Mercurial > libavcodec.hg
annotate wmadec.c @ 3645:47821be55b6c libavcodec
mmx implementation of deblocking strength decision.
2-3% faster h264.
author | lorenm |
---|---|
date | Mon, 28 Aug 2006 09:33:01 +0000 |
parents | e28285ddde8d |
children | 44fdf93a0901 |
rev | line source |
---|---|
783 | 1 /* |
2 * WMA compatible decoder | |
3 * Copyright (c) 2002 The FFmpeg Project. | |
4 * | |
5 * This library is free software; you can redistribute it and/or | |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
9 * | |
10 * This library is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 * Lesser General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU Lesser General Public | |
16 * License along with this library; if not, write to the Free Software | |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
3022
diff
changeset
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
783 | 18 */ |
1106 | 19 |
20 /** | |
21 * @file wmadec.c | |
22 * WMA compatible decoder. | |
1967
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
23 * This decoder handles Microsoft Windows Media Audio data, versions 1 & 2. |
2967 | 24 * 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
|
25 * (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
|
26 * |
2b0fc6b25ab8
add the minimal documentation to make this decoder useful
melanson
parents:
1750
diff
changeset
|
27 * 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
|
28 * bytes provided with the WMA data. These are the extra, codec-specific |
2967 | 29 * bytes at the end of a WAVEFORMATEX data structure. Transmit these bytes |
30 * 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
|
31 * should be 4 extra bytes for v1 data and 6 extra bytes for v2 data. |
1106 | 32 */ |
33 | |
783 | 34 #include "avcodec.h" |
2398
582e635cfa08
common.c -> bitstream.c (and the single non bitstream func -> utils.c)
michael
parents:
2370
diff
changeset
|
35 #include "bitstream.h" |
783 | 36 #include "dsputil.h" |
37 | |
38 /* size of blocks */ | |
39 #define BLOCK_MIN_BITS 7 | |
40 #define BLOCK_MAX_BITS 11 | |
41 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS) | |
42 | |
43 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1) | |
44 | |
45 /* XXX: find exact max size */ | |
46 #define HIGH_BAND_MAX_SIZE 16 | |
47 | |
48 #define NB_LSP_COEFS 10 | |
49 | |
817 | 50 /* XXX: is it a suitable value ? */ |
2775
f3cdd51c9e16
WMA MAX_CODED_SUPERFRAME_SIZE too small patch by (Mark Weaver: mark-clist, npsl co uk)
michael
parents:
2398
diff
changeset
|
51 #define MAX_CODED_SUPERFRAME_SIZE 16384 |
783 | 52 |
53 #define MAX_CHANNELS 2 | |
54 | |
55 #define NOISE_TAB_SIZE 8192 | |
56 | |
57 #define LSP_POW_BITS 7 | |
58 | |
3022 | 59 #define VLCBITS 9 |
3113 | 60 #define VLCMAX ((22+VLCBITS-1)/VLCBITS) |
61 | |
62 #define EXPVLCBITS 8 | |
63 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS) | |
64 | |
65 #define HGAINVLCBITS 9 | |
66 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS) | |
3022 | 67 |
783 | 68 typedef struct WMADecodeContext { |
69 GetBitContext gb; | |
70 int sample_rate; | |
71 int nb_channels; | |
72 int bit_rate; | |
73 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */ | |
74 int block_align; | |
75 int use_bit_reservoir; | |
76 int use_variable_block_len; | |
77 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */ | |
78 int use_noise_coding; /* true if perceptual noise is added */ | |
79 int byte_offset_bits; | |
80 VLC exp_vlc; | |
81 int exponent_sizes[BLOCK_NB_SIZES]; | |
82 uint16_t exponent_bands[BLOCK_NB_SIZES][25]; | |
83 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */ | |
84 int coefs_start; /* first coded coef */ | |
85 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */ | |
86 int exponent_high_sizes[BLOCK_NB_SIZES]; | |
2967 | 87 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE]; |
783 | 88 VLC hgain_vlc; |
2967 | 89 |
783 | 90 /* coded values in high bands */ |
91 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE]; | |
92 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE]; | |
93 | |
94 /* there are two possible tables for spectral coefficients */ | |
95 VLC coef_vlc[2]; | |
96 uint16_t *run_table[2]; | |
97 uint16_t *level_table[2]; | |
98 /* frame info */ | |
99 int frame_len; /* frame length in samples */ | |
100 int frame_len_bits; /* frame_len = 1 << frame_len_bits */ | |
101 int nb_block_sizes; /* number of block sizes */ | |
102 /* block info */ | |
103 int reset_block_lengths; | |
104 int block_len_bits; /* log2 of current block length */ | |
105 int next_block_len_bits; /* log2 of next block length */ | |
106 int prev_block_len_bits; /* log2 of prev block length */ | |
107 int block_len; /* block length in samples */ | |
108 int block_num; /* block number in current frame */ | |
109 int block_pos; /* current position in frame */ | |
110 uint8_t ms_stereo; /* true if mid/side stereo mode */ | |
111 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */ | |
3089 | 112 DECLARE_ALIGNED_16(float, exponents[MAX_CHANNELS][BLOCK_MAX_SIZE]); |
783 | 113 float max_exponent[MAX_CHANNELS]; |
114 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE]; | |
3089 | 115 DECLARE_ALIGNED_16(float, coefs[MAX_CHANNELS][BLOCK_MAX_SIZE]); |
783 | 116 MDCTContext mdct_ctx[BLOCK_NB_SIZES]; |
1031
19de1445beb2
use av_malloc() functions - added av_strdup and av_realloc()
bellard
parents:
1025
diff
changeset
|
117 float *windows[BLOCK_NB_SIZES]; |
3089 | 118 DECLARE_ALIGNED_16(FFTSample, mdct_tmp[BLOCK_MAX_SIZE]); /* temporary storage for imdct */ |
783 | 119 /* output buffer for one frame and the last for IMDCT windowing */ |
3089 | 120 DECLARE_ALIGNED_16(float, frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2]); |
783 | 121 /* last frame info */ |
122 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */ | |
123 int last_bitoffset; | |
124 int last_superframe_len; | |
125 float noise_table[NOISE_TAB_SIZE]; | |
126 int noise_index; | |
127 float noise_mult; /* XXX: suppress that and integrate it in the noise array */ | |
128 /* lsp_to_curve tables */ | |
129 float lsp_cos_table[BLOCK_MAX_SIZE]; | |
130 float lsp_pow_e_table[256]; | |
131 float lsp_pow_m_table1[(1 << LSP_POW_BITS)]; | |
132 float lsp_pow_m_table2[(1 << LSP_POW_BITS)]; | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
133 DSPContext dsp; |
1343 | 134 |
135 #ifdef TRACE | |
136 int frame_count; | |
137 #endif | |
783 | 138 } WMADecodeContext; |
139 | |
140 typedef struct CoefVLCTable { | |
141 int n; /* total number of codes */ | |
142 const uint32_t *huffcodes; /* VLC bit values */ | |
143 const uint8_t *huffbits; /* VLC bit size */ | |
144 const uint16_t *levels; /* table to build run/level tables */ | |
145 } CoefVLCTable; | |
146 | |
147 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len); | |
148 | |
149 #include "wmadata.h" | |
150 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
151 #ifdef TRACE |
783 | 152 static void dump_shorts(const char *name, const short *tab, int n) |
153 { | |
154 int i; | |
155 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
156 tprintf("%s[%d]:\n", name, n); |
783 | 157 for(i=0;i<n;i++) { |
158 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
|
159 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
160 tprintf(" %5d.0", tab[i]); |
783 | 161 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
|
162 tprintf("\n"); |
783 | 163 } |
164 } | |
165 | |
166 static void dump_floats(const char *name, int prec, const float *tab, int n) | |
167 { | |
168 int i; | |
169 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
170 tprintf("%s[%d]:\n", name, n); |
783 | 171 for(i=0;i<n;i++) { |
172 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
|
173 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
174 tprintf(" %8.*f", prec, tab[i]); |
783 | 175 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
|
176 tprintf("\n"); |
783 | 177 } |
178 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
|
179 tprintf("\n"); |
783 | 180 } |
181 #endif | |
182 | |
183 /* XXX: use same run/length optimization as mpeg decoders */ | |
2967 | 184 static void init_coef_vlc(VLC *vlc, |
783 | 185 uint16_t **prun_table, uint16_t **plevel_table, |
186 const CoefVLCTable *vlc_table) | |
187 { | |
188 int n = vlc_table->n; | |
189 const uint8_t *table_bits = vlc_table->huffbits; | |
190 const uint32_t *table_codes = vlc_table->huffcodes; | |
191 const uint16_t *levels_table = vlc_table->levels; | |
192 uint16_t *run_table, *level_table; | |
193 const uint16_t *p; | |
194 int i, l, j, level; | |
195 | |
3113 | 196 init_vlc(vlc, VLCBITS, n, table_bits, 1, 1, table_codes, 4, 4, 0); |
783 | 197 |
1031
19de1445beb2
use av_malloc() functions - added av_strdup and av_realloc()
bellard
parents:
1025
diff
changeset
|
198 run_table = av_malloc(n * sizeof(uint16_t)); |
19de1445beb2
use av_malloc() functions - added av_strdup and av_realloc()
bellard
parents:
1025
diff
changeset
|
199 level_table = av_malloc(n * sizeof(uint16_t)); |
783 | 200 p = levels_table; |
201 i = 2; | |
202 level = 1; | |
203 while (i < n) { | |
204 l = *p++; | |
205 for(j=0;j<l;j++) { | |
206 run_table[i] = j; | |
207 level_table[i] = level; | |
208 i++; | |
209 } | |
210 level++; | |
211 } | |
212 *prun_table = run_table; | |
213 *plevel_table = level_table; | |
214 } | |
215 | |
216 static int wma_decode_init(AVCodecContext * avctx) | |
217 { | |
218 WMADecodeContext *s = avctx->priv_data; | |
219 int i, flags1, flags2; | |
220 float *window; | |
221 uint8_t *extradata; | |
3235 | 222 float bps1, high_freq; |
223 volatile float bps; | |
783 | 224 int sample_rate1; |
225 int coef_vlc_table; | |
2967 | 226 |
783 | 227 s->sample_rate = avctx->sample_rate; |
228 s->nb_channels = avctx->channels; | |
229 s->bit_rate = avctx->bit_rate; | |
230 s->block_align = avctx->block_align; | |
231 | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
232 dsputil_init(&s->dsp, avctx); |
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
233 |
808
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
234 if (avctx->codec->id == CODEC_ID_WMAV1) { |
783 | 235 s->version = 1; |
236 } else { | |
237 s->version = 2; | |
238 } | |
2967 | 239 |
783 | 240 /* extract flag infos */ |
241 flags1 = 0; | |
242 flags2 = 0; | |
243 extradata = avctx->extradata; | |
244 if (s->version == 1 && avctx->extradata_size >= 4) { | |
245 flags1 = extradata[0] | (extradata[1] << 8); | |
246 flags2 = extradata[2] | (extradata[3] << 8); | |
247 } else if (s->version == 2 && avctx->extradata_size >= 6) { | |
2967 | 248 flags1 = extradata[0] | (extradata[1] << 8) | |
783 | 249 (extradata[2] << 16) | (extradata[3] << 24); |
250 flags2 = extradata[4] | (extradata[5] << 8); | |
251 } | |
252 s->use_exp_vlc = flags2 & 0x0001; | |
253 s->use_bit_reservoir = flags2 & 0x0002; | |
254 s->use_variable_block_len = flags2 & 0x0004; | |
255 | |
256 /* compute MDCT block size */ | |
257 if (s->sample_rate <= 16000) { | |
258 s->frame_len_bits = 9; | |
2967 | 259 } else if (s->sample_rate <= 22050 || |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
260 (s->sample_rate <= 32000 && s->version == 1)) { |
783 | 261 s->frame_len_bits = 10; |
262 } else { | |
263 s->frame_len_bits = 11; | |
264 } | |
265 s->frame_len = 1 << s->frame_len_bits; | |
266 if (s->use_variable_block_len) { | |
808
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
267 int nb_max, nb; |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
268 nb = ((flags2 >> 3) & 3) + 1; |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
269 if ((s->bit_rate / s->nb_channels) >= 32000) |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
270 nb += 2; |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
271 nb_max = s->frame_len_bits - BLOCK_MIN_BITS; |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
272 if (nb > nb_max) |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
273 nb = nb_max; |
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
274 s->nb_block_sizes = nb + 1; |
783 | 275 } else { |
276 s->nb_block_sizes = 1; | |
277 } | |
278 | |
279 /* init rate dependant parameters */ | |
280 s->use_noise_coding = 1; | |
3235 | 281 high_freq = s->sample_rate * 0.5; |
783 | 282 |
283 /* if version 2, then the rates are normalized */ | |
284 sample_rate1 = s->sample_rate; | |
285 if (s->version == 2) { | |
2967 | 286 if (sample_rate1 >= 44100) |
783 | 287 sample_rate1 = 44100; |
2967 | 288 else if (sample_rate1 >= 22050) |
783 | 289 sample_rate1 = 22050; |
2967 | 290 else if (sample_rate1 >= 16000) |
783 | 291 sample_rate1 = 16000; |
2967 | 292 else if (sample_rate1 >= 11025) |
783 | 293 sample_rate1 = 11025; |
2967 | 294 else if (sample_rate1 >= 8000) |
783 | 295 sample_rate1 = 8000; |
296 } | |
297 | |
298 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate); | |
2992 | 299 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0 + 0.5)) + 2; |
783 | 300 |
301 /* compute high frequency value and choose if noise coding should | |
302 be activated */ | |
303 bps1 = bps; | |
304 if (s->nb_channels == 2) | |
305 bps1 = bps * 1.6; | |
306 if (sample_rate1 == 44100) { | |
307 if (bps1 >= 0.61) | |
308 s->use_noise_coding = 0; | |
309 else | |
3235 | 310 high_freq = high_freq * 0.4; |
783 | 311 } else if (sample_rate1 == 22050) { |
312 if (bps1 >= 1.16) | |
313 s->use_noise_coding = 0; | |
2967 | 314 else if (bps1 >= 0.72) |
3235 | 315 high_freq = high_freq * 0.7; |
783 | 316 else |
3235 | 317 high_freq = high_freq * 0.6; |
783 | 318 } else if (sample_rate1 == 16000) { |
319 if (bps > 0.5) | |
3235 | 320 high_freq = high_freq * 0.5; |
783 | 321 else |
3235 | 322 high_freq = high_freq * 0.3; |
783 | 323 } else if (sample_rate1 == 11025) { |
3235 | 324 high_freq = high_freq * 0.7; |
783 | 325 } else if (sample_rate1 == 8000) { |
326 if (bps <= 0.625) { | |
3235 | 327 high_freq = high_freq * 0.5; |
783 | 328 } else if (bps > 0.75) { |
329 s->use_noise_coding = 0; | |
330 } else { | |
3235 | 331 high_freq = high_freq * 0.65; |
783 | 332 } |
333 } else { | |
334 if (bps >= 0.8) { | |
3235 | 335 high_freq = high_freq * 0.75; |
783 | 336 } else if (bps >= 0.6) { |
3235 | 337 high_freq = high_freq * 0.6; |
783 | 338 } else { |
3235 | 339 high_freq = high_freq * 0.5; |
783 | 340 } |
341 } | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
342 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
343 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n", |
2967 | 344 s->version, s->nb_channels, s->sample_rate, s->bit_rate, |
783 | 345 s->block_align); |
3235 | 346 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n", |
347 bps, bps1, high_freq, s->byte_offset_bits); | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
348 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n", |
808
e9bfaabcf07d
fixed nb_block_sizes detection - fixed codec_id test (avctx->codec_id does not need to be initialized)
bellard
parents:
797
diff
changeset
|
349 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes); |
783 | 350 |
351 /* compute the scale factor band sizes for each MDCT block size */ | |
352 { | |
353 int a, b, pos, lpos, k, block_len, i, j, n; | |
354 const uint8_t *table; | |
2967 | 355 |
783 | 356 if (s->version == 1) { |
357 s->coefs_start = 3; | |
358 } else { | |
359 s->coefs_start = 0; | |
360 } | |
361 for(k = 0; k < s->nb_block_sizes; k++) { | |
362 block_len = s->frame_len >> k; | |
363 | |
364 if (s->version == 1) { | |
365 lpos = 0; | |
366 for(i=0;i<25;i++) { | |
367 a = wma_critical_freqs[i]; | |
368 b = s->sample_rate; | |
369 pos = ((block_len * 2 * a) + (b >> 1)) / b; | |
2967 | 370 if (pos > block_len) |
783 | 371 pos = block_len; |
372 s->exponent_bands[0][i] = pos - lpos; | |
373 if (pos >= block_len) { | |
374 i++; | |
375 break; | |
376 } | |
377 lpos = pos; | |
378 } | |
379 s->exponent_sizes[0] = i; | |
380 } else { | |
381 /* hardcoded tables */ | |
382 table = NULL; | |
383 a = s->frame_len_bits - BLOCK_MIN_BITS - k; | |
384 if (a < 3) { | |
385 if (s->sample_rate >= 44100) | |
386 table = exponent_band_44100[a]; | |
387 else if (s->sample_rate >= 32000) | |
388 table = exponent_band_32000[a]; | |
389 else if (s->sample_rate >= 22050) | |
390 table = exponent_band_22050[a]; | |
391 } | |
392 if (table) { | |
393 n = *table++; | |
394 for(i=0;i<n;i++) | |
395 s->exponent_bands[k][i] = table[i]; | |
396 s->exponent_sizes[k] = n; | |
397 } else { | |
398 j = 0; | |
399 lpos = 0; | |
400 for(i=0;i<25;i++) { | |
401 a = wma_critical_freqs[i]; | |
402 b = s->sample_rate; | |
403 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b); | |
404 pos <<= 2; | |
2967 | 405 if (pos > block_len) |
783 | 406 pos = block_len; |
407 if (pos > lpos) | |
408 s->exponent_bands[k][j++] = pos - lpos; | |
409 if (pos >= block_len) | |
410 break; | |
411 lpos = pos; | |
412 } | |
413 s->exponent_sizes[k] = j; | |
414 } | |
415 } | |
416 | |
417 /* max number of coefs */ | |
418 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k; | |
419 /* high freq computation */ | |
3235 | 420 s->high_band_start[k] = (int)((block_len * 2 * high_freq) / |
421 s->sample_rate + 0.5); | |
783 | 422 n = s->exponent_sizes[k]; |
423 j = 0; | |
424 pos = 0; | |
425 for(i=0;i<n;i++) { | |
426 int start, end; | |
427 start = pos; | |
428 pos += s->exponent_bands[k][i]; | |
429 end = pos; | |
430 if (start < s->high_band_start[k]) | |
431 start = s->high_band_start[k]; | |
432 if (end > s->coefs_end[k]) | |
433 end = s->coefs_end[k]; | |
434 if (end > start) | |
435 s->exponent_high_bands[k][j++] = end - start; | |
436 } | |
437 s->exponent_high_sizes[k] = j; | |
438 #if 0 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
439 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ", |
2967 | 440 s->frame_len >> k, |
783 | 441 s->coefs_end[k], |
442 s->high_band_start[k], | |
443 s->exponent_high_sizes[k]); | |
444 for(j=0;j<s->exponent_high_sizes[k];j++) | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
445 tprintf(" %d", s->exponent_high_bands[k][j]); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
446 tprintf("\n"); |
783 | 447 #endif |
448 } | |
449 } | |
450 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
451 #ifdef TRACE |
783 | 452 { |
453 int i, j; | |
454 for(i = 0; i < s->nb_block_sizes; i++) { | |
2967 | 455 tprintf("%5d: n=%2d:", |
456 s->frame_len >> i, | |
783 | 457 s->exponent_sizes[i]); |
458 for(j=0;j<s->exponent_sizes[i];j++) | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
459 tprintf(" %d", s->exponent_bands[i][j]); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
460 tprintf("\n"); |
783 | 461 } |
462 } | |
463 #endif | |
464 | |
465 /* init MDCT */ | |
466 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
467 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1); |
2967 | 468 |
783 | 469 /* init MDCT windows : simple sinus window */ |
470 for(i = 0; i < s->nb_block_sizes; i++) { | |
471 int n, j; | |
472 float alpha; | |
473 n = 1 << (s->frame_len_bits - i); | |
474 window = av_malloc(sizeof(float) * n); | |
475 alpha = M_PI / (2.0 * n); | |
476 for(j=0;j<n;j++) { | |
477 window[n - j - 1] = sin((j + 0.5) * alpha); | |
478 } | |
479 s->windows[i] = window; | |
480 } | |
481 | |
482 s->reset_block_lengths = 1; | |
2967 | 483 |
783 | 484 if (s->use_noise_coding) { |
485 | |
486 /* init the noise generator */ | |
487 if (s->use_exp_vlc) | |
488 s->noise_mult = 0.02; | |
489 else | |
490 s->noise_mult = 0.04; | |
2967 | 491 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
492 #ifdef TRACE |
783 | 493 for(i=0;i<NOISE_TAB_SIZE;i++) |
494 s->noise_table[i] = 1.0 * s->noise_mult; | |
495 #else | |
496 { | |
497 unsigned int seed; | |
498 float norm; | |
499 seed = 1; | |
500 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult; | |
501 for(i=0;i<NOISE_TAB_SIZE;i++) { | |
502 seed = seed * 314159 + 1; | |
503 s->noise_table[i] = (float)((int)seed) * norm; | |
504 } | |
505 } | |
506 #endif | |
3113 | 507 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(hgain_huffbits), |
783 | 508 hgain_huffbits, 1, 1, |
2370
26560d4fdb1f
Memory leak fix patch by (Burkhard Plaum <plaum >at< ipf.uni-stuttgart )dot( de>)
michael
parents:
2180
diff
changeset
|
509 hgain_huffcodes, 2, 2, 0); |
783 | 510 } |
511 | |
512 if (s->use_exp_vlc) { | |
3113 | 513 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(scale_huffbits), |
783 | 514 scale_huffbits, 1, 1, |
2370
26560d4fdb1f
Memory leak fix patch by (Burkhard Plaum <plaum >at< ipf.uni-stuttgart )dot( de>)
michael
parents:
2180
diff
changeset
|
515 scale_huffcodes, 4, 4, 0); |
783 | 516 } else { |
517 wma_lsp_to_curve_init(s, s->frame_len); | |
518 } | |
519 | |
520 /* choose the VLC tables for the coefficients */ | |
521 coef_vlc_table = 2; | |
522 if (s->sample_rate >= 32000) { | |
523 if (bps1 < 0.72) | |
524 coef_vlc_table = 0; | |
525 else if (bps1 < 1.16) | |
526 coef_vlc_table = 1; | |
527 } | |
528 | |
529 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0], | |
530 &coef_vlcs[coef_vlc_table * 2]); | |
531 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1], | |
532 &coef_vlcs[coef_vlc_table * 2 + 1]); | |
533 return 0; | |
534 } | |
535 | |
536 /* interpolate values for a bigger or smaller block. The block must | |
537 have multiple sizes */ | |
538 static void interpolate_array(float *scale, int old_size, int new_size) | |
539 { | |
540 int i, j, jincr, k; | |
541 float v; | |
542 | |
543 if (new_size > old_size) { | |
544 jincr = new_size / old_size; | |
545 j = new_size; | |
546 for(i = old_size - 1; i >=0; i--) { | |
547 v = scale[i]; | |
548 k = jincr; | |
549 do { | |
550 scale[--j] = v; | |
551 } while (--k); | |
552 } | |
553 } else if (new_size < old_size) { | |
554 j = 0; | |
555 jincr = old_size / new_size; | |
556 for(i = 0; i < new_size; i++) { | |
557 scale[i] = scale[j]; | |
558 j += jincr; | |
559 } | |
560 } | |
561 } | |
562 | |
563 /* compute x^-0.25 with an exponent and mantissa table. We use linear | |
564 interpolation to reduce the mantissa table size at a small speed | |
565 expense (linear interpolation approximately doubles the number of | |
566 bits of precision). */ | |
567 static inline float pow_m1_4(WMADecodeContext *s, float x) | |
568 { | |
569 union { | |
570 float f; | |
571 unsigned int v; | |
572 } u, t; | |
573 unsigned int e, m; | |
574 float a, b; | |
575 | |
576 u.f = x; | |
577 e = u.v >> 23; | |
578 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1); | |
579 /* build interpolation scale: 1 <= t < 2. */ | |
580 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23); | |
581 a = s->lsp_pow_m_table1[m]; | |
582 b = s->lsp_pow_m_table2[m]; | |
583 return s->lsp_pow_e_table[e] * (a + b * t.f); | |
584 } | |
585 | |
586 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len) | |
2967 | 587 { |
783 | 588 float wdel, a, b; |
589 int i, e, m; | |
590 | |
591 wdel = M_PI / frame_len; | |
592 for(i=0;i<frame_len;i++) | |
593 s->lsp_cos_table[i] = 2.0f * cos(wdel * i); | |
594 | |
595 /* tables for x^-0.25 computation */ | |
596 for(i=0;i<256;i++) { | |
597 e = i - 126; | |
598 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25); | |
599 } | |
600 | |
601 /* NOTE: these two tables are needed to avoid two operations in | |
602 pow_m1_4 */ | |
603 b = 1.0; | |
604 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) { | |
605 m = (1 << LSP_POW_BITS) + i; | |
606 a = (float)m * (0.5 / (1 << LSP_POW_BITS)); | |
607 a = pow(a, -0.25); | |
608 s->lsp_pow_m_table1[i] = 2 * a - b; | |
609 s->lsp_pow_m_table2[i] = b - a; | |
610 b = a; | |
611 } | |
612 #if 0 | |
613 for(i=1;i<20;i++) { | |
614 float v, r1, r2; | |
615 v = 5.0 / i; | |
616 r1 = pow_m1_4(s, v); | |
617 r2 = pow(v,-0.25); | |
618 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1); | |
619 } | |
620 #endif | |
621 } | |
622 | |
623 /* NOTE: We use the same code as Vorbis here */ | |
624 /* XXX: optimize it further with SSE/3Dnow */ | |
2967 | 625 static void wma_lsp_to_curve(WMADecodeContext *s, |
626 float *out, float *val_max_ptr, | |
783 | 627 int n, float *lsp) |
628 { | |
629 int i, j; | |
630 float p, q, w, v, val_max; | |
631 | |
632 val_max = 0; | |
633 for(i=0;i<n;i++) { | |
634 p = 0.5f; | |
635 q = 0.5f; | |
636 w = s->lsp_cos_table[i]; | |
637 for(j=1;j<NB_LSP_COEFS;j+=2){ | |
638 q *= w - lsp[j - 1]; | |
639 p *= w - lsp[j]; | |
640 } | |
641 p *= p * (2.0f - w); | |
642 q *= q * (2.0f + w); | |
643 v = p + q; | |
644 v = pow_m1_4(s, v); | |
645 if (v > val_max) | |
646 val_max = v; | |
647 out[i] = v; | |
648 } | |
649 *val_max_ptr = val_max; | |
650 } | |
651 | |
652 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */ | |
653 static void decode_exp_lsp(WMADecodeContext *s, int ch) | |
654 { | |
655 float lsp_coefs[NB_LSP_COEFS]; | |
656 int val, i; | |
657 | |
658 for(i = 0; i < NB_LSP_COEFS; i++) { | |
659 if (i == 0 || i >= 8) | |
660 val = get_bits(&s->gb, 3); | |
661 else | |
662 val = get_bits(&s->gb, 4); | |
663 lsp_coefs[i] = lsp_codebook[i][val]; | |
664 } | |
665 | |
666 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch], | |
667 s->block_len, lsp_coefs); | |
668 } | |
669 | |
670 /* decode exponents coded with VLC codes */ | |
671 static int decode_exp_vlc(WMADecodeContext *s, int ch) | |
672 { | |
673 int last_exp, n, code; | |
674 const uint16_t *ptr, *band_ptr; | |
675 float v, *q, max_scale, *q_end; | |
2967 | 676 |
783 | 677 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits]; |
678 ptr = band_ptr; | |
679 q = s->exponents[ch]; | |
680 q_end = q + s->block_len; | |
681 max_scale = 0; | |
682 if (s->version == 1) { | |
683 last_exp = get_bits(&s->gb, 5) + 10; | |
3235 | 684 /* XXX: use a table */ |
685 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 686 max_scale = v; |
687 n = *ptr++; | |
688 do { | |
689 *q++ = v; | |
690 } while (--n); | |
691 } | |
692 last_exp = 36; | |
693 while (q < q_end) { | |
3113 | 694 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX); |
783 | 695 if (code < 0) |
696 return -1; | |
697 /* NOTE: this offset is the same as MPEG4 AAC ! */ | |
698 last_exp += code - 60; | |
3235 | 699 /* XXX: use a table */ |
700 v = pow(10, last_exp * (1.0 / 16.0)); | |
783 | 701 if (v > max_scale) |
702 max_scale = v; | |
703 n = *ptr++; | |
704 do { | |
705 *q++ = v; | |
706 } while (--n); | |
707 } | |
708 s->max_exponent[ch] = max_scale; | |
709 return 0; | |
710 } | |
711 | |
712 /* return 0 if OK. return 1 if last block of frame. return -1 if | |
713 unrecorrable error. */ | |
714 static int wma_decode_block(WMADecodeContext *s) | |
715 { | |
716 int n, v, a, ch, code, bsize; | |
717 int coef_nb_bits, total_gain, parse_exponents; | |
3606
e28285ddde8d
Alignment of the LT window, segfault fix patch by Baptiste Coudurier.
banan
parents:
3592
diff
changeset
|
718 DECLARE_ALIGNED_16(float, window[BLOCK_MAX_SIZE * 2]); |
2101 | 719 // XXX: FIXME!! there's a bug somewhere which makes this mandatory under altivec |
720 #ifdef HAVE_ALTIVEC | |
721 volatile int nb_coefs[MAX_CHANNELS] __attribute__((aligned(16))); | |
722 #else | |
783 | 723 int nb_coefs[MAX_CHANNELS]; |
2101 | 724 #endif |
783 | 725 float mdct_norm; |
726 | |
1343 | 727 #ifdef TRACE |
728 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num); | |
729 #endif | |
783 | 730 |
731 /* compute current block length */ | |
732 if (s->use_variable_block_len) { | |
733 n = av_log2(s->nb_block_sizes - 1) + 1; | |
2967 | 734 |
783 | 735 if (s->reset_block_lengths) { |
736 s->reset_block_lengths = 0; | |
737 v = get_bits(&s->gb, n); | |
738 if (v >= s->nb_block_sizes) | |
739 return -1; | |
740 s->prev_block_len_bits = s->frame_len_bits - v; | |
741 v = get_bits(&s->gb, n); | |
742 if (v >= s->nb_block_sizes) | |
743 return -1; | |
744 s->block_len_bits = s->frame_len_bits - v; | |
745 } else { | |
746 /* update block lengths */ | |
747 s->prev_block_len_bits = s->block_len_bits; | |
748 s->block_len_bits = s->next_block_len_bits; | |
749 } | |
750 v = get_bits(&s->gb, n); | |
751 if (v >= s->nb_block_sizes) | |
752 return -1; | |
753 s->next_block_len_bits = s->frame_len_bits - v; | |
754 } else { | |
755 /* fixed block len */ | |
756 s->next_block_len_bits = s->frame_len_bits; | |
757 s->prev_block_len_bits = s->frame_len_bits; | |
758 s->block_len_bits = s->frame_len_bits; | |
759 } | |
760 | |
761 /* now check if the block length is coherent with the frame length */ | |
762 s->block_len = 1 << s->block_len_bits; | |
763 if ((s->block_pos + s->block_len) > s->frame_len) | |
764 return -1; | |
765 | |
766 if (s->nb_channels == 2) { | |
767 s->ms_stereo = get_bits(&s->gb, 1); | |
768 } | |
769 v = 0; | |
770 for(ch = 0; ch < s->nb_channels; ch++) { | |
771 a = get_bits(&s->gb, 1); | |
772 s->channel_coded[ch] = a; | |
773 v |= a; | |
774 } | |
775 /* if no channel coded, no need to go further */ | |
776 /* XXX: fix potential framing problems */ | |
777 if (!v) | |
778 goto next; | |
779 | |
780 bsize = s->frame_len_bits - s->block_len_bits; | |
781 | |
782 /* read total gain and extract corresponding number of bits for | |
783 coef escape coding */ | |
784 total_gain = 1; | |
785 for(;;) { | |
786 a = get_bits(&s->gb, 7); | |
787 total_gain += a; | |
788 if (a != 127) | |
789 break; | |
790 } | |
2967 | 791 |
783 | 792 if (total_gain < 15) |
793 coef_nb_bits = 13; | |
794 else if (total_gain < 32) | |
795 coef_nb_bits = 12; | |
796 else if (total_gain < 40) | |
797 coef_nb_bits = 11; | |
798 else if (total_gain < 45) | |
799 coef_nb_bits = 10; | |
800 else | |
801 coef_nb_bits = 9; | |
802 | |
803 /* compute number of coefficients */ | |
804 n = s->coefs_end[bsize] - s->coefs_start; | |
805 for(ch = 0; ch < s->nb_channels; ch++) | |
806 nb_coefs[ch] = n; | |
807 | |
808 /* complex coding */ | |
809 if (s->use_noise_coding) { | |
810 | |
811 for(ch = 0; ch < s->nb_channels; ch++) { | |
812 if (s->channel_coded[ch]) { | |
813 int i, n, a; | |
814 n = s->exponent_high_sizes[bsize]; | |
815 for(i=0;i<n;i++) { | |
816 a = get_bits(&s->gb, 1); | |
817 s->high_band_coded[ch][i] = a; | |
818 /* if noise coding, the coefficients are not transmitted */ | |
819 if (a) | |
820 nb_coefs[ch] -= s->exponent_high_bands[bsize][i]; | |
821 } | |
822 } | |
823 } | |
824 for(ch = 0; ch < s->nb_channels; ch++) { | |
825 if (s->channel_coded[ch]) { | |
826 int i, n, val, code; | |
827 | |
828 n = s->exponent_high_sizes[bsize]; | |
829 val = (int)0x80000000; | |
830 for(i=0;i<n;i++) { | |
831 if (s->high_band_coded[ch][i]) { | |
832 if (val == (int)0x80000000) { | |
833 val = get_bits(&s->gb, 7) - 19; | |
834 } else { | |
3113 | 835 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX); |
783 | 836 if (code < 0) |
837 return -1; | |
838 val += code - 18; | |
839 } | |
840 s->high_band_values[ch][i] = val; | |
841 } | |
842 } | |
843 } | |
844 } | |
845 } | |
2967 | 846 |
783 | 847 /* exposant can be interpolated in short blocks. */ |
848 parse_exponents = 1; | |
849 if (s->block_len_bits != s->frame_len_bits) { | |
850 parse_exponents = get_bits(&s->gb, 1); | |
851 } | |
2967 | 852 |
783 | 853 if (parse_exponents) { |
854 for(ch = 0; ch < s->nb_channels; ch++) { | |
855 if (s->channel_coded[ch]) { | |
856 if (s->use_exp_vlc) { | |
857 if (decode_exp_vlc(s, ch) < 0) | |
858 return -1; | |
859 } else { | |
860 decode_exp_lsp(s, ch); | |
861 } | |
862 } | |
863 } | |
864 } else { | |
865 for(ch = 0; ch < s->nb_channels; ch++) { | |
866 if (s->channel_coded[ch]) { | |
2967 | 867 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits, |
783 | 868 s->block_len); |
869 } | |
870 } | |
871 } | |
872 | |
873 /* parse spectral coefficients : just RLE encoding */ | |
874 for(ch = 0; ch < s->nb_channels; ch++) { | |
875 if (s->channel_coded[ch]) { | |
876 VLC *coef_vlc; | |
877 int level, run, sign, tindex; | |
878 int16_t *ptr, *eptr; | |
879 const int16_t *level_table, *run_table; | |
880 | |
881 /* special VLC tables are used for ms stereo because | |
882 there is potentially less energy there */ | |
883 tindex = (ch == 1 && s->ms_stereo); | |
884 coef_vlc = &s->coef_vlc[tindex]; | |
885 run_table = s->run_table[tindex]; | |
886 level_table = s->level_table[tindex]; | |
887 /* XXX: optimize */ | |
888 ptr = &s->coefs1[ch][0]; | |
889 eptr = ptr + nb_coefs[ch]; | |
890 memset(ptr, 0, s->block_len * sizeof(int16_t)); | |
891 for(;;) { | |
3113 | 892 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX); |
783 | 893 if (code < 0) |
894 return -1; | |
895 if (code == 1) { | |
896 /* EOB */ | |
897 break; | |
898 } else if (code == 0) { | |
899 /* escape */ | |
900 level = get_bits(&s->gb, coef_nb_bits); | |
901 /* NOTE: this is rather suboptimal. reading | |
902 block_len_bits would be better */ | |
903 run = get_bits(&s->gb, s->frame_len_bits); | |
904 } else { | |
905 /* normal code */ | |
906 run = run_table[code]; | |
907 level = level_table[code]; | |
908 } | |
909 sign = get_bits(&s->gb, 1); | |
910 if (!sign) | |
911 level = -level; | |
912 ptr += run; | |
913 if (ptr >= eptr) | |
3361 | 914 { |
915 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n"); | |
916 break; | |
917 } | |
783 | 918 *ptr++ = level; |
919 /* NOTE: EOB can be omitted */ | |
920 if (ptr >= eptr) | |
921 break; | |
922 } | |
923 } | |
924 if (s->version == 1 && s->nb_channels >= 2) { | |
925 align_get_bits(&s->gb); | |
926 } | |
927 } | |
2967 | 928 |
783 | 929 /* normalize */ |
930 { | |
931 int n4 = s->block_len / 2; | |
932 mdct_norm = 1.0 / (float)n4; | |
933 if (s->version == 1) { | |
934 mdct_norm *= sqrt(n4); | |
935 } | |
936 } | |
937 | |
938 /* finally compute the MDCT coefficients */ | |
939 for(ch = 0; ch < s->nb_channels; ch++) { | |
940 if (s->channel_coded[ch]) { | |
941 int16_t *coefs1; | |
942 float *coefs, *exponents, mult, mult1, noise, *exp_ptr; | |
943 int i, j, n, n1, last_high_band; | |
944 float exp_power[HIGH_BAND_MAX_SIZE]; | |
945 | |
946 coefs1 = s->coefs1[ch]; | |
947 exponents = s->exponents[ch]; | |
3235 | 948 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch]; |
783 | 949 mult *= mdct_norm; |
950 coefs = s->coefs[ch]; | |
951 if (s->use_noise_coding) { | |
952 mult1 = mult; | |
953 /* very low freqs : noise */ | |
954 for(i = 0;i < s->coefs_start; i++) { | |
955 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1; | |
956 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
957 } | |
2967 | 958 |
783 | 959 n1 = s->exponent_high_sizes[bsize]; |
960 | |
961 /* compute power of high bands */ | |
2967 | 962 exp_ptr = exponents + |
963 s->high_band_start[bsize] - | |
783 | 964 s->coefs_start; |
965 last_high_band = 0; /* avoid warning */ | |
966 for(j=0;j<n1;j++) { | |
2967 | 967 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 968 s->block_len_bits][j]; |
969 if (s->high_band_coded[ch][j]) { | |
970 float e2, v; | |
971 e2 = 0; | |
972 for(i = 0;i < n; i++) { | |
973 v = exp_ptr[i]; | |
974 e2 += v * v; | |
975 } | |
976 exp_power[j] = e2 / n; | |
977 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
|
978 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n); |
783 | 979 } |
980 exp_ptr += n; | |
981 } | |
982 | |
983 /* main freqs and high freqs */ | |
984 for(j=-1;j<n1;j++) { | |
985 if (j < 0) { | |
2967 | 986 n = s->high_band_start[bsize] - |
783 | 987 s->coefs_start; |
988 } else { | |
2967 | 989 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 990 s->block_len_bits][j]; |
991 } | |
992 if (j >= 0 && s->high_band_coded[ch][j]) { | |
993 /* use noise with specified power */ | |
994 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]); | |
3235 | 995 /* XXX: use a table */ |
996 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05); | |
783 | 997 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult); |
998 mult1 *= mdct_norm; | |
999 for(i = 0;i < n; i++) { | |
1000 noise = s->noise_table[s->noise_index]; | |
1001 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
1002 *coefs++ = (*exponents++) * noise * mult1; | |
1003 } | |
1004 } else { | |
1005 /* coded values + small noise */ | |
1006 for(i = 0;i < n; i++) { | |
1007 noise = s->noise_table[s->noise_index]; | |
1008 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
1009 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult; | |
1010 } | |
1011 } | |
1012 } | |
1013 | |
1014 /* very high freqs : noise */ | |
1015 n = s->block_len - s->coefs_end[bsize]; | |
1016 mult1 = mult * exponents[-1]; | |
1017 for(i = 0; i < n; i++) { | |
1018 *coefs++ = s->noise_table[s->noise_index] * mult1; | |
1019 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
1020 } | |
1021 } else { | |
1022 /* XXX: optimize more */ | |
1023 for(i = 0;i < s->coefs_start; i++) | |
1024 *coefs++ = 0.0; | |
1025 n = nb_coefs[ch]; | |
1026 for(i = 0;i < n; i++) { | |
1027 *coefs++ = coefs1[i] * exponents[i] * mult; | |
1028 } | |
1029 n = s->block_len - s->coefs_end[bsize]; | |
1030 for(i = 0;i < n; i++) | |
1031 *coefs++ = 0.0; | |
1032 } | |
1033 } | |
1034 } | |
1035 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
1036 #ifdef TRACE |
783 | 1037 for(ch = 0; ch < s->nb_channels; ch++) { |
1038 if (s->channel_coded[ch]) { | |
1039 dump_floats("exponents", 3, s->exponents[ch], s->block_len); | |
1040 dump_floats("coefs", 1, s->coefs[ch], s->block_len); | |
1041 } | |
1042 } | |
1043 #endif | |
2967 | 1044 |
783 | 1045 if (s->ms_stereo && s->channel_coded[1]) { |
1046 float a, b; | |
1047 int i; | |
1048 | |
1049 /* nominal case for ms stereo: we do it before mdct */ | |
1050 /* no need to optimize this case because it should almost | |
1051 never happen */ | |
1052 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
|
1053 tprintf("rare ms-stereo case happened\n"); |
783 | 1054 memset(s->coefs[0], 0, sizeof(float) * s->block_len); |
1055 s->channel_coded[0] = 1; | |
1056 } | |
2967 | 1057 |
783 | 1058 for(i = 0; i < s->block_len; i++) { |
1059 a = s->coefs[0][i]; | |
1060 b = s->coefs[1][i]; | |
1061 s->coefs[0][i] = a + b; | |
1062 s->coefs[1][i] = a - b; | |
1063 } | |
1064 } | |
1065 | |
1066 /* build the window : we ensure that when the windows overlap | |
1067 their squared sum is always 1 (MDCT reconstruction rule) */ | |
1068 /* XXX: merge with output */ | |
1069 { | |
1070 int i, next_block_len, block_len, prev_block_len, n; | |
1071 float *wptr; | |
1072 | |
1073 block_len = s->block_len; | |
1074 prev_block_len = 1 << s->prev_block_len_bits; | |
1075 next_block_len = 1 << s->next_block_len_bits; | |
1076 | |
1077 /* right part */ | |
1078 wptr = window + block_len; | |
1079 if (block_len <= next_block_len) { | |
1080 for(i=0;i<block_len;i++) | |
1081 *wptr++ = s->windows[bsize][i]; | |
1082 } else { | |
1083 /* overlap */ | |
1084 n = (block_len / 2) - (next_block_len / 2); | |
1085 for(i=0;i<n;i++) | |
1086 *wptr++ = 1.0; | |
1087 for(i=0;i<next_block_len;i++) | |
1088 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i]; | |
1089 for(i=0;i<n;i++) | |
1090 *wptr++ = 0.0; | |
1091 } | |
1092 | |
1093 /* left part */ | |
1094 wptr = window + block_len; | |
1095 if (block_len <= prev_block_len) { | |
1096 for(i=0;i<block_len;i++) | |
1097 *--wptr = s->windows[bsize][i]; | |
1098 } else { | |
1099 /* overlap */ | |
1100 n = (block_len / 2) - (prev_block_len / 2); | |
1101 for(i=0;i<n;i++) | |
1102 *--wptr = 1.0; | |
1103 for(i=0;i<prev_block_len;i++) | |
1104 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i]; | |
1105 for(i=0;i<n;i++) | |
1106 *--wptr = 0.0; | |
1107 } | |
1108 } | |
1109 | |
2967 | 1110 |
783 | 1111 for(ch = 0; ch < s->nb_channels; ch++) { |
1112 if (s->channel_coded[ch]) { | |
3089 | 1113 DECLARE_ALIGNED_16(FFTSample, output[BLOCK_MAX_SIZE * 2]); |
783 | 1114 float *ptr; |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
1115 int n4, index, n; |
783 | 1116 |
1117 n = s->block_len; | |
1118 n4 = s->block_len / 2; | |
3555 | 1119 s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize], |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
1120 output, s->coefs[ch], s->mdct_tmp); |
783 | 1121 |
1122 /* XXX: optimize all that by build the window and | |
1123 multipying/adding at the same time */ | |
1124 | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
1125 /* multiply by the window and add in the frame */ |
783 | 1126 index = (s->frame_len / 2) + s->block_pos - n4; |
1127 ptr = &s->frame_out[ch][index]; | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
1128 s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1); |
783 | 1129 |
1130 /* specific fast case for ms-stereo : add to second | |
1131 channel if it is not coded */ | |
1132 if (s->ms_stereo && !s->channel_coded[1]) { | |
1133 ptr = &s->frame_out[1][index]; | |
3592
6a358dccf2ab
SIMD vector optimizations. 3% faster overall decoding.
banan
parents:
3555
diff
changeset
|
1134 s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1); |
783 | 1135 } |
1136 } | |
1137 } | |
1138 next: | |
1139 /* update block number */ | |
1140 s->block_num++; | |
1141 s->block_pos += s->block_len; | |
1142 if (s->block_pos >= s->frame_len) | |
1143 return 1; | |
1144 else | |
1145 return 0; | |
1146 } | |
1147 | |
1148 /* decode a frame of frame_len samples */ | |
1149 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples) | |
1150 { | |
1151 int ret, i, n, a, ch, incr; | |
1152 int16_t *ptr; | |
1153 float *iptr; | |
1154 | |
1343 | 1155 #ifdef TRACE |
1156 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len); | |
1157 #endif | |
783 | 1158 |
1159 /* read each block */ | |
1160 s->block_num = 0; | |
1161 s->block_pos = 0; | |
1162 for(;;) { | |
1163 ret = wma_decode_block(s); | |
2967 | 1164 if (ret < 0) |
783 | 1165 return -1; |
1166 if (ret) | |
1167 break; | |
1168 } | |
1169 | |
1170 /* convert frame to integer */ | |
1171 n = s->frame_len; | |
1172 incr = s->nb_channels; | |
1173 for(ch = 0; ch < s->nb_channels; ch++) { | |
1174 ptr = samples + ch; | |
1175 iptr = s->frame_out[ch]; | |
1176 | |
1177 for(i=0;i<n;i++) { | |
797 | 1178 a = lrintf(*iptr++); |
783 | 1179 if (a > 32767) |
1180 a = 32767; | |
1181 else if (a < -32768) | |
1182 a = -32768; | |
1183 *ptr = a; | |
1184 ptr += incr; | |
1185 } | |
1186 /* prepare for next block */ | |
1187 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len], | |
1188 s->frame_len * sizeof(float)); | |
1189 /* XXX: suppress this */ | |
2967 | 1190 memset(&s->frame_out[ch][s->frame_len], 0, |
783 | 1191 s->frame_len * sizeof(float)); |
1192 } | |
1193 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
1194 #ifdef TRACE |
783 | 1195 dump_shorts("samples", samples, n * s->nb_channels); |
1196 #endif | |
1197 return 0; | |
1198 } | |
1199 | |
2967 | 1200 static int wma_decode_superframe(AVCodecContext *avctx, |
783 | 1201 void *data, int *data_size, |
1064 | 1202 uint8_t *buf, int buf_size) |
783 | 1203 { |
1204 WMADecodeContext *s = avctx->priv_data; | |
1205 int nb_frames, bit_offset, i, pos, len; | |
1206 uint8_t *q; | |
1207 int16_t *samples; | |
2967 | 1208 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
1209 tprintf("***decode_superframe:\n"); |
783 | 1210 |
1750 | 1211 if(buf_size==0){ |
1212 s->last_superframe_len = 0; | |
1213 return 0; | |
1214 } | |
2967 | 1215 |
783 | 1216 samples = data; |
1217 | |
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
|
1218 init_get_bits(&s->gb, buf, buf_size*8); |
2967 | 1219 |
783 | 1220 if (s->use_bit_reservoir) { |
1221 /* read super frame header */ | |
1222 get_bits(&s->gb, 4); /* super frame index */ | |
1223 nb_frames = get_bits(&s->gb, 4) - 1; | |
1224 | |
1225 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3); | |
1226 | |
1227 if (s->last_superframe_len > 0) { | |
1228 // printf("skip=%d\n", s->last_bitoffset); | |
1229 /* add bit_offset bits to last frame */ | |
2967 | 1230 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) > |
783 | 1231 MAX_CODED_SUPERFRAME_SIZE) |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1232 goto fail; |
783 | 1233 q = s->last_superframe + s->last_superframe_len; |
1234 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
|
1235 while (len > 7) { |
783 | 1236 *q++ = (get_bits)(&s->gb, 8); |
1237 len -= 8; | |
1238 } | |
1239 if (len > 0) { | |
1240 *q++ = (get_bits)(&s->gb, len) << (8 - len); | |
1241 } | |
2967 | 1242 |
783 | 1243 /* 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
|
1244 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8); |
783 | 1245 /* skip unused bits */ |
1246 if (s->last_bitoffset > 0) | |
1247 skip_bits(&s->gb, s->last_bitoffset); | |
1248 /* this frame is stored in the last superframe and in the | |
1249 current one */ | |
1250 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1251 goto fail; |
783 | 1252 samples += s->nb_channels * s->frame_len; |
1253 } | |
1254 | |
1255 /* read each frame starting from bit_offset */ | |
1256 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
|
1257 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8); |
783 | 1258 len = pos & 7; |
1259 if (len > 0) | |
1260 skip_bits(&s->gb, len); | |
2967 | 1261 |
783 | 1262 s->reset_block_lengths = 1; |
1263 for(i=0;i<nb_frames;i++) { | |
1264 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1265 goto fail; |
783 | 1266 samples += s->nb_channels * s->frame_len; |
1267 } | |
1268 | |
1269 /* we copy the end of the frame in the last frame buffer */ | |
1270 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7); | |
1271 s->last_bitoffset = pos & 7; | |
1272 pos >>= 3; | |
1273 len = buf_size - pos; | |
819 | 1274 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
|
1275 goto fail; |
783 | 1276 } |
1277 s->last_superframe_len = len; | |
1278 memcpy(s->last_superframe, buf + pos, len); | |
1279 } else { | |
1280 /* single frame decode */ | |
1281 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1282 goto fail; |
783 | 1283 samples += s->nb_channels * s->frame_len; |
1284 } | |
1285 *data_size = (int8_t *)samples - (int8_t *)data; | |
1286 return s->block_align; | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1287 fail: |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1288 /* when error, we reset the bit reservoir */ |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1289 s->last_superframe_len = 0; |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1290 return -1; |
783 | 1291 } |
1292 | |
1293 static int wma_decode_end(AVCodecContext *avctx) | |
1294 { | |
1295 WMADecodeContext *s = avctx->priv_data; | |
1296 int i; | |
1297 | |
1298 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
1299 ff_mdct_end(&s->mdct_ctx[i]); |
783 | 1300 for(i = 0; i < s->nb_block_sizes; i++) |
1301 av_free(s->windows[i]); | |
1302 | |
1303 if (s->use_exp_vlc) { | |
1304 free_vlc(&s->exp_vlc); | |
1305 } | |
1306 if (s->use_noise_coding) { | |
1307 free_vlc(&s->hgain_vlc); | |
1308 } | |
1309 for(i = 0;i < 2; i++) { | |
1310 free_vlc(&s->coef_vlc[i]); | |
1311 av_free(s->run_table[i]); | |
1312 av_free(s->level_table[i]); | |
1313 } | |
2967 | 1314 |
783 | 1315 return 0; |
1316 } | |
1317 | |
1318 AVCodec wmav1_decoder = | |
1319 { | |
1320 "wmav1", | |
1321 CODEC_TYPE_AUDIO, | |
1322 CODEC_ID_WMAV1, | |
1323 sizeof(WMADecodeContext), | |
1324 wma_decode_init, | |
1325 NULL, | |
1326 wma_decode_end, | |
1327 wma_decode_superframe, | |
1328 }; | |
1329 | |
1330 AVCodec wmav2_decoder = | |
1331 { | |
1332 "wmav2", | |
1333 CODEC_TYPE_AUDIO, | |
1334 CODEC_ID_WMAV2, | |
1335 sizeof(WMADecodeContext), | |
1336 wma_decode_init, | |
1337 NULL, | |
1338 wma_decode_end, | |
1339 wma_decode_superframe, | |
1340 }; |