr, AVPacket *avpkt)
813 AVFrame *frame = data;
814 const uint8_t *buf = avpkt->data;
815 int buf_size = avpkt->size;
816 TwinContext *tctx = avctx->priv_data;
818 const ModeTab *mtab = tctx->mtab;
820 enum FrameType ftype;
821 int window_type, ret;
822 static const enum FrameType wtype_to_ftype_table[] = {
823 FT_LONG, FT_LONG, FT_SHORT, FT_LONG,
824 FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM
827 if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
828 av_log(avctx, AV_LOG_ERROR,
829 "Frame too small (%d bytes). Truncated file?\n", buf_size);
830 return AVERROR(EINVAL);
833 /* get output buffer */
834 if (tctx->discarded_packets >= 2) {
835 frame->nb_samples = mtab->size;
836 if ((ret = ff_get_buffer(avctx, frame)) < 0) {
837 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
840 out = (float **)frame->extended_data;
843 init_get_bits(&gb, buf, buf_size * 8);
844 skip_bits(&gb, get_bits(&gb, 8));
845 window_type = get_bits(&gb, WINDOW_TYPE_BITS);
847 if (window_type > 8) {
848 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
852 ftype = wtype_to_ftype_table[window_type];
854 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
856 imdct_output(tctx, ftype, window_type, out);
858 FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
860 if (tctx->discarded_packets < 2) {
861 tctx->discarded_packets++;
872 * Init IMDCT and windowing tables
874 static av_cold int init_mdct_win(TwinContext *tctx)
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.;
883 for (i = 0; i < 3; i++) {
884 int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
885 if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
886 -sqrt(norm/bsize) / (1<<15))))
890 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
891 mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
893 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
894 2 * mtab->size * channels * sizeof(*tctx->spectrum),
896 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
897 2 * mtab->size * channels * sizeof(*tctx->curr_frame),
899 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
900 2 * mtab->size * channels * sizeof(*tctx->prev_frame),
903 for (i = 0; i < 3; i++) {
904 int m = 4*mtab->size/mtab->fmode[i].sub;
905 double freq = 2*M_PI/m;
906 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
907 (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
909 for (j = 0; j <= m/8; j++)
910 tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
911 for (j = 1; j < m/8; j++)
912 tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
916 ff_init_ff_sine_windows(av_log2(size_m));
917 ff_init_ff_sine_windows(av_log2(size_s/2));
918 ff_init_ff_sine_windows(av_log2(mtab->size));
922 return AVERROR(ENOMEM);
926 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
927 * each line do a cyclic permutation, i.e.
928 * abcdefghijklm -> defghijklmabc
929 * where the amount to be shifted is evaluated depending on the column.
931 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
933 const uint8_t line_len[2], int length_div,
934 enum FrameType ftype)
939 for (i = 0; i < line_len[0]; i++) {
942 if (num_blocks == 1 ||
943 (ftype == FT_LONG && num_vect % num_blocks) ||
944 (ftype != FT_LONG && num_vect & 1 ) ||
947 } else if (ftype == FT_LONG) {
952 for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
953 tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
958 * Interpret the input data as in the following table:
969 * and transpose it, giving the output
970 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
972 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
973 const uint8_t line_len[2], int length_div)
977 for (i = 0; i < num_vect; i++)
978 for (j = 0; j < line_len[i >= length_div]; j++)
979 out[cont++] = in[j*num_vect + i];
982 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
984 int block_size = size/n_blocks;
987 for (i = 0; i < size; i++)
988 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
991 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
994 const ModeTab *mtab = tctx->mtab;
996 int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
998 if (ftype == FT_PPC) {
999 size = tctx->avctx->channels;
1000 block_size = mtab->ppc_shape_len;
1002 size = tctx->avctx->channels * mtab->fmode[ftype].sub;
1003 block_size = mtab->size / mtab->fmode[ftype].sub;
1006 permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
1007 block_size, tctx->length[ftype],
1008 tctx->length_change[ftype], ftype);
1010 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
1011 tctx->length[ftype], tctx->length_change[ftype]);
1013 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
1017 static av_cold void init_bitstream_params(TwinContext *tctx)
1019 const ModeTab *mtab = tctx->mtab;
1020 int n_ch = tctx->avctx->channels;
1021 int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
1022 tctx->avctx->sample_rate;
1024 int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
1025 mtab->lsp_split*mtab->lsp_bit2);
1027 int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
1028 mtab->ppc_period_bit);
1030 int bsize_no_main_cb[3];
1033 enum FrameType frametype;
1035 for (i = 0; i < 3; i++)
1036 // +1 for history usage switch
1037 bse_bits[i] = n_ch *
1038 (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
1040 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1041 WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
1043 for (i = 0; i < 2; i++)
1044 bsize_no_main_cb[i] =
1045 lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
1046 mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
1048 // The remaining bits are all used for the main spectrum coefficients
1049 for (i = 0; i < 4; i++) {
1052 int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1054 bit_size = n_ch * mtab->ppc_shape_bit;
1055 vect_size = n_ch * mtab->ppc_shape_len;
1057 bit_size = total_fr_bits - bsize_no_main_cb[i];
1058 vect_size = n_ch * mtab->size;
1061 tctx->n_div[i] = (bit_size + 13) / 14;
1063 rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1064 rounded_down = (bit_size )/tctx->n_div[i];
1065 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1066 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1067 tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2;
1068 tctx->bits_main_spec[1][i][0] = (rounded_up )/2;
1069 tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
1070 tctx->bits_main_spec[1][i][1] = (rounded_down )/2;
1071 tctx->bits_main_spec_change[i] = num_rounded_up;
1073 rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1074 rounded_down = (vect_size )/tctx->n_div[i];
1075 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1076 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1077 tctx->length[i][0] = rounded_up;
1078 tctx->length[i][1] = rounded_down;
1079 tctx->length_change[i] = num_rounded_up;
1082 for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
1083 construct_perm_table(tctx, frametype);
1086 static av_cold int twin_decode_close(AVCodecContext *avctx)
1088 TwinContext *tctx = avctx->priv_data;
1091 for (i = 0; i < 3; i++) {
1092 ff_mdct_end(&tctx->mdct_ctx[i]);
1093 av_free(tctx->cos_tabs[i]);
1097 av_free(tctx->curr_frame);
1098 av_free(tctx->spectrum);
1099 av_free(tctx->prev_frame);
1100 av_free(tctx->tmp_buf);
1105 static av_cold int twin_decode_init(AVCodecContext *avctx)
1108 TwinContext *tctx = avctx->priv_data;
1111 tctx->avctx = avctx;
1112 avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
1114 if (!avctx->extradata || avctx->extradata_size < 12) {
1115 av_log(avctx, AV_LOG_ERROR, "Missing or incomplete extradata\n");
1116 return AVERROR_INVALIDDATA;
1118 avctx->channels = AV_RB32(avctx->extradata ) + 1;
1119 avctx->bit_rate = AV_RB32(avctx->extradata + 4) * 1000;
1120 isampf = AV_RB32(avctx->extradata + 8);
1122 if (isampf < 8 || isampf > 44) {
1123 av_log(avctx, AV_LOG_ERROR, "Unsupported sample rate\n");
1124 return AVERROR_INVALIDDATA;
1127 case 44: avctx->sample_rate = 44100; break;
1128 case 22: avctx->sample_rate = 22050; break;
1129 case 11: avctx->sample_rate = 11025; break;
1130 default: avctx->sample_rate = isampf * 1000; break;
1133 if (avctx->channels <= 0 || avctx->channels > CHANNELS_MAX) {
1134 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1138 avctx->channel_layout = avctx->channels == 1 ? AV_CH_LAYOUT_MONO :
1139 AV_CH_LAYOUT_STEREO;
1141 ibps = avctx->bit_rate / (1000 * avctx->channels);
1144 av_log(avctx, AV_LOG_ERROR, "unsupported per channel bitrate %dkbps\n", ibps);
1145 return AVERROR_INVALIDDATA;
1148 switch ((isampf << 8) + ibps) {
1149 case (8 <<8) + 8: tctx->mtab = &mode_08_08; break;
1150 case (11<<8) + 8: tctx->mtab = &mode_11_08; break;
1151 case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
1152 case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
1153 case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
1154 case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
1155 case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
1156 case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
1157 case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
1159 av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
1163 avpriv_float_dsp_init(&tctx->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
1164 if ((ret = init_mdct_win(tctx))) {
1165 av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
1166 twin_decode_close(avctx);
1169 init_bitstream_params(tctx);
1171 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1176 AVCodec ff_twinvq_decoder = {
1178 .type = AVMEDIA_TYPE_AUDIO,
1179 .id = AV_CODEC_ID_TWINVQ,
1180 .priv_data_size = sizeof(TwinContext),
1181 .init = twin_decode_init,
1182 .close = twin_decode_close,
1183 .decode = twin_decode_frame,
1184 .capabilities = CODEC_CAP_DR1,
1185 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
1186 .sample_fmts = (const enum AVSampleFormat[]) { AV_SAMPLE_FMT_FLTP,
1187 AV_SAMPLE_FMT_NONE },