2 * IMC compatible decoder
3 * Copyright (c) 2002-2004 Maxim Poliakovski
4 * Copyright (c) 2006 Benjamin Larsson
5 * Copyright (c) 2006 Konstantin Shishkov
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * @file imc.c IMC - Intel Music Coder
26 * A mdct based codec using a 256 points large transform
27 * divied into 32 bands with some mix of scale factors.
28 * Only mono is supported.
37 #define ALT_BITSTREAM_READER
39 #include "bitstream.h"
44 #define IMC_BLOCK_SIZE 64
45 #define IMC_FRAME_ID 0x21
50 float old_floor[BANDS];
51 float flcoeffs1[BANDS];
52 float flcoeffs2[BANDS];
53 float flcoeffs3[BANDS];
54 float flcoeffs4[BANDS];
55 float flcoeffs5[BANDS];
56 float flcoeffs6[BANDS];
57 float CWdecoded[COEFFS];
61 float mdct_sine_window[COEFFS];
62 float post_cos[COEFFS];
63 float post_sin[COEFFS];
64 float pre_coef1[COEFFS];
65 float pre_coef2[COEFFS];
66 float last_fft_im[COEFFS];
69 int bandWidthT[BANDS]; ///< codewords per band
70 int bitsBandT[BANDS]; ///< how many bits per codeword in band
71 int CWlengthT[COEFFS]; ///< how many bits in each codeword
72 int levlCoeffBuf[BANDS];
73 int bandFlagsBuf[BANDS]; ///< flags for each band
74 int sumLenArr[BANDS]; ///< bits for all coeffs in band
75 int skipFlagRaw[BANDS]; ///< skip flags are stored in raw form or not
76 int skipFlagBits[BANDS]; ///< bits used to code skip flags
77 int skipFlagCount[BANDS]; ///< skipped coeffients per band
78 int skipFlags[COEFFS]; ///< skip coefficient decoding or not
79 int codewords[COEFFS]; ///< raw codewords read from bitstream
82 VLC huffman_vlc[4][4];
88 DECLARE_ALIGNED_16(FFTComplex, samples[COEFFS/2]);
89 DECLARE_ALIGNED_16(float, out_samples[COEFFS]);
93 static av_cold int imc_decode_init(AVCodecContext * avctx)
96 IMCContext *q = avctx->priv_data;
101 for(i = 0; i < BANDS; i++)
102 q->old_floor[i] = 1.0;
104 /* Build mdct window, a simple sine window normalized with sqrt(2) */
105 for(i = 0; i < COEFFS; i++)
106 q->mdct_sine_window[i] = sin((i + 0.5) / 512.0 * M_PI) * sqrt(2.0);
107 for(i = 0; i < COEFFS/2; i++){
108 q->post_cos[i] = cos(i / 256.0 * M_PI);
109 q->post_sin[i] = sin(i / 256.0 * M_PI);
111 r1 = sin((i * 4.0 + 1.0) / 1024.0 * M_PI);
112 r2 = cos((i * 4.0 + 1.0) / 1024.0 * M_PI);
116 q->pre_coef1[i] = (r1 + r2) * sqrt(2.0);
117 q->pre_coef2[i] = -(r1 - r2) * sqrt(2.0);
121 q->pre_coef1[i] = -(r1 + r2) * sqrt(2.0);
122 q->pre_coef2[i] = (r1 - r2) * sqrt(2.0);
125 q->last_fft_im[i] = 0;
128 /* Generate a square root table */
130 for(i = 0; i < 30; i++) {
131 q->sqrt_tab[i] = sqrt(i);
134 /* initialize the VLC tables */
135 for(i = 0; i < 4 ; i++) {
136 for(j = 0; j < 4; j++) {
137 init_vlc (&q->huffman_vlc[i][j], 9, imc_huffman_sizes[i],
138 imc_huffman_lens[i][j], 1, 1,
139 imc_huffman_bits[i][j], 2, 2, 1);
142 q->one_div_log2 = 1/log(2);
144 ff_fft_init(&q->fft, 7, 1);
145 dsputil_init(&q->dsp, avctx);
149 static void imc_calculate_coeffs(IMCContext* q, float* flcoeffs1, float* flcoeffs2, int* bandWidthT,
150 float* flcoeffs3, float* flcoeffs5)
155 float snr_limit = 1.e-30;
159 for(i = 0; i < BANDS; i++) {
160 flcoeffs5[i] = workT2[i] = 0.0;
162 workT1[i] = flcoeffs1[i] * flcoeffs1[i];
163 flcoeffs3[i] = 2.0 * flcoeffs2[i];
166 flcoeffs3[i] = -30000.0;
168 workT3[i] = bandWidthT[i] * workT1[i] * 0.01;
169 if (workT3[i] <= snr_limit)
173 for(i = 0; i < BANDS; i++) {
174 for(cnt2 = i; cnt2 < cyclTab[i]; cnt2++)
175 flcoeffs5[cnt2] = flcoeffs5[cnt2] + workT3[i];
176 workT2[cnt2-1] = workT2[cnt2-1] + workT3[i];
179 for(i = 1; i < BANDS; i++) {
180 accum = (workT2[i-1] + accum) * imc_weights1[i-1];
181 flcoeffs5[i] += accum;
184 for(i = 0; i < BANDS; i++)
187 for(i = 0; i < BANDS; i++) {
188 for(cnt2 = i-1; cnt2 > cyclTab2[i]; cnt2--)
189 flcoeffs5[cnt2] += workT3[i];
190 workT2[cnt2+1] += workT3[i];
195 for(i = BANDS-2; i >= 0; i--) {
196 accum = (workT2[i+1] + accum) * imc_weights2[i];
197 flcoeffs5[i] += accum;
198 //there is missing code here, but it seems to never be triggered
203 static void imc_read_level_coeffs(IMCContext* q, int stream_format_code, int* levlCoeffs)
208 const uint8_t *cb_sel;
211 s = stream_format_code >> 1;
212 hufftab[0] = &q->huffman_vlc[s][0];
213 hufftab[1] = &q->huffman_vlc[s][1];
214 hufftab[2] = &q->huffman_vlc[s][2];
215 hufftab[3] = &q->huffman_vlc[s][3];
216 cb_sel = imc_cb_select[s];
218 if(stream_format_code & 4)
221 levlCoeffs[0] = get_bits(&q->gb, 7);
222 for(i = start; i < BANDS; i++){
223 levlCoeffs[i] = get_vlc2(&q->gb, hufftab[cb_sel[i]]->table, hufftab[cb_sel[i]]->bits, 2);
224 if(levlCoeffs[i] == 17)
225 levlCoeffs[i] += get_bits(&q->gb, 4);
229 static void imc_decode_level_coefficients(IMCContext* q, int* levlCoeffBuf, float* flcoeffs1,
234 //maybe some frequency division thingy
236 flcoeffs1[0] = 20000.0 / pow (2, levlCoeffBuf[0] * 0.18945); // 0.18945 = log2(10) * 0.05703125
237 flcoeffs2[0] = log(flcoeffs1[0])/log(2);
241 for(i = 1; i < BANDS; i++) {
242 level = levlCoeffBuf[i];
249 else if (level <= 24)
254 tmp *= imc_exp_tab[15 + level];
255 tmp2 += 0.83048 * level; // 0.83048 = log2(10) * 0.25
263 static void imc_decode_level_coefficients2(IMCContext* q, int* levlCoeffBuf, float* old_floor, float* flcoeffs1,
266 //FIXME maybe flag_buf = noise coding and flcoeffs1 = new scale factors
267 // and flcoeffs2 old scale factors
268 // might be incomplete due to a missing table that is in the binary code
269 for(i = 0; i < BANDS; i++) {
271 if(levlCoeffBuf[i] < 16) {
272 flcoeffs1[i] = imc_exp_tab2[levlCoeffBuf[i]] * old_floor[i];
273 flcoeffs2[i] = (levlCoeffBuf[i]-7) * 0.83048 + flcoeffs2[i]; // 0.83048 = log2(10) * 0.25
275 flcoeffs1[i] = old_floor[i];
281 * Perform bit allocation depending on bits available
283 static int bit_allocation (IMCContext* q, int stream_format_code, int freebits, int flag) {
285 const float limit = -1.e20;
294 float lowest = 1.e10;
300 for(i = 0; i < BANDS; i++)
301 highest = FFMAX(highest, q->flcoeffs1[i]);
303 for(i = 0; i < BANDS-1; i++) {
304 q->flcoeffs4[i] = q->flcoeffs3[i] - log(q->flcoeffs5[i])/log(2);
306 q->flcoeffs4[BANDS - 1] = limit;
308 highest = highest * 0.25;
310 for(i = 0; i < BANDS; i++) {
312 if ((band_tab[i+1] - band_tab[i]) == q->bandWidthT[i])
315 if ((band_tab[i+1] - band_tab[i]) > q->bandWidthT[i])
318 if (((band_tab[i+1] - band_tab[i])/2) >= q->bandWidthT[i])
324 q->flcoeffs4[i] = q->flcoeffs4[i] + xTab[(indx*2 + (q->flcoeffs1[i] < highest)) * 2 + flag];
327 if (stream_format_code & 0x2) {
328 q->flcoeffs4[0] = limit;
329 q->flcoeffs4[1] = limit;
330 q->flcoeffs4[2] = limit;
331 q->flcoeffs4[3] = limit;
334 for(i = (stream_format_code & 0x2)?4:0; i < BANDS-1; i++) {
335 iacc += q->bandWidthT[i];
336 summa += q->bandWidthT[i] * q->flcoeffs4[i];
338 q->bandWidthT[BANDS-1] = 0;
339 summa = (summa * 0.5 - freebits) / iacc;
342 for(i = 0; i < BANDS/2; i++) {
343 rres = summer - freebits;
344 if((rres >= -8) && (rres <= 8)) break;
349 for(j = (stream_format_code & 0x2)?4:0; j < BANDS; j++) {
350 cwlen = av_clip((int)((q->flcoeffs4[j] * 0.5) - summa + 0.5), 0, 6);
352 q->bitsBandT[j] = cwlen;
353 summer += q->bandWidthT[j] * cwlen;
356 iacc += q->bandWidthT[j];
361 if (freebits < summer)
368 summa = (float)(summer - freebits) / ((t1 + 1) * iacc) + summa;
371 for(i = (stream_format_code & 0x2)?4:0; i < BANDS; i++) {
372 for(j = band_tab[i]; j < band_tab[i+1]; j++)
373 q->CWlengthT[j] = q->bitsBandT[i];
376 if (freebits > summer) {
377 for(i = 0; i < BANDS; i++) {
378 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
384 if (highest <= -1.e20)
390 for(i = 0; i < BANDS; i++) {
391 if (workT[i] > highest) {
397 if (highest > -1.e20) {
398 workT[found_indx] -= 2.0;
399 if (++(q->bitsBandT[found_indx]) == 6)
400 workT[found_indx] = -1.e20;
402 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (freebits > summer); j++){
407 }while (freebits > summer);
409 if (freebits < summer) {
410 for(i = 0; i < BANDS; i++) {
411 workT[i] = q->bitsBandT[i] ? (q->bitsBandT[i] * -2 + q->flcoeffs4[i] + 1.585) : 1.e20;
413 if (stream_format_code & 0x2) {
419 while (freebits < summer){
422 for(i = 0; i < BANDS; i++) {
423 if (workT[i] < lowest) {
428 //if(lowest >= 1.e10) break;
429 workT[low_indx] = lowest + 2.0;
431 if (!(--q->bitsBandT[low_indx]))
432 workT[low_indx] = 1.e20;
434 for(j = band_tab[low_indx]; j < band_tab[low_indx+1] && (freebits < summer); j++){
435 if(q->CWlengthT[j] > 0){
445 static void imc_get_skip_coeff(IMCContext* q) {
448 memset(q->skipFlagBits, 0, sizeof(q->skipFlagBits));
449 memset(q->skipFlagCount, 0, sizeof(q->skipFlagCount));
450 for(i = 0; i < BANDS; i++) {
451 if (!q->bandFlagsBuf[i] || !q->bandWidthT[i])
454 if (!q->skipFlagRaw[i]) {
455 q->skipFlagBits[i] = band_tab[i+1] - band_tab[i];
457 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
458 if ((q->skipFlags[j] = get_bits1(&q->gb)))
459 q->skipFlagCount[i]++;
462 for(j = band_tab[i]; j < (band_tab[i+1]-1); j += 2) {
463 if(!get_bits1(&q->gb)){//0
464 q->skipFlagBits[i]++;
467 q->skipFlagCount[i] += 2;
469 if(get_bits1(&q->gb)){//11
470 q->skipFlagBits[i] +=2;
473 q->skipFlagCount[i]++;
475 q->skipFlagBits[i] +=3;
477 if(!get_bits1(&q->gb)){//100
479 q->skipFlagCount[i]++;
487 if (j < band_tab[i+1]) {
488 q->skipFlagBits[i]++;
489 if ((q->skipFlags[j] = get_bits1(&q->gb)))
490 q->skipFlagCount[i]++;
497 * Increase highest' band coefficient sizes as some bits won't be used
499 static void imc_adjust_bit_allocation (IMCContext* q, int summer) {
506 for(i = 0; i < BANDS; i++) {
507 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
510 while (corrected < summer) {
511 if(highest <= -1.e20)
516 for(i = 0; i < BANDS; i++) {
517 if (workT[i] > highest) {
523 if (highest > -1.e20) {
524 workT[found_indx] -= 2.0;
525 if (++(q->bitsBandT[found_indx]) == 6)
526 workT[found_indx] = -1.e20;
528 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (corrected < summer); j++) {
529 if (!q->skipFlags[j] && (q->CWlengthT[j] < 6)) {
538 static void imc_imdct256(IMCContext *q) {
543 for(i=0; i < COEFFS/2; i++){
544 q->samples[i].re = -(q->pre_coef1[i] * q->CWdecoded[COEFFS-1-i*2]) -
545 (q->pre_coef2[i] * q->CWdecoded[i*2]);
546 q->samples[i].im = (q->pre_coef2[i] * q->CWdecoded[COEFFS-1-i*2]) -
547 (q->pre_coef1[i] * q->CWdecoded[i*2]);
551 ff_fft_permute(&q->fft, q->samples);
552 ff_fft_calc (&q->fft, q->samples);
554 /* postrotation, window and reorder */
555 for(i = 0; i < COEFFS/2; i++){
556 re = (q->samples[i].re * q->post_cos[i]) + (-q->samples[i].im * q->post_sin[i]);
557 im = (-q->samples[i].im * q->post_cos[i]) - (q->samples[i].re * q->post_sin[i]);
558 q->out_samples[i*2] = (q->mdct_sine_window[COEFFS-1-i*2] * q->last_fft_im[i]) + (q->mdct_sine_window[i*2] * re);
559 q->out_samples[COEFFS-1-i*2] = (q->mdct_sine_window[i*2] * q->last_fft_im[i]) - (q->mdct_sine_window[COEFFS-1-i*2] * re);
560 q->last_fft_im[i] = im;
564 static int inverse_quant_coeff (IMCContext* q, int stream_format_code) {
566 int middle_value, cw_len, max_size;
567 const float* quantizer;
569 for(i = 0; i < BANDS; i++) {
570 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
572 cw_len = q->CWlengthT[j];
574 if (cw_len <= 0 || q->skipFlags[j])
577 max_size = 1 << cw_len;
578 middle_value = max_size >> 1;
580 if (q->codewords[j] >= max_size || q->codewords[j] < 0)
584 quantizer = imc_quantizer2[(stream_format_code & 2) >> 1];
585 if (q->codewords[j] >= middle_value)
586 q->CWdecoded[j] = quantizer[q->codewords[j] - 8] * q->flcoeffs6[i];
588 q->CWdecoded[j] = -quantizer[max_size - q->codewords[j] - 8 - 1] * q->flcoeffs6[i];
590 quantizer = imc_quantizer1[((stream_format_code & 2) >> 1) | (q->bandFlagsBuf[i] << 1)];
591 if (q->codewords[j] >= middle_value)
592 q->CWdecoded[j] = quantizer[q->codewords[j] - 1] * q->flcoeffs6[i];
594 q->CWdecoded[j] = -quantizer[max_size - 2 - q->codewords[j]] * q->flcoeffs6[i];
602 static int imc_get_coeffs (IMCContext* q) {
603 int i, j, cw_len, cw;
605 for(i = 0; i < BANDS; i++) {
606 if(!q->sumLenArr[i]) continue;
607 if (q->bandFlagsBuf[i] || q->bandWidthT[i]) {
608 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
609 cw_len = q->CWlengthT[j];
612 if (get_bits_count(&q->gb) + cw_len > 512){
613 //av_log(NULL,0,"Band %i coeff %i cw_len %i\n",i,j,cw_len);
617 if(cw_len && (!q->bandFlagsBuf[i] || !q->skipFlags[j]))
618 cw = get_bits(&q->gb, cw_len);
620 q->codewords[j] = cw;
627 static int imc_decode_frame(AVCodecContext * avctx,
628 void *data, int *data_size,
629 const uint8_t * buf, int buf_size)
632 IMCContext *q = avctx->priv_data;
634 int stream_format_code;
638 int counter, bitscount;
639 uint16_t buf16[IMC_BLOCK_SIZE / 2];
641 if (buf_size < IMC_BLOCK_SIZE) {
642 av_log(avctx, AV_LOG_ERROR, "imc frame too small!\n");
645 for(i = 0; i < IMC_BLOCK_SIZE / 2; i++)
646 buf16[i] = bswap_16(((const uint16_t*)buf)[i]);
648 init_get_bits(&q->gb, (const uint8_t*)buf16, IMC_BLOCK_SIZE * 8);
650 /* Check the frame header */
651 imc_hdr = get_bits(&q->gb, 9);
652 if (imc_hdr != IMC_FRAME_ID) {
653 av_log(avctx, AV_LOG_ERROR, "imc frame header check failed!\n");
654 av_log(avctx, AV_LOG_ERROR, "got %x instead of 0x21.\n", imc_hdr);
657 stream_format_code = get_bits(&q->gb, 3);
659 if(stream_format_code & 1){
660 av_log(avctx, AV_LOG_ERROR, "Stream code format %X is not supported\n", stream_format_code);
664 // av_log(avctx, AV_LOG_DEBUG, "stream_format_code = %d\n", stream_format_code);
666 if (stream_format_code & 0x04)
667 q->decoder_reset = 1;
669 if(q->decoder_reset) {
670 memset(q->out_samples, 0, sizeof(q->out_samples));
671 for(i = 0; i < BANDS; i++)q->old_floor[i] = 1.0;
672 for(i = 0; i < COEFFS; i++)q->CWdecoded[i] = 0;
673 q->decoder_reset = 0;
676 flag = get_bits1(&q->gb);
677 imc_read_level_coeffs(q, stream_format_code, q->levlCoeffBuf);
679 if (stream_format_code & 0x4)
680 imc_decode_level_coefficients(q, q->levlCoeffBuf, q->flcoeffs1, q->flcoeffs2);
682 imc_decode_level_coefficients2(q, q->levlCoeffBuf, q->old_floor, q->flcoeffs1, q->flcoeffs2);
684 memcpy(q->old_floor, q->flcoeffs1, 32 * sizeof(float));
687 for (i=0 ; i<BANDS ; i++) {
688 if (q->levlCoeffBuf[i] == 16) {
689 q->bandWidthT[i] = 0;
692 q->bandWidthT[i] = band_tab[i+1] - band_tab[i];
694 memset(q->bandFlagsBuf, 0, BANDS * sizeof(int));
695 for(i = 0; i < BANDS-1; i++) {
696 if (q->bandWidthT[i])
697 q->bandFlagsBuf[i] = get_bits1(&q->gb);
700 imc_calculate_coeffs(q, q->flcoeffs1, q->flcoeffs2, q->bandWidthT, q->flcoeffs3, q->flcoeffs5);
703 /* first 4 bands will be assigned 5 bits per coefficient */
704 if (stream_format_code & 0x2) {
711 for(i = 1; i < 4; i++){
712 bits = (q->levlCoeffBuf[i] == 16) ? 0 : 5;
713 q->bitsBandT[i] = bits;
714 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
715 q->CWlengthT[j] = bits;
721 if(bit_allocation (q, stream_format_code, 512 - bitscount - get_bits_count(&q->gb), flag) < 0) {
722 av_log(avctx, AV_LOG_ERROR, "Bit allocations failed\n");
723 q->decoder_reset = 1;
727 for(i = 0; i < BANDS; i++) {
729 q->skipFlagRaw[i] = 0;
730 for(j = band_tab[i]; j < band_tab[i+1]; j++)
731 q->sumLenArr[i] += q->CWlengthT[j];
732 if (q->bandFlagsBuf[i])
733 if( (((band_tab[i+1] - band_tab[i]) * 1.5) > q->sumLenArr[i]) && (q->sumLenArr[i] > 0))
734 q->skipFlagRaw[i] = 1;
737 imc_get_skip_coeff(q);
739 for(i = 0; i < BANDS; i++) {
740 q->flcoeffs6[i] = q->flcoeffs1[i];
741 /* band has flag set and at least one coded coefficient */
742 if (q->bandFlagsBuf[i] && (band_tab[i+1] - band_tab[i]) != q->skipFlagCount[i]){
743 q->flcoeffs6[i] *= q->sqrt_tab[band_tab[i+1] - band_tab[i]] /
744 q->sqrt_tab[(band_tab[i+1] - band_tab[i] - q->skipFlagCount[i])];
748 /* calculate bits left, bits needed and adjust bit allocation */
751 for(i = 0; i < BANDS; i++) {
752 if (q->bandFlagsBuf[i]) {
753 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
754 if(q->skipFlags[j]) {
755 summer += q->CWlengthT[j];
759 bits += q->skipFlagBits[i];
760 summer -= q->skipFlagBits[i];
763 imc_adjust_bit_allocation(q, summer);
765 for(i = 0; i < BANDS; i++) {
768 for(j = band_tab[i]; j < band_tab[i+1]; j++)
769 if (!q->skipFlags[j])
770 q->sumLenArr[i] += q->CWlengthT[j];
773 memset(q->codewords, 0, sizeof(q->codewords));
775 if(imc_get_coeffs(q) < 0) {
776 av_log(avctx, AV_LOG_ERROR, "Read coefficients failed\n");
777 q->decoder_reset = 1;
781 if(inverse_quant_coeff(q, stream_format_code) < 0) {
782 av_log(avctx, AV_LOG_ERROR, "Inverse quantization of coefficients failed\n");
783 q->decoder_reset = 1;
787 memset(q->skipFlags, 0, sizeof(q->skipFlags));
791 q->dsp.float_to_int16(data, q->out_samples, COEFFS);
793 *data_size = COEFFS * sizeof(int16_t);
795 return IMC_BLOCK_SIZE;
799 static av_cold int imc_decode_close(AVCodecContext * avctx)
801 IMCContext *q = avctx->priv_data;
808 AVCodec imc_decoder = {
810 .type = CODEC_TYPE_AUDIO,
812 .priv_data_size = sizeof(IMCContext),
813 .init = imc_decode_init,
814 .close = imc_decode_close,
815 .decode = imc_decode_frame,