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
  43 /** Frequency in Hz for lower limit of noise substitution **/
  44 #define NOISE_LOW_LIMIT 4000
  45
  46 /** Total number of usable codebooks **/
  47 #define CB_TOT 13
  48
  49 /** bits needed to code codebook run value for long windows */
  50 static const uint8_t run_value_bits_long[64] = {
  51      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
  52      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
  53     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
  54     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
  55 };
  56
  57 /** bits needed to code codebook run value for short windows */
  58 static const uint8_t run_value_bits_short[16] = {
  59     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
  60 };
  61
  62 static const uint8_t * const run_value_bits[2] = {
  63     run_value_bits_long, run_value_bits_short
  64 };
  65
  66 /** Map to convert values from BandCodingPath index to a codebook index **/
  67 static const uint8_t aac_cb_out_map[CB_TOT]  = {0,1,2,3,4,5,6,7,8,9,10,11,13};
  68 /** Inverse map to convert from codebooks to BandCodingPath indices **/
  69 static const uint8_t aac_cb_in_map[CB_TOT+1] = {0,1,2,3,4,5,6,7,8,9,10,11,0,12};
  70
  71 /**
  72  * Quantize one coefficient.
  73  * @return absolute value of the quantized coefficient
  74  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
  75  */
  76 static av_always_inline int quant(float coef, const float Q)
  77 {
  78     float a = coef * Q;
  79     return sqrtf(a * sqrtf(a)) + 0.4054;
  80 }
  81
  82 static void quantize_bands(int *out, const float *in, const float *scaled,
  83                            int size, float Q34, int is_signed, int maxval)
  84 {
  85     int i;
  86     double qc;
  87     for (i = 0; i < size; i++) {
  88         qc = scaled[i] * Q34;
  89         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
  90         if (is_signed && in[i] < 0.0f) {
  91             out[i] = -out[i];
  92         }
  93     }
  94 }
  95
  96 static void abs_pow34_v(float *out, const float *in, const int size)
  97 {
  98 #ifndef USE_REALLY_FULL_SEARCH
  99     int i;
 100     for (i = 0; i < size; i++) {
 101         float a = fabsf(in[i]);
 102         out[i] = sqrtf(a * sqrtf(a));
 103     }
 104 #endif /* USE_REALLY_FULL_SEARCH */
 105 }
 106
 107 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
 108 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
 109
 110 /**
 111  * Calculate rate distortion cost for quantizing with given codebook
 112  *
 113  * @return quantization distortion
 114  */
 115 static av_always_inline float quantize_and_encode_band_cost_template(
 116                                 struct AACEncContext *s,
 117                                 PutBitContext *pb, const float *in,
 118                                 const float *scaled, int size, int scale_idx,
 119                                 int cb, const float lambda, const float uplim,
 120                                 int *bits, int BT_ZERO, int BT_UNSIGNED,
 121                                 int BT_PAIR, int BT_ESC, int BT_NOISE)
 122 {
 123     const int q_idx = POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512;
 124     const float Q   = ff_aac_pow2sf_tab [q_idx];
 125     const float Q34 = ff_aac_pow34sf_tab[q_idx];
 126     const float IQ  = ff_aac_pow2sf_tab [POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
 127     const float CLIPPED_ESCAPE = 165140.0f*IQ;
 128     int i, j;
 129     float cost = 0;
 130     const int dim = BT_PAIR ? 2 : 4;
 131     int resbits = 0;
 132     int off;
 133
 134     if (BT_ZERO) {
 135         for (i = 0; i < size; i++)
 136             cost += in[i]*in[i];
 137         if (bits)
 138             *bits = 0;
 139         return cost * lambda;
 140     }
 141     if (BT_NOISE) {
 142         for (i = 0; i < size; i++)
 143             cost += in[i]*in[i];
 144         if (bits)
 145             *bits = 0;
 146         return cost * lambda;
 147     }
 148     if (!scaled) {
 149         abs_pow34_v(s->scoefs, in, size);
 150         scaled = s->scoefs;
 151     }
 152     quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, aac_cb_maxval[cb]);
 153     if (BT_UNSIGNED) {
 154         off = 0;
 155     } else {
 156         off = aac_cb_maxval[cb];
 157     }
 158     for (i = 0; i < size; i += dim) {
 159         const float *vec;
 160         int *quants = s->qcoefs + i;
 161         int curidx = 0;
 162         int curbits;
 163         float rd = 0.0f;
 164         for (j = 0; j < dim; j++) {
 165             curidx *= aac_cb_range[cb];
 166             curidx += quants[j] + off;
 167         }
 168         curbits =  ff_aac_spectral_bits[cb-1][curidx];
 169         vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
 170         if (BT_UNSIGNED) {
 171             for (j = 0; j < dim; j++) {
 172                 float t = fabsf(in[i+j]);
 173                 float di;
 174                 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
 175                     if (t >= CLIPPED_ESCAPE) {
 176                         di = t - CLIPPED_ESCAPE;
 177                         curbits += 21;
 178                     } else {
 179                         int c = av_clip_uintp2(quant(t, Q), 13);
 180                         di = t - c*cbrtf(c)*IQ;
 181                         curbits += av_log2(c)*2 - 4 + 1;
 182                     }
 183                 } else {
 184                     di = t - vec[j]*IQ;
 185                 }
 186                 if (vec[j] != 0.0f)
 187                     curbits++;
 188                 rd += di*di;
 189             }
 190         } else {
 191             for (j = 0; j < dim; j++) {
 192                 float di = in[i+j] - vec[j]*IQ;
 193                 rd += di*di;
 194             }
 195         }
 196         cost    += rd * lambda + curbits;
 197         resbits += curbits;
 198         if (cost >= uplim)
 199             return uplim;
 200         if (pb) {
 201             put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
 202             if (BT_UNSIGNED)
 203                 for (j = 0; j < dim; j++)
 204                     if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
 205                         put_bits(pb, 1, in[i+j] < 0.0f);
 206             if (BT_ESC) {
 207                 for (j = 0; j < 2; j++) {
 208                     if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
 209                         int coef = av_clip_uintp2(quant(fabsf(in[i+j]), Q), 13);
 210                         int len = av_log2(coef);
 211
 212                         put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
 213                         put_sbits(pb, len, coef);
 214                     }
 215                 }
 216             }
 217         }
 218     }
 219
 220     if (bits)
 221         *bits = resbits;
 222     return cost;
 223 }
 224
 225 static float quantize_and_encode_band_cost_NONE(struct AACEncContext *s, PutBitContext *pb,
 226                                                 const float *in, const float *scaled,
 227                                                 int size, int scale_idx, int cb,
 228                                                 const float lambda, const float uplim,
 229                                                 int *bits) {
 230     av_assert0(0);
 231     return 0.0f;
 232 }
 233
 234 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE) \
 235 static float quantize_and_encode_band_cost_ ## NAME(                                    \
 236                                 struct AACEncContext *s,                                \
 237                                 PutBitContext *pb, const float *in,                     \
 238                                 const float *scaled, int size, int scale_idx,           \
 239                                 int cb, const float lambda, const float uplim,          \
 240                                 int *bits) {                                            \
 241     return quantize_and_encode_band_cost_template(                                      \
 242                                 s, pb, in, scaled, size, scale_idx,                     \
 243                                 BT_ESC ? ESC_BT : cb, lambda, uplim, bits,              \
 244                                 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE);       \
 245 }
 246
 247 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO,  1, 0, 0, 0, 0)
 248 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0, 0)
 249 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0, 0)
 250 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0, 0)
 251 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0, 0)
 252 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC,   0, 1, 1, 1, 0)
 253 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NOISE, 0, 0, 0, 0, 1)
 254
 255 static float (*const quantize_and_encode_band_cost_arr[])(
 256                                 struct AACEncContext *s,
 257                                 PutBitContext *pb, const float *in,
 258                                 const float *scaled, int size, int scale_idx,
 259                                 int cb, const float lambda, const float uplim,
 260                                 int *bits) = {
 261     quantize_and_encode_band_cost_ZERO,
 262     quantize_and_encode_band_cost_SQUAD,
 263     quantize_and_encode_band_cost_SQUAD,
 264     quantize_and_encode_band_cost_UQUAD,
 265     quantize_and_encode_band_cost_UQUAD,
 266     quantize_and_encode_band_cost_SPAIR,
 267     quantize_and_encode_band_cost_SPAIR,
 268     quantize_and_encode_band_cost_UPAIR,
 269     quantize_and_encode_band_cost_UPAIR,
 270     quantize_and_encode_band_cost_UPAIR,
 271     quantize_and_encode_band_cost_UPAIR,
 272     quantize_and_encode_band_cost_ESC,
 273     quantize_and_encode_band_cost_NONE,     /* CB 12 doesn't exist */
 274     quantize_and_encode_band_cost_NOISE,
 275 };
 276
 277 #define quantize_and_encode_band_cost(                                  \
 278                                 s, pb, in, scaled, size, scale_idx, cb, \
 279                                 lambda, uplim, bits)                    \
 280     quantize_and_encode_band_cost_arr[cb](                              \
 281                                 s, pb, in, scaled, size, scale_idx, cb, \
 282                                 lambda, uplim, bits)
 283
 284 static float quantize_band_cost(struct AACEncContext *s, const float *in,
 285                                 const float *scaled, int size, int scale_idx,
 286                                 int cb, const float lambda, const float uplim,
 287                                 int *bits)
 288 {
 289     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
 290                                          cb, lambda, uplim, bits);
 291 }
 292
 293 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
 294                                      const float *in, int size, int scale_idx,
 295                                      int cb, const float lambda)
 296 {
 297     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
 298                                   INFINITY, NULL);
 299 }
 300
 301 static float find_max_val(int group_len, int swb_size, const float *scaled) {
 302     float maxval = 0.0f;
 303     int w2, i;
 304     for (w2 = 0; w2 < group_len; w2++) {
 305         for (i = 0; i < swb_size; i++) {
 306             maxval = FFMAX(maxval, scaled[w2*128+i]);
 307         }
 308     }
 309     return maxval;
 310 }
 311
 312 static int find_min_book(float maxval, int sf) {
 313     float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
 314     float Q34 = sqrtf(Q * sqrtf(Q));
 315     int qmaxval, cb;
 316     qmaxval = maxval * Q34 + 0.4054f;
 317     if      (qmaxval ==  0) cb = 0;
 318     else if (qmaxval ==  1) cb = 1;
 319     else if (qmaxval ==  2) cb = 3;
 320     else if (qmaxval <=  4) cb = 5;
 321     else if (qmaxval <=  7) cb = 7;
 322     else if (qmaxval <= 12) cb = 9;
 323     else                    cb = 11;
 324     return cb;
 325 }
 326
 327 /**
 328  * structure used in optimal codebook search
 329  */
 330 typedef struct BandCodingPath {
 331     int prev_idx; ///< pointer to the previous path point
 332     float cost;   ///< path cost
 333     int run;
 334 } BandCodingPath;
 335
 336 /**
 337  * Encode band info for single window group bands.
 338  */
 339 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
 340                                      int win, int group_len, const float lambda)
 341 {
 342     BandCodingPath path[120][CB_TOT];
 343     int w, swb, cb, start, size;
 344     int i, j;
 345     const int max_sfb  = sce->ics.max_sfb;
 346     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 347     const int run_esc  = (1 << run_bits) - 1;
 348     int idx, ppos, count;
 349     int stackrun[120], stackcb[120], stack_len;
 350     float next_minrd = INFINITY;
 351     int next_mincb = 0;
 352
 353     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 354     start = win*128;
 355     for (cb = 0; cb < CB_TOT; cb++) {
 356         path[0][cb].cost     = 0.0f;
 357         path[0][cb].prev_idx = -1;
 358         path[0][cb].run      = 0;
 359     }
 360     for (swb = 0; swb < max_sfb; swb++) {
 361         size = sce->ics.swb_sizes[swb];
 362         if (sce->zeroes[win*16 + swb]) {
 363             for (cb = 0; cb < CB_TOT; cb++) {
 364                 path[swb+1][cb].prev_idx = cb;
 365                 path[swb+1][cb].cost     = path[swb][cb].cost;
 366                 path[swb+1][cb].run      = path[swb][cb].run + 1;
 367             }
 368         } else {
 369             float minrd = next_minrd;
 370             int mincb = next_mincb;
 371             next_minrd = INFINITY;
 372             next_mincb = 0;
 373             for (cb = 0; cb < CB_TOT; cb++) {
 374                 float cost_stay_here, cost_get_here;
 375                 float rd = 0.0f;
 376                 for (w = 0; w < group_len; w++) {
 377                     FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
 378                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
 379                                              s->scoefs + start + w*128, size,
 380                                              sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
 381                                              lambda / band->threshold, INFINITY, NULL);
 382                 }
 383                 cost_stay_here = path[swb][cb].cost + rd;
 384                 cost_get_here  = minrd              + rd + run_bits + 4;
 385                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 386                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 387                     cost_stay_here += run_bits;
 388                 if (cost_get_here < cost_stay_here) {
 389                     path[swb+1][cb].prev_idx = mincb;
 390                     path[swb+1][cb].cost     = cost_get_here;
 391                     path[swb+1][cb].run      = 1;
 392                 } else {
 393                     path[swb+1][cb].prev_idx = cb;
 394                     path[swb+1][cb].cost     = cost_stay_here;
 395                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 396                 }
 397                 if (path[swb+1][cb].cost < next_minrd) {
 398                     next_minrd = path[swb+1][cb].cost;
 399                     next_mincb = cb;
 400                 }
 401             }
 402         }
 403         start += sce->ics.swb_sizes[swb];
 404     }
 405
 406     //convert resulting path from backward-linked list
 407     stack_len = 0;
 408     idx       = 0;
 409     for (cb = 1; cb < CB_TOT; cb++)
 410         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 411             idx = cb;
 412     ppos = max_sfb;
 413     while (ppos > 0) {
 414         cb = idx;
 415         stackrun[stack_len] = path[ppos][cb].run;
 416         stackcb [stack_len] = cb;
 417         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 418         ppos -= path[ppos][cb].run;
 419         stack_len++;
 420     }
 421     //perform actual band info encoding
 422     start = 0;
 423     for (i = stack_len - 1; i >= 0; i--) {
 424         cb = aac_cb_out_map[stackcb[i]];
 425         put_bits(&s->pb, 4, cb);
 426         count = stackrun[i];
 427         memset(sce->zeroes + win*16 + start, !cb, count);
 428         //XXX: memset when band_type is also uint8_t
 429         for (j = 0; j < count; j++) {
 430             sce->band_type[win*16 + start] = cb;
 431             start++;
 432         }
 433         while (count >= run_esc) {
 434             put_bits(&s->pb, run_bits, run_esc);
 435             count -= run_esc;
 436         }
 437         put_bits(&s->pb, run_bits, count);
 438     }
 439 }
 440
 441 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
 442                                   int win, int group_len, const float lambda)
 443 {
 444     BandCodingPath path[120][CB_TOT];
 445     int w, swb, cb, start, size;
 446     int i, j;
 447     const int max_sfb  = sce->ics.max_sfb;
 448     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
 449     const int run_esc  = (1 << run_bits) - 1;
 450     int idx, ppos, count;
 451     int stackrun[120], stackcb[120], stack_len;
 452     float next_minbits = INFINITY;
 453     int next_mincb = 0;
 454
 455     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 456     start = win*128;
 457     for (cb = 0; cb < CB_TOT; cb++) {
 458         path[0][cb].cost     = run_bits+4;
 459         path[0][cb].prev_idx = -1;
 460         path[0][cb].run      = 0;
 461     }
 462     for (swb = 0; swb < max_sfb; swb++) {
 463         size = sce->ics.swb_sizes[swb];
 464         if (sce->zeroes[win*16 + swb]) {
 465             float cost_stay_here = path[swb][0].cost;
 466             float cost_get_here  = next_minbits + run_bits + 4;
 467             if (   run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
 468                 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
 469                 cost_stay_here += run_bits;
 470             if (cost_get_here < cost_stay_here) {
 471                 path[swb+1][0].prev_idx = next_mincb;
 472                 path[swb+1][0].cost     = cost_get_here;
 473                 path[swb+1][0].run      = 1;
 474             } else {
 475                 path[swb+1][0].prev_idx = 0;
 476                 path[swb+1][0].cost     = cost_stay_here;
 477                 path[swb+1][0].run      = path[swb][0].run + 1;
 478             }
 479             next_minbits = path[swb+1][0].cost;
 480             next_mincb = 0;
 481             for (cb = 1; cb < CB_TOT; cb++) {
 482                 path[swb+1][cb].cost = 61450;
 483                 path[swb+1][cb].prev_idx = -1;
 484                 path[swb+1][cb].run = 0;
 485             }
 486         } else {
 487             float minbits = next_minbits;
 488             int mincb = next_mincb;
 489             int startcb = sce->band_type[win*16+swb];
 490             startcb = aac_cb_in_map[startcb];
 491             next_minbits = INFINITY;
 492             next_mincb = 0;
 493             for (cb = 0; cb < startcb; cb++) {
 494                 path[swb+1][cb].cost = 61450;
 495                 path[swb+1][cb].prev_idx = -1;
 496                 path[swb+1][cb].run = 0;
 497             }
 498             for (cb = startcb; cb < CB_TOT; cb++) {
 499                 float cost_stay_here, cost_get_here;
 500                 float bits = 0.0f;
 501                 if (cb == 12 && sce->band_type[win*16+swb] != NOISE_BT) {
 502                     path[swb+1][cb].cost = 61450;
 503                     path[swb+1][cb].prev_idx = -1;
 504                     path[swb+1][cb].run = 0;
 505                     continue;
 506                 }
 507                 for (w = 0; w < group_len; w++) {
 508                     bits += quantize_band_cost(s, sce->coeffs + start + w*128,
 509                                                s->scoefs + start + w*128, size,
 510                                                sce->sf_idx[(win+w)*16+swb],
 511                                                aac_cb_out_map[cb],
 512                                                0, INFINITY, NULL);
 513                 }
 514                 cost_stay_here = path[swb][cb].cost + bits;
 515                 cost_get_here  = minbits            + bits + run_bits + 4;
 516                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
 517                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
 518                     cost_stay_here += run_bits;
 519                 if (cost_get_here < cost_stay_here) {
 520                     path[swb+1][cb].prev_idx = mincb;
 521                     path[swb+1][cb].cost     = cost_get_here;
 522                     path[swb+1][cb].run      = 1;
 523                 } else {
 524                     path[swb+1][cb].prev_idx = cb;
 525                     path[swb+1][cb].cost     = cost_stay_here;
 526                     path[swb+1][cb].run      = path[swb][cb].run + 1;
 527                 }
 528                 if (path[swb+1][cb].cost < next_minbits) {
 529                     next_minbits = path[swb+1][cb].cost;
 530                     next_mincb = cb;
 531                 }
 532             }
 533         }
 534         start += sce->ics.swb_sizes[swb];
 535     }
 536
 537     //convert resulting path from backward-linked list
 538     stack_len = 0;
 539     idx       = 0;
 540     for (cb = 1; cb < CB_TOT; cb++)
 541         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
 542             idx = cb;
 543     ppos = max_sfb;
 544     while (ppos > 0) {
 545         av_assert1(idx >= 0);
 546         cb = idx;
 547         stackrun[stack_len] = path[ppos][cb].run;
 548         stackcb [stack_len] = cb;
 549         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
 550         ppos -= path[ppos][cb].run;
 551         stack_len++;
 552     }
 553     //perform actual band info encoding
 554     start = 0;
 555     for (i = stack_len - 1; i >= 0; i--) {
 556         cb = aac_cb_out_map[stackcb[i]];
 557         put_bits(&s->pb, 4, cb);
 558         count = stackrun[i];
 559         memset(sce->zeroes + win*16 + start, !cb, count);
 560         //XXX: memset when band_type is also uint8_t
 561         for (j = 0; j < count; j++) {
 562             sce->band_type[win*16 + start] = cb;
 563             start++;
 564         }
 565         while (count >= run_esc) {
 566             put_bits(&s->pb, run_bits, run_esc);
 567             count -= run_esc;
 568         }
 569         put_bits(&s->pb, run_bits, count);
 570     }
 571 }
 572
 573 /** Return the minimum scalefactor where the quantized coef does not clip. */
 574 static av_always_inline uint8_t coef2minsf(float coef) {
 575     return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 576 }
 577
 578 /** Return the maximum scalefactor where the quantized coef is not zero. */
 579 static av_always_inline uint8_t coef2maxsf(float coef) {
 580     return av_clip_uint8(log2f(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 581 }
 582
 583 typedef struct TrellisPath {
 584     float cost;
 585     int prev;
 586 } TrellisPath;
 587
 588 #define TRELLIS_STAGES 121
 589 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
 590
 591 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
 592                                        SingleChannelElement *sce,
 593                                        const float lambda)
 594 {
 595     int q, w, w2, g, start = 0;
 596     int i, j;
 597     int idx;
 598     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
 599     int bandaddr[TRELLIS_STAGES];
 600     int minq;
 601     float mincost;
 602     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
 603     int q0, q1, qcnt = 0;
 604
 605     for (i = 0; i < 1024; i++) {
 606         float t = fabsf(sce->coeffs[i]);
 607         if (t > 0.0f) {
 608             q0f = FFMIN(q0f, t);
 609             q1f = FFMAX(q1f, t);
 610             qnrgf += t*t;
 611             qcnt++;
 612         }
 613     }
 614
 615     if (!qcnt) {
 616         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
 617         memset(sce->zeroes, 1, sizeof(sce->zeroes));
 618         return;
 619     }
 620
 621     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 622     q0 = coef2minsf(q0f);
 623     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 624     q1 = coef2maxsf(q1f);
 625     if (q1 - q0 > 60) {
 626         int q0low  = q0;
 627         int q1high = q1;
 628         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
 629         int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
 630         q1 = qnrg + 30;
 631         q0 = qnrg - 30;
 632         if (q0 < q0low) {
 633             q1 += q0low - q0;
 634             q0  = q0low;
 635         } else if (q1 > q1high) {
 636             q0 -= q1 - q1high;
 637             q1  = q1high;
 638         }
 639     }
 640
 641     for (i = 0; i < TRELLIS_STATES; i++) {
 642         paths[0][i].cost    = 0.0f;
 643         paths[0][i].prev    = -1;
 644     }
 645     for (j = 1; j < TRELLIS_STAGES; j++) {
 646         for (i = 0; i < TRELLIS_STATES; i++) {
 647             paths[j][i].cost    = INFINITY;
 648             paths[j][i].prev    = -2;
 649         }
 650     }
 651     idx = 1;
 652     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 653     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 654         start = w*128;
 655         for (g = 0; g < sce->ics.num_swb; g++) {
 656             const float *coefs = sce->coeffs + start;
 657             float qmin, qmax;
 658             int nz = 0;
 659
 660             bandaddr[idx] = w * 16 + g;
 661             qmin = INT_MAX;
 662             qmax = 0.0f;
 663             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 664                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 665                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 666                     sce->zeroes[(w+w2)*16+g] = 1;
 667                     continue;
 668                 }
 669                 sce->zeroes[(w+w2)*16+g] = 0;
 670                 nz = 1;
 671                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
 672                     float t = fabsf(coefs[w2*128+i]);
 673                     if (t > 0.0f)
 674                         qmin = FFMIN(qmin, t);
 675                     qmax = FFMAX(qmax, t);
 676                 }
 677             }
 678             if (nz) {
 679                 int minscale, maxscale;
 680                 float minrd = INFINITY;
 681                 float maxval;
 682                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
 683                 minscale = coef2minsf(qmin);
 684                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
 685                 maxscale = coef2maxsf(qmax);
 686                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
 687                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
 688                 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
 689                 for (q = minscale; q < maxscale; q++) {
 690                     float dist = 0;
 691                     int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
 692                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 693                         FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 694                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
 695                                                    q + q0, cb, lambda / band->threshold, INFINITY, NULL);
 696                     }
 697                     minrd = FFMIN(minrd, dist);
 698
 699                     for (i = 0; i < q1 - q0; i++) {
 700                         float cost;
 701                         cost = paths[idx - 1][i].cost + dist
 702                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
 703                         if (cost < paths[idx][q].cost) {
 704                             paths[idx][q].cost    = cost;
 705                             paths[idx][q].prev    = i;
 706                         }
 707                     }
 708                 }
 709             } else {
 710                 for (q = 0; q < q1 - q0; q++) {
 711                     paths[idx][q].cost = paths[idx - 1][q].cost + 1;
 712                     paths[idx][q].prev = q;
 713                 }
 714             }
 715             sce->zeroes[w*16+g] = !nz;
 716             start += sce->ics.swb_sizes[g];
 717             idx++;
 718         }
 719     }
 720     idx--;
 721     mincost = paths[idx][0].cost;
 722     minq    = 0;
 723     for (i = 1; i < TRELLIS_STATES; i++) {
 724         if (paths[idx][i].cost < mincost) {
 725             mincost = paths[idx][i].cost;
 726             minq = i;
 727         }
 728     }
 729     while (idx) {
 730         sce->sf_idx[bandaddr[idx]] = minq + q0;
 731         minq = paths[idx][minq].prev;
 732         idx--;
 733     }
 734     //set the same quantizers inside window groups
 735     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 736         for (g = 0;  g < sce->ics.num_swb; g++)
 737             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
 738                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 739 }
 740
 741 /**
 742  * two-loop quantizers search taken from ISO 13818-7 Appendix C
 743  */
 744 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
 745                                           AACEncContext *s,
 746                                           SingleChannelElement *sce,
 747                                           const float lambda)
 748 {
 749     int start = 0, i, w, w2, g;
 750     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels * (lambda / 120.f);
 751     const float freq_mult = avctx->sample_rate/(1024.0f/sce->ics.num_windows)/2.0f;
 752     float dists[128] = { 0 }, uplims[128] = { 0 };
 753     float maxvals[128];
 754     int noise_sf[128] = { 0 };
 755     int fflag, minscaler, minscaler_n;
 756     int its  = 0;
 757     int allz = 0;
 758     float minthr = INFINITY;
 759
 760     // for values above this the decoder might end up in an endless loop
 761     // due to always having more bits than what can be encoded.
 762     destbits = FFMIN(destbits, 5800);
 763     //XXX: some heuristic to determine initial quantizers will reduce search time
 764     //determine zero bands and upper limits
 765     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 766         start = 0;
 767         for (g = 0;  g < sce->ics.num_swb; g++) {
 768             int nz = 0;
 769             float uplim = 0.0f, energy = 0.0f;
 770             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 771                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
 772                 uplim += band->threshold;
 773                 energy += band->energy;
 774                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
 775                     sce->zeroes[(w+w2)*16+g] = 1;
 776                     continue;
 777                 }
 778                 nz = 1;
 779             }
 780             uplims[w*16+g] = uplim *512;
 781             if (s->options.pns && start*freq_mult > NOISE_LOW_LIMIT && energy < uplim * 1.2f) {
 782                 noise_sf[w*16+g] = av_clip(4+FFMIN(log2f(energy)*2,255), -100, 155);
 783                 sce->band_type[w*16+g] = NOISE_BT;
 784                 nz= 1;
 785             } else { /** Band type will be determined by the twoloop algorithm */
 786                 sce->band_type[w*16+g] = 0;
 787             }
 788             sce->zeroes[w*16+g] = !nz;
 789             if (nz)
 790                 minthr = FFMIN(minthr, uplim);
 791             allz |= nz;
 792             start += sce->ics.swb_sizes[g];
 793         }
 794     }
 795     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 796         for (g = 0;  g < sce->ics.num_swb; g++) {
 797             if (sce->zeroes[w*16+g]) {
 798                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
 799                 continue;
 800             }
 801             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
 802         }
 803     }
 804
 805     if (!allz)
 806         return;
 807     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
 808
 809     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 810         start = w*128;
 811         for (g = 0;  g < sce->ics.num_swb; g++) {
 812             const float *scaled = s->scoefs + start;
 813             maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
 814             start += sce->ics.swb_sizes[g];
 815         }
 816     }
 817
 818     //perform two-loop search
 819     //outer loop - improve quality
 820     do {
 821         int tbits, qstep;
 822         minscaler = sce->sf_idx[0];
 823         minscaler_n = sce->sf_idx[0];
 824         //inner loop - quantize spectrum to fit into given number of bits
 825         qstep = its ? 1 : 32;
 826         do {
 827             int prev = -1;
 828             tbits = 0;
 829             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 830                 start = w*128;
 831                 for (g = 0;  g < sce->ics.num_swb; g++) {
 832                     const float *coefs = sce->coeffs + start;
 833                     const float *scaled = s->scoefs + start;
 834                     int bits = 0;
 835                     int cb;
 836                     float dist = 0.0f;
 837
 838                     if (sce->band_type[w*16+g] == NOISE_BT) {
 839                         minscaler_n = FFMIN(minscaler_n, noise_sf[w*16+g]);
 840                         start += sce->ics.swb_sizes[g];
 841                         continue;
 842                     } else if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
 843                         start += sce->ics.swb_sizes[g];
 844                         continue;
 845                     }
 846                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
 847                     cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
 848                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 849                         int b;
 850                         dist += quantize_band_cost(s, coefs + w2*128,
 851                                                    scaled + w2*128,
 852                                                    sce->ics.swb_sizes[g],
 853                                                    sce->sf_idx[w*16+g],
 854                                                    cb,
 855                                                    1.0f,
 856                                                    INFINITY,
 857                                                    &b);
 858                         bits += b;
 859                     }
 860                     dists[w*16+g] = dist - bits;
 861                     if (prev != -1) {
 862                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
 863                     }
 864                     tbits += bits;
 865                     start += sce->ics.swb_sizes[g];
 866                     prev = sce->sf_idx[w*16+g];
 867                 }
 868             }
 869             if (tbits > destbits) {
 870                 for (i = 0; i < 128; i++)
 871                     if (sce->sf_idx[i] < 218 - qstep)
 872                         sce->sf_idx[i] += qstep;
 873             } else {
 874                 for (i = 0; i < 128; i++)
 875                     if (sce->sf_idx[i] > 60 - qstep)
 876                         sce->sf_idx[i] -= qstep;
 877             }
 878             qstep >>= 1;
 879             if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
 880                 qstep = 1;
 881         } while (qstep);
 882
 883         fflag = 0;
 884         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
 885
 886         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
 887             for (g = 0; g < sce->ics.num_swb; g++)
 888                 if (sce->band_type[w*16+g] == NOISE_BT)
 889                     sce->sf_idx[w*16+g] = av_clip(noise_sf[w*16+g], minscaler_n, minscaler_n + SCALE_MAX_DIFF);
 890
 891         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 892             for (g = 0; g < sce->ics.num_swb; g++) {
 893                 int prevsc = sce->sf_idx[w*16+g];
 894                 if (sce->band_type[w*16+g] == NOISE_BT)
 895                     continue;
 896                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
 897                     if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
 898                         sce->sf_idx[w*16+g]--;
 899                     else //Try to make sure there is some energy in every band
 900                         sce->sf_idx[w*16+g]-=2;
 901                 }
 902                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
 903                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
 904                 if (sce->sf_idx[w*16+g] != prevsc)
 905                     fflag = 1;
 906                 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
 907             }
 908         }
 909         its++;
 910     } while (fflag && its < 10);
 911 }
 912
 913 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
 914                                        SingleChannelElement *sce,
 915                                        const float lambda)
 916 {
 917     int start = 0, i, w, w2, g;
 918     float uplim[128], maxq[128];
 919     int minq, maxsf;
 920     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
 921     int last = 0, lastband = 0, curband = 0;
 922     float avg_energy = 0.0;
 923     if (sce->ics.num_windows == 1) {
 924         start = 0;
 925         for (i = 0; i < 1024; i++) {
 926             if (i - start >= sce->ics.swb_sizes[curband]) {
 927                 start += sce->ics.swb_sizes[curband];
 928                 curband++;
 929             }
 930             if (sce->coeffs[i]) {
 931                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
 932                 last = i;
 933                 lastband = curband;
 934             }
 935         }
 936     } else {
 937         for (w = 0; w < 8; w++) {
 938             const float *coeffs = sce->coeffs + w*128;
 939             curband = start = 0;
 940             for (i = 0; i < 128; i++) {
 941                 if (i - start >= sce->ics.swb_sizes[curband]) {
 942                     start += sce->ics.swb_sizes[curband];
 943                     curband++;
 944                 }
 945                 if (coeffs[i]) {
 946                     avg_energy += coeffs[i] * coeffs[i];
 947                     last = FFMAX(last, i);
 948                     lastband = FFMAX(lastband, curband);
 949                 }
 950             }
 951         }
 952     }
 953     last++;
 954     avg_energy /= last;
 955     if (avg_energy == 0.0f) {
 956         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
 957             sce->sf_idx[i] = SCALE_ONE_POS;
 958         return;
 959     }
 960     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
 961         start = w*128;
 962         for (g = 0; g < sce->ics.num_swb; g++) {
 963             float *coefs   = sce->coeffs + start;
 964             const int size = sce->ics.swb_sizes[g];
 965             int start2 = start, end2 = start + size, peakpos = start;
 966             float maxval = -1, thr = 0.0f, t;
 967             maxq[w*16+g] = 0.0f;
 968             if (g > lastband) {
 969                 maxq[w*16+g] = 0.0f;
 970                 start += size;
 971                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
 972                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
 973                 continue;
 974             }
 975             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
 976                 for (i = 0; i < size; i++) {
 977                     float t = coefs[w2*128+i]*coefs[w2*128+i];
 978                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
 979                     thr += t;
 980                     if (sce->ics.num_windows == 1 && maxval < t) {
 981                         maxval  = t;
 982                         peakpos = start+i;
 983                     }
 984                 }
 985             }
 986             if (sce->ics.num_windows == 1) {
 987                 start2 = FFMAX(peakpos - 2, start2);
 988                 end2   = FFMIN(peakpos + 3, end2);
 989             } else {
 990                 start2 -= start;
 991                 end2   -= start;
 992             }
 993             start += size;
 994             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
 995             t   = 1.0 - (1.0 * start2 / last);
 996             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
 997         }
 998     }
 999     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1000     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
1001     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1002         start = w*128;
1003         for (g = 0;  g < sce->ics.num_swb; g++) {
1004             const float *coefs  = sce->coeffs + start;
1005             const float *scaled = s->scoefs   + start;
1006             const int size      = sce->ics.swb_sizes[g];
1007             int scf, prev_scf, step;
1008             int min_scf = -1, max_scf = 256;
1009             float curdiff;
1010             if (maxq[w*16+g] < 21.544) {
1011                 sce->zeroes[w*16+g] = 1;
1012                 start += size;
1013                 continue;
1014             }
1015             sce->zeroes[w*16+g] = 0;
1016             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
1017             for (;;) {
1018                 float dist = 0.0f;
1019                 int quant_max;
1020
1021                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1022                     int b;
1023                     dist += quantize_band_cost(s, coefs + w2*128,
1024                                                scaled + w2*128,
1025                                                sce->ics.swb_sizes[g],
1026                                                scf,
1027                                                ESC_BT,
1028                                                lambda,
1029                                                INFINITY,
1030                                                &b);
1031                     dist -= b;
1032                 }
1033                 dist *= 1.0f / 512.0f / lambda;
1034                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
1035                 if (quant_max >= 8191) { // too much, return to the previous quantizer
1036                     sce->sf_idx[w*16+g] = prev_scf;
1037                     break;
1038                 }
1039                 prev_scf = scf;
1040                 curdiff = fabsf(dist - uplim[w*16+g]);
1041                 if (curdiff <= 1.0f)
1042                     step = 0;
1043                 else
1044                     step = log2f(curdiff);
1045                 if (dist > uplim[w*16+g])
1046                     step = -step;
1047                 scf += step;
1048                 scf = av_clip_uint8(scf);
1049                 step = scf - prev_scf;
1050                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
1051                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
1052                     break;
1053                 }
1054                 if (step > 0)
1055                     min_scf = prev_scf;
1056                 else
1057                     max_scf = prev_scf;
1058             }
1059             start += size;
1060         }
1061     }
1062     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1063     for (i = 1; i < 128; i++) {
1064         if (!sce->sf_idx[i])
1065             sce->sf_idx[i] = sce->sf_idx[i-1];
1066         else
1067             minq = FFMIN(minq, sce->sf_idx[i]);
1068     }
1069     if (minq == INT_MAX)
1070         minq = 0;
1071     minq = FFMIN(minq, SCALE_MAX_POS);
1072     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1073     for (i = 126; i >= 0; i--) {
1074         if (!sce->sf_idx[i])
1075             sce->sf_idx[i] = sce->sf_idx[i+1];
1076         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1077     }
1078 }
1079
1080 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1081                                        SingleChannelElement *sce,
1082                                        const float lambda)
1083 {
1084     int i, w, w2, g;
1085     int minq = 255;
1086
1087     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1088     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1089         for (g = 0; g < sce->ics.num_swb; g++) {
1090             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1091                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1092                 if (band->energy <= band->threshold) {
1093                     sce->sf_idx[(w+w2)*16+g] = 218;
1094                     sce->zeroes[(w+w2)*16+g] = 1;
1095                 } else {
1096                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1097                     sce->zeroes[(w+w2)*16+g] = 0;
1098                 }
1099                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1100             }
1101         }
1102     }
1103     for (i = 0; i < 128; i++) {
1104         sce->sf_idx[i] = 140;
1105         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1106     }
1107     //set the same quantizers inside window groups
1108     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1109         for (g = 0;  g < sce->ics.num_swb; g++)
1110             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1111                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1112 }
1113
1114 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1115                           const float lambda)
1116 {
1117     int start = 0, i, w, w2, g;
1118     float M[128], S[128];
1119     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1120     SingleChannelElement *sce0 = &cpe->ch[0];
1121     SingleChannelElement *sce1 = &cpe->ch[1];
1122     if (!cpe->common_window)
1123         return;
1124     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1125         for (g = 0;  g < sce0->ics.num_swb; g++) {
1126             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1127                 float dist1 = 0.0f, dist2 = 0.0f;
1128                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1129                     FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1130                     FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1131                     float minthr = FFMIN(band0->threshold, band1->threshold);
1132                     float maxthr = FFMAX(band0->threshold, band1->threshold);
1133                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1134                         M[i] = (sce0->pcoeffs[start+w2*128+i]
1135                               + sce1->pcoeffs[start+w2*128+i]) * 0.5;
1136                         S[i] =  M[i]
1137                               - sce1->pcoeffs[start+w2*128+i];
1138                     }
1139                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1140                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1141                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
1142                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
1143                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1144                                                 L34,
1145                                                 sce0->ics.swb_sizes[g],
1146                                                 sce0->sf_idx[(w+w2)*16+g],
1147                                                 sce0->band_type[(w+w2)*16+g],
1148                                                 lambda / band0->threshold, INFINITY, NULL);
1149                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1150                                                 R34,
1151                                                 sce1->ics.swb_sizes[g],
1152                                                 sce1->sf_idx[(w+w2)*16+g],
1153                                                 sce1->band_type[(w+w2)*16+g],
1154                                                 lambda / band1->threshold, INFINITY, NULL);
1155                     dist2 += quantize_band_cost(s, M,
1156                                                 M34,
1157                                                 sce0->ics.swb_sizes[g],
1158                                                 sce0->sf_idx[(w+w2)*16+g],
1159                                                 sce0->band_type[(w+w2)*16+g],
1160                                                 lambda / maxthr, INFINITY, NULL);
1161                     dist2 += quantize_band_cost(s, S,
1162                                                 S34,
1163                                                 sce1->ics.swb_sizes[g],
1164                                                 sce1->sf_idx[(w+w2)*16+g],
1165                                                 sce1->band_type[(w+w2)*16+g],
1166                                                 lambda / minthr, INFINITY, NULL);
1167                 }
1168                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1169             }
1170             start += sce0->ics.swb_sizes[g];
1171         }
1172     }
1173 }
1174
1175 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
1176     [AAC_CODER_FAAC] = {
1177         search_for_quantizers_faac,
1178         encode_window_bands_info,
1179         quantize_and_encode_band,
1180         search_for_ms,
1181     },
1182     [AAC_CODER_ANMR] = {
1183         search_for_quantizers_anmr,
1184         encode_window_bands_info,
1185         quantize_and_encode_band,
1186         search_for_ms,
1187     },
1188     [AAC_CODER_TWOLOOP] = {
1189         search_for_quantizers_twoloop,
1190         codebook_trellis_rate,
1191         quantize_and_encode_band,
1192         search_for_ms,
1193     },
1194     [AAC_CODER_FAST] = {
1195         search_for_quantizers_fast,
1196         encode_window_bands_info,
1197         quantize_and_encode_band,
1198         search_for_ms,
1199     },
1200 };