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 "internal.h"
00024 #include "get_bits.h"
00025 #include "dsputil.h"
00026 #include "fft.h"
00027 #include "lsp.h"
00028 #include "sinewin.h"
00029 
00030 #include <math.h>
00031 #include <stdint.h>
00032 
00033 #include "twinvq_data.h"
00034 
00035 enum FrameType {
00036     FT_SHORT = 0,  
00037     FT_MEDIUM,     
00038     FT_LONG,       
00039     FT_PPC,        
00040 };
00041 
00045 struct FrameMode {
00046     uint8_t         sub;      
00047     const uint16_t *bark_tab;
00048 
00050     uint8_t         bark_env_size;
00051 
00052     const int16_t  *bark_cb;    
00053     uint8_t         bark_n_coef;
00054     uint8_t         bark_n_bit; 
00055 
00057 
00058     const int16_t    *cb0;
00059     const int16_t    *cb1;
00061 
00062     uint8_t         cb_len_read; 
00063 };
00064 
00069 typedef struct {
00070     struct FrameMode fmode[3]; 
00071 
00072     uint16_t     size;        
00073     uint8_t      n_lsp;       
00074     const float *lspcodebook;
00075 
00076     /* number of bits of the different LSP CB coefficients */
00077     uint8_t      lsp_bit0;
00078     uint8_t      lsp_bit1;
00079     uint8_t      lsp_bit2;
00080 
00081     uint8_t      lsp_split;      
00082     const int16_t *ppc_shape_cb; 
00083 
00085     uint8_t      ppc_period_bit;
00086 
00087     uint8_t      ppc_shape_bit;  
00088     uint8_t      ppc_shape_len;  
00089     uint8_t      pgain_bit;      
00090 
00092     uint16_t     peak_per2wid;
00093 } ModeTab;
00094 
00095 static const ModeTab mode_08_08 = {
00096     {
00097         { 8, bark_tab_s08_64,  10, tab.fcb08s  , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
00098         { 2, bark_tab_m08_256, 20, tab.fcb08m  , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
00099         { 1, bark_tab_l08_512, 30, tab.fcb08l  , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
00100     },
00101     512 , 12, tab.lsp08,   1, 5, 3, 3, tab.shape08  , 8, 28, 20, 6, 40
00102 };
00103 
00104 static const ModeTab mode_11_08 = {
00105     {
00106         { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
00107         { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
00108         { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
00109     },
00110     512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
00111 };
00112 
00113 static const ModeTab mode_11_10 = {
00114     {
00115         { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
00116         { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
00117         { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
00118     },
00119     512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
00120 };
00121 
00122 static const ModeTab mode_16_16 = {
00123     {
00124         { 8, bark_tab_s16_128, 10, tab.fcb16s  , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
00125         { 2, bark_tab_m16_512, 20, tab.fcb16m  , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
00126         { 1, bark_tab_l16_1024,30, tab.fcb16l  , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
00127     },
00128     1024, 16, tab.lsp16,   1, 6, 4, 3, tab.shape16  , 9, 56, 60, 7, 180
00129 };
00130 
00131 static const ModeTab mode_22_20 = {
00132     {
00133         { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
00134         { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
00135         { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
00136     },
00137     1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
00138 };
00139 
00140 static const ModeTab mode_22_24 = {
00141     {
00142         { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
00143         { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
00144         { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
00145     },
00146     1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
00147 };
00148 
00149 static const ModeTab mode_22_32 = {
00150     {
00151         { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
00152         { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
00153         { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
00154     },
00155     512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
00156 };
00157 
00158 static const ModeTab mode_44_40 = {
00159     {
00160         {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
00161         { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
00162         { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
00163     },
00164     2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
00165 };
00166 
00167 static const ModeTab mode_44_48 = {
00168     {
00169         {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
00170         { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
00171         { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
00172     },
00173     2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
00174 };
00175 
00176 typedef struct TwinContext {
00177     AVCodecContext *avctx;
00178     AVFrame frame;
00179     DSPContext      dsp;
00180     FFTContext mdct_ctx[3];
00181 
00182     const ModeTab *mtab;
00183 
00184     // history
00185     float lsp_hist[2][20];           
00186     float bark_hist[3][2][40];       
00187 
00188     // bitstream parameters
00189     int16_t permut[4][4096];
00190     uint8_t length[4][2];            
00191     uint8_t length_change[4];
00192     uint8_t bits_main_spec[2][4][2]; 
00193     int bits_main_spec_change[4];
00194     int n_div[4];
00195 
00196     float *spectrum;
00197     float *curr_frame;               
00198     float *prev_frame;               
00199     int last_block_pos[2];
00200     int discarded_packets;
00201 
00202     float *cos_tabs[3];
00203 
00204     // scratch buffers
00205     float *tmp_buf;
00206 } TwinContext;
00207 
00208 #define PPC_SHAPE_CB_SIZE 64
00209 #define PPC_SHAPE_LEN_MAX 60
00210 #define SUB_AMP_MAX       4500.0
00211 #define MULAW_MU          100.0
00212 #define GAIN_BITS         8
00213 #define AMP_MAX           13000.0
00214 #define SUB_GAIN_BITS     5
00215 #define WINDOW_TYPE_BITS  4
00216 #define PGAIN_MU          200
00217 #define LSP_COEFS_MAX     20
00218 #define LSP_SPLIT_MAX     4
00219 #define CHANNELS_MAX      2
00220 #define SUBBLOCKS_MAX     16
00221 #define BARK_N_COEF_MAX   4
00222 
00224 static void memset_float(float *buf, float val, int size)
00225 {
00226     while (size--)
00227         *buf++ = val;
00228 }
00229 
00242 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
00243 {
00244     int j;
00245     float p = 0.5f;
00246     float q = 0.5f;
00247     float two_cos_w = 2.0f*cos_val;
00248 
00249     for (j = 0; j + 1 < order; j += 2*2) {
00250         // Unroll the loop once since order is a multiple of four
00251         q *= lsp[j  ] - two_cos_w;
00252         p *= lsp[j+1] - two_cos_w;
00253 
00254         q *= lsp[j+2] - two_cos_w;
00255         p *= lsp[j+3] - two_cos_w;
00256     }
00257 
00258     p *= p * (2.0f - two_cos_w);
00259     q *= q * (2.0f + two_cos_w);
00260 
00261     return 0.5 / (p + q);
00262 }
00263 
00267 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
00268 {
00269     int i;
00270     const ModeTab *mtab = tctx->mtab;
00271     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00272 
00273     for (i = 0; i < size_s/2; i++) {
00274         float cos_i = tctx->cos_tabs[0][i];
00275         lpc[i]          = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
00276         lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
00277     }
00278 }
00279 
00280 static void interpolate(float *out, float v1, float v2, int size)
00281 {
00282     int i;
00283     float step = (v1 - v2)/(size + 1);
00284 
00285     for (i = 0; i < size; i++) {
00286         v2 += step;
00287         out[i] = v2;
00288     }
00289 }
00290 
00291 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
00292 {
00293     return part ? -cos_tab[size - idx - 1] :
00294                    cos_tab[       idx    ];
00295 }
00296 
00311 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
00312                                          enum FrameType ftype,
00313                                          float *out, const float *in,
00314                                          int size, int step, int part)
00315 {
00316     int i;
00317     const ModeTab *mtab = tctx->mtab;
00318     const float *cos_tab = tctx->cos_tabs[ftype];
00319 
00320     // Fill the 's'
00321     for (i = 0; i < size; i += step)
00322         out[i] =
00323             eval_lpc_spectrum(in,
00324                               get_cos(i, part, cos_tab, size),
00325                               mtab->n_lsp);
00326 
00327     // Fill the 'iiiibiiii'
00328     for (i = step; i <= size - 2*step; i += step) {
00329         if (out[i + step] + out[i - step] >  1.95*out[i] ||
00330             out[i + step]                 >=  out[i - step]) {
00331             interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
00332         } else {
00333             out[i - step/2] =
00334                 eval_lpc_spectrum(in,
00335                                   get_cos(i-step/2, part, cos_tab, size),
00336                                   mtab->n_lsp);
00337             interpolate(out + i - step   + 1, out[i-step/2], out[i-step  ], step/2 - 1);
00338             interpolate(out + i - step/2 + 1, out[i       ], out[i-step/2], step/2 - 1);
00339         }
00340     }
00341 
00342     interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
00343 }
00344 
00345 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
00346                                const float *buf, float *lpc,
00347                                int size, int step)
00348 {
00349     eval_lpcenv_or_interp(tctx, ftype, lpc         , buf, size/2,   step, 0);
00350     eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
00351 
00352     interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
00353 
00354     memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
00355 }
00356 
00362 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
00363                     enum FrameType ftype,
00364                     const int16_t *cb0, const int16_t *cb1, int cb_len)
00365 {
00366     int pos = 0;
00367     int i, j;
00368 
00369     for (i = 0; i < tctx->n_div[ftype]; i++) {
00370         int tmp0, tmp1;
00371         int sign0 = 1;
00372         int sign1 = 1;
00373         const int16_t *tab0, *tab1;
00374         int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
00375         int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
00376 
00377         int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
00378         if (bits == 7) {
00379             if (get_bits1(gb))
00380                 sign0 = -1;
00381             bits = 6;
00382         }
00383         tmp0 = get_bits(gb, bits);
00384 
00385         bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
00386 
00387         if (bits == 7) {
00388             if (get_bits1(gb))
00389                 sign1 = -1;
00390 
00391             bits = 6;
00392         }
00393         tmp1 = get_bits(gb, bits);
00394 
00395         tab0 = cb0 + tmp0*cb_len;
00396         tab1 = cb1 + tmp1*cb_len;
00397 
00398         for (j = 0; j < length; j++)
00399             out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
00400 
00401         pos += length;
00402     }
00403 
00404 }
00405 
00406 static inline float mulawinv(float y, float clip, float mu)
00407 {
00408     y = av_clipf(y/clip, -1, 1);
00409     return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
00410 }
00411 
00432 static int very_broken_op(int a, int b)
00433 {
00434     int x = a*b + 200;
00435     int size;
00436     const uint8_t *rtab;
00437 
00438     if (x%400 || b%5)
00439         return x/400;
00440 
00441     x /= 400;
00442 
00443     size = tabs[b/5].size;
00444     rtab = tabs[b/5].tab;
00445     return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
00446 }
00447 
00453 static void add_peak(int period, int width, const float *shape,
00454                      float ppc_gain, float *speech, int len)
00455 {
00456     int i, j;
00457 
00458     const float *shape_end = shape + len;
00459     int center;
00460 
00461     // First peak centered around zero
00462     for (i = 0; i < width/2; i++)
00463         speech[i] += ppc_gain * *shape++;
00464 
00465     for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
00466         center = very_broken_op(period, i);
00467         for (j = -width/2; j < (width+1)/2; j++)
00468             speech[j+center] += ppc_gain * *shape++;
00469     }
00470 
00471     // For the last block, be careful not to go beyond the end of the buffer
00472     center = very_broken_op(period, i);
00473     for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
00474         speech[j+center] += ppc_gain * *shape++;
00475 }
00476 
00477 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
00478                        float ppc_gain, float *speech)
00479 {
00480     const ModeTab *mtab = tctx->mtab;
00481     int isampf = tctx->avctx->sample_rate/1000;
00482     int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
00483     int min_period = ROUNDED_DIV(  40*2*mtab->size, isampf);
00484     int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
00485     int period_range = max_period - min_period;
00486 
00487     // This is actually the period multiplied by 400. It is just linearly coded
00488     // between its maximum and minimum value.
00489     int period = min_period +
00490         ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
00491     int width;
00492 
00493     if (isampf == 22 && ibps == 32) {
00494         // For some unknown reason, NTT decided to code this case differently...
00495         width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
00496     } else
00497         width =             (period      )* mtab->peak_per2wid/(400*mtab->size);
00498 
00499     add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
00500 }
00501 
00502 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
00503                      float *out)
00504 {
00505     const ModeTab *mtab = tctx->mtab;
00506     int i, j;
00507     int sub = mtab->fmode[ftype].sub;
00508     float step     = AMP_MAX     / ((1 <<     GAIN_BITS) - 1);
00509     float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
00510 
00511     if (ftype == FT_LONG) {
00512         for (i = 0; i < tctx->avctx->channels; i++)
00513             out[i] = (1./(1<<13)) *
00514                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00515                          AMP_MAX, MULAW_MU);
00516     } else {
00517         for (i = 0; i < tctx->avctx->channels; i++) {
00518             float val = (1./(1<<23)) *
00519                 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
00520                          AMP_MAX, MULAW_MU);
00521 
00522             for (j = 0; j < sub; j++) {
00523                 out[i*sub + j] =
00524                     val*mulawinv(sub_step* 0.5 +
00525                                  sub_step* get_bits(gb, SUB_GAIN_BITS),
00526                                  SUB_AMP_MAX, MULAW_MU);
00527             }
00528         }
00529     }
00530 }
00531 
00538 static void rearrange_lsp(int order, float *lsp, float min_dist)
00539 {
00540     int i;
00541     float min_dist2 = min_dist * 0.5;
00542     for (i = 1; i < order; i++)
00543         if (lsp[i] - lsp[i-1] < min_dist) {
00544             float avg = (lsp[i] + lsp[i-1]) * 0.5;
00545 
00546             lsp[i-1] = avg - min_dist2;
00547             lsp[i  ] = avg + min_dist2;
00548         }
00549 }
00550 
00551 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
00552                        int lpc_hist_idx, float *lsp, float *hist)
00553 {
00554     const ModeTab *mtab = tctx->mtab;
00555     int i, j;
00556 
00557     const float *cb  =  mtab->lspcodebook;
00558     const float *cb2 =  cb  + (1 << mtab->lsp_bit1)*mtab->n_lsp;
00559     const float *cb3 =  cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
00560 
00561     const int8_t funny_rounding[4] = {
00562         -2,
00563         mtab->lsp_split == 4 ? -2 : 1,
00564         mtab->lsp_split == 4 ? -2 : 1,
00565         0
00566     };
00567 
00568     j = 0;
00569     for (i = 0; i < mtab->lsp_split; i++) {
00570         int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
00571         for (; j < chunk_end; j++)
00572             lsp[j] = cb [lpc_idx1    * mtab->n_lsp + j] +
00573                      cb2[lpc_idx2[i] * mtab->n_lsp + j];
00574     }
00575 
00576     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00577 
00578     for (i = 0; i < mtab->n_lsp; i++) {
00579         float tmp1 = 1. -          cb3[lpc_hist_idx*mtab->n_lsp + i];
00580         float tmp2 =     hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
00581         hist[i] = lsp[i];
00582         lsp[i]  = lsp[i] * tmp1 + tmp2;
00583     }
00584 
00585     rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
00586     rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
00587     ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
00588 }
00589 
00590 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
00591                                  enum FrameType ftype, float *lpc)
00592 {
00593     int i;
00594     int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
00595 
00596     for (i = 0; i < tctx->mtab->n_lsp; i++)
00597         lsp[i] =  2*cos(lsp[i]);
00598 
00599     switch (ftype) {
00600     case FT_LONG:
00601         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
00602         break;
00603     case FT_MEDIUM:
00604         eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
00605         break;
00606     case FT_SHORT:
00607         eval_lpcenv(tctx, lsp, lpc);
00608         break;
00609     }
00610 }
00611 
00612 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
00613                             float *in, float *prev, int ch)
00614 {
00615     FFTContext *mdct = &tctx->mdct_ctx[ftype];
00616     const ModeTab *mtab = tctx->mtab;
00617     int bsize = mtab->size / mtab->fmode[ftype].sub;
00618     int size  = mtab->size;
00619     float *buf1 = tctx->tmp_buf;
00620     int j;
00621     int wsize; // Window size
00622     float *out = tctx->curr_frame + 2*ch*mtab->size;
00623     float *out2 = out;
00624     float *prev_buf;
00625     int first_wsize;
00626 
00627     static const uint8_t wtype_to_wsize[]      = {0, 0, 2, 2, 2, 1, 0, 1, 1};
00628     int types_sizes[] = {
00629         mtab->size /    mtab->fmode[FT_LONG  ].sub,
00630         mtab->size /    mtab->fmode[FT_MEDIUM].sub,
00631         mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
00632     };
00633 
00634     wsize = types_sizes[wtype_to_wsize[wtype]];
00635     first_wsize = wsize;
00636     prev_buf = prev + (size - bsize)/2;
00637 
00638     for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00639         int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
00640 
00641         if (!j && wtype == 4)
00642             sub_wtype = 4;
00643         else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
00644             sub_wtype = 7;
00645 
00646         wsize = types_sizes[wtype_to_wsize[sub_wtype]];
00647 
00648         mdct->imdct_half(mdct, buf1 + bsize*j, in + bsize*j);
00649 
00650         tctx->dsp.vector_fmul_window(out2,
00651                                      prev_buf + (bsize-wsize)/2,
00652                                      buf1 + bsize*j,
00653                                      ff_sine_windows[av_log2(wsize)],
00654                                      wsize/2);
00655         out2 += wsize;
00656 
00657         memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
00658 
00659         out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
00660 
00661         prev_buf = buf1 + bsize*j + bsize/2;
00662     }
00663 
00664     tctx->last_block_pos[ch] = (size + first_wsize)/2;
00665 }
00666 
00667 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
00668                          float *out)
00669 {
00670     const ModeTab *mtab = tctx->mtab;
00671     int size1, size2;
00672     float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
00673     int i;
00674 
00675     for (i = 0; i < tctx->avctx->channels; i++) {
00676         imdct_and_window(tctx, ftype, wtype,
00677                          tctx->spectrum + i*mtab->size,
00678                          prev_buf + 2*i*mtab->size,
00679                          i);
00680     }
00681 
00682     if (!out)
00683         return;
00684 
00685     size2 = tctx->last_block_pos[0];
00686     size1 = mtab->size - size2;
00687     if (tctx->avctx->channels == 2) {
00688         tctx->dsp.butterflies_float_interleave(out, prev_buf,
00689                                                &prev_buf[2*mtab->size],
00690                                                size1);
00691 
00692         out += 2 * size1;
00693 
00694         tctx->dsp.butterflies_float_interleave(out, tctx->curr_frame,
00695                                                &tctx->curr_frame[2*mtab->size],
00696                                                size2);
00697     } else {
00698         memcpy(out, prev_buf, size1 * sizeof(*out));
00699 
00700         out += size1;
00701 
00702         memcpy(out, tctx->curr_frame, size2 * sizeof(*out));
00703     }
00704 
00705 }
00706 
00707 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
00708                          int ch, float *out, float gain, enum FrameType ftype)
00709 {
00710     const ModeTab *mtab = tctx->mtab;
00711     int i,j;
00712     float *hist = tctx->bark_hist[ftype][ch];
00713     float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
00714     int bark_n_coef  = mtab->fmode[ftype].bark_n_coef;
00715     int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
00716     int idx = 0;
00717 
00718     for (i = 0; i < fw_cb_len; i++)
00719         for (j = 0; j < bark_n_coef; j++, idx++) {
00720             float tmp2 =
00721                 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
00722             float st = use_hist ?
00723                 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
00724 
00725             hist[idx] = tmp2;
00726             if (st < -1.) st = 1.;
00727 
00728             memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
00729             out += mtab->fmode[ftype].bark_tab[idx];
00730         }
00731 
00732 }
00733 
00734 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
00735                                      float *out, enum FrameType ftype)
00736 {
00737     const ModeTab *mtab = tctx->mtab;
00738     int channels = tctx->avctx->channels;
00739     int sub = mtab->fmode[ftype].sub;
00740     int block_size = mtab->size / sub;
00741     float gain[CHANNELS_MAX*SUBBLOCKS_MAX];
00742     float ppc_shape[PPC_SHAPE_LEN_MAX * CHANNELS_MAX * 4];
00743     uint8_t bark1[CHANNELS_MAX][SUBBLOCKS_MAX][BARK_N_COEF_MAX];
00744     uint8_t bark_use_hist[CHANNELS_MAX][SUBBLOCKS_MAX];
00745 
00746     uint8_t lpc_idx1[CHANNELS_MAX];
00747     uint8_t lpc_idx2[CHANNELS_MAX][LSP_SPLIT_MAX];
00748     uint8_t lpc_hist_idx[CHANNELS_MAX];
00749 
00750     int i, j, k;
00751 
00752     dequant(tctx, gb, out, ftype,
00753             mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
00754             mtab->fmode[ftype].cb_len_read);
00755 
00756     for (i = 0; i < channels; i++)
00757         for (j = 0; j < sub; j++)
00758             for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
00759                 bark1[i][j][k] =
00760                     get_bits(gb, mtab->fmode[ftype].bark_n_bit);
00761 
00762     for (i = 0; i < channels; i++)
00763         for (j = 0; j < sub; j++)
00764             bark_use_hist[i][j] = get_bits1(gb);
00765 
00766     dec_gain(tctx, gb, ftype, gain);
00767 
00768     for (i = 0; i < channels; i++) {
00769         lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
00770         lpc_idx1    [i] = get_bits(gb, tctx->mtab->lsp_bit1);
00771 
00772         for (j = 0; j < tctx->mtab->lsp_split; j++)
00773             lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
00774     }
00775 
00776     if (ftype == FT_LONG) {
00777         int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
00778             tctx->n_div[3];
00779         dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
00780                 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
00781     }
00782 
00783     for (i = 0; i < channels; i++) {
00784         float *chunk = out + mtab->size * i;
00785         float lsp[LSP_COEFS_MAX];
00786 
00787         for (j = 0; j < sub; j++) {
00788             dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
00789                          tctx->tmp_buf, gain[sub*i+j], ftype);
00790 
00791             tctx->dsp.vector_fmul(chunk + block_size*j, chunk + block_size*j, tctx->tmp_buf,
00792                                   block_size);
00793 
00794         }
00795 
00796         if (ftype == FT_LONG) {
00797             float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
00798             int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
00799             int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
00800             float v = 1./8192*
00801                 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
00802 
00803             decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
00804                        chunk);
00805         }
00806 
00807         decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
00808                    tctx->lsp_hist[i]);
00809 
00810         dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
00811 
00812         for (j = 0; j < mtab->fmode[ftype].sub; j++) {
00813             tctx->dsp.vector_fmul(chunk, chunk, tctx->tmp_buf, block_size);
00814             chunk += block_size;
00815         }
00816     }
00817 }
00818 
00819 static int twin_decode_frame(AVCodecContext * avctx, void *data,
00820                              int *got_frame_ptr, AVPacket *avpkt)
00821 {
00822     const uint8_t *buf = avpkt->data;
00823     int buf_size = avpkt->size;
00824     TwinContext *tctx = avctx->priv_data;
00825     GetBitContext gb;
00826     const ModeTab *mtab = tctx->mtab;
00827     float *out = NULL;
00828     enum FrameType ftype;
00829     int window_type, ret;
00830     static const enum FrameType wtype_to_ftype_table[] = {
00831         FT_LONG,   FT_LONG, FT_SHORT, FT_LONG,
00832         FT_MEDIUM, FT_LONG, FT_LONG,  FT_MEDIUM, FT_MEDIUM
00833     };
00834 
00835     if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
00836         av_log(avctx, AV_LOG_ERROR,
00837                "Frame too small (%d bytes). Truncated file?\n", buf_size);
00838         return AVERROR(EINVAL);
00839     }
00840 
00841     /* get output buffer */
00842     if (tctx->discarded_packets >= 2) {
00843         tctx->frame.nb_samples = mtab->size;
00844         if ((ret = ff_get_buffer(avctx, &tctx->frame)) < 0) {
00845             av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
00846             return ret;
00847         }
00848         out = (float *)tctx->frame.data[0];
00849     }
00850 
00851     init_get_bits(&gb, buf, buf_size * 8);
00852     skip_bits(&gb, get_bits(&gb, 8));
00853     window_type = get_bits(&gb, WINDOW_TYPE_BITS);
00854 
00855     if (window_type > 8) {
00856         av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
00857         return -1;
00858     }
00859 
00860     ftype = wtype_to_ftype_table[window_type];
00861 
00862     read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
00863 
00864     imdct_output(tctx, ftype, window_type, out);
00865 
00866     FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
00867 
00868     if (tctx->discarded_packets < 2) {
00869         tctx->discarded_packets++;
00870         *got_frame_ptr = 0;
00871         return buf_size;
00872     }
00873 
00874     *got_frame_ptr   = 1;
00875     *(AVFrame *)data = tctx->frame;;
00876 
00877     return buf_size;
00878 }
00879 
00883 static av_cold int init_mdct_win(TwinContext *tctx)
00884 {
00885     int i, j, ret;
00886     const ModeTab *mtab = tctx->mtab;
00887     int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
00888     int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
00889     int channels = tctx->avctx->channels;
00890     float norm = channels == 1 ? 2. : 1.;
00891 
00892     for (i = 0; i < 3; i++) {
00893         int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
00894         if ((ret = ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
00895                                 -sqrt(norm/bsize) / (1<<15))))
00896             return ret;
00897     }
00898 
00899     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->tmp_buf,
00900                      mtab->size * sizeof(*tctx->tmp_buf), alloc_fail);
00901 
00902     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->spectrum,
00903                      2 * mtab->size * channels * sizeof(*tctx->spectrum),
00904                      alloc_fail);
00905     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->curr_frame,
00906                      2 * mtab->size * channels * sizeof(*tctx->curr_frame),
00907                      alloc_fail);
00908     FF_ALLOC_OR_GOTO(tctx->avctx, tctx->prev_frame,
00909                      2 * mtab->size * channels * sizeof(*tctx->prev_frame),
00910                      alloc_fail);
00911 
00912     for (i = 0; i < 3; i++) {
00913         int m = 4*mtab->size/mtab->fmode[i].sub;
00914         double freq = 2*M_PI/m;
00915         FF_ALLOC_OR_GOTO(tctx->avctx, tctx->cos_tabs[i],
00916                          (m / 4) * sizeof(*tctx->cos_tabs[i]), alloc_fail);
00917 
00918         for (j = 0; j <= m/8; j++)
00919             tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
00920         for (j = 1; j <  m/8; j++)
00921             tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
00922     }
00923 
00924 
00925     ff_init_ff_sine_windows(av_log2(size_m));
00926     ff_init_ff_sine_windows(av_log2(size_s/2));
00927     ff_init_ff_sine_windows(av_log2(mtab->size));
00928 
00929     return 0;
00930 alloc_fail:
00931     return AVERROR(ENOMEM);
00932 }
00933 
00940 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
00941                               int block_size,
00942                               const uint8_t line_len[2], int length_div,
00943                               enum FrameType ftype)
00944 
00945 {
00946     int i,j;
00947 
00948     for (i = 0; i < line_len[0]; i++) {
00949         int shift;
00950 
00951         if (num_blocks == 1 ||
00952             (ftype == FT_LONG && num_vect % num_blocks) ||
00953             (ftype != FT_LONG && num_vect & 1         ) ||
00954             i == line_len[1]) {
00955             shift = 0;
00956         } else if (ftype == FT_LONG) {
00957             shift = i;
00958         } else
00959             shift = i*i;
00960 
00961         for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
00962             tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
00963     }
00964 }
00965 
00981 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
00982                            const uint8_t line_len[2], int length_div)
00983 {
00984     int i,j;
00985     int cont= 0;
00986     for (i = 0; i < num_vect; i++)
00987         for (j = 0; j < line_len[i >= length_div]; j++)
00988             out[cont++] = in[j*num_vect + i];
00989 }
00990 
00991 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
00992 {
00993     int block_size = size/n_blocks;
00994     int i;
00995 
00996     for (i = 0; i < size; i++)
00997         out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
00998 }
00999 
01000 static av_cold void construct_perm_table(TwinContext *tctx, int ftype)
01001 {
01002     int block_size;
01003     const ModeTab *mtab = tctx->mtab;
01004     int size;
01005     int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
01006 
01007     if (ftype == FT_PPC) {
01008         size  = tctx->avctx->channels;
01009         block_size = mtab->ppc_shape_len;
01010     } else {
01011         size       = tctx->avctx->channels * mtab->fmode[ftype].sub;
01012         block_size = mtab->size / mtab->fmode[ftype].sub;
01013     }
01014 
01015     permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
01016                       block_size, tctx->length[ftype],
01017                       tctx->length_change[ftype], ftype);
01018 
01019     transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
01020                    tctx->length[ftype], tctx->length_change[ftype]);
01021 
01022     linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
01023                 size*block_size);
01024 }
01025 
01026 static av_cold void init_bitstream_params(TwinContext *tctx)
01027 {
01028     const ModeTab *mtab = tctx->mtab;
01029     int n_ch = tctx->avctx->channels;
01030     int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
01031                              tctx->avctx->sample_rate;
01032 
01033     int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
01034                                    mtab->lsp_split*mtab->lsp_bit2);
01035 
01036     int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
01037                          mtab->ppc_period_bit);
01038 
01039     int bsize_no_main_cb[3];
01040     int bse_bits[3];
01041     int i;
01042     enum FrameType frametype;
01043 
01044     for (i = 0; i < 3; i++)
01045         // +1 for history usage switch
01046         bse_bits[i] = n_ch *
01047             (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
01048 
01049     bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
01050                           WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
01051 
01052     for (i = 0; i < 2; i++)
01053         bsize_no_main_cb[i] =
01054             lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
01055             mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
01056 
01057     // The remaining bits are all used for the main spectrum coefficients
01058     for (i = 0; i < 4; i++) {
01059         int bit_size;
01060         int vect_size;
01061         int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
01062         if (i == 3) {
01063             bit_size  = n_ch * mtab->ppc_shape_bit;
01064             vect_size = n_ch * mtab->ppc_shape_len;
01065         } else {
01066             bit_size = total_fr_bits - bsize_no_main_cb[i];
01067             vect_size = n_ch * mtab->size;
01068         }
01069 
01070         tctx->n_div[i] = (bit_size + 13) / 14;
01071 
01072         rounded_up   = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01073         rounded_down = (bit_size           )/tctx->n_div[i];
01074         num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
01075         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01076         tctx->bits_main_spec[0][i][0] = (rounded_up   + 1)/2;
01077         tctx->bits_main_spec[1][i][0] = (rounded_up      )/2;
01078         tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
01079         tctx->bits_main_spec[1][i][1] = (rounded_down    )/2;
01080         tctx->bits_main_spec_change[i] = num_rounded_up;
01081 
01082         rounded_up   = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
01083         rounded_down = (vect_size                     )/tctx->n_div[i];
01084         num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
01085         num_rounded_up = tctx->n_div[i] - num_rounded_down;
01086         tctx->length[i][0] = rounded_up;
01087         tctx->length[i][1] = rounded_down;
01088         tctx->length_change[i] = num_rounded_up;
01089     }
01090 
01091     for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
01092         construct_perm_table(tctx, frametype);
01093 }
01094 
01095 static av_cold int twin_decode_close(AVCodecContext *avctx)
01096 {
01097     TwinContext *tctx = avctx->priv_data;
01098     int i;
01099 
01100     for (i = 0; i < 3; i++) {
01101         ff_mdct_end(&tctx->mdct_ctx[i]);
01102         av_free(tctx->cos_tabs[i]);
01103     }
01104 
01105 
01106     av_free(tctx->curr_frame);
01107     av_free(tctx->spectrum);
01108     av_free(tctx->prev_frame);
01109     av_free(tctx->tmp_buf);
01110 
01111     return 0;
01112 }
01113 
01114 static av_cold int twin_decode_init(AVCodecContext *avctx)
01115 {
01116     int ret;
01117     TwinContext *tctx = avctx->priv_data;
01118     int isampf, ibps;
01119 
01120     tctx->avctx       = avctx;
01121     avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
01122 
01123     if (!avctx->extradata || avctx->extradata_size < 12) {
01124         av_log(avctx, AV_LOG_ERROR, "Missing or incomplete extradata\n");
01125         return AVERROR_INVALIDDATA;
01126     }
01127     avctx->channels = AV_RB32(avctx->extradata    ) + 1;
01128     avctx->bit_rate = AV_RB32(avctx->extradata + 4) * 1000;
01129     isampf          = AV_RB32(avctx->extradata + 8);
01130     switch (isampf) {
01131     case 44: avctx->sample_rate = 44100;         break;
01132     case 22: avctx->sample_rate = 22050;         break;
01133     case 11: avctx->sample_rate = 11025;         break;
01134     default: avctx->sample_rate = isampf * 1000; break;
01135     }
01136 
01137     if (avctx->channels > CHANNELS_MAX) {
01138         av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
01139                avctx->channels);
01140         return -1;
01141     }
01142     ibps = avctx->bit_rate / (1000 * avctx->channels);
01143     if (ibps < 8 || ibps > 48) {
01144         av_log(avctx, AV_LOG_ERROR, "Bad bitrate per channel value %d\n", ibps);
01145         return AVERROR_INVALIDDATA;
01146     }
01147 
01148     switch ((isampf << 8) +  ibps) {
01149     case (8 <<8) +  8: tctx->mtab = &mode_08_08; break;
01150     case (11<<8) +  8: tctx->mtab = &mode_11_08; break;
01151     case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
01152     case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
01153     case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
01154     case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
01155     case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
01156     case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
01157     case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
01158     default:
01159         av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
01160         return -1;
01161     }
01162 
01163     dsputil_init(&tctx->dsp, avctx);
01164     if ((ret = init_mdct_win(tctx))) {
01165         av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
01166         twin_decode_close(avctx);
01167         return ret;
01168     }
01169     init_bitstream_params(tctx);
01170 
01171     memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
01172 
01173     avcodec_get_frame_defaults(&tctx->frame);
01174     avctx->coded_frame = &tctx->frame;
01175 
01176     return 0;
01177 }
01178 
01179 AVCodec ff_twinvq_decoder = {
01180     .name           = "twinvq",
01181     .type           = AVMEDIA_TYPE_AUDIO,
01182     .id             = CODEC_ID_TWINVQ,
01183     .priv_data_size = sizeof(TwinContext),
01184     .init           = twin_decode_init,
01185     .close          = twin_decode_close,
01186     .decode         = twin_decode_frame,
01187     .capabilities   = CODEC_CAP_DR1,
01188     .long_name      = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
01189 };