2 * WMA compatible decoder
3 * Copyright (c) 2002 The FFmpeg Project.
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * WMA compatible decoder.
25 * This decoder handles Microsoft Windows Media Audio data, versions 1 & 2.
26 * WMA v1 is identified by audio format 0x160 in Microsoft media files
27 * (ASF/AVI/WAV). WMA v2 is identified by audio format 0x161.
29 * To use this decoder, a calling application must supply the extra data
30 * bytes provided with the WMA data. These are the extra, codec-specific
31 * bytes at the end of a WAVEFORMATEX data structure. Transmit these bytes
32 * to the decoder using the extradata[_size] fields in AVCodecContext. There
33 * should be 4 extra bytes for v1 data and 6 extra bytes for v2 data.
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);
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);
133 * compute x^-0.25 with an exponent and mantissa table. We use linear
134 * interpolation to reduce the mantissa table size at a small speed
135 * expense (linear interpolation approximately doubles the number of
136 * bits of precision).
138 static inline float pow_m1_4(WMACodecContext *s, float x)
149 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
150 /* build interpolation scale: 1 <= t < 2. */
151 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
152 a = s->lsp_pow_m_table1[m];
153 b = s->lsp_pow_m_table2[m];
154 return s->lsp_pow_e_table[e] * (a + b * t.f);
157 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len)
162 wdel = M_PI / frame_len;
163 for(i=0;i<frame_len;i++)
164 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
166 /* tables for x^-0.25 computation */
169 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
172 /* NOTE: these two tables are needed to avoid two operations in
175 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
176 m = (1 << LSP_POW_BITS) + i;
177 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
179 s->lsp_pow_m_table1[i] = 2 * a - b;
180 s->lsp_pow_m_table2[i] = b - a;
189 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
195 * NOTE: We use the same code as Vorbis here
196 * @todo optimize it further with SSE/3Dnow
198 static void wma_lsp_to_curve(WMACodecContext *s,
199 float *out, float *val_max_ptr,
203 float p, q, w, v, val_max;
209 w = s->lsp_cos_table[i];
210 for(j=1;j<NB_LSP_COEFS;j+=2){
222 *val_max_ptr = val_max;
226 * decode exponents coded with LSP coefficients (same idea as Vorbis)
228 static void decode_exp_lsp(WMACodecContext *s, int ch)
230 float lsp_coefs[NB_LSP_COEFS];
233 for(i = 0; i < NB_LSP_COEFS; i++) {
234 if (i == 0 || i >= 8)
235 val = get_bits(&s->gb, 3);
237 val = get_bits(&s->gb, 4);
238 lsp_coefs[i] = ff_wma_lsp_codebook[i][val];
241 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
242 s->block_len, lsp_coefs);
246 * decode exponents coded with VLC codes
248 static int decode_exp_vlc(WMACodecContext *s, int ch)
250 int last_exp, n, code;
251 const uint16_t *ptr, *band_ptr;
252 float v, *q, max_scale, *q_end;
254 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
256 q = s->exponents[ch];
257 q_end = q + s->block_len;
259 if (s->version == 1) {
260 last_exp = get_bits(&s->gb, 5) + 10;
261 /* XXX: use a table */
262 v = pow(10, last_exp * (1.0 / 16.0));
272 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
275 /* NOTE: this offset is the same as MPEG4 AAC ! */
276 last_exp += code - 60;
277 /* XXX: use a table */
278 v = pow(10, last_exp * (1.0 / 16.0));
286 s->max_exponent[ch] = max_scale;
292 * Apply MDCT window and add into output.
294 * We ensure that when the windows overlap their squared sum
295 * is always 1 (MDCT reconstruction rule).
297 static void wma_window(WMACodecContext *s, float *out)
299 float *in = s->output;
300 int block_len, bsize, n;
303 if (s->block_len_bits <= s->prev_block_len_bits) {
304 block_len = s->block_len;
305 bsize = s->frame_len_bits - s->block_len_bits;
307 s->dsp.vector_fmul_add_add(out, in, s->windows[bsize],
308 out, 0, block_len, 1);
311 block_len = 1 << s->prev_block_len_bits;
312 n = (s->block_len - block_len) / 2;
313 bsize = s->frame_len_bits - s->prev_block_len_bits;
315 s->dsp.vector_fmul_add_add(out+n, in+n, s->windows[bsize],
316 out+n, 0, block_len, 1);
318 memcpy(out+n+block_len, in+n+block_len, n*sizeof(float));
325 if (s->block_len_bits <= s->next_block_len_bits) {
326 block_len = s->block_len;
327 bsize = s->frame_len_bits - s->block_len_bits;
329 s->dsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
332 block_len = 1 << s->next_block_len_bits;
333 n = (s->block_len - block_len) / 2;
334 bsize = s->frame_len_bits - s->next_block_len_bits;
336 memcpy(out, in, n*sizeof(float));
338 s->dsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
340 memset(out+n+block_len, 0, n*sizeof(float));
346 * @return 0 if OK. 1 if last block of frame. return -1 if
347 * unrecorrable error.
349 static int wma_decode_block(WMACodecContext *s)
351 int n, v, a, ch, code, bsize;
352 int coef_nb_bits, total_gain;
353 int nb_coefs[MAX_CHANNELS];
357 tprintf(s->avctx, "***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
360 /* compute current block length */
361 if (s->use_variable_block_len) {
362 n = av_log2(s->nb_block_sizes - 1) + 1;
364 if (s->reset_block_lengths) {
365 s->reset_block_lengths = 0;
366 v = get_bits(&s->gb, n);
367 if (v >= s->nb_block_sizes)
369 s->prev_block_len_bits = s->frame_len_bits - v;
370 v = get_bits(&s->gb, n);
371 if (v >= s->nb_block_sizes)
373 s->block_len_bits = s->frame_len_bits - v;
375 /* update block lengths */
376 s->prev_block_len_bits = s->block_len_bits;
377 s->block_len_bits = s->next_block_len_bits;
379 v = get_bits(&s->gb, n);
380 if (v >= s->nb_block_sizes)
382 s->next_block_len_bits = s->frame_len_bits - v;
384 /* fixed block len */
385 s->next_block_len_bits = s->frame_len_bits;
386 s->prev_block_len_bits = s->frame_len_bits;
387 s->block_len_bits = s->frame_len_bits;
390 /* now check if the block length is coherent with the frame length */
391 s->block_len = 1 << s->block_len_bits;
392 if ((s->block_pos + s->block_len) > s->frame_len)
395 if (s->nb_channels == 2) {
396 s->ms_stereo = get_bits1(&s->gb);
399 for(ch = 0; ch < s->nb_channels; ch++) {
400 a = get_bits1(&s->gb);
401 s->channel_coded[ch] = a;
404 /* if no channel coded, no need to go further */
405 /* XXX: fix potential framing problems */
409 bsize = s->frame_len_bits - s->block_len_bits;
411 /* read total gain and extract corresponding number of bits for
412 coef escape coding */
415 a = get_bits(&s->gb, 7);
421 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain);
423 /* compute number of coefficients */
424 n = s->coefs_end[bsize] - s->coefs_start;
425 for(ch = 0; ch < s->nb_channels; ch++)
429 if (s->use_noise_coding) {
431 for(ch = 0; ch < s->nb_channels; ch++) {
432 if (s->channel_coded[ch]) {
434 n = s->exponent_high_sizes[bsize];
436 a = get_bits1(&s->gb);
437 s->high_band_coded[ch][i] = a;
438 /* if noise coding, the coefficients are not transmitted */
440 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
444 for(ch = 0; ch < s->nb_channels; ch++) {
445 if (s->channel_coded[ch]) {
448 n = s->exponent_high_sizes[bsize];
449 val = (int)0x80000000;
451 if (s->high_band_coded[ch][i]) {
452 if (val == (int)0x80000000) {
453 val = get_bits(&s->gb, 7) - 19;
455 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX);
460 s->high_band_values[ch][i] = val;
467 /* exponents can be reused in short blocks. */
468 if ((s->block_len_bits == s->frame_len_bits) ||
470 for(ch = 0; ch < s->nb_channels; ch++) {
471 if (s->channel_coded[ch]) {
472 if (s->use_exp_vlc) {
473 if (decode_exp_vlc(s, ch) < 0)
476 decode_exp_lsp(s, ch);
478 s->exponents_bsize[ch] = bsize;
483 /* parse spectral coefficients : just RLE encoding */
484 for(ch = 0; ch < s->nb_channels; ch++) {
485 if (s->channel_coded[ch]) {
487 int level, run, sign, tindex;
489 const uint16_t *level_table, *run_table;
491 /* special VLC tables are used for ms stereo because
492 there is potentially less energy there */
493 tindex = (ch == 1 && s->ms_stereo);
494 coef_vlc = &s->coef_vlc[tindex];
495 run_table = s->run_table[tindex];
496 level_table = s->level_table[tindex];
498 ptr = &s->coefs1[ch][0];
499 eptr = ptr + nb_coefs[ch];
500 memset(ptr, 0, s->block_len * sizeof(int16_t));
502 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX);
508 } else if (code == 0) {
510 level = get_bits(&s->gb, coef_nb_bits);
511 /* NOTE: this is rather suboptimal. reading
512 block_len_bits would be better */
513 run = get_bits(&s->gb, s->frame_len_bits);
516 run = run_table[code];
517 level = level_table[code];
519 sign = get_bits1(&s->gb);
525 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n");
529 /* NOTE: EOB can be omitted */
534 if (s->version == 1 && s->nb_channels >= 2) {
535 align_get_bits(&s->gb);
541 int n4 = s->block_len / 2;
542 mdct_norm = 1.0 / (float)n4;
543 if (s->version == 1) {
544 mdct_norm *= sqrt(n4);
548 /* finally compute the MDCT coefficients */
549 for(ch = 0; ch < s->nb_channels; ch++) {
550 if (s->channel_coded[ch]) {
552 float *coefs, *exponents, mult, mult1, noise;
553 int i, j, n, n1, last_high_band, esize;
554 float exp_power[HIGH_BAND_MAX_SIZE];
556 coefs1 = s->coefs1[ch];
557 exponents = s->exponents[ch];
558 esize = s->exponents_bsize[ch];
559 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
561 coefs = s->coefs[ch];
562 if (s->use_noise_coding) {
564 /* very low freqs : noise */
565 for(i = 0;i < s->coefs_start; i++) {
566 *coefs++ = s->noise_table[s->noise_index] *
567 exponents[i<<bsize>>esize] * mult1;
568 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
571 n1 = s->exponent_high_sizes[bsize];
573 /* compute power of high bands */
574 exponents = s->exponents[ch] +
575 (s->high_band_start[bsize]<<bsize);
576 last_high_band = 0; /* avoid warning */
578 n = s->exponent_high_bands[s->frame_len_bits -
579 s->block_len_bits][j];
580 if (s->high_band_coded[ch][j]) {
583 for(i = 0;i < n; i++) {
584 v = exponents[i<<bsize>>esize];
587 exp_power[j] = e2 / n;
589 tprintf(s->avctx, "%d: power=%f (%d)\n", j, exp_power[j], n);
591 exponents += n<<bsize;
594 /* main freqs and high freqs */
595 exponents = s->exponents[ch] + (s->coefs_start<<bsize);
598 n = s->high_band_start[bsize] -
601 n = s->exponent_high_bands[s->frame_len_bits -
602 s->block_len_bits][j];
604 if (j >= 0 && s->high_band_coded[ch][j]) {
605 /* use noise with specified power */
606 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
607 /* XXX: use a table */
608 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
609 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
611 for(i = 0;i < n; i++) {
612 noise = s->noise_table[s->noise_index];
613 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
615 exponents[i<<bsize>>esize] * mult1;
617 exponents += n<<bsize;
619 /* coded values + small noise */
620 for(i = 0;i < n; i++) {
621 noise = s->noise_table[s->noise_index];
622 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
623 *coefs++ = ((*coefs1++) + noise) *
624 exponents[i<<bsize>>esize] * mult;
626 exponents += n<<bsize;
630 /* very high freqs : noise */
631 n = s->block_len - s->coefs_end[bsize];
632 mult1 = mult * exponents[((-1<<bsize))>>esize];
633 for(i = 0; i < n; i++) {
634 *coefs++ = s->noise_table[s->noise_index] * mult1;
635 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
638 /* XXX: optimize more */
639 for(i = 0;i < s->coefs_start; i++)
642 for(i = 0;i < n; i++) {
643 *coefs++ = coefs1[i] * exponents[i<<bsize>>esize] * mult;
645 n = s->block_len - s->coefs_end[bsize];
646 for(i = 0;i < n; i++)
653 for(ch = 0; ch < s->nb_channels; ch++) {
654 if (s->channel_coded[ch]) {
655 dump_floats(s, "exponents", 3, s->exponents[ch], s->block_len);
656 dump_floats(s, "coefs", 1, s->coefs[ch], s->block_len);
661 if (s->ms_stereo && s->channel_coded[1]) {
665 /* nominal case for ms stereo: we do it before mdct */
666 /* no need to optimize this case because it should almost
668 if (!s->channel_coded[0]) {
669 tprintf(s->avctx, "rare ms-stereo case happened\n");
670 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
671 s->channel_coded[0] = 1;
674 for(i = 0; i < s->block_len; i++) {
677 s->coefs[0][i] = a + b;
678 s->coefs[1][i] = a - b;
682 for(ch = 0; ch < s->nb_channels; ch++) {
683 if (s->channel_coded[ch]) {
687 n4 = s->block_len / 2;
688 s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize],
689 s->output, s->coefs[ch], s->mdct_tmp);
691 /* multiply by the window and add in the frame */
692 index = (s->frame_len / 2) + s->block_pos - n4;
693 wma_window(s, &s->frame_out[ch][index]);
695 /* specific fast case for ms-stereo : add to second
696 channel if it is not coded */
697 if (s->ms_stereo && !s->channel_coded[1]) {
698 wma_window(s, &s->frame_out[1][index]);
703 /* update block number */
705 s->block_pos += s->block_len;
706 if (s->block_pos >= s->frame_len)
712 /* decode a frame of frame_len samples */
713 static int wma_decode_frame(WMACodecContext *s, int16_t *samples)
715 int ret, i, n, ch, incr;
720 tprintf(s->avctx, "***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
723 /* read each block */
727 ret = wma_decode_block(s);
734 /* convert frame to integer */
736 incr = s->nb_channels;
737 for(ch = 0; ch < s->nb_channels; ch++) {
739 iptr = s->frame_out[ch];
742 *ptr = av_clip_int16(lrintf(*iptr++));
745 /* prepare for next block */
746 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
747 s->frame_len * sizeof(float));
751 dump_shorts(s, "samples", samples, n * s->nb_channels);
756 static int wma_decode_superframe(AVCodecContext *avctx,
757 void *data, int *data_size,
758 uint8_t *buf, int buf_size)
760 WMACodecContext *s = avctx->priv_data;
761 int nb_frames, bit_offset, i, pos, len;
765 tprintf(avctx, "***decode_superframe:\n");
768 s->last_superframe_len = 0;
774 init_get_bits(&s->gb, buf, buf_size*8);
776 if (s->use_bit_reservoir) {
777 /* read super frame header */
778 skip_bits(&s->gb, 4); /* super frame index */
779 nb_frames = get_bits(&s->gb, 4) - 1;
781 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
783 if (s->last_superframe_len > 0) {
784 // printf("skip=%d\n", s->last_bitoffset);
785 /* add bit_offset bits to last frame */
786 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
787 MAX_CODED_SUPERFRAME_SIZE)
789 q = s->last_superframe + s->last_superframe_len;
792 *q++ = (get_bits)(&s->gb, 8);
796 *q++ = (get_bits)(&s->gb, len) << (8 - len);
799 /* XXX: bit_offset bits into last frame */
800 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
801 /* skip unused bits */
802 if (s->last_bitoffset > 0)
803 skip_bits(&s->gb, s->last_bitoffset);
804 /* this frame is stored in the last superframe and in the
806 if (wma_decode_frame(s, samples) < 0)
808 samples += s->nb_channels * s->frame_len;
811 /* read each frame starting from bit_offset */
812 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
813 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
816 skip_bits(&s->gb, len);
818 s->reset_block_lengths = 1;
819 for(i=0;i<nb_frames;i++) {
820 if (wma_decode_frame(s, samples) < 0)
822 samples += s->nb_channels * s->frame_len;
825 /* we copy the end of the frame in the last frame buffer */
826 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
827 s->last_bitoffset = pos & 7;
829 len = buf_size - pos;
830 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
833 s->last_superframe_len = len;
834 memcpy(s->last_superframe, buf + pos, len);
836 /* single frame decode */
837 if (wma_decode_frame(s, samples) < 0)
839 samples += s->nb_channels * s->frame_len;
842 //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);
844 *data_size = (int8_t *)samples - (int8_t *)data;
845 return s->block_align;
847 /* when error, we reset the bit reservoir */
848 s->last_superframe_len = 0;
852 AVCodec wmav1_decoder =
857 sizeof(WMACodecContext),
861 wma_decode_superframe,
864 AVCodec wmav2_decoder =
869 sizeof(WMACodecContext),
873 wma_decode_superframe,