2
* WMA compatible decoder
3
* Copyright (c) 2002 The FFmpeg Project.
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.
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.
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
17
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22
* WMA compatible decoder.
29
#define BLOCK_MIN_BITS 7
30
#define BLOCK_MAX_BITS 11
31
#define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
33
#define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
35
/* XXX: find exact max size */
36
#define HIGH_BAND_MAX_SIZE 16
38
#define NB_LSP_COEFS 10
40
/* XXX: is it a suitable value ? */
41
#define MAX_CODED_SUPERFRAME_SIZE 4096
43
#define MAX_CHANNELS 2
45
#define NOISE_TAB_SIZE 8192
47
#define LSP_POW_BITS 7
49
typedef struct WMADecodeContext {
54
int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
56
int use_bit_reservoir;
57
int use_variable_block_len;
58
int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
59
int use_noise_coding; /* true if perceptual noise is added */
62
int exponent_sizes[BLOCK_NB_SIZES];
63
uint16_t exponent_bands[BLOCK_NB_SIZES][25];
64
int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
65
int coefs_start; /* first coded coef */
66
int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
67
int exponent_high_sizes[BLOCK_NB_SIZES];
68
int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
71
/* coded values in high bands */
72
int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
73
int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
75
/* there are two possible tables for spectral coefficients */
77
uint16_t *run_table[2];
78
uint16_t *level_table[2];
80
int frame_len; /* frame length in samples */
81
int frame_len_bits; /* frame_len = 1 << frame_len_bits */
82
int nb_block_sizes; /* number of block sizes */
84
int reset_block_lengths;
85
int block_len_bits; /* log2 of current block length */
86
int next_block_len_bits; /* log2 of next block length */
87
int prev_block_len_bits; /* log2 of prev block length */
88
int block_len; /* block length in samples */
89
int block_num; /* block number in current frame */
90
int block_pos; /* current position in frame */
91
uint8_t ms_stereo; /* true if mid/side stereo mode */
92
uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
93
float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
94
float max_exponent[MAX_CHANNELS];
95
int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
96
float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
97
MDCTContext mdct_ctx[BLOCK_NB_SIZES];
98
float *windows[BLOCK_NB_SIZES];
99
FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
100
/* output buffer for one frame and the last for IMDCT windowing */
101
float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
102
/* last frame info */
103
uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
105
int last_superframe_len;
106
float noise_table[NOISE_TAB_SIZE];
108
float noise_mult; /* XXX: suppress that and integrate it in the noise array */
109
/* lsp_to_curve tables */
110
float lsp_cos_table[BLOCK_MAX_SIZE];
111
float lsp_pow_e_table[256];
112
float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
113
float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
120
typedef struct CoefVLCTable {
121
int n; /* total number of codes */
122
const uint32_t *huffcodes; /* VLC bit values */
123
const uint8_t *huffbits; /* VLC bit size */
124
const uint16_t *levels; /* table to build run/level tables */
127
static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
132
static void dump_shorts(const char *name, const short *tab, int n)
136
tprintf("%s[%d]:\n", name, n);
140
tprintf(" %5d.0", tab[i]);
146
static void dump_floats(const char *name, int prec, const float *tab, int n)
150
tprintf("%s[%d]:\n", name, n);
154
tprintf(" %8.*f", prec, tab[i]);
163
/* XXX: use same run/length optimization as mpeg decoders */
164
static void init_coef_vlc(VLC *vlc,
165
uint16_t **prun_table, uint16_t **plevel_table,
166
const CoefVLCTable *vlc_table)
168
int n = vlc_table->n;
169
const uint8_t *table_bits = vlc_table->huffbits;
170
const uint32_t *table_codes = vlc_table->huffcodes;
171
const uint16_t *levels_table = vlc_table->levels;
172
uint16_t *run_table, *level_table;
176
init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4);
178
run_table = av_malloc(n * sizeof(uint16_t));
179
level_table = av_malloc(n * sizeof(uint16_t));
187
level_table[i] = level;
192
*prun_table = run_table;
193
*plevel_table = level_table;
196
static int wma_decode_init(AVCodecContext * avctx)
198
WMADecodeContext *s = avctx->priv_data;
199
int i, flags1, flags2;
202
float bps1, high_freq, bps;
206
s->sample_rate = avctx->sample_rate;
207
s->nb_channels = avctx->channels;
208
s->bit_rate = avctx->bit_rate;
209
s->block_align = avctx->block_align;
211
if (avctx->codec->id == CODEC_ID_WMAV1) {
217
/* extract flag infos */
220
extradata = avctx->extradata;
221
if (s->version == 1 && avctx->extradata_size >= 4) {
222
flags1 = extradata[0] | (extradata[1] << 8);
223
flags2 = extradata[2] | (extradata[3] << 8);
224
} else if (s->version == 2 && avctx->extradata_size >= 6) {
225
flags1 = extradata[0] | (extradata[1] << 8) |
226
(extradata[2] << 16) | (extradata[3] << 24);
227
flags2 = extradata[4] | (extradata[5] << 8);
229
s->use_exp_vlc = flags2 & 0x0001;
230
s->use_bit_reservoir = flags2 & 0x0002;
231
s->use_variable_block_len = flags2 & 0x0004;
233
/* compute MDCT block size */
234
if (s->sample_rate <= 16000) {
235
s->frame_len_bits = 9;
236
} else if (s->sample_rate <= 22050 ||
237
(s->sample_rate <= 32000 && s->version == 1)) {
238
s->frame_len_bits = 10;
240
s->frame_len_bits = 11;
242
s->frame_len = 1 << s->frame_len_bits;
243
if (s->use_variable_block_len) {
245
nb = ((flags2 >> 3) & 3) + 1;
246
if ((s->bit_rate / s->nb_channels) >= 32000)
248
nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
251
s->nb_block_sizes = nb + 1;
253
s->nb_block_sizes = 1;
256
/* init rate dependant parameters */
257
s->use_noise_coding = 1;
258
high_freq = s->sample_rate * 0.5;
260
/* if version 2, then the rates are normalized */
261
sample_rate1 = s->sample_rate;
262
if (s->version == 2) {
263
if (sample_rate1 >= 44100)
264
sample_rate1 = 44100;
265
else if (sample_rate1 >= 22050)
266
sample_rate1 = 22050;
267
else if (sample_rate1 >= 16000)
268
sample_rate1 = 16000;
269
else if (sample_rate1 >= 11025)
270
sample_rate1 = 11025;
271
else if (sample_rate1 >= 8000)
275
bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
276
s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
278
/* compute high frequency value and choose if noise coding should
281
if (s->nb_channels == 2)
283
if (sample_rate1 == 44100) {
285
s->use_noise_coding = 0;
287
high_freq = high_freq * 0.4;
288
} else if (sample_rate1 == 22050) {
290
s->use_noise_coding = 0;
291
else if (bps1 >= 0.72)
292
high_freq = high_freq * 0.7;
294
high_freq = high_freq * 0.6;
295
} else if (sample_rate1 == 16000) {
297
high_freq = high_freq * 0.5;
299
high_freq = high_freq * 0.3;
300
} else if (sample_rate1 == 11025) {
301
high_freq = high_freq * 0.7;
302
} else if (sample_rate1 == 8000) {
304
high_freq = high_freq * 0.5;
305
} else if (bps > 0.75) {
306
s->use_noise_coding = 0;
308
high_freq = high_freq * 0.65;
312
high_freq = high_freq * 0.75;
313
} else if (bps >= 0.6) {
314
high_freq = high_freq * 0.6;
316
high_freq = high_freq * 0.5;
319
dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
320
dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
321
s->version, s->nb_channels, s->sample_rate, s->bit_rate,
323
dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
324
bps, bps1, high_freq, s->byte_offset_bits);
325
dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
326
s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
328
/* compute the scale factor band sizes for each MDCT block size */
330
int a, b, pos, lpos, k, block_len, i, j, n;
331
const uint8_t *table;
333
if (s->version == 1) {
338
for(k = 0; k < s->nb_block_sizes; k++) {
339
block_len = s->frame_len >> k;
341
if (s->version == 1) {
344
a = wma_critical_freqs[i];
346
pos = ((block_len * 2 * a) + (b >> 1)) / b;
349
s->exponent_bands[0][i] = pos - lpos;
350
if (pos >= block_len) {
356
s->exponent_sizes[0] = i;
358
/* hardcoded tables */
360
a = s->frame_len_bits - BLOCK_MIN_BITS - k;
362
if (s->sample_rate >= 44100)
363
table = exponent_band_44100[a];
364
else if (s->sample_rate >= 32000)
365
table = exponent_band_32000[a];
366
else if (s->sample_rate >= 22050)
367
table = exponent_band_22050[a];
372
s->exponent_bands[k][i] = table[i];
373
s->exponent_sizes[k] = n;
378
a = wma_critical_freqs[i];
380
pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
385
s->exponent_bands[k][j++] = pos - lpos;
386
if (pos >= block_len)
390
s->exponent_sizes[k] = j;
394
/* max number of coefs */
395
s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
396
/* high freq computation */
397
s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
398
s->sample_rate + 0.5);
399
n = s->exponent_sizes[k];
405
pos += s->exponent_bands[k][i];
407
if (start < s->high_band_start[k])
408
start = s->high_band_start[k];
409
if (end > s->coefs_end[k])
410
end = s->coefs_end[k];
412
s->exponent_high_bands[k][j++] = end - start;
414
s->exponent_high_sizes[k] = j;
416
tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
419
s->high_band_start[k],
420
s->exponent_high_sizes[k]);
421
for(j=0;j<s->exponent_high_sizes[k];j++)
422
tprintf(" %d", s->exponent_high_bands[k][j]);
431
for(i = 0; i < s->nb_block_sizes; i++) {
432
tprintf("%5d: n=%2d:",
434
s->exponent_sizes[i]);
435
for(j=0;j<s->exponent_sizes[i];j++)
436
tprintf(" %d", s->exponent_bands[i][j]);
443
for(i = 0; i < s->nb_block_sizes; i++)
444
ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
446
/* init MDCT windows : simple sinus window */
447
for(i = 0; i < s->nb_block_sizes; i++) {
450
n = 1 << (s->frame_len_bits - i);
451
window = av_malloc(sizeof(float) * n);
452
alpha = M_PI / (2.0 * n);
454
window[n - j - 1] = sin((j + 0.5) * alpha);
456
s->windows[i] = window;
459
s->reset_block_lengths = 1;
461
if (s->use_noise_coding) {
463
/* init the noise generator */
465
s->noise_mult = 0.02;
467
s->noise_mult = 0.04;
470
for(i=0;i<NOISE_TAB_SIZE;i++)
471
s->noise_table[i] = 1.0 * s->noise_mult;
477
norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
478
for(i=0;i<NOISE_TAB_SIZE;i++) {
479
seed = seed * 314159 + 1;
480
s->noise_table[i] = (float)((int)seed) * norm;
484
init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
485
hgain_huffbits, 1, 1,
486
hgain_huffcodes, 2, 2);
489
if (s->use_exp_vlc) {
490
init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
491
scale_huffbits, 1, 1,
492
scale_huffcodes, 4, 4);
494
wma_lsp_to_curve_init(s, s->frame_len);
497
/* choose the VLC tables for the coefficients */
499
if (s->sample_rate >= 32000) {
502
else if (bps1 < 1.16)
506
init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
507
&coef_vlcs[coef_vlc_table * 2]);
508
init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
509
&coef_vlcs[coef_vlc_table * 2 + 1]);
513
/* interpolate values for a bigger or smaller block. The block must
514
have multiple sizes */
515
static void interpolate_array(float *scale, int old_size, int new_size)
520
if (new_size > old_size) {
521
jincr = new_size / old_size;
523
for(i = old_size - 1; i >=0; i--) {
530
} else if (new_size < old_size) {
532
jincr = old_size / new_size;
533
for(i = 0; i < new_size; i++) {
540
/* compute x^-0.25 with an exponent and mantissa table. We use linear
541
interpolation to reduce the mantissa table size at a small speed
542
expense (linear interpolation approximately doubles the number of
543
bits of precision). */
544
static inline float pow_m1_4(WMADecodeContext *s, float x)
555
m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
556
/* build interpolation scale: 1 <= t < 2. */
557
t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
558
a = s->lsp_pow_m_table1[m];
559
b = s->lsp_pow_m_table2[m];
560
return s->lsp_pow_e_table[e] * (a + b * t.f);
563
static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
568
wdel = M_PI / frame_len;
569
for(i=0;i<frame_len;i++)
570
s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
572
/* tables for x^-0.25 computation */
575
s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
578
/* NOTE: these two tables are needed to avoid two operations in
581
for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
582
m = (1 << LSP_POW_BITS) + i;
583
a = (float)m * (0.5 / (1 << LSP_POW_BITS));
585
s->lsp_pow_m_table1[i] = 2 * a - b;
586
s->lsp_pow_m_table2[i] = b - a;
595
printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
600
/* NOTE: We use the same code as Vorbis here */
601
/* XXX: optimize it further with SSE/3Dnow */
602
static void wma_lsp_to_curve(WMADecodeContext *s,
603
float *out, float *val_max_ptr,
607
float p, q, w, v, val_max;
613
w = s->lsp_cos_table[i];
614
for(j=1;j<NB_LSP_COEFS;j+=2){
626
*val_max_ptr = val_max;
629
/* decode exponents coded with LSP coefficients (same idea as Vorbis) */
630
static void decode_exp_lsp(WMADecodeContext *s, int ch)
632
float lsp_coefs[NB_LSP_COEFS];
635
for(i = 0; i < NB_LSP_COEFS; i++) {
636
if (i == 0 || i >= 8)
637
val = get_bits(&s->gb, 3);
639
val = get_bits(&s->gb, 4);
640
lsp_coefs[i] = lsp_codebook[i][val];
643
wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
644
s->block_len, lsp_coefs);
647
/* decode exponents coded with VLC codes */
648
static int decode_exp_vlc(WMADecodeContext *s, int ch)
650
int last_exp, n, code;
651
const uint16_t *ptr, *band_ptr;
652
float v, *q, max_scale, *q_end;
654
band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
656
q = s->exponents[ch];
657
q_end = q + s->block_len;
659
if (s->version == 1) {
660
last_exp = get_bits(&s->gb, 5) + 10;
661
/* XXX: use a table */
662
v = pow(10, last_exp * (1.0 / 16.0));
671
code = get_vlc(&s->gb, &s->exp_vlc);
674
/* NOTE: this offset is the same as MPEG4 AAC ! */
675
last_exp += code - 60;
676
/* XXX: use a table */
677
v = pow(10, last_exp * (1.0 / 16.0));
685
s->max_exponent[ch] = max_scale;
689
/* return 0 if OK. return 1 if last block of frame. return -1 if
690
unrecorrable error. */
691
static int wma_decode_block(WMADecodeContext *s)
693
int n, v, a, ch, code, bsize;
694
int coef_nb_bits, total_gain, parse_exponents;
695
float window[BLOCK_MAX_SIZE * 2];
696
int nb_coefs[MAX_CHANNELS];
700
tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
703
/* compute current block length */
704
if (s->use_variable_block_len) {
705
n = av_log2(s->nb_block_sizes - 1) + 1;
707
if (s->reset_block_lengths) {
708
s->reset_block_lengths = 0;
709
v = get_bits(&s->gb, n);
710
if (v >= s->nb_block_sizes)
712
s->prev_block_len_bits = s->frame_len_bits - v;
713
v = get_bits(&s->gb, n);
714
if (v >= s->nb_block_sizes)
716
s->block_len_bits = s->frame_len_bits - v;
718
/* update block lengths */
719
s->prev_block_len_bits = s->block_len_bits;
720
s->block_len_bits = s->next_block_len_bits;
722
v = get_bits(&s->gb, n);
723
if (v >= s->nb_block_sizes)
725
s->next_block_len_bits = s->frame_len_bits - v;
727
/* fixed block len */
728
s->next_block_len_bits = s->frame_len_bits;
729
s->prev_block_len_bits = s->frame_len_bits;
730
s->block_len_bits = s->frame_len_bits;
733
/* now check if the block length is coherent with the frame length */
734
s->block_len = 1 << s->block_len_bits;
735
if ((s->block_pos + s->block_len) > s->frame_len)
738
if (s->nb_channels == 2) {
739
s->ms_stereo = get_bits(&s->gb, 1);
742
for(ch = 0; ch < s->nb_channels; ch++) {
743
a = get_bits(&s->gb, 1);
744
s->channel_coded[ch] = a;
747
/* if no channel coded, no need to go further */
748
/* XXX: fix potential framing problems */
752
bsize = s->frame_len_bits - s->block_len_bits;
754
/* read total gain and extract corresponding number of bits for
755
coef escape coding */
758
a = get_bits(&s->gb, 7);
766
else if (total_gain < 32)
768
else if (total_gain < 40)
770
else if (total_gain < 45)
775
/* compute number of coefficients */
776
n = s->coefs_end[bsize] - s->coefs_start;
777
for(ch = 0; ch < s->nb_channels; ch++)
781
if (s->use_noise_coding) {
783
for(ch = 0; ch < s->nb_channels; ch++) {
784
if (s->channel_coded[ch]) {
786
n = s->exponent_high_sizes[bsize];
788
a = get_bits(&s->gb, 1);
789
s->high_band_coded[ch][i] = a;
790
/* if noise coding, the coefficients are not transmitted */
792
nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
796
for(ch = 0; ch < s->nb_channels; ch++) {
797
if (s->channel_coded[ch]) {
800
n = s->exponent_high_sizes[bsize];
801
val = (int)0x80000000;
803
if (s->high_band_coded[ch][i]) {
804
if (val == (int)0x80000000) {
805
val = get_bits(&s->gb, 7) - 19;
807
code = get_vlc(&s->gb, &s->hgain_vlc);
812
s->high_band_values[ch][i] = val;
819
/* exposant can be interpolated in short blocks. */
821
if (s->block_len_bits != s->frame_len_bits) {
822
parse_exponents = get_bits(&s->gb, 1);
825
if (parse_exponents) {
826
for(ch = 0; ch < s->nb_channels; ch++) {
827
if (s->channel_coded[ch]) {
828
if (s->use_exp_vlc) {
829
if (decode_exp_vlc(s, ch) < 0)
832
decode_exp_lsp(s, ch);
837
for(ch = 0; ch < s->nb_channels; ch++) {
838
if (s->channel_coded[ch]) {
839
interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
845
/* parse spectral coefficients : just RLE encoding */
846
for(ch = 0; ch < s->nb_channels; ch++) {
847
if (s->channel_coded[ch]) {
849
int level, run, sign, tindex;
851
const int16_t *level_table, *run_table;
853
/* special VLC tables are used for ms stereo because
854
there is potentially less energy there */
855
tindex = (ch == 1 && s->ms_stereo);
856
coef_vlc = &s->coef_vlc[tindex];
857
run_table = s->run_table[tindex];
858
level_table = s->level_table[tindex];
860
ptr = &s->coefs1[ch][0];
861
eptr = ptr + nb_coefs[ch];
862
memset(ptr, 0, s->block_len * sizeof(int16_t));
864
code = get_vlc(&s->gb, coef_vlc);
870
} else if (code == 0) {
872
level = get_bits(&s->gb, coef_nb_bits);
873
/* NOTE: this is rather suboptimal. reading
874
block_len_bits would be better */
875
run = get_bits(&s->gb, s->frame_len_bits);
878
run = run_table[code];
879
level = level_table[code];
881
sign = get_bits(&s->gb, 1);
888
/* NOTE: EOB can be omitted */
893
if (s->version == 1 && s->nb_channels >= 2) {
894
align_get_bits(&s->gb);
900
int n4 = s->block_len / 2;
901
mdct_norm = 1.0 / (float)n4;
902
if (s->version == 1) {
903
mdct_norm *= sqrt(n4);
907
/* finally compute the MDCT coefficients */
908
for(ch = 0; ch < s->nb_channels; ch++) {
909
if (s->channel_coded[ch]) {
911
float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
912
int i, j, n, n1, last_high_band;
913
float exp_power[HIGH_BAND_MAX_SIZE];
915
coefs1 = s->coefs1[ch];
916
exponents = s->exponents[ch];
917
mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
919
coefs = s->coefs[ch];
920
if (s->use_noise_coding) {
922
/* very low freqs : noise */
923
for(i = 0;i < s->coefs_start; i++) {
924
*coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
925
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
928
n1 = s->exponent_high_sizes[bsize];
930
/* compute power of high bands */
931
exp_ptr = exponents +
932
s->high_band_start[bsize] -
934
last_high_band = 0; /* avoid warning */
936
n = s->exponent_high_bands[s->frame_len_bits -
937
s->block_len_bits][j];
938
if (s->high_band_coded[ch][j]) {
941
for(i = 0;i < n; i++) {
945
exp_power[j] = e2 / n;
947
tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
952
/* main freqs and high freqs */
955
n = s->high_band_start[bsize] -
958
n = s->exponent_high_bands[s->frame_len_bits -
959
s->block_len_bits][j];
961
if (j >= 0 && s->high_band_coded[ch][j]) {
962
/* use noise with specified power */
963
mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
964
/* XXX: use a table */
965
mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
966
mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
968
for(i = 0;i < n; i++) {
969
noise = s->noise_table[s->noise_index];
970
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
971
*coefs++ = (*exponents++) * noise * mult1;
974
/* coded values + small noise */
975
for(i = 0;i < n; i++) {
976
noise = s->noise_table[s->noise_index];
977
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
978
*coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
983
/* very high freqs : noise */
984
n = s->block_len - s->coefs_end[bsize];
985
mult1 = mult * exponents[-1];
986
for(i = 0; i < n; i++) {
987
*coefs++ = s->noise_table[s->noise_index] * mult1;
988
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
991
/* XXX: optimize more */
992
for(i = 0;i < s->coefs_start; i++)
995
for(i = 0;i < n; i++) {
996
*coefs++ = coefs1[i] * exponents[i] * mult;
998
n = s->block_len - s->coefs_end[bsize];
999
for(i = 0;i < n; i++)
1006
for(ch = 0; ch < s->nb_channels; ch++) {
1007
if (s->channel_coded[ch]) {
1008
dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1009
dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1014
if (s->ms_stereo && s->channel_coded[1]) {
1018
/* nominal case for ms stereo: we do it before mdct */
1019
/* no need to optimize this case because it should almost
1021
if (!s->channel_coded[0]) {
1022
tprintf("rare ms-stereo case happened\n");
1023
memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1024
s->channel_coded[0] = 1;
1027
for(i = 0; i < s->block_len; i++) {
1030
s->coefs[0][i] = a + b;
1031
s->coefs[1][i] = a - b;
1035
/* build the window : we ensure that when the windows overlap
1036
their squared sum is always 1 (MDCT reconstruction rule) */
1037
/* XXX: merge with output */
1039
int i, next_block_len, block_len, prev_block_len, n;
1042
block_len = s->block_len;
1043
prev_block_len = 1 << s->prev_block_len_bits;
1044
next_block_len = 1 << s->next_block_len_bits;
1047
wptr = window + block_len;
1048
if (block_len <= next_block_len) {
1049
for(i=0;i<block_len;i++)
1050
*wptr++ = s->windows[bsize][i];
1053
n = (block_len / 2) - (next_block_len / 2);
1056
for(i=0;i<next_block_len;i++)
1057
*wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1063
wptr = window + block_len;
1064
if (block_len <= prev_block_len) {
1065
for(i=0;i<block_len;i++)
1066
*--wptr = s->windows[bsize][i];
1069
n = (block_len / 2) - (prev_block_len / 2);
1072
for(i=0;i<prev_block_len;i++)
1073
*--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1080
for(ch = 0; ch < s->nb_channels; ch++) {
1081
if (s->channel_coded[ch]) {
1082
FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1084
int i, n4, index, n;
1087
n4 = s->block_len / 2;
1088
ff_imdct_calc(&s->mdct_ctx[bsize],
1089
output, s->coefs[ch], s->mdct_tmp);
1091
/* XXX: optimize all that by build the window and
1092
multipying/adding at the same time */
1093
/* multiply by the window */
1094
for(i=0;i<n * 2;i++) {
1095
output[i] *= window[i];
1098
/* add in the frame */
1099
index = (s->frame_len / 2) + s->block_pos - n4;
1100
ptr = &s->frame_out[ch][index];
1101
for(i=0;i<n * 2;i++) {
1106
/* specific fast case for ms-stereo : add to second
1107
channel if it is not coded */
1108
if (s->ms_stereo && !s->channel_coded[1]) {
1109
ptr = &s->frame_out[1][index];
1110
for(i=0;i<n * 2;i++) {
1118
/* update block number */
1120
s->block_pos += s->block_len;
1121
if (s->block_pos >= s->frame_len)
1127
/* decode a frame of frame_len samples */
1128
static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1130
int ret, i, n, a, ch, incr;
1135
tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
1138
/* read each block */
1142
ret = wma_decode_block(s);
1149
/* convert frame to integer */
1151
incr = s->nb_channels;
1152
for(ch = 0; ch < s->nb_channels; ch++) {
1154
iptr = s->frame_out[ch];
1157
a = lrintf(*iptr++);
1160
else if (a < -32768)
1165
/* prepare for next block */
1166
memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1167
s->frame_len * sizeof(float));
1168
/* XXX: suppress this */
1169
memset(&s->frame_out[ch][s->frame_len], 0,
1170
s->frame_len * sizeof(float));
1174
dump_shorts("samples", samples, n * s->nb_channels);
1179
static int wma_decode_superframe(AVCodecContext *avctx,
1180
void *data, int *data_size,
1181
uint8_t *buf, int buf_size)
1183
WMADecodeContext *s = avctx->priv_data;
1184
int nb_frames, bit_offset, i, pos, len;
1188
tprintf("***decode_superframe:\n");
1192
init_get_bits(&s->gb, buf, buf_size*8);
1194
if (s->use_bit_reservoir) {
1195
/* read super frame header */
1196
get_bits(&s->gb, 4); /* super frame index */
1197
nb_frames = get_bits(&s->gb, 4) - 1;
1199
bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1201
if (s->last_superframe_len > 0) {
1202
// printf("skip=%d\n", s->last_bitoffset);
1203
/* add bit_offset bits to last frame */
1204
if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1205
MAX_CODED_SUPERFRAME_SIZE)
1207
q = s->last_superframe + s->last_superframe_len;
1210
*q++ = (get_bits)(&s->gb, 8);
1214
*q++ = (get_bits)(&s->gb, len) << (8 - len);
1217
/* XXX: bit_offset bits into last frame */
1218
init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1219
/* skip unused bits */
1220
if (s->last_bitoffset > 0)
1221
skip_bits(&s->gb, s->last_bitoffset);
1222
/* this frame is stored in the last superframe and in the
1224
if (wma_decode_frame(s, samples) < 0)
1226
samples += s->nb_channels * s->frame_len;
1229
/* read each frame starting from bit_offset */
1230
pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1231
init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1234
skip_bits(&s->gb, len);
1236
s->reset_block_lengths = 1;
1237
for(i=0;i<nb_frames;i++) {
1238
if (wma_decode_frame(s, samples) < 0)
1240
samples += s->nb_channels * s->frame_len;
1243
/* we copy the end of the frame in the last frame buffer */
1244
pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1245
s->last_bitoffset = pos & 7;
1247
len = buf_size - pos;
1248
if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1251
s->last_superframe_len = len;
1252
memcpy(s->last_superframe, buf + pos, len);
1254
/* single frame decode */
1255
if (wma_decode_frame(s, samples) < 0)
1257
samples += s->nb_channels * s->frame_len;
1259
*data_size = (int8_t *)samples - (int8_t *)data;
1260
return s->block_align;
1262
/* when error, we reset the bit reservoir */
1263
s->last_superframe_len = 0;
1267
static int wma_decode_end(AVCodecContext *avctx)
1269
WMADecodeContext *s = avctx->priv_data;
1272
for(i = 0; i < s->nb_block_sizes; i++)
1273
ff_mdct_end(&s->mdct_ctx[i]);
1274
for(i = 0; i < s->nb_block_sizes; i++)
1275
av_free(s->windows[i]);
1277
if (s->use_exp_vlc) {
1278
free_vlc(&s->exp_vlc);
1280
if (s->use_noise_coding) {
1281
free_vlc(&s->hgain_vlc);
1283
for(i = 0;i < 2; i++) {
1284
free_vlc(&s->coef_vlc[i]);
1285
av_free(s->run_table[i]);
1286
av_free(s->level_table[i]);
1292
AVCodec wmav1_decoder =
1297
sizeof(WMADecodeContext),
1301
wma_decode_superframe,
1304
AVCodec wmav2_decoder =
1309
sizeof(WMADecodeContext),
1313
wma_decode_superframe,
added
removed
ffmpeg/libavcodec/wmadec.c
