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;
85 int i, flags1, flags2;
90 /* extract flag infos */
93 extradata = avctx->extradata;
94 if (avctx->codec->id == CODEC_ID_WMAV1 && avctx->extradata_size >= 4) {
95 flags1 = AV_RL16(extradata);
96 flags2 = AV_RL16(extradata+2);
97 } else if (avctx->codec->id == CODEC_ID_WMAV2 && avctx->extradata_size >= 6) {
98 flags1 = AV_RL32(extradata);
99 flags2 = AV_RL16(extradata+4);
101 // for(i=0; i<avctx->extradata_size; i++)
102 // av_log(NULL, AV_LOG_ERROR, "%02X ", extradata[i]);
104 s->use_exp_vlc = flags2 & 0x0001;
105 s->use_bit_reservoir = flags2 & 0x0002;
106 s->use_variable_block_len = flags2 & 0x0004;
108 if(ff_wma_init(avctx, flags2)<0)
112 for(i = 0; i < s->nb_block_sizes; i++)
113 ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1, 1.0);
115 if (s->use_noise_coding) {
116 init_vlc(&s->hgain_vlc, HGAINVLCBITS, sizeof(ff_wma_hgain_huffbits),
117 ff_wma_hgain_huffbits, 1, 1,
118 ff_wma_hgain_huffcodes, 2, 2, 0);
121 if (s->use_exp_vlc) {
122 init_vlc(&s->exp_vlc, EXPVLCBITS, sizeof(ff_wma_scale_huffbits), //FIXME move out of context
123 ff_wma_scale_huffbits, 1, 1,
124 ff_wma_scale_huffcodes, 4, 4, 0);
126 wma_lsp_to_curve_init(s, s->frame_len);
129 avctx->sample_fmt = SAMPLE_FMT_S16;
134 * compute x^-0.25 with an exponent and mantissa table. We use linear
135 * interpolation to reduce the mantissa table size at a small speed
136 * expense (linear interpolation approximately doubles the number of
137 * bits of precision).
139 static inline float pow_m1_4(WMACodecContext *s, float x)
150 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
151 /* build interpolation scale: 1 <= t < 2. */
152 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
153 a = s->lsp_pow_m_table1[m];
154 b = s->lsp_pow_m_table2[m];
155 return s->lsp_pow_e_table[e] * (a + b * t.f);
158 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len)
163 wdel = M_PI / frame_len;
164 for(i=0;i<frame_len;i++)
165 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
167 /* tables for x^-0.25 computation */
170 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
173 /* NOTE: these two tables are needed to avoid two operations in
176 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
177 m = (1 << LSP_POW_BITS) + i;
178 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
180 s->lsp_pow_m_table1[i] = 2 * a - b;
181 s->lsp_pow_m_table2[i] = b - a;
190 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
196 * NOTE: We use the same code as Vorbis here
197 * @todo optimize it further with SSE/3Dnow
199 static void wma_lsp_to_curve(WMACodecContext *s,
200 float *out, float *val_max_ptr,
204 float p, q, w, v, val_max;
210 w = s->lsp_cos_table[i];
211 for(j=1;j<NB_LSP_COEFS;j+=2){
223 *val_max_ptr = val_max;
227 * decode exponents coded with LSP coefficients (same idea as Vorbis)
229 static void decode_exp_lsp(WMACodecContext *s, int ch)
231 float lsp_coefs[NB_LSP_COEFS];
234 for(i = 0; i < NB_LSP_COEFS; i++) {
235 if (i == 0 || i >= 8)
236 val = get_bits(&s->gb, 3);
238 val = get_bits(&s->gb, 4);
239 lsp_coefs[i] = ff_wma_lsp_codebook[i][val];
242 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
243 s->block_len, lsp_coefs);
247 * decode exponents coded with VLC codes
249 static int decode_exp_vlc(WMACodecContext *s, int ch)
251 int last_exp, n, code;
252 const uint16_t *ptr, *band_ptr;
253 float v, *q, max_scale, *q_end;
255 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
257 q = s->exponents[ch];
258 q_end = q + s->block_len;
260 if (s->version == 1) {
261 last_exp = get_bits(&s->gb, 5) + 10;
262 /* XXX: use a table */
263 v = pow(10, last_exp * (1.0 / 16.0));
273 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
276 /* NOTE: this offset is the same as MPEG4 AAC ! */
277 last_exp += code - 60;
278 /* XXX: use a table */
279 v = pow(10, last_exp * (1.0 / 16.0));
287 s->max_exponent[ch] = max_scale;
293 * Apply MDCT window and add into output.
295 * We ensure that when the windows overlap their squared sum
296 * is always 1 (MDCT reconstruction rule).
298 static void wma_window(WMACodecContext *s, float *out)
300 float *in = s->output;
301 int block_len, bsize, n;
304 if (s->block_len_bits <= s->prev_block_len_bits) {
305 block_len = s->block_len;
306 bsize = s->frame_len_bits - s->block_len_bits;
308 s->dsp.vector_fmul_add_add(out, in, s->windows[bsize],
309 out, 0, block_len, 1);
312 block_len = 1 << s->prev_block_len_bits;
313 n = (s->block_len - block_len) / 2;
314 bsize = s->frame_len_bits - s->prev_block_len_bits;
316 s->dsp.vector_fmul_add_add(out+n, in+n, s->windows[bsize],
317 out+n, 0, block_len, 1);
319 memcpy(out+n+block_len, in+n+block_len, n*sizeof(float));
326 if (s->block_len_bits <= s->next_block_len_bits) {
327 block_len = s->block_len;
328 bsize = s->frame_len_bits - s->block_len_bits;
330 s->dsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
333 block_len = 1 << s->next_block_len_bits;
334 n = (s->block_len - block_len) / 2;
335 bsize = s->frame_len_bits - s->next_block_len_bits;
337 memcpy(out, in, n*sizeof(float));
339 s->dsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
341 memset(out+n+block_len, 0, n*sizeof(float));
347 * @return 0 if OK. 1 if last block of frame. return -1 if
348 * unrecorrable error.
350 static int wma_decode_block(WMACodecContext *s)
352 int n, v, a, ch, code, bsize;
353 int coef_nb_bits, total_gain;
354 int nb_coefs[MAX_CHANNELS];
358 tprintf(s->avctx, "***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
361 /* compute current block length */
362 if (s->use_variable_block_len) {
363 n = av_log2(s->nb_block_sizes - 1) + 1;
365 if (s->reset_block_lengths) {
366 s->reset_block_lengths = 0;
367 v = get_bits(&s->gb, n);
368 if (v >= s->nb_block_sizes)
370 s->prev_block_len_bits = s->frame_len_bits - v;
371 v = get_bits(&s->gb, n);
372 if (v >= s->nb_block_sizes)
374 s->block_len_bits = s->frame_len_bits - v;
376 /* update block lengths */
377 s->prev_block_len_bits = s->block_len_bits;
378 s->block_len_bits = s->next_block_len_bits;
380 v = get_bits(&s->gb, n);
381 if (v >= s->nb_block_sizes)
383 s->next_block_len_bits = s->frame_len_bits - v;
385 /* fixed block len */
386 s->next_block_len_bits = s->frame_len_bits;
387 s->prev_block_len_bits = s->frame_len_bits;
388 s->block_len_bits = s->frame_len_bits;
391 /* now check if the block length is coherent with the frame length */
392 s->block_len = 1 << s->block_len_bits;
393 if ((s->block_pos + s->block_len) > s->frame_len)
396 if (s->nb_channels == 2) {
397 s->ms_stereo = get_bits1(&s->gb);
400 for(ch = 0; ch < s->nb_channels; ch++) {
401 a = get_bits1(&s->gb);
402 s->channel_coded[ch] = a;
406 bsize = s->frame_len_bits - s->block_len_bits;
408 /* if no channel coded, no need to go further */
409 /* XXX: fix potential framing problems */
413 /* read total gain and extract corresponding number of bits for
414 coef escape coding */
417 a = get_bits(&s->gb, 7);
423 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain);
425 /* compute number of coefficients */
426 n = s->coefs_end[bsize] - s->coefs_start;
427 for(ch = 0; ch < s->nb_channels; ch++)
431 if (s->use_noise_coding) {
433 for(ch = 0; ch < s->nb_channels; ch++) {
434 if (s->channel_coded[ch]) {
436 n = s->exponent_high_sizes[bsize];
438 a = get_bits1(&s->gb);
439 s->high_band_coded[ch][i] = a;
440 /* if noise coding, the coefficients are not transmitted */
442 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
446 for(ch = 0; ch < s->nb_channels; ch++) {
447 if (s->channel_coded[ch]) {
450 n = s->exponent_high_sizes[bsize];
451 val = (int)0x80000000;
453 if (s->high_band_coded[ch][i]) {
454 if (val == (int)0x80000000) {
455 val = get_bits(&s->gb, 7) - 19;
457 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX);
462 s->high_band_values[ch][i] = val;
469 /* exponents can be reused in short blocks. */
470 if ((s->block_len_bits == s->frame_len_bits) ||
472 for(ch = 0; ch < s->nb_channels; ch++) {
473 if (s->channel_coded[ch]) {
474 if (s->use_exp_vlc) {
475 if (decode_exp_vlc(s, ch) < 0)
478 decode_exp_lsp(s, ch);
480 s->exponents_bsize[ch] = bsize;
485 /* parse spectral coefficients : just RLE encoding */
486 for(ch = 0; ch < s->nb_channels; ch++) {
487 if (s->channel_coded[ch]) {
489 int level, run, sign, tindex;
491 const uint16_t *level_table, *run_table;
493 /* special VLC tables are used for ms stereo because
494 there is potentially less energy there */
495 tindex = (ch == 1 && s->ms_stereo);
496 coef_vlc = &s->coef_vlc[tindex];
497 run_table = s->run_table[tindex];
498 level_table = s->level_table[tindex];
500 ptr = &s->coefs1[ch][0];
501 eptr = ptr + nb_coefs[ch];
502 memset(ptr, 0, s->block_len * sizeof(int16_t));
504 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX);
510 } else if (code == 0) {
512 level = get_bits(&s->gb, coef_nb_bits);
513 /* NOTE: this is rather suboptimal. reading
514 block_len_bits would be better */
515 run = get_bits(&s->gb, s->frame_len_bits);
518 run = run_table[code];
519 level = level_table[code];
521 sign = get_bits1(&s->gb);
527 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n");
531 /* NOTE: EOB can be omitted */
536 if (s->version == 1 && s->nb_channels >= 2) {
537 align_get_bits(&s->gb);
543 int n4 = s->block_len / 2;
544 mdct_norm = 1.0 / (float)n4;
545 if (s->version == 1) {
546 mdct_norm *= sqrt(n4);
550 /* finally compute the MDCT coefficients */
551 for(ch = 0; ch < s->nb_channels; ch++) {
552 if (s->channel_coded[ch]) {
554 float *coefs, *exponents, mult, mult1, noise;
555 int i, j, n, n1, last_high_band, esize;
556 float exp_power[HIGH_BAND_MAX_SIZE];
558 coefs1 = s->coefs1[ch];
559 exponents = s->exponents[ch];
560 esize = s->exponents_bsize[ch];
561 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
563 coefs = s->coefs[ch];
564 if (s->use_noise_coding) {
566 /* very low freqs : noise */
567 for(i = 0;i < s->coefs_start; i++) {
568 *coefs++ = s->noise_table[s->noise_index] *
569 exponents[i<<bsize>>esize] * mult1;
570 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
573 n1 = s->exponent_high_sizes[bsize];
575 /* compute power of high bands */
576 exponents = s->exponents[ch] +
577 (s->high_band_start[bsize]<<bsize);
578 last_high_band = 0; /* avoid warning */
580 n = s->exponent_high_bands[s->frame_len_bits -
581 s->block_len_bits][j];
582 if (s->high_band_coded[ch][j]) {
585 for(i = 0;i < n; i++) {
586 v = exponents[i<<bsize>>esize];
589 exp_power[j] = e2 / n;
591 tprintf(s->avctx, "%d: power=%f (%d)\n", j, exp_power[j], n);
593 exponents += n<<bsize;
596 /* main freqs and high freqs */
597 exponents = s->exponents[ch] + (s->coefs_start<<bsize);
600 n = s->high_band_start[bsize] -
603 n = s->exponent_high_bands[s->frame_len_bits -
604 s->block_len_bits][j];
606 if (j >= 0 && s->high_band_coded[ch][j]) {
607 /* use noise with specified power */
608 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
609 /* XXX: use a table */
610 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
611 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
613 for(i = 0;i < n; i++) {
614 noise = s->noise_table[s->noise_index];
615 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
617 exponents[i<<bsize>>esize] * mult1;
619 exponents += n<<bsize;
621 /* coded values + small noise */
622 for(i = 0;i < n; i++) {
623 noise = s->noise_table[s->noise_index];
624 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
625 *coefs++ = ((*coefs1++) + noise) *
626 exponents[i<<bsize>>esize] * mult;
628 exponents += n<<bsize;
632 /* very high freqs : noise */
633 n = s->block_len - s->coefs_end[bsize];
634 mult1 = mult * exponents[((-1<<bsize))>>esize];
635 for(i = 0; i < n; i++) {
636 *coefs++ = s->noise_table[s->noise_index] * mult1;
637 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
640 /* XXX: optimize more */
641 for(i = 0;i < s->coefs_start; i++)
644 for(i = 0;i < n; i++) {
645 *coefs++ = coefs1[i] * exponents[i<<bsize>>esize] * mult;
647 n = s->block_len - s->coefs_end[bsize];
648 for(i = 0;i < n; i++)
655 for(ch = 0; ch < s->nb_channels; ch++) {
656 if (s->channel_coded[ch]) {
657 dump_floats(s, "exponents", 3, s->exponents[ch], s->block_len);
658 dump_floats(s, "coefs", 1, s->coefs[ch], s->block_len);
663 if (s->ms_stereo && s->channel_coded[1]) {
667 /* nominal case for ms stereo: we do it before mdct */
668 /* no need to optimize this case because it should almost
670 if (!s->channel_coded[0]) {
671 tprintf(s->avctx, "rare ms-stereo case happened\n");
672 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
673 s->channel_coded[0] = 1;
676 for(i = 0; i < s->block_len; i++) {
679 s->coefs[0][i] = a + b;
680 s->coefs[1][i] = a - b;
685 for(ch = 0; ch < s->nb_channels; ch++) {
688 n4 = s->block_len / 2;
689 if(s->channel_coded[ch]){
690 ff_imdct_calc(&s->mdct_ctx[bsize], s->output, s->coefs[ch]);
691 }else if(!(s->ms_stereo && ch==1))
692 memset(s->output, 0, sizeof(s->output));
694 /* multiply by the window and add in the frame */
695 index = (s->frame_len / 2) + s->block_pos - n4;
696 wma_window(s, &s->frame_out[ch][index]);
699 /* update block number */
701 s->block_pos += s->block_len;
702 if (s->block_pos >= s->frame_len)
708 /* decode a frame of frame_len samples */
709 static int wma_decode_frame(WMACodecContext *s, int16_t *samples)
711 int ret, i, n, ch, incr;
716 tprintf(s->avctx, "***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
719 /* read each block */
723 ret = wma_decode_block(s);
730 /* convert frame to integer */
732 incr = s->nb_channels;
733 for(ch = 0; ch < s->nb_channels; ch++) {
735 iptr = s->frame_out[ch];
738 *ptr = av_clip_int16(lrintf(*iptr++));
741 /* prepare for next block */
742 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
743 s->frame_len * sizeof(float));
747 dump_shorts(s, "samples", samples, n * s->nb_channels);
752 static int wma_decode_superframe(AVCodecContext *avctx,
753 void *data, int *data_size,
756 const uint8_t *buf = avpkt->data;
757 int buf_size = avpkt->size;
758 WMACodecContext *s = avctx->priv_data;
759 int nb_frames, bit_offset, i, pos, len;
763 tprintf(avctx, "***decode_superframe:\n");
766 s->last_superframe_len = 0;
769 if (buf_size < s->block_align)
771 buf_size = s->block_align;
775 init_get_bits(&s->gb, buf, buf_size*8);
777 if (s->use_bit_reservoir) {
778 /* read super frame header */
779 skip_bits(&s->gb, 4); /* super frame index */
780 nb_frames = get_bits(&s->gb, 4) - 1;
782 if((nb_frames+1) * s->nb_channels * s->frame_len * sizeof(int16_t) > *data_size){
783 av_log(s->avctx, AV_LOG_ERROR, "Insufficient output space\n");
787 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
789 if (s->last_superframe_len > 0) {
790 // printf("skip=%d\n", s->last_bitoffset);
791 /* add bit_offset bits to last frame */
792 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
793 MAX_CODED_SUPERFRAME_SIZE)
795 q = s->last_superframe + s->last_superframe_len;
798 *q++ = (get_bits)(&s->gb, 8);
802 *q++ = (get_bits)(&s->gb, len) << (8 - len);
805 /* XXX: bit_offset bits into last frame */
806 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
807 /* skip unused bits */
808 if (s->last_bitoffset > 0)
809 skip_bits(&s->gb, s->last_bitoffset);
810 /* this frame is stored in the last superframe and in the
812 if (wma_decode_frame(s, samples) < 0)
814 samples += s->nb_channels * s->frame_len;
817 /* read each frame starting from bit_offset */
818 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
819 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
822 skip_bits(&s->gb, len);
824 s->reset_block_lengths = 1;
825 for(i=0;i<nb_frames;i++) {
826 if (wma_decode_frame(s, samples) < 0)
828 samples += s->nb_channels * s->frame_len;
831 /* we copy the end of the frame in the last frame buffer */
832 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
833 s->last_bitoffset = pos & 7;
835 len = buf_size - pos;
836 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
839 s->last_superframe_len = len;
840 memcpy(s->last_superframe, buf + pos, len);
842 if(s->nb_channels * s->frame_len * sizeof(int16_t) > *data_size){
843 av_log(s->avctx, AV_LOG_ERROR, "Insufficient output space\n");
846 /* single frame decode */
847 if (wma_decode_frame(s, samples) < 0)
849 samples += s->nb_channels * s->frame_len;
852 //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);
854 *data_size = (int8_t *)samples - (int8_t *)data;
855 return s->block_align;
857 /* when error, we reset the bit reservoir */
858 s->last_superframe_len = 0;
862 AVCodec wmav1_decoder =
867 sizeof(WMACodecContext),
871 wma_decode_superframe,
872 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio 1"),
875 AVCodec wmav2_decoder =
880 sizeof(WMACodecContext),
884 wma_decode_superframe,
885 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio 2"),