2 * WMA compatible decoder
3 * Copyright (c) 2002 The FFmpeg Project.
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * WMA compatible decoder.
25 * This decoder handles Microsoft Windows Media Audio data, versions 1 & 2.
26 * WMA v1 is identified by audio format 0x160 in Microsoft media files
27 * (ASF/AVI/WAV). WMA v2 is identified by audio format 0x161.
29 * To use this decoder, a calling application must supply the extra data
30 * bytes provided with the WMA data. These are the extra, codec-specific
31 * bytes at the end of a WAVEFORMATEX data structure. Transmit these bytes
32 * to the decoder using the extradata[_size] fields in AVCodecContext. There
33 * should be 4 extra bytes for v1 data and 6 extra bytes for v2 data.
37 #include "bitstream.h"
41 #define BLOCK_MIN_BITS 7
42 #define BLOCK_MAX_BITS 11
43 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
45 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
47 /* XXX: find exact max size */
48 #define HIGH_BAND_MAX_SIZE 16
50 #define NB_LSP_COEFS 10
52 /* XXX: is it a suitable value ? */
53 #define MAX_CODED_SUPERFRAME_SIZE 16384
55 #define MAX_CHANNELS 2
57 #define NOISE_TAB_SIZE 8192
59 #define LSP_POW_BITS 7
62 #define VLCMAX ((22+VLCBITS-1)/VLCBITS)
65 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS)
67 #define HGAINVLCBITS 9
68 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS)
70 typedef struct WMADecodeContext {
75 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
77 int use_bit_reservoir;
78 int use_variable_block_len;
79 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
80 int use_noise_coding; /* true if perceptual noise is added */
83 int exponent_sizes[BLOCK_NB_SIZES];
84 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
85 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
86 int coefs_start; /* first coded coef */
87 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
88 int exponent_high_sizes[BLOCK_NB_SIZES];
89 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
92 /* coded values in high bands */
93 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
94 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
96 /* there are two possible tables for spectral coefficients */
98 uint16_t *run_table[2];
99 uint16_t *level_table[2];
101 int frame_len; /* frame length in samples */
102 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
103 int nb_block_sizes; /* number of block sizes */
105 int reset_block_lengths;
106 int block_len_bits; /* log2 of current block length */
107 int next_block_len_bits; /* log2 of next block length */
108 int prev_block_len_bits; /* log2 of prev block length */
109 int block_len; /* block length in samples */
110 int block_num; /* block number in current frame */
111 int block_pos; /* current position in frame */
112 uint8_t ms_stereo; /* true if mid/side stereo mode */
113 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
114 DECLARE_ALIGNED_16(float, exponents[MAX_CHANNELS][BLOCK_MAX_SIZE]);
115 float max_exponent[MAX_CHANNELS];
116 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
117 DECLARE_ALIGNED_16(float, coefs[MAX_CHANNELS][BLOCK_MAX_SIZE]);
118 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
119 float *windows[BLOCK_NB_SIZES];
120 DECLARE_ALIGNED_16(FFTSample, mdct_tmp[BLOCK_MAX_SIZE]); /* temporary storage for imdct */
121 /* output buffer for one frame and the last for IMDCT windowing */
122 DECLARE_ALIGNED_16(float, frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2]);
123 /* last frame info */
124 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
126 int last_superframe_len;
127 float noise_table[NOISE_TAB_SIZE];
129 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
130 /* lsp_to_curve tables */
131 float lsp_cos_table[BLOCK_MAX_SIZE];
132 float lsp_pow_e_table[256];
133 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
134 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
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 */
149 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
154 static void dump_shorts(const char *name, const short *tab, int n)
158 tprintf("%s[%d]:\n", name, n);
162 tprintf(" %5d.0", tab[i]);
168 static void dump_floats(const char *name, int prec, const float *tab, int n)
172 tprintf("%s[%d]:\n", name, n);
176 tprintf(" %8.*f", prec, tab[i]);
185 /* XXX: use same run/length optimization as mpeg decoders */
186 static void init_coef_vlc(VLC *vlc,
187 uint16_t **prun_table, uint16_t **plevel_table,
188 const CoefVLCTable *vlc_table)
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;
198 init_vlc(vlc, VLCBITS, n, table_bits, 1, 1, table_codes, 4, 4, 0);
200 run_table = av_malloc(n * sizeof(uint16_t));
201 level_table = av_malloc(n * sizeof(uint16_t));
209 level_table[i] = level;
214 *prun_table = run_table;
215 *plevel_table = level_table;
218 static int wma_decode_init(AVCodecContext * avctx)
220 WMADecodeContext *s = avctx->priv_data;
221 int i, flags1, flags2;
224 float bps1, high_freq;
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;
234 dsputil_init(&s->dsp, avctx);
236 if (avctx->codec->id == CODEC_ID_WMAV1) {
242 /* extract flag infos */
245 extradata = avctx->extradata;
246 if (s->version == 1 && avctx->extradata_size >= 4) {
247 flags1 = extradata[0] | (extradata[1] << 8);
248 flags2 = extradata[2] | (extradata[3] << 8);
249 } else if (s->version == 2 && avctx->extradata_size >= 6) {
250 flags1 = extradata[0] | (extradata[1] << 8) |
251 (extradata[2] << 16) | (extradata[3] << 24);
252 flags2 = extradata[4] | (extradata[5] << 8);
254 s->use_exp_vlc = flags2 & 0x0001;
255 s->use_bit_reservoir = flags2 & 0x0002;
256 s->use_variable_block_len = flags2 & 0x0004;
258 /* compute MDCT block size */
259 if (s->sample_rate <= 16000) {
260 s->frame_len_bits = 9;
261 } else if (s->sample_rate <= 22050 ||
262 (s->sample_rate <= 32000 && s->version == 1)) {
263 s->frame_len_bits = 10;
265 s->frame_len_bits = 11;
267 s->frame_len = 1 << s->frame_len_bits;
268 if (s->use_variable_block_len) {
270 nb = ((flags2 >> 3) & 3) + 1;
271 if ((s->bit_rate / s->nb_channels) >= 32000)
273 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
276 s->nb_block_sizes = nb + 1;
278 s->nb_block_sizes = 1;
281 /* init rate dependant parameters */
282 s->use_noise_coding = 1;
283 high_freq = s->sample_rate * 0.5;
285 /* if version 2, then the rates are normalized */
286 sample_rate1 = s->sample_rate;
287 if (s->version == 2) {
288 if (sample_rate1 >= 44100)
289 sample_rate1 = 44100;
290 else if (sample_rate1 >= 22050)
291 sample_rate1 = 22050;
292 else if (sample_rate1 >= 16000)
293 sample_rate1 = 16000;
294 else if (sample_rate1 >= 11025)
295 sample_rate1 = 11025;
296 else if (sample_rate1 >= 8000)
300 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
301 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0 + 0.5)) + 2;
303 /* compute high frequency value and choose if noise coding should
306 if (s->nb_channels == 2)
308 if (sample_rate1 == 44100) {
310 s->use_noise_coding = 0;
312 high_freq = high_freq * 0.4;
313 } else if (sample_rate1 == 22050) {
315 s->use_noise_coding = 0;
316 else if (bps1 >= 0.72)
317 high_freq = high_freq * 0.7;
319 high_freq = high_freq * 0.6;
320 } else if (sample_rate1 == 16000) {
322 high_freq = high_freq * 0.5;
324 high_freq = high_freq * 0.3;
325 } else if (sample_rate1 == 11025) {
326 high_freq = high_freq * 0.7;
327 } else if (sample_rate1 == 8000) {
329 high_freq = high_freq * 0.5;
330 } else if (bps > 0.75) {
331 s->use_noise_coding = 0;
333 high_freq = high_freq * 0.65;
337 high_freq = high_freq * 0.75;
338 } else if (bps >= 0.6) {
339 high_freq = high_freq * 0.6;
341 high_freq = high_freq * 0.5;
344 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
345 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
346 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
348 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
349 bps, bps1, high_freq, s->byte_offset_bits);
350 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
351 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
353 /* compute the scale factor band sizes for each MDCT block size */
355 int a, b, pos, lpos, k, block_len, i, j, n;
356 const uint8_t *table;
358 if (s->version == 1) {
363 for(k = 0; k < s->nb_block_sizes; k++) {
364 block_len = s->frame_len >> k;
366 if (s->version == 1) {
369 a = wma_critical_freqs[i];
371 pos = ((block_len * 2 * a) + (b >> 1)) / b;
374 s->exponent_bands[0][i] = pos - lpos;
375 if (pos >= block_len) {
381 s->exponent_sizes[0] = i;
383 /* hardcoded tables */
385 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
387 if (s->sample_rate >= 44100)
388 table = exponent_band_44100[a];
389 else if (s->sample_rate >= 32000)
390 table = exponent_band_32000[a];
391 else if (s->sample_rate >= 22050)
392 table = exponent_band_22050[a];
397 s->exponent_bands[k][i] = table[i];
398 s->exponent_sizes[k] = n;
403 a = wma_critical_freqs[i];
405 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
410 s->exponent_bands[k][j++] = pos - lpos;
411 if (pos >= block_len)
415 s->exponent_sizes[k] = j;
419 /* max number of coefs */
420 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
421 /* high freq computation */
422 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
423 s->sample_rate + 0.5);
424 n = s->exponent_sizes[k];
430 pos += s->exponent_bands[k][i];
432 if (start < s->high_band_start[k])
433 start = s->high_band_start[k];
434 if (end > s->coefs_end[k])
435 end = s->coefs_end[k];
437 s->exponent_high_bands[k][j++] = end - start;
439 s->exponent_high_sizes[k] = j;
441 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
444 s->high_band_start[k],
445 s->exponent_high_sizes[k]);
446 for(j=0;j<s->exponent_high_sizes[k];j++)
447 tprintf(" %d", s->exponent_high_bands[k][j]);
456 for(i = 0; i < s->nb_block_sizes; i++) {
457 tprintf("%5d: n=%2d:",
459 s->exponent_sizes[i]);
460 for(j=0;j<s->exponent_sizes[i];j++)
461 tprintf(" %d", s->exponent_bands[i][j]);
468 for(i = 0; i < s->nb_block_sizes; i++)
469 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
471 /* init MDCT windows : simple sinus window */
472 for(i = 0; i < s->nb_block_sizes; i++) {
475 n = 1 << (s->frame_len_bits - i);
476 window = av_malloc(sizeof(float) * n);
477 alpha = M_PI / (2.0 * n);
479 window[n - j - 1] = sin((j + 0.5) * alpha);
481 s->windows[i] = window;
484 s->reset_block_lengths = 1;
486 if (s->use_noise_coding) {
488 /* init the noise generator */
490 s->noise_mult = 0.02;
492 s->noise_mult = 0.04;
495 for(i=0;i<NOISE_TAB_SIZE;i++)
496 s->noise_table[i] = 1.0 * s->noise_mult;
502 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
503 for(i=0;i<NOISE_TAB_SIZE;i++) {
504 seed = seed * 314159 + 1;
505 s->noise_table[i] = (float)((int)seed) * norm;
509 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(hgain_huffbits),
510 hgain_huffbits, 1, 1,
511 hgain_huffcodes, 2, 2, 0);
514 if (s->use_exp_vlc) {
515 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(scale_huffbits),
516 scale_huffbits, 1, 1,
517 scale_huffcodes, 4, 4, 0);
519 wma_lsp_to_curve_init(s, s->frame_len);
522 /* choose the VLC tables for the coefficients */
524 if (s->sample_rate >= 32000) {
527 else if (bps1 < 1.16)
531 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
532 &coef_vlcs[coef_vlc_table * 2]);
533 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
534 &coef_vlcs[coef_vlc_table * 2 + 1]);
538 /* interpolate values for a bigger or smaller block. The block must
539 have multiple sizes */
540 static void interpolate_array(float *scale, int old_size, int new_size)
545 if (new_size > old_size) {
546 jincr = new_size / old_size;
548 for(i = old_size - 1; i >=0; i--) {
555 } else if (new_size < old_size) {
557 jincr = old_size / new_size;
558 for(i = 0; i < new_size; i++) {
565 /* compute x^-0.25 with an exponent and mantissa table. We use linear
566 interpolation to reduce the mantissa table size at a small speed
567 expense (linear interpolation approximately doubles the number of
568 bits of precision). */
569 static inline float pow_m1_4(WMADecodeContext *s, float x)
580 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
581 /* build interpolation scale: 1 <= t < 2. */
582 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
583 a = s->lsp_pow_m_table1[m];
584 b = s->lsp_pow_m_table2[m];
585 return s->lsp_pow_e_table[e] * (a + b * t.f);
588 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
593 wdel = M_PI / frame_len;
594 for(i=0;i<frame_len;i++)
595 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
597 /* tables for x^-0.25 computation */
600 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
603 /* NOTE: these two tables are needed to avoid two operations in
606 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
607 m = (1 << LSP_POW_BITS) + i;
608 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
610 s->lsp_pow_m_table1[i] = 2 * a - b;
611 s->lsp_pow_m_table2[i] = b - a;
620 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
625 /* NOTE: We use the same code as Vorbis here */
626 /* XXX: optimize it further with SSE/3Dnow */
627 static void wma_lsp_to_curve(WMADecodeContext *s,
628 float *out, float *val_max_ptr,
632 float p, q, w, v, val_max;
638 w = s->lsp_cos_table[i];
639 for(j=1;j<NB_LSP_COEFS;j+=2){
651 *val_max_ptr = val_max;
654 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
655 static void decode_exp_lsp(WMADecodeContext *s, int ch)
657 float lsp_coefs[NB_LSP_COEFS];
660 for(i = 0; i < NB_LSP_COEFS; i++) {
661 if (i == 0 || i >= 8)
662 val = get_bits(&s->gb, 3);
664 val = get_bits(&s->gb, 4);
665 lsp_coefs[i] = lsp_codebook[i][val];
668 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
669 s->block_len, lsp_coefs);
672 /* decode exponents coded with VLC codes */
673 static int decode_exp_vlc(WMADecodeContext *s, int ch)
675 int last_exp, n, code;
676 const uint16_t *ptr, *band_ptr;
677 float v, *q, max_scale, *q_end;
679 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
681 q = s->exponents[ch];
682 q_end = q + s->block_len;
684 if (s->version == 1) {
685 last_exp = get_bits(&s->gb, 5) + 10;
686 /* XXX: use a table */
687 v = pow(10, last_exp * (1.0 / 16.0));
696 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
699 /* NOTE: this offset is the same as MPEG4 AAC ! */
700 last_exp += code - 60;
701 /* XXX: use a table */
702 v = pow(10, last_exp * (1.0 / 16.0));
710 s->max_exponent[ch] = max_scale;
714 /* return 0 if OK. return 1 if last block of frame. return -1 if
715 unrecorrable error. */
716 static int wma_decode_block(WMADecodeContext *s)
718 int n, v, a, ch, code, bsize;
719 int coef_nb_bits, total_gain, parse_exponents;
720 DECLARE_ALIGNED_16(float, window[BLOCK_MAX_SIZE * 2]);
721 int nb_coefs[MAX_CHANNELS];
725 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
728 /* compute current block length */
729 if (s->use_variable_block_len) {
730 n = av_log2(s->nb_block_sizes - 1) + 1;
732 if (s->reset_block_lengths) {
733 s->reset_block_lengths = 0;
734 v = get_bits(&s->gb, n);
735 if (v >= s->nb_block_sizes)
737 s->prev_block_len_bits = s->frame_len_bits - v;
738 v = get_bits(&s->gb, n);
739 if (v >= s->nb_block_sizes)
741 s->block_len_bits = s->frame_len_bits - v;
743 /* update block lengths */
744 s->prev_block_len_bits = s->block_len_bits;
745 s->block_len_bits = s->next_block_len_bits;
747 v = get_bits(&s->gb, n);
748 if (v >= s->nb_block_sizes)
750 s->next_block_len_bits = s->frame_len_bits - v;
752 /* fixed block len */
753 s->next_block_len_bits = s->frame_len_bits;
754 s->prev_block_len_bits = s->frame_len_bits;
755 s->block_len_bits = s->frame_len_bits;
758 /* now check if the block length is coherent with the frame length */
759 s->block_len = 1 << s->block_len_bits;
760 if ((s->block_pos + s->block_len) > s->frame_len)
763 if (s->nb_channels == 2) {
764 s->ms_stereo = get_bits(&s->gb, 1);
767 for(ch = 0; ch < s->nb_channels; ch++) {
768 a = get_bits(&s->gb, 1);
769 s->channel_coded[ch] = a;
772 /* if no channel coded, no need to go further */
773 /* XXX: fix potential framing problems */
777 bsize = s->frame_len_bits - s->block_len_bits;
779 /* read total gain and extract corresponding number of bits for
780 coef escape coding */
783 a = get_bits(&s->gb, 7);
791 else if (total_gain < 32)
793 else if (total_gain < 40)
795 else if (total_gain < 45)
800 /* compute number of coefficients */
801 n = s->coefs_end[bsize] - s->coefs_start;
802 for(ch = 0; ch < s->nb_channels; ch++)
806 if (s->use_noise_coding) {
808 for(ch = 0; ch < s->nb_channels; ch++) {
809 if (s->channel_coded[ch]) {
811 n = s->exponent_high_sizes[bsize];
813 a = get_bits(&s->gb, 1);
814 s->high_band_coded[ch][i] = a;
815 /* if noise coding, the coefficients are not transmitted */
817 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
821 for(ch = 0; ch < s->nb_channels; ch++) {
822 if (s->channel_coded[ch]) {
825 n = s->exponent_high_sizes[bsize];
826 val = (int)0x80000000;
828 if (s->high_band_coded[ch][i]) {
829 if (val == (int)0x80000000) {
830 val = get_bits(&s->gb, 7) - 19;
832 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX);
837 s->high_band_values[ch][i] = val;
844 /* exposant can be interpolated in short blocks. */
846 if (s->block_len_bits != s->frame_len_bits) {
847 parse_exponents = get_bits(&s->gb, 1);
850 if (parse_exponents) {
851 for(ch = 0; ch < s->nb_channels; ch++) {
852 if (s->channel_coded[ch]) {
853 if (s->use_exp_vlc) {
854 if (decode_exp_vlc(s, ch) < 0)
857 decode_exp_lsp(s, ch);
862 for(ch = 0; ch < s->nb_channels; ch++) {
863 if (s->channel_coded[ch]) {
864 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
870 /* parse spectral coefficients : just RLE encoding */
871 for(ch = 0; ch < s->nb_channels; ch++) {
872 if (s->channel_coded[ch]) {
874 int level, run, sign, tindex;
876 const uint16_t *level_table, *run_table;
878 /* special VLC tables are used for ms stereo because
879 there is potentially less energy there */
880 tindex = (ch == 1 && s->ms_stereo);
881 coef_vlc = &s->coef_vlc[tindex];
882 run_table = s->run_table[tindex];
883 level_table = s->level_table[tindex];
885 ptr = &s->coefs1[ch][0];
886 eptr = ptr + nb_coefs[ch];
887 memset(ptr, 0, s->block_len * sizeof(int16_t));
889 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX);
895 } else if (code == 0) {
897 level = get_bits(&s->gb, coef_nb_bits);
898 /* NOTE: this is rather suboptimal. reading
899 block_len_bits would be better */
900 run = get_bits(&s->gb, s->frame_len_bits);
903 run = run_table[code];
904 level = level_table[code];
906 sign = get_bits(&s->gb, 1);
912 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n");
916 /* NOTE: EOB can be omitted */
921 if (s->version == 1 && s->nb_channels >= 2) {
922 align_get_bits(&s->gb);
928 int n4 = s->block_len / 2;
929 mdct_norm = 1.0 / (float)n4;
930 if (s->version == 1) {
931 mdct_norm *= sqrt(n4);
935 /* finally compute the MDCT coefficients */
936 for(ch = 0; ch < s->nb_channels; ch++) {
937 if (s->channel_coded[ch]) {
939 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
940 int i, j, n, n1, last_high_band;
941 float exp_power[HIGH_BAND_MAX_SIZE];
943 coefs1 = s->coefs1[ch];
944 exponents = s->exponents[ch];
945 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
947 coefs = s->coefs[ch];
948 if (s->use_noise_coding) {
950 /* very low freqs : noise */
951 for(i = 0;i < s->coefs_start; i++) {
952 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
953 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
956 n1 = s->exponent_high_sizes[bsize];
958 /* compute power of high bands */
959 exp_ptr = exponents +
960 s->high_band_start[bsize] -
962 last_high_band = 0; /* avoid warning */
964 n = s->exponent_high_bands[s->frame_len_bits -
965 s->block_len_bits][j];
966 if (s->high_band_coded[ch][j]) {
969 for(i = 0;i < n; i++) {
973 exp_power[j] = e2 / n;
975 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
980 /* main freqs and high freqs */
983 n = s->high_band_start[bsize] -
986 n = s->exponent_high_bands[s->frame_len_bits -
987 s->block_len_bits][j];
989 if (j >= 0 && s->high_band_coded[ch][j]) {
990 /* use noise with specified power */
991 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
992 /* XXX: use a table */
993 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
994 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
996 for(i = 0;i < n; i++) {
997 noise = s->noise_table[s->noise_index];
998 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
999 *coefs++ = (*exponents++) * noise * mult1;
1002 /* coded values + small noise */
1003 for(i = 0;i < n; i++) {
1004 noise = s->noise_table[s->noise_index];
1005 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1006 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
1011 /* very high freqs : noise */
1012 n = s->block_len - s->coefs_end[bsize];
1013 mult1 = mult * exponents[-1];
1014 for(i = 0; i < n; i++) {
1015 *coefs++ = s->noise_table[s->noise_index] * mult1;
1016 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1019 /* XXX: optimize more */
1020 for(i = 0;i < s->coefs_start; i++)
1023 for(i = 0;i < n; i++) {
1024 *coefs++ = coefs1[i] * exponents[i] * mult;
1026 n = s->block_len - s->coefs_end[bsize];
1027 for(i = 0;i < n; i++)
1034 for(ch = 0; ch < s->nb_channels; ch++) {
1035 if (s->channel_coded[ch]) {
1036 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1037 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1042 if (s->ms_stereo && s->channel_coded[1]) {
1046 /* nominal case for ms stereo: we do it before mdct */
1047 /* no need to optimize this case because it should almost
1049 if (!s->channel_coded[0]) {
1050 tprintf("rare ms-stereo case happened\n");
1051 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1052 s->channel_coded[0] = 1;
1055 for(i = 0; i < s->block_len; i++) {
1058 s->coefs[0][i] = a + b;
1059 s->coefs[1][i] = a - b;
1063 /* build the window : we ensure that when the windows overlap
1064 their squared sum is always 1 (MDCT reconstruction rule) */
1065 /* XXX: merge with output */
1067 int i, next_block_len, block_len, prev_block_len, n;
1070 block_len = s->block_len;
1071 prev_block_len = 1 << s->prev_block_len_bits;
1072 next_block_len = 1 << s->next_block_len_bits;
1075 wptr = window + block_len;
1076 if (block_len <= next_block_len) {
1077 for(i=0;i<block_len;i++)
1078 *wptr++ = s->windows[bsize][i];
1081 n = (block_len / 2) - (next_block_len / 2);
1084 for(i=0;i<next_block_len;i++)
1085 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1091 wptr = window + block_len;
1092 if (block_len <= prev_block_len) {
1093 for(i=0;i<block_len;i++)
1094 *--wptr = s->windows[bsize][i];
1097 n = (block_len / 2) - (prev_block_len / 2);
1100 for(i=0;i<prev_block_len;i++)
1101 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1108 for(ch = 0; ch < s->nb_channels; ch++) {
1109 if (s->channel_coded[ch]) {
1110 DECLARE_ALIGNED_16(FFTSample, output[BLOCK_MAX_SIZE * 2]);
1115 n4 = s->block_len / 2;
1116 s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize],
1117 output, s->coefs[ch], s->mdct_tmp);
1119 /* XXX: optimize all that by build the window and
1120 multipying/adding at the same time */
1122 /* multiply by the window and add in the frame */
1123 index = (s->frame_len / 2) + s->block_pos - n4;
1124 ptr = &s->frame_out[ch][index];
1125 s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1);
1127 /* specific fast case for ms-stereo : add to second
1128 channel if it is not coded */
1129 if (s->ms_stereo && !s->channel_coded[1]) {
1130 ptr = &s->frame_out[1][index];
1131 s->dsp.vector_fmul_add_add(ptr,window,output,ptr,0,2*n,1);
1136 /* update block number */
1138 s->block_pos += s->block_len;
1139 if (s->block_pos >= s->frame_len)
1145 /* decode a frame of frame_len samples */
1146 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1148 int ret, i, n, a, ch, incr;
1153 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
1156 /* read each block */
1160 ret = wma_decode_block(s);
1167 /* convert frame to integer */
1169 incr = s->nb_channels;
1170 for(ch = 0; ch < s->nb_channels; ch++) {
1172 iptr = s->frame_out[ch];
1175 a = lrintf(*iptr++);
1178 else if (a < -32768)
1183 /* prepare for next block */
1184 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1185 s->frame_len * sizeof(float));
1186 /* XXX: suppress this */
1187 memset(&s->frame_out[ch][s->frame_len], 0,
1188 s->frame_len * sizeof(float));
1192 dump_shorts("samples", samples, n * s->nb_channels);
1197 static int wma_decode_superframe(AVCodecContext *avctx,
1198 void *data, int *data_size,
1199 uint8_t *buf, int buf_size)
1201 WMADecodeContext *s = avctx->priv_data;
1202 int nb_frames, bit_offset, i, pos, len;
1206 tprintf("***decode_superframe:\n");
1209 s->last_superframe_len = 0;
1215 init_get_bits(&s->gb, buf, buf_size*8);
1217 if (s->use_bit_reservoir) {
1218 /* read super frame header */
1219 get_bits(&s->gb, 4); /* super frame index */
1220 nb_frames = get_bits(&s->gb, 4) - 1;
1222 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1224 if (s->last_superframe_len > 0) {
1225 // printf("skip=%d\n", s->last_bitoffset);
1226 /* add bit_offset bits to last frame */
1227 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1228 MAX_CODED_SUPERFRAME_SIZE)
1230 q = s->last_superframe + s->last_superframe_len;
1233 *q++ = (get_bits)(&s->gb, 8);
1237 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1240 /* XXX: bit_offset bits into last frame */
1241 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1242 /* skip unused bits */
1243 if (s->last_bitoffset > 0)
1244 skip_bits(&s->gb, s->last_bitoffset);
1245 /* this frame is stored in the last superframe and in the
1247 if (wma_decode_frame(s, samples) < 0)
1249 samples += s->nb_channels * s->frame_len;
1252 /* read each frame starting from bit_offset */
1253 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1254 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1257 skip_bits(&s->gb, len);
1259 s->reset_block_lengths = 1;
1260 for(i=0;i<nb_frames;i++) {
1261 if (wma_decode_frame(s, samples) < 0)
1263 samples += s->nb_channels * s->frame_len;
1266 /* we copy the end of the frame in the last frame buffer */
1267 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1268 s->last_bitoffset = pos & 7;
1270 len = buf_size - pos;
1271 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1274 s->last_superframe_len = len;
1275 memcpy(s->last_superframe, buf + pos, len);
1277 /* single frame decode */
1278 if (wma_decode_frame(s, samples) < 0)
1280 samples += s->nb_channels * s->frame_len;
1282 *data_size = (int8_t *)samples - (int8_t *)data;
1283 return s->block_align;
1285 /* when error, we reset the bit reservoir */
1286 s->last_superframe_len = 0;
1290 static int wma_decode_end(AVCodecContext *avctx)
1292 WMADecodeContext *s = avctx->priv_data;
1295 for(i = 0; i < s->nb_block_sizes; i++)
1296 ff_mdct_end(&s->mdct_ctx[i]);
1297 for(i = 0; i < s->nb_block_sizes; i++)
1298 av_free(s->windows[i]);
1300 if (s->use_exp_vlc) {
1301 free_vlc(&s->exp_vlc);
1303 if (s->use_noise_coding) {
1304 free_vlc(&s->hgain_vlc);
1306 for(i = 0;i < 2; i++) {
1307 free_vlc(&s->coef_vlc[i]);
1308 av_free(s->run_table[i]);
1309 av_free(s->level_table[i]);
1315 AVCodec wmav1_decoder =
1320 sizeof(WMADecodeContext),
1324 wma_decode_superframe,
1327 AVCodec wmav2_decoder =
1332 sizeof(WMADecodeContext),
1336 wma_decode_superframe,