• Main Page
  • Related Pages
  • Modules
  • Data Structures
  • Files
  • Examples
  • File List
  • Globals

libavcodec/twinvq.c

Go to the documentation of this file.
00001 /*
00002  * TwinVQ decoder
00003  * Copyright (c) 2009 Vitor Sessak
00004  *
00005  * This file is part of Libav.
00006  *
00007  * Libav is free software; you can redistribute it and/or
00008  * modify it under the terms of the GNU Lesser General Public
00009  * License as published by the Free Software Foundation; either
00010  * version 2.1 of the License, or (at your option) any later version.
00011  *
00012  * Libav is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00015  * Lesser General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU Lesser General Public
00018  * License along with Libav; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00022 #include "avcodec.h"
00023 #include "get_bits.h"
00024 #include "dsputil.h"
00025 #include "fft.h"
00026 #include "lsp.h"
00027 #include "sinewin.h"
00028 
00029 #include <math.h>
00030 #include <stdint.h>
00031 
00032 #include "twinvq_data.h"
00033 
00034 enum FrameType {
00035     FT_SHORT = 0,  
00036     FT_MEDIUM,     
00037     FT_LONG,       
00038     FT_PPC,        
00039 };
00040 
00044 struct FrameMode {
00045     uint8_t         sub;      
00046     const uint16_t *bark_tab;
00047 
00049     uint8_t         bark_env_size;
00050 
00051     const int16_t  *bark_cb;    
00052     uint8_t         bark_n_coef;
00053     uint8_t         bark_n_bit; 
00054 
00056 
00057     const int16_t    *cb0;
00058     const int16_t    *cb1;
00060 
00061     uint8_t         cb_len_read; 
00062 };
00063 
00068 typedef struct {
00069     struct FrameMode fmode[3]; 
00070 
00071     uint16_t     size;        
00072     uint8_t      n_lsp;       
00073     const float *lspcodebook;
00074 
00075     /* number of bits of the different LSP CB coefficients */
00076     uint8_t      lsp_bit0;
00077     uint8_t      lsp_bit1;
00078     uint8_t      lsp_bit2;
00079 
00080     uint8_t      lsp_split;      
00081     const int16_t *ppc_shape_cb; 
00082 
00084     uint8_t      ppc_period_bit;
00085 
00086     uint8_t      ppc_shape_bit;  
00087     uint8_t      ppc_shape_len;  
00088     uint8_t      pgain_bit;      
00089 
00091     uint16_t     peak_per2wid;
00092 } ModeTab;
00093 
00094 static const ModeTab mode_08_08 = {
00095     {
00096         { 8, bark_tab_s08_64,  10, tab.fcb08s  , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
00097         { 2, bark_tab_m08_256, 20, tab.fcb08m  , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
00098         { 1, bark_tab_l08_512, 30, tab.fcb08l  , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
00099     },
00100     512 , 12, tab.lsp08,   1, 5, 3, 3, tab.shape08  , 8, 28, 20, 6, 40
00101 };
00102 
00103 static const ModeTab mode_11_08 = {
00104     {
00105         { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
00106         { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
00107         { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
00108     },
00109     512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
00110 };
00111 
00112 static const ModeTab mode_11_10 = {
00113     {
00114         { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
00115         { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
00116         { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
00117     },
00118     512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
00119 };
00120 
00121 static const ModeTab mode_16_16 = {
00122     {
00123         { 8, bark_tab_s16_128, 10, tab.fcb16s  , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
00124         { 2, bark_tab_m16_512, 20, tab.fcb16m  , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
00125         { 1, bark_tab_l16_1024,30, tab.fcb16l  , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
00126     },
00127     1024, 16, tab.lsp16,   1, 6, 4, 3, tab.shape16  , 9, 56, 60, 7, 180
00128 };
00129 
00130 static const ModeTab mode_22_20 = {
00131     {
00132         { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
00133         { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
00134         { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
00135     },
00136     1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
00137 };
00138 
00139 static const ModeTab mode_22_24 = {
00140     {
00141         { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
00142         { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
00143         { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
00144     },
00145     1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
00146 };
00147 
00148 static const ModeTab mode_22_32 = {
00149     {
00150         { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
00151         { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
00152         { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
00153     },
00154     512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
00155 };
00156 
00157 static const ModeTab mode_44_40 = {
00158     {
00159         {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
00160         { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
00161         { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
00162     },
00163     2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
00164 };
00165 
00166 static const ModeTab mode_44_48 = {
00167     {
00168         {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
00169         { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
00170         { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
00171     },
00172     2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
00173 };
00174 
00175 typedef struct TwinContext {
00176     AVCodecContext *avctx;
00177     AVFrame frame;
00178     DSPContext      dsp;
00179     FFTContext mdct_ctx[3];
00180 
00181     const ModeTab *mtab;
00182 
00183     // history
00184     float lsp_hist[2][20];           
00185     float bark_hist[3][2][40];       
00186 
00187     // bitstream parameters
00188     int16_t permut[4][4096];
00189     uint8_t length[4][2];            
00190     uint8_t length_change[4];
00191     uint8_t bits_main_spec[2][4][2]; 
00192     int bits_main_spec_change[4];
00193     int n_div[4];
00194 
00195     float *spectrum;
00196     float *curr_frame;               
00197     float *prev_frame;               
00198     int last_block_pos[2];
00199     int discarded_packets;
00200 
00201     float *cos_tabs[3];
00202 
00203     // scratch buffers
00204     float *tmp_buf;
00205 } TwinContext;
00206 
00207 #define PPC_SHAPE_CB_SIZE 64
00208 #define PPC_SHAPE_LEN_MAX 60
00209 #define SUB_AMP_MAX       4500.0
00210 #define MULAW_MU          100.0
00211 #define GAIN_BITS         8
00212 #define AMP_MAX           13000.0
00213 #define SUB_GAIN_BITS     5
00214 #define WINDOW_TYPE_BITS  4
00215 #define PGAIN_MU          200
00216 #define LSP_COEFS_MAX     20
00217 #define LSP_SPLIT_MAX     4
00218 #define CHANNELS_MAX      2
00219 #define SUBBLOCKS_MAX     16
00220 #define BARK_N_COEF_MAX   4
00221 
00223 static void memset_float(float *buf, float val, int size)
00224 {
00225     while (size--)
00226         *buf++ = val;
00227 }
00228 
00241 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
00242 {
00243     int j;
00244     float p = 0.5f;
00245     float q = 0.5f;
00246     float two_cos_w = 2.0f*cos_val;
00247 
00248     for (j = 0; j + 1 < order; j += 2*2) {
00249         // Unroll the loop once since order is a multiple of four
00250         q *= lsp[j  ] - two_cos_w;
00251         p *= lsp[j+1] - two_cos_w;
00252 
00253         q *= lsp[j+2] - two_cos_w;
00254         p *= lsp[j+3] - two_cos_w;
00255     }
00256 
00257     p *= p * (2.0f - two_cos_w);
00258     q *= q * (2.0f + two_cos_w);
00259 
00260     return 0.5 / (p + q);
00261 }
00262 
00266 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
00267 {
00268     int i;
00269     const ModeTab *mtab = tctx->mtab;
00270     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00271 
00272     for (i = 0; i < size_s/2; i++) {
00273         float cos_i = tctx->cos_tabs[0][i];
00274         lpc[i]          = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
00275         lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
00276     }
00277 }
00278 
00279 static void interpolate(float *out, float v1, float v2, int size)
00280 {
00281     int i;
00282     float step = (v1 - v2)/(size + 1);
00283 
00284     for (i = 0; i < size; i++) {
00285         v2 += step;
00286         out[i] = v2;
00287     }
00288 }
00289 
00290 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
00291 {
00292     return part ? -cos_tab[size - idx - 1] :
00293                    cos_tab[       idx    ];
00294 }
00295 
00310 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
00311                                          enum FrameType ftype,
00312                                          float *out, const float *in,
00313                                          int size, int step, int part)
00314 {
00315     int i;
00316     const ModeTab *mtab = tctx->mtab;
00317     const float *cos_tab = tctx->cos_tabs[ftype];
00318 
00319     // Fill the 's'
00320     for (i = 0; i < size; i += step)
00321         out[i] =
00322             eval_lpc_spectrum(in,
00323                               get_cos(i, part, cos_tab, size),
00324                               mtab->n_lsp);
00325 
00326     // Fill the 'iiiibiiii'
00327     for (i = step; i <= size - 2*step; i += step) {
00328         if (out[i + step] + out[i - step] >  1.95*out[i] ||
00329             out[i + step]                 >=  out[i - step]) {
00330             interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
00331         } else {
00332             out[i - step/2] =
00333                 eval_lpc_spectrum(in,
00334                                   get_cos(i-step/2, part, cos_tab, size),
00335                                   mtab->n_lsp);
00336             interpolate(out + i - step   + 1, out[i-step/2], out[i-step  ], step/2 - 1);
00337             interpolate(out + i - step/2 + 1, out[i       ], out[i-step/2], step/2 - 1);
00338         }
00339     }
00340 
00341     interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
00342 }
00343 
00344 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
00345                                const float *buf, float *lpc,
00346                                int size, int step)
00347 {
00348     eval_lpcenv_or_interp(tctx, ftype, lpc         , buf, size/2,   step, 0);
00349     eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
00350 
00351     interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
00352 
00353     memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
00354 }
00355 
00361 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
00362                     enum FrameType ftype,
00363                     const int16_t *cb0, const int16_t *cb1, int cb_len)
00364 {
00365     int pos = 0;
00366     int i, j;
00367 
00368     for (i = 0; i < tctx->n_div[ftype]; i++) {
00369         int tmp0, tmp1;
00370         int sign0 = 1;
00371         int sign1 = 1;
00372         const int16_t *tab0, *tab1;
00373         int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
00374         int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
00375 
00376         int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
00377         if (bits == 7) {
00378             if (get_bits1(gb))
00379                 sign0 = -1;
00380             bits = 6;
00381         }
00382         tmp0 = get_bits(gb, bits);
00383 
00384         bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
00385 
00386         if (bits == 7) {
00387             if (get_bits1(gb))
00388                 sign1 = -1;
00389 
00390             bits = 6;
00391         }
00392         tmp1 = get_bits(gb, bits);
00393 
00394         tab0 = cb0 + tmp0*cb_len;
00395         tab1 = cb1 + tmp1*cb_len;
00396 
00397         for (j = 0; j < length; j++)
00398             out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
00399 
00400         pos += length;
00401     }
00402 
00403 }
00404 
00405 static inline float mulawinv(float y, float clip, float mu)
00406 {
00407     y = av_clipf(y/clip, -1, 1);
00408     return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
00409 }
00410 
00431 static int very_broken_op(int a, int b)
00432 {
00433     int x = a*b + 200;
00434     int size;
00435     const uint8_t *rtab;
00436 
00437     if (x%400 || b%5)
00438         return x/400;
00439 
00440     x /= 400;
00441 
00442     size = tabs[b/5].size;
00443     rtab = tabs[b/5].tab;
00444     return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
00445 }
00446 
00452 static void add_peak(int period, int width, const float *shape,
00453                      float ppc_gain, float *speech, int len)
00454 {
00455     int i, j;
00456 
00457     const float *shape_end = shape + len;
00458     int center;
00459 
00460     // First peak centered around zero
00461     for (i = 0; i < width/2; i++)
00462         speech[i] += ppc_gain * *shape++;
00463 
00464     for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
00465         center = very_broken_op(period, i);
00466         for (j = -width/2; j < (width+1)/2; j++)
00467             speech[j+center] += ppc_gain * *shape++;
00468     }
00469 
00470     // For the last block, be careful not to go beyond the end of the buffer
00471     center = very_broken_op(period, i);
00472     for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
00473         speech[j+center] += ppc_gain * *shape++;
00474 }
00475 
00476 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
00477                        float ppc_gain, float *speech)
00478 {
00479     const ModeTab *mtab = tctx->mtab;
00480     int isampf = tctx->avctx->sample_rate/1000;
00481     int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
00482     int min_period = ROUNDED_DIV(  40*2*mtab->size, isampf);
00483     int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
00484     int period_range = max_period - min_period;
00485 
00486     // This is actually the period multiplied by 400. It is just linearly coded
00487     // between its maximum and minimum value.
00488     int period = min_period +
00489         ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
00490     int width;
00491 
00492     if (isampf == 22 && ibps == 32) {
00493         // For some unknown reason, NTT decided to code this case differently...
00494         width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
00495     } else
00496         width =             (period      )* mtab->peak_per2wid/(400*mtab->size);
00497 
00498     add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
00499 }
00500 
00501 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
00502                      float *out)
00503 {
00504     const ModeTab *mtab = tctx->mtab;
00505     int i, j;
00506     int sub = mtab->fmode[ftype].sub;
00507     float step     = AMP_MAX     / ((1 <<     GAIN_BITS) - 1);
00508     float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
00509 
00510     if (ftype == FT_LONG) {
00511         for (i = 0; i < tctx->avctx->channels; i++)
00512             out[i] = (1./(1<<13)) *
00513                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00514                          AMP_MAX, MULAW_MU);
00515     } else {
00516         for (i = 0; i < tctx->avctx->channels; i++) {
00517             float val = (1./(1<<23)) *
00518                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00519                          AMP_MAX, MULAW_MU);
00520 
00521             for (j = 0; j < sub; j++) {
00522                 out[i*sub + j] =
00523                     val*mulawinv(sub_step* 0.5 +
00524                                  sub_step* get_bits(gb, SUB_GAIN_BITS),
00525                                  SUB_AMP_MAX, MULAW_MU);
00526             }
00527         }
00528     }
00529 }
00530 
00537 static void rearrange_lsp(int order, float *lsp, float min_dist)
00538 {
00539     int i;
00540     float min_dist2 = min_dist * 0.5;
00541     for (i = 1; i < order; i++)
00542         if (lsp[i] - lsp[i-1] < min_dist) {
00543             float avg = (lsp[i] + lsp[i-1]) * 0.5;
00544 
00545             lsp[i-1] = avg - min_dist2;
00546             lsp[i  ] = avg + min_dist2;
00547         }
00548 }
00549 
00550 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
00551                        int lpc_hist_idx, float *lsp, float *hist)
00552 {
00553     const ModeTab *mtab = tctx->mtab;
00554     int i, j;
00555 
00556     const float *cb  =  mtab->lspcodebook;
00557     const float *cb2 =  cb  + (1 << mtab->lsp_bit1)*mtab->n_lsp;
00558     const float *cb3 =  cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
00559 
00560     const int8_t funny_rounding[4] = {
00561         -2,
00562         mtab->lsp_split == 4 ? -2 : 1,
00563         mtab->lsp_split == 4 ? -2 : 1,
00564         0
00565     };
00566 
00567     j = 0;
00568     for (i = 0; i < mtab->lsp_split; i++) {
00569         int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
00570         for (; j < chunk_end; j++)
00571             lsp[j] = cb [lpc_idx1    * mtab->n_lsp + j] +
00572                      cb2[lpc_idx2[i] * mtab->n_lsp + j];
00573     }
00574 
00575     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00576 
00577     for (i = 0; i < mtab->n_lsp; i++) {
00578         float tmp1 = 1. -          cb3[lpc_hist_idx*mtab->n_lsp + i];
00579         float tmp2 =     hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
00580         hist[i] = lsp[i];
00581         lsp[i]  = lsp[i] * tmp1 + tmp2;
00582     }
00583 
00584     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00585     rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
00586     ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
00587 }
00588 
00589 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
00590                                  enum FrameType ftype, float *lpc)
00591 {
00592     int i;
00593     int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
00594 
00595     for (i = 0; i < tctx->mtab->n_lsp; i++)
00596         lsp[i] =  2*cos(lsp[i]);
00597 
00598     switch (ftype) {
00599     case FT_LONG:
00600         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
00601         break;
00602     case FT_MEDIUM:
00603         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
00604         break;
00605     case FT_SHORT:
00606         eval_lpcenv(tctx, lsp, lpc);
00607         break;
00608     }
00609 }
00610 
00611 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
00612                             float *in, float *prev, int ch)
00613 {
00614     FFTContext *mdct = &tctx->mdct_ctx[ftype];
00615     const ModeTab *mtab = tctx->mtab;
00616     int bsize = mtab->size / mtab->fmode[ftype].sub;
00617     int size  = mtab->size;
00618     float *buf1 = tctx->tmp_buf;
00619     int j;
00620     int wsize; // Window size
00621     float *out = tctx->curr_frame + 2*ch*mtab->size;
00622     float *out2 = out;
00623     float *prev_buf;
00624     int first_wsize;
00625 
00626     static const uint8_t wtype_to_wsize[]      = {0, 0, 2, 2, 2, 1, 0, 1, 1};
00627     int types_sizes[] = {
00628         mtab->size /    mtab->fmode[FT_LONG  ].sub,
00629         mtab->size /    mtab->fmode[FT_MEDIUM].sub,
00630         mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
00631     };
00632 
00633     wsize = types_sizes[wtype_to_wsize[wtype]];
00634     first_wsize = wsize;
00635     prev_buf = prev + (size - bsize)/2;
00636 
00637     for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00638         int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
00639 
00640         if (!j && wtype == 4)
00641             sub_wtype = 4;
00642         else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
00643             sub_wtype = 7;
00644 
00645         wsize = types_sizes[wtype_to_wsize[sub_wtype]];
00646 
00647         mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
00648 
00649         tctx->dsp.vector_fmul_window(out2,
00650                                      prev_buf + (bsize-wsize)/2,
00651                                      buf1 + bsize*j,
00652                                      ff_sine_windows[av_log2(wsize)],
00653                                      wsize/2);
00654         out2 += wsize;
00655 
00656         memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
00657 
00658         out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
00659 
00660         prev_buf = buf1 + bsize*j + bsize/2;
00661     }
00662 
00663     tctx->last_block_pos[ch] = (size + first_wsize)/2;
00664 }
00665 
00666 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
00667                          float *out)
00668 {
00669     const ModeTab *mtab = tctx->mtab;
00670     int size1, size2;
00671     float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
00672     int i;
00673 
00674     for (i = 0; i < tctx->avctx->channels; i++) {
00675         imdct_and_window(tctx, ftype, wtype,
00676                          tctx->spectrum + i*mtab->size,
00677                          prev_buf + 2*i*mtab->size,
00678                          i);
00679     }
00680 
00681     if (!out)
00682         return;
00683 
00684     size2 = tctx->last_block_pos[0];
00685     size1 = mtab->size - size2;
00686     if (tctx->avctx->channels == 2) {
00687         tctx->dsp.butterflies_float_interleave(out, prev_buf,
00688                                                &prev_buf[2*mtab->size],
00689                                                size1);
00690 
00691         out += 2 * size1;
00692 
00693         tctx->dsp.butterflies_float_interleave(out, tctx->curr_frame,
00694                                                &tctx->curr_frame[2*mtab->size],
00695                                                size2);
00696     } else {
00697         memcpy(out, prev_buf, size1 * sizeof(*out));
00698 
00699         out += size1;
00700 
00701         memcpy(out, tctx->curr_frame, size2 * sizeof(*out));
00702     }
00703 
00704 }
00705 
00706 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
00707                          int ch, float *out, float gain, enum FrameType ftype)
00708 {
00709     const ModeTab *mtab = tctx->mtab;
00710     int i,j;
00711     float *hist = tctx->bark_hist[ftype][ch];
00712     float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
00713     int bark_n_coef  = mtab->fmode[ftype].bark_n_coef;
00714     int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
00715     int idx = 0;
00716 
00717     for (i = 0; i < fw_cb_len; i++)
00718         for (j = 0; j < bark_n_coef; j++, idx++) {
00719             float tmp2 =
00720                 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
00721             float st = use_hist ?
00722                 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
00723 
00724             hist[idx] = tmp2;
00725             if (st < -1.) st = 1.;
00726 
00727             memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
00728             out += mtab->fmode[ftype].bark_tab[idx];
00729         }
00730 
00731 }
00732 
00733 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
00734                                      float *out, enum FrameType ftype)
00735 {
00736     const ModeTab *mtab = tctx->mtab;
00737     int channels = tctx->avctx->channels;
00738     int sub = mtab->fmode[ftype].sub;
00739     int block_size = mtab->size / sub;
00740     float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
00741     float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
00742     uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
00743     uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
00744 
00745     uint8_t lpc_idx1[CHANNELS_MAX];
00746     uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
00747     uint8_t lpc_hist_idx[CHANNELS_MAX];
00748 
00749     int i, j, k;
00750 
00751     dequant(tctx, gb, out, ftype,
00752             mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
00753             mtab->fmode[ftype].cb_len_read);
00754 
00755     for (i = 0; i < channels; i++)
00756         for (j = 0; j < sub; j++)
00757             for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
00758                 bark1[i][j][k] =
00759                     get_bits(gb, mtab->fmode[ftype].bark_n_bit);
00760 
00761     for (i = 0; i < channels; i++)
00762         for (j = 0; j < sub; j++)
00763             bark_use_hist[i][j] = get_bits1(gb);
00764 
00765     dec_gain(tctx, gb, ftype, gain);
00766 
00767     for (i = 0; i < channels; i++) {
00768         lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
00769         lpc_idx1    [i] = get_bits(gb, tctx->mtab->lsp_bit1);
00770 
00771         for (j = 0; j < tctx->mtab->lsp_split; j++)
00772             lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
00773     }
00774 
00775     if (ftype == FT_LONG) {
00776         int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
00777             tctx->n_div[3];
00778         dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
00779                 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
00780     }
00781 
00782     for (i = 0; i < channels; i++) {
00783         float *chunk = out + mtab->size * i;
00784         float lsp[LSP_COEFS_MAX];
00785 
00786         for (j = 0; j < sub; j++) {
00787             dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
00788                          tctx->tmp_buf, gain[sub*i+j], ftype);
00789 
00790             tctx->dsp.vector_fmul(chunk + block_size*j, chunk + block_size*j, tctx->tmp_buf,
00791                                   block_size);
00792 
00793         }
00794 
00795         if (ftype == FT_LONG) {
00796             float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
00797             int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
00798             int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
00799             float v = 1./8192*
00800                 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
00801 
00802             decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
00803                        chunk);
00804         }
00805 
00806         decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
00807                    tctx->lsp_hist[i]);
00808 
00809         dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
00810 
00811         for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00812             tctx->dsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
00813             chunk += block_size;
00814         }
00815     }
00816 }
00817 
00818 static int twin_decode_frame(AVCodecContext * avctx, void *data,
00819                              int *got_frame_ptr, AVPacket *avpkt)
00820 {
00821     const uint8_t *buf = avpkt->data;
00822     int buf_size = avpkt->size;
00823     TwinContext *tctx = avctx->priv_data;
00824     GetBitContext gb;
00825     const ModeTab *mtab = tctx->mtab;
00826     float *out = NULL;
00827     enum FrameType ftype;
00828     int window_type, ret;
00829     static const enum FrameType wtype_to_ftype_table[] = {
00830         FT_LONG,   FT_LONG, FT_SHORT, FT_LONG,
00831         FT_MEDIUM, FT_LONG, FT_LONG,  FT_MEDIUM, FT_MEDIUM
00832     };
00833 
00834     if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
00835         av_log(avctx, AV_LOG_ERROR,
00836                "Frame too small (%d bytes). Truncated file?\n", buf_size);
00837         return AVERROR(EINVAL);
00838     }
00839 
00840     /* get output buffer */
00841     if (tctx->discarded_packets >= 2) {
00842         tctx->frame.nb_samples = mtab->size;
00843         if ((ret = avctx->get_buffer(avctx, &tctx->frame)) < 0) {
00844             av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
00845             return ret;
00846         }
00847         out = (float *)tctx->frame.data[0];
00848     }
00849 
00850     init_get_bits(&gb, buf, buf_size * 8);
00851     skip_bits(&gb, get_bits(&gb, 8));
00852     window_type = get_bits(&gb, WINDOW_TYPE_BITS);
00853 
00854     if (window_type > 8) {
00855         av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
00856         return -1;
00857     }
00858 
00859     ftype = wtype_to_ftype_table[window_type];
00860 
00861     read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
00862 
00863     imdct_output(tctx, ftype, window_type, out);
00864 
00865     FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
00866 
00867     if (tctx->discarded_packets < 2) {
00868         tctx->discarded_packets++;
00869         *got_frame_ptr = 0;
00870         return buf_size;
00871     }
00872 
00873     *got_frame_ptr   = 1;
00874     *(AVFrame *)data = tctx->frame;;
00875 
00876     return buf_size;
00877 }
00878 
00882 static av_cold int init_mdct_win(TwinContext *tctx)
00883 {
00884     int i, j, ret;
00885     const ModeTab *mtab = tctx->mtab;
00886     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00887     int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
00888     int channels = tctx->avctx->channels;
00889     float norm = channels == 1 ? 2. : 1.;
00890 
00891     for (i = 0; i < 3; i++) {
00892         int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
00893         if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
00894                                 -sqrt(norm/bsize) / (1<<15))))
00895             return ret;
00896     }
00897 
00898     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
00899                      mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
00900 
00901     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
00902                      2 * mtab->size * channels * sizeof(*tctx->spectrum),
00903                      alloc_fail);
00904     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
00905                      2 * mtab->size * channels * sizeof(*tctx->curr_frame),
00906                      alloc_fail);
00907     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
00908                      2 * mtab->size * channels * sizeof(*tctx->prev_frame),
00909                      alloc_fail);
00910 
00911     for (i = 0; i < 3; i++) {
00912         int m = 4*mtab->size/mtab->fmode[i].sub;
00913         double freq = 2*M_PI/m;
00914         FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
00915                          (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
00916 
00917         for (j = 0; j <= m/8; j++)
00918             tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
00919         for (j = 1; j <  m/8; j++)
00920             tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
00921     }
00922 
00923 
00924     ff_init_ff_sine_windows(av_log2(size_m));
00925     ff_init_ff_sine_windows(av_log2(size_s/2));
00926     ff_init_ff_sine_windows(av_log2(mtab->size));
00927 
00928     return 0;
00929 alloc_fail:
00930     return AVERROR(ENOMEM);
00931 }
00932 
00939 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
00940                               int block_size,
00941                               const uint8_t line_len[2], int length_div,
00942                               enum FrameType ftype)
00943 
00944 {
00945     int i,j;
00946 
00947     for (i = 0; i < line_len[0]; i++) {
00948         int shift;
00949 
00950         if (num_blocks == 1 ||
00951             (ftype == FT_LONG && num_vect % num_blocks) ||
00952             (ftype != FT_LONG && num_vect & 1         ) ||
00953             i == line_len[1]) {
00954             shift = 0;
00955         } else if (ftype == FT_LONG) {
00956             shift = i;
00957         } else
00958             shift = i*i;
00959 
00960         for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
00961             tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
00962     }
00963 }
00964 
00980 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
00981                            const uint8_t line_len[2], int length_div)
00982 {
00983     int i,j;
00984     int cont= 0;
00985     for (i = 0; i < num_vect; i++)
00986         for (j = 0; j < line_len[i >= length_div]; j++)
00987             out[cont++] = in[j*num_vect + i];
00988 }
00989 
00990 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
00991 {
00992     int block_size = size/n_blocks;
00993     int i;
00994 
00995     for (i = 0; i < size; i++)
00996         out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
00997 }
00998 
00999 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
01000 {
01001     int block_size;
01002     const ModeTab *mtab = tctx->mtab;
01003     int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
01004     int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
01005 
01006     if (ftype == FT_PPC) {
01007         size  = tctx->avctx->channels;
01008         block_size = mtab->ppc_shape_len;
01009     } else
01010         block_size = mtab->size / mtab->fmode[ftype].sub;
01011 
01012     permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
01013                       block_size, tctx->length[ftype],
01014                       tctx->length_change[ftype], ftype);
01015 
01016     transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
01017                    tctx->length[ftype], tctx->length_change[ftype]);
01018 
01019     linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
01020                 size*block_size);
01021 }
01022 
01023 static av_cold void init_bitstream_params(TwinContext *tctx)
01024 {
01025     const ModeTab *mtab = tctx->mtab;
01026     int n_ch = tctx->avctx->channels;
01027     int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
01028                              tctx->avctx->sample_rate;
01029 
01030     int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
01031                                    mtab->lsp_split*mtab->lsp_bit2);
01032 
01033     int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
01034                          mtab->ppc_period_bit);
01035 
01036     int bsize_no_main_cb[3];
01037     int bse_bits[3];
01038     int i;
01039     enum FrameType frametype;
01040 
01041     for (i = 0; i < 3; i++)
01042         // +1 for history usage switch
01043         bse_bits[i] = n_ch *
01044             (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
01045 
01046     bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
01047                           WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
01048 
01049     for (i = 0; i < 2; i++)
01050         bsize_no_main_cb[i] =
01051             lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
01052             mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
01053 
01054     // The remaining bits are all used for the main spectrum coefficients
01055     for (i = 0; i < 4; i++) {
01056         int bit_size;
01057         int vect_size;
01058         int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
01059         if (i == 3) {
01060             bit_size  = n_ch * mtab->ppc_shape_bit;
01061             vect_size = n_ch * mtab->ppc_shape_len;
01062         } else {
01063             bit_size = total_fr_bits - bsize_no_main_cb[i];
01064             vect_size = n_ch * mtab->size;
01065         }
01066 
01067         tctx->n_div[i] = (bit_size + 13) / 14;
01068 
01069         rounded_up   = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01070         rounded_down = (bit_size           )/tctx->n_div[i];
01071         num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
01072         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01073         tctx->bits_main_spec[0][i][0] = (rounded_up   + 1)/2;
01074         tctx->bits_main_spec[1][i][0] = (rounded_up      )/2;
01075         tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
01076         tctx->bits_main_spec[1][i][1] = (rounded_down    )/2;
01077         tctx->bits_main_spec_change[i] = num_rounded_up;
01078 
01079         rounded_up   = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01080         rounded_down = (vect_size                     )/tctx->n_div[i];
01081         num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
01082         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01083         tctx->length[i][0] = rounded_up;
01084         tctx->length[i][1] = rounded_down;
01085         tctx->length_change[i] = num_rounded_up;
01086     }
01087 
01088     for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
01089         construct_perm_table(tctx, frametype);
01090 }
01091 
01092 static av_cold int twin_decode_close(AVCodecContext *avctx)
01093 {
01094     TwinContext *tctx = avctx->priv_data;
01095     int i;
01096 
01097     for (i = 0; i < 3; i++) {
01098         ff_mdct_end(&tctx->mdct_ctx[i]);
01099         av_free(tctx->cos_tabs[i]);
01100     }
01101 
01102 
01103     av_free(tctx->curr_frame);
01104     av_free(tctx->spectrum);
01105     av_free(tctx->prev_frame);
01106     av_free(tctx->tmp_buf);
01107 
01108     return 0;
01109 }
01110 
01111 static av_cold int twin_decode_init(AVCodecContext *avctx)
01112 {
01113     int ret;
01114     TwinContext *tctx = avctx->priv_data;
01115     int isampf, ibps;
01116 
01117     tctx->avctx       = avctx;
01118     avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
01119 
01120     if (!avctx->extradata || avctx->extradata_size < 12) {
01121         av_log(avctx, AV_LOG_ERROR, "Missing or incomplete extradata\n");
01122         return AVERROR_INVALIDDATA;
01123     }
01124     avctx->channels = AV_RB32(avctx->extradata    ) + 1;
01125     avctx->bit_rate = AV_RB32(avctx->extradata + 4) * 1000;
01126     isampf          = AV_RB32(avctx->extradata + 8);
01127     switch (isampf) {
01128     case 44: avctx->sample_rate = 44100;         break;
01129     case 22: avctx->sample_rate = 22050;         break;
01130     case 11: avctx->sample_rate = 11025;         break;
01131     default: avctx->sample_rate = isampf * 1000; break;
01132     }
01133 
01134     if (avctx->channels > CHANNELS_MAX) {
01135         av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
01136                avctx->channels);
01137         return -1;
01138     }
01139     ibps = avctx->bit_rate / (1000 * avctx->channels);
01140 
01141     switch ((isampf << 8) +  ibps) {
01142     case (8 <<8) +  8: tctx->mtab = &mode_08_08; break;
01143     case (11<<8) +  8: tctx->mtab = &mode_11_08; break;
01144     case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
01145     case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
01146     case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
01147     case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
01148     case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
01149     case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
01150     case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
01151     default:
01152         av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
01153         return -1;
01154     }
01155 
01156     dsputil_init(&tctx->dsp, avctx);
01157     if ((ret = init_mdct_win(tctx))) {
01158         av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
01159         twin_decode_close(avctx);
01160         return ret;
01161     }
01162     init_bitstream_params(tctx);
01163 
01164     memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
01165 
01166     avcodec_get_frame_defaults(&tctx->frame);
01167     avctx->coded_frame = &tctx->frame;
01168 
01169     return 0;
01170 }
01171 
01172 AVCodec ff_twinvq_decoder = {
01173     .name           = "twinvq",
01174     .type           = AVMEDIA_TYPE_AUDIO,
01175     .id             = CODEC_ID_TWINVQ,
01176     .priv_data_size = sizeof(TwinContext),
01177     .init           = twin_decode_init,
01178     .close          = twin_decode_close,
01179     .decode         = twin_decode_frame,
01180     .capabilities   = CODEC_CAP_DR1,
01181     .long_name      = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
01182 };
Generated on Sat Mar 17 2012 12:57:50 for Libav by doxygen 1.7.1