3 * Copyright (c) 2001, 2002 Fabrice Bellard.
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 #include "mpegaudio.h"
25 * - in low precision mode, use more 16 bit multiplies in synth filter
26 * - test lsf / mpeg25 extensively.
29 /* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
31 #ifdef CONFIG_MPEGAUDIO_HP
32 #define USE_HIGHPRECISION
35 #ifdef USE_HIGHPRECISION
36 #define FRAC_BITS 23 /* fractional bits for sb_samples and dct */
37 #define WFRAC_BITS 16 /* fractional bits for window */
39 #define FRAC_BITS 15 /* fractional bits for sb_samples and dct */
40 #define WFRAC_BITS 14 /* fractional bits for window */
43 #define FRAC_ONE (1 << FRAC_BITS)
45 #define MULL(a,b) (((INT64)(a) * (INT64)(b)) >> FRAC_BITS)
46 #define MUL64(a,b) ((INT64)(a) * (INT64)(b))
47 #define FIX(a) ((int)((a) * FRAC_ONE))
48 /* WARNING: only correct for posititive numbers */
49 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
50 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
53 typedef INT16 MPA_INT;
55 typedef INT32 MPA_INT;
61 #define BACKSTEP_SIZE 512
63 typedef struct MPADecodeContext {
64 UINT8 inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
66 UINT8 *inbuf_ptr, *inbuf;
68 int free_format_frame_size; /* frame size in case of free format
69 (zero if currently unknown) */
70 /* next header (used in free format parsing) */
71 UINT32 free_format_next_header;
75 int sample_rate_index; /* between 0 and 8 */
83 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2];
84 int synth_buf_offset[MPA_MAX_CHANNELS];
85 INT32 sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT];
86 INT32 mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
92 /* layer 3 "granule" */
93 typedef struct GranuleDef {
98 int scalefac_compress;
102 int subblock_gain[3];
103 UINT8 scalefac_scale;
104 UINT8 count1table_select;
105 int region_size[3]; /* number of huffman codes in each region */
107 int short_start, long_end; /* long/short band indexes */
108 UINT8 scale_factors[40];
109 INT32 sb_hybrid[SBLIMIT * 18]; /* 576 samples */
112 #define MODE_EXT_MS_STEREO 2
113 #define MODE_EXT_I_STEREO 1
115 /* layer 3 huffman tables */
116 typedef struct HuffTable {
122 #include "mpegaudiodectab.h"
124 /* vlc structure for decoding layer 3 huffman tables */
125 static VLC huff_vlc[16];
126 static UINT8 *huff_code_table[16];
127 static VLC huff_quad_vlc[2];
128 /* computed from band_size_long */
129 static UINT16 band_index_long[9][23];
130 /* XXX: free when all decoders are closed */
131 #define TABLE_4_3_SIZE (8191 + 16)
132 static INT8 *table_4_3_exp;
134 static UINT16 *table_4_3_value;
136 static UINT32 *table_4_3_value;
138 /* intensity stereo coef table */
139 static INT32 is_table[2][16];
140 static INT32 is_table_lsf[2][2][16];
141 static INT32 csa_table[8][2];
142 static INT32 mdct_win[8][36];
144 /* lower 2 bits: modulo 3, higher bits: shift */
145 static UINT16 scale_factor_modshift[64];
146 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
147 static INT32 scale_factor_mult[15][3];
148 /* mult table for layer 2 group quantization */
150 #define SCALE_GEN(v) \
151 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
153 static INT32 scale_factor_mult2[3][3] = {
154 SCALE_GEN(4.0 / 3.0), /* 3 steps */
155 SCALE_GEN(4.0 / 5.0), /* 5 steps */
156 SCALE_GEN(4.0 / 9.0), /* 9 steps */
160 static UINT32 scale_factor_mult3[4] = {
162 FIXR(1.18920711500272106671),
163 FIXR(1.41421356237309504880),
164 FIXR(1.68179283050742908605),
167 static MPA_INT window[512];
169 /* layer 1 unscaling */
170 /* n = number of bits of the mantissa minus 1 */
171 static inline int l1_unscale(int n, int mant, int scale_factor)
176 shift = scale_factor_modshift[scale_factor];
179 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
181 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
182 return (int)((val + (1LL << (shift - 1))) >> shift);
185 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
189 shift = scale_factor_modshift[scale_factor];
193 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
194 /* NOTE: at this point, 0 <= shift <= 21 */
196 val = (val + (1 << (shift - 1))) >> shift;
200 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
201 static inline int l3_unscale(int value, int exponent)
210 e = table_4_3_exp[value];
211 e += (exponent >> 2);
217 m = table_4_3_value[value];
219 m = (m * scale_factor_mult3[exponent & 3]);
220 m = (m + (1 << (e-1))) >> e;
223 m = MUL64(m, scale_factor_mult3[exponent & 3]);
224 m = (m + (UINT64_C(1) << (e-1))) >> e;
229 /* all integer n^(4/3) computation code */
232 #define POW_FRAC_BITS 24
233 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
234 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
235 #define POW_MULL(a,b) (((INT64)(a) * (INT64)(b)) >> POW_FRAC_BITS)
237 static int dev_4_3_coefs[DEV_ORDER];
239 static int pow_mult3[3] = {
241 POW_FIX(1.25992104989487316476),
242 POW_FIX(1.58740105196819947474),
245 static void int_pow_init(void)
250 for(i=0;i<DEV_ORDER;i++) {
251 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
252 dev_4_3_coefs[i] = a;
256 /* return the mantissa and the binary exponent */
257 static int int_pow(int i, int *exp_ptr)
265 while (a < (1 << (POW_FRAC_BITS - 1))) {
269 a -= (1 << POW_FRAC_BITS);
271 for(j = DEV_ORDER - 1; j >= 0; j--)
272 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
273 a = (1 << POW_FRAC_BITS) + a1;
274 /* exponent compute (exact) */
278 a = POW_MULL(a, pow_mult3[er]);
279 while (a >= 2 * POW_FRAC_ONE) {
283 /* convert to float */
284 while (a < POW_FRAC_ONE) {
288 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
289 #if POW_FRAC_BITS > FRAC_BITS
290 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
291 /* correct overflow */
292 if (a >= 2 * (1 << FRAC_BITS)) {
301 static int decode_init(AVCodecContext * avctx)
303 MPADecodeContext *s = avctx->priv_data;
308 /* scale factors table for layer 1/2 */
311 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
314 scale_factor_modshift[i] = mod | (shift << 2);
317 /* scale factor multiply for layer 1 */
321 norm = ((INT64_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
322 scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
323 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
324 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
325 dprintf("%d: norm=%x s=%x %x %x\n",
327 scale_factor_mult[i][0],
328 scale_factor_mult[i][1],
329 scale_factor_mult[i][2]);
333 /* max = 18760, max sum over all 16 coefs : 44736 */
338 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
347 /* huffman decode tables */
348 huff_code_table[0] = NULL;
350 const HuffTable *h = &mpa_huff_tables[i];
358 init_vlc(&huff_vlc[i], 8, n,
359 h->bits, 1, 1, h->codes, 2, 2);
361 code_table = av_mallocz(n);
363 for(x=0;x<xsize;x++) {
365 code_table[j++] = (x << 4) | y;
367 huff_code_table[i] = code_table;
370 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
371 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1);
377 band_index_long[i][j] = k;
378 k += band_size_long[i][j];
380 band_index_long[i][22] = k;
383 /* compute n ^ (4/3) and store it in mantissa/exp format */
384 if (!av_mallocz_static(&table_4_3_exp,
385 TABLE_4_3_SIZE * sizeof(table_4_3_exp[0])))
387 if (!av_mallocz_static(&table_4_3_value,
388 TABLE_4_3_SIZE * sizeof(table_4_3_value[0])))
392 for(i=1;i<TABLE_4_3_SIZE;i++) {
400 f = pow((double)i, 4.0 / 3.0);
404 if ((unsigned short)m1 != m1) {
410 if (m != m1 || e != e1) {
411 printf("%4d: m=%x m1=%x e=%d e1=%d\n",
416 /* normalized to FRAC_BITS */
417 table_4_3_value[i] = m;
418 table_4_3_exp[i] = e;
425 f = tan((double)i * M_PI / 12.0);
426 v = FIXR(f / (1.0 + f));
431 is_table[1][6 - i] = v;
435 is_table[0][i] = is_table[1][i] = 0.0;
442 e = -(j + 1) * ((i + 1) >> 1);
443 f = pow(2.0, e / 4.0);
445 is_table_lsf[j][k ^ 1][i] = FIXR(f);
446 is_table_lsf[j][k][i] = FIXR(1.0);
447 dprintf("is_table_lsf %d %d: %x %x\n",
448 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
455 cs = 1.0 / sqrt(1.0 + ci * ci);
457 csa_table[i][0] = FIX(cs);
458 csa_table[i][1] = FIX(ca);
461 /* compute mdct windows */
464 v = FIXR(sin(M_PI * (i + 0.5) / 36.0));
470 mdct_win[1][18 + i] = FIXR(1.0);
471 mdct_win[1][24 + i] = FIXR(sin(M_PI * ((i + 6) + 0.5) / 12.0));
472 mdct_win[1][30 + i] = FIXR(0.0);
474 mdct_win[3][i] = FIXR(0.0);
475 mdct_win[3][6 + i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
476 mdct_win[3][12 + i] = FIXR(1.0);
480 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
482 /* NOTE: we do frequency inversion adter the MDCT by changing
483 the sign of the right window coefs */
486 mdct_win[j + 4][i] = mdct_win[j][i];
487 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
493 printf("win%d=\n", j);
495 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
503 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
504 s->inbuf_ptr = s->inbuf;
511 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
515 #define COS0_0 FIXR(0.50060299823519630134)
516 #define COS0_1 FIXR(0.50547095989754365998)
517 #define COS0_2 FIXR(0.51544730992262454697)
518 #define COS0_3 FIXR(0.53104259108978417447)
519 #define COS0_4 FIXR(0.55310389603444452782)
520 #define COS0_5 FIXR(0.58293496820613387367)
521 #define COS0_6 FIXR(0.62250412303566481615)
522 #define COS0_7 FIXR(0.67480834145500574602)
523 #define COS0_8 FIXR(0.74453627100229844977)
524 #define COS0_9 FIXR(0.83934964541552703873)
525 #define COS0_10 FIXR(0.97256823786196069369)
526 #define COS0_11 FIXR(1.16943993343288495515)
527 #define COS0_12 FIXR(1.48416461631416627724)
528 #define COS0_13 FIXR(2.05778100995341155085)
529 #define COS0_14 FIXR(3.40760841846871878570)
530 #define COS0_15 FIXR(10.19000812354805681150)
532 #define COS1_0 FIXR(0.50241928618815570551)
533 #define COS1_1 FIXR(0.52249861493968888062)
534 #define COS1_2 FIXR(0.56694403481635770368)
535 #define COS1_3 FIXR(0.64682178335999012954)
536 #define COS1_4 FIXR(0.78815462345125022473)
537 #define COS1_5 FIXR(1.06067768599034747134)
538 #define COS1_6 FIXR(1.72244709823833392782)
539 #define COS1_7 FIXR(5.10114861868916385802)
541 #define COS2_0 FIXR(0.50979557910415916894)
542 #define COS2_1 FIXR(0.60134488693504528054)
543 #define COS2_2 FIXR(0.89997622313641570463)
544 #define COS2_3 FIXR(2.56291544774150617881)
546 #define COS3_0 FIXR(0.54119610014619698439)
547 #define COS3_1 FIXR(1.30656296487637652785)
549 #define COS4_0 FIXR(0.70710678118654752439)
551 /* butterfly operator */
554 tmp0 = tab[a] + tab[b];\
555 tmp1 = tab[a] - tab[b];\
557 tab[b] = MULL(tmp1, c);\
560 #define BF1(a, b, c, d)\
567 #define BF2(a, b, c, d)\
577 #define ADD(a, b) tab[a] += tab[b]
579 /* DCT32 without 1/sqrt(2) coef zero scaling. */
580 static void dct32(INT32 *out, INT32 *tab)
712 out[ 1] = tab[16] + tab[24];
713 out[17] = tab[17] + tab[25];
714 out[ 9] = tab[18] + tab[26];
715 out[25] = tab[19] + tab[27];
716 out[ 5] = tab[20] + tab[28];
717 out[21] = tab[21] + tab[29];
718 out[13] = tab[22] + tab[30];
719 out[29] = tab[23] + tab[31];
720 out[ 3] = tab[24] + tab[20];
721 out[19] = tab[25] + tab[21];
722 out[11] = tab[26] + tab[22];
723 out[27] = tab[27] + tab[23];
724 out[ 7] = tab[28] + tab[18];
725 out[23] = tab[29] + tab[19];
726 out[15] = tab[30] + tab[17];
730 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
734 #define OUT_SAMPLE(sum)\
737 sum1 = (sum + (1 << (OUT_SHIFT - 1))) >> OUT_SHIFT;\
740 else if (sum1 > 32767)\
746 #define SUM8(off, op) \
748 sum op w[0 * 64 + off] * p[0 * 64];\
749 sum op w[1 * 64 + off] * p[1 * 64];\
750 sum op w[2 * 64 + off] * p[2 * 64];\
751 sum op w[3 * 64 + off] * p[3 * 64];\
752 sum op w[4 * 64 + off] * p[4 * 64];\
753 sum op w[5 * 64 + off] * p[5 * 64];\
754 sum op w[6 * 64 + off] * p[6 * 64];\
755 sum op w[7 * 64 + off] * p[7 * 64];\
760 #define OUT_SAMPLE(sum)\
763 sum1 = (int)((sum + (INT64_C(1) << (OUT_SHIFT - 1))) >> OUT_SHIFT);\
766 else if (sum1 > 32767)\
772 #define SUM8(off, op) \
774 sum op MUL64(w[0 * 64 + off], p[0 * 64]);\
775 sum op MUL64(w[1 * 64 + off], p[1 * 64]);\
776 sum op MUL64(w[2 * 64 + off], p[2 * 64]);\
777 sum op MUL64(w[3 * 64 + off], p[3 * 64]);\
778 sum op MUL64(w[4 * 64 + off], p[4 * 64]);\
779 sum op MUL64(w[5 * 64 + off], p[5 * 64]);\
780 sum op MUL64(w[6 * 64 + off], p[6 * 64]);\
781 sum op MUL64(w[7 * 64 + off], p[7 * 64]);\
786 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
788 /* XXX: optimize by avoiding ring buffer usage */
789 static void synth_filter(MPADecodeContext *s1,
790 int ch, INT16 *samples, int incr,
791 INT32 sb_samples[SBLIMIT])
794 register MPA_INT *synth_buf, *p;
803 dct32(tmp, sb_samples);
805 offset = s1->synth_buf_offset[ch];
806 synth_buf = s1->synth_buf[ch] + offset;
811 /* NOTE: can cause a loss in precision if very high amplitude
820 /* copy to avoid wrap */
821 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
826 p = synth_buf + 16 + j; /* 0-15 */
828 p = synth_buf + 48 - j; /* 32-47 */
834 p = synth_buf + 32; /* 48 */
842 p = synth_buf + 48 - j; /* 17-31 */
844 p = synth_buf + 16 + j; /* 49-63 */
849 offset = (offset - 32) & 511;
850 s1->synth_buf_offset[ch] = offset;
854 #define C1 FIXR(0.99144486137381041114)
855 #define C3 FIXR(0.92387953251128675612)
856 #define C5 FIXR(0.79335334029123516458)
857 #define C7 FIXR(0.60876142900872063941)
858 #define C9 FIXR(0.38268343236508977173)
859 #define C11 FIXR(0.13052619222005159154)
861 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
863 static void imdct12(int *out, int *in)
866 INT64 in1_3, in1_9, in4_3, in4_9;
868 in1_3 = MUL64(in[1], C3);
869 in1_9 = MUL64(in[1], C9);
870 in4_3 = MUL64(in[4], C3);
871 in4_9 = MUL64(in[4], C9);
873 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
874 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
877 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
878 MUL64(in[2] + in[5], C3) - in4_9);
881 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
882 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
885 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
886 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
889 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
890 MUL64(in[2] + in[5], C9) + in4_3);
893 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
894 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
907 #define C1 FIXR(0.98480775301220805936)
908 #define C2 FIXR(0.93969262078590838405)
909 #define C3 FIXR(0.86602540378443864676)
910 #define C4 FIXR(0.76604444311897803520)
911 #define C5 FIXR(0.64278760968653932632)
913 #define C7 FIXR(0.34202014332566873304)
914 #define C8 FIXR(0.17364817766693034885)
916 /* 0.5 / cos(pi*(2*i+1)/36) */
917 static const int icos36[9] = {
918 FIXR(0.50190991877167369479),
919 FIXR(0.51763809020504152469),
920 FIXR(0.55168895948124587824),
921 FIXR(0.61038729438072803416),
922 FIXR(0.70710678118654752439),
923 FIXR(0.87172339781054900991),
924 FIXR(1.18310079157624925896),
925 FIXR(1.93185165257813657349),
926 FIXR(5.73685662283492756461),
929 static const int icos72[18] = {
930 /* 0.5 / cos(pi*(2*i+19)/72) */
931 FIXR(0.74009361646113053152),
932 FIXR(0.82133981585229078570),
933 FIXR(0.93057949835178895673),
934 FIXR(1.08284028510010010928),
935 FIXR(1.30656296487637652785),
936 FIXR(1.66275476171152078719),
937 FIXR(2.31011315767264929558),
938 FIXR(3.83064878777019433457),
939 FIXR(11.46279281302667383546),
941 /* 0.5 / cos(pi*(2*(i + 18) +19)/72) */
942 FIXR(-0.67817085245462840086),
943 FIXR(-0.63023620700513223342),
944 FIXR(-0.59284452371708034528),
945 FIXR(-0.56369097343317117734),
946 FIXR(-0.54119610014619698439),
947 FIXR(-0.52426456257040533932),
948 FIXR(-0.51213975715725461845),
949 FIXR(-0.50431448029007636036),
950 FIXR(-0.50047634258165998492),
953 /* using Lee like decomposition followed by hand coded 9 points DCT */
954 static void imdct36(int *out, int *in)
956 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
957 int tmp[18], *tmp1, *in1;
969 in3_3 = MUL64(in1[2*3], C3);
970 in6_6 = MUL64(in1[2*6], C6);
972 tmp1[0] = FRAC_RND(MUL64(in1[2*1], C1) + in3_3 +
973 MUL64(in1[2*5], C5) + MUL64(in1[2*7], C7));
974 tmp1[2] = in1[2*0] + FRAC_RND(MUL64(in1[2*2], C2) +
975 MUL64(in1[2*4], C4) + in6_6 +
976 MUL64(in1[2*8], C8));
977 tmp1[4] = FRAC_RND(MUL64(in1[2*1] - in1[2*5] - in1[2*7], C3));
978 tmp1[6] = FRAC_RND(MUL64(in1[2*2] - in1[2*4] - in1[2*8], C6)) -
980 tmp1[8] = FRAC_RND(MUL64(in1[2*1], C5) - in3_3 -
981 MUL64(in1[2*5], C7) + MUL64(in1[2*7], C1));
982 tmp1[10] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C8) -
983 MUL64(in1[2*4], C2) + in6_6 +
984 MUL64(in1[2*8], C4));
985 tmp1[12] = FRAC_RND(MUL64(in1[2*1], C7) - in3_3 +
986 MUL64(in1[2*5], C1) -
987 MUL64(in1[2*7], C5));
988 tmp1[14] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C4) +
989 MUL64(in1[2*4], C8) + in6_6 -
990 MUL64(in1[2*8], C2));
991 tmp1[16] = in1[2*0] - in1[2*2] + in1[2*4] - in1[2*6] + in1[2*8];
1003 s1 = MULL(t3 + t2, icos36[j]);
1004 s3 = MULL(t3 - t2, icos36[8 - j]);
1006 t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
1007 t1 = MULL(s0 - s1, icos72[8 - j]);
1008 out[18 + 9 + j] = t0;
1009 out[18 + 8 - j] = t0;
1013 t0 = MULL(s2 + s3, icos72[9+j]);
1014 t1 = MULL(s2 - s3, icos72[j]);
1015 out[18 + 9 + (8 - j)] = t0;
1017 out[9 + (8 - j)] = -t1;
1023 s1 = MULL(tmp[17], icos36[4]);
1024 t0 = MULL(s0 + s1, icos72[9 + 4]);
1025 t1 = MULL(s0 - s1, icos72[4]);
1026 out[18 + 9 + 4] = t0;
1027 out[18 + 8 - 4] = t0;
1032 /* fast header check for resync */
1033 static int check_header(UINT32 header)
1036 if ((header & 0xffe00000) != 0xffe00000)
1039 if (((header >> 17) & 3) == 0)
1042 if (((header >> 12) & 0xf) == 0xf)
1045 if (((header >> 10) & 3) == 3)
1050 /* header + layer + bitrate + freq + lsf/mpeg25 */
1051 #define SAME_HEADER_MASK \
1052 (0xffe00000 | (3 << 17) | (0xf << 12) | (3 << 10) | (3 << 19))
1054 /* header decoding. MUST check the header before because no
1055 consistency check is done there. Return 1 if free format found and
1056 that the frame size must be computed externally */
1057 static int decode_header(MPADecodeContext *s, UINT32 header)
1059 int sample_rate, frame_size, mpeg25, padding;
1060 int sample_rate_index, bitrate_index;
1061 if (header & (1<<20)) {
1062 s->lsf = (header & (1<<19)) ? 0 : 1;
1069 s->layer = 4 - ((header >> 17) & 3);
1070 /* extract frequency */
1071 sample_rate_index = (header >> 10) & 3;
1072 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1073 sample_rate_index += 3 * (s->lsf + mpeg25);
1074 s->sample_rate_index = sample_rate_index;
1075 s->error_protection = ((header >> 16) & 1) ^ 1;
1076 s->sample_rate = sample_rate;
1078 bitrate_index = (header >> 12) & 0xf;
1079 padding = (header >> 9) & 1;
1080 //extension = (header >> 8) & 1;
1081 s->mode = (header >> 6) & 3;
1082 s->mode_ext = (header >> 4) & 3;
1083 //copyright = (header >> 3) & 1;
1084 //original = (header >> 2) & 1;
1085 //emphasis = header & 3;
1087 if (s->mode == MPA_MONO)
1092 if (bitrate_index != 0) {
1093 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1094 s->bit_rate = frame_size * 1000;
1097 frame_size = (frame_size * 12000) / sample_rate;
1098 frame_size = (frame_size + padding) * 4;
1101 frame_size = (frame_size * 144000) / sample_rate;
1102 frame_size += padding;
1106 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1107 frame_size += padding;
1110 s->frame_size = frame_size;
1112 /* if no frame size computed, signal it */
1113 if (!s->free_format_frame_size)
1115 /* free format: compute bitrate and real frame size from the
1116 frame size we extracted by reading the bitstream */
1117 s->frame_size = s->free_format_frame_size;
1120 s->frame_size += padding * 4;
1121 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1124 s->frame_size += padding;
1125 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1129 s->frame_size += padding;
1130 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1136 printf("layer%d, %d Hz, %d kbits/s, ",
1137 s->layer, s->sample_rate, s->bit_rate);
1138 if (s->nb_channels == 2) {
1139 if (s->layer == 3) {
1140 if (s->mode_ext & MODE_EXT_MS_STEREO)
1142 if (s->mode_ext & MODE_EXT_I_STEREO)
1154 /* return the number of decoded frames */
1155 static int mp_decode_layer1(MPADecodeContext *s)
1157 int bound, i, v, n, ch, j, mant;
1158 UINT8 allocation[MPA_MAX_CHANNELS][SBLIMIT];
1159 UINT8 scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1161 if (s->mode == MPA_JSTEREO)
1162 bound = (s->mode_ext + 1) * 4;
1166 /* allocation bits */
1167 for(i=0;i<bound;i++) {
1168 for(ch=0;ch<s->nb_channels;ch++) {
1169 allocation[ch][i] = get_bits(&s->gb, 4);
1172 for(i=bound;i<SBLIMIT;i++) {
1173 allocation[0][i] = get_bits(&s->gb, 4);
1177 for(i=0;i<bound;i++) {
1178 for(ch=0;ch<s->nb_channels;ch++) {
1179 if (allocation[ch][i])
1180 scale_factors[ch][i] = get_bits(&s->gb, 6);
1183 for(i=bound;i<SBLIMIT;i++) {
1184 if (allocation[0][i]) {
1185 scale_factors[0][i] = get_bits(&s->gb, 6);
1186 scale_factors[1][i] = get_bits(&s->gb, 6);
1190 /* compute samples */
1192 for(i=0;i<bound;i++) {
1193 for(ch=0;ch<s->nb_channels;ch++) {
1194 n = allocation[ch][i];
1196 mant = get_bits(&s->gb, n + 1);
1197 v = l1_unscale(n, mant, scale_factors[ch][i]);
1201 s->sb_samples[ch][j][i] = v;
1204 for(i=bound;i<SBLIMIT;i++) {
1205 n = allocation[0][i];
1207 mant = get_bits(&s->gb, n + 1);
1208 v = l1_unscale(n, mant, scale_factors[0][i]);
1209 s->sb_samples[0][j][i] = v;
1210 v = l1_unscale(n, mant, scale_factors[1][i]);
1211 s->sb_samples[1][j][i] = v;
1213 s->sb_samples[0][j][i] = 0;
1214 s->sb_samples[1][j][i] = 0;
1221 /* bitrate is in kb/s */
1222 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1224 int ch_bitrate, table;
1226 ch_bitrate = bitrate / nb_channels;
1228 if ((freq == 48000 && ch_bitrate >= 56) ||
1229 (ch_bitrate >= 56 && ch_bitrate <= 80))
1231 else if (freq != 48000 && ch_bitrate >= 96)
1233 else if (freq != 32000 && ch_bitrate <= 48)
1243 static int mp_decode_layer2(MPADecodeContext *s)
1245 int sblimit; /* number of used subbands */
1246 const unsigned char *alloc_table;
1247 int table, bit_alloc_bits, i, j, ch, bound, v;
1248 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1249 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1250 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1251 int scale, qindex, bits, steps, k, l, m, b;
1253 /* select decoding table */
1254 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1255 s->sample_rate, s->lsf);
1256 sblimit = sblimit_table[table];
1257 alloc_table = alloc_tables[table];
1259 if (s->mode == MPA_JSTEREO)
1260 bound = (s->mode_ext + 1) * 4;
1264 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1265 /* parse bit allocation */
1267 for(i=0;i<bound;i++) {
1268 bit_alloc_bits = alloc_table[j];
1269 for(ch=0;ch<s->nb_channels;ch++) {
1270 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1272 j += 1 << bit_alloc_bits;
1274 for(i=bound;i<sblimit;i++) {
1275 bit_alloc_bits = alloc_table[j];
1276 v = get_bits(&s->gb, bit_alloc_bits);
1277 bit_alloc[0][i] = v;
1278 bit_alloc[1][i] = v;
1279 j += 1 << bit_alloc_bits;
1284 for(ch=0;ch<s->nb_channels;ch++) {
1285 for(i=0;i<sblimit;i++)
1286 printf(" %d", bit_alloc[ch][i]);
1293 for(i=0;i<sblimit;i++) {
1294 for(ch=0;ch<s->nb_channels;ch++) {
1295 if (bit_alloc[ch][i])
1296 scale_code[ch][i] = get_bits(&s->gb, 2);
1301 for(i=0;i<sblimit;i++) {
1302 for(ch=0;ch<s->nb_channels;ch++) {
1303 if (bit_alloc[ch][i]) {
1304 sf = scale_factors[ch][i];
1305 switch(scale_code[ch][i]) {
1308 sf[0] = get_bits(&s->gb, 6);
1309 sf[1] = get_bits(&s->gb, 6);
1310 sf[2] = get_bits(&s->gb, 6);
1313 sf[0] = get_bits(&s->gb, 6);
1318 sf[0] = get_bits(&s->gb, 6);
1319 sf[2] = get_bits(&s->gb, 6);
1323 sf[0] = get_bits(&s->gb, 6);
1324 sf[2] = get_bits(&s->gb, 6);
1333 for(ch=0;ch<s->nb_channels;ch++) {
1334 for(i=0;i<sblimit;i++) {
1335 if (bit_alloc[ch][i]) {
1336 sf = scale_factors[ch][i];
1337 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1348 for(l=0;l<12;l+=3) {
1350 for(i=0;i<bound;i++) {
1351 bit_alloc_bits = alloc_table[j];
1352 for(ch=0;ch<s->nb_channels;ch++) {
1353 b = bit_alloc[ch][i];
1355 scale = scale_factors[ch][i][k];
1356 qindex = alloc_table[j+b];
1357 bits = quant_bits[qindex];
1359 /* 3 values at the same time */
1360 v = get_bits(&s->gb, -bits);
1361 steps = quant_steps[qindex];
1362 s->sb_samples[ch][k * 12 + l + 0][i] =
1363 l2_unscale_group(steps, v % steps, scale);
1365 s->sb_samples[ch][k * 12 + l + 1][i] =
1366 l2_unscale_group(steps, v % steps, scale);
1368 s->sb_samples[ch][k * 12 + l + 2][i] =
1369 l2_unscale_group(steps, v, scale);
1372 v = get_bits(&s->gb, bits);
1373 v = l1_unscale(bits - 1, v, scale);
1374 s->sb_samples[ch][k * 12 + l + m][i] = v;
1378 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1379 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1380 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1383 /* next subband in alloc table */
1384 j += 1 << bit_alloc_bits;
1386 /* XXX: find a way to avoid this duplication of code */
1387 for(i=bound;i<sblimit;i++) {
1388 bit_alloc_bits = alloc_table[j];
1389 b = bit_alloc[0][i];
1391 int mant, scale0, scale1;
1392 scale0 = scale_factors[0][i][k];
1393 scale1 = scale_factors[1][i][k];
1394 qindex = alloc_table[j+b];
1395 bits = quant_bits[qindex];
1397 /* 3 values at the same time */
1398 v = get_bits(&s->gb, -bits);
1399 steps = quant_steps[qindex];
1402 s->sb_samples[0][k * 12 + l + 0][i] =
1403 l2_unscale_group(steps, mant, scale0);
1404 s->sb_samples[1][k * 12 + l + 0][i] =
1405 l2_unscale_group(steps, mant, scale1);
1408 s->sb_samples[0][k * 12 + l + 1][i] =
1409 l2_unscale_group(steps, mant, scale0);
1410 s->sb_samples[1][k * 12 + l + 1][i] =
1411 l2_unscale_group(steps, mant, scale1);
1412 s->sb_samples[0][k * 12 + l + 2][i] =
1413 l2_unscale_group(steps, v, scale0);
1414 s->sb_samples[1][k * 12 + l + 2][i] =
1415 l2_unscale_group(steps, v, scale1);
1418 mant = get_bits(&s->gb, bits);
1419 s->sb_samples[0][k * 12 + l + m][i] =
1420 l1_unscale(bits - 1, mant, scale0);
1421 s->sb_samples[1][k * 12 + l + m][i] =
1422 l1_unscale(bits - 1, mant, scale1);
1426 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1427 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1428 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1429 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1430 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1431 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1433 /* next subband in alloc table */
1434 j += 1 << bit_alloc_bits;
1436 /* fill remaining samples to zero */
1437 for(i=sblimit;i<SBLIMIT;i++) {
1438 for(ch=0;ch<s->nb_channels;ch++) {
1439 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1440 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1441 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1450 * Seek back in the stream for backstep bytes (at most 511 bytes)
1452 static void seek_to_maindata(MPADecodeContext *s, unsigned int backstep)
1456 /* compute current position in stream */
1457 ptr = s->gb.buffer + (get_bits_count(&s->gb)>>3);
1459 /* copy old data before current one */
1461 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1462 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1463 /* init get bits again */
1464 init_get_bits(&s->gb, ptr, (s->frame_size + backstep)*8);
1466 /* prepare next buffer */
1467 s->inbuf_index ^= 1;
1468 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1469 s->old_frame_size = s->frame_size;
1472 static inline void lsf_sf_expand(int *slen,
1473 int sf, int n1, int n2, int n3)
1492 static void exponents_from_scale_factors(MPADecodeContext *s,
1496 const UINT8 *bstab, *pretab;
1497 int len, i, j, k, l, v0, shift, gain, gains[3];
1500 exp_ptr = exponents;
1501 gain = g->global_gain - 210;
1502 shift = g->scalefac_scale + 1;
1504 bstab = band_size_long[s->sample_rate_index];
1505 pretab = mpa_pretab[g->preflag];
1506 for(i=0;i<g->long_end;i++) {
1507 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1513 if (g->short_start < 13) {
1514 bstab = band_size_short[s->sample_rate_index];
1515 gains[0] = gain - (g->subblock_gain[0] << 3);
1516 gains[1] = gain - (g->subblock_gain[1] << 3);
1517 gains[2] = gain - (g->subblock_gain[2] << 3);
1519 for(i=g->short_start;i<13;i++) {
1522 v0 = gains[l] - (g->scale_factors[k++] << shift);
1530 /* handle n = 0 too */
1531 static inline int get_bitsz(GetBitContext *s, int n)
1536 return get_bits(s, n);
1539 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1540 INT16 *exponents, int end_pos)
1543 int linbits, code, x, y, l, v, i, j, k, pos;
1544 GetBitContext last_gb;
1548 /* low frequencies (called big values) */
1551 j = g->region_size[i];
1554 /* select vlc table */
1555 k = g->table_select[i];
1556 l = mpa_huff_data[k][0];
1557 linbits = mpa_huff_data[k][1];
1559 code_table = huff_code_table[l];
1561 /* read huffcode and compute each couple */
1563 if (get_bits_count(&s->gb) >= end_pos)
1566 code = get_vlc(&s->gb, vlc);
1569 y = code_table[code];
1576 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1577 i, g->region_size[i] - j, x, y, exponents[s_index]);
1580 x += get_bitsz(&s->gb, linbits);
1581 v = l3_unscale(x, exponents[s_index]);
1582 if (get_bits1(&s->gb))
1587 g->sb_hybrid[s_index++] = v;
1590 y += get_bitsz(&s->gb, linbits);
1591 v = l3_unscale(y, exponents[s_index]);
1592 if (get_bits1(&s->gb))
1597 g->sb_hybrid[s_index++] = v;
1601 /* high frequencies */
1602 vlc = &huff_quad_vlc[g->count1table_select];
1603 last_gb.buffer = NULL;
1604 while (s_index <= 572) {
1605 pos = get_bits_count(&s->gb);
1606 if (pos >= end_pos) {
1607 if (pos > end_pos && last_gb.buffer != NULL) {
1608 /* some encoders generate an incorrect size for this
1609 part. We must go back into the data */
1617 code = get_vlc(&s->gb, vlc);
1618 dprintf("t=%d code=%d\n", g->count1table_select, code);
1622 if (code & (8 >> i)) {
1623 /* non zero value. Could use a hand coded function for
1625 v = l3_unscale(1, exponents[s_index]);
1626 if(get_bits1(&s->gb))
1631 g->sb_hybrid[s_index++] = v;
1634 while (s_index < 576)
1635 g->sb_hybrid[s_index++] = 0;
1639 /* Reorder short blocks from bitstream order to interleaved order. It
1640 would be faster to do it in parsing, but the code would be far more
1642 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1645 INT32 *ptr, *dst, *ptr1;
1648 if (g->block_type != 2)
1651 if (g->switch_point) {
1652 if (s->sample_rate_index != 8) {
1653 ptr = g->sb_hybrid + 36;
1655 ptr = g->sb_hybrid + 48;
1661 for(i=g->short_start;i<13;i++) {
1662 len = band_size_short[s->sample_rate_index][i];
1666 for(j=len;j>0;j--) {
1671 memcpy(ptr1, tmp, len * 3 * sizeof(INT32));
1675 #define ISQRT2 FIXR(0.70710678118654752440)
1677 static void compute_stereo(MPADecodeContext *s,
1678 GranuleDef *g0, GranuleDef *g1)
1682 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1683 INT32 (*is_tab)[16];
1685 int non_zero_found_short[3];
1687 /* intensity stereo */
1688 if (s->mode_ext & MODE_EXT_I_STEREO) {
1693 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1697 tab0 = g0->sb_hybrid + 576;
1698 tab1 = g1->sb_hybrid + 576;
1700 non_zero_found_short[0] = 0;
1701 non_zero_found_short[1] = 0;
1702 non_zero_found_short[2] = 0;
1703 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1704 for(i = 12;i >= g1->short_start;i--) {
1705 /* for last band, use previous scale factor */
1708 len = band_size_short[s->sample_rate_index][i];
1712 if (!non_zero_found_short[l]) {
1713 /* test if non zero band. if so, stop doing i-stereo */
1714 for(j=0;j<len;j++) {
1716 non_zero_found_short[l] = 1;
1720 sf = g1->scale_factors[k + l];
1726 for(j=0;j<len;j++) {
1728 tab0[j] = MULL(tmp0, v1);
1729 tab1[j] = MULL(tmp0, v2);
1733 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1734 /* lower part of the spectrum : do ms stereo
1736 for(j=0;j<len;j++) {
1739 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1740 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1747 non_zero_found = non_zero_found_short[0] |
1748 non_zero_found_short[1] |
1749 non_zero_found_short[2];
1751 for(i = g1->long_end - 1;i >= 0;i--) {
1752 len = band_size_long[s->sample_rate_index][i];
1755 /* test if non zero band. if so, stop doing i-stereo */
1756 if (!non_zero_found) {
1757 for(j=0;j<len;j++) {
1763 /* for last band, use previous scale factor */
1764 k = (i == 21) ? 20 : i;
1765 sf = g1->scale_factors[k];
1770 for(j=0;j<len;j++) {
1772 tab0[j] = MULL(tmp0, v1);
1773 tab1[j] = MULL(tmp0, v2);
1777 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1778 /* lower part of the spectrum : do ms stereo
1780 for(j=0;j<len;j++) {
1783 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1784 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1789 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1790 /* ms stereo ONLY */
1791 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1793 tab0 = g0->sb_hybrid;
1794 tab1 = g1->sb_hybrid;
1795 for(i=0;i<576;i++) {
1798 tab0[i] = tmp0 + tmp1;
1799 tab1[i] = tmp0 - tmp1;
1804 static void compute_antialias(MPADecodeContext *s,
1807 INT32 *ptr, *p0, *p1, *csa;
1808 int n, tmp0, tmp1, i, j;
1810 /* we antialias only "long" bands */
1811 if (g->block_type == 2) {
1812 if (!g->switch_point)
1814 /* XXX: check this for 8000Hz case */
1820 ptr = g->sb_hybrid + 18;
1821 for(i = n;i > 0;i--) {
1824 csa = &csa_table[0][0];
1828 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1829 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1838 static void compute_imdct(MPADecodeContext *s,
1843 INT32 *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
1847 int i, j, k, mdct_long_end, v, sblimit;
1849 /* find last non zero block */
1850 ptr = g->sb_hybrid + 576;
1851 ptr1 = g->sb_hybrid + 2 * 18;
1852 while (ptr >= ptr1) {
1854 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
1858 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
1860 if (g->block_type == 2) {
1861 /* XXX: check for 8000 Hz */
1862 if (g->switch_point)
1867 mdct_long_end = sblimit;
1872 for(j=0;j<mdct_long_end;j++) {
1874 /* apply window & overlap with previous buffer */
1875 out_ptr = sb_samples + j;
1877 if (g->switch_point && j < 2)
1880 win1 = mdct_win[g->block_type];
1881 /* select frequency inversion */
1882 win = win1 + ((4 * 36) & -(j & 1));
1884 *out_ptr = MULL(out[i], win[i]) + buf[i];
1885 buf[i] = MULL(out[i + 18], win[i + 18]);
1891 for(j=mdct_long_end;j<sblimit;j++) {
1897 /* select frequency inversion */
1898 win = mdct_win[2] + ((4 * 36) & -(j & 1));
1901 /* reorder input for short mdct */
1908 /* apply 12 point window and do small overlap */
1910 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
1911 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
1916 out_ptr = sb_samples + j;
1918 *out_ptr = out[i] + buf[i];
1919 buf[i] = out[i + 18];
1926 for(j=sblimit;j<SBLIMIT;j++) {
1928 out_ptr = sb_samples + j;
1939 void sample_dump(int fnum, INT32 *tab, int n)
1941 static FILE *files[16], *f;
1948 sprintf(buf, "/tmp/out%d.%s.pcm",
1950 #ifdef USE_HIGHPRECISION
1956 f = fopen(buf, "w");
1964 printf("pos=%d\n", pos);
1966 printf(" %0.4f", (double)tab[i] / FRAC_ONE);
1973 /* normalize to 23 frac bits */
1974 v = tab[i] << (23 - FRAC_BITS);
1975 fwrite(&v, 1, sizeof(INT32), f);
1981 /* main layer3 decoding function */
1982 static int mp_decode_layer3(MPADecodeContext *s)
1984 int nb_granules, main_data_begin, private_bits;
1985 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
1986 GranuleDef granules[2][2], *g;
1987 INT16 exponents[576];
1989 /* read side info */
1991 main_data_begin = get_bits(&s->gb, 8);
1992 if (s->nb_channels == 2)
1993 private_bits = get_bits(&s->gb, 2);
1995 private_bits = get_bits(&s->gb, 1);
1998 main_data_begin = get_bits(&s->gb, 9);
1999 if (s->nb_channels == 2)
2000 private_bits = get_bits(&s->gb, 3);
2002 private_bits = get_bits(&s->gb, 5);
2004 for(ch=0;ch<s->nb_channels;ch++) {
2005 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2006 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2010 for(gr=0;gr<nb_granules;gr++) {
2011 for(ch=0;ch<s->nb_channels;ch++) {
2012 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2013 g = &granules[ch][gr];
2014 g->part2_3_length = get_bits(&s->gb, 12);
2015 g->big_values = get_bits(&s->gb, 9);
2016 g->global_gain = get_bits(&s->gb, 8);
2017 /* if MS stereo only is selected, we precompute the
2018 1/sqrt(2) renormalization factor */
2019 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2021 g->global_gain -= 2;
2023 g->scalefac_compress = get_bits(&s->gb, 9);
2025 g->scalefac_compress = get_bits(&s->gb, 4);
2026 blocksplit_flag = get_bits(&s->gb, 1);
2027 if (blocksplit_flag) {
2028 g->block_type = get_bits(&s->gb, 2);
2029 if (g->block_type == 0)
2031 g->switch_point = get_bits(&s->gb, 1);
2033 g->table_select[i] = get_bits(&s->gb, 5);
2035 g->subblock_gain[i] = get_bits(&s->gb, 3);
2036 /* compute huffman coded region sizes */
2037 if (g->block_type == 2)
2038 g->region_size[0] = (36 / 2);
2040 if (s->sample_rate_index <= 2)
2041 g->region_size[0] = (36 / 2);
2042 else if (s->sample_rate_index != 8)
2043 g->region_size[0] = (54 / 2);
2045 g->region_size[0] = (108 / 2);
2047 g->region_size[1] = (576 / 2);
2049 int region_address1, region_address2, l;
2051 g->switch_point = 0;
2053 g->table_select[i] = get_bits(&s->gb, 5);
2054 /* compute huffman coded region sizes */
2055 region_address1 = get_bits(&s->gb, 4);
2056 region_address2 = get_bits(&s->gb, 3);
2057 dprintf("region1=%d region2=%d\n",
2058 region_address1, region_address2);
2060 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2061 l = region_address1 + region_address2 + 2;
2062 /* should not overflow */
2066 band_index_long[s->sample_rate_index][l] >> 1;
2068 /* convert region offsets to region sizes and truncate
2069 size to big_values */
2070 g->region_size[2] = (576 / 2);
2073 k = g->region_size[i];
2074 if (k > g->big_values)
2076 g->region_size[i] = k - j;
2080 /* compute band indexes */
2081 if (g->block_type == 2) {
2082 if (g->switch_point) {
2083 /* if switched mode, we handle the 36 first samples as
2084 long blocks. For 8000Hz, we handle the 48 first
2085 exponents as long blocks (XXX: check this!) */
2086 if (s->sample_rate_index <= 2)
2088 else if (s->sample_rate_index != 8)
2091 g->long_end = 4; /* 8000 Hz */
2093 if (s->sample_rate_index != 8)
2102 g->short_start = 13;
2108 g->preflag = get_bits(&s->gb, 1);
2109 g->scalefac_scale = get_bits(&s->gb, 1);
2110 g->count1table_select = get_bits(&s->gb, 1);
2111 dprintf("block_type=%d switch_point=%d\n",
2112 g->block_type, g->switch_point);
2116 /* now we get bits from the main_data_begin offset */
2117 dprintf("seekback: %d\n", main_data_begin);
2118 seek_to_maindata(s, main_data_begin);
2120 for(gr=0;gr<nb_granules;gr++) {
2121 for(ch=0;ch<s->nb_channels;ch++) {
2122 g = &granules[ch][gr];
2124 bits_pos = get_bits_count(&s->gb);
2128 int slen, slen1, slen2;
2130 /* MPEG1 scale factors */
2131 slen1 = slen_table[0][g->scalefac_compress];
2132 slen2 = slen_table[1][g->scalefac_compress];
2133 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2134 if (g->block_type == 2) {
2135 n = g->switch_point ? 17 : 18;
2138 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2140 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2142 g->scale_factors[j++] = 0;
2144 sc = granules[ch][0].scale_factors;
2147 n = (k == 0 ? 6 : 5);
2148 if ((g->scfsi & (0x8 >> k)) == 0) {
2149 slen = (k < 2) ? slen1 : slen2;
2151 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2153 /* simply copy from last granule */
2155 g->scale_factors[j] = sc[j];
2160 g->scale_factors[j++] = 0;
2164 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2167 printf(" %d", g->scale_factors[i]);
2172 int tindex, tindex2, slen[4], sl, sf;
2174 /* LSF scale factors */
2175 if (g->block_type == 2) {
2176 tindex = g->switch_point ? 2 : 1;
2180 sf = g->scalefac_compress;
2181 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2182 /* intensity stereo case */
2185 lsf_sf_expand(slen, sf, 6, 6, 0);
2187 } else if (sf < 244) {
2188 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2191 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2197 lsf_sf_expand(slen, sf, 5, 4, 4);
2199 } else if (sf < 500) {
2200 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2203 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2211 n = lsf_nsf_table[tindex2][tindex][k];
2214 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2216 /* XXX: should compute exact size */
2218 g->scale_factors[j] = 0;
2221 printf("gr=%d ch=%d scale_factors:\n",
2224 printf(" %d", g->scale_factors[i]);
2230 exponents_from_scale_factors(s, g, exponents);
2232 /* read Huffman coded residue */
2233 if (huffman_decode(s, g, exponents,
2234 bits_pos + g->part2_3_length) < 0)
2237 sample_dump(0, g->sb_hybrid, 576);
2240 /* skip extension bits */
2241 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2242 if (bits_left < 0) {
2243 dprintf("bits_left=%d\n", bits_left);
2246 while (bits_left >= 16) {
2247 skip_bits(&s->gb, 16);
2251 skip_bits(&s->gb, bits_left);
2254 if (s->nb_channels == 2)
2255 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2257 for(ch=0;ch<s->nb_channels;ch++) {
2258 g = &granules[ch][gr];
2260 reorder_block(s, g);
2262 sample_dump(0, g->sb_hybrid, 576);
2264 compute_antialias(s, g);
2266 sample_dump(1, g->sb_hybrid, 576);
2268 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2270 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2274 return nb_granules * 18;
2277 static int mp_decode_frame(MPADecodeContext *s,
2280 int i, nb_frames, ch;
2283 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2284 (s->inbuf_ptr - s->inbuf - HEADER_SIZE)*8);
2286 /* skip error protection field */
2287 if (s->error_protection)
2288 get_bits(&s->gb, 16);
2290 dprintf("frame %d:\n", s->frame_count);
2293 nb_frames = mp_decode_layer1(s);
2296 nb_frames = mp_decode_layer2(s);
2300 nb_frames = mp_decode_layer3(s);
2304 for(i=0;i<nb_frames;i++) {
2305 for(ch=0;ch<s->nb_channels;ch++) {
2307 printf("%d-%d:", i, ch);
2308 for(j=0;j<SBLIMIT;j++)
2309 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2314 /* apply the synthesis filter */
2315 for(ch=0;ch<s->nb_channels;ch++) {
2316 samples_ptr = samples + ch;
2317 for(i=0;i<nb_frames;i++) {
2318 synth_filter(s, ch, samples_ptr, s->nb_channels,
2319 s->sb_samples[ch][i]);
2320 samples_ptr += 32 * s->nb_channels;
2326 return nb_frames * 32 * sizeof(short) * s->nb_channels;
2329 static int decode_frame(AVCodecContext * avctx,
2330 void *data, int *data_size,
2331 UINT8 * buf, int buf_size)
2333 MPADecodeContext *s = avctx->priv_data;
2337 short *out_samples = data;
2341 while (buf_size > 0) {
2342 len = s->inbuf_ptr - s->inbuf;
2343 if (s->frame_size == 0) {
2344 /* special case for next header for first frame in free
2345 format case (XXX: find a simpler method) */
2346 if (s->free_format_next_header != 0) {
2347 s->inbuf[0] = s->free_format_next_header >> 24;
2348 s->inbuf[1] = s->free_format_next_header >> 16;
2349 s->inbuf[2] = s->free_format_next_header >> 8;
2350 s->inbuf[3] = s->free_format_next_header;
2351 s->inbuf_ptr = s->inbuf + 4;
2352 s->free_format_next_header = 0;
2355 /* no header seen : find one. We need at least HEADER_SIZE
2356 bytes to parse it */
2357 len = HEADER_SIZE - len;
2361 memcpy(s->inbuf_ptr, buf_ptr, len);
2364 s->inbuf_ptr += len;
2366 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2368 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2369 (s->inbuf[2] << 8) | s->inbuf[3];
2371 if (check_header(header) < 0) {
2372 /* no sync found : move by one byte (inefficient, but simple!) */
2373 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2375 dprintf("skip %x\n", header);
2376 /* reset free format frame size to give a chance
2377 to get a new bitrate */
2378 s->free_format_frame_size = 0;
2380 if (decode_header(s, header) == 1) {
2381 /* free format: prepare to compute frame size */
2384 /* update codec info */
2385 avctx->sample_rate = s->sample_rate;
2386 avctx->channels = s->nb_channels;
2387 avctx->bit_rate = s->bit_rate;
2388 avctx->frame_size = s->frame_size;
2391 } else if (s->frame_size == -1) {
2392 /* free format : find next sync to compute frame size */
2393 len = MPA_MAX_CODED_FRAME_SIZE - len;
2397 /* frame too long: resync */
2399 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2406 memcpy(s->inbuf_ptr, buf_ptr, len);
2407 /* check for header */
2408 p = s->inbuf_ptr - 3;
2409 pend = s->inbuf_ptr + len - 4;
2411 header = (p[0] << 24) | (p[1] << 16) |
2413 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2414 (s->inbuf[2] << 8) | s->inbuf[3];
2415 /* check with high probability that we have a
2417 if ((header & SAME_HEADER_MASK) ==
2418 (header1 & SAME_HEADER_MASK)) {
2419 /* header found: update pointers */
2420 len = (p + 4) - s->inbuf_ptr;
2424 /* compute frame size */
2425 s->free_format_next_header = header;
2426 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2427 padding = (header1 >> 9) & 1;
2429 s->free_format_frame_size -= padding * 4;
2431 s->free_format_frame_size -= padding;
2432 dprintf("free frame size=%d padding=%d\n",
2433 s->free_format_frame_size, padding);
2434 decode_header(s, header1);
2439 /* not found: simply increase pointers */
2441 s->inbuf_ptr += len;
2444 } else if (len < s->frame_size) {
2445 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2446 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2447 len = s->frame_size - len;
2450 memcpy(s->inbuf_ptr, buf_ptr, len);
2452 s->inbuf_ptr += len;
2455 out_size = mp_decode_frame(s, out_samples);
2456 s->inbuf_ptr = s->inbuf;
2458 *data_size = out_size;
2464 return buf_ptr - buf;
2467 AVCodec mp2_decoder =
2472 sizeof(MPADecodeContext),
2479 AVCodec mp3_decoder =
2484 sizeof(MPADecodeContext),