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(WMADecodeContext *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(WMADecodeContext *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 = extradata[0] | (extradata[1] << 8);
96 flags2 = extradata[2] | (extradata[3] << 8);
97 } else if (avctx->codec->id == CODEC_ID_WMAV2 && avctx->extradata_size >= 6) {
98 flags1 = extradata[0] | (extradata[1] << 8) |
99 (extradata[2] << 16) | (extradata[3] << 24);
100 flags2 = extradata[4] | (extradata[5] << 8);
102 // for(i=0; i<avctx->extradata_size; i++)
103 // av_log(NULL, AV_LOG_ERROR, "%02X ", extradata[i]);
105 s->use_exp_vlc = flags2 & 0x0001;
106 s->use_bit_reservoir = flags2 & 0x0002;
107 s->use_variable_block_len = flags2 & 0x0004;
109 ff_wma_init(avctx, flags2);
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 * interpolate values for a bigger or smaller block. The block must
134 * have multiple sizes
136 static void interpolate_array(float *scale, int old_size, int new_size)
141 if (new_size > old_size) {
142 jincr = new_size / old_size;
144 for(i = old_size - 1; i >=0; i--) {
151 } else if (new_size < old_size) {
153 jincr = old_size / new_size;
154 for(i = 0; i < new_size; i++) {
162 * compute x^-0.25 with an exponent and mantissa table. We use linear
163 * interpolation to reduce the mantissa table size at a small speed
164 * expense (linear interpolation approximately doubles the number of
165 * bits of precision).
167 static inline float pow_m1_4(WMACodecContext *s, float x)
178 m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
179 /* build interpolation scale: 1 <= t < 2. */
180 t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
181 a = s->lsp_pow_m_table1[m];
182 b = s->lsp_pow_m_table2[m];
183 return s->lsp_pow_e_table[e] * (a + b * t.f);
186 static void wma_lsp_to_curve_init(WMACodecContext *s, int frame_len)
191 wdel = M_PI / frame_len;
192 for(i=0;i<frame_len;i++)
193 s->lsp_cos_table[i] = 2.0f * cos(wdel * i);
195 /* tables for x^-0.25 computation */
198 s->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
201 /* NOTE: these two tables are needed to avoid two operations in
204 for(i=(1 << LSP_POW_BITS) - 1;i>=0;i--) {
205 m = (1 << LSP_POW_BITS) + i;
206 a = (float)m * (0.5 / (1 << LSP_POW_BITS));
208 s->lsp_pow_m_table1[i] = 2 * a - b;
209 s->lsp_pow_m_table2[i] = b - a;
218 printf("%f^-0.25=%f e=%f\n", v, r1, r2 - r1);
224 * NOTE: We use the same code as Vorbis here
225 * @todo optimize it further with SSE/3Dnow
227 static void wma_lsp_to_curve(WMACodecContext *s,
228 float *out, float *val_max_ptr,
232 float p, q, w, v, val_max;
238 w = s->lsp_cos_table[i];
239 for(j=1;j<NB_LSP_COEFS;j+=2){
251 *val_max_ptr = val_max;
255 * decode exponents coded with LSP coefficients (same idea as Vorbis)
257 static void decode_exp_lsp(WMACodecContext *s, int ch)
259 float lsp_coefs[NB_LSP_COEFS];
262 for(i = 0; i < NB_LSP_COEFS; i++) {
263 if (i == 0 || i >= 8)
264 val = get_bits(&s->gb, 3);
266 val = get_bits(&s->gb, 4);
267 lsp_coefs[i] = ff_wma_lsp_codebook[i][val];
270 wma_lsp_to_curve(s, s->exponents[ch], &s->max_exponent[ch],
271 s->block_len, lsp_coefs);
275 * decode exponents coded with VLC codes
277 static int decode_exp_vlc(WMACodecContext *s, int ch)
279 int last_exp, n, code;
280 const uint16_t *ptr, *band_ptr;
281 float v, *q, max_scale, *q_end;
283 band_ptr = s->exponent_bands[s->frame_len_bits - s->block_len_bits];
285 q = s->exponents[ch];
286 q_end = q + s->block_len;
288 if (s->version == 1) {
289 last_exp = get_bits(&s->gb, 5) + 10;
290 /* XXX: use a table */
291 v = pow(10, last_exp * (1.0 / 16.0));
301 code = get_vlc2(&s->gb, s->exp_vlc.table, EXPVLCBITS, EXPMAX);
304 /* NOTE: this offset is the same as MPEG4 AAC ! */
305 last_exp += code - 60;
306 /* XXX: use a table */
307 v = pow(10, last_exp * (1.0 / 16.0));
315 s->max_exponent[ch] = max_scale;
321 * Apply MDCT window and add into output.
323 * We ensure that when the windows overlap their squared sum
324 * is always 1 (MDCT reconstruction rule).
326 static void wma_window(WMACodecContext *s, float *out)
328 float *in = s->output;
329 int block_len, bsize, n;
332 if (s->block_len_bits <= s->prev_block_len_bits) {
333 block_len = s->block_len;
334 bsize = s->frame_len_bits - s->block_len_bits;
336 s->dsp.vector_fmul_add_add(out, in, s->windows[bsize],
337 out, 0, block_len, 1);
340 block_len = 1 << s->prev_block_len_bits;
341 n = (s->block_len - block_len) / 2;
342 bsize = s->frame_len_bits - s->prev_block_len_bits;
344 s->dsp.vector_fmul_add_add(out+n, in+n, s->windows[bsize],
345 out+n, 0, block_len, 1);
347 memcpy(out+n+block_len, in+n+block_len, n*sizeof(float));
354 if (s->block_len_bits <= s->next_block_len_bits) {
355 block_len = s->block_len;
356 bsize = s->frame_len_bits - s->block_len_bits;
358 s->dsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
361 block_len = 1 << s->next_block_len_bits;
362 n = (s->block_len - block_len) / 2;
363 bsize = s->frame_len_bits - s->next_block_len_bits;
365 memcpy(out, in, n*sizeof(float));
367 s->dsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
369 memset(out+n+block_len, 0, n*sizeof(float));
375 * @return 0 if OK. 1 if last block of frame. return -1 if
376 * unrecorrable error.
378 static int wma_decode_block(WMACodecContext *s)
380 int n, v, a, ch, code, bsize;
381 int coef_nb_bits, total_gain, parse_exponents;
382 int nb_coefs[MAX_CHANNELS];
386 tprintf(s->avctx, "***decode_block: %d:%d\n", s->frame_count - 1, s->block_num);
389 /* compute current block length */
390 if (s->use_variable_block_len) {
391 n = av_log2(s->nb_block_sizes - 1) + 1;
393 if (s->reset_block_lengths) {
394 s->reset_block_lengths = 0;
395 v = get_bits(&s->gb, n);
396 if (v >= s->nb_block_sizes)
398 s->prev_block_len_bits = s->frame_len_bits - v;
399 v = get_bits(&s->gb, n);
400 if (v >= s->nb_block_sizes)
402 s->block_len_bits = s->frame_len_bits - v;
404 /* update block lengths */
405 s->prev_block_len_bits = s->block_len_bits;
406 s->block_len_bits = s->next_block_len_bits;
408 v = get_bits(&s->gb, n);
409 if (v >= s->nb_block_sizes)
411 s->next_block_len_bits = s->frame_len_bits - v;
413 /* fixed block len */
414 s->next_block_len_bits = s->frame_len_bits;
415 s->prev_block_len_bits = s->frame_len_bits;
416 s->block_len_bits = s->frame_len_bits;
419 /* now check if the block length is coherent with the frame length */
420 s->block_len = 1 << s->block_len_bits;
421 if ((s->block_pos + s->block_len) > s->frame_len)
424 if (s->nb_channels == 2) {
425 s->ms_stereo = get_bits(&s->gb, 1);
428 for(ch = 0; ch < s->nb_channels; ch++) {
429 a = get_bits(&s->gb, 1);
430 s->channel_coded[ch] = a;
433 /* if no channel coded, no need to go further */
434 /* XXX: fix potential framing problems */
438 bsize = s->frame_len_bits - s->block_len_bits;
440 /* read total gain and extract corresponding number of bits for
441 coef escape coding */
444 a = get_bits(&s->gb, 7);
450 coef_nb_bits= ff_wma_total_gain_to_bits(total_gain);
452 /* compute number of coefficients */
453 n = s->coefs_end[bsize] - s->coefs_start;
454 for(ch = 0; ch < s->nb_channels; ch++)
458 if (s->use_noise_coding) {
460 for(ch = 0; ch < s->nb_channels; ch++) {
461 if (s->channel_coded[ch]) {
463 n = s->exponent_high_sizes[bsize];
465 a = get_bits(&s->gb, 1);
466 s->high_band_coded[ch][i] = a;
467 /* if noise coding, the coefficients are not transmitted */
469 nb_coefs[ch] -= s->exponent_high_bands[bsize][i];
473 for(ch = 0; ch < s->nb_channels; ch++) {
474 if (s->channel_coded[ch]) {
477 n = s->exponent_high_sizes[bsize];
478 val = (int)0x80000000;
480 if (s->high_band_coded[ch][i]) {
481 if (val == (int)0x80000000) {
482 val = get_bits(&s->gb, 7) - 19;
484 code = get_vlc2(&s->gb, s->hgain_vlc.table, HGAINVLCBITS, HGAINMAX);
489 s->high_band_values[ch][i] = val;
496 /* exposant can be interpolated in short blocks. */
498 if (s->block_len_bits != s->frame_len_bits) {
499 parse_exponents = get_bits(&s->gb, 1);
502 if (parse_exponents) {
503 for(ch = 0; ch < s->nb_channels; ch++) {
504 if (s->channel_coded[ch]) {
505 if (s->use_exp_vlc) {
506 if (decode_exp_vlc(s, ch) < 0)
509 decode_exp_lsp(s, ch);
514 for(ch = 0; ch < s->nb_channels; ch++) {
515 if (s->channel_coded[ch]) {
516 interpolate_array(s->exponents[ch], 1 << s->prev_block_len_bits,
522 /* parse spectral coefficients : just RLE encoding */
523 for(ch = 0; ch < s->nb_channels; ch++) {
524 if (s->channel_coded[ch]) {
526 int level, run, sign, tindex;
528 const uint16_t *level_table, *run_table;
530 /* special VLC tables are used for ms stereo because
531 there is potentially less energy there */
532 tindex = (ch == 1 && s->ms_stereo);
533 coef_vlc = &s->coef_vlc[tindex];
534 run_table = s->run_table[tindex];
535 level_table = s->level_table[tindex];
537 ptr = &s->coefs1[ch][0];
538 eptr = ptr + nb_coefs[ch];
539 memset(ptr, 0, s->block_len * sizeof(int16_t));
541 code = get_vlc2(&s->gb, coef_vlc->table, VLCBITS, VLCMAX);
547 } else if (code == 0) {
549 level = get_bits(&s->gb, coef_nb_bits);
550 /* NOTE: this is rather suboptimal. reading
551 block_len_bits would be better */
552 run = get_bits(&s->gb, s->frame_len_bits);
555 run = run_table[code];
556 level = level_table[code];
558 sign = get_bits(&s->gb, 1);
564 av_log(NULL, AV_LOG_ERROR, "overflow in spectral RLE, ignoring\n");
568 /* NOTE: EOB can be omitted */
573 if (s->version == 1 && s->nb_channels >= 2) {
574 align_get_bits(&s->gb);
580 int n4 = s->block_len / 2;
581 mdct_norm = 1.0 / (float)n4;
582 if (s->version == 1) {
583 mdct_norm *= sqrt(n4);
587 /* finally compute the MDCT coefficients */
588 for(ch = 0; ch < s->nb_channels; ch++) {
589 if (s->channel_coded[ch]) {
591 float *coefs, *exponents, mult, mult1, noise, *exp_ptr;
592 int i, j, n, n1, last_high_band;
593 float exp_power[HIGH_BAND_MAX_SIZE];
595 coefs1 = s->coefs1[ch];
596 exponents = s->exponents[ch];
597 mult = pow(10, total_gain * 0.05) / s->max_exponent[ch];
599 coefs = s->coefs[ch];
600 if (s->use_noise_coding) {
602 /* very low freqs : noise */
603 for(i = 0;i < s->coefs_start; i++) {
604 *coefs++ = s->noise_table[s->noise_index] * (*exponents++) * mult1;
605 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
608 n1 = s->exponent_high_sizes[bsize];
610 /* compute power of high bands */
611 exp_ptr = exponents +
612 s->high_band_start[bsize] -
614 last_high_band = 0; /* avoid warning */
616 n = s->exponent_high_bands[s->frame_len_bits -
617 s->block_len_bits][j];
618 if (s->high_band_coded[ch][j]) {
621 for(i = 0;i < n; i++) {
625 exp_power[j] = e2 / n;
627 tprintf(s->avctx, "%d: power=%f (%d)\n", j, exp_power[j], n);
632 /* main freqs and high freqs */
635 n = s->high_band_start[bsize] -
638 n = s->exponent_high_bands[s->frame_len_bits -
639 s->block_len_bits][j];
641 if (j >= 0 && s->high_band_coded[ch][j]) {
642 /* use noise with specified power */
643 mult1 = sqrt(exp_power[j] / exp_power[last_high_band]);
644 /* XXX: use a table */
645 mult1 = mult1 * pow(10, s->high_band_values[ch][j] * 0.05);
646 mult1 = mult1 / (s->max_exponent[ch] * s->noise_mult);
648 for(i = 0;i < n; i++) {
649 noise = s->noise_table[s->noise_index];
650 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
651 *coefs++ = (*exponents++) * noise * mult1;
654 /* coded values + small noise */
655 for(i = 0;i < n; i++) {
656 noise = s->noise_table[s->noise_index];
657 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
658 *coefs++ = ((*coefs1++) + noise) * (*exponents++) * mult;
663 /* very high freqs : noise */
664 n = s->block_len - s->coefs_end[bsize];
665 mult1 = mult * exponents[-1];
666 for(i = 0; i < n; i++) {
667 *coefs++ = s->noise_table[s->noise_index] * mult1;
668 s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
671 /* XXX: optimize more */
672 for(i = 0;i < s->coefs_start; i++)
675 for(i = 0;i < n; i++) {
676 *coefs++ = coefs1[i] * exponents[i] * mult;
678 n = s->block_len - s->coefs_end[bsize];
679 for(i = 0;i < n; i++)
686 for(ch = 0; ch < s->nb_channels; ch++) {
687 if (s->channel_coded[ch]) {
688 dump_floats(s, "exponents", 3, s->exponents[ch], s->block_len);
689 dump_floats(s, "coefs", 1, s->coefs[ch], s->block_len);
694 if (s->ms_stereo && s->channel_coded[1]) {
698 /* nominal case for ms stereo: we do it before mdct */
699 /* no need to optimize this case because it should almost
701 if (!s->channel_coded[0]) {
702 tprintf(s->avctx, "rare ms-stereo case happened\n");
703 memset(s->coefs[0], 0, sizeof(float) * s->block_len);
704 s->channel_coded[0] = 1;
707 for(i = 0; i < s->block_len; i++) {
710 s->coefs[0][i] = a + b;
711 s->coefs[1][i] = a - b;
715 for(ch = 0; ch < s->nb_channels; ch++) {
716 if (s->channel_coded[ch]) {
720 n4 = s->block_len / 2;
721 s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize],
722 s->output, s->coefs[ch], s->mdct_tmp);
724 /* multiply by the window and add in the frame */
725 index = (s->frame_len / 2) + s->block_pos - n4;
726 wma_window(s, &s->frame_out[ch][index]);
728 /* specific fast case for ms-stereo : add to second
729 channel if it is not coded */
730 if (s->ms_stereo && !s->channel_coded[1]) {
731 wma_window(s, &s->frame_out[1][index]);
736 /* update block number */
738 s->block_pos += s->block_len;
739 if (s->block_pos >= s->frame_len)
745 /* decode a frame of frame_len samples */
746 static int wma_decode_frame(WMACodecContext *s, int16_t *samples)
748 int ret, i, n, a, ch, incr;
753 tprintf(s->avctx, "***decode_frame: %d size=%d\n", s->frame_count++, s->frame_len);
756 /* read each block */
760 ret = wma_decode_block(s);
767 /* convert frame to integer */
769 incr = s->nb_channels;
770 for(ch = 0; ch < s->nb_channels; ch++) {
772 iptr = s->frame_out[ch];
783 /* prepare for next block */
784 memmove(&s->frame_out[ch][0], &s->frame_out[ch][s->frame_len],
785 s->frame_len * sizeof(float));
789 dump_shorts(s, "samples", samples, n * s->nb_channels);
794 static int wma_decode_superframe(AVCodecContext *avctx,
795 void *data, int *data_size,
796 uint8_t *buf, int buf_size)
798 WMACodecContext *s = avctx->priv_data;
799 int nb_frames, bit_offset, i, pos, len;
803 tprintf(avctx, "***decode_superframe:\n");
806 s->last_superframe_len = 0;
812 init_get_bits(&s->gb, buf, buf_size*8);
814 if (s->use_bit_reservoir) {
815 /* read super frame header */
816 get_bits(&s->gb, 4); /* super frame index */
817 nb_frames = get_bits(&s->gb, 4) - 1;
819 bit_offset = get_bits(&s->gb, s->byte_offset_bits + 3);
821 if (s->last_superframe_len > 0) {
822 // printf("skip=%d\n", s->last_bitoffset);
823 /* add bit_offset bits to last frame */
824 if ((s->last_superframe_len + ((bit_offset + 7) >> 3)) >
825 MAX_CODED_SUPERFRAME_SIZE)
827 q = s->last_superframe + s->last_superframe_len;
830 *q++ = (get_bits)(&s->gb, 8);
834 *q++ = (get_bits)(&s->gb, len) << (8 - len);
837 /* XXX: bit_offset bits into last frame */
838 init_get_bits(&s->gb, s->last_superframe, MAX_CODED_SUPERFRAME_SIZE*8);
839 /* skip unused bits */
840 if (s->last_bitoffset > 0)
841 skip_bits(&s->gb, s->last_bitoffset);
842 /* this frame is stored in the last superframe and in the
844 if (wma_decode_frame(s, samples) < 0)
846 samples += s->nb_channels * s->frame_len;
849 /* read each frame starting from bit_offset */
850 pos = bit_offset + 4 + 4 + s->byte_offset_bits + 3;
851 init_get_bits(&s->gb, buf + (pos >> 3), (MAX_CODED_SUPERFRAME_SIZE - (pos >> 3))*8);
854 skip_bits(&s->gb, len);
856 s->reset_block_lengths = 1;
857 for(i=0;i<nb_frames;i++) {
858 if (wma_decode_frame(s, samples) < 0)
860 samples += s->nb_channels * s->frame_len;
863 /* we copy the end of the frame in the last frame buffer */
864 pos = get_bits_count(&s->gb) + ((bit_offset + 4 + 4 + s->byte_offset_bits + 3) & ~7);
865 s->last_bitoffset = pos & 7;
867 len = buf_size - pos;
868 if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0) {
871 s->last_superframe_len = len;
872 memcpy(s->last_superframe, buf + pos, len);
874 /* single frame decode */
875 if (wma_decode_frame(s, samples) < 0)
877 samples += s->nb_channels * s->frame_len;
880 //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);
882 *data_size = (int8_t *)samples - (int8_t *)data;
883 return s->block_align;
885 /* when error, we reset the bit reservoir */
886 s->last_superframe_len = 0;
890 AVCodec wmav1_decoder =
895 sizeof(WMACodecContext),
899 wma_decode_superframe,
902 AVCodec wmav2_decoder =
907 sizeof(WMACodecContext),
911 wma_decode_superframe,