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.
26 #include <codecs/lib/codeclib.h>
27 #include <codecs/libasf/asf.h>
32 static void wma_lsp_to_curve_init(WMADecodeContext
*s
, int frame_len
);
34 /*declarations of statically allocated variables used to remove malloc calls*/
36 fixed32 coefsarray
[MAX_CHANNELS
][BLOCK_MAX_SIZE
] IBSS_ATTR MEM_ALIGN_ATTR
;
37 /*decode and window into IRAM on targets with at least 80KB of codec IRAM*/
38 fixed32 frame_out_buf
[MAX_CHANNELS
][BLOCK_MAX_SIZE
* 2] IBSS_ATTR_WMA_LARGE_IRAM MEM_ALIGN_ATTR
;
40 /*MDCT reconstruction windows*/
41 fixed32 stat0
[2048] MEM_ALIGN_ATTR
, stat1
[1024] MEM_ALIGN_ATTR
,
42 stat2
[512] MEM_ALIGN_ATTR
, stat3
[256] MEM_ALIGN_ATTR
, stat4
[128] MEM_ALIGN_ATTR
;
45 uint16_t *runtabarray
[2], *levtabarray
[2];
47 uint16_t runtab_big
[1336] MEM_ALIGN_ATTR
, runtab_small
[1072] MEM_ALIGN_ATTR
,
48 levtab_big
[1336] MEM_ALIGN_ATTR
, levtab_small
[1072] MEM_ALIGN_ATTR
;
50 #define VLCBUF1SIZE 4598
51 #define VLCBUF2SIZE 3574
52 #define VLCBUF3SIZE 360
53 #define VLCBUF4SIZE 540
55 /*putting these in IRAM actually makes PP slower*/
57 VLC_TYPE vlcbuf1
[VLCBUF1SIZE
][2] MEM_ALIGN_ATTR
;
58 VLC_TYPE vlcbuf2
[VLCBUF2SIZE
][2] MEM_ALIGN_ATTR
;
59 /* This buffer gets reused for lsp tables */
60 VLC_TYPE vlcbuf3
[VLCBUF3SIZE
][2] MEM_ALIGN_ATTR
;
61 VLC_TYPE vlcbuf4
[VLCBUF4SIZE
][2] MEM_ALIGN_ATTR
;
67 * Apply MDCT window and add into output.
69 * We ensure that when the windows overlap their squared sum
70 * is always 1 (MDCT reconstruction rule).
72 * The Vorbis I spec has a great diagram explaining this process.
73 * See section 1.3.2.3 of http://xiph.org/vorbis/doc/Vorbis_I_spec.html
75 static void wma_window(WMADecodeContext
*s
, fixed32
*in
, fixed32
*out
)
77 //float *in = s->output;
78 int block_len
, bsize
, n
;
82 /* previous block was larger, so we'll use the size of the current
83 * block to set the window size*/
84 if (s
->block_len_bits
<= s
->prev_block_len_bits
) {
85 block_len
= s
->block_len
;
86 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
88 vector_fmul_add_add(out
, in
, s
->windows
[bsize
], block_len
);
91 /*previous block was smaller or the same size, so use it's size to set the window length*/
92 block_len
= 1 << s
->prev_block_len_bits
;
93 /*find the middle of the two overlapped blocks, this will be the first overlapped sample*/
94 n
= (s
->block_len
- block_len
) / 2;
95 bsize
= s
->frame_len_bits
- s
->prev_block_len_bits
;
97 vector_fmul_add_add(out
+n
, in
+n
, s
->windows
[bsize
], block_len
);
99 memcpy(out
+n
+block_len
, in
+n
+block_len
, n
*sizeof(fixed32
));
101 /* Advance to the end of the current block and prepare to window it for the next block.
102 * Since the window function needs to be reversed, we do it backwards starting with the
103 * last sample and moving towards the first
109 if (s
->block_len_bits
<= s
->next_block_len_bits
) {
110 block_len
= s
->block_len
;
111 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
113 vector_fmul_reverse(out
, in
, s
->windows
[bsize
], block_len
);
116 block_len
= 1 << s
->next_block_len_bits
;
117 n
= (s
->block_len
- block_len
) / 2;
118 bsize
= s
->frame_len_bits
- s
->next_block_len_bits
;
120 memcpy(out
, in
, n
*sizeof(fixed32
));
122 vector_fmul_reverse(out
+n
, in
+n
, s
->windows
[bsize
], block_len
);
124 memset(out
+n
+block_len
, 0, n
*sizeof(fixed32
));
131 /* XXX: use same run/length optimization as mpeg decoders */
132 static void init_coef_vlc(VLC
*vlc
,
133 uint16_t **prun_table
, uint16_t **plevel_table
,
134 const CoefVLCTable
*vlc_table
, int tab
)
136 int n
= vlc_table
->n
;
137 const uint8_t *table_bits
= vlc_table
->huffbits
;
138 const uint32_t *table_codes
= vlc_table
->huffcodes
;
139 const uint16_t *levels_table
= vlc_table
->levels
;
140 uint16_t *run_table
, *level_table
;
145 init_vlc(vlc
, VLCBITS
, n
, table_bits
, 1, 1, table_codes
, 4, 4, INIT_VLC_USE_NEW_STATIC
);
147 run_table
= runtabarray
[tab
];
148 level_table
= levtabarray
[tab
];
159 level_table
[i
] = level
;
164 *prun_table
= run_table
;
165 *plevel_table
= level_table
;
168 int wma_decode_init(WMADecodeContext
* s
, asf_waveformatex_t
*wfx
)
181 coldfire_set_macsr(EMAC_FRACTIONAL
| EMAC_SATURATE
);
184 /*clear stereo setting to avoid glitches when switching stereo->mono*/
185 s
->channel_coded
[0]=0;
186 s
->channel_coded
[1]=0;
189 s
->sample_rate
= wfx
->rate
;
190 s
->nb_channels
= wfx
->channels
;
191 s
->bit_rate
= wfx
->bitrate
;
192 s
->block_align
= wfx
->blockalign
;
194 s
->coefs
= &coefsarray
;
195 s
->frame_out
= &frame_out_buf
;
197 if (wfx
->codec_id
== ASF_CODEC_ID_WMAV1
) {
199 } else if (wfx
->codec_id
== ASF_CODEC_ID_WMAV2
) {
202 /*one of those other wma flavors that don't have GPLed decoders */
206 /* extract flag infos */
208 extradata
= wfx
->data
;
209 if (s
->version
== 1 && wfx
->datalen
>= 4) {
210 flags2
= extradata
[2] | (extradata
[3] << 8);
211 }else if (s
->version
== 2 && wfx
->datalen
>= 6){
212 flags2
= extradata
[4] | (extradata
[5] << 8);
214 s
->use_exp_vlc
= flags2
& 0x0001;
215 s
->use_bit_reservoir
= flags2
& 0x0002;
216 s
->use_variable_block_len
= flags2
& 0x0004;
218 /* compute MDCT block size */
219 if (s
->sample_rate
<= 16000){
220 s
->frame_len_bits
= 9;
221 }else if (s
->sample_rate
<= 22050 ||
222 (s
->sample_rate
<= 32000 && s
->version
== 1)){
223 s
->frame_len_bits
= 10;
225 s
->frame_len_bits
= 11;
227 s
->frame_len
= 1 << s
->frame_len_bits
;
228 if (s
-> use_variable_block_len
)
231 nb
= ((flags2
>> 3) & 3) + 1;
232 if ((s
->bit_rate
/ s
->nb_channels
) >= 32000)
236 nb_max
= s
->frame_len_bits
- BLOCK_MIN_BITS
; //max is 11-7
239 s
->nb_block_sizes
= nb
+ 1;
243 s
->nb_block_sizes
= 1;
246 /* init rate dependant parameters */
247 s
->use_noise_coding
= 1;
248 high_freq
= itofix64(s
->sample_rate
) >> 1;
251 /* if version 2, then the rates are normalized */
252 sample_rate1
= s
->sample_rate
;
255 if (sample_rate1
>= 44100)
256 sample_rate1
= 44100;
257 else if (sample_rate1
>= 22050)
258 sample_rate1
= 22050;
259 else if (sample_rate1
>= 16000)
260 sample_rate1
= 16000;
261 else if (sample_rate1
>= 11025)
262 sample_rate1
= 11025;
263 else if (sample_rate1
>= 8000)
267 fixed64 tmp
= itofix64(s
->bit_rate
);
268 fixed64 tmp2
= itofix64(s
->nb_channels
* s
->sample_rate
);
269 bps
= fixdiv64(tmp
, tmp2
);
270 fixed64 tim
= bps
* s
->frame_len
;
271 fixed64 tmpi
= fixdiv64(tim
,itofix64(8));
272 s
->byte_offset_bits
= av_log2(fixtoi64(tmpi
+0x8000)) + 2;
274 /* compute high frequency value and choose if noise coding should
277 if (s
->nb_channels
== 2)
278 bps1
= fixmul32(bps
,0x1999a);
279 if (sample_rate1
== 44100)
282 s
->use_noise_coding
= 0;
284 high_freq
= fixmul32(high_freq
,0x6666);
286 else if (sample_rate1
== 22050)
289 s
->use_noise_coding
= 0;
290 else if (bps1
>= 0xb852)
291 high_freq
= fixmul32(high_freq
,0xb333);
293 high_freq
= fixmul32(high_freq
,0x999a);
295 else if (sample_rate1
== 16000)
298 high_freq
= fixmul32(high_freq
,0x8000);
300 high_freq
= fixmul32(high_freq
,0x4ccd);
302 else if (sample_rate1
== 11025)
304 high_freq
= fixmul32(high_freq
,0xb333);
306 else if (sample_rate1
== 8000)
310 high_freq
= fixmul32(high_freq
,0x8000);
312 else if (bps
> 0xc000)
314 s
->use_noise_coding
= 0;
318 high_freq
= fixmul32(high_freq
,0xa666);
325 high_freq
= fixmul32(high_freq
,0xc000);
327 else if (bps
>= 0x999a)
329 high_freq
= fixmul32(high_freq
,0x999a);
333 high_freq
= fixmul32(high_freq
,0x8000);
337 /* compute the scale factor band sizes for each MDCT block size */
339 int a
, b
, pos
, lpos
, k
, block_len
, i
, j
, n
;
340 const uint8_t *table
;
350 for(k
= 0; k
< s
->nb_block_sizes
; ++k
)
352 block_len
= s
->frame_len
>> k
;
359 a
= wma_critical_freqs
[i
];
361 pos
= ((block_len
* 2 * a
) + (b
>> 1)) / b
;
364 s
->exponent_bands
[0][i
] = pos
- lpos
;
365 if (pos
>= block_len
)
372 s
->exponent_sizes
[0] = i
;
376 /* hardcoded tables */
378 a
= s
->frame_len_bits
- BLOCK_MIN_BITS
- k
;
381 if (s
->sample_rate
>= 44100)
382 table
= exponent_band_44100
[a
];
383 else if (s
->sample_rate
>= 32000)
384 table
= exponent_band_32000
[a
];
385 else if (s
->sample_rate
>= 22050)
386 table
= exponent_band_22050
[a
];
392 s
->exponent_bands
[k
][i
] = table
[i
];
393 s
->exponent_sizes
[k
] = n
;
401 a
= wma_critical_freqs
[i
];
403 pos
= ((block_len
* 2 * a
) + (b
<< 1)) / (4 * b
);
408 s
->exponent_bands
[k
][j
++] = pos
- lpos
;
409 if (pos
>= block_len
)
413 s
->exponent_sizes
[k
] = j
;
417 /* max number of coefs */
418 s
->coefs_end
[k
] = (s
->frame_len
- ((s
->frame_len
* 9) / 100)) >> k
;
419 /* high freq computation */
421 fixed32 tmp1
= high_freq
*2; /* high_freq is a fixed32!*/
422 fixed32 tmp2
=itofix32(s
->sample_rate
>>1);
423 s
->high_band_start
[k
] = fixtoi32( fixdiv32(tmp1
, tmp2
) * (block_len
>>1) +0x8000);
426 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
427 s->sample_rate + 0.5);*/
429 n
= s
->exponent_sizes
[k
];
436 pos
+= s
->exponent_bands
[k
][i
];
438 if (start
< s
->high_band_start
[k
])
439 start
= s
->high_band_start
[k
];
440 if (end
> s
->coefs_end
[k
])
441 end
= s
->coefs_end
[k
];
443 s
->exponent_high_bands
[k
][j
++] = end
- start
;
445 s
->exponent_high_sizes
[k
] = j
;
449 /* ffmpeg uses malloc to only allocate as many window sizes as needed.
450 * However, we're really only interested in the worst case memory usage.
451 * In the worst case you can have 5 window sizes, 128 doubling up 2048
452 * Smaller windows are handled differently.
453 * Since we don't have malloc, just statically allocate this
462 /* init MDCT windows : simple sinus window */
463 for(i
= 0; i
< s
->nb_block_sizes
; i
++)
467 n
= 1 << (s
->frame_len_bits
- i
);
470 /* this calculates 0.5/(2*n) */
471 alpha
= (1<<15)>>(s
->frame_len_bits
- i
+1);
474 fixed32 j2
= itofix32(j
) + 0x8000;
475 /*alpha between 0 and pi/2*/
476 window
[j
] = fsincos(fixmul32(j2
,alpha
)<<16, 0);
478 s
->windows
[i
] = window
;
482 s
->reset_block_lengths
= 1;
484 if (s
->use_noise_coding
)
486 /* init the noise generator */
489 s
->noise_mult
= 0x51f;
490 s
->noise_table
= noisetable_exp
;
494 s
->noise_mult
= 0xa3d;
495 /* LSP values are simply 2x the EXP values */
496 for (i
=0;i
<NOISE_TAB_SIZE
;++i
)
497 noisetable_exp
[i
] = noisetable_exp
[i
]<< 1;
498 s
->noise_table
= noisetable_exp
;
501 /* We use a lookup table computered in advance, so no need to do this*/
506 norm
= 0; // PJJ: near as makes any diff to 0!
507 for (i
=0;i
<NOISE_TAB_SIZE
;++i
)
509 seed
= seed
* 314159 + 1;
510 s
->noise_table
[i
] = itofix32((int)seed
) * norm
;
515 s
->hgain_vlc
.table
= vlcbuf4
;
516 s
->hgain_vlc
.table_allocated
= VLCBUF4SIZE
;
517 init_vlc(&s
->hgain_vlc
, HGAINVLCBITS
, sizeof(hgain_huffbits
),
518 hgain_huffbits
, 1, 1,
519 hgain_huffcodes
, 2, 2, INIT_VLC_USE_NEW_STATIC
);
525 s
->exp_vlc
.table
= vlcbuf3
;
526 s
->exp_vlc
.table_allocated
= VLCBUF3SIZE
;
528 init_vlc(&s
->exp_vlc
, EXPVLCBITS
, sizeof(scale_huffbits
),
529 scale_huffbits
, 1, 1,
530 scale_huffcodes
, 4, 4, INIT_VLC_USE_NEW_STATIC
);
534 wma_lsp_to_curve_init(s
, s
->frame_len
);
537 /* choose the VLC tables for the coefficients */
539 if (s
->sample_rate
>= 32000)
543 else if (bps1
< 0x128f6)
547 /* since the coef2 table is the biggest and that has index 2 in coef_vlcs
548 it's safe to always assign like this */
549 runtabarray
[0] = runtab_big
; runtabarray
[1] = runtab_small
;
550 levtabarray
[0] = levtab_big
; levtabarray
[1] = levtab_small
;
552 s
->coef_vlc
[0].table
= vlcbuf1
;
553 s
->coef_vlc
[0].table_allocated
= VLCBUF1SIZE
;
554 s
->coef_vlc
[1].table
= vlcbuf2
;
555 s
->coef_vlc
[1].table_allocated
= VLCBUF2SIZE
;
558 init_coef_vlc(&s
->coef_vlc
[0], &s
->run_table
[0], &s
->level_table
[0],
559 &coef_vlcs
[coef_vlc_table
* 2], 0);
560 init_coef_vlc(&s
->coef_vlc
[1], &s
->run_table
[1], &s
->level_table
[1],
561 &coef_vlcs
[coef_vlc_table
* 2 + 1], 1);
563 s
->last_superframe_len
= 0;
564 s
->last_bitoffset
= 0;
570 /* compute x^-0.25 with an exponent and mantissa table. We use linear
571 interpolation to reduce the mantissa table size at a small speed
572 expense (linear interpolation approximately doubles the number of
573 bits of precision). */
574 static inline fixed32
pow_m1_4(WMADecodeContext
*s
, fixed32 x
)
585 m
= (u
.v
>> (23 - LSP_POW_BITS
)) & ((1 << LSP_POW_BITS
) - 1);
586 /* build interpolation scale: 1 <= t < 2. */
587 t
.v
= ((u
.v
<< LSP_POW_BITS
) & ((1 << 23) - 1)) | (127 << 23);
588 a
= ((fixed32
*)s
->lsp_pow_m_table1
)[m
];
589 b
= ((fixed32
*)s
->lsp_pow_m_table2
)[m
];
591 /* lsp_pow_e_table contains 32.32 format */
592 /* TODO: Since we're unlikely have value that cover the whole
593 * IEEE754 range, we probably don't need to have all possible exponents */
595 return (lsp_pow_e_table
[e
] * (a
+ fixmul32(b
, ftofix32(t
.f
))) >>32);
598 static void wma_lsp_to_curve_init(WMADecodeContext
*s
, int frame_len
)
600 fixed32 wdel
, a
, b
, temp2
;
603 wdel
= fixdiv32(itofix32(1), itofix32(frame_len
));
604 for (i
=0; i
<frame_len
; ++i
)
606 /* TODO: can probably reuse the trig_init values here */
607 fsincos((wdel
*i
)<<15, &temp2
);
608 /* get 3 bits headroom + 1 bit from not doubleing the values */
609 s
->lsp_cos_table
[i
] = temp2
>>3;
612 /* NOTE: these two tables are needed to avoid two operations in
617 s
->lsp_pow_m_table1
= &vlcbuf3
[0];
618 s
->lsp_pow_m_table2
= &vlcbuf3
[VLCBUF3SIZE
];
620 /*double check this later*/
621 for(i
=(1 << LSP_POW_BITS
) - 1;i
>=0;i
--)
623 m
= (1 << LSP_POW_BITS
) + i
;
624 a
= pow_a_table
[ix
++]<<4;
625 ((fixed32
*)s
->lsp_pow_m_table1
)[i
] = 2 * a
- b
;
626 ((fixed32
*)s
->lsp_pow_m_table2
)[i
] = b
- a
;
632 /* NOTE: We use the same code as Vorbis here */
633 /* XXX: optimize it further with SSE/3Dnow */
634 static void wma_lsp_to_curve(WMADecodeContext
*s
,
636 fixed32
*val_max_ptr
,
641 fixed32 p
, q
, w
, v
, val_max
, temp2
;
646 /* shift by 2 now to reduce rounding error,
647 * we can renormalize right before pow_m1_4
652 w
= s
->lsp_cos_table
[i
];
654 for (j
=1;j
<NB_LSP_COEFS
;j
+=2)
656 /* w is 5.27 format, lsp is in 16.16, temp2 becomes 5.27 format */
657 temp2
= ((w
- (lsp
[j
- 1]<<11)));
659 /* q is 16.16 format, temp2 is 5.27, q becomes 16.16 */
660 q
= fixmul32b(q
, temp2
)<<4;
661 p
= fixmul32b(p
, (w
- (lsp
[j
]<<11)))<<4;
664 /* 2 in 5.27 format is 0x10000000 */
665 p
= fixmul32(p
, fixmul32b(p
, (0x10000000 - w
)))<<3;
666 q
= fixmul32(q
, fixmul32b(q
, (0x10000000 + w
)))<<3;
668 v
= (p
+ q
) >>9; /* p/q end up as 16.16 */
675 *val_max_ptr
= val_max
;
678 /* decode exponents coded with LSP coefficients (same idea as Vorbis)
679 * only used for low bitrate (< 16kbps) files
681 static void decode_exp_lsp(WMADecodeContext
*s
, int ch
)
683 fixed32 lsp_coefs
[NB_LSP_COEFS
];
686 for (i
= 0; i
< NB_LSP_COEFS
; ++i
)
688 if (i
== 0 || i
>= 8)
689 val
= get_bits(&s
->gb
, 3);
691 val
= get_bits(&s
->gb
, 4);
692 lsp_coefs
[i
] = lsp_codebook
[i
][val
];
697 &s
->max_exponent
[ch
],
702 /* decode exponents coded with VLC codes - used for bitrate >= 32kbps*/
703 static int decode_exp_vlc(WMADecodeContext
*s
, int ch
)
705 int last_exp
, n
, code
;
706 const uint16_t *ptr
, *band_ptr
;
707 fixed32 v
, max_scale
;
710 /*accommodate the 60 negative indices */
711 const fixed32
*pow_10_to_yover16_ptr
= &pow_10_to_yover16
[61];
713 band_ptr
= s
->exponent_bands
[s
->frame_len_bits
- s
->block_len_bits
];
715 q
= s
->exponents
[ch
];
716 q_end
= q
+ s
->block_len
;
720 if (s
->version
== 1) //wmav1 only
722 last_exp
= get_bits(&s
->gb
, 5) + 10;
724 v
= pow_10_to_yover16_ptr
[last_exp
];
732 } while ((n
-= 4) > 0);
739 code
= get_vlc2(&s
->gb
, s
->exp_vlc
.table
, EXPVLCBITS
, EXPMAX
);
744 /* NOTE: this offset is the same as MPEG4 AAC ! */
745 last_exp
+= code
- 60;
747 v
= pow_10_to_yover16_ptr
[last_exp
];
758 } while ((n
-= 4) > 0);
761 s
->max_exponent
[ch
] = max_scale
;
765 /* return 0 if OK. return 1 if last block of frame. return -1 if
766 unrecorrable error. */
767 static int wma_decode_block(WMADecodeContext
*s
)
769 int n
, v
, a
, ch
, code
, bsize
;
770 int coef_nb_bits
, total_gain
;
771 int nb_coefs
[MAX_CHANNELS
];
774 /*DEBUGF("***decode_block: %d (%d samples of %d in frame)\n", s->block_num, s->block_len, s->frame_len);*/
776 /* compute current block length */
777 if (s
->use_variable_block_len
)
779 n
= av_log2(s
->nb_block_sizes
- 1) + 1;
781 if (s
->reset_block_lengths
)
783 s
->reset_block_lengths
= 0;
784 v
= get_bits(&s
->gb
, n
);
785 if (v
>= s
->nb_block_sizes
)
789 s
->prev_block_len_bits
= s
->frame_len_bits
- v
;
790 v
= get_bits(&s
->gb
, n
);
791 if (v
>= s
->nb_block_sizes
)
795 s
->block_len_bits
= s
->frame_len_bits
- v
;
799 /* update block lengths */
800 s
->prev_block_len_bits
= s
->block_len_bits
;
801 s
->block_len_bits
= s
->next_block_len_bits
;
803 v
= get_bits(&s
->gb
, n
);
805 if (v
>= s
->nb_block_sizes
)
807 // rb->splash(HZ*4, "v was %d", v); //5, 7
808 return -4; //this is it
811 //rb->splash(HZ, "passed v block (%d)!", v);
813 s
->next_block_len_bits
= s
->frame_len_bits
- v
;
817 /* fixed block len */
818 s
->next_block_len_bits
= s
->frame_len_bits
;
819 s
->prev_block_len_bits
= s
->frame_len_bits
;
820 s
->block_len_bits
= s
->frame_len_bits
;
822 /* now check if the block length is coherent with the frame length */
823 s
->block_len
= 1 << s
->block_len_bits
;
825 if ((s
->block_pos
+ s
->block_len
) > s
->frame_len
)
827 return -5; //oddly 32k sample from tracker fails here
830 if (s
->nb_channels
== 2)
832 s
->ms_stereo
= get_bits1(&s
->gb
);
835 for (ch
= 0; ch
< s
->nb_channels
; ++ch
)
837 a
= get_bits1(&s
->gb
);
838 s
->channel_coded
[ch
] = a
;
841 /* if no channel coded, no need to go further */
842 /* XXX: fix potential framing problems */
848 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
850 /* read total gain and extract corresponding number of bits for
851 coef escape coding */
855 a
= get_bits(&s
->gb
, 7);
865 else if (total_gain
< 32)
867 else if (total_gain
< 40)
869 else if (total_gain
< 45)
874 /* compute number of coefficients */
875 n
= s
->coefs_end
[bsize
] - s
->coefs_start
;
877 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
882 if (s
->use_noise_coding
)
885 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
887 if (s
->channel_coded
[ch
])
890 n
= s
->exponent_high_sizes
[bsize
];
893 a
= get_bits1(&s
->gb
);
894 s
->high_band_coded
[ch
][i
] = a
;
895 /* if noise coding, the coefficients are not transmitted */
897 nb_coefs
[ch
] -= s
->exponent_high_bands
[bsize
][i
];
901 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
903 if (s
->channel_coded
[ch
])
907 n
= s
->exponent_high_sizes
[bsize
];
908 val
= (int)0x80000000;
911 if (s
->high_band_coded
[ch
][i
])
913 if (val
== (int)0x80000000)
915 val
= get_bits(&s
->gb
, 7) - 19;
919 //code = get_vlc(&s->gb, &s->hgain_vlc);
920 code
= get_vlc2(&s
->gb
, s
->hgain_vlc
.table
, HGAINVLCBITS
, HGAINMAX
);
927 s
->high_band_values
[ch
][i
] = val
;
934 /* exponents can be reused in short blocks. */
935 if ((s
->block_len_bits
== s
->frame_len_bits
) || get_bits1(&s
->gb
))
937 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
939 if (s
->channel_coded
[ch
])
943 if (decode_exp_vlc(s
, ch
) < 0)
950 decode_exp_lsp(s
, ch
);
952 s
->exponents_bsize
[ch
] = bsize
;
957 /* parse spectral coefficients : just RLE encoding */
958 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
960 if (s
->channel_coded
[ch
])
963 int level
, run
, sign
, tindex
;
965 const int16_t *level_table
, *run_table
;
967 /* special VLC tables are used for ms stereo because
968 there is potentially less energy there */
969 tindex
= (ch
== 1 && s
->ms_stereo
);
970 coef_vlc
= &s
->coef_vlc
[tindex
];
971 run_table
= s
->run_table
[tindex
];
972 level_table
= s
->level_table
[tindex
];
974 ptr
= &s
->coefs1
[ch
][0];
975 eptr
= ptr
+ nb_coefs
[ch
];
976 memset(ptr
, 0, s
->block_len
* sizeof(int16_t));
980 code
= get_vlc2(&s
->gb
, coef_vlc
->table
, VLCBITS
, VLCMAX
);
994 level
= get_bits(&s
->gb
, coef_nb_bits
);
995 /* NOTE: this is rather suboptimal. reading
996 block_len_bits would be better */
997 run
= get_bits(&s
->gb
, s
->frame_len_bits
);
1002 run
= run_table
[code
];
1003 level
= level_table
[code
];
1005 sign
= get_bits1(&s
->gb
);
1016 /* NOTE: EOB can be omitted */
1021 if (s
->version
== 1 && s
->nb_channels
>= 2)
1023 align_get_bits(&s
->gb
);
1028 int n4
= s
->block_len
>> 1;
1031 mdct_norm
= 0x10000>>(s
->block_len_bits
-1);
1033 if (s
->version
== 1)
1035 mdct_norm
*= fixtoi32(fixsqrt32(itofix32(n4
)));
1040 /* finally compute the MDCT coefficients */
1041 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1043 if (s
->channel_coded
[ch
])
1047 fixed32
*coefs
, atemp
;
1050 fixed32 noise
, temp1
, temp2
, mult2
;
1051 int i
, j
, n
, n1
, last_high_band
, esize
;
1052 fixed32 exp_power
[HIGH_BAND_MAX_SIZE
];
1054 //total_gain, coefs1, mdctnorm are lossless
1056 coefs1
= s
->coefs1
[ch
];
1057 exponents
= s
->exponents
[ch
];
1058 esize
= s
->exponents_bsize
[ch
];
1059 coefs
= (*(s
->coefs
))[ch
];
1063 * The calculation of coefs has a shift right by 2 built in. This
1064 * prepares samples for the Tremor IMDCT which uses a slightly
1065 * different fixed format then the ffmpeg one. If the old ffmpeg
1066 * imdct is used, each shift storing into coefs should be reduced
1068 * See SVN logs for details.
1072 if (s
->use_noise_coding
)
1074 /*This case is only used for low bitrates (typically less then 32kbps)*/
1076 /*TODO: mult should be converted to 32 bit to speed up noise coding*/
1078 mult
= fixdiv64(pow_table
[total_gain
+20],Fixed32To64(s
->max_exponent
[ch
]));
1079 mult
= mult
* mdct_norm
;
1082 /* very low freqs : noise */
1083 for(i
= 0;i
< s
->coefs_start
; ++i
)
1085 *coefs
++ = fixmul32( (fixmul32(s
->noise_table
[s
->noise_index
],
1086 exponents
[i
<<bsize
>>esize
])>>4),Fixed32From64(mult1
)) >>2;
1087 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1090 n1
= s
->exponent_high_sizes
[bsize
];
1092 /* compute power of high bands */
1093 exponents
= s
->exponents
[ch
] +(s
->high_band_start
[bsize
]<<bsize
);
1094 last_high_band
= 0; /* avoid warning */
1097 n
= s
->exponent_high_bands
[s
->frame_len_bits
-
1098 s
->block_len_bits
][j
];
1099 if (s
->high_band_coded
[ch
][j
])
1103 for(i
= 0;i
< n
; ++i
)
1105 /*v is normalized later on so its fixed format is irrelevant*/
1106 v
= exponents
[i
<<bsize
>>esize
]>>4;
1107 e2
+= fixmul32(v
, v
)>>3;
1109 exp_power
[j
] = e2
/n
; /*n is an int...*/
1112 exponents
+= n
<<bsize
;
1115 /* main freqs and high freqs */
1116 exponents
= s
->exponents
[ch
] + (s
->coefs_start
<<bsize
);
1121 n
= s
->high_band_start
[bsize
] -
1126 n
= s
->exponent_high_bands
[s
->frame_len_bits
-
1127 s
->block_len_bits
][j
];
1129 if (j
>= 0 && s
->high_band_coded
[ch
][j
])
1131 /* use noise with specified power */
1132 fixed32 tmp
= fixdiv32(exp_power
[j
],exp_power
[last_high_band
]);
1134 /*mult1 is 48.16, pow_table is 48.16*/
1135 mult1
= fixmul32(fixsqrt32(tmp
),
1136 pow_table
[s
->high_band_values
[ch
][j
]+20]) >> 16;
1138 /*this step has a fairly high degree of error for some reason*/
1139 mult1
= fixdiv64(mult1
,fixmul32(s
->max_exponent
[ch
],s
->noise_mult
));
1140 mult1
= mult1
*mdct_norm
>>PRECISION
;
1141 for(i
= 0;i
< n
; ++i
)
1143 noise
= s
->noise_table
[s
->noise_index
];
1144 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1145 *coefs
++ = fixmul32((fixmul32(exponents
[i
<<bsize
>>esize
],noise
)>>4),
1146 Fixed32From64(mult1
)) >>2;
1149 exponents
+= n
<<bsize
;
1153 /* coded values + small noise */
1154 for(i
= 0;i
< n
; ++i
)
1156 noise
= s
->noise_table
[s
->noise_index
];
1157 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1159 /*don't forget to renormalize the noise*/
1160 temp1
= (((int32_t)*coefs1
++)<<16) + (noise
>>4);
1161 temp2
= fixmul32(exponents
[i
<<bsize
>>esize
], mult
>>18);
1162 *coefs
++ = fixmul32(temp1
, temp2
);
1164 exponents
+= n
<<bsize
;
1168 /* very high freqs : noise */
1169 n
= s
->block_len
- s
->coefs_end
[bsize
];
1170 mult2
= fixmul32(mult
>>16,exponents
[((-1<<bsize
))>>esize
]) ;
1171 for (i
= 0; i
< n
; ++i
)
1173 /*renormalize the noise product and then reduce to 14.18 precison*/
1174 *coefs
++ = fixmul32(s
->noise_table
[s
->noise_index
],mult2
) >>6;
1176 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1181 /*Noise coding not used, simply convert from exp to fixed representation*/
1183 fixed32 mult3
= (fixed32
)(fixdiv64(pow_table
[total_gain
+20],
1184 Fixed32To64(s
->max_exponent
[ch
])));
1185 mult3
= fixmul32(mult3
, mdct_norm
);
1187 /*zero the first 3 coefficients for WMA V1, does nothing otherwise*/
1188 for(i
=0; i
<s
->coefs_start
; i
++)
1193 /* XXX: optimize more, unrolling this loop in asm
1194 might be a good idea */
1196 for(i
= 0;i
< n
; ++i
)
1198 /*ffmpeg imdct needs 15.17, while tremor 14.18*/
1199 atemp
= (coefs1
[i
] * mult3
)>>2;
1200 *coefs
++=fixmul32(atemp
,exponents
[i
<<bsize
>>esize
]);
1202 n
= s
->block_len
- s
->coefs_end
[bsize
];
1203 memset(coefs
, 0, n
*sizeof(fixed32
));
1210 if (s
->ms_stereo
&& s
->channel_coded
[1])
1214 fixed32 (*coefs
)[MAX_CHANNELS
][BLOCK_MAX_SIZE
] = (s
->coefs
);
1216 /* nominal case for ms stereo: we do it before mdct */
1217 /* no need to optimize this case because it should almost
1219 if (!s
->channel_coded
[0])
1221 memset((*(s
->coefs
))[0], 0, sizeof(fixed32
) * s
->block_len
);
1222 s
->channel_coded
[0] = 1;
1225 for(i
= 0; i
< s
->block_len
; ++i
)
1229 (*coefs
)[0][i
] = a
+ b
;
1230 (*coefs
)[1][i
] = a
- b
;
1234 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1236 /* BLOCK_MAX_SIZE is 2048 (samples) and MAX_CHANNELS is 2. */
1237 static uint32_t scratch_buf
[BLOCK_MAX_SIZE
* MAX_CHANNELS
] IBSS_ATTR MEM_ALIGN_ATTR
;
1238 if (s
->channel_coded
[ch
])
1242 n4
= s
->block_len
>>1;
1244 ff_imdct_calc((s
->frame_len_bits
- bsize
+ 1),
1248 /* add in the frame */
1249 index
= (s
->frame_len
/ 2) + s
->block_pos
- n4
;
1250 wma_window(s
, scratch_buf
, &((*s
->frame_out
)[ch
][index
]));
1254 /* specific fast case for ms-stereo : add to second
1255 channel if it is not coded */
1256 if (s
->ms_stereo
&& !s
->channel_coded
[1])
1258 wma_window(s
, scratch_buf
, &((*s
->frame_out
)[1][index
]));
1263 /* update block number */
1265 s
->block_pos
+= s
->block_len
;
1266 if (s
->block_pos
>= s
->frame_len
)
1276 /* decode a frame of frame_len samples */
1277 static int wma_decode_frame(WMADecodeContext
*s
)
1281 /* read each block */
1288 ret
= wma_decode_block(s
);
1292 DEBUGF("wma_decode_block failed with code %d\n", ret
);
1304 /* Initialise the superframe decoding */
1306 int wma_decode_superframe_init(WMADecodeContext
* s
,
1307 const uint8_t *buf
, /*input*/
1312 s
->last_superframe_len
= 0;
1316 s
->current_frame
= 0;
1318 init_get_bits(&s
->gb
, buf
, buf_size
*8);
1320 if (s
->use_bit_reservoir
)
1322 /* read super frame header */
1323 skip_bits(&s
->gb
, 4); /* super frame index */
1324 s
->nb_frames
= get_bits(&s
->gb
, 4);
1326 if (s
->last_superframe_len
== 0)
1328 else if (s
->nb_frames
== 0)
1331 s
->bit_offset
= get_bits(&s
->gb
, s
->byte_offset_bits
+ 3);
1340 /* Decode a single frame in the current superframe - return -1 if
1341 there was a decoding error, or the number of samples decoded.
1344 int wma_decode_superframe_frame(WMADecodeContext
* s
,
1345 const uint8_t *buf
, /*input*/
1352 for(ch
= 0; ch
< s
->nb_channels
; ch
++)
1353 memmove(&((*s
->frame_out
)[ch
][0]),
1354 &((*s
->frame_out
)[ch
][s
->frame_len
]),
1355 s
->frame_len
* sizeof(fixed32
));
1357 if ((s
->use_bit_reservoir
) && (s
->current_frame
== 0))
1359 if (s
->last_superframe_len
> 0)
1361 /* add s->bit_offset bits to last frame */
1362 if ((s
->last_superframe_len
+ ((s
->bit_offset
+ 7) >> 3)) >
1363 MAX_CODED_SUPERFRAME_SIZE
)
1365 DEBUGF("superframe size too large error\n");
1368 q
= s
->last_superframe
+ s
->last_superframe_len
;
1369 len
= s
->bit_offset
;
1372 *q
++ = (get_bits
)(&s
->gb
, 8);
1377 *q
++ = (get_bits
)(&s
->gb
, len
) << (8 - len
);
1380 /* XXX: s->bit_offset bits into last frame */
1381 init_get_bits(&s
->gb
, s
->last_superframe
, MAX_CODED_SUPERFRAME_SIZE
*8);
1382 /* skip unused bits */
1383 if (s
->last_bitoffset
> 0)
1384 skip_bits(&s
->gb
, s
->last_bitoffset
);
1386 /* this frame is stored in the last superframe and in the
1388 if (wma_decode_frame(s
) < 0)
1395 /* read each frame starting from s->bit_offset */
1396 pos
= s
->bit_offset
+ 4 + 4 + s
->byte_offset_bits
+ 3;
1397 init_get_bits(&s
->gb
, buf
+ (pos
>> 3), (MAX_CODED_SUPERFRAME_SIZE
- (pos
>> 3))*8);
1400 skip_bits(&s
->gb
, len
);
1402 s
->reset_block_lengths
= 1;
1405 /* If we haven't decoded a frame yet, do it now */
1408 if (wma_decode_frame(s
) < 0)
1416 if ((s
->use_bit_reservoir
) && (s
->current_frame
== s
->nb_frames
))
1418 /* we copy the end of the frame in the last frame buffer */
1419 pos
= get_bits_count(&s
->gb
) + ((s
->bit_offset
+ 4 + 4 + s
->byte_offset_bits
+ 3) & ~7);
1420 s
->last_bitoffset
= pos
& 7;
1422 len
= buf_size
- pos
;
1423 if (len
> MAX_CODED_SUPERFRAME_SIZE
|| len
< 0)
1425 DEBUGF("superframe size too large error after decoding\n");
1428 s
->last_superframe_len
= len
;
1429 memcpy(s
->last_superframe
, buf
+ pos
, len
);
1432 return s
->frame_len
;
1435 /* when error, we reset the bit reservoir */
1437 s
->last_superframe_len
= 0;