Mercurial > libavcodec.hg
annotate wmadec.c @ 4604:40cd5f7f34e7 libavcodec
WMV3 Complex Profile initial support
author | kostya |
---|---|
date | Wed, 28 Feb 2007 06:39:58 +0000 |
parents | 7b9ce6f729ae |
children | 99d9dd34903b |
rev | line source |
---|---|
783 | 1 /* |
2 * WMA compatible decoder | |
3 * Copyright (c) 2002 The FFmpeg Project. | |
4 * | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
783 | 8 * modify it under the terms of the GNU Lesser General Public |
9 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
783 | 11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
783 | 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 * Lesser General Public License for more details. | |
16 * | |
17 * You should have received a copy of the GNU Lesser General Public | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3776
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
3022
diff
changeset
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
783 | 20 */ |
1106 | 21 |
22 /** | |
23 * @file wmadec.c | |
24 * WMA compatible decoder. | |
1967
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
25 * This decoder handles Microsoft Windows Media Audio data, versions 1 & 2. |
2967 | 26 * WMA v1 is identified by audio format 0x160 in Microsoft media files |
1967
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
27 * (ASF/AVI/WAV). WMA v2 is identified by audio format 0x161. |
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
28 * |
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
29 * To use this decoder, a calling application must supply the extra data |
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
30 * bytes provided with the WMA data. These are the extra, codec-specific |
2967 | 31 * bytes at the end of a WAVEFORMATEX data structure. Transmit these bytes |
32 * to the decoder using the extradata[_size] fields in AVCodecContext. There | |
1967
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
33 * should be 4 extra bytes for v1 data and 6 extra bytes for v2 data. |
1106 | 34 */ |
35 | |
783 | 36 #include "avcodec.h" |
4490 | 37 #include "wma.h" |
783 | 38 |
4490 | 39 #undef NDEBUG |
40 #include <assert.h> | |
3113 | 41 |
42 #define EXPVLCBITS 8 | |
43 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS) | |
44 | |
45 #define HGAINVLCBITS 9 | |
46 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS) | |
3022 | 47 |
4601 | 48 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len); |
783 | 49 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
50 #ifdef TRACE |
4600 | 51 static void dump_shorts(WMADecodeContext *s, const char *name, const short *tab, int n) |
783 | 52 { |
53 int i; | |
54 | |
4600 | 55 tprintf(s->avctx, "%s[%d]:\n", name, n); |
783 | 56 for(i=0;i<n;i++) { |
57 if ((i & 7) == 0) | |
4600 | 58 tprintf(s->avctx, "%4d: ", i); |
59 tprintf(s->avctx, " %5d.0", tab[i]); | |
783 | 60 if ((i & 7) == 7) |
4600 | 61 tprintf(s->avctx, "\n"); |
783 | 62 } |
63 } | |
64 | |
4600 | 65 static void dump_floats(WMADecodeContext *s, const char *name, int prec, const float *tab, int n) |
783 | 66 { |
67 int i; | |
68 | |
4600 | 69 tprintf(s->avctx, "%s[%d]:\n", name, n); |
783 | 70 for(i=0;i<n;i++) { |
71 if ((i & 7) == 0) | |
4600 | 72 tprintf(s->avctx, "%4d: ", i); |
73 tprintf(s->avctx, " %8.*f", prec, tab[i]); | |
783 | 74 if ((i & 7) == 7) |
4600 | 75 tprintf(s->avctx, "\n"); |
783 | 76 } |
77 if ((i & 7) != 0) | |
4600 | 78 tprintf(s->avctx, "\n"); |
783 | 79 } |
80 #endif | |
81 | |
82 static int wma_decode_init(AVCodecContext * avctx) | |
83 { | |
4601 | 84 WMACodecContext *s = avctx->priv_data; |
783 | 85 int i, flags1, flags2; |
86 uint8_t *extradata; | |
2967 | 87 |
4600 | 88 s->avctx = avctx; |
89 | |
783 | 90 /* extract flag infos */ |
91 flags1 = 0; | |
92 flags2 = 0; | |
93 extradata = avctx->extradata; | |
4490 | 94 if (avctx->codec->id == CODEC_ID_WMAV1 && avctx->extradata_size >= 4) { |
783 | 95 flags1 = extradata[0] | (extradata[1] << 8); |
96 flags2 = extradata[2] | (extradata[3] << 8); | |
4490 | 97 } else if (avctx->codec->id == CODEC_ID_WMAV2 && avctx->extradata_size >= 6) { |
2967 | 98 flags1 = extradata[0] | (extradata[1] << 8) | |
783 | 99 (extradata[2] << 16) | (extradata[3] << 24); |
100 flags2 = extradata[4] | (extradata[5] << 8); | |
101 } | |
4491 | 102 // for(i=0; i<avctx->extradata_size; i++) |
103 // av_log(NULL, AV_LOG_ERROR, "%02X ", extradata[i]); | |
4490 | 104 |
783 | 105 s->use_exp_vlc = flags2 & 0x0001; |
106 s->use_bit_reservoir = flags2 & 0x0002; | |
107 s->use_variable_block_len = flags2 & 0x0004; | |
108 | |
4490 | 109 ff_wma_init(avctx, flags2); |
783 | 110 |
111 /* init MDCT */ | |
112 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
113 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1); |
2967 | 114 |
783 | 115 if (s->use_noise_coding) { |
4490 | 116 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(ff_wma_hgain_huffbits), |
117 ff_wma_hgain_huffbits, 1, 1, | |
118 ff_wma_hgain_huffcodes, 2, 2, 0); | |
783 | 119 } |
120 | |
121 if (s->use_exp_vlc) { | |
4490 | 122 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(ff_wma_scale_huffbits), //FIXME move out of context |
123 ff_wma_scale_huffbits, 1, 1, | |
124 ff_wma_scale_huffcodes, 4, 4, 0); | |
783 | 125 } else { |
126 wma_lsp_to_curve_init(s, s->frame_len); | |
127 } | |
128 | |
129 return 0; | |
130 } | |
131 | |
4497 | 132 /** |
133 * interpolate values for a bigger or smaller block. The block must | |
134 * have multiple sizes | |
135 */ | |
783 | 136 static void interpolate_array(float *scale, int old_size, int new_size) |
137 { | |
138 int i, j, jincr, k; | |
139 float v; | |
140 | |
141 if (new_size > old_size) { | |
142 jincr = new_size / old_size; | |
143 j = new_size; | |
144 for(i = old_size - 1; i >=0; i--) { | |
145 v = scale[i]; | |
146 k = jincr; | |
147 do { | |
148 scale[--j] = v; | |
149 } while (--k); | |
150 } | |
151 } else if (new_size < old_size) { | |
152 j = 0; | |
153 jincr = old_size / new_size; | |
154 for(i = 0; i < new_size; i++) { | |
155 scale[i] = scale[j]; | |
156 j += jincr; | |
157 } | |
158 } | |
159 } | |
160 | |
4497 | 161 /** |
162 * compute x^-0.25 with an exponent and mantissa table. We use linear | |
163 * interpolation to reduce the mantissa table size at a small speed | |
164 * expense (linear interpolation approximately doubles the number of | |
165 * bits of precision). | |
166 */ | |
4601 | 167 static inline float pow_m1_4(WMACodecContext *s, float x) |
783 | 168 { |
169 union { | |
170 float f; | |
171 unsigned int v; | |
172 } u, t; | |
173 unsigned int e, m; | |
174 float a, b; | |
175 | |
176 u.f = x; | |
177 e = u.v >> 23; | |
178 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1); | |
179 /* build interpolation scale: 1 <= t < 2. */ | |
180 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23); | |
181 a = s->lsp_pow_m_table1[m]; | |
182 b = s->lsp_pow_m_table2[m]; | |
183 return s->lsp_pow_e_table[e] * (a + b * t.f); | |
184 } | |
185 | |
4601 | 186 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len) |
2967 | 187 { |
783 | 188 float wdel, a, b; |
189 int i, e, m; | |
190 | |
191 wdel = M_PI / frame_len; | |
192 for(i=0;i<frame_len;i++) | |
193 s->lsp_cos_table[i] = 2.0f * cos(wdel * i); | |
194 | |
195 /* tables for x^-0.25 computation */ | |
196 for(i=0;i<256;i++) { | |
197 e = i - 126; | |
198 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25); | |
199 } | |
200 | |
201 /* NOTE: these two tables are needed to avoid two operations in | |
202 pow_m1_4 */ | |
203 b = 1.0; | |
204 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) { | |
205 m = (1 << LSP_POW_BITS) + i; | |
206 a = (float)m * (0.5 / (1 << LSP_POW_BITS)); | |
207 a = pow(a, -0.25); | |
208 s->lsp_pow_m_table1[i] = 2 * a - b; | |
209 s->lsp_pow_m_table2[i] = b - a; | |
210 b = a; | |
211 } | |
212 #if 0 | |
213 for(i=1;i<20;i++) { | |
214 float v, r1, r2; | |
215 v = 5.0 / i; | |
216 r1 = pow_m1_4(s, v); | |
217 r2 = pow(v,-0.25); | |
218 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1); | |
219 } | |
220 #endif | |
221 } | |
222 | |
4497 | 223 /** |
224 * NOTE: We use the same code as Vorbis here | |
225 * @todo optimize it further with SSE/3Dnow | |
226 */ | |
4601 | 227 static void wma_lsp_to_curve(WMACodecContext *s, |
2967 | 228 float *out, float *val_max_ptr, |
783 | 229 int n, float *lsp) |
230 { | |
231 int i, j; | |
232 float p, q, w, v, val_max; | |
233 | |
234 val_max = 0; | |
235 for(i=0;i<n;i++) { | |
236 p = 0.5f; | |
237 q = 0.5f; | |
238 w = s->lsp_cos_table[i]; | |
239 for(j=1;j<NB_LSP_COEFS;j+=2){ | |
240 q *= w - lsp[j - 1]; | |
241 p *= w - lsp[j]; | |
242 } | |
243 p *= p * (2.0f - w); | |
244 q *= q * (2.0f + w); | |
245 v = p + q; | |
246 v = pow_m1_4(s, v); | |
247 if (v > val_max) | |
248 val_max = v; | |
249 out[i] = v; | |
250 } | |
251 *val_max_ptr = val_max; | |
252 } | |
253 | |
4497 | 254 /** |
255 * decode exponents coded with LSP coefficients (same idea as Vorbis) | |
256 */ | |
4601 | 257 static void decode_exp_lsp(WMACodecContext *s, int ch) |
783 | 258 { |
259 float lsp_coefs[NB_LSP_COEFS]; | |
260 int val, i; | |
261 | |
262 for(i = 0; i < NB_LSP_COEFS; i++) { | |
263 if (i == 0 || i >= 8) | |
264 val = get_bits(&s->gb, 3); | |
265 else | |
266 val = get_bits(&s->gb, 4); | |
4490 | 267 lsp_coefs[i] = ff_wma_lsp_codebook[i][val]; |
783 | 268 } |
269 | |
270 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch], | |
271 s->block_len, lsp_coefs); | |
272 } | |
273 | |
4497 | 274 /** |
275 * decode exponents coded with VLC codes | |
276 */ | |
4601 | 277 static int decode_exp_vlc(WMACodecContext *s, int ch) |
783 | 278 { |
279 int last_exp, n, code; | |
280 const uint16_t *ptr, *band_ptr; | |
281 float v, *q, max_scale, *q_end; | |
2967 | 282 |
783 | 283 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits]; |
284 ptr = band_ptr; | |
285 q = s->exponents[ch]; | |
286 q_end = q + s->block_len; | |
287 max_scale = 0; | |
288 if (s->version == 1) { | |
289 last_exp = get_bits(&s->gb, 5) + 10; | |
3235 | 290 /* XXX: use a table */ |
291 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 292 max_scale = v; |
293 n = *ptr++; | |
294 do { | |
295 *q++ = v; | |
296 } while (--n); | |
4490 | 297 }else |
298 last_exp = 36; | |
299 | |
783 | 300 while (q < q_end) { |
3113 | 301 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX); |
783 | 302 if (code < 0) |
303 return -1; | |
304 /* NOTE: this offset is the same as MPEG4 AAC ! */ | |
305 last_exp += code - 60; | |
3235 | 306 /* XXX: use a table */ |
307 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 308 if (v > max_scale) |
309 max_scale = v; | |
310 n = *ptr++; | |
311 do { | |
312 *q++ = v; | |
313 } while (--n); | |
314 } | |
315 s->max_exponent[ch] = max_scale; | |
316 return 0; | |
317 } | |
318 | |
4497 | 319 /** |
320 * @return 0 if OK. 1 if last block of frame. return -1 if | |
321 * unrecorrable error. | |
322 */ | |
4601 | 323 static int wma_decode_block(WMACodecContext *s) |
783 | 324 { |
325 int n, v, a, ch, code, bsize; | |
326 int coef_nb_bits, total_gain, parse_exponents; | |
327 int nb_coefs[MAX_CHANNELS]; | |
328 float mdct_norm; | |
329 | |
1343 | 330 #ifdef TRACE |
4600 | 331 tprintf(s->avctx, "***decode_block: %d:%d\n", s->frame_count - 1, s->block_num); |
1343 | 332 #endif |
783 | 333 |
334 /* compute current block length */ | |
335 if (s->use_variable_block_len) { | |
336 n = av_log2(s->nb_block_sizes - 1) + 1; | |
2967 | 337 |
783 | 338 if (s->reset_block_lengths) { |
339 s->reset_block_lengths = 0; | |
340 v = get_bits(&s->gb, n); | |
341 if (v >= s->nb_block_sizes) | |
342 return -1; | |
343 s->prev_block_len_bits = s->frame_len_bits - v; | |
344 v = get_bits(&s->gb, n); | |
345 if (v >= s->nb_block_sizes) | |
346 return -1; | |
347 s->block_len_bits = s->frame_len_bits - v; | |
348 } else { | |
349 /* update block lengths */ | |
350 s->prev_block_len_bits = s->block_len_bits; | |
351 s->block_len_bits = s->next_block_len_bits; | |
352 } | |
353 v = get_bits(&s->gb, n); | |
354 if (v >= s->nb_block_sizes) | |
355 return -1; | |
356 s->next_block_len_bits = s->frame_len_bits - v; | |
357 } else { | |
358 /* fixed block len */ | |
359 s->next_block_len_bits = s->frame_len_bits; | |
360 s->prev_block_len_bits = s->frame_len_bits; | |
361 s->block_len_bits = s->frame_len_bits; | |
362 } | |
363 | |
364 /* now check if the block length is coherent with the frame length */ | |
365 s->block_len = 1 << s->block_len_bits; | |
366 if ((s->block_pos + s->block_len) > s->frame_len) | |
367 return -1; | |
368 | |
369 if (s->nb_channels == 2) { | |
370 s->ms_stereo = get_bits(&s->gb, 1); | |
371 } | |
372 v = 0; | |
373 for(ch = 0; ch < s->nb_channels; ch++) { | |
374 a = get_bits(&s->gb, 1); | |
375 s->channel_coded[ch] = a; | |
376 v |= a; | |
377 } | |
378 /* if no channel coded, no need to go further */ | |
379 /* XXX: fix potential framing problems */ | |
380 if (!v) | |
381 goto next; | |
382 | |
383 bsize = s->frame_len_bits - s->block_len_bits; | |
384 | |
385 /* read total gain and extract corresponding number of bits for | |
386 coef escape coding */ | |
387 total_gain = 1; | |
388 for(;;) { | |
389 a = get_bits(&s->gb, 7); | |
390 total_gain += a; | |
391 if (a != 127) | |
392 break; | |
393 } | |
2967 | 394 |
4490 | 395 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain); |
783 | 396 |
397 /* compute number of coefficients */ | |
398 n = s->coefs_end[bsize] - s->coefs_start; | |
399 for(ch = 0; ch < s->nb_channels; ch++) | |
400 nb_coefs[ch] = n; | |
401 | |
402 /* complex coding */ | |
403 if (s->use_noise_coding) { | |
404 | |
405 for(ch = 0; ch < s->nb_channels; ch++) { | |
406 if (s->channel_coded[ch]) { | |
407 int i, n, a; | |
408 n = s->exponent_high_sizes[bsize]; | |
409 for(i=0;i<n;i++) { | |
410 a = get_bits(&s->gb, 1); | |
411 s->high_band_coded[ch][i] = a; | |
412 /* if noise coding, the coefficients are not transmitted */ | |
413 if (a) | |
414 nb_coefs[ch] -= s->exponent_high_bands[bsize][i]; | |
415 } | |
416 } | |
417 } | |
418 for(ch = 0; ch < s->nb_channels; ch++) { | |
419 if (s->channel_coded[ch]) { | |
420 int i, n, val, code; | |
421 | |
422 n = s->exponent_high_sizes[bsize]; | |
423 val = (int)0x80000000; | |
424 for(i=0;i<n;i++) { | |
425 if (s->high_band_coded[ch][i]) { | |
426 if (val == (int)0x80000000) { | |
427 val = get_bits(&s->gb, 7) - 19; | |
428 } else { | |
3113 | 429 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX); |
783 | 430 if (code < 0) |
431 return -1; | |
432 val += code - 18; | |
433 } | |
434 s->high_band_values[ch][i] = val; | |
435 } | |
436 } | |
437 } | |
438 } | |
439 } | |
2967 | 440 |
783 | 441 /* exposant can be interpolated in short blocks. */ |
442 parse_exponents = 1; | |
443 if (s->block_len_bits != s->frame_len_bits) { | |
444 parse_exponents = get_bits(&s->gb, 1); | |
445 } | |
2967 | 446 |
783 | 447 if (parse_exponents) { |
448 for(ch = 0; ch < s->nb_channels; ch++) { | |
449 if (s->channel_coded[ch]) { | |
450 if (s->use_exp_vlc) { | |
451 if (decode_exp_vlc(s, ch) < 0) | |
452 return -1; | |
453 } else { | |
454 decode_exp_lsp(s, ch); | |
455 } | |
456 } | |
457 } | |
458 } else { | |
459 for(ch = 0; ch < s->nb_channels; ch++) { | |
460 if (s->channel_coded[ch]) { | |
2967 | 461 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits, |
783 | 462 s->block_len); |
463 } | |
464 } | |
465 } | |
466 | |
467 /* parse spectral coefficients : just RLE encoding */ | |
468 for(ch = 0; ch < s->nb_channels; ch++) { | |
469 if (s->channel_coded[ch]) { | |
470 VLC *coef_vlc; | |
471 int level, run, sign, tindex; | |
472 int16_t *ptr, *eptr; | |
3776 | 473 const uint16_t *level_table, *run_table; |
783 | 474 |
475 /* special VLC tables are used for ms stereo because | |
476 there is potentially less energy there */ | |
477 tindex = (ch == 1 && s->ms_stereo); | |
478 coef_vlc = &s->coef_vlc[tindex]; | |
479 run_table = s->run_table[tindex]; | |
480 level_table = s->level_table[tindex]; | |
481 /* XXX: optimize */ | |
482 ptr = &s->coefs1[ch][0]; | |
483 eptr = ptr + nb_coefs[ch]; | |
484 memset(ptr, 0, s->block_len * sizeof(int16_t)); | |
485 for(;;) { | |
3113 | 486 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX); |
783 | 487 if (code < 0) |
488 return -1; | |
489 if (code == 1) { | |
490 /* EOB */ | |
491 break; | |
492 } else if (code == 0) { | |
493 /* escape */ | |
494 level = get_bits(&s->gb, coef_nb_bits); | |
495 /* NOTE: this is rather suboptimal. reading | |
496 block_len_bits would be better */ | |
497 run = get_bits(&s->gb, s->frame_len_bits); | |
498 } else { | |
499 /* normal code */ | |
500 run = run_table[code]; | |
501 level = level_table[code]; | |
502 } | |
503 sign = get_bits(&s->gb, 1); | |
504 if (!sign) | |
505 level = -level; | |
506 ptr += run; | |
507 if (ptr >= eptr) | |
3361 | 508 { |
509 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n"); | |
510 break; | |
511 } | |
783 | 512 *ptr++ = level; |
513 /* NOTE: EOB can be omitted */ | |
514 if (ptr >= eptr) | |
515 break; | |
516 } | |
517 } | |
518 if (s->version == 1 && s->nb_channels >= 2) { | |
519 align_get_bits(&s->gb); | |
520 } | |
521 } | |
2967 | 522 |
783 | 523 /* normalize */ |
524 { | |
525 int n4 = s->block_len / 2; | |
526 mdct_norm = 1.0 / (float)n4; | |
527 if (s->version == 1) { | |
528 mdct_norm *= sqrt(n4); | |
529 } | |
530 } | |
531 | |
532 /* finally compute the MDCT coefficients */ | |
533 for(ch = 0; ch < s->nb_channels; ch++) { | |
534 if (s->channel_coded[ch]) { | |
535 int16_t *coefs1; | |
536 float *coefs, *exponents, mult, mult1, noise, *exp_ptr; | |
537 int i, j, n, n1, last_high_band; | |
538 float exp_power[HIGH_BAND_MAX_SIZE]; | |
539 | |
540 coefs1 = s->coefs1[ch]; | |
541 exponents = s->exponents[ch]; | |
3235 | 542 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch]; |
783 | 543 mult *= mdct_norm; |
544 coefs = s->coefs[ch]; | |
545 if (s->use_noise_coding) { | |
546 mult1 = mult; | |
547 /* very low freqs : noise */ | |
548 for(i = 0;i < s->coefs_start; i++) { | |
549 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1; | |
550 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
551 } | |
2967 | 552 |
783 | 553 n1 = s->exponent_high_sizes[bsize]; |
554 | |
555 /* compute power of high bands */ | |
2967 | 556 exp_ptr = exponents + |
557 s->high_band_start[bsize] - | |
783 | 558 s->coefs_start; |
559 last_high_band = 0; /* avoid warning */ | |
560 for(j=0;j<n1;j++) { | |
2967 | 561 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 562 s->block_len_bits][j]; |
563 if (s->high_band_coded[ch][j]) { | |
564 float e2, v; | |
565 e2 = 0; | |
566 for(i = 0;i < n; i++) { | |
567 v = exp_ptr[i]; | |
568 e2 += v * v; | |
569 } | |
570 exp_power[j] = e2 / n; | |
571 last_high_band = j; | |
4600 | 572 tprintf(s->avctx, "%d: power=%f (%d)\n", j, exp_power[j], n); |
783 | 573 } |
574 exp_ptr += n; | |
575 } | |
576 | |
577 /* main freqs and high freqs */ | |
578 for(j=-1;j<n1;j++) { | |
579 if (j < 0) { | |
2967 | 580 n = s->high_band_start[bsize] - |
783 | 581 s->coefs_start; |
582 } else { | |
2967 | 583 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 584 s->block_len_bits][j]; |
585 } | |
586 if (j >= 0 && s->high_band_coded[ch][j]) { | |
587 /* use noise with specified power */ | |
588 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]); | |
3235 | 589 /* XXX: use a table */ |
590 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05); | |
783 | 591 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult); |
592 mult1 *= mdct_norm; | |
593 for(i = 0;i < n; i++) { | |
594 noise = s->noise_table[s->noise_index]; | |
595 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
596 *coefs++ = (*exponents++) * noise * mult1; | |
597 } | |
598 } else { | |
599 /* coded values + small noise */ | |
600 for(i = 0;i < n; i++) { | |
601 noise = s->noise_table[s->noise_index]; | |
602 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
603 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult; | |
604 } | |
605 } | |
606 } | |
607 | |
608 /* very high freqs : noise */ | |
609 n = s->block_len - s->coefs_end[bsize]; | |
610 mult1 = mult * exponents[-1]; | |
611 for(i = 0; i < n; i++) { | |
612 *coefs++ = s->noise_table[s->noise_index] * mult1; | |
613 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
614 } | |
615 } else { | |
616 /* XXX: optimize more */ | |
617 for(i = 0;i < s->coefs_start; i++) | |
618 *coefs++ = 0.0; | |
619 n = nb_coefs[ch]; | |
620 for(i = 0;i < n; i++) { | |
621 *coefs++ = coefs1[i] * exponents[i] * mult; | |
622 } | |
623 n = s->block_len - s->coefs_end[bsize]; | |
624 for(i = 0;i < n; i++) | |
625 *coefs++ = 0.0; | |
626 } | |
627 } | |
628 } | |
629 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
630 #ifdef TRACE |
783 | 631 for(ch = 0; ch < s->nb_channels; ch++) { |
632 if (s->channel_coded[ch]) { | |
4600 | 633 dump_floats(s, "exponents", 3, s->exponents[ch], s->block_len); |
634 dump_floats(s, "coefs", 1, s->coefs[ch], s->block_len); | |
783 | 635 } |
636 } | |
637 #endif | |
2967 | 638 |
783 | 639 if (s->ms_stereo && s->channel_coded[1]) { |
640 float a, b; | |
641 int i; | |
642 | |
643 /* nominal case for ms stereo: we do it before mdct */ | |
644 /* no need to optimize this case because it should almost | |
645 never happen */ | |
646 if (!s->channel_coded[0]) { | |
4600 | 647 tprintf(s->avctx, "rare ms-stereo case happened\n"); |
783 | 648 memset(s->coefs[0], 0, sizeof(float) * s->block_len); |
649 s->channel_coded[0] = 1; | |
650 } | |
2967 | 651 |
783 | 652 for(i = 0; i < s->block_len; i++) { |
653 a = s->coefs[0][i]; | |
654 b = s->coefs[1][i]; | |
655 s->coefs[0][i] = a + b; | |
656 s->coefs[1][i] = a - b; | |
657 } | |
658 } | |
659 | |
660 /* build the window : we ensure that when the windows overlap | |
661 their squared sum is always 1 (MDCT reconstruction rule) */ | |
662 /* XXX: merge with output */ | |
663 { | |
664 int i, next_block_len, block_len, prev_block_len, n; | |
665 float *wptr; | |
666 | |
667 block_len = s->block_len; | |
668 prev_block_len = 1 << s->prev_block_len_bits; | |
669 next_block_len = 1 << s->next_block_len_bits; | |
670 | |
671 /* right part */ | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
672 wptr = s->window + block_len; |
783 | 673 if (block_len <= next_block_len) { |
674 for(i=0;i<block_len;i++) | |
675 *wptr++ = s->windows[bsize][i]; | |
676 } else { | |
677 /* overlap */ | |
678 n = (block_len / 2) - (next_block_len / 2); | |
679 for(i=0;i<n;i++) | |
680 *wptr++ = 1.0; | |
681 for(i=0;i<next_block_len;i++) | |
682 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i]; | |
683 for(i=0;i<n;i++) | |
684 *wptr++ = 0.0; | |
685 } | |
686 | |
687 /* left part */ | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
688 wptr = s->window + block_len; |
783 | 689 if (block_len <= prev_block_len) { |
690 for(i=0;i<block_len;i++) | |
691 *--wptr = s->windows[bsize][i]; | |
692 } else { | |
693 /* overlap */ | |
694 n = (block_len / 2) - (prev_block_len / 2); | |
695 for(i=0;i<n;i++) | |
696 *--wptr = 1.0; | |
697 for(i=0;i<prev_block_len;i++) | |
698 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i]; | |
699 for(i=0;i<n;i++) | |
700 *--wptr = 0.0; | |
701 } | |
702 } | |
703 | |
2967 | 704 |
783 | 705 for(ch = 0; ch < s->nb_channels; ch++) { |
706 if (s->channel_coded[ch]) { | |
707 float *ptr; | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
708 int n4, index, n; |
783 | 709 |
710 n = s->block_len; | |
711 n4 = s->block_len / 2; | |
3555 | 712 s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize], |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
713 s->output, s->coefs[ch], s->mdct_tmp); |
783 | 714 |
715 /* XXX: optimize all that by build the window and | |
716 multipying/adding at the same time */ | |
717 | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
718 /* multiply by the window and add in the frame */ |
783 | 719 index = (s->frame_len / 2) + s->block_pos - n4; |
720 ptr = &s->frame_out[ch][index]; | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
721 s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1); |
783 | 722 |
723 /* specific fast case for ms-stereo : add to second | |
724 channel if it is not coded */ | |
725 if (s->ms_stereo && !s->channel_coded[1]) { | |
726 ptr = &s->frame_out[1][index]; | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
727 s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1); |
783 | 728 } |
729 } | |
730 } | |
731 next: | |
732 /* update block number */ | |
733 s->block_num++; | |
734 s->block_pos += s->block_len; | |
735 if (s->block_pos >= s->frame_len) | |
736 return 1; | |
737 else | |
738 return 0; | |
739 } | |
740 | |
741 /* decode a frame of frame_len samples */ | |
4601 | 742 static int wma_decode_frame(WMACodecContext *s, int16_t *samples) |
783 | 743 { |
744 int ret, i, n, a, ch, incr; | |
745 int16_t *ptr; | |
746 float *iptr; | |
747 | |
1343 | 748 #ifdef TRACE |
4600 | 749 tprintf(s->avctx, "***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len); |
1343 | 750 #endif |
783 | 751 |
752 /* read each block */ | |
753 s->block_num = 0; | |
754 s->block_pos = 0; | |
755 for(;;) { | |
756 ret = wma_decode_block(s); | |
2967 | 757 if (ret < 0) |
783 | 758 return -1; |
759 if (ret) | |
760 break; | |
761 } | |
762 | |
763 /* convert frame to integer */ | |
764 n = s->frame_len; | |
765 incr = s->nb_channels; | |
766 for(ch = 0; ch < s->nb_channels; ch++) { | |
767 ptr = samples + ch; | |
768 iptr = s->frame_out[ch]; | |
769 | |
770 for(i=0;i<n;i++) { | |
797 | 771 a = lrintf(*iptr++); |
783 | 772 if (a > 32767) |
773 a = 32767; | |
774 else if (a < -32768) | |
775 a = -32768; | |
776 *ptr = a; | |
777 ptr += incr; | |
778 } | |
779 /* prepare for next block */ | |
780 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len], | |
781 s->frame_len * sizeof(float)); | |
782 /* XXX: suppress this */ | |
2967 | 783 memset(&s->frame_out[ch][s->frame_len], 0, |
783 | 784 s->frame_len * sizeof(float)); |
785 } | |
786 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
787 #ifdef TRACE |
4600 | 788 dump_shorts(s, "samples", samples, n * s->nb_channels); |
783 | 789 #endif |
790 return 0; | |
791 } | |
792 | |
2967 | 793 static int wma_decode_superframe(AVCodecContext *avctx, |
783 | 794 void *data, int *data_size, |
1064 | 795 uint8_t *buf, int buf_size) |
783 | 796 { |
4601 | 797 WMACodecContext *s = avctx->priv_data; |
783 | 798 int nb_frames, bit_offset, i, pos, len; |
799 uint8_t *q; | |
800 int16_t *samples; | |
2967 | 801 |
4600 | 802 tprintf(avctx, "***decode_superframe:\n"); |
783 | 803 |
1750 | 804 if(buf_size==0){ |
805 s->last_superframe_len = 0; | |
806 return 0; | |
807 } | |
2967 | 808 |
783 | 809 samples = data; |
810 | |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
972
diff
changeset
|
811 init_get_bits(&s->gb, buf, buf_size*8); |
2967 | 812 |
783 | 813 if (s->use_bit_reservoir) { |
814 /* read super frame header */ | |
815 get_bits(&s->gb, 4); /* super frame index */ | |
816 nb_frames = get_bits(&s->gb, 4) - 1; | |
817 | |
818 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3); | |
819 | |
820 if (s->last_superframe_len > 0) { | |
821 // printf("skip=%d\n", s->last_bitoffset); | |
822 /* add bit_offset bits to last frame */ | |
2967 | 823 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) > |
783 | 824 MAX_CODED_SUPERFRAME_SIZE) |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
825 goto fail; |
783 | 826 q = s->last_superframe + s->last_superframe_len; |
827 len = bit_offset; | |
3362
c43fcf831f7c
Do not read full byte when less than 8 bits are still to be read.
reimar
parents:
3361
diff
changeset
|
828 while (len > 7) { |
783 | 829 *q++ = (get_bits)(&s->gb, 8); |
830 len -= 8; | |
831 } | |
832 if (len > 0) { | |
833 *q++ = (get_bits)(&s->gb, len) << (8 - len); | |
834 } | |
2967 | 835 |
783 | 836 /* XXX: bit_offset bits into last frame */ |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
972
diff
changeset
|
837 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8); |
783 | 838 /* skip unused bits */ |
839 if (s->last_bitoffset > 0) | |
840 skip_bits(&s->gb, s->last_bitoffset); | |
841 /* this frame is stored in the last superframe and in the | |
842 current one */ | |
843 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
844 goto fail; |
783 | 845 samples += s->nb_channels * s->frame_len; |
846 } | |
847 | |
848 /* read each frame starting from bit_offset */ | |
849 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3; | |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
972
diff
changeset
|
850 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8); |
783 | 851 len = pos & 7; |
852 if (len > 0) | |
853 skip_bits(&s->gb, len); | |
2967 | 854 |
783 | 855 s->reset_block_lengths = 1; |
856 for(i=0;i<nb_frames;i++) { | |
857 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
858 goto fail; |
783 | 859 samples += s->nb_channels * s->frame_len; |
860 } | |
861 | |
862 /* we copy the end of the frame in the last frame buffer */ | |
863 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7); | |
864 s->last_bitoffset = pos & 7; | |
865 pos >>= 3; | |
866 len = buf_size - pos; | |
819 | 867 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) { |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
868 goto fail; |
783 | 869 } |
870 s->last_superframe_len = len; | |
871 memcpy(s->last_superframe, buf + pos, len); | |
872 } else { | |
873 /* single frame decode */ | |
874 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
875 goto fail; |
783 | 876 samples += s->nb_channels * s->frame_len; |
877 } | |
4490 | 878 |
879 //av_log(NULL, AV_LOG_ERROR, "%d %d %d %d outbytes:%d eaten:%d\n", s->frame_len_bits, s->block_len_bits, s->frame_len, s->block_len, (int8_t *)samples - (int8_t *)data, s->block_align); | |
880 | |
783 | 881 *data_size = (int8_t *)samples - (int8_t *)data; |
882 return s->block_align; | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
883 fail: |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
884 /* when error, we reset the bit reservoir */ |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
885 s->last_superframe_len = 0; |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
886 return -1; |
783 | 887 } |
888 | |
889 AVCodec wmav1_decoder = | |
890 { | |
891 "wmav1", | |
892 CODEC_TYPE_AUDIO, | |
893 CODEC_ID_WMAV1, | |
4601 | 894 sizeof(WMACodecContext), |
783 | 895 wma_decode_init, |
896 NULL, | |
4490 | 897 ff_wma_end, |
783 | 898 wma_decode_superframe, |
899 }; | |
900 | |
901 AVCodec wmav2_decoder = | |
902 { | |
903 "wmav2", | |
904 CODEC_TYPE_AUDIO, | |
905 CODEC_ID_WMAV2, | |
4601 | 906 sizeof(WMACodecContext), |
783 | 907 wma_decode_init, |
908 NULL, | |
4490 | 909 ff_wma_end, |
783 | 910 wma_decode_superframe, |
911 }; |