git.sesse.net Git - ffmpeg/blob - libavcodec/aaccoder.c

   1 /*
   2  * AAC coefficients encoder
   3  * Copyright (C) 2008-2009 Konstantin Shishkov
   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 /**
  23  * @file
  24  * AAC coefficients encoder
  25  */
  26
  27 /***********************************
  28  *              TODOs:
  29  * speedup quantizer selection
  30  * add sane pulse detection
  31  ***********************************/
  32
  33 #include "avcodec.h"
  34 #include "put_bits.h"
  35 #include "aac.h"
  36 #include "aacenc.h"
  37 #include "aactab.h"
  38
  39 /** bits needed to code codebook run value for long windows */
  40 static const uint8_t run_value_bits_long[64] = {
  41      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
  42      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
  43     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
  44     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
  45 };
  46
  47 /** bits needed to code codebook run value for short windows */
  48 static const uint8_t run_value_bits_short[16] = {
  49     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
  50 };
  51
  52 static const uint8_t *run_value_bits[2] = {
  53     run_value_bits_long, run_value_bits_short
  54 };
  55
  56
  57 /**
  58  * Quantize one coefficient.
  59  * @return absolute value of the quantized coefficient
  60  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
  61  */
  62 static av_always_inline int quant(float coef, const float Q)
  63 {
  64     float a = coef * Q;
  65     return sqrtf(a * sqrtf(a)) + 0.4054;
  66 }
  67
  68 static void quantize_bands(int *out, const float *in, const float *scaled,
  69                            int size, float Q34, int is_signed, int maxval)
  70 {
  71     int i;
  72     double qc;
  73     for (i = 0; i < size; i++) {
  74         qc = scaled[i] * Q34;
  75         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
  76         if (is_signed && in[i] < 0.0f) {
  77             out[i] = -out[i];
  78         }
  79     }
  80 }
  81
  82 static void abs_pow34_v(float *out, const float *in, const int size)
  83 {
  84 #ifndef USE_REALLY_FULL_SEARCH
  85     int i;
  86     for (i = 0; i < size; i++) {
  87         float a = fabsf(in[i]);
  88         out[i] = sqrtf(a * sqrtf(a));
  89     }
  90 #endif /* USE_REALLY_FULL_SEARCH */
  91 }
  92
  93 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
  94 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
  95
  96 /**
  97  * Calculate rate distortion cost for quantizing with given codebook
  98  *
  99  * @return quantization distortion
 100  */
 101 static float quantize_and_encode_band_cost(struct AACEncContext *s,
 102                                 PutBitContext *pb, const float *in,
 103                                 const float *scaled, int size, int scale_idx,
 104                                 int cb, const float lambda, const float uplim,
 105                                 int *bits)
 106 {
 107     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
 108     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
 109     const float CLIPPED_ESCAPE = 165140.0f*IQ;
 110     int i, j, k;
 111     float cost = 0;
 112     const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
 113     int resbits = 0;
 114     const float  Q34 = sqrtf(Q * sqrtf(Q));
 115     const int range  = aac_cb_range[cb];
 116     const int maxval = aac_cb_maxval[cb];
 117     int off;
 118
 119     if (!cb) {
 120         for (i = 0; i < size; i++)
 121             cost += in[i]*in[i];
 122         if (bits)
 123             *bits = 0;
 124         return cost * lambda;
 125     }
 126     if (!scaled) {
 127         abs_pow34_v(s->scoefs, in, size);
 128         scaled = s->scoefs;
 129     }
 130     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
 131     if (IS_CODEBOOK_UNSIGNED(cb)) {
 132         off = 0;
 133     } else {
 134         off = maxval;
 135     }
 136     for (i = 0; i < size; i += dim) {
 137         const float *vec;
 138         int *quants = s->qcoefs + i;
 139         int curidx = 0;
 140         int curbits;
 141         float rd = 0.0f;
 142         for (j = 0; j < dim; j++) {
 143             curidx *= range;
 144             curidx += quants[j] + off;
 145         }
 146             curbits =  ff_aac_spectral_bits[cb-1][curidx];
 147             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
 148             if (IS_CODEBOOK_UNSIGNED(cb)) {
 149                 for (k = 0; k < dim; k++) {
 150                     float t = fabsf(in[i+k]);
 151                     float di;
 152                     if (vec[k] == 64.0f) { //FIXME: slow
 153                         if (t >= CLIPPED_ESCAPE) {
 154                             di = t - CLIPPED_ESCAPE;
 155                             curbits += 21;
 156                         } else {
 157                             int c = av_clip(quant(t, Q), 0, 8191);
 158                             di = t - c*cbrtf(c)*IQ;
 159                             curbits += av_log2(c)*2 - 4 + 1;
 160                         }
 161                     } else {
 162                         di = t - vec[k]*IQ;
 163                     }
 164                     if (vec[k] != 0.0f)
 165                         curbits++;
 166                     rd += di*di;
 167                 }
 168             } else {
 169                 for (k = 0; k < dim; k++) {
 170                     float di = in[i+k] - vec[k]*IQ;
 171                     rd += di*di;
 172                 }
 173             }
 174         cost    += rd * lambda + curbits;
 175         resbits += curbits;
 176         if (cost >= uplim)
 177             return uplim;
 178         if (pb) {
 179         put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
 180         if (IS_CODEBOOK_UNSIGNED(cb))
 181             for (j = 0; j < dim; j++)
 182                 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
 183                     put_bits(pb, 1, in[i+j] < 0.0f);
 184         if (cb == ESC_BT) {
 185             for (j = 0; j < 2; j++) {
 186                 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
 187                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
 188                     int len = av_log2(coef);
 189
 190                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
 191                     put_bits(pb, len, coef & ((1 << len) - 1));
 192                 }
 193             }
 194         }
 195         }
 196     }
 197
 198     if (bits)
 199         *bits = resbits;
 200     return cost;
 201 }
 202 static float quantize_band_cost(struct AACEncContext *s, const float *in,
 203                                 const float *scaled, int size, int scale_idx,
 204                                 int cb, const float lambda, const float uplim,
 205                                 int *bits)
 206 {
 207     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
 208                                          cb, lambda, uplim, bits);
 209 }
 210
 211 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
 212                                      const float *in, int size, int scale_idx,
 213                                      int cb, const float lambda)
 214 {
 215     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
 216                                   INFINITY, NULL);
 217 }
 218
 219 /**
 220  * structure used in optimal codebook search
 221  */
 222 typedef struct BandCodingPath {
 223     int prev_idx; ///< pointer to the previous path point
 224     float cost;   ///< path cost
 225     int run;
 226 } BandCodingPath;
 227
 228 /**
 229  * Encode band info for single window group bands.
 230  */
 231 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
 232                                      int win, int group_len, const float lambda)
 233 {
 234     BandCodingPath path[120][12];
 235     int w, swb, cb, start, start2, size;
 236     int i, j;
 237     const int max_sfb  = sce->ics.max_sfb;
 238     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 239     const int run_esc  = (1 << run_bits) - 1;
 240     int idx, ppos, count;
 241     int stackrun[120], stackcb[120], stack_len;
 242     float next_minrd = INFINITY;
 243     int next_mincb = 0;
 244
 245     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 246     start = win*128;
 247     for (cb = 0; cb < 12; cb++) {
 248         path[0][cb].cost     = 0.0f;
 249         path[0][cb].prev_idx = -1;
 250         path[0][cb].run      = 0;
 251     }
 252     for (swb = 0; swb < max_sfb; swb++) {
 253         start2 = start;
 254         size = sce->ics.swb_sizes[swb];
 255         if (sce->zeroes[win*16 + swb]) {
 256             for (cb = 0; cb < 12; cb++) {
 257                 path[swb+1][cb].prev_idx = cb;
 258                 path[swb+1][cb].cost     = path[swb][cb].cost;
 259                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 260             }
 261         } else {
 262             float minrd = next_minrd;
 263             int mincb = next_mincb;
 264             next_minrd = INFINITY;
 265             next_mincb = 0;
 266             for (cb = 0; cb < 12; cb++) {
 267                 float cost_stay_here, cost_get_here;
 268                 float rd = 0.0f;
 269                 for (w = 0; w < group_len; w++) {
 270                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
 271                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 272                                              s->scoefs + start + w*128, size,
 273                                              sce->sf_idx[(win+w)*16+swb], cb,
 274                                              lambda / band->threshold, INFINITY, NULL);
 275                 }
 276                 cost_stay_here = path[swb][cb].cost + rd;
 277                 cost_get_here  = minrd              + rd + run_bits + 4;
 278                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 279                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 280                     cost_stay_here += run_bits;
 281                 if (cost_get_here < cost_stay_here) {
 282                     path[swb+1][cb].prev_idx = mincb;
 283                     path[swb+1][cb].cost     = cost_get_here;
 284                     path[swb+1][cb].run      = 1;
 285                 } else {
 286                     path[swb+1][cb].prev_idx = cb;
 287                     path[swb+1][cb].cost     = cost_stay_here;
 288                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 289                 }
 290                 if (path[swb+1][cb].cost < next_minrd) {
 291                     next_minrd = path[swb+1][cb].cost;
 292                     next_mincb = cb;
 293                 }
 294             }
 295         }
 296         start += sce->ics.swb_sizes[swb];
 297     }
 298
 299     //convert resulting path from backward-linked list
 300     stack_len = 0;
 301     idx       = 0;
 302     for (cb = 1; cb < 12; cb++)
 303         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 304             idx = cb;
 305     ppos = max_sfb;
 306     while (ppos > 0) {
 307         cb = idx;
 308         stackrun[stack_len] = path[ppos][cb].run;
 309         stackcb [stack_len] = cb;
 310         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 311         ppos -= path[ppos][cb].run;
 312         stack_len++;
 313     }
 314     //perform actual band info encoding
 315     start = 0;
 316     for (i = stack_len - 1; i >= 0; i--) {
 317         put_bits(&s->pb, 4, stackcb[i]);
 318         count = stackrun[i];
 319         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
 320         //XXX: memset when band_type is also uint8_t
 321         for (j = 0; j < count; j++) {
 322             sce->band_type[win*16 + start] =  stackcb[i];
 323             start++;
 324         }
 325         while (count >= run_esc) {
 326             put_bits(&s->pb, run_bits, run_esc);
 327             count -= run_esc;
 328         }
 329         put_bits(&s->pb, run_bits, count);
 330     }
 331 }
 332
 333 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
 334                                   int win, int group_len, const float lambda)
 335 {
 336     BandCodingPath path[120][12];
 337     int w, swb, cb, start, start2, size;
 338     int i, j;
 339     const int max_sfb  = sce->ics.max_sfb;
 340     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 341     const int run_esc  = (1 << run_bits) - 1;
 342     int idx, ppos, count;
 343     int stackrun[120], stackcb[120], stack_len;
 344     float next_minrd = INFINITY;
 345     int next_mincb = 0;
 346
 347     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 348     start = win*128;
 349     for (cb = 0; cb < 12; cb++) {
 350         path[0][cb].cost     = run_bits+4;
 351         path[0][cb].prev_idx = -1;
 352         path[0][cb].run      = 0;
 353     }
 354     for (swb = 0; swb < max_sfb; swb++) {
 355         start2 = start;
 356         size = sce->ics.swb_sizes[swb];
 357         if (sce->zeroes[win*16 + swb]) {
 358             for (cb = 0; cb < 12; cb++) {
 359                 path[swb+1][cb].prev_idx = cb;
 360                 path[swb+1][cb].cost     = path[swb][cb].cost;
 361                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 362             }
 363         } else {
 364             float minrd = next_minrd;
 365             int mincb = next_mincb;
 366             int startcb = sce->band_type[win*16+swb];
 367             next_minrd = INFINITY;
 368             next_mincb = 0;
 369             for (cb = 0; cb < startcb; cb++) {
 370                 path[swb+1][cb].cost = 61450;
 371                 path[swb+1][cb].prev_idx = -1;
 372                 path[swb+1][cb].run = 0;
 373             }
 374             for (cb = startcb; cb < 12; cb++) {
 375                 float cost_stay_here, cost_get_here;
 376                 float rd = 0.0f;
 377                 for (w = 0; w < group_len; w++) {
 378                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 379                                              s->scoefs + start + w*128, size,
 380                                              sce->sf_idx[(win+w)*16+swb], cb,
 381                                              0, INFINITY, NULL);
 382                 }
 383                 cost_stay_here = path[swb][cb].cost + rd;
 384                 cost_get_here  = minrd              + rd + run_bits + 4;
 385                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 386                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 387                     cost_stay_here += run_bits;
 388                 if (cost_get_here < cost_stay_here) {
 389                     path[swb+1][cb].prev_idx = mincb;
 390                     path[swb+1][cb].cost     = cost_get_here;
 391                     path[swb+1][cb].run      = 1;
 392                 } else {
 393                     path[swb+1][cb].prev_idx = cb;
 394                     path[swb+1][cb].cost     = cost_stay_here;
 395                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 396                 }
 397                 if (path[swb+1][cb].cost < next_minrd) {
 398                     next_minrd = path[swb+1][cb].cost;
 399                     next_mincb = cb;
 400                 }
 401             }
 402         }
 403         start += sce->ics.swb_sizes[swb];
 404     }
 405
 406     //convert resulting path from backward-linked list
 407     stack_len = 0;
 408     idx       = 0;
 409     for (cb = 1; cb < 12; cb++)
 410         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 411             idx = cb;
 412     ppos = max_sfb;
 413     while (ppos > 0) {
 414         if (idx < 0) abort();
 415         cb = idx;
 416         stackrun[stack_len] = path[ppos][cb].run;
 417         stackcb [stack_len] = cb;
 418         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 419         ppos -= path[ppos][cb].run;
 420         stack_len++;
 421     }
 422     //perform actual band info encoding
 423     start = 0;
 424     for (i = stack_len - 1; i >= 0; i--) {
 425         put_bits(&s->pb, 4, stackcb[i]);
 426         count = stackrun[i];
 427         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
 428         //XXX: memset when band_type is also uint8_t
 429         for (j = 0; j < count; j++) {
 430             sce->band_type[win*16 + start] =  stackcb[i];
 431             start++;
 432         }
 433         while (count >= run_esc) {
 434             put_bits(&s->pb, run_bits, run_esc);
 435             count -= run_esc;
 436         }
 437         put_bits(&s->pb, run_bits, count);
 438     }
 439 }
 440
 441 typedef struct TrellisPath {
 442     float cost;
 443     int prev;
 444     int min_val;
 445     int max_val;
 446 } TrellisPath;
 447
 448 #define TRELLIS_STAGES 121
 449 #define TRELLIS_STATES 256
 450
 451 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
 452                                        SingleChannelElement *sce,
 453                                        const float lambda)
 454 {
 455     int q, w, w2, g, start = 0;
 456     int i, j;
 457     int idx;
 458     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
 459     int bandaddr[TRELLIS_STAGES];
 460     int minq;
 461     float mincost;
 462
 463     for (i = 0; i < TRELLIS_STATES; i++) {
 464         paths[0][i].cost    = 0.0f;
 465         paths[0][i].prev    = -1;
 466         paths[0][i].min_val = i;
 467         paths[0][i].max_val = i;
 468     }
 469     for (j = 1; j < TRELLIS_STAGES; j++) {
 470         for (i = 0; i < TRELLIS_STATES; i++) {
 471             paths[j][i].cost    = INFINITY;
 472             paths[j][i].prev    = -2;
 473             paths[j][i].min_val = INT_MAX;
 474             paths[j][i].max_val = 0;
 475         }
 476     }
 477     idx = 1;
 478     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 479     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 480         start = w*128;
 481         for (g = 0; g < sce->ics.num_swb; g++) {
 482             const float *coefs = sce->coeffs + start;
 483             float qmin, qmax;
 484             int nz = 0;
 485
 486             bandaddr[idx] = w * 16 + g;
 487             qmin = INT_MAX;
 488             qmax = 0.0f;
 489             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 490                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 491                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 492                     sce->zeroes[(w+w2)*16+g] = 1;
 493                     continue;
 494                 }
 495                 sce->zeroes[(w+w2)*16+g] = 0;
 496                 nz = 1;
 497                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
 498                     float t = fabsf(coefs[w2*128+i]);
 499                     if (t > 0.0f)
 500                         qmin = FFMIN(qmin, t);
 501                     qmax = FFMAX(qmax, t);
 502                 }
 503             }
 504             if (nz) {
 505                 int minscale, maxscale;
 506                 float minrd = INFINITY;
 507                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 508                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 509                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 510                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 511                 for (q = minscale; q < maxscale; q++) {
 512                     float dists[12], dist;
 513                     memset(dists, 0, sizeof(dists));
 514                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 515                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 516                         int cb;
 517                         for (cb = 0; cb <= ESC_BT; cb++)
 518                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
 519                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
 520                     }
 521                     dist = dists[0];
 522                     for (i = 1; i <= ESC_BT; i++)
 523                         dist = FFMIN(dist, dists[i]);
 524                     minrd = FFMIN(minrd, dist);
 525
 526                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
 527                         float cost;
 528                         int minv, maxv;
 529                         if (isinf(paths[idx - 1][i].cost))
 530                             continue;
 531                         cost = paths[idx - 1][i].cost + dist
 532                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 533                         minv = FFMIN(paths[idx - 1][i].min_val, q);
 534                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
 535                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
 536                             paths[idx][q].cost    = cost;
 537                             paths[idx][q].prev    = i;
 538                             paths[idx][q].min_val = minv;
 539                             paths[idx][q].max_val = maxv;
 540                         }
 541                     }
 542                 }
 543             } else {
 544                 for (q = 0; q < TRELLIS_STATES; q++) {
 545                     if (!isinf(paths[idx - 1][q].cost)) {
 546                         paths[idx][q].cost = paths[idx - 1][q].cost + 1;
 547                         paths[idx][q].prev = q;
 548                         paths[idx][q].min_val = FFMIN(paths[idx - 1][q].min_val, q);
 549                         paths[idx][q].max_val = FFMAX(paths[idx - 1][q].max_val, q);
 550                         continue;
 551                     }
 552                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
 553                         float cost;
 554                         int minv, maxv;
 555                         if (isinf(paths[idx - 1][i].cost))
 556                             continue;
 557                         cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 558                         minv = FFMIN(paths[idx - 1][i].min_val, q);
 559                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
 560                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
 561                             paths[idx][q].cost    = cost;
 562                             paths[idx][q].prev    = i;
 563                             paths[idx][q].min_val = minv;
 564                             paths[idx][q].max_val = maxv;
 565                         }
 566                     }
 567                 }
 568             }
 569             sce->zeroes[w*16+g] = !nz;
 570             start += sce->ics.swb_sizes[g];
 571             idx++;
 572         }
 573     }
 574     idx--;
 575     mincost = paths[idx][0].cost;
 576     minq    = 0;
 577     for (i = 1; i < TRELLIS_STATES; i++) {
 578         if (paths[idx][i].cost < mincost) {
 579             mincost = paths[idx][i].cost;
 580             minq = i;
 581         }
 582     }
 583     while (idx) {
 584         sce->sf_idx[bandaddr[idx]] = minq;
 585         minq = paths[idx][minq].prev;
 586         idx--;
 587     }
 588     //set the same quantizers inside window groups
 589     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 590         for (g = 0;  g < sce->ics.num_swb; g++)
 591             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 592                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 593 }
 594
 595 /**
 596  * two-loop quantizers search taken from ISO 13818-7 Appendix C
 597  */
 598 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
 599                                           AACEncContext *s,
 600                                           SingleChannelElement *sce,
 601                                           const float lambda)
 602 {
 603     int start = 0, i, w, w2, g;
 604     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
 605     float dists[128], uplims[128];
 606     int fflag, minscaler;
 607     int its  = 0;
 608     int allz = 0;
 609     float minthr = INFINITY;
 610
 611     //XXX: some heuristic to determine initial quantizers will reduce search time
 612     memset(dists, 0, sizeof(dists));
 613     //determine zero bands and upper limits
 614     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 615         for (g = 0;  g < sce->ics.num_swb; g++) {
 616             int nz = 0;
 617             float uplim = 0.0f;
 618             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 619                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 620                 uplim += band->threshold;
 621                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 622                     sce->zeroes[(w+w2)*16+g] = 1;
 623                     continue;
 624                 }
 625                 nz = 1;
 626             }
 627             uplims[w*16+g] = uplim *512;
 628             sce->zeroes[w*16+g] = !nz;
 629             if (nz)
 630                 minthr = FFMIN(minthr, uplim);
 631             allz = FFMAX(allz, nz);
 632         }
 633     }
 634     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 635         for (g = 0;  g < sce->ics.num_swb; g++) {
 636             if (sce->zeroes[w*16+g]) {
 637                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
 638                 continue;
 639             }
 640             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
 641         }
 642     }
 643
 644     if (!allz)
 645         return;
 646     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 647     //perform two-loop search
 648     //outer loop - improve quality
 649     do {
 650         int tbits, qstep;
 651         minscaler = sce->sf_idx[0];
 652         //inner loop - quantize spectrum to fit into given number of bits
 653         qstep = its ? 1 : 32;
 654         do {
 655             int prev = -1;
 656             tbits = 0;
 657             fflag = 0;
 658             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 659                 start = w*128;
 660                 for (g = 0;  g < sce->ics.num_swb; g++) {
 661                     const float *coefs = sce->coeffs + start;
 662                     const float *scaled = s->scoefs + start;
 663                     int bits = 0;
 664                     int cb;
 665                     float mindist = INFINITY;
 666                     int minbits = 0;
 667
 668                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
 669                         start += sce->ics.swb_sizes[g];
 670                         continue;
 671                     }
 672                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
 673                     {
 674                         float dist = 0.0f;
 675                         int bb = 0;
 676                         float maxval = 0.0f;
 677                         float Q = ff_aac_pow2sf_tab[200 - sce->sf_idx[w*16+g] + SCALE_ONE_POS - SCALE_DIV_512];
 678                         float Q34 = sqrtf(Q * sqrtf(Q));
 679                         int qmaxval;
 680                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 681                             for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
 682                                 maxval = FFMAX(maxval, scaled[w2*128+i]);
 683                             }
 684                         }
 685                         qmaxval = maxval * Q34 + 0.4054;
 686                         if      (qmaxval ==  0) cb = 0;
 687                         else if (qmaxval ==  1) cb = 1;
 688                         else if (qmaxval ==  2) cb = 3;
 689                         else if (qmaxval <=  4) cb = 5;
 690                         else if (qmaxval <=  7) cb = 7;
 691                         else if (qmaxval <= 12) cb = 9;
 692                         else                    cb = 11;
 693                         sce->band_type[w*16+g] = cb;
 694                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 695                             int b;
 696                             dist += quantize_band_cost(s, coefs + w2*128,
 697                                                        scaled + w2*128,
 698                                                        sce->ics.swb_sizes[g],
 699                                                        sce->sf_idx[w*16+g],
 700                                                        cb,
 701                                                        lambda,
 702                                                        INFINITY,
 703                                                        &b);
 704                             bb += b;
 705                         }
 706                             mindist = dist;
 707                             minbits = bb;
 708                     }
 709                     dists[w*16+g] = (mindist - minbits) / lambda;
 710                     bits = minbits;
 711                     if (prev != -1) {
 712                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
 713                     }
 714                     tbits += bits;
 715                     start += sce->ics.swb_sizes[g];
 716                     prev = sce->sf_idx[w*16+g];
 717                 }
 718             }
 719             if (tbits > destbits) {
 720                 for (i = 0; i < 128; i++)
 721                     if (sce->sf_idx[i] < 218 - qstep)
 722                         sce->sf_idx[i] += qstep;
 723             } else {
 724                 for (i = 0; i < 128; i++)
 725                     if (sce->sf_idx[i] > 60 - qstep)
 726                         sce->sf_idx[i] -= qstep;
 727             }
 728             qstep >>= 1;
 729             if (!qstep && tbits > destbits*1.02)
 730                 qstep = 1;
 731             if (sce->sf_idx[0] >= 217)
 732                 break;
 733         } while (qstep);
 734
 735         fflag = 0;
 736         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
 737         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 738             start = w*128;
 739             for (g = 0; g < sce->ics.num_swb; g++) {
 740                 int prevsc = sce->sf_idx[w*16+g];
 741                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
 742                     sce->sf_idx[w*16+g]--;
 743                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
 744                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
 745                 if (sce->sf_idx[w*16+g] != prevsc)
 746                     fflag = 1;
 747             }
 748         }
 749         its++;
 750     } while (fflag && its < 10);
 751 }
 752
 753 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
 754                                        SingleChannelElement *sce,
 755                                        const float lambda)
 756 {
 757     int start = 0, i, w, w2, g;
 758     float uplim[128], maxq[128];
 759     int minq, maxsf;
 760     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
 761     int last = 0, lastband = 0, curband = 0;
 762     float avg_energy = 0.0;
 763     if (sce->ics.num_windows == 1) {
 764         start = 0;
 765         for (i = 0; i < 1024; i++) {
 766             if (i - start >= sce->ics.swb_sizes[curband]) {
 767                 start += sce->ics.swb_sizes[curband];
 768                 curband++;
 769             }
 770             if (sce->coeffs[i]) {
 771                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
 772                 last = i;
 773                 lastband = curband;
 774             }
 775         }
 776     } else {
 777         for (w = 0; w < 8; w++) {
 778             const float *coeffs = sce->coeffs + w*128;
 779             start = 0;
 780             for (i = 0; i < 128; i++) {
 781                 if (i - start >= sce->ics.swb_sizes[curband]) {
 782                     start += sce->ics.swb_sizes[curband];
 783                     curband++;
 784                 }
 785                 if (coeffs[i]) {
 786                     avg_energy += coeffs[i] * coeffs[i];
 787                     last = FFMAX(last, i);
 788                     lastband = FFMAX(lastband, curband);
 789                 }
 790             }
 791         }
 792     }
 793     last++;
 794     avg_energy /= last;
 795     if (avg_energy == 0.0f) {
 796         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
 797             sce->sf_idx[i] = SCALE_ONE_POS;
 798         return;
 799     }
 800     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 801         start = w*128;
 802         for (g = 0; g < sce->ics.num_swb; g++) {
 803             float *coefs   = sce->coeffs + start;
 804             const int size = sce->ics.swb_sizes[g];
 805             int start2 = start, end2 = start + size, peakpos = start;
 806             float maxval = -1, thr = 0.0f, t;
 807             maxq[w*16+g] = 0.0f;
 808             if (g > lastband) {
 809                 maxq[w*16+g] = 0.0f;
 810                 start += size;
 811                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
 812                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
 813                 continue;
 814             }
 815             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 816                 for (i = 0; i < size; i++) {
 817                     float t = coefs[w2*128+i]*coefs[w2*128+i];
 818                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
 819                     thr += t;
 820                     if (sce->ics.num_windows == 1 && maxval < t) {
 821                         maxval  = t;
 822                         peakpos = start+i;
 823                     }
 824                 }
 825             }
 826             if (sce->ics.num_windows == 1) {
 827                 start2 = FFMAX(peakpos - 2, start2);
 828                 end2   = FFMIN(peakpos + 3, end2);
 829             } else {
 830                 start2 -= start;
 831                 end2   -= start;
 832             }
 833             start += size;
 834             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
 835             t   = 1.0 - (1.0 * start2 / last);
 836             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
 837         }
 838     }
 839     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 840     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 841     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 842         start = w*128;
 843         for (g = 0;  g < sce->ics.num_swb; g++) {
 844             const float *coefs  = sce->coeffs + start;
 845             const float *scaled = s->scoefs   + start;
 846             const int size      = sce->ics.swb_sizes[g];
 847             int scf, prev_scf, step;
 848             int min_scf = -1, max_scf = 256;
 849             float curdiff;
 850             if (maxq[w*16+g] < 21.544) {
 851                 sce->zeroes[w*16+g] = 1;
 852                 start += size;
 853                 continue;
 854             }
 855             sce->zeroes[w*16+g] = 0;
 856             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
 857             step = 16;
 858             for (;;) {
 859                 float dist = 0.0f;
 860                 int quant_max;
 861
 862                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 863                     int b;
 864                     dist += quantize_band_cost(s, coefs + w2*128,
 865                                                scaled + w2*128,
 866                                                sce->ics.swb_sizes[g],
 867                                                scf,
 868                                                ESC_BT,
 869                                                lambda,
 870                                                INFINITY,
 871                                                &b);
 872                     dist -= b;
 873                 }
 874                 dist *= 1.0f / 512.0f / lambda;
 875                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
 876                 if (quant_max >= 8191) { // too much, return to the previous quantizer
 877                     sce->sf_idx[w*16+g] = prev_scf;
 878                     break;
 879                 }
 880                 prev_scf = scf;
 881                 curdiff = fabsf(dist - uplim[w*16+g]);
 882                 if (curdiff <= 1.0f)
 883                     step = 0;
 884                 else
 885                     step = log2(curdiff);
 886                 if (dist > uplim[w*16+g])
 887                     step = -step;
 888                 scf += step;
 889                 scf = av_clip_uint8(scf);
 890                 step = scf - prev_scf;
 891                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
 892                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
 893                     break;
 894                 }
 895                 if (step > 0)
 896                     min_scf = prev_scf;
 897                 else
 898                     max_scf = prev_scf;
 899             }
 900             start += size;
 901         }
 902     }
 903     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
 904     for (i = 1; i < 128; i++) {
 905         if (!sce->sf_idx[i])
 906             sce->sf_idx[i] = sce->sf_idx[i-1];
 907         else
 908             minq = FFMIN(minq, sce->sf_idx[i]);
 909     }
 910     if (minq == INT_MAX)
 911         minq = 0;
 912     minq = FFMIN(minq, SCALE_MAX_POS);
 913     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
 914     for (i = 126; i >= 0; i--) {
 915         if (!sce->sf_idx[i])
 916             sce->sf_idx[i] = sce->sf_idx[i+1];
 917         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
 918     }
 919 }
 920
 921 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
 922                                        SingleChannelElement *sce,
 923                                        const float lambda)
 924 {
 925     int start = 0, i, w, w2, g;
 926     int minq = 255;
 927
 928     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 929     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 930         start = w*128;
 931         for (g = 0; g < sce->ics.num_swb; g++) {
 932             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 933                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 934                 if (band->energy <= band->threshold) {
 935                     sce->sf_idx[(w+w2)*16+g] = 218;
 936                     sce->zeroes[(w+w2)*16+g] = 1;
 937                 } else {
 938                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
 939                     sce->zeroes[(w+w2)*16+g] = 0;
 940                 }
 941                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
 942             }
 943         }
 944     }
 945     for (i = 0; i < 128; i++) {
 946         sce->sf_idx[i] = 140;
 947         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
 948     }
 949     //set the same quantizers inside window groups
 950     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 951         for (g = 0;  g < sce->ics.num_swb; g++)
 952             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 953                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 954 }
 955
 956 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
 957                           const float lambda)
 958 {
 959     int start = 0, i, w, w2, g;
 960     float M[128], S[128];
 961     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
 962     SingleChannelElement *sce0 = &cpe->ch[0];
 963     SingleChannelElement *sce1 = &cpe->ch[1];
 964     if (!cpe->common_window)
 965         return;
 966     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
 967         for (g = 0;  g < sce0->ics.num_swb; g++) {
 968             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
 969                 float dist1 = 0.0f, dist2 = 0.0f;
 970                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
 971                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
 972                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
 973                     float minthr = FFMIN(band0->threshold, band1->threshold);
 974                     float maxthr = FFMAX(band0->threshold, band1->threshold);
 975                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
 976                         M[i] = (sce0->coeffs[start+w2*128+i]
 977                               + sce1->coeffs[start+w2*128+i]) * 0.5;
 978                         S[i] =  sce0->coeffs[start+w2*128+i]
 979                               - sce1->coeffs[start+w2*128+i];
 980                     }
 981                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
 982                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
 983                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
 984                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
 985                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
 986                                                 L34,
 987                                                 sce0->ics.swb_sizes[g],
 988                                                 sce0->sf_idx[(w+w2)*16+g],
 989                                                 sce0->band_type[(w+w2)*16+g],
 990                                                 lambda / band0->threshold, INFINITY, NULL);
 991                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
 992                                                 R34,
 993                                                 sce1->ics.swb_sizes[g],
 994                                                 sce1->sf_idx[(w+w2)*16+g],
 995                                                 sce1->band_type[(w+w2)*16+g],
 996                                                 lambda / band1->threshold, INFINITY, NULL);
 997                     dist2 += quantize_band_cost(s, M,
 998                                                 M34,
 999                                                 sce0->ics.swb_sizes[g],
1000                                                 sce0->sf_idx[(w+w2)*16+g],
1001                                                 sce0->band_type[(w+w2)*16+g],
1002                                                 lambda / maxthr, INFINITY, NULL);
1003                     dist2 += quantize_band_cost(s, S,
1004                                                 S34,
1005                                                 sce1->ics.swb_sizes[g],
1006                                                 sce1->sf_idx[(w+w2)*16+g],
1007                                                 sce1->band_type[(w+w2)*16+g],
1008                                                 lambda / minthr, INFINITY, NULL);
1009                 }
1010                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1011             }
1012             start += sce0->ics.swb_sizes[g];
1013         }
1014     }
1015 }
1016
1017 AACCoefficientsEncoder ff_aac_coders[] = {
1018     {
1019         search_for_quantizers_faac,
1020         encode_window_bands_info,
1021         quantize_and_encode_band,
1022         search_for_ms,
1023     },
1024     {
1025         search_for_quantizers_anmr,
1026         encode_window_bands_info,
1027         quantize_and_encode_band,
1028         search_for_ms,
1029     },
1030     {
1031         search_for_quantizers_twoloop,
1032         codebook_trellis_rate,
1033         quantize_and_encode_band,
1034         search_for_ms,
1035     },
1036     {
1037         search_for_quantizers_fast,
1038         encode_window_bands_info,
1039         quantize_and_encode_band,
1040         search_for_ms,
1041     },
1042 };