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.
23 * This decoder handles Microsoft Windows Media Audio data, versions 1 & 2.
24 * WMA v1 is identified by audio format 0x160 in Microsoft media files
25 * (ASF/AVI/WAV). WMA v2 is identified by audio format 0x161.
27 * To use this decoder, a calling application must supply the extra data
28 * bytes provided with the WMA data. These are the extra, codec-specific
29 * bytes at the end of a WAVEFORMATEX data structure. Transmit these bytes
30 * to the decoder using the extradata[_size] fields in AVCodecContext. There
31 * should be 4 extra bytes for v1 data and 6 extra bytes for v2 data.
38 #define BLOCK_MIN_BITS 7
39 #define BLOCK_MAX_BITS 11
40 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
42 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
44 /* XXX: find exact max size */
45 #define HIGH_BAND_MAX_SIZE 16
47 #define NB_LSP_COEFS 10
49 /* XXX: is it a suitable value ? */
50 #define MAX_CODED_SUPERFRAME_SIZE 4096
52 #define MAX_CHANNELS 2
54 #define NOISE_TAB_SIZE 8192
56 #define LSP_POW_BITS 7
58 typedef struct WMADecodeContext {
63 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
65 int use_bit_reservoir;
66 int use_variable_block_len;
67 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
68 int use_noise_coding; /* true if perceptual noise is added */
71 int exponent_sizes[BLOCK_NB_SIZES];
72 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
73 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
74 int coefs_start; /* first coded coef */
75 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
76 int exponent_high_sizes[BLOCK_NB_SIZES];
77 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
80 /* coded values in high bands */
81 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
82 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
84 /* there are two possible tables for spectral coefficients */
86 uint16_t *run_table[2];
87 uint16_t *level_table[2];
89 int frame_len; /* frame length in samples */
90 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
91 int nb_block_sizes; /* number of block sizes */
93 int reset_block_lengths;
94 int block_len_bits; /* log2 of current block length */
95 int next_block_len_bits; /* log2 of next block length */
96 int prev_block_len_bits; /* log2 of prev block length */
97 int block_len; /* block length in samples */
98 int block_num; /* block number in current frame */
99 int block_pos; /* current position in frame */
100 uint8_t ms_stereo; /* true if mid/side stereo mode */
101 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
102 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
103 float max_exponent[MAX_CHANNELS];
104 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
105 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
106 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
107 float *windows[BLOCK_NB_SIZES];
108 FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
109 /* output buffer for one frame and the last for IMDCT windowing */
110 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
111 /* last frame info */
112 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
114 int last_superframe_len;
115 float noise_table[NOISE_TAB_SIZE];
117 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
118 /* lsp_to_curve tables */
119 float lsp_cos_table[BLOCK_MAX_SIZE];
120 float lsp_pow_e_table[256];
121 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
122 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
129 typedef struct CoefVLCTable {
130 int n; /* total number of codes */
131 const uint32_t *huffcodes; /* VLC bit values */
132 const uint8_t *huffbits; /* VLC bit size */
133 const uint16_t *levels; /* table to build run/level tables */
136 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
141 static void dump_shorts(const char *name, const short *tab, int n)
145 tprintf("%s[%d]:\n", name, n);
149 tprintf(" %5d.0", tab[i]);
155 static void dump_floats(const char *name, int prec, const float *tab, int n)
159 tprintf("%s[%d]:\n", name, n);
163 tprintf(" %8.*f", prec, tab[i]);
172 /* XXX: use same run/length optimization as mpeg decoders */
173 static void init_coef_vlc(VLC *vlc,
174 uint16_t **prun_table, uint16_t **plevel_table,
175 const CoefVLCTable *vlc_table)
177 int n = vlc_table->n;
178 const uint8_t *table_bits = vlc_table->huffbits;
179 const uint32_t *table_codes = vlc_table->huffcodes;
180 const uint16_t *levels_table = vlc_table->levels;
181 uint16_t *run_table, *level_table;
185 init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4);
187 run_table = av_malloc(n * sizeof(uint16_t));
188 level_table = av_malloc(n * sizeof(uint16_t));
196 level_table[i] = level;
201 *prun_table = run_table;
202 *plevel_table = level_table;
205 static int wma_decode_init(AVCodecContext * avctx)
207 WMADecodeContext *s = avctx->priv_data;
208 int i, flags1, flags2;
211 float bps1, high_freq;
216 s->sample_rate = avctx->sample_rate;
217 s->nb_channels = avctx->channels;
218 s->bit_rate = avctx->bit_rate;
219 s->block_align = avctx->block_align;
221 if (avctx->codec->id == CODEC_ID_WMAV1) {
227 /* extract flag infos */
230 extradata = avctx->extradata;
231 if (s->version == 1 && avctx->extradata_size >= 4) {
232 flags1 = extradata[0] | (extradata[1] << 8);
233 flags2 = extradata[2] | (extradata[3] << 8);
234 } else if (s->version == 2 && avctx->extradata_size >= 6) {
235 flags1 = extradata[0] | (extradata[1] << 8) |
236 (extradata[2] << 16) | (extradata[3] << 24);
237 flags2 = extradata[4] | (extradata[5] << 8);
239 s->use_exp_vlc = flags2 & 0x0001;
240 s->use_bit_reservoir = flags2 & 0x0002;
241 s->use_variable_block_len = flags2 & 0x0004;
243 /* compute MDCT block size */
244 if (s->sample_rate <= 16000) {
245 s->frame_len_bits = 9;
246 } else if (s->sample_rate <= 22050 ||
247 (s->sample_rate <= 32000 && s->version == 1)) {
248 s->frame_len_bits = 10;
250 s->frame_len_bits = 11;
252 s->frame_len = 1 << s->frame_len_bits;
253 if (s->use_variable_block_len) {
255 nb = ((flags2 >> 3) & 3) + 1;
256 if ((s->bit_rate / s->nb_channels) >= 32000)
258 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
261 s->nb_block_sizes = nb + 1;
263 s->nb_block_sizes = 1;
266 /* init rate dependant parameters */
267 s->use_noise_coding = 1;
268 high_freq = s->sample_rate * 0.5;
270 /* if version 2, then the rates are normalized */
271 sample_rate1 = s->sample_rate;
272 if (s->version == 2) {
273 if (sample_rate1 >= 44100)
274 sample_rate1 = 44100;
275 else if (sample_rate1 >= 22050)
276 sample_rate1 = 22050;
277 else if (sample_rate1 >= 16000)
278 sample_rate1 = 16000;
279 else if (sample_rate1 >= 11025)
280 sample_rate1 = 11025;
281 else if (sample_rate1 >= 8000)
285 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
286 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0)) + 2;
288 /* compute high frequency value and choose if noise coding should
291 if (s->nb_channels == 2)
293 if (sample_rate1 == 44100) {
295 s->use_noise_coding = 0;
297 high_freq = high_freq * 0.4;
298 } else if (sample_rate1 == 22050) {
300 s->use_noise_coding = 0;
301 else if (bps1 >= 0.72)
302 high_freq = high_freq * 0.7;
304 high_freq = high_freq * 0.6;
305 } else if (sample_rate1 == 16000) {
307 high_freq = high_freq * 0.5;
309 high_freq = high_freq * 0.3;
310 } else if (sample_rate1 == 11025) {
311 high_freq = high_freq * 0.7;
312 } else if (sample_rate1 == 8000) {
314 high_freq = high_freq * 0.5;
315 } else if (bps > 0.75) {
316 s->use_noise_coding = 0;
318 high_freq = high_freq * 0.65;
322 high_freq = high_freq * 0.75;
323 } else if (bps >= 0.6) {
324 high_freq = high_freq * 0.6;
326 high_freq = high_freq * 0.5;
329 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
330 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
331 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
333 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
334 bps, bps1, high_freq, s->byte_offset_bits);
335 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
336 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
338 /* compute the scale factor band sizes for each MDCT block size */
340 int a, b, pos, lpos, k, block_len, i, j, n;
341 const uint8_t *table;
343 if (s->version == 1) {
348 for(k = 0; k < s->nb_block_sizes; k++) {
349 block_len = s->frame_len >> k;
351 if (s->version == 1) {
354 a = wma_critical_freqs[i];
356 pos = ((block_len * 2 * a) + (b >> 1)) / b;
359 s->exponent_bands[0][i] = pos - lpos;
360 if (pos >= block_len) {
366 s->exponent_sizes[0] = i;
368 /* hardcoded tables */
370 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
372 if (s->sample_rate >= 44100)
373 table = exponent_band_44100[a];
374 else if (s->sample_rate >= 32000)
375 table = exponent_band_32000[a];
376 else if (s->sample_rate >= 22050)
377 table = exponent_band_22050[a];
382 s->exponent_bands[k][i] = table[i];
383 s->exponent_sizes[k] = n;
388 a = wma_critical_freqs[i];
390 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
395 s->exponent_bands[k][j++] = pos - lpos;
396 if (pos >= block_len)
400 s->exponent_sizes[k] = j;
404 /* max number of coefs */
405 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
406 /* high freq computation */
407 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
408 s->sample_rate + 0.5);
409 n = s->exponent_sizes[k];
415 pos += s->exponent_bands[k][i];
417 if (start < s->high_band_start[k])
418 start = s->high_band_start[k];
419 if (end > s->coefs_end[k])
420 end = s->coefs_end[k];
422 s->exponent_high_bands[k][j++] = end - start;
424 s->exponent_high_sizes[k] = j;
426 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
429 s->high_band_start[k],
430 s->exponent_high_sizes[k]);
431 for(j=0;j<s->exponent_high_sizes[k];j++)
432 tprintf(" %d", s->exponent_high_bands[k][j]);
441 for(i = 0; i < s->nb_block_sizes; i++) {
442 tprintf("%5d: n=%2d:",
444 s->exponent_sizes[i]);
445 for(j=0;j<s->exponent_sizes[i];j++)
446 tprintf(" %d", s->exponent_bands[i][j]);
453 for(i = 0; i < s->nb_block_sizes; i++)
454 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
456 /* init MDCT windows : simple sinus window */
457 for(i = 0; i < s->nb_block_sizes; i++) {
460 n = 1 << (s->frame_len_bits - i);
461 window = av_malloc(sizeof(float) * n);
462 alpha = M_PI / (2.0 * n);
464 window[n - j - 1] = sin((j + 0.5) * alpha);
466 s->windows[i] = window;
469 s->reset_block_lengths = 1;
471 if (s->use_noise_coding) {
473 /* init the noise generator */
475 s->noise_mult = 0.02;
477 s->noise_mult = 0.04;
480 for(i=0;i<NOISE_TAB_SIZE;i++)
481 s->noise_table[i] = 1.0 * s->noise_mult;
487 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
488 for(i=0;i<NOISE_TAB_SIZE;i++) {
489 seed = seed * 314159 + 1;
490 s->noise_table[i] = (float)((int)seed) * norm;
494 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
495 hgain_huffbits, 1, 1,
496 hgain_huffcodes, 2, 2);
499 if (s->use_exp_vlc) {
500 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
501 scale_huffbits, 1, 1,
502 scale_huffcodes, 4, 4);
504 wma_lsp_to_curve_init(s, s->frame_len);
507 /* choose the VLC tables for the coefficients */
509 if (s->sample_rate >= 32000) {
512 else if (bps1 < 1.16)
516 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
517 &coef_vlcs[coef_vlc_table * 2]);
518 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
519 &coef_vlcs[coef_vlc_table * 2 + 1]);
523 /* interpolate values for a bigger or smaller block. The block must
524 have multiple sizes */
525 static void interpolate_array(float *scale, int old_size, int new_size)
530 if (new_size > old_size) {
531 jincr = new_size / old_size;
533 for(i = old_size - 1; i >=0; i--) {
540 } else if (new_size < old_size) {
542 jincr = old_size / new_size;
543 for(i = 0; i < new_size; i++) {
550 /* compute x^-0.25 with an exponent and mantissa table. We use linear
551 interpolation to reduce the mantissa table size at a small speed
552 expense (linear interpolation approximately doubles the number of
553 bits of precision). */
554 static inline float pow_m1_4(WMADecodeContext *s, float x)
565 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
566 /* build interpolation scale: 1 <= t < 2. */
567 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
568 a = s->lsp_pow_m_table1[m];
569 b = s->lsp_pow_m_table2[m];
570 return s->lsp_pow_e_table[e] * (a + b * t.f);
573 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
578 wdel = M_PI / frame_len;
579 for(i=0;i<frame_len;i++)
580 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
582 /* tables for x^-0.25 computation */
585 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
588 /* NOTE: these two tables are needed to avoid two operations in
591 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
592 m = (1 << LSP_POW_BITS) + i;
593 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
595 s->lsp_pow_m_table1[i] = 2 * a - b;
596 s->lsp_pow_m_table2[i] = b - a;
605 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
610 /* NOTE: We use the same code as Vorbis here */
611 /* XXX: optimize it further with SSE/3Dnow */
612 static void wma_lsp_to_curve(WMADecodeContext *s,
613 float *out, float *val_max_ptr,
617 float p, q, w, v, val_max;
623 w = s->lsp_cos_table[i];
624 for(j=1;j<NB_LSP_COEFS;j+=2){
636 *val_max_ptr = val_max;
639 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
640 static void decode_exp_lsp(WMADecodeContext *s, int ch)
642 float lsp_coefs[NB_LSP_COEFS];
645 for(i = 0; i < NB_LSP_COEFS; i++) {
646 if (i == 0 || i >= 8)
647 val = get_bits(&s->gb, 3);
649 val = get_bits(&s->gb, 4);
650 lsp_coefs[i] = lsp_codebook[i][val];
653 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
654 s->block_len, lsp_coefs);
657 /* decode exponents coded with VLC codes */
658 static int decode_exp_vlc(WMADecodeContext *s, int ch)
660 int last_exp, n, code;
661 const uint16_t *ptr, *band_ptr;
662 float v, *q, max_scale, *q_end;
664 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
666 q = s->exponents[ch];
667 q_end = q + s->block_len;
669 if (s->version == 1) {
670 last_exp = get_bits(&s->gb, 5) + 10;
671 /* XXX: use a table */
672 v = pow(10, last_exp * (1.0 / 16.0));
681 code = get_vlc(&s->gb, &s->exp_vlc);
684 /* NOTE: this offset is the same as MPEG4 AAC ! */
685 last_exp += code - 60;
686 /* XXX: use a table */
687 v = pow(10, last_exp * (1.0 / 16.0));
695 s->max_exponent[ch] = max_scale;
699 /* return 0 if OK. return 1 if last block of frame. return -1 if
700 unrecorrable error. */
701 static int wma_decode_block(WMADecodeContext *s)
703 int n, v, a, ch, code, bsize;
704 int coef_nb_bits, total_gain, parse_exponents;
705 float window[BLOCK_MAX_SIZE * 2];
706 // XXX: FIXME!! there's a bug somewhere which makes this mandatory under altivec
708 volatile int nb_coefs[MAX_CHANNELS] __attribute__((aligned(16)));
710 int nb_coefs[MAX_CHANNELS];
715 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
718 /* compute current block length */
719 if (s->use_variable_block_len) {
720 n = av_log2(s->nb_block_sizes - 1) + 1;
722 if (s->reset_block_lengths) {
723 s->reset_block_lengths = 0;
724 v = get_bits(&s->gb, n);
725 if (v >= s->nb_block_sizes)
727 s->prev_block_len_bits = s->frame_len_bits - v;
728 v = get_bits(&s->gb, n);
729 if (v >= s->nb_block_sizes)
731 s->block_len_bits = s->frame_len_bits - v;
733 /* update block lengths */
734 s->prev_block_len_bits = s->block_len_bits;
735 s->block_len_bits = s->next_block_len_bits;
737 v = get_bits(&s->gb, n);
738 if (v >= s->nb_block_sizes)
740 s->next_block_len_bits = s->frame_len_bits - v;
742 /* fixed block len */
743 s->next_block_len_bits = s->frame_len_bits;
744 s->prev_block_len_bits = s->frame_len_bits;
745 s->block_len_bits = s->frame_len_bits;
748 /* now check if the block length is coherent with the frame length */
749 s->block_len = 1 << s->block_len_bits;
750 if ((s->block_pos + s->block_len) > s->frame_len)
753 if (s->nb_channels == 2) {
754 s->ms_stereo = get_bits(&s->gb, 1);
757 for(ch = 0; ch < s->nb_channels; ch++) {
758 a = get_bits(&s->gb, 1);
759 s->channel_coded[ch] = a;
762 /* if no channel coded, no need to go further */
763 /* XXX: fix potential framing problems */
767 bsize = s->frame_len_bits - s->block_len_bits;
769 /* read total gain and extract corresponding number of bits for
770 coef escape coding */
773 a = get_bits(&s->gb, 7);
781 else if (total_gain < 32)
783 else if (total_gain < 40)
785 else if (total_gain < 45)
790 /* compute number of coefficients */
791 n = s->coefs_end[bsize] - s->coefs_start;
792 for(ch = 0; ch < s->nb_channels; ch++)
796 if (s->use_noise_coding) {
798 for(ch = 0; ch < s->nb_channels; ch++) {
799 if (s->channel_coded[ch]) {
801 n = s->exponent_high_sizes[bsize];
803 a = get_bits(&s->gb, 1);
804 s->high_band_coded[ch][i] = a;
805 /* if noise coding, the coefficients are not transmitted */
807 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
811 for(ch = 0; ch < s->nb_channels; ch++) {
812 if (s->channel_coded[ch]) {
815 n = s->exponent_high_sizes[bsize];
816 val = (int)0x80000000;
818 if (s->high_band_coded[ch][i]) {
819 if (val == (int)0x80000000) {
820 val = get_bits(&s->gb, 7) - 19;
822 code = get_vlc(&s->gb, &s->hgain_vlc);
827 s->high_band_values[ch][i] = val;
834 /* exposant can be interpolated in short blocks. */
836 if (s->block_len_bits != s->frame_len_bits) {
837 parse_exponents = get_bits(&s->gb, 1);
840 if (parse_exponents) {
841 for(ch = 0; ch < s->nb_channels; ch++) {
842 if (s->channel_coded[ch]) {
843 if (s->use_exp_vlc) {
844 if (decode_exp_vlc(s, ch) < 0)
847 decode_exp_lsp(s, ch);
852 for(ch = 0; ch < s->nb_channels; ch++) {
853 if (s->channel_coded[ch]) {
854 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
860 /* parse spectral coefficients : just RLE encoding */
861 for(ch = 0; ch < s->nb_channels; ch++) {
862 if (s->channel_coded[ch]) {
864 int level, run, sign, tindex;
866 const int16_t *level_table, *run_table;
868 /* special VLC tables are used for ms stereo because
869 there is potentially less energy there */
870 tindex = (ch == 1 && s->ms_stereo);
871 coef_vlc = &s->coef_vlc[tindex];
872 run_table = s->run_table[tindex];
873 level_table = s->level_table[tindex];
875 ptr = &s->coefs1[ch][0];
876 eptr = ptr + nb_coefs[ch];
877 memset(ptr, 0, s->block_len * sizeof(int16_t));
879 code = get_vlc(&s->gb, coef_vlc);
885 } else if (code == 0) {
887 level = get_bits(&s->gb, coef_nb_bits);
888 /* NOTE: this is rather suboptimal. reading
889 block_len_bits would be better */
890 run = get_bits(&s->gb, s->frame_len_bits);
893 run = run_table[code];
894 level = level_table[code];
896 sign = get_bits(&s->gb, 1);
903 /* NOTE: EOB can be omitted */
908 if (s->version == 1 && s->nb_channels >= 2) {
909 align_get_bits(&s->gb);
915 int n4 = s->block_len / 2;
916 mdct_norm = 1.0 / (float)n4;
917 if (s->version == 1) {
918 mdct_norm *= sqrt(n4);
922 /* finally compute the MDCT coefficients */
923 for(ch = 0; ch < s->nb_channels; ch++) {
924 if (s->channel_coded[ch]) {
926 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
927 int i, j, n, n1, last_high_band;
928 float exp_power[HIGH_BAND_MAX_SIZE];
930 coefs1 = s->coefs1[ch];
931 exponents = s->exponents[ch];
932 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
934 coefs = s->coefs[ch];
935 if (s->use_noise_coding) {
937 /* very low freqs : noise */
938 for(i = 0;i < s->coefs_start; i++) {
939 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
940 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
943 n1 = s->exponent_high_sizes[bsize];
945 /* compute power of high bands */
946 exp_ptr = exponents +
947 s->high_band_start[bsize] -
949 last_high_band = 0; /* avoid warning */
951 n = s->exponent_high_bands[s->frame_len_bits -
952 s->block_len_bits][j];
953 if (s->high_band_coded[ch][j]) {
956 for(i = 0;i < n; i++) {
960 exp_power[j] = e2 / n;
962 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
967 /* main freqs and high freqs */
970 n = s->high_band_start[bsize] -
973 n = s->exponent_high_bands[s->frame_len_bits -
974 s->block_len_bits][j];
976 if (j >= 0 && s->high_band_coded[ch][j]) {
977 /* use noise with specified power */
978 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
979 /* XXX: use a table */
980 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
981 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
983 for(i = 0;i < n; i++) {
984 noise = s->noise_table[s->noise_index];
985 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
986 *coefs++ = (*exponents++) * noise * mult1;
989 /* coded values + small noise */
990 for(i = 0;i < n; i++) {
991 noise = s->noise_table[s->noise_index];
992 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
993 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
998 /* very high freqs : noise */
999 n = s->block_len - s->coefs_end[bsize];
1000 mult1 = mult * exponents[-1];
1001 for(i = 0; i < n; i++) {
1002 *coefs++ = s->noise_table[s->noise_index] * mult1;
1003 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1006 /* XXX: optimize more */
1007 for(i = 0;i < s->coefs_start; i++)
1010 for(i = 0;i < n; i++) {
1011 *coefs++ = coefs1[i] * exponents[i] * mult;
1013 n = s->block_len - s->coefs_end[bsize];
1014 for(i = 0;i < n; i++)
1021 for(ch = 0; ch < s->nb_channels; ch++) {
1022 if (s->channel_coded[ch]) {
1023 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1024 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1029 if (s->ms_stereo && s->channel_coded[1]) {
1033 /* nominal case for ms stereo: we do it before mdct */
1034 /* no need to optimize this case because it should almost
1036 if (!s->channel_coded[0]) {
1037 tprintf("rare ms-stereo case happened\n");
1038 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1039 s->channel_coded[0] = 1;
1042 for(i = 0; i < s->block_len; i++) {
1045 s->coefs[0][i] = a + b;
1046 s->coefs[1][i] = a - b;
1050 /* build the window : we ensure that when the windows overlap
1051 their squared sum is always 1 (MDCT reconstruction rule) */
1052 /* XXX: merge with output */
1054 int i, next_block_len, block_len, prev_block_len, n;
1057 block_len = s->block_len;
1058 prev_block_len = 1 << s->prev_block_len_bits;
1059 next_block_len = 1 << s->next_block_len_bits;
1062 wptr = window + block_len;
1063 if (block_len <= next_block_len) {
1064 for(i=0;i<block_len;i++)
1065 *wptr++ = s->windows[bsize][i];
1068 n = (block_len / 2) - (next_block_len / 2);
1071 for(i=0;i<next_block_len;i++)
1072 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1078 wptr = window + block_len;
1079 if (block_len <= prev_block_len) {
1080 for(i=0;i<block_len;i++)
1081 *--wptr = s->windows[bsize][i];
1084 n = (block_len / 2) - (prev_block_len / 2);
1087 for(i=0;i<prev_block_len;i++)
1088 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1095 for(ch = 0; ch < s->nb_channels; ch++) {
1096 if (s->channel_coded[ch]) {
1097 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1099 int i, n4, index, n;
1102 n4 = s->block_len / 2;
1103 ff_imdct_calc(&s->mdct_ctx[bsize],
1104 output, s->coefs[ch], s->mdct_tmp);
1106 /* XXX: optimize all that by build the window and
1107 multipying/adding at the same time */
1108 /* multiply by the window */
1109 for(i=0;i<n * 2;i++) {
1110 output[i] *= window[i];
1113 /* add in the frame */
1114 index = (s->frame_len / 2) + s->block_pos - n4;
1115 ptr = &s->frame_out[ch][index];
1116 for(i=0;i<n * 2;i++) {
1121 /* specific fast case for ms-stereo : add to second
1122 channel if it is not coded */
1123 if (s->ms_stereo && !s->channel_coded[1]) {
1124 ptr = &s->frame_out[1][index];
1125 for(i=0;i<n * 2;i++) {
1133 /* update block number */
1135 s->block_pos += s->block_len;
1136 if (s->block_pos >= s->frame_len)
1142 /* decode a frame of frame_len samples */
1143 static int wma_decode_frame(WMADecodeContext *s, int16_t *samples)
1145 int ret, i, n, a, ch, incr;
1150 tprintf("***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
1153 /* read each block */
1157 ret = wma_decode_block(s);
1164 /* convert frame to integer */
1166 incr = s->nb_channels;
1167 for(ch = 0; ch < s->nb_channels; ch++) {
1169 iptr = s->frame_out[ch];
1172 a = lrintf(*iptr++);
1175 else if (a < -32768)
1180 /* prepare for next block */
1181 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
1182 s->frame_len * sizeof(float));
1183 /* XXX: suppress this */
1184 memset(&s->frame_out[ch][s->frame_len], 0,
1185 s->frame_len * sizeof(float));
1189 dump_shorts("samples", samples, n * s->nb_channels);
1194 static int wma_decode_superframe(AVCodecContext *avctx,
1195 void *data, int *data_size,
1196 uint8_t *buf, int buf_size)
1198 WMADecodeContext *s = avctx->priv_data;
1199 int nb_frames, bit_offset, i, pos, len;
1203 tprintf("***decode_superframe:\n");
1206 s->last_superframe_len = 0;
1212 init_get_bits(&s->gb, buf, buf_size*8);
1214 if (s->use_bit_reservoir) {
1215 /* read super frame header */
1216 get_bits(&s->gb, 4); /* super frame index */
1217 nb_frames = get_bits(&s->gb, 4) - 1;
1219 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
1221 if (s->last_superframe_len > 0) {
1222 // printf("skip=%d\n", s->last_bitoffset);
1223 /* add bit_offset bits to last frame */
1224 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
1225 MAX_CODED_SUPERFRAME_SIZE)
1227 q = s->last_superframe + s->last_superframe_len;
1230 *q++ = (get_bits)(&s->gb, 8);
1234 *q++ = (get_bits)(&s->gb, len) << (8 - len);
1237 /* XXX: bit_offset bits into last frame */
1238 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
1239 /* skip unused bits */
1240 if (s->last_bitoffset > 0)
1241 skip_bits(&s->gb, s->last_bitoffset);
1242 /* this frame is stored in the last superframe and in the
1244 if (wma_decode_frame(s, samples) < 0)
1246 samples += s->nb_channels * s->frame_len;
1249 /* read each frame starting from bit_offset */
1250 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
1251 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
1254 skip_bits(&s->gb, len);
1256 s->reset_block_lengths = 1;
1257 for(i=0;i<nb_frames;i++) {
1258 if (wma_decode_frame(s, samples) < 0)
1260 samples += s->nb_channels * s->frame_len;
1263 /* we copy the end of the frame in the last frame buffer */
1264 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
1265 s->last_bitoffset = pos & 7;
1267 len = buf_size - pos;
1268 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
1271 s->last_superframe_len = len;
1272 memcpy(s->last_superframe, buf + pos, len);
1274 /* single frame decode */
1275 if (wma_decode_frame(s, samples) < 0)
1277 samples += s->nb_channels * s->frame_len;
1279 *data_size = (int8_t *)samples - (int8_t *)data;
1280 return s->block_align;
1282 /* when error, we reset the bit reservoir */
1283 s->last_superframe_len = 0;
1287 static int wma_decode_end(AVCodecContext *avctx)
1289 WMADecodeContext *s = avctx->priv_data;
1292 for(i = 0; i < s->nb_block_sizes; i++)
1293 ff_mdct_end(&s->mdct_ctx[i]);
1294 for(i = 0; i < s->nb_block_sizes; i++)
1295 av_free(s->windows[i]);
1297 if (s->use_exp_vlc) {
1298 free_vlc(&s->exp_vlc);
1300 if (s->use_noise_coding) {
1301 free_vlc(&s->hgain_vlc);
1303 for(i = 0;i < 2; i++) {
1304 free_vlc(&s->coef_vlc[i]);
1305 av_free(s->run_table[i]);
1306 av_free(s->level_table[i]);
1312 AVCodec wmav1_decoder =
1317 sizeof(WMADecodeContext),
1321 wma_decode_superframe,
1324 AVCodec wmav2_decoder =
1329 sizeof(WMADecodeContext),
1333 wma_decode_superframe,