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_FRAME_ID 0x21
49 float old_floor[BANDS];
50 float flcoeffs1[BANDS];
51 float flcoeffs2[BANDS];
52 float flcoeffs3[BANDS];
53 float flcoeffs4[BANDS];
54 float flcoeffs5[BANDS];
55 float flcoeffs6[BANDS];
56 float CWdecoded[COEFFS];
60 float mdct_sine_window[COEFFS];
61 float post_cos[COEFFS];
62 float post_sin[COEFFS];
63 float pre_coef1[COEFFS];
64 float pre_coef2[COEFFS];
65 float last_fft_im[COEFFS];
68 int bandWidthT[BANDS]; ///< codewords per band
69 int bitsBandT[BANDS]; ///< how many bits per codeword in band
70 int CWlengthT[COEFFS]; ///< how many bits in each codeword
71 int levlCoeffBuf[BANDS];
72 int bandFlagsBuf[BANDS]; ///< flags for each band
73 int sumLenArr[BANDS]; ///< bits for all coeffs in band
74 int skipFlagRaw[BANDS]; ///< skip flags are stored in raw form or not
75 int skipFlagBits[BANDS]; ///< bits used to code skip flags
76 int skipFlagCount[BANDS]; ///< skipped coeffients per band
77 int skipFlags[COEFFS]; ///< skip coefficient decoding or not
78 int codewords[COEFFS]; ///< raw codewords read from bitstream
81 VLC huffman_vlc[4][4];
87 DECLARE_ALIGNED_16(FFTComplex, samples[COEFFS/2]);
88 DECLARE_ALIGNED_16(float, out_samples[COEFFS]);
92 static int imc_decode_init(AVCodecContext * avctx)
95 IMCContext *q = avctx->priv_data;
100 for(i = 0; i < BANDS; i++)
101 q->old_floor[i] = 1.0;
103 /* Build mdct window, a simple sine window normalized with sqrt(2) */
104 for(i = 0; i < COEFFS; i++)
105 q->mdct_sine_window[i] = sin((i + 0.5) / 512.0 * M_PI) * sqrt(2.0);
106 for(i = 0; i < COEFFS/2; i++){
107 q->post_cos[i] = cos(i / 256.0 * M_PI);
108 q->post_sin[i] = sin(i / 256.0 * M_PI);
110 r1 = sin((i * 4.0 + 1.0) / 1024.0 * M_PI);
111 r2 = cos((i * 4.0 + 1.0) / 1024.0 * M_PI);
115 q->pre_coef1[i] = (r1 + r2) * sqrt(2.0);
116 q->pre_coef2[i] = -(r1 - r2) * sqrt(2.0);
120 q->pre_coef1[i] = -(r1 + r2) * sqrt(2.0);
121 q->pre_coef2[i] = (r1 - r2) * sqrt(2.0);
124 q->last_fft_im[i] = 0;
127 /* Generate a square root table */
129 for(i = 0; i < 30; i++) {
130 q->sqrt_tab[i] = sqrt(i);
133 /* initialize the VLC tables */
134 for(i = 0; i < 4 ; i++) {
135 for(j = 0; j < 4; j++) {
136 init_vlc (&q->huffman_vlc[i][j], 9, imc_huffman_sizes[i],
137 imc_huffman_lens[i][j], 1, 1,
138 imc_huffman_bits[i][j], 2, 2, 1);
141 q->one_div_log2 = 1/log(2);
143 ff_fft_init(&q->fft, 7, 1);
144 dsputil_init(&q->dsp, avctx);
148 static void imc_calculate_coeffs(IMCContext* q, float* flcoeffs1, float* flcoeffs2, int* bandWidthT,
149 float* flcoeffs3, float* flcoeffs5)
154 float snr_limit = 1.e-30;
158 for(i = 0; i < BANDS; i++) {
159 flcoeffs5[i] = workT2[i] = 0.0;
161 workT1[i] = flcoeffs1[i] * flcoeffs1[i];
162 flcoeffs3[i] = 2.0 * flcoeffs2[i];
165 flcoeffs3[i] = -30000.0;
167 workT3[i] = bandWidthT[i] * workT1[i] * 0.01;
168 if (workT3[i] <= snr_limit)
172 for(i = 0; i < BANDS; i++) {
173 for(cnt2 = i; cnt2 < cyclTab[i]; cnt2++)
174 flcoeffs5[cnt2] = flcoeffs5[cnt2] + workT3[i];
175 workT2[cnt2-1] = workT2[cnt2-1] + workT3[i];
178 for(i = 1; i < BANDS; i++) {
179 accum = (workT2[i-1] + accum) * imc_weights1[i-1];
180 flcoeffs5[i] += accum;
183 for(i = 0; i < BANDS; i++)
186 for(i = 0; i < BANDS; i++) {
187 for(cnt2 = i-1; cnt2 > cyclTab2[i]; cnt2--)
188 flcoeffs5[cnt2] += workT3[i];
189 workT2[cnt2+1] += workT3[i];
194 for(i = BANDS-2; i >= 0; i--) {
195 accum = (workT2[i+1] + accum) * imc_weights2[i];
196 flcoeffs5[i] += accum;
197 //there is missing code here, but it seems to never be triggered
202 static void imc_read_level_coeffs(IMCContext* q, int stream_format_code, int* levlCoeffs)
207 const uint8_t *cb_sel;
210 s = stream_format_code >> 1;
211 hufftab[0] = &q->huffman_vlc[s][0];
212 hufftab[1] = &q->huffman_vlc[s][1];
213 hufftab[2] = &q->huffman_vlc[s][2];
214 hufftab[3] = &q->huffman_vlc[s][3];
215 cb_sel = imc_cb_select[s];
217 if(stream_format_code & 4)
220 levlCoeffs[0] = get_bits(&q->gb, 7);
221 for(i = start; i < BANDS; i++){
222 levlCoeffs[i] = get_vlc2(&q->gb, hufftab[cb_sel[i]]->table, hufftab[cb_sel[i]]->bits, 2);
223 if(levlCoeffs[i] == 17)
224 levlCoeffs[i] += get_bits(&q->gb, 4);
228 static void imc_decode_level_coefficients(IMCContext* q, int* levlCoeffBuf, float* flcoeffs1,
233 //maybe some frequency division thingy
235 flcoeffs1[0] = 20000.0 / pow (2, levlCoeffBuf[0] * 0.18945); // 0.18945 = log2(10) * 0.05703125
236 flcoeffs2[0] = log(flcoeffs1[0])/log(2);
240 for(i = 1; i < BANDS; i++) {
241 level = levlCoeffBuf[i];
248 else if (level <= 24)
253 tmp *= imc_exp_tab[15 + level];
254 tmp2 += 0.83048 * level; // 0.83048 = log2(10) * 0.25
262 static void imc_decode_level_coefficients2(IMCContext* q, int* levlCoeffBuf, float* old_floor, float* flcoeffs1,
265 //FIXME maybe flag_buf = noise coding and flcoeffs1 = new scale factors
266 // and flcoeffs2 old scale factors
267 // might be incomplete due to a missing table that is in the binary code
268 for(i = 0; i < BANDS; i++) {
270 if(levlCoeffBuf[i] < 16) {
271 flcoeffs1[i] = imc_exp_tab2[levlCoeffBuf[i]] * old_floor[i];
272 flcoeffs2[i] = (levlCoeffBuf[i]-7) * 0.83048 + flcoeffs2[i]; // 0.83048 = log2(10) * 0.25
274 flcoeffs1[i] = old_floor[i];
280 * Perform bit allocation depending on bits available
282 static int bit_allocation (IMCContext* q, int stream_format_code, int freebits, int flag) {
284 const float limit = -1.e20;
293 float lowest = 1.e10;
299 for(i = 0; i < BANDS; i++)
300 highest = FFMAX(highest, q->flcoeffs1[i]);
302 for(i = 0; i < BANDS-1; i++) {
303 q->flcoeffs4[i] = q->flcoeffs3[i] - log(q->flcoeffs5[i])/log(2);
305 q->flcoeffs4[BANDS - 1] = limit;
307 highest = highest * 0.25;
309 for(i = 0; i < BANDS; i++) {
311 if ((band_tab[i+1] - band_tab[i]) == q->bandWidthT[i])
314 if ((band_tab[i+1] - band_tab[i]) > q->bandWidthT[i])
317 if (((band_tab[i+1] - band_tab[i])/2) >= q->bandWidthT[i])
323 q->flcoeffs4[i] = q->flcoeffs4[i] + xTab[(indx*2 + (q->flcoeffs1[i] < highest)) * 2 + flag];
326 if (stream_format_code & 0x2) {
327 q->flcoeffs4[0] = limit;
328 q->flcoeffs4[1] = limit;
329 q->flcoeffs4[2] = limit;
330 q->flcoeffs4[3] = limit;
333 for(i = (stream_format_code & 0x2)?4:0; i < BANDS-1; i++) {
334 iacc += q->bandWidthT[i];
335 summa += q->bandWidthT[i] * q->flcoeffs4[i];
337 q->bandWidthT[BANDS-1] = 0;
338 summa = (summa * 0.5 - freebits) / iacc;
341 for(i = 0; i < BANDS/2; i++) {
342 rres = summer - freebits;
343 if((rres >= -8) && (rres <= 8)) break;
348 for(j = (stream_format_code & 0x2)?4:0; j < BANDS; j++) {
349 cwlen = av_clip((int)((q->flcoeffs4[j] * 0.5) - summa + 0.5), 0, 6);
351 q->bitsBandT[j] = cwlen;
352 summer += q->bandWidthT[j] * cwlen;
355 iacc += q->bandWidthT[j];
360 if (freebits < summer)
367 summa = (float)(summer - freebits) / ((t1 + 1) * iacc) + summa;
370 for(i = (stream_format_code & 0x2)?4:0; i < BANDS; i++) {
371 for(j = band_tab[i]; j < band_tab[i+1]; j++)
372 q->CWlengthT[j] = q->bitsBandT[i];
375 if (freebits > summer) {
376 for(i = 0; i < BANDS; i++) {
377 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
383 if (highest <= -1.e20)
389 for(i = 0; i < BANDS; i++) {
390 if (workT[i] > highest) {
396 if (highest > -1.e20) {
397 workT[found_indx] -= 2.0;
398 if (++(q->bitsBandT[found_indx]) == 6)
399 workT[found_indx] = -1.e20;
401 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (freebits > summer); j++){
406 }while (freebits > summer);
408 if (freebits < summer) {
409 for(i = 0; i < BANDS; i++) {
410 workT[i] = q->bitsBandT[i] ? (q->bitsBandT[i] * -2 + q->flcoeffs4[i] + 1.585) : 1.e20;
412 if (stream_format_code & 0x2) {
418 while (freebits < summer){
421 for(i = 0; i < BANDS; i++) {
422 if (workT[i] < lowest) {
427 //if(lowest >= 1.e10) break;
428 workT[low_indx] = lowest + 2.0;
430 if (!(--q->bitsBandT[low_indx]))
431 workT[low_indx] = 1.e20;
433 for(j = band_tab[low_indx]; j < band_tab[low_indx+1] && (freebits < summer); j++){
434 if(q->CWlengthT[j] > 0){
444 static void imc_get_skip_coeff(IMCContext* q) {
447 memset(q->skipFlagBits, 0, sizeof(q->skipFlagBits));
448 memset(q->skipFlagCount, 0, sizeof(q->skipFlagCount));
449 for(i = 0; i < BANDS; i++) {
450 if (!q->bandFlagsBuf[i] || !q->bandWidthT[i])
453 if (!q->skipFlagRaw[i]) {
454 q->skipFlagBits[i] = band_tab[i+1] - band_tab[i];
456 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
457 if ((q->skipFlags[j] = get_bits1(&q->gb)))
458 q->skipFlagCount[i]++;
461 for(j = band_tab[i]; j < (band_tab[i+1]-1); j += 2) {
462 if(!get_bits1(&q->gb)){//0
463 q->skipFlagBits[i]++;
466 q->skipFlagCount[i] += 2;
468 if(get_bits1(&q->gb)){//11
469 q->skipFlagBits[i] +=2;
472 q->skipFlagCount[i]++;
474 q->skipFlagBits[i] +=3;
476 if(!get_bits1(&q->gb)){//100
478 q->skipFlagCount[i]++;
486 if (j < band_tab[i+1]) {
487 q->skipFlagBits[i]++;
488 if ((q->skipFlags[j] = get_bits1(&q->gb)))
489 q->skipFlagCount[i]++;
496 * Increase highest' band coefficient sizes as some bits won't be used
498 static void imc_adjust_bit_allocation (IMCContext* q, int summer) {
505 for(i = 0; i < BANDS; i++) {
506 workT[i] = (q->bitsBandT[i] == 6) ? -1.e20 : (q->bitsBandT[i] * -2 + q->flcoeffs4[i] - 0.415);
509 while (corrected < summer) {
510 if(highest <= -1.e20)
515 for(i = 0; i < BANDS; i++) {
516 if (workT[i] > highest) {
522 if (highest > -1.e20) {
523 workT[found_indx] -= 2.0;
524 if (++(q->bitsBandT[found_indx]) == 6)
525 workT[found_indx] = -1.e20;
527 for(j = band_tab[found_indx]; j < band_tab[found_indx+1] && (corrected < summer); j++) {
528 if (!q->skipFlags[j] && (q->CWlengthT[j] < 6)) {
537 static void imc_imdct256(IMCContext *q) {
542 for(i=0; i < COEFFS/2; i++){
543 q->samples[i].re = -(q->pre_coef1[i] * q->CWdecoded[COEFFS-1-i*2]) -
544 (q->pre_coef2[i] * q->CWdecoded[i*2]);
545 q->samples[i].im = (q->pre_coef2[i] * q->CWdecoded[COEFFS-1-i*2]) -
546 (q->pre_coef1[i] * q->CWdecoded[i*2]);
550 ff_fft_permute(&q->fft, q->samples);
551 ff_fft_calc (&q->fft, q->samples);
553 /* postrotation, window and reorder */
554 for(i = 0; i < COEFFS/2; i++){
555 re = (q->samples[i].re * q->post_cos[i]) + (-q->samples[i].im * q->post_sin[i]);
556 im = (-q->samples[i].im * q->post_cos[i]) - (q->samples[i].re * q->post_sin[i]);
557 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);
558 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);
559 q->last_fft_im[i] = im;
563 static int inverse_quant_coeff (IMCContext* q, int stream_format_code) {
565 int middle_value, cw_len, max_size;
566 const float* quantizer;
568 for(i = 0; i < BANDS; i++) {
569 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
571 cw_len = q->CWlengthT[j];
573 if (cw_len <= 0 || q->skipFlags[j])
576 max_size = 1 << cw_len;
577 middle_value = max_size >> 1;
579 if (q->codewords[j] >= max_size || q->codewords[j] < 0)
583 quantizer = imc_quantizer2[(stream_format_code & 2) >> 1];
584 if (q->codewords[j] >= middle_value)
585 q->CWdecoded[j] = quantizer[q->codewords[j] - 8] * q->flcoeffs6[i];
587 q->CWdecoded[j] = -quantizer[max_size - q->codewords[j] - 8 - 1] * q->flcoeffs6[i];
589 quantizer = imc_quantizer1[((stream_format_code & 2) >> 1) | (q->bandFlagsBuf[i] << 1)];
590 if (q->codewords[j] >= middle_value)
591 q->CWdecoded[j] = quantizer[q->codewords[j] - 1] * q->flcoeffs6[i];
593 q->CWdecoded[j] = -quantizer[max_size - 2 - q->codewords[j]] * q->flcoeffs6[i];
601 static int imc_get_coeffs (IMCContext* q) {
602 int i, j, cw_len, cw;
604 for(i = 0; i < BANDS; i++) {
605 if(!q->sumLenArr[i]) continue;
606 if (q->bandFlagsBuf[i] || q->bandWidthT[i]) {
607 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
608 cw_len = q->CWlengthT[j];
611 if (get_bits_count(&q->gb) + cw_len > 512){
612 //av_log(NULL,0,"Band %i coeff %i cw_len %i\n",i,j,cw_len);
616 if(cw_len && (!q->bandFlagsBuf[i] || !q->skipFlags[j]))
617 cw = get_bits(&q->gb, cw_len);
619 q->codewords[j] = cw;
626 static int imc_decode_frame(AVCodecContext * avctx,
627 void *data, int *data_size,
628 const uint8_t * buf, int buf_size)
631 IMCContext *q = avctx->priv_data;
633 int stream_format_code;
637 int counter, bitscount;
638 uint16_t *buf16 = (uint16_t *) buf;
640 /* FIXME: input should not be modified */
641 for(i = 0; i < FFMIN(buf_size, avctx->block_align) / 2; i++)
642 buf16[i] = bswap_16(buf16[i]);
644 init_get_bits(&q->gb, buf, 512);
646 /* Check the frame header */
647 imc_hdr = get_bits(&q->gb, 9);
648 if (imc_hdr != IMC_FRAME_ID) {
649 av_log(avctx, AV_LOG_ERROR, "imc frame header check failed!\n");
650 av_log(avctx, AV_LOG_ERROR, "got %x instead of 0x21.\n", imc_hdr);
653 stream_format_code = get_bits(&q->gb, 3);
655 if(stream_format_code & 1){
656 av_log(avctx, AV_LOG_ERROR, "Stream code format %X is not supported\n", stream_format_code);
660 // av_log(avctx, AV_LOG_DEBUG, "stream_format_code = %d\n", stream_format_code);
662 if (stream_format_code & 0x04)
663 q->decoder_reset = 1;
665 if(q->decoder_reset) {
666 memset(q->out_samples, 0, sizeof(q->out_samples));
667 for(i = 0; i < BANDS; i++)q->old_floor[i] = 1.0;
668 for(i = 0; i < COEFFS; i++)q->CWdecoded[i] = 0;
669 q->decoder_reset = 0;
672 flag = get_bits1(&q->gb);
673 imc_read_level_coeffs(q, stream_format_code, q->levlCoeffBuf);
675 if (stream_format_code & 0x4)
676 imc_decode_level_coefficients(q, q->levlCoeffBuf, q->flcoeffs1, q->flcoeffs2);
678 imc_decode_level_coefficients2(q, q->levlCoeffBuf, q->old_floor, q->flcoeffs1, q->flcoeffs2);
680 memcpy(q->old_floor, q->flcoeffs1, 32 * sizeof(float));
683 for (i=0 ; i<BANDS ; i++) {
684 if (q->levlCoeffBuf[i] == 16) {
685 q->bandWidthT[i] = 0;
688 q->bandWidthT[i] = band_tab[i+1] - band_tab[i];
690 memset(q->bandFlagsBuf, 0, BANDS * sizeof(int));
691 for(i = 0; i < BANDS-1; i++) {
692 if (q->bandWidthT[i])
693 q->bandFlagsBuf[i] = get_bits1(&q->gb);
696 imc_calculate_coeffs(q, q->flcoeffs1, q->flcoeffs2, q->bandWidthT, q->flcoeffs3, q->flcoeffs5);
699 /* first 4 bands will be assigned 5 bits per coefficient */
700 if (stream_format_code & 0x2) {
707 for(i = 1; i < 4; i++){
708 bits = (q->levlCoeffBuf[i] == 16) ? 0 : 5;
709 q->bitsBandT[i] = bits;
710 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
711 q->CWlengthT[j] = bits;
717 if(bit_allocation (q, stream_format_code, 512 - bitscount - get_bits_count(&q->gb), flag) < 0) {
718 av_log(avctx, AV_LOG_ERROR, "Bit allocations failed\n");
719 q->decoder_reset = 1;
723 for(i = 0; i < BANDS; i++) {
725 q->skipFlagRaw[i] = 0;
726 for(j = band_tab[i]; j < band_tab[i+1]; j++)
727 q->sumLenArr[i] += q->CWlengthT[j];
728 if (q->bandFlagsBuf[i])
729 if( (((band_tab[i+1] - band_tab[i]) * 1.5) > q->sumLenArr[i]) && (q->sumLenArr[i] > 0))
730 q->skipFlagRaw[i] = 1;
733 imc_get_skip_coeff(q);
735 for(i = 0; i < BANDS; i++) {
736 q->flcoeffs6[i] = q->flcoeffs1[i];
737 /* band has flag set and at least one coded coefficient */
738 if (q->bandFlagsBuf[i] && (band_tab[i+1] - band_tab[i]) != q->skipFlagCount[i]){
739 q->flcoeffs6[i] *= q->sqrt_tab[band_tab[i+1] - band_tab[i]] /
740 q->sqrt_tab[(band_tab[i+1] - band_tab[i] - q->skipFlagCount[i])];
744 /* calculate bits left, bits needed and adjust bit allocation */
747 for(i = 0; i < BANDS; i++) {
748 if (q->bandFlagsBuf[i]) {
749 for(j = band_tab[i]; j < band_tab[i+1]; j++) {
750 if(q->skipFlags[j]) {
751 summer += q->CWlengthT[j];
755 bits += q->skipFlagBits[i];
756 summer -= q->skipFlagBits[i];
759 imc_adjust_bit_allocation(q, summer);
761 for(i = 0; i < BANDS; i++) {
764 for(j = band_tab[i]; j < band_tab[i+1]; j++)
765 if (!q->skipFlags[j])
766 q->sumLenArr[i] += q->CWlengthT[j];
769 memset(q->codewords, 0, sizeof(q->codewords));
771 if(imc_get_coeffs(q) < 0) {
772 av_log(avctx, AV_LOG_ERROR, "Read coefficients failed\n");
773 q->decoder_reset = 1;
777 if(inverse_quant_coeff(q, stream_format_code) < 0) {
778 av_log(avctx, AV_LOG_ERROR, "Inverse quantization of coefficients failed\n");
779 q->decoder_reset = 1;
783 memset(q->skipFlags, 0, sizeof(q->skipFlags));
787 q->dsp.float_to_int16(data, q->out_samples, COEFFS);
789 *data_size = COEFFS * sizeof(int16_t);
791 return avctx->block_align;
795 static int imc_decode_close(AVCodecContext * avctx)
797 IMCContext *q = avctx->priv_data;
804 AVCodec imc_decoder = {
806 .type = CODEC_TYPE_AUDIO,
808 .priv_data_size = sizeof(IMCContext),
809 .init = imc_decode_init,
810 .close = imc_decode_close,
811 .decode = imc_decode_frame,