libavcodec/aacsbr_fixed.c

   1 /*
   2  * Copyright (c) 2013
   3  *      MIPS Technologies, Inc., California.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
  14  *    contributors may be used to endorse or promote products derived from
  15  *    this software without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  * AAC Spectral Band Replication decoding functions (fixed-point)
  30  * Copyright (c) 2008-2009 Robert Swain ( rob opendot cl )
  31  * Copyright (c) 2009-2010 Alex Converse <alex.converse@gmail.com>
  32  *
  33  * This file is part of FFmpeg.
  34  *
  35  * FFmpeg is free software; you can redistribute it and/or
  36  * modify it under the terms of the GNU Lesser General Public
  37  * License as published by the Free Software Foundation; either
  38  * version 2.1 of the License, or (at your option) any later version.
  39  *
  40  * FFmpeg is distributed in the hope that it will be useful,
  41  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  42  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  43  * Lesser General Public License for more details.
  44  *
  45  * You should have received a copy of the GNU Lesser General Public
  46  * License along with FFmpeg; if not, write to the Free Software
  47  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  48  */
  49
  50 /**
  51  * @file
  52  * AAC Spectral Band Replication decoding functions (fixed-point)
  53  * Note: Rounding-to-nearest used unless otherwise stated
  54  * @author Robert Swain ( rob opendot cl )
  55  * @author Stanislav Ocovaj ( stanislav.ocovaj imgtec com )
  56  */
  57 #define USE_FIXED 1
  58
  59 #include "aac.h"
  60 #include "sbr.h"
  61 #include "aacsbr.h"
  62 #include "aacsbrdata.h"
  63 #include "aacps.h"
  64 #include "sbrdsp.h"
  65 #include "libavutil/internal.h"
  66 #include "libavutil/libm.h"
  67 #include "libavutil/avassert.h"
  68
  69 #include <stdint.h>
  70 #include <float.h>
  71 #include <math.h>
  72
  73 static void aacsbr_func_ptr_init(AACSBRContext *c);
  74 static const int CONST_LN2       = Q31(0.6931471806/256);  // ln(2)/256
  75 static const int CONST_RECIP_LN2 = Q31(0.7213475204);      // 0.5/ln(2)
  76 static const int CONST_076923    = Q31(0.76923076923076923077f);
  77
  78 static const int fixed_log_table[10] =
  79 {
  80     Q31(1.0/2), Q31(1.0/3), Q31(1.0/4), Q31(1.0/5), Q31(1.0/6),
  81     Q31(1.0/7), Q31(1.0/8), Q31(1.0/9), Q31(1.0/10), Q31(1.0/11)
  82 };
  83
  84 static int fixed_log(int x)
  85 {
  86     int i, ret, xpow, tmp;
  87
  88     ret = x;
  89     xpow = x;
  90     for (i=0; i<10; i+=2){
  91         xpow = (int)(((int64_t)xpow * x + 0x40000000) >> 31);
  92         tmp = (int)(((int64_t)xpow * fixed_log_table[i] + 0x40000000) >> 31);
  93         ret -= tmp;
  94
  95         xpow = (int)(((int64_t)xpow * x + 0x40000000) >> 31);
  96         tmp = (int)(((int64_t)xpow * fixed_log_table[i+1] + 0x40000000) >> 31);
  97         ret += tmp;
  98     }
  99
 100     return ret;
 101 }
 102
 103 static const int fixed_exp_table[7] =
 104 {
 105     Q31(1.0/2), Q31(1.0/6), Q31(1.0/24), Q31(1.0/120),
 106     Q31(1.0/720), Q31(1.0/5040), Q31(1.0/40320)
 107 };
 108
 109 static int fixed_exp(int x)
 110 {
 111     int i, ret, xpow, tmp;
 112
 113     ret = 0x800000 + x;
 114     xpow = x;
 115     for (i=0; i<7; i++){
 116         xpow = (int)(((int64_t)xpow * x + 0x400000) >> 23);
 117         tmp = (int)(((int64_t)xpow * fixed_exp_table[i] + 0x40000000) >> 31);
 118         ret += tmp;
 119     }
 120
 121     return ret;
 122 }
 123
 124 static void make_bands(int16_t* bands, int start, int stop, int num_bands)
 125 {
 126     int k, previous, present;
 127     int base, prod, nz = 0;
 128
 129     base = (stop << 23) / start;
 130     while (base < 0x40000000){
 131         base <<= 1;
 132         nz++;
 133     }
 134     base = fixed_log(base - 0x80000000);
 135     base = (((base + 0x80) >> 8) + (8-nz)*CONST_LN2) / num_bands;
 136     base = fixed_exp(base);
 137
 138     previous = start;
 139     prod = start << 23;
 140
 141     for (k = 0; k < num_bands-1; k++) {
 142         prod = (int)(((int64_t)prod * base + 0x400000) >> 23);
 143         present = (prod + 0x400000) >> 23;
 144         bands[k] = present - previous;
 145         previous = present;
 146     }
 147     bands[num_bands-1] = stop - previous;
 148 }
 149
 150 /// Dequantization and stereo decoding (14496-3 sp04 p203)
 151 static void sbr_dequant(SpectralBandReplication *sbr, int id_aac)
 152 {
 153     int k, e;
 154     int ch;
 155
 156     if (id_aac == TYPE_CPE && sbr->bs_coupling) {
 157         int alpha      = sbr->data[0].bs_amp_res ?  2 :  1;
 158         int pan_offset = sbr->data[0].bs_amp_res ? 12 : 24;
 159         for (e = 1; e <= sbr->data[0].bs_num_env; e++) {
 160             for (k = 0; k < sbr->n[sbr->data[0].bs_freq_res[e]]; k++) {
 161                 SoftFloat temp1, temp2, fac;
 162
 163                 temp1.exp = sbr->data[0].env_facs_q[e][k] * alpha + 14;
 164                 if (temp1.exp & 1)
 165                   temp1.mant = 759250125;
 166                 else
 167                   temp1.mant = 0x20000000;
 168                 temp1.exp = (temp1.exp >> 1) + 1;
 169                 if (temp1.exp > 66) { // temp1 > 1E20
 170                     av_log(NULL, AV_LOG_ERROR, "envelope scalefactor overflow in dequant\n");
 171                     temp1 = FLOAT_1;
 172                 }
 173
 174                 temp2.exp = (pan_offset - sbr->data[1].env_facs_q[e][k]) * alpha;
 175                 if (temp2.exp & 1)
 176                   temp2.mant = 759250125;
 177                 else
 178                   temp2.mant = 0x20000000;
 179                 temp2.exp = (temp2.exp >> 1) + 1;
 180                 fac   = av_div_sf(temp1, av_add_sf(FLOAT_1, temp2));
 181                 sbr->data[0].env_facs[e][k] = fac;
 182                 sbr->data[1].env_facs[e][k] = av_mul_sf(fac, temp2);
 183             }
 184         }
 185         for (e = 1; e <= sbr->data[0].bs_num_noise; e++) {
 186             for (k = 0; k < sbr->n_q; k++) {
 187                 SoftFloat temp1, temp2, fac;
 188
 189                 temp1.exp = NOISE_FLOOR_OFFSET - \
 190                     sbr->data[0].noise_facs_q[e][k] + 2;
 191                 temp1.mant = 0x20000000;
 192                 av_assert0(temp1.exp <= 66);
 193                 temp2.exp = 12 - sbr->data[1].noise_facs_q[e][k] + 1;
 194                 temp2.mant = 0x20000000;
 195                 fac   = av_div_sf(temp1, av_add_sf(FLOAT_1, temp2));
 196                 sbr->data[0].noise_facs[e][k] = fac;
 197                 sbr->data[1].noise_facs[e][k] = av_mul_sf(fac, temp2);
 198             }
 199         }
 200     } else { // SCE or one non-coupled CPE
 201         for (ch = 0; ch < (id_aac == TYPE_CPE) + 1; ch++) {
 202             int alpha = sbr->data[ch].bs_amp_res ? 2 : 1;
 203             for (e = 1; e <= sbr->data[ch].bs_num_env; e++)
 204                 for (k = 0; k < sbr->n[sbr->data[ch].bs_freq_res[e]]; k++){
 205                     SoftFloat temp1;
 206
 207                     temp1.exp = alpha * sbr->data[ch].env_facs_q[e][k] + 12;
 208                     if (temp1.exp & 1)
 209                         temp1.mant = 759250125;
 210                     else
 211                         temp1.mant = 0x20000000;
 212                     temp1.exp = (temp1.exp >> 1) + 1;
 213                     if (temp1.exp > 66) { // temp1 > 1E20
 214                         av_log(NULL, AV_LOG_ERROR, "envelope scalefactor overflow in dequant\n");
 215                         temp1 = FLOAT_1;
 216                     }
 217                     sbr->data[ch].env_facs[e][k] = temp1;
 218                 }
 219             for (e = 1; e <= sbr->data[ch].bs_num_noise; e++)
 220                 for (k = 0; k < sbr->n_q; k++){
 221                     sbr->data[ch].noise_facs[e][k].exp = NOISE_FLOOR_OFFSET - \
 222                         sbr->data[ch].noise_facs_q[e][k] + 1;
 223                     sbr->data[ch].noise_facs[e][k].mant = 0x20000000;
 224                 }
 225         }
 226     }
 227 }
 228
 229 /** High Frequency Generation (14496-3 sp04 p214+) and Inverse Filtering
 230  * (14496-3 sp04 p214)
 231  * Warning: This routine does not seem numerically stable.
 232  */
 233 static void sbr_hf_inverse_filter(SBRDSPContext *dsp,
 234                                   int (*alpha0)[2], int (*alpha1)[2],
 235                                   const int X_low[32][40][2], int k0)
 236 {
 237     int k;
 238     int shift, round;
 239
 240     for (k = 0; k < k0; k++) {
 241         SoftFloat phi[3][2][2];
 242         SoftFloat a00, a01, a10, a11;
 243         SoftFloat dk;
 244
 245         dsp->autocorrelate(X_low[k], phi);
 246
 247         dk = av_sub_sf(av_mul_sf(phi[2][1][0], phi[1][0][0]),
 248              av_mul_sf(av_add_sf(av_mul_sf(phi[1][1][0], phi[1][1][0]),
 249              av_mul_sf(phi[1][1][1], phi[1][1][1])), FLOAT_0999999));
 250
 251         if (!dk.mant) {
 252             a10 = FLOAT_0;
 253             a11 = FLOAT_0;
 254         } else {
 255             SoftFloat temp_real, temp_im;
 256             temp_real = av_sub_sf(av_sub_sf(av_mul_sf(phi[0][0][0], phi[1][1][0]),
 257                                             av_mul_sf(phi[0][0][1], phi[1][1][1])),
 258                                   av_mul_sf(phi[0][1][0], phi[1][0][0]));
 259             temp_im   = av_sub_sf(av_add_sf(av_mul_sf(phi[0][0][0], phi[1][1][1]),
 260                                             av_mul_sf(phi[0][0][1], phi[1][1][0])),
 261                                   av_mul_sf(phi[0][1][1], phi[1][0][0]));
 262
 263             a10 = av_div_sf(temp_real, dk);
 264             a11 = av_div_sf(temp_im,   dk);
 265         }
 266
 267         if (!phi[1][0][0].mant) {
 268             a00 = FLOAT_0;
 269             a01 = FLOAT_0;
 270         } else {
 271             SoftFloat temp_real, temp_im;
 272             temp_real = av_add_sf(phi[0][0][0],
 273                                   av_add_sf(av_mul_sf(a10, phi[1][1][0]),
 274                                             av_mul_sf(a11, phi[1][1][1])));
 275             temp_im   = av_add_sf(phi[0][0][1],
 276                                   av_sub_sf(av_mul_sf(a11, phi[1][1][0]),
 277                                             av_mul_sf(a10, phi[1][1][1])));
 278
 279             temp_real.mant = -temp_real.mant;
 280             temp_im.mant   = -temp_im.mant;
 281             a00 = av_div_sf(temp_real, phi[1][0][0]);
 282             a01 = av_div_sf(temp_im,   phi[1][0][0]);
 283         }
 284
 285         shift = a00.exp;
 286         if (shift >= 3)
 287             alpha0[k][0] = 0x7fffffff;
 288         else if (shift <= -30)
 289             alpha0[k][0] = 0;
 290         else {
 291             shift = 1-shift;
 292             if (shift <= 0)
 293                 alpha0[k][0] = a00.mant * (1<<-shift);
 294             else {
 295                 round = 1 << (shift-1);
 296                 alpha0[k][0] = (a00.mant + round) >> shift;
 297             }
 298         }
 299
 300         shift = a01.exp;
 301         if (shift >= 3)
 302             alpha0[k][1] = 0x7fffffff;
 303         else if (shift <= -30)
 304             alpha0[k][1] = 0;
 305         else {
 306             shift = 1-shift;
 307             if (shift <= 0)
 308                 alpha0[k][1] = a01.mant * (1<<-shift);
 309             else {
 310                 round = 1 << (shift-1);
 311                 alpha0[k][1] = (a01.mant + round) >> shift;
 312             }
 313         }
 314         shift = a10.exp;
 315         if (shift >= 3)
 316             alpha1[k][0] = 0x7fffffff;
 317         else if (shift <= -30)
 318             alpha1[k][0] = 0;
 319         else {
 320             shift = 1-shift;
 321             if (shift <= 0)
 322                 alpha1[k][0] = a10.mant * (1<<-shift);
 323             else {
 324                 round = 1 << (shift-1);
 325                 alpha1[k][0] = (a10.mant + round) >> shift;
 326             }
 327         }
 328
 329         shift = a11.exp;
 330         if (shift >= 3)
 331             alpha1[k][1] = 0x7fffffff;
 332         else if (shift <= -30)
 333             alpha1[k][1] = 0;
 334         else {
 335             shift = 1-shift;
 336             if (shift <= 0)
 337                 alpha1[k][1] = a11.mant * (1<<-shift);
 338             else {
 339                 round = 1 << (shift-1);
 340                 alpha1[k][1] = (a11.mant + round) >> shift;
 341             }
 342         }
 343
 344         shift = (int)(((int64_t)(alpha1[k][0]>>1) * (alpha1[k][0]>>1) + \
 345                        (int64_t)(alpha1[k][1]>>1) * (alpha1[k][1]>>1) + \
 346                        0x40000000) >> 31);
 347         if (shift >= 0x20000000){
 348             alpha1[k][0] = 0;
 349             alpha1[k][1] = 0;
 350             alpha0[k][0] = 0;
 351             alpha0[k][1] = 0;
 352         }
 353
 354         shift = (int)(((int64_t)(alpha0[k][0]>>1) * (alpha0[k][0]>>1) + \
 355                        (int64_t)(alpha0[k][1]>>1) * (alpha0[k][1]>>1) + \
 356                        0x40000000) >> 31);
 357         if (shift >= 0x20000000){
 358             alpha1[k][0] = 0;
 359             alpha1[k][1] = 0;
 360             alpha0[k][0] = 0;
 361             alpha0[k][1] = 0;
 362         }
 363     }
 364 }
 365
 366 /// Chirp Factors (14496-3 sp04 p214)
 367 static void sbr_chirp(SpectralBandReplication *sbr, SBRData *ch_data)
 368 {
 369     int i;
 370     int new_bw;
 371     static const int bw_tab[] = { 0, 1610612736, 1932735283, 2104533975 };
 372     int64_t accu;
 373
 374     for (i = 0; i < sbr->n_q; i++) {
 375         if (ch_data->bs_invf_mode[0][i] + ch_data->bs_invf_mode[1][i] == 1)
 376             new_bw = 1288490189;
 377         else
 378             new_bw = bw_tab[ch_data->bs_invf_mode[0][i]];
 379
 380         if (new_bw < ch_data->bw_array[i]){
 381             accu  = (int64_t)new_bw * 1610612736;
 382             accu += (int64_t)ch_data->bw_array[i] * 0x20000000;
 383             new_bw = (int)((accu + 0x40000000) >> 31);
 384         } else {
 385             accu  = (int64_t)new_bw * 1946157056;
 386             accu += (int64_t)ch_data->bw_array[i] * 201326592;
 387             new_bw = (int)((accu + 0x40000000) >> 31);
 388         }
 389         ch_data->bw_array[i] = new_bw < 0x2000000 ? 0 : new_bw;
 390     }
 391 }
 392
 393 /**
 394  * Calculation of levels of additional HF signal components (14496-3 sp04 p219)
 395  * and Calculation of gain (14496-3 sp04 p219)
 396  */
 397 static void sbr_gain_calc(SpectralBandReplication *sbr,
 398                           SBRData *ch_data, const int e_a[2])
 399 {
 400     int e, k, m;
 401     // max gain limits : -3dB, 0dB, 3dB, inf dB (limiter off)
 402     static const SoftFloat limgain[4] = { { 760155524,  0 }, { 0x20000000,  1 },
 403                                             { 758351638,  1 }, { 625000000, 34 } };
 404
 405     for (e = 0; e < ch_data->bs_num_env; e++) {
 406         int delta = !((e == e_a[1]) || (e == e_a[0]));
 407         for (k = 0; k < sbr->n_lim; k++) {
 408             SoftFloat gain_boost, gain_max;
 409             SoftFloat sum[2];
 410             sum[0] = sum[1] = FLOAT_0;
 411             for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) {
 412                 const SoftFloat temp = av_div_sf(sbr->e_origmapped[e][m],
 413                                             av_add_sf(FLOAT_1, sbr->q_mapped[e][m]));
 414                 sbr->q_m[e][m] = av_sqrt_sf(av_mul_sf(temp, sbr->q_mapped[e][m]));
 415                 sbr->s_m[e][m] = av_sqrt_sf(av_mul_sf(temp, av_int2sf(ch_data->s_indexmapped[e + 1][m], 0)));
 416                 if (!sbr->s_mapped[e][m]) {
 417                     if (delta) {
 418                       sbr->gain[e][m] = av_sqrt_sf(av_div_sf(sbr->e_origmapped[e][m],
 419                                             av_mul_sf(av_add_sf(FLOAT_1, sbr->e_curr[e][m]),
 420                                             av_add_sf(FLOAT_1, sbr->q_mapped[e][m]))));
 421                     } else {
 422                       sbr->gain[e][m] = av_sqrt_sf(av_div_sf(sbr->e_origmapped[e][m],
 423                                             av_add_sf(FLOAT_1, sbr->e_curr[e][m])));
 424                     }
 425                 } else {
 426                     sbr->gain[e][m] = av_sqrt_sf(
 427                                         av_div_sf(
 428                                             av_mul_sf(sbr->e_origmapped[e][m], sbr->q_mapped[e][m]),
 429                                             av_mul_sf(
 430                                                 av_add_sf(FLOAT_1, sbr->e_curr[e][m]),
 431                                                 av_add_sf(FLOAT_1, sbr->q_mapped[e][m]))));
 432                 }
 433                 sbr->gain[e][m] = av_add_sf(sbr->gain[e][m], FLOAT_MIN);
 434             }
 435             for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) {
 436                 sum[0] = av_add_sf(sum[0], sbr->e_origmapped[e][m]);
 437                 sum[1] = av_add_sf(sum[1], sbr->e_curr[e][m]);
 438             }
 439             gain_max = av_mul_sf(limgain[sbr->bs_limiter_gains],
 440                             av_sqrt_sf(
 441                                 av_div_sf(
 442                                     av_add_sf(FLOAT_EPSILON, sum[0]),
 443                                     av_add_sf(FLOAT_EPSILON, sum[1]))));
 444             if (av_gt_sf(gain_max, FLOAT_100000))
 445               gain_max = FLOAT_100000;
 446             for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) {
 447                 SoftFloat q_m_max = av_div_sf(
 448                                         av_mul_sf(sbr->q_m[e][m], gain_max),
 449                                         sbr->gain[e][m]);
 450                 if (av_gt_sf(sbr->q_m[e][m], q_m_max))
 451                   sbr->q_m[e][m] = q_m_max;
 452                 if (av_gt_sf(sbr->gain[e][m], gain_max))
 453                   sbr->gain[e][m] = gain_max;
 454             }
 455             sum[0] = sum[1] = FLOAT_0;
 456             for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) {
 457                 sum[0] = av_add_sf(sum[0], sbr->e_origmapped[e][m]);
 458                 sum[1] = av_add_sf(sum[1],
 459                             av_mul_sf(
 460                                 av_mul_sf(sbr->e_curr[e][m],
 461                                           sbr->gain[e][m]),
 462                                 sbr->gain[e][m]));
 463                 sum[1] = av_add_sf(sum[1],
 464                             av_mul_sf(sbr->s_m[e][m], sbr->s_m[e][m]));
 465                 if (delta && !sbr->s_m[e][m].mant)
 466                   sum[1] = av_add_sf(sum[1],
 467                                 av_mul_sf(sbr->q_m[e][m], sbr->q_m[e][m]));
 468             }
 469             gain_boost = av_sqrt_sf(
 470                             av_div_sf(
 471                                 av_add_sf(FLOAT_EPSILON, sum[0]),
 472                                 av_add_sf(FLOAT_EPSILON, sum[1])));
 473             if (av_gt_sf(gain_boost, FLOAT_1584893192))
 474               gain_boost = FLOAT_1584893192;
 475
 476             for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) {
 477                 sbr->gain[e][m] = av_mul_sf(sbr->gain[e][m], gain_boost);
 478                 sbr->q_m[e][m]  = av_mul_sf(sbr->q_m[e][m], gain_boost);
 479                 sbr->s_m[e][m]  = av_mul_sf(sbr->s_m[e][m], gain_boost);
 480             }
 481         }
 482     }
 483 }
 484
 485 /// Assembling HF Signals (14496-3 sp04 p220)
 486 static void sbr_hf_assemble(int Y1[38][64][2],
 487                             const int X_high[64][40][2],
 488                             SpectralBandReplication *sbr, SBRData *ch_data,
 489                             const int e_a[2])
 490 {
 491     int e, i, j, m;
 492     const int h_SL = 4 * !sbr->bs_smoothing_mode;
 493     const int kx = sbr->kx[1];
 494     const int m_max = sbr->m[1];
 495     static const SoftFloat h_smooth[5] = {
 496       { 715827883, -1 },
 497       { 647472402, -1 },
 498       { 937030863, -2 },
 499       { 989249804, -3 },
 500       { 546843842, -4 },
 501     };
 502     SoftFloat (*g_temp)[48] = ch_data->g_temp, (*q_temp)[48] = ch_data->q_temp;
 503     int indexnoise = ch_data->f_indexnoise;
 504     int indexsine  = ch_data->f_indexsine;
 505
 506     if (sbr->reset) {
 507         for (i = 0; i < h_SL; i++) {
 508             memcpy(g_temp[i + 2*ch_data->t_env[0]], sbr->gain[0], m_max * sizeof(sbr->gain[0][0]));
 509             memcpy(q_temp[i + 2*ch_data->t_env[0]], sbr->q_m[0],  m_max * sizeof(sbr->q_m[0][0]));
 510         }
 511     } else if (h_SL) {
 512         for (i = 0; i < 4; i++) {
 513             memcpy(g_temp[i + 2 * ch_data->t_env[0]],
 514                    g_temp[i + 2 * ch_data->t_env_num_env_old],
 515                    sizeof(g_temp[0]));
 516             memcpy(q_temp[i + 2 * ch_data->t_env[0]],
 517                    q_temp[i + 2 * ch_data->t_env_num_env_old],
 518                    sizeof(q_temp[0]));
 519         }
 520     }
 521
 522     for (e = 0; e < ch_data->bs_num_env; e++) {
 523         for (i = 2 * ch_data->t_env[e]; i < 2 * ch_data->t_env[e + 1]; i++) {
 524             memcpy(g_temp[h_SL + i], sbr->gain[e], m_max * sizeof(sbr->gain[0][0]));
 525             memcpy(q_temp[h_SL + i], sbr->q_m[e],  m_max * sizeof(sbr->q_m[0][0]));
 526         }
 527     }
 528
 529     for (e = 0; e < ch_data->bs_num_env; e++) {
 530         for (i = 2 * ch_data->t_env[e]; i < 2 * ch_data->t_env[e + 1]; i++) {
 531             SoftFloat g_filt_tab[48];
 532             SoftFloat q_filt_tab[48];
 533             SoftFloat *g_filt, *q_filt;
 534
 535             if (h_SL && e != e_a[0] && e != e_a[1]) {
 536                 g_filt = g_filt_tab;
 537                 q_filt = q_filt_tab;
 538                 for (m = 0; m < m_max; m++) {
 539                     const int idx1 = i + h_SL;
 540                     g_filt[m].mant = g_filt[m].exp = 0;
 541                     q_filt[m].mant = q_filt[m].exp = 0;
 542                     for (j = 0; j <= h_SL; j++) {
 543                         g_filt[m] = av_add_sf(g_filt[m],
 544                                         av_mul_sf(g_temp[idx1 - j][m],
 545                                             h_smooth[j]));
 546                         q_filt[m] = av_add_sf(q_filt[m],
 547                                         av_mul_sf(q_temp[idx1 - j][m],
 548                                             h_smooth[j]));
 549                     }
 550                 }
 551             } else {
 552                 g_filt = g_temp[i + h_SL];
 553                 q_filt = q_temp[i];
 554             }
 555
 556             sbr->dsp.hf_g_filt(Y1[i] + kx, X_high + kx, g_filt, m_max,
 557                                i + ENVELOPE_ADJUSTMENT_OFFSET);
 558
 559             if (e != e_a[0] && e != e_a[1]) {
 560                 sbr->dsp.hf_apply_noise[indexsine](Y1[i] + kx, sbr->s_m[e],
 561                                                    q_filt, indexnoise,
 562                                                    kx, m_max);
 563             } else {
 564                 int idx = indexsine&1;
 565                 int A = (1-((indexsine+(kx & 1))&2));
 566                 int B = (A^(-idx)) + idx;
 567                 unsigned *out = &Y1[i][kx][idx];
 568                 int shift;
 569                 unsigned round;
 570
 571                 SoftFloat *in  = sbr->s_m[e];
 572                 for (m = 0; m+1 < m_max; m+=2) {
 573                     int shift2;
 574                     shift = 22 - in[m  ].exp;
 575                     shift2= 22 - in[m+1].exp;
 576                     if (shift < 1 || shift2 < 1) {
 577                         av_log(NULL, AV_LOG_ERROR, "Overflow in sbr_hf_assemble, shift=%d,%d\n", shift, shift2);
 578                         return;
 579                     }
 580                     if (shift < 32) {
 581                         round = 1 << (shift-1);
 582                         out[2*m  ] += (int)(in[m  ].mant * A + round) >> shift;
 583                     }
 584
 585                     if (shift2 < 32) {
 586                         round = 1 << (shift2-1);
 587                         out[2*m+2] += (int)(in[m+1].mant * B + round) >> shift2;
 588                     }
 589                 }
 590                 if(m_max&1)
 591                 {
 592                     shift = 22 - in[m  ].exp;
 593                     if (shift < 1) {
 594                         av_log(NULL, AV_LOG_ERROR, "Overflow in sbr_hf_assemble, shift=%d\n", shift);
 595                         return;
 596                     } else if (shift < 32) {
 597                         round = 1 << (shift-1);
 598                         out[2*m  ] += (int)(in[m  ].mant * A + round) >> shift;
 599                     }
 600                 }
 601             }
 602             indexnoise = (indexnoise + m_max) & 0x1ff;
 603             indexsine = (indexsine + 1) & 3;
 604         }
 605     }
 606     ch_data->f_indexnoise = indexnoise;
 607     ch_data->f_indexsine  = indexsine;
 608 }
 609
 610 #include "aacsbr_template.c"