Libav 0.7.1
libavcodec/sipr.c
Go to the documentation of this file.
00001 /*
00002  * SIPR / ACELP.NET decoder
00003  *
00004  * Copyright (c) 2008 Vladimir Voroshilov
00005  * Copyright (c) 2009 Vitor Sessak
00006  *
00007  * This file is part of Libav.
00008  *
00009  * Libav is free software; you can redistribute it and/or
00010  * modify it under the terms of the GNU Lesser General Public
00011  * License as published by the Free Software Foundation; either
00012  * version 2.1 of the License, or (at your option) any later version.
00013  *
00014  * Libav is distributed in the hope that it will be useful,
00015  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00016  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00017  * Lesser General Public License for more details.
00018  *
00019  * You should have received a copy of the GNU Lesser General Public
00020  * License along with Libav; if not, write to the Free Software
00021  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00022  */
00023 
00024 #include <math.h>
00025 #include <stdint.h>
00026 #include <string.h>
00027 
00028 #include "libavutil/mathematics.h"
00029 #include "avcodec.h"
00030 #define ALT_BITSTREAM_READER_LE
00031 #include "get_bits.h"
00032 #include "dsputil.h"
00033 
00034 #include "lsp.h"
00035 #include "celp_math.h"
00036 #include "acelp_vectors.h"
00037 #include "acelp_pitch_delay.h"
00038 #include "acelp_filters.h"
00039 #include "celp_filters.h"
00040 
00041 #define MAX_SUBFRAME_COUNT   5
00042 
00043 #include "sipr.h"
00044 #include "siprdata.h"
00045 
00046 typedef struct {
00047     const char *mode_name;
00048     uint16_t bits_per_frame;
00049     uint8_t subframe_count;
00050     uint8_t frames_per_packet;
00051     float pitch_sharp_factor;
00052 
00053     /* bitstream parameters */
00054     uint8_t number_of_fc_indexes;
00055     uint8_t ma_predictor_bits;  
00056 
00058     uint8_t vq_indexes_bits[5];
00059 
00061     uint8_t pitch_delay_bits[5];
00062 
00063     uint8_t gp_index_bits;
00064     uint8_t fc_index_bits[10]; 
00065     uint8_t gc_index_bits;     
00066 } SiprModeParam;
00067 
00068 static const SiprModeParam modes[MODE_COUNT] = {
00069     [MODE_16k] = {
00070         .mode_name          = "16k",
00071         .bits_per_frame     = 160,
00072         .subframe_count     = SUBFRAME_COUNT_16k,
00073         .frames_per_packet  = 1,
00074         .pitch_sharp_factor = 0.00,
00075 
00076         .number_of_fc_indexes = 10,
00077         .ma_predictor_bits    = 1,
00078         .vq_indexes_bits      = {7, 8, 7, 7, 7},
00079         .pitch_delay_bits     = {9, 6},
00080         .gp_index_bits        = 4,
00081         .fc_index_bits        = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5},
00082         .gc_index_bits        = 5
00083     },
00084 
00085     [MODE_8k5] = {
00086         .mode_name          = "8k5",
00087         .bits_per_frame     = 152,
00088         .subframe_count     = 3,
00089         .frames_per_packet  = 1,
00090         .pitch_sharp_factor = 0.8,
00091 
00092         .number_of_fc_indexes = 3,
00093         .ma_predictor_bits    = 0,
00094         .vq_indexes_bits      = {6, 7, 7, 7, 5},
00095         .pitch_delay_bits     = {8, 5, 5},
00096         .gp_index_bits        = 0,
00097         .fc_index_bits        = {9, 9, 9},
00098         .gc_index_bits        = 7
00099     },
00100 
00101     [MODE_6k5] = {
00102         .mode_name          = "6k5",
00103         .bits_per_frame     = 232,
00104         .subframe_count     = 3,
00105         .frames_per_packet  = 2,
00106         .pitch_sharp_factor = 0.8,
00107 
00108         .number_of_fc_indexes = 3,
00109         .ma_predictor_bits    = 0,
00110         .vq_indexes_bits      = {6, 7, 7, 7, 5},
00111         .pitch_delay_bits     = {8, 5, 5},
00112         .gp_index_bits        = 0,
00113         .fc_index_bits        = {5, 5, 5},
00114         .gc_index_bits        = 7
00115     },
00116 
00117     [MODE_5k0] = {
00118         .mode_name          = "5k0",
00119         .bits_per_frame     = 296,
00120         .subframe_count     = 5,
00121         .frames_per_packet  = 2,
00122         .pitch_sharp_factor = 0.85,
00123 
00124         .number_of_fc_indexes = 1,
00125         .ma_predictor_bits    = 0,
00126         .vq_indexes_bits      = {6, 7, 7, 7, 5},
00127         .pitch_delay_bits     = {8, 5, 8, 5, 5},
00128         .gp_index_bits        = 0,
00129         .fc_index_bits        = {10},
00130         .gc_index_bits        = 7
00131     }
00132 };
00133 
00134 const float ff_pow_0_5[] = {
00135     1.0/(1 <<  1), 1.0/(1 <<  2), 1.0/(1 <<  3), 1.0/(1 <<  4),
00136     1.0/(1 <<  5), 1.0/(1 <<  6), 1.0/(1 <<  7), 1.0/(1 <<  8),
00137     1.0/(1 <<  9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12),
00138     1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16)
00139 };
00140 
00141 static void dequant(float *out, const int *idx, const float *cbs[])
00142 {
00143     int i;
00144     int stride  = 2;
00145     int num_vec = 5;
00146 
00147     for (i = 0; i < num_vec; i++)
00148         memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float));
00149 
00150 }
00151 
00152 static void lsf_decode_fp(float *lsfnew, float *lsf_history,
00153                           const SiprParameters *parm)
00154 {
00155     int i;
00156     float lsf_tmp[LP_FILTER_ORDER];
00157 
00158     dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks);
00159 
00160     for (i = 0; i < LP_FILTER_ORDER; i++)
00161         lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i];
00162 
00163     ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1);
00164 
00165     /* Note that a minimum distance is not enforced between the last value and
00166        the previous one, contrary to what is done in ff_acelp_reorder_lsf() */
00167     ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1);
00168     lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI);
00169 
00170     memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history));
00171 
00172     for (i = 0; i < LP_FILTER_ORDER - 1; i++)
00173         lsfnew[i] = cos(lsfnew[i]);
00174     lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI;
00175 }
00176 
00178 static void pitch_sharpening(int pitch_lag_int, float beta,
00179                              float *fixed_vector)
00180 {
00181     int i;
00182 
00183     for (i = pitch_lag_int; i < SUBFR_SIZE; i++)
00184         fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int];
00185 }
00186 
00192 static void decode_parameters(SiprParameters* parms, GetBitContext *pgb,
00193                               const SiprModeParam *p)
00194 {
00195     int i, j;
00196 
00197     parms->ma_pred_switch           = get_bits(pgb, p->ma_predictor_bits);
00198 
00199     for (i = 0; i < 5; i++)
00200         parms->vq_indexes[i]        = get_bits(pgb, p->vq_indexes_bits[i]);
00201 
00202     for (i = 0; i < p->subframe_count; i++) {
00203         parms->pitch_delay[i]       = get_bits(pgb, p->pitch_delay_bits[i]);
00204         parms->gp_index[i]          = get_bits(pgb, p->gp_index_bits);
00205 
00206         for (j = 0; j < p->number_of_fc_indexes; j++)
00207             parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]);
00208 
00209         parms->gc_index[i]          = get_bits(pgb, p->gc_index_bits);
00210     }
00211 }
00212 
00213 static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az,
00214                            int num_subfr)
00215 {
00216     double lsfint[LP_FILTER_ORDER];
00217     int i,j;
00218     float t, t0 = 1.0 / num_subfr;
00219 
00220     t = t0 * 0.5;
00221     for (i = 0; i < num_subfr; i++) {
00222         for (j = 0; j < LP_FILTER_ORDER; j++)
00223             lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j];
00224 
00225         ff_amrwb_lsp2lpc(lsfint, Az, LP_FILTER_ORDER);
00226         Az += LP_FILTER_ORDER;
00227         t += t0;
00228     }
00229 }
00230 
00234 static void eval_ir(const float *Az, int pitch_lag, float *freq,
00235                     float pitch_sharp_factor)
00236 {
00237     float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1];
00238     int i;
00239 
00240     tmp1[0] = 1.;
00241     for (i = 0; i < LP_FILTER_ORDER; i++) {
00242         tmp1[i+1] = Az[i] * ff_pow_0_55[i];
00243         tmp2[i  ] = Az[i] * ff_pow_0_7 [i];
00244     }
00245     memset(tmp1 + 11, 0, 37 * sizeof(float));
00246 
00247     ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE,
00248                                  LP_FILTER_ORDER);
00249 
00250     pitch_sharpening(pitch_lag, pitch_sharp_factor, freq);
00251 }
00252 
00256 static void convolute_with_sparse(float *out, const AMRFixed *pulses,
00257                                   const float *shape, int length)
00258 {
00259     int i, j;
00260 
00261     memset(out, 0, length*sizeof(float));
00262     for (i = 0; i < pulses->n; i++)
00263         for (j = pulses->x[i]; j < length; j++)
00264             out[j] += pulses->y[i] * shape[j - pulses->x[i]];
00265 }
00266 
00270 static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples)
00271 {
00272     float buf[SUBFR_SIZE + LP_FILTER_ORDER];
00273     float *pole_out = buf + LP_FILTER_ORDER;
00274     float lpc_n[LP_FILTER_ORDER];
00275     float lpc_d[LP_FILTER_ORDER];
00276     int i;
00277 
00278     for (i = 0; i < LP_FILTER_ORDER; i++) {
00279         lpc_d[i] = lpc[i] * ff_pow_0_75[i];
00280         lpc_n[i] = lpc[i] * ff_pow_0_5 [i];
00281     };
00282 
00283     memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem,
00284            LP_FILTER_ORDER*sizeof(float));
00285 
00286     ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE,
00287                                  LP_FILTER_ORDER);
00288 
00289     memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
00290            LP_FILTER_ORDER*sizeof(float));
00291 
00292     ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE);
00293 
00294     memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0,
00295            LP_FILTER_ORDER*sizeof(*pole_out));
00296 
00297     memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
00298            LP_FILTER_ORDER*sizeof(*pole_out));
00299 
00300     ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE,
00301                                       LP_FILTER_ORDER);
00302 
00303 }
00304 
00305 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses,
00306                                 SiprMode mode, int low_gain)
00307 {
00308     int i;
00309 
00310     switch (mode) {
00311     case MODE_6k5:
00312         for (i = 0; i < 3; i++) {
00313             fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i;
00314             fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1;
00315         }
00316         fixed_sparse->n = 3;
00317         break;
00318     case MODE_8k5:
00319         for (i = 0; i < 3; i++) {
00320             fixed_sparse->x[2*i    ] = 3 * ((pulses[i] >> 4) & 0xf) + i;
00321             fixed_sparse->x[2*i + 1] = 3 * ( pulses[i]       & 0xf) + i;
00322 
00323             fixed_sparse->y[2*i    ] = (pulses[i] & 0x100) ? -1.0: 1.0;
00324 
00325             fixed_sparse->y[2*i + 1] =
00326                 (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ?
00327                 -fixed_sparse->y[2*i    ] : fixed_sparse->y[2*i];
00328         }
00329 
00330         fixed_sparse->n = 6;
00331         break;
00332     case MODE_5k0:
00333     default:
00334         if (low_gain) {
00335             int offset = (pulses[0] & 0x200) ? 2 : 0;
00336             int val = pulses[0];
00337 
00338             for (i = 0; i < 3; i++) {
00339                 int index = (val & 0x7) * 6 + 4 - i*2;
00340 
00341                 fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1;
00342                 fixed_sparse->x[i] = index;
00343 
00344                 val >>= 3;
00345             }
00346             fixed_sparse->n = 3;
00347         } else {
00348             int pulse_subset = (pulses[0] >> 8) & 1;
00349 
00350             fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset;
00351             fixed_sparse->x[1] = ( pulses[0]       & 15) * 3 + pulse_subset + 1;
00352 
00353             fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1;
00354             fixed_sparse->y[1] = -fixed_sparse->y[0];
00355             fixed_sparse->n = 2;
00356         }
00357         break;
00358     }
00359 }
00360 
00361 static void decode_frame(SiprContext *ctx, SiprParameters *params,
00362                          float *out_data)
00363 {
00364     int i, j;
00365     int subframe_count = modes[ctx->mode].subframe_count;
00366     int frame_size = subframe_count * SUBFR_SIZE;
00367     float Az[LP_FILTER_ORDER * MAX_SUBFRAME_COUNT];
00368     float *excitation;
00369     float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER];
00370     float lsf_new[LP_FILTER_ORDER];
00371     float *impulse_response = ir_buf + LP_FILTER_ORDER;
00372     float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for
00373                                         // memory alignment
00374     int t0_first = 0;
00375     AMRFixed fixed_cb;
00376 
00377     memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float));
00378     lsf_decode_fp(lsf_new, ctx->lsf_history, params);
00379 
00380     sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count);
00381 
00382     memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float));
00383 
00384     excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL;
00385 
00386     for (i = 0; i < subframe_count; i++) {
00387         float *pAz = Az + i*LP_FILTER_ORDER;
00388         float fixed_vector[SUBFR_SIZE];
00389         int T0,T0_frac;
00390         float pitch_gain, gain_code, avg_energy;
00391 
00392         ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i,
00393                             ctx->mode == MODE_5k0, 6);
00394 
00395         if (i == 0 || (i == 2 && ctx->mode == MODE_5k0))
00396             t0_first = T0;
00397 
00398         ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0),
00399                               ff_b60_sinc, 6,
00400                               2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER,
00401                               SUBFR_SIZE);
00402 
00403         decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode,
00404                             ctx->past_pitch_gain < 0.8);
00405 
00406         eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor);
00407 
00408         convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response,
00409                               SUBFR_SIZE);
00410 
00411         avg_energy =
00412             (0.01 + ff_dot_productf(fixed_vector, fixed_vector, SUBFR_SIZE))/
00413                 SUBFR_SIZE;
00414 
00415         ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0];
00416 
00417         gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1],
00418                                           avg_energy, ctx->energy_history,
00419                                           34 - 15.0/(0.05*M_LN10/M_LN2),
00420                                           pred);
00421 
00422         ff_weighted_vector_sumf(excitation, excitation, fixed_vector,
00423                                 pitch_gain, gain_code, SUBFR_SIZE);
00424 
00425         pitch_gain *= 0.5 * pitch_gain;
00426         pitch_gain = FFMIN(pitch_gain, 0.4);
00427 
00428         ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain;
00429         ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain);
00430         gain_code *= ctx->gain_mem;
00431 
00432         for (j = 0; j < SUBFR_SIZE; j++)
00433             fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j];
00434 
00435         if (ctx->mode == MODE_5k0) {
00436             postfilter_5k0(ctx, pAz, fixed_vector);
00437 
00438             ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
00439                                          pAz, excitation, SUBFR_SIZE,
00440                                          LP_FILTER_ORDER);
00441         }
00442 
00443         ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector,
00444                                      SUBFR_SIZE, LP_FILTER_ORDER);
00445 
00446         excitation += SUBFR_SIZE;
00447     }
00448 
00449     memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER,
00450            LP_FILTER_ORDER * sizeof(float));
00451 
00452     if (ctx->mode == MODE_5k0) {
00453         for (i = 0; i < subframe_count; i++) {
00454             float energy = ff_dot_productf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
00455                                            ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
00456                                            SUBFR_SIZE);
00457             ff_adaptive_gain_control(&synth[i * SUBFR_SIZE],
00458                                      &synth[i * SUBFR_SIZE], energy,
00459                                      SUBFR_SIZE, 0.9, &ctx->postfilter_agc);
00460         }
00461 
00462         memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size,
00463                LP_FILTER_ORDER*sizeof(float));
00464     }
00465     memmove(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL,
00466            (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float));
00467 
00468     ff_acelp_apply_order_2_transfer_function(out_data, synth,
00469                                              (const float[2]) {-1.99997   , 1.000000000},
00470                                              (const float[2]) {-1.93307352, 0.935891986},
00471                                              0.939805806,
00472                                              ctx->highpass_filt_mem,
00473                                              frame_size);
00474 }
00475 
00476 static av_cold int sipr_decoder_init(AVCodecContext * avctx)
00477 {
00478     SiprContext *ctx = avctx->priv_data;
00479     int i;
00480 
00481     if      (avctx->bit_rate > 12200) ctx->mode = MODE_16k;
00482     else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5;
00483     else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5;
00484     else                              ctx->mode = MODE_5k0;
00485 
00486     av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name);
00487 
00488     if (ctx->mode == MODE_16k)
00489         ff_sipr_init_16k(ctx);
00490 
00491     for (i = 0; i < LP_FILTER_ORDER; i++)
00492         ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1));
00493 
00494     for (i = 0; i < 4; i++)
00495         ctx->energy_history[i] = -14;
00496 
00497     avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
00498 
00499     return 0;
00500 }
00501 
00502 static int sipr_decode_frame(AVCodecContext *avctx, void *datap,
00503                              int *data_size, AVPacket *avpkt)
00504 {
00505     SiprContext *ctx = avctx->priv_data;
00506     const uint8_t *buf=avpkt->data;
00507     SiprParameters parm;
00508     const SiprModeParam *mode_par = &modes[ctx->mode];
00509     GetBitContext gb;
00510     float *data = datap;
00511     int subframe_size = ctx->mode == MODE_16k ? L_SUBFR_16k : SUBFR_SIZE;
00512     int i;
00513 
00514     ctx->avctx = avctx;
00515     if (avpkt->size < (mode_par->bits_per_frame >> 3)) {
00516         av_log(avctx, AV_LOG_ERROR,
00517                "Error processing packet: packet size (%d) too small\n",
00518                avpkt->size);
00519 
00520         *data_size = 0;
00521         return -1;
00522     }
00523     if (*data_size < subframe_size * mode_par->subframe_count * sizeof(float)) {
00524         av_log(avctx, AV_LOG_ERROR,
00525                "Error processing packet: output buffer (%d) too small\n",
00526                *data_size);
00527 
00528         *data_size = 0;
00529         return -1;
00530     }
00531 
00532     init_get_bits(&gb, buf, mode_par->bits_per_frame);
00533 
00534     for (i = 0; i < mode_par->frames_per_packet; i++) {
00535         decode_parameters(&parm, &gb, mode_par);
00536 
00537         if (ctx->mode == MODE_16k)
00538             ff_sipr_decode_frame_16k(ctx, &parm, data);
00539         else
00540             decode_frame(ctx, &parm, data);
00541 
00542         data += subframe_size * mode_par->subframe_count;
00543     }
00544 
00545     *data_size = mode_par->frames_per_packet * subframe_size *
00546         mode_par->subframe_count * sizeof(float);
00547 
00548     return mode_par->bits_per_frame >> 3;
00549 }
00550 
00551 AVCodec ff_sipr_decoder = {
00552     "sipr",
00553     AVMEDIA_TYPE_AUDIO,
00554     CODEC_ID_SIPR,
00555     sizeof(SiprContext),
00556     sipr_decoder_init,
00557     NULL,
00558     NULL,
00559     sipr_decode_frame,
00560     .long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"),
00561 };