Mercurial > libavcodec.hg
annotate wmadec.c @ 3227:ea0ca798d389 libavcodec
avoid i++ in decode_block()
author | michael |
---|---|
date | Wed, 29 Mar 2006 09:13:43 +0000 |
parents | 10cda832bd0f |
children | d3c05c7dabcd |
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)]; | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
133 /* pow tables */ |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
134 float pow_005_10[121]; |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
135 float pow_00625_10[121]; |
1343 | 136 |
137 #ifdef TRACE | |
138 int frame_count; | |
139 #endif | |
783 | 140 } WMADecodeContext; |
141 | |
142 typedef struct CoefVLCTable { | |
143 int n; /* total number of codes */ | |
144 const uint32_t *huffcodes; /* VLC bit values */ | |
145 const uint8_t *huffbits; /* VLC bit size */ | |
146 const uint16_t *levels; /* table to build run/level tables */ | |
147 } CoefVLCTable; | |
148 | |
149 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len); | |
150 | |
151 #include "wmadata.h" | |
152 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
153 #ifdef TRACE |
783 | 154 static void dump_shorts(const char *name, const short *tab, int n) |
155 { | |
156 int i; | |
157 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
158 tprintf("%s[%d]:\n", name, n); |
783 | 159 for(i=0;i<n;i++) { |
160 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
|
161 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
162 tprintf(" %5d.0", tab[i]); |
783 | 163 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
|
164 tprintf("\n"); |
783 | 165 } |
166 } | |
167 | |
168 static void dump_floats(const char *name, int prec, const float *tab, int n) | |
169 { | |
170 int i; | |
171 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
172 tprintf("%s[%d]:\n", name, n); |
783 | 173 for(i=0;i<n;i++) { |
174 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
|
175 tprintf("%4d: ", i); |
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
176 tprintf(" %8.*f", prec, tab[i]); |
783 | 177 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
|
178 tprintf("\n"); |
783 | 179 } |
180 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
|
181 tprintf("\n"); |
783 | 182 } |
183 #endif | |
184 | |
185 /* XXX: use same run/length optimization as mpeg decoders */ | |
2967 | 186 static void init_coef_vlc(VLC *vlc, |
783 | 187 uint16_t **prun_table, uint16_t **plevel_table, |
188 const CoefVLCTable *vlc_table) | |
189 { | |
190 int n = vlc_table->n; | |
191 const uint8_t *table_bits = vlc_table->huffbits; | |
192 const uint32_t *table_codes = vlc_table->huffcodes; | |
193 const uint16_t *levels_table = vlc_table->levels; | |
194 uint16_t *run_table, *level_table; | |
195 const uint16_t *p; | |
196 int i, l, j, level; | |
197 | |
3113 | 198 init_vlc(vlc, VLCBITS, n, table_bits, 1, 1, table_codes, 4, 4, 0); |
783 | 199 |
1031
19de1445beb2
use av_malloc() functions - added av_strdup and av_realloc()
bellard
parents:
1025
diff
changeset
|
200 run_table = av_malloc(n * sizeof(uint16_t)); |
19de1445beb2
use av_malloc() functions - added av_strdup and av_realloc()
bellard
parents:
1025
diff
changeset
|
201 level_table = av_malloc(n * sizeof(uint16_t)); |
783 | 202 p = levels_table; |
203 i = 2; | |
204 level = 1; | |
205 while (i < n) { | |
206 l = *p++; | |
207 for(j=0;j<l;j++) { | |
208 run_table[i] = j; | |
209 level_table[i] = level; | |
210 i++; | |
211 } | |
212 level++; | |
213 } | |
214 *prun_table = run_table; | |
215 *plevel_table = level_table; | |
216 } | |
217 | |
218 static int wma_decode_init(AVCodecContext * avctx) | |
219 { | |
220 WMADecodeContext *s = avctx->priv_data; | |
221 int i, flags1, flags2; | |
222 float *window; | |
223 uint8_t *extradata; | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
224 float bps, bps1; |
3195
10cda832bd0f
fix coverity warning CID: 255 (uninitalized variable used to build tables which arent used, well there is a slight change of a FPU exception maybe ...)
michael
parents:
3176
diff
changeset
|
225 volatile float high_freq_factor= 0; //initial value should not matter as the tables build from this are unused if !use_noise_coding |
783 | 226 int sample_rate1; |
227 int coef_vlc_table; | |
2967 | 228 |
783 | 229 s->sample_rate = avctx->sample_rate; |
230 s->nb_channels = avctx->channels; | |
231 s->bit_rate = avctx->bit_rate; | |
232 s->block_align = avctx->block_align; | |
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; | |
281 | |
282 /* if version 2, then the rates are normalized */ | |
283 sample_rate1 = s->sample_rate; | |
284 if (s->version == 2) { | |
2967 | 285 if (sample_rate1 >= 44100) |
783 | 286 sample_rate1 = 44100; |
2967 | 287 else if (sample_rate1 >= 22050) |
783 | 288 sample_rate1 = 22050; |
2967 | 289 else if (sample_rate1 >= 16000) |
783 | 290 sample_rate1 = 16000; |
2967 | 291 else if (sample_rate1 >= 11025) |
783 | 292 sample_rate1 = 11025; |
2967 | 293 else if (sample_rate1 >= 8000) |
783 | 294 sample_rate1 = 8000; |
295 } | |
296 | |
297 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate); | |
2992 | 298 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0 + 0.5)) + 2; |
783 | 299 |
300 /* compute high frequency value and choose if noise coding should | |
301 be activated */ | |
302 bps1 = bps; | |
303 if (s->nb_channels == 2) | |
304 bps1 = bps * 1.6; | |
305 if (sample_rate1 == 44100) { | |
306 if (bps1 >= 0.61) | |
307 s->use_noise_coding = 0; | |
308 else | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
309 high_freq_factor = 0.4; |
783 | 310 } else if (sample_rate1 == 22050) { |
311 if (bps1 >= 1.16) | |
312 s->use_noise_coding = 0; | |
2967 | 313 else if (bps1 >= 0.72) |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
314 high_freq_factor = 0.7; |
783 | 315 else |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
316 high_freq_factor = 0.6; |
783 | 317 } else if (sample_rate1 == 16000) { |
318 if (bps > 0.5) | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
319 high_freq_factor = 0.5; |
783 | 320 else |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
321 high_freq_factor = 0.3; |
783 | 322 } else if (sample_rate1 == 11025) { |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
323 high_freq_factor = 0.7; |
783 | 324 } else if (sample_rate1 == 8000) { |
325 if (bps <= 0.625) { | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
326 high_freq_factor = 0.5; |
783 | 327 } else if (bps > 0.75) { |
328 s->use_noise_coding = 0; | |
329 } else { | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
330 high_freq_factor = 0.65; |
783 | 331 } |
332 } else { | |
333 if (bps >= 0.8) { | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
334 high_freq_factor = 0.75; |
783 | 335 } else if (bps >= 0.6) { |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
336 high_freq_factor = 0.6; |
783 | 337 } else { |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
338 high_freq_factor = 0.5; |
783 | 339 } |
340 } | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
341 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
|
342 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n", |
2967 | 343 s->version, s->nb_channels, s->sample_rate, s->bit_rate, |
783 | 344 s->block_align); |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
345 dprintf("bps=%f bps1=%f bitoffset=%d\n", |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
346 bps, bps1, 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
|
347 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
|
348 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes); |
783 | 349 |
350 /* compute the scale factor band sizes for each MDCT block size */ | |
351 { | |
352 int a, b, pos, lpos, k, block_len, i, j, n; | |
353 const uint8_t *table; | |
2967 | 354 |
783 | 355 if (s->version == 1) { |
356 s->coefs_start = 3; | |
357 } else { | |
358 s->coefs_start = 0; | |
359 } | |
360 for(k = 0; k < s->nb_block_sizes; k++) { | |
361 block_len = s->frame_len >> k; | |
362 | |
363 if (s->version == 1) { | |
364 lpos = 0; | |
365 for(i=0;i<25;i++) { | |
366 a = wma_critical_freqs[i]; | |
367 b = s->sample_rate; | |
368 pos = ((block_len * 2 * a) + (b >> 1)) / b; | |
2967 | 369 if (pos > block_len) |
783 | 370 pos = block_len; |
371 s->exponent_bands[0][i] = pos - lpos; | |
372 if (pos >= block_len) { | |
373 i++; | |
374 break; | |
375 } | |
376 lpos = pos; | |
377 } | |
378 s->exponent_sizes[0] = i; | |
379 } else { | |
380 /* hardcoded tables */ | |
381 table = NULL; | |
382 a = s->frame_len_bits - BLOCK_MIN_BITS - k; | |
383 if (a < 3) { | |
384 if (s->sample_rate >= 44100) | |
385 table = exponent_band_44100[a]; | |
386 else if (s->sample_rate >= 32000) | |
387 table = exponent_band_32000[a]; | |
388 else if (s->sample_rate >= 22050) | |
389 table = exponent_band_22050[a]; | |
390 } | |
391 if (table) { | |
392 n = *table++; | |
393 for(i=0;i<n;i++) | |
394 s->exponent_bands[k][i] = table[i]; | |
395 s->exponent_sizes[k] = n; | |
396 } else { | |
397 j = 0; | |
398 lpos = 0; | |
399 for(i=0;i<25;i++) { | |
400 a = wma_critical_freqs[i]; | |
401 b = s->sample_rate; | |
402 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b); | |
403 pos <<= 2; | |
2967 | 404 if (pos > block_len) |
783 | 405 pos = block_len; |
406 if (pos > lpos) | |
407 s->exponent_bands[k][j++] = pos - lpos; | |
408 if (pos >= block_len) | |
409 break; | |
410 lpos = pos; | |
411 } | |
412 s->exponent_sizes[k] = j; | |
413 } | |
414 } | |
415 | |
416 /* max number of coefs */ | |
417 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k; | |
418 /* high freq computation */ | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
419 s->high_band_start[k] = (int)((block_len * high_freq_factor) + 0.5); |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
420 |
783 | 421 n = s->exponent_sizes[k]; |
422 j = 0; | |
423 pos = 0; | |
424 for(i=0;i<n;i++) { | |
425 int start, end; | |
426 start = pos; | |
427 pos += s->exponent_bands[k][i]; | |
428 end = pos; | |
429 if (start < s->high_band_start[k]) | |
430 start = s->high_band_start[k]; | |
431 if (end > s->coefs_end[k]) | |
432 end = s->coefs_end[k]; | |
433 if (end > start) | |
434 s->exponent_high_bands[k][j++] = end - start; | |
435 } | |
436 s->exponent_high_sizes[k] = j; | |
437 #if 0 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
438 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ", |
2967 | 439 s->frame_len >> k, |
783 | 440 s->coefs_end[k], |
441 s->high_band_start[k], | |
442 s->exponent_high_sizes[k]); | |
443 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
|
444 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
|
445 tprintf("\n"); |
783 | 446 #endif |
447 } | |
448 } | |
449 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
450 #ifdef TRACE |
783 | 451 { |
452 int i, j; | |
453 for(i = 0; i < s->nb_block_sizes; i++) { | |
2967 | 454 tprintf("%5d: n=%2d:", |
455 s->frame_len >> i, | |
783 | 456 s->exponent_sizes[i]); |
457 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
|
458 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
|
459 tprintf("\n"); |
783 | 460 } |
461 } | |
462 #endif | |
463 | |
464 /* init MDCT */ | |
465 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
466 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1); |
2967 | 467 |
783 | 468 /* init MDCT windows : simple sinus window */ |
469 for(i = 0; i < s->nb_block_sizes; i++) { | |
470 int n, j; | |
471 float alpha; | |
472 n = 1 << (s->frame_len_bits - i); | |
473 window = av_malloc(sizeof(float) * n); | |
474 alpha = M_PI / (2.0 * n); | |
475 for(j=0;j<n;j++) { | |
476 window[n - j - 1] = sin((j + 0.5) * alpha); | |
477 } | |
478 s->windows[i] = window; | |
479 } | |
480 | |
481 s->reset_block_lengths = 1; | |
2967 | 482 |
783 | 483 if (s->use_noise_coding) { |
484 | |
485 /* init the noise generator */ | |
486 if (s->use_exp_vlc) | |
487 s->noise_mult = 0.02; | |
488 else | |
489 s->noise_mult = 0.04; | |
2967 | 490 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
491 #ifdef TRACE |
783 | 492 for(i=0;i<NOISE_TAB_SIZE;i++) |
493 s->noise_table[i] = 1.0 * s->noise_mult; | |
494 #else | |
495 { | |
496 unsigned int seed; | |
497 float norm; | |
498 seed = 1; | |
499 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult; | |
500 for(i=0;i<NOISE_TAB_SIZE;i++) { | |
501 seed = seed * 314159 + 1; | |
502 s->noise_table[i] = (float)((int)seed) * norm; | |
503 } | |
504 } | |
505 #endif | |
3113 | 506 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(hgain_huffbits), |
783 | 507 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
|
508 hgain_huffcodes, 2, 2, 0); |
783 | 509 } |
510 | |
511 if (s->use_exp_vlc) { | |
3113 | 512 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(scale_huffbits), |
783 | 513 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
|
514 scale_huffcodes, 4, 4, 0); |
783 | 515 } else { |
516 wma_lsp_to_curve_init(s, s->frame_len); | |
517 } | |
518 | |
519 /* choose the VLC tables for the coefficients */ | |
520 coef_vlc_table = 2; | |
521 if (s->sample_rate >= 32000) { | |
522 if (bps1 < 0.72) | |
523 coef_vlc_table = 0; | |
524 else if (bps1 < 1.16) | |
525 coef_vlc_table = 1; | |
526 } | |
527 | |
528 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0], | |
529 &coef_vlcs[coef_vlc_table * 2]); | |
530 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1], | |
531 &coef_vlcs[coef_vlc_table * 2 + 1]); | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
532 |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
533 /* init pow tables */ |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
534 for (i=0 ; i<121 ; i++) { |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
535 s->pow_005_10[i] = pow(10, i * 0.05); |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
536 s->pow_00625_10[i] = pow(10, i * (1.0 / 16.0)); |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
537 } |
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
538 |
783 | 539 return 0; |
540 } | |
541 | |
542 /* interpolate values for a bigger or smaller block. The block must | |
543 have multiple sizes */ | |
544 static void interpolate_array(float *scale, int old_size, int new_size) | |
545 { | |
546 int i, j, jincr, k; | |
547 float v; | |
548 | |
549 if (new_size > old_size) { | |
550 jincr = new_size / old_size; | |
551 j = new_size; | |
552 for(i = old_size - 1; i >=0; i--) { | |
553 v = scale[i]; | |
554 k = jincr; | |
555 do { | |
556 scale[--j] = v; | |
557 } while (--k); | |
558 } | |
559 } else if (new_size < old_size) { | |
560 j = 0; | |
561 jincr = old_size / new_size; | |
562 for(i = 0; i < new_size; i++) { | |
563 scale[i] = scale[j]; | |
564 j += jincr; | |
565 } | |
566 } | |
567 } | |
568 | |
569 /* compute x^-0.25 with an exponent and mantissa table. We use linear | |
570 interpolation to reduce the mantissa table size at a small speed | |
571 expense (linear interpolation approximately doubles the number of | |
572 bits of precision). */ | |
573 static inline float pow_m1_4(WMADecodeContext *s, float x) | |
574 { | |
575 union { | |
576 float f; | |
577 unsigned int v; | |
578 } u, t; | |
579 unsigned int e, m; | |
580 float a, b; | |
581 | |
582 u.f = x; | |
583 e = u.v >> 23; | |
584 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1); | |
585 /* build interpolation scale: 1 <= t < 2. */ | |
586 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23); | |
587 a = s->lsp_pow_m_table1[m]; | |
588 b = s->lsp_pow_m_table2[m]; | |
589 return s->lsp_pow_e_table[e] * (a + b * t.f); | |
590 } | |
591 | |
592 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len) | |
2967 | 593 { |
783 | 594 float wdel, a, b; |
595 int i, e, m; | |
596 | |
597 wdel = M_PI / frame_len; | |
598 for(i=0;i<frame_len;i++) | |
599 s->lsp_cos_table[i] = 2.0f * cos(wdel * i); | |
600 | |
601 /* tables for x^-0.25 computation */ | |
602 for(i=0;i<256;i++) { | |
603 e = i - 126; | |
604 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25); | |
605 } | |
606 | |
607 /* NOTE: these two tables are needed to avoid two operations in | |
608 pow_m1_4 */ | |
609 b = 1.0; | |
610 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) { | |
611 m = (1 << LSP_POW_BITS) + i; | |
612 a = (float)m * (0.5 / (1 << LSP_POW_BITS)); | |
613 a = pow(a, -0.25); | |
614 s->lsp_pow_m_table1[i] = 2 * a - b; | |
615 s->lsp_pow_m_table2[i] = b - a; | |
616 b = a; | |
617 } | |
618 #if 0 | |
619 for(i=1;i<20;i++) { | |
620 float v, r1, r2; | |
621 v = 5.0 / i; | |
622 r1 = pow_m1_4(s, v); | |
623 r2 = pow(v,-0.25); | |
624 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1); | |
625 } | |
626 #endif | |
627 } | |
628 | |
629 /* NOTE: We use the same code as Vorbis here */ | |
630 /* XXX: optimize it further with SSE/3Dnow */ | |
2967 | 631 static void wma_lsp_to_curve(WMADecodeContext *s, |
632 float *out, float *val_max_ptr, | |
783 | 633 int n, float *lsp) |
634 { | |
635 int i, j; | |
636 float p, q, w, v, val_max; | |
637 | |
638 val_max = 0; | |
639 for(i=0;i<n;i++) { | |
640 p = 0.5f; | |
641 q = 0.5f; | |
642 w = s->lsp_cos_table[i]; | |
643 for(j=1;j<NB_LSP_COEFS;j+=2){ | |
644 q *= w - lsp[j - 1]; | |
645 p *= w - lsp[j]; | |
646 } | |
647 p *= p * (2.0f - w); | |
648 q *= q * (2.0f + w); | |
649 v = p + q; | |
650 v = pow_m1_4(s, v); | |
651 if (v > val_max) | |
652 val_max = v; | |
653 out[i] = v; | |
654 } | |
655 *val_max_ptr = val_max; | |
656 } | |
657 | |
658 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */ | |
659 static void decode_exp_lsp(WMADecodeContext *s, int ch) | |
660 { | |
661 float lsp_coefs[NB_LSP_COEFS]; | |
662 int val, i; | |
663 | |
664 for(i = 0; i < NB_LSP_COEFS; i++) { | |
665 if (i == 0 || i >= 8) | |
666 val = get_bits(&s->gb, 3); | |
667 else | |
668 val = get_bits(&s->gb, 4); | |
669 lsp_coefs[i] = lsp_codebook[i][val]; | |
670 } | |
671 | |
672 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch], | |
673 s->block_len, lsp_coefs); | |
674 } | |
675 | |
676 /* decode exponents coded with VLC codes */ | |
677 static int decode_exp_vlc(WMADecodeContext *s, int ch) | |
678 { | |
679 int last_exp, n, code; | |
680 const uint16_t *ptr, *band_ptr; | |
681 float v, *q, max_scale, *q_end; | |
2967 | 682 |
783 | 683 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits]; |
684 ptr = band_ptr; | |
685 q = s->exponents[ch]; | |
686 q_end = q + s->block_len; | |
687 max_scale = 0; | |
688 if (s->version == 1) { | |
689 last_exp = get_bits(&s->gb, 5) + 10; | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
690 v = s->pow_00625_10[last_exp]; |
783 | 691 max_scale = v; |
692 n = *ptr++; | |
693 do { | |
694 *q++ = v; | |
695 } while (--n); | |
696 } | |
697 last_exp = 36; | |
698 while (q < q_end) { | |
3113 | 699 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX); |
783 | 700 if (code < 0) |
701 return -1; | |
702 /* NOTE: this offset is the same as MPEG4 AAC ! */ | |
703 last_exp += code - 60; | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
704 v = s->pow_00625_10[last_exp]; |
783 | 705 if (v > max_scale) |
706 max_scale = v; | |
707 n = *ptr++; | |
708 do { | |
709 *q++ = v; | |
710 } while (--n); | |
711 } | |
712 s->max_exponent[ch] = max_scale; | |
713 return 0; | |
714 } | |
715 | |
716 /* return 0 if OK. return 1 if last block of frame. return -1 if | |
717 unrecorrable error. */ | |
718 static int wma_decode_block(WMADecodeContext *s) | |
719 { | |
720 int n, v, a, ch, code, bsize; | |
721 int coef_nb_bits, total_gain, parse_exponents; | |
722 float window[BLOCK_MAX_SIZE * 2]; | |
2101 | 723 // XXX: FIXME!! there's a bug somewhere which makes this mandatory under altivec |
724 #ifdef HAVE_ALTIVEC | |
725 volatile int nb_coefs[MAX_CHANNELS] __attribute__((aligned(16))); | |
726 #else | |
783 | 727 int nb_coefs[MAX_CHANNELS]; |
2101 | 728 #endif |
783 | 729 float mdct_norm; |
730 | |
1343 | 731 #ifdef TRACE |
732 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num); | |
733 #endif | |
783 | 734 |
735 /* compute current block length */ | |
736 if (s->use_variable_block_len) { | |
737 n = av_log2(s->nb_block_sizes - 1) + 1; | |
2967 | 738 |
783 | 739 if (s->reset_block_lengths) { |
740 s->reset_block_lengths = 0; | |
741 v = get_bits(&s->gb, n); | |
742 if (v >= s->nb_block_sizes) | |
743 return -1; | |
744 s->prev_block_len_bits = s->frame_len_bits - v; | |
745 v = get_bits(&s->gb, n); | |
746 if (v >= s->nb_block_sizes) | |
747 return -1; | |
748 s->block_len_bits = s->frame_len_bits - v; | |
749 } else { | |
750 /* update block lengths */ | |
751 s->prev_block_len_bits = s->block_len_bits; | |
752 s->block_len_bits = s->next_block_len_bits; | |
753 } | |
754 v = get_bits(&s->gb, n); | |
755 if (v >= s->nb_block_sizes) | |
756 return -1; | |
757 s->next_block_len_bits = s->frame_len_bits - v; | |
758 } else { | |
759 /* fixed block len */ | |
760 s->next_block_len_bits = s->frame_len_bits; | |
761 s->prev_block_len_bits = s->frame_len_bits; | |
762 s->block_len_bits = s->frame_len_bits; | |
763 } | |
764 | |
765 /* now check if the block length is coherent with the frame length */ | |
766 s->block_len = 1 << s->block_len_bits; | |
767 if ((s->block_pos + s->block_len) > s->frame_len) | |
768 return -1; | |
769 | |
770 if (s->nb_channels == 2) { | |
771 s->ms_stereo = get_bits(&s->gb, 1); | |
772 } | |
773 v = 0; | |
774 for(ch = 0; ch < s->nb_channels; ch++) { | |
775 a = get_bits(&s->gb, 1); | |
776 s->channel_coded[ch] = a; | |
777 v |= a; | |
778 } | |
779 /* if no channel coded, no need to go further */ | |
780 /* XXX: fix potential framing problems */ | |
781 if (!v) | |
782 goto next; | |
783 | |
784 bsize = s->frame_len_bits - s->block_len_bits; | |
785 | |
786 /* read total gain and extract corresponding number of bits for | |
787 coef escape coding */ | |
788 total_gain = 1; | |
789 for(;;) { | |
790 a = get_bits(&s->gb, 7); | |
791 total_gain += a; | |
792 if (a != 127) | |
793 break; | |
794 } | |
2967 | 795 |
783 | 796 if (total_gain < 15) |
797 coef_nb_bits = 13; | |
798 else if (total_gain < 32) | |
799 coef_nb_bits = 12; | |
800 else if (total_gain < 40) | |
801 coef_nb_bits = 11; | |
802 else if (total_gain < 45) | |
803 coef_nb_bits = 10; | |
804 else | |
805 coef_nb_bits = 9; | |
806 | |
807 /* compute number of coefficients */ | |
808 n = s->coefs_end[bsize] - s->coefs_start; | |
809 for(ch = 0; ch < s->nb_channels; ch++) | |
810 nb_coefs[ch] = n; | |
811 | |
812 /* complex coding */ | |
813 if (s->use_noise_coding) { | |
814 | |
815 for(ch = 0; ch < s->nb_channels; ch++) { | |
816 if (s->channel_coded[ch]) { | |
817 int i, n, a; | |
818 n = s->exponent_high_sizes[bsize]; | |
819 for(i=0;i<n;i++) { | |
820 a = get_bits(&s->gb, 1); | |
821 s->high_band_coded[ch][i] = a; | |
822 /* if noise coding, the coefficients are not transmitted */ | |
823 if (a) | |
824 nb_coefs[ch] -= s->exponent_high_bands[bsize][i]; | |
825 } | |
826 } | |
827 } | |
828 for(ch = 0; ch < s->nb_channels; ch++) { | |
829 if (s->channel_coded[ch]) { | |
830 int i, n, val, code; | |
831 | |
832 n = s->exponent_high_sizes[bsize]; | |
833 val = (int)0x80000000; | |
834 for(i=0;i<n;i++) { | |
835 if (s->high_band_coded[ch][i]) { | |
836 if (val == (int)0x80000000) { | |
837 val = get_bits(&s->gb, 7) - 19; | |
838 } else { | |
3113 | 839 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX); |
783 | 840 if (code < 0) |
841 return -1; | |
842 val += code - 18; | |
843 } | |
844 s->high_band_values[ch][i] = val; | |
845 } | |
846 } | |
847 } | |
848 } | |
849 } | |
2967 | 850 |
783 | 851 /* exposant can be interpolated in short blocks. */ |
852 parse_exponents = 1; | |
853 if (s->block_len_bits != s->frame_len_bits) { | |
854 parse_exponents = get_bits(&s->gb, 1); | |
855 } | |
2967 | 856 |
783 | 857 if (parse_exponents) { |
858 for(ch = 0; ch < s->nb_channels; ch++) { | |
859 if (s->channel_coded[ch]) { | |
860 if (s->use_exp_vlc) { | |
861 if (decode_exp_vlc(s, ch) < 0) | |
862 return -1; | |
863 } else { | |
864 decode_exp_lsp(s, ch); | |
865 } | |
866 } | |
867 } | |
868 } else { | |
869 for(ch = 0; ch < s->nb_channels; ch++) { | |
870 if (s->channel_coded[ch]) { | |
2967 | 871 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits, |
783 | 872 s->block_len); |
873 } | |
874 } | |
875 } | |
876 | |
877 /* parse spectral coefficients : just RLE encoding */ | |
878 for(ch = 0; ch < s->nb_channels; ch++) { | |
879 if (s->channel_coded[ch]) { | |
880 VLC *coef_vlc; | |
881 int level, run, sign, tindex; | |
882 int16_t *ptr, *eptr; | |
883 const int16_t *level_table, *run_table; | |
884 | |
885 /* special VLC tables are used for ms stereo because | |
886 there is potentially less energy there */ | |
887 tindex = (ch == 1 && s->ms_stereo); | |
888 coef_vlc = &s->coef_vlc[tindex]; | |
889 run_table = s->run_table[tindex]; | |
890 level_table = s->level_table[tindex]; | |
891 /* XXX: optimize */ | |
892 ptr = &s->coefs1[ch][0]; | |
893 eptr = ptr + nb_coefs[ch]; | |
894 memset(ptr, 0, s->block_len * sizeof(int16_t)); | |
895 for(;;) { | |
3113 | 896 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX); |
783 | 897 if (code < 0) |
898 return -1; | |
899 if (code == 1) { | |
900 /* EOB */ | |
901 break; | |
902 } else if (code == 0) { | |
903 /* escape */ | |
904 level = get_bits(&s->gb, coef_nb_bits); | |
905 /* NOTE: this is rather suboptimal. reading | |
906 block_len_bits would be better */ | |
907 run = get_bits(&s->gb, s->frame_len_bits); | |
908 } else { | |
909 /* normal code */ | |
910 run = run_table[code]; | |
911 level = level_table[code]; | |
912 } | |
913 sign = get_bits(&s->gb, 1); | |
914 if (!sign) | |
915 level = -level; | |
916 ptr += run; | |
917 if (ptr >= eptr) | |
918 return -1; | |
919 *ptr++ = level; | |
920 /* NOTE: EOB can be omitted */ | |
921 if (ptr >= eptr) | |
922 break; | |
923 } | |
924 } | |
925 if (s->version == 1 && s->nb_channels >= 2) { | |
926 align_get_bits(&s->gb); | |
927 } | |
928 } | |
2967 | 929 |
783 | 930 /* normalize */ |
931 { | |
932 int n4 = s->block_len / 2; | |
933 mdct_norm = 1.0 / (float)n4; | |
934 if (s->version == 1) { | |
935 mdct_norm *= sqrt(n4); | |
936 } | |
937 } | |
938 | |
939 /* finally compute the MDCT coefficients */ | |
940 for(ch = 0; ch < s->nb_channels; ch++) { | |
941 if (s->channel_coded[ch]) { | |
942 int16_t *coefs1; | |
943 float *coefs, *exponents, mult, mult1, noise, *exp_ptr; | |
944 int i, j, n, n1, last_high_band; | |
945 float exp_power[HIGH_BAND_MAX_SIZE]; | |
946 | |
947 coefs1 = s->coefs1[ch]; | |
948 exponents = s->exponents[ch]; | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
949 mult = s->pow_005_10[total_gain] / s->max_exponent[ch]; |
783 | 950 mult *= mdct_norm; |
951 coefs = s->coefs[ch]; | |
952 if (s->use_noise_coding) { | |
953 mult1 = mult; | |
954 /* very low freqs : noise */ | |
955 for(i = 0;i < s->coefs_start; i++) { | |
956 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1; | |
957 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1); | |
958 } | |
2967 | 959 |
783 | 960 n1 = s->exponent_high_sizes[bsize]; |
961 | |
962 /* compute power of high bands */ | |
2967 | 963 exp_ptr = exponents + |
964 s->high_band_start[bsize] - | |
783 | 965 s->coefs_start; |
966 last_high_band = 0; /* avoid warning */ | |
967 for(j=0;j<n1;j++) { | |
2967 | 968 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 969 s->block_len_bits][j]; |
970 if (s->high_band_coded[ch][j]) { | |
971 float e2, v; | |
972 e2 = 0; | |
973 for(i = 0;i < n; i++) { | |
974 v = exp_ptr[i]; | |
975 e2 += v * v; | |
976 } | |
977 exp_power[j] = e2 / n; | |
978 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
|
979 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n); |
783 | 980 } |
981 exp_ptr += n; | |
982 } | |
983 | |
984 /* main freqs and high freqs */ | |
985 for(j=-1;j<n1;j++) { | |
986 if (j < 0) { | |
2967 | 987 n = s->high_band_start[bsize] - |
783 | 988 s->coefs_start; |
989 } else { | |
2967 | 990 n = s->exponent_high_bands[s->frame_len_bits - |
783 | 991 s->block_len_bits][j]; |
992 } | |
993 if (j >= 0 && s->high_band_coded[ch][j]) { | |
994 /* use noise with specified power */ | |
995 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]); | |
3176
babf844e1308
Init simplification and 2% faster wma_decode_block on amd64 with tables use
banan
parents:
3113
diff
changeset
|
996 mult1 = mult1 * s->pow_005_10[s->high_band_values[ch][j]]; |
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; |
1115 int i, n4, index, n; | |
1116 | |
1117 n = s->block_len; | |
1118 n4 = s->block_len / 2; | |
2967 | 1119 ff_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 /* multiply by the window */ | |
1125 for(i=0;i<n * 2;i++) { | |
1126 output[i] *= window[i]; | |
1127 } | |
1128 | |
1129 /* add in the frame */ | |
1130 index = (s->frame_len / 2) + s->block_pos - n4; | |
1131 ptr = &s->frame_out[ch][index]; | |
1132 for(i=0;i<n * 2;i++) { | |
1133 *ptr += output[i]; | |
1134 ptr++; | |
1135 } | |
1136 | |
1137 /* specific fast case for ms-stereo : add to second | |
1138 channel if it is not coded */ | |
1139 if (s->ms_stereo && !s->channel_coded[1]) { | |
1140 ptr = &s->frame_out[1][index]; | |
1141 for(i=0;i<n * 2;i++) { | |
1142 *ptr += output[i]; | |
1143 ptr++; | |
1144 } | |
1145 } | |
1146 } | |
1147 } | |
1148 next: | |
1149 /* update block number */ | |
1150 s->block_num++; | |
1151 s->block_pos += s->block_len; | |
1152 if (s->block_pos >= s->frame_len) | |
1153 return 1; | |
1154 else | |
1155 return 0; | |
1156 } | |
1157 | |
1158 /* decode a frame of frame_len samples */ | |
1159 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples) | |
1160 { | |
1161 int ret, i, n, a, ch, incr; | |
1162 int16_t *ptr; | |
1163 float *iptr; | |
1164 | |
1343 | 1165 #ifdef TRACE |
1166 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len); | |
1167 #endif | |
783 | 1168 |
1169 /* read each block */ | |
1170 s->block_num = 0; | |
1171 s->block_pos = 0; | |
1172 for(;;) { | |
1173 ret = wma_decode_block(s); | |
2967 | 1174 if (ret < 0) |
783 | 1175 return -1; |
1176 if (ret) | |
1177 break; | |
1178 } | |
1179 | |
1180 /* convert frame to integer */ | |
1181 n = s->frame_len; | |
1182 incr = s->nb_channels; | |
1183 for(ch = 0; ch < s->nb_channels; ch++) { | |
1184 ptr = samples + ch; | |
1185 iptr = s->frame_out[ch]; | |
1186 | |
1187 for(i=0;i<n;i++) { | |
797 | 1188 a = lrintf(*iptr++); |
783 | 1189 if (a > 32767) |
1190 a = 32767; | |
1191 else if (a < -32768) | |
1192 a = -32768; | |
1193 *ptr = a; | |
1194 ptr += incr; | |
1195 } | |
1196 /* prepare for next block */ | |
1197 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len], | |
1198 s->frame_len * sizeof(float)); | |
1199 /* XXX: suppress this */ | |
2967 | 1200 memset(&s->frame_out[ch][s->frame_len], 0, |
783 | 1201 s->frame_len * sizeof(float)); |
1202 } | |
1203 | |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
1204 #ifdef TRACE |
783 | 1205 dump_shorts("samples", samples, n * s->nb_channels); |
1206 #endif | |
1207 return 0; | |
1208 } | |
1209 | |
2967 | 1210 static int wma_decode_superframe(AVCodecContext *avctx, |
783 | 1211 void *data, int *data_size, |
1064 | 1212 uint8_t *buf, int buf_size) |
783 | 1213 { |
1214 WMADecodeContext *s = avctx->priv_data; | |
1215 int nb_frames, bit_offset, i, pos, len; | |
1216 uint8_t *q; | |
1217 int16_t *samples; | |
2967 | 1218 |
1342
f574934c4219
uniformization (now it uses the same trace functions as h264, defined in common.h)
al3x
parents:
1303
diff
changeset
|
1219 tprintf("***decode_superframe:\n"); |
783 | 1220 |
1750 | 1221 if(buf_size==0){ |
1222 s->last_superframe_len = 0; | |
1223 return 0; | |
1224 } | |
2967 | 1225 |
783 | 1226 samples = data; |
1227 | |
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
|
1228 init_get_bits(&s->gb, buf, buf_size*8); |
2967 | 1229 |
783 | 1230 if (s->use_bit_reservoir) { |
1231 /* read super frame header */ | |
1232 get_bits(&s->gb, 4); /* super frame index */ | |
1233 nb_frames = get_bits(&s->gb, 4) - 1; | |
1234 | |
1235 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3); | |
1236 | |
1237 if (s->last_superframe_len > 0) { | |
1238 // printf("skip=%d\n", s->last_bitoffset); | |
1239 /* add bit_offset bits to last frame */ | |
2967 | 1240 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) > |
783 | 1241 MAX_CODED_SUPERFRAME_SIZE) |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1242 goto fail; |
783 | 1243 q = s->last_superframe + s->last_superframe_len; |
1244 len = bit_offset; | |
1245 while (len > 0) { | |
1246 *q++ = (get_bits)(&s->gb, 8); | |
1247 len -= 8; | |
1248 } | |
1249 if (len > 0) { | |
1250 *q++ = (get_bits)(&s->gb, len) << (8 - len); | |
1251 } | |
2967 | 1252 |
783 | 1253 /* 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
|
1254 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8); |
783 | 1255 /* skip unused bits */ |
1256 if (s->last_bitoffset > 0) | |
1257 skip_bits(&s->gb, s->last_bitoffset); | |
1258 /* this frame is stored in the last superframe and in the | |
1259 current one */ | |
1260 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1261 goto fail; |
783 | 1262 samples += s->nb_channels * s->frame_len; |
1263 } | |
1264 | |
1265 /* read each frame starting from bit_offset */ | |
1266 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
|
1267 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8); |
783 | 1268 len = pos & 7; |
1269 if (len > 0) | |
1270 skip_bits(&s->gb, len); | |
2967 | 1271 |
783 | 1272 s->reset_block_lengths = 1; |
1273 for(i=0;i<nb_frames;i++) { | |
1274 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1275 goto fail; |
783 | 1276 samples += s->nb_channels * s->frame_len; |
1277 } | |
1278 | |
1279 /* we copy the end of the frame in the last frame buffer */ | |
1280 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7); | |
1281 s->last_bitoffset = pos & 7; | |
1282 pos >>= 3; | |
1283 len = buf_size - pos; | |
819 | 1284 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
|
1285 goto fail; |
783 | 1286 } |
1287 s->last_superframe_len = len; | |
1288 memcpy(s->last_superframe, buf + pos, len); | |
1289 } else { | |
1290 /* single frame decode */ | |
1291 if (wma_decode_frame(s, samples) < 0) | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1292 goto fail; |
783 | 1293 samples += s->nb_channels * s->frame_len; |
1294 } | |
1295 *data_size = (int8_t *)samples - (int8_t *)data; | |
1296 return s->block_align; | |
964
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1297 fail: |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1298 /* when error, we reset the bit reservoir */ |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1299 s->last_superframe_len = 0; |
6e6773512288
oops : better error resilience - should fix most wma decoding problems
bellard
parents:
819
diff
changeset
|
1300 return -1; |
783 | 1301 } |
1302 | |
1303 static int wma_decode_end(AVCodecContext *avctx) | |
1304 { | |
1305 WMADecodeContext *s = avctx->priv_data; | |
1306 int i; | |
1307 | |
1308 for(i = 0; i < s->nb_block_sizes; i++) | |
795
55add0e7eafb
avoid name clash - fixed again block size selection
bellard
parents:
785
diff
changeset
|
1309 ff_mdct_end(&s->mdct_ctx[i]); |
783 | 1310 for(i = 0; i < s->nb_block_sizes; i++) |
1311 av_free(s->windows[i]); | |
1312 | |
1313 if (s->use_exp_vlc) { | |
1314 free_vlc(&s->exp_vlc); | |
1315 } | |
1316 if (s->use_noise_coding) { | |
1317 free_vlc(&s->hgain_vlc); | |
1318 } | |
1319 for(i = 0;i < 2; i++) { | |
1320 free_vlc(&s->coef_vlc[i]); | |
1321 av_free(s->run_table[i]); | |
1322 av_free(s->level_table[i]); | |
1323 } | |
2967 | 1324 |
783 | 1325 return 0; |
1326 } | |
1327 | |
1328 AVCodec wmav1_decoder = | |
1329 { | |
1330 "wmav1", | |
1331 CODEC_TYPE_AUDIO, | |
1332 CODEC_ID_WMAV1, | |
1333 sizeof(WMADecodeContext), | |
1334 wma_decode_init, | |
1335 NULL, | |
1336 wma_decode_end, | |
1337 wma_decode_superframe, | |
1338 }; | |
1339 | |
1340 AVCodec wmav2_decoder = | |
1341 { | |
1342 "wmav2", | |
1343 CODEC_TYPE_AUDIO, | |
1344 CODEC_ID_WMAV2, | |
1345 sizeof(WMADecodeContext), | |
1346 wma_decode_init, | |
1347 NULL, | |
1348 wma_decode_end, | |
1349 wma_decode_superframe, | |
1350 }; |