• 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 FFmpeg.
00006  *
00007  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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     DSPContext      dsp;
00178     FFTContext mdct_ctx[3];
00179 
00180     const ModeTab *mtab;
00181 
00182     // history
00183     float lsp_hist[2][20];           
00184     float bark_hist[3][2][40];       
00185 
00186     // bitstream parameters
00187     int16_t permut[4][4096];
00188     uint8_t length[4][2];            
00189     uint8_t length_change[4];
00190     uint8_t bits_main_spec[2][4][2]; 
00191     int bits_main_spec_change[4];
00192     int n_div[4];
00193 
00194     float *spectrum;
00195     float *curr_frame;               
00196     float *prev_frame;               
00197     int last_block_pos[2];
00198 
00199     float *cos_tabs[3];
00200 
00201     // scratch buffers
00202     float *tmp_buf;
00203 } TwinContext;
00204 
00205 #define PPC_SHAPE_CB_SIZE 64
00206 #define PPC_SHAPE_LEN_MAX 60
00207 #define SUB_AMP_MAX       4500.0
00208 #define MULAW_MU          100.0
00209 #define GAIN_BITS         8
00210 #define AMP_MAX           13000.0
00211 #define SUB_GAIN_BITS     5
00212 #define WINDOW_TYPE_BITS  4
00213 #define PGAIN_MU          200
00214 #define LSP_COEFS_MAX     20
00215 #define LSP_SPLIT_MAX     4
00216 #define CHANNELS_MAX      2
00217 #define SUBBLOCKS_MAX     16
00218 #define BARK_N_COEF_MAX   4
00219 
00221 static void memset_float(float *buf, float val, int size)
00222 {
00223     while (size--)
00224         *buf++ = val;
00225 }
00226 
00239 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
00240 {
00241     int j;
00242     float p = 0.5f;
00243     float q = 0.5f;
00244     float two_cos_w = 2.0f*cos_val;
00245 
00246     for (j = 0; j + 1 < order; j += 2*2) {
00247         // Unroll the loop once since order is a multiple of four
00248         q *= lsp[j  ] - two_cos_w;
00249         p *= lsp[j+1] - two_cos_w;
00250 
00251         q *= lsp[j+2] - two_cos_w;
00252         p *= lsp[j+3] - two_cos_w;
00253     }
00254 
00255     p *= p * (2.0f - two_cos_w);
00256     q *= q * (2.0f + two_cos_w);
00257 
00258     return 0.5 / (p + q);
00259 }
00260 
00264 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
00265 {
00266     int i;
00267     const ModeTab *mtab = tctx->mtab;
00268     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00269 
00270     for (i = 0; i < size_s/2; i++) {
00271         float cos_i = tctx->cos_tabs[0][i];
00272         lpc[i]          = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
00273         lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
00274     }
00275 }
00276 
00277 static void interpolate(float *out, float v1, float v2, int size)
00278 {
00279     int i;
00280     float step = (v1 - v2)/(size + 1);
00281 
00282     for (i = 0; i < size; i++) {
00283         v2 += step;
00284         out[i] = v2;
00285     }
00286 }
00287 
00288 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
00289 {
00290     return part ? -cos_tab[size - idx - 1] :
00291                    cos_tab[       idx    ];
00292 }
00293 
00308 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
00309                                          enum FrameType ftype,
00310                                          float *out, const float *in,
00311                                          int size, int step, int part)
00312 {
00313     int i;
00314     const ModeTab *mtab = tctx->mtab;
00315     const float *cos_tab = tctx->cos_tabs[ftype];
00316 
00317     // Fill the 's'
00318     for (i = 0; i < size; i += step)
00319         out[i] =
00320             eval_lpc_spectrum(in,
00321                               get_cos(i, part, cos_tab, size),
00322                               mtab->n_lsp);
00323 
00324     // Fill the 'iiiibiiii'
00325     for (i = step; i <= size - 2*step; i += step) {
00326         if (out[i + step] + out[i - step] >  1.95*out[i] ||
00327             out[i + step]                 >=  out[i - step]) {
00328             interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
00329         } else {
00330             out[i - step/2] =
00331                 eval_lpc_spectrum(in,
00332                                   get_cos(i-step/2, part, cos_tab, size),
00333                                   mtab->n_lsp);
00334             interpolate(out + i - step   + 1, out[i-step/2], out[i-step  ], step/2 - 1);
00335             interpolate(out + i - step/2 + 1, out[i       ], out[i-step/2], step/2 - 1);
00336         }
00337     }
00338 
00339     interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
00340 }
00341 
00342 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
00343                                const float *buf, float *lpc,
00344                                int size, int step)
00345 {
00346     eval_lpcenv_or_interp(tctx, ftype, lpc         , buf, size/2,   step, 0);
00347     eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
00348 
00349     interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
00350 
00351     memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
00352 }
00353 
00359 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
00360                     enum FrameType ftype,
00361                     const int16_t *cb0, const int16_t *cb1, int cb_len)
00362 {
00363     int pos = 0;
00364     int i, j;
00365 
00366     for (i = 0; i < tctx->n_div[ftype]; i++) {
00367         int tmp0, tmp1;
00368         int sign0 = 1;
00369         int sign1 = 1;
00370         const int16_t *tab0, *tab1;
00371         int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
00372         int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
00373 
00374         int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
00375         if (bits == 7) {
00376             if (get_bits1(gb))
00377                 sign0 = -1;
00378             bits = 6;
00379         }
00380         tmp0 = get_bits(gb, bits);
00381 
00382         bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
00383 
00384         if (bits == 7) {
00385             if (get_bits1(gb))
00386                 sign1 = -1;
00387 
00388             bits = 6;
00389         }
00390         tmp1 = get_bits(gb, bits);
00391 
00392         tab0 = cb0 + tmp0*cb_len;
00393         tab1 = cb1 + tmp1*cb_len;
00394 
00395         for (j = 0; j < length; j++)
00396             out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
00397 
00398         pos += length;
00399     }
00400 
00401 }
00402 
00403 static inline float mulawinv(float y, float clip, float mu)
00404 {
00405     y = av_clipf(y/clip, -1, 1);
00406     return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
00407 }
00408 
00429 static int very_broken_op(int a, int b)
00430 {
00431     int x = a*b + 200;
00432     int size;
00433     const uint8_t *rtab;
00434 
00435     if (x%400 || b%5)
00436         return x/400;
00437 
00438     x /= 400;
00439 
00440     size = tabs[b/5].size;
00441     rtab = tabs[b/5].tab;
00442     return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
00443 }
00444 
00450 static void add_peak(int period, int width, const float *shape,
00451                      float ppc_gain, float *speech, int len)
00452 {
00453     int i, j;
00454 
00455     const float *shape_end = shape + len;
00456     int center;
00457 
00458     // First peak centered around zero
00459     for (i = 0; i < width/2; i++)
00460         speech[i] += ppc_gain * *shape++;
00461 
00462     for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
00463         center = very_broken_op(period, i);
00464         for (j = -width/2; j < (width+1)/2; j++)
00465             speech[j+center] += ppc_gain * *shape++;
00466     }
00467 
00468     // For the last block, be careful not to go beyond the end of the buffer
00469     center = very_broken_op(period, i);
00470     for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
00471         speech[j+center] += ppc_gain * *shape++;
00472 }
00473 
00474 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
00475                        float ppc_gain, float *speech)
00476 {
00477     const ModeTab *mtab = tctx->mtab;
00478     int isampf = tctx->avctx->sample_rate/1000;
00479     int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
00480     int min_period = ROUNDED_DIV(  40*2*mtab->size, isampf);
00481     int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
00482     int period_range = max_period - min_period;
00483 
00484     // This is actually the period multiplied by 400. It is just linearly coded
00485     // between its maximum and minimum value.
00486     int period = min_period +
00487         ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
00488     int width;
00489 
00490     if (isampf == 22 && ibps == 32) {
00491         // For some unknown reason, NTT decided to code this case differently...
00492         width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
00493     } else
00494         width =             (period      )* mtab->peak_per2wid/(400*mtab->size);
00495 
00496     add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
00497 }
00498 
00499 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
00500                      float *out)
00501 {
00502     const ModeTab *mtab = tctx->mtab;
00503     int i, j;
00504     int sub = mtab->fmode[ftype].sub;
00505     float step     = AMP_MAX     / ((1 <<     GAIN_BITS) - 1);
00506     float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
00507 
00508     if (ftype == FT_LONG) {
00509         for (i = 0; i < tctx->avctx->channels; i++)
00510             out[i] = (1./(1<<13)) *
00511                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00512                          AMP_MAX, MULAW_MU);
00513     } else {
00514         for (i = 0; i < tctx->avctx->channels; i++) {
00515             float val = (1./(1<<23)) *
00516                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00517                          AMP_MAX, MULAW_MU);
00518 
00519             for (j = 0; j < sub; j++) {
00520                 out[i*sub + j] =
00521                     val*mulawinv(sub_step* 0.5 +
00522                                  sub_step* get_bits(gb, SUB_GAIN_BITS),
00523                                  SUB_AMP_MAX, MULAW_MU);
00524             }
00525         }
00526     }
00527 }
00528 
00535 static void rearrange_lsp(int order, float *lsp, float min_dist)
00536 {
00537     int i;
00538     float min_dist2 = min_dist * 0.5;
00539     for (i = 1; i < order; i++)
00540         if (lsp[i] - lsp[i-1] < min_dist) {
00541             float avg = (lsp[i] + lsp[i-1]) * 0.5;
00542 
00543             lsp[i-1] = avg - min_dist2;
00544             lsp[i  ] = avg + min_dist2;
00545         }
00546 }
00547 
00548 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
00549                        int lpc_hist_idx, float *lsp, float *hist)
00550 {
00551     const ModeTab *mtab = tctx->mtab;
00552     int i, j;
00553 
00554     const float *cb  =  mtab->lspcodebook;
00555     const float *cb2 =  cb  + (1 << mtab->lsp_bit1)*mtab->n_lsp;
00556     const float *cb3 =  cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
00557 
00558     const int8_t funny_rounding[4] = {
00559         -2,
00560         mtab->lsp_split == 4 ? -2 : 1,
00561         mtab->lsp_split == 4 ? -2 : 1,
00562         0
00563     };
00564 
00565     j = 0;
00566     for (i = 0; i < mtab->lsp_split; i++) {
00567         int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
00568         for (; j < chunk_end; j++)
00569             lsp[j] = cb [lpc_idx1    * mtab->n_lsp + j] +
00570                      cb2[lpc_idx2[i] * mtab->n_lsp + j];
00571     }
00572 
00573     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00574 
00575     for (i = 0; i < mtab->n_lsp; i++) {
00576         float tmp1 = 1. -          cb3[lpc_hist_idx*mtab->n_lsp + i];
00577         float tmp2 =     hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
00578         hist[i] = lsp[i];
00579         lsp[i]  = lsp[i] * tmp1 + tmp2;
00580     }
00581 
00582     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00583     rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
00584     ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
00585 }
00586 
00587 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
00588                                  enum FrameType ftype, float *lpc)
00589 {
00590     int i;
00591     int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
00592 
00593     for (i = 0; i < tctx->mtab->n_lsp; i++)
00594         lsp[i] =  2*cos(lsp[i]);
00595 
00596     switch (ftype) {
00597     case FT_LONG:
00598         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
00599         break;
00600     case FT_MEDIUM:
00601         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
00602         break;
00603     case FT_SHORT:
00604         eval_lpcenv(tctx, lsp, lpc);
00605         break;
00606     }
00607 }
00608 
00609 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
00610                             float *in, float *prev, int ch)
00611 {
00612     FFTContext *mdct = &tctx->mdct_ctx[ftype];
00613     const ModeTab *mtab = tctx->mtab;
00614     int bsize = mtab->size / mtab->fmode[ftype].sub;
00615     int size  = mtab->size;
00616     float *buf1 = tctx->tmp_buf;
00617     int j;
00618     int wsize; // Window size
00619     float *out = tctx->curr_frame + 2*ch*mtab->size;
00620     float *out2 = out;
00621     float *prev_buf;
00622     int first_wsize;
00623 
00624     static const uint8_t wtype_to_wsize[]      = {0, 0, 2, 2, 2, 1, 0, 1, 1};
00625     int types_sizes[] = {
00626         mtab->size /    mtab->fmode[FT_LONG  ].sub,
00627         mtab->size /    mtab->fmode[FT_MEDIUM].sub,
00628         mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
00629     };
00630 
00631     wsize = types_sizes[wtype_to_wsize[wtype]];
00632     first_wsize = wsize;
00633     prev_buf = prev + (size - bsize)/2;
00634 
00635     for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00636         int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
00637 
00638         if (!j && wtype == 4)
00639             sub_wtype = 4;
00640         else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
00641             sub_wtype = 7;
00642 
00643         wsize = types_sizes[wtype_to_wsize[sub_wtype]];
00644 
00645         mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
00646 
00647         tctx->dsp.vector_fmul_window(out2,
00648                                      prev_buf + (bsize-wsize)/2,
00649                                      buf1 + bsize*j,
00650                                      ff_sine_windows[av_log2(wsize)],
00651                                      wsize/2);
00652         out2 += wsize;
00653 
00654         memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
00655 
00656         out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
00657 
00658         prev_buf = buf1 + bsize*j + bsize/2;
00659     }
00660 
00661     tctx->last_block_pos[ch] = (size + first_wsize)/2;
00662 }
00663 
00664 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
00665                          float *out)
00666 {
00667     const ModeTab *mtab = tctx->mtab;
00668     float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
00669     int i, j;
00670 
00671     for (i = 0; i < tctx->avctx->channels; i++) {
00672         imdct_and_window(tctx, ftype, wtype,
00673                          tctx->spectrum + i*mtab->size,
00674                          prev_buf + 2*i*mtab->size,
00675                          i);
00676     }
00677 
00678     if (tctx->avctx->channels == 2) {
00679         for (i = 0; i < mtab->size - tctx->last_block_pos[0]; i++) {
00680             float f1 = prev_buf[               i];
00681             float f2 = prev_buf[2*mtab->size + i];
00682             out[2*i    ] = f1 + f2;
00683             out[2*i + 1] = f1 - f2;
00684         }
00685         for (j = 0; i < mtab->size; j++,i++) {
00686             float f1 = tctx->curr_frame[               j];
00687             float f2 = tctx->curr_frame[2*mtab->size + j];
00688             out[2*i    ] = f1 + f2;
00689             out[2*i + 1] = f1 - f2;
00690         }
00691     } else {
00692         memcpy(out, prev_buf,
00693                (mtab->size - tctx->last_block_pos[0]) * sizeof(*out));
00694 
00695         out +=  mtab->size - tctx->last_block_pos[0];
00696 
00697         memcpy(out, tctx->curr_frame,
00698                (tctx->last_block_pos[0]) * sizeof(*out));
00699     }
00700 
00701 }
00702 
00703 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
00704                          int ch, float *out, float gain, enum FrameType ftype)
00705 {
00706     const ModeTab *mtab = tctx->mtab;
00707     int i,j;
00708     float *hist = tctx->bark_hist[ftype][ch];
00709     float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
00710     int bark_n_coef  = mtab->fmode[ftype].bark_n_coef;
00711     int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
00712     int idx = 0;
00713 
00714     for (i = 0; i < fw_cb_len; i++)
00715         for (j = 0; j < bark_n_coef; j++, idx++) {
00716             float tmp2 =
00717                 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
00718             float st = use_hist ?
00719                 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
00720 
00721             hist[idx] = tmp2;
00722             if (st < -1.) st = 1.;
00723 
00724             memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
00725             out += mtab->fmode[ftype].bark_tab[idx];
00726         }
00727 
00728 }
00729 
00730 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
00731                                      float *out, enum FrameType ftype)
00732 {
00733     const ModeTab *mtab = tctx->mtab;
00734     int channels = tctx->avctx->channels;
00735     int sub = mtab->fmode[ftype].sub;
00736     int block_size = mtab->size / sub;
00737     float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
00738     float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
00739     uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
00740     uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
00741 
00742     uint8_t lpc_idx1[CHANNELS_MAX];
00743     uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
00744     uint8_t lpc_hist_idx[CHANNELS_MAX];
00745 
00746     int i, j, k;
00747 
00748     dequant(tctx, gb, out, ftype,
00749             mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
00750             mtab->fmode[ftype].cb_len_read);
00751 
00752     for (i = 0; i < channels; i++)
00753         for (j = 0; j < sub; j++)
00754             for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
00755                 bark1[i][j][k] =
00756                     get_bits(gb, mtab->fmode[ftype].bark_n_bit);
00757 
00758     for (i = 0; i < channels; i++)
00759         for (j = 0; j < sub; j++)
00760             bark_use_hist[i][j] = get_bits1(gb);
00761 
00762     dec_gain(tctx, gb, ftype, gain);
00763 
00764     for (i = 0; i < channels; i++) {
00765         lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
00766         lpc_idx1    [i] = get_bits(gb, tctx->mtab->lsp_bit1);
00767 
00768         for (j = 0; j < tctx->mtab->lsp_split; j++)
00769             lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
00770     }
00771 
00772     if (ftype == FT_LONG) {
00773         int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
00774             tctx->n_div[3];
00775         dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
00776                 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
00777     }
00778 
00779     for (i = 0; i < channels; i++) {
00780         float *chunk = out + mtab->size * i;
00781         float lsp[LSP_COEFS_MAX];
00782 
00783         for (j = 0; j < sub; j++) {
00784             dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
00785                          tctx->tmp_buf, gain[sub*i+j], ftype);
00786 
00787             tctx->dsp.vector_fmul(chunk + block_size*j, chunk + block_size*j, tctx->tmp_buf,
00788                                   block_size);
00789 
00790         }
00791 
00792         if (ftype == FT_LONG) {
00793             float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
00794             int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
00795             int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
00796             float v = 1./8192*
00797                 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
00798 
00799             decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
00800                        chunk);
00801         }
00802 
00803         decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
00804                    tctx->lsp_hist[i]);
00805 
00806         dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
00807 
00808         for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00809             tctx->dsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
00810             chunk += block_size;
00811         }
00812     }
00813 }
00814 
00815 static int twin_decode_frame(AVCodecContext * avctx, void *data,
00816                              int *data_size, AVPacket *avpkt)
00817 {
00818     const uint8_t *buf = avpkt->data;
00819     int buf_size = avpkt->size;
00820     TwinContext *tctx = avctx->priv_data;
00821     GetBitContext gb;
00822     const ModeTab *mtab = tctx->mtab;
00823     float *out = data;
00824     enum FrameType ftype;
00825     int window_type, out_size;
00826     static const enum FrameType wtype_to_ftype_table[] = {
00827         FT_LONG,   FT_LONG, FT_SHORT, FT_LONG,
00828         FT_MEDIUM, FT_LONG, FT_LONG,  FT_MEDIUM, FT_MEDIUM
00829     };
00830 
00831     if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
00832         av_log(avctx, AV_LOG_ERROR,
00833                "Frame too small (%d bytes). Truncated file?\n", buf_size);
00834         *data_size = 0;
00835         return buf_size;
00836     }
00837 
00838     out_size = mtab->size * avctx->channels *
00839                av_get_bytes_per_sample(avctx->sample_fmt);
00840     if (*data_size < out_size) {
00841         av_log(avctx, AV_LOG_ERROR, "output buffer is too small\n");
00842         return AVERROR(EINVAL);
00843     }
00844 
00845     init_get_bits(&gb, buf, buf_size * 8);
00846     skip_bits(&gb, get_bits(&gb, 8));
00847     window_type = get_bits(&gb, WINDOW_TYPE_BITS);
00848 
00849     if (window_type > 8) {
00850         av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
00851         return -1;
00852     }
00853 
00854     ftype = wtype_to_ftype_table[window_type];
00855 
00856     read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
00857 
00858     imdct_output(tctx, ftype, window_type, out);
00859 
00860     FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
00861 
00862     if (tctx->avctx->frame_number < 2) {
00863         *data_size=0;
00864         return buf_size;
00865     }
00866 
00867     *data_size = out_size;
00868 
00869     return buf_size;
00870 }
00871 
00875 static av_cold void init_mdct_win(TwinContext *tctx)
00876 {
00877     int i,j;
00878     const ModeTab *mtab = tctx->mtab;
00879     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00880     int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
00881     int channels = tctx->avctx->channels;
00882     float norm = channels == 1 ? 2. : 1.;
00883 
00884     for (i = 0; i < 3; i++) {
00885         int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
00886         ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
00887                      -sqrt(norm/bsize) / (1<<15));
00888     }
00889 
00890     tctx->tmp_buf  = av_malloc(mtab->size            * sizeof(*tctx->tmp_buf));
00891 
00892     tctx->spectrum  = av_malloc(2*mtab->size*channels*sizeof(float));
00893     tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float));
00894     tctx->prev_frame  = av_malloc(2*mtab->size*channels*sizeof(float));
00895 
00896     for (i = 0; i < 3; i++) {
00897         int m = 4*mtab->size/mtab->fmode[i].sub;
00898         double freq = 2*M_PI/m;
00899         tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs));
00900 
00901         for (j = 0; j <= m/8; j++)
00902             tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
00903         for (j = 1; j <  m/8; j++)
00904             tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
00905     }
00906 
00907 
00908     ff_init_ff_sine_windows(av_log2(size_m));
00909     ff_init_ff_sine_windows(av_log2(size_s/2));
00910     ff_init_ff_sine_windows(av_log2(mtab->size));
00911 }
00912 
00919 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
00920                               int block_size,
00921                               const uint8_t line_len[2], int length_div,
00922                               enum FrameType ftype)
00923 
00924 {
00925     int i,j;
00926 
00927     for (i = 0; i < line_len[0]; i++) {
00928         int shift;
00929 
00930         if (num_blocks == 1 ||
00931             (ftype == FT_LONG && num_vect % num_blocks) ||
00932             (ftype != FT_LONG && num_vect & 1         ) ||
00933             i == line_len[1]) {
00934             shift = 0;
00935         } else if (ftype == FT_LONG) {
00936             shift = i;
00937         } else
00938             shift = i*i;
00939 
00940         for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
00941             tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
00942     }
00943 }
00944 
00960 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
00961                            const uint8_t line_len[2], int length_div)
00962 {
00963     int i,j;
00964     int cont= 0;
00965     for (i = 0; i < num_vect; i++)
00966         for (j = 0; j < line_len[i >= length_div]; j++)
00967             out[cont++] = in[j*num_vect + i];
00968 }
00969 
00970 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
00971 {
00972     int block_size = size/n_blocks;
00973     int i;
00974 
00975     for (i = 0; i < size; i++)
00976         out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
00977 }
00978 
00979 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
00980 {
00981     int block_size;
00982     const ModeTab *mtab = tctx->mtab;
00983     int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
00984     int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
00985 
00986     if (ftype == FT_PPC) {
00987         size  = tctx->avctx->channels;
00988         block_size = mtab->ppc_shape_len;
00989     } else
00990         block_size = mtab->size / mtab->fmode[ftype].sub;
00991 
00992     permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
00993                       block_size, tctx->length[ftype],
00994                       tctx->length_change[ftype], ftype);
00995 
00996     transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
00997                    tctx->length[ftype], tctx->length_change[ftype]);
00998 
00999     linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
01000                 size*block_size);
01001 }
01002 
01003 static av_cold void init_bitstream_params(TwinContext *tctx)
01004 {
01005     const ModeTab *mtab = tctx->mtab;
01006     int n_ch = tctx->avctx->channels;
01007     int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
01008                              tctx->avctx->sample_rate;
01009 
01010     int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
01011                                    mtab->lsp_split*mtab->lsp_bit2);
01012 
01013     int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
01014                          mtab->ppc_period_bit);
01015 
01016     int bsize_no_main_cb[3];
01017     int bse_bits[3];
01018     int i;
01019     enum FrameType frametype;
01020 
01021     for (i = 0; i < 3; i++)
01022         // +1 for history usage switch
01023         bse_bits[i] = n_ch *
01024             (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
01025 
01026     bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
01027                           WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
01028 
01029     for (i = 0; i < 2; i++)
01030         bsize_no_main_cb[i] =
01031             lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
01032             mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
01033 
01034     // The remaining bits are all used for the main spectrum coefficients
01035     for (i = 0; i < 4; i++) {
01036         int bit_size;
01037         int vect_size;
01038         int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
01039         if (i == 3) {
01040             bit_size  = n_ch * mtab->ppc_shape_bit;
01041             vect_size = n_ch * mtab->ppc_shape_len;
01042         } else {
01043             bit_size = total_fr_bits - bsize_no_main_cb[i];
01044             vect_size = n_ch * mtab->size;
01045         }
01046 
01047         tctx->n_div[i] = (bit_size + 13) / 14;
01048 
01049         rounded_up   = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01050         rounded_down = (bit_size           )/tctx->n_div[i];
01051         num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
01052         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01053         tctx->bits_main_spec[0][i][0] = (rounded_up   + 1)/2;
01054         tctx->bits_main_spec[1][i][0] = (rounded_up      )/2;
01055         tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
01056         tctx->bits_main_spec[1][i][1] = (rounded_down    )/2;
01057         tctx->bits_main_spec_change[i] = num_rounded_up;
01058 
01059         rounded_up   = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01060         rounded_down = (vect_size                     )/tctx->n_div[i];
01061         num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
01062         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01063         tctx->length[i][0] = rounded_up;
01064         tctx->length[i][1] = rounded_down;
01065         tctx->length_change[i] = num_rounded_up;
01066     }
01067 
01068     for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
01069         construct_perm_table(tctx, frametype);
01070 }
01071 
01072 static av_cold int twin_decode_init(AVCodecContext *avctx)
01073 {
01074     TwinContext *tctx = avctx->priv_data;
01075     int isampf = avctx->sample_rate/1000;
01076     int ibps = avctx->bit_rate/(1000 * avctx->channels);
01077 
01078     tctx->avctx       = avctx;
01079     avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
01080 
01081     if (avctx->channels > CHANNELS_MAX) {
01082         av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
01083                avctx->channels);
01084         return -1;
01085     }
01086 
01087     switch ((isampf << 8) +  ibps) {
01088     case (8 <<8) +  8: tctx->mtab = &mode_08_08; break;
01089     case (11<<8) +  8: tctx->mtab = &mode_11_08; break;
01090     case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
01091     case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
01092     case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
01093     case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
01094     case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
01095     case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
01096     case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
01097     default:
01098         av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
01099         return -1;
01100     }
01101 
01102     dsputil_init(&tctx->dsp, avctx);
01103     init_mdct_win(tctx);
01104     init_bitstream_params(tctx);
01105 
01106     memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
01107 
01108     return 0;
01109 }
01110 
01111 static av_cold int twin_decode_close(AVCodecContext *avctx)
01112 {
01113     TwinContext *tctx = avctx->priv_data;
01114     int i;
01115 
01116     for (i = 0; i < 3; i++) {
01117         ff_mdct_end(&tctx->mdct_ctx[i]);
01118         av_free(tctx->cos_tabs[i]);
01119     }
01120 
01121 
01122     av_free(tctx->curr_frame);
01123     av_free(tctx->spectrum);
01124     av_free(tctx->prev_frame);
01125     av_free(tctx->tmp_buf);
01126 
01127     return 0;
01128 }
01129 
01130 AVCodec ff_twinvq_decoder =
01131 {
01132     "twinvq",
01133     AVMEDIA_TYPE_AUDIO,
01134     CODEC_ID_TWINVQ,
01135     sizeof(TwinContext),
01136     twin_decode_init,
01137     NULL,
01138     twin_decode_close,
01139     twin_decode_frame,
01140     .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
01141 };

Generated on Fri Feb 22 2013 07:24:29 for FFmpeg by  doxygen 1.7.1