Mercurial > libavcodec.hg
annotate wmadec.c @ 4490:0efc832d9102 libavcodec
wma encoder
author | michael |
---|---|
date | Tue, 06 Feb 2007 20:19:04 +0000 |
parents | b43bd0c56eaa |
children | a6ca201dd590 |
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 |
783 | 48 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len); |
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 |
783 | 51 static void dump_shorts(const char *name, const short *tab, int n) |
52 { | |
53 int i; | |
54 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
55 tprintf("%s[%d]:\n", name, n); |
783 | 56 for(i=0;i<n;i++) { |
57 if ((i & 7) == 0) | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
58 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
59 tprintf(" %5d.0", tab[i]); |
783 | 60 if ((i & 7) == 7) |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
61 tprintf("\n"); |
783 | 62 } |
63 } | |
64 | |
65 static void dump_floats(const char *name, int prec, const float *tab, int n) | |
66 { | |
67 int i; | |
68 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
69 tprintf("%s[%d]:\n", name, n); |
783 | 70 for(i=0;i<n;i++) { |
71 if ((i & 7) == 0) | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
72 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
73 tprintf(" %8.*f", prec, tab[i]); |
783 | 74 if ((i & 7) == 7) |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
75 tprintf("\n"); |
783 | 76 } |
77 if ((i & 7) != 0) | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
78 tprintf("\n"); |
783 | 79 } |
80 #endif | |
81 | |
82 static int wma_decode_init(AVCodecContext * avctx) | |
83 { | |
84 WMADecodeContext *s = avctx->priv_data; | |
85 int i, flags1, flags2; | |
86 uint8_t *extradata; | |
2967 | 87 |
783 | 88 /* extract flag infos */ |
89 flags1 = 0; | |
90 flags2 = 0; | |
91 extradata = avctx->extradata; | |
4490 | 92 if (avctx->codec->id == CODEC_ID_WMAV1 && avctx->extradata_size >= 4) { |
783 | 93 flags1 = extradata[0] | (extradata[1] << 8); |
94 flags2 = extradata[2] | (extradata[3] << 8); | |
4490 | 95 } else if (avctx->codec->id == CODEC_ID_WMAV2 && avctx->extradata_size >= 6) { |
2967 | 96 flags1 = extradata[0] | (extradata[1] << 8) | |
783 | 97 (extradata[2] << 16) | (extradata[3] << 24); |
98 flags2 = extradata[4] | (extradata[5] << 8); | |
99 } | |
4490 | 100 for(i=0; i<avctx->extradata_size; i++) |
101 av_log(NULL, AV_LOG_ERROR, "%02X ", extradata[i]); | |
102 | |
783 | 103 s->use_exp_vlc = flags2 & 0x0001; |
104 s->use_bit_reservoir = flags2 & 0x0002; | |
105 s->use_variable_block_len = flags2 & 0x0004; | |
106 | |
4490 | 107 ff_wma_init(avctx, flags2); |
783 | 108 |
109 /* init MDCT */ | |
110 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
111 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1); |
2967 | 112 |
783 | 113 if (s->use_noise_coding) { |
4490 | 114 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(ff_wma_hgain_huffbits), |
115 ff_wma_hgain_huffbits, 1, 1, | |
116 ff_wma_hgain_huffcodes, 2, 2, 0); | |
783 | 117 } |
118 | |
119 if (s->use_exp_vlc) { | |
4490 | 120 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(ff_wma_scale_huffbits), //FIXME move out of context |
121 ff_wma_scale_huffbits, 1, 1, | |
122 ff_wma_scale_huffcodes, 4, 4, 0); | |
783 | 123 } else { |
124 wma_lsp_to_curve_init(s, s->frame_len); | |
125 } | |
126 | |
127 return 0; | |
128 } | |
129 | |
130 /* interpolate values for a bigger or smaller block. The block must | |
131 have multiple sizes */ | |
132 static void interpolate_array(float *scale, int old_size, int new_size) | |
133 { | |
134 int i, j, jincr, k; | |
135 float v; | |
136 | |
137 if (new_size > old_size) { | |
138 jincr = new_size / old_size; | |
139 j = new_size; | |
140 for(i = old_size - 1; i >=0; i--) { | |
141 v = scale[i]; | |
142 k = jincr; | |
143 do { | |
144 scale[--j] = v; | |
145 } while (--k); | |
146 } | |
147 } else if (new_size < old_size) { | |
148 j = 0; | |
149 jincr = old_size / new_size; | |
150 for(i = 0; i < new_size; i++) { | |
151 scale[i] = scale[j]; | |
152 j += jincr; | |
153 } | |
154 } | |
155 } | |
156 | |
157 /* compute x^-0.25 with an exponent and mantissa table. We use linear | |
158 interpolation to reduce the mantissa table size at a small speed | |
159 expense (linear interpolation approximately doubles the number of | |
160 bits of precision). */ | |
161 static inline float pow_m1_4(WMADecodeContext *s, float x) | |
162 { | |
163 union { | |
164 float f; | |
165 unsigned int v; | |
166 } u, t; | |
167 unsigned int e, m; | |
168 float a, b; | |
169 | |
170 u.f = x; | |
171 e = u.v >> 23; | |
172 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1); | |
173 /* build interpolation scale: 1 <= t < 2. */ | |
174 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23); | |
175 a = s->lsp_pow_m_table1[m]; | |
176 b = s->lsp_pow_m_table2[m]; | |
177 return s->lsp_pow_e_table[e] * (a + b * t.f); | |
178 } | |
179 | |
180 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len) | |
2967 | 181 { |
783 | 182 float wdel, a, b; |
183 int i, e, m; | |
184 | |
185 wdel = M_PI / frame_len; | |
186 for(i=0;i<frame_len;i++) | |
187 s->lsp_cos_table[i] = 2.0f * cos(wdel * i); | |
188 | |
189 /* tables for x^-0.25 computation */ | |
190 for(i=0;i<256;i++) { | |
191 e = i - 126; | |
192 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25); | |
193 } | |
194 | |
195 /* NOTE: these two tables are needed to avoid two operations in | |
196 pow_m1_4 */ | |
197 b = 1.0; | |
198 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) { | |
199 m = (1 << LSP_POW_BITS) + i; | |
200 a = (float)m * (0.5 / (1 << LSP_POW_BITS)); | |
201 a = pow(a, -0.25); | |
202 s->lsp_pow_m_table1[i] = 2 * a - b; | |
203 s->lsp_pow_m_table2[i] = b - a; | |
204 b = a; | |
205 } | |
206 #if 0 | |
207 for(i=1;i<20;i++) { | |
208 float v, r1, r2; | |
209 v = 5.0 / i; | |
210 r1 = pow_m1_4(s, v); | |
211 r2 = pow(v,-0.25); | |
212 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1); | |
213 } | |
214 #endif | |
215 } | |
216 | |
217 /* NOTE: We use the same code as Vorbis here */ | |
218 /* XXX: optimize it further with SSE/3Dnow */ | |
2967 | 219 static void wma_lsp_to_curve(WMADecodeContext *s, |
220 float *out, float *val_max_ptr, | |
783 | 221 int n, float *lsp) |
222 { | |
223 int i, j; | |
224 float p, q, w, v, val_max; | |
225 | |
226 val_max = 0; | |
227 for(i=0;i<n;i++) { | |
228 p = 0.5f; | |
229 q = 0.5f; | |
230 w = s->lsp_cos_table[i]; | |
231 for(j=1;j<NB_LSP_COEFS;j+=2){ | |
232 q *= w - lsp[j - 1]; | |
233 p *= w - lsp[j]; | |
234 } | |
235 p *= p * (2.0f - w); | |
236 q *= q * (2.0f + w); | |
237 v = p + q; | |
238 v = pow_m1_4(s, v); | |
239 if (v > val_max) | |
240 val_max = v; | |
241 out[i] = v; | |
242 } | |
243 *val_max_ptr = val_max; | |
244 } | |
245 | |
246 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */ | |
247 static void decode_exp_lsp(WMADecodeContext *s, int ch) | |
248 { | |
249 float lsp_coefs[NB_LSP_COEFS]; | |
250 int val, i; | |
251 | |
252 for(i = 0; i < NB_LSP_COEFS; i++) { | |
253 if (i == 0 || i >= 8) | |
254 val = get_bits(&s->gb, 3); | |
255 else | |
256 val = get_bits(&s->gb, 4); | |
4490 | 257 lsp_coefs[i] = ff_wma_lsp_codebook[i][val]; |
783 | 258 } |
259 | |
260 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch], | |
261 s->block_len, lsp_coefs); | |
262 } | |
263 | |
264 /* decode exponents coded with VLC codes */ | |
265 static int decode_exp_vlc(WMADecodeContext *s, int ch) | |
266 { | |
267 int last_exp, n, code; | |
268 const uint16_t *ptr, *band_ptr; | |
269 float v, *q, max_scale, *q_end; | |
2967 | 270 |
783 | 271 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits]; |
272 ptr = band_ptr; | |
273 q = s->exponents[ch]; | |
274 q_end = q + s->block_len; | |
275 max_scale = 0; | |
276 if (s->version == 1) { | |
277 last_exp = get_bits(&s->gb, 5) + 10; | |
3235 | 278 /* XXX: use a table */ |
279 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 280 max_scale = v; |
281 n = *ptr++; | |
282 do { | |
283 *q++ = v; | |
284 } while (--n); | |
4490 | 285 }else |
286 last_exp = 36; | |
287 | |
783 | 288 while (q < q_end) { |
3113 | 289 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX); |
783 | 290 if (code < 0) |
291 return -1; | |
292 /* NOTE: this offset is the same as MPEG4 AAC ! */ | |
293 last_exp += code - 60; | |
3235 | 294 /* XXX: use a table */ |
295 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 296 if (v > max_scale) |
297 max_scale = v; | |
298 n = *ptr++; | |
299 do { | |
300 *q++ = v; | |
301 } while (--n); | |
302 } | |
303 s->max_exponent[ch] = max_scale; | |
304 return 0; | |
305 } | |
306 | |
307 /* return 0 if OK. return 1 if last block of frame. return -1 if | |
308 unrecorrable error. */ | |
309 static int wma_decode_block(WMADecodeContext *s) | |
310 { | |
311 int n, v, a, ch, code, bsize; | |
312 int coef_nb_bits, total_gain, parse_exponents; | |
313 int nb_coefs[MAX_CHANNELS]; | |
314 float mdct_norm; | |
315 | |
1343 | 316 #ifdef TRACE |
317 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num); | |
318 #endif | |
783 | 319 |
320 /* compute current block length */ | |
321 if (s->use_variable_block_len) { | |
322 n = av_log2(s->nb_block_sizes - 1) + 1; | |
2967 | 323 |
783 | 324 if (s->reset_block_lengths) { |
325 s->reset_block_lengths = 0; | |
326 v = get_bits(&s->gb, n); | |
327 if (v >= s->nb_block_sizes) | |
328 return -1; | |
329 s->prev_block_len_bits = s->frame_len_bits - v; | |
330 v = get_bits(&s->gb, n); | |
331 if (v >= s->nb_block_sizes) | |
332 return -1; | |
333 s->block_len_bits = s->frame_len_bits - v; | |
334 } else { | |
335 /* update block lengths */ | |
336 s->prev_block_len_bits = s->block_len_bits; | |
337 s->block_len_bits = s->next_block_len_bits; | |
338 } | |
339 v = get_bits(&s->gb, n); | |
340 if (v >= s->nb_block_sizes) | |
341 return -1; | |
342 s->next_block_len_bits = s->frame_len_bits - v; | |
343 } else { | |
344 /* fixed block len */ | |
345 s->next_block_len_bits = s->frame_len_bits; | |
346 s->prev_block_len_bits = s->frame_len_bits; | |
347 s->block_len_bits = s->frame_len_bits; | |
348 } | |
349 | |
350 /* now check if the block length is coherent with the frame length */ | |
351 s->block_len = 1 << s->block_len_bits; | |
352 if ((s->block_pos + s->block_len) > s->frame_len) | |
353 return -1; | |
354 | |
355 if (s->nb_channels == 2) { | |
356 s->ms_stereo = get_bits(&s->gb, 1); | |
357 } | |
358 v = 0; | |
359 for(ch = 0; ch < s->nb_channels; ch++) { | |
360 a = get_bits(&s->gb, 1); | |
361 s->channel_coded[ch] = a; | |
362 v |= a; | |
363 } | |
364 /* if no channel coded, no need to go further */ | |
365 /* XXX: fix potential framing problems */ | |
366 if (!v) | |
367 goto next; | |
368 | |
369 bsize = s->frame_len_bits - s->block_len_bits; | |
370 | |
371 /* read total gain and extract corresponding number of bits for | |
372 coef escape coding */ | |
373 total_gain = 1; | |
374 for(;;) { | |
375 a = get_bits(&s->gb, 7); | |
376 total_gain += a; | |
377 if (a != 127) | |
378 break; | |
379 } | |
2967 | 380 |
4490 | 381 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain); |
783 | 382 |
383 /* compute number of coefficients */ | |
384 n = s->coefs_end[bsize] - s->coefs_start; | |
385 for(ch = 0; ch < s->nb_channels; ch++) | |
386 nb_coefs[ch] = n; | |
387 | |
388 /* complex coding */ | |
389 if (s->use_noise_coding) { | |
390 | |
391 for(ch = 0; ch < s->nb_channels; ch++) { | |
392 if (s->channel_coded[ch]) { | |
393 int i, n, a; | |
394 n = s->exponent_high_sizes[bsize]; | |
395 for(i=0;i<n;i++) { | |
396 a = get_bits(&s->gb, 1); | |
397 s->high_band_coded[ch][i] = a; | |
398 /* if noise coding, the coefficients are not transmitted */ | |
399 if (a) | |
400 nb_coefs[ch] -= s->exponent_high_bands[bsize][i]; | |
401 } | |
402 } | |
403 } | |
404 for(ch = 0; ch < s->nb_channels; ch++) { | |
405 if (s->channel_coded[ch]) { | |
406 int i, n, val, code; | |
407 | |
408 n = s->exponent_high_sizes[bsize]; | |
409 val = (int)0x80000000; | |
410 for(i=0;i<n;i++) { | |
411 if (s->high_band_coded[ch][i]) { | |
412 if (val == (int)0x80000000) { | |
413 val = get_bits(&s->gb, 7) - 19; | |
414 } else { | |
3113 | 415 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX); |
783 | 416 if (code < 0) |
417 return -1; | |
418 val += code - 18; | |
419 } | |
420 s->high_band_values[ch][i] = val; | |
421 } | |
422 } | |
423 } | |
424 } | |
425 } | |
2967 | 426 |
783 | 427 /* exposant can be interpolated in short blocks. */ |
428 parse_exponents = 1; | |
429 if (s->block_len_bits != s->frame_len_bits) { | |
430 parse_exponents = get_bits(&s->gb, 1); | |
431 } | |
2967 | 432 |
783 | 433 if (parse_exponents) { |
434 for(ch = 0; ch < s->nb_channels; ch++) { | |
435 if (s->channel_coded[ch]) { | |
436 if (s->use_exp_vlc) { | |
437 if (decode_exp_vlc(s, ch) < 0) | |
438 return -1; | |
439 } else { | |
440 decode_exp_lsp(s, ch); | |
441 } | |
442 } | |
443 } | |
444 } else { | |
445 for(ch = 0; ch < s->nb_channels; ch++) { | |
446 if (s->channel_coded[ch]) { | |
2967 | 447 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits, |
783 | 448 s->block_len); |
449 } | |
450 } | |
451 } | |
452 | |
453 /* parse spectral coefficients : just RLE encoding */ | |
454 for(ch = 0; ch < s->nb_channels; ch++) { | |
455 if (s->channel_coded[ch]) { | |
456 VLC *coef_vlc; | |
457 int level, run, sign, tindex; | |
458 int16_t *ptr, *eptr; | |
3776 | 459 const uint16_t *level_table, *run_table; |
783 | 460 |
461 /* special VLC tables are used for ms stereo because | |
462 there is potentially less energy there */ | |
463 tindex = (ch == 1 && s->ms_stereo); | |
464 coef_vlc = &s->coef_vlc[tindex]; | |
465 run_table = s->run_table[tindex]; | |
466 level_table = s->level_table[tindex]; | |
467 /* XXX: optimize */ | |
468 ptr = &s->coefs1[ch][0]; | |
469 eptr = ptr + nb_coefs[ch]; | |
470 memset(ptr, 0, s->block_len * sizeof(int16_t)); | |
471 for(;;) { | |
3113 | 472 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX); |
783 | 473 if (code < 0) |
474 return -1; | |
475 if (code == 1) { | |
476 /* EOB */ | |
477 break; | |
478 } else if (code == 0) { | |
479 /* escape */ | |
480 level = get_bits(&s->gb, coef_nb_bits); | |
481 /* NOTE: this is rather suboptimal. reading | |
482 block_len_bits would be better */ | |
483 run = get_bits(&s->gb, s->frame_len_bits); | |
484 } else { | |
485 /* normal code */ | |
486 run = run_table[code]; | |
487 level = level_table[code]; | |
488 } | |
489 sign = get_bits(&s->gb, 1); | |
490 if (!sign) | |
491 level = -level; | |
492 ptr += run; | |
493 if (ptr >= eptr) | |
3361 | 494 { |
495 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n"); | |
496 break; | |
497 } | |
783 | 498 *ptr++ = level; |
499 /* NOTE: EOB can be omitted */ | |
500 if (ptr >= eptr) | |
501 break; | |
502 } | |
503 } | |
504 if (s->version == 1 && s->nb_channels >= 2) { | |
505 align_get_bits(&s->gb); | |
506 } | |
507 } | |
2967 | 508 |
783 | 509 /* normalize */ |
510 { | |
511 int n4 = s->block_len / 2; | |
512 mdct_norm = 1.0 / (float)n4; | |
513 if (s->version == 1) { | |
514 mdct_norm *= sqrt(n4); | |
515 } | |
516 } | |
517 | |
518 /* finally compute the MDCT coefficients */ | |
519 for(ch = 0; ch < s->nb_channels; ch++) { | |
520 if (s->channel_coded[ch]) { | |
521 int16_t *coefs1; | |
522 float *coefs, *exponents, mult, mult1, noise, *exp_ptr; | |
523 int i, j, n, n1, last_high_band; | |
524 float exp_power[HIGH_BAND_MAX_SIZE]; | |
525 | |
526 coefs1 = s->coefs1[ch]; | |
527 exponents = s->exponents[ch]; | |
3235 | 528 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch]; |
783 | 529 mult *= mdct_norm; |
530 coefs = s->coefs[ch]; | |
531 if (s->use_noise_coding) { | |
532 mult1 = mult; | |
533 /* very low freqs : noise */ | |
534 for(i = 0;i < s->coefs_start; i++) { | |
535 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1; | |
536 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
537 } | |
2967 | 538 |
783 | 539 n1 = s->exponent_high_sizes[bsize]; |
540 | |
541 /* compute power of high bands */ | |
2967 | 542 exp_ptr = exponents + |
543 s->high_band_start[bsize] - | |
783 | 544 s->coefs_start; |
545 last_high_band = 0; /* avoid warning */ | |
546 for(j=0;j<n1;j++) { | |
2967 | 547 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 548 s->block_len_bits][j]; |
549 if (s->high_band_coded[ch][j]) { | |
550 float e2, v; | |
551 e2 = 0; | |
552 for(i = 0;i < n; i++) { | |
553 v = exp_ptr[i]; | |
554 e2 += v * v; | |
555 } | |
556 exp_power[j] = e2 / n; | |
557 last_high_band = j; | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
558 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n); |
783 | 559 } |
560 exp_ptr += n; | |
561 } | |
562 | |
563 /* main freqs and high freqs */ | |
564 for(j=-1;j<n1;j++) { | |
565 if (j < 0) { | |
2967 | 566 n = s->high_band_start[bsize] - |
783 | 567 s->coefs_start; |
568 } else { | |
2967 | 569 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 570 s->block_len_bits][j]; |
571 } | |
572 if (j >= 0 && s->high_band_coded[ch][j]) { | |
573 /* use noise with specified power */ | |
574 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]); | |
3235 | 575 /* XXX: use a table */ |
576 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05); | |
783 | 577 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult); |
578 mult1 *= mdct_norm; | |
579 for(i = 0;i < n; i++) { | |
580 noise = s->noise_table[s->noise_index]; | |
581 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
582 *coefs++ = (*exponents++) * noise * mult1; | |
583 } | |
584 } else { | |
585 /* coded values + small noise */ | |
586 for(i = 0;i < n; i++) { | |
587 noise = s->noise_table[s->noise_index]; | |
588 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
589 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult; | |
590 } | |
591 } | |
592 } | |
593 | |
594 /* very high freqs : noise */ | |
595 n = s->block_len - s->coefs_end[bsize]; | |
596 mult1 = mult * exponents[-1]; | |
597 for(i = 0; i < n; i++) { | |
598 *coefs++ = s->noise_table[s->noise_index] * mult1; | |
599 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
600 } | |
601 } else { | |
602 /* XXX: optimize more */ | |
603 for(i = 0;i < s->coefs_start; i++) | |
604 *coefs++ = 0.0; | |
605 n = nb_coefs[ch]; | |
606 for(i = 0;i < n; i++) { | |
607 *coefs++ = coefs1[i] * exponents[i] * mult; | |
608 } | |
609 n = s->block_len - s->coefs_end[bsize]; | |
610 for(i = 0;i < n; i++) | |
611 *coefs++ = 0.0; | |
612 } | |
613 } | |
614 } | |
615 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
616 #ifdef TRACE |
783 | 617 for(ch = 0; ch < s->nb_channels; ch++) { |
618 if (s->channel_coded[ch]) { | |
619 dump_floats("exponents", 3, s->exponents[ch], s->block_len); | |
620 dump_floats("coefs", 1, s->coefs[ch], s->block_len); | |
621 } | |
622 } | |
623 #endif | |
2967 | 624 |
783 | 625 if (s->ms_stereo && s->channel_coded[1]) { |
626 float a, b; | |
627 int i; | |
628 | |
629 /* nominal case for ms stereo: we do it before mdct */ | |
630 /* no need to optimize this case because it should almost | |
631 never happen */ | |
632 if (!s->channel_coded[0]) { | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
633 tprintf("rare ms-stereo case happened\n"); |
783 | 634 memset(s->coefs[0], 0, sizeof(float) * s->block_len); |
635 s->channel_coded[0] = 1; | |
636 } | |
2967 | 637 |
783 | 638 for(i = 0; i < s->block_len; i++) { |
639 a = s->coefs[0][i]; | |
640 b = s->coefs[1][i]; | |
641 s->coefs[0][i] = a + b; | |
642 s->coefs[1][i] = a - b; | |
643 } | |
644 } | |
645 | |
646 /* build the window : we ensure that when the windows overlap | |
647 their squared sum is always 1 (MDCT reconstruction rule) */ | |
648 /* XXX: merge with output */ | |
649 { | |
650 int i, next_block_len, block_len, prev_block_len, n; | |
651 float *wptr; | |
652 | |
653 block_len = s->block_len; | |
654 prev_block_len = 1 << s->prev_block_len_bits; | |
655 next_block_len = 1 << s->next_block_len_bits; | |
656 | |
657 /* right part */ | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
658 wptr = s->window + block_len; |
783 | 659 if (block_len <= next_block_len) { |
660 for(i=0;i<block_len;i++) | |
661 *wptr++ = s->windows[bsize][i]; | |
662 } else { | |
663 /* overlap */ | |
664 n = (block_len / 2) - (next_block_len / 2); | |
665 for(i=0;i<n;i++) | |
666 *wptr++ = 1.0; | |
667 for(i=0;i<next_block_len;i++) | |
668 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i]; | |
669 for(i=0;i<n;i++) | |
670 *wptr++ = 0.0; | |
671 } | |
672 | |
673 /* left part */ | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
674 wptr = s->window + block_len; |
783 | 675 if (block_len <= prev_block_len) { |
676 for(i=0;i<block_len;i++) | |
677 *--wptr = s->windows[bsize][i]; | |
678 } else { | |
679 /* overlap */ | |
680 n = (block_len / 2) - (prev_block_len / 2); | |
681 for(i=0;i<n;i++) | |
682 *--wptr = 1.0; | |
683 for(i=0;i<prev_block_len;i++) | |
684 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i]; | |
685 for(i=0;i<n;i++) | |
686 *--wptr = 0.0; | |
687 } | |
688 } | |
689 | |
2967 | 690 |
783 | 691 for(ch = 0; ch < s->nb_channels; ch++) { |
692 if (s->channel_coded[ch]) { | |
693 float *ptr; | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
694 int n4, index, n; |
783 | 695 |
696 n = s->block_len; | |
697 n4 = s->block_len / 2; | |
3555 | 698 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
|
699 s->output, s->coefs[ch], s->mdct_tmp); |
783 | 700 |
701 /* XXX: optimize all that by build the window and | |
702 multipying/adding at the same time */ | |
703 | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
704 /* multiply by the window and add in the frame */ |
783 | 705 index = (s->frame_len / 2) + s->block_pos - n4; |
706 ptr = &s->frame_out[ch][index]; | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
707 s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1); |
783 | 708 |
709 /* specific fast case for ms-stereo : add to second | |
710 channel if it is not coded */ | |
711 if (s->ms_stereo && !s->channel_coded[1]) { | |
712 ptr = &s->frame_out[1][index]; | |
4301
b43bd0c56eaa
Bug fix for crashes when SSE is used on unaligned arrays.
banan
parents:
3947
diff
changeset
|
713 s->dsp.vector_fmul_add_add(ptr,s->window,s->output,ptr,0,2*n,1); |
783 | 714 } |
715 } | |
716 } | |
717 next: | |
718 /* update block number */ | |
719 s->block_num++; | |
720 s->block_pos += s->block_len; | |
721 if (s->block_pos >= s->frame_len) | |
722 return 1; | |
723 else | |
724 return 0; | |
725 } | |
726 | |
727 /* decode a frame of frame_len samples */ | |
728 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples) | |
729 { | |
730 int ret, i, n, a, ch, incr; | |
731 int16_t *ptr; | |
732 float *iptr; | |
733 | |
1343 | 734 #ifdef TRACE |
735 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len); | |
736 #endif | |
783 | 737 |
738 /* read each block */ | |
739 s->block_num = 0; | |
740 s->block_pos = 0; | |
741 for(;;) { | |
742 ret = wma_decode_block(s); | |
2967 | 743 if (ret < 0) |
783 | 744 return -1; |
745 if (ret) | |
746 break; | |
747 } | |
748 | |
749 /* convert frame to integer */ | |
750 n = s->frame_len; | |
751 incr = s->nb_channels; | |
752 for(ch = 0; ch < s->nb_channels; ch++) { | |
753 ptr = samples + ch; | |
754 iptr = s->frame_out[ch]; | |
755 | |
756 for(i=0;i<n;i++) { | |
797 | 757 a = lrintf(*iptr++); |
783 | 758 if (a > 32767) |
759 a = 32767; | |
760 else if (a < -32768) | |
761 a = -32768; | |
762 *ptr = a; | |
763 ptr += incr; | |
764 } | |
765 /* prepare for next block */ | |
766 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len], | |
767 s->frame_len * sizeof(float)); | |
768 /* XXX: suppress this */ | |
2967 | 769 memset(&s->frame_out[ch][s->frame_len], 0, |
783 | 770 s->frame_len * sizeof(float)); |
771 } | |
772 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
773 #ifdef TRACE |
783 | 774 dump_shorts("samples", samples, n * s->nb_channels); |
775 #endif | |
776 return 0; | |
777 } | |
778 | |
2967 | 779 static int wma_decode_superframe(AVCodecContext *avctx, |
783 | 780 void *data, int *data_size, |
1064 | 781 uint8_t *buf, int buf_size) |
783 | 782 { |
783 WMADecodeContext *s = avctx->priv_data; | |
784 int nb_frames, bit_offset, i, pos, len; | |
785 uint8_t *q; | |
786 int16_t *samples; | |
2967 | 787 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
788 tprintf("***decode_superframe:\n"); |
783 | 789 |
1750 | 790 if(buf_size==0){ |
791 s->last_superframe_len = 0; | |
792 return 0; | |
793 } | |
2967 | 794 |
783 | 795 samples = data; |
796 | |
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
|
797 init_get_bits(&s->gb, buf, buf_size*8); |
2967 | 798 |
783 | 799 if (s->use_bit_reservoir) { |
800 /* read super frame header */ | |
801 get_bits(&s->gb, 4); /* super frame index */ | |
802 nb_frames = get_bits(&s->gb, 4) - 1; | |
803 | |
804 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3); | |
805 | |
806 if (s->last_superframe_len > 0) { | |
807 // printf("skip=%d\n", s->last_bitoffset); | |
808 /* add bit_offset bits to last frame */ | |
2967 | 809 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) > |
783 | 810 MAX_CODED_SUPERFRAME_SIZE) |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
811 goto fail; |
783 | 812 q = s->last_superframe + s->last_superframe_len; |
813 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
|
814 while (len > 7) { |
783 | 815 *q++ = (get_bits)(&s->gb, 8); |
816 len -= 8; | |
817 } | |
818 if (len > 0) { | |
819 *q++ = (get_bits)(&s->gb, len) << (8 - len); | |
820 } | |
2967 | 821 |
783 | 822 /* 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
|
823 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8); |
783 | 824 /* skip unused bits */ |
825 if (s->last_bitoffset > 0) | |
826 skip_bits(&s->gb, s->last_bitoffset); | |
827 /* this frame is stored in the last superframe and in the | |
828 current one */ | |
829 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
830 goto fail; |
783 | 831 samples += s->nb_channels * s->frame_len; |
832 } | |
833 | |
834 /* read each frame starting from bit_offset */ | |
835 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
|
836 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8); |
783 | 837 len = pos & 7; |
838 if (len > 0) | |
839 skip_bits(&s->gb, len); | |
2967 | 840 |
783 | 841 s->reset_block_lengths = 1; |
842 for(i=0;i<nb_frames;i++) { | |
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 /* we copy the end of the frame in the last frame buffer */ | |
849 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7); | |
850 s->last_bitoffset = pos & 7; | |
851 pos >>= 3; | |
852 len = buf_size - pos; | |
819 | 853 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
|
854 goto fail; |
783 | 855 } |
856 s->last_superframe_len = len; | |
857 memcpy(s->last_superframe, buf + pos, len); | |
858 } else { | |
859 /* single frame decode */ | |
860 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
861 goto fail; |
783 | 862 samples += s->nb_channels * s->frame_len; |
863 } | |
4490 | 864 |
865 //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); | |
866 | |
783 | 867 *data_size = (int8_t *)samples - (int8_t *)data; |
868 return s->block_align; | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
869 fail: |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
870 /* when error, we reset the bit reservoir */ |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
871 s->last_superframe_len = 0; |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
872 return -1; |
783 | 873 } |
874 | |
875 AVCodec wmav1_decoder = | |
876 { | |
877 "wmav1", | |
878 CODEC_TYPE_AUDIO, | |
879 CODEC_ID_WMAV1, | |
880 sizeof(WMADecodeContext), | |
881 wma_decode_init, | |
882 NULL, | |
4490 | 883 ff_wma_end, |
783 | 884 wma_decode_superframe, |
885 }; | |
886 | |
887 AVCodec wmav2_decoder = | |
888 { | |
889 "wmav2", | |
890 CODEC_TYPE_AUDIO, | |
891 CODEC_ID_WMAV2, | |
892 sizeof(WMADecodeContext), | |
893 wma_decode_init, | |
894 NULL, | |
4490 | 895 ff_wma_end, |
783 | 896 wma_decode_superframe, |
897 }; |