3 * Copyright (c) 2009 Vitor Sessak
5 * This file is part of Libav.
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #include "libavutil/channel_layout.h"
26 #include "libavutil/float_dsp.h"
33 #include "twinvq_data.h"
36 FT_SHORT = 0, ///< Short frame (divided in n sub-blocks)
37 FT_MEDIUM, ///< Medium frame (divided in m<n sub-blocks)
38 FT_LONG, ///< Long frame (single sub-block + PPC)
39 FT_PPC, ///< Periodic Peak Component (part of the long frame)
43 * Parameters and tables that are different for each frame type
46 uint8_t sub; ///< Number subblocks in each frame
47 const uint16_t *bark_tab;
49 /** number of distinct bark scale envelope values */
50 uint8_t bark_env_size;
52 const int16_t *bark_cb; ///< codebook for the bark scale envelope (BSE)
53 uint8_t bark_n_coef;///< number of BSE CB coefficients to read
54 uint8_t bark_n_bit; ///< number of bits of the BSE coefs
57 /** main codebooks for spectrum data */
62 uint8_t cb_len_read; ///< number of spectrum coefficients to read
66 * Parameters and tables that are different for every combination of
70 struct FrameMode fmode[3]; ///< frame type-dependant parameters
72 uint16_t size; ///< frame size in samples
73 uint8_t n_lsp; ///< number of lsp coefficients
74 const float *lspcodebook;
76 /* number of bits of the different LSP CB coefficients */
81 uint8_t lsp_split; ///< number of CB entries for the LSP decoding
82 const int16_t *ppc_shape_cb; ///< PPC shape CB
84 /** number of the bits for the PPC period value */
85 uint8_t ppc_period_bit;
87 uint8_t ppc_shape_bit; ///< number of bits of the PPC shape CB coeffs
88 uint8_t ppc_shape_len; ///< size of PPC shape CB
89 uint8_t pgain_bit; ///< bits for PPC gain
91 /** constant for peak period to peak width conversion */
92 uint16_t peak_per2wid;
95 static const ModeTab mode_08_08 = {
97 { 8, bark_tab_s08_64, 10, tab.fcb08s, 1, 5, tab.cb0808s0, tab.cb0808s1, 18 },
98 { 2, bark_tab_m08_256, 20, tab.fcb08m, 2, 5, tab.cb0808m0, tab.cb0808m1, 16 },
99 { 1, bark_tab_l08_512, 30, tab.fcb08l, 3, 6, tab.cb0808l0, tab.cb0808l1, 17 }
101 512, 12, tab.lsp08, 1, 5, 3, 3, tab.shape08, 8, 28, 20, 6, 40
104 static const ModeTab mode_11_08 = {
106 { 8, bark_tab_s11_64, 10, tab.fcb11s, 1, 5, tab.cb1108s0, tab.cb1108s1, 29 },
107 { 2, bark_tab_m11_256, 20, tab.fcb11m, 2, 5, tab.cb1108m0, tab.cb1108m1, 24 },
108 { 1, bark_tab_l11_512, 30, tab.fcb11l, 3, 6, tab.cb1108l0, tab.cb1108l1, 27 }
110 512, 16, tab.lsp11, 1, 6, 4, 3, tab.shape11, 9, 36, 30, 7, 90
113 static const ModeTab mode_11_10 = {
115 { 8, bark_tab_s11_64, 10, tab.fcb11s, 1, 5, tab.cb1110s0, tab.cb1110s1, 21 },
116 { 2, bark_tab_m11_256, 20, tab.fcb11m, 2, 5, tab.cb1110m0, tab.cb1110m1, 18 },
117 { 1, bark_tab_l11_512, 30, tab.fcb11l, 3, 6, tab.cb1110l0, tab.cb1110l1, 20 }
119 512, 16, tab.lsp11, 1, 6, 4, 3, tab.shape11, 9, 36, 30, 7, 90
122 static const ModeTab mode_16_16 = {
124 { 8, bark_tab_s16_128, 10, tab.fcb16s, 1, 5, tab.cb1616s0, tab.cb1616s1, 16 },
125 { 2, bark_tab_m16_512, 20, tab.fcb16m, 2, 5, tab.cb1616m0, tab.cb1616m1, 15 },
126 { 1, bark_tab_l16_1024, 30, tab.fcb16l, 3, 6, tab.cb1616l0, tab.cb1616l1, 16 }
128 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16, 9, 56, 60, 7, 180
131 static const ModeTab mode_22_20 = {
133 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18 },
134 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17 },
135 { 1, bark_tab_l22_1024, 32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18 }
137 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
140 static const ModeTab mode_22_24 = {
142 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15 },
143 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14 },
144 { 1, bark_tab_l22_1024, 32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15 }
146 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
149 static const ModeTab mode_22_32 = {
151 { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11 },
152 { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11 },
153 { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12 }
155 512, 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
158 static const ModeTab mode_44_40 = {
160 { 16, bark_tab_s44_128, 10, tab.fcb44s, 1, 6, tab.cb4440s0, tab.cb4440s1, 18 },
161 { 4, bark_tab_m44_512, 20, tab.fcb44m, 2, 6, tab.cb4440m0, tab.cb4440m1, 17 },
162 { 1, bark_tab_l44_2048, 40, tab.fcb44l, 4, 6, tab.cb4440l0, tab.cb4440l1, 17 }
164 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44, 9, 84, 54, 7, 432
167 static const ModeTab mode_44_48 = {
169 { 16, bark_tab_s44_128, 10, tab.fcb44s, 1, 6, tab.cb4448s0, tab.cb4448s1, 15 },
170 { 4, bark_tab_m44_512, 20, tab.fcb44m, 2, 6, tab.cb4448m0, tab.cb4448m1, 14 },
171 { 1, bark_tab_l44_2048, 40, tab.fcb44l, 4, 6, tab.cb4448l0, tab.cb4448l1, 14 }
173 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44, 9, 84, 54, 7, 432
176 typedef struct TwinContext {
177 AVCodecContext *avctx;
178 AVFloatDSPContext fdsp;
179 FFTContext mdct_ctx[3];
184 float lsp_hist[2][20]; ///< LSP coefficients of the last frame
185 float bark_hist[3][2][40]; ///< BSE coefficients of last frame
187 // bitstream parameters
188 int16_t permut[4][4096];
189 uint8_t length[4][2]; ///< main codebook stride
190 uint8_t length_change[4];
191 uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
192 int bits_main_spec_change[4];
196 float *curr_frame; ///< non-interleaved output
197 float *prev_frame; ///< non-interleaved previous frame
198 int last_block_pos[2];
199 int discarded_packets;
207 #define PPC_SHAPE_CB_SIZE 64
208 #define PPC_SHAPE_LEN_MAX 60
209 #define SUB_AMP_MAX 4500.0
210 #define MULAW_MU 100.0
212 #define AMP_MAX 13000.0
213 #define SUB_GAIN_BITS 5
214 #define WINDOW_TYPE_BITS 4
216 #define LSP_COEFS_MAX 20
217 #define LSP_SPLIT_MAX 4
218 #define CHANNELS_MAX 2
219 #define SUBBLOCKS_MAX 16
220 #define BARK_N_COEF_MAX 4
222 /** @note not speed critical, hence not optimized */
223 static void memset_float(float *buf, float val, int size)
230 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
233 * @param lsp a vector of the cosine of the LSP values
234 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
235 * @param order the order of the LSP (and the size of the *lsp buffer). Must
236 * be a multiple of four.
237 * @return the LPC value
239 * @todo reuse code from Vorbis decoder: vorbis_floor0_decode
241 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
246 float two_cos_w = 2.0f * cos_val;
248 for (j = 0; j + 1 < order; j += 2 * 2) {
249 // Unroll the loop once since order is a multiple of four
250 q *= lsp[j] - two_cos_w;
251 p *= lsp[j + 1] - two_cos_w;
253 q *= lsp[j + 2] - two_cos_w;
254 p *= lsp[j + 3] - two_cos_w;
257 p *= p * (2.0f - two_cos_w);
258 q *= q * (2.0f + two_cos_w);
260 return 0.5 / (p + q);
264 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
266 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
269 const ModeTab *mtab = tctx->mtab;
270 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
272 for (i = 0; i < size_s / 2; i++) {
273 float cos_i = tctx->cos_tabs[0][i];
274 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
275 lpc[size_s - i - 1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
279 static void interpolate(float *out, float v1, float v2, int size)
282 float step = (v1 - v2) / (size + 1);
284 for (i = 0; i < size; i++) {
290 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
292 return part ? -cos_tab[size - idx - 1]
297 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
298 * Probably for speed reasons, the coefficients are evaluated as
299 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
300 * where s is an evaluated value, i is a value interpolated from the others
301 * and b might be either calculated or interpolated, depending on an
302 * unexplained condition.
304 * @param step the size of a block "siiiibiiii"
305 * @param in the cosine of the LSP data
306 * @param part is 0 for 0...PI (positive cosine values) and 1 for PI...2PI
307 * (negative cosine values)
308 * @param size the size of the whole output
310 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
311 enum FrameType ftype,
312 float *out, const float *in,
313 int size, int step, int part)
316 const ModeTab *mtab = tctx->mtab;
317 const float *cos_tab = tctx->cos_tabs[ftype];
320 for (i = 0; i < size; i += step)
322 eval_lpc_spectrum(in,
323 get_cos(i, part, cos_tab, size),
326 // Fill the 'iiiibiiii'
327 for (i = step; i <= size - 2 * step; i += step) {
328 if (out[i + step] + out[i - step] > 1.95 * out[i] ||
329 out[i + step] >= out[i - step]) {
330 interpolate(out + i - step + 1, out[i], out[i - step], step - 1);
333 eval_lpc_spectrum(in,
334 get_cos(i - step / 2, part, cos_tab, size),
336 interpolate(out + i - step + 1, out[i - step / 2],
337 out[i - step], step / 2 - 1);
338 interpolate(out + i - step / 2 + 1, out[i],
339 out[i - step / 2], step / 2 - 1);
343 interpolate(out + size - 2 * step + 1, out[size - step],
344 out[size - 2 * step], step - 1);
347 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
348 const float *buf, float *lpc,
351 eval_lpcenv_or_interp(tctx, ftype, lpc, buf, size / 2, step, 0);
352 eval_lpcenv_or_interp(tctx, ftype, lpc + size / 2, buf, size / 2,
355 interpolate(lpc + size / 2 - step + 1, lpc[size / 2],
356 lpc[size / 2 - step], step);
358 memset_float(lpc + size - 2 * step + 1, lpc[size - 2 * step], 2 * step - 1);
362 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
363 * bitstream, sum the corresponding vectors and write the result to *out
366 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
367 enum FrameType ftype,
368 const int16_t *cb0, const int16_t *cb1, int cb_len)
373 for (i = 0; i < tctx->n_div[ftype]; i++) {
377 const int16_t *tab0, *tab1;
378 int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
379 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
381 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
387 tmp0 = get_bits(gb, bits);
389 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
397 tmp1 = get_bits(gb, bits);
399 tab0 = cb0 + tmp0 * cb_len;
400 tab1 = cb1 + tmp1 * cb_len;
402 for (j = 0; j < length; j++)
403 out[tctx->permut[ftype][pos + j]] = sign0 * tab0[j] +
410 static inline float mulawinv(float y, float clip, float mu)
412 y = av_clipf(y / clip, -1, 1);
413 return clip * FFSIGN(y) * (exp(log(1 + mu) * fabs(y)) - 1) / mu;
417 * Evaluate a * b / 400 rounded to the nearest integer. When, for example,
418 * a * b == 200 and the nearest integer is ill-defined, use a table to emulate
419 * the following broken float-based implementation used by the binary decoder:
422 * static int very_broken_op(int a, int b)
424 * static float test; // Ugh, force gcc to do the division first...
427 * return b * test + 0.5;
431 * @note if this function is replaced by just ROUNDED_DIV(a * b, 400.0), the
432 * stddev between the original file (before encoding with Yamaha encoder) and
433 * the decoded output increases, which leads one to believe that the encoder
434 * expects exactly this broken calculation.
436 static int very_broken_op(int a, int b)
442 if (x % 400 || b % 5)
447 size = tabs[b / 5].size;
448 rtab = tabs[b / 5].tab;
449 return x - rtab[size * av_log2(2 * (x - 1) / size) + (x - 1) % size];
453 * Sum to data a periodic peak of a given period, width and shape.
455 * @param period the period of the peak divised by 400.0
457 static void add_peak(int period, int width, const float *shape,
458 float ppc_gain, float *speech, int len)
462 const float *shape_end = shape + len;
465 // First peak centered around zero
466 for (i = 0; i < width / 2; i++)
467 speech[i] += ppc_gain * *shape++;
469 for (i = 1; i < ROUNDED_DIV(len, width); i++) {
470 center = very_broken_op(period, i);
471 for (j = -width / 2; j < (width + 1) / 2; j++)
472 speech[j + center] += ppc_gain * *shape++;
475 // For the last block, be careful not to go beyond the end of the buffer
476 center = very_broken_op(period, i);
477 for (j = -width / 2; j < (width + 1) / 2 && shape < shape_end; j++)
478 speech[j + center] += ppc_gain * *shape++;
481 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
482 float ppc_gain, float *speech)
484 const ModeTab *mtab = tctx->mtab;
485 int isampf = tctx->avctx->sample_rate / 1000;
486 int ibps = tctx->avctx->bit_rate / (1000 * tctx->avctx->channels);
487 int min_period = ROUNDED_DIV(40 * 2 * mtab->size, isampf);
488 int max_period = ROUNDED_DIV(40 * 2 * mtab->size * 6, isampf);
489 int period_range = max_period - min_period;
491 // This is actually the period multiplied by 400. It is just linearly coded
492 // between its maximum and minimum value.
493 int period = min_period +
494 ROUNDED_DIV(period_coef * period_range,
495 (1 << mtab->ppc_period_bit) - 1);
498 if (isampf == 22 && ibps == 32) {
499 // For some unknown reason, NTT decided to code this case differently...
500 width = ROUNDED_DIV((period + 800) * mtab->peak_per2wid,
503 width = period * mtab->peak_per2wid / (400 * mtab->size);
505 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
508 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
511 const ModeTab *mtab = tctx->mtab;
513 int sub = mtab->fmode[ftype].sub;
514 float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
515 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
517 if (ftype == FT_LONG) {
518 for (i = 0; i < tctx->avctx->channels; i++)
519 out[i] = (1.0 / (1 << 13)) *
520 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
523 for (i = 0; i < tctx->avctx->channels; i++) {
524 float val = (1.0 / (1 << 23)) *
525 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
528 for (j = 0; j < sub; j++)
530 val * mulawinv(sub_step * 0.5 +
531 sub_step * get_bits(gb, SUB_GAIN_BITS),
532 SUB_AMP_MAX, MULAW_MU);
538 * Rearrange the LSP coefficients so that they have a minimum distance of
539 * min_dist. This function does it exactly as described in section of 3.2.4
540 * of the G.729 specification (but interestingly is different from what the
541 * reference decoder actually does).
543 static void rearrange_lsp(int order, float *lsp, float min_dist)
546 float min_dist2 = min_dist * 0.5;
547 for (i = 1; i < order; i++)
548 if (lsp[i] - lsp[i - 1] < min_dist) {
549 float avg = (lsp[i] + lsp[i - 1]) * 0.5;
551 lsp[i - 1] = avg - min_dist2;
552 lsp[i] = avg + min_dist2;
556 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
557 int lpc_hist_idx, float *lsp, float *hist)
559 const ModeTab *mtab = tctx->mtab;
562 const float *cb = mtab->lspcodebook;
563 const float *cb2 = cb + (1 << mtab->lsp_bit1) * mtab->n_lsp;
564 const float *cb3 = cb2 + (1 << mtab->lsp_bit2) * mtab->n_lsp;
566 const int8_t funny_rounding[4] = {
568 mtab->lsp_split == 4 ? -2 : 1,
569 mtab->lsp_split == 4 ? -2 : 1,
574 for (i = 0; i < mtab->lsp_split; i++) {
575 int chunk_end = ((i + 1) * mtab->n_lsp + funny_rounding[i]) /
577 for (; j < chunk_end; j++)
578 lsp[j] = cb[lpc_idx1 * mtab->n_lsp + j] +
579 cb2[lpc_idx2[i] * mtab->n_lsp + j];
582 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
584 for (i = 0; i < mtab->n_lsp; i++) {
585 float tmp1 = 1.0 - cb3[lpc_hist_idx * mtab->n_lsp + i];
586 float tmp2 = hist[i] * cb3[lpc_hist_idx * mtab->n_lsp + i];
588 lsp[i] = lsp[i] * tmp1 + tmp2;
591 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
592 rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
593 ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
596 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
597 enum FrameType ftype, float *lpc)
600 int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
602 for (i = 0; i < tctx->mtab->n_lsp; i++)
603 lsp[i] = 2 * cos(lsp[i]);
607 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
610 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
613 eval_lpcenv(tctx, lsp, lpc);
618 static const uint8_t wtype_to_wsize[] = { 0, 0, 2, 2, 2, 1, 0, 1, 1 };
620 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
621 float *in, float *prev, int ch)
623 FFTContext *mdct = &tctx->mdct_ctx[ftype];
624 const ModeTab *mtab = tctx->mtab;
625 int bsize = mtab->size / mtab->fmode[ftype].sub;
626 int size = mtab->size;
627 float *buf1 = tctx->tmp_buf;
628 int j, first_wsize, wsize; // Window size
629 float *out = tctx->curr_frame + 2 * ch * mtab->size;
632 int types_sizes[] = {
633 mtab->size / mtab->fmode[FT_LONG].sub,
634 mtab->size / mtab->fmode[FT_MEDIUM].sub,
635 mtab->size / (mtab->fmode[FT_SHORT].sub * 2),
638 wsize = types_sizes[wtype_to_wsize[wtype]];
640 prev_buf = prev + (size - bsize) / 2;
642 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
643 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
645 if (!j && wtype == 4)
647 else if (j == mtab->fmode[ftype].sub - 1 && wtype == 7)
650 wsize = types_sizes[wtype_to_wsize[sub_wtype]];
652 mdct->imdct_half(mdct, buf1 + bsize * j, in + bsize * j);
654 tctx->fdsp.vector_fmul_window(out2, prev_buf + (bsize - wsize) / 2,
656 ff_sine_windows[av_log2(wsize)],
660 memcpy(out2, buf1 + bsize * j + wsize / 2,
661 (bsize - wsize / 2) * sizeof(float));
663 out2 += ftype == FT_MEDIUM ? (bsize - wsize) / 2 : bsize - wsize;
665 prev_buf = buf1 + bsize * j + bsize / 2;
668 tctx->last_block_pos[ch] = (size + first_wsize) / 2;
671 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
674 const ModeTab *mtab = tctx->mtab;
675 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
678 for (i = 0; i < tctx->avctx->channels; i++)
679 imdct_and_window(tctx, ftype, wtype,
680 tctx->spectrum + i * mtab->size,
681 prev_buf + 2 * i * mtab->size,
687 size2 = tctx->last_block_pos[0];
688 size1 = mtab->size - size2;
690 memcpy(&out[0][0], prev_buf, size1 * sizeof(out[0][0]));
691 memcpy(&out[0][size1], tctx->curr_frame, size2 * sizeof(out[0][0]));
693 if (tctx->avctx->channels == 2) {
694 memcpy(&out[1][0], &prev_buf[2 * mtab->size],
695 size1 * sizeof(out[1][0]));
696 memcpy(&out[1][size1], &tctx->curr_frame[2 * mtab->size],
697 size2 * sizeof(out[1][0]));
698 tctx->fdsp.butterflies_float(out[0], out[1], mtab->size);
702 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
703 int ch, float *out, float gain, enum FrameType ftype)
705 const ModeTab *mtab = tctx->mtab;
707 float *hist = tctx->bark_hist[ftype][ch];
708 float val = ((const float []) { 0.4, 0.35, 0.28 })[ftype];
709 int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
710 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
713 for (i = 0; i < fw_cb_len; i++)
714 for (j = 0; j < bark_n_coef; j++, idx++) {
715 float tmp2 = mtab->fmode[ftype].bark_cb[fw_cb_len * in[j] + i] *
717 float st = use_hist ? (1.0 - val) * tmp2 + val * hist[idx] + 1.0
724 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
725 out += mtab->fmode[ftype].bark_tab[idx];
729 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
730 float *out, enum FrameType ftype)
732 const ModeTab *mtab = tctx->mtab;
733 int channels = tctx->avctx->channels;
734 int sub = mtab->fmode[ftype].sub;
735 int block_size = mtab->size / sub;
736 float gain[CHANNELS_MAX * SUBBLOCKS_MAX];
737 float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
738 uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
739 uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
741 uint8_t lpc_idx1[CHANNELS_MAX];
742 uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
743 uint8_t lpc_hist_idx[CHANNELS_MAX];
747 dequant(tctx, gb, out, ftype,
748 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
749 mtab->fmode[ftype].cb_len_read);
751 for (i = 0; i < channels; i++)
752 for (j = 0; j < sub; j++)
753 for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
755 get_bits(gb, mtab->fmode[ftype].bark_n_bit);
757 for (i = 0; i < channels; i++)
758 for (j = 0; j < sub; j++)
759 bark_use_hist[i][j] = get_bits1(gb);
761 dec_gain(tctx, gb, ftype, gain);
763 for (i = 0; i < channels; i++) {
764 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
765 lpc_idx1[i] = get_bits(gb, tctx->mtab->lsp_bit1);
767 for (j = 0; j < tctx->mtab->lsp_split; j++)
768 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
771 if (ftype == FT_LONG) {
772 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len * channels - 1) /
774 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
775 mtab->ppc_shape_cb + cb_len_p * PPC_SHAPE_CB_SIZE, cb_len_p);
778 for (i = 0; i < channels; i++) {
779 float *chunk = out + mtab->size * i;
780 float lsp[LSP_COEFS_MAX];
782 for (j = 0; j < sub; j++) {
783 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
784 tctx->tmp_buf, gain[sub * i + j], ftype);
786 tctx->fdsp.vector_fmul(chunk + block_size * j,
787 chunk + block_size * j,
788 tctx->tmp_buf, block_size);
791 if (ftype == FT_LONG) {
792 float pgain_step = 25000.0 / ((1 << mtab->pgain_bit) - 1);
793 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
794 int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
795 float v = 1.0 / 8192 *
796 mulawinv(pgain_step * g_coef + pgain_step / 2,
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->fdsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
815 static int twin_decode_frame(AVCodecContext *avctx, void *data,
816 int *got_frame_ptr, AVPacket *avpkt)
818 AVFrame *frame = data;
819 const uint8_t *buf = avpkt->data;
820 int buf_size = avpkt->size;
821 TwinContext *tctx = avctx->priv_data;
823 const ModeTab *mtab = tctx->mtab;
825 enum FrameType ftype;
826 int window_type, ret;
827 static const enum FrameType wtype_to_ftype_table[] = {
828 FT_LONG, FT_LONG, FT_SHORT, FT_LONG,
829 FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM
832 if (buf_size * 8 < avctx->bit_rate * mtab->size / avctx->sample_rate + 8) {
833 av_log(avctx, AV_LOG_ERROR,
834 "Frame too small (%d bytes). Truncated file?\n", buf_size);
835 return AVERROR(EINVAL);
838 /* get output buffer */
839 if (tctx->discarded_packets >= 2) {
840 frame->nb_samples = mtab->size;
841 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
842 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
845 out = (float **)frame->extended_data;
848 init_get_bits(&gb, buf, buf_size * 8);
849 skip_bits(&gb, get_bits(&gb, 8));
850 window_type = get_bits(&gb, WINDOW_TYPE_BITS);
852 if (window_type > 8) {
853 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
857 ftype = wtype_to_ftype_table[window_type];
859 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
861 imdct_output(tctx, ftype, window_type, out);
863 FFSWAP(float *, tctx->curr_frame, tctx->prev_frame);
865 if (tctx->discarded_packets < 2) {
866 tctx->discarded_packets++;
877 * Init IMDCT and windowing tables
879 static av_cold int init_mdct_win(TwinContext *tctx)
882 const ModeTab *mtab = tctx->mtab;
883 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
884 int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
885 int channels = tctx->avctx->channels;
886 float norm = channels == 1 ? 2.0 : 1.0;
888 for (i = 0; i < 3; i++) {
889 int bsize = tctx->mtab->size / tctx->mtab->fmode[i].sub;
890 if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
891 -sqrt(norm / bsize) / (1 << 15))))
895 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
896 mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
898 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
899 2 * mtab->size * channels * sizeof(*tctx->spectrum),
901 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
902 2 * mtab->size * channels * sizeof(*tctx->curr_frame),
904 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
905 2 * mtab->size * channels * sizeof(*tctx->prev_frame),
908 for (i = 0; i < 3; i++) {
909 int m = 4 * mtab->size / mtab->fmode[i].sub;
910 double freq = 2 * M_PI / m;
911 FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
912 (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
914 for (j = 0; j <= m / 8; j++)
915 tctx->cos_tabs[i][j] = cos((2 * j + 1) * freq);
916 for (j = 1; j < m / 8; j++)
917 tctx->cos_tabs[i][m / 4 - j] = tctx->cos_tabs[i][j];
920 ff_init_ff_sine_windows(av_log2(size_m));
921 ff_init_ff_sine_windows(av_log2(size_s / 2));
922 ff_init_ff_sine_windows(av_log2(mtab->size));
927 return AVERROR(ENOMEM);
931 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
932 * each line do a cyclic permutation, i.e.
933 * abcdefghijklm -> defghijklmabc
934 * where the amount to be shifted is evaluated depending on the column.
936 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
938 const uint8_t line_len[2], int length_div,
939 enum FrameType ftype)
943 for (i = 0; i < line_len[0]; i++) {
946 if (num_blocks == 1 ||
947 (ftype == FT_LONG && num_vect % num_blocks) ||
948 (ftype != FT_LONG && num_vect & 1) ||
951 } else if (ftype == FT_LONG) {
956 for (j = 0; j < num_vect && (j + num_vect * i < block_size * num_blocks); j++)
957 tab[i * num_vect + j] = i * num_vect + (j + shift) % num_vect;
962 * Interpret the input data as in the following table:
973 * and transpose it, giving the output
974 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
976 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
977 const uint8_t line_len[2], int length_div)
982 for (i = 0; i < num_vect; i++)
983 for (j = 0; j < line_len[i >= length_div]; j++)
984 out[cont++] = in[j * num_vect + i];
987 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
989 int block_size = size / n_blocks;
992 for (i = 0; i < size; i++)
993 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
996 static av_cold void construct_perm_table(TwinContext *tctx,
997 enum FrameType ftype)
999 int block_size, size;
1000 const ModeTab *mtab = tctx->mtab;
1001 int16_t *tmp_perm = (int16_t *)tctx->tmp_buf;
1003 if (ftype == FT_PPC) {
1004 size = tctx->avctx->channels;
1005 block_size = mtab->ppc_shape_len;
1007 size = tctx->avctx->channels * mtab->fmode[ftype].sub;
1008 block_size = mtab->size / mtab->fmode[ftype].sub;
1011 permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
1012 block_size, tctx->length[ftype],
1013 tctx->length_change[ftype], ftype);
1015 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
1016 tctx->length[ftype], tctx->length_change[ftype]);
1018 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
1022 static av_cold void init_bitstream_params(TwinContext *tctx)
1024 const ModeTab *mtab = tctx->mtab;
1025 int n_ch = tctx->avctx->channels;
1026 int total_fr_bits = tctx->avctx->bit_rate * mtab->size /
1027 tctx->avctx->sample_rate;
1029 int lsp_bits_per_block = n_ch * (mtab->lsp_bit0 + mtab->lsp_bit1 +
1030 mtab->lsp_split * mtab->lsp_bit2);
1032 int ppc_bits = n_ch * (mtab->pgain_bit + mtab->ppc_shape_bit +
1033 mtab->ppc_period_bit);
1035 int bsize_no_main_cb[3], bse_bits[3], i;
1036 enum FrameType frametype;
1038 for (i = 0; i < 3; i++)
1039 // +1 for history usage switch
1040 bse_bits[i] = n_ch *
1041 (mtab->fmode[i].bark_n_coef *
1042 mtab->fmode[i].bark_n_bit + 1);
1044 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1045 WINDOW_TYPE_BITS + n_ch * GAIN_BITS;
1047 for (i = 0; i < 2; i++)
1048 bsize_no_main_cb[i] =
1049 lsp_bits_per_block + n_ch * GAIN_BITS + WINDOW_TYPE_BITS +
1050 mtab->fmode[i].sub * (bse_bits[i] + n_ch * SUB_GAIN_BITS);
1052 // The remaining bits are all used for the main spectrum coefficients
1053 for (i = 0; i < 4; i++) {
1054 int bit_size, vect_size;
1055 int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1057 bit_size = n_ch * mtab->ppc_shape_bit;
1058 vect_size = n_ch * mtab->ppc_shape_len;
1060 bit_size = total_fr_bits - bsize_no_main_cb[i];
1061 vect_size = n_ch * mtab->size;
1064 tctx->n_div[i] = (bit_size + 13) / 14;
1066 rounded_up = (bit_size + tctx->n_div[i] - 1) /
1068 rounded_down = (bit_size) / tctx->n_div[i];
1069 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1070 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1071 tctx->bits_main_spec[0][i][0] = (rounded_up + 1) / 2;
1072 tctx->bits_main_spec[1][i][0] = rounded_up / 2;
1073 tctx->bits_main_spec[0][i][1] = (rounded_down + 1) / 2;
1074 tctx->bits_main_spec[1][i][1] = rounded_down / 2;
1075 tctx->bits_main_spec_change[i] = num_rounded_up;
1077 rounded_up = (vect_size + tctx->n_div[i] - 1) /
1079 rounded_down = (vect_size) / tctx->n_div[i];
1080 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1081 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1082 tctx->length[i][0] = rounded_up;
1083 tctx->length[i][1] = rounded_down;
1084 tctx->length_change[i] = num_rounded_up;
1087 for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
1088 construct_perm_table(tctx, frametype);
1091 static av_cold int twin_decode_close(AVCodecContext *avctx)
1093 TwinContext *tctx = avctx->priv_data;
1096 for (i = 0; i < 3; i++) {
1097 ff_mdct_end(&tctx->mdct_ctx[i]);
1098 av_free(tctx->cos_tabs[i]);
1101 av_free(tctx->curr_frame);
1102 av_free(tctx->spectrum);
1103 av_free(tctx->prev_frame);
1104 av_free(tctx->tmp_buf);
1109 static av_cold int twin_decode_init(AVCodecContext *avctx)
1111 int ret, isampf, ibps;
1112 TwinContext *tctx = avctx->priv_data;
1114 tctx->avctx = avctx;
1115 avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
1117 if (!avctx->extradata || avctx->extradata_size < 12) {
1118 av_log(avctx, AV_LOG_ERROR, "Missing or incomplete extradata\n");
1119 return AVERROR_INVALIDDATA;
1121 avctx->channels = AV_RB32(avctx->extradata) + 1;
1122 avctx->bit_rate = AV_RB32(avctx->extradata + 4) * 1000;
1123 isampf = AV_RB32(avctx->extradata + 8);
1125 if (isampf < 8 || isampf > 44) {
1126 av_log(avctx, AV_LOG_ERROR, "Unsupported sample rate\n");
1127 return AVERROR_INVALIDDATA;
1131 avctx->sample_rate = 44100;
1134 avctx->sample_rate = 22050;
1137 avctx->sample_rate = 11025;
1140 avctx->sample_rate = isampf * 1000;
1144 if (avctx->channels <= 0 || avctx->channels > CHANNELS_MAX) {
1145 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1149 avctx->channel_layout = avctx->channels == 1 ? AV_CH_LAYOUT_MONO
1150 : AV_CH_LAYOUT_STEREO;
1152 ibps = avctx->bit_rate / (1000 * avctx->channels);
1154 switch ((isampf << 8) + ibps) {
1156 tctx->mtab = &mode_08_08;
1159 tctx->mtab = &mode_11_08;
1161 case (11 << 8) + 10:
1162 tctx->mtab = &mode_11_10;
1164 case (16 << 8) + 16:
1165 tctx->mtab = &mode_16_16;
1167 case (22 << 8) + 20:
1168 tctx->mtab = &mode_22_20;
1170 case (22 << 8) + 24:
1171 tctx->mtab = &mode_22_24;
1173 case (22 << 8) + 32:
1174 tctx->mtab = &mode_22_32;
1176 case (44 << 8) + 40:
1177 tctx->mtab = &mode_44_40;
1179 case (44 << 8) + 48:
1180 tctx->mtab = &mode_44_48;
1183 av_log(avctx, AV_LOG_ERROR,
1184 "This version does not support %d kHz - %d kbit/s/ch mode.\n",
1189 avpriv_float_dsp_init(&tctx->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
1190 if ((ret = init_mdct_win(tctx))) {
1191 av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
1192 twin_decode_close(avctx);
1195 init_bitstream_params(tctx);
1197 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1202 AVCodec ff_twinvq_decoder = {
1204 .type = AVMEDIA_TYPE_AUDIO,
1205 .id = AV_CODEC_ID_TWINVQ,
1206 .priv_data_size = sizeof(TwinContext),
1207 .init = twin_decode_init,
1208 .close = twin_decode_close,
1209 .decode = twin_decode_frame,
1210 .capabilities = CODEC_CAP_DR1,
1211 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
1212 .sample_fmts = (const enum AVSampleFormat[]) { AV_SAMPLE_FMT_FLTP,
1213 AV_SAMPLE_FMT_NONE },