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
23 * @file libavcodec/wmadec.c
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.
43 #define EXPMAX ((19+EXPVLCBITS-1)/EXPVLCBITS)
45 #define HGAINVLCBITS 9
46 #define HGAINMAX ((13+HGAINVLCBITS-1)/HGAINVLCBITS)
48 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len);
51 static void dump_shorts(WMACodecContext *s, const char *name, const short *tab, int n)
55 tprintf(s->avctx, "%s[%d]:\n", name, n);
58 tprintf(s->avctx, "%4d: ", i);
59 tprintf(s->avctx, " %5d.0", tab[i]);
61 tprintf(s->avctx, "\n");
65 static void dump_floats(WMACodecContext *s, const char *name, int prec, const float *tab, int n)
69 tprintf(s->avctx, "%s[%d]:\n", name, n);
72 tprintf(s->avctx, "%4d: ", i);
73 tprintf(s->avctx, " %8.*f", prec, tab[i]);
75 tprintf(s->avctx, "\n");
78 tprintf(s->avctx, "\n");
82 static int wma_decode_init(AVCodecContext * avctx)
84 WMACodecContext *s = avctx->priv_data;
90 /* extract flag infos */
92 extradata = avctx->extradata;
93 if (avctx->codec->id == CODEC_ID_WMAV1 && avctx->extradata_size >= 4) {
94 flags2 = AV_RL16(extradata+2);
95 } else if (avctx->codec->id == CODEC_ID_WMAV2 && avctx->extradata_size >= 6) {
96 flags2 = AV_RL16(extradata+4);
98 // for(i=0; i<avctx->extradata_size; i++)
99 // av_log(NULL, AV_LOG_ERROR, "%02X ", extradata[i]);
101 s->use_exp_vlc = flags2 & 0x0001;
102 s->use_bit_reservoir = flags2 & 0x0002;
103 s->use_variable_block_len = flags2 & 0x0004;
105 if(ff_wma_init(avctx, flags2)<0)
109 for(i = 0; i < s->nb_block_sizes; i++)
110 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1, 1.0);
112 if (s->use_noise_coding) {
113 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(ff_wma_hgain_huffbits),
114 ff_wma_hgain_huffbits, 1, 1,
115 ff_wma_hgain_huffcodes, 2, 2, 0);
118 if (s->use_exp_vlc) {
119 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(ff_wma_scale_huffbits), //FIXME move out of context
120 ff_wma_scale_huffbits, 1, 1,
121 ff_wma_scale_huffcodes, 4, 4, 0);
123 wma_lsp_to_curve_init(s, s->frame_len);
126 avctx->sample_fmt = SAMPLE_FMT_S16;
131 * compute x^-0.25 with an exponent and mantissa table. We use linear
132 * interpolation to reduce the mantissa table size at a small speed
133 * expense (linear interpolation approximately doubles the number of
134 * bits of precision).
136 static inline float pow_m1_4(WMACodecContext *s, float x)
147 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
148 /* build interpolation scale: 1 <= t < 2. */
149 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
150 a = s->lsp_pow_m_table1[m];
151 b = s->lsp_pow_m_table2[m];
152 return s->lsp_pow_e_table[e] * (a + b * t.f);
155 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len)
160 wdel = M_PI / frame_len;
161 for(i=0;i<frame_len;i++)
162 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
164 /* tables for x^-0.25 computation */
167 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
170 /* NOTE: these two tables are needed to avoid two operations in
173 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
174 m = (1 << LSP_POW_BITS) + i;
175 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
177 s->lsp_pow_m_table1[i] = 2 * a - b;
178 s->lsp_pow_m_table2[i] = b - a;
187 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
193 * NOTE: We use the same code as Vorbis here
194 * @todo optimize it further with SSE/3Dnow
196 static void wma_lsp_to_curve(WMACodecContext *s,
197 float *out, float *val_max_ptr,
201 float p, q, w, v, val_max;
207 w = s->lsp_cos_table[i];
208 for(j=1;j<NB_LSP_COEFS;j+=2){
220 *val_max_ptr = val_max;
224 * decode exponents coded with LSP coefficients (same idea as Vorbis)
226 static void decode_exp_lsp(WMACodecContext *s, int ch)
228 float lsp_coefs[NB_LSP_COEFS];
231 for(i = 0; i < NB_LSP_COEFS; i++) {
232 if (i == 0 || i >= 8)
233 val = get_bits(&s->gb, 3);
235 val = get_bits(&s->gb, 4);
236 lsp_coefs[i] = ff_wma_lsp_codebook[i][val];
239 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
240 s->block_len, lsp_coefs);
244 * decode exponents coded with VLC codes
246 static int decode_exp_vlc(WMACodecContext *s, int ch)
248 int last_exp, n, code;
250 float v, *q, max_scale, *q_end;
252 ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
253 q = s->exponents[ch];
254 q_end = q + s->block_len;
256 if (s->version == 1) {
257 last_exp = get_bits(&s->gb, 5) + 10;
258 /* XXX: use a table */
259 v = pow(10, last_exp * (1.0 / 16.0));
269 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
272 /* NOTE: this offset is the same as MPEG4 AAC ! */
273 last_exp += code - 60;
274 /* XXX: use a table */
275 v = pow(10, last_exp * (1.0 / 16.0));
283 s->max_exponent[ch] = max_scale;
289 * Apply MDCT window and add into output.
291 * We ensure that when the windows overlap their squared sum
292 * is always 1 (MDCT reconstruction rule).
294 static void wma_window(WMACodecContext *s, float *out)
296 float *in = s->output;
297 int block_len, bsize, n;
300 if (s->block_len_bits <= s->prev_block_len_bits) {
301 block_len = s->block_len;
302 bsize = s->frame_len_bits - s->block_len_bits;
304 s->dsp.vector_fmul_add(out, in, s->windows[bsize],
308 block_len = 1 << s->prev_block_len_bits;
309 n = (s->block_len - block_len) / 2;
310 bsize = s->frame_len_bits - s->prev_block_len_bits;
312 s->dsp.vector_fmul_add(out+n, in+n, s->windows[bsize],
315 memcpy(out+n+block_len, in+n+block_len, n*sizeof(float));
322 if (s->block_len_bits <= s->next_block_len_bits) {
323 block_len = s->block_len;
324 bsize = s->frame_len_bits - s->block_len_bits;
326 s->dsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
329 block_len = 1 << s->next_block_len_bits;
330 n = (s->block_len - block_len) / 2;
331 bsize = s->frame_len_bits - s->next_block_len_bits;
333 memcpy(out, in, n*sizeof(float));
335 s->dsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
337 memset(out+n+block_len, 0, n*sizeof(float));
343 * @return 0 if OK. 1 if last block of frame. return -1 if
344 * unrecorrable error.
346 static int wma_decode_block(WMACodecContext *s)
348 int n, v, a, ch, bsize;
349 int coef_nb_bits, total_gain;
350 int nb_coefs[MAX_CHANNELS];
354 tprintf(s->avctx, "***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
357 /* compute current block length */
358 if (s->use_variable_block_len) {
359 n = av_log2(s->nb_block_sizes - 1) + 1;
361 if (s->reset_block_lengths) {
362 s->reset_block_lengths = 0;
363 v = get_bits(&s->gb, n);
364 if (v >= s->nb_block_sizes)
366 s->prev_block_len_bits = s->frame_len_bits - v;
367 v = get_bits(&s->gb, n);
368 if (v >= s->nb_block_sizes)
370 s->block_len_bits = s->frame_len_bits - v;
372 /* update block lengths */
373 s->prev_block_len_bits = s->block_len_bits;
374 s->block_len_bits = s->next_block_len_bits;
376 v = get_bits(&s->gb, n);
377 if (v >= s->nb_block_sizes)
379 s->next_block_len_bits = s->frame_len_bits - v;
381 /* fixed block len */
382 s->next_block_len_bits = s->frame_len_bits;
383 s->prev_block_len_bits = s->frame_len_bits;
384 s->block_len_bits = s->frame_len_bits;
387 /* now check if the block length is coherent with the frame length */
388 s->block_len = 1 << s->block_len_bits;
389 if ((s->block_pos + s->block_len) > s->frame_len)
392 if (s->nb_channels == 2) {
393 s->ms_stereo = get_bits1(&s->gb);
396 for(ch = 0; ch < s->nb_channels; ch++) {
397 a = get_bits1(&s->gb);
398 s->channel_coded[ch] = a;
402 bsize = s->frame_len_bits - s->block_len_bits;
404 /* if no channel coded, no need to go further */
405 /* XXX: fix potential framing problems */
409 /* read total gain and extract corresponding number of bits for
410 coef escape coding */
413 a = get_bits(&s->gb, 7);
419 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain);
421 /* compute number of coefficients */
422 n = s->coefs_end[bsize] - s->coefs_start;
423 for(ch = 0; ch < s->nb_channels; ch++)
427 if (s->use_noise_coding) {
429 for(ch = 0; ch < s->nb_channels; ch++) {
430 if (s->channel_coded[ch]) {
432 n = s->exponent_high_sizes[bsize];
434 a = get_bits1(&s->gb);
435 s->high_band_coded[ch][i] = a;
436 /* if noise coding, the coefficients are not transmitted */
438 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
442 for(ch = 0; ch < s->nb_channels; ch++) {
443 if (s->channel_coded[ch]) {
446 n = s->exponent_high_sizes[bsize];
447 val = (int)0x80000000;
449 if (s->high_band_coded[ch][i]) {
450 if (val == (int)0x80000000) {
451 val = get_bits(&s->gb, 7) - 19;
453 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX);
458 s->high_band_values[ch][i] = val;
465 /* exponents can be reused in short blocks. */
466 if ((s->block_len_bits == s->frame_len_bits) ||
468 for(ch = 0; ch < s->nb_channels; ch++) {
469 if (s->channel_coded[ch]) {
470 if (s->use_exp_vlc) {
471 if (decode_exp_vlc(s, ch) < 0)
474 decode_exp_lsp(s, ch);
476 s->exponents_bsize[ch] = bsize;
481 /* parse spectral coefficients : just RLE encoding */
482 for(ch = 0; ch < s->nb_channels; ch++) {
483 if (s->channel_coded[ch]) {
485 WMACoef* ptr = &s->coefs1[ch][0];
487 /* special VLC tables are used for ms stereo because
488 there is potentially less energy there */
489 tindex = (ch == 1 && s->ms_stereo);
490 memset(ptr, 0, s->block_len * sizeof(WMACoef));
491 ff_wma_run_level_decode(s->avctx, &s->gb, &s->coef_vlc[tindex],
492 s->level_table[tindex], s->run_table[tindex],
493 0, ptr, 0, nb_coefs[ch],
494 s->block_len, s->frame_len_bits, coef_nb_bits);
496 if (s->version == 1 && s->nb_channels >= 2) {
497 align_get_bits(&s->gb);
503 int n4 = s->block_len / 2;
504 mdct_norm = 1.0 / (float)n4;
505 if (s->version == 1) {
506 mdct_norm *= sqrt(n4);
510 /* finally compute the MDCT coefficients */
511 for(ch = 0; ch < s->nb_channels; ch++) {
512 if (s->channel_coded[ch]) {
514 float *coefs, *exponents, mult, mult1, noise;
515 int i, j, n, n1, last_high_band, esize;
516 float exp_power[HIGH_BAND_MAX_SIZE];
518 coefs1 = s->coefs1[ch];
519 exponents = s->exponents[ch];
520 esize = s->exponents_bsize[ch];
521 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
523 coefs = s->coefs[ch];
524 if (s->use_noise_coding) {
526 /* very low freqs : noise */
527 for(i = 0;i < s->coefs_start; i++) {
528 *coefs++ = s->noise_table[s->noise_index] *
529 exponents[i<<bsize>>esize] * mult1;
530 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
533 n1 = s->exponent_high_sizes[bsize];
535 /* compute power of high bands */
536 exponents = s->exponents[ch] +
537 (s->high_band_start[bsize]<<bsize);
538 last_high_band = 0; /* avoid warning */
540 n = s->exponent_high_bands[s->frame_len_bits -
541 s->block_len_bits][j];
542 if (s->high_band_coded[ch][j]) {
545 for(i = 0;i < n; i++) {
546 v = exponents[i<<bsize>>esize];
549 exp_power[j] = e2 / n;
551 tprintf(s->avctx, "%d: power=%f (%d)\n", j, exp_power[j], n);
553 exponents += n<<bsize;
556 /* main freqs and high freqs */
557 exponents = s->exponents[ch] + (s->coefs_start<<bsize);
560 n = s->high_band_start[bsize] -
563 n = s->exponent_high_bands[s->frame_len_bits -
564 s->block_len_bits][j];
566 if (j >= 0 && s->high_band_coded[ch][j]) {
567 /* use noise with specified power */
568 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
569 /* XXX: use a table */
570 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
571 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
573 for(i = 0;i < n; i++) {
574 noise = s->noise_table[s->noise_index];
575 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
577 exponents[i<<bsize>>esize] * mult1;
579 exponents += n<<bsize;
581 /* coded values + small noise */
582 for(i = 0;i < n; i++) {
583 noise = s->noise_table[s->noise_index];
584 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
585 *coefs++ = ((*coefs1++) + noise) *
586 exponents[i<<bsize>>esize] * mult;
588 exponents += n<<bsize;
592 /* very high freqs : noise */
593 n = s->block_len - s->coefs_end[bsize];
594 mult1 = mult * exponents[((-1<<bsize))>>esize];
595 for(i = 0; i < n; i++) {
596 *coefs++ = s->noise_table[s->noise_index] * mult1;
597 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
600 /* XXX: optimize more */
601 for(i = 0;i < s->coefs_start; i++)
604 for(i = 0;i < n; i++) {
605 *coefs++ = coefs1[i] * exponents[i<<bsize>>esize] * mult;
607 n = s->block_len - s->coefs_end[bsize];
608 for(i = 0;i < n; i++)
615 for(ch = 0; ch < s->nb_channels; ch++) {
616 if (s->channel_coded[ch]) {
617 dump_floats(s, "exponents", 3, s->exponents[ch], s->block_len);
618 dump_floats(s, "coefs", 1, s->coefs[ch], s->block_len);
623 if (s->ms_stereo && s->channel_coded[1]) {
624 /* nominal case for ms stereo: we do it before mdct */
625 /* no need to optimize this case because it should almost
627 if (!s->channel_coded[0]) {
628 tprintf(s->avctx, "rare ms-stereo case happened\n");
629 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
630 s->channel_coded[0] = 1;
633 s->dsp.butterflies_float(s->coefs[0], s->coefs[1], s->block_len);
637 for(ch = 0; ch < s->nb_channels; ch++) {
640 n4 = s->block_len / 2;
641 if(s->channel_coded[ch]){
642 ff_imdct_calc(&s->mdct_ctx[bsize], s->output, s->coefs[ch]);
643 }else if(!(s->ms_stereo && ch==1))
644 memset(s->output, 0, sizeof(s->output));
646 /* multiply by the window and add in the frame */
647 index = (s->frame_len / 2) + s->block_pos - n4;
648 wma_window(s, &s->frame_out[ch][index]);
651 /* update block number */
653 s->block_pos += s->block_len;
654 if (s->block_pos >= s->frame_len)
660 /* decode a frame of frame_len samples */
661 static int wma_decode_frame(WMACodecContext *s, int16_t *samples)
663 int ret, i, n, ch, incr;
668 tprintf(s->avctx, "***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
671 /* read each block */
675 ret = wma_decode_block(s);
682 /* convert frame to integer */
684 incr = s->nb_channels;
685 for(ch = 0; ch < s->nb_channels; ch++) {
687 iptr = s->frame_out[ch];
690 *ptr = av_clip_int16(lrintf(*iptr++));
693 /* prepare for next block */
694 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
695 s->frame_len * sizeof(float));
699 dump_shorts(s, "samples", samples, n * s->nb_channels);
704 static int wma_decode_superframe(AVCodecContext *avctx,
705 void *data, int *data_size,
708 const uint8_t *buf = avpkt->data;
709 int buf_size = avpkt->size;
710 WMACodecContext *s = avctx->priv_data;
711 int nb_frames, bit_offset, i, pos, len;
715 tprintf(avctx, "***decode_superframe:\n");
718 s->last_superframe_len = 0;
721 if (buf_size < s->block_align)
723 buf_size = s->block_align;
727 init_get_bits(&s->gb, buf, buf_size*8);
729 if (s->use_bit_reservoir) {
730 /* read super frame header */
731 skip_bits(&s->gb, 4); /* super frame index */
732 nb_frames = get_bits(&s->gb, 4) - 1;
734 if((nb_frames+1) * s->nb_channels * s->frame_len * sizeof(int16_t) > *data_size){
735 av_log(s->avctx, AV_LOG_ERROR, "Insufficient output space\n");
739 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
741 if (s->last_superframe_len > 0) {
742 // printf("skip=%d\n", s->last_bitoffset);
743 /* add bit_offset bits to last frame */
744 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
745 MAX_CODED_SUPERFRAME_SIZE)
747 q = s->last_superframe + s->last_superframe_len;
750 *q++ = (get_bits)(&s->gb, 8);
754 *q++ = (get_bits)(&s->gb, len) << (8 - len);
757 /* XXX: bit_offset bits into last frame */
758 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
759 /* skip unused bits */
760 if (s->last_bitoffset > 0)
761 skip_bits(&s->gb, s->last_bitoffset);
762 /* this frame is stored in the last superframe and in the
764 if (wma_decode_frame(s, samples) < 0)
766 samples += s->nb_channels * s->frame_len;
769 /* read each frame starting from bit_offset */
770 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
771 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
774 skip_bits(&s->gb, len);
776 s->reset_block_lengths = 1;
777 for(i=0;i<nb_frames;i++) {
778 if (wma_decode_frame(s, samples) < 0)
780 samples += s->nb_channels * s->frame_len;
783 /* we copy the end of the frame in the last frame buffer */
784 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
785 s->last_bitoffset = pos & 7;
787 len = buf_size - pos;
788 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
791 s->last_superframe_len = len;
792 memcpy(s->last_superframe, buf + pos, len);
794 if(s->nb_channels * s->frame_len * sizeof(int16_t) > *data_size){
795 av_log(s->avctx, AV_LOG_ERROR, "Insufficient output space\n");
798 /* single frame decode */
799 if (wma_decode_frame(s, samples) < 0)
801 samples += s->nb_channels * s->frame_len;
804 //av_log(NULL, AV_LOG_ERROR, "%d %d %d %d outbytes:%d eaten:%d\n", s->frame_len_bits, s->block_len_bits, s->frame_len, s->block_len, (int8_t *)samples - (int8_t *)data, s->block_align);
806 *data_size = (int8_t *)samples - (int8_t *)data;
807 return s->block_align;
809 /* when error, we reset the bit reservoir */
810 s->last_superframe_len = 0;
814 AVCodec wmav1_decoder =
819 sizeof(WMACodecContext),
823 wma_decode_superframe,
824 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio 1"),
827 AVCodec wmav2_decoder =
832 sizeof(WMACodecContext),
836 wma_decode_superframe,
837 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio 2"),