works with old gcc versions too
[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 //#define DEBUG_PARAMS
29 //#define DEBUG_TRACE
30
31 /* size of blocks */
32 #define BLOCK_MIN_BITS 7
33 #define BLOCK_MAX_BITS 11
34 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
35
36 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
37
38 /* XXX: find exact max size */
39 #define HIGH_BAND_MAX_SIZE 16
40
41 #define NB_LSP_COEFS 10
42
43 /* XXX: is it a suitable value ? */
44 #define MAX_CODED_SUPERFRAME_SIZE 4096
45
46 #define MAX_CHANNELS 2
47
48 #define NOISE_TAB_SIZE 8192
49
50 #define LSP_POW_BITS 7
51
52 typedef struct WMADecodeContext {
53 GetBitContext gb;
54 int sample_rate;
55 int nb_channels;
56 int bit_rate;
57 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
58 int block_align;
59 int use_bit_reservoir;
60 int use_variable_block_len;
61 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
62 int use_noise_coding; /* true if perceptual noise is added */
63 int byte_offset_bits;
64 VLC exp_vlc;
65 int exponent_sizes[BLOCK_NB_SIZES];
66 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
67 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
68 int coefs_start; /* first coded coef */
69 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
70 int exponent_high_sizes[BLOCK_NB_SIZES];
71 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
72 VLC hgain_vlc;
73
74 /* coded values in high bands */
75 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
76 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
77
78 /* there are two possible tables for spectral coefficients */
79 VLC coef_vlc[2];
80 uint16_t *run_table[2];
81 uint16_t *level_table[2];
82 /* frame info */
83 int frame_len; /* frame length in samples */
84 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
85 int nb_block_sizes; /* number of block sizes */
86 /* block info */
87 int reset_block_lengths;
88 int block_len_bits; /* log2 of current block length */
89 int next_block_len_bits; /* log2 of next block length */
90 int prev_block_len_bits; /* log2 of prev block length */
91 int block_len; /* block length in samples */
92 int block_num; /* block number in current frame */
93 int block_pos; /* current position in frame */
94 uint8_t ms_stereo; /* true if mid/side stereo mode */
95 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
96 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
97 float max_exponent[MAX_CHANNELS];
98 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
99 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
100 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
101 float *windows[BLOCK_NB_SIZES];
102 FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
103 /* output buffer for one frame and the last for IMDCT windowing */
104 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
105 /* last frame info */
106 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
107 int last_bitoffset;
108 int last_superframe_len;
109 float noise_table[NOISE_TAB_SIZE];
110 int noise_index;
111 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
112 /* lsp_to_curve tables */
113 float lsp_cos_table[BLOCK_MAX_SIZE];
114 float lsp_pow_e_table[256];
115 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
116 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
117 } WMADecodeContext;
118
119 typedef struct CoefVLCTable {
120 int n; /* total number of codes */
121 const uint32_t *huffcodes; /* VLC bit values */
122 const uint8_t *huffbits; /* VLC bit size */
123 const uint16_t *levels; /* table to build run/level tables */
124 } CoefVLCTable;
125
126 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
127
128 #include "wmadata.h"
129
130 #ifdef DEBUG_TRACE
131 #include <stdarg.h>
132 int frame_count;
133
134 static FILE *flog;
135
136 void trace(const char *fmt, ...)
137 {
138 va_list ap;
139
140
141 if (!flog) {
142 flog = fopen("/tmp/out.log", "w");
143 setlinebuf(flog);
144 }
145
146 va_start(ap, fmt);
147 vfprintf(flog, fmt, ap);
148 va_end(ap);
149 }
150
151 #define get_bits(s, n) get_bits_trace(s, n)
152 #define get_vlc(s, vlc) get_vlc_trace(s, vlc)
153
154 unsigned int get_bits_trace(GetBitContext *s, int n)
155 {
156 unsigned int val;
157 val = (get_bits)(s, n);
158 trace("get_bits(%d) : 0x%x\n", n, val);
159 return val;
160 }
161
162 static int get_vlc_trace(GetBitContext *s, VLC *vlc)
163 {
164 int code;
165 code = (get_vlc)(s, vlc);
166 trace("get_vlc() : %d\n", code);
167 return code;
168 }
169
170 static void dump_shorts(const char *name, const short *tab, int n)
171 {
172 int i;
173
174 trace("%s[%d]:\n", name, n);
175 for(i=0;i<n;i++) {
176 if ((i & 7) == 0)
177 trace("%4d: ", i);
178 trace(" %5d.0", tab[i]);
179 if ((i & 7) == 7)
180 trace("\n");
181 }
182 }
183
184 static void dump_floats(const char *name, int prec, const float *tab, int n)
185 {
186 int i;
187
188 trace("%s[%d]:\n", name, n);
189 for(i=0;i<n;i++) {
190 if ((i & 7) == 0)
191 trace("%4d: ", i);
192 trace(" %8.*f", prec, tab[i]);
193 if ((i & 7) == 7)
194 trace("\n");
195 }
196 if ((i & 7) != 0)
197 trace("\n");
198 }
199
200 #else
201
202 #define trace(fmt, args...)
203
204 #endif
205
206 /* XXX: use same run/length optimization as mpeg decoders */
207 static void init_coef_vlc(VLC *vlc,
208 uint16_t **prun_table, uint16_t **plevel_table,
209 const CoefVLCTable *vlc_table)
210 {
211 int n = vlc_table->n;
212 const uint8_t *table_bits = vlc_table->huffbits;
213 const uint32_t *table_codes = vlc_table->huffcodes;
214 const uint16_t *levels_table = vlc_table->levels;
215 uint16_t *run_table, *level_table;
216 const uint16_t *p;
217 int i, l, j, level;
218
219 init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4);
220
221 run_table = av_malloc(n * sizeof(uint16_t));
222 level_table = av_malloc(n * sizeof(uint16_t));
223 p = levels_table;
224 i = 2;
225 level = 1;
226 while (i < n) {
227 l = *p++;
228 for(j=0;j<l;j++) {
229 run_table[i] = j;
230 level_table[i] = level;
231 i++;
232 }
233 level++;
234 }
235 *prun_table = run_table;
236 *plevel_table = level_table;
237 }
238
239 static int wma_decode_init(AVCodecContext * avctx)
240 {
241 WMADecodeContext *s = avctx->priv_data;
242 int i, flags1, flags2;
243 float *window;
244 uint8_t *extradata;
245 float bps1, high_freq, bps;
246 int sample_rate1;
247 int coef_vlc_table;
248
249 s->sample_rate = avctx->sample_rate;
250 s->nb_channels = avctx->channels;
251 s->bit_rate = avctx->bit_rate;
252 s->block_align = avctx->block_align;
253
254 if (avctx->codec->id == CODEC_ID_WMAV1) {
255 s->version = 1;
256 } else {
257 s->version = 2;
258 }
259
260 /* extract flag infos */
261 flags1 = 0;
262 flags2 = 0;
263 extradata = avctx->extradata;
264 if (s->version == 1 && avctx->extradata_size >= 4) {
265 flags1 = extradata[0] | (extradata[1] << 8);
266 flags2 = extradata[2] | (extradata[3] << 8);
267 } else if (s->version == 2 && avctx->extradata_size >= 6) {
268 flags1 = extradata[0] | (extradata[1] << 8) |
269 (extradata[2] << 16) | (extradata[3] << 24);
270 flags2 = extradata[4] | (extradata[5] << 8);
271 }
272 s->use_exp_vlc = flags2 & 0x0001;
273 s->use_bit_reservoir = flags2 & 0x0002;
274 s->use_variable_block_len = flags2 & 0x0004;
275
276 /* compute MDCT block size */
277 if (s->sample_rate <= 16000) {
278 s->frame_len_bits = 9;
279 } else if (s->sample_rate <= 22050 ||
280 (s->sample_rate <= 32000 && s->version == 1)) {
281 s->frame_len_bits = 10;
282 } else {
283 s->frame_len_bits = 11;
284 }
285 s->frame_len = 1 << s->frame_len_bits;
286 if (s->use_variable_block_len) {
287 int nb_max, nb;
288 nb = ((flags2 >> 3) & 3) + 1;
289 if ((s->bit_rate / s->nb_channels) >= 32000)
290 nb += 2;
291 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
292 if (nb > nb_max)
293 nb = nb_max;
294 s->nb_block_sizes = nb + 1;
295 } else {
296 s->nb_block_sizes = 1;
297 }
298
299 /* init rate dependant parameters */
300 s->use_noise_coding = 1;
301 high_freq = s->sample_rate * 0.5;
302
303 /* if version 2, then the rates are normalized */
304 sample_rate1 = s->sample_rate;
305 if (s->version == 2) {
306 if (sample_rate1 >= 44100)
307 sample_rate1 = 44100;
308 else if (sample_rate1 >= 22050)
309 sample_rate1 = 22050;
310 else if (sample_rate1 >= 16000)
311 sample_rate1 = 16000;
312 else if (sample_rate1 >= 11025)
313 sample_rate1 = 11025;
314 else if (sample_rate1 >= 8000)
315 sample_rate1 = 8000;
316 }
317
318 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
319 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
320
321 /* compute high frequency value and choose if noise coding should
322 be activated */
323 bps1 = bps;
324 if (s->nb_channels == 2)
325 bps1 = bps * 1.6;
326 if (sample_rate1 == 44100) {
327 if (bps1 >= 0.61)
328 s->use_noise_coding = 0;
329 else
330 high_freq = high_freq * 0.4;
331 } else if (sample_rate1 == 22050) {
332 if (bps1 >= 1.16)
333 s->use_noise_coding = 0;
334 else if (bps1 >= 0.72)
335 high_freq = high_freq * 0.7;
336 else
337 high_freq = high_freq * 0.6;
338 } else if (sample_rate1 == 16000) {
339 if (bps > 0.5)
340 high_freq = high_freq * 0.5;
341 else
342 high_freq = high_freq * 0.3;
343 } else if (sample_rate1 == 11025) {
344 high_freq = high_freq * 0.7;
345 } else if (sample_rate1 == 8000) {
346 if (bps <= 0.625) {
347 high_freq = high_freq * 0.5;
348 } else if (bps > 0.75) {
349 s->use_noise_coding = 0;
350 } else {
351 high_freq = high_freq * 0.65;
352 }
353 } else {
354 if (bps >= 0.8) {
355 high_freq = high_freq * 0.75;
356 } else if (bps >= 0.6) {
357 high_freq = high_freq * 0.6;
358 } else {
359 high_freq = high_freq * 0.5;
360 }
361 }
362 #ifdef DEBUG_PARAMS
363 printf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
364 printf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
365 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
366 s->block_align);
367 printf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
368 bps, bps1, high_freq, s->byte_offset_bits);
369 printf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
370 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
371 #endif
372
373 /* compute the scale factor band sizes for each MDCT block size */
374 {
375 int a, b, pos, lpos, k, block_len, i, j, n;
376 const uint8_t *table;
377
378 if (s->version == 1) {
379 s->coefs_start = 3;
380 } else {
381 s->coefs_start = 0;
382 }
383 for(k = 0; k < s->nb_block_sizes; k++) {
384 block_len = s->frame_len >> k;
385
386 if (s->version == 1) {
387 lpos = 0;
388 for(i=0;i<25;i++) {
389 a = wma_critical_freqs[i];
390 b = s->sample_rate;
391 pos = ((block_len * 2 * a) + (b >> 1)) / b;
392 if (pos > block_len)
393 pos = block_len;
394 s->exponent_bands[0][i] = pos - lpos;
395 if (pos >= block_len) {
396 i++;
397 break;
398 }
399 lpos = pos;
400 }
401 s->exponent_sizes[0] = i;
402 } else {
403 /* hardcoded tables */
404 table = NULL;
405 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
406 if (a < 3) {
407 if (s->sample_rate >= 44100)
408 table = exponent_band_44100[a];
409 else if (s->sample_rate >= 32000)
410 table = exponent_band_32000[a];
411 else if (s->sample_rate >= 22050)
412 table = exponent_band_22050[a];
413 }
414 if (table) {
415 n = *table++;
416 for(i=0;i<n;i++)
417 s->exponent_bands[k][i] = table[i];
418 s->exponent_sizes[k] = n;
419 } else {
420 j = 0;
421 lpos = 0;
422 for(i=0;i<25;i++) {
423 a = wma_critical_freqs[i];
424 b = s->sample_rate;
425 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
426 pos <<= 2;
427 if (pos > block_len)
428 pos = block_len;
429 if (pos > lpos)
430 s->exponent_bands[k][j++] = pos - lpos;
431 if (pos >= block_len)
432 break;
433 lpos = pos;
434 }
435 s->exponent_sizes[k] = j;
436 }
437 }
438
439 /* max number of coefs */
440 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
441 /* high freq computation */
442 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
443 s->sample_rate + 0.5);
444 n = s->exponent_sizes[k];
445 j = 0;
446 pos = 0;
447 for(i=0;i<n;i++) {
448 int start, end;
449 start = pos;
450 pos += s->exponent_bands[k][i];
451 end = pos;
452 if (start < s->high_band_start[k])
453 start = s->high_band_start[k];
454 if (end > s->coefs_end[k])
455 end = s->coefs_end[k];
456 if (end > start)
457 s->exponent_high_bands[k][j++] = end - start;
458 }
459 s->exponent_high_sizes[k] = j;
460 #if 0
461 trace("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
462 s->frame_len >> k,
463 s->coefs_end[k],
464 s->high_band_start[k],
465 s->exponent_high_sizes[k]);
466 for(j=0;j<s->exponent_high_sizes[k];j++)
467 trace(" %d", s->exponent_high_bands[k][j]);
468 trace("\n");
469 #endif
470 }
471 }
472
473 #ifdef DEBUG_TRACE
474 {
475 int i, j;
476 for(i = 0; i < s->nb_block_sizes; i++) {
477 trace("%5d: n=%2d:",
478 s->frame_len >> i,
479 s->exponent_sizes[i]);
480 for(j=0;j<s->exponent_sizes[i];j++)
481 trace(" %d", s->exponent_bands[i][j]);
482 trace("\n");
483 }
484 }
485 #endif
486
487 /* init MDCT */
488 for(i = 0; i < s->nb_block_sizes; i++)
489 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
490
491 /* init MDCT windows : simple sinus window */
492 for(i = 0; i < s->nb_block_sizes; i++) {
493 int n, j;
494 float alpha;
495 n = 1 << (s->frame_len_bits - i);
496 window = av_malloc(sizeof(float) * n);
497 alpha = M_PI / (2.0 * n);
498 for(j=0;j<n;j++) {
499 window[n - j - 1] = sin((j + 0.5) * alpha);
500 }
501 s->windows[i] = window;
502 }
503
504 s->reset_block_lengths = 1;
505
506 if (s->use_noise_coding) {
507
508 /* init the noise generator */
509 if (s->use_exp_vlc)
510 s->noise_mult = 0.02;
511 else
512 s->noise_mult = 0.04;
513
514 #if defined(DEBUG_TRACE)
515 for(i=0;i<NOISE_TAB_SIZE;i++)
516 s->noise_table[i] = 1.0 * s->noise_mult;
517 #else
518 {
519 unsigned int seed;
520 float norm;
521 seed = 1;
522 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
523 for(i=0;i<NOISE_TAB_SIZE;i++) {
524 seed = seed * 314159 + 1;
525 s->noise_table[i] = (float)((int)seed) * norm;
526 }
527 }
528 #endif
529 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
530 hgain_huffbits, 1, 1,
531 hgain_huffcodes, 2, 2);
532 }
533
534 if (s->use_exp_vlc) {
535 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
536 scale_huffbits, 1, 1,
537 scale_huffcodes, 4, 4);
538 } else {
539 wma_lsp_to_curve_init(s, s->frame_len);
540 }
541
542 /* choose the VLC tables for the coefficients */
543 coef_vlc_table = 2;
544 if (s->sample_rate >= 32000) {
545 if (bps1 < 0.72)
546 coef_vlc_table = 0;
547 else if (bps1 < 1.16)
548 coef_vlc_table = 1;
549 }
550
551 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
552 &coef_vlcs[coef_vlc_table * 2]);
553 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
554 &coef_vlcs[coef_vlc_table * 2 + 1]);
555 return 0;
556 }
557
558 /* interpolate values for a bigger or smaller block. The block must
559 have multiple sizes */
560 static void interpolate_array(float *scale, int old_size, int new_size)
561 {
562 int i, j, jincr, k;
563 float v;
564
565 if (new_size > old_size) {
566 jincr = new_size / old_size;
567 j = new_size;
568 for(i = old_size - 1; i >=0; i--) {
569 v = scale[i];
570 k = jincr;
571 do {
572 scale[--j] = v;
573 } while (--k);
574 }
575 } else if (new_size < old_size) {
576 j = 0;
577 jincr = old_size / new_size;
578 for(i = 0; i < new_size; i++) {
579 scale[i] = scale[j];
580 j += jincr;
581 }
582 }
583 }
584
585 /* compute x^-0.25 with an exponent and mantissa table. We use linear
586 interpolation to reduce the mantissa table size at a small speed
587 expense (linear interpolation approximately doubles the number of
588 bits of precision). */
589 static inline float pow_m1_4(WMADecodeContext *s, float x)
590 {
591 union {
592 float f;
593 unsigned int v;
594 } u, t;
595 unsigned int e, m;
596 float a, b;
597
598 u.f = x;
599 e = u.v >> 23;
600 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
601 /* build interpolation scale: 1 <= t < 2. */
602 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
603 a = s->lsp_pow_m_table1[m];
604 b = s->lsp_pow_m_table2[m];
605 return s->lsp_pow_e_table[e] * (a + b * t.f);
606 }
607
608 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
609 {
610 float wdel, a, b;
611 int i, e, m;
612
613 wdel = M_PI / frame_len;
614 for(i=0;i<frame_len;i++)
615 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
616
617 /* tables for x^-0.25 computation */
618 for(i=0;i<256;i++) {
619 e = i - 126;
620 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
621 }
622
623 /* NOTE: these two tables are needed to avoid two operations in
624 pow_m1_4 */
625 b = 1.0;
626 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
627 m = (1 << LSP_POW_BITS) + i;
628 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
629 a = pow(a, -0.25);
630 s->lsp_pow_m_table1[i] = 2 * a - b;
631 s->lsp_pow_m_table2[i] = b - a;
632 b = a;
633 }
634 #if 0
635 for(i=1;i<20;i++) {
636 float v, r1, r2;
637 v = 5.0 / i;
638 r1 = pow_m1_4(s, v);
639 r2 = pow(v,-0.25);
640 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
641 }
642 #endif
643 }
644
645 /* NOTE: We use the same code as Vorbis here */
646 /* XXX: optimize it further with SSE/3Dnow */
647 static void wma_lsp_to_curve(WMADecodeContext *s,
648 float *out, float *val_max_ptr,
649 int n, float *lsp)
650 {
651 int i, j;
652 float p, q, w, v, val_max;
653
654 val_max = 0;
655 for(i=0;i<n;i++) {
656 p = 0.5f;
657 q = 0.5f;
658 w = s->lsp_cos_table[i];
659 for(j=1;j<NB_LSP_COEFS;j+=2){
660 q *= w - lsp[j - 1];
661 p *= w - lsp[j];
662 }
663 p *= p * (2.0f - w);
664 q *= q * (2.0f + w);
665 v = p + q;
666 v = pow_m1_4(s, v);
667 if (v > val_max)
668 val_max = v;
669 out[i] = v;
670 }
671 *val_max_ptr = val_max;
672 }
673
674 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
675 static void decode_exp_lsp(WMADecodeContext *s, int ch)
676 {
677 float lsp_coefs[NB_LSP_COEFS];
678 int val, i;
679
680 for(i = 0; i < NB_LSP_COEFS; i++) {
681 if (i == 0 || i >= 8)
682 val = get_bits(&s->gb, 3);
683 else
684 val = get_bits(&s->gb, 4);
685 lsp_coefs[i] = lsp_codebook[i][val];
686 }
687
688 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
689 s->block_len, lsp_coefs);
690 }
691
692 /* decode exponents coded with VLC codes */
693 static int decode_exp_vlc(WMADecodeContext *s, int ch)
694 {
695 int last_exp, n, code;
696 const uint16_t *ptr, *band_ptr;
697 float v, *q, max_scale, *q_end;
698
699 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
700 ptr = band_ptr;
701 q = s->exponents[ch];
702 q_end = q + s->block_len;
703 max_scale = 0;
704 if (s->version == 1) {
705 last_exp = get_bits(&s->gb, 5) + 10;
706 /* XXX: use a table */
707 v = pow(10, last_exp * (1.0 / 16.0));
708 max_scale = v;
709 n = *ptr++;
710 do {
711 *q++ = v;
712 } while (--n);
713 }
714 last_exp = 36;
715 while (q < q_end) {
716 code = get_vlc(&s->gb, &s->exp_vlc);
717 if (code < 0)
718 return -1;
719 /* NOTE: this offset is the same as MPEG4 AAC ! */
720 last_exp += code - 60;
721 /* XXX: use a table */
722 v = pow(10, last_exp * (1.0 / 16.0));
723 if (v > max_scale)
724 max_scale = v;
725 n = *ptr++;
726 do {
727 *q++ = v;
728 } while (--n);
729 }
730 s->max_exponent[ch] = max_scale;
731 return 0;
732 }
733
734 /* return 0 if OK. return 1 if last block of frame. return -1 if
735 unrecorrable error. */
736 static int wma_decode_block(WMADecodeContext *s)
737 {
738 int n, v, a, ch, code, bsize;
739 int coef_nb_bits, total_gain, parse_exponents;
740 float window[BLOCK_MAX_SIZE * 2];
741 int nb_coefs[MAX_CHANNELS];
742 float mdct_norm;
743
744 trace("***decode_block: %d:%d\n", frame_count - 1, s->block_num);
745
746 /* compute current block length */
747 if (s->use_variable_block_len) {
748 n = av_log2(s->nb_block_sizes - 1) + 1;
749
750 if (s->reset_block_lengths) {
751 s->reset_block_lengths = 0;
752 v = get_bits(&s->gb, n);
753 if (v >= s->nb_block_sizes)
754 return -1;
755 s->prev_block_len_bits = s->frame_len_bits - v;
756 v = get_bits(&s->gb, n);
757 if (v >= s->nb_block_sizes)
758 return -1;
759 s->block_len_bits = s->frame_len_bits - v;
760 } else {
761 /* update block lengths */
762 s->prev_block_len_bits = s->block_len_bits;
763 s->block_len_bits = s->next_block_len_bits;
764 }
765 v = get_bits(&s->gb, n);
766 if (v >= s->nb_block_sizes)
767 return -1;
768 s->next_block_len_bits = s->frame_len_bits - v;
769 } else {
770 /* fixed block len */
771 s->next_block_len_bits = s->frame_len_bits;
772 s->prev_block_len_bits = s->frame_len_bits;
773 s->block_len_bits = s->frame_len_bits;
774 }
775
776 /* now check if the block length is coherent with the frame length */
777 s->block_len = 1 << s->block_len_bits;
778 if ((s->block_pos + s->block_len) > s->frame_len)
779 return -1;
780
781 if (s->nb_channels == 2) {
782 s->ms_stereo = get_bits(&s->gb, 1);
783 }
784 v = 0;
785 for(ch = 0; ch < s->nb_channels; ch++) {
786 a = get_bits(&s->gb, 1);
787 s->channel_coded[ch] = a;
788 v |= a;
789 }
790 /* if no channel coded, no need to go further */
791 /* XXX: fix potential framing problems */
792 if (!v)
793 goto next;
794
795 bsize = s->frame_len_bits - s->block_len_bits;
796
797 /* read total gain and extract corresponding number of bits for
798 coef escape coding */
799 total_gain = 1;
800 for(;;) {
801 a = get_bits(&s->gb, 7);
802 total_gain += a;
803 if (a != 127)
804 break;
805 }
806
807 if (total_gain < 15)
808 coef_nb_bits = 13;
809 else if (total_gain < 32)
810 coef_nb_bits = 12;
811 else if (total_gain < 40)
812 coef_nb_bits = 11;
813 else if (total_gain < 45)
814 coef_nb_bits = 10;
815 else
816 coef_nb_bits = 9;
817
818 /* compute number of coefficients */
819 n = s->coefs_end[bsize] - s->coefs_start;
820 for(ch = 0; ch < s->nb_channels; ch++)
821 nb_coefs[ch] = n;
822
823 /* complex coding */
824 if (s->use_noise_coding) {
825
826 for(ch = 0; ch < s->nb_channels; ch++) {
827 if (s->channel_coded[ch]) {
828 int i, n, a;
829 n = s->exponent_high_sizes[bsize];
830 for(i=0;i<n;i++) {
831 a = get_bits(&s->gb, 1);
832 s->high_band_coded[ch][i] = a;
833 /* if noise coding, the coefficients are not transmitted */
834 if (a)
835 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
836 }
837 }
838 }
839 for(ch = 0; ch < s->nb_channels; ch++) {
840 if (s->channel_coded[ch]) {
841 int i, n, val, code;
842
843 n = s->exponent_high_sizes[bsize];
844 val = (int)0x80000000;
845 for(i=0;i<n;i++) {
846 if (s->high_band_coded[ch][i]) {
847 if (val == (int)0x80000000) {
848 val = get_bits(&s->gb, 7) - 19;
849 } else {
850 code = get_vlc(&s->gb, &s->hgain_vlc);
851 if (code < 0)
852 return -1;
853 val += code - 18;
854 }
855 s->high_band_values[ch][i] = val;
856 }
857 }
858 }
859 }
860 }
861
862 /* exposant can be interpolated in short blocks. */
863 parse_exponents = 1;
864 if (s->block_len_bits != s->frame_len_bits) {
865 parse_exponents = get_bits(&s->gb, 1);
866 }
867
868 if (parse_exponents) {
869 for(ch = 0; ch < s->nb_channels; ch++) {
870 if (s->channel_coded[ch]) {
871 if (s->use_exp_vlc) {
872 if (decode_exp_vlc(s, ch) < 0)
873 return -1;
874 } else {
875 decode_exp_lsp(s, ch);
876 }
877 }
878 }
879 } else {
880 for(ch = 0; ch < s->nb_channels; ch++) {
881 if (s->channel_coded[ch]) {
882 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
883 s->block_len);
884 }
885 }
886 }
887
888 /* parse spectral coefficients : just RLE encoding */
889 for(ch = 0; ch < s->nb_channels; ch++) {
890 if (s->channel_coded[ch]) {
891 VLC *coef_vlc;
892 int level, run, sign, tindex;
893 int16_t *ptr, *eptr;
894 const int16_t *level_table, *run_table;
895
896 /* special VLC tables are used for ms stereo because
897 there is potentially less energy there */
898 tindex = (ch == 1 && s->ms_stereo);
899 coef_vlc = &s->coef_vlc[tindex];
900 run_table = s->run_table[tindex];
901 level_table = s->level_table[tindex];
902 /* XXX: optimize */
903 ptr = &s->coefs1[ch][0];
904 eptr = ptr + nb_coefs[ch];
905 memset(ptr, 0, s->block_len * sizeof(int16_t));
906 for(;;) {
907 code = get_vlc(&s->gb, coef_vlc);
908 if (code < 0)
909 return -1;
910 if (code == 1) {
911 /* EOB */
912 break;
913 } else if (code == 0) {
914 /* escape */
915 level = get_bits(&s->gb, coef_nb_bits);
916 /* NOTE: this is rather suboptimal. reading
917 block_len_bits would be better */
918 run = get_bits(&s->gb, s->frame_len_bits);
919 } else {
920 /* normal code */
921 run = run_table[code];
922 level = level_table[code];
923 }
924 sign = get_bits(&s->gb, 1);
925 if (!sign)
926 level = -level;
927 ptr += run;
928 if (ptr >= eptr)
929 return -1;
930 *ptr++ = level;
931 /* NOTE: EOB can be omitted */
932 if (ptr >= eptr)
933 break;
934 }
935 }
936 if (s->version == 1 && s->nb_channels >= 2) {
937 align_get_bits(&s->gb);
938 }
939 }
940
941 /* normalize */
942 {
943 int n4 = s->block_len / 2;
944 mdct_norm = 1.0 / (float)n4;
945 if (s->version == 1) {
946 mdct_norm *= sqrt(n4);
947 }
948 }
949
950 /* finally compute the MDCT coefficients */
951 for(ch = 0; ch < s->nb_channels; ch++) {
952 if (s->channel_coded[ch]) {
953 int16_t *coefs1;
954 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
955 int i, j, n, n1, last_high_band;
956 float exp_power[HIGH_BAND_MAX_SIZE];
957
958 coefs1 = s->coefs1[ch];
959 exponents = s->exponents[ch];
960 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
961 mult *= mdct_norm;
962 coefs = s->coefs[ch];
963 if (s->use_noise_coding) {
964 mult1 = mult;
965 /* very low freqs : noise */
966 for(i = 0;i < s->coefs_start; i++) {
967 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
968 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
969 }
970
971 n1 = s->exponent_high_sizes[bsize];
972
973 /* compute power of high bands */
974 exp_ptr = exponents +
975 s->high_band_start[bsize] -
976 s->coefs_start;
977 last_high_band = 0; /* avoid warning */
978 for(j=0;j<n1;j++) {
979 n = s->exponent_high_bands[s->frame_len_bits -
980 s->block_len_bits][j];
981 if (s->high_band_coded[ch][j]) {
982 float e2, v;
983 e2 = 0;
984 for(i = 0;i < n; i++) {
985 v = exp_ptr[i];
986 e2 += v * v;
987 }
988 exp_power[j] = e2 / n;
989 last_high_band = j;
990 trace("%d: power=%f (%d)\n", j, exp_power[j], n);
991 }
992 exp_ptr += n;
993 }
994
995 /* main freqs and high freqs */
996 for(j=-1;j<n1;j++) {
997 if (j < 0) {
998 n = s->high_band_start[bsize] -
999 s->coefs_start;
1000 } else {
1001 n = s->exponent_high_bands[s->frame_len_bits -
1002 s->block_len_bits][j];
1003 }
1004 if (j >= 0 && s->high_band_coded[ch][j]) {
1005 /* use noise with specified power */
1006 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
1007 /* XXX: use a table */
1008 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
1009 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
1010 mult1 *= mdct_norm;
1011 for(i = 0;i < n; i++) {
1012 noise = s->noise_table[s->noise_index];
1013 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1014 *coefs++ = (*exponents++) * noise * mult1;
1015 }
1016 } else {
1017 /* coded values + small noise */
1018 for(i = 0;i < n; i++) {
1019 noise = s->noise_table[s->noise_index];
1020 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1021 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
1022 }
1023 }
1024 }
1025
1026 /* very high freqs : noise */
1027 n = s->block_len - s->coefs_end[bsize];
1028 mult1 = mult * exponents[-1];
1029 for(i = 0; i < n; i++) {
1030 *coefs++ = s->noise_table[s->noise_index] * mult1;
1031 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1032 }
1033 } else {
1034 /* XXX: optimize more */
1035 for(i = 0;i < s->coefs_start; i++)
1036 *coefs++ = 0.0;
1037 n = nb_coefs[ch];
1038 for(i = 0;i < n; i++) {
1039 *coefs++ = coefs1[i] * exponents[i] * mult;
1040 }
1041 n = s->block_len - s->coefs_end[bsize];
1042 for(i = 0;i < n; i++)
1043 *coefs++ = 0.0;
1044 }
1045 }
1046 }
1047
1048 #ifdef DEBUG_TRACE
1049 for(ch = 0; ch < s->nb_channels; ch++) {
1050 if (s->channel_coded[ch]) {
1051 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1052 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1053 }
1054 }
1055 #endif
1056
1057 if (s->ms_stereo && s->channel_coded[1]) {
1058 float a, b;
1059 int i;
1060
1061 /* nominal case for ms stereo: we do it before mdct */
1062 /* no need to optimize this case because it should almost
1063 never happen */
1064 if (!s->channel_coded[0]) {
1065 #ifdef DEBUG_TRACE
1066 trace("rare ms-stereo case happened\n");
1067 #endif
1068 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1069 s->channel_coded[0] = 1;
1070 }
1071
1072 for(i = 0; i < s->block_len; i++) {
1073 a = s->coefs[0][i];
1074 b = s->coefs[1][i];
1075 s->coefs[0][i] = a + b;
1076 s->coefs[1][i] = a - b;
1077 }
1078 }
1079
1080 /* build the window : we ensure that when the windows overlap
1081 their squared sum is always 1 (MDCT reconstruction rule) */
1082 /* XXX: merge with output */
1083 {
1084 int i, next_block_len, block_len, prev_block_len, n;
1085 float *wptr;
1086
1087 block_len = s->block_len;
1088 prev_block_len = 1 << s->prev_block_len_bits;
1089 next_block_len = 1 << s->next_block_len_bits;
1090
1091 /* right part */
1092 wptr = window + block_len;
1093 if (block_len <= next_block_len) {
1094 for(i=0;i<block_len;i++)
1095 *wptr++ = s->windows[bsize][i];
1096 } else {
1097 /* overlap */
1098 n = (block_len / 2) - (next_block_len / 2);
1099 for(i=0;i<n;i++)
1100 *wptr++ = 1.0;
1101 for(i=0;i<next_block_len;i++)
1102 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1103 for(i=0;i<n;i++)
1104 *wptr++ = 0.0;
1105 }
1106
1107 /* left part */
1108 wptr = window + block_len;
1109 if (block_len <= prev_block_len) {
1110 for(i=0;i<block_len;i++)
1111 *--wptr = s->windows[bsize][i];
1112 } else {
1113 /* overlap */
1114 n = (block_len / 2) - (prev_block_len / 2);
1115 for(i=0;i<n;i++)
1116 *--wptr = 1.0;
1117 for(i=0;i<prev_block_len;i++)
1118 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1119 for(i=0;i<n;i++)
1120 *--wptr = 0.0;
1121 }
1122 }
1123
1124
1125 for(ch = 0; ch < s->nb_channels; ch++) {
1126 if (s->channel_coded[ch]) {
1127 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1128 float *ptr;
1129 int i, n4, index, n;
1130
1131 n = s->block_len;
1132 n4 = s->block_len / 2;
1133 ff_imdct_calc(&s->mdct_ctx[bsize],
1134 output, s->coefs[ch], s->mdct_tmp);
1135
1136 /* XXX: optimize all that by build the window and
1137 multipying/adding at the same time */
1138 /* multiply by the window */
1139 for(i=0;i<n * 2;i++) {
1140 output[i] *= window[i];
1141 }
1142
1143 /* add in the frame */
1144 index = (s->frame_len / 2) + s->block_pos - n4;
1145 ptr = &s->frame_out[ch][index];
1146 for(i=0;i<n * 2;i++) {
1147 *ptr += output[i];
1148 ptr++;
1149 }
1150
1151 /* specific fast case for ms-stereo : add to second
1152 channel if it is not coded */
1153 if (s->ms_stereo && !s->channel_coded[1]) {
1154 ptr = &s->frame_out[1][index];
1155 for(i=0;i<n * 2;i++) {
1156 *ptr += output[i];
1157 ptr++;
1158 }
1159 }
1160 }
1161 }
1162 next:
1163 /* update block number */
1164 s->block_num++;
1165 s->block_pos += s->block_len;
1166 if (s->block_pos >= s->frame_len)
1167 return 1;
1168 else
1169 return 0;
1170 }
1171
1172 /* decode a frame of frame_len samples */
1173 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1174 {
1175 int ret, i, n, a, ch, incr;
1176 int16_t *ptr;
1177 float *iptr;
1178
1179 trace("***decode_frame: %d size=%d\n", frame_count++, s->frame_len);
1180
1181 /* read each block */
1182 s->block_num = 0;
1183 s->block_pos = 0;
1184 for(;;) {
1185 ret = wma_decode_block(s);
1186 if (ret < 0)
1187 return -1;
1188 if (ret)
1189 break;
1190 }
1191
1192 /* convert frame to integer */
1193 n = s->frame_len;
1194 incr = s->nb_channels;
1195 for(ch = 0; ch < s->nb_channels; ch++) {
1196 ptr = samples + ch;
1197 iptr = s->frame_out[ch];
1198
1199 for(i=0;i<n;i++) {
1200 a = lrintf(*iptr++);
1201 if (a > 32767)
1202 a = 32767;
1203 else if (a < -32768)
1204 a = -32768;
1205 *ptr = a;
1206 ptr += incr;
1207 }
1208 /* prepare for next block */
1209 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1210 s->frame_len * sizeof(float));
1211 /* XXX: suppress this */
1212 memset(&s->frame_out[ch][s->frame_len], 0,
1213 s->frame_len * sizeof(float));
1214 }
1215
1216 #ifdef DEBUG_TRACE
1217 dump_shorts("samples", samples, n * s->nb_channels);
1218 #endif
1219 return 0;
1220 }
1221
1222 static int wma_decode_superframe(AVCodecContext *avctx,
1223 void *data, int *data_size,
1224 uint8_t *buf, int buf_size)
1225 {
1226 WMADecodeContext *s = avctx->priv_data;
1227 int nb_frames, bit_offset, i, pos, len;
1228 uint8_t *q;
1229 int16_t *samples;
1230
1231 trace("***decode_superframe:\n");
1232
1233 samples = data;
1234
1235 init_get_bits(&s->gb, buf, buf_size*8);
1236
1237 if (s->use_bit_reservoir) {
1238 /* read super frame header */
1239 get_bits(&s->gb, 4); /* super frame index */
1240 nb_frames = get_bits(&s->gb, 4) - 1;
1241
1242 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1243
1244 if (s->last_superframe_len > 0) {
1245 // printf("skip=%d\n", s->last_bitoffset);
1246 /* add bit_offset bits to last frame */
1247 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1248 MAX_CODED_SUPERFRAME_SIZE)
1249 goto fail;
1250 q = s->last_superframe + s->last_superframe_len;
1251 len = bit_offset;
1252 while (len > 0) {
1253 *q++ = (get_bits)(&s->gb, 8);
1254 len -= 8;
1255 }
1256 if (len > 0) {
1257 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1258 }
1259
1260 /* XXX: bit_offset bits into last frame */
1261 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1262 /* skip unused bits */
1263 if (s->last_bitoffset > 0)
1264 skip_bits(&s->gb, s->last_bitoffset);
1265 /* this frame is stored in the last superframe and in the
1266 current one */
1267 if (wma_decode_frame(s, samples) < 0)
1268 goto fail;
1269 samples += s->nb_channels * s->frame_len;
1270 }
1271
1272 /* read each frame starting from bit_offset */
1273 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1274 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1275 len = pos & 7;
1276 if (len > 0)
1277 skip_bits(&s->gb, len);
1278
1279 s->reset_block_lengths = 1;
1280 for(i=0;i<nb_frames;i++) {
1281 if (wma_decode_frame(s, samples) < 0)
1282 goto fail;
1283 samples += s->nb_channels * s->frame_len;
1284 }
1285
1286 /* we copy the end of the frame in the last frame buffer */
1287 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1288 s->last_bitoffset = pos & 7;
1289 pos >>= 3;
1290 len = buf_size - pos;
1291 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1292 goto fail;
1293 }
1294 s->last_superframe_len = len;
1295 memcpy(s->last_superframe, buf + pos, len);
1296 } else {
1297 /* single frame decode */
1298 if (wma_decode_frame(s, samples) < 0)
1299 goto fail;
1300 samples += s->nb_channels * s->frame_len;
1301 }
1302 *data_size = (int8_t *)samples - (int8_t *)data;
1303 return s->block_align;
1304 fail:
1305 /* when error, we reset the bit reservoir */
1306 s->last_superframe_len = 0;
1307 return -1;
1308 }
1309
1310 static int wma_decode_end(AVCodecContext *avctx)
1311 {
1312 WMADecodeContext *s = avctx->priv_data;
1313 int i;
1314
1315 for(i = 0; i < s->nb_block_sizes; i++)
1316 ff_mdct_end(&s->mdct_ctx[i]);
1317 for(i = 0; i < s->nb_block_sizes; i++)
1318 av_free(s->windows[i]);
1319
1320 if (s->use_exp_vlc) {
1321 free_vlc(&s->exp_vlc);
1322 }
1323 if (s->use_noise_coding) {
1324 free_vlc(&s->hgain_vlc);
1325 }
1326 for(i = 0;i < 2; i++) {
1327 free_vlc(&s->coef_vlc[i]);
1328 av_free(s->run_table[i]);
1329 av_free(s->level_table[i]);
1330 }
1331
1332 return 0;
1333 }
1334
1335 AVCodec wmav1_decoder =
1336 {
1337 "wmav1",
1338 CODEC_TYPE_AUDIO,
1339 CODEC_ID_WMAV1,
1340 sizeof(WMADecodeContext),
1341 wma_decode_init,
1342 NULL,
1343 wma_decode_end,
1344 wma_decode_superframe,
1345 };
1346
1347 AVCodec wmav2_decoder =
1348 {
1349 "wmav2",
1350 CODEC_TYPE_AUDIO,
1351 CODEC_ID_WMAV2,
1352 sizeof(WMADecodeContext),
1353 wma_decode_init,
1354 NULL,
1355 wma_decode_end,
1356 wma_decode_superframe,
1357 };