xref: /third_party/ffmpeg/libavcodec/twinvq.c (revision cabdff1a)
1/*
2 * TwinVQ decoder
3 * Copyright (c) 2009 Vitor Sessak
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22#include <math.h>
23#include <stdint.h>
24
25#include "libavutil/channel_layout.h"
26#include "libavutil/float_dsp.h"
27#include "avcodec.h"
28#include "fft.h"
29#include "internal.h"
30#include "lsp.h"
31#include "sinewin.h"
32#include "twinvq.h"
33
34/**
35 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
36 * spectrum pairs.
37 *
38 * @param lsp a vector of the cosine of the LSP values
39 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
40 * @param order the order of the LSP (and the size of the *lsp buffer). Must
41 *        be a multiple of four.
42 * @return the LPC value
43 *
44 * @todo reuse code from Vorbis decoder: vorbis_floor0_decode
45 */
46static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
47{
48    int j;
49    float p         = 0.5f;
50    float q         = 0.5f;
51    float two_cos_w = 2.0f * cos_val;
52
53    for (j = 0; j + 1 < order; j += 2 * 2) {
54        // Unroll the loop once since order is a multiple of four
55        q *= lsp[j]     - two_cos_w;
56        p *= lsp[j + 1] - two_cos_w;
57
58        q *= lsp[j + 2] - two_cos_w;
59        p *= lsp[j + 3] - two_cos_w;
60    }
61
62    p *= p * (2.0f - two_cos_w);
63    q *= q * (2.0f + two_cos_w);
64
65    return 0.5 / (p + q);
66}
67
68/**
69 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
70 */
71static void eval_lpcenv(TwinVQContext *tctx, const float *cos_vals, float *lpc)
72{
73    int i;
74    const TwinVQModeTab *mtab = tctx->mtab;
75    int size_s = mtab->size / mtab->fmode[TWINVQ_FT_SHORT].sub;
76
77    for (i = 0; i < size_s / 2; i++) {
78        float cos_i = tctx->cos_tabs[0][i];
79        lpc[i]              = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
80        lpc[size_s - i - 1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
81    }
82}
83
84static void interpolate(float *out, float v1, float v2, int size)
85{
86    int i;
87    float step = (v1 - v2) / (size + 1);
88
89    for (i = 0; i < size; i++) {
90        v2    += step;
91        out[i] = v2;
92    }
93}
94
95static inline float get_cos(int idx, int part, const float *cos_tab, int size)
96{
97    return part ? -cos_tab[size - idx - 1]
98                :  cos_tab[idx];
99}
100
101/**
102 * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
103 * Probably for speed reasons, the coefficients are evaluated as
104 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
105 * where s is an evaluated value, i is a value interpolated from the others
106 * and b might be either calculated or interpolated, depending on an
107 * unexplained condition.
108 *
109 * @param step the size of a block "siiiibiiii"
110 * @param in the cosine of the LSP data
111 * @param part is 0 for 0...PI (positive cosine values) and 1 for PI...2PI
112 *        (negative cosine values)
113 * @param size the size of the whole output
114 */
115static inline void eval_lpcenv_or_interp(TwinVQContext *tctx,
116                                         enum TwinVQFrameType ftype,
117                                         float *out, const float *in,
118                                         int size, int step, int part)
119{
120    int i;
121    const TwinVQModeTab *mtab = tctx->mtab;
122    const float *cos_tab      = tctx->cos_tabs[ftype];
123
124    // Fill the 's'
125    for (i = 0; i < size; i += step)
126        out[i] =
127            eval_lpc_spectrum(in,
128                              get_cos(i, part, cos_tab, size),
129                              mtab->n_lsp);
130
131    // Fill the 'iiiibiiii'
132    for (i = step; i <= size - 2 * step; i += step) {
133        if (out[i + step] + out[i - step] > 1.95 * out[i] ||
134            out[i + step]                 >= out[i - step]) {
135            interpolate(out + i - step + 1, out[i], out[i - step], step - 1);
136        } else {
137            out[i - step / 2] =
138                eval_lpc_spectrum(in,
139                                  get_cos(i - step / 2, part, cos_tab, size),
140                                  mtab->n_lsp);
141            interpolate(out + i - step + 1, out[i - step / 2],
142                        out[i - step], step / 2 - 1);
143            interpolate(out + i - step / 2 + 1, out[i],
144                        out[i - step / 2], step / 2 - 1);
145        }
146    }
147
148    interpolate(out + size - 2 * step + 1, out[size - step],
149                out[size - 2 * step], step - 1);
150}
151
152static void eval_lpcenv_2parts(TwinVQContext *tctx, enum TwinVQFrameType ftype,
153                               const float *buf, float *lpc,
154                               int size, int step)
155{
156    eval_lpcenv_or_interp(tctx, ftype, lpc, buf, size / 2, step, 0);
157    eval_lpcenv_or_interp(tctx, ftype, lpc + size / 2, buf, size / 2,
158                          2 * step, 1);
159
160    interpolate(lpc + size / 2 - step + 1, lpc[size / 2],
161                lpc[size / 2 - step], step);
162
163    twinvq_memset_float(lpc + size - 2 * step + 1, lpc[size - 2 * step],
164                        2 * step - 1);
165}
166
167/**
168 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
169 * bitstream, sum the corresponding vectors and write the result to *out
170 * after permutation.
171 */
172static void dequant(TwinVQContext *tctx, const uint8_t *cb_bits, float *out,
173                    enum TwinVQFrameType ftype,
174                    const int16_t *cb0, const int16_t *cb1, int cb_len)
175{
176    int pos = 0;
177    int i, j;
178
179    for (i = 0; i < tctx->n_div[ftype]; i++) {
180        int tmp0, tmp1;
181        int sign0 = 1;
182        int sign1 = 1;
183        const int16_t *tab0, *tab1;
184        int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
185        int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
186
187        int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
188        tmp0 = *cb_bits++;
189        if (bits == 7) {
190            if (tmp0 & 0x40)
191                sign0 = -1;
192            tmp0 &= 0x3F;
193        }
194
195        bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
196        tmp1 = *cb_bits++;
197        if (bits == 7) {
198            if (tmp1 & 0x40)
199                sign1 = -1;
200            tmp1 &= 0x3F;
201        }
202
203        tab0 = cb0 + tmp0 * cb_len;
204        tab1 = cb1 + tmp1 * cb_len;
205
206        for (j = 0; j < length; j++)
207            out[tctx->permut[ftype][pos + j]] = sign0 * tab0[j] +
208                                                sign1 * tab1[j];
209
210        pos += length;
211    }
212}
213
214static void dec_gain(TwinVQContext *tctx,
215                     enum TwinVQFrameType ftype, float *out)
216{
217    const TwinVQModeTab   *mtab =  tctx->mtab;
218    const TwinVQFrameData *bits = &tctx->bits[tctx->cur_frame];
219    int i, j;
220    int channels   = tctx->avctx->ch_layout.nb_channels;
221    int sub        = mtab->fmode[ftype].sub;
222    float step     = TWINVQ_AMP_MAX     / ((1 << TWINVQ_GAIN_BITS)     - 1);
223    float sub_step = TWINVQ_SUB_AMP_MAX / ((1 << TWINVQ_SUB_GAIN_BITS) - 1);
224
225    if (ftype == TWINVQ_FT_LONG) {
226        for (i = 0; i < channels; i++)
227            out[i] = (1.0 / (1 << 13)) *
228                     twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
229                                     TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
230    } else {
231        for (i = 0; i < channels; i++) {
232            float val = (1.0 / (1 << 23)) *
233                        twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
234                                        TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
235
236            for (j = 0; j < sub; j++)
237                out[i * sub + j] =
238                    val * twinvq_mulawinv(sub_step * 0.5 +
239                                          sub_step * bits->sub_gain_bits[i * sub + j],
240                                          TWINVQ_SUB_AMP_MAX, TWINVQ_MULAW_MU);
241        }
242    }
243}
244
245/**
246 * Rearrange the LSP coefficients so that they have a minimum distance of
247 * min_dist. This function does it exactly as described in section of 3.2.4
248 * of the G.729 specification (but interestingly is different from what the
249 * reference decoder actually does).
250 */
251static void rearrange_lsp(int order, float *lsp, float min_dist)
252{
253    int i;
254    float min_dist2 = min_dist * 0.5;
255    for (i = 1; i < order; i++)
256        if (lsp[i] - lsp[i - 1] < min_dist) {
257            float avg = (lsp[i] + lsp[i - 1]) * 0.5;
258
259            lsp[i - 1] = avg - min_dist2;
260            lsp[i]     = avg + min_dist2;
261        }
262}
263
264static void decode_lsp(TwinVQContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
265                       int lpc_hist_idx, float *lsp, float *hist)
266{
267    const TwinVQModeTab *mtab = tctx->mtab;
268    int i, j;
269
270    const float *cb  = mtab->lspcodebook;
271    const float *cb2 = cb  + (1 << mtab->lsp_bit1) * mtab->n_lsp;
272    const float *cb3 = cb2 + (1 << mtab->lsp_bit2) * mtab->n_lsp;
273
274    const int8_t funny_rounding[4] = {
275        -2,
276        mtab->lsp_split == 4 ? -2 : 1,
277        mtab->lsp_split == 4 ? -2 : 1,
278        0
279    };
280
281    j = 0;
282    for (i = 0; i < mtab->lsp_split; i++) {
283        int chunk_end = ((i + 1) * mtab->n_lsp + funny_rounding[i]) /
284                        mtab->lsp_split;
285        for (; j < chunk_end; j++)
286            lsp[j] = cb[lpc_idx1     * mtab->n_lsp + j] +
287                     cb2[lpc_idx2[i] * mtab->n_lsp + j];
288    }
289
290    rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
291
292    for (i = 0; i < mtab->n_lsp; i++) {
293        float tmp1 = 1.0     - cb3[lpc_hist_idx * mtab->n_lsp + i];
294        float tmp2 = hist[i] * cb3[lpc_hist_idx * mtab->n_lsp + i];
295        hist[i] = lsp[i];
296        lsp[i]  = lsp[i] * tmp1 + tmp2;
297    }
298
299    rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
300    rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
301    ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
302}
303
304static void dec_lpc_spectrum_inv(TwinVQContext *tctx, float *lsp,
305                                 enum TwinVQFrameType ftype, float *lpc)
306{
307    int i;
308    int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
309
310    for (i = 0; i < tctx->mtab->n_lsp; i++)
311        lsp[i] = 2 * cos(lsp[i]);
312
313    switch (ftype) {
314    case TWINVQ_FT_LONG:
315        eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
316        break;
317    case TWINVQ_FT_MEDIUM:
318        eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
319        break;
320    case TWINVQ_FT_SHORT:
321        eval_lpcenv(tctx, lsp, lpc);
322        break;
323    }
324}
325
326static const uint8_t wtype_to_wsize[] = { 0, 0, 2, 2, 2, 1, 0, 1, 1 };
327
328static void imdct_and_window(TwinVQContext *tctx, enum TwinVQFrameType ftype,
329                             int wtype, float *in, float *prev, int ch)
330{
331    FFTContext *mdct = &tctx->mdct_ctx[ftype];
332    const TwinVQModeTab *mtab = tctx->mtab;
333    int bsize = mtab->size / mtab->fmode[ftype].sub;
334    int size  = mtab->size;
335    float *buf1 = tctx->tmp_buf;
336    int j, first_wsize, wsize; // Window size
337    float *out  = tctx->curr_frame + 2 * ch * mtab->size;
338    float *out2 = out;
339    float *prev_buf;
340    int types_sizes[] = {
341        mtab->size /  mtab->fmode[TWINVQ_FT_LONG].sub,
342        mtab->size /  mtab->fmode[TWINVQ_FT_MEDIUM].sub,
343        mtab->size / (mtab->fmode[TWINVQ_FT_SHORT].sub * 2),
344    };
345
346    wsize       = types_sizes[wtype_to_wsize[wtype]];
347    first_wsize = wsize;
348    prev_buf    = prev + (size - bsize) / 2;
349
350    for (j = 0; j < mtab->fmode[ftype].sub; j++) {
351        int sub_wtype = ftype == TWINVQ_FT_MEDIUM ? 8 : wtype;
352
353        if (!j && wtype == 4)
354            sub_wtype = 4;
355        else if (j == mtab->fmode[ftype].sub - 1 && wtype == 7)
356            sub_wtype = 7;
357
358        wsize = types_sizes[wtype_to_wsize[sub_wtype]];
359
360        mdct->imdct_half(mdct, buf1 + bsize * j, in + bsize * j);
361
362        tctx->fdsp->vector_fmul_window(out2, prev_buf + (bsize - wsize) / 2,
363                                      buf1 + bsize * j,
364                                      ff_sine_windows[av_log2(wsize)],
365                                      wsize / 2);
366        out2 += wsize;
367
368        memcpy(out2, buf1 + bsize * j + wsize / 2,
369               (bsize - wsize / 2) * sizeof(float));
370
371        out2 += ftype == TWINVQ_FT_MEDIUM ? (bsize - wsize) / 2 : bsize - wsize;
372
373        prev_buf = buf1 + bsize * j + bsize / 2;
374    }
375
376    tctx->last_block_pos[ch] = (size + first_wsize) / 2;
377}
378
379static void imdct_output(TwinVQContext *tctx, enum TwinVQFrameType ftype,
380                         int wtype, float **out, int offset)
381{
382    const TwinVQModeTab *mtab = tctx->mtab;
383    float *prev_buf           = tctx->prev_frame + tctx->last_block_pos[0];
384    int channels              = tctx->avctx->ch_layout.nb_channels;
385    int size1, size2, i;
386    float *out1, *out2;
387
388    for (i = 0; i < channels; i++)
389        imdct_and_window(tctx, ftype, wtype,
390                         tctx->spectrum + i * mtab->size,
391                         prev_buf + 2 * i * mtab->size,
392                         i);
393
394    if (!out)
395        return;
396
397    size2 = tctx->last_block_pos[0];
398    size1 = mtab->size - size2;
399
400    out1 = &out[0][0] + offset;
401    memcpy(out1,         prev_buf,         size1 * sizeof(*out1));
402    memcpy(out1 + size1, tctx->curr_frame, size2 * sizeof(*out1));
403
404    if (channels == 2) {
405        out2 = &out[1][0] + offset;
406        memcpy(out2, &prev_buf[2 * mtab->size],
407               size1 * sizeof(*out2));
408        memcpy(out2 + size1, &tctx->curr_frame[2 * mtab->size],
409               size2 * sizeof(*out2));
410        tctx->fdsp->butterflies_float(out1, out2, mtab->size);
411    }
412}
413
414static void read_and_decode_spectrum(TwinVQContext *tctx, float *out,
415                                     enum TwinVQFrameType ftype)
416{
417    const TwinVQModeTab *mtab = tctx->mtab;
418    TwinVQFrameData *bits     = &tctx->bits[tctx->cur_frame];
419    int channels              = tctx->avctx->ch_layout.nb_channels;
420    int sub        = mtab->fmode[ftype].sub;
421    int block_size = mtab->size / sub;
422    float gain[TWINVQ_CHANNELS_MAX * TWINVQ_SUBBLOCKS_MAX];
423    float ppc_shape[TWINVQ_PPC_SHAPE_LEN_MAX * TWINVQ_CHANNELS_MAX * 4];
424
425    int i, j;
426
427    dequant(tctx, bits->main_coeffs, out, ftype,
428            mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
429            mtab->fmode[ftype].cb_len_read);
430
431    dec_gain(tctx, ftype, gain);
432
433    if (ftype == TWINVQ_FT_LONG) {
434        int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len * channels - 1) /
435                       tctx->n_div[3];
436        dequant(tctx, bits->ppc_coeffs, ppc_shape,
437                TWINVQ_FT_PPC, mtab->ppc_shape_cb,
438                mtab->ppc_shape_cb + cb_len_p * TWINVQ_PPC_SHAPE_CB_SIZE,
439                cb_len_p);
440    }
441
442    for (i = 0; i < channels; i++) {
443        float *chunk = out + mtab->size * i;
444        float lsp[TWINVQ_LSP_COEFS_MAX];
445
446        for (j = 0; j < sub; j++) {
447            tctx->dec_bark_env(tctx, bits->bark1[i][j],
448                               bits->bark_use_hist[i][j], i,
449                               tctx->tmp_buf, gain[sub * i + j], ftype);
450
451            tctx->fdsp->vector_fmul(chunk + block_size * j,
452                                   chunk + block_size * j,
453                                   tctx->tmp_buf, block_size);
454        }
455
456        if (ftype == TWINVQ_FT_LONG)
457            tctx->decode_ppc(tctx, bits->p_coef[i], bits->g_coef[i],
458                             ppc_shape + i * mtab->ppc_shape_len, chunk);
459
460        decode_lsp(tctx, bits->lpc_idx1[i], bits->lpc_idx2[i],
461                   bits->lpc_hist_idx[i], lsp, tctx->lsp_hist[i]);
462
463        dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
464
465        for (j = 0; j < mtab->fmode[ftype].sub; j++) {
466            tctx->fdsp->vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
467            chunk += block_size;
468        }
469    }
470}
471
472const enum TwinVQFrameType ff_twinvq_wtype_to_ftype_table[] = {
473    TWINVQ_FT_LONG,   TWINVQ_FT_LONG, TWINVQ_FT_SHORT, TWINVQ_FT_LONG,
474    TWINVQ_FT_MEDIUM, TWINVQ_FT_LONG, TWINVQ_FT_LONG,  TWINVQ_FT_MEDIUM,
475    TWINVQ_FT_MEDIUM
476};
477
478int ff_twinvq_decode_frame(AVCodecContext *avctx, AVFrame *frame,
479                           int *got_frame_ptr, AVPacket *avpkt)
480{
481    const uint8_t *buf = avpkt->data;
482    int buf_size       = avpkt->size;
483    TwinVQContext *tctx = avctx->priv_data;
484    const TwinVQModeTab *mtab = tctx->mtab;
485    float **out = NULL;
486    int ret;
487
488    /* get output buffer */
489    if (tctx->discarded_packets >= 2) {
490        frame->nb_samples = mtab->size * tctx->frames_per_packet;
491        if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
492            return ret;
493        out = (float **)frame->extended_data;
494    }
495
496    if (buf_size < avctx->block_align) {
497        av_log(avctx, AV_LOG_ERROR,
498               "Frame too small (%d bytes). Truncated file?\n", buf_size);
499        return AVERROR(EINVAL);
500    }
501
502    if ((ret = tctx->read_bitstream(avctx, tctx, buf, buf_size)) < 0)
503        return ret;
504
505    for (tctx->cur_frame = 0; tctx->cur_frame < tctx->frames_per_packet;
506         tctx->cur_frame++) {
507        read_and_decode_spectrum(tctx, tctx->spectrum,
508                                 tctx->bits[tctx->cur_frame].ftype);
509
510        imdct_output(tctx, tctx->bits[tctx->cur_frame].ftype,
511                     tctx->bits[tctx->cur_frame].window_type, out,
512                     tctx->cur_frame * mtab->size);
513
514        FFSWAP(float *, tctx->curr_frame, tctx->prev_frame);
515    }
516
517    if (tctx->discarded_packets < 2) {
518        tctx->discarded_packets++;
519        *got_frame_ptr = 0;
520        return buf_size;
521    }
522
523    *got_frame_ptr = 1;
524
525    // VQF can deliver packets 1 byte greater than block align
526    if (buf_size == avctx->block_align + 1)
527        return buf_size;
528    return avctx->block_align;
529}
530
531/**
532 * Init IMDCT and windowing tables
533 */
534static av_cold int init_mdct_win(TwinVQContext *tctx)
535{
536    int i, j, ret;
537    const TwinVQModeTab *mtab = tctx->mtab;
538    int size_s = mtab->size / mtab->fmode[TWINVQ_FT_SHORT].sub;
539    int size_m = mtab->size / mtab->fmode[TWINVQ_FT_MEDIUM].sub;
540    int channels = tctx->avctx->ch_layout.nb_channels;
541    float norm = channels == 1 ? 2.0 : 1.0;
542    int table_size = 2 * mtab->size * channels;
543
544    for (i = 0; i < 3; i++) {
545        int bsize = tctx->mtab->size / tctx->mtab->fmode[i].sub;
546        if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
547                                -sqrt(norm / bsize) / (1 << 15))))
548            return ret;
549    }
550
551    if (!FF_ALLOC_TYPED_ARRAY(tctx->tmp_buf,    mtab->size) ||
552        !FF_ALLOC_TYPED_ARRAY(tctx->spectrum,   table_size) ||
553        !FF_ALLOC_TYPED_ARRAY(tctx->curr_frame, table_size) ||
554        !FF_ALLOC_TYPED_ARRAY(tctx->prev_frame, table_size))
555        return AVERROR(ENOMEM);
556
557    for (i = 0; i < 3; i++) {
558        int m       = 4 * mtab->size / mtab->fmode[i].sub;
559        double freq = 2 * M_PI / m;
560        if (!FF_ALLOC_TYPED_ARRAY(tctx->cos_tabs[i], m / 4))
561            return AVERROR(ENOMEM);
562        for (j = 0; j <= m / 8; j++)
563            tctx->cos_tabs[i][j] = cos((2 * j + 1) * freq);
564        for (j = 1; j < m / 8; j++)
565            tctx->cos_tabs[i][m / 4 - j] = tctx->cos_tabs[i][j];
566    }
567
568    ff_init_ff_sine_windows(av_log2(size_m));
569    ff_init_ff_sine_windows(av_log2(size_s / 2));
570    ff_init_ff_sine_windows(av_log2(mtab->size));
571
572    return 0;
573}
574
575/**
576 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
577 * each line do a cyclic permutation, i.e.
578 * abcdefghijklm -> defghijklmabc
579 * where the amount to be shifted is evaluated depending on the column.
580 */
581static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
582                              int block_size,
583                              const uint8_t line_len[2], int length_div,
584                              enum TwinVQFrameType ftype)
585{
586    int i, j;
587
588    for (i = 0; i < line_len[0]; i++) {
589        int shift;
590
591        if (num_blocks == 1                                    ||
592            (ftype == TWINVQ_FT_LONG && num_vect % num_blocks) ||
593            (ftype != TWINVQ_FT_LONG && num_vect & 1)          ||
594            i == line_len[1]) {
595            shift = 0;
596        } else if (ftype == TWINVQ_FT_LONG) {
597            shift = i;
598        } else
599            shift = i * i;
600
601        for (j = 0; j < num_vect && (j + num_vect * i < block_size * num_blocks); j++)
602            tab[i * num_vect + j] = i * num_vect + (j + shift) % num_vect;
603    }
604}
605
606/**
607 * Interpret the input data as in the following table:
608 *
609 * @verbatim
610 *
611 * abcdefgh
612 * ijklmnop
613 * qrstuvw
614 * x123456
615 *
616 * @endverbatim
617 *
618 * and transpose it, giving the output
619 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
620 */
621static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
622                           const uint8_t line_len[2], int length_div)
623{
624    int i, j;
625    int cont = 0;
626
627    for (i = 0; i < num_vect; i++)
628        for (j = 0; j < line_len[i >= length_div]; j++)
629            out[cont++] = in[j * num_vect + i];
630}
631
632static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
633{
634    int block_size = size / n_blocks;
635    int i;
636
637    for (i = 0; i < size; i++)
638        out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
639}
640
641static av_cold void construct_perm_table(TwinVQContext *tctx,
642                                         enum TwinVQFrameType ftype)
643{
644    int block_size, size;
645    const TwinVQModeTab *mtab = tctx->mtab;
646    int16_t *tmp_perm = (int16_t *)tctx->tmp_buf;
647
648    if (ftype == TWINVQ_FT_PPC) {
649        size       = tctx->avctx->ch_layout.nb_channels;
650        block_size = mtab->ppc_shape_len;
651    } else {
652        size       = tctx->avctx->ch_layout.nb_channels * mtab->fmode[ftype].sub;
653        block_size = mtab->size / mtab->fmode[ftype].sub;
654    }
655
656    permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
657                      block_size, tctx->length[ftype],
658                      tctx->length_change[ftype], ftype);
659
660    transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
661                   tctx->length[ftype], tctx->length_change[ftype]);
662
663    linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
664                size * block_size);
665}
666
667static av_cold void init_bitstream_params(TwinVQContext *tctx)
668{
669    const TwinVQModeTab *mtab = tctx->mtab;
670    int n_ch                  = tctx->avctx->ch_layout.nb_channels;
671    int total_fr_bits         = tctx->avctx->bit_rate * mtab->size /
672                                tctx->avctx->sample_rate;
673
674    int lsp_bits_per_block = n_ch * (mtab->lsp_bit0 + mtab->lsp_bit1 +
675                                     mtab->lsp_split * mtab->lsp_bit2);
676
677    int ppc_bits = n_ch * (mtab->pgain_bit + mtab->ppc_shape_bit +
678                           mtab->ppc_period_bit);
679
680    int bsize_no_main_cb[3], bse_bits[3], i;
681    enum TwinVQFrameType frametype;
682
683    for (i = 0; i < 3; i++)
684        // +1 for history usage switch
685        bse_bits[i] = n_ch *
686                      (mtab->fmode[i].bark_n_coef *
687                       mtab->fmode[i].bark_n_bit + 1);
688
689    bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
690                          TWINVQ_WINDOW_TYPE_BITS + n_ch * TWINVQ_GAIN_BITS;
691
692    for (i = 0; i < 2; i++)
693        bsize_no_main_cb[i] =
694            lsp_bits_per_block + n_ch * TWINVQ_GAIN_BITS +
695            TWINVQ_WINDOW_TYPE_BITS +
696            mtab->fmode[i].sub * (bse_bits[i] + n_ch * TWINVQ_SUB_GAIN_BITS);
697
698    if (tctx->codec == TWINVQ_CODEC_METASOUND && !tctx->is_6kbps) {
699        bsize_no_main_cb[1] += 2;
700        bsize_no_main_cb[2] += 2;
701    }
702
703    // The remaining bits are all used for the main spectrum coefficients
704    for (i = 0; i < 4; i++) {
705        int bit_size, vect_size;
706        int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
707        if (i == 3) {
708            bit_size  = n_ch * mtab->ppc_shape_bit;
709            vect_size = n_ch * mtab->ppc_shape_len;
710        } else {
711            bit_size  = total_fr_bits - bsize_no_main_cb[i];
712            vect_size = n_ch * mtab->size;
713        }
714
715        tctx->n_div[i] = (bit_size + 13) / 14;
716
717        rounded_up                     = (bit_size + tctx->n_div[i] - 1) /
718                                         tctx->n_div[i];
719        rounded_down                   = (bit_size) / tctx->n_div[i];
720        num_rounded_down               = rounded_up * tctx->n_div[i] - bit_size;
721        num_rounded_up                 = tctx->n_div[i] - num_rounded_down;
722        tctx->bits_main_spec[0][i][0]  = (rounded_up + 1)   / 2;
723        tctx->bits_main_spec[1][i][0]  =  rounded_up        / 2;
724        tctx->bits_main_spec[0][i][1]  = (rounded_down + 1) / 2;
725        tctx->bits_main_spec[1][i][1]  =  rounded_down      / 2;
726        tctx->bits_main_spec_change[i] = num_rounded_up;
727
728        rounded_up             = (vect_size + tctx->n_div[i] - 1) /
729                                 tctx->n_div[i];
730        rounded_down           = (vect_size) / tctx->n_div[i];
731        num_rounded_down       = rounded_up * tctx->n_div[i] - vect_size;
732        num_rounded_up         = tctx->n_div[i] - num_rounded_down;
733        tctx->length[i][0]     = rounded_up;
734        tctx->length[i][1]     = rounded_down;
735        tctx->length_change[i] = num_rounded_up;
736    }
737
738    for (frametype = TWINVQ_FT_SHORT; frametype <= TWINVQ_FT_PPC; frametype++)
739        construct_perm_table(tctx, frametype);
740}
741
742av_cold int ff_twinvq_decode_close(AVCodecContext *avctx)
743{
744    TwinVQContext *tctx = avctx->priv_data;
745    int i;
746
747    for (i = 0; i < 3; i++) {
748        ff_mdct_end(&tctx->mdct_ctx[i]);
749        av_freep(&tctx->cos_tabs[i]);
750    }
751
752    av_freep(&tctx->curr_frame);
753    av_freep(&tctx->spectrum);
754    av_freep(&tctx->prev_frame);
755    av_freep(&tctx->tmp_buf);
756    av_freep(&tctx->fdsp);
757
758    return 0;
759}
760
761av_cold int ff_twinvq_decode_init(AVCodecContext *avctx)
762{
763    int ret;
764    TwinVQContext *tctx = avctx->priv_data;
765    int64_t frames_per_packet;
766
767    tctx->avctx       = avctx;
768    avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
769
770    if (!avctx->block_align) {
771        avctx->block_align = tctx->frame_size + 7 >> 3;
772    }
773    frames_per_packet = avctx->block_align * 8LL / tctx->frame_size;
774    if (frames_per_packet <= 0) {
775        av_log(avctx, AV_LOG_ERROR, "Block align is %"PRId64" bits, expected %d\n",
776               avctx->block_align * (int64_t)8, tctx->frame_size);
777        return AVERROR_INVALIDDATA;
778    }
779    if (frames_per_packet > TWINVQ_MAX_FRAMES_PER_PACKET) {
780        av_log(avctx, AV_LOG_ERROR, "Too many frames per packet (%"PRId64")\n",
781               frames_per_packet);
782        return AVERROR_INVALIDDATA;
783    }
784    tctx->frames_per_packet = frames_per_packet;
785
786    tctx->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
787    if (!tctx->fdsp)
788        return AVERROR(ENOMEM);
789    if ((ret = init_mdct_win(tctx))) {
790        av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
791        return ret;
792    }
793    init_bitstream_params(tctx);
794
795    twinvq_memset_float(tctx->bark_hist[0][0], 0.1,
796                        FF_ARRAY_ELEMS(tctx->bark_hist));
797
798    return 0;
799}
800