libavcodec/twinvq.c

   1 /*
   2  * TwinVQ decoder
   3  * Copyright (c) 2009 Vitor Sessak
   4  *
   5  * This file is part of Libav.
   6  *
   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.
  11  *
  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.
  16  *
  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
  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  */
  46 static 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  */
  71 static 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
  84 static 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
  95 static 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  */
 115 static 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
 152 static 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  */
 172 static 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
 214 static 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 sub        = mtab->fmode[ftype].sub;
 221     float step     = TWINVQ_AMP_MAX     / ((1 << TWINVQ_GAIN_BITS)     - 1);
 222     float sub_step = TWINVQ_SUB_AMP_MAX / ((1 << TWINVQ_SUB_GAIN_BITS) - 1);
 223
 224     if (ftype == TWINVQ_FT_LONG) {
 225         for (i = 0; i < tctx->avctx->channels; i++)
 226             out[i] = (1.0 / (1 << 13)) *
 227                      twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
 228                                      TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
 229     } else {
 230         for (i = 0; i < tctx->avctx->channels; i++) {
 231             float val = (1.0 / (1 << 23)) *
 232                         twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
 233                                         TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
 234
 235             for (j = 0; j < sub; j++)
 236                 out[i * sub + j] =
 237                     val * twinvq_mulawinv(sub_step * 0.5 +
 238                                           sub_step * bits->sub_gain_bits[i * sub + j],
 239                                           TWINVQ_SUB_AMP_MAX, TWINVQ_MULAW_MU);
 240         }
 241     }
 242 }
 243
 244 /**
 245  * Rearrange the LSP coefficients so that they have a minimum distance of
 246  * min_dist. This function does it exactly as described in section of 3.2.4
 247  * of the G.729 specification (but interestingly is different from what the
 248  * reference decoder actually does).
 249  */
 250 static void rearrange_lsp(int order, float *lsp, float min_dist)
 251 {
 252     int i;
 253     float min_dist2 = min_dist * 0.5;
 254     for (i = 1; i < order; i++)
 255         if (lsp[i] - lsp[i - 1] < min_dist) {
 256             float avg = (lsp[i] + lsp[i - 1]) * 0.5;
 257
 258             lsp[i - 1] = avg - min_dist2;
 259             lsp[i]     = avg + min_dist2;
 260         }
 261 }
 262
 263 static void decode_lsp(TwinVQContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
 264                        int lpc_hist_idx, float *lsp, float *hist)
 265 {
 266     const TwinVQModeTab *mtab = tctx->mtab;
 267     int i, j;
 268
 269     const float *cb  = mtab->lspcodebook;
 270     const float *cb2 = cb  + (1 << mtab->lsp_bit1) * mtab->n_lsp;
 271     const float *cb3 = cb2 + (1 << mtab->lsp_bit2) * mtab->n_lsp;
 272
 273     const int8_t funny_rounding[4] = {
 274         -2,
 275         mtab->lsp_split == 4 ? -2 : 1,
 276         mtab->lsp_split == 4 ? -2 : 1,
 277         0
 278     };
 279
 280     j = 0;
 281     for (i = 0; i < mtab->lsp_split; i++) {
 282         int chunk_end = ((i + 1) * mtab->n_lsp + funny_rounding[i]) /
 283                         mtab->lsp_split;
 284         for (; j < chunk_end; j++)
 285             lsp[j] = cb[lpc_idx1     * mtab->n_lsp + j] +
 286                      cb2[lpc_idx2[i] * mtab->n_lsp + j];
 287     }
 288
 289     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
 290
 291     for (i = 0; i < mtab->n_lsp; i++) {
 292         float tmp1 = 1.0     - cb3[lpc_hist_idx * mtab->n_lsp + i];
 293         float tmp2 = hist[i] * cb3[lpc_hist_idx * mtab->n_lsp + i];
 294         hist[i] = lsp[i];
 295         lsp[i]  = lsp[i] * tmp1 + tmp2;
 296     }
 297
 298     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
 299     rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
 300     ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
 301 }
 302
 303 static void dec_lpc_spectrum_inv(TwinVQContext *tctx, float *lsp,
 304                                  enum TwinVQFrameType ftype, float *lpc)
 305 {
 306     int i;
 307     int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
 308
 309     for (i = 0; i < tctx->mtab->n_lsp; i++)
 310         lsp[i] = 2 * cos(lsp[i]);
 311
 312     switch (ftype) {
 313     case TWINVQ_FT_LONG:
 314         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
 315         break;
 316     case TWINVQ_FT_MEDIUM:
 317         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
 318         break;
 319     case TWINVQ_FT_SHORT:
 320         eval_lpcenv(tctx, lsp, lpc);
 321         break;
 322     }
 323 }
 324
 325 static const uint8_t wtype_to_wsize[] = { 0, 0, 2, 2, 2, 1, 0, 1, 1 };
 326
 327 static void imdct_and_window(TwinVQContext *tctx, enum TwinVQFrameType ftype,
 328                              int wtype, float *in, float *prev, int ch)
 329 {
 330     FFTContext *mdct = &tctx->mdct_ctx[ftype];
 331     const TwinVQModeTab *mtab = tctx->mtab;
 332     int bsize = mtab->size / mtab->fmode[ftype].sub;
 333     int size  = mtab->size;
 334     float *buf1 = tctx->tmp_buf;
 335     int j, first_wsize, wsize; // Window size
 336     float *out  = tctx->curr_frame + 2 * ch * mtab->size;
 337     float *out2 = out;
 338     float *prev_buf;
 339     int types_sizes[] = {
 340         mtab->size /  mtab->fmode[TWINVQ_FT_LONG].sub,
 341         mtab->size /  mtab->fmode[TWINVQ_FT_MEDIUM].sub,
 342         mtab->size / (mtab->fmode[TWINVQ_FT_SHORT].sub * 2),
 343     };
 344
 345     wsize       = types_sizes[wtype_to_wsize[wtype]];
 346     first_wsize = wsize;
 347     prev_buf    = prev + (size - bsize) / 2;
 348
 349     for (j = 0; j < mtab->fmode[ftype].sub; j++) {
 350         int sub_wtype = ftype == TWINVQ_FT_MEDIUM ? 8 : wtype;
 351
 352         if (!j && wtype == 4)
 353             sub_wtype = 4;
 354         else if (j == mtab->fmode[ftype].sub - 1 && wtype == 7)
 355             sub_wtype = 7;
 356
 357         wsize = types_sizes[wtype_to_wsize[sub_wtype]];
 358
 359         mdct->imdct_half(mdct, buf1 + bsize * j, in + bsize * j);
 360
 361         tctx->fdsp.vector_fmul_window(out2, prev_buf + (bsize - wsize) / 2,
 362                                       buf1 + bsize * j,
 363                                       ff_sine_windows[av_log2(wsize)],
 364                                       wsize / 2);
 365         out2 += wsize;
 366
 367         memcpy(out2, buf1 + bsize * j + wsize / 2,
 368                (bsize - wsize / 2) * sizeof(float));
 369
 370         out2 += ftype == TWINVQ_FT_MEDIUM ? (bsize - wsize) / 2 : bsize - wsize;
 371
 372         prev_buf = buf1 + bsize * j + bsize / 2;
 373     }
 374
 375     tctx->last_block_pos[ch] = (size + first_wsize) / 2;
 376 }
 377
 378 static void imdct_output(TwinVQContext *tctx, enum TwinVQFrameType ftype,
 379                          int wtype, float **out, int offset)
 380 {
 381     const TwinVQModeTab *mtab = tctx->mtab;
 382     float *prev_buf           = tctx->prev_frame + tctx->last_block_pos[0];
 383     int size1, size2, i;
 384     float *out1, *out2;
 385
 386     for (i = 0; i < tctx->avctx->channels; i++)
 387         imdct_and_window(tctx, ftype, wtype,
 388                          tctx->spectrum + i * mtab->size,
 389                          prev_buf + 2 * i * mtab->size,
 390                          i);
 391
 392     if (!out)
 393         return;
 394
 395     size2 = tctx->last_block_pos[0];
 396     size1 = mtab->size - size2;
 397
 398     out1 = &out[0][0] + offset;
 399     memcpy(out1,         prev_buf,         size1 * sizeof(*out1));
 400     memcpy(out1 + size1, tctx->curr_frame, size2 * sizeof(*out1));
 401
 402     if (tctx->avctx->channels == 2) {
 403         out2 = &out[1][0] + offset;
 404         memcpy(out2, &prev_buf[2 * mtab->size],
 405                size1 * sizeof(*out2));
 406         memcpy(out2 + size1, &tctx->curr_frame[2 * mtab->size],
 407                size2 * sizeof(*out2));
 408         tctx->fdsp.butterflies_float(out1, out2, mtab->size);
 409     }
 410 }
 411
 412 static void read_and_decode_spectrum(TwinVQContext *tctx, float *out,
 413                                      enum TwinVQFrameType ftype)
 414 {
 415     const TwinVQModeTab *mtab = tctx->mtab;
 416     TwinVQFrameData *bits     = &tctx->bits[tctx->cur_frame];
 417     int channels              = tctx->avctx->channels;
 418     int sub        = mtab->fmode[ftype].sub;
 419     int block_size = mtab->size / sub;
 420     float gain[TWINVQ_CHANNELS_MAX * TWINVQ_SUBBLOCKS_MAX];
 421     float ppc_shape[TWINVQ_PPC_SHAPE_LEN_MAX * TWINVQ_CHANNELS_MAX * 4];
 422
 423     int i, j;
 424
 425     dequant(tctx, bits->main_coeffs, out, ftype,
 426             mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
 427             mtab->fmode[ftype].cb_len_read);
 428
 429     dec_gain(tctx, ftype, gain);
 430
 431     if (ftype == TWINVQ_FT_LONG) {
 432         int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len * channels - 1) /
 433                        tctx->n_div[3];
 434         dequant(tctx, bits->ppc_coeffs, ppc_shape,
 435                 TWINVQ_FT_PPC, mtab->ppc_shape_cb,
 436                 mtab->ppc_shape_cb + cb_len_p * TWINVQ_PPC_SHAPE_CB_SIZE,
 437                 cb_len_p);
 438     }
 439
 440     for (i = 0; i < channels; i++) {
 441         float *chunk = out + mtab->size * i;
 442         float lsp[TWINVQ_LSP_COEFS_MAX];
 443
 444         for (j = 0; j < sub; j++) {
 445             tctx->dec_bark_env(tctx, bits->bark1[i][j],
 446                                bits->bark_use_hist[i][j], i,
 447                                tctx->tmp_buf, gain[sub * i + j], ftype);
 448
 449             tctx->fdsp.vector_fmul(chunk + block_size * j,
 450                                    chunk + block_size * j,
 451                                    tctx->tmp_buf, block_size);
 452         }
 453
 454         if (ftype == TWINVQ_FT_LONG)
 455             tctx->decode_ppc(tctx, bits->p_coef[i], bits->g_coef[i],
 456                              ppc_shape + i * mtab->ppc_shape_len, chunk);
 457
 458         decode_lsp(tctx, bits->lpc_idx1[i], bits->lpc_idx2[i],
 459                    bits->lpc_hist_idx[i], lsp, tctx->lsp_hist[i]);
 460
 461         dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
 462
 463         for (j = 0; j < mtab->fmode[ftype].sub; j++) {
 464             tctx->fdsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
 465             chunk += block_size;
 466         }
 467     }
 468 }
 469
 470 const enum TwinVQFrameType ff_twinvq_wtype_to_ftype_table[] = {
 471     TWINVQ_FT_LONG,   TWINVQ_FT_LONG, TWINVQ_FT_SHORT, TWINVQ_FT_LONG,
 472     TWINVQ_FT_MEDIUM, TWINVQ_FT_LONG, TWINVQ_FT_LONG,  TWINVQ_FT_MEDIUM,
 473     TWINVQ_FT_MEDIUM
 474 };
 475
 476 int ff_twinvq_decode_frame(AVCodecContext *avctx, void *data,
 477                            int *got_frame_ptr, AVPacket *avpkt)
 478 {
 479     AVFrame *frame     = data;
 480     const uint8_t *buf = avpkt->data;
 481     int buf_size       = avpkt->size;
 482     TwinVQContext *tctx = avctx->priv_data;
 483     const TwinVQModeTab *mtab = tctx->mtab;
 484     float **out = NULL;
 485     int ret;
 486
 487     /* get output buffer */
 488     if (tctx->discarded_packets >= 2) {
 489         frame->nb_samples = mtab->size * tctx->frames_per_packet;
 490         if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
 491             av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
 492             return ret;
 493         }
 494         out = (float **)frame->extended_data;
 495     }
 496
 497     if (buf_size < avctx->block_align) {
 498         av_log(avctx, AV_LOG_ERROR,
 499                "Frame too small (%d bytes). Truncated file?\n", buf_size);
 500         return AVERROR(EINVAL);
 501     }
 502
 503     if ((ret = tctx->read_bitstream(avctx, tctx, buf, buf_size)) < 0)
 504         return ret;
 505
 506     for (tctx->cur_frame = 0; tctx->cur_frame < tctx->frames_per_packet;
 507          tctx->cur_frame++) {
 508         read_and_decode_spectrum(tctx, tctx->spectrum,
 509                                  tctx->bits[tctx->cur_frame].ftype);
 510
 511         imdct_output(tctx, tctx->bits[tctx->cur_frame].ftype,
 512                      tctx->bits[tctx->cur_frame].window_type, out,
 513                      tctx->cur_frame * mtab->size);
 514
 515         FFSWAP(float *, tctx->curr_frame, tctx->prev_frame);
 516     }
 517
 518     if (tctx->discarded_packets < 2) {
 519         tctx->discarded_packets++;
 520         *got_frame_ptr = 0;
 521         return buf_size;
 522     }
 523
 524     *got_frame_ptr = 1;
 525
 526     // VQF can deliver packets 1 byte greater than block align
 527     if (buf_size == avctx->block_align + 1)
 528         return buf_size;
 529     return avctx->block_align;
 530 }
 531
 532 /**
 533  * Init IMDCT and windowing tables
 534  */
 535 static av_cold int init_mdct_win(TwinVQContext *tctx)
 536 {
 537     int i, j, ret;
 538     const TwinVQModeTab *mtab = tctx->mtab;
 539     int size_s = mtab->size / mtab->fmode[TWINVQ_FT_SHORT].sub;
 540     int size_m = mtab->size / mtab->fmode[TWINVQ_FT_MEDIUM].sub;
 541     int channels = tctx->avctx->channels;
 542     float norm = channels == 1 ? 2.0 : 1.0;
 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     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
 552                      mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
 553
 554     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
 555                      2 * mtab->size * channels * sizeof(*tctx->spectrum),
 556                      alloc_fail);
 557     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
 558                      2 * mtab->size * channels * sizeof(*tctx->curr_frame),
 559                      alloc_fail);
 560     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
 561                      2 * mtab->size * channels * sizeof(*tctx->prev_frame),
 562                      alloc_fail);
 563
 564     for (i = 0; i < 3; i++) {
 565         int m       = 4 * mtab->size / mtab->fmode[i].sub;
 566         double freq = 2 * M_PI / m;
 567         FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
 568                          (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
 569
 570         for (j = 0; j <= m / 8; j++)
 571             tctx->cos_tabs[i][j] = cos((2 * j + 1) * freq);
 572         for (j = 1; j < m / 8; j++)
 573             tctx->cos_tabs[i][m / 4 - j] = tctx->cos_tabs[i][j];
 574     }
 575
 576     ff_init_ff_sine_windows(av_log2(size_m));
 577     ff_init_ff_sine_windows(av_log2(size_s / 2));
 578     ff_init_ff_sine_windows(av_log2(mtab->size));
 579
 580     return 0;
 581
 582 alloc_fail:
 583     return AVERROR(ENOMEM);
 584 }
 585
 586 /**
 587  * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
 588  * each line do a cyclic permutation, i.e.
 589  * abcdefghijklm -> defghijklmabc
 590  * where the amount to be shifted is evaluated depending on the column.
 591  */
 592 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
 593                               int block_size,
 594                               const uint8_t line_len[2], int length_div,
 595                               enum TwinVQFrameType ftype)
 596 {
 597     int i, j;
 598
 599     for (i = 0; i < line_len[0]; i++) {
 600         int shift;
 601
 602         if (num_blocks == 1                                    ||
 603             (ftype == TWINVQ_FT_LONG && num_vect % num_blocks) ||
 604             (ftype != TWINVQ_FT_LONG && num_vect & 1)          ||
 605             i == line_len[1]) {
 606             shift = 0;
 607         } else if (ftype == TWINVQ_FT_LONG) {
 608             shift = i;
 609         } else
 610             shift = i * i;
 611
 612         for (j = 0; j < num_vect && (j + num_vect * i < block_size * num_blocks); j++)
 613             tab[i * num_vect + j] = i * num_vect + (j + shift) % num_vect;
 614     }
 615 }
 616
 617 /**
 618  * Interpret the input data as in the following table:
 619  *
 620  * @verbatim
 621  *
 622  * abcdefgh
 623  * ijklmnop
 624  * qrstuvw
 625  * x123456
 626  *
 627  * @endverbatim
 628  *
 629  * and transpose it, giving the output
 630  * aiqxbjr1cks2dlt3emu4fvn5gow6hp
 631  */
 632 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
 633                            const uint8_t line_len[2], int length_div)
 634 {
 635     int i, j;
 636     int cont = 0;
 637
 638     for (i = 0; i < num_vect; i++)
 639         for (j = 0; j < line_len[i >= length_div]; j++)
 640             out[cont++] = in[j * num_vect + i];
 641 }
 642
 643 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
 644 {
 645     int block_size = size / n_blocks;
 646     int i;
 647
 648     for (i = 0; i < size; i++)
 649         out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
 650 }
 651
 652 static av_cold void construct_perm_table(TwinVQContext *tctx,
 653                                          enum TwinVQFrameType ftype)
 654 {
 655     int block_size, size;
 656     const TwinVQModeTab *mtab = tctx->mtab;
 657     int16_t *tmp_perm = (int16_t *)tctx->tmp_buf;
 658
 659     if (ftype == TWINVQ_FT_PPC) {
 660         size       = tctx->avctx->channels;
 661         block_size = mtab->ppc_shape_len;
 662     } else {
 663         size       = tctx->avctx->channels * mtab->fmode[ftype].sub;
 664         block_size = mtab->size / mtab->fmode[ftype].sub;
 665     }
 666
 667     permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
 668                       block_size, tctx->length[ftype],
 669                       tctx->length_change[ftype], ftype);
 670
 671     transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
 672                    tctx->length[ftype], tctx->length_change[ftype]);
 673
 674     linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
 675                 size * block_size);
 676 }
 677
 678 static av_cold void init_bitstream_params(TwinVQContext *tctx)
 679 {
 680     const TwinVQModeTab *mtab = tctx->mtab;
 681     int n_ch                  = tctx->avctx->channels;
 682     int total_fr_bits         = tctx->avctx->bit_rate * mtab->size /
 683                                 tctx->avctx->sample_rate;
 684
 685     int lsp_bits_per_block = n_ch * (mtab->lsp_bit0 + mtab->lsp_bit1 +
 686                                      mtab->lsp_split * mtab->lsp_bit2);
 687
 688     int ppc_bits = n_ch * (mtab->pgain_bit + mtab->ppc_shape_bit +
 689                            mtab->ppc_period_bit);
 690
 691     int bsize_no_main_cb[3], bse_bits[3], i;
 692     enum TwinVQFrameType frametype;
 693
 694     for (i = 0; i < 3; i++)
 695         // +1 for history usage switch
 696         bse_bits[i] = n_ch *
 697                       (mtab->fmode[i].bark_n_coef *
 698                        mtab->fmode[i].bark_n_bit + 1);
 699
 700     bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
 701                           TWINVQ_WINDOW_TYPE_BITS + n_ch * TWINVQ_GAIN_BITS;
 702
 703     for (i = 0; i < 2; i++)
 704         bsize_no_main_cb[i] =
 705             lsp_bits_per_block + n_ch * TWINVQ_GAIN_BITS +
 706             TWINVQ_WINDOW_TYPE_BITS +
 707             mtab->fmode[i].sub * (bse_bits[i] + n_ch * TWINVQ_SUB_GAIN_BITS);
 708
 709     if (tctx->codec == TWINVQ_CODEC_METASOUND && !tctx->is_6kbps) {
 710         bsize_no_main_cb[1] += 2;
 711         bsize_no_main_cb[2] += 2;
 712     }
 713
 714     // The remaining bits are all used for the main spectrum coefficients
 715     for (i = 0; i < 4; i++) {
 716         int bit_size, vect_size;
 717         int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
 718         if (i == 3) {
 719             bit_size  = n_ch * mtab->ppc_shape_bit;
 720             vect_size = n_ch * mtab->ppc_shape_len;
 721         } else {
 722             bit_size  = total_fr_bits - bsize_no_main_cb[i];
 723             vect_size = n_ch * mtab->size;
 724         }
 725
 726         tctx->n_div[i] = (bit_size + 13) / 14;
 727
 728         rounded_up                     = (bit_size + tctx->n_div[i] - 1) /
 729                                          tctx->n_div[i];
 730         rounded_down                   = (bit_size) / tctx->n_div[i];
 731         num_rounded_down               = rounded_up * tctx->n_div[i] - bit_size;
 732         num_rounded_up                 = tctx->n_div[i] - num_rounded_down;
 733         tctx->bits_main_spec[0][i][0]  = (rounded_up + 1)   / 2;
 734         tctx->bits_main_spec[1][i][0]  =  rounded_up        / 2;
 735         tctx->bits_main_spec[0][i][1]  = (rounded_down + 1) / 2;
 736         tctx->bits_main_spec[1][i][1]  =  rounded_down      / 2;
 737         tctx->bits_main_spec_change[i] = num_rounded_up;
 738
 739         rounded_up             = (vect_size + tctx->n_div[i] - 1) /
 740                                  tctx->n_div[i];
 741         rounded_down           = (vect_size) / tctx->n_div[i];
 742         num_rounded_down       = rounded_up * tctx->n_div[i] - vect_size;
 743         num_rounded_up         = tctx->n_div[i] - num_rounded_down;
 744         tctx->length[i][0]     = rounded_up;
 745         tctx->length[i][1]     = rounded_down;
 746         tctx->length_change[i] = num_rounded_up;
 747     }
 748
 749     for (frametype = TWINVQ_FT_SHORT; frametype <= TWINVQ_FT_PPC; frametype++)
 750         construct_perm_table(tctx, frametype);
 751 }
 752
 753 av_cold int ff_twinvq_decode_close(AVCodecContext *avctx)
 754 {
 755     TwinVQContext *tctx = avctx->priv_data;
 756     int i;
 757
 758     for (i = 0; i < 3; i++) {
 759         ff_mdct_end(&tctx->mdct_ctx[i]);
 760         av_free(tctx->cos_tabs[i]);
 761     }
 762
 763     av_free(tctx->curr_frame);
 764     av_free(tctx->spectrum);
 765     av_free(tctx->prev_frame);
 766     av_free(tctx->tmp_buf);
 767
 768     return 0;
 769 }
 770
 771 av_cold int ff_twinvq_decode_init(AVCodecContext *avctx)
 772 {
 773     int ret;
 774     TwinVQContext *tctx = avctx->priv_data;
 775
 776     tctx->avctx       = avctx;
 777     avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
 778
 779     if (!avctx->block_align) {
 780         avctx->block_align = tctx->frame_size + 7 >> 3;
 781     } else if (avctx->block_align * 8 < tctx->frame_size) {
 782         av_log(avctx, AV_LOG_ERROR, "Block align is %d bits, expected %d\n",
 783                avctx->block_align * 8, tctx->frame_size);
 784         return AVERROR_INVALIDDATA;
 785     }
 786     tctx->frames_per_packet = avctx->block_align * 8 / tctx->frame_size;
 787     if (tctx->frames_per_packet > TWINVQ_MAX_FRAMES_PER_PACKET) {
 788         av_log(avctx, AV_LOG_ERROR, "Too many frames per packet (%d)\n",
 789                tctx->frames_per_packet);
 790         return AVERROR_INVALIDDATA;
 791     }
 792
 793     avpriv_float_dsp_init(&tctx->fdsp, avctx->flags & AV_CODEC_FLAG_BITEXACT);
 794     if ((ret = init_mdct_win(tctx))) {
 795         av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
 796         ff_twinvq_decode_close(avctx);
 797         return ret;
 798     }
 799     init_bitstream_params(tctx);
 800
 801     twinvq_memset_float(tctx->bark_hist[0][0], 0.1,
 802                         FF_ARRAY_ELEMS(tctx->bark_hist));
 803
 804     return 0;
 805 }