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 //#define USE_HIGHPRECISION
33 #ifdef USE_HIGHPRECISION
34 #define FRAC_BITS 23 /* fractional bits for sb_samples and dct */
35 #define WFRAC_BITS 16 /* fractional bits for window */
37 #define FRAC_BITS 15 /* fractional bits for sb_samples and dct */
38 #define WFRAC_BITS 14 /* fractional bits for window */
41 #define FRAC_ONE (1 << FRAC_BITS)
43 #define MULL(a,b) (((INT64)(a) * (INT64)(b)) >> FRAC_BITS)
44 #define MUL64(a,b) ((INT64)(a) * (INT64)(b))
45 #define FIX(a) ((int)((a) * FRAC_ONE))
46 /* WARNING: only correct for posititive numbers */
47 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
48 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
51 typedef INT16 MPA_INT;
53 typedef INT32 MPA_INT;
59 #define BACKSTEP_SIZE 512
61 typedef struct MPADecodeContext {
62 UINT8 inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
64 UINT8 *inbuf_ptr, *inbuf;
66 int free_format_frame_size; /* frame size in case of free format
67 (zero if currently unknown) */
68 /* next header (used in free format parsing) */
69 UINT32 free_format_next_header;
73 int sample_rate_index; /* between 0 and 8 */
81 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2];
82 int synth_buf_offset[MPA_MAX_CHANNELS];
83 INT32 sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT];
84 INT32 mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
90 /* layer 3 "granule" */
91 typedef struct GranuleDef {
96 int scalefac_compress;
100 int subblock_gain[3];
101 UINT8 scalefac_scale;
102 UINT8 count1table_select;
103 int region_size[3]; /* number of huffman codes in each region */
105 int short_start, long_end; /* long/short band indexes */
106 UINT8 scale_factors[40];
107 INT32 sb_hybrid[SBLIMIT * 18]; /* 576 samples */
110 #define MODE_EXT_MS_STEREO 2
111 #define MODE_EXT_I_STEREO 1
113 /* layer 3 huffman tables */
114 typedef struct HuffTable {
120 #include "mpegaudiodectab.h"
122 /* vlc structure for decoding layer 3 huffman tables */
123 static VLC huff_vlc[16];
124 static UINT8 *huff_code_table[16];
125 static VLC huff_quad_vlc[2];
126 /* computed from band_size_long */
127 static UINT16 band_index_long[9][23];
128 /* XXX: free when all decoders are closed */
129 #define TABLE_4_3_SIZE (8191 + 16)
130 static INT8 *table_4_3_exp;
132 static UINT16 *table_4_3_value;
134 static UINT32 *table_4_3_value;
136 /* intensity stereo coef table */
137 static INT32 is_table[2][16];
138 static INT32 is_table_lsf[2][2][16];
139 static INT32 csa_table[8][2];
140 static INT32 mdct_win[8][36];
142 /* lower 2 bits: modulo 3, higher bits: shift */
143 static UINT16 scale_factor_modshift[64];
144 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
145 static INT32 scale_factor_mult[15][3];
146 /* mult table for layer 2 group quantization */
148 #define SCALE_GEN(v) \
149 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
151 static INT32 scale_factor_mult2[3][3] = {
152 SCALE_GEN(1.0 / 3.0), /* 3 steps */
153 SCALE_GEN(1.0 / 5.0), /* 5 steps */
154 SCALE_GEN(1.0 / 9.0), /* 9 steps */
158 static UINT32 scale_factor_mult3[4] = {
160 FIXR(1.18920711500272106671),
161 FIXR(1.41421356237309504880),
162 FIXR(1.68179283050742908605),
165 static MPA_INT window[512];
167 /* layer 1 unscaling */
168 /* n = number of bits of the mantissa minus 1 */
169 static inline int l1_unscale(int n, int mant, int scale_factor)
174 shift = scale_factor_modshift[scale_factor];
177 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
179 return (int)((val + (1 << (shift - 1))) >> shift);
182 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
186 shift = scale_factor_modshift[scale_factor];
189 /* XXX: store the result directly */
190 val = (2 * (mant - (steps >> 1))) * scale_factor_mult2[steps >> 2][mod];
191 return (val + (1 << (shift - 1))) >> shift;
194 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
195 static inline int l3_unscale(int value, int exponent)
204 e = table_4_3_exp[value];
205 e += (exponent >> 2);
211 m = table_4_3_value[value];
213 m = (m * scale_factor_mult3[exponent & 3]);
214 m = (m + (1 << (e-1))) >> e;
217 m = MUL64(m, scale_factor_mult3[exponent & 3]);
218 m = (m + (UINT64_C(1) << (e-1))) >> e;
223 /* all integer n^(4/3) computation code */
226 #define POW_FRAC_BITS 24
227 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
228 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
229 #define POW_MULL(a,b) (((INT64)(a) * (INT64)(b)) >> POW_FRAC_BITS)
231 static int dev_4_3_coefs[DEV_ORDER];
233 static int pow_mult3[3] = {
235 POW_FIX(1.25992104989487316476),
236 POW_FIX(1.58740105196819947474),
239 static void int_pow_init(void)
244 for(i=0;i<DEV_ORDER;i++) {
245 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
246 dev_4_3_coefs[i] = a;
250 /* return the mantissa and the binary exponent */
251 static int int_pow(int i, int *exp_ptr)
259 while (a < (1 << (POW_FRAC_BITS - 1))) {
263 a -= (1 << POW_FRAC_BITS);
265 for(j = DEV_ORDER - 1; j >= 0; j--)
266 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
267 a = (1 << POW_FRAC_BITS) + a1;
268 /* exponent compute (exact) */
272 a = POW_MULL(a, pow_mult3[er]);
273 while (a >= 2 * POW_FRAC_ONE) {
277 /* convert to float */
278 while (a < POW_FRAC_ONE) {
282 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
283 #if (POW_FRAC_BITS - 1) > FRAC_BITS
284 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
285 /* correct overflow */
286 if (a >= 2 * (1 << FRAC_BITS)) {
295 static int decode_init(AVCodecContext * avctx)
297 MPADecodeContext *s = avctx->priv_data;
302 /* scale factors table for layer 1/2 */
305 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
312 scale_factor_modshift[i] = mod | (shift << 2);
315 /* scale factor multiply for layer 1 */
319 norm = ((INT64_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
320 scale_factor_mult[i][0] = MULL(FIXR(1.0), norm);
321 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259), norm);
322 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249), norm);
323 dprintf("%d: norm=%x s=%x %x %x\n",
325 scale_factor_mult[i][0],
326 scale_factor_mult[i][1],
327 scale_factor_mult[i][2]);
331 /* max = 18760, max sum over all 16 coefs : 44736 */
336 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
345 /* huffman decode tables */
346 huff_code_table[0] = NULL;
348 const HuffTable *h = &mpa_huff_tables[i];
355 init_vlc(&huff_vlc[i], 8, n,
356 h->bits, 1, 1, h->codes, 2, 2);
358 code_table = av_mallocz(n);
360 for(x=0;x<xsize;x++) {
362 code_table[j++] = (x << 4) | y;
364 huff_code_table[i] = code_table;
367 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
368 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1);
374 band_index_long[i][j] = k;
375 k += band_size_long[i][j];
377 band_index_long[i][22] = k;
380 /* compute n ^ (4/3) and store it in mantissa/exp format */
381 table_4_3_exp = av_mallocz(TABLE_4_3_SIZE *
382 sizeof(table_4_3_exp[0]));
385 table_4_3_value = av_mallocz(TABLE_4_3_SIZE *
386 sizeof(table_4_3_value[0]));
387 if (!table_4_3_value) {
388 av_free(table_4_3_exp);
393 for(i=1;i<TABLE_4_3_SIZE;i++) {
401 f = pow((double)i, 4.0 / 3.0);
405 if ((unsigned short)m1 != m1) {
411 if (m != m1 || e != e1) {
412 printf("%4d: m=%x m1=%x e=%d e1=%d\n",
417 /* normalized to FRAC_BITS */
418 table_4_3_value[i] = m;
419 table_4_3_exp[i] = e;
426 f = tan((double)i * M_PI / 12.0);
427 v = FIXR(f / (1.0 + f));
432 is_table[1][6 - i] = v;
436 is_table[0][i] = is_table[1][i] = 0.0;
443 e = -(j + 1) * ((i + 1) >> 1);
444 f = pow(2.0, e / 4.0);
446 is_table_lsf[j][k ^ 1][i] = FIXR(f);
447 is_table_lsf[j][k][i] = FIXR(1.0);
448 dprintf("is_table_lsf %d %d: %x %x\n",
449 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
456 cs = 1.0 / sqrt(1.0 + ci * ci);
458 csa_table[i][0] = FIX(cs);
459 csa_table[i][1] = FIX(ca);
462 /* compute mdct windows */
465 v = FIXR(sin(M_PI * (i + 0.5) / 36.0));
471 mdct_win[1][18 + i] = FIXR(1.0);
472 mdct_win[1][24 + i] = FIXR(sin(M_PI * ((i + 6) + 0.5) / 12.0));
473 mdct_win[1][30 + i] = FIXR(0.0);
475 mdct_win[3][i] = FIXR(0.0);
476 mdct_win[3][6 + i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
477 mdct_win[3][12 + i] = FIXR(1.0);
481 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
483 /* NOTE: we do frequency inversion adter the MDCT by changing
484 the sign of the right window coefs */
487 mdct_win[j + 4][i] = mdct_win[j][i];
488 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
494 printf("win%d=\n", j);
496 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
504 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
505 s->inbuf_ptr = s->inbuf;
512 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */;
516 #define COS0_0 FIXR(0.50060299823519630134)
517 #define COS0_1 FIXR(0.50547095989754365998)
518 #define COS0_2 FIXR(0.51544730992262454697)
519 #define COS0_3 FIXR(0.53104259108978417447)
520 #define COS0_4 FIXR(0.55310389603444452782)
521 #define COS0_5 FIXR(0.58293496820613387367)
522 #define COS0_6 FIXR(0.62250412303566481615)
523 #define COS0_7 FIXR(0.67480834145500574602)
524 #define COS0_8 FIXR(0.74453627100229844977)
525 #define COS0_9 FIXR(0.83934964541552703873)
526 #define COS0_10 FIXR(0.97256823786196069369)
527 #define COS0_11 FIXR(1.16943993343288495515)
528 #define COS0_12 FIXR(1.48416461631416627724)
529 #define COS0_13 FIXR(2.05778100995341155085)
530 #define COS0_14 FIXR(3.40760841846871878570)
531 #define COS0_15 FIXR(10.19000812354805681150)
533 #define COS1_0 FIXR(0.50241928618815570551)
534 #define COS1_1 FIXR(0.52249861493968888062)
535 #define COS1_2 FIXR(0.56694403481635770368)
536 #define COS1_3 FIXR(0.64682178335999012954)
537 #define COS1_4 FIXR(0.78815462345125022473)
538 #define COS1_5 FIXR(1.06067768599034747134)
539 #define COS1_6 FIXR(1.72244709823833392782)
540 #define COS1_7 FIXR(5.10114861868916385802)
542 #define COS2_0 FIXR(0.50979557910415916894)
543 #define COS2_1 FIXR(0.60134488693504528054)
544 #define COS2_2 FIXR(0.89997622313641570463)
545 #define COS2_3 FIXR(2.56291544774150617881)
547 #define COS3_0 FIXR(0.54119610014619698439)
548 #define COS3_1 FIXR(1.30656296487637652785)
550 #define COS4_0 FIXR(0.70710678118654752439)
552 /* butterfly operator */
555 tmp0 = tab[a] + tab[b];\
556 tmp1 = tab[a] - tab[b];\
558 tab[b] = MULL(tmp1, c);\
561 #define BF1(a, b, c, d)\
568 #define BF2(a, b, c, d)\
578 #define ADD(a, b) tab[a] += tab[b]
580 /* DCT32 without 1/sqrt(2) coef zero scaling. */
581 static void dct32(INT32 *out, INT32 *tab)
713 out[ 1] = tab[16] + tab[24];
714 out[17] = tab[17] + tab[25];
715 out[ 9] = tab[18] + tab[26];
716 out[25] = tab[19] + tab[27];
717 out[ 5] = tab[20] + tab[28];
718 out[21] = tab[21] + tab[29];
719 out[13] = tab[22] + tab[30];
720 out[29] = tab[23] + tab[31];
721 out[ 3] = tab[24] + tab[20];
722 out[19] = tab[25] + tab[21];
723 out[11] = tab[26] + tab[22];
724 out[27] = tab[27] + tab[23];
725 out[ 7] = tab[28] + tab[18];
726 out[23] = tab[29] + tab[19];
727 out[15] = tab[30] + tab[17];
731 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
735 #define OUT_SAMPLE(sum)\
738 sum1 = (sum + (1 << (OUT_SHIFT - 1))) >> OUT_SHIFT;\
741 else if (sum1 > 32767)\
747 #define SUM8(off, op) \
749 sum op w[0 * 64 + off] * p[0 * 64];\
750 sum op w[1 * 64 + off] * p[1 * 64];\
751 sum op w[2 * 64 + off] * p[2 * 64];\
752 sum op w[3 * 64 + off] * p[3 * 64];\
753 sum op w[4 * 64 + off] * p[4 * 64];\
754 sum op w[5 * 64 + off] * p[5 * 64];\
755 sum op w[6 * 64 + off] * p[6 * 64];\
756 sum op w[7 * 64 + off] * p[7 * 64];\
761 #define OUT_SAMPLE(sum)\
764 sum1 = (int)((sum + (INT64_C(1) << (OUT_SHIFT - 1))) >> OUT_SHIFT);\
767 else if (sum1 > 32767)\
773 #define SUM8(off, op) \
775 sum op MUL64(w[0 * 64 + off], p[0 * 64]);\
776 sum op MUL64(w[1 * 64 + off], p[1 * 64]);\
777 sum op MUL64(w[2 * 64 + off], p[2 * 64]);\
778 sum op MUL64(w[3 * 64 + off], p[3 * 64]);\
779 sum op MUL64(w[4 * 64 + off], p[4 * 64]);\
780 sum op MUL64(w[5 * 64 + off], p[5 * 64]);\
781 sum op MUL64(w[6 * 64 + off], p[6 * 64]);\
782 sum op MUL64(w[7 * 64 + off], p[7 * 64]);\
787 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
789 /* XXX: optimize by avoiding ring buffer usage */
790 static void synth_filter(MPADecodeContext *s1,
791 int ch, INT16 *samples, int incr,
792 INT32 sb_samples[SBLIMIT])
795 register MPA_INT *synth_buf, *p;
804 dct32(tmp, sb_samples);
806 offset = s1->synth_buf_offset[ch];
807 synth_buf = s1->synth_buf[ch] + offset;
819 /* copy to avoid wrap */
820 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
825 p = synth_buf + 16 + j; /* 0-15 */
827 p = synth_buf + 48 - j; /* 32-47 */
833 p = synth_buf + 32; /* 48 */
841 p = synth_buf + 48 - j; /* 17-31 */
843 p = synth_buf + 16 + j; /* 49-63 */
848 offset = (offset - 32) & 511;
849 s1->synth_buf_offset[ch] = offset;
853 #define C1 FIXR(0.99144486137381041114)
854 #define C3 FIXR(0.92387953251128675612)
855 #define C5 FIXR(0.79335334029123516458)
856 #define C7 FIXR(0.60876142900872063941)
857 #define C9 FIXR(0.38268343236508977173)
858 #define C11 FIXR(0.13052619222005159154)
860 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
862 static void imdct12(int *out, int *in)
865 INT64 in1_3, in1_9, in4_3, in4_9;
867 in1_3 = MUL64(in[1], C3);
868 in1_9 = MUL64(in[1], C9);
869 in4_3 = MUL64(in[4], C3);
870 in4_9 = MUL64(in[4], C9);
872 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
873 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
876 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
877 MUL64(in[2] + in[5], C3) - in4_9);
880 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
881 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
884 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
885 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
888 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
889 MUL64(in[2] + in[5], C9) + in4_3);
892 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
893 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
906 #define C1 FIXR(0.98480775301220805936)
907 #define C2 FIXR(0.93969262078590838405)
908 #define C3 FIXR(0.86602540378443864676)
909 #define C4 FIXR(0.76604444311897803520)
910 #define C5 FIXR(0.64278760968653932632)
912 #define C7 FIXR(0.34202014332566873304)
913 #define C8 FIXR(0.17364817766693034885)
915 /* 0.5 / cos(pi*(2*i+1)/36) */
916 static const int icos36[9] = {
917 FIXR(0.50190991877167369479),
918 FIXR(0.51763809020504152469),
919 FIXR(0.55168895948124587824),
920 FIXR(0.61038729438072803416),
921 FIXR(0.70710678118654752439),
922 FIXR(0.87172339781054900991),
923 FIXR(1.18310079157624925896),
924 FIXR(1.93185165257813657349),
925 FIXR(5.73685662283492756461),
928 static const int icos72[18] = {
929 /* 0.5 / cos(pi*(2*i+19)/72) */
930 FIXR(0.74009361646113053152),
931 FIXR(0.82133981585229078570),
932 FIXR(0.93057949835178895673),
933 FIXR(1.08284028510010010928),
934 FIXR(1.30656296487637652785),
935 FIXR(1.66275476171152078719),
936 FIXR(2.31011315767264929558),
937 FIXR(3.83064878777019433457),
938 FIXR(11.46279281302667383546),
940 /* 0.5 / cos(pi*(2*(i + 18) +19)/72) */
941 FIXR(-0.67817085245462840086),
942 FIXR(-0.63023620700513223342),
943 FIXR(-0.59284452371708034528),
944 FIXR(-0.56369097343317117734),
945 FIXR(-0.54119610014619698439),
946 FIXR(-0.52426456257040533932),
947 FIXR(-0.51213975715725461845),
948 FIXR(-0.50431448029007636036),
949 FIXR(-0.50047634258165998492),
952 /* using Lee like decomposition followed by hand coded 9 points DCT */
953 static void imdct36(int *out, int *in)
955 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
956 int tmp[18], *tmp1, *in1;
968 in3_3 = MUL64(in1[2*3], C3);
969 in6_6 = MUL64(in1[2*6], C6);
971 tmp1[0] = FRAC_RND(MUL64(in1[2*1], C1) + in3_3 +
972 MUL64(in1[2*5], C5) + MUL64(in1[2*7], C7));
973 tmp1[2] = in1[2*0] + FRAC_RND(MUL64(in1[2*2], C2) +
974 MUL64(in1[2*4], C4) + in6_6 +
975 MUL64(in1[2*8], C8));
976 tmp1[4] = FRAC_RND(MUL64(in1[2*1] - in1[2*5] - in1[2*7], C3));
977 tmp1[6] = FRAC_RND(MUL64(in1[2*2] - in1[2*4] - in1[2*8], C6)) -
979 tmp1[8] = FRAC_RND(MUL64(in1[2*1], C5) - in3_3 -
980 MUL64(in1[2*5], C7) + MUL64(in1[2*7], C1));
981 tmp1[10] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C8) -
982 MUL64(in1[2*4], C2) + in6_6 +
983 MUL64(in1[2*8], C4));
984 tmp1[12] = FRAC_RND(MUL64(in1[2*1], C7) - in3_3 +
985 MUL64(in1[2*5], C1) -
986 MUL64(in1[2*7], C5));
987 tmp1[14] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C4) +
988 MUL64(in1[2*4], C8) + in6_6 -
989 MUL64(in1[2*8], C2));
990 tmp1[16] = in1[2*0] - in1[2*2] + in1[2*4] - in1[2*6] + in1[2*8];
1002 s1 = MULL(t3 + t2, icos36[j]);
1003 s3 = MULL(t3 - t2, icos36[8 - j]);
1005 t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
1006 t1 = MULL(s0 - s1, icos72[8 - j]);
1007 out[18 + 9 + j] = t0;
1008 out[18 + 8 - j] = t0;
1012 t0 = MULL(s2 + s3, icos72[9+j]);
1013 t1 = MULL(s2 - s3, icos72[j]);
1014 out[18 + 9 + (8 - j)] = t0;
1016 out[9 + (8 - j)] = -t1;
1022 s1 = MULL(tmp[17], icos36[4]);
1023 t0 = MULL(s0 + s1, icos72[9 + 4]);
1024 t1 = MULL(s0 - s1, icos72[4]);
1025 out[18 + 9 + 4] = t0;
1026 out[18 + 8 - 4] = t0;
1031 /* fast header check for resync */
1032 static int check_header(UINT32 header)
1035 if ((header & 0xffe00000) != 0xffe00000)
1038 if (((header >> 17) & 3) == 0)
1041 if (((header >> 12) & 0xf) == 0xf)
1044 if (((header >> 10) & 3) == 3)
1049 /* header + layer + bitrate + freq + lsf/mpeg25 */
1050 #define SAME_HEADER_MASK \
1051 (0xffe00000 | (3 << 17) | (0xf << 12) | (3 << 10) | (3 << 19))
1053 /* header decoding. MUST check the header before because no
1054 consistency check is done there. Return 1 if free format found and
1055 that the frame size must be computed externally */
1056 static int decode_header(MPADecodeContext *s, UINT32 header)
1058 int sample_rate, frame_size, mpeg25, padding;
1059 int sample_rate_index, bitrate_index;
1060 if (header & (1<<20)) {
1061 s->lsf = (header & (1<<19)) ? 0 : 1;
1068 s->layer = 4 - ((header >> 17) & 3);
1069 /* extract frequency */
1070 sample_rate_index = (header >> 10) & 3;
1071 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1072 if (sample_rate == 0)
1074 sample_rate_index += 3 * (s->lsf + mpeg25);
1075 s->sample_rate_index = sample_rate_index;
1076 s->error_protection = ((header >> 16) & 1) ^ 1;
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;
1134 s->sample_rate = sample_rate;
1137 printf("layer%d, %d Hz, %d kbits/s, ",
1138 s->layer, s->sample_rate, s->bit_rate);
1139 if (s->nb_channels == 2) {
1140 if (s->layer == 3) {
1141 if (s->mode_ext & MODE_EXT_MS_STEREO)
1143 if (s->mode_ext & MODE_EXT_I_STEREO)
1155 /* return the number of decoded frames */
1156 static int mp_decode_layer1(MPADecodeContext *s)
1158 int bound, i, v, n, ch, j, mant;
1159 UINT8 allocation[MPA_MAX_CHANNELS][SBLIMIT];
1160 UINT8 scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1162 if (s->mode == MPA_JSTEREO)
1163 bound = (s->mode_ext + 1) * 4;
1167 /* allocation bits */
1168 for(i=0;i<bound;i++) {
1169 for(ch=0;ch<s->nb_channels;ch++) {
1170 allocation[ch][i] = get_bits(&s->gb, 4);
1173 for(i=bound;i<SBLIMIT;i++) {
1174 allocation[0][i] = get_bits(&s->gb, 4);
1178 for(i=0;i<bound;i++) {
1179 for(ch=0;ch<s->nb_channels;ch++) {
1180 if (allocation[ch][i])
1181 scale_factors[ch][i] = get_bits(&s->gb, 6);
1184 for(i=bound;i<SBLIMIT;i++) {
1185 if (allocation[0][i]) {
1186 scale_factors[0][i] = get_bits(&s->gb, 6);
1187 scale_factors[1][i] = get_bits(&s->gb, 6);
1191 /* compute samples */
1193 for(i=0;i<bound;i++) {
1194 for(ch=0;ch<s->nb_channels;ch++) {
1195 n = allocation[ch][i];
1197 mant = get_bits(&s->gb, n + 1);
1198 v = l1_unscale(n, mant, scale_factors[ch][i]);
1202 s->sb_samples[ch][j][i] = v;
1205 for(i=bound;i<SBLIMIT;i++) {
1206 n = allocation[0][i];
1208 mant = get_bits(&s->gb, n + 1);
1209 v = l1_unscale(n, mant, scale_factors[0][i]);
1210 s->sb_samples[0][j][i] = v;
1211 v = l1_unscale(n, mant, scale_factors[1][i]);
1212 s->sb_samples[1][j][i] = v;
1214 s->sb_samples[0][j][i] = 0;
1215 s->sb_samples[1][j][i] = 0;
1222 /* bitrate is in kb/s */
1223 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1225 int ch_bitrate, table;
1227 ch_bitrate = bitrate / nb_channels;
1229 if ((freq == 48000 && ch_bitrate >= 56) ||
1230 (ch_bitrate >= 56 && ch_bitrate <= 80))
1232 else if (freq != 48000 && ch_bitrate >= 96)
1234 else if (freq != 32000 && ch_bitrate <= 48)
1244 static int mp_decode_layer2(MPADecodeContext *s)
1246 int sblimit; /* number of used subbands */
1247 const unsigned char *alloc_table;
1248 int table, bit_alloc_bits, i, j, ch, bound, v;
1249 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1250 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1251 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1252 int scale, qindex, bits, steps, k, l, m, b;
1254 /* select decoding table */
1255 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1256 s->sample_rate, s->lsf);
1257 sblimit = sblimit_table[table];
1258 alloc_table = alloc_tables[table];
1260 if (s->mode == MPA_JSTEREO)
1261 bound = (s->mode_ext + 1) * 4;
1265 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1266 /* parse bit allocation */
1268 for(i=0;i<bound;i++) {
1269 bit_alloc_bits = alloc_table[j];
1270 for(ch=0;ch<s->nb_channels;ch++) {
1271 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1273 j += 1 << bit_alloc_bits;
1275 for(i=bound;i<sblimit;i++) {
1276 bit_alloc_bits = alloc_table[j];
1277 v = get_bits(&s->gb, bit_alloc_bits);
1278 bit_alloc[0][i] = v;
1279 bit_alloc[1][i] = v;
1280 j += 1 << bit_alloc_bits;
1285 for(ch=0;ch<s->nb_channels;ch++) {
1286 for(i=0;i<sblimit;i++)
1287 printf(" %d", bit_alloc[ch][i]);
1294 for(i=0;i<sblimit;i++) {
1295 for(ch=0;ch<s->nb_channels;ch++) {
1296 if (bit_alloc[ch][i])
1297 scale_code[ch][i] = get_bits(&s->gb, 2);
1302 for(i=0;i<sblimit;i++) {
1303 for(ch=0;ch<s->nb_channels;ch++) {
1304 if (bit_alloc[ch][i]) {
1305 sf = scale_factors[ch][i];
1306 switch(scale_code[ch][i]) {
1309 sf[0] = get_bits(&s->gb, 6);
1310 sf[1] = get_bits(&s->gb, 6);
1311 sf[2] = get_bits(&s->gb, 6);
1314 sf[0] = get_bits(&s->gb, 6);
1319 sf[0] = get_bits(&s->gb, 6);
1320 sf[2] = get_bits(&s->gb, 6);
1324 sf[0] = get_bits(&s->gb, 6);
1325 sf[2] = get_bits(&s->gb, 6);
1334 for(ch=0;ch<s->nb_channels;ch++) {
1335 for(i=0;i<sblimit;i++) {
1336 if (bit_alloc[ch][i]) {
1337 sf = scale_factors[ch][i];
1338 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1349 for(l=0;l<12;l+=3) {
1351 for(i=0;i<bound;i++) {
1352 bit_alloc_bits = alloc_table[j];
1353 for(ch=0;ch<s->nb_channels;ch++) {
1354 b = bit_alloc[ch][i];
1356 scale = scale_factors[ch][i][k];
1357 qindex = alloc_table[j+b];
1358 bits = quant_bits[qindex];
1360 /* 3 values at the same time */
1361 v = get_bits(&s->gb, -bits);
1362 steps = quant_steps[qindex];
1363 s->sb_samples[ch][k * 12 + l + 0][i] =
1364 l2_unscale_group(steps, v % steps, scale);
1366 s->sb_samples[ch][k * 12 + l + 1][i] =
1367 l2_unscale_group(steps, v % steps, scale);
1369 s->sb_samples[ch][k * 12 + l + 2][i] =
1370 l2_unscale_group(steps, v, scale);
1373 v = get_bits(&s->gb, bits);
1374 v = l1_unscale(bits - 1, v, scale);
1375 s->sb_samples[ch][k * 12 + l + m][i] = v;
1379 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1380 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1381 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1384 /* next subband in alloc table */
1385 j += 1 << bit_alloc_bits;
1387 /* XXX: find a way to avoid this duplication of code */
1388 for(i=bound;i<sblimit;i++) {
1389 bit_alloc_bits = alloc_table[j];
1390 b = bit_alloc[0][i];
1392 int mant, scale0, scale1;
1393 scale0 = scale_factors[0][i][k];
1394 scale1 = scale_factors[1][i][k];
1395 qindex = alloc_table[j+b];
1396 bits = quant_bits[qindex];
1398 /* 3 values at the same time */
1399 v = get_bits(&s->gb, -bits);
1400 steps = quant_steps[qindex];
1403 s->sb_samples[0][k * 12 + l + 0][i] =
1404 l2_unscale_group(steps, mant, scale0);
1405 s->sb_samples[1][k * 12 + l + 0][i] =
1406 l2_unscale_group(steps, mant, scale1);
1409 s->sb_samples[0][k * 12 + l + 1][i] =
1410 l2_unscale_group(steps, mant, scale0);
1411 s->sb_samples[1][k * 12 + l + 1][i] =
1412 l2_unscale_group(steps, mant, scale1);
1413 s->sb_samples[0][k * 12 + l + 2][i] =
1414 l2_unscale_group(steps, v, scale0);
1415 s->sb_samples[1][k * 12 + l + 2][i] =
1416 l2_unscale_group(steps, v, scale1);
1419 mant = get_bits(&s->gb, bits);
1420 s->sb_samples[0][k * 12 + l + m][i] =
1421 l1_unscale(bits - 1, mant, scale0);
1422 s->sb_samples[1][k * 12 + l + m][i] =
1423 l1_unscale(bits - 1, mant, scale1);
1427 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1428 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1429 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1430 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1431 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1432 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1434 /* next subband in alloc table */
1435 j += 1 << bit_alloc_bits;
1437 /* fill remaining samples to zero */
1438 for(i=sblimit;i<SBLIMIT;i++) {
1439 for(ch=0;ch<s->nb_channels;ch++) {
1440 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1441 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1442 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1451 * Seek back in the stream for backstep bytes (at most 511 bytes)
1453 static void seek_to_maindata(MPADecodeContext *s, long backstep)
1457 /* compute current position in stream */
1458 #ifdef ALT_BITSTREAM_READER
1459 ptr = s->gb.buffer + (s->gb.index>>3);
1461 ptr = s->gb.buf_ptr - (s->gb.bit_cnt >> 3);
1463 /* copy old data before current one */
1465 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1466 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1467 /* init get bits again */
1468 init_get_bits(&s->gb, ptr, s->frame_size + backstep);
1470 /* prepare next buffer */
1471 s->inbuf_index ^= 1;
1472 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1473 s->old_frame_size = s->frame_size;
1476 static inline void lsf_sf_expand(int *slen,
1477 int sf, int n1, int n2, int n3)
1496 static void exponents_from_scale_factors(MPADecodeContext *s,
1500 const UINT8 *bstab, *pretab;
1501 int len, i, j, k, l, v0, shift, gain, gains[3];
1504 exp_ptr = exponents;
1505 gain = g->global_gain - 210;
1506 shift = g->scalefac_scale + 1;
1508 bstab = band_size_long[s->sample_rate_index];
1509 pretab = mpa_pretab[g->preflag];
1510 for(i=0;i<g->long_end;i++) {
1511 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1517 if (g->short_start < 13) {
1518 bstab = band_size_short[s->sample_rate_index];
1519 gains[0] = gain - (g->subblock_gain[0] << 3);
1520 gains[1] = gain - (g->subblock_gain[1] << 3);
1521 gains[2] = gain - (g->subblock_gain[2] << 3);
1523 for(i=g->short_start;i<13;i++) {
1526 v0 = gains[l] - (g->scale_factors[k++] << shift);
1534 /* handle n = 0 too */
1535 static inline int get_bitsz(GetBitContext *s, int n)
1540 return get_bits(s, n);
1543 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1544 INT16 *exponents, int end_pos)
1547 int linbits, code, x, y, l, v, i, j, k, pos;
1548 UINT8 *last_buf_ptr;
1549 UINT32 last_bit_buf;
1554 /* low frequencies (called big values) */
1557 j = g->region_size[i];
1560 /* select vlc table */
1561 k = g->table_select[i];
1562 l = mpa_huff_data[k][0];
1563 linbits = mpa_huff_data[k][1];
1565 code_table = huff_code_table[l];
1567 /* read huffcode and compute each couple */
1569 if (get_bits_count(&s->gb) >= end_pos)
1572 code = get_vlc(&s->gb, vlc);
1575 y = code_table[code];
1582 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1583 i, g->region_size[i] - j, x, y, exponents[s_index]);
1586 x += get_bitsz(&s->gb, linbits);
1587 v = l3_unscale(x, exponents[s_index]);
1588 if (get_bits1(&s->gb))
1593 g->sb_hybrid[s_index++] = v;
1596 y += get_bitsz(&s->gb, linbits);
1597 v = l3_unscale(y, exponents[s_index]);
1598 if (get_bits1(&s->gb))
1603 g->sb_hybrid[s_index++] = v;
1607 /* high frequencies */
1608 vlc = &huff_quad_vlc[g->count1table_select];
1609 last_buf_ptr = NULL;
1612 while (s_index <= 572) {
1613 pos = get_bits_count(&s->gb);
1614 if (pos >= end_pos) {
1615 if (pos > end_pos && last_buf_ptr != NULL) {
1616 /* some encoders generate an incorrect size for this
1617 part. We must go back into the data */
1619 #ifdef ALT_BITSTREAM_READER
1620 s->gb.buffer = last_buf_ptr;
1621 s->gb.index = last_bit_cnt;
1623 s->gb.buf_ptr = last_buf_ptr;
1624 s->gb.bit_buf = last_bit_buf;
1625 s->gb.bit_cnt = last_bit_cnt;
1630 #ifdef ALT_BITSTREAM_READER
1631 last_buf_ptr = s->gb.buffer;
1632 last_bit_cnt = s->gb.index;
1634 last_buf_ptr = s->gb.buf_ptr;
1635 last_bit_buf = s->gb.bit_buf;
1636 last_bit_cnt = s->gb.bit_cnt;
1639 code = get_vlc(&s->gb, vlc);
1640 dprintf("t=%d code=%d\n", g->count1table_select, code);
1644 if (code & (8 >> i)) {
1645 /* non zero value. Could use a hand coded function for
1647 v = l3_unscale(1, exponents[s_index]);
1648 if(get_bits1(&s->gb))
1653 g->sb_hybrid[s_index++] = v;
1656 while (s_index < 576)
1657 g->sb_hybrid[s_index++] = 0;
1661 /* Reorder short blocks from bitstream order to interleaved order. It
1662 would be faster to do it in parsing, but the code would be far more
1664 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1667 INT32 *ptr, *dst, *ptr1;
1670 if (g->block_type != 2)
1673 if (g->switch_point) {
1674 if (s->sample_rate_index != 8) {
1675 ptr = g->sb_hybrid + 36;
1677 ptr = g->sb_hybrid + 48;
1683 for(i=g->short_start;i<13;i++) {
1684 len = band_size_short[s->sample_rate_index][i];
1688 for(j=len;j>0;j--) {
1693 memcpy(ptr1, tmp, len * 3 * sizeof(INT32));
1697 #define ISQRT2 FIXR(0.70710678118654752440)
1699 static void compute_stereo(MPADecodeContext *s,
1700 GranuleDef *g0, GranuleDef *g1)
1704 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1705 INT32 (*is_tab)[16];
1707 int non_zero_found_short[3];
1709 /* intensity stereo */
1710 if (s->mode_ext & MODE_EXT_I_STEREO) {
1715 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1719 tab0 = g0->sb_hybrid + 576;
1720 tab1 = g1->sb_hybrid + 576;
1722 non_zero_found_short[0] = 0;
1723 non_zero_found_short[1] = 0;
1724 non_zero_found_short[2] = 0;
1725 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1726 for(i = 12;i >= g1->short_start;i--) {
1727 /* for last band, use previous scale factor */
1730 len = band_size_short[s->sample_rate_index][i];
1734 if (!non_zero_found_short[l]) {
1735 /* test if non zero band. if so, stop doing i-stereo */
1736 for(j=0;j<len;j++) {
1738 non_zero_found_short[l] = 1;
1742 sf = g1->scale_factors[k + l];
1748 for(j=0;j<len;j++) {
1750 tab0[j] = MULL(tmp0, v1);
1751 tab1[j] = MULL(tmp0, v2);
1755 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1756 /* lower part of the spectrum : do ms stereo
1758 for(j=0;j<len;j++) {
1761 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1762 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1769 non_zero_found = non_zero_found_short[0] |
1770 non_zero_found_short[1] |
1771 non_zero_found_short[2];
1773 for(i = g1->long_end - 1;i >= 0;i--) {
1774 len = band_size_long[s->sample_rate_index][i];
1777 /* test if non zero band. if so, stop doing i-stereo */
1778 if (!non_zero_found) {
1779 for(j=0;j<len;j++) {
1785 /* for last band, use previous scale factor */
1786 k = (i == 21) ? 20 : i;
1787 sf = g1->scale_factors[k];
1792 for(j=0;j<len;j++) {
1794 tab0[j] = MULL(tmp0, v1);
1795 tab1[j] = MULL(tmp0, v2);
1799 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1800 /* lower part of the spectrum : do ms stereo
1802 for(j=0;j<len;j++) {
1805 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1806 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1811 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1812 /* ms stereo ONLY */
1813 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1815 tab0 = g0->sb_hybrid;
1816 tab1 = g1->sb_hybrid;
1817 for(i=0;i<576;i++) {
1820 tab0[i] = tmp0 + tmp1;
1821 tab1[i] = tmp0 - tmp1;
1826 static void compute_antialias(MPADecodeContext *s,
1829 INT32 *ptr, *p0, *p1, *csa;
1830 int n, tmp0, tmp1, i, j;
1832 /* we antialias only "long" bands */
1833 if (g->block_type == 2) {
1834 if (!g->switch_point)
1836 /* XXX: check this for 8000Hz case */
1842 ptr = g->sb_hybrid + 18;
1843 for(i = n;i > 0;i--) {
1846 csa = &csa_table[0][0];
1850 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1851 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1860 static void compute_imdct(MPADecodeContext *s,
1865 INT32 *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
1869 int i, j, k, mdct_long_end, v, sblimit;
1871 /* find last non zero block */
1872 ptr = g->sb_hybrid + 576;
1873 ptr1 = g->sb_hybrid + 2 * 18;
1874 while (ptr >= ptr1) {
1876 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
1880 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
1882 if (g->block_type == 2) {
1883 /* XXX: check for 8000 Hz */
1884 if (g->switch_point)
1889 mdct_long_end = sblimit;
1894 for(j=0;j<mdct_long_end;j++) {
1896 /* apply window & overlap with previous buffer */
1897 out_ptr = sb_samples + j;
1899 if (g->switch_point && j < 2)
1902 win1 = mdct_win[g->block_type];
1903 /* select frequency inversion */
1904 win = win1 + ((4 * 36) & -(j & 1));
1906 *out_ptr = MULL(out[i], win[i]) + buf[i];
1907 buf[i] = MULL(out[i + 18], win[i + 18]);
1913 for(j=mdct_long_end;j<sblimit;j++) {
1919 /* select frequency inversion */
1920 win = mdct_win[2] + ((4 * 36) & -(j & 1));
1923 /* reorder input for short mdct */
1930 /* apply 12 point window and do small overlap */
1932 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
1933 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
1938 out_ptr = sb_samples + j;
1940 *out_ptr = out[i] + buf[i];
1941 buf[i] = out[i + 18];
1948 for(j=sblimit;j<SBLIMIT;j++) {
1950 out_ptr = sb_samples + j;
1961 void sample_dump(int fnum, INT32 *tab, int n)
1963 static FILE *files[16], *f;
1968 sprintf(buf, "/tmp/out%d.pcm", fnum);
1969 f = fopen(buf, "w");
1978 printf("pos=%d\n", pos);
1980 printf(" %f", (double)tab[i] / 32768.0);
1987 fwrite(tab, 1, n * sizeof(INT32), f);
1992 /* main layer3 decoding function */
1993 static int mp_decode_layer3(MPADecodeContext *s)
1995 int nb_granules, main_data_begin, private_bits;
1996 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
1997 GranuleDef granules[2][2], *g;
1998 INT16 exponents[576];
2000 /* read side info */
2002 main_data_begin = get_bits(&s->gb, 8);
2003 if (s->nb_channels == 2)
2004 private_bits = get_bits(&s->gb, 2);
2006 private_bits = get_bits(&s->gb, 1);
2009 main_data_begin = get_bits(&s->gb, 9);
2010 if (s->nb_channels == 2)
2011 private_bits = get_bits(&s->gb, 3);
2013 private_bits = get_bits(&s->gb, 5);
2015 for(ch=0;ch<s->nb_channels;ch++) {
2016 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2017 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2021 for(gr=0;gr<nb_granules;gr++) {
2022 for(ch=0;ch<s->nb_channels;ch++) {
2023 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2024 g = &granules[ch][gr];
2025 g->part2_3_length = get_bits(&s->gb, 12);
2026 g->big_values = get_bits(&s->gb, 9);
2027 g->global_gain = get_bits(&s->gb, 8);
2028 /* if MS stereo only is selected, we precompute the
2029 1/sqrt(2) renormalization factor */
2030 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2032 g->global_gain -= 2;
2034 g->scalefac_compress = get_bits(&s->gb, 9);
2036 g->scalefac_compress = get_bits(&s->gb, 4);
2037 blocksplit_flag = get_bits(&s->gb, 1);
2038 if (blocksplit_flag) {
2039 g->block_type = get_bits(&s->gb, 2);
2040 if (g->block_type == 0)
2042 g->switch_point = get_bits(&s->gb, 1);
2044 g->table_select[i] = get_bits(&s->gb, 5);
2046 g->subblock_gain[i] = get_bits(&s->gb, 3);
2047 /* compute huffman coded region sizes */
2048 if (g->block_type == 2)
2049 g->region_size[0] = (36 / 2);
2051 if (s->sample_rate_index <= 2)
2052 g->region_size[0] = (36 / 2);
2053 else if (s->sample_rate_index != 8)
2054 g->region_size[0] = (54 / 2);
2056 g->region_size[0] = (108 / 2);
2058 g->region_size[1] = (576 / 2);
2060 int region_address1, region_address2, l;
2062 g->switch_point = 0;
2064 g->table_select[i] = get_bits(&s->gb, 5);
2065 /* compute huffman coded region sizes */
2066 region_address1 = get_bits(&s->gb, 4);
2067 region_address2 = get_bits(&s->gb, 3);
2068 dprintf("region1=%d region2=%d\n",
2069 region_address1, region_address2);
2071 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2072 l = region_address1 + region_address2 + 2;
2073 /* should not overflow */
2077 band_index_long[s->sample_rate_index][l] >> 1;
2079 /* convert region offsets to region sizes and truncate
2080 size to big_values */
2081 g->region_size[2] = (576 / 2);
2084 k = g->region_size[i];
2085 if (k > g->big_values)
2087 g->region_size[i] = k - j;
2091 /* compute band indexes */
2092 if (g->block_type == 2) {
2093 if (g->switch_point) {
2094 /* if switched mode, we handle the 36 first samples as
2095 long blocks. For 8000Hz, we handle the 48 first
2096 exponents as long blocks (XXX: check this!) */
2097 if (s->sample_rate_index <= 2)
2099 else if (s->sample_rate_index != 8)
2102 g->long_end = 4; /* 8000 Hz */
2104 if (s->sample_rate_index != 8)
2113 g->short_start = 13;
2119 g->preflag = get_bits(&s->gb, 1);
2120 g->scalefac_scale = get_bits(&s->gb, 1);
2121 g->count1table_select = get_bits(&s->gb, 1);
2122 dprintf("block_type=%d switch_point=%d\n",
2123 g->block_type, g->switch_point);
2127 /* now we get bits from the main_data_begin offset */
2128 dprintf("seekback: %d\n", main_data_begin);
2129 seek_to_maindata(s, main_data_begin);
2131 for(gr=0;gr<nb_granules;gr++) {
2132 for(ch=0;ch<s->nb_channels;ch++) {
2133 g = &granules[ch][gr];
2135 bits_pos = get_bits_count(&s->gb);
2139 int slen, slen1, slen2;
2141 /* MPEG1 scale factors */
2142 slen1 = slen_table[0][g->scalefac_compress];
2143 slen2 = slen_table[1][g->scalefac_compress];
2144 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2145 if (g->block_type == 2) {
2146 n = g->switch_point ? 17 : 18;
2149 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2151 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2153 g->scale_factors[j++] = 0;
2155 sc = granules[ch][0].scale_factors;
2158 n = (k == 0 ? 6 : 5);
2159 if ((g->scfsi & (0x8 >> k)) == 0) {
2160 slen = (k < 2) ? slen1 : slen2;
2162 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2164 /* simply copy from last granule */
2166 g->scale_factors[j] = sc[j];
2171 g->scale_factors[j++] = 0;
2175 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2178 printf(" %d", g->scale_factors[i]);
2183 int tindex, tindex2, slen[4], sl, sf;
2185 /* LSF scale factors */
2186 if (g->block_type == 2) {
2187 tindex = g->switch_point ? 2 : 1;
2191 sf = g->scalefac_compress;
2192 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2193 /* intensity stereo case */
2196 lsf_sf_expand(slen, sf, 6, 6, 0);
2198 } else if (sf < 244) {
2199 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2202 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2208 lsf_sf_expand(slen, sf, 5, 4, 4);
2210 } else if (sf < 500) {
2211 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2214 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2222 n = lsf_nsf_table[tindex2][tindex][k];
2225 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2227 /* XXX: should compute exact size */
2229 g->scale_factors[j] = 0;
2232 printf("gr=%d ch=%d scale_factors:\n",
2235 printf(" %d", g->scale_factors[i]);
2241 exponents_from_scale_factors(s, g, exponents);
2243 /* read Huffman coded residue */
2244 if (huffman_decode(s, g, exponents,
2245 bits_pos + g->part2_3_length) < 0)
2248 sample_dump(0, g->sb_hybrid, 576);
2251 /* skip extension bits */
2252 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2253 if (bits_left < 0) {
2254 dprintf("bits_left=%d\n", bits_left);
2257 while (bits_left >= 16) {
2258 skip_bits(&s->gb, 16);
2262 skip_bits(&s->gb, bits_left);
2265 if (s->nb_channels == 2)
2266 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2268 for(ch=0;ch<s->nb_channels;ch++) {
2269 g = &granules[ch][gr];
2271 reorder_block(s, g);
2273 sample_dump(0, g->sb_hybrid, 576);
2275 compute_antialias(s, g);
2277 sample_dump(1, g->sb_hybrid, 576);
2279 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2281 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2285 return nb_granules * 18;
2288 static int mp_decode_frame(MPADecodeContext *s,
2291 int i, nb_frames, ch;
2294 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2295 s->inbuf_ptr - s->inbuf - HEADER_SIZE);
2297 /* skip error protection field */
2298 if (s->error_protection)
2299 get_bits(&s->gb, 16);
2301 dprintf("frame %d:\n", s->frame_count);
2304 nb_frames = mp_decode_layer1(s);
2307 nb_frames = mp_decode_layer2(s);
2311 nb_frames = mp_decode_layer3(s);
2315 for(i=0;i<nb_frames;i++) {
2316 for(ch=0;ch<s->nb_channels;ch++) {
2318 printf("%d-%d:", i, ch);
2319 for(j=0;j<SBLIMIT;j++)
2320 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2325 /* apply the synthesis filter */
2326 for(ch=0;ch<s->nb_channels;ch++) {
2327 samples_ptr = samples + ch;
2328 for(i=0;i<nb_frames;i++) {
2329 synth_filter(s, ch, samples_ptr, s->nb_channels,
2330 s->sb_samples[ch][i]);
2331 samples_ptr += 32 * s->nb_channels;
2337 return nb_frames * 32 * sizeof(short) * s->nb_channels;
2340 static int decode_frame(AVCodecContext * avctx,
2341 void *data, int *data_size,
2342 UINT8 * buf, int buf_size)
2344 MPADecodeContext *s = avctx->priv_data;
2348 short *out_samples = data;
2352 while (buf_size > 0) {
2353 len = s->inbuf_ptr - s->inbuf;
2354 if (s->frame_size == 0) {
2355 /* special case for next header for first frame in free
2356 format case (XXX: find a simpler method) */
2357 if (s->free_format_next_header != 0) {
2358 s->inbuf[0] = s->free_format_next_header >> 24;
2359 s->inbuf[1] = s->free_format_next_header >> 16;
2360 s->inbuf[2] = s->free_format_next_header >> 8;
2361 s->inbuf[3] = s->free_format_next_header;
2362 s->inbuf_ptr = s->inbuf + 4;
2363 s->free_format_next_header = 0;
2366 /* no header seen : find one. We need at least HEADER_SIZE
2367 bytes to parse it */
2368 len = HEADER_SIZE - len;
2372 memcpy(s->inbuf_ptr, buf_ptr, len);
2375 s->inbuf_ptr += len;
2377 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2379 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2380 (s->inbuf[2] << 8) | s->inbuf[3];
2382 if (check_header(header) < 0) {
2383 /* no sync found : move by one byte (inefficient, but simple!) */
2384 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2386 dprintf("skip %x\n", header);
2387 /* reset free format frame size to give a chance
2388 to get a new bitrate */
2389 s->free_format_frame_size = 0;
2391 if (decode_header(s, header) == 1) {
2392 /* free format: compute frame size */
2394 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2397 /* update codec info */
2398 avctx->sample_rate = s->sample_rate;
2399 avctx->channels = s->nb_channels;
2400 avctx->bit_rate = s->bit_rate;
2401 avctx->frame_size = s->frame_size;
2405 } else if (s->frame_size == -1) {
2406 /* free format : find next sync to compute frame size */
2407 len = MPA_MAX_CODED_FRAME_SIZE - len;
2411 /* frame too long: resync */
2418 memcpy(s->inbuf_ptr, buf_ptr, len);
2419 /* check for header */
2420 p = s->inbuf_ptr - 3;
2421 pend = s->inbuf_ptr + len - 4;
2423 header = (p[0] << 24) | (p[1] << 16) |
2425 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2426 (s->inbuf[2] << 8) | s->inbuf[3];
2427 /* check with high probability that we have a
2429 if ((header & SAME_HEADER_MASK) ==
2430 (header1 & SAME_HEADER_MASK)) {
2431 /* header found: update pointers */
2432 len = (p + 4) - s->inbuf_ptr;
2436 /* compute frame size */
2437 s->free_format_next_header = header;
2438 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2439 padding = (header1 >> 9) & 1;
2441 s->free_format_frame_size -= padding * 4;
2443 s->free_format_frame_size -= padding;
2444 dprintf("free frame size=%d padding=%d\n",
2445 s->free_format_frame_size, padding);
2446 decode_header(s, header1);
2451 /* not found: simply increase pointers */
2453 s->inbuf_ptr += len;
2456 } else if (len < s->frame_size) {
2457 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2458 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2459 len = s->frame_size - len;
2463 len = buf_size > 4 ? 4 : buf_size;
2464 memcpy(s->inbuf_ptr, buf_ptr, len);
2466 s->inbuf_ptr += len;
2469 out_size = mp_decode_frame(s, out_samples);
2470 s->inbuf_ptr = s->inbuf;
2472 *data_size = out_size;
2477 return buf_ptr - buf;
2480 AVCodec mp2_decoder =
2485 sizeof(MPADecodeContext),
2492 AVCodec mp3_decoder =
2497 sizeof(MPADecodeContext),