projects / libav.git / blob
? search:


25498c4d28cf80cad58cc64550f99941b92c6097
[libav.git] / libavcodec / wmadec.c
1 /*
2 * WMA compatible decoder
3 * Copyright (c) 2002 The FFmpeg Project.
4 *
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
9 *
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
14 *
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20 /**
21 * @file wmadec.c
22 * WMA compatible decoder.
23 */
24
25 #include "avcodec.h"
26 #include "dsputil.h"
27
28 /* size of blocks */
29 #define BLOCK_MIN_BITS 7
30 #define BLOCK_MAX_BITS 11
31 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
32
33 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
34
35 /* XXX: find exact max size */
36 #define HIGH_BAND_MAX_SIZE 16
37
38 #define NB_LSP_COEFS 10
39
40 /* XXX: is it a suitable value ? */
41 #define MAX_CODED_SUPERFRAME_SIZE 4096
42
43 #define MAX_CHANNELS 2
44
45 #define NOISE_TAB_SIZE 8192
46
47 #define LSP_POW_BITS 7
48
49 typedef struct WMADecodeContext {
50 GetBitContext gb;
51 int sample_rate;
52 int nb_channels;
53 int bit_rate;
54 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
55 int block_align;
56 int use_bit_reservoir;
57 int use_variable_block_len;
58 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
59 int use_noise_coding; /* true if perceptual noise is added */
60 int byte_offset_bits;
61 VLC exp_vlc;
62 int exponent_sizes[BLOCK_NB_SIZES];
63 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
64 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
65 int coefs_start; /* first coded coef */
66 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
67 int exponent_high_sizes[BLOCK_NB_SIZES];
68 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
69 VLC hgain_vlc;
70
71 /* coded values in high bands */
72 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
73 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
74
75 /* there are two possible tables for spectral coefficients */
76 VLC coef_vlc[2];
77 uint16_t *run_table[2];
78 uint16_t *level_table[2];
79 /* frame info */
80 int frame_len; /* frame length in samples */
81 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
82 int nb_block_sizes; /* number of block sizes */
83 /* block info */
84 int reset_block_lengths;
85 int block_len_bits; /* log2 of current block length */
86 int next_block_len_bits; /* log2 of next block length */
87 int prev_block_len_bits; /* log2 of prev block length */
88 int block_len; /* block length in samples */
89 int block_num; /* block number in current frame */
90 int block_pos; /* current position in frame */
91 uint8_t ms_stereo; /* true if mid/side stereo mode */
92 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
93 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
94 float max_exponent[MAX_CHANNELS];
95 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
96 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
97 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
98 float *windows[BLOCK_NB_SIZES];
99 FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
100 /* output buffer for one frame and the last for IMDCT windowing */
101 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
102 /* last frame info */
103 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
104 int last_bitoffset;
105 int last_superframe_len;
106 float noise_table[NOISE_TAB_SIZE];
107 int noise_index;
108 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
109 /* lsp_to_curve tables */
110 float lsp_cos_table[BLOCK_MAX_SIZE];
111 float lsp_pow_e_table[256];
112 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
113 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
114
115 #ifdef TRACE
116 int frame_count;
117 #endif
118 } WMADecodeContext;
119
120 typedef struct CoefVLCTable {
121 int n; /* total number of codes */
122 const uint32_t *huffcodes; /* VLC bit values */
123 const uint8_t *huffbits; /* VLC bit size */
124 const uint16_t *levels; /* table to build run/level tables */
125 } CoefVLCTable;
126
127 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
128
129 #include "wmadata.h"
130
131 #ifdef TRACE
132 static void dump_shorts(const char *name, const short *tab, int n)
133 {
134 int i;
135
136 tprintf("%s[%d]:\n", name, n);
137 for(i=0;i<n;i++) {
138 if ((i & 7) == 0)
139 tprintf("%4d: ", i);
140 tprintf(" %5d.0", tab[i]);
141 if ((i & 7) == 7)
142 tprintf("\n");
143 }
144 }
145
146 static void dump_floats(const char *name, int prec, const float *tab, int n)
147 {
148 int i;
149
150 tprintf("%s[%d]:\n", name, n);
151 for(i=0;i<n;i++) {
152 if ((i & 7) == 0)
153 tprintf("%4d: ", i);
154 tprintf(" %8.*f", prec, tab[i]);
155 if ((i & 7) == 7)
156 tprintf("\n");
157 }
158 if ((i & 7) != 0)
159 tprintf("\n");
160 }
161 #endif
162
163 /* XXX: use same run/length optimization as mpeg decoders */
164 static void init_coef_vlc(VLC *vlc,
165 uint16_t **prun_table, uint16_t **plevel_table,
166 const CoefVLCTable *vlc_table)
167 {
168 int n = vlc_table->n;
169 const uint8_t *table_bits = vlc_table->huffbits;
170 const uint32_t *table_codes = vlc_table->huffcodes;
171 const uint16_t *levels_table = vlc_table->levels;
172 uint16_t *run_table, *level_table;
173 const uint16_t *p;
174 int i, l, j, level;
175
176 init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4);
177
178 run_table = av_malloc(n * sizeof(uint16_t));
179 level_table = av_malloc(n * sizeof(uint16_t));
180 p = levels_table;
181 i = 2;
182 level = 1;
183 while (i < n) {
184 l = *p++;
185 for(j=0;j<l;j++) {
186 run_table[i] = j;
187 level_table[i] = level;
188 i++;
189 }
190 level++;
191 }
192 *prun_table = run_table;
193 *plevel_table = level_table;
194 }
195
196 static int wma_decode_init(AVCodecContext * avctx)
197 {
198 WMADecodeContext *s = avctx->priv_data;
199 int i, flags1, flags2;
200 float *window;
201 uint8_t *extradata;
202 float bps1, high_freq, bps;
203 int sample_rate1;
204 int coef_vlc_table;
205
206 s->sample_rate = avctx->sample_rate;
207 s->nb_channels = avctx->channels;
208 s->bit_rate = avctx->bit_rate;
209 s->block_align = avctx->block_align;
210
211 if (avctx->codec->id == CODEC_ID_WMAV1) {
212 s->version = 1;
213 } else {
214 s->version = 2;
215 }
216
217 /* extract flag infos */
218 flags1 = 0;
219 flags2 = 0;
220 extradata = avctx->extradata;
221 if (s->version == 1 && avctx->extradata_size >= 4) {
222 flags1 = extradata[0] | (extradata[1] << 8);
223 flags2 = extradata[2] | (extradata[3] << 8);
224 } else if (s->version == 2 && avctx->extradata_size >= 6) {
225 flags1 = extradata[0] | (extradata[1] << 8) |
226 (extradata[2] << 16) | (extradata[3] << 24);
227 flags2 = extradata[4] | (extradata[5] << 8);
228 }
229 s->use_exp_vlc = flags2 & 0x0001;
230 s->use_bit_reservoir = flags2 & 0x0002;
231 s->use_variable_block_len = flags2 & 0x0004;
232
233 /* compute MDCT block size */
234 if (s->sample_rate <= 16000) {
235 s->frame_len_bits = 9;
236 } else if (s->sample_rate <= 22050 ||
237 (s->sample_rate <= 32000 && s->version == 1)) {
238 s->frame_len_bits = 10;
239 } else {
240 s->frame_len_bits = 11;
241 }
242 s->frame_len = 1 << s->frame_len_bits;
243 if (s->use_variable_block_len) {
244 int nb_max, nb;
245 nb = ((flags2 >> 3) & 3) + 1;
246 if ((s->bit_rate / s->nb_channels) >= 32000)
247 nb += 2;
248 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
249 if (nb > nb_max)
250 nb = nb_max;
251 s->nb_block_sizes = nb + 1;
252 } else {
253 s->nb_block_sizes = 1;
254 }
255
256 /* init rate dependant parameters */
257 s->use_noise_coding = 1;
258 high_freq = s->sample_rate * 0.5;
259
260 /* if version 2, then the rates are normalized */
261 sample_rate1 = s->sample_rate;
262 if (s->version == 2) {
263 if (sample_rate1 >= 44100)
264 sample_rate1 = 44100;
265 else if (sample_rate1 >= 22050)
266 sample_rate1 = 22050;
267 else if (sample_rate1 >= 16000)
268 sample_rate1 = 16000;
269 else if (sample_rate1 >= 11025)
270 sample_rate1 = 11025;
271 else if (sample_rate1 >= 8000)
272 sample_rate1 = 8000;
273 }
274
275 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
276 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
277
278 /* compute high frequency value and choose if noise coding should
279 be activated */
280 bps1 = bps;
281 if (s->nb_channels == 2)
282 bps1 = bps * 1.6;
283 if (sample_rate1 == 44100) {
284 if (bps1 >= 0.61)
285 s->use_noise_coding = 0;
286 else
287 high_freq = high_freq * 0.4;
288 } else if (sample_rate1 == 22050) {
289 if (bps1 >= 1.16)
290 s->use_noise_coding = 0;
291 else if (bps1 >= 0.72)
292 high_freq = high_freq * 0.7;
293 else
294 high_freq = high_freq * 0.6;
295 } else if (sample_rate1 == 16000) {
296 if (bps > 0.5)
297 high_freq = high_freq * 0.5;
298 else
299 high_freq = high_freq * 0.3;
300 } else if (sample_rate1 == 11025) {
301 high_freq = high_freq * 0.7;
302 } else if (sample_rate1 == 8000) {
303 if (bps <= 0.625) {
304 high_freq = high_freq * 0.5;
305 } else if (bps > 0.75) {
306 s->use_noise_coding = 0;
307 } else {
308 high_freq = high_freq * 0.65;
309 }
310 } else {
311 if (bps >= 0.8) {
312 high_freq = high_freq * 0.75;
313 } else if (bps >= 0.6) {
314 high_freq = high_freq * 0.6;
315 } else {
316 high_freq = high_freq * 0.5;
317 }
318 }
319 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
320 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
321 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
322 s->block_align);
323 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
324 bps, bps1, high_freq, s->byte_offset_bits);
325 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
326 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
327
328 /* compute the scale factor band sizes for each MDCT block size */
329 {
330 int a, b, pos, lpos, k, block_len, i, j, n;
331 const uint8_t *table;
332
333 if (s->version == 1) {
334 s->coefs_start = 3;
335 } else {
336 s->coefs_start = 0;
337 }
338 for(k = 0; k < s->nb_block_sizes; k++) {
339 block_len = s->frame_len >> k;
340
341 if (s->version == 1) {
342 lpos = 0;
343 for(i=0;i<25;i++) {
344 a = wma_critical_freqs[i];
345 b = s->sample_rate;
346 pos = ((block_len * 2 * a) + (b >> 1)) / b;
347 if (pos > block_len)
348 pos = block_len;
349 s->exponent_bands[0][i] = pos - lpos;
350 if (pos >= block_len) {
351 i++;
352 break;
353 }
354 lpos = pos;
355 }
356 s->exponent_sizes[0] = i;
357 } else {
358 /* hardcoded tables */
359 table = NULL;
360 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
361 if (a < 3) {
362 if (s->sample_rate >= 44100)
363 table = exponent_band_44100[a];
364 else if (s->sample_rate >= 32000)
365 table = exponent_band_32000[a];
366 else if (s->sample_rate >= 22050)
367 table = exponent_band_22050[a];
368 }
369 if (table) {
370 n = *table++;
371 for(i=0;i<n;i++)
372 s->exponent_bands[k][i] = table[i];
373 s->exponent_sizes[k] = n;
374 } else {
375 j = 0;
376 lpos = 0;
377 for(i=0;i<25;i++) {
378 a = wma_critical_freqs[i];
379 b = s->sample_rate;
380 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
381 pos <<= 2;
382 if (pos > block_len)
383 pos = block_len;
384 if (pos > lpos)
385 s->exponent_bands[k][j++] = pos - lpos;
386 if (pos >= block_len)
387 break;
388 lpos = pos;
389 }
390 s->exponent_sizes[k] = j;
391 }
392 }
393
394 /* max number of coefs */
395 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
396 /* high freq computation */
397 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
398 s->sample_rate + 0.5);
399 n = s->exponent_sizes[k];
400 j = 0;
401 pos = 0;
402 for(i=0;i<n;i++) {
403 int start, end;
404 start = pos;
405 pos += s->exponent_bands[k][i];
406 end = pos;
407 if (start < s->high_band_start[k])
408 start = s->high_band_start[k];
409 if (end > s->coefs_end[k])
410 end = s->coefs_end[k];
411 if (end > start)
412 s->exponent_high_bands[k][j++] = end - start;
413 }
414 s->exponent_high_sizes[k] = j;
415 #if 0
416 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
417 s->frame_len >> k,
418 s->coefs_end[k],
419 s->high_band_start[k],
420 s->exponent_high_sizes[k]);
421 for(j=0;j<s->exponent_high_sizes[k];j++)
422 tprintf(" %d", s->exponent_high_bands[k][j]);
423 tprintf("\n");
424 #endif
425 }
426 }
427
428 #ifdef TRACE
429 {
430 int i, j;
431 for(i = 0; i < s->nb_block_sizes; i++) {
432 tprintf("%5d: n=%2d:",
433 s->frame_len >> i,
434 s->exponent_sizes[i]);
435 for(j=0;j<s->exponent_sizes[i];j++)
436 tprintf(" %d", s->exponent_bands[i][j]);
437 tprintf("\n");
438 }
439 }
440 #endif
441
442 /* init MDCT */
443 for(i = 0; i < s->nb_block_sizes; i++)
444 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
445
446 /* init MDCT windows : simple sinus window */
447 for(i = 0; i < s->nb_block_sizes; i++) {
448 int n, j;
449 float alpha;
450 n = 1 << (s->frame_len_bits - i);
451 window = av_malloc(sizeof(float) * n);
452 alpha = M_PI / (2.0 * n);
453 for(j=0;j<n;j++) {
454 window[n - j - 1] = sin((j + 0.5) * alpha);
455 }
456 s->windows[i] = window;
457 }
458
459 s->reset_block_lengths = 1;
460
461 if (s->use_noise_coding) {
462
463 /* init the noise generator */
464 if (s->use_exp_vlc)
465 s->noise_mult = 0.02;
466 else
467 s->noise_mult = 0.04;
468
469 #ifdef TRACE
470 for(i=0;i<NOISE_TAB_SIZE;i++)
471 s->noise_table[i] = 1.0 * s->noise_mult;
472 #else
473 {
474 unsigned int seed;
475 float norm;
476 seed = 1;
477 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
478 for(i=0;i<NOISE_TAB_SIZE;i++) {
479 seed = seed * 314159 + 1;
480 s->noise_table[i] = (float)((int)seed) * norm;
481 }
482 }
483 #endif
484 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
485 hgain_huffbits, 1, 1,
486 hgain_huffcodes, 2, 2);
487 }
488
489 if (s->use_exp_vlc) {
490 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
491 scale_huffbits, 1, 1,
492 scale_huffcodes, 4, 4);
493 } else {
494 wma_lsp_to_curve_init(s, s->frame_len);
495 }
496
497 /* choose the VLC tables for the coefficients */
498 coef_vlc_table = 2;
499 if (s->sample_rate >= 32000) {
500 if (bps1 < 0.72)
501 coef_vlc_table = 0;
502 else if (bps1 < 1.16)
503 coef_vlc_table = 1;
504 }
505
506 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
507 &coef_vlcs[coef_vlc_table * 2]);
508 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
509 &coef_vlcs[coef_vlc_table * 2 + 1]);
510 return 0;
511 }
512
513 /* interpolate values for a bigger or smaller block. The block must
514 have multiple sizes */
515 static void interpolate_array(float *scale, int old_size, int new_size)
516 {
517 int i, j, jincr, k;
518 float v;
519
520 if (new_size > old_size) {
521 jincr = new_size / old_size;
522 j = new_size;
523 for(i = old_size - 1; i >=0; i--) {
524 v = scale[i];
525 k = jincr;
526 do {
527 scale[--j] = v;
528 } while (--k);
529 }
530 } else if (new_size < old_size) {
531 j = 0;
532 jincr = old_size / new_size;
533 for(i = 0; i < new_size; i++) {
534 scale[i] = scale[j];
535 j += jincr;
536 }
537 }
538 }
539
540 /* compute x^-0.25 with an exponent and mantissa table. We use linear
541 interpolation to reduce the mantissa table size at a small speed
542 expense (linear interpolation approximately doubles the number of
543 bits of precision). */
544 static inline float pow_m1_4(WMADecodeContext *s, float x)
545 {
546 union {
547 float f;
548 unsigned int v;
549 } u, t;
550 unsigned int e, m;
551 float a, b;
552
553 u.f = x;
554 e = u.v >> 23;
555 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
556 /* build interpolation scale: 1 <= t < 2. */
557 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
558 a = s->lsp_pow_m_table1[m];
559 b = s->lsp_pow_m_table2[m];
560 return s->lsp_pow_e_table[e] * (a + b * t.f);
561 }
562
563 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
564 {
565 float wdel, a, b;
566 int i, e, m;
567
568 wdel = M_PI / frame_len;
569 for(i=0;i<frame_len;i++)
570 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
571
572 /* tables for x^-0.25 computation */
573 for(i=0;i<256;i++) {
574 e = i - 126;
575 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
576 }
577
578 /* NOTE: these two tables are needed to avoid two operations in
579 pow_m1_4 */
580 b = 1.0;
581 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
582 m = (1 << LSP_POW_BITS) + i;
583 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
584 a = pow(a, -0.25);
585 s->lsp_pow_m_table1[i] = 2 * a - b;
586 s->lsp_pow_m_table2[i] = b - a;
587 b = a;
588 }
589 #if 0
590 for(i=1;i<20;i++) {
591 float v, r1, r2;
592 v = 5.0 / i;
593 r1 = pow_m1_4(s, v);
594 r2 = pow(v,-0.25);
595 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
596 }
597 #endif
598 }
599
600 /* NOTE: We use the same code as Vorbis here */
601 /* XXX: optimize it further with SSE/3Dnow */
602 static void wma_lsp_to_curve(WMADecodeContext *s,
603 float *out, float *val_max_ptr,
604 int n, float *lsp)
605 {
606 int i, j;
607 float p, q, w, v, val_max;
608
609 val_max = 0;
610 for(i=0;i<n;i++) {
611 p = 0.5f;
612 q = 0.5f;
613 w = s->lsp_cos_table[i];
614 for(j=1;j<NB_LSP_COEFS;j+=2){
615 q *= w - lsp[j - 1];
616 p *= w - lsp[j];
617 }
618 p *= p * (2.0f - w);
619 q *= q * (2.0f + w);
620 v = p + q;
621 v = pow_m1_4(s, v);
622 if (v > val_max)
623 val_max = v;
624 out[i] = v;
625 }
626 *val_max_ptr = val_max;
627 }
628
629 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
630 static void decode_exp_lsp(WMADecodeContext *s, int ch)
631 {
632 float lsp_coefs[NB_LSP_COEFS];
633 int val, i;
634
635 for(i = 0; i < NB_LSP_COEFS; i++) {
636 if (i == 0 || i >= 8)
637 val = get_bits(&s->gb, 3);
638 else
639 val = get_bits(&s->gb, 4);
640 lsp_coefs[i] = lsp_codebook[i][val];
641 }
642
643 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
644 s->block_len, lsp_coefs);
645 }
646
647 /* decode exponents coded with VLC codes */
648 static int decode_exp_vlc(WMADecodeContext *s, int ch)
649 {
650 int last_exp, n, code;
651 const uint16_t *ptr, *band_ptr;
652 float v, *q, max_scale, *q_end;
653
654 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
655 ptr = band_ptr;
656 q = s->exponents[ch];
657 q_end = q + s->block_len;
658 max_scale = 0;
659 if (s->version == 1) {
660 last_exp = get_bits(&s->gb, 5) + 10;
661 /* XXX: use a table */
662 v = pow(10, last_exp * (1.0 / 16.0));
663 max_scale = v;
664 n = *ptr++;
665 do {
666 *q++ = v;
667 } while (--n);
668 }
669 last_exp = 36;
670 while (q < q_end) {
671 code = get_vlc(&s->gb, &s->exp_vlc);
672 if (code < 0)
673 return -1;
674 /* NOTE: this offset is the same as MPEG4 AAC ! */
675 last_exp += code - 60;
676 /* XXX: use a table */
677 v = pow(10, last_exp * (1.0 / 16.0));
678 if (v > max_scale)
679 max_scale = v;
680 n = *ptr++;
681 do {
682 *q++ = v;
683 } while (--n);
684 }
685 s->max_exponent[ch] = max_scale;
686 return 0;
687 }
688
689 /* return 0 if OK. return 1 if last block of frame. return -1 if
690 unrecorrable error. */
691 static int wma_decode_block(WMADecodeContext *s)
692 {
693 int n, v, a, ch, code, bsize;
694 int coef_nb_bits, total_gain, parse_exponents;
695 float window[BLOCK_MAX_SIZE * 2];
696 int nb_coefs[MAX_CHANNELS];
697 float mdct_norm;
698
699 #ifdef TRACE
700 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
701 #endif
702
703 /* compute current block length */
704 if (s->use_variable_block_len) {
705 n = av_log2(s->nb_block_sizes - 1) + 1;
706
707 if (s->reset_block_lengths) {
708 s->reset_block_lengths = 0;
709 v = get_bits(&s->gb, n);
710 if (v >= s->nb_block_sizes)
711 return -1;
712 s->prev_block_len_bits = s->frame_len_bits - v;
713 v = get_bits(&s->gb, n);
714 if (v >= s->nb_block_sizes)
715 return -1;
716 s->block_len_bits = s->frame_len_bits - v;
717 } else {
718 /* update block lengths */
719 s->prev_block_len_bits = s->block_len_bits;
720 s->block_len_bits = s->next_block_len_bits;
721 }
722 v = get_bits(&s->gb, n);
723 if (v >= s->nb_block_sizes)
724 return -1;
725 s->next_block_len_bits = s->frame_len_bits - v;
726 } else {
727 /* fixed block len */
728 s->next_block_len_bits = s->frame_len_bits;
729 s->prev_block_len_bits = s->frame_len_bits;
730 s->block_len_bits = s->frame_len_bits;
731 }
732
733 /* now check if the block length is coherent with the frame length */
734 s->block_len = 1 << s->block_len_bits;
735 if ((s->block_pos + s->block_len) > s->frame_len)
736 return -1;
737
738 if (s->nb_channels == 2) {
739 s->ms_stereo = get_bits(&s->gb, 1);
740 }
741 v = 0;
742 for(ch = 0; ch < s->nb_channels; ch++) {
743 a = get_bits(&s->gb, 1);
744 s->channel_coded[ch] = a;
745 v |= a;
746 }
747 /* if no channel coded, no need to go further */
748 /* XXX: fix potential framing problems */
749 if (!v)
750 goto next;
751
752 bsize = s->frame_len_bits - s->block_len_bits;
753
754 /* read total gain and extract corresponding number of bits for
755 coef escape coding */
756 total_gain = 1;
757 for(;;) {
758 a = get_bits(&s->gb, 7);
759 total_gain += a;
760 if (a != 127)
761 break;
762 }
763
764 if (total_gain < 15)
765 coef_nb_bits = 13;
766 else if (total_gain < 32)
767 coef_nb_bits = 12;
768 else if (total_gain < 40)
769 coef_nb_bits = 11;
770 else if (total_gain < 45)
771 coef_nb_bits = 10;
772 else
773 coef_nb_bits = 9;
774
775 /* compute number of coefficients */
776 n = s->coefs_end[bsize] - s->coefs_start;
777 for(ch = 0; ch < s->nb_channels; ch++)
778 nb_coefs[ch] = n;
779
780 /* complex coding */
781 if (s->use_noise_coding) {
782
783 for(ch = 0; ch < s->nb_channels; ch++) {
784 if (s->channel_coded[ch]) {
785 int i, n, a;
786 n = s->exponent_high_sizes[bsize];
787 for(i=0;i<n;i++) {
788 a = get_bits(&s->gb, 1);
789 s->high_band_coded[ch][i] = a;
790 /* if noise coding, the coefficients are not transmitted */
791 if (a)
792 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
793 }
794 }
795 }
796 for(ch = 0; ch < s->nb_channels; ch++) {
797 if (s->channel_coded[ch]) {
798 int i, n, val, code;
799
800 n = s->exponent_high_sizes[bsize];
801 val = (int)0x80000000;
802 for(i=0;i<n;i++) {
803 if (s->high_band_coded[ch][i]) {
804 if (val == (int)0x80000000) {
805 val = get_bits(&s->gb, 7) - 19;
806 } else {
807 code = get_vlc(&s->gb, &s->hgain_vlc);
808 if (code < 0)
809 return -1;
810 val += code - 18;
811 }
812 s->high_band_values[ch][i] = val;
813 }
814 }
815 }
816 }
817 }
818
819 /* exposant can be interpolated in short blocks. */
820 parse_exponents = 1;
821 if (s->block_len_bits != s->frame_len_bits) {
822 parse_exponents = get_bits(&s->gb, 1);
823 }
824
825 if (parse_exponents) {
826 for(ch = 0; ch < s->nb_channels; ch++) {
827 if (s->channel_coded[ch]) {
828 if (s->use_exp_vlc) {
829 if (decode_exp_vlc(s, ch) < 0)
830 return -1;
831 } else {
832 decode_exp_lsp(s, ch);
833 }
834 }
835 }
836 } else {
837 for(ch = 0; ch < s->nb_channels; ch++) {
838 if (s->channel_coded[ch]) {
839 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
840 s->block_len);
841 }
842 }
843 }
844
845 /* parse spectral coefficients : just RLE encoding */
846 for(ch = 0; ch < s->nb_channels; ch++) {
847 if (s->channel_coded[ch]) {
848 VLC *coef_vlc;
849 int level, run, sign, tindex;
850 int16_t *ptr, *eptr;
851 const int16_t *level_table, *run_table;
852
853 /* special VLC tables are used for ms stereo because
854 there is potentially less energy there */
855 tindex = (ch == 1 && s->ms_stereo);
856 coef_vlc = &s->coef_vlc[tindex];
857 run_table = s->run_table[tindex];
858 level_table = s->level_table[tindex];
859 /* XXX: optimize */
860 ptr = &s->coefs1[ch][0];
861 eptr = ptr + nb_coefs[ch];
862 memset(ptr, 0, s->block_len * sizeof(int16_t));
863 for(;;) {
864 code = get_vlc(&s->gb, coef_vlc);
865 if (code < 0)
866 return -1;
867 if (code == 1) {
868 /* EOB */
869 break;
870 } else if (code == 0) {
871 /* escape */
872 level = get_bits(&s->gb, coef_nb_bits);
873 /* NOTE: this is rather suboptimal. reading
874 block_len_bits would be better */
875 run = get_bits(&s->gb, s->frame_len_bits);
876 } else {
877 /* normal code */
878 run = run_table[code];
879 level = level_table[code];
880 }
881 sign = get_bits(&s->gb, 1);
882 if (!sign)
883 level = -level;
884 ptr += run;
885 if (ptr >= eptr)
886 return -1;
887 *ptr++ = level;
888 /* NOTE: EOB can be omitted */
889 if (ptr >= eptr)
890 break;
891 }
892 }
893 if (s->version == 1 && s->nb_channels >= 2) {
894 align_get_bits(&s->gb);
895 }
896 }
897
898 /* normalize */
899 {
900 int n4 = s->block_len / 2;
901 mdct_norm = 1.0 / (float)n4;
902 if (s->version == 1) {
903 mdct_norm *= sqrt(n4);
904 }
905 }
906
907 /* finally compute the MDCT coefficients */
908 for(ch = 0; ch < s->nb_channels; ch++) {
909 if (s->channel_coded[ch]) {
910 int16_t *coefs1;
911 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
912 int i, j, n, n1, last_high_band;
913 float exp_power[HIGH_BAND_MAX_SIZE];
914
915 coefs1 = s->coefs1[ch];
916 exponents = s->exponents[ch];
917 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
918 mult *= mdct_norm;
919 coefs = s->coefs[ch];
920 if (s->use_noise_coding) {
921 mult1 = mult;
922 /* very low freqs : noise */
923 for(i = 0;i < s->coefs_start; i++) {
924 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
925 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
926 }
927
928 n1 = s->exponent_high_sizes[bsize];
929
930 /* compute power of high bands */
931 exp_ptr = exponents +
932 s->high_band_start[bsize] -
933 s->coefs_start;
934 last_high_band = 0; /* avoid warning */
935 for(j=0;j<n1;j++) {
936 n = s->exponent_high_bands[s->frame_len_bits -
937 s->block_len_bits][j];
938 if (s->high_band_coded[ch][j]) {
939 float e2, v;
940 e2 = 0;
941 for(i = 0;i < n; i++) {
942 v = exp_ptr[i];
943 e2 += v * v;
944 }
945 exp_power[j] = e2 / n;
946 last_high_band = j;
947 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
948 }
949 exp_ptr += n;
950 }
951
952 /* main freqs and high freqs */
953 for(j=-1;j<n1;j++) {
954 if (j < 0) {
955 n = s->high_band_start[bsize] -
956 s->coefs_start;
957 } else {
958 n = s->exponent_high_bands[s->frame_len_bits -
959 s->block_len_bits][j];
960 }
961 if (j >= 0 && s->high_band_coded[ch][j]) {
962 /* use noise with specified power */
963 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
964 /* XXX: use a table */
965 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
966 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
967 mult1 *= mdct_norm;
968 for(i = 0;i < n; i++) {
969 noise = s->noise_table[s->noise_index];
970 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
971 *coefs++ = (*exponents++) * noise * mult1;
972 }
973 } else {
974 /* coded values + small noise */
975 for(i = 0;i < n; i++) {
976 noise = s->noise_table[s->noise_index];
977 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
978 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
979 }
980 }
981 }
982
983 /* very high freqs : noise */
984 n = s->block_len - s->coefs_end[bsize];
985 mult1 = mult * exponents[-1];
986 for(i = 0; i < n; i++) {
987 *coefs++ = s->noise_table[s->noise_index] * mult1;
988 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
989 }
990 } else {
991 /* XXX: optimize more */
992 for(i = 0;i < s->coefs_start; i++)
993 *coefs++ = 0.0;
994 n = nb_coefs[ch];
995 for(i = 0;i < n; i++) {
996 *coefs++ = coefs1[i] * exponents[i] * mult;
997 }
998 n = s->block_len - s->coefs_end[bsize];
999 for(i = 0;i < n; i++)
1000 *coefs++ = 0.0;
1001 }
1002 }
1003 }
1004
1005 #ifdef TRACE
1006 for(ch = 0; ch < s->nb_channels; ch++) {
1007 if (s->channel_coded[ch]) {
1008 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1009 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1010 }
1011 }
1012 #endif
1013
1014 if (s->ms_stereo && s->channel_coded[1]) {
1015 float a, b;
1016 int i;
1017
1018 /* nominal case for ms stereo: we do it before mdct */
1019 /* no need to optimize this case because it should almost
1020 never happen */
1021 if (!s->channel_coded[0]) {
1022 tprintf("rare ms-stereo case happened\n");
1023 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1024 s->channel_coded[0] = 1;
1025 }
1026
1027 for(i = 0; i < s->block_len; i++) {
1028 a = s->coefs[0][i];
1029 b = s->coefs[1][i];
1030 s->coefs[0][i] = a + b;
1031 s->coefs[1][i] = a - b;
1032 }
1033 }
1034
1035 /* build the window : we ensure that when the windows overlap
1036 their squared sum is always 1 (MDCT reconstruction rule) */
1037 /* XXX: merge with output */
1038 {
1039 int i, next_block_len, block_len, prev_block_len, n;
1040 float *wptr;
1041
1042 block_len = s->block_len;
1043 prev_block_len = 1 << s->prev_block_len_bits;
1044 next_block_len = 1 << s->next_block_len_bits;
1045
1046 /* right part */
1047 wptr = window + block_len;
1048 if (block_len <= next_block_len) {
1049 for(i=0;i<block_len;i++)
1050 *wptr++ = s->windows[bsize][i];
1051 } else {
1052 /* overlap */
1053 n = (block_len / 2) - (next_block_len / 2);
1054 for(i=0;i<n;i++)
1055 *wptr++ = 1.0;
1056 for(i=0;i<next_block_len;i++)
1057 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1058 for(i=0;i<n;i++)
1059 *wptr++ = 0.0;
1060 }
1061
1062 /* left part */
1063 wptr = window + block_len;
1064 if (block_len <= prev_block_len) {
1065 for(i=0;i<block_len;i++)
1066 *--wptr = s->windows[bsize][i];
1067 } else {
1068 /* overlap */
1069 n = (block_len / 2) - (prev_block_len / 2);
1070 for(i=0;i<n;i++)
1071 *--wptr = 1.0;
1072 for(i=0;i<prev_block_len;i++)
1073 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1074 for(i=0;i<n;i++)
1075 *--wptr = 0.0;
1076 }
1077 }
1078
1079
1080 for(ch = 0; ch < s->nb_channels; ch++) {
1081 if (s->channel_coded[ch]) {
1082 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1083 float *ptr;
1084 int i, n4, index, n;
1085
1086 n = s->block_len;
1087 n4 = s->block_len / 2;
1088 ff_imdct_calc(&s->mdct_ctx[bsize],
1089 output, s->coefs[ch], s->mdct_tmp);
1090
1091 /* XXX: optimize all that by build the window and
1092 multipying/adding at the same time */
1093 /* multiply by the window */
1094 for(i=0;i<n * 2;i++) {
1095 output[i] *= window[i];
1096 }
1097
1098 /* add in the frame */
1099 index = (s->frame_len / 2) + s->block_pos - n4;
1100 ptr = &s->frame_out[ch][index];
1101 for(i=0;i<n * 2;i++) {
1102 *ptr += output[i];
1103 ptr++;
1104 }
1105
1106 /* specific fast case for ms-stereo : add to second
1107 channel if it is not coded */
1108 if (s->ms_stereo && !s->channel_coded[1]) {
1109 ptr = &s->frame_out[1][index];
1110 for(i=0;i<n * 2;i++) {
1111 *ptr += output[i];
1112 ptr++;
1113 }
1114 }
1115 }
1116 }
1117 next:
1118 /* update block number */
1119 s->block_num++;
1120 s->block_pos += s->block_len;
1121 if (s->block_pos >= s->frame_len)
1122 return 1;
1123 else
1124 return 0;
1125 }
1126
1127 /* decode a frame of frame_len samples */
1128 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1129 {
1130 int ret, i, n, a, ch, incr;
1131 int16_t *ptr;
1132 float *iptr;
1133
1134 #ifdef TRACE
1135 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
1136 #endif
1137
1138 /* read each block */
1139 s->block_num = 0;
1140 s->block_pos = 0;
1141 for(;;) {
1142 ret = wma_decode_block(s);
1143 if (ret < 0)
1144 return -1;
1145 if (ret)
1146 break;
1147 }
1148
1149 /* convert frame to integer */
1150 n = s->frame_len;
1151 incr = s->nb_channels;
1152 for(ch = 0; ch < s->nb_channels; ch++) {
1153 ptr = samples + ch;
1154 iptr = s->frame_out[ch];
1155
1156 for(i=0;i<n;i++) {
1157 a = lrintf(*iptr++);
1158 if (a > 32767)
1159 a = 32767;
1160 else if (a < -32768)
1161 a = -32768;
1162 *ptr = a;
1163 ptr += incr;
1164 }
1165 /* prepare for next block */
1166 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1167 s->frame_len * sizeof(float));
1168 /* XXX: suppress this */
1169 memset(&s->frame_out[ch][s->frame_len], 0,
1170 s->frame_len * sizeof(float));
1171 }
1172
1173 #ifdef TRACE
1174 dump_shorts("samples", samples, n * s->nb_channels);
1175 #endif
1176 return 0;
1177 }
1178
1179 static int wma_decode_superframe(AVCodecContext *avctx,
1180 void *data, int *data_size,
1181 uint8_t *buf, int buf_size)
1182 {
1183 WMADecodeContext *s = avctx->priv_data;
1184 int nb_frames, bit_offset, i, pos, len;
1185 uint8_t *q;
1186 int16_t *samples;
1187
1188 tprintf("***decode_superframe:\n");
1189
1190 if(buf_size==0){
1191 s->last_superframe_len = 0;
1192 return 0;
1193 }
1194
1195 samples = data;
1196
1197 init_get_bits(&s->gb, buf, buf_size*8);
1198
1199 if (s->use_bit_reservoir) {
1200 /* read super frame header */
1201 get_bits(&s->gb, 4); /* super frame index */
1202 nb_frames = get_bits(&s->gb, 4) - 1;
1203
1204 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1205
1206 if (s->last_superframe_len > 0) {
1207 // printf("skip=%d\n", s->last_bitoffset);
1208 /* add bit_offset bits to last frame */
1209 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1210 MAX_CODED_SUPERFRAME_SIZE)
1211 goto fail;
1212 q = s->last_superframe + s->last_superframe_len;
1213 len = bit_offset;
1214 while (len > 0) {
1215 *q++ = (get_bits)(&s->gb, 8);
1216 len -= 8;
1217 }
1218 if (len > 0) {
1219 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1220 }
1221
1222 /* XXX: bit_offset bits into last frame */
1223 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1224 /* skip unused bits */
1225 if (s->last_bitoffset > 0)
1226 skip_bits(&s->gb, s->last_bitoffset);
1227 /* this frame is stored in the last superframe and in the
1228 current one */
1229 if (wma_decode_frame(s, samples) < 0)
1230 goto fail;
1231 samples += s->nb_channels * s->frame_len;
1232 }
1233
1234 /* read each frame starting from bit_offset */
1235 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1236 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1237 len = pos & 7;
1238 if (len > 0)
1239 skip_bits(&s->gb, len);
1240
1241 s->reset_block_lengths = 1;
1242 for(i=0;i<nb_frames;i++) {
1243 if (wma_decode_frame(s, samples) < 0)
1244 goto fail;
1245 samples += s->nb_channels * s->frame_len;
1246 }
1247
1248 /* we copy the end of the frame in the last frame buffer */
1249 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1250 s->last_bitoffset = pos & 7;
1251 pos >>= 3;
1252 len = buf_size - pos;
1253 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1254 goto fail;
1255 }
1256 s->last_superframe_len = len;
1257 memcpy(s->last_superframe, buf + pos, len);
1258 } else {
1259 /* single frame decode */
1260 if (wma_decode_frame(s, samples) < 0)
1261 goto fail;
1262 samples += s->nb_channels * s->frame_len;
1263 }
1264 *data_size = (int8_t *)samples - (int8_t *)data;
1265 return s->block_align;
1266 fail:
1267 /* when error, we reset the bit reservoir */
1268 s->last_superframe_len = 0;
1269 return -1;
1270 }
1271
1272 static int wma_decode_end(AVCodecContext *avctx)
1273 {
1274 WMADecodeContext *s = avctx->priv_data;
1275 int i;
1276
1277 for(i = 0; i < s->nb_block_sizes; i++)
1278 ff_mdct_end(&s->mdct_ctx[i]);
1279 for(i = 0; i < s->nb_block_sizes; i++)
1280 av_free(s->windows[i]);
1281
1282 if (s->use_exp_vlc) {
1283 free_vlc(&s->exp_vlc);
1284 }
1285 if (s->use_noise_coding) {
1286 free_vlc(&s->hgain_vlc);
1287 }
1288 for(i = 0;i < 2; i++) {
1289 free_vlc(&s->coef_vlc[i]);
1290 av_free(s->run_table[i]);
1291 av_free(s->level_table[i]);
1292 }
1293
1294 return 0;
1295 }
1296
1297 AVCodec wmav1_decoder =
1298 {
1299 "wmav1",
1300 CODEC_TYPE_AUDIO,
1301 CODEC_ID_WMAV1,
1302 sizeof(WMADecodeContext),
1303 wma_decode_init,
1304 NULL,
1305 wma_decode_end,
1306 wma_decode_superframe,
1307 };
1308
1309 AVCodec wmav2_decoder =
1310 {
1311 "wmav2",
1312 CODEC_TYPE_AUDIO,
1313 CODEC_ID_WMAV2,
1314 sizeof(WMADecodeContext),
1315 wma_decode_init,
1316 NULL,
1317 wma_decode_end,
1318 wma_decode_superframe,
1319 };
Libav master