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 16384
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
dump_shorts(const char * name,const short * tab,int n)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
dump_floats(const char * name,int prec,const float * tab,int n)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 */
init_coef_vlc(VLC * vlc,uint16_t ** prun_table,uint16_t ** plevel_table,const CoefVLCTable * vlc_table)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
wma_decode_init(AVCodecContext * avctx)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
207 s->sample_rate = avctx->sample_rate;
208 s->nb_channels = avctx->channels;
209 s->bit_rate = avctx->bit_rate;
210 s->block_align = avctx->block_align;
211
212 if (avctx->codec->id == CODEC_ID_WMAV1) {
213 s->version = 1;
214 } else {
215 s->version = 2;
216 }
217
218 /* extract flag infos */
219 flags1 = 0;
220 flags2 = 0;
221 extradata = avctx->extradata;
222 if (s->version == 1 && avctx->extradata_size >= 4) {
223 flags1 = extradata[0] | (extradata[1] << 8);
224 flags2 = extradata[2] | (extradata[3] << 8);
225 } else if (s->version == 2 && avctx->extradata_size >= 6) {
226 flags1 = extradata[0] | (extradata[1] << 8) |
227 (extradata[2] << 16) | (extradata[3] << 24);
228 flags2 = extradata[4] | (extradata[5] << 8);
229 }
230 s->use_exp_vlc = flags2 & 0x0001;
231 s->use_bit_reservoir = flags2 & 0x0002;
232 s->use_variable_block_len = flags2 & 0x0004;
233
234 /* compute MDCT block size */
235 if (s->sample_rate <= 16000) {
236 s->frame_len_bits = 9;
237 } else if (s->sample_rate <= 22050 ||
238 (s->sample_rate <= 32000 && s->version == 1)) {
239 s->frame_len_bits = 10;
240 } else {
241 s->frame_len_bits = 11;
242 }
243 s->frame_len = 1 << s->frame_len_bits;
244 if (s->use_variable_block_len) {
245 int nb_max, nb;
246 nb = ((flags2 >> 3) & 3) + 1;
247 if ((s->bit_rate / s->nb_channels) >= 32000)
248 nb += 2;
249 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
250 if (nb > nb_max)
251 nb = nb_max;
252 s->nb_block_sizes = nb + 1;
253 } else {
254 s->nb_block_sizes = 1;
255 }
256
257 /* init rate dependant parameters */
258 s->use_noise_coding = 1;
259 high_freq = s->sample_rate * 0.5;
260
261 /* if version 2, then the rates are normalized */
262 sample_rate1 = s->sample_rate;
263 if (s->version == 2) {
264 if (sample_rate1 >= 44100)
265 sample_rate1 = 44100;
266 else if (sample_rate1 >= 22050)
267 sample_rate1 = 22050;
268 else if (sample_rate1 >= 16000)
269 sample_rate1 = 16000;
270 else if (sample_rate1 >= 11025)
271 sample_rate1 = 11025;
272 else if (sample_rate1 >= 8000)
273 sample_rate1 = 8000;
274 }
275
276 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
277 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
278
279 /* compute high frequency value and choose if noise coding should
280 be activated */
281 bps1 = bps;
282 if (s->nb_channels == 2)
283 bps1 = bps * 1.6;
284 if (sample_rate1 == 44100) {
285 if (bps1 >= 0.61)
286 s->use_noise_coding = 0;
287 else
288 high_freq = high_freq * 0.4;
289 } else if (sample_rate1 == 22050) {
290 if (bps1 >= 1.16)
291 s->use_noise_coding = 0;
292 else if (bps1 >= 0.72)
293 high_freq = high_freq * 0.7;
294 else
295 high_freq = high_freq * 0.6;
296 } else if (sample_rate1 == 16000) {
297 if (bps > 0.5)
298 high_freq = high_freq * 0.5;
299 else
300 high_freq = high_freq * 0.3;
301 } else if (sample_rate1 == 11025) {
302 high_freq = high_freq * 0.7;
303 } else if (sample_rate1 == 8000) {
304 if (bps <= 0.625) {
305 high_freq = high_freq * 0.5;
306 } else if (bps > 0.75) {
307 s->use_noise_coding = 0;
308 } else {
309 high_freq = high_freq * 0.65;
310 }
311 } else {
312 if (bps >= 0.8) {
313 high_freq = high_freq * 0.75;
314 } else if (bps >= 0.6) {
315 high_freq = high_freq * 0.6;
316 } else {
317 high_freq = high_freq * 0.5;
318 }
319 }
320 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
321 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
322 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
323 s->block_align);
324 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
325 bps, bps1, high_freq, s->byte_offset_bits);
326 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
327 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
328
329 /* compute the scale factor band sizes for each MDCT block size */
330 {
331 int a, b, pos, lpos, k, block_len, i, j, n;
332 const uint8_t *table;
333
334 if (s->version == 1) {
335 s->coefs_start = 3;
336 } else {
337 s->coefs_start = 0;
338 }
339 for(k = 0; k < s->nb_block_sizes; k++) {
340 block_len = s->frame_len >> k;
341
342 if (s->version == 1) {
343 lpos = 0;
344 for(i=0;i<25;i++) {
345 a = wma_critical_freqs[i];
346 b = s->sample_rate;
347 pos = ((block_len * 2 * a) + (b >> 1)) / b;
348 if (pos > block_len)
349 pos = block_len;
350 s->exponent_bands[0][i] = pos - lpos;
351 if (pos >= block_len) {
352 i++;
353 break;
354 }
355 lpos = pos;
356 }
357 s->exponent_sizes[0] = i;
358 } else {
359 /* hardcoded tables */
360 table = NULL;
361 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
362 if (a < 3) {
363 if (s->sample_rate >= 44100)
364 table = exponent_band_44100[a];
365 else if (s->sample_rate >= 32000)
366 table = exponent_band_32000[a];
367 else if (s->sample_rate >= 22050)
368 table = exponent_band_22050[a];
369 }
370 if (table) {
371 n = *table++;
372 for(i=0;i<n;i++)
373 s->exponent_bands[k][i] = table[i];
374 s->exponent_sizes[k] = n;
375 } else {
376 j = 0;
377 lpos = 0;
378 for(i=0;i<25;i++) {
379 a = wma_critical_freqs[i];
380 b = s->sample_rate;
381 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
382 pos <<= 2;
383 if (pos > block_len)
384 pos = block_len;
385 if (pos > lpos)
386 s->exponent_bands[k][j++] = pos - lpos;
387 if (pos >= block_len)
388 break;
389 lpos = pos;
390 }
391 s->exponent_sizes[k] = j;
392 }
393 }
394
395 /* max number of coefs */
396 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
397 /* high freq computation */
398 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
399 s->sample_rate + 0.5);
400 n = s->exponent_sizes[k];
401 j = 0;
402 pos = 0;
403 for(i=0;i<n;i++) {
404 int start, end;
405 start = pos;
406 pos += s->exponent_bands[k][i];
407 end = pos;
408 if (start < s->high_band_start[k])
409 start = s->high_band_start[k];
410 if (end > s->coefs_end[k])
411 end = s->coefs_end[k];
412 if (end > start)
413 s->exponent_high_bands[k][j++] = end - start;
414 }
415 s->exponent_high_sizes[k] = j;
416 #if 0
417 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
418 s->frame_len >> k,
419 s->coefs_end[k],
420 s->high_band_start[k],
421 s->exponent_high_sizes[k]);
422 for(j=0;j<s->exponent_high_sizes[k];j++)
423 tprintf(" %d", s->exponent_high_bands[k][j]);
424 tprintf("\n");
425 #endif
426 }
427 }
428
429 #ifdef TRACE
430 {
431 int i, j;
432 for(i = 0; i < s->nb_block_sizes; i++) {
433 tprintf("%5d: n=%2d:",
434 s->frame_len >> i,
435 s->exponent_sizes[i]);
436 for(j=0;j<s->exponent_sizes[i];j++)
437 tprintf(" %d", s->exponent_bands[i][j]);
438 tprintf("\n");
439 }
440 }
441 #endif
442
443 /* init MDCT */
444 for(i = 0; i < s->nb_block_sizes; i++)
445 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
446
447 /* init MDCT windows : simple sinus window */
448 for(i = 0; i < s->nb_block_sizes; i++) {
449 int n, j;
450 float alpha;
451 n = 1 << (s->frame_len_bits - i);
452 window = av_malloc(sizeof(float) * n);
453 alpha = M_PI / (2.0 * n);
454 for(j=0;j<n;j++) {
455 window[n - j - 1] = sin((j + 0.5) * alpha);
456 }
457 s->windows[i] = window;
458 }
459
460 s->reset_block_lengths = 1;
461
462 if (s->use_noise_coding) {
463
464 /* init the noise generator */
465 if (s->use_exp_vlc)
466 s->noise_mult = 0.02;
467 else
468 s->noise_mult = 0.04;
469
470 #ifdef TRACE
471 for(i=0;i<NOISE_TAB_SIZE;i++)
472 s->noise_table[i] = 1.0 * s->noise_mult;
473 #else
474 {
475 unsigned int seed;
476 float norm;
477 seed = 1;
478 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
479 for(i=0;i<NOISE_TAB_SIZE;i++) {
480 seed = seed * 314159 + 1;
481 s->noise_table[i] = (float)((int)seed) * norm;
482 }
483 }
484 #endif
485 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
486 hgain_huffbits, 1, 1,
487 hgain_huffcodes, 2, 2);
488 }
489
490 if (s->use_exp_vlc) {
491 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
492 scale_huffbits, 1, 1,
493 scale_huffcodes, 4, 4);
494 } else {
495 wma_lsp_to_curve_init(s, s->frame_len);
496 }
497
498 /* choose the VLC tables for the coefficients */
499 coef_vlc_table = 2;
500 if (s->sample_rate >= 32000) {
501 if (bps1 < 0.72)
502 coef_vlc_table = 0;
503 else if (bps1 < 1.16)
504 coef_vlc_table = 1;
505 }
506
507 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
508 &coef_vlcs[coef_vlc_table * 2]);
509 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
510 &coef_vlcs[coef_vlc_table * 2 + 1]);
511 return 0;
512 }
513
514 /* interpolate values for a bigger or smaller block. The block must
515 have multiple sizes */
interpolate_array(float * scale,int old_size,int new_size)516 static void interpolate_array(float *scale, int old_size, int new_size)
517 {
518 int i, j, jincr, k;
519 float v;
520
521 if (new_size > old_size) {
522 jincr = new_size / old_size;
523 j = new_size;
524 for(i = old_size - 1; i >=0; i--) {
525 v = scale[i];
526 k = jincr;
527 do {
528 scale[--j] = v;
529 } while (--k);
530 }
531 } else if (new_size < old_size) {
532 j = 0;
533 jincr = old_size / new_size;
534 for(i = 0; i < new_size; i++) {
535 scale[i] = scale[j];
536 j += jincr;
537 }
538 }
539 }
540
541 /* compute x^-0.25 with an exponent and mantissa table. We use linear
542 interpolation to reduce the mantissa table size at a small speed
543 expense (linear interpolation approximately doubles the number of
544 bits of precision). */
pow_m1_4(WMADecodeContext * s,float x)545 static inline float pow_m1_4(WMADecodeContext *s, float x)
546 {
547 union {
548 float f;
549 unsigned int v;
550 } u, t;
551 unsigned int e, m;
552 float a, b;
553
554 u.f = x;
555 e = u.v >> 23;
556 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
557 /* build interpolation scale: 1 <= t < 2. */
558 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
559 a = s->lsp_pow_m_table1[m];
560 b = s->lsp_pow_m_table2[m];
561 return s->lsp_pow_e_table[e] * (a + b * t.f);
562 }
563
wma_lsp_to_curve_init(WMADecodeContext * s,int frame_len)564 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
565 {
566 float wdel, a, b;
567 int i, e, m;
568
569 wdel = M_PI / frame_len;
570 for(i=0;i<frame_len;i++)
571 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
572
573 /* tables for x^-0.25 computation */
574 for(i=0;i<256;i++) {
575 e = i - 126;
576 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
577 }
578
579 /* NOTE: these two tables are needed to avoid two operations in
580 pow_m1_4 */
581 b = 1.0;
582 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
583 m = (1 << LSP_POW_BITS) + i;
584 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
585 a = pow(a, -0.25);
586 s->lsp_pow_m_table1[i] = 2 * a - b;
587 s->lsp_pow_m_table2[i] = b - a;
588 b = a;
589 }
590 #if 0
591 for(i=1;i<20;i++) {
592 float v, r1, r2;
593 v = 5.0 / i;
594 r1 = pow_m1_4(s, v);
595 r2 = pow(v,-0.25);
596 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
597 }
598 #endif
599 }
600
601 /* NOTE: We use the same code as Vorbis here */
602 /* XXX: optimize it further with SSE/3Dnow */
wma_lsp_to_curve(WMADecodeContext * s,float * out,float * val_max_ptr,int n,float * lsp)603 static void wma_lsp_to_curve(WMADecodeContext *s,
604 float *out, float *val_max_ptr,
605 int n, float *lsp)
606 {
607 int i, j;
608 float p, q, w, v, val_max;
609
610 val_max = 0;
611 for(i=0;i<n;i++) {
612 p = 0.5f;
613 q = 0.5f;
614 w = s->lsp_cos_table[i];
615 for(j=1;j<NB_LSP_COEFS;j+=2){
616 q *= w - lsp[j - 1];
617 p *= w - lsp[j];
618 }
619 p *= p * (2.0f - w);
620 q *= q * (2.0f + w);
621 v = p + q;
622 v = pow_m1_4(s, v);
623 if (v > val_max)
624 val_max = v;
625 out[i] = v;
626 }
627 *val_max_ptr = val_max;
628 }
629
630 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
decode_exp_lsp(WMADecodeContext * s,int ch)631 static void decode_exp_lsp(WMADecodeContext *s, int ch)
632 {
633 float lsp_coefs[NB_LSP_COEFS];
634 int val, i;
635
636 for(i = 0; i < NB_LSP_COEFS; i++) {
637 if (i == 0 || i >= 8)
638 val = get_bits(&s->gb, 3);
639 else
640 val = get_bits(&s->gb, 4);
641 lsp_coefs[i] = lsp_codebook[i][val];
642 }
643
644 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
645 s->block_len, lsp_coefs);
646 }
647
648 /* decode exponents coded with VLC codes */
decode_exp_vlc(WMADecodeContext * s,int ch)649 static int decode_exp_vlc(WMADecodeContext *s, int ch)
650 {
651 int last_exp, n, code;
652 const uint16_t *ptr, *band_ptr;
653 float v, *q, max_scale, *q_end;
654
655 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
656 ptr = band_ptr;
657 q = s->exponents[ch];
658 q_end = q + s->block_len;
659 max_scale = 0;
660 if (s->version == 1) {
661 last_exp = get_bits(&s->gb, 5) + 10;
662 /* XXX: use a table */
663 v = pow(10, last_exp * (1.0 / 16.0));
664 max_scale = v;
665 n = *ptr++;
666 do {
667 *q++ = v;
668 } while (--n);
669 }
670 last_exp = 36;
671 while (q < q_end) {
672 code = get_vlc(&s->gb, &s->exp_vlc);
673 if (code < 0)
674 return -1;
675 /* NOTE: this offset is the same as MPEG4 AAC ! */
676 last_exp += code - 60;
677 /* XXX: use a table */
678 v = pow(10, last_exp * (1.0 / 16.0));
679 if (v > max_scale)
680 max_scale = v;
681 n = *ptr++;
682 do {
683 *q++ = v;
684 } while (--n);
685 }
686 s->max_exponent[ch] = max_scale;
687 return 0;
688 }
689
690 /* return 0 if OK. return 1 if last block of frame. return -1 if
691 unrecorrable error. */
wma_decode_block(WMADecodeContext * s)692 static int wma_decode_block(WMADecodeContext *s)
693 {
694 int n, v, a, ch, code, bsize;
695 int coef_nb_bits, total_gain, parse_exponents;
696 float window[BLOCK_MAX_SIZE * 2];
697 int nb_coefs[MAX_CHANNELS];
698 float mdct_norm;
699
700 #ifdef TRACE
701 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
702 #endif
703
704 /* compute current block length */
705 if (s->use_variable_block_len) {
706 n = av_log2(s->nb_block_sizes - 1) + 1;
707
708 if (s->reset_block_lengths) {
709 s->reset_block_lengths = 0;
710 v = get_bits(&s->gb, n);
711 if (v >= s->nb_block_sizes)
712 return -1;
713 s->prev_block_len_bits = s->frame_len_bits - v;
714 v = get_bits(&s->gb, n);
715 if (v >= s->nb_block_sizes)
716 return -1;
717 s->block_len_bits = s->frame_len_bits - v;
718 } else {
719 /* update block lengths */
720 s->prev_block_len_bits = s->block_len_bits;
721 s->block_len_bits = s->next_block_len_bits;
722 }
723 v = get_bits(&s->gb, n);
724 if (v >= s->nb_block_sizes)
725 return -1;
726 s->next_block_len_bits = s->frame_len_bits - v;
727 } else {
728 /* fixed block len */
729 s->next_block_len_bits = s->frame_len_bits;
730 s->prev_block_len_bits = s->frame_len_bits;
731 s->block_len_bits = s->frame_len_bits;
732 }
733
734 /* now check if the block length is coherent with the frame length */
735 s->block_len = 1 << s->block_len_bits;
736 if ((s->block_pos + s->block_len) > s->frame_len)
737 return -1;
738
739 if (s->nb_channels == 2) {
740 s->ms_stereo = get_bits(&s->gb, 1);
741 }
742 v = 0;
743 for(ch = 0; ch < s->nb_channels; ch++) {
744 a = get_bits(&s->gb, 1);
745 s->channel_coded[ch] = a;
746 v |= a;
747 }
748 /* if no channel coded, no need to go further */
749 /* XXX: fix potential framing problems */
750 if (!v)
751 goto next;
752
753 bsize = s->frame_len_bits - s->block_len_bits;
754
755 /* read total gain and extract corresponding number of bits for
756 coef escape coding */
757 total_gain = 1;
758 for(;;) {
759 a = get_bits(&s->gb, 7);
760 total_gain += a;
761 if (a != 127)
762 break;
763 }
764
765 if (total_gain < 15)
766 coef_nb_bits = 13;
767 else if (total_gain < 32)
768 coef_nb_bits = 12;
769 else if (total_gain < 40)
770 coef_nb_bits = 11;
771 else if (total_gain < 45)
772 coef_nb_bits = 10;
773 else
774 coef_nb_bits = 9;
775
776 /* compute number of coefficients */
777 n = s->coefs_end[bsize] - s->coefs_start;
778 for(ch = 0; ch < s->nb_channels; ch++)
779 nb_coefs[ch] = n;
780
781 /* complex coding */
782 if (s->use_noise_coding) {
783
784 for(ch = 0; ch < s->nb_channels; ch++) {
785 if (s->channel_coded[ch]) {
786 int i, n, a;
787 n = s->exponent_high_sizes[bsize];
788 for(i=0;i<n;i++) {
789 a = get_bits(&s->gb, 1);
790 s->high_band_coded[ch][i] = a;
791 /* if noise coding, the coefficients are not transmitted */
792 if (a)
793 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
794 }
795 }
796 }
797 for(ch = 0; ch < s->nb_channels; ch++) {
798 if (s->channel_coded[ch]) {
799 int i, n, val, code;
800
801 n = s->exponent_high_sizes[bsize];
802 val = (int)0x80000000;
803 for(i=0;i<n;i++) {
804 if (s->high_band_coded[ch][i]) {
805 if (val == (int)0x80000000) {
806 val = get_bits(&s->gb, 7) - 19;
807 } else {
808 code = get_vlc(&s->gb, &s->hgain_vlc);
809 if (code < 0)
810 return -1;
811 val += code - 18;
812 }
813 s->high_band_values[ch][i] = val;
814 }
815 }
816 }
817 }
818 }
819
820 /* exposant can be interpolated in short blocks. */
821 parse_exponents = 1;
822 if (s->block_len_bits != s->frame_len_bits) {
823 parse_exponents = get_bits(&s->gb, 1);
824 }
825
826 if (parse_exponents) {
827 for(ch = 0; ch < s->nb_channels; ch++) {
828 if (s->channel_coded[ch]) {
829 if (s->use_exp_vlc) {
830 if (decode_exp_vlc(s, ch) < 0)
831 return -1;
832 } else {
833 decode_exp_lsp(s, ch);
834 }
835 }
836 }
837 } else {
838 for(ch = 0; ch < s->nb_channels; ch++) {
839 if (s->channel_coded[ch]) {
840 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
841 s->block_len);
842 }
843 }
844 }
845
846 /* parse spectral coefficients : just RLE encoding */
847 for(ch = 0; ch < s->nb_channels; ch++) {
848 if (s->channel_coded[ch]) {
849 VLC *coef_vlc;
850 int level, run, sign, tindex;
851 int16_t *ptr, *eptr;
852 const uint16_t *level_table, *run_table;
853
854 /* special VLC tables are used for ms stereo because
855 there is potentially less energy there */
856 tindex = (ch == 1 && s->ms_stereo);
857 coef_vlc = &s->coef_vlc[tindex];
858 run_table = s->run_table[tindex];
859 level_table = s->level_table[tindex];
860 /* XXX: optimize */
861 ptr = &s->coefs1[ch][0];
862 eptr = ptr + nb_coefs[ch];
863 memset(ptr, 0, s->block_len * sizeof(int16_t));
864 for(;;) {
865 code = get_vlc(&s->gb, coef_vlc);
866 if (code < 0)
867 return -1;
868 if (code == 1) {
869 /* EOB */
870 break;
871 } else if (code == 0) {
872 /* escape */
873 level = get_bits(&s->gb, coef_nb_bits);
874 /* NOTE: this is rather suboptimal. reading
875 block_len_bits would be better */
876 run = get_bits(&s->gb, s->frame_len_bits);
877 } else {
878 /* normal code */
879 run = run_table[code];
880 level = level_table[code];
881 }
882 sign = get_bits(&s->gb, 1);
883 if (!sign)
884 level = -level;
885 ptr += run;
886 if (ptr >= eptr)
887 return -1;
888 *ptr++ = level;
889 /* NOTE: EOB can be omitted */
890 if (ptr >= eptr)
891 break;
892 }
893 }
894 if (s->version == 1 && s->nb_channels >= 2) {
895 align_get_bits(&s->gb);
896 }
897 }
898
899 /* normalize */
900 {
901 int n4 = s->block_len / 2;
902 mdct_norm = 1.0 / (float)n4;
903 if (s->version == 1) {
904 mdct_norm *= sqrt(n4);
905 }
906 }
907
908 /* finally compute the MDCT coefficients */
909 for(ch = 0; ch < s->nb_channels; ch++) {
910 if (s->channel_coded[ch]) {
911 int16_t *coefs1;
912 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
913 int i, j, n, n1, last_high_band;
914 float exp_power[HIGH_BAND_MAX_SIZE];
915
916 coefs1 = s->coefs1[ch];
917 exponents = s->exponents[ch];
918 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
919 mult *= mdct_norm;
920 coefs = s->coefs[ch];
921 if (s->use_noise_coding) {
922 mult1 = mult;
923 /* very low freqs : noise */
924 for(i = 0;i < s->coefs_start; i++) {
925 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
926 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
927 }
928
929 n1 = s->exponent_high_sizes[bsize];
930
931 /* compute power of high bands */
932 exp_ptr = exponents +
933 s->high_band_start[bsize] -
934 s->coefs_start;
935 last_high_band = 0; /* avoid warning */
936 for(j=0;j<n1;j++) {
937 n = s->exponent_high_bands[s->frame_len_bits -
938 s->block_len_bits][j];
939 if (s->high_band_coded[ch][j]) {
940 float e2, v;
941 e2 = 0;
942 for(i = 0;i < n; i++) {
943 v = exp_ptr[i];
944 e2 += v * v;
945 }
946 exp_power[j] = e2 / n;
947 last_high_band = j;
948 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
949 }
950 exp_ptr += n;
951 }
952
953 /* main freqs and high freqs */
954 for(j=-1;j<n1;j++) {
955 if (j < 0) {
956 n = s->high_band_start[bsize] -
957 s->coefs_start;
958 } else {
959 n = s->exponent_high_bands[s->frame_len_bits -
960 s->block_len_bits][j];
961 }
962 if (j >= 0 && s->high_band_coded[ch][j]) {
963 /* use noise with specified power */
964 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
965 /* XXX: use a table */
966 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
967 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
968 mult1 *= mdct_norm;
969 for(i = 0;i < n; i++) {
970 noise = s->noise_table[s->noise_index];
971 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
972 *coefs++ = (*exponents++) * noise * mult1;
973 }
974 } else {
975 /* coded values + small noise */
976 for(i = 0;i < n; i++) {
977 noise = s->noise_table[s->noise_index];
978 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
979 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
980 }
981 }
982 }
983
984 /* very high freqs : noise */
985 n = s->block_len - s->coefs_end[bsize];
986 mult1 = mult * exponents[-1];
987 for(i = 0; i < n; i++) {
988 *coefs++ = s->noise_table[s->noise_index] * mult1;
989 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
990 }
991 } else {
992 /* XXX: optimize more */
993 for(i = 0;i < s->coefs_start; i++)
994 *coefs++ = 0.0;
995 n = nb_coefs[ch];
996 for(i = 0;i < n; i++) {
997 *coefs++ = coefs1[i] * exponents[i] * mult;
998 }
999 n = s->block_len - s->coefs_end[bsize];
1000 for(i = 0;i < n; i++)
1001 *coefs++ = 0.0;
1002 }
1003 }
1004 }
1005
1006 #ifdef TRACE
1007 for(ch = 0; ch < s->nb_channels; ch++) {
1008 if (s->channel_coded[ch]) {
1009 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1010 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1011 }
1012 }
1013 #endif
1014
1015 if (s->ms_stereo && s->channel_coded[1]) {
1016 float a, b;
1017 int i;
1018
1019 /* nominal case for ms stereo: we do it before mdct */
1020 /* no need to optimize this case because it should almost
1021 never happen */
1022 if (!s->channel_coded[0]) {
1023 tprintf("rare ms-stereo case happened\n");
1024 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1025 s->channel_coded[0] = 1;
1026 }
1027
1028 for(i = 0; i < s->block_len; i++) {
1029 a = s->coefs[0][i];
1030 b = s->coefs[1][i];
1031 s->coefs[0][i] = a + b;
1032 s->coefs[1][i] = a - b;
1033 }
1034 }
1035
1036 /* build the window : we ensure that when the windows overlap
1037 their squared sum is always 1 (MDCT reconstruction rule) */
1038 /* XXX: merge with output */
1039 {
1040 int i, next_block_len, block_len, prev_block_len, n;
1041 float *wptr;
1042
1043 block_len = s->block_len;
1044 prev_block_len = 1 << s->prev_block_len_bits;
1045 next_block_len = 1 << s->next_block_len_bits;
1046
1047 /* right part */
1048 wptr = window + block_len;
1049 if (block_len <= next_block_len) {
1050 for(i=0;i<block_len;i++)
1051 *wptr++ = s->windows[bsize][i];
1052 } else {
1053 /* overlap */
1054 n = (block_len / 2) - (next_block_len / 2);
1055 for(i=0;i<n;i++)
1056 *wptr++ = 1.0;
1057 for(i=0;i<next_block_len;i++)
1058 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1059 for(i=0;i<n;i++)
1060 *wptr++ = 0.0;
1061 }
1062
1063 /* left part */
1064 wptr = window + block_len;
1065 if (block_len <= prev_block_len) {
1066 for(i=0;i<block_len;i++)
1067 *--wptr = s->windows[bsize][i];
1068 } else {
1069 /* overlap */
1070 n = (block_len / 2) - (prev_block_len / 2);
1071 for(i=0;i<n;i++)
1072 *--wptr = 1.0;
1073 for(i=0;i<prev_block_len;i++)
1074 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1075 for(i=0;i<n;i++)
1076 *--wptr = 0.0;
1077 }
1078 }
1079
1080
1081 for(ch = 0; ch < s->nb_channels; ch++) {
1082 if (s->channel_coded[ch]) {
1083 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1084 float *ptr;
1085 int i, n4, index, n;
1086
1087 n = s->block_len;
1088 n4 = s->block_len / 2;
1089 ff_imdct_calc(&s->mdct_ctx[bsize],
1090 output, s->coefs[ch], s->mdct_tmp);
1091
1092 /* XXX: optimize all that by build the window and
1093 multipying/adding at the same time */
1094 /* multiply by the window */
1095 for(i=0;i<n * 2;i++) {
1096 output[i] *= window[i];
1097 }
1098
1099 /* add in the frame */
1100 index = (s->frame_len / 2) + s->block_pos - n4;
1101 ptr = &s->frame_out[ch][index];
1102 for(i=0;i<n * 2;i++) {
1103 *ptr += output[i];
1104 ptr++;
1105 }
1106
1107 /* specific fast case for ms-stereo : add to second
1108 channel if it is not coded */
1109 if (s->ms_stereo && !s->channel_coded[1]) {
1110 ptr = &s->frame_out[1][index];
1111 for(i=0;i<n * 2;i++) {
1112 *ptr += output[i];
1113 ptr++;
1114 }
1115 }
1116 }
1117 }
1118 next:
1119 /* update block number */
1120 s->block_num++;
1121 s->block_pos += s->block_len;
1122 if (s->block_pos >= s->frame_len)
1123 return 1;
1124 else
1125 return 0;
1126 }
1127
1128 /* decode a frame of frame_len samples */
wma_decode_frame(WMADecodeContext * s,int16_t * samples)1129 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1130 {
1131 int ret, i, n, a, ch, incr;
1132 int16_t *ptr;
1133 float *iptr;
1134
1135 #ifdef TRACE
1136 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
1137 #endif
1138
1139 /* read each block */
1140 s->block_num = 0;
1141 s->block_pos = 0;
1142 for(;;) {
1143 ret = wma_decode_block(s);
1144 if (ret < 0)
1145 return -1;
1146 if (ret)
1147 break;
1148 }
1149
1150 /* convert frame to integer */
1151 n = s->frame_len;
1152 incr = s->nb_channels;
1153 for(ch = 0; ch < s->nb_channels; ch++) {
1154 ptr = samples + ch;
1155 iptr = s->frame_out[ch];
1156
1157 for(i=0;i<n;i++) {
1158 a = rintf(*iptr++);
1159 if (a > 32767)
1160 a = 32767;
1161 else if (a < -32768)
1162 a = -32768;
1163 *ptr = a;
1164 ptr += incr;
1165 }
1166 /* prepare for next block */
1167 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1168 s->frame_len * sizeof(float));
1169 /* XXX: suppress this */
1170 memset(&s->frame_out[ch][s->frame_len], 0,
1171 s->frame_len * sizeof(float));
1172 }
1173
1174 #ifdef TRACE
1175 dump_shorts("samples", samples, n * s->nb_channels);
1176 #endif
1177 return 0;
1178 }
1179
wma_decode_superframe(AVCodecContext * avctx,void * data,int * data_size,uint8_t * buf,int buf_size)1180 static int wma_decode_superframe(AVCodecContext *avctx,
1181 void *data, int *data_size,
1182 uint8_t *buf, int buf_size)
1183 {
1184 WMADecodeContext *s = avctx->priv_data;
1185 int nb_frames, bit_offset, i, pos, len;
1186 uint8_t *q;
1187 int16_t *samples;
1188
1189 tprintf("***decode_superframe:\n");
1190
1191 if(buf_size==0){
1192 s->last_superframe_len = 0;
1193 return 0;
1194 }
1195
1196 samples = data;
1197
1198 init_get_bits(&s->gb, buf, buf_size*8);
1199
1200 if (s->use_bit_reservoir) {
1201 /* read super frame header */
1202 get_bits(&s->gb, 4); /* super frame index */
1203 nb_frames = get_bits(&s->gb, 4) - 1;
1204
1205 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1206
1207 if (s->last_superframe_len > 0) {
1208 // printf("skip=%d\n", s->last_bitoffset);
1209 /* add bit_offset bits to last frame */
1210 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1211 MAX_CODED_SUPERFRAME_SIZE)
1212 goto fail;
1213 q = s->last_superframe + s->last_superframe_len;
1214 len = bit_offset;
1215 while (len > 0) {
1216 *q++ = (get_bits)(&s->gb, 8);
1217 len -= 8;
1218 }
1219 if (len > 0) {
1220 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1221 }
1222
1223 /* XXX: bit_offset bits into last frame */
1224 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1225 /* skip unused bits */
1226 if (s->last_bitoffset > 0)
1227 skip_bits(&s->gb, s->last_bitoffset);
1228 /* this frame is stored in the last superframe and in the
1229 current one */
1230 if (wma_decode_frame(s, samples) < 0)
1231 goto fail;
1232 samples += s->nb_channels * s->frame_len;
1233 }
1234
1235 /* read each frame starting from bit_offset */
1236 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1237 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1238 len = pos & 7;
1239 if (len > 0)
1240 skip_bits(&s->gb, len);
1241
1242 s->reset_block_lengths = 1;
1243 for(i=0;i<nb_frames;i++) {
1244 if (wma_decode_frame(s, samples) < 0)
1245 goto fail;
1246 samples += s->nb_channels * s->frame_len;
1247 }
1248
1249 /* we copy the end of the frame in the last frame buffer */
1250 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1251 s->last_bitoffset = pos & 7;
1252 pos >>= 3;
1253 len = buf_size - pos;
1254 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1255 goto fail;
1256 }
1257 s->last_superframe_len = len;
1258 memcpy(s->last_superframe, buf + pos, len);
1259 } else {
1260 /* single frame decode */
1261 if (wma_decode_frame(s, samples) < 0)
1262 goto fail;
1263 samples += s->nb_channels * s->frame_len;
1264 }
1265 *data_size = (int8_t *)samples - (int8_t *)data;
1266 return s->block_align;
1267 fail:
1268 /* when error, we reset the bit reservoir */
1269 s->last_superframe_len = 0;
1270 return -1;
1271 }
1272
wma_decode_end(AVCodecContext * avctx)1273 static int wma_decode_end(AVCodecContext *avctx)
1274 {
1275 WMADecodeContext *s = avctx->priv_data;
1276 int i;
1277
1278 for(i = 0; i < s->nb_block_sizes; i++)
1279 ff_mdct_end(&s->mdct_ctx[i]);
1280 for(i = 0; i < s->nb_block_sizes; i++)
1281 av_free(s->windows[i]);
1282
1283 if (s->use_exp_vlc) {
1284 free_vlc(&s->exp_vlc);
1285 }
1286 if (s->use_noise_coding) {
1287 free_vlc(&s->hgain_vlc);
1288 }
1289 for(i = 0;i < 2; i++) {
1290 free_vlc(&s->coef_vlc[i]);
1291 av_free(s->run_table[i]);
1292 av_free(s->level_table[i]);
1293 }
1294
1295 return 0;
1296 }
1297
1298 AVCodec wmav1_decoder =
1299 {
1300 "wmav1",
1301 CODEC_TYPE_AUDIO,
1302 CODEC_ID_WMAV1,
1303 sizeof(WMADecodeContext),
1304 wma_decode_init,
1305 NULL,
1306 wma_decode_end,
1307 wma_decode_superframe,
1308 };
1309
1310 AVCodec wmav2_decoder =
1311 {
1312 "wmav2",
1313 CODEC_TYPE_AUDIO,
1314 CODEC_ID_WMAV2,
1315 sizeof(WMADecodeContext),
1316 wma_decode_init,
1317 NULL,
1318 wma_decode_end,
1319 wma_decode_superframe,
1320 };
1321