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>
30 #include "bitstream.h"
33 #define VLCBITS 7 /*7 is the lowest without glitching*/
34 #define VLCMAX ((22+VLCBITS-1)/VLCBITS)
37 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS)
39 #define HGAINVLCBITS 9
40 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS)
43 typedef struct CoefVLCTable
45 int n
; /* total number of codes */
46 const uint32_t *huffcodes
; /* VLC bit values */
47 const uint8_t *huffbits
; /* VLC bit size */
48 const uint16_t *levels
; /* table to build run/level tables */
52 static void wma_lsp_to_curve_init(WMADecodeContext
*s
, int frame_len
);
54 fixed32 coefsarray
[MAX_CHANNELS
][BLOCK_MAX_SIZE
] IBSS_ATTR
;
56 //static variables that replace malloced stuff
57 fixed32 stat0
[2048], stat1
[1024], stat2
[512], stat3
[256], stat4
[128]; //these are the MDCT reconstruction windows
59 uint16_t *runtabarray
[2], *levtabarray
[2]; //these are VLC lookup tables
61 uint16_t runtab0
[1336], runtab1
[1336], levtab0
[1336], levtab1
[1336]; //these could be made smaller since only one can be 1336
63 #define VLCBUF1SIZE 4598
64 #define VLCBUF2SIZE 3574
65 #define VLCBUF3SIZE 360
66 #define VLCBUF4SIZE 540
68 /*putting these in IRAM actually makes PP slower*/
70 VLC_TYPE vlcbuf1
[VLCBUF1SIZE
][2];
71 VLC_TYPE vlcbuf2
[VLCBUF2SIZE
][2];
72 VLC_TYPE vlcbuf3
[VLCBUF3SIZE
][2];
73 VLC_TYPE vlcbuf4
[VLCBUF4SIZE
][2];
77 #include "wmadata.h" // PJJ
82 * Helper functions for wma_window.
89 void vector_fmul_add_add(fixed32
*dst
, const fixed32
*data
,
90 const fixed32
*window
, int n
)
92 /* Block sizes are always power of two */
95 "ldmia %[d]!, {r0, r1};"
96 "ldmia %[w]!, {r4, r5};"
97 /* consume the first data and window value so we can use those
99 "smull r8, r9, r0, r4;"
100 "ldmia %[dst], {r0, r4};"
101 "add r0, r0, r9, lsl #1;" /* *dst=*dst+(r9<<1)*/
102 "smull r8, r9, r1, r5;"
103 "add r1, r4, r9, lsl #1;"
104 "stmia %[dst]!, {r0, r1};"
105 "subs %[n], %[n], #2;"
107 : [d
] "+r" (data
), [w
] "+r" (window
), [dst
] "+r" (dst
), [n
] "+r" (n
)
108 : : "r0", "r1", "r4", "r5", "r8", "r9", "memory", "cc");
112 void vector_fmul_reverse(fixed32
*dst
, const fixed32
*src0
, const fixed32
*src1
,
115 /* Block sizes are always power of two */
117 "add %[s1], %[s1], %[n], lsl #2;"
119 "ldmia %[s0]!, {r0, r1};"
120 "ldmdb %[s1]!, {r4, r5};"
121 "smull r8, r9, r0, r5;"
122 "mov r0, r9, lsl #1;"
123 "smull r8, r9, r1, r4;"
124 "mov r1, r9, lsl #1;"
125 "stmia %[dst]!, {r0, r1};"
126 "subs %[n], %[n], #2;"
128 : [s0
] "+r" (src0
), [s1
] "+r" (src1
), [dst
] "+r" (dst
), [n
] "+r" (len
)
129 : : "r0", "r1", "r4", "r5", "r8", "r9", "memory", "cc");
132 #elif defined(CPU_COLDFIRE)
135 void vector_fmul_add_add(fixed32
*dst
, const fixed32
*data
,
136 const fixed32
*window
, int n
)
138 /* Block sizes are always power of two. Smallest block is always way bigger
142 "movem.l (%[d]), %%d0-%%d3;"
143 "movem.l (%[w]), %%d4-%%d5/%%a0-%%a1;"
144 "mac.l %%d0, %%d4, %%acc0;"
145 "mac.l %%d1, %%d5, %%acc1;"
146 "mac.l %%d2, %%a0, %%acc2;"
147 "mac.l %%d3, %%a1, %%acc3;"
148 "lea.l (16, %[d]), %[d];"
149 "lea.l (16, %[w]), %[w];"
150 "movclr.l %%acc0, %%d0;"
151 "movclr.l %%acc1, %%d1;"
152 "movclr.l %%acc2, %%d2;"
153 "movclr.l %%acc3, %%d3;"
154 "add.l %%d0, (%[dst])+;"
155 "add.l %%d1, (%[dst])+;"
156 "add.l %%d2, (%[dst])+;"
157 "add.l %%d3, (%[dst])+;"
160 : [d
] "+a" (data
), [w
] "+a" (window
), [dst
] "+a" (dst
), [n
] "+d" (n
)
161 : : "d0", "d1", "d2", "d3", "d4", "d5", "a0", "a1", "memory", "cc");
165 void vector_fmul_reverse(fixed32
*dst
, const fixed32
*src0
, const fixed32
*src1
,
168 /* Block sizes are always power of two. Smallest block is always way bigger
171 "lea.l (-16, %[s1], %[n]*4), %[s1];"
173 "movem.l (%[s0]), %%d0-%%d3;"
174 "movem.l (%[s1]), %%d4-%%d5/%%a0-%%a1;"
175 "mac.l %%d0, %%a1, %%acc0;"
176 "mac.l %%d1, %%a0, %%acc1;"
177 "mac.l %%d2, %%d5, %%acc2;"
178 "mac.l %%d3, %%d4, %%acc3;"
179 "lea.l (16, %[s0]), %[s0];"
180 "lea.l (-16, %[s1]), %[s1];"
181 "movclr.l %%acc0, %%d0;"
182 "movclr.l %%acc1, %%d1;"
183 "movclr.l %%acc2, %%d2;"
184 "movclr.l %%acc3, %%d3;"
185 "movem.l %%d0-%%d3, (%[dst]);"
186 "lea.l (16, %[dst]), %[dst];"
189 : [s0
] "+a" (src0
), [s1
] "+a" (src1
), [dst
] "+a" (dst
), [n
] "+d" (len
)
190 : : "d0", "d1", "d2", "d3", "d4", "d5", "a0", "a1", "memory", "cc");
195 static inline void vector_fmul_add_add(fixed32
*dst
, const fixed32
*src0
, const fixed32
*src1
, int len
){
198 dst
[i
] = fixmul32b(src0
[i
], src1
[i
]) + dst
[i
];
201 static inline void vector_fmul_reverse(fixed32
*dst
, const fixed32
*src0
, const fixed32
*src1
, int len
){
205 dst
[i
] = fixmul32b(src0
[i
], src1
[-i
]);
211 * Apply MDCT window and add into output.
213 * We ensure that when the windows overlap their squared sum
214 * is always 1 (MDCT reconstruction rule).
216 * The Vorbis I spec has a great diagram explaining this process.
217 * See section 1.3.2.3 of http://xiph.org/vorbis/doc/Vorbis_I_spec.html
219 static void wma_window(WMADecodeContext
*s
, fixed32
*in
, fixed32
*out
)
221 //float *in = s->output;
222 int block_len
, bsize
, n
;
225 /*previous block was larger, so we'll use the size of the current block to set the window size*/
226 if (s
->block_len_bits
<= s
->prev_block_len_bits
) {
227 block_len
= s
->block_len
;
228 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
230 vector_fmul_add_add(out
, in
, s
->windows
[bsize
], block_len
);
233 /*previous block was smaller or the same size, so use it's size to set the window length*/
234 block_len
= 1 << s
->prev_block_len_bits
;
235 /*find the middle of the two overlapped blocks, this will be the first overlapped sample*/
236 n
= (s
->block_len
- block_len
) / 2;
237 bsize
= s
->frame_len_bits
- s
->prev_block_len_bits
;
239 vector_fmul_add_add(out
+n
, in
+n
, s
->windows
[bsize
], block_len
);
241 memcpy(out
+n
+block_len
, in
+n
+block_len
, n
*sizeof(fixed32
));
243 /* Advance to the end of the current block and prepare to window it for the next block.
244 * Since the window function needs to be reversed, we do it backwards starting with the
245 * last sample and moving towards the first
251 if (s
->block_len_bits
<= s
->next_block_len_bits
) {
252 block_len
= s
->block_len
;
253 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
255 vector_fmul_reverse(out
, in
, s
->windows
[bsize
], block_len
);
258 block_len
= 1 << s
->next_block_len_bits
;
259 n
= (s
->block_len
- block_len
) / 2;
260 bsize
= s
->frame_len_bits
- s
->next_block_len_bits
;
262 memcpy(out
, in
, n
*sizeof(fixed32
));
264 vector_fmul_reverse(out
+n
, in
+n
, s
->windows
[bsize
], block_len
);
266 memset(out
+n
+block_len
, 0, n
*sizeof(fixed32
));
273 /* XXX: use same run/length optimization as mpeg decoders */
274 static void init_coef_vlc(VLC
*vlc
,
275 uint16_t **prun_table
, uint16_t **plevel_table
,
276 const CoefVLCTable
*vlc_table
, int tab
)
278 int n
= vlc_table
->n
;
279 const uint8_t *table_bits
= vlc_table
->huffbits
;
280 const uint32_t *table_codes
= vlc_table
->huffcodes
;
281 const uint16_t *levels_table
= vlc_table
->levels
;
282 uint16_t *run_table
, *level_table
;
287 init_vlc(vlc
, VLCBITS
, n
, table_bits
, 1, 1, table_codes
, 4, 4, 0);
289 run_table
= runtabarray
[tab
];
290 level_table
= levtabarray
[tab
];
301 level_table
[i
] = level
;
306 *prun_table
= run_table
;
307 *plevel_table
= level_table
;
310 int wma_decode_init(WMADecodeContext
* s
, asf_waveformatex_t
*wfx
)
312 //WMADecodeContext *s = avctx->priv_data;
313 int i
, flags1
, flags2
;
323 coldfire_set_macsr(EMAC_FRACTIONAL
| EMAC_SATURATE
);
326 s
->sample_rate
= wfx
->rate
;
327 s
->nb_channels
= wfx
->channels
;
328 s
->bit_rate
= wfx
->bitrate
;
329 s
->block_align
= wfx
->blockalign
;
331 s
->coefs
= &coefsarray
;
333 if (wfx
->codec_id
== ASF_CODEC_ID_WMAV1
) {
335 } else if (wfx
->codec_id
== ASF_CODEC_ID_WMAV2
) {
338 /*one of those other wma flavors that don't have GPLed decoders */
342 /* extract flag infos */
345 extradata
= wfx
->data
;
346 if (s
->version
== 1 && wfx
->datalen
>= 4) {
347 flags1
= extradata
[0] | (extradata
[1] << 8);
348 flags2
= extradata
[2] | (extradata
[3] << 8);
349 }else if (s
->version
== 2 && wfx
->datalen
>= 6){
350 flags1
= extradata
[0] | (extradata
[1] << 8) |
351 (extradata
[2] << 16) | (extradata
[3] << 24);
352 flags2
= extradata
[4] | (extradata
[5] << 8);
354 s
->use_exp_vlc
= flags2
& 0x0001;
355 s
->use_bit_reservoir
= flags2
& 0x0002;
356 s
->use_variable_block_len
= flags2
& 0x0004;
358 /* compute MDCT block size */
359 if (s
->sample_rate
<= 16000){
360 s
->frame_len_bits
= 9;
361 }else if (s
->sample_rate
<= 22050 ||
362 (s
->sample_rate
<= 32000 && s
->version
== 1)){
363 s
->frame_len_bits
= 10;
365 s
->frame_len_bits
= 11;
367 s
->frame_len
= 1 << s
->frame_len_bits
;
368 if (s
-> use_variable_block_len
)
371 nb
= ((flags2
>> 3) & 3) + 1;
372 if ((s
->bit_rate
/ s
->nb_channels
) >= 32000)
376 nb_max
= s
->frame_len_bits
- BLOCK_MIN_BITS
; //max is 11-7
379 s
->nb_block_sizes
= nb
+ 1;
383 s
->nb_block_sizes
= 1;
386 /* init rate dependant parameters */
387 s
->use_noise_coding
= 1;
388 high_freq
= itofix64(s
->sample_rate
) >> 1;
391 /* if version 2, then the rates are normalized */
392 sample_rate1
= s
->sample_rate
;
395 if (sample_rate1
>= 44100)
396 sample_rate1
= 44100;
397 else if (sample_rate1
>= 22050)
398 sample_rate1
= 22050;
399 else if (sample_rate1
>= 16000)
400 sample_rate1
= 16000;
401 else if (sample_rate1
>= 11025)
402 sample_rate1
= 11025;
403 else if (sample_rate1
>= 8000)
407 fixed64 tmp
= itofix64(s
->bit_rate
);
408 fixed64 tmp2
= itofix64(s
->nb_channels
* s
->sample_rate
);
409 bps
= fixdiv64(tmp
, tmp2
);
410 fixed64 tim
= bps
* s
->frame_len
;
411 fixed64 tmpi
= fixdiv64(tim
,itofix64(8));
412 s
->byte_offset_bits
= av_log2(fixtoi64(tmpi
+0x8000)) + 2;
414 /* compute high frequency value and choose if noise coding should
417 if (s
->nb_channels
== 2)
418 bps1
= fixmul32(bps
,0x1999a);
419 if (sample_rate1
== 44100)
422 s
->use_noise_coding
= 0;
424 high_freq
= fixmul32(high_freq
,0x6666);
426 else if (sample_rate1
== 22050)
429 s
->use_noise_coding
= 0;
430 else if (bps1
>= 0xb852)
431 high_freq
= fixmul32(high_freq
,0xb333);
433 high_freq
= fixmul32(high_freq
,0x999a);
435 else if (sample_rate1
== 16000)
438 high_freq
= fixmul32(high_freq
,0x8000);
440 high_freq
= fixmul32(high_freq
,0x4ccd);
442 else if (sample_rate1
== 11025)
444 high_freq
= fixmul32(high_freq
,0xb333);
446 else if (sample_rate1
== 8000)
450 high_freq
= fixmul32(high_freq
,0x8000);
452 else if (bps
> 0xc000)
454 s
->use_noise_coding
= 0;
458 high_freq
= fixmul32(high_freq
,0xa666);
465 high_freq
= fixmul32(high_freq
,0xc000);
467 else if (bps
>= 0x999a)
469 high_freq
= fixmul32(high_freq
,0x999a);
473 high_freq
= fixmul32(high_freq
,0x8000);
477 /* compute the scale factor band sizes for each MDCT block size */
479 int a
, b
, pos
, lpos
, k
, block_len
, i
, j
, n
;
480 const uint8_t *table
;
490 for(k
= 0; k
< s
->nb_block_sizes
; ++k
)
492 block_len
= s
->frame_len
>> k
;
499 a
= wma_critical_freqs
[i
];
501 pos
= ((block_len
* 2 * a
) + (b
>> 1)) / b
;
504 s
->exponent_bands
[0][i
] = pos
- lpos
;
505 if (pos
>= block_len
)
512 s
->exponent_sizes
[0] = i
;
516 /* hardcoded tables */
518 a
= s
->frame_len_bits
- BLOCK_MIN_BITS
- k
;
521 if (s
->sample_rate
>= 44100)
522 table
= exponent_band_44100
[a
];
523 else if (s
->sample_rate
>= 32000)
524 table
= exponent_band_32000
[a
];
525 else if (s
->sample_rate
>= 22050)
526 table
= exponent_band_22050
[a
];
532 s
->exponent_bands
[k
][i
] = table
[i
];
533 s
->exponent_sizes
[k
] = n
;
541 a
= wma_critical_freqs
[i
];
543 pos
= ((block_len
* 2 * a
) + (b
<< 1)) / (4 * b
);
548 s
->exponent_bands
[k
][j
++] = pos
- lpos
;
549 if (pos
>= block_len
)
553 s
->exponent_sizes
[k
] = j
;
557 /* max number of coefs */
558 s
->coefs_end
[k
] = (s
->frame_len
- ((s
->frame_len
* 9) / 100)) >> k
;
559 /* high freq computation */
561 fixed32 tmp1
= high_freq
*2; /* high_freq is a fixed32!*/
562 fixed32 tmp2
=itofix32(s
->sample_rate
>>1);
563 s
->high_band_start
[k
] = fixtoi32( fixdiv32(tmp1
, tmp2
) * (block_len
>>1) +0x8000);
566 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
567 s->sample_rate + 0.5);*/
569 n
= s
->exponent_sizes
[k
];
576 pos
+= s
->exponent_bands
[k
][i
];
578 if (start
< s
->high_band_start
[k
])
579 start
= s
->high_band_start
[k
];
580 if (end
> s
->coefs_end
[k
])
581 end
= s
->coefs_end
[k
];
583 s
->exponent_high_bands
[k
][j
++] = end
- start
;
585 s
->exponent_high_sizes
[k
] = j
;
591 for(i
= 0; i
< s
->nb_block_sizes
; ++i
)
593 ff_mdct_init(&s
->mdct_ctx
[i
], s
->frame_len_bits
- i
+ 1, 1);
596 /*ffmpeg uses malloc to only allocate as many window sizes as needed. However, we're really only interested in the worst case memory usage.
597 * In the worst case you can have 5 window sizes, 128 doubling up 2048
598 * Smaller windows are handled differently.
599 * Since we don't have malloc, just statically allocate this
608 /* init MDCT windows : simple sinus window */
609 for(i
= 0; i
< s
->nb_block_sizes
; i
++)
613 n
= 1 << (s
->frame_len_bits
- i
);
614 //window = av_malloc(sizeof(fixed32) * n);
617 //fixed32 n2 = itofix32(n<<1); //2x the window length
618 //alpha = fixdiv32(M_PI_F, n2); //PI / (2x Window length) == PI<<(s->frame_len_bits - i+1)
620 //alpha = M_PI_F>>(s->frame_len_bits - i+1);
621 alpha
= (1<<15)>>(s
->frame_len_bits
- i
+1); /* this calculates 0.5/(2*n) */
624 fixed32 j2
= itofix32(j
) + 0x8000;
625 window
[j
] = fsincos(fixmul32(j2
,alpha
)<<16, 0); //alpha between 0 and pi/2
628 s
->windows
[i
] = window
;
632 s
->reset_block_lengths
= 1;
634 if (s
->use_noise_coding
)
636 /* init the noise generator */
639 s
->noise_mult
= 0x51f;
640 s
->noise_table
= noisetable_exp
;
644 s
->noise_mult
= 0xa3d;
645 /* LSP values are simply 2x the EXP values */
646 for (i
=0;i
<NOISE_TAB_SIZE
;++i
)
647 noisetable_exp
[i
] = noisetable_exp
[i
]<< 1;
648 s
->noise_table
= noisetable_exp
;
655 norm
= 0; // PJJ: near as makes any diff to 0!
656 for (i
=0;i
<NOISE_TAB_SIZE
;++i
)
658 seed
= seed
* 314159 + 1;
659 s
->noise_table
[i
] = itofix32((int)seed
) * norm
;
664 s
->hgain_vlc
.table
= vlcbuf4
;
665 s
->hgain_vlc
.table_allocated
= VLCBUF4SIZE
;
666 init_vlc(&s
->hgain_vlc
, HGAINVLCBITS
, sizeof(hgain_huffbits
),
667 hgain_huffbits
, 1, 1,
668 hgain_huffcodes
, 2, 2, 0);
674 s
->exp_vlc
.table
= vlcbuf3
;
675 s
->exp_vlc
.table_allocated
= VLCBUF3SIZE
;
677 init_vlc(&s
->exp_vlc
, EXPVLCBITS
, sizeof(scale_huffbits
),
678 scale_huffbits
, 1, 1,
679 scale_huffcodes
, 4, 4, 0);
683 wma_lsp_to_curve_init(s
, s
->frame_len
);
686 /* choose the VLC tables for the coefficients */
688 if (s
->sample_rate
>= 32000)
692 else if (bps1
< 0x128f6)
696 runtabarray
[0] = runtab0
; runtabarray
[1] = runtab1
;
697 levtabarray
[0] = levtab0
; levtabarray
[1] = levtab1
;
699 s
->coef_vlc
[0].table
= vlcbuf1
;
700 s
->coef_vlc
[0].table_allocated
= VLCBUF1SIZE
;
701 s
->coef_vlc
[1].table
= vlcbuf2
;
702 s
->coef_vlc
[1].table_allocated
= VLCBUF2SIZE
;
705 init_coef_vlc(&s
->coef_vlc
[0], &s
->run_table
[0], &s
->level_table
[0],
706 &coef_vlcs
[coef_vlc_table
* 2], 0);
707 init_coef_vlc(&s
->coef_vlc
[1], &s
->run_table
[1], &s
->level_table
[1],
708 &coef_vlcs
[coef_vlc_table
* 2 + 1], 1);
710 s
->last_superframe_len
= 0;
711 s
->last_bitoffset
= 0;
717 /* compute x^-0.25 with an exponent and mantissa table. We use linear
718 interpolation to reduce the mantissa table size at a small speed
719 expense (linear interpolation approximately doubles the number of
720 bits of precision). */
721 static inline fixed32
pow_m1_4(WMADecodeContext
*s
, fixed32 x
)
732 m
= (u
.v
>> (23 - LSP_POW_BITS
)) & ((1 << LSP_POW_BITS
) - 1);
733 /* build interpolation scale: 1 <= t < 2. */
734 t
.v
= ((u
.v
<< LSP_POW_BITS
) & ((1 << 23) - 1)) | (127 << 23);
735 a
= s
->lsp_pow_m_table1
[m
];
736 b
= s
->lsp_pow_m_table2
[m
];
738 /* lsp_pow_e_table contains 32.32 format */
739 /* TODO: Since we're unlikely have value that cover the whole
740 * IEEE754 range, we probably don't need to have all possible exponents */
742 return (lsp_pow_e_table
[e
] * (a
+ fixmul32(b
, ftofix32(t
.f
))) >>32);
745 static void wma_lsp_to_curve_init(WMADecodeContext
*s
, int frame_len
)
747 fixed32 wdel
, a
, b
, temp
, temp2
;
750 wdel
= fixdiv32(M_PI_F
, itofix32(frame_len
));
751 temp
= fixdiv32(itofix32(1), itofix32(frame_len
));
752 for (i
=0; i
<frame_len
; ++i
)
754 /* TODO: can probably reuse the trig_init values here */
755 fsincos((temp
*i
)<<15, &temp2
);
756 /* get 3 bits headroom + 1 bit from not doubleing the values */
757 s
->lsp_cos_table
[i
] = temp2
>>3;
760 /* NOTE: these two tables are needed to avoid two operations in
765 /*double check this later*/
766 for(i
=(1 << LSP_POW_BITS
) - 1;i
>=0;i
--)
768 m
= (1 << LSP_POW_BITS
) + i
;
769 a
= pow_a_table
[ix
++]<<4;
770 s
->lsp_pow_m_table1
[i
] = 2 * a
- b
;
771 s
->lsp_pow_m_table2
[i
] = b
- a
;
777 /* NOTE: We use the same code as Vorbis here */
778 /* XXX: optimize it further with SSE/3Dnow */
779 static void wma_lsp_to_curve(WMADecodeContext
*s
,
781 fixed32
*val_max_ptr
,
786 fixed32 p
, q
, w
, v
, val_max
, temp
, temp2
;
791 /* shift by 2 now to reduce rounding error,
792 * we can renormalize right before pow_m1_4
797 w
= s
->lsp_cos_table
[i
];
799 for (j
=1;j
<NB_LSP_COEFS
;j
+=2)
801 /* w is 5.27 format, lsp is in 16.16, temp2 becomes 5.27 format */
802 temp2
= ((w
- (lsp
[j
- 1]<<11)));
805 /* q is 16.16 format, temp2 is 5.27, q becomes 16.16 */
806 q
= fixmul32b(q
, temp2
)<<4;
807 p
= fixmul32b(p
, (w
- (lsp
[j
]<<11)))<<4;
810 /* 2 in 5.27 format is 0x10000000 */
811 p
= fixmul32(p
, fixmul32b(p
, (0x10000000 - w
)))<<3;
812 q
= fixmul32(q
, fixmul32b(q
, (0x10000000 + w
)))<<3;
814 v
= (p
+ q
) >>9; /* p/q end up as 16.16 */
821 *val_max_ptr
= val_max
;
824 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
825 static void decode_exp_lsp(WMADecodeContext
*s
, int ch
)
827 fixed32 lsp_coefs
[NB_LSP_COEFS
];
830 for (i
= 0; i
< NB_LSP_COEFS
; ++i
)
832 if (i
== 0 || i
>= 8)
833 val
= get_bits(&s
->gb
, 3);
835 val
= get_bits(&s
->gb
, 4);
836 lsp_coefs
[i
] = lsp_codebook
[i
][val
];
841 &s
->max_exponent
[ch
],
846 /* decode exponents coded with VLC codes */
847 static int decode_exp_vlc(WMADecodeContext
*s
, int ch
)
849 int last_exp
, n
, code
;
850 const uint16_t *ptr
, *band_ptr
;
851 fixed32 v
, max_scale
;
854 /*accommodate the 60 negative indices */
855 const fixed32
*pow_10_to_yover16_ptr
= &pow_10_to_yover16
[61];
857 band_ptr
= s
->exponent_bands
[s
->frame_len_bits
- s
->block_len_bits
];
859 q
= s
->exponents
[ch
];
860 q_end
= q
+ s
->block_len
;
864 if (s
->version
== 1) //wmav1 only
866 last_exp
= get_bits(&s
->gb
, 5) + 10;
867 /* XXX: use a table */
868 v
= pow_10_to_yover16_ptr
[last_exp
];
881 code
= get_vlc2(&s
->gb
, s
->exp_vlc
.table
, EXPVLCBITS
, EXPMAX
);
886 /* NOTE: this offset is the same as MPEG4 AAC ! */
887 last_exp
+= code
- 60;
888 /* XXX: use a table */
889 v
= pow_10_to_yover16_ptr
[last_exp
];
903 s
->max_exponent
[ch
] = max_scale
;
907 /* return 0 if OK. return 1 if last block of frame. return -1 if
908 unrecorrable error. */
909 static int wma_decode_block(WMADecodeContext
*s
)
911 int n
, v
, a
, ch
, code
, bsize
;
912 int coef_nb_bits
, total_gain
;
913 int nb_coefs
[MAX_CHANNELS
];
916 /*DEBUGF("***decode_block: %d (%d samples of %d in frame)\n", s->block_num, s->block_len, s->frame_len);*/
918 /* compute current block length */
919 if (s
->use_variable_block_len
)
921 n
= av_log2(s
->nb_block_sizes
- 1) + 1;
923 if (s
->reset_block_lengths
)
925 s
->reset_block_lengths
= 0;
926 v
= get_bits(&s
->gb
, n
);
927 if (v
>= s
->nb_block_sizes
)
931 s
->prev_block_len_bits
= s
->frame_len_bits
- v
;
932 v
= get_bits(&s
->gb
, n
);
933 if (v
>= s
->nb_block_sizes
)
937 s
->block_len_bits
= s
->frame_len_bits
- v
;
941 /* update block lengths */
942 s
->prev_block_len_bits
= s
->block_len_bits
;
943 s
->block_len_bits
= s
->next_block_len_bits
;
945 v
= get_bits(&s
->gb
, n
);
947 if (v
>= s
->nb_block_sizes
)
949 // rb->splash(HZ*4, "v was %d", v); //5, 7
950 return -4; //this is it
953 //rb->splash(HZ, "passed v block (%d)!", v);
955 s
->next_block_len_bits
= s
->frame_len_bits
- v
;
959 /* fixed block len */
960 s
->next_block_len_bits
= s
->frame_len_bits
;
961 s
->prev_block_len_bits
= s
->frame_len_bits
;
962 s
->block_len_bits
= s
->frame_len_bits
;
964 /* now check if the block length is coherent with the frame length */
965 s
->block_len
= 1 << s
->block_len_bits
;
967 if ((s
->block_pos
+ s
->block_len
) > s
->frame_len
)
969 return -5; //oddly 32k sample from tracker fails here
972 if (s
->nb_channels
== 2)
974 s
->ms_stereo
= get_bits(&s
->gb
, 1);
977 for (ch
= 0; ch
< s
->nb_channels
; ++ch
)
979 a
= get_bits(&s
->gb
, 1);
980 s
->channel_coded
[ch
] = a
;
983 /* if no channel coded, no need to go further */
984 /* XXX: fix potential framing problems */
990 bsize
= s
->frame_len_bits
- s
->block_len_bits
;
992 /* read total gain and extract corresponding number of bits for
993 coef escape coding */
997 a
= get_bits(&s
->gb
, 7);
1005 if (total_gain
< 15)
1007 else if (total_gain
< 32)
1009 else if (total_gain
< 40)
1011 else if (total_gain
< 45)
1016 /* compute number of coefficients */
1017 n
= s
->coefs_end
[bsize
] - s
->coefs_start
;
1019 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1023 /* complex coding */
1024 if (s
->use_noise_coding
)
1027 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1029 if (s
->channel_coded
[ch
])
1032 n
= s
->exponent_high_sizes
[bsize
];
1035 a
= get_bits(&s
->gb
, 1);
1036 s
->high_band_coded
[ch
][i
] = a
;
1037 /* if noise coding, the coefficients are not transmitted */
1039 nb_coefs
[ch
] -= s
->exponent_high_bands
[bsize
][i
];
1043 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1045 if (s
->channel_coded
[ch
])
1047 int i
, n
, val
, code
;
1049 n
= s
->exponent_high_sizes
[bsize
];
1050 val
= (int)0x80000000;
1053 if (s
->high_band_coded
[ch
][i
])
1055 if (val
== (int)0x80000000)
1057 val
= get_bits(&s
->gb
, 7) - 19;
1061 //code = get_vlc(&s->gb, &s->hgain_vlc);
1062 code
= get_vlc2(&s
->gb
, s
->hgain_vlc
.table
, HGAINVLCBITS
, HGAINMAX
);
1069 s
->high_band_values
[ch
][i
] = val
;
1076 /* exponents can be reused in short blocks. */
1077 if ((s
->block_len_bits
== s
->frame_len_bits
) || get_bits(&s
->gb
, 1))
1079 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1081 if (s
->channel_coded
[ch
])
1085 if (decode_exp_vlc(s
, ch
) < 0)
1092 decode_exp_lsp(s
, ch
);
1094 s
->exponents_bsize
[ch
] = bsize
;
1099 /* parse spectral coefficients : just RLE encoding */
1100 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1102 if (s
->channel_coded
[ch
])
1105 int level
, run
, sign
, tindex
;
1106 int16_t *ptr
, *eptr
;
1107 const int16_t *level_table
, *run_table
;
1109 /* special VLC tables are used for ms stereo because
1110 there is potentially less energy there */
1111 tindex
= (ch
== 1 && s
->ms_stereo
);
1112 coef_vlc
= &s
->coef_vlc
[tindex
];
1113 run_table
= s
->run_table
[tindex
];
1114 level_table
= s
->level_table
[tindex
];
1116 ptr
= &s
->coefs1
[ch
][0];
1117 eptr
= ptr
+ nb_coefs
[ch
];
1118 memset(ptr
, 0, s
->block_len
* sizeof(int16_t));
1122 code
= get_vlc2(&s
->gb
, coef_vlc
->table
, VLCBITS
, VLCMAX
);
1123 //code = get_vlc(&s->gb, coef_vlc);
1136 level
= get_bits(&s
->gb
, coef_nb_bits
);
1137 /* NOTE: this is rather suboptimal. reading
1138 block_len_bits would be better */
1139 run
= get_bits(&s
->gb
, s
->frame_len_bits
);
1144 run
= run_table
[code
];
1145 level
= level_table
[code
];
1147 sign
= get_bits(&s
->gb
, 1);
1158 /* NOTE: EOB can be omitted */
1163 if (s
->version
== 1 && s
->nb_channels
>= 2)
1165 align_get_bits(&s
->gb
);
1170 int n4
= s
->block_len
>> 1;
1171 //mdct_norm = 0x10000;
1172 //mdct_norm = fixdiv32(mdct_norm,itofix32(n4));
1174 mdct_norm
= 0x10000>>(s
->block_len_bits
-1); //theres no reason to do a divide by two in fixed precision ...
1176 if (s
->version
== 1)
1178 fixed32 tmp
= fixtoi32(fixsqrt32(itofix32(n4
)));
1179 mdct_norm
*= tmp
; // PJJ : exercise this path
1184 /* finally compute the MDCT coefficients */
1185 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1187 if (s
->channel_coded
[ch
])
1190 fixed32
*exponents
, *exp_ptr
;
1191 fixed32
*coefs
, atemp
;
1194 fixed32 noise
, temp1
, temp2
, mult2
;
1195 int i
, j
, n
, n1
, last_high_band
, esize
;
1196 fixed32 exp_power
[HIGH_BAND_MAX_SIZE
];
1198 //total_gain, coefs1, mdctnorm are lossless
1200 coefs1
= s
->coefs1
[ch
];
1201 exponents
= s
->exponents
[ch
];
1202 esize
= s
->exponents_bsize
[ch
];
1203 coefs
= (*(s
->coefs
))[ch
];
1208 * Previously the IMDCT was run in 17.15 precision to avoid overflow. However rare files could
1209 * overflow here as well, so switch to 17.15 during coefs calculation.
1213 if (s
->use_noise_coding
)
1215 /*TODO: mult should be converted to 32 bit to speed up noise coding*/
1217 mult
= fixdiv64(pow_table
[total_gain
+20],Fixed32To64(s
->max_exponent
[ch
]));
1218 mult
= mult
* mdct_norm
; //what the hell? This is actually fixed64*2^16!
1221 /* very low freqs : noise */
1222 for(i
= 0;i
< s
->coefs_start
; ++i
)
1224 *coefs
++ = fixmul32( (fixmul32(s
->noise_table
[s
->noise_index
],(*exponents
++))>>4),Fixed32From64(mult1
)) >>1;
1225 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1228 n1
= s
->exponent_high_sizes
[bsize
];
1230 /* compute power of high bands */
1231 exp_ptr
= exponents
+
1232 s
->high_band_start
[bsize
] -
1234 last_high_band
= 0; /* avoid warning */
1237 n
= s
->exponent_high_bands
[s
->frame_len_bits
-
1238 s
->block_len_bits
][j
];
1239 if (s
->high_band_coded
[ch
][j
])
1243 for(i
= 0;i
< n
; ++i
)
1245 /*v is noramlized later on so its fixed format is irrelevant*/
1247 e2
+= fixmul32(v
, v
)>>3;
1249 exp_power
[j
] = e2
/n
; /*n is an int...*/
1255 /* main freqs and high freqs */
1260 n
= s
->high_band_start
[bsize
] -
1265 n
= s
->exponent_high_bands
[s
->frame_len_bits
-
1266 s
->block_len_bits
][j
];
1268 if (j
>= 0 && s
->high_band_coded
[ch
][j
])
1270 /* use noise with specified power */
1271 fixed32 tmp
= fixdiv32(exp_power
[j
],exp_power
[last_high_band
]);
1272 mult1
= (fixed64
)fixsqrt32(tmp
);
1273 /* XXX: use a table */
1274 /*mult1 is 48.16, pow_table is 48.16*/
1275 mult1
= mult1
* pow_table
[s
->high_band_values
[ch
][j
]+20] >> PRECISION
;
1277 /*this step has a fairly high degree of error for some reason*/
1278 mult1
= fixdiv64(mult1
,fixmul32(s
->max_exponent
[ch
],s
->noise_mult
));
1280 mult1
= mult1
*mdct_norm
>>PRECISION
;
1281 for(i
= 0;i
< n
; ++i
)
1283 noise
= s
->noise_table
[s
->noise_index
];
1284 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1285 *coefs
++ = fixmul32((fixmul32(*exponents
,noise
)>>4),Fixed32From64(mult1
)) >>1;
1291 /* coded values + small noise */
1292 for(i
= 0;i
< n
; ++i
)
1294 // PJJ: check code path
1295 noise
= s
->noise_table
[s
->noise_index
];
1296 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1298 /*don't forget to renormalize the noise*/
1299 temp1
= (((int32_t)*coefs1
++)<<16) + (noise
>>4);
1300 temp2
= fixmul32(*exponents
, mult
>>17);
1301 *coefs
++ = fixmul32(temp1
, temp2
);
1307 /* very high freqs : noise */
1308 n
= s
->block_len
- s
->coefs_end
[bsize
];
1309 mult2
= fixmul32(mult
>>16,exponents
[-1]) ; /*the work around for 32.32 vars are getting stupid*/
1310 for (i
= 0; i
< n
; ++i
)
1312 /*renormalize the noise product and then reduce to 17.15 precison*/
1313 *coefs
++ = fixmul32(s
->noise_table
[s
->noise_index
],mult2
) >>5;
1315 s
->noise_index
= (s
->noise_index
+ 1) & (NOISE_TAB_SIZE
- 1);
1320 /*Noise coding not used, simply convert from exp to fixed representation*/
1323 fixed32 mult3
= (fixed32
)(fixdiv64(pow_table
[total_gain
+20],Fixed32To64(s
->max_exponent
[ch
])));
1324 mult3
= fixmul32(mult3
, mdct_norm
);
1328 /* XXX: optimize more, unrolling this loop in asm might be a good idea */
1330 for(i
= 0;i
< n
; ++i
)
1332 atemp
= (coefs1
[i
] * mult3
)>>1;
1333 *coefs
++=fixmul32(atemp
,exponents
[i
<<bsize
>>esize
]);
1335 n
= s
->block_len
- s
->coefs_end
[bsize
];
1336 memset(coefs
, 0, n
*sizeof(fixed32
));
1343 if (s
->ms_stereo
&& s
->channel_coded
[1])
1347 fixed32 (*coefs
)[MAX_CHANNELS
][BLOCK_MAX_SIZE
] = (s
->coefs
);
1349 /* nominal case for ms stereo: we do it before mdct */
1350 /* no need to optimize this case because it should almost
1352 if (!s
->channel_coded
[0])
1354 memset((*(s
->coefs
))[0], 0, sizeof(fixed32
) * s
->block_len
);
1355 s
->channel_coded
[0] = 1;
1358 for(i
= 0; i
< s
->block_len
; ++i
)
1362 (*coefs
)[0][i
] = a
+ b
;
1363 (*coefs
)[1][i
] = a
- b
;
1367 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1369 if (s
->channel_coded
[ch
])
1371 static fixed32 output
[BLOCK_MAX_SIZE
* 2] IBSS_ATTR
;
1376 n4
= s
->block_len
>>1;
1378 ff_imdct_calc(&s
->mdct_ctx
[bsize
],
1383 /* add in the frame */
1384 index
= (s
->frame_len
/ 2) + s
->block_pos
- n4
;
1386 wma_window(s
, output
, &s
->frame_out
[ch
][index
]);
1390 /* specific fast case for ms-stereo : add to second
1391 channel if it is not coded */
1392 if (s
->ms_stereo
&& !s
->channel_coded
[1])
1394 wma_window(s
, output
, &s
->frame_out
[1][index
]);
1399 /* update block number */
1401 s
->block_pos
+= s
->block_len
;
1402 if (s
->block_pos
>= s
->frame_len
)
1412 /* decode a frame of frame_len samples */
1413 static int wma_decode_frame(WMADecodeContext
*s
, int32_t *samples
)
1415 int ret
, i
, n
, ch
, incr
;
1418 // rb->splash(HZ, "in wma_decode_frame");
1420 /* read each block */
1427 ret
= wma_decode_block(s
);
1431 DEBUGF("wma_decode_block failed with code %d\n", ret
);
1440 /* return frame with full 30-bit precision */
1442 incr
= s
->nb_channels
;
1443 for(ch
= 0; ch
< s
->nb_channels
; ++ch
)
1446 iptr
= s
->frame_out
[ch
];
1453 /* prepare for next block */
1454 memmove(&s
->frame_out
[ch
][0], &s
->frame_out
[ch
][s
->frame_len
],
1455 s
->frame_len
* sizeof(fixed32
));
1462 /* Initialise the superframe decoding */
1464 int wma_decode_superframe_init(WMADecodeContext
* s
,
1465 uint8_t *buf
, /*input*/
1470 s
->last_superframe_len
= 0;
1474 s
->current_frame
= 0;
1476 init_get_bits(&s
->gb
, buf
, buf_size
*8);
1478 if (s
->use_bit_reservoir
)
1480 /* read super frame header */
1481 get_bits(&s
->gb
, 4); /* super frame index */
1482 s
->nb_frames
= get_bits(&s
->gb
, 4);
1484 if (s
->last_superframe_len
== 0)
1486 else if (s
->nb_frames
== 0)
1489 s
->bit_offset
= get_bits(&s
->gb
, s
->byte_offset_bits
+ 3);
1498 /* Decode a single frame in the current superframe - return -1 if
1499 there was a decoding error, or the number of samples decoded.
1502 int wma_decode_superframe_frame(WMADecodeContext
* s
,
1503 int32_t* samples
, /*output*/
1504 uint8_t *buf
, /*input*/
1510 if ((s
->use_bit_reservoir
) && (s
->current_frame
== 0))
1512 if (s
->last_superframe_len
> 0)
1514 /* add s->bit_offset bits to last frame */
1515 if ((s
->last_superframe_len
+ ((s
->bit_offset
+ 7) >> 3)) >
1516 MAX_CODED_SUPERFRAME_SIZE
)
1518 DEBUGF("superframe size too large error\n");
1521 q
= s
->last_superframe
+ s
->last_superframe_len
;
1522 len
= s
->bit_offset
;
1525 *q
++ = (get_bits
)(&s
->gb
, 8);
1530 *q
++ = (get_bits
)(&s
->gb
, len
) << (8 - len
);
1533 /* XXX: s->bit_offset bits into last frame */
1534 init_get_bits(&s
->gb
, s
->last_superframe
, MAX_CODED_SUPERFRAME_SIZE
*8);
1535 /* skip unused bits */
1536 if (s
->last_bitoffset
> 0)
1537 skip_bits(&s
->gb
, s
->last_bitoffset
);
1539 /* this frame is stored in the last superframe and in the
1541 if (wma_decode_frame(s
, samples
) < 0)
1548 /* read each frame starting from s->bit_offset */
1549 pos
= s
->bit_offset
+ 4 + 4 + s
->byte_offset_bits
+ 3;
1550 init_get_bits(&s
->gb
, buf
+ (pos
>> 3), (MAX_CODED_SUPERFRAME_SIZE
- (pos
>> 3))*8);
1553 skip_bits(&s
->gb
, len
);
1555 s
->reset_block_lengths
= 1;
1558 /* If we haven't decoded a frame yet, do it now */
1561 if (wma_decode_frame(s
, samples
) < 0)
1569 if ((s
->use_bit_reservoir
) && (s
->current_frame
== s
->nb_frames
))
1571 /* we copy the end of the frame in the last frame buffer */
1572 pos
= get_bits_count(&s
->gb
) + ((s
->bit_offset
+ 4 + 4 + s
->byte_offset_bits
+ 3) & ~7);
1573 s
->last_bitoffset
= pos
& 7;
1575 len
= buf_size
- pos
;
1576 if (len
> MAX_CODED_SUPERFRAME_SIZE
|| len
< 0)
1578 DEBUGF("superframe size too large error after decodeing\n");
1581 s
->last_superframe_len
= len
;
1582 memcpy(s
->last_superframe
, buf
+ pos
, len
);
1585 return s
->frame_len
;
1588 /* when error, we reset the bit reservoir */
1590 s
->last_superframe_len
= 0;