libavcodec/twinvq.c

   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 "libavutil/mem.h"
  28 #include "avcodec.h"
  29 #include "decode.h"
  30 #include "lsp.h"
  31 #include "metasound_twinvq_data.h"
  32 #include "sinewin.h"
  33 #include "twinvq.h"
  34
  35 /**
  36  * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
  37  * spectrum pairs.
  38  *
  39  * @param lsp a vector of the cosine of the LSP values
  40  * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
  41  * @param order the order of the LSP (and the size of the *lsp buffer). Must
  42  *        be a multiple of four.
  43  * @return the LPC value
  44  *
  45  * @todo reuse code from Vorbis decoder: vorbis_floor0_decode
  46  */
  47 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
  48 {
  49     int j;
  50     float p         = 0.5f;
  51     float q         = 0.5f;
  52     float two_cos_w = 2.0f * cos_val;
  53
  54     for (j = 0; j + 1 < order; j += 2 * 2) {
  55         // Unroll the loop once since order is a multiple of four
  56         q *= lsp[j]     - two_cos_w;
  57         p *= lsp[j + 1] - two_cos_w;
  58
  59         q *= lsp[j + 2] - two_cos_w;
  60         p *= lsp[j + 3] - two_cos_w;
  61     }
  62
  63     p *= p * (2.0f - two_cos_w);
  64     q *= q * (2.0f + two_cos_w);
  65
  66     return 0.5 / (p + q);
  67 }
  68
  69 /**
  70  * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
  71  */
  72 static void eval_lpcenv(TwinVQContext *tctx, const float *cos_vals, float *lpc)
  73 {
  74     int i;
  75     const TwinVQModeTab *mtab = tctx->mtab;
  76     int size_s = mtab->size / mtab->fmode[TWINVQ_FT_SHORT].sub;
  77
  78     for (i = 0; i < size_s / 2; i++) {
  79         float cos_i = tctx->cos_tabs[0][i];
  80         lpc[i]              = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
  81         lpc[size_s - i - 1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
  82     }
  83 }
  84
  85 static void interpolate(float *out, float v1, float v2, int size)
  86 {
  87     int i;
  88     float step = (v1 - v2) / (size + 1);
  89
  90     for (i = 0; i < size; i++) {
  91         v2    += step;
  92         out[i] = v2;
  93     }
  94 }
  95
  96 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
  97 {
  98     return part ? -cos_tab[size - idx - 1]
  99                 :  cos_tab[idx];
 100 }
 101
 102 /**
 103  * Evaluate the LPC amplitude spectrum envelope from the line spectrum pairs.
 104  * Probably for speed reasons, the coefficients are evaluated as
 105  * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
 106  * where s is an evaluated value, i is a value interpolated from the others
 107  * and b might be either calculated or interpolated, depending on an
 108  * unexplained condition.
 109  *
 110  * @param step the size of a block "siiiibiiii"
 111  * @param in the cosine of the LSP data
 112  * @param part is 0 for 0...PI (positive cosine values) and 1 for PI...2PI
 113  *        (negative cosine values)
 114  * @param size the size of the whole output
 115  */
 116 static inline void eval_lpcenv_or_interp(TwinVQContext *tctx,
 117                                          enum TwinVQFrameType ftype,
 118                                          float *out, const float *in,
 119                                          int size, int step, int part)
 120 {
 121     int i;
 122     const TwinVQModeTab *mtab = tctx->mtab;
 123     const float *cos_tab      = tctx->cos_tabs[ftype];
 124
 125     // Fill the 's'
 126     for (i = 0; i < size; i += step)
 127         out[i] =
 128             eval_lpc_spectrum(in,
 129                               get_cos(i, part, cos_tab, size),
 130                               mtab->n_lsp);
 131
 132     // Fill the 'iiiibiiii'
 133     for (i = step; i <= size - 2 * step; i += step) {
 134         if (out[i + step] + out[i - step] > 1.95 * out[i] ||
 135             out[i + step]                 >= out[i - step]) {
 136             interpolate(out + i - step + 1, out[i], out[i - step], step - 1);
 137         } else {
 138             out[i - step / 2] =
 139                 eval_lpc_spectrum(in,
 140                                   get_cos(i - step / 2, part, cos_tab, size),
 141                                   mtab->n_lsp);
 142             interpolate(out + i - step + 1, out[i - step / 2],
 143                         out[i - step], step / 2 - 1);
 144             interpolate(out + i - step / 2 + 1, out[i],
 145                         out[i - step / 2], step / 2 - 1);
 146         }
 147     }
 148
 149     interpolate(out + size - 2 * step + 1, out[size - step],
 150                 out[size - 2 * step], step - 1);
 151 }
 152
 153 static void eval_lpcenv_2parts(TwinVQContext *tctx, enum TwinVQFrameType ftype,
 154                                const float *buf, float *lpc,
 155                                int size, int step)
 156 {
 157     eval_lpcenv_or_interp(tctx, ftype, lpc, buf, size / 2, step, 0);
 158     eval_lpcenv_or_interp(tctx, ftype, lpc + size / 2, buf, size / 2,
 159                           2 * step, 1);
 160
 161     interpolate(lpc + size / 2 - step + 1, lpc[size / 2],
 162                 lpc[size / 2 - step], step);
 163
 164     twinvq_memset_float(lpc + size - 2 * step + 1, lpc[size - 2 * step],
 165                         2 * step - 1);
 166 }
 167
 168 /**
 169  * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
 170  * bitstream, sum the corresponding vectors and write the result to *out
 171  * after permutation.
 172  */
 173 static void dequant(TwinVQContext *tctx, const uint8_t *cb_bits, float *out,
 174                     enum TwinVQFrameType ftype,
 175                     const int16_t *cb0, const int16_t *cb1, int cb_len)
 176 {
 177     int pos = 0;
 178     int i, j;
 179
 180     for (i = 0; i < tctx->n_div[ftype]; i++) {
 181         int tmp0, tmp1;
 182         int sign0 = 1;
 183         int sign1 = 1;
 184         const int16_t *tab0, *tab1;
 185         int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
 186         int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
 187
 188         int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
 189         tmp0 = *cb_bits++;
 190         if (bits == 7) {
 191             if (tmp0 & 0x40)
 192                 sign0 = -1;
 193             tmp0 &= 0x3F;
 194         }
 195
 196         bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
 197         tmp1 = *cb_bits++;
 198         if (bits == 7) {
 199             if (tmp1 & 0x40)
 200                 sign1 = -1;
 201             tmp1 &= 0x3F;
 202         }
 203
 204         tab0 = cb0 + tmp0 * cb_len;
 205         tab1 = cb1 + tmp1 * cb_len;
 206
 207         for (j = 0; j < length; j++)
 208             out[tctx->permut[ftype][pos + j]] = sign0 * tab0[j] +
 209                                                 sign1 * tab1[j];
 210
 211         pos += length;
 212     }
 213 }
 214
 215 static void dec_gain(TwinVQContext *tctx,
 216                      enum TwinVQFrameType ftype, float *out)
 217 {
 218     const TwinVQModeTab   *mtab =  tctx->mtab;
 219     const TwinVQFrameData *bits = &tctx->bits[tctx->cur_frame];
 220     int i, j;
 221     int channels   = tctx->avctx->ch_layout.nb_channels;
 222     int sub        = mtab->fmode[ftype].sub;
 223     float step     = TWINVQ_AMP_MAX     / ((1 << TWINVQ_GAIN_BITS)     - 1);
 224     float sub_step = TWINVQ_SUB_AMP_MAX / ((1 << TWINVQ_SUB_GAIN_BITS) - 1);
 225
 226     if (ftype == TWINVQ_FT_LONG) {
 227         for (i = 0; i < channels; i++)
 228             out[i] = (1.0 / (1 << 13)) *
 229                      twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
 230                                      TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
 231     } else {
 232         for (i = 0; i < channels; i++) {
 233             float val = (1.0 / (1 << 23)) *
 234                         twinvq_mulawinv(step * 0.5 + step * bits->gain_bits[i],
 235                                         TWINVQ_AMP_MAX, TWINVQ_MULAW_MU);
 236
 237             for (j = 0; j < sub; j++)
 238                 out[i * sub + j] =
 239                     val * twinvq_mulawinv(sub_step * 0.5 +
 240                                           sub_step * bits->sub_gain_bits[i * sub + j],
 241                                           TWINVQ_SUB_AMP_MAX, TWINVQ_MULAW_MU);
 242         }
 243     }
 244 }
 245
 246 /**
 247  * Rearrange the LSP coefficients so that they have a minimum distance of
 248  * min_dist. This function does it exactly as described in section of 3.2.4
 249  * of the G.729 specification (but interestingly is different from what the
 250  * reference decoder actually does).
 251  */
 252 static void rearrange_lsp(int order, float *lsp, float min_dist)
 253 {
 254     int i;
 255     float min_dist2 = min_dist * 0.5;
 256     for (i = 1; i < order; i++)
 257         if (lsp[i] - lsp[i - 1] < min_dist) {
 258             float avg = (lsp[i] + lsp[i - 1]) * 0.5;
 259
 260             lsp[i - 1] = avg - min_dist2;
 261             lsp[i]     = avg + min_dist2;
 262         }
 263 }
 264
 265 static void decode_lsp(TwinVQContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
 266                        int lpc_hist_idx, float *lsp, float *hist)
 267 {
 268     const TwinVQModeTab *mtab = tctx->mtab;
 269     int i, j;
 270
 271     const float *cb  = mtab->lspcodebook;
 272     const float *cb2 = cb  + (1 << mtab->lsp_bit1) * mtab->n_lsp;
 273     const float *cb3 = cb2 + (1 << mtab->lsp_bit2) * mtab->n_lsp;
 274
 275     const int8_t funny_rounding[4] = {
 276         -2,
 277         mtab->lsp_split == 4 ? -2 : 1,
 278         mtab->lsp_split == 4 ? -2 : 1,
 279         0
 280     };
 281
 282     j = 0;
 283     for (i = 0; i < mtab->lsp_split; i++) {
 284         int chunk_end = ((i + 1) * mtab->n_lsp + funny_rounding[i]) /
 285                         mtab->lsp_split;
 286         for (; j < chunk_end; j++)
 287             lsp[j] = cb[lpc_idx1     * mtab->n_lsp + j] +
 288                      cb2[lpc_idx2[i] * mtab->n_lsp + j];
 289     }
 290
 291     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
 292
 293     for (i = 0; i < mtab->n_lsp; i++) {
 294         float tmp1 = 1.0     - cb3[lpc_hist_idx * mtab->n_lsp + i];
 295         float tmp2 = hist[i] * cb3[lpc_hist_idx * mtab->n_lsp + i];
 296         hist[i] = lsp[i];
 297         lsp[i]  = lsp[i] * tmp1 + tmp2;
 298     }
 299
 300     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
 301     rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
 302     ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
 303 }
 304
 305 static void dec_lpc_spectrum_inv(TwinVQContext *tctx, float *lsp,
 306                                  enum TwinVQFrameType ftype, float *lpc)
 307 {
 308     int i;
 309     int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
 310
 311     for (i = 0; i < tctx->mtab->n_lsp; i++)
 312         lsp[i] = 2 * cos(lsp[i]);
 313
 314     switch (ftype) {
 315     case TWINVQ_FT_LONG:
 316         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
 317         break;
 318     case TWINVQ_FT_MEDIUM:
 319         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
 320         break;
 321     case TWINVQ_FT_SHORT:
 322         eval_lpcenv(tctx, lsp, lpc);
 323         break;
 324     }
 325 }
 326
 327 static const uint8_t wtype_to_wsize[] = { 0, 0, 2, 2, 2, 1, 0, 1, 1 };
 328
 329 static void imdct_and_window(TwinVQContext *tctx, enum TwinVQFrameType ftype,
 330                              int wtype, float *in, float *prev, int ch)
 331 {
 332     AVTXContext *tx = tctx->tx[ftype];
 333     av_tx_fn tx_fn = tctx->tx_fn[ftype];
 334     const TwinVQModeTab *mtab = tctx->mtab;
 335     int bsize = mtab->size / mtab->fmode[ftype].sub;
 336     int size  = mtab->size;
 337     float *buf1 = tctx->tmp_buf;
 338     int j, first_wsize, wsize; // Window size
 339     float *out  = tctx->curr_frame + 2 * ch * mtab->size;
 340     float *out2 = out;
 341     float *prev_buf;
 342     int types_sizes[] = {
 343         mtab->size /  mtab->fmode[TWINVQ_FT_LONG].sub,
 344         mtab->size /  mtab->fmode[TWINVQ_FT_MEDIUM].sub,
 345         mtab->size / (mtab->fmode[TWINVQ_FT_SHORT].sub * 2),
 346     };
 347
 348     wsize       = types_sizes[wtype_to_wsize[wtype]];
 349     first_wsize = wsize;
 350     prev_buf    = prev + (size - bsize) / 2;
 351
 352     for (j = 0; j < mtab->fmode[ftype].sub; j++) {
 353         int sub_wtype = ftype == TWINVQ_FT_MEDIUM ? 8 : wtype;
 354
 355         if (!j && wtype == 4)
 356             sub_wtype = 4;
 357         else if (j == mtab->fmode[ftype].sub - 1 && wtype == 7)
 358             sub_wtype = 7;
 359
 360         wsize = types_sizes[wtype_to_wsize[sub_wtype]];
 361
 362         tx_fn(tx, buf1 + bsize * j, in + bsize * j, sizeof(float));
 363
 364         tctx->fdsp->vector_fmul_window(out2, prev_buf + (bsize - wsize) / 2,
 365                                       buf1 + bsize * j,
 366                                       ff_sine_windows[av_log2(wsize)],
 367                                       wsize / 2);
 368         out2 += wsize;
 369
 370         memcpy(out2, buf1 + bsize * j + wsize / 2,
 371                (bsize - wsize / 2) * sizeof(float));
 372
 373         out2 += ftype == TWINVQ_FT_MEDIUM ? (bsize - wsize) / 2 : bsize - wsize;
 374
 375         prev_buf = buf1 + bsize * j + bsize / 2;
 376     }
 377
 378     tctx->last_block_pos[ch] = (size + first_wsize) / 2;
 379 }
 380
 381 static void imdct_output(TwinVQContext *tctx, enum TwinVQFrameType ftype,
 382                          int wtype, float **out, int offset)
 383 {
 384     const TwinVQModeTab *mtab = tctx->mtab;
 385     float *prev_buf           = tctx->prev_frame + tctx->last_block_pos[0];
 386     int channels              = tctx->avctx->ch_layout.nb_channels;
 387     int size1, size2, i;
 388     float *out1, *out2;
 389
 390     for (i = 0; i < channels; i++)
 391         imdct_and_window(tctx, ftype, wtype,
 392                          tctx->spectrum + i * mtab->size,
 393                          prev_buf + 2 * i * mtab->size,
 394                          i);
 395
 396     if (!out)
 397         return;
 398
 399     size2 = tctx->last_block_pos[0];
 400     size1 = mtab->size - size2;
 401
 402     out1 = &out[0][0] + offset;
 403     memcpy(out1,         prev_buf,         size1 * sizeof(*out1));
 404     memcpy(out1 + size1, tctx->curr_frame, size2 * sizeof(*out1));
 405
 406     if (channels == 2) {
 407         out2 = &out[1][0] + offset;
 408         memcpy(out2, &prev_buf[2 * mtab->size],
 409                size1 * sizeof(*out2));
 410         memcpy(out2 + size1, &tctx->curr_frame[2 * mtab->size],
 411                size2 * sizeof(*out2));
 412         tctx->fdsp->butterflies_float(out1, out2, mtab->size);
 413     }
 414 }
 415
 416 static void read_and_decode_spectrum(TwinVQContext *tctx, float *out,
 417                                      enum TwinVQFrameType ftype)
 418 {
 419     const TwinVQModeTab *mtab = tctx->mtab;
 420     TwinVQFrameData *bits     = &tctx->bits[tctx->cur_frame];
 421     int channels              = tctx->avctx->ch_layout.nb_channels;
 422     int sub        = mtab->fmode[ftype].sub;
 423     int block_size = mtab->size / sub;
 424     float gain[TWINVQ_CHANNELS_MAX * TWINVQ_SUBBLOCKS_MAX];
 425     float ppc_shape[TWINVQ_PPC_SHAPE_LEN_MAX * TWINVQ_CHANNELS_MAX * 4];
 426
 427     int i, j;
 428
 429     dequant(tctx, bits->main_coeffs, out, ftype,
 430             mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
 431             mtab->fmode[ftype].cb_len_read);
 432
 433     dec_gain(tctx, ftype, gain);
 434
 435     if (ftype == TWINVQ_FT_LONG) {
 436         int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len * channels - 1) /
 437                        tctx->n_div[3];
 438         dequant(tctx, bits->ppc_coeffs, ppc_shape,
 439                 TWINVQ_FT_PPC, mtab->ppc_shape_cb,
 440                 mtab->ppc_shape_cb + cb_len_p * TWINVQ_PPC_SHAPE_CB_SIZE,
 441                 cb_len_p);
 442     }
 443
 444     for (i = 0; i < channels; i++) {
 445         float *chunk = out + mtab->size * i;
 446         float lsp[TWINVQ_LSP_COEFS_MAX];
 447
 448         for (j = 0; j < sub; j++) {
 449             tctx->dec_bark_env(tctx, bits->bark1[i][j],
 450                                bits->bark_use_hist[i][j], i,
 451                                tctx->tmp_buf, gain[sub * i + j], ftype);
 452
 453             tctx->fdsp->vector_fmul(chunk + block_size * j,
 454                                    chunk + block_size * j,
 455                                    tctx->tmp_buf, block_size);
 456         }
 457
 458         if (ftype == TWINVQ_FT_LONG)
 459             tctx->decode_ppc(tctx, bits->p_coef[i], bits->g_coef[i],
 460                              ppc_shape + i * mtab->ppc_shape_len, chunk);
 461
 462         decode_lsp(tctx, bits->lpc_idx1[i], bits->lpc_idx2[i],
 463                    bits->lpc_hist_idx[i], lsp, tctx->lsp_hist[i]);
 464
 465         dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
 466
 467         for (j = 0; j < mtab->fmode[ftype].sub; j++) {
 468             tctx->fdsp->vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
 469             chunk += block_size;
 470         }
 471     }
 472 }
 473
 474 const enum TwinVQFrameType ff_twinvq_wtype_to_ftype_table[] = {
 475     TWINVQ_FT_LONG,   TWINVQ_FT_LONG, TWINVQ_FT_SHORT, TWINVQ_FT_LONG,
 476     TWINVQ_FT_MEDIUM, TWINVQ_FT_LONG, TWINVQ_FT_LONG,  TWINVQ_FT_MEDIUM,
 477     TWINVQ_FT_MEDIUM
 478 };
 479
 480 int ff_twinvq_decode_frame(AVCodecContext *avctx, AVFrame *frame,
 481                            int *got_frame_ptr, AVPacket *avpkt)
 482 {
 483     const uint8_t *buf = avpkt->data;
 484     int buf_size       = avpkt->size;
 485     TwinVQContext *tctx = avctx->priv_data;
 486     const TwinVQModeTab *mtab = tctx->mtab;
 487     float **out = NULL;
 488     int ret;
 489
 490     /* get output buffer */
 491     if (tctx->discarded_packets >= 2) {
 492         frame->nb_samples = mtab->size * tctx->frames_per_packet;
 493         if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
 494             return ret;
 495         out = (float **)frame->extended_data;
 496     }
 497
 498     if (buf_size < avctx->block_align) {
 499         av_log(avctx, AV_LOG_ERROR,
 500                "Frame too small (%d bytes). Truncated file?\n", buf_size);
 501         return AVERROR(EINVAL);
 502     }
 503
 504     if ((ret = tctx->read_bitstream(avctx, tctx, buf, buf_size)) < 0)
 505         return ret;
 506
 507     for (tctx->cur_frame = 0; tctx->cur_frame < tctx->frames_per_packet;
 508          tctx->cur_frame++) {
 509         read_and_decode_spectrum(tctx, tctx->spectrum,
 510                                  tctx->bits[tctx->cur_frame].ftype);
 511
 512         imdct_output(tctx, tctx->bits[tctx->cur_frame].ftype,
 513                      tctx->bits[tctx->cur_frame].window_type, out,
 514                      tctx->cur_frame * mtab->size);
 515
 516         FFSWAP(float *, tctx->curr_frame, tctx->prev_frame);
 517     }
 518
 519     if (tctx->discarded_packets < 2) {
 520         tctx->discarded_packets++;
 521         *got_frame_ptr = 0;
 522         return buf_size;
 523     }
 524
 525     *got_frame_ptr = 1;
 526
 527     // VQF can deliver packets 1 byte greater than block align
 528     if (buf_size == avctx->block_align + 1)
 529         return buf_size;
 530     return avctx->block_align;
 531 }
 532
 533 /**
 534  * Init IMDCT and windowing tables
 535  */
 536 static av_cold int init_mdct_win(TwinVQContext *tctx)
 537 {
 538     int i, j, ret;
 539     const TwinVQModeTab *mtab = tctx->mtab;
 540     int size_s = mtab->size / mtab->fmode[TWINVQ_FT_SHORT].sub;
 541     int size_m = mtab->size / mtab->fmode[TWINVQ_FT_MEDIUM].sub;
 542     int channels = tctx->avctx->ch_layout.nb_channels;
 543     float norm = channels == 1 ? 2.0 : 1.0;
 544     int table_size = 2 * mtab->size * channels;
 545
 546     for (i = 0; i < 3; i++) {
 547         int bsize = tctx->mtab->size / tctx->mtab->fmode[i].sub;
 548         const float scale = -sqrt(norm / bsize) / (1 << 15);
 549         if ((ret = av_tx_init(&tctx->tx[i], &tctx->tx_fn[i], AV_TX_FLOAT_MDCT,
 550                               1, bsize, &scale, 0)))
 551             return ret;
 552     }
 553
 554     if (!FF_ALLOC_TYPED_ARRAY(tctx->tmp_buf,    mtab->size) ||
 555         !FF_ALLOC_TYPED_ARRAY(tctx->spectrum,   table_size) ||
 556         !FF_ALLOC_TYPED_ARRAY(tctx->curr_frame, table_size) ||
 557         !FF_ALLOC_TYPED_ARRAY(tctx->prev_frame, table_size))
 558         return AVERROR(ENOMEM);
 559
 560     for (i = 0; i < 3; i++) {
 561         int m       = 4 * mtab->size / mtab->fmode[i].sub;
 562         double freq = 2 * M_PI / m;
 563         if (!FF_ALLOC_TYPED_ARRAY(tctx->cos_tabs[i], m / 4))
 564             return AVERROR(ENOMEM);
 565         for (j = 0; j <= m / 8; j++)
 566             tctx->cos_tabs[i][j] = cos((2 * j + 1) * freq);
 567         for (j = 1; j < m / 8; j++)
 568             tctx->cos_tabs[i][m / 4 - j] = tctx->cos_tabs[i][j];
 569     }
 570
 571     ff_init_ff_sine_windows(av_log2(size_m));
 572     ff_init_ff_sine_windows(av_log2(size_s / 2));
 573     ff_init_ff_sine_windows(av_log2(mtab->size));
 574
 575     return 0;
 576 }
 577
 578 /**
 579  * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
 580  * each line do a cyclic permutation, i.e.
 581  * abcdefghijklm -> defghijklmabc
 582  * where the amount to be shifted is evaluated depending on the column.
 583  */
 584 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
 585                               int block_size,
 586                               const uint8_t line_len[2], int length_div,
 587                               enum TwinVQFrameType ftype)
 588 {
 589     int i, j;
 590
 591     for (i = 0; i < line_len[0]; i++) {
 592         int shift;
 593
 594         if (num_blocks == 1                                    ||
 595             (ftype == TWINVQ_FT_LONG && num_vect % num_blocks) ||
 596             (ftype != TWINVQ_FT_LONG && num_vect & 1)          ||
 597             i == line_len[1]) {
 598             shift = 0;
 599         } else if (ftype == TWINVQ_FT_LONG) {
 600             shift = i;
 601         } else
 602             shift = i * i;
 603
 604         for (j = 0; j < num_vect && (j + num_vect * i < block_size * num_blocks); j++)
 605             tab[i * num_vect + j] = i * num_vect + (j + shift) % num_vect;
 606     }
 607 }
 608
 609 /**
 610  * Interpret the input data as in the following table:
 611  *
 612  * @verbatim
 613  *
 614  * abcdefgh
 615  * ijklmnop
 616  * qrstuvw
 617  * x123456
 618  *
 619  * @endverbatim
 620  *
 621  * and transpose it, giving the output
 622  * aiqxbjr1cks2dlt3emu4fvn5gow6hp
 623  */
 624 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
 625                            const uint8_t line_len[2], int length_div)
 626 {
 627     int i, j;
 628     int cont = 0;
 629
 630     for (i = 0; i < num_vect; i++)
 631         for (j = 0; j < line_len[i >= length_div]; j++)
 632             out[cont++] = in[j * num_vect + i];
 633 }
 634
 635 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
 636 {
 637     int block_size = size / n_blocks;
 638     int i;
 639
 640     for (i = 0; i < size; i++)
 641         out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
 642 }
 643
 644 static av_cold void construct_perm_table(TwinVQContext *tctx,
 645                                          enum TwinVQFrameType ftype)
 646 {
 647     int block_size, size;
 648     const TwinVQModeTab *mtab = tctx->mtab;
 649     int16_t *tmp_perm = (int16_t *)tctx->tmp_buf;
 650
 651     if (ftype == TWINVQ_FT_PPC) {
 652         size       = tctx->avctx->ch_layout.nb_channels;
 653         block_size = mtab->ppc_shape_len;
 654     } else {
 655         size       = tctx->avctx->ch_layout.nb_channels * mtab->fmode[ftype].sub;
 656         block_size = mtab->size / mtab->fmode[ftype].sub;
 657     }
 658
 659     permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
 660                       block_size, tctx->length[ftype],
 661                       tctx->length_change[ftype], ftype);
 662
 663     transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
 664                    tctx->length[ftype], tctx->length_change[ftype]);
 665
 666     linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
 667                 size * block_size);
 668 }
 669
 670 static av_cold void init_bitstream_params(TwinVQContext *tctx)
 671 {
 672     const TwinVQModeTab *mtab = tctx->mtab;
 673     int n_ch                  = tctx->avctx->ch_layout.nb_channels;
 674     int total_fr_bits         = tctx->avctx->bit_rate * mtab->size /
 675                                 tctx->avctx->sample_rate;
 676
 677     int lsp_bits_per_block = n_ch * (mtab->lsp_bit0 + mtab->lsp_bit1 +
 678                                      mtab->lsp_split * mtab->lsp_bit2);
 679
 680     int ppc_bits = n_ch * (mtab->pgain_bit + mtab->ppc_shape_bit +
 681                            mtab->ppc_period_bit);
 682
 683     int bsize_no_main_cb[3], bse_bits[3], i;
 684     enum TwinVQFrameType frametype;
 685
 686     for (i = 0; i < 3; i++)
 687         // +1 for history usage switch
 688         bse_bits[i] = n_ch *
 689                       (mtab->fmode[i].bark_n_coef *
 690                        mtab->fmode[i].bark_n_bit + 1);
 691
 692     bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
 693                           TWINVQ_WINDOW_TYPE_BITS + n_ch * TWINVQ_GAIN_BITS;
 694
 695     for (i = 0; i < 2; i++)
 696         bsize_no_main_cb[i] =
 697             lsp_bits_per_block + n_ch * TWINVQ_GAIN_BITS +
 698             TWINVQ_WINDOW_TYPE_BITS +
 699             mtab->fmode[i].sub * (bse_bits[i] + n_ch * TWINVQ_SUB_GAIN_BITS);
 700
 701     if (tctx->codec == TWINVQ_CODEC_METASOUND && !tctx->is_6kbps) {
 702         bsize_no_main_cb[1] += 2;
 703         bsize_no_main_cb[2] += 2;
 704     }
 705
 706     // The remaining bits are all used for the main spectrum coefficients
 707     for (i = 0; i < 4; i++) {
 708         int bit_size, vect_size;
 709         int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
 710         if (i == 3) {
 711             bit_size  = n_ch * mtab->ppc_shape_bit;
 712             vect_size = n_ch * mtab->ppc_shape_len;
 713         } else {
 714             bit_size  = total_fr_bits - bsize_no_main_cb[i];
 715             vect_size = n_ch * mtab->size;
 716         }
 717
 718         tctx->n_div[i] = (bit_size + 13) / 14;
 719
 720         rounded_up                     = (bit_size + tctx->n_div[i] - 1) /
 721                                          tctx->n_div[i];
 722         rounded_down                   = (bit_size) / tctx->n_div[i];
 723         num_rounded_down               = rounded_up * tctx->n_div[i] - bit_size;
 724         num_rounded_up                 = tctx->n_div[i] - num_rounded_down;
 725         tctx->bits_main_spec[0][i][0]  = (rounded_up + 1)   / 2;
 726         tctx->bits_main_spec[1][i][0]  =  rounded_up        / 2;
 727         tctx->bits_main_spec[0][i][1]  = (rounded_down + 1) / 2;
 728         tctx->bits_main_spec[1][i][1]  =  rounded_down      / 2;
 729         tctx->bits_main_spec_change[i] = num_rounded_up;
 730
 731         rounded_up             = (vect_size + tctx->n_div[i] - 1) /
 732                                  tctx->n_div[i];
 733         rounded_down           = (vect_size) / tctx->n_div[i];
 734         num_rounded_down       = rounded_up * tctx->n_div[i] - vect_size;
 735         num_rounded_up         = tctx->n_div[i] - num_rounded_down;
 736         tctx->length[i][0]     = rounded_up;
 737         tctx->length[i][1]     = rounded_down;
 738         tctx->length_change[i] = num_rounded_up;
 739     }
 740
 741     for (frametype = TWINVQ_FT_SHORT; frametype <= TWINVQ_FT_PPC; frametype++)
 742         construct_perm_table(tctx, frametype);
 743 }
 744
 745 av_cold int ff_twinvq_decode_close(AVCodecContext *avctx)
 746 {
 747     TwinVQContext *tctx = avctx->priv_data;
 748     int i;
 749
 750     for (i = 0; i < 3; i++) {
 751         av_tx_uninit(&tctx->tx[i]);
 752         av_freep(&tctx->cos_tabs[i]);
 753     }
 754
 755     av_freep(&tctx->curr_frame);
 756     av_freep(&tctx->spectrum);
 757     av_freep(&tctx->prev_frame);
 758     av_freep(&tctx->tmp_buf);
 759     av_freep(&tctx->fdsp);
 760
 761     return 0;
 762 }
 763
 764 av_cold int ff_twinvq_decode_init(AVCodecContext *avctx)
 765 {
 766     int ret;
 767     TwinVQContext *tctx = avctx->priv_data;
 768     int64_t frames_per_packet;
 769
 770     tctx->avctx       = avctx;
 771     avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
 772
 773     if (!avctx->block_align) {
 774         avctx->block_align = tctx->frame_size + 7 >> 3;
 775     }
 776     frames_per_packet = avctx->block_align * 8LL / tctx->frame_size;
 777     if (frames_per_packet <= 0) {
 778         av_log(avctx, AV_LOG_ERROR, "Block align is %"PRId64" bits, expected %d\n",
 779                avctx->block_align * (int64_t)8, tctx->frame_size);
 780         return AVERROR_INVALIDDATA;
 781     }
 782     if (frames_per_packet > TWINVQ_MAX_FRAMES_PER_PACKET) {
 783         av_log(avctx, AV_LOG_ERROR, "Too many frames per packet (%"PRId64")\n",
 784                frames_per_packet);
 785         return AVERROR_INVALIDDATA;
 786     }
 787     tctx->frames_per_packet = frames_per_packet;
 788
 789     tctx->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
 790     if (!tctx->fdsp)
 791         return AVERROR(ENOMEM);
 792     if ((ret = init_mdct_win(tctx))) {
 793         av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
 794         return ret;
 795     }
 796     init_bitstream_params(tctx);
 797
 798     twinvq_memset_float(tctx->bark_hist[0][0], 0.1,
 799                         FF_ARRAY_ELEMS(tctx->bark_hist));
 800
 801     return 0;
 802 }