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
26 * @file imc.c IMC - Intel Music Coder
27 * A mdct based codec using a 256 points large transform
28 * divied into 32 bands with some mix of scale factors.
29 * Only mono is supported.
38 #define ALT_BITSTREAM_READER
40 #include "bitstream.h"
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];
89 DECLARE_ALIGNED_16(FFTComplex, samples[COEFFS/2]);
90 DECLARE_ALIGNED_16(float, out_samples[COEFFS]);
94 static int imc_decode_init(AVCodecContext * avctx)
97 IMCContext *q = avctx->priv_data;
100 q->decoder_reset = 1;
102 for(i = 0; i < BANDS; i++)
103 q->old_floor[i] = 1.0;
105 /* Build mdct window, a simple sine window normalized with sqrt(2) */
106 for(i = 0; i < COEFFS; i++)
107 q->mdct_sine_window[i] = sin((i + 0.5) / 512.0 * M_PI) * sqrt(2.0);
108 for(i = 0; i < COEFFS/2; i++){
109 q->post_cos[i] = cos(i / 256.0 * M_PI);
110 q->post_sin[i] = sin(i / 256.0 * M_PI);
112 r1 = sin((i * 4.0 + 1.0) / 1024.0 * M_PI);
113 r2 = cos((i * 4.0 + 1.0) / 1024.0 * M_PI);
117 q->pre_coef1[i] = (r1 + r2) * sqrt(2.0);
118 q->pre_coef2[i] = -(r1 - r2) * sqrt(2.0);
122 q->pre_coef1[i] = -(r1 + r2) * sqrt(2.0);
123 q->pre_coef2[i] = (r1 - r2) * sqrt(2.0);
126 q->last_fft_im[i] = 0;
128 q->flcf1 = log2(10) * 0.05703125;
129 q->flcf2 = log2(10) * 0.25;
131 /* Generate a square root table */
133 for(i = 0; i < 30; i++) {
134 q->sqrt_tab[i] = sqrt(i);
137 /* initialize the VLC tables */
138 for(i = 0; i < 4 ; i++) {
139 for(j = 0; j < 4; j++) {
140 init_vlc (&q->huffman_vlc[i][j], 9, imc_huffman_sizes[i],
141 imc_huffman_lens[i][j], 1, 1,
142 imc_huffman_bits[i][j], 2, 2, 1);
145 q->one_div_log2 = 1/log(2);
147 ff_fft_init(&q->fft, 7, 1);
148 dsputil_init(&q->dsp, avctx);
152 static void imc_calculate_coeffs(IMCContext* q, float* flcoeffs1, float* flcoeffs2, int* bandWidthT,
153 float* flcoeffs3, float* flcoeffs5)
158 float snr_limit = 1.e-30;
162 for(i = 0; i < BANDS; i++) {
163 flcoeffs5[i] = workT2[i] = 0.0;
165 workT1[i] = flcoeffs1[i] * flcoeffs1[i];
166 flcoeffs3[i] = 2.0 * flcoeffs2[i];
169 flcoeffs3[i] = -30000.0;
171 workT3[i] = bandWidthT[i] * workT1[i] * 0.01;
172 if (workT3[i] <= snr_limit)
176 for(i = 0; i < BANDS; i++) {
177 for(cnt2 = i; cnt2 < cyclTab[i]; cnt2++)
178 flcoeffs5[cnt2] = flcoeffs5[cnt2] + workT3[i];
179 workT2[cnt2-1] = workT2[cnt2-1] + workT3[i];
182 for(i = 1; i < BANDS; i++) {
183 accum = (workT2[i-1] + accum) * imc_weights1[i-1];
184 flcoeffs5[i] += accum;
187 for(i = 0; i < BANDS; i++)
190 for(i = 0; i < BANDS; i++) {
191 for(cnt2 = i-1; cnt2 > cyclTab2[i]; cnt2--)
192 flcoeffs5[cnt2] += workT3[i];
193 workT2[cnt2+1] += workT3[i];
198 for(i = BANDS-2; i >= 0; i--) {
199 accum = (workT2[i+1] + accum) * imc_weights2[i];
200 flcoeffs5[i] += accum;
201 //there is missing code here, but it seems to never be triggered
206 static void imc_read_level_coeffs(IMCContext* q, int stream_format_code, int* levlCoeffs)
211 const uint8_t *cb_sel;
214 s = stream_format_code >> 1;
215 hufftab[0] = &q->huffman_vlc[s][0];
216 hufftab[1] = &q->huffman_vlc[s][1];
217 hufftab[2] = &q->huffman_vlc[s][2];
218 hufftab[3] = &q->huffman_vlc[s][3];
219 cb_sel = imc_cb_select[s];
221 if(stream_format_code & 4)
224 levlCoeffs[0] = get_bits(&q->gb, 7);
225 for(i = start; i < BANDS; i++){
226 levlCoeffs[i] = get_vlc2(&q->gb, hufftab[cb_sel[i]]->table, hufftab[cb_sel[i]]->bits, 2);
227 if(levlCoeffs[i] == 17)
228 levlCoeffs[i] += get_bits(&q->gb, 4);
232 static void imc_decode_level_coefficients(IMCContext* q, int* levlCoeffBuf, float* flcoeffs1,
237 //maybe some frequency division thingy
239 flcoeffs1[0] = 20000.0 / pow (2, levlCoeffBuf[0] * q->flcf1);
240 flcoeffs2[0] = log2(flcoeffs1[0]);
244 for(i = 1; i < BANDS; i++) {
245 level = levlCoeffBuf[i];
252 else if (level <= 24)
257 tmp *= imc_exp_tab[15 + level];
258 tmp2 += q->flcf2 * level;
266 static void imc_decode_level_coefficients2(IMCContext* q, int* levlCoeffBuf, float* old_floor, float* flcoeffs1,
269 //FIXME maybe flag_buf = noise coding and flcoeffs1 = new scale factors
270 // and flcoeffs2 old scale factors
271 // might be incomplete due to a missing table that is in the binary code
272 for(i = 0; i < BANDS; i++) {
274 if(levlCoeffBuf[i] < 16) {
275 flcoeffs1[i] = imc_exp_tab2[levlCoeffBuf[i]] * old_floor[i];
276 flcoeffs2[i] = (levlCoeffBuf[i]-7) * q->flcf2 + flcoeffs2[i];
278 flcoeffs1[i] = old_floor[i];
284 * Perform bit allocation depending on bits available
286 static int bit_allocation (IMCContext* q, int stream_format_code, int freebits, int flag) {
288 const float limit = -1.e20;
297 float lowest = 1.e10;
303 for(i = 0; i < BANDS; i++)
304 highest = FFMAX(highest, q->flcoeffs1[i]);
306 for(i = 0; i < BANDS-1; i++) {
307 q->flcoeffs4[i] = q->flcoeffs3[i] - log2(q->flcoeffs5[i]);
309 q->flcoeffs4[BANDS - 1] = limit;
311 highest = highest * 0.25;
313 for(i = 0; i < BANDS; i++) {
315 if ((band_tab[i+1] - band_tab[i]) == q->bandWidthT[i])
318 if ((band_tab[i+1] - band_tab[i]) > q->bandWidthT[i])
321 if (((band_tab[i+1] - band_tab[i])/2) >= q->bandWidthT[i])
327 q->flcoeffs4[i] = q->flcoeffs4[i] + xTab[(indx*2 + (q->flcoeffs1[i] < highest)) * 2 + flag];
330 if (stream_format_code & 0x2) {
331 q->flcoeffs4[0] = limit;
332 q->flcoeffs4[1] = limit;
333 q->flcoeffs4[2] = limit;
334 q->flcoeffs4[3] = limit;
337 for(i = (stream_format_code & 0x2)?4:0; i < BANDS-1; i++) {
338 iacc += q->bandWidthT[i];
339 summa += q->bandWidthT[i] * q->flcoeffs4[i];
341 q->bandWidthT[BANDS-1] = 0;
342 summa = (summa * 0.5 - freebits) / iacc;
345 for(i = 0; i < BANDS/2; i++) {
346 rres = summer - freebits;
347 if((rres >= -8) && (rres <= 8)) break;
352 for(j = (stream_format_code & 0x2)?4:0; j < BANDS; j++) {
353 cwlen = clip((int)((q->flcoeffs4[j] * 0.5) - summa + 0.5), 0, 6);
355 q->bitsBandT[j] = cwlen;
356 summer += q->bandWidthT[j] * cwlen;
359 iacc += q->bandWidthT[j];
364 if (freebits < summer)
371 summa = (float)(summer - freebits) / ((t1 + 1) * iacc) + summa;
374 for(i = (stream_format_code & 0x2)?4:0; i < BANDS; i++) {
375 for(j = band_tab[i]; j < band_tab[i+1]; j++)
376 q->CWlengthT[j] = q->bitsBandT[i];
379 if (freebits > summer) {
380 for(i = 0; i < BANDS; i++) {
381 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
387 if (highest <= -1.e20)
393 for(i = 0; i < BANDS; i++) {
394 if (workT[i] > highest) {
400 if (highest > -1.e20) {
401 workT[found_indx] -= 2.0;
402 if (++(q->bitsBandT[found_indx]) == 6)
403 workT[found_indx] = -1.e20;
405 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (freebits > summer); j++){
410 }while (freebits > summer);
412 if (freebits < summer) {
413 for(i = 0; i < BANDS; i++) {
414 workT[i] = q->bitsBandT[i] ? (q->bitsBandT[i] * -2 + q->flcoeffs4[i] + 1.585) : 1.e20;
416 if (stream_format_code & 0x2) {
422 while (freebits < summer){
425 for(i = 0; i < BANDS; i++) {
426 if (workT[i] < lowest) {
431 //if(lowest >= 1.e10) break;
432 workT[low_indx] = lowest + 2.0;
434 if (!(--q->bitsBandT[low_indx]))
435 workT[low_indx] = 1.e20;
437 for(j = band_tab[low_indx]; j < band_tab[low_indx+1] && (freebits < summer); j++){
438 if(q->CWlengthT[j] > 0){
448 static void imc_get_skip_coeff(IMCContext* q) {
451 memset(q->skipFlagBits, 0, sizeof(q->skipFlagBits));
452 memset(q->skipFlagCount, 0, sizeof(q->skipFlagCount));
453 for(i = 0; i < BANDS; i++) {
454 if (!q->bandFlagsBuf[i] || !q->bandWidthT[i])
457 if (!q->skipFlagRaw[i]) {
458 q->skipFlagBits[i] = band_tab[i+1] - band_tab[i];
460 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
461 if ((q->skipFlags[j] = get_bits(&q->gb,1)))
462 q->skipFlagCount[i]++;
465 for(j = band_tab[i]; j < (band_tab[i+1]-1); j += 2) {
466 if(!get_bits1(&q->gb)){//0
467 q->skipFlagBits[i]++;
470 q->skipFlagCount[i] += 2;
472 if(get_bits1(&q->gb)){//11
473 q->skipFlagBits[i] +=2;
476 q->skipFlagCount[i]++;
478 q->skipFlagBits[i] +=3;
480 if(!get_bits1(&q->gb)){//100
482 q->skipFlagCount[i]++;
490 if (j < band_tab[i+1]) {
491 q->skipFlagBits[i]++;
492 if ((q->skipFlags[j] = get_bits(&q->gb,1)))
493 q->skipFlagCount[i]++;
500 * Increase highest' band coefficient sizes as some bits won't be used
502 static void imc_adjust_bit_allocation (IMCContext* q, int summer) {
509 for(i = 0; i < BANDS; i++) {
510 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
513 while (corrected < summer) {
514 if(highest <= -1.e20)
519 for(i = 0; i < BANDS; i++) {
520 if (workT[i] > highest) {
526 if (highest > -1.e20) {
527 workT[found_indx] -= 2.0;
528 if (++(q->bitsBandT[found_indx]) == 6)
529 workT[found_indx] = -1.e20;
531 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (corrected < summer); j++) {
532 if (!q->skipFlags[j] && (q->CWlengthT[j] < 6)) {
541 static void imc_imdct256(IMCContext *q) {
546 for(i=0; i < COEFFS/2; i++){
547 q->samples[i].re = -(q->pre_coef1[i] * q->CWdecoded[COEFFS-1-i*2]) -
548 (q->pre_coef2[i] * q->CWdecoded[i*2]);
549 q->samples[i].im = (q->pre_coef2[i] * q->CWdecoded[COEFFS-1-i*2]) -
550 (q->pre_coef1[i] * q->CWdecoded[i*2]);
554 ff_fft_permute(&q->fft, q->samples);
555 ff_fft_calc (&q->fft, q->samples);
557 /* postrotation, window and reorder */
558 for(i = 0; i < COEFFS/2; i++){
559 re = (q->samples[i].re * q->post_cos[i]) + (-q->samples[i].im * q->post_sin[i]);
560 im = (-q->samples[i].im * q->post_cos[i]) - (q->samples[i].re * q->post_sin[i]);
561 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);
562 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);
563 q->last_fft_im[i] = im;
567 static int inverse_quant_coeff (IMCContext* q, int stream_format_code) {
569 int middle_value, cw_len, max_size;
570 const float* quantizer;
572 for(i = 0; i < BANDS; i++) {
573 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
575 cw_len = q->CWlengthT[j];
577 if (cw_len <= 0 || q->skipFlags[j])
580 max_size = 1 << cw_len;
581 middle_value = max_size >> 1;
583 if (q->codewords[j] >= max_size || q->codewords[j] < 0)
587 quantizer = imc_quantizer2[(stream_format_code & 2) >> 1];
588 if (q->codewords[j] >= middle_value)
589 q->CWdecoded[j] = quantizer[q->codewords[j] - 8] * q->flcoeffs6[i];
591 q->CWdecoded[j] = -quantizer[max_size - q->codewords[j] - 8 - 1] * q->flcoeffs6[i];
593 quantizer = imc_quantizer1[((stream_format_code & 2) >> 1) | (q->bandFlagsBuf[i] << 1)];
594 if (q->codewords[j] >= middle_value)
595 q->CWdecoded[j] = quantizer[q->codewords[j] - 1] * q->flcoeffs6[i];
597 q->CWdecoded[j] = -quantizer[max_size - 2 - q->codewords[j]] * q->flcoeffs6[i];
605 static int imc_get_coeffs (IMCContext* q) {
606 int i, j, cw_len, cw;
608 for(i = 0; i < BANDS; i++) {
609 if(!q->sumLenArr[i]) continue;
610 if (q->bandFlagsBuf[i] || q->bandWidthT[i]) {
611 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
612 cw_len = q->CWlengthT[j];
615 if (get_bits_count(&q->gb) + cw_len > 512){
616 //av_log(NULL,0,"Band %i coeff %i cw_len %i\n",i,j,cw_len);
620 if(cw_len && (!q->bandFlagsBuf[i] || !q->skipFlags[j]))
621 cw = get_bits(&q->gb, cw_len);
623 q->codewords[j] = cw;
630 static int imc_decode_frame(AVCodecContext * avctx,
631 void *data, int *data_size,
632 uint8_t * buf, int buf_size)
635 IMCContext *q = avctx->priv_data;
637 int stream_format_code;
641 int counter, bitscount;
642 uint16_t *buf16 = (uint16_t *) buf;
644 /* FIXME: input should not be modified */
645 for(i = 0; i < FFMIN(buf_size, avctx->block_align) / 2; i++)
646 buf16[i] = bswap_16(buf16[i]);
648 init_get_bits(&q->gb, buf, 512);
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 avctx->block_align;
799 static 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,