a6fa2f8b2a1049057e0cb9fa9b0adaf6f1a22f4c
[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 #include "avcodec.h"
20 #include "dsputil.h"
21
22 //#define DEBUG_PARAMS
23 //#define DEBUG_TRACE
24
25 /* size of blocks */
26 #define BLOCK_MIN_BITS 7
27 #define BLOCK_MAX_BITS 11
28 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
29
30 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
31
32 /* XXX: find exact max size */
33 #define HIGH_BAND_MAX_SIZE 16
34
35 #define NB_LSP_COEFS 10
36
37 /* XXX: is it a suitable value ? */
38 #define MAX_CODED_SUPERFRAME_SIZE 4096
39
40 #define MAX_CHANNELS 2
41
42 #define NOISE_TAB_SIZE 8192
43
44 #define LSP_POW_BITS 7
45
46 typedef struct WMADecodeContext {
47 GetBitContext gb;
48 int sample_rate;
49 int nb_channels;
50 int bit_rate;
51 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
52 int block_align;
53 int use_bit_reservoir;
54 int use_variable_block_len;
55 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
56 int use_noise_coding; /* true if perceptual noise is added */
57 int byte_offset_bits;
58 VLC exp_vlc;
59 int exponent_sizes[BLOCK_NB_SIZES];
60 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
61 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
62 int coefs_start; /* first coded coef */
63 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
64 int exponent_high_sizes[BLOCK_NB_SIZES];
65 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
66 VLC hgain_vlc;
67
68 /* coded values in high bands */
69 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
70 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
71
72 /* there are two possible tables for spectral coefficients */
73 VLC coef_vlc[2];
74 uint16_t *run_table[2];
75 uint16_t *level_table[2];
76 /* frame info */
77 int frame_len; /* frame length in samples */
78 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
79 int nb_block_sizes; /* number of block sizes */
80 /* block info */
81 int reset_block_lengths;
82 int block_len_bits; /* log2 of current block length */
83 int next_block_len_bits; /* log2 of next block length */
84 int prev_block_len_bits; /* log2 of prev block length */
85 int block_len; /* block length in samples */
86 int block_num; /* block number in current frame */
87 int block_pos; /* current position in frame */
88 uint8_t ms_stereo; /* true if mid/side stereo mode */
89 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
90 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
91 float max_exponent[MAX_CHANNELS];
92 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
93 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
94 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
95 float *windows[BLOCK_NB_SIZES] __attribute__((aligned(16)));
96 FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
97 /* output buffer for one frame and the last for IMDCT windowing */
98 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
99 /* last frame info */
100 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
101 int last_bitoffset;
102 int last_superframe_len;
103 float noise_table[NOISE_TAB_SIZE];
104 int noise_index;
105 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
106 /* lsp_to_curve tables */
107 float lsp_cos_table[BLOCK_MAX_SIZE];
108 float lsp_pow_e_table[256];
109 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
110 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
111 } WMADecodeContext;
112
113 typedef struct CoefVLCTable {
114 int n; /* total number of codes */
115 const uint32_t *huffcodes; /* VLC bit values */
116 const uint8_t *huffbits; /* VLC bit size */
117 const uint16_t *levels; /* table to build run/level tables */
118 } CoefVLCTable;
119
120 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
121
122 #include "wmadata.h"
123
124 #ifdef DEBUG_TRACE
125 #include <stdarg.h>
126 int frame_count;
127
128 static FILE *flog;
129
130 void trace(const char *fmt, ...)
131 {
132 va_list ap;
133
134
135 if (!flog) {
136 flog = fopen("/tmp/out.log", "w");
137 setlinebuf(flog);
138 }
139
140 va_start(ap, fmt);
141 vfprintf(flog, fmt, ap);
142 va_end(ap);
143 }
144
145 #define get_bits(s, n) get_bits_trace(s, n)
146 #define get_vlc(s, vlc) get_vlc_trace(s, vlc)
147
148 unsigned int get_bits_trace(GetBitContext *s, int n)
149 {
150 unsigned int val;
151 val = (get_bits)(s, n);
152 trace("get_bits(%d) : 0x%x\n", n, val);
153 return val;
154 }
155
156 static int get_vlc_trace(GetBitContext *s, VLC *vlc)
157 {
158 int code;
159 code = (get_vlc)(s, vlc);
160 trace("get_vlc() : %d\n", code);
161 return code;
162 }
163
164 static void dump_shorts(const char *name, const short *tab, int n)
165 {
166 int i;
167
168 trace("%s[%d]:\n", name, n);
169 for(i=0;i<n;i++) {
170 if ((i & 7) == 0)
171 trace("%4d: ", i);
172 trace(" %5d.0", tab[i]);
173 if ((i & 7) == 7)
174 trace("\n");
175 }
176 }
177
178 static void dump_floats(const char *name, int prec, const float *tab, int n)
179 {
180 int i;
181
182 trace("%s[%d]:\n", name, n);
183 for(i=0;i<n;i++) {
184 if ((i & 7) == 0)
185 trace("%4d: ", i);
186 trace(" %8.*f", prec, tab[i]);
187 if ((i & 7) == 7)
188 trace("\n");
189 }
190 if ((i & 7) != 0)
191 trace("\n");
192 }
193
194 #else
195
196 #define trace(fmt, ...)
197
198 #endif
199
200 /* XXX: use same run/length optimization as mpeg decoders */
201 static void init_coef_vlc(VLC *vlc,
202 uint16_t **prun_table, uint16_t **plevel_table,
203 const CoefVLCTable *vlc_table)
204 {
205 int n = vlc_table->n;
206 const uint8_t *table_bits = vlc_table->huffbits;
207 const uint32_t *table_codes = vlc_table->huffcodes;
208 const uint16_t *levels_table = vlc_table->levels;
209 uint16_t *run_table, *level_table;
210 const uint16_t *p;
211 int i, l, j, level;
212
213 init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4);
214
215 run_table = malloc(n * sizeof(uint16_t));
216 level_table = malloc(n * sizeof(uint16_t));
217 p = levels_table;
218 i = 2;
219 level = 1;
220 while (i < n) {
221 l = *p++;
222 for(j=0;j<l;j++) {
223 run_table[i] = j;
224 level_table[i] = level;
225 i++;
226 }
227 level++;
228 }
229 *prun_table = run_table;
230 *plevel_table = level_table;
231 }
232
233 static int wma_decode_init(AVCodecContext * avctx)
234 {
235 WMADecodeContext *s = avctx->priv_data;
236 int i, flags1, flags2;
237 float *window;
238 uint8_t *extradata;
239 float bps1, high_freq, bps;
240 int sample_rate1;
241 int coef_vlc_table;
242
243 s->sample_rate = avctx->sample_rate;
244 s->nb_channels = avctx->channels;
245 s->bit_rate = avctx->bit_rate;
246 s->block_align = avctx->block_align;
247
248 if (avctx->codec->id == CODEC_ID_WMAV1) {
249 s->version = 1;
250 } else {
251 s->version = 2;
252 }
253
254 /* extract flag infos */
255 flags1 = 0;
256 flags2 = 0;
257 extradata = avctx->extradata;
258 if (s->version == 1 && avctx->extradata_size >= 4) {
259 flags1 = extradata[0] | (extradata[1] << 8);
260 flags2 = extradata[2] | (extradata[3] << 8);
261 } else if (s->version == 2 && avctx->extradata_size >= 6) {
262 flags1 = extradata[0] | (extradata[1] << 8) |
263 (extradata[2] << 16) | (extradata[3] << 24);
264 flags2 = extradata[4] | (extradata[5] << 8);
265 }
266 s->use_exp_vlc = flags2 & 0x0001;
267 s->use_bit_reservoir = flags2 & 0x0002;
268 s->use_variable_block_len = flags2 & 0x0004;
269
270 /* compute MDCT block size */
271 if (s->sample_rate <= 16000) {
272 s->frame_len_bits = 9;
273 } else if (s->sample_rate <= 22050 ||
274 (s->sample_rate <= 32000 && s->version == 1)) {
275 s->frame_len_bits = 10;
276 } else {
277 s->frame_len_bits = 11;
278 }
279 s->frame_len = 1 << s->frame_len_bits;
280 if (s->use_variable_block_len) {
281 int nb_max, nb;
282 nb = ((flags2 >> 3) & 3) + 1;
283 if ((s->bit_rate / s->nb_channels) >= 32000)
284 nb += 2;
285 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
286 if (nb > nb_max)
287 nb = nb_max;
288 s->nb_block_sizes = nb + 1;
289 } else {
290 s->nb_block_sizes = 1;
291 }
292
293 /* init rate dependant parameters */
294 s->use_noise_coding = 1;
295 high_freq = s->sample_rate * 0.5;
296
297 /* if version 2, then the rates are normalized */
298 sample_rate1 = s->sample_rate;
299 if (s->version == 2) {
300 if (sample_rate1 >= 44100)
301 sample_rate1 = 44100;
302 else if (sample_rate1 >= 22050)
303 sample_rate1 = 22050;
304 else if (sample_rate1 >= 16000)
305 sample_rate1 = 16000;
306 else if (sample_rate1 >= 11025)
307 sample_rate1 = 11025;
308 else if (sample_rate1 >= 8000)
309 sample_rate1 = 8000;
310 }
311
312 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
313 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
314
315 /* compute high frequency value and choose if noise coding should
316 be activated */
317 bps1 = bps;
318 if (s->nb_channels == 2)
319 bps1 = bps * 1.6;
320 if (sample_rate1 == 44100) {
321 if (bps1 >= 0.61)
322 s->use_noise_coding = 0;
323 else
324 high_freq = high_freq * 0.4;
325 } else if (sample_rate1 == 22050) {
326 if (bps1 >= 1.16)
327 s->use_noise_coding = 0;
328 else if (bps1 >= 0.72)
329 high_freq = high_freq * 0.7;
330 else
331 high_freq = high_freq * 0.6;
332 } else if (sample_rate1 == 16000) {
333 if (bps > 0.5)
334 high_freq = high_freq * 0.5;
335 else
336 high_freq = high_freq * 0.3;
337 } else if (sample_rate1 == 11025) {
338 high_freq = high_freq * 0.7;
339 } else if (sample_rate1 == 8000) {
340 if (bps <= 0.625) {
341 high_freq = high_freq * 0.5;
342 } else if (bps > 0.75) {
343 s->use_noise_coding = 0;
344 } else {
345 high_freq = high_freq * 0.65;
346 }
347 } else {
348 if (bps >= 0.8) {
349 high_freq = high_freq * 0.75;
350 } else if (bps >= 0.6) {
351 high_freq = high_freq * 0.6;
352 } else {
353 high_freq = high_freq * 0.5;
354 }
355 }
356 #ifdef DEBUG_PARAMS
357 printf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
358 printf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
359 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
360 s->block_align);
361 printf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
362 bps, bps1, high_freq, s->byte_offset_bits);
363 printf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
364 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
365 #endif
366
367 /* compute the scale factor band sizes for each MDCT block size */
368 {
369 int a, b, pos, lpos, k, block_len, i, j, n;
370 const uint8_t *table;
371
372 if (s->version == 1) {
373 s->coefs_start = 3;
374 } else {
375 s->coefs_start = 0;
376 }
377 for(k = 0; k < s->nb_block_sizes; k++) {
378 block_len = s->frame_len >> k;
379
380 if (s->version == 1) {
381 lpos = 0;
382 for(i=0;i<25;i++) {
383 a = wma_critical_freqs[i];
384 b = s->sample_rate;
385 pos = ((block_len * 2 * a) + (b >> 1)) / b;
386 if (pos > block_len)
387 pos = block_len;
388 s->exponent_bands[0][i] = pos - lpos;
389 if (pos >= block_len) {
390 i++;
391 break;
392 }
393 lpos = pos;
394 }
395 s->exponent_sizes[0] = i;
396 } else {
397 /* hardcoded tables */
398 table = NULL;
399 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
400 if (a < 3) {
401 if (s->sample_rate >= 44100)
402 table = exponent_band_44100[a];
403 else if (s->sample_rate >= 32000)
404 table = exponent_band_32000[a];
405 else if (s->sample_rate >= 22050)
406 table = exponent_band_22050[a];
407 }
408 if (table) {
409 n = *table++;
410 for(i=0;i<n;i++)
411 s->exponent_bands[k][i] = table[i];
412 s->exponent_sizes[k] = n;
413 } else {
414 j = 0;
415 lpos = 0;
416 for(i=0;i<25;i++) {
417 a = wma_critical_freqs[i];
418 b = s->sample_rate;
419 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
420 pos <<= 2;
421 if (pos > block_len)
422 pos = block_len;
423 if (pos > lpos)
424 s->exponent_bands[k][j++] = pos - lpos;
425 if (pos >= block_len)
426 break;
427 lpos = pos;
428 }
429 s->exponent_sizes[k] = j;
430 }
431 }
432
433 /* max number of coefs */
434 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
435 /* high freq computation */
436 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
437 s->sample_rate + 0.5);
438 n = s->exponent_sizes[k];
439 j = 0;
440 pos = 0;
441 for(i=0;i<n;i++) {
442 int start, end;
443 start = pos;
444 pos += s->exponent_bands[k][i];
445 end = pos;
446 if (start < s->high_band_start[k])
447 start = s->high_band_start[k];
448 if (end > s->coefs_end[k])
449 end = s->coefs_end[k];
450 if (end > start)
451 s->exponent_high_bands[k][j++] = end - start;
452 }
453 s->exponent_high_sizes[k] = j;
454 #if 0
455 trace("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
456 s->frame_len >> k,
457 s->coefs_end[k],
458 s->high_band_start[k],
459 s->exponent_high_sizes[k]);
460 for(j=0;j<s->exponent_high_sizes[k];j++)
461 trace(" %d", s->exponent_high_bands[k][j]);
462 trace("\n");
463 #endif
464 }
465 }
466
467 #ifdef DEBUG_TRACE
468 {
469 int i, j;
470 for(i = 0; i < s->nb_block_sizes; i++) {
471 trace("%5d: n=%2d:",
472 s->frame_len >> i,
473 s->exponent_sizes[i]);
474 for(j=0;j<s->exponent_sizes[i];j++)
475 trace(" %d", s->exponent_bands[i][j]);
476 trace("\n");
477 }
478 }
479 #endif
480
481 /* init MDCT */
482 for(i = 0; i < s->nb_block_sizes; i++)
483 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
484
485 /* init MDCT windows : simple sinus window */
486 for(i = 0; i < s->nb_block_sizes; i++) {
487 int n, j;
488 float alpha;
489 n = 1 << (s->frame_len_bits - i);
490 window = av_malloc(sizeof(float) * n);
491 alpha = M_PI / (2.0 * n);
492 for(j=0;j<n;j++) {
493 window[n - j - 1] = sin((j + 0.5) * alpha);
494 }
495 s->windows[i] = window;
496 }
497
498 s->reset_block_lengths = 1;
499
500 if (s->use_noise_coding) {
501
502 /* init the noise generator */
503 if (s->use_exp_vlc)
504 s->noise_mult = 0.02;
505 else
506 s->noise_mult = 0.04;
507
508 #if defined(DEBUG_TRACE)
509 for(i=0;i<NOISE_TAB_SIZE;i++)
510 s->noise_table[i] = 1.0 * s->noise_mult;
511 #else
512 {
513 unsigned int seed;
514 float norm;
515 seed = 1;
516 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
517 for(i=0;i<NOISE_TAB_SIZE;i++) {
518 seed = seed * 314159 + 1;
519 s->noise_table[i] = (float)((int)seed) * norm;
520 }
521 }
522 #endif
523 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
524 hgain_huffbits, 1, 1,
525 hgain_huffcodes, 2, 2);
526 }
527
528 if (s->use_exp_vlc) {
529 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
530 scale_huffbits, 1, 1,
531 scale_huffcodes, 4, 4);
532 } else {
533 wma_lsp_to_curve_init(s, s->frame_len);
534 }
535
536 /* choose the VLC tables for the coefficients */
537 coef_vlc_table = 2;
538 if (s->sample_rate >= 32000) {
539 if (bps1 < 0.72)
540 coef_vlc_table = 0;
541 else if (bps1 < 1.16)
542 coef_vlc_table = 1;
543 }
544
545 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
546 &coef_vlcs[coef_vlc_table * 2]);
547 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
548 &coef_vlcs[coef_vlc_table * 2 + 1]);
549 return 0;
550 }
551
552 /* interpolate values for a bigger or smaller block. The block must
553 have multiple sizes */
554 static void interpolate_array(float *scale, int old_size, int new_size)
555 {
556 int i, j, jincr, k;
557 float v;
558
559 if (new_size > old_size) {
560 jincr = new_size / old_size;
561 j = new_size;
562 for(i = old_size - 1; i >=0; i--) {
563 v = scale[i];
564 k = jincr;
565 do {
566 scale[--j] = v;
567 } while (--k);
568 }
569 } else if (new_size < old_size) {
570 j = 0;
571 jincr = old_size / new_size;
572 for(i = 0; i < new_size; i++) {
573 scale[i] = scale[j];
574 j += jincr;
575 }
576 }
577 }
578
579 /* compute x^-0.25 with an exponent and mantissa table. We use linear
580 interpolation to reduce the mantissa table size at a small speed
581 expense (linear interpolation approximately doubles the number of
582 bits of precision). */
583 static inline float pow_m1_4(WMADecodeContext *s, float x)
584 {
585 union {
586 float f;
587 unsigned int v;
588 } u, t;
589 unsigned int e, m;
590 float a, b;
591
592 u.f = x;
593 e = u.v >> 23;
594 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
595 /* build interpolation scale: 1 <= t < 2. */
596 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
597 a = s->lsp_pow_m_table1[m];
598 b = s->lsp_pow_m_table2[m];
599 return s->lsp_pow_e_table[e] * (a + b * t.f);
600 }
601
602 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
603 {
604 float wdel, a, b;
605 int i, e, m;
606
607 wdel = M_PI / frame_len;
608 for(i=0;i<frame_len;i++)
609 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
610
611 /* tables for x^-0.25 computation */
612 for(i=0;i<256;i++) {
613 e = i - 126;
614 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
615 }
616
617 /* NOTE: these two tables are needed to avoid two operations in
618 pow_m1_4 */
619 b = 1.0;
620 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
621 m = (1 << LSP_POW_BITS) + i;
622 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
623 a = pow(a, -0.25);
624 s->lsp_pow_m_table1[i] = 2 * a - b;
625 s->lsp_pow_m_table2[i] = b - a;
626 b = a;
627 }
628 #if 0
629 for(i=1;i<20;i++) {
630 float v, r1, r2;
631 v = 5.0 / i;
632 r1 = pow_m1_4(s, v);
633 r2 = pow(v,-0.25);
634 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
635 }
636 #endif
637 }
638
639 /* NOTE: We use the same code as Vorbis here */
640 /* XXX: optimize it further with SSE/3Dnow */
641 static void wma_lsp_to_curve(WMADecodeContext *s,
642 float *out, float *val_max_ptr,
643 int n, float *lsp)
644 {
645 int i, j;
646 float p, q, w, v, val_max;
647
648 val_max = 0;
649 for(i=0;i<n;i++) {
650 p = 0.5f;
651 q = 0.5f;
652 w = s->lsp_cos_table[i];
653 for(j=1;j<NB_LSP_COEFS;j+=2){
654 q *= w - lsp[j - 1];
655 p *= w - lsp[j];
656 }
657 p *= p * (2.0f - w);
658 q *= q * (2.0f + w);
659 v = p + q;
660 v = pow_m1_4(s, v);
661 if (v > val_max)
662 val_max = v;
663 out[i] = v;
664 }
665 *val_max_ptr = val_max;
666 }
667
668 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
669 static void decode_exp_lsp(WMADecodeContext *s, int ch)
670 {
671 float lsp_coefs[NB_LSP_COEFS];
672 int val, i;
673
674 for(i = 0; i < NB_LSP_COEFS; i++) {
675 if (i == 0 || i >= 8)
676 val = get_bits(&s->gb, 3);
677 else
678 val = get_bits(&s->gb, 4);
679 lsp_coefs[i] = lsp_codebook[i][val];
680 }
681
682 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
683 s->block_len, lsp_coefs);
684 }
685
686 /* decode exponents coded with VLC codes */
687 static int decode_exp_vlc(WMADecodeContext *s, int ch)
688 {
689 int last_exp, n, code;
690 const uint16_t *ptr, *band_ptr;
691 float v, *q, max_scale, *q_end;
692
693 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
694 ptr = band_ptr;
695 q = s->exponents[ch];
696 q_end = q + s->block_len;
697 max_scale = 0;
698 if (s->version == 1) {
699 last_exp = get_bits(&s->gb, 5) + 10;
700 /* XXX: use a table */
701 v = pow(10, last_exp * (1.0 / 16.0));
702 max_scale = v;
703 n = *ptr++;
704 do {
705 *q++ = v;
706 } while (--n);
707 }
708 last_exp = 36;
709 while (q < q_end) {
710 code = get_vlc(&s->gb, &s->exp_vlc);
711 if (code < 0)
712 return -1;
713 /* NOTE: this offset is the same as MPEG4 AAC ! */
714 last_exp += code - 60;
715 /* XXX: use a table */
716 v = pow(10, last_exp * (1.0 / 16.0));
717 if (v > max_scale)
718 max_scale = v;
719 n = *ptr++;
720 do {
721 *q++ = v;
722 } while (--n);
723 }
724 s->max_exponent[ch] = max_scale;
725 return 0;
726 }
727
728 /* return 0 if OK. return 1 if last block of frame. return -1 if
729 unrecorrable error. */
730 static int wma_decode_block(WMADecodeContext *s)
731 {
732 int n, v, a, ch, code, bsize;
733 int coef_nb_bits, total_gain, parse_exponents;
734 float window[BLOCK_MAX_SIZE * 2];
735 int nb_coefs[MAX_CHANNELS];
736 float mdct_norm;
737
738 trace("***decode_block: %d:%d\n", frame_count - 1, s->block_num);
739
740 /* compute current block length */
741 if (s->use_variable_block_len) {
742 n = av_log2(s->nb_block_sizes - 1) + 1;
743
744 if (s->reset_block_lengths) {
745 s->reset_block_lengths = 0;
746 v = get_bits(&s->gb, n);
747 if (v >= s->nb_block_sizes)
748 return -1;
749 s->prev_block_len_bits = s->frame_len_bits - v;
750 v = get_bits(&s->gb, n);
751 if (v >= s->nb_block_sizes)
752 return -1;
753 s->block_len_bits = s->frame_len_bits - v;
754 } else {
755 /* update block lengths */
756 s->prev_block_len_bits = s->block_len_bits;
757 s->block_len_bits = s->next_block_len_bits;
758 }
759 v = get_bits(&s->gb, n);
760 if (v >= s->nb_block_sizes)
761 return -1;
762 s->next_block_len_bits = s->frame_len_bits - v;
763 } else {
764 /* fixed block len */
765 s->next_block_len_bits = s->frame_len_bits;
766 s->prev_block_len_bits = s->frame_len_bits;
767 s->block_len_bits = s->frame_len_bits;
768 }
769
770 /* now check if the block length is coherent with the frame length */
771 s->block_len = 1 << s->block_len_bits;
772 if ((s->block_pos + s->block_len) > s->frame_len)
773 return -1;
774
775 if (s->nb_channels == 2) {
776 s->ms_stereo = get_bits(&s->gb, 1);
777 }
778 v = 0;
779 for(ch = 0; ch < s->nb_channels; ch++) {
780 a = get_bits(&s->gb, 1);
781 s->channel_coded[ch] = a;
782 v |= a;
783 }
784 /* if no channel coded, no need to go further */
785 /* XXX: fix potential framing problems */
786 if (!v)
787 goto next;
788
789 bsize = s->frame_len_bits - s->block_len_bits;
790
791 /* read total gain and extract corresponding number of bits for
792 coef escape coding */
793 total_gain = 1;
794 for(;;) {
795 a = get_bits(&s->gb, 7);
796 total_gain += a;
797 if (a != 127)
798 break;
799 }
800
801 if (total_gain < 15)
802 coef_nb_bits = 13;
803 else if (total_gain < 32)
804 coef_nb_bits = 12;
805 else if (total_gain < 40)
806 coef_nb_bits = 11;
807 else if (total_gain < 45)
808 coef_nb_bits = 10;
809 else
810 coef_nb_bits = 9;
811
812 /* compute number of coefficients */
813 n = s->coefs_end[bsize] - s->coefs_start;
814 for(ch = 0; ch < s->nb_channels; ch++)
815 nb_coefs[ch] = n;
816
817 /* complex coding */
818 if (s->use_noise_coding) {
819
820 for(ch = 0; ch < s->nb_channels; ch++) {
821 if (s->channel_coded[ch]) {
822 int i, n, a;
823 n = s->exponent_high_sizes[bsize];
824 for(i=0;i<n;i++) {
825 a = get_bits(&s->gb, 1);
826 s->high_band_coded[ch][i] = a;
827 /* if noise coding, the coefficients are not transmitted */
828 if (a)
829 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
830 }
831 }
832 }
833 for(ch = 0; ch < s->nb_channels; ch++) {
834 if (s->channel_coded[ch]) {
835 int i, n, val, code;
836
837 n = s->exponent_high_sizes[bsize];
838 val = (int)0x80000000;
839 for(i=0;i<n;i++) {
840 if (s->high_band_coded[ch][i]) {
841 if (val == (int)0x80000000) {
842 val = get_bits(&s->gb, 7) - 19;
843 } else {
844 code = get_vlc(&s->gb, &s->hgain_vlc);
845 if (code < 0)
846 return -1;
847 val += code - 18;
848 }
849 s->high_band_values[ch][i] = val;
850 }
851 }
852 }
853 }
854 }
855
856 /* exposant can be interpolated in short blocks. */
857 parse_exponents = 1;
858 if (s->block_len_bits != s->frame_len_bits) {
859 parse_exponents = get_bits(&s->gb, 1);
860 }
861
862 if (parse_exponents) {
863 for(ch = 0; ch < s->nb_channels; ch++) {
864 if (s->channel_coded[ch]) {
865 if (s->use_exp_vlc) {
866 if (decode_exp_vlc(s, ch) < 0)
867 return -1;
868 } else {
869 decode_exp_lsp(s, ch);
870 }
871 }
872 }
873 } else {
874 for(ch = 0; ch < s->nb_channels; ch++) {
875 if (s->channel_coded[ch]) {
876 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
877 s->block_len);
878 }
879 }
880 }
881
882 /* parse spectral coefficients : just RLE encoding */
883 for(ch = 0; ch < s->nb_channels; ch++) {
884 if (s->channel_coded[ch]) {
885 VLC *coef_vlc;
886 int level, run, sign, tindex;
887 int16_t *ptr, *eptr;
888 const int16_t *level_table, *run_table;
889
890 /* special VLC tables are used for ms stereo because
891 there is potentially less energy there */
892 tindex = (ch == 1 && s->ms_stereo);
893 coef_vlc = &s->coef_vlc[tindex];
894 run_table = s->run_table[tindex];
895 level_table = s->level_table[tindex];
896 /* XXX: optimize */
897 ptr = &s->coefs1[ch][0];
898 eptr = ptr + nb_coefs[ch];
899 memset(ptr, 0, s->block_len * sizeof(int16_t));
900 for(;;) {
901 code = get_vlc(&s->gb, coef_vlc);
902 if (code < 0)
903 return -1;
904 if (code == 1) {
905 /* EOB */
906 break;
907 } else if (code == 0) {
908 /* escape */
909 level = get_bits(&s->gb, coef_nb_bits);
910 /* NOTE: this is rather suboptimal. reading
911 block_len_bits would be better */
912 run = get_bits(&s->gb, s->frame_len_bits);
913 } else {
914 /* normal code */
915 run = run_table[code];
916 level = level_table[code];
917 }
918 sign = get_bits(&s->gb, 1);
919 if (!sign)
920 level = -level;
921 ptr += run;
922 if (ptr >= eptr)
923 return -1;
924 *ptr++ = level;
925 /* NOTE: EOB can be omitted */
926 if (ptr >= eptr)
927 break;
928 }
929 }
930 if (s->version == 1 && s->nb_channels >= 2) {
931 align_get_bits(&s->gb);
932 }
933 }
934
935 /* normalize */
936 {
937 int n4 = s->block_len / 2;
938 mdct_norm = 1.0 / (float)n4;
939 if (s->version == 1) {
940 mdct_norm *= sqrt(n4);
941 }
942 }
943
944 /* finally compute the MDCT coefficients */
945 for(ch = 0; ch < s->nb_channels; ch++) {
946 if (s->channel_coded[ch]) {
947 int16_t *coefs1;
948 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
949 int i, j, n, n1, last_high_band;
950 float exp_power[HIGH_BAND_MAX_SIZE];
951
952 coefs1 = s->coefs1[ch];
953 exponents = s->exponents[ch];
954 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
955 mult *= mdct_norm;
956 coefs = s->coefs[ch];
957 if (s->use_noise_coding) {
958 mult1 = mult;
959 /* very low freqs : noise */
960 for(i = 0;i < s->coefs_start; i++) {
961 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
962 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
963 }
964
965 n1 = s->exponent_high_sizes[bsize];
966
967 /* compute power of high bands */
968 exp_ptr = exponents +
969 s->high_band_start[bsize] -
970 s->coefs_start;
971 last_high_band = 0; /* avoid warning */
972 for(j=0;j<n1;j++) {
973 n = s->exponent_high_bands[s->frame_len_bits -
974 s->block_len_bits][j];
975 if (s->high_band_coded[ch][j]) {
976 float e2, v;
977 e2 = 0;
978 for(i = 0;i < n; i++) {
979 v = exp_ptr[i];
980 e2 += v * v;
981 }
982 exp_power[j] = e2 / n;
983 last_high_band = j;
984 trace("%d: power=%f (%d)\n", j, exp_power[j], n);
985 }
986 exp_ptr += n;
987 }
988
989 /* main freqs and high freqs */
990 for(j=-1;j<n1;j++) {
991 if (j < 0) {
992 n = s->high_band_start[bsize] -
993 s->coefs_start;
994 } else {
995 n = s->exponent_high_bands[s->frame_len_bits -
996 s->block_len_bits][j];
997 }
998 if (j >= 0 && s->high_band_coded[ch][j]) {
999 /* use noise with specified power */
1000 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
1001 /* XXX: use a table */
1002 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
1003 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
1004 mult1 *= mdct_norm;
1005 for(i = 0;i < n; i++) {
1006 noise = s->noise_table[s->noise_index];
1007 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1008 *coefs++ = (*exponents++) * noise * mult1;
1009 }
1010 } else {
1011 /* coded values + small noise */
1012 for(i = 0;i < n; i++) {
1013 noise = s->noise_table[s->noise_index];
1014 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1015 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
1016 }
1017 }
1018 }
1019
1020 /* very high freqs : noise */
1021 n = s->block_len - s->coefs_end[bsize];
1022 mult1 = mult * exponents[-1];
1023 for(i = 0; i < n; i++) {
1024 *coefs++ = s->noise_table[s->noise_index] * mult1;
1025 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1026 }
1027 } else {
1028 /* XXX: optimize more */
1029 for(i = 0;i < s->coefs_start; i++)
1030 *coefs++ = 0.0;
1031 n = nb_coefs[ch];
1032 for(i = 0;i < n; i++) {
1033 *coefs++ = coefs1[i] * exponents[i] * mult;
1034 }
1035 n = s->block_len - s->coefs_end[bsize];
1036 for(i = 0;i < n; i++)
1037 *coefs++ = 0.0;
1038 }
1039 }
1040 }
1041
1042 #ifdef DEBUG_TRACE
1043 for(ch = 0; ch < s->nb_channels; ch++) {
1044 if (s->channel_coded[ch]) {
1045 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1046 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1047 }
1048 }
1049 #endif
1050
1051 if (s->ms_stereo && s->channel_coded[1]) {
1052 float a, b;
1053 int i;
1054
1055 /* nominal case for ms stereo: we do it before mdct */
1056 /* no need to optimize this case because it should almost
1057 never happen */
1058 if (!s->channel_coded[0]) {
1059 #ifdef DEBUG_TRACE
1060 trace("rare ms-stereo case happened\n");
1061 #endif
1062 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1063 s->channel_coded[0] = 1;
1064 }
1065
1066 for(i = 0; i < s->block_len; i++) {
1067 a = s->coefs[0][i];
1068 b = s->coefs[1][i];
1069 s->coefs[0][i] = a + b;
1070 s->coefs[1][i] = a - b;
1071 }
1072 }
1073
1074 /* build the window : we ensure that when the windows overlap
1075 their squared sum is always 1 (MDCT reconstruction rule) */
1076 /* XXX: merge with output */
1077 {
1078 int i, next_block_len, block_len, prev_block_len, n;
1079 float *wptr;
1080
1081 block_len = s->block_len;
1082 prev_block_len = 1 << s->prev_block_len_bits;
1083 next_block_len = 1 << s->next_block_len_bits;
1084
1085 /* right part */
1086 wptr = window + block_len;
1087 if (block_len <= next_block_len) {
1088 for(i=0;i<block_len;i++)
1089 *wptr++ = s->windows[bsize][i];
1090 } else {
1091 /* overlap */
1092 n = (block_len / 2) - (next_block_len / 2);
1093 for(i=0;i<n;i++)
1094 *wptr++ = 1.0;
1095 for(i=0;i<next_block_len;i++)
1096 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1097 for(i=0;i<n;i++)
1098 *wptr++ = 0.0;
1099 }
1100
1101 /* left part */
1102 wptr = window + block_len;
1103 if (block_len <= prev_block_len) {
1104 for(i=0;i<block_len;i++)
1105 *--wptr = s->windows[bsize][i];
1106 } else {
1107 /* overlap */
1108 n = (block_len / 2) - (prev_block_len / 2);
1109 for(i=0;i<n;i++)
1110 *--wptr = 1.0;
1111 for(i=0;i<prev_block_len;i++)
1112 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1113 for(i=0;i<n;i++)
1114 *--wptr = 0.0;
1115 }
1116 }
1117
1118
1119 for(ch = 0; ch < s->nb_channels; ch++) {
1120 if (s->channel_coded[ch]) {
1121 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1122 float *ptr;
1123 int i, n4, index, n;
1124
1125 n = s->block_len;
1126 n4 = s->block_len / 2;
1127 ff_imdct_calc(&s->mdct_ctx[bsize],
1128 output, s->coefs[ch], s->mdct_tmp);
1129
1130 /* XXX: optimize all that by build the window and
1131 multipying/adding at the same time */
1132 /* multiply by the window */
1133 for(i=0;i<n * 2;i++) {
1134 output[i] *= window[i];
1135 }
1136
1137 /* add in the frame */
1138 index = (s->frame_len / 2) + s->block_pos - n4;
1139 ptr = &s->frame_out[ch][index];
1140 for(i=0;i<n * 2;i++) {
1141 *ptr += output[i];
1142 ptr++;
1143 }
1144
1145 /* specific fast case for ms-stereo : add to second
1146 channel if it is not coded */
1147 if (s->ms_stereo && !s->channel_coded[1]) {
1148 ptr = &s->frame_out[1][index];
1149 for(i=0;i<n * 2;i++) {
1150 *ptr += output[i];
1151 ptr++;
1152 }
1153 }
1154 }
1155 }
1156 next:
1157 /* update block number */
1158 s->block_num++;
1159 s->block_pos += s->block_len;
1160 if (s->block_pos >= s->frame_len)
1161 return 1;
1162 else
1163 return 0;
1164 }
1165
1166 /* decode a frame of frame_len samples */
1167 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1168 {
1169 int ret, i, n, a, ch, incr;
1170 int16_t *ptr;
1171 float *iptr;
1172
1173 trace("***decode_frame: %d size=%d\n", frame_count++, s->frame_len);
1174
1175 /* read each block */
1176 s->block_num = 0;
1177 s->block_pos = 0;
1178 for(;;) {
1179 ret = wma_decode_block(s);
1180 if (ret < 0)
1181 return -1;
1182 if (ret)
1183 break;
1184 }
1185
1186 /* convert frame to integer */
1187 n = s->frame_len;
1188 incr = s->nb_channels;
1189 for(ch = 0; ch < s->nb_channels; ch++) {
1190 ptr = samples + ch;
1191 iptr = s->frame_out[ch];
1192
1193 for(i=0;i<n;i++) {
1194 a = lrintf(*iptr++);
1195 if (a > 32767)
1196 a = 32767;
1197 else if (a < -32768)
1198 a = -32768;
1199 *ptr = a;
1200 ptr += incr;
1201 }
1202 /* prepare for next block */
1203 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1204 s->frame_len * sizeof(float));
1205 /* XXX: suppress this */
1206 memset(&s->frame_out[ch][s->frame_len], 0,
1207 s->frame_len * sizeof(float));
1208 }
1209
1210 #ifdef DEBUG_TRACE
1211 dump_shorts("samples", samples, n * s->nb_channels);
1212 #endif
1213 return 0;
1214 }
1215
1216 static int wma_decode_superframe(AVCodecContext *avctx,
1217 void *data, int *data_size,
1218 UINT8 *buf, int buf_size)
1219 {
1220 WMADecodeContext *s = avctx->priv_data;
1221 int nb_frames, bit_offset, i, pos, len;
1222 uint8_t *q;
1223 int16_t *samples;
1224
1225 trace("***decode_superframe:\n");
1226
1227 samples = data;
1228
1229 init_get_bits(&s->gb, buf, buf_size);
1230
1231 if (s->use_bit_reservoir) {
1232 /* read super frame header */
1233 get_bits(&s->gb, 4); /* super frame index */
1234 nb_frames = get_bits(&s->gb, 4) - 1;
1235
1236 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1237
1238 if (s->last_superframe_len > 0) {
1239 // printf("skip=%d\n", s->last_bitoffset);
1240 /* add bit_offset bits to last frame */
1241 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1242 MAX_CODED_SUPERFRAME_SIZE)
1243 goto fail;
1244 q = s->last_superframe + s->last_superframe_len;
1245 len = bit_offset;
1246 while (len > 0) {
1247 *q++ = (get_bits)(&s->gb, 8);
1248 len -= 8;
1249 }
1250 if (len > 0) {
1251 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1252 }
1253
1254 /* XXX: bit_offset bits into last frame */
1255 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE);
1256 /* skip unused bits */
1257 if (s->last_bitoffset > 0)
1258 skip_bits(&s->gb, s->last_bitoffset);
1259 /* this frame is stored in the last superframe and in the
1260 current one */
1261 if (wma_decode_frame(s, samples) < 0)
1262 goto fail;
1263 samples += s->nb_channels * s->frame_len;
1264 }
1265
1266 /* read each frame starting from bit_offset */
1267 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1268 init_get_bits(&s->gb, buf + (pos >> 3), MAX_CODED_SUPERFRAME_SIZE - (pos >> 3));
1269 len = pos & 7;
1270 if (len > 0)
1271 skip_bits(&s->gb, len);
1272
1273 s->reset_block_lengths = 1;
1274 for(i=0;i<nb_frames;i++) {
1275 if (wma_decode_frame(s, samples) < 0)
1276 goto fail;
1277 samples += s->nb_channels * s->frame_len;
1278 }
1279
1280 /* we copy the end of the frame in the last frame buffer */
1281 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1282 s->last_bitoffset = pos & 7;
1283 pos >>= 3;
1284 len = buf_size - pos;
1285 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1286 goto fail;
1287 }
1288 s->last_superframe_len = len;
1289 memcpy(s->last_superframe, buf + pos, len);
1290 } else {
1291 /* single frame decode */
1292 if (wma_decode_frame(s, samples) < 0)
1293 goto fail;
1294 samples += s->nb_channels * s->frame_len;
1295 }
1296 *data_size = (int8_t *)samples - (int8_t *)data;
1297 return s->block_align;
1298 fail:
1299 /* when error, we reset the bit reservoir */
1300 s->last_superframe_len = 0;
1301 return -1;
1302 }
1303
1304 static int wma_decode_end(AVCodecContext *avctx)
1305 {
1306 WMADecodeContext *s = avctx->priv_data;
1307 int i;
1308
1309 for(i = 0; i < s->nb_block_sizes; i++)
1310 ff_mdct_end(&s->mdct_ctx[i]);
1311 for(i = 0; i < s->nb_block_sizes; i++)
1312 av_free(s->windows[i]);
1313
1314 if (s->use_exp_vlc) {
1315 free_vlc(&s->exp_vlc);
1316 }
1317 if (s->use_noise_coding) {
1318 free_vlc(&s->hgain_vlc);
1319 }
1320 for(i = 0;i < 2; i++) {
1321 free_vlc(&s->coef_vlc[i]);
1322 av_free(s->run_table[i]);
1323 av_free(s->level_table[i]);
1324 }
1325
1326 return 0;
1327 }
1328
1329 AVCodec wmav1_decoder =
1330 {
1331 "wmav1",
1332 CODEC_TYPE_AUDIO,
1333 CODEC_ID_WMAV1,
1334 sizeof(WMADecodeContext),
1335 wma_decode_init,
1336 NULL,
1337 wma_decode_end,
1338 wma_decode_superframe,
1339 };
1340
1341 AVCodec wmav2_decoder =
1342 {
1343 "wmav2",
1344 CODEC_TYPE_AUDIO,
1345 CODEC_ID_WMAV2,
1346 sizeof(WMADecodeContext),
1347 wma_decode_init,
1348 NULL,
1349 wma_decode_end,
1350 wma_decode_superframe,
1351 };