remove step parameter from vector_fmul_add().
[paraslash.git] / wmadec_filter.c
1 /*
2 * WMA compatible decoder
3 *
4 * Extracted 2009 from the mplayer source code 2009-02-10.
5 *
6 * Copyright (c) 2002 The FFmpeg Project
7 *
8 * Licensed under the GNU Lesser General Public License.
9 * For licencing details see COPYING.LIB.
10 */
11
12 /** * \file wmadec_filter.c paraslash's WMA decoder. */
13
14 /*
15 * This decoder handles Microsoft Windows Media Audio data version 2.
16 */
17
18 #define _XOPEN_SOURCE 600
19
20 #include <sys/time.h>
21 #include <inttypes.h>
22 #include <stdio.h>
23 #include <stdlib.h>
24 #include <math.h>
25 #include <string.h>
26 #include <regex.h>
27
28 #include "para.h"
29 #include "error.h"
30 #include "list.h"
31 #include "ggo.h"
32 #include "string.h"
33 #include "sched.h"
34 #include "filter.h"
35 #include "bitstream.h"
36 #include "imdct.h"
37 #include "wma.h"
38 #include "wmadata.h"
39
40
41 /* size of blocks */
42 #define BLOCK_MIN_BITS 7
43 #define BLOCK_MAX_BITS 11
44 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
45
46 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
47
48 /* XXX: find exact max size */
49 #define HIGH_BAND_MAX_SIZE 16
50
51 /* XXX: is it a suitable value ? */
52 #define MAX_CODED_SUPERFRAME_SIZE 16384
53
54 #define MAX_CHANNELS 2
55
56 #define NOISE_TAB_SIZE 8192
57
58 #define LSP_POW_BITS 7
59
60 struct private_wmadec_data {
61 struct asf_header_info ahi;
62 struct getbit_context gb;
63 int use_bit_reservoir;
64 int use_variable_block_len;
65 int use_exp_vlc; ///< exponent coding: 0 = lsp, 1 = vlc + delta
66 int use_noise_coding; ///< true if perceptual noise is added
67 int byte_offset_bits;
68 struct vlc exp_vlc;
69 int exponent_sizes[BLOCK_NB_SIZES];
70 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
71 int high_band_start[BLOCK_NB_SIZES]; ///< index of first coef in high band
72 int coefs_start; ///< first coded coef
73 int coefs_end[BLOCK_NB_SIZES]; ///< max number of coded coefficients
74 int exponent_high_sizes[BLOCK_NB_SIZES];
75 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
76 struct vlc hgain_vlc;
77
78 /* coded values in high bands */
79 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
80 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
81
82 /* there are two possible tables for spectral coefficients */
83 struct vlc coef_vlc[2];
84 uint16_t *run_table[2];
85 uint16_t *level_table[2];
86 uint16_t *int_table[2];
87 const struct coef_vlc_table *coef_vlcs[2];
88 /* frame info */
89 int frame_len; ///< frame length in samples
90 int frame_len_bits; ///< frame_len = 1 << frame_len_bits
91 int nb_block_sizes; ///< number of block sizes
92 /* block info */
93 int reset_block_lengths;
94 int block_len_bits; ///< log2 of current block length
95 int next_block_len_bits; ///< log2 of next block length
96 int prev_block_len_bits; ///< log2 of prev block length
97 int block_len; ///< block length in samples
98 int block_pos; ///< current position in frame
99 uint8_t ms_stereo; ///< true if mid/side stereo mode
100 uint8_t channel_coded[MAX_CHANNELS]; ///< true if channel is coded
101 int exponents_bsize[MAX_CHANNELS]; ///< log2 ratio frame/exp. length
102 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE];
103 float max_exponent[MAX_CHANNELS];
104 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
105 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE];
106 float output[BLOCK_MAX_SIZE * 2];
107 struct mdct_context *mdct_ctx[BLOCK_NB_SIZES];
108 float *windows[BLOCK_NB_SIZES];
109 /* output buffer for one frame and the last for IMDCT windowing */
110 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
111 /* last frame info */
112 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
113 int last_bitoffset;
114 int last_superframe_len;
115 float noise_table[NOISE_TAB_SIZE];
116 int noise_index;
117 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
118 /* lsp_to_curve tables */
119 float lsp_cos_table[BLOCK_MAX_SIZE];
120 float lsp_pow_e_table[256];
121 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
122 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
123 };
124
125 #define EXPVLCBITS 8
126 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS)
127
128 #define HGAINVLCBITS 9
129 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS)
130
131 #define VLCBITS 9
132 #define VLCMAX ((22+VLCBITS-1)/VLCBITS)
133
134 DECLARE_ALIGNED(16, float, ff_sine_128[128]);
135 DECLARE_ALIGNED(16, float, ff_sine_256[256]);
136 DECLARE_ALIGNED(16, float, ff_sine_512[512]);
137 DECLARE_ALIGNED(16, float, ff_sine_1024[1024]);
138 DECLARE_ALIGNED(16, float, ff_sine_2048[2048]);
139 DECLARE_ALIGNED(16, float, ff_sine_4096[4096]);
140
141 static float *ff_sine_windows[6] = {
142 ff_sine_128, ff_sine_256, ff_sine_512, ff_sine_1024,
143 ff_sine_2048, ff_sine_4096
144 };
145
146 /* Generate a sine window. */
147 static void sine_window_init(float *window, int n)
148 {
149 int i;
150
151 for (i = 0; i < n; i++)
152 window[i] = sinf((i + 0.5) * (M_PI / (2.0 * n)));
153 }
154
155 static int wmadec_cleanup(struct private_wmadec_data *s)
156 {
157 int i;
158
159 for (i = 0; i < s->nb_block_sizes; i++)
160 imdct_end(s->mdct_ctx[i]);
161
162 if (s->use_exp_vlc)
163 free_vlc(&s->exp_vlc);
164 if (s->use_noise_coding)
165 free_vlc(&s->hgain_vlc);
166 for (i = 0; i < 2; i++) {
167 free_vlc(&s->coef_vlc[i]);
168 free(s->run_table[i]);
169 free(s->level_table[i]);
170 free(s->int_table[i]);
171 }
172 return 0;
173 }
174
175 /* XXX: use same run/length optimization as mpeg decoders */
176 //FIXME maybe split decode / encode or pass flag
177 static void init_coef_vlc(struct vlc *vlc, uint16_t **prun_table,
178 uint16_t **plevel_table, uint16_t **pint_table,
179 const struct coef_vlc_table *vlc_table)
180 {
181 int n = vlc_table->n;
182 const uint8_t *table_bits = vlc_table->huffbits;
183 const uint32_t *table_codes = vlc_table->huffcodes;
184 const uint16_t *levels_table = vlc_table->levels;
185 uint16_t *run_table, *level_table, *int_table;
186 int i, l, j, k, level;
187
188 init_vlc(vlc, VLCBITS, n, table_bits, 1, 1, table_codes, 4, 4);
189
190 run_table = para_malloc(n * sizeof(uint16_t));
191 level_table = para_malloc(n * sizeof(uint16_t));
192 int_table = para_malloc(n * sizeof(uint16_t));
193 i = 2;
194 level = 1;
195 k = 0;
196 while (i < n) {
197 int_table[k] = i;
198 l = levels_table[k++];
199 for (j = 0; j < l; j++) {
200 run_table[i] = j;
201 level_table[i] = level;
202 i++;
203 }
204 level++;
205 }
206 *prun_table = run_table;
207 *plevel_table = level_table;
208 *pint_table = int_table;
209 }
210
211 /* compute the scale factor band sizes for each MDCT block size */
212 static void compute_scale_factor_band_sizes(struct private_wmadec_data *s,
213 float high_freq)
214 {
215 struct asf_header_info *ahi = &s->ahi;
216 int a, b, pos, lpos, k, block_len, i, j, n;
217 const uint8_t *table;
218
219 s->coefs_start = 0;
220 for (k = 0; k < s->nb_block_sizes; k++) {
221 block_len = s->frame_len >> k;
222
223 table = NULL;
224 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
225 if (a < 3) {
226 if (ahi->sample_rate >= 44100)
227 table = exponent_band_44100[a];
228 else if (ahi->sample_rate >= 32000)
229 table = exponent_band_32000[a];
230 else if (ahi->sample_rate >= 22050)
231 table = exponent_band_22050[a];
232 }
233 if (table) {
234 n = *table++;
235 for (i = 0; i < n; i++)
236 s->exponent_bands[k][i] = table[i];
237 s->exponent_sizes[k] = n;
238 } else {
239 j = 0;
240 lpos = 0;
241 for (i = 0; i < 25; i++) {
242 a = wma_critical_freqs[i];
243 b = ahi->sample_rate;
244 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
245 pos <<= 2;
246 if (pos > block_len)
247 pos = block_len;
248 if (pos > lpos)
249 s->exponent_bands[k][j++] = pos - lpos;
250 if (pos >= block_len)
251 break;
252 lpos = pos;
253 }
254 s->exponent_sizes[k] = j;
255 }
256
257 /* max number of coefs */
258 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
259 /* high freq computation */
260 s->high_band_start[k] = (int) ((block_len * 2 * high_freq)
261 / ahi->sample_rate + 0.5);
262 n = s->exponent_sizes[k];
263 j = 0;
264 pos = 0;
265 for (i = 0; i < n; i++) {
266 int start, end;
267 start = pos;
268 pos += s->exponent_bands[k][i];
269 end = pos;
270 if (start < s->high_band_start[k])
271 start = s->high_band_start[k];
272 if (end > s->coefs_end[k])
273 end = s->coefs_end[k];
274 if (end > start)
275 s->exponent_high_bands[k][j++] = end - start;
276 }
277 s->exponent_high_sizes[k] = j;
278 }
279 }
280
281 static int wma_init(struct private_wmadec_data *s, int flags2, struct asf_header_info *ahi)
282 {
283 int i;
284 float bps1, high_freq;
285 volatile float bps;
286 int sample_rate1;
287 int coef_vlc_table;
288
289 if (ahi->sample_rate <= 0 || ahi->sample_rate > 50000
290 || ahi->channels <= 0 || ahi->channels > 8
291 || ahi->bit_rate <= 0)
292 return -E_WMA_BAD_PARAMS;
293
294 /* compute MDCT block size */
295 if (ahi->sample_rate <= 16000) {
296 s->frame_len_bits = 9;
297 } else if (ahi->sample_rate <= 22050) {
298 s->frame_len_bits = 10;
299 } else {
300 s->frame_len_bits = 11;
301 }
302 s->frame_len = 1 << s->frame_len_bits;
303 if (s->use_variable_block_len) {
304 int nb_max, nb;
305 nb = ((flags2 >> 3) & 3) + 1;
306 if ((ahi->bit_rate / ahi->channels) >= 32000)
307 nb += 2;
308 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
309 if (nb > nb_max)
310 nb = nb_max;
311 s->nb_block_sizes = nb + 1;
312 } else {
313 s->nb_block_sizes = 1;
314 }
315
316 /* init rate dependent parameters */
317 s->use_noise_coding = 1;
318 high_freq = ahi->sample_rate * 0.5;
319
320 /* wma2 rates are normalized */
321 sample_rate1 = ahi->sample_rate;
322 if (sample_rate1 >= 44100)
323 sample_rate1 = 44100;
324 else if (sample_rate1 >= 22050)
325 sample_rate1 = 22050;
326 else if (sample_rate1 >= 16000)
327 sample_rate1 = 16000;
328 else if (sample_rate1 >= 11025)
329 sample_rate1 = 11025;
330 else if (sample_rate1 >= 8000)
331 sample_rate1 = 8000;
332
333 bps = (float) ahi->bit_rate / (float) (ahi->channels * ahi->sample_rate);
334 s->byte_offset_bits = wma_log2((int) (bps * s->frame_len / 8.0 + 0.5)) + 2;
335 /*
336 * Compute high frequency value and choose if noise coding should be
337 * activated.
338 */
339 bps1 = bps;
340 if (ahi->channels == 2)
341 bps1 = bps * 1.6;
342 if (sample_rate1 == 44100) {
343 if (bps1 >= 0.61)
344 s->use_noise_coding = 0;
345 else
346 high_freq = high_freq * 0.4;
347 } else if (sample_rate1 == 22050) {
348 if (bps1 >= 1.16)
349 s->use_noise_coding = 0;
350 else if (bps1 >= 0.72)
351 high_freq = high_freq * 0.7;
352 else
353 high_freq = high_freq * 0.6;
354 } else if (sample_rate1 == 16000) {
355 if (bps > 0.5)
356 high_freq = high_freq * 0.5;
357 else
358 high_freq = high_freq * 0.3;
359 } else if (sample_rate1 == 11025) {
360 high_freq = high_freq * 0.7;
361 } else if (sample_rate1 == 8000) {
362 if (bps <= 0.625) {
363 high_freq = high_freq * 0.5;
364 } else if (bps > 0.75) {
365 s->use_noise_coding = 0;
366 } else {
367 high_freq = high_freq * 0.65;
368 }
369 } else {
370 if (bps >= 0.8) {
371 high_freq = high_freq * 0.75;
372 } else if (bps >= 0.6) {
373 high_freq = high_freq * 0.6;
374 } else {
375 high_freq = high_freq * 0.5;
376 }
377 }
378 PARA_INFO_LOG("channels=%d sample_rate=%d "
379 "bitrate=%d block_align=%d\n",
380 ahi->channels, ahi->sample_rate,
381 ahi->bit_rate, ahi->block_align);
382 PARA_INFO_LOG("frame_len=%d, bps=%f bps1=%f "
383 "high_freq=%f bitoffset=%d\n",
384 s->frame_len, bps, bps1,
385 high_freq, s->byte_offset_bits);
386 PARA_INFO_LOG("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
387 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
388
389 compute_scale_factor_band_sizes(s, high_freq);
390 /* init MDCT windows : simple sinus window */
391 for (i = 0; i < s->nb_block_sizes; i++) {
392 int n;
393 n = 1 << (s->frame_len_bits - i);
394 sine_window_init(ff_sine_windows[s->frame_len_bits - i - 7], n);
395 s->windows[i] = ff_sine_windows[s->frame_len_bits - i - 7];
396 }
397
398 s->reset_block_lengths = 1;
399
400 if (s->use_noise_coding) {
401 /* init the noise generator */
402 if (s->use_exp_vlc)
403 s->noise_mult = 0.02;
404 else
405 s->noise_mult = 0.04;
406
407 {
408 unsigned int seed;
409 float norm;
410 seed = 1;
411 norm = (1.0 / (float) (1LL << 31)) * sqrt(3) * s->noise_mult;
412 for (i = 0; i < NOISE_TAB_SIZE; i++) {
413 seed = seed * 314159 + 1;
414 s->noise_table[i] = (float) ((int) seed) * norm;
415 }
416 }
417 }
418
419 /* choose the VLC tables for the coefficients */
420 coef_vlc_table = 2;
421 if (ahi->sample_rate >= 32000) {
422 if (bps1 < 0.72)
423 coef_vlc_table = 0;
424 else if (bps1 < 1.16)
425 coef_vlc_table = 1;
426 }
427 s->coef_vlcs[0] = &coef_vlcs[coef_vlc_table * 2];
428 s->coef_vlcs[1] = &coef_vlcs[coef_vlc_table * 2 + 1];
429 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
430 &s->int_table[0], s->coef_vlcs[0]);
431 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
432 &s->int_table[1], s->coef_vlcs[1]);
433 return 0;
434 }
435
436 static void wma_lsp_to_curve_init(struct private_wmadec_data *s, int frame_len)
437 {
438 float wdel, a, b;
439 int i, e, m;
440
441 wdel = M_PI / frame_len;
442 for (i = 0; i < frame_len; i++)
443 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
444
445 /* tables for x^-0.25 computation */
446 for (i = 0; i < 256; i++) {
447 e = i - 126;
448 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
449 }
450
451 /* These two tables are needed to avoid two operations in pow_m1_4. */
452 b = 1.0;
453 for (i = (1 << LSP_POW_BITS) - 1; i >= 0; i--) {
454 m = (1 << LSP_POW_BITS) + i;
455 a = (float) m *(0.5 / (1 << LSP_POW_BITS));
456 a = pow(a, -0.25);
457 s->lsp_pow_m_table1[i] = 2 * a - b;
458 s->lsp_pow_m_table2[i] = b - a;
459 b = a;
460 }
461 }
462
463 static int wma_decode_init(char *initial_buf, int len, struct private_wmadec_data **result)
464 {
465 struct private_wmadec_data *s;
466 int ret, i;
467
468 PARA_NOTICE_LOG("initial buf: %d bytes\n", len);
469 s = para_calloc(sizeof(*s));
470 ret = read_asf_header(initial_buf, len, &s->ahi);
471 if (ret <= 0) {
472 free(s);
473 return ret;
474 }
475
476 s->use_exp_vlc = s->ahi.flags2 & 0x0001;
477 s->use_bit_reservoir = s->ahi.flags2 & 0x0002;
478 s->use_variable_block_len = s->ahi.flags2 & 0x0004;
479
480 ret = wma_init(s, s->ahi.flags2, &s->ahi);
481 if (ret < 0)
482 return ret;
483 /* init MDCT */
484 for (i = 0; i < s->nb_block_sizes; i++) {
485 ret = imdct_init(s->frame_len_bits - i + 1, &s->mdct_ctx[i]);
486 if (ret < 0)
487 return ret;
488 }
489 if (s->use_noise_coding) {
490 PARA_INFO_LOG("using noise coding\n");
491 init_vlc(&s->hgain_vlc, HGAINVLCBITS,
492 sizeof(ff_wma_hgain_huffbits), ff_wma_hgain_huffbits,
493 1, 1, ff_wma_hgain_huffcodes, 2, 2);
494 }
495
496 if (s->use_exp_vlc) {
497 PARA_INFO_LOG("using exp_vlc\n");
498 init_vlc(&s->exp_vlc, EXPVLCBITS,
499 sizeof(ff_wma_scale_huffbits), ff_wma_scale_huffbits,
500 1, 1, ff_wma_scale_huffcodes, 4, 4);
501 } else {
502 PARA_INFO_LOG("using curve\n");
503 wma_lsp_to_curve_init(s, s->frame_len);
504 }
505 *result = s;
506 return s->ahi.header_len;
507 }
508
509 /**
510 * compute x^-0.25 with an exponent and mantissa table. We use linear
511 * interpolation to reduce the mantissa table size at a small speed
512 * expense (linear interpolation approximately doubles the number of
513 * bits of precision).
514 */
515 static inline float pow_m1_4(struct private_wmadec_data *s, float x)
516 {
517 union {
518 float f;
519 unsigned int v;
520 } u, t;
521 unsigned int e, m;
522 float a, b;
523
524 u.f = x;
525 e = u.v >> 23;
526 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
527 /* build interpolation scale: 1 <= t < 2. */
528 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
529 a = s->lsp_pow_m_table1[m];
530 b = s->lsp_pow_m_table2[m];
531 return s->lsp_pow_e_table[e] * (a + b * t.f);
532 }
533
534 static void wma_lsp_to_curve(struct private_wmadec_data *s,
535 float *out, float *val_max_ptr, int n, float *lsp)
536 {
537 int i, j;
538 float p, q, w, v, val_max;
539
540 val_max = 0;
541 for (i = 0; i < n; i++) {
542 p = 0.5f;
543 q = 0.5f;
544 w = s->lsp_cos_table[i];
545 for (j = 1; j < NB_LSP_COEFS; j += 2) {
546 q *= w - lsp[j - 1];
547 p *= w - lsp[j];
548 }
549 p *= p * (2.0f - w);
550 q *= q * (2.0f + w);
551 v = p + q;
552 v = pow_m1_4(s, v);
553 if (v > val_max)
554 val_max = v;
555 out[i] = v;
556 }
557 *val_max_ptr = val_max;
558 }
559
560 /* Decode exponents coded with LSP coefficients (same idea as Vorbis). */
561 static void decode_exp_lsp(struct private_wmadec_data *s, int ch)
562 {
563 float lsp_coefs[NB_LSP_COEFS];
564 int val, i;
565
566 for (i = 0; i < NB_LSP_COEFS; i++) {
567 if (i == 0 || i >= 8)
568 val = get_bits(&s->gb, 3);
569 else
570 val = get_bits(&s->gb, 4);
571 lsp_coefs[i] = ff_wma_lsp_codebook[i][val];
572 }
573
574 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
575 s->block_len, lsp_coefs);
576 }
577
578 /*
579 * Parse a vlc code, faster then get_vlc().
580 *
581 * \param bits The number of bits which will be read at once, must be
582 * identical to nb_bits in init_vlc()
583 *
584 * \param max_depth The number of times bits bits must be read to completely
585 * read the longest vlc code = (max_vlc_length + bits - 1) / bits.
586 */
587 static int get_vlc2(struct getbit_context *s, VLC_TYPE(*table)[2],
588 int bits, int max_depth)
589 {
590 int code;
591
592 OPEN_READER(re, s)
593 UPDATE_CACHE(re, s)
594 GET_VLC(code, re, s, table, bits, max_depth)
595 CLOSE_READER(re, s)
596 return code;
597 }
598
599 /* Decode exponents coded with VLC codes. */
600 static int decode_exp_vlc(struct private_wmadec_data *s, int ch)
601 {
602 int last_exp, n, code;
603 const uint16_t *ptr, *band_ptr;
604 float v, *q, max_scale, *q_end;
605
606 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
607 ptr = band_ptr;
608 q = s->exponents[ch];
609 q_end = q + s->block_len;
610 max_scale = 0;
611 last_exp = 36;
612
613 while (q < q_end) {
614 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
615 if (code < 0)
616 return -1;
617 /* NOTE: this offset is the same as MPEG4 AAC ! */
618 last_exp += code - 60;
619 /* XXX: use a table */
620 v = pow(10, last_exp * (1.0 / 16.0));
621 if (v > max_scale)
622 max_scale = v;
623 n = *ptr++;
624 do {
625 *q++ = v;
626 } while (--n);
627 }
628 s->max_exponent[ch] = max_scale;
629 return 0;
630 }
631
632 static void vector_fmul_add(float *dst, const float *src0, const float *src1,
633 const float *src2, int src3, int len)
634 {
635 int i;
636 for (i = 0; i < len; i++)
637 dst[i] = src0[i] * src1[i] + src2[i] + src3;
638 }
639
640 static void vector_fmul_reverse_c(float *dst, const float *src0,
641 const float *src1, int len)
642 {
643 int i;
644 src1 += len - 1;
645 for (i = 0; i < len; i++)
646 dst[i] = src0[i] * src1[-i];
647 }
648
649 /**
650 * Apply MDCT window and add into output.
651 *
652 * We ensure that when the windows overlap their squared sum
653 * is always 1 (MDCT reconstruction rule).
654 */
655 static void wma_window(struct private_wmadec_data *s, float *out)
656 {
657 float *in = s->output;
658 int block_len, bsize, n;
659
660 /* left part */
661 if (s->block_len_bits <= s->prev_block_len_bits) {
662 block_len = s->block_len;
663 bsize = s->frame_len_bits - s->block_len_bits;
664
665 vector_fmul_add(out, in, s->windows[bsize],
666 out, 0, block_len);
667
668 } else {
669 block_len = 1 << s->prev_block_len_bits;
670 n = (s->block_len - block_len) / 2;
671 bsize = s->frame_len_bits - s->prev_block_len_bits;
672
673 vector_fmul_add(out + n, in + n, s->windows[bsize],
674 out + n, 0, block_len);
675
676 memcpy(out + n + block_len, in + n + block_len,
677 n * sizeof(float));
678 }
679
680 out += s->block_len;
681 in += s->block_len;
682
683 /* right part */
684 if (s->block_len_bits <= s->next_block_len_bits) {
685 block_len = s->block_len;
686 bsize = s->frame_len_bits - s->block_len_bits;
687
688 vector_fmul_reverse_c(out, in, s->windows[bsize], block_len);
689
690 } else {
691 block_len = 1 << s->next_block_len_bits;
692 n = (s->block_len - block_len) / 2;
693 bsize = s->frame_len_bits - s->next_block_len_bits;
694
695 memcpy(out, in, n * sizeof(float));
696
697 vector_fmul_reverse_c(out + n, in + n, s->windows[bsize],
698 block_len);
699
700 memset(out + n + block_len, 0, n * sizeof(float));
701 }
702 }
703
704 static int wma_total_gain_to_bits(int total_gain)
705 {
706 if (total_gain < 15)
707 return 13;
708 else if (total_gain < 32)
709 return 12;
710 else if (total_gain < 40)
711 return 11;
712 else if (total_gain < 45)
713 return 10;
714 else
715 return 9;
716 }
717
718 /**
719 * @return 0 if OK. 1 if last block of frame. return -1 if
720 * unrecorrable error.
721 */
722 static int wma_decode_block(struct private_wmadec_data *s)
723 {
724 int n, v, ch, code, bsize;
725 int coef_nb_bits, total_gain;
726 int nb_coefs[MAX_CHANNELS];
727 float mdct_norm;
728
729 /* compute current block length */
730 if (s->use_variable_block_len) {
731 n = wma_log2(s->nb_block_sizes - 1) + 1;
732
733 if (s->reset_block_lengths) {
734 s->reset_block_lengths = 0;
735 v = get_bits(&s->gb, n);
736 if (v >= s->nb_block_sizes)
737 return -1;
738 s->prev_block_len_bits = s->frame_len_bits - v;
739 v = get_bits(&s->gb, n);
740 if (v >= s->nb_block_sizes)
741 return -1;
742 s->block_len_bits = s->frame_len_bits - v;
743 } else {
744 /* update block lengths */
745 s->prev_block_len_bits = s->block_len_bits;
746 s->block_len_bits = s->next_block_len_bits;
747 }
748 v = get_bits(&s->gb, n);
749 if (v >= s->nb_block_sizes)
750 return -1;
751 s->next_block_len_bits = s->frame_len_bits - v;
752 } else {
753 /* fixed block len */
754 s->next_block_len_bits = s->frame_len_bits;
755 s->prev_block_len_bits = s->frame_len_bits;
756 s->block_len_bits = s->frame_len_bits;
757 }
758
759 /* now check if the block length is coherent with the frame length */
760 s->block_len = 1 << s->block_len_bits;
761 if ((s->block_pos + s->block_len) > s->frame_len)
762 return -E_INCOHERENT_BLOCK_LEN;
763
764 if (s->ahi.channels == 2) {
765 s->ms_stereo = get_bits1(&s->gb);
766 }
767 v = 0;
768 for (ch = 0; ch < s->ahi.channels; ch++) {
769 int a = get_bits1(&s->gb);
770 s->channel_coded[ch] = a;
771 v |= a;
772 }
773
774 bsize = s->frame_len_bits - s->block_len_bits;
775
776 /* if no channel coded, no need to go further */
777 /* XXX: fix potential framing problems */
778 if (!v)
779 goto next;
780
781 /* read total gain and extract corresponding number of bits for
782 coef escape coding */
783 total_gain = 1;
784 for (;;) {
785 int a = get_bits(&s->gb, 7);
786 total_gain += a;
787 if (a != 127)
788 break;
789 }
790
791 coef_nb_bits = wma_total_gain_to_bits(total_gain);
792
793 /* compute number of coefficients */
794 n = s->coefs_end[bsize] - s->coefs_start;
795 for (ch = 0; ch < s->ahi.channels; ch++)
796 nb_coefs[ch] = n;
797
798 /* complex coding */
799 if (s->use_noise_coding) {
800 for (ch = 0; ch < s->ahi.channels; ch++) {
801 if (s->channel_coded[ch]) {
802 int i, m, a;
803 m = s->exponent_high_sizes[bsize];
804 for (i = 0; i < m; i++) {
805 a = get_bits1(&s->gb);
806 s->high_band_coded[ch][i] = a;
807 /* if noise coding, the coefficients are not transmitted */
808 if (a)
809 nb_coefs[ch] -=
810 s->
811 exponent_high_bands[bsize]
812 [i];
813 }
814 }
815 }
816 for (ch = 0; ch < s->ahi.channels; ch++) {
817 if (s->channel_coded[ch]) {
818 int i, val;
819
820 n = s->exponent_high_sizes[bsize];
821 val = (int) 0x80000000;
822 for (i = 0; i < n; i++) {
823 if (s->high_band_coded[ch][i]) {
824 if (val == (int) 0x80000000) {
825 val =
826 get_bits(&s->gb,
827 7) - 19;
828 } else {
829 code =
830 get_vlc2(&s->gb,
831 s->
832 hgain_vlc.
833 table,
834 HGAINVLCBITS,
835 HGAINMAX);
836 if (code < 0)
837 return -1;
838 val += code - 18;
839 }
840 s->high_band_values[ch][i] =
841 val;
842 }
843 }
844 }
845 }
846 }
847
848 /* exponents can be reused in short blocks. */
849 if ((s->block_len_bits == s->frame_len_bits) || get_bits1(&s->gb)) {
850 for (ch = 0; ch < s->ahi.channels; ch++) {
851 if (s->channel_coded[ch]) {
852 if (s->use_exp_vlc) {
853 if (decode_exp_vlc(s, ch) < 0)
854 return -1;
855 } else {
856 decode_exp_lsp(s, ch);
857 }
858 s->exponents_bsize[ch] = bsize;
859 }
860 }
861 }
862
863 /* parse spectral coefficients : just RLE encoding */
864 for (ch = 0; ch < s->ahi.channels; ch++) {
865 if (s->channel_coded[ch]) {
866 struct vlc *coef_vlc;
867 int level, run, sign, tindex;
868 int16_t *ptr, *eptr;
869 const uint16_t *level_table, *run_table;
870
871 /* special VLC tables are used for ms stereo because
872 there is potentially less energy there */
873 tindex = (ch == 1 && s->ms_stereo);
874 coef_vlc = &s->coef_vlc[tindex];
875 run_table = s->run_table[tindex];
876 level_table = s->level_table[tindex];
877 /* XXX: optimize */
878 ptr = &s->coefs1[ch][0];
879 eptr = ptr + nb_coefs[ch];
880 memset(ptr, 0, s->block_len * sizeof(int16_t));
881 for (;;) {
882 code =
883 get_vlc2(&s->gb, coef_vlc->table, VLCBITS,
884 VLCMAX);
885 if (code < 0)
886 return -1;
887 if (code == 1) {
888 /* EOB */
889 break;
890 } else if (code == 0) {
891 /* escape */
892 level = get_bits(&s->gb, coef_nb_bits);
893 /* NOTE: this is rather suboptimal. reading
894 block_len_bits would be better */
895 run =
896 get_bits(&s->gb, s->frame_len_bits);
897 } else {
898 /* normal code */
899 run = run_table[code];
900 level = level_table[code];
901 }
902 sign = get_bits1(&s->gb);
903 if (!sign)
904 level = -level;
905 ptr += run;
906 if (ptr >= eptr) {
907 PARA_ERROR_LOG("overflow in spectral RLE, ignoring\n");
908 break;
909 }
910 *ptr++ = level;
911 /* NOTE: EOB can be omitted */
912 if (ptr >= eptr)
913 break;
914 }
915 }
916 }
917
918 /* normalize */
919 {
920 int n4 = s->block_len / 2;
921 mdct_norm = 1.0 / (float) n4;
922 }
923
924 /* finally compute the MDCT coefficients */
925 for (ch = 0; ch < s->ahi.channels; ch++) {
926 if (s->channel_coded[ch]) {
927 int16_t *coefs1;
928 float *coefs, *exponents, mult, mult1, noise;
929 int i, j, n1, last_high_band, esize;
930 float exp_power[HIGH_BAND_MAX_SIZE];
931
932 coefs1 = s->coefs1[ch];
933 exponents = s->exponents[ch];
934 esize = s->exponents_bsize[ch];
935 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
936 mult *= mdct_norm;
937 coefs = s->coefs[ch];
938 if (s->use_noise_coding) {
939 mult1 = mult;
940 /* very low freqs : noise */
941 for (i = 0; i < s->coefs_start; i++) {
942 *coefs++ =
943 s->noise_table[s->noise_index] *
944 exponents[i << bsize >> esize] *
945 mult1;
946 s->noise_index =
947 (s->noise_index +
948 1) & (NOISE_TAB_SIZE - 1);
949 }
950
951 n1 = s->exponent_high_sizes[bsize];
952
953 /* compute power of high bands */
954 exponents = s->exponents[ch] +
955 (s->high_band_start[bsize] << bsize);
956 last_high_band = 0; /* avoid warning */
957 for (j = 0; j < n1; j++) {
958 n = s->exponent_high_bands[s->
959 frame_len_bits
960 -
961 s->
962 block_len_bits]
963 [j];
964 if (s->high_band_coded[ch][j]) {
965 float e2, val;
966 e2 = 0;
967 for (i = 0; i < n; i++) {
968 val = exponents[i << bsize
969 >> esize];
970 e2 += val * val;
971 }
972 exp_power[j] = e2 / n;
973 last_high_band = j;
974 }
975 exponents += n << bsize;
976 }
977
978 /* main freqs and high freqs */
979 exponents =
980 s->exponents[ch] +
981 (s->coefs_start << bsize);
982 for (j = -1; j < n1; j++) {
983 if (j < 0) {
984 n = s->high_band_start[bsize] -
985 s->coefs_start;
986 } else {
987 n = s->exponent_high_bands[s->
988 frame_len_bits
989 -
990 s->
991 block_len_bits]
992 [j];
993 }
994 if (j >= 0 && s->high_band_coded[ch][j]) {
995 /* use noise with specified power */
996 mult1 =
997 sqrt(exp_power[j] /
998 exp_power
999 [last_high_band]);
1000 /* XXX: use a table */
1001 mult1 =
1002 mult1 * pow(10,
1003 s->
1004 high_band_values
1005 [ch][j] * 0.05);
1006 mult1 =
1007 mult1 /
1008 (s->max_exponent[ch] *
1009 s->noise_mult);
1010 mult1 *= mdct_norm;
1011 for (i = 0; i < n; i++) {
1012 noise =
1013 s->noise_table[s->
1014 noise_index];
1015 s->noise_index =
1016 (s->noise_index +
1017 1) &
1018 (NOISE_TAB_SIZE -
1019 1);
1020 *coefs++ =
1021 noise *
1022 exponents[i << bsize
1023 >> esize]
1024 * mult1;
1025 }
1026 exponents += n << bsize;
1027 } else {
1028 /* coded values + small noise */
1029 for (i = 0; i < n; i++) {
1030 noise =
1031 s->noise_table[s->
1032 noise_index];
1033 s->noise_index =
1034 (s->noise_index +
1035 1) &
1036 (NOISE_TAB_SIZE -
1037 1);
1038 *coefs++ =
1039 ((*coefs1++) +
1040 noise) *
1041 exponents[i << bsize
1042 >> esize]
1043 * mult;
1044 }
1045 exponents += n << bsize;
1046 }
1047 }
1048
1049 /* very high freqs : noise */
1050 n = s->block_len - s->coefs_end[bsize];
1051 mult1 =
1052 mult * exponents[((-1 << bsize)) >> esize];
1053 for (i = 0; i < n; i++) {
1054 *coefs++ =
1055 s->noise_table[s->noise_index] *
1056 mult1;
1057 s->noise_index =
1058 (s->noise_index +
1059 1) & (NOISE_TAB_SIZE - 1);
1060 }
1061 } else {
1062 /* XXX: optimize more */
1063 for (i = 0; i < s->coefs_start; i++)
1064 *coefs++ = 0.0;
1065 n = nb_coefs[ch];
1066 for (i = 0; i < n; i++) {
1067 *coefs++ =
1068 coefs1[i] *
1069 exponents[i << bsize >> esize] *
1070 mult;
1071 }
1072 n = s->block_len - s->coefs_end[bsize];
1073 for (i = 0; i < n; i++)
1074 *coefs++ = 0.0;
1075 }
1076 }
1077 }
1078
1079 if (s->ms_stereo && s->channel_coded[1]) {
1080 float a, b;
1081 int i;
1082
1083 /*
1084 * Nominal case for ms stereo: we do it before mdct.
1085 *
1086 * No need to optimize this case because it should almost never
1087 * happen.
1088 */
1089 if (!s->channel_coded[0]) {
1090 PARA_NOTICE_LOG("rare ms-stereo\n");
1091 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1092 s->channel_coded[0] = 1;
1093 }
1094 for (i = 0; i < s->block_len; i++) {
1095 a = s->coefs[0][i];
1096 b = s->coefs[1][i];
1097 s->coefs[0][i] = a + b;
1098 s->coefs[1][i] = a - b;
1099 }
1100 }
1101
1102 next:
1103 for (ch = 0; ch < s->ahi.channels; ch++) {
1104 int n4, index;
1105
1106 n = s->block_len;
1107 n4 = s->block_len / 2;
1108 if (s->channel_coded[ch])
1109 imdct(s->mdct_ctx[bsize], s->output, s->coefs[ch]);
1110 else if (!(s->ms_stereo && ch == 1))
1111 memset(s->output, 0, sizeof(s->output));
1112
1113 /* multiply by the window and add in the frame */
1114 index = (s->frame_len / 2) + s->block_pos - n4;
1115 wma_window(s, &s->frame_out[ch][index]);
1116 }
1117
1118 /* update block number */
1119 s->block_pos += s->block_len;
1120 if (s->block_pos >= s->frame_len)
1121 return 1;
1122 else
1123 return 0;
1124 }
1125
1126 /*
1127 * Clip a signed integer value into the -32768,32767 range.
1128 *
1129 * \param a The value to clip.
1130 *
1131 * \return The clipped value.
1132 */
1133 static inline int16_t av_clip_int16(int a)
1134 {
1135 if ((a + 32768) & ~65535)
1136 return (a >> 31) ^ 32767;
1137 else
1138 return a;
1139 }
1140
1141 /* Decode a frame of frame_len samples. */
1142 static int wma_decode_frame(struct private_wmadec_data *s, int16_t * samples)
1143 {
1144 int ret, i, n, ch, incr;
1145 int16_t *ptr;
1146 float *iptr;
1147
1148 /* read each block */
1149 s->block_pos = 0;
1150 for (;;) {
1151 ret = wma_decode_block(s);
1152 if (ret < 0)
1153 return -1;
1154 if (ret)
1155 break;
1156 }
1157
1158 /* convert frame to integer */
1159 n = s->frame_len;
1160 incr = s->ahi.channels;
1161 for (ch = 0; ch < s->ahi.channels; ch++) {
1162 ptr = samples + ch;
1163 iptr = s->frame_out[ch];
1164
1165 for (i = 0; i < n; i++) {
1166 *ptr = av_clip_int16(lrintf(*iptr++));
1167 ptr += incr;
1168 }
1169 /* prepare for next block */
1170 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1171 s->frame_len * sizeof(float));
1172 }
1173 return 0;
1174 }
1175
1176 static int wma_decode_superframe(struct private_wmadec_data *s, void *data,
1177 int *data_size, const uint8_t *buf, int buf_size)
1178 {
1179 int ret, nb_frames, bit_offset, i, pos, len;
1180 uint8_t *q;
1181 int16_t *samples;
1182 static int frame_count;
1183
1184 if (buf_size == 0) {
1185 s->last_superframe_len = 0;
1186 return 0;
1187 }
1188 if (buf_size < s->ahi.block_align)
1189 return 0;
1190 buf_size = s->ahi.block_align;
1191 samples = data;
1192 init_get_bits(&s->gb, buf, buf_size * 8);
1193 if (s->use_bit_reservoir) {
1194 /* read super frame header */
1195 skip_bits(&s->gb, 4); /* super frame index */
1196 nb_frames = get_bits(&s->gb, 4) - 1;
1197 // PARA_DEBUG_LOG("have %d frames\n", nb_frames);
1198 ret = -E_WMA_OUTPUT_SPACE;
1199 if ((nb_frames + 1) * s->ahi.channels * s->frame_len
1200 * sizeof(int16_t) > *data_size)
1201 goto fail;
1202
1203 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1204
1205 if (s->last_superframe_len > 0) {
1206 /* add bit_offset bits to last frame */
1207 ret = -E_WMA_BAD_SUPERFRAME;
1208 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1209 MAX_CODED_SUPERFRAME_SIZE)
1210 goto fail;
1211 q = s->last_superframe + s->last_superframe_len;
1212 len = bit_offset;
1213 while (len > 7) {
1214 *q++ = get_bits(&s->gb, 8);
1215 len -= 8;
1216 }
1217 if (len > 0) {
1218 *q++ = get_bits(&s->gb, len) << (8 - len);
1219 }
1220
1221 /* XXX: bit_offset bits into last frame */
1222 init_get_bits(&s->gb, s->last_superframe,
1223 MAX_CODED_SUPERFRAME_SIZE * 8);
1224 /* skip unused bits */
1225 if (s->last_bitoffset > 0)
1226 skip_bits(&s->gb, s->last_bitoffset);
1227 /*
1228 * This frame is stored in the last superframe and in
1229 * the current one.
1230 */
1231 ret = -E_WMA_DECODE;
1232 if (wma_decode_frame(s, samples) < 0)
1233 goto fail;
1234 frame_count++;
1235 samples += s->ahi.channels * s->frame_len;
1236 }
1237
1238 /* read each frame starting from bit_offset */
1239 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1240 init_get_bits(&s->gb, buf + (pos >> 3),
1241 (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3)) * 8);
1242 len = pos & 7;
1243 if (len > 0)
1244 skip_bits(&s->gb, len);
1245
1246 s->reset_block_lengths = 1;
1247 for (i = 0; i < nb_frames; i++) {
1248 ret = -E_WMA_DECODE;
1249 if (wma_decode_frame(s, samples) < 0)
1250 goto fail;
1251 frame_count++;
1252 samples += s->ahi.channels * s->frame_len;
1253 }
1254
1255 /* we copy the end of the frame in the last frame buffer */
1256 pos = get_bits_count(&s->gb) +
1257 ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1258 s->last_bitoffset = pos & 7;
1259 pos >>= 3;
1260 len = buf_size - pos;
1261 ret = -E_WMA_BAD_SUPERFRAME;
1262 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1263 goto fail;
1264 }
1265 s->last_superframe_len = len;
1266 memcpy(s->last_superframe, buf + pos, len);
1267 } else {
1268 PARA_DEBUG_LOG("not using bit reservoir\n");
1269 ret = -E_WMA_OUTPUT_SPACE;
1270 if (s->ahi.channels * s->frame_len * sizeof(int16_t) > *data_size)
1271 goto fail;
1272 /* single frame decode */
1273 ret = -E_WMA_DECODE;
1274 if (wma_decode_frame(s, samples) < 0)
1275 goto fail;
1276 frame_count++;
1277 samples += s->ahi.channels * s->frame_len;
1278 }
1279 PARA_DEBUG_LOG("frame_count: %d frame_len: %d, block_len: %d, "
1280 "outbytes: %d, eaten: %d\n",
1281 frame_count, s->frame_len, s->block_len,
1282 (int8_t *) samples - (int8_t *) data, s->ahi.block_align);
1283 *data_size = (int8_t *)samples - (int8_t *)data;
1284 return s->ahi.block_align;
1285 fail:
1286 /* reset the bit reservoir on errors */
1287 s->last_superframe_len = 0;
1288 return ret;
1289 }
1290
1291 static ssize_t wmadec_convert(char *inbuffer, size_t len,
1292 struct filter_node *fn)
1293 {
1294 int ret, out_size = fn->bufsize - fn->loaded;
1295 struct private_wmadec_data *pwd = fn->private_data;
1296
1297 if (out_size < 128 * 1024)
1298 return 0;
1299 if (!pwd) {
1300 ret = wma_decode_init(inbuffer, len, &pwd);
1301 if (ret <= 0)
1302 return ret;
1303 fn->private_data = pwd;
1304 fn->fc->channels = pwd->ahi.channels;
1305 fn->fc->samplerate = pwd->ahi.sample_rate;
1306 return pwd->ahi.header_len;
1307 }
1308 /* skip 31 bytes */
1309 if (len <= WMA_FRAME_SKIP + pwd->ahi.block_align)
1310 return 0;
1311 ret = wma_decode_superframe(pwd, fn->buf + fn->loaded,
1312 &out_size, (uint8_t *)inbuffer + WMA_FRAME_SKIP,
1313 len - WMA_FRAME_SKIP);
1314 if (ret < 0)
1315 return ret;
1316 fn->loaded += out_size;
1317 return ret + WMA_FRAME_SKIP;
1318 }
1319
1320 static void wmadec_close(struct filter_node *fn)
1321 {
1322 struct private_wmadec_data *pwd = fn->private_data;
1323 if (!pwd)
1324 return;
1325 wmadec_cleanup(pwd);
1326 free(fn->buf);
1327 fn->buf = NULL;
1328 free(fn->private_data);
1329 fn->private_data = NULL;
1330 }
1331
1332 static void wmadec_open(struct filter_node *fn)
1333 {
1334 fn->bufsize = 1024 * 1024;
1335 fn->buf = para_malloc(fn->bufsize);
1336 fn->private_data = NULL;
1337 fn->loaded = 0;
1338 }
1339
1340 /**
1341 * The init function of the wma decoder.
1342 *
1343 * \param f Its fields are filled in by the function.
1344 */
1345 void wmadec_filter_init(struct filter *f)
1346 {
1347 f->open = wmadec_open;
1348 f->close = wmadec_close;
1349 f->convert = wmadec_convert;
1350 }