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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 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.
35 #include "bitstream.h"
39 #define BLOCK_MIN_BITS 7
40 #define BLOCK_MAX_BITS 11
41 #define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)
43 #define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)
45 /* XXX: find exact max size */
46 #define HIGH_BAND_MAX_SIZE 16
48 #define NB_LSP_COEFS 10
50 /* XXX: is it a suitable value ? */
51 #define MAX_CODED_SUPERFRAME_SIZE 16384
53 #define MAX_CHANNELS 2
55 #define NOISE_TAB_SIZE 8192
57 #define LSP_POW_BITS 7
61 typedef struct WMADecodeContext {
66 int version; /* 1 = 0x160 (WMAV1), 2 = 0x161 (WMAV2) */
68 int use_bit_reservoir;
69 int use_variable_block_len;
70 int use_exp_vlc; /* exponent coding: 0 = lsp, 1 = vlc + delta */
71 int use_noise_coding; /* true if perceptual noise is added */
74 int exponent_sizes[BLOCK_NB_SIZES];
75 uint16_t exponent_bands[BLOCK_NB_SIZES][25];
76 int high_band_start[BLOCK_NB_SIZES]; /* index of first coef in high band */
77 int coefs_start; /* first coded coef */
78 int coefs_end[BLOCK_NB_SIZES]; /* max number of coded coefficients */
79 int exponent_high_sizes[BLOCK_NB_SIZES];
80 int exponent_high_bands[BLOCK_NB_SIZES][HIGH_BAND_MAX_SIZE];
83 /* coded values in high bands */
84 int high_band_coded[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
85 int high_band_values[MAX_CHANNELS][HIGH_BAND_MAX_SIZE];
87 /* there are two possible tables for spectral coefficients */
89 uint16_t *run_table[2];
90 uint16_t *level_table[2];
92 int frame_len; /* frame length in samples */
93 int frame_len_bits; /* frame_len = 1 << frame_len_bits */
94 int nb_block_sizes; /* number of block sizes */
96 int reset_block_lengths;
97 int block_len_bits; /* log2 of current block length */
98 int next_block_len_bits; /* log2 of next block length */
99 int prev_block_len_bits; /* log2 of prev block length */
100 int block_len; /* block length in samples */
101 int block_num; /* block number in current frame */
102 int block_pos; /* current position in frame */
103 uint8_t ms_stereo; /* true if mid/side stereo mode */
104 uint8_t channel_coded[MAX_CHANNELS]; /* true if channel is coded */
105 float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
106 float max_exponent[MAX_CHANNELS];
107 int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
108 float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE] __attribute__((aligned(16)));
109 MDCTContext mdct_ctx[BLOCK_NB_SIZES];
110 float *windows[BLOCK_NB_SIZES];
111 FFTSample mdct_tmp[BLOCK_MAX_SIZE] __attribute__((aligned(16))); /* temporary storage for imdct */
112 /* output buffer for one frame and the last for IMDCT windowing */
113 float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
114 /* last frame info */
115 uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
117 int last_superframe_len;
118 float noise_table[NOISE_TAB_SIZE];
120 float noise_mult; /* XXX: suppress that and integrate it in the noise array */
121 /* lsp_to_curve tables */
122 float lsp_cos_table[BLOCK_MAX_SIZE];
123 float lsp_pow_e_table[256];
124 float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
125 float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
132 typedef struct CoefVLCTable {
133 int n; /* total number of codes */
134 const uint32_t *huffcodes; /* VLC bit values */
135 const uint8_t *huffbits; /* VLC bit size */
136 const uint16_t *levels; /* table to build run/level tables */
139 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len);
144 static void dump_shorts(const char *name, const short *tab, int n)
148 tprintf("%s[%d]:\n", name, n);
152 tprintf(" %5d.0", tab[i]);
158 static void dump_floats(const char *name, int prec, const float *tab, int n)
162 tprintf("%s[%d]:\n", name, n);
166 tprintf(" %8.*f", prec, tab[i]);
175 /* XXX: use same run/length optimization as mpeg decoders */
176 static void init_coef_vlc(VLC *vlc,
177 uint16_t **prun_table, uint16_t **plevel_table,
178 const CoefVLCTable *vlc_table)
180 int n = vlc_table->n;
181 const uint8_t *table_bits = vlc_table->huffbits;
182 const uint32_t *table_codes = vlc_table->huffcodes;
183 const uint16_t *levels_table = vlc_table->levels;
184 uint16_t *run_table, *level_table;
188 init_vlc(vlc, 9, n, table_bits, 1, 1, table_codes, 4, 4, 0);
190 run_table = av_malloc(n * sizeof(uint16_t));
191 level_table = av_malloc(n * sizeof(uint16_t));
199 level_table[i] = level;
204 *prun_table = run_table;
205 *plevel_table = level_table;
208 static int wma_decode_init(AVCodecContext * avctx)
210 WMADecodeContext *s = avctx->priv_data;
211 int i, flags1, flags2;
214 float bps1, high_freq;
219 s->sample_rate = avctx->sample_rate;
220 s->nb_channels = avctx->channels;
221 s->bit_rate = avctx->bit_rate;
222 s->block_align = avctx->block_align;
224 if (avctx->codec->id == CODEC_ID_WMAV1) {
230 /* extract flag infos */
233 extradata = avctx->extradata;
234 if (s->version == 1 && avctx->extradata_size >= 4) {
235 flags1 = extradata[0] | (extradata[1] << 8);
236 flags2 = extradata[2] | (extradata[3] << 8);
237 } else if (s->version == 2 && avctx->extradata_size >= 6) {
238 flags1 = extradata[0] | (extradata[1] << 8) |
239 (extradata[2] << 16) | (extradata[3] << 24);
240 flags2 = extradata[4] | (extradata[5] << 8);
242 s->use_exp_vlc = flags2 & 0x0001;
243 s->use_bit_reservoir = flags2 & 0x0002;
244 s->use_variable_block_len = flags2 & 0x0004;
246 /* compute MDCT block size */
247 if (s->sample_rate <= 16000) {
248 s->frame_len_bits = 9;
249 } else if (s->sample_rate <= 22050 ||
250 (s->sample_rate <= 32000 && s->version == 1)) {
251 s->frame_len_bits = 10;
253 s->frame_len_bits = 11;
255 s->frame_len = 1 << s->frame_len_bits;
256 if (s->use_variable_block_len) {
258 nb = ((flags2 >> 3) & 3) + 1;
259 if ((s->bit_rate / s->nb_channels) >= 32000)
261 nb_max = s->frame_len_bits - BLOCK_MIN_BITS;
264 s->nb_block_sizes = nb + 1;
266 s->nb_block_sizes = 1;
269 /* init rate dependant parameters */
270 s->use_noise_coding = 1;
271 high_freq = s->sample_rate * 0.5;
273 /* if version 2, then the rates are normalized */
274 sample_rate1 = s->sample_rate;
275 if (s->version == 2) {
276 if (sample_rate1 >= 44100)
277 sample_rate1 = 44100;
278 else if (sample_rate1 >= 22050)
279 sample_rate1 = 22050;
280 else if (sample_rate1 >= 16000)
281 sample_rate1 = 16000;
282 else if (sample_rate1 >= 11025)
283 sample_rate1 = 11025;
284 else if (sample_rate1 >= 8000)
288 bps = (float)s->bit_rate / (float)(s->nb_channels * s->sample_rate);
289 s->byte_offset_bits = av_log2((int)(bps * s->frame_len / 8.0 + 0.5)) + 2;
291 /* compute high frequency value and choose if noise coding should
294 if (s->nb_channels == 2)
296 if (sample_rate1 == 44100) {
298 s->use_noise_coding = 0;
300 high_freq = high_freq * 0.4;
301 } else if (sample_rate1 == 22050) {
303 s->use_noise_coding = 0;
304 else if (bps1 >= 0.72)
305 high_freq = high_freq * 0.7;
307 high_freq = high_freq * 0.6;
308 } else if (sample_rate1 == 16000) {
310 high_freq = high_freq * 0.5;
312 high_freq = high_freq * 0.3;
313 } else if (sample_rate1 == 11025) {
314 high_freq = high_freq * 0.7;
315 } else if (sample_rate1 == 8000) {
317 high_freq = high_freq * 0.5;
318 } else if (bps > 0.75) {
319 s->use_noise_coding = 0;
321 high_freq = high_freq * 0.65;
325 high_freq = high_freq * 0.75;
326 } else if (bps >= 0.6) {
327 high_freq = high_freq * 0.6;
329 high_freq = high_freq * 0.5;
332 dprintf("flags1=0x%x flags2=0x%x\n", flags1, flags2);
333 dprintf("version=%d channels=%d sample_rate=%d bitrate=%d block_align=%d\n",
334 s->version, s->nb_channels, s->sample_rate, s->bit_rate,
336 dprintf("bps=%f bps1=%f high_freq=%f bitoffset=%d\n",
337 bps, bps1, high_freq, s->byte_offset_bits);
338 dprintf("use_noise_coding=%d use_exp_vlc=%d nb_block_sizes=%d\n",
339 s->use_noise_coding, s->use_exp_vlc, s->nb_block_sizes);
341 /* compute the scale factor band sizes for each MDCT block size */
343 int a, b, pos, lpos, k, block_len, i, j, n;
344 const uint8_t *table;
346 if (s->version == 1) {
351 for(k = 0; k < s->nb_block_sizes; k++) {
352 block_len = s->frame_len >> k;
354 if (s->version == 1) {
357 a = wma_critical_freqs[i];
359 pos = ((block_len * 2 * a) + (b >> 1)) / b;
362 s->exponent_bands[0][i] = pos - lpos;
363 if (pos >= block_len) {
369 s->exponent_sizes[0] = i;
371 /* hardcoded tables */
373 a = s->frame_len_bits - BLOCK_MIN_BITS - k;
375 if (s->sample_rate >= 44100)
376 table = exponent_band_44100[a];
377 else if (s->sample_rate >= 32000)
378 table = exponent_band_32000[a];
379 else if (s->sample_rate >= 22050)
380 table = exponent_band_22050[a];
385 s->exponent_bands[k][i] = table[i];
386 s->exponent_sizes[k] = n;
391 a = wma_critical_freqs[i];
393 pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
398 s->exponent_bands[k][j++] = pos - lpos;
399 if (pos >= block_len)
403 s->exponent_sizes[k] = j;
407 /* max number of coefs */
408 s->coefs_end[k] = (s->frame_len - ((s->frame_len * 9) / 100)) >> k;
409 /* high freq computation */
410 s->high_band_start[k] = (int)((block_len * 2 * high_freq) /
411 s->sample_rate + 0.5);
412 n = s->exponent_sizes[k];
418 pos += s->exponent_bands[k][i];
420 if (start < s->high_band_start[k])
421 start = s->high_band_start[k];
422 if (end > s->coefs_end[k])
423 end = s->coefs_end[k];
425 s->exponent_high_bands[k][j++] = end - start;
427 s->exponent_high_sizes[k] = j;
429 tprintf("%5d: coefs_end=%d high_band_start=%d nb_high_bands=%d: ",
432 s->high_band_start[k],
433 s->exponent_high_sizes[k]);
434 for(j=0;j<s->exponent_high_sizes[k];j++)
435 tprintf(" %d", s->exponent_high_bands[k][j]);
444 for(i = 0; i < s->nb_block_sizes; i++) {
445 tprintf("%5d: n=%2d:",
447 s->exponent_sizes[i]);
448 for(j=0;j<s->exponent_sizes[i];j++)
449 tprintf(" %d", s->exponent_bands[i][j]);
456 for(i = 0; i < s->nb_block_sizes; i++)
457 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
459 /* init MDCT windows : simple sinus window */
460 for(i = 0; i < s->nb_block_sizes; i++) {
463 n = 1 << (s->frame_len_bits - i);
464 window = av_malloc(sizeof(float) * n);
465 alpha = M_PI / (2.0 * n);
467 window[n - j - 1] = sin((j + 0.5) * alpha);
469 s->windows[i] = window;
472 s->reset_block_lengths = 1;
474 if (s->use_noise_coding) {
476 /* init the noise generator */
478 s->noise_mult = 0.02;
480 s->noise_mult = 0.04;
483 for(i=0;i<NOISE_TAB_SIZE;i++)
484 s->noise_table[i] = 1.0 * s->noise_mult;
490 norm = (1.0 / (float)(1LL << 31)) * sqrt(3) * s->noise_mult;
491 for(i=0;i<NOISE_TAB_SIZE;i++) {
492 seed = seed * 314159 + 1;
493 s->noise_table[i] = (float)((int)seed) * norm;
497 init_vlc(&s->hgain_vlc, 9, sizeof(hgain_huffbits),
498 hgain_huffbits, 1, 1,
499 hgain_huffcodes, 2, 2, 0);
502 if (s->use_exp_vlc) {
503 init_vlc(&s->exp_vlc, 9, sizeof(scale_huffbits),
504 scale_huffbits, 1, 1,
505 scale_huffcodes, 4, 4, 0);
507 wma_lsp_to_curve_init(s, s->frame_len);
510 /* choose the VLC tables for the coefficients */
512 if (s->sample_rate >= 32000) {
515 else if (bps1 < 1.16)
519 init_coef_vlc(&s->coef_vlc[0], &s->run_table[0], &s->level_table[0],
520 &coef_vlcs[coef_vlc_table * 2]);
521 init_coef_vlc(&s->coef_vlc[1], &s->run_table[1], &s->level_table[1],
522 &coef_vlcs[coef_vlc_table * 2 + 1]);
526 /* interpolate values for a bigger or smaller block. The block must
527 have multiple sizes */
528 static void interpolate_array(float *scale, int old_size, int new_size)
533 if (new_size > old_size) {
534 jincr = new_size / old_size;
536 for(i = old_size - 1; i >=0; i--) {
543 } else if (new_size < old_size) {
545 jincr = old_size / new_size;
546 for(i = 0; i < new_size; i++) {
553 /* compute x^-0.25 with an exponent and mantissa table. We use linear
554 interpolation to reduce the mantissa table size at a small speed
555 expense (linear interpolation approximately doubles the number of
556 bits of precision). */
557 static inline float pow_m1_4(WMADecodeContext *s, float x)
568 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
569 /* build interpolation scale: 1 <= t < 2. */
570 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
571 a = s->lsp_pow_m_table1[m];
572 b = s->lsp_pow_m_table2[m];
573 return s->lsp_pow_e_table[e] * (a + b * t.f);
576 static void wma_lsp_to_curve_init(WMADecodeContext *s, int frame_len)
581 wdel = M_PI / frame_len;
582 for(i=0;i<frame_len;i++)
583 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
585 /* tables for x^-0.25 computation */
588 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
591 /* NOTE: these two tables are needed to avoid two operations in
594 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
595 m = (1 << LSP_POW_BITS) + i;
596 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
598 s->lsp_pow_m_table1[i] = 2 * a - b;
599 s->lsp_pow_m_table2[i] = b - a;
608 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
613 /* NOTE: We use the same code as Vorbis here */
614 /* XXX: optimize it further with SSE/3Dnow */
615 static void wma_lsp_to_curve(WMADecodeContext *s,
616 float *out, float *val_max_ptr,
620 float p, q, w, v, val_max;
626 w = s->lsp_cos_table[i];
627 for(j=1;j<NB_LSP_COEFS;j+=2){
639 *val_max_ptr = val_max;
642 /* decode exponents coded with LSP coefficients (same idea as Vorbis) */
643 static void decode_exp_lsp(WMADecodeContext *s, int ch)
645 float lsp_coefs[NB_LSP_COEFS];
648 for(i = 0; i < NB_LSP_COEFS; i++) {
649 if (i == 0 || i >= 8)
650 val = get_bits(&s->gb, 3);
652 val = get_bits(&s->gb, 4);
653 lsp_coefs[i] = lsp_codebook[i][val];
656 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
657 s->block_len, lsp_coefs);
660 /* decode exponents coded with VLC codes */
661 static int decode_exp_vlc(WMADecodeContext *s, int ch)
663 int last_exp, n, code;
664 const uint16_t *ptr, *band_ptr;
665 float v, *q, max_scale, *q_end;
667 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
669 q = s->exponents[ch];
670 q_end = q + s->block_len;
672 if (s->version == 1) {
673 last_exp = get_bits(&s->gb, 5) + 10;
674 /* XXX: use a table */
675 v = pow(10, last_exp * (1.0 / 16.0));
684 code = get_vlc2(&s->gb, s->exp_vlc.table, VLCBITS, 2);
687 /* NOTE: this offset is the same as MPEG4 AAC ! */
688 last_exp += code - 60;
689 /* XXX: use a table */
690 v = pow(10, last_exp * (1.0 / 16.0));
698 s->max_exponent[ch] = max_scale;
702 /* return 0 if OK. return 1 if last block of frame. return -1 if
703 unrecorrable error. */
704 static int wma_decode_block(WMADecodeContext *s)
706 int n, v, a, ch, code, bsize;
707 int coef_nb_bits, total_gain, parse_exponents;
708 float window[BLOCK_MAX_SIZE * 2];
709 // XXX: FIXME!! there's a bug somewhere which makes this mandatory under altivec
711 volatile int nb_coefs[MAX_CHANNELS] __attribute__((aligned(16)));
713 int nb_coefs[MAX_CHANNELS];
718 tprintf("***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
721 /* compute current block length */
722 if (s->use_variable_block_len) {
723 n = av_log2(s->nb_block_sizes - 1) + 1;
725 if (s->reset_block_lengths) {
726 s->reset_block_lengths = 0;
727 v = get_bits(&s->gb, n);
728 if (v >= s->nb_block_sizes)
730 s->prev_block_len_bits = s->frame_len_bits - v;
731 v = get_bits(&s->gb, n);
732 if (v >= s->nb_block_sizes)
734 s->block_len_bits = s->frame_len_bits - v;
736 /* update block lengths */
737 s->prev_block_len_bits = s->block_len_bits;
738 s->block_len_bits = s->next_block_len_bits;
740 v = get_bits(&s->gb, n);
741 if (v >= s->nb_block_sizes)
743 s->next_block_len_bits = s->frame_len_bits - v;
745 /* fixed block len */
746 s->next_block_len_bits = s->frame_len_bits;
747 s->prev_block_len_bits = s->frame_len_bits;
748 s->block_len_bits = s->frame_len_bits;
751 /* now check if the block length is coherent with the frame length */
752 s->block_len = 1 << s->block_len_bits;
753 if ((s->block_pos + s->block_len) > s->frame_len)
756 if (s->nb_channels == 2) {
757 s->ms_stereo = get_bits(&s->gb, 1);
760 for(ch = 0; ch < s->nb_channels; ch++) {
761 a = get_bits(&s->gb, 1);
762 s->channel_coded[ch] = a;
765 /* if no channel coded, no need to go further */
766 /* XXX: fix potential framing problems */
770 bsize = s->frame_len_bits - s->block_len_bits;
772 /* read total gain and extract corresponding number of bits for
773 coef escape coding */
776 a = get_bits(&s->gb, 7);
784 else if (total_gain < 32)
786 else if (total_gain < 40)
788 else if (total_gain < 45)
793 /* compute number of coefficients */
794 n = s->coefs_end[bsize] - s->coefs_start;
795 for(ch = 0; ch < s->nb_channels; ch++)
799 if (s->use_noise_coding) {
801 for(ch = 0; ch < s->nb_channels; ch++) {
802 if (s->channel_coded[ch]) {
804 n = s->exponent_high_sizes[bsize];
806 a = get_bits(&s->gb, 1);
807 s->high_band_coded[ch][i] = a;
808 /* if noise coding, the coefficients are not transmitted */
810 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
814 for(ch = 0; ch < s->nb_channels; ch++) {
815 if (s->channel_coded[ch]) {
818 n = s->exponent_high_sizes[bsize];
819 val = (int)0x80000000;
821 if (s->high_band_coded[ch][i]) {
822 if (val == (int)0x80000000) {
823 val = get_bits(&s->gb, 7) - 19;
825 code = get_vlc2(&s->gb, s->hgain_vlc.table, VLCBITS, 2);
830 s->high_band_values[ch][i] = val;
837 /* exposant can be interpolated in short blocks. */
839 if (s->block_len_bits != s->frame_len_bits) {
840 parse_exponents = get_bits(&s->gb, 1);
843 if (parse_exponents) {
844 for(ch = 0; ch < s->nb_channels; ch++) {
845 if (s->channel_coded[ch]) {
846 if (s->use_exp_vlc) {
847 if (decode_exp_vlc(s, ch) < 0)
850 decode_exp_lsp(s, ch);
855 for(ch = 0; ch < s->nb_channels; ch++) {
856 if (s->channel_coded[ch]) {
857 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
863 /* parse spectral coefficients : just RLE encoding */
864 for(ch = 0; ch < s->nb_channels; ch++) {
865 if (s->channel_coded[ch]) {
867 int level, run, sign, tindex;
869 const int16_t *level_table, *run_table;
871 /* special VLC tables are used for ms stereo because
872 there is potentially less energy there */
873 tindex = (ch == 1 && s->ms_stereo);
874 coef_vlc = &s->coef_vlc[tindex];
875 run_table = s->run_table[tindex];
876 level_table = s->level_table[tindex];
878 ptr = &s->coefs1[ch][0];
879 eptr = ptr + nb_coefs[ch];
880 memset(ptr, 0, s->block_len * sizeof(int16_t));
882 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, 3);
888 } else if (code == 0) {
890 level = get_bits(&s->gb, coef_nb_bits);
891 /* NOTE: this is rather suboptimal. reading
892 block_len_bits would be better */
893 run = get_bits(&s->gb, s->frame_len_bits);
896 run = run_table[code];
897 level = level_table[code];
899 sign = get_bits(&s->gb, 1);
906 /* NOTE: EOB can be omitted */
911 if (s->version == 1 && s->nb_channels >= 2) {
912 align_get_bits(&s->gb);
918 int n4 = s->block_len / 2;
919 mdct_norm = 1.0 / (float)n4;
920 if (s->version == 1) {
921 mdct_norm *= sqrt(n4);
925 /* finally compute the MDCT coefficients */
926 for(ch = 0; ch < s->nb_channels; ch++) {
927 if (s->channel_coded[ch]) {
929 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
930 int i, j, n, n1, last_high_band;
931 float exp_power[HIGH_BAND_MAX_SIZE];
933 coefs1 = s->coefs1[ch];
934 exponents = s->exponents[ch];
935 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
937 coefs = s->coefs[ch];
938 if (s->use_noise_coding) {
940 /* very low freqs : noise */
941 for(i = 0;i < s->coefs_start; i++) {
942 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
943 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
946 n1 = s->exponent_high_sizes[bsize];
948 /* compute power of high bands */
949 exp_ptr = exponents +
950 s->high_band_start[bsize] -
952 last_high_band = 0; /* avoid warning */
954 n = s->exponent_high_bands[s->frame_len_bits -
955 s->block_len_bits][j];
956 if (s->high_band_coded[ch][j]) {
959 for(i = 0;i < n; i++) {
963 exp_power[j] = e2 / n;
965 tprintf("%d: power=%f (%d)\n", j, exp_power[j], n);
970 /* main freqs and high freqs */
973 n = s->high_band_start[bsize] -
976 n = s->exponent_high_bands[s->frame_len_bits -
977 s->block_len_bits][j];
979 if (j >= 0 && s->high_band_coded[ch][j]) {
980 /* use noise with specified power */
981 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
982 /* XXX: use a table */
983 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
984 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
986 for(i = 0;i < n; i++) {
987 noise = s->noise_table[s->noise_index];
988 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
989 *coefs++ = (*exponents++) * noise * mult1;
992 /* coded values + small noise */
993 for(i = 0;i < n; i++) {
994 noise = s->noise_table[s->noise_index];
995 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
996 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
1001 /* very high freqs : noise */
1002 n = s->block_len - s->coefs_end[bsize];
1003 mult1 = mult * exponents[-1];
1004 for(i = 0; i < n; i++) {
1005 *coefs++ = s->noise_table[s->noise_index] * mult1;
1006 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
1009 /* XXX: optimize more */
1010 for(i = 0;i < s->coefs_start; i++)
1013 for(i = 0;i < n; i++) {
1014 *coefs++ = coefs1[i] * exponents[i] * mult;
1016 n = s->block_len - s->coefs_end[bsize];
1017 for(i = 0;i < n; i++)
1024 for(ch = 0; ch < s->nb_channels; ch++) {
1025 if (s->channel_coded[ch]) {
1026 dump_floats("exponents", 3, s->exponents[ch], s->block_len);
1027 dump_floats("coefs", 1, s->coefs[ch], s->block_len);
1032 if (s->ms_stereo && s->channel_coded[1]) {
1036 /* nominal case for ms stereo: we do it before mdct */
1037 /* no need to optimize this case because it should almost
1039 if (!s->channel_coded[0]) {
1040 tprintf("rare ms-stereo case happened\n");
1041 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
1042 s->channel_coded[0] = 1;
1045 for(i = 0; i < s->block_len; i++) {
1048 s->coefs[0][i] = a + b;
1049 s->coefs[1][i] = a - b;
1053 /* build the window : we ensure that when the windows overlap
1054 their squared sum is always 1 (MDCT reconstruction rule) */
1055 /* XXX: merge with output */
1057 int i, next_block_len, block_len, prev_block_len, n;
1060 block_len = s->block_len;
1061 prev_block_len = 1 << s->prev_block_len_bits;
1062 next_block_len = 1 << s->next_block_len_bits;
1065 wptr = window + block_len;
1066 if (block_len <= next_block_len) {
1067 for(i=0;i<block_len;i++)
1068 *wptr++ = s->windows[bsize][i];
1071 n = (block_len / 2) - (next_block_len / 2);
1074 for(i=0;i<next_block_len;i++)
1075 *wptr++ = s->windows[s->frame_len_bits - s->next_block_len_bits][i];
1081 wptr = window + block_len;
1082 if (block_len <= prev_block_len) {
1083 for(i=0;i<block_len;i++)
1084 *--wptr = s->windows[bsize][i];
1087 n = (block_len / 2) - (prev_block_len / 2);
1090 for(i=0;i<prev_block_len;i++)
1091 *--wptr = s->windows[s->frame_len_bits - s->prev_block_len_bits][i];
1098 for(ch = 0; ch < s->nb_channels; ch++) {
1099 if (s->channel_coded[ch]) {
1100 FFTSample output[BLOCK_MAX_SIZE * 2] __attribute__((aligned(16)));
1102 int i, n4, index, n;
1105 n4 = s->block_len / 2;
1106 ff_imdct_calc(&s->mdct_ctx[bsize],
1107 output, s->coefs[ch], s->mdct_tmp);
1109 /* XXX: optimize all that by build the window and
1110 multipying/adding at the same time */
1111 /* multiply by the window */
1112 for(i=0;i<n * 2;i++) {
1113 output[i] *= window[i];
1116 /* add in the frame */
1117 index = (s->frame_len / 2) + s->block_pos - n4;
1118 ptr = &s->frame_out[ch][index];
1119 for(i=0;i<n * 2;i++) {
1124 /* specific fast case for ms-stereo : add to second
1125 channel if it is not coded */
1126 if (s->ms_stereo && !s->channel_coded[1]) {
1127 ptr = &s->frame_out[1][index];
1128 for(i=0;i<n * 2;i++) {
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,