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 <float.h>
  34 #include "avcodec.h"
  35 #include "put_bits.h"
  36 #include "aac.h"
  37 #include "aacenc.h"
  38 #include "aactab.h"
  39
  40 /** bits needed to code codebook run value for long windows */
  41 static const uint8_t run_value_bits_long[64] = {
  42      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
  43      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
  44     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
  45     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
  46 };
  47
  48 /** bits needed to code codebook run value for short windows */
  49 static const uint8_t run_value_bits_short[16] = {
  50     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
  51 };
  52
  53 static const uint8_t *run_value_bits[2] = {
  54     run_value_bits_long, run_value_bits_short
  55 };
  56
  57
  58 /**
  59  * Quantize one coefficient.
  60  * @return absolute value of the quantized coefficient
  61  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
  62  */
  63 static av_always_inline int quant(float coef, const float Q)
  64 {
  65     float a = coef * Q;
  66     return sqrtf(a * sqrtf(a)) + 0.4054;
  67 }
  68
  69 static void quantize_bands(int *out, const float *in, const float *scaled,
  70                            int size, float Q34, int is_signed, int maxval)
  71 {
  72     int i;
  73     double qc;
  74     for (i = 0; i < size; i++) {
  75         qc = scaled[i] * Q34;
  76         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
  77         if (is_signed && in[i] < 0.0f) {
  78             out[i] = -out[i];
  79         }
  80     }
  81 }
  82
  83 static void abs_pow34_v(float *out, const float *in, const int size)
  84 {
  85 #ifndef USE_REALLY_FULL_SEARCH
  86     int i;
  87     for (i = 0; i < size; i++) {
  88         float a = fabsf(in[i]);
  89         out[i] = sqrtf(a * sqrtf(a));
  90     }
  91 #endif /* USE_REALLY_FULL_SEARCH */
  92 }
  93
  94 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
  95 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
  96
  97 /**
  98  * Calculate rate distortion cost for quantizing with given codebook
  99  *
 100  * @return quantization distortion
 101  */
 102 static float quantize_and_encode_band_cost(struct AACEncContext *s,
 103                                 PutBitContext *pb, const float *in,
 104                                 const float *scaled, int size, int scale_idx,
 105                                 int cb, const float lambda, const float uplim,
 106                                 int *bits)
 107 {
 108     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
 109     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
 110     const float CLIPPED_ESCAPE = 165140.0f*IQ;
 111     int i, j, k;
 112     float cost = 0;
 113     const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
 114     int resbits = 0;
 115     const float  Q34 = sqrtf(Q * sqrtf(Q));
 116     const int range  = aac_cb_range[cb];
 117     const int maxval = aac_cb_maxval[cb];
 118     int off;
 119
 120     if (!cb) {
 121         for (i = 0; i < size; i++)
 122             cost += in[i]*in[i];
 123         if (bits)
 124             *bits = 0;
 125         return cost * lambda;
 126     }
 127     if (!scaled) {
 128         abs_pow34_v(s->scoefs, in, size);
 129         scaled = s->scoefs;
 130     }
 131     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
 132     if (IS_CODEBOOK_UNSIGNED(cb)) {
 133         off = 0;
 134     } else {
 135         off = maxval;
 136     }
 137     for (i = 0; i < size; i += dim) {
 138         const float *vec;
 139         int *quants = s->qcoefs + i;
 140         int curidx = 0;
 141         int curbits;
 142         float rd = 0.0f;
 143         for (j = 0; j < dim; j++) {
 144             curidx *= range;
 145             curidx += quants[j] + off;
 146         }
 147             curbits =  ff_aac_spectral_bits[cb-1][curidx];
 148             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
 149             if (IS_CODEBOOK_UNSIGNED(cb)) {
 150                 for (k = 0; k < dim; k++) {
 151                     float t = fabsf(in[i+k]);
 152                     float di;
 153                     if (vec[k] == 64.0f) { //FIXME: slow
 154                         if (t >= CLIPPED_ESCAPE) {
 155                             di = t - CLIPPED_ESCAPE;
 156                             curbits += 21;
 157                         } else {
 158                             int c = av_clip(quant(t, Q), 0, 8191);
 159                             di = t - c*cbrtf(c)*IQ;
 160                             curbits += av_log2(c)*2 - 4 + 1;
 161                         }
 162                     } else {
 163                         di = t - vec[k]*IQ;
 164                     }
 165                     if (vec[k] != 0.0f)
 166                         curbits++;
 167                     rd += di*di;
 168                 }
 169             } else {
 170                 for (k = 0; k < dim; k++) {
 171                     float di = in[i+k] - vec[k]*IQ;
 172                     rd += di*di;
 173                 }
 174             }
 175         cost    += rd * lambda + curbits;
 176         resbits += curbits;
 177         if (cost >= uplim)
 178             return uplim;
 179         if (pb) {
 180         put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
 181         if (IS_CODEBOOK_UNSIGNED(cb))
 182             for (j = 0; j < dim; j++)
 183                 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
 184                     put_bits(pb, 1, in[i+j] < 0.0f);
 185         if (cb == ESC_BT) {
 186             for (j = 0; j < 2; j++) {
 187                 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
 188                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
 189                     int len = av_log2(coef);
 190
 191                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
 192                     put_bits(pb, len, coef & ((1 << len) - 1));
 193                 }
 194             }
 195         }
 196         }
 197     }
 198
 199     if (bits)
 200         *bits = resbits;
 201     return cost;
 202 }
 203 static float quantize_band_cost(struct AACEncContext *s, const float *in,
 204                                 const float *scaled, int size, int scale_idx,
 205                                 int cb, const float lambda, const float uplim,
 206                                 int *bits)
 207 {
 208     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
 209                                          cb, lambda, uplim, bits);
 210 }
 211
 212 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
 213                                      const float *in, int size, int scale_idx,
 214                                      int cb, const float lambda)
 215 {
 216     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
 217                                   INFINITY, NULL);
 218 }
 219
 220 static int find_min_book(int sf, int group_len, int swb_size, const float *scaled) {
 221     float maxval = 0.0f;
 222     float Q = ff_aac_pow2sf_tab[200 - sf + SCALE_ONE_POS - SCALE_DIV_512];
 223     float Q34 = sqrtf(Q * sqrtf(Q));
 224     int qmaxval, cb, w2, i;
 225     for (w2 = 0; w2 < group_len; w2++) {
 226         for (i = 0; i < swb_size; i++) {
 227             maxval = FFMAX(maxval, scaled[w2*128+i]);
 228         }
 229     }
 230     qmaxval = maxval * Q34 + 0.4054f;
 231     if      (qmaxval ==  0) cb = 0;
 232     else if (qmaxval ==  1) cb = 1;
 233     else if (qmaxval ==  2) cb = 3;
 234     else if (qmaxval <=  4) cb = 5;
 235     else if (qmaxval <=  7) cb = 7;
 236     else if (qmaxval <= 12) cb = 9;
 237     else                    cb = 11;
 238     return cb;
 239 }
 240
 241 /**
 242  * structure used in optimal codebook search
 243  */
 244 typedef struct BandCodingPath {
 245     int prev_idx; ///< pointer to the previous path point
 246     float cost;   ///< path cost
 247     int run;
 248 } BandCodingPath;
 249
 250 /**
 251  * Encode band info for single window group bands.
 252  */
 253 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
 254                                      int win, int group_len, const float lambda)
 255 {
 256     BandCodingPath path[120][12];
 257     int w, swb, cb, start, start2, size;
 258     int i, j;
 259     const int max_sfb  = sce->ics.max_sfb;
 260     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 261     const int run_esc  = (1 << run_bits) - 1;
 262     int idx, ppos, count;
 263     int stackrun[120], stackcb[120], stack_len;
 264     float next_minrd = INFINITY;
 265     int next_mincb = 0;
 266
 267     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 268     start = win*128;
 269     for (cb = 0; cb < 12; cb++) {
 270         path[0][cb].cost     = 0.0f;
 271         path[0][cb].prev_idx = -1;
 272         path[0][cb].run      = 0;
 273     }
 274     for (swb = 0; swb < max_sfb; swb++) {
 275         start2 = start;
 276         size = sce->ics.swb_sizes[swb];
 277         if (sce->zeroes[win*16 + swb]) {
 278             for (cb = 0; cb < 12; cb++) {
 279                 path[swb+1][cb].prev_idx = cb;
 280                 path[swb+1][cb].cost     = path[swb][cb].cost;
 281                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 282             }
 283         } else {
 284             float minrd = next_minrd;
 285             int mincb = next_mincb;
 286             next_minrd = INFINITY;
 287             next_mincb = 0;
 288             for (cb = 0; cb < 12; cb++) {
 289                 float cost_stay_here, cost_get_here;
 290                 float rd = 0.0f;
 291                 for (w = 0; w < group_len; w++) {
 292                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
 293                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 294                                              s->scoefs + start + w*128, size,
 295                                              sce->sf_idx[(win+w)*16+swb], cb,
 296                                              lambda / band->threshold, INFINITY, NULL);
 297                 }
 298                 cost_stay_here = path[swb][cb].cost + rd;
 299                 cost_get_here  = minrd              + rd + run_bits + 4;
 300                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 301                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 302                     cost_stay_here += run_bits;
 303                 if (cost_get_here < cost_stay_here) {
 304                     path[swb+1][cb].prev_idx = mincb;
 305                     path[swb+1][cb].cost     = cost_get_here;
 306                     path[swb+1][cb].run      = 1;
 307                 } else {
 308                     path[swb+1][cb].prev_idx = cb;
 309                     path[swb+1][cb].cost     = cost_stay_here;
 310                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 311                 }
 312                 if (path[swb+1][cb].cost < next_minrd) {
 313                     next_minrd = path[swb+1][cb].cost;
 314                     next_mincb = cb;
 315                 }
 316             }
 317         }
 318         start += sce->ics.swb_sizes[swb];
 319     }
 320
 321     //convert resulting path from backward-linked list
 322     stack_len = 0;
 323     idx       = 0;
 324     for (cb = 1; cb < 12; cb++)
 325         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 326             idx = cb;
 327     ppos = max_sfb;
 328     while (ppos > 0) {
 329         cb = idx;
 330         stackrun[stack_len] = path[ppos][cb].run;
 331         stackcb [stack_len] = cb;
 332         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 333         ppos -= path[ppos][cb].run;
 334         stack_len++;
 335     }
 336     //perform actual band info encoding
 337     start = 0;
 338     for (i = stack_len - 1; i >= 0; i--) {
 339         put_bits(&s->pb, 4, stackcb[i]);
 340         count = stackrun[i];
 341         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
 342         //XXX: memset when band_type is also uint8_t
 343         for (j = 0; j < count; j++) {
 344             sce->band_type[win*16 + start] =  stackcb[i];
 345             start++;
 346         }
 347         while (count >= run_esc) {
 348             put_bits(&s->pb, run_bits, run_esc);
 349             count -= run_esc;
 350         }
 351         put_bits(&s->pb, run_bits, count);
 352     }
 353 }
 354
 355 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
 356                                   int win, int group_len, const float lambda)
 357 {
 358     BandCodingPath path[120][12];
 359     int w, swb, cb, start, start2, size;
 360     int i, j;
 361     const int max_sfb  = sce->ics.max_sfb;
 362     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 363     const int run_esc  = (1 << run_bits) - 1;
 364     int idx, ppos, count;
 365     int stackrun[120], stackcb[120], stack_len;
 366     float next_minrd = INFINITY;
 367     int next_mincb = 0;
 368
 369     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 370     start = win*128;
 371     for (cb = 0; cb < 12; cb++) {
 372         path[0][cb].cost     = run_bits+4;
 373         path[0][cb].prev_idx = -1;
 374         path[0][cb].run      = 0;
 375     }
 376     for (swb = 0; swb < max_sfb; swb++) {
 377         start2 = start;
 378         size = sce->ics.swb_sizes[swb];
 379         if (sce->zeroes[win*16 + swb]) {
 380             for (cb = 0; cb < 12; cb++) {
 381                 path[swb+1][cb].prev_idx = cb;
 382                 path[swb+1][cb].cost     = path[swb][cb].cost;
 383                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 384             }
 385         } else {
 386             float minrd = next_minrd;
 387             int mincb = next_mincb;
 388             int startcb = sce->band_type[win*16+swb];
 389             next_minrd = INFINITY;
 390             next_mincb = 0;
 391             for (cb = 0; cb < startcb; cb++) {
 392                 path[swb+1][cb].cost = 61450;
 393                 path[swb+1][cb].prev_idx = -1;
 394                 path[swb+1][cb].run = 0;
 395             }
 396             for (cb = startcb; cb < 12; cb++) {
 397                 float cost_stay_here, cost_get_here;
 398                 float rd = 0.0f;
 399                 for (w = 0; w < group_len; w++) {
 400                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 401                                              s->scoefs + start + w*128, size,
 402                                              sce->sf_idx[(win+w)*16+swb], cb,
 403                                              0, INFINITY, NULL);
 404                 }
 405                 cost_stay_here = path[swb][cb].cost + rd;
 406                 cost_get_here  = minrd              + rd + run_bits + 4;
 407                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 408                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 409                     cost_stay_here += run_bits;
 410                 if (cost_get_here < cost_stay_here) {
 411                     path[swb+1][cb].prev_idx = mincb;
 412                     path[swb+1][cb].cost     = cost_get_here;
 413                     path[swb+1][cb].run      = 1;
 414                 } else {
 415                     path[swb+1][cb].prev_idx = cb;
 416                     path[swb+1][cb].cost     = cost_stay_here;
 417                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 418                 }
 419                 if (path[swb+1][cb].cost < next_minrd) {
 420                     next_minrd = path[swb+1][cb].cost;
 421                     next_mincb = cb;
 422                 }
 423             }
 424         }
 425         start += sce->ics.swb_sizes[swb];
 426     }
 427
 428     //convert resulting path from backward-linked list
 429     stack_len = 0;
 430     idx       = 0;
 431     for (cb = 1; cb < 12; cb++)
 432         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 433             idx = cb;
 434     ppos = max_sfb;
 435     while (ppos > 0) {
 436         if (idx < 0) abort();
 437         cb = idx;
 438         stackrun[stack_len] = path[ppos][cb].run;
 439         stackcb [stack_len] = cb;
 440         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 441         ppos -= path[ppos][cb].run;
 442         stack_len++;
 443     }
 444     //perform actual band info encoding
 445     start = 0;
 446     for (i = stack_len - 1; i >= 0; i--) {
 447         put_bits(&s->pb, 4, stackcb[i]);
 448         count = stackrun[i];
 449         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
 450         //XXX: memset when band_type is also uint8_t
 451         for (j = 0; j < count; j++) {
 452             sce->band_type[win*16 + start] =  stackcb[i];
 453             start++;
 454         }
 455         while (count >= run_esc) {
 456             put_bits(&s->pb, run_bits, run_esc);
 457             count -= run_esc;
 458         }
 459         put_bits(&s->pb, run_bits, count);
 460     }
 461 }
 462
 463 typedef struct TrellisPath {
 464     float cost;
 465     int prev;
 466 } TrellisPath;
 467
 468 #define TRELLIS_STAGES 121
 469 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
 470
 471 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
 472                                        SingleChannelElement *sce,
 473                                        const float lambda)
 474 {
 475     int q, w, w2, g, start = 0;
 476     int i, j;
 477     int idx;
 478     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
 479     int bandaddr[TRELLIS_STAGES];
 480     int minq;
 481     float mincost;
 482     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
 483     int q0, q1, qcnt = 0;
 484
 485     for (i = 0; i < 1024; i++) {
 486         float t = fabsf(sce->coeffs[i]);
 487         if (t > 0.0f) {
 488             q0f = FFMIN(q0f, t);
 489             q1f = FFMAX(q1f, t);
 490             qnrgf += t*t;
 491             qcnt++;
 492         }
 493     }
 494
 495     if (!qcnt) {
 496         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 497         memset(sce->zeroes, 1, sizeof(sce->zeroes));
 498         return;
 499     }
 500
 501     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 502     q0 = av_clip_uint8(log2(q0f)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 503     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 504     q1 = av_clip_uint8(log2(q1f)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 505     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
 506     if (q1 - q0 > 60) {
 507         int q0low  = q0;
 508         int q1high = q1;
 509         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
 510         int qnrg = av_clip_uint8(log2(sqrt(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
 511         q1 = qnrg + 30;
 512         q0 = qnrg - 30;
 513     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
 514         if (q0 < q0low) {
 515             q1 += q0low - q0;
 516             q0  = q0low;
 517         } else if (q1 > q1high) {
 518             q0 -= q1 - q1high;
 519             q1  = q1high;
 520         }
 521     }
 522     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
 523
 524     for (i = 0; i < TRELLIS_STATES; i++) {
 525         paths[0][i].cost    = 0.0f;
 526         paths[0][i].prev    = -1;
 527     }
 528     for (j = 1; j < TRELLIS_STAGES; j++) {
 529         for (i = 0; i < TRELLIS_STATES; i++) {
 530             paths[j][i].cost    = INFINITY;
 531             paths[j][i].prev    = -2;
 532         }
 533     }
 534     idx = 1;
 535     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 536     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 537         start = w*128;
 538         for (g = 0; g < sce->ics.num_swb; g++) {
 539             const float *coefs = sce->coeffs + start;
 540             float qmin, qmax;
 541             int nz = 0;
 542
 543             bandaddr[idx] = w * 16 + g;
 544             qmin = INT_MAX;
 545             qmax = 0.0f;
 546             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 547                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 548                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 549                     sce->zeroes[(w+w2)*16+g] = 1;
 550                     continue;
 551                 }
 552                 sce->zeroes[(w+w2)*16+g] = 0;
 553                 nz = 1;
 554                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
 555                     float t = fabsf(coefs[w2*128+i]);
 556                     if (t > 0.0f)
 557                         qmin = FFMIN(qmin, t);
 558                     qmax = FFMAX(qmax, t);
 559                 }
 560             }
 561             if (nz) {
 562                 int minscale, maxscale;
 563                 float minrd = INFINITY;
 564                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 565                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 566                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 567                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 568                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
 569                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
 570                 for (q = minscale; q < maxscale; q++) {
 571                     float dist = 0;
 572                     int cb = find_min_book(sce->sf_idx[w*16+g], sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
 573                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 574                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 575                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
 576                                                             q + q0, cb, lambda / band->threshold, INFINITY, NULL);
 577                     }
 578                     minrd = FFMIN(minrd, dist);
 579
 580                     for (i = 0; i < q1 - q0; i++) {
 581                         float cost;
 582                         if (isinf(paths[idx - 1][i].cost))
 583                             continue;
 584                         cost = paths[idx - 1][i].cost + dist
 585                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 586                         if (cost < paths[idx][q].cost) {
 587                             paths[idx][q].cost    = cost;
 588                             paths[idx][q].prev    = i;
 589                         }
 590                     }
 591                 }
 592             } else {
 593                 for (q = 0; q < q1 - q0; q++) {
 594                     if (!isinf(paths[idx - 1][q].cost)) {
 595                         paths[idx][q].cost = paths[idx - 1][q].cost + 1;
 596                         paths[idx][q].prev = q;
 597                         continue;
 598                     }
 599                     for (i = 0; i < q1 - q0; i++) {
 600                         float cost;
 601                         if (isinf(paths[idx - 1][i].cost))
 602                             continue;
 603                         cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 604                         if (cost < paths[idx][q].cost) {
 605                             paths[idx][q].cost    = cost;
 606                             paths[idx][q].prev    = i;
 607                         }
 608                     }
 609                 }
 610             }
 611             sce->zeroes[w*16+g] = !nz;
 612             start += sce->ics.swb_sizes[g];
 613             idx++;
 614         }
 615     }
 616     idx--;
 617     mincost = paths[idx][0].cost;
 618     minq    = 0;
 619     for (i = 1; i < TRELLIS_STATES; i++) {
 620         if (paths[idx][i].cost < mincost) {
 621             mincost = paths[idx][i].cost;
 622             minq = i;
 623         }
 624     }
 625     while (idx) {
 626         sce->sf_idx[bandaddr[idx]] = minq + q0;
 627         minq = paths[idx][minq].prev;
 628         idx--;
 629     }
 630     //set the same quantizers inside window groups
 631     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 632         for (g = 0;  g < sce->ics.num_swb; g++)
 633             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 634                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 635 }
 636
 637 /**
 638  * two-loop quantizers search taken from ISO 13818-7 Appendix C
 639  */
 640 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
 641                                           AACEncContext *s,
 642                                           SingleChannelElement *sce,
 643                                           const float lambda)
 644 {
 645     int start = 0, i, w, w2, g;
 646     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
 647     float dists[128], uplims[128];
 648     int fflag, minscaler;
 649     int its  = 0;
 650     int allz = 0;
 651     float minthr = INFINITY;
 652
 653     //XXX: some heuristic to determine initial quantizers will reduce search time
 654     memset(dists, 0, sizeof(dists));
 655     //determine zero bands and upper limits
 656     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 657         for (g = 0;  g < sce->ics.num_swb; g++) {
 658             int nz = 0;
 659             float uplim = 0.0f;
 660             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 661                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 662                 uplim += band->threshold;
 663                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 664                     sce->zeroes[(w+w2)*16+g] = 1;
 665                     continue;
 666                 }
 667                 nz = 1;
 668             }
 669             uplims[w*16+g] = uplim *512;
 670             sce->zeroes[w*16+g] = !nz;
 671             if (nz)
 672                 minthr = FFMIN(minthr, uplim);
 673             allz = FFMAX(allz, nz);
 674         }
 675     }
 676     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 677         for (g = 0;  g < sce->ics.num_swb; g++) {
 678             if (sce->zeroes[w*16+g]) {
 679                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
 680                 continue;
 681             }
 682             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
 683         }
 684     }
 685
 686     if (!allz)
 687         return;
 688     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 689     //perform two-loop search
 690     //outer loop - improve quality
 691     do {
 692         int tbits, qstep;
 693         minscaler = sce->sf_idx[0];
 694         //inner loop - quantize spectrum to fit into given number of bits
 695         qstep = its ? 1 : 32;
 696         do {
 697             int prev = -1;
 698             tbits = 0;
 699             fflag = 0;
 700             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 701                 start = w*128;
 702                 for (g = 0;  g < sce->ics.num_swb; g++) {
 703                     const float *coefs = sce->coeffs + start;
 704                     const float *scaled = s->scoefs + start;
 705                     int bits = 0;
 706                     int cb;
 707                     float mindist = INFINITY;
 708                     int minbits = 0;
 709
 710                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
 711                         start += sce->ics.swb_sizes[g];
 712                         continue;
 713                     }
 714                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
 715                     {
 716                         float dist = 0.0f;
 717                         int bb = 0;
 718                         cb = find_min_book(sce->sf_idx[w*16+g], sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
 719                         sce->band_type[w*16+g] = cb;
 720                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 721                             int b;
 722                             dist += quantize_band_cost(s, coefs + w2*128,
 723                                                        scaled + w2*128,
 724                                                        sce->ics.swb_sizes[g],
 725                                                        sce->sf_idx[w*16+g],
 726                                                        cb,
 727                                                        lambda,
 728                                                        INFINITY,
 729                                                        &b);
 730                             bb += b;
 731                         }
 732                             mindist = dist;
 733                             minbits = bb;
 734                     }
 735                     dists[w*16+g] = (mindist - minbits) / lambda;
 736                     bits = minbits;
 737                     if (prev != -1) {
 738                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
 739                     }
 740                     tbits += bits;
 741                     start += sce->ics.swb_sizes[g];
 742                     prev = sce->sf_idx[w*16+g];
 743                 }
 744             }
 745             if (tbits > destbits) {
 746                 for (i = 0; i < 128; i++)
 747                     if (sce->sf_idx[i] < 218 - qstep)
 748                         sce->sf_idx[i] += qstep;
 749             } else {
 750                 for (i = 0; i < 128; i++)
 751                     if (sce->sf_idx[i] > 60 - qstep)
 752                         sce->sf_idx[i] -= qstep;
 753             }
 754             qstep >>= 1;
 755             if (!qstep && tbits > destbits*1.02)
 756                 qstep = 1;
 757             if (sce->sf_idx[0] >= 217)
 758                 break;
 759         } while (qstep);
 760
 761         fflag = 0;
 762         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
 763         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 764             start = w*128;
 765             for (g = 0; g < sce->ics.num_swb; g++) {
 766                 int prevsc = sce->sf_idx[w*16+g];
 767                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
 768                     sce->sf_idx[w*16+g]--;
 769                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
 770                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
 771                 if (sce->sf_idx[w*16+g] != prevsc)
 772                     fflag = 1;
 773             }
 774         }
 775         its++;
 776     } while (fflag && its < 10);
 777 }
 778
 779 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
 780                                        SingleChannelElement *sce,
 781                                        const float lambda)
 782 {
 783     int start = 0, i, w, w2, g;
 784     float uplim[128], maxq[128];
 785     int minq, maxsf;
 786     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
 787     int last = 0, lastband = 0, curband = 0;
 788     float avg_energy = 0.0;
 789     if (sce->ics.num_windows == 1) {
 790         start = 0;
 791         for (i = 0; i < 1024; i++) {
 792             if (i - start >= sce->ics.swb_sizes[curband]) {
 793                 start += sce->ics.swb_sizes[curband];
 794                 curband++;
 795             }
 796             if (sce->coeffs[i]) {
 797                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
 798                 last = i;
 799                 lastband = curband;
 800             }
 801         }
 802     } else {
 803         for (w = 0; w < 8; w++) {
 804             const float *coeffs = sce->coeffs + w*128;
 805             start = 0;
 806             for (i = 0; i < 128; i++) {
 807                 if (i - start >= sce->ics.swb_sizes[curband]) {
 808                     start += sce->ics.swb_sizes[curband];
 809                     curband++;
 810                 }
 811                 if (coeffs[i]) {
 812                     avg_energy += coeffs[i] * coeffs[i];
 813                     last = FFMAX(last, i);
 814                     lastband = FFMAX(lastband, curband);
 815                 }
 816             }
 817         }
 818     }
 819     last++;
 820     avg_energy /= last;
 821     if (avg_energy == 0.0f) {
 822         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
 823             sce->sf_idx[i] = SCALE_ONE_POS;
 824         return;
 825     }
 826     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 827         start = w*128;
 828         for (g = 0; g < sce->ics.num_swb; g++) {
 829             float *coefs   = sce->coeffs + start;
 830             const int size = sce->ics.swb_sizes[g];
 831             int start2 = start, end2 = start + size, peakpos = start;
 832             float maxval = -1, thr = 0.0f, t;
 833             maxq[w*16+g] = 0.0f;
 834             if (g > lastband) {
 835                 maxq[w*16+g] = 0.0f;
 836                 start += size;
 837                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
 838                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
 839                 continue;
 840             }
 841             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 842                 for (i = 0; i < size; i++) {
 843                     float t = coefs[w2*128+i]*coefs[w2*128+i];
 844                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
 845                     thr += t;
 846                     if (sce->ics.num_windows == 1 && maxval < t) {
 847                         maxval  = t;
 848                         peakpos = start+i;
 849                     }
 850                 }
 851             }
 852             if (sce->ics.num_windows == 1) {
 853                 start2 = FFMAX(peakpos - 2, start2);
 854                 end2   = FFMIN(peakpos + 3, end2);
 855             } else {
 856                 start2 -= start;
 857                 end2   -= start;
 858             }
 859             start += size;
 860             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
 861             t   = 1.0 - (1.0 * start2 / last);
 862             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
 863         }
 864     }
 865     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 866     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 867     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 868         start = w*128;
 869         for (g = 0;  g < sce->ics.num_swb; g++) {
 870             const float *coefs  = sce->coeffs + start;
 871             const float *scaled = s->scoefs   + start;
 872             const int size      = sce->ics.swb_sizes[g];
 873             int scf, prev_scf, step;
 874             int min_scf = -1, max_scf = 256;
 875             float curdiff;
 876             if (maxq[w*16+g] < 21.544) {
 877                 sce->zeroes[w*16+g] = 1;
 878                 start += size;
 879                 continue;
 880             }
 881             sce->zeroes[w*16+g] = 0;
 882             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
 883             step = 16;
 884             for (;;) {
 885                 float dist = 0.0f;
 886                 int quant_max;
 887
 888                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 889                     int b;
 890                     dist += quantize_band_cost(s, coefs + w2*128,
 891                                                scaled + w2*128,
 892                                                sce->ics.swb_sizes[g],
 893                                                scf,
 894                                                ESC_BT,
 895                                                lambda,
 896                                                INFINITY,
 897                                                &b);
 898                     dist -= b;
 899                 }
 900                 dist *= 1.0f / 512.0f / lambda;
 901                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
 902                 if (quant_max >= 8191) { // too much, return to the previous quantizer
 903                     sce->sf_idx[w*16+g] = prev_scf;
 904                     break;
 905                 }
 906                 prev_scf = scf;
 907                 curdiff = fabsf(dist - uplim[w*16+g]);
 908                 if (curdiff <= 1.0f)
 909                     step = 0;
 910                 else
 911                     step = log2(curdiff);
 912                 if (dist > uplim[w*16+g])
 913                     step = -step;
 914                 scf += step;
 915                 scf = av_clip_uint8(scf);
 916                 step = scf - prev_scf;
 917                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
 918                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
 919                     break;
 920                 }
 921                 if (step > 0)
 922                     min_scf = prev_scf;
 923                 else
 924                     max_scf = prev_scf;
 925             }
 926             start += size;
 927         }
 928     }
 929     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
 930     for (i = 1; i < 128; i++) {
 931         if (!sce->sf_idx[i])
 932             sce->sf_idx[i] = sce->sf_idx[i-1];
 933         else
 934             minq = FFMIN(minq, sce->sf_idx[i]);
 935     }
 936     if (minq == INT_MAX)
 937         minq = 0;
 938     minq = FFMIN(minq, SCALE_MAX_POS);
 939     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
 940     for (i = 126; i >= 0; i--) {
 941         if (!sce->sf_idx[i])
 942             sce->sf_idx[i] = sce->sf_idx[i+1];
 943         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
 944     }
 945 }
 946
 947 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
 948                                        SingleChannelElement *sce,
 949                                        const float lambda)
 950 {
 951     int start = 0, i, w, w2, g;
 952     int minq = 255;
 953
 954     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 955     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 956         start = w*128;
 957         for (g = 0; g < sce->ics.num_swb; g++) {
 958             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 959                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
 960                 if (band->energy <= band->threshold) {
 961                     sce->sf_idx[(w+w2)*16+g] = 218;
 962                     sce->zeroes[(w+w2)*16+g] = 1;
 963                 } else {
 964                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
 965                     sce->zeroes[(w+w2)*16+g] = 0;
 966                 }
 967                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
 968             }
 969         }
 970     }
 971     for (i = 0; i < 128; i++) {
 972         sce->sf_idx[i] = 140;
 973         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
 974     }
 975     //set the same quantizers inside window groups
 976     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 977         for (g = 0;  g < sce->ics.num_swb; g++)
 978             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 979                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 980 }
 981
 982 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
 983                           const float lambda)
 984 {
 985     int start = 0, i, w, w2, g;
 986     float M[128], S[128];
 987     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
 988     SingleChannelElement *sce0 = &cpe->ch[0];
 989     SingleChannelElement *sce1 = &cpe->ch[1];
 990     if (!cpe->common_window)
 991         return;
 992     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
 993         for (g = 0;  g < sce0->ics.num_swb; g++) {
 994             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
 995                 float dist1 = 0.0f, dist2 = 0.0f;
 996                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
 997                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
 998                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
 999                     float minthr = FFMIN(band0->threshold, band1->threshold);
1000                     float maxthr = FFMAX(band0->threshold, band1->threshold);
1001                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1002                         M[i] = (sce0->coeffs[start+w2*128+i]
1003                               + sce1->coeffs[start+w2*128+i]) * 0.5;
1004                         S[i] =  sce0->coeffs[start+w2*128+i]
1005                               - sce1->coeffs[start+w2*128+i];
1006                     }
1007                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1008                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1009                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
1010                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
1011                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1012                                                 L34,
1013                                                 sce0->ics.swb_sizes[g],
1014                                                 sce0->sf_idx[(w+w2)*16+g],
1015                                                 sce0->band_type[(w+w2)*16+g],
1016                                                 lambda / band0->threshold, INFINITY, NULL);
1017                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1018                                                 R34,
1019                                                 sce1->ics.swb_sizes[g],
1020                                                 sce1->sf_idx[(w+w2)*16+g],
1021                                                 sce1->band_type[(w+w2)*16+g],
1022                                                 lambda / band1->threshold, INFINITY, NULL);
1023                     dist2 += quantize_band_cost(s, M,
1024                                                 M34,
1025                                                 sce0->ics.swb_sizes[g],
1026                                                 sce0->sf_idx[(w+w2)*16+g],
1027                                                 sce0->band_type[(w+w2)*16+g],
1028                                                 lambda / maxthr, INFINITY, NULL);
1029                     dist2 += quantize_band_cost(s, S,
1030                                                 S34,
1031                                                 sce1->ics.swb_sizes[g],
1032                                                 sce1->sf_idx[(w+w2)*16+g],
1033                                                 sce1->band_type[(w+w2)*16+g],
1034                                                 lambda / minthr, INFINITY, NULL);
1035                 }
1036                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1037             }
1038             start += sce0->ics.swb_sizes[g];
1039         }
1040     }
1041 }
1042
1043 AACCoefficientsEncoder ff_aac_coders[] = {
1044     {
1045         search_for_quantizers_faac,
1046         encode_window_bands_info,
1047         quantize_and_encode_band,
1048         search_for_ms,
1049     },
1050     {
1051         search_for_quantizers_anmr,
1052         encode_window_bands_info,
1053         quantize_and_encode_band,
1054         search_for_ms,
1055     },
1056     {
1057         search_for_quantizers_twoloop,
1058         codebook_trellis_rate,
1059         quantize_and_encode_band,
1060         search_for_ms,
1061     },
1062     {
1063         search_for_quantizers_fast,
1064         encode_window_bands_info,
1065         quantize_and_encode_band,
1066         search_for_ms,
1067     },
1068 };