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 "libavutil/libm.h" // brought forward to work around cygwin header breakage
  34
  35 #include <float.h>
  36 #include "libavutil/mathematics.h"
  37 #include "avcodec.h"
  38 #include "put_bits.h"
  39 #include "aac.h"
  40 #include "aacenc.h"
  41 #include "aactab.h"
  42 #include "aac_tablegen_decl.h"
  43
  44 /** Frequency in Hz for lower limit of noise substitution **/
  45 #define NOISE_LOW_LIMIT 4500
  46
  47 /* Energy spread threshold value below which no PNS is used, this corresponds to
  48  * typically around 17Khz, after which PNS usage decays ending at 19Khz */
  49 #define NOISE_SPREAD_THRESHOLD 0.5f
  50
  51 /* This constant gets divided by lambda to return ~1.65 which when multiplied
  52  * by the band->threshold and compared to band->energy is the boundary between
  53  * excessive PNS and little PNS usage. */
  54 #define NOISE_LAMBDA_NUMERATOR 252.1f
  55
  56 /** Frequency in Hz for lower limit of intensity stereo   **/
  57 #define INT_STEREO_LOW_LIMIT 6100
  58
  59 /** Total number of usable codebooks **/
  60 #define CB_TOT 12
  61
  62 /** Total number of codebooks, including special ones **/
  63 #define CB_TOT_ALL 15
  64
  65 /** bits needed to code codebook run value for long windows */
  66 static const uint8_t run_value_bits_long[64] = {
  67      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
  68      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
  69     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
  70     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
  71 };
  72
  73 /** bits needed to code codebook run value for short windows */
  74 static const uint8_t run_value_bits_short[16] = {
  75     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
  76 };
  77
  78 static const uint8_t * const run_value_bits[2] = {
  79     run_value_bits_long, run_value_bits_short
  80 };
  81
  82 /** Map to convert values from BandCodingPath index to a codebook index **/
  83 static const uint8_t aac_cb_out_map[CB_TOT_ALL]  = {0,1,2,3,4,5,6,7,8,9,10,11,13,14,15};
  84 /** Inverse map to convert from codebooks to BandCodingPath indices **/
  85 static const uint8_t aac_cb_in_map[CB_TOT_ALL+1] = {0,1,2,3,4,5,6,7,8,9,10,11,0,12,13,14};
  86
  87 /**
  88  * Quantize one coefficient.
  89  * @return absolute value of the quantized coefficient
  90  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
  91  */
  92 static av_always_inline int quant(float coef, const float Q)
  93 {
  94     float a = coef * Q;
  95     return sqrtf(a * sqrtf(a)) + 0.4054;
  96 }
  97
  98 static void quantize_bands(int *out, const float *in, const float *scaled,
  99                            int size, float Q34, int is_signed, int maxval)
 100 {
 101     int i;
 102     double qc;
 103     for (i = 0; i < size; i++) {
 104         qc = scaled[i] * Q34;
 105         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
 106         if (is_signed && in[i] < 0.0f) {
 107             out[i] = -out[i];
 108         }
 109     }
 110 }
 111
 112 static void abs_pow34_v(float *out, const float *in, const int size)
 113 {
 114 #ifndef USE_REALLY_FULL_SEARCH
 115     int i;
 116     for (i = 0; i < size; i++) {
 117         float a = fabsf(in[i]);
 118         out[i] = sqrtf(a * sqrtf(a));
 119     }
 120 #endif /* USE_REALLY_FULL_SEARCH */
 121 }
 122
 123 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
 124 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
 125
 126 /**
 127  * Calculate rate distortion cost for quantizing with given codebook
 128  *
 129  * @return quantization distortion
 130  */
 131 static av_always_inline float quantize_and_encode_band_cost_template(
 132                                 struct AACEncContext *s,
 133                                 PutBitContext *pb, const float *in,
 134                                 const float *scaled, int size, int scale_idx,
 135                                 int cb, const float lambda, const float uplim,
 136                                 int *bits, int BT_ZERO, int BT_UNSIGNED,
 137                                 int BT_PAIR, int BT_ESC, int BT_NOISE, int BT_STEREO)
 138 {
 139     const int q_idx = POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512;
 140     const float Q   = ff_aac_pow2sf_tab [q_idx];
 141     const float Q34 = ff_aac_pow34sf_tab[q_idx];
 142     const float IQ  = ff_aac_pow2sf_tab [POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
 143     const float CLIPPED_ESCAPE = 165140.0f*IQ;
 144     int i, j;
 145     float cost = 0;
 146     const int dim = BT_PAIR ? 2 : 4;
 147     int resbits = 0;
 148     int off;
 149
 150     if (BT_ZERO || BT_NOISE || BT_STEREO) {
 151         for (i = 0; i < size; i++)
 152             cost += in[i]*in[i];
 153         if (bits)
 154             *bits = 0;
 155         return cost * lambda;
 156     }
 157     if (!scaled) {
 158         abs_pow34_v(s->scoefs, in, size);
 159         scaled = s->scoefs;
 160     }
 161     quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, aac_cb_maxval[cb]);
 162     if (BT_UNSIGNED) {
 163         off = 0;
 164     } else {
 165         off = aac_cb_maxval[cb];
 166     }
 167     for (i = 0; i < size; i += dim) {
 168         const float *vec;
 169         int *quants = s->qcoefs + i;
 170         int curidx = 0;
 171         int curbits;
 172         float rd = 0.0f;
 173         for (j = 0; j < dim; j++) {
 174             curidx *= aac_cb_range[cb];
 175             curidx += quants[j] + off;
 176         }
 177         curbits =  ff_aac_spectral_bits[cb-1][curidx];
 178         vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
 179         if (BT_UNSIGNED) {
 180             for (j = 0; j < dim; j++) {
 181                 float t = fabsf(in[i+j]);
 182                 float di;
 183                 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
 184                     if (t >= CLIPPED_ESCAPE) {
 185                         di = t - CLIPPED_ESCAPE;
 186                         curbits += 21;
 187                     } else {
 188                         int c = av_clip_uintp2(quant(t, Q), 13);
 189                         di = t - c*cbrtf(c)*IQ;
 190                         curbits += av_log2(c)*2 - 4 + 1;
 191                     }
 192                 } else {
 193                     di = t - vec[j]*IQ;
 194                 }
 195                 if (vec[j] != 0.0f)
 196                     curbits++;
 197                 rd += di*di;
 198             }
 199         } else {
 200             for (j = 0; j < dim; j++) {
 201                 float di = in[i+j] - vec[j]*IQ;
 202                 rd += di*di;
 203             }
 204         }
 205         cost    += rd * lambda + curbits;
 206         resbits += curbits;
 207         if (cost >= uplim)
 208             return uplim;
 209         if (pb) {
 210             put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
 211             if (BT_UNSIGNED)
 212                 for (j = 0; j < dim; j++)
 213                     if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
 214                         put_bits(pb, 1, in[i+j] < 0.0f);
 215             if (BT_ESC) {
 216                 for (j = 0; j < 2; j++) {
 217                     if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
 218                         int coef = av_clip_uintp2(quant(fabsf(in[i+j]), Q), 13);
 219                         int len = av_log2(coef);
 220
 221                         put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
 222                         put_sbits(pb, len, coef);
 223                     }
 224                 }
 225             }
 226         }
 227     }
 228
 229     if (bits)
 230         *bits = resbits;
 231     return cost;
 232 }
 233
 234 static float quantize_and_encode_band_cost_NONE(struct AACEncContext *s, PutBitContext *pb,
 235                                                 const float *in, const float *scaled,
 236                                                 int size, int scale_idx, int cb,
 237                                                 const float lambda, const float uplim,
 238                                                 int *bits) {
 239     av_assert0(0);
 240     return 0.0f;
 241 }
 242
 243 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE, BT_STEREO) \
 244 static float quantize_and_encode_band_cost_ ## NAME(                                         \
 245                                 struct AACEncContext *s,                                     \
 246                                 PutBitContext *pb, const float *in,                          \
 247                                 const float *scaled, int size, int scale_idx,                \
 248                                 int cb, const float lambda, const float uplim,               \
 249                                 int *bits) {                                                 \
 250     return quantize_and_encode_band_cost_template(                                           \
 251                                 s, pb, in, scaled, size, scale_idx,                          \
 252                                 BT_ESC ? ESC_BT : cb, lambda, uplim, bits,                   \
 253                                 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE, BT_STEREO); \
 254 }
 255
 256 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO,  1, 0, 0, 0, 0, 0)
 257 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0, 0, 0)
 258 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0, 0, 0)
 259 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0, 0, 0)
 260 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0, 0, 0)
 261 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC,   0, 1, 1, 1, 0, 0)
 262 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NOISE, 0, 0, 0, 0, 1, 0)
 263 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(STEREO,0, 0, 0, 0, 0, 1)
 264
 265 static float (*const quantize_and_encode_band_cost_arr[])(
 266                                 struct AACEncContext *s,
 267                                 PutBitContext *pb, const float *in,
 268                                 const float *scaled, int size, int scale_idx,
 269                                 int cb, const float lambda, const float uplim,
 270                                 int *bits) = {
 271     quantize_and_encode_band_cost_ZERO,
 272     quantize_and_encode_band_cost_SQUAD,
 273     quantize_and_encode_band_cost_SQUAD,
 274     quantize_and_encode_band_cost_UQUAD,
 275     quantize_and_encode_band_cost_UQUAD,
 276     quantize_and_encode_band_cost_SPAIR,
 277     quantize_and_encode_band_cost_SPAIR,
 278     quantize_and_encode_band_cost_UPAIR,
 279     quantize_and_encode_band_cost_UPAIR,
 280     quantize_and_encode_band_cost_UPAIR,
 281     quantize_and_encode_band_cost_UPAIR,
 282     quantize_and_encode_band_cost_ESC,
 283     quantize_and_encode_band_cost_NONE,     /* CB 12 doesn't exist */
 284     quantize_and_encode_band_cost_NOISE,
 285     quantize_and_encode_band_cost_STEREO,
 286     quantize_and_encode_band_cost_STEREO,
 287 };
 288
 289 #define quantize_and_encode_band_cost(                                  \
 290                                 s, pb, in, scaled, size, scale_idx, cb, \
 291                                 lambda, uplim, bits)                    \
 292     quantize_and_encode_band_cost_arr[cb](                              \
 293                                 s, pb, in, scaled, size, scale_idx, cb, \
 294                                 lambda, uplim, bits)
 295
 296 static float quantize_band_cost(struct AACEncContext *s, const float *in,
 297                                 const float *scaled, int size, int scale_idx,
 298                                 int cb, const float lambda, const float uplim,
 299                                 int *bits)
 300 {
 301     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
 302                                          cb, lambda, uplim, bits);
 303 }
 304
 305 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
 306                                      const float *in, int size, int scale_idx,
 307                                      int cb, const float lambda)
 308 {
 309     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
 310                                   INFINITY, NULL);
 311 }
 312
 313 static float find_max_val(int group_len, int swb_size, const float *scaled) {
 314     float maxval = 0.0f;
 315     int w2, i;
 316     for (w2 = 0; w2 < group_len; w2++) {
 317         for (i = 0; i < swb_size; i++) {
 318             maxval = FFMAX(maxval, scaled[w2*128+i]);
 319         }
 320     }
 321     return maxval;
 322 }
 323
 324 static int find_min_book(float maxval, int sf) {
 325     float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
 326     float Q34 = sqrtf(Q * sqrtf(Q));
 327     int qmaxval, cb;
 328     qmaxval = maxval * Q34 + 0.4054f;
 329     if      (qmaxval ==  0) cb = 0;
 330     else if (qmaxval ==  1) cb = 1;
 331     else if (qmaxval ==  2) cb = 3;
 332     else if (qmaxval <=  4) cb = 5;
 333     else if (qmaxval <=  7) cb = 7;
 334     else if (qmaxval <= 12) cb = 9;
 335     else                    cb = 11;
 336     return cb;
 337 }
 338
 339 /**
 340  * structure used in optimal codebook search
 341  */
 342 typedef struct BandCodingPath {
 343     int prev_idx; ///< pointer to the previous path point
 344     float cost;   ///< path cost
 345     int run;
 346 } BandCodingPath;
 347
 348 /**
 349  * Encode band info for single window group bands.
 350  */
 351 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
 352                                      int win, int group_len, const float lambda)
 353 {
 354     BandCodingPath path[120][CB_TOT_ALL];
 355     int w, swb, cb, start, size;
 356     int i, j;
 357     const int max_sfb  = sce->ics.max_sfb;
 358     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 359     const int run_esc  = (1 << run_bits) - 1;
 360     int idx, ppos, count;
 361     int stackrun[120], stackcb[120], stack_len;
 362     float next_minrd = INFINITY;
 363     int next_mincb = 0;
 364
 365     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 366     start = win*128;
 367     for (cb = 0; cb < CB_TOT_ALL; cb++) {
 368         path[0][cb].cost     = 0.0f;
 369         path[0][cb].prev_idx = -1;
 370         path[0][cb].run      = 0;
 371     }
 372     for (swb = 0; swb < max_sfb; swb++) {
 373         size = sce->ics.swb_sizes[swb];
 374         if (sce->zeroes[win*16 + swb]) {
 375             for (cb = 0; cb < CB_TOT_ALL; cb++) {
 376                 path[swb+1][cb].prev_idx = cb;
 377                 path[swb+1][cb].cost     = path[swb][cb].cost;
 378                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 379             }
 380         } else {
 381             float minrd = next_minrd;
 382             int mincb = next_mincb;
 383             next_minrd = INFINITY;
 384             next_mincb = 0;
 385             for (cb = 0; cb < CB_TOT_ALL; cb++) {
 386                 float cost_stay_here, cost_get_here;
 387                 float rd = 0.0f;
 388                 if (cb >= 12 && sce->band_type[win*16+swb] < aac_cb_out_map[cb] ||
 389                     cb  < aac_cb_in_map[sce->band_type[win*16+swb]] && sce->band_type[win*16+swb] > aac_cb_out_map[cb]) {
 390                     path[swb+1][cb].prev_idx = -1;
 391                     path[swb+1][cb].cost     = INFINITY;
 392                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 393                     continue;
 394                 }
 395                 for (w = 0; w < group_len; w++) {
 396                     FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
 397                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 398                                              s->scoefs + start + w*128, size,
 399                                              sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
 400                                              lambda / band->threshold, INFINITY, NULL);
 401                 }
 402                 cost_stay_here = path[swb][cb].cost + rd;
 403                 cost_get_here  = minrd              + rd + run_bits + 4;
 404                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 405                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 406                     cost_stay_here += run_bits;
 407                 if (cost_get_here < cost_stay_here) {
 408                     path[swb+1][cb].prev_idx = mincb;
 409                     path[swb+1][cb].cost     = cost_get_here;
 410                     path[swb+1][cb].run      = 1;
 411                 } else {
 412                     path[swb+1][cb].prev_idx = cb;
 413                     path[swb+1][cb].cost     = cost_stay_here;
 414                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 415                 }
 416                 if (path[swb+1][cb].cost < next_minrd) {
 417                     next_minrd = path[swb+1][cb].cost;
 418                     next_mincb = cb;
 419                 }
 420             }
 421         }
 422         start += sce->ics.swb_sizes[swb];
 423     }
 424
 425     //convert resulting path from backward-linked list
 426     stack_len = 0;
 427     idx       = 0;
 428     for (cb = 1; cb < CB_TOT_ALL; cb++)
 429         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 430             idx = cb;
 431     ppos = max_sfb;
 432     while (ppos > 0) {
 433         av_assert1(idx >= 0);
 434         cb = idx;
 435         stackrun[stack_len] = path[ppos][cb].run;
 436         stackcb [stack_len] = cb;
 437         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 438         ppos -= path[ppos][cb].run;
 439         stack_len++;
 440     }
 441     //perform actual band info encoding
 442     start = 0;
 443     for (i = stack_len - 1; i >= 0; i--) {
 444         cb = aac_cb_out_map[stackcb[i]];
 445         put_bits(&s->pb, 4, cb);
 446         count = stackrun[i];
 447         memset(sce->zeroes + win*16 + start, !cb, count);
 448         //XXX: memset when band_type is also uint8_t
 449         for (j = 0; j < count; j++) {
 450             sce->band_type[win*16 + start] = cb;
 451             start++;
 452         }
 453         while (count >= run_esc) {
 454             put_bits(&s->pb, run_bits, run_esc);
 455             count -= run_esc;
 456         }
 457         put_bits(&s->pb, run_bits, count);
 458     }
 459 }
 460
 461 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
 462                                   int win, int group_len, const float lambda)
 463 {
 464     BandCodingPath path[120][CB_TOT_ALL];
 465     int w, swb, cb, start, size;
 466     int i, j;
 467     const int max_sfb  = sce->ics.max_sfb;
 468     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 469     const int run_esc  = (1 << run_bits) - 1;
 470     int idx, ppos, count;
 471     int stackrun[120], stackcb[120], stack_len;
 472     float next_minbits = INFINITY;
 473     int next_mincb = 0;
 474
 475     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 476     start = win*128;
 477     for (cb = 0; cb < CB_TOT_ALL; cb++) {
 478         path[0][cb].cost     = run_bits+4;
 479         path[0][cb].prev_idx = -1;
 480         path[0][cb].run      = 0;
 481     }
 482     for (swb = 0; swb < max_sfb; swb++) {
 483         size = sce->ics.swb_sizes[swb];
 484         if (sce->zeroes[win*16 + swb]) {
 485             float cost_stay_here = path[swb][0].cost;
 486             float cost_get_here  = next_minbits + run_bits + 4;
 487             if (   run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
 488                 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
 489                 cost_stay_here += run_bits;
 490             if (cost_get_here < cost_stay_here) {
 491                 path[swb+1][0].prev_idx = next_mincb;
 492                 path[swb+1][0].cost     = cost_get_here;
 493                 path[swb+1][0].run      = 1;
 494             } else {
 495                 path[swb+1][0].prev_idx = 0;
 496                 path[swb+1][0].cost     = cost_stay_here;
 497                 path[swb+1][0].run      = path[swb][0].run + 1;
 498             }
 499             next_minbits = path[swb+1][0].cost;
 500             next_mincb = 0;
 501             for (cb = 1; cb < CB_TOT_ALL; cb++) {
 502                 path[swb+1][cb].cost = 61450;
 503                 path[swb+1][cb].prev_idx = -1;
 504                 path[swb+1][cb].run = 0;
 505             }
 506         } else {
 507             float minbits = next_minbits;
 508             int mincb = next_mincb;
 509             int startcb = sce->band_type[win*16+swb];
 510             startcb = aac_cb_in_map[startcb];
 511             next_minbits = INFINITY;
 512             next_mincb = 0;
 513             for (cb = 0; cb < startcb; cb++) {
 514                 path[swb+1][cb].cost = 61450;
 515                 path[swb+1][cb].prev_idx = -1;
 516                 path[swb+1][cb].run = 0;
 517             }
 518             for (cb = startcb; cb < CB_TOT_ALL; cb++) {
 519                 float cost_stay_here, cost_get_here;
 520                 float bits = 0.0f;
 521                 if (cb >= 12 && sce->band_type[win*16+swb] != aac_cb_out_map[cb]) {
 522                     path[swb+1][cb].cost = 61450;
 523                     path[swb+1][cb].prev_idx = -1;
 524                     path[swb+1][cb].run = 0;
 525                     continue;
 526                 }
 527                 for (w = 0; w < group_len; w++) {
 528                     bits += quantize_band_cost(s, sce->coeffs + start + w*128,
 529                                                s->scoefs + start + w*128, size,
 530                                                sce->sf_idx[(win+w)*16+swb],
 531                                                aac_cb_out_map[cb],
 532                                                0, INFINITY, NULL);
 533                 }
 534                 cost_stay_here = path[swb][cb].cost + bits;
 535                 cost_get_here  = minbits            + bits + run_bits + 4;
 536                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 537                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 538                     cost_stay_here += run_bits;
 539                 if (cost_get_here < cost_stay_here) {
 540                     path[swb+1][cb].prev_idx = mincb;
 541                     path[swb+1][cb].cost     = cost_get_here;
 542                     path[swb+1][cb].run      = 1;
 543                 } else {
 544                     path[swb+1][cb].prev_idx = cb;
 545                     path[swb+1][cb].cost     = cost_stay_here;
 546                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 547                 }
 548                 if (path[swb+1][cb].cost < next_minbits) {
 549                     next_minbits = path[swb+1][cb].cost;
 550                     next_mincb = cb;
 551                 }
 552             }
 553         }
 554         start += sce->ics.swb_sizes[swb];
 555     }
 556
 557     //convert resulting path from backward-linked list
 558     stack_len = 0;
 559     idx       = 0;
 560     for (cb = 1; cb < CB_TOT_ALL; cb++)
 561         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 562             idx = cb;
 563     ppos = max_sfb;
 564     while (ppos > 0) {
 565         av_assert1(idx >= 0);
 566         cb = idx;
 567         stackrun[stack_len] = path[ppos][cb].run;
 568         stackcb [stack_len] = cb;
 569         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 570         ppos -= path[ppos][cb].run;
 571         stack_len++;
 572     }
 573     //perform actual band info encoding
 574     start = 0;
 575     for (i = stack_len - 1; i >= 0; i--) {
 576         cb = aac_cb_out_map[stackcb[i]];
 577         put_bits(&s->pb, 4, cb);
 578         count = stackrun[i];
 579         memset(sce->zeroes + win*16 + start, !cb, count);
 580         //XXX: memset when band_type is also uint8_t
 581         for (j = 0; j < count; j++) {
 582             sce->band_type[win*16 + start] = cb;
 583             start++;
 584         }
 585         while (count >= run_esc) {
 586             put_bits(&s->pb, run_bits, run_esc);
 587             count -= run_esc;
 588         }
 589         put_bits(&s->pb, run_bits, count);
 590     }
 591 }
 592
 593 /** Return the minimum scalefactor where the quantized coef does not clip. */
 594 static av_always_inline uint8_t coef2minsf(float coef) {
 595     return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 596 }
 597
 598 /** Return the maximum scalefactor where the quantized coef is not zero. */
 599 static av_always_inline uint8_t coef2maxsf(float coef) {
 600     return av_clip_uint8(log2f(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 601 }
 602
 603 typedef struct TrellisPath {
 604     float cost;
 605     int prev;
 606 } TrellisPath;
 607
 608 #define TRELLIS_STAGES 121
 609 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
 610
 611 static void set_special_band_scalefactors(AACEncContext *s, SingleChannelElement *sce)
 612 {
 613     int w, g, start = 0;
 614     int minscaler_n = sce->sf_idx[0], minscaler_i = sce->sf_idx[0];
 615     int bands = 0;
 616
 617     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 618         start = 0;
 619         for (g = 0;  g < sce->ics.num_swb; g++) {
 620             if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
 621                 sce->sf_idx[w*16+g] = av_clip(ceilf(log2f(sce->is_ener[w*16+g])*2), -155, 100);
 622                 minscaler_i = FFMIN(minscaler_i, sce->sf_idx[w*16+g]);
 623                 bands++;
 624             } else if (sce->band_type[w*16+g] == NOISE_BT) {
 625                 sce->sf_idx[w*16+g] = av_clip(4+log2f(sce->pns_ener[w*16+g])*2, -100, 155);
 626                 minscaler_n = FFMIN(minscaler_n, sce->sf_idx[w*16+g]);
 627                 bands++;
 628             }
 629             start += sce->ics.swb_sizes[g];
 630         }
 631     }
 632
 633     if (!bands)
 634         return;
 635
 636     /* Clip the scalefactor indices */
 637     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 638         for (g = 0;  g < sce->ics.num_swb; g++) {
 639             if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
 640                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_i, minscaler_i + SCALE_MAX_DIFF);
 641             } else if (sce->band_type[w*16+g] == NOISE_BT) {
 642                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_n, minscaler_n + SCALE_MAX_DIFF);
 643             }
 644         }
 645     }
 646 }
 647
 648 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
 649                                        SingleChannelElement *sce,
 650                                        const float lambda)
 651 {
 652     int q, w, w2, g, start = 0;
 653     int i, j;
 654     int idx;
 655     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
 656     int bandaddr[TRELLIS_STAGES];
 657     int minq;
 658     float mincost;
 659     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
 660     int q0, q1, qcnt = 0;
 661
 662     for (i = 0; i < 1024; i++) {
 663         float t = fabsf(sce->coeffs[i]);
 664         if (t > 0.0f) {
 665             q0f = FFMIN(q0f, t);
 666             q1f = FFMAX(q1f, t);
 667             qnrgf += t*t;
 668             qcnt++;
 669         }
 670     }
 671
 672     if (!qcnt) {
 673         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 674         memset(sce->zeroes, 1, sizeof(sce->zeroes));
 675         return;
 676     }
 677
 678     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 679     q0 = coef2minsf(q0f);
 680     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 681     q1 = coef2maxsf(q1f);
 682     if (q1 - q0 > 60) {
 683         int q0low  = q0;
 684         int q1high = q1;
 685         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
 686         int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
 687         q1 = qnrg + 30;
 688         q0 = qnrg - 30;
 689         if (q0 < q0low) {
 690             q1 += q0low - q0;
 691             q0  = q0low;
 692         } else if (q1 > q1high) {
 693             q0 -= q1 - q1high;
 694             q1  = q1high;
 695         }
 696     }
 697
 698     for (i = 0; i < TRELLIS_STATES; i++) {
 699         paths[0][i].cost    = 0.0f;
 700         paths[0][i].prev    = -1;
 701     }
 702     for (j = 1; j < TRELLIS_STAGES; j++) {
 703         for (i = 0; i < TRELLIS_STATES; i++) {
 704             paths[j][i].cost    = INFINITY;
 705             paths[j][i].prev    = -2;
 706         }
 707     }
 708     idx = 1;
 709     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 710     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 711         start = w*128;
 712         for (g = 0; g < sce->ics.num_swb; g++) {
 713             const float *coefs = sce->coeffs + start;
 714             float qmin, qmax;
 715             int nz = 0;
 716
 717             bandaddr[idx] = w * 16 + g;
 718             qmin = INT_MAX;
 719             qmax = 0.0f;
 720             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 721                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 722                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 723                     sce->zeroes[(w+w2)*16+g] = 1;
 724                     continue;
 725                 }
 726                 sce->zeroes[(w+w2)*16+g] = 0;
 727                 nz = 1;
 728                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
 729                     float t = fabsf(coefs[w2*128+i]);
 730                     if (t > 0.0f)
 731                         qmin = FFMIN(qmin, t);
 732                     qmax = FFMAX(qmax, t);
 733                 }
 734             }
 735             if (nz) {
 736                 int minscale, maxscale;
 737                 float minrd = INFINITY;
 738                 float maxval;
 739                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 740                 minscale = coef2minsf(qmin);
 741                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 742                 maxscale = coef2maxsf(qmax);
 743                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
 744                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
 745                 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
 746                 for (q = minscale; q < maxscale; q++) {
 747                     float dist = 0;
 748                     int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
 749                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 750                         FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 751                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
 752                                                    q + q0, cb, lambda / band->threshold, INFINITY, NULL);
 753                     }
 754                     minrd = FFMIN(minrd, dist);
 755
 756                     for (i = 0; i < q1 - q0; i++) {
 757                         float cost;
 758                         cost = paths[idx - 1][i].cost + dist
 759                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 760                         if (cost < paths[idx][q].cost) {
 761                             paths[idx][q].cost    = cost;
 762                             paths[idx][q].prev    = i;
 763                         }
 764                     }
 765                 }
 766             } else {
 767                 for (q = 0; q < q1 - q0; q++) {
 768                     paths[idx][q].cost = paths[idx - 1][q].cost + 1;
 769                     paths[idx][q].prev = q;
 770                 }
 771             }
 772             sce->zeroes[w*16+g] = !nz;
 773             start += sce->ics.swb_sizes[g];
 774             idx++;
 775         }
 776     }
 777     idx--;
 778     mincost = paths[idx][0].cost;
 779     minq    = 0;
 780     for (i = 1; i < TRELLIS_STATES; i++) {
 781         if (paths[idx][i].cost < mincost) {
 782             mincost = paths[idx][i].cost;
 783             minq = i;
 784         }
 785     }
 786     while (idx) {
 787         sce->sf_idx[bandaddr[idx]] = minq + q0;
 788         minq = paths[idx][minq].prev;
 789         idx--;
 790     }
 791     //set the same quantizers inside window groups
 792     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 793         for (g = 0;  g < sce->ics.num_swb; g++)
 794             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 795                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 796 }
 797
 798 /**
 799  * two-loop quantizers search taken from ISO 13818-7 Appendix C
 800  */
 801 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
 802                                           AACEncContext *s,
 803                                           SingleChannelElement *sce,
 804                                           const float lambda)
 805 {
 806     int start = 0, i, w, w2, g;
 807     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels * (lambda / 120.f);
 808     float dists[128] = { 0 }, uplims[128] = { 0 };
 809     float maxvals[128];
 810     int fflag, minscaler;
 811     int its  = 0;
 812     int allz = 0;
 813     float minthr = INFINITY;
 814
 815     // for values above this the decoder might end up in an endless loop
 816     // due to always having more bits than what can be encoded.
 817     destbits = FFMIN(destbits, 5800);
 818     //XXX: some heuristic to determine initial quantizers will reduce search time
 819     //determine zero bands and upper limits
 820     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 821         for (g = 0;  g < sce->ics.num_swb; g++) {
 822             int nz = 0;
 823             float uplim = 0.0f, energy = 0.0f;
 824             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 825                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 826                 uplim  += band->threshold;
 827                 energy += band->energy;
 828                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 829                     sce->zeroes[(w+w2)*16+g] = 1;
 830                     continue;
 831                 }
 832                 nz = 1;
 833             }
 834             uplims[w*16+g] = uplim *512;
 835             sce->zeroes[w*16+g] = !nz;
 836             if (nz)
 837                 minthr = FFMIN(minthr, uplim);
 838             allz |= nz;
 839         }
 840     }
 841     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 842         for (g = 0;  g < sce->ics.num_swb; g++) {
 843             if (sce->zeroes[w*16+g]) {
 844                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
 845                 continue;
 846             }
 847             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
 848         }
 849     }
 850
 851     if (!allz)
 852         return;
 853     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 854
 855     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 856         start = w*128;
 857         for (g = 0;  g < sce->ics.num_swb; g++) {
 858             const float *scaled = s->scoefs + start;
 859             maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
 860             start += sce->ics.swb_sizes[g];
 861         }
 862     }
 863
 864     //perform two-loop search
 865     //outer loop - improve quality
 866     do {
 867         int tbits, qstep;
 868         minscaler = sce->sf_idx[0];
 869         //inner loop - quantize spectrum to fit into given number of bits
 870         qstep = its ? 1 : 32;
 871         do {
 872             int prev = -1;
 873             tbits = 0;
 874             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 875                 start = w*128;
 876                 for (g = 0;  g < sce->ics.num_swb; g++) {
 877                     const float *coefs = sce->coeffs + start;
 878                     const float *scaled = s->scoefs + start;
 879                     int bits = 0;
 880                     int cb;
 881                     float dist = 0.0f;
 882
 883                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
 884                         start += sce->ics.swb_sizes[g];
 885                         continue;
 886                     }
 887                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
 888                     cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
 889                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 890                         int b;
 891                         dist += quantize_band_cost(s, coefs + w2*128,
 892                                                    scaled + w2*128,
 893                                                    sce->ics.swb_sizes[g],
 894                                                    sce->sf_idx[w*16+g],
 895                                                    cb,
 896                                                    1.0f,
 897                                                    INFINITY,
 898                                                    &b);
 899                         bits += b;
 900                     }
 901                     dists[w*16+g] = dist - bits;
 902                     if (prev != -1) {
 903                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
 904                     }
 905                     tbits += bits;
 906                     start += sce->ics.swb_sizes[g];
 907                     prev = sce->sf_idx[w*16+g];
 908                 }
 909             }
 910             if (tbits > destbits) {
 911                 for (i = 0; i < 128; i++)
 912                     if (sce->sf_idx[i] < 218 - qstep)
 913                         sce->sf_idx[i] += qstep;
 914             } else {
 915                 for (i = 0; i < 128; i++)
 916                     if (sce->sf_idx[i] > 60 - qstep)
 917                         sce->sf_idx[i] -= qstep;
 918             }
 919             qstep >>= 1;
 920             if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
 921                 qstep = 1;
 922         } while (qstep);
 923
 924         fflag = 0;
 925         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
 926
 927         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 928             for (g = 0; g < sce->ics.num_swb; g++) {
 929                 int prevsc = sce->sf_idx[w*16+g];
 930                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
 931                     if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
 932                         sce->sf_idx[w*16+g]--;
 933                     else //Try to make sure there is some energy in every band
 934                         sce->sf_idx[w*16+g]-=2;
 935                 }
 936                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
 937                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
 938                 if (sce->sf_idx[w*16+g] != prevsc)
 939                     fflag = 1;
 940                 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
 941             }
 942         }
 943         its++;
 944     } while (fflag && its < 10);
 945 }
 946
 947 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
 948                                        SingleChannelElement *sce,
 949                                        const float lambda)
 950 {
 951     int start = 0, i, w, w2, g;
 952     float uplim[128], maxq[128];
 953     int minq, maxsf;
 954     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
 955     int last = 0, lastband = 0, curband = 0;
 956     float avg_energy = 0.0;
 957     if (sce->ics.num_windows == 1) {
 958         start = 0;
 959         for (i = 0; i < 1024; i++) {
 960             if (i - start >= sce->ics.swb_sizes[curband]) {
 961                 start += sce->ics.swb_sizes[curband];
 962                 curband++;
 963             }
 964             if (sce->coeffs[i]) {
 965                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
 966                 last = i;
 967                 lastband = curband;
 968             }
 969         }
 970     } else {
 971         for (w = 0; w < 8; w++) {
 972             const float *coeffs = sce->coeffs + w*128;
 973             curband = start = 0;
 974             for (i = 0; i < 128; i++) {
 975                 if (i - start >= sce->ics.swb_sizes[curband]) {
 976                     start += sce->ics.swb_sizes[curband];
 977                     curband++;
 978                 }
 979                 if (coeffs[i]) {
 980                     avg_energy += coeffs[i] * coeffs[i];
 981                     last = FFMAX(last, i);
 982                     lastband = FFMAX(lastband, curband);
 983                 }
 984             }
 985         }
 986     }
 987     last++;
 988     avg_energy /= last;
 989     if (avg_energy == 0.0f) {
 990         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
 991             sce->sf_idx[i] = SCALE_ONE_POS;
 992         return;
 993     }
 994     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 995         start = w*128;
 996         for (g = 0; g < sce->ics.num_swb; g++) {
 997             float *coefs   = sce->coeffs + start;
 998             const int size = sce->ics.swb_sizes[g];
 999             int start2 = start, end2 = start + size, peakpos = start;
1000             float maxval = -1, thr = 0.0f, t;
1001             maxq[w*16+g] = 0.0f;
1002             if (g > lastband) {
1003                 maxq[w*16+g] = 0.0f;
1004                 start += size;
1005                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
1006                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
1007                 continue;
1008             }
1009             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1010                 for (i = 0; i < size; i++) {
1011                     float t = coefs[w2*128+i]*coefs[w2*128+i];
1012                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
1013                     thr += t;
1014                     if (sce->ics.num_windows == 1 && maxval < t) {
1015                         maxval  = t;
1016                         peakpos = start+i;
1017                     }
1018                 }
1019             }
1020             if (sce->ics.num_windows == 1) {
1021                 start2 = FFMAX(peakpos - 2, start2);
1022                 end2   = FFMIN(peakpos + 3, end2);
1023             } else {
1024                 start2 -= start;
1025                 end2   -= start;
1026             }
1027             start += size;
1028             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
1029             t   = 1.0 - (1.0 * start2 / last);
1030             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
1031         }
1032     }
1033     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1034     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
1035     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1036         start = w*128;
1037         for (g = 0;  g < sce->ics.num_swb; g++) {
1038             const float *coefs  = sce->coeffs + start;
1039             const float *scaled = s->scoefs   + start;
1040             const int size      = sce->ics.swb_sizes[g];
1041             int scf, prev_scf, step;
1042             int min_scf = -1, max_scf = 256;
1043             float curdiff;
1044             if (maxq[w*16+g] < 21.544) {
1045                 sce->zeroes[w*16+g] = 1;
1046                 start += size;
1047                 continue;
1048             }
1049             sce->zeroes[w*16+g] = 0;
1050             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
1051             for (;;) {
1052                 float dist = 0.0f;
1053                 int quant_max;
1054
1055                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1056                     int b;
1057                     dist += quantize_band_cost(s, coefs + w2*128,
1058                                                scaled + w2*128,
1059                                                sce->ics.swb_sizes[g],
1060                                                scf,
1061                                                ESC_BT,
1062                                                lambda,
1063                                                INFINITY,
1064                                                &b);
1065                     dist -= b;
1066                 }
1067                 dist *= 1.0f / 512.0f / lambda;
1068                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
1069                 if (quant_max >= 8191) { // too much, return to the previous quantizer
1070                     sce->sf_idx[w*16+g] = prev_scf;
1071                     break;
1072                 }
1073                 prev_scf = scf;
1074                 curdiff = fabsf(dist - uplim[w*16+g]);
1075                 if (curdiff <= 1.0f)
1076                     step = 0;
1077                 else
1078                     step = log2f(curdiff);
1079                 if (dist > uplim[w*16+g])
1080                     step = -step;
1081                 scf += step;
1082                 scf = av_clip_uint8(scf);
1083                 step = scf - prev_scf;
1084                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
1085                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
1086                     break;
1087                 }
1088                 if (step > 0)
1089                     min_scf = prev_scf;
1090                 else
1091                     max_scf = prev_scf;
1092             }
1093             start += size;
1094         }
1095     }
1096     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1097     for (i = 1; i < 128; i++) {
1098         if (!sce->sf_idx[i])
1099             sce->sf_idx[i] = sce->sf_idx[i-1];
1100         else
1101             minq = FFMIN(minq, sce->sf_idx[i]);
1102     }
1103     if (minq == INT_MAX)
1104         minq = 0;
1105     minq = FFMIN(minq, SCALE_MAX_POS);
1106     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1107     for (i = 126; i >= 0; i--) {
1108         if (!sce->sf_idx[i])
1109             sce->sf_idx[i] = sce->sf_idx[i+1];
1110         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1111     }
1112 }
1113
1114 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1115                                        SingleChannelElement *sce,
1116                                        const float lambda)
1117 {
1118     int i, w, w2, g;
1119     int minq = 255;
1120
1121     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1122     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1123         for (g = 0; g < sce->ics.num_swb; g++) {
1124             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1125                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1126                 if (band->energy <= band->threshold) {
1127                     sce->sf_idx[(w+w2)*16+g] = 218;
1128                     sce->zeroes[(w+w2)*16+g] = 1;
1129                 } else {
1130                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1131                     sce->zeroes[(w+w2)*16+g] = 0;
1132                 }
1133                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1134             }
1135         }
1136     }
1137     for (i = 0; i < 128; i++) {
1138         sce->sf_idx[i] = 140;
1139         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1140     }
1141     //set the same quantizers inside window groups
1142     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1143         for (g = 0;  g < sce->ics.num_swb; g++)
1144             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1145                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1146 }
1147
1148 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce,
1149                            const float lambda)
1150 {
1151     int start = 0, w, w2, g;
1152     const float freq_mult = avctx->sample_rate/(1024.0f/sce->ics.num_windows)/2.0f;
1153     const float spread_threshold = NOISE_SPREAD_THRESHOLD*(lambda/120.f);
1154     const float thr_mult = NOISE_LAMBDA_NUMERATOR/lambda;
1155
1156     /* Coders !twoloop don't reset the band_types */
1157     for (w = 0; w < 128; w++)
1158         if (sce->band_type[w] == NOISE_BT)
1159             sce->band_type[w] = 0;
1160
1161     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1162         start = 0;
1163         for (g = 0;  g < sce->ics.num_swb; g++) {
1164             if (start*freq_mult > NOISE_LOW_LIMIT*(lambda/170.0f)) {
1165                 float energy = 0.0f, threshold = 0.0f, spread = 0.0f;
1166                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1167                     FFPsyBand *band = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1168                     energy += band->energy;
1169                     threshold += band->threshold;
1170                     spread += band->spread;
1171                 }
1172                 if (spread > spread_threshold*sce->ics.group_len[w] &&
1173                     ((sce->zeroes[w*16+g] && energy >= threshold) ||
1174                     energy < threshold*thr_mult*sce->ics.group_len[w])) {
1175                     sce->band_type[w*16+g] = NOISE_BT;
1176                     sce->pns_ener[w*16+g] = energy / sce->ics.group_len[w];
1177                     sce->zeroes[w*16+g] = 0;
1178                 }
1179             }
1180             start += sce->ics.swb_sizes[g];
1181         }
1182     }
1183 }
1184
1185 static void search_for_is(AACEncContext *s, AVCodecContext *avctx, ChannelElement *cpe,
1186                           const float lambda)
1187 {
1188     float IS[128];
1189     float *L34  = s->scoefs + 128*0, *R34  = s->scoefs + 128*1;
1190     float *I34  = s->scoefs + 128*2;
1191     SingleChannelElement *sce0 = &cpe->ch[0];
1192     SingleChannelElement *sce1 = &cpe->ch[1];
1193     int start = 0, count = 0, i, w, w2, g;
1194     const float freq_mult = avctx->sample_rate/(1024.0f/sce0->ics.num_windows)/2.0f;
1195
1196     for (w = 0; w < 128; w++)
1197         if (sce1->band_type[w] >= INTENSITY_BT2)
1198             sce1->band_type[w] = 0;
1199
1200     if (!cpe->common_window)
1201         return;
1202     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1203         start = 0;
1204         for (g = 0;  g < sce0->ics.num_swb; g++) {
1205             if (start*freq_mult > INT_STEREO_LOW_LIMIT*(lambda/170.0f) &&
1206                 cpe->ch[0].band_type[w*16+g] != NOISE_BT && !cpe->ch[0].zeroes[w*16+g] &&
1207                 cpe->ch[1].band_type[w*16+g] != NOISE_BT && !cpe->ch[1].zeroes[w*16+g]) {
1208                 int phase = 0;
1209                 float ener0 = 0.0f, ener1 = 0.0f, ener01 = 0.0f;
1210                 float dist1 = 0.0f, dist2 = 0.0f;
1211                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1212                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1213                         float coef0 = sce0->pcoeffs[start+(w+w2)*128+i];
1214                         float coef1 = sce1->pcoeffs[start+(w+w2)*128+i];
1215                         phase += coef0*coef1 >= 0.0f ? 1 : -1;
1216                         ener0 += coef0*coef0;
1217                         ener1 += coef1*coef1;
1218                         ener01 += (coef0 + coef1)*(coef0 + coef1);
1219                     }
1220                 }
1221                 if (!phase) { /* Too much phase difference between channels */
1222                     start += sce0->ics.swb_sizes[g];
1223                     continue;
1224                 }
1225                 phase = av_clip(phase, -1, 1);
1226                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1227                     FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1228                     FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1229                     int is_band_type, is_sf_idx = FFMAX(1, sce0->sf_idx[(w+w2)*16+g]-4);
1230                     float e01_34 = phase*pow(sqrt(ener1/ener0), 3.0/4.0);
1231                     float maxval, dist_spec_err = 0.0f;
1232                     float minthr = FFMIN(band0->threshold, band1->threshold);
1233                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++)
1234                         IS[i] = (sce0->pcoeffs[start+(w+w2)*128+i] + phase*sce1->pcoeffs[start+(w+w2)*128+i]) * sqrt(ener0/ener01);
1235                     abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
1236                     abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
1237                     abs_pow34_v(I34, IS,                            sce0->ics.swb_sizes[g]);
1238                     maxval = find_max_val(1, sce0->ics.swb_sizes[g], I34);
1239                     is_band_type = find_min_book(maxval, is_sf_idx);
1240                     dist1 += quantize_band_cost(s, sce0->coeffs + start + (w+w2)*128,
1241                                                 L34,
1242                                                 sce0->ics.swb_sizes[g],
1243                                                 sce0->sf_idx[(w+w2)*16+g],
1244                                                 sce0->band_type[(w+w2)*16+g],
1245                                                 lambda / band0->threshold, INFINITY, NULL);
1246                     dist1 += quantize_band_cost(s, sce1->coeffs + start + (w+w2)*128,
1247                                                 R34,
1248                                                 sce1->ics.swb_sizes[g],
1249                                                 sce1->sf_idx[(w+w2)*16+g],
1250                                                 sce1->band_type[(w+w2)*16+g],
1251                                                 lambda / band1->threshold, INFINITY, NULL);
1252                     dist2 += quantize_band_cost(s, IS,
1253                                                 I34,
1254                                                 sce0->ics.swb_sizes[g],
1255                                                 is_sf_idx,
1256                                                 is_band_type,
1257                                                 lambda / minthr, INFINITY, NULL);
1258                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1259                         dist_spec_err += (L34[i] - I34[i])*(L34[i] - I34[i]);
1260                         dist_spec_err += (R34[i] - I34[i]*e01_34)*(R34[i] - I34[i]*e01_34);
1261                     }
1262                     dist_spec_err *= lambda / minthr;
1263                     dist2 += dist_spec_err;
1264                 }
1265                 if (dist2 <= dist1) {
1266                     cpe->is_mask[w*16+g] = 1;
1267                     cpe->ms_mask[w*16+g] = 0;
1268                     cpe->ch[0].is_ener[w*16+g] = sqrt(ener0/ener01);
1269                     cpe->ch[1].is_ener[w*16+g] = ener0/ener1;
1270                     if (phase)
1271                         cpe->ch[1].band_type[w*16+g] = INTENSITY_BT;
1272                     else
1273                         cpe->ch[1].band_type[w*16+g] = INTENSITY_BT2;
1274                     count++;
1275                 }
1276             }
1277             start += sce0->ics.swb_sizes[g];
1278         }
1279     }
1280     cpe->is_mode = !!count;
1281 }
1282
1283 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1284                           const float lambda)
1285 {
1286     int start = 0, i, w, w2, g;
1287     float M[128], S[128];
1288     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1289     SingleChannelElement *sce0 = &cpe->ch[0];
1290     SingleChannelElement *sce1 = &cpe->ch[1];
1291     if (!cpe->common_window)
1292         return;
1293     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1294         start = 0;
1295         for (g = 0;  g < sce0->ics.num_swb; g++) {
1296             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g] && !cpe->is_mask[w*16+g]) {
1297                 float dist1 = 0.0f, dist2 = 0.0f;
1298                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1299                     FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1300                     FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1301                     float minthr = FFMIN(band0->threshold, band1->threshold);
1302                     float maxthr = FFMAX(band0->threshold, band1->threshold);
1303                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1304                         M[i] = (sce0->pcoeffs[start+(w+w2)*128+i]
1305                               + sce1->pcoeffs[start+(w+w2)*128+i]) * 0.5;
1306                         S[i] =  M[i]
1307                               - sce1->pcoeffs[start+(w+w2)*128+i];
1308                     }
1309                     abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
1310                     abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
1311                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
1312                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
1313                     dist1 += quantize_band_cost(s, sce0->coeffs + start + (w+w2)*128,
1314                                                 L34,
1315                                                 sce0->ics.swb_sizes[g],
1316                                                 sce0->sf_idx[(w+w2)*16+g],
1317                                                 sce0->band_type[(w+w2)*16+g],
1318                                                 lambda / band0->threshold, INFINITY, NULL);
1319                     dist1 += quantize_band_cost(s, sce1->coeffs + start + (w+w2)*128,
1320                                                 R34,
1321                                                 sce1->ics.swb_sizes[g],
1322                                                 sce1->sf_idx[(w+w2)*16+g],
1323                                                 sce1->band_type[(w+w2)*16+g],
1324                                                 lambda / band1->threshold, INFINITY, NULL);
1325                     dist2 += quantize_band_cost(s, M,
1326                                                 M34,
1327                                                 sce0->ics.swb_sizes[g],
1328                                                 sce0->sf_idx[(w+w2)*16+g],
1329                                                 sce0->band_type[(w+w2)*16+g],
1330                                                 lambda / maxthr, INFINITY, NULL);
1331                     dist2 += quantize_band_cost(s, S,
1332                                                 S34,
1333                                                 sce1->ics.swb_sizes[g],
1334                                                 sce1->sf_idx[(w+w2)*16+g],
1335                                                 sce1->band_type[(w+w2)*16+g],
1336                                                 lambda / minthr, INFINITY, NULL);
1337                 }
1338                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1339             }
1340             start += sce0->ics.swb_sizes[g];
1341         }
1342     }
1343 }
1344
1345 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
1346     [AAC_CODER_FAAC] = {
1347         search_for_quantizers_faac,
1348         encode_window_bands_info,
1349         quantize_and_encode_band,
1350         set_special_band_scalefactors,
1351         search_for_pns,
1352         search_for_ms,
1353         search_for_is,
1354     },
1355     [AAC_CODER_ANMR] = {
1356         search_for_quantizers_anmr,
1357         encode_window_bands_info,
1358         quantize_and_encode_band,
1359         set_special_band_scalefactors,
1360         search_for_pns,
1361         search_for_ms,
1362         search_for_is,
1363     },
1364     [AAC_CODER_TWOLOOP] = {
1365         search_for_quantizers_twoloop,
1366         codebook_trellis_rate,
1367         quantize_and_encode_band,
1368         set_special_band_scalefactors,
1369         search_for_pns,
1370         search_for_ms,
1371         search_for_is,
1372     },
1373     [AAC_CODER_FAST] = {
1374         search_for_quantizers_fast,
1375         encode_window_bands_info,
1376         quantize_and_encode_band,
1377         set_special_band_scalefactors,
1378         search_for_pns,
1379         search_for_ms,
1380         search_for_is,
1381     },
1382 };