Libav
|
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 00028 #include <math.h> 00029 #include <stdint.h> 00030 00031 #include "twinvq_data.h" 00032 00033 enum FrameType { 00034 FT_SHORT = 0, 00035 FT_MEDIUM, 00036 FT_LONG, 00037 FT_PPC, 00038 }; 00039 00043 struct FrameMode { 00044 uint8_t sub; 00045 const uint16_t *bark_tab; 00046 00048 uint8_t bark_env_size; 00049 00050 const int16_t *bark_cb; 00051 uint8_t bark_n_coef; 00052 uint8_t bark_n_bit; 00053 00055 00056 const int16_t *cb0; 00057 const int16_t *cb1; 00059 00060 uint8_t cb_len_read; 00061 }; 00062 00067 typedef struct { 00068 struct FrameMode fmode[3]; 00069 00070 uint16_t size; 00071 uint8_t n_lsp; 00072 const float *lspcodebook; 00073 00074 /* number of bits of the different LSP CB coefficients */ 00075 uint8_t lsp_bit0; 00076 uint8_t lsp_bit1; 00077 uint8_t lsp_bit2; 00078 00079 uint8_t lsp_split; 00080 const int16_t *ppc_shape_cb; 00081 00083 uint8_t ppc_period_bit; 00084 00085 uint8_t ppc_shape_bit; 00086 uint8_t ppc_shape_len; 00087 uint8_t pgain_bit; 00088 00090 uint16_t peak_per2wid; 00091 } ModeTab; 00092 00093 static const ModeTab mode_08_08 = { 00094 { 00095 { 8, bark_tab_s08_64, 10, tab.fcb08s , 1, 5, tab.cb0808s0, tab.cb0808s1, 18}, 00096 { 2, bark_tab_m08_256, 20, tab.fcb08m , 2, 5, tab.cb0808m0, tab.cb0808m1, 16}, 00097 { 1, bark_tab_l08_512, 30, tab.fcb08l , 3, 6, tab.cb0808l0, tab.cb0808l1, 17} 00098 }, 00099 512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40 00100 }; 00101 00102 static const ModeTab mode_11_08 = { 00103 { 00104 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1108s0, tab.cb1108s1, 29}, 00105 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1108m0, tab.cb1108m1, 24}, 00106 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1108l0, tab.cb1108l1, 27} 00107 }, 00108 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90 00109 }; 00110 00111 static const ModeTab mode_11_10 = { 00112 { 00113 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1110s0, tab.cb1110s1, 21}, 00114 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1110m0, tab.cb1110m1, 18}, 00115 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1110l0, tab.cb1110l1, 20} 00116 }, 00117 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90 00118 }; 00119 00120 static const ModeTab mode_16_16 = { 00121 { 00122 { 8, bark_tab_s16_128, 10, tab.fcb16s , 1, 5, tab.cb1616s0, tab.cb1616s1, 16}, 00123 { 2, bark_tab_m16_512, 20, tab.fcb16m , 2, 5, tab.cb1616m0, tab.cb1616m1, 15}, 00124 { 1, bark_tab_l16_1024,30, tab.fcb16l , 3, 6, tab.cb1616l0, tab.cb1616l1, 16} 00125 }, 00126 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180 00127 }; 00128 00129 static const ModeTab mode_22_20 = { 00130 { 00131 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18}, 00132 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17}, 00133 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18} 00134 }, 00135 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144 00136 }; 00137 00138 static const ModeTab mode_22_24 = { 00139 { 00140 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15}, 00141 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14}, 00142 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15} 00143 }, 00144 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144 00145 }; 00146 00147 static const ModeTab mode_22_32 = { 00148 { 00149 { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11}, 00150 { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11}, 00151 { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12} 00152 }, 00153 512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72 00154 }; 00155 00156 static const ModeTab mode_44_40 = { 00157 { 00158 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4440s0, tab.cb4440s1, 18}, 00159 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4440m0, tab.cb4440m1, 17}, 00160 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4440l0, tab.cb4440l1, 17} 00161 }, 00162 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432 00163 }; 00164 00165 static const ModeTab mode_44_48 = { 00166 { 00167 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4448s0, tab.cb4448s1, 15}, 00168 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4448m0, tab.cb4448m1, 14}, 00169 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4448l0, tab.cb4448l1, 14} 00170 }, 00171 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432 00172 }; 00173 00174 typedef struct TwinContext { 00175 AVCodecContext *avctx; 00176 DSPContext dsp; 00177 FFTContext mdct_ctx[3]; 00178 00179 const ModeTab *mtab; 00180 00181 // history 00182 float lsp_hist[2][20]; 00183 float bark_hist[3][2][40]; 00184 00185 // bitstream parameters 00186 int16_t permut[4][4096]; 00187 uint8_t length[4][2]; 00188 uint8_t length_change[4]; 00189 uint8_t bits_main_spec[2][4][2]; 00190 int bits_main_spec_change[4]; 00191 int n_div[4]; 00192 00193 float *spectrum; 00194 float *curr_frame; 00195 float *prev_frame; 00196 int last_block_pos[2]; 00197 00198 float *cos_tabs[3]; 00199 00200 // scratch buffers 00201 float *tmp_buf; 00202 } TwinContext; 00203 00204 #define PPC_SHAPE_CB_SIZE 64 00205 #define SUB_AMP_MAX 4500.0 00206 #define MULAW_MU 100.0 00207 #define GAIN_BITS 8 00208 #define AMP_MAX 13000.0 00209 #define SUB_GAIN_BITS 5 00210 #define WINDOW_TYPE_BITS 4 00211 #define PGAIN_MU 200 00212 00214 static void memset_float(float *buf, float val, int size) 00215 { 00216 while (size--) 00217 *buf++ = val; 00218 } 00219 00232 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order) 00233 { 00234 int j; 00235 float p = 0.5f; 00236 float q = 0.5f; 00237 float two_cos_w = 2.0f*cos_val; 00238 00239 for (j = 0; j + 1 < order; j += 2*2) { 00240 // Unroll the loop once since order is a multiple of four 00241 q *= lsp[j ] - two_cos_w; 00242 p *= lsp[j+1] - two_cos_w; 00243 00244 q *= lsp[j+2] - two_cos_w; 00245 p *= lsp[j+3] - two_cos_w; 00246 } 00247 00248 p *= p * (2.0f - two_cos_w); 00249 q *= q * (2.0f + two_cos_w); 00250 00251 return 0.5 / (p + q); 00252 } 00253 00257 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc) 00258 { 00259 int i; 00260 const ModeTab *mtab = tctx->mtab; 00261 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub; 00262 00263 for (i = 0; i < size_s/2; i++) { 00264 float cos_i = tctx->cos_tabs[0][i]; 00265 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp); 00266 lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp); 00267 } 00268 } 00269 00270 static void interpolate(float *out, float v1, float v2, int size) 00271 { 00272 int i; 00273 float step = (v1 - v2)/(size + 1); 00274 00275 for (i = 0; i < size; i++) { 00276 v2 += step; 00277 out[i] = v2; 00278 } 00279 } 00280 00281 static inline float get_cos(int idx, int part, const float *cos_tab, int size) 00282 { 00283 return part ? -cos_tab[size - idx - 1] : 00284 cos_tab[ idx ]; 00285 } 00286 00301 static inline void eval_lpcenv_or_interp(TwinContext *tctx, 00302 enum FrameType ftype, 00303 float *out, const float *in, 00304 int size, int step, int part) 00305 { 00306 int i; 00307 const ModeTab *mtab = tctx->mtab; 00308 const float *cos_tab = tctx->cos_tabs[ftype]; 00309 00310 // Fill the 's' 00311 for (i = 0; i < size; i += step) 00312 out[i] = 00313 eval_lpc_spectrum(in, 00314 get_cos(i, part, cos_tab, size), 00315 mtab->n_lsp); 00316 00317 // Fill the 'iiiibiiii' 00318 for (i = step; i <= size - 2*step; i += step) { 00319 if (out[i + step] + out[i - step] > 1.95*out[i] || 00320 out[i + step] >= out[i - step]) { 00321 interpolate(out + i - step + 1, out[i], out[i-step], step - 1); 00322 } else { 00323 out[i - step/2] = 00324 eval_lpc_spectrum(in, 00325 get_cos(i-step/2, part, cos_tab, size), 00326 mtab->n_lsp); 00327 interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1); 00328 interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1); 00329 } 00330 } 00331 00332 interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1); 00333 } 00334 00335 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype, 00336 const float *buf, float *lpc, 00337 int size, int step) 00338 { 00339 eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0); 00340 eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1); 00341 00342 interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step); 00343 00344 memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1); 00345 } 00346 00352 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out, 00353 enum FrameType ftype, 00354 const int16_t *cb0, const int16_t *cb1, int cb_len) 00355 { 00356 int pos = 0; 00357 int i, j; 00358 00359 for (i = 0; i < tctx->n_div[ftype]; i++) { 00360 int tmp0, tmp1; 00361 int sign0 = 1; 00362 int sign1 = 1; 00363 const int16_t *tab0, *tab1; 00364 int length = tctx->length[ftype][i >= tctx->length_change[ftype]]; 00365 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]); 00366 00367 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part]; 00368 if (bits == 7) { 00369 if (get_bits1(gb)) 00370 sign0 = -1; 00371 bits = 6; 00372 } 00373 tmp0 = get_bits(gb, bits); 00374 00375 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part]; 00376 00377 if (bits == 7) { 00378 if (get_bits1(gb)) 00379 sign1 = -1; 00380 00381 bits = 6; 00382 } 00383 tmp1 = get_bits(gb, bits); 00384 00385 tab0 = cb0 + tmp0*cb_len; 00386 tab1 = cb1 + tmp1*cb_len; 00387 00388 for (j = 0; j < length; j++) 00389 out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j]; 00390 00391 pos += length; 00392 } 00393 00394 } 00395 00396 static inline float mulawinv(float y, float clip, float mu) 00397 { 00398 y = av_clipf(y/clip, -1, 1); 00399 return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu; 00400 } 00401 00422 static int very_broken_op(int a, int b) 00423 { 00424 int x = a*b + 200; 00425 int size; 00426 const uint8_t *rtab; 00427 00428 if (x%400 || b%5) 00429 return x/400; 00430 00431 x /= 400; 00432 00433 size = tabs[b/5].size; 00434 rtab = tabs[b/5].tab; 00435 return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size]; 00436 } 00437 00443 static void add_peak(int period, int width, const float *shape, 00444 float ppc_gain, float *speech, int len) 00445 { 00446 int i, j; 00447 00448 const float *shape_end = shape + len; 00449 int center; 00450 00451 // First peak centered around zero 00452 for (i = 0; i < width/2; i++) 00453 speech[i] += ppc_gain * *shape++; 00454 00455 for (i = 1; i < ROUNDED_DIV(len,width) ; i++) { 00456 center = very_broken_op(period, i); 00457 for (j = -width/2; j < (width+1)/2; j++) 00458 speech[j+center] += ppc_gain * *shape++; 00459 } 00460 00461 // For the last block, be careful not to go beyond the end of the buffer 00462 center = very_broken_op(period, i); 00463 for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++) 00464 speech[j+center] += ppc_gain * *shape++; 00465 } 00466 00467 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape, 00468 float ppc_gain, float *speech) 00469 { 00470 const ModeTab *mtab = tctx->mtab; 00471 int isampf = tctx->avctx->sample_rate/1000; 00472 int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels); 00473 int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf); 00474 int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf); 00475 int period_range = max_period - min_period; 00476 00477 // This is actually the period multiplied by 400. It is just linearly coded 00478 // between its maximum and minimum value. 00479 int period = min_period + 00480 ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1); 00481 int width; 00482 00483 if (isampf == 22 && ibps == 32) { 00484 // For some unknown reason, NTT decided to code this case differently... 00485 width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size); 00486 } else 00487 width = (period )* mtab->peak_per2wid/(400*mtab->size); 00488 00489 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len); 00490 } 00491 00492 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype, 00493 float *out) 00494 { 00495 const ModeTab *mtab = tctx->mtab; 00496 int i, j; 00497 int sub = mtab->fmode[ftype].sub; 00498 float step = AMP_MAX / ((1 << GAIN_BITS) - 1); 00499 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1); 00500 00501 if (ftype == FT_LONG) { 00502 for (i = 0; i < tctx->avctx->channels; i++) 00503 out[i] = (1./(1<<13)) * 00504 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS), 00505 AMP_MAX, MULAW_MU); 00506 } else { 00507 for (i = 0; i < tctx->avctx->channels; i++) { 00508 float val = (1./(1<<23)) * 00509 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS), 00510 AMP_MAX, MULAW_MU); 00511 00512 for (j = 0; j < sub; j++) { 00513 out[i*sub + j] = 00514 val*mulawinv(sub_step* 0.5 + 00515 sub_step* get_bits(gb, SUB_GAIN_BITS), 00516 SUB_AMP_MAX, MULAW_MU); 00517 } 00518 } 00519 } 00520 } 00521 00528 static void rearrange_lsp(int order, float *lsp, float min_dist) 00529 { 00530 int i; 00531 float min_dist2 = min_dist * 0.5; 00532 for (i = 1; i < order; i++) 00533 if (lsp[i] - lsp[i-1] < min_dist) { 00534 float avg = (lsp[i] + lsp[i-1]) * 0.5; 00535 00536 lsp[i-1] = avg - min_dist2; 00537 lsp[i ] = avg + min_dist2; 00538 } 00539 } 00540 00541 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2, 00542 int lpc_hist_idx, float *lsp, float *hist) 00543 { 00544 const ModeTab *mtab = tctx->mtab; 00545 int i, j; 00546 00547 const float *cb = mtab->lspcodebook; 00548 const float *cb2 = cb + (1 << mtab->lsp_bit1)*mtab->n_lsp; 00549 const float *cb3 = cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp; 00550 00551 const int8_t funny_rounding[4] = { 00552 -2, 00553 mtab->lsp_split == 4 ? -2 : 1, 00554 mtab->lsp_split == 4 ? -2 : 1, 00555 0 00556 }; 00557 00558 j = 0; 00559 for (i = 0; i < mtab->lsp_split; i++) { 00560 int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split; 00561 for (; j < chunk_end; j++) 00562 lsp[j] = cb [lpc_idx1 * mtab->n_lsp + j] + 00563 cb2[lpc_idx2[i] * mtab->n_lsp + j]; 00564 } 00565 00566 rearrange_lsp(mtab->n_lsp, lsp, 0.0001); 00567 00568 for (i = 0; i < mtab->n_lsp; i++) { 00569 float tmp1 = 1. - cb3[lpc_hist_idx*mtab->n_lsp + i]; 00570 float tmp2 = hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i]; 00571 hist[i] = lsp[i]; 00572 lsp[i] = lsp[i] * tmp1 + tmp2; 00573 } 00574 00575 rearrange_lsp(mtab->n_lsp, lsp, 0.0001); 00576 rearrange_lsp(mtab->n_lsp, lsp, 0.000095); 00577 ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp); 00578 } 00579 00580 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp, 00581 enum FrameType ftype, float *lpc) 00582 { 00583 int i; 00584 int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub; 00585 00586 for (i = 0; i < tctx->mtab->n_lsp; i++) 00587 lsp[i] = 2*cos(lsp[i]); 00588 00589 switch (ftype) { 00590 case FT_LONG: 00591 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8); 00592 break; 00593 case FT_MEDIUM: 00594 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2); 00595 break; 00596 case FT_SHORT: 00597 eval_lpcenv(tctx, lsp, lpc); 00598 break; 00599 } 00600 } 00601 00602 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype, 00603 float *in, float *prev, int ch) 00604 { 00605 const ModeTab *mtab = tctx->mtab; 00606 int bsize = mtab->size / mtab->fmode[ftype].sub; 00607 int size = mtab->size; 00608 float *buf1 = tctx->tmp_buf; 00609 int j; 00610 int wsize; // Window size 00611 float *out = tctx->curr_frame + 2*ch*mtab->size; 00612 float *out2 = out; 00613 float *prev_buf; 00614 int first_wsize; 00615 00616 static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1}; 00617 int types_sizes[] = { 00618 mtab->size / mtab->fmode[FT_LONG ].sub, 00619 mtab->size / mtab->fmode[FT_MEDIUM].sub, 00620 mtab->size / (2*mtab->fmode[FT_SHORT ].sub), 00621 }; 00622 00623 wsize = types_sizes[wtype_to_wsize[wtype]]; 00624 first_wsize = wsize; 00625 prev_buf = prev + (size - bsize)/2; 00626 00627 for (j = 0; j < mtab->fmode[ftype].sub; j++) { 00628 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype; 00629 00630 if (!j && wtype == 4) 00631 sub_wtype = 4; 00632 else if (j == mtab->fmode[ftype].sub-1 && wtype == 7) 00633 sub_wtype = 7; 00634 00635 wsize = types_sizes[wtype_to_wsize[sub_wtype]]; 00636 00637 ff_imdct_half(&tctx->mdct_ctx[ftype], buf1 + bsize*j, in + bsize*j); 00638 00639 tctx->dsp.vector_fmul_window(out2, 00640 prev_buf + (bsize-wsize)/2, 00641 buf1 + bsize*j, 00642 ff_sine_windows[av_log2(wsize)], 00643 0.0, 00644 wsize/2); 00645 out2 += wsize; 00646 00647 memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float)); 00648 00649 out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize; 00650 00651 prev_buf = buf1 + bsize*j + bsize/2; 00652 } 00653 00654 tctx->last_block_pos[ch] = (size + first_wsize)/2; 00655 } 00656 00657 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype, 00658 float *out) 00659 { 00660 const ModeTab *mtab = tctx->mtab; 00661 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0]; 00662 int i, j; 00663 00664 for (i = 0; i < tctx->avctx->channels; i++) { 00665 imdct_and_window(tctx, ftype, wtype, 00666 tctx->spectrum + i*mtab->size, 00667 prev_buf + 2*i*mtab->size, 00668 i); 00669 } 00670 00671 if (tctx->avctx->channels == 2) { 00672 for (i = 0; i < mtab->size - tctx->last_block_pos[0]; i++) { 00673 float f1 = prev_buf[ i]; 00674 float f2 = prev_buf[2*mtab->size + i]; 00675 out[2*i ] = f1 + f2; 00676 out[2*i + 1] = f1 - f2; 00677 } 00678 for (j = 0; i < mtab->size; j++,i++) { 00679 float f1 = tctx->curr_frame[ j]; 00680 float f2 = tctx->curr_frame[2*mtab->size + j]; 00681 out[2*i ] = f1 + f2; 00682 out[2*i + 1] = f1 - f2; 00683 } 00684 } else { 00685 memcpy(out, prev_buf, 00686 (mtab->size - tctx->last_block_pos[0]) * sizeof(*out)); 00687 00688 out += mtab->size - tctx->last_block_pos[0]; 00689 00690 memcpy(out, tctx->curr_frame, 00691 (tctx->last_block_pos[0]) * sizeof(*out)); 00692 } 00693 00694 } 00695 00696 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist, 00697 int ch, float *out, float gain, enum FrameType ftype) 00698 { 00699 const ModeTab *mtab = tctx->mtab; 00700 int i,j; 00701 float *hist = tctx->bark_hist[ftype][ch]; 00702 float val = ((const float []) {0.4, 0.35, 0.28})[ftype]; 00703 int bark_n_coef = mtab->fmode[ftype].bark_n_coef; 00704 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef; 00705 int idx = 0; 00706 00707 for (i = 0; i < fw_cb_len; i++) 00708 for (j = 0; j < bark_n_coef; j++, idx++) { 00709 float tmp2 = 00710 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096); 00711 float st = use_hist ? 00712 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.; 00713 00714 hist[idx] = tmp2; 00715 if (st < -1.) st = 1.; 00716 00717 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]); 00718 out += mtab->fmode[ftype].bark_tab[idx]; 00719 } 00720 00721 } 00722 00723 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb, 00724 float *out, enum FrameType ftype) 00725 { 00726 const ModeTab *mtab = tctx->mtab; 00727 int channels = tctx->avctx->channels; 00728 int sub = mtab->fmode[ftype].sub; 00729 int block_size = mtab->size / sub; 00730 float gain[channels*sub]; 00731 float ppc_shape[mtab->ppc_shape_len * channels * 4]; 00732 uint8_t bark1[channels][sub][mtab->fmode[ftype].bark_n_coef]; 00733 uint8_t bark_use_hist[channels][sub]; 00734 00735 uint8_t lpc_idx1[channels]; 00736 uint8_t lpc_idx2[channels][tctx->mtab->lsp_split]; 00737 uint8_t lpc_hist_idx[channels]; 00738 00739 int i, j, k; 00740 00741 dequant(tctx, gb, out, ftype, 00742 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1, 00743 mtab->fmode[ftype].cb_len_read); 00744 00745 for (i = 0; i < channels; i++) 00746 for (j = 0; j < sub; j++) 00747 for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++) 00748 bark1[i][j][k] = 00749 get_bits(gb, mtab->fmode[ftype].bark_n_bit); 00750 00751 for (i = 0; i < channels; i++) 00752 for (j = 0; j < sub; j++) 00753 bark_use_hist[i][j] = get_bits1(gb); 00754 00755 dec_gain(tctx, gb, ftype, gain); 00756 00757 for (i = 0; i < channels; i++) { 00758 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0); 00759 lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1); 00760 00761 for (j = 0; j < tctx->mtab->lsp_split; j++) 00762 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2); 00763 } 00764 00765 if (ftype == FT_LONG) { 00766 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/ 00767 tctx->n_div[3]; 00768 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb, 00769 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p); 00770 } 00771 00772 for (i = 0; i < channels; i++) { 00773 float *chunk = out + mtab->size * i; 00774 float lsp[tctx->mtab->n_lsp]; 00775 00776 for (j = 0; j < sub; j++) { 00777 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i, 00778 tctx->tmp_buf, gain[sub*i+j], ftype); 00779 00780 tctx->dsp.vector_fmul(chunk + block_size*j, tctx->tmp_buf, 00781 block_size); 00782 00783 } 00784 00785 if (ftype == FT_LONG) { 00786 float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1); 00787 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit); 00788 int g_coef = get_bits(gb, tctx->mtab->pgain_bit); 00789 float v = 1./8192* 00790 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU); 00791 00792 decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v, 00793 chunk); 00794 } 00795 00796 decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp, 00797 tctx->lsp_hist[i]); 00798 00799 dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf); 00800 00801 for (j = 0; j < mtab->fmode[ftype].sub; j++) { 00802 tctx->dsp.vector_fmul(chunk, tctx->tmp_buf, block_size); 00803 chunk += block_size; 00804 } 00805 } 00806 } 00807 00808 static int twin_decode_frame(AVCodecContext * avctx, void *data, 00809 int *data_size, AVPacket *avpkt) 00810 { 00811 const uint8_t *buf = avpkt->data; 00812 int buf_size = avpkt->size; 00813 TwinContext *tctx = avctx->priv_data; 00814 GetBitContext gb; 00815 const ModeTab *mtab = tctx->mtab; 00816 float *out = data; 00817 enum FrameType ftype; 00818 int window_type; 00819 static const enum FrameType wtype_to_ftype_table[] = { 00820 FT_LONG, FT_LONG, FT_SHORT, FT_LONG, 00821 FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM 00822 }; 00823 00824 if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) { 00825 av_log(avctx, AV_LOG_ERROR, 00826 "Frame too small (%d bytes). Truncated file?\n", buf_size); 00827 *data_size = 0; 00828 return buf_size; 00829 } 00830 00831 init_get_bits(&gb, buf, buf_size * 8); 00832 skip_bits(&gb, get_bits(&gb, 8)); 00833 window_type = get_bits(&gb, WINDOW_TYPE_BITS); 00834 00835 if (window_type > 8) { 00836 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n"); 00837 return -1; 00838 } 00839 00840 ftype = wtype_to_ftype_table[window_type]; 00841 00842 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype); 00843 00844 imdct_output(tctx, ftype, window_type, out); 00845 00846 FFSWAP(float*, tctx->curr_frame, tctx->prev_frame); 00847 00848 if (tctx->avctx->frame_number < 2) { 00849 *data_size=0; 00850 return buf_size; 00851 } 00852 00853 *data_size = mtab->size*avctx->channels*4; 00854 00855 return buf_size; 00856 } 00857 00861 static av_cold void init_mdct_win(TwinContext *tctx) 00862 { 00863 int i,j; 00864 const ModeTab *mtab = tctx->mtab; 00865 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub; 00866 int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub; 00867 int channels = tctx->avctx->channels; 00868 float norm = channels == 1 ? 2. : 1.; 00869 00870 for (i = 0; i < 3; i++) { 00871 int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub; 00872 ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1, 00873 -sqrt(norm/bsize) / (1<<15)); 00874 } 00875 00876 tctx->tmp_buf = av_malloc(mtab->size * sizeof(*tctx->tmp_buf)); 00877 00878 tctx->spectrum = av_malloc(2*mtab->size*channels*sizeof(float)); 00879 tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float)); 00880 tctx->prev_frame = av_malloc(2*mtab->size*channels*sizeof(float)); 00881 00882 for (i = 0; i < 3; i++) { 00883 int m = 4*mtab->size/mtab->fmode[i].sub; 00884 double freq = 2*M_PI/m; 00885 tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs)); 00886 00887 for (j = 0; j <= m/8; j++) 00888 tctx->cos_tabs[i][j] = cos((2*j + 1)*freq); 00889 for (j = 1; j < m/8; j++) 00890 tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j]; 00891 } 00892 00893 00894 ff_init_ff_sine_windows(av_log2(size_m)); 00895 ff_init_ff_sine_windows(av_log2(size_s/2)); 00896 ff_init_ff_sine_windows(av_log2(mtab->size)); 00897 } 00898 00905 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks, 00906 int block_size, 00907 const uint8_t line_len[2], int length_div, 00908 enum FrameType ftype) 00909 00910 { 00911 int i,j; 00912 00913 for (i = 0; i < line_len[0]; i++) { 00914 int shift; 00915 00916 if (num_blocks == 1 || 00917 (ftype == FT_LONG && num_vect % num_blocks) || 00918 (ftype != FT_LONG && num_vect & 1 ) || 00919 i == line_len[1]) { 00920 shift = 0; 00921 } else if (ftype == FT_LONG) { 00922 shift = i; 00923 } else 00924 shift = i*i; 00925 00926 for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++) 00927 tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect; 00928 } 00929 } 00930 00946 static void transpose_perm(int16_t *out, int16_t *in, int num_vect, 00947 const uint8_t line_len[2], int length_div) 00948 { 00949 int i,j; 00950 int cont= 0; 00951 for (i = 0; i < num_vect; i++) 00952 for (j = 0; j < line_len[i >= length_div]; j++) 00953 out[cont++] = in[j*num_vect + i]; 00954 } 00955 00956 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size) 00957 { 00958 int block_size = size/n_blocks; 00959 int i; 00960 00961 for (i = 0; i < size; i++) 00962 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks; 00963 } 00964 00965 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype) 00966 { 00967 int block_size; 00968 const ModeTab *mtab = tctx->mtab; 00969 int size = tctx->avctx->channels*mtab->fmode[ftype].sub; 00970 int16_t *tmp_perm = (int16_t *) tctx->tmp_buf; 00971 00972 if (ftype == FT_PPC) { 00973 size = tctx->avctx->channels; 00974 block_size = mtab->ppc_shape_len; 00975 } else 00976 block_size = mtab->size / mtab->fmode[ftype].sub; 00977 00978 permutate_in_line(tmp_perm, tctx->n_div[ftype], size, 00979 block_size, tctx->length[ftype], 00980 tctx->length_change[ftype], ftype); 00981 00982 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype], 00983 tctx->length[ftype], tctx->length_change[ftype]); 00984 00985 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size, 00986 size*block_size); 00987 } 00988 00989 static av_cold void init_bitstream_params(TwinContext *tctx) 00990 { 00991 const ModeTab *mtab = tctx->mtab; 00992 int n_ch = tctx->avctx->channels; 00993 int total_fr_bits = tctx->avctx->bit_rate*mtab->size/ 00994 tctx->avctx->sample_rate; 00995 00996 int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 + 00997 mtab->lsp_split*mtab->lsp_bit2); 00998 00999 int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit + 01000 mtab->ppc_period_bit); 01001 01002 int bsize_no_main_cb[3]; 01003 int bse_bits[3]; 01004 int i; 01005 enum FrameType frametype; 01006 01007 for (i = 0; i < 3; i++) 01008 // +1 for history usage switch 01009 bse_bits[i] = n_ch * 01010 (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1); 01011 01012 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits + 01013 WINDOW_TYPE_BITS + n_ch*GAIN_BITS; 01014 01015 for (i = 0; i < 2; i++) 01016 bsize_no_main_cb[i] = 01017 lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS + 01018 mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS); 01019 01020 // The remaining bits are all used for the main spectrum coefficients 01021 for (i = 0; i < 4; i++) { 01022 int bit_size; 01023 int vect_size; 01024 int rounded_up, rounded_down, num_rounded_down, num_rounded_up; 01025 if (i == 3) { 01026 bit_size = n_ch * mtab->ppc_shape_bit; 01027 vect_size = n_ch * mtab->ppc_shape_len; 01028 } else { 01029 bit_size = total_fr_bits - bsize_no_main_cb[i]; 01030 vect_size = n_ch * mtab->size; 01031 } 01032 01033 tctx->n_div[i] = (bit_size + 13) / 14; 01034 01035 rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i]; 01036 rounded_down = (bit_size )/tctx->n_div[i]; 01037 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size; 01038 num_rounded_up = tctx->n_div[i] - num_rounded_down; 01039 tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2; 01040 tctx->bits_main_spec[1][i][0] = (rounded_up )/2; 01041 tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2; 01042 tctx->bits_main_spec[1][i][1] = (rounded_down )/2; 01043 tctx->bits_main_spec_change[i] = num_rounded_up; 01044 01045 rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i]; 01046 rounded_down = (vect_size )/tctx->n_div[i]; 01047 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size; 01048 num_rounded_up = tctx->n_div[i] - num_rounded_down; 01049 tctx->length[i][0] = rounded_up; 01050 tctx->length[i][1] = rounded_down; 01051 tctx->length_change[i] = num_rounded_up; 01052 } 01053 01054 for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++) 01055 construct_perm_table(tctx, frametype); 01056 } 01057 01058 static av_cold int twin_decode_init(AVCodecContext *avctx) 01059 { 01060 TwinContext *tctx = avctx->priv_data; 01061 int isampf = avctx->sample_rate/1000; 01062 int ibps = avctx->bit_rate/(1000 * avctx->channels); 01063 01064 tctx->avctx = avctx; 01065 avctx->sample_fmt = SAMPLE_FMT_FLT; 01066 01067 if (avctx->channels > 2) { 01068 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n", 01069 avctx->channels); 01070 return -1; 01071 } 01072 01073 switch ((isampf << 8) + ibps) { 01074 case (8 <<8) + 8: tctx->mtab = &mode_08_08; break; 01075 case (11<<8) + 8: tctx->mtab = &mode_11_08; break; 01076 case (11<<8) + 10: tctx->mtab = &mode_11_10; break; 01077 case (16<<8) + 16: tctx->mtab = &mode_16_16; break; 01078 case (22<<8) + 20: tctx->mtab = &mode_22_20; break; 01079 case (22<<8) + 24: tctx->mtab = &mode_22_24; break; 01080 case (22<<8) + 32: tctx->mtab = &mode_22_32; break; 01081 case (44<<8) + 40: tctx->mtab = &mode_44_40; break; 01082 case (44<<8) + 48: tctx->mtab = &mode_44_48; break; 01083 default: 01084 av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf); 01085 return -1; 01086 } 01087 01088 dsputil_init(&tctx->dsp, avctx); 01089 init_mdct_win(tctx); 01090 init_bitstream_params(tctx); 01091 01092 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist)); 01093 01094 return 0; 01095 } 01096 01097 static av_cold int twin_decode_close(AVCodecContext *avctx) 01098 { 01099 TwinContext *tctx = avctx->priv_data; 01100 int i; 01101 01102 for (i = 0; i < 3; i++) { 01103 ff_mdct_end(&tctx->mdct_ctx[i]); 01104 av_free(tctx->cos_tabs[i]); 01105 } 01106 01107 01108 av_free(tctx->curr_frame); 01109 av_free(tctx->spectrum); 01110 av_free(tctx->prev_frame); 01111 av_free(tctx->tmp_buf); 01112 01113 return 0; 01114 } 01115 01116 AVCodec twinvq_decoder = 01117 { 01118 "twinvq", 01119 AVMEDIA_TYPE_AUDIO, 01120 CODEC_ID_TWINVQ, 01121 sizeof(TwinContext), 01122 twin_decode_init, 01123 NULL, 01124 twin_decode_close, 01125 twin_decode_frame, 01126 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"), 01127 };