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 * @file mpegaudiodec.c
27 #include "mpegaudio.h"
31 * - in low precision mode, use more 16 bit multiplies in synth filter
32 * - test lsf / mpeg25 extensively.
35 /* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
37 #ifdef CONFIG_MPEGAUDIO_HP
38 #define USE_HIGHPRECISION
41 #ifdef USE_HIGHPRECISION
42 #define FRAC_BITS 23 /* fractional bits for sb_samples and dct */
43 #define WFRAC_BITS 16 /* fractional bits for window */
45 #define FRAC_BITS 15 /* fractional bits for sb_samples and dct */
46 #define WFRAC_BITS 14 /* fractional bits for window */
49 #define FRAC_ONE (1 << FRAC_BITS)
51 #define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
52 #define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
53 #define FIX(a) ((int)((a) * FRAC_ONE))
54 /* WARNING: only correct for posititive numbers */
55 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
56 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
59 typedef int16_t MPA_INT;
61 typedef int32_t MPA_INT;
67 #define BACKSTEP_SIZE 512
69 typedef struct MPADecodeContext {
70 uint8_t inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
72 uint8_t *inbuf_ptr, *inbuf;
74 int free_format_frame_size; /* frame size in case of free format
75 (zero if currently unknown) */
76 /* next header (used in free format parsing) */
77 uint32_t free_format_next_header;
81 int sample_rate_index; /* between 0 and 8 */
89 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
90 int synth_buf_offset[MPA_MAX_CHANNELS];
91 int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
92 int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
98 /* layer 3 "granule" */
99 typedef struct GranuleDef {
104 int scalefac_compress;
106 uint8_t switch_point;
108 int subblock_gain[3];
109 uint8_t scalefac_scale;
110 uint8_t count1table_select;
111 int region_size[3]; /* number of huffman codes in each region */
113 int short_start, long_end; /* long/short band indexes */
114 uint8_t scale_factors[40];
115 int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
118 #define MODE_EXT_MS_STEREO 2
119 #define MODE_EXT_I_STEREO 1
121 /* layer 3 huffman tables */
122 typedef struct HuffTable {
125 const uint16_t *codes;
128 #include "mpegaudiodectab.h"
130 /* vlc structure for decoding layer 3 huffman tables */
131 static VLC huff_vlc[16];
132 static uint8_t *huff_code_table[16];
133 static VLC huff_quad_vlc[2];
134 /* computed from band_size_long */
135 static uint16_t band_index_long[9][23];
136 /* XXX: free when all decoders are closed */
137 #define TABLE_4_3_SIZE (8191 + 16)
138 static int8_t *table_4_3_exp;
140 static uint16_t *table_4_3_value;
142 static uint32_t *table_4_3_value;
144 /* intensity stereo coef table */
145 static int32_t is_table[2][16];
146 static int32_t is_table_lsf[2][2][16];
147 static int32_t csa_table[8][2];
148 static int32_t mdct_win[8][36];
150 /* lower 2 bits: modulo 3, higher bits: shift */
151 static uint16_t scale_factor_modshift[64];
152 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
153 static int32_t scale_factor_mult[15][3];
154 /* mult table for layer 2 group quantization */
156 #define SCALE_GEN(v) \
157 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
159 static int32_t scale_factor_mult2[3][3] = {
160 SCALE_GEN(4.0 / 3.0), /* 3 steps */
161 SCALE_GEN(4.0 / 5.0), /* 5 steps */
162 SCALE_GEN(4.0 / 9.0), /* 9 steps */
166 static uint32_t scale_factor_mult3[4] = {
168 FIXR(1.18920711500272106671),
169 FIXR(1.41421356237309504880),
170 FIXR(1.68179283050742908605),
173 static MPA_INT window[512] __attribute__((aligned(16)));
175 /* layer 1 unscaling */
176 /* n = number of bits of the mantissa minus 1 */
177 static inline int l1_unscale(int n, int mant, int scale_factor)
182 shift = scale_factor_modshift[scale_factor];
185 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
187 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
188 return (int)((val + (1LL << (shift - 1))) >> shift);
191 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
195 shift = scale_factor_modshift[scale_factor];
199 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
200 /* NOTE: at this point, 0 <= shift <= 21 */
202 val = (val + (1 << (shift - 1))) >> shift;
206 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
207 static inline int l3_unscale(int value, int exponent)
216 e = table_4_3_exp[value];
217 e += (exponent >> 2);
223 m = table_4_3_value[value];
225 m = (m * scale_factor_mult3[exponent & 3]);
226 m = (m + (1 << (e-1))) >> e;
229 m = MUL64(m, scale_factor_mult3[exponent & 3]);
230 m = (m + (uint64_t_C(1) << (e-1))) >> e;
235 /* all integer n^(4/3) computation code */
238 #define POW_FRAC_BITS 24
239 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
240 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
241 #define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)
243 static int dev_4_3_coefs[DEV_ORDER];
245 static int pow_mult3[3] = {
247 POW_FIX(1.25992104989487316476),
248 POW_FIX(1.58740105196819947474),
251 static void int_pow_init(void)
256 for(i=0;i<DEV_ORDER;i++) {
257 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
258 dev_4_3_coefs[i] = a;
262 /* return the mantissa and the binary exponent */
263 static int int_pow(int i, int *exp_ptr)
271 while (a < (1 << (POW_FRAC_BITS - 1))) {
275 a -= (1 << POW_FRAC_BITS);
277 for(j = DEV_ORDER - 1; j >= 0; j--)
278 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
279 a = (1 << POW_FRAC_BITS) + a1;
280 /* exponent compute (exact) */
284 a = POW_MULL(a, pow_mult3[er]);
285 while (a >= 2 * POW_FRAC_ONE) {
289 /* convert to float */
290 while (a < POW_FRAC_ONE) {
294 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
295 #if POW_FRAC_BITS > FRAC_BITS
296 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
297 /* correct overflow */
298 if (a >= 2 * (1 << FRAC_BITS)) {
307 static int decode_init(AVCodecContext * avctx)
309 MPADecodeContext *s = avctx->priv_data;
314 /* scale factors table for layer 1/2 */
317 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
320 scale_factor_modshift[i] = mod | (shift << 2);
323 /* scale factor multiply for layer 1 */
327 norm = ((int64_t_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
328 scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
329 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
330 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
331 dprintf("%d: norm=%x s=%x %x %x\n",
333 scale_factor_mult[i][0],
334 scale_factor_mult[i][1],
335 scale_factor_mult[i][2]);
339 /* max = 18760, max sum over all 16 coefs : 44736 */
344 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
353 /* huffman decode tables */
354 huff_code_table[0] = NULL;
356 const HuffTable *h = &mpa_huff_tables[i];
364 init_vlc(&huff_vlc[i], 8, n,
365 h->bits, 1, 1, h->codes, 2, 2);
367 code_table = av_mallocz(n);
369 for(x=0;x<xsize;x++) {
371 code_table[j++] = (x << 4) | y;
373 huff_code_table[i] = code_table;
376 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
377 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1);
383 band_index_long[i][j] = k;
384 k += band_size_long[i][j];
386 band_index_long[i][22] = k;
389 /* compute n ^ (4/3) and store it in mantissa/exp format */
390 if (!av_mallocz_static(&table_4_3_exp,
391 TABLE_4_3_SIZE * sizeof(table_4_3_exp[0])))
393 if (!av_mallocz_static(&table_4_3_value,
394 TABLE_4_3_SIZE * sizeof(table_4_3_value[0])))
398 for(i=1;i<TABLE_4_3_SIZE;i++) {
406 f = pow((double)i, 4.0 / 3.0);
410 if ((unsigned short)m1 != m1) {
416 if (m != m1 || e != e1) {
417 printf("%4d: m=%x m1=%x e=%d e1=%d\n",
422 /* normalized to FRAC_BITS */
423 table_4_3_value[i] = m;
424 table_4_3_exp[i] = e;
431 f = tan((double)i * M_PI / 12.0);
432 v = FIXR(f / (1.0 + f));
437 is_table[1][6 - i] = v;
441 is_table[0][i] = is_table[1][i] = 0.0;
448 e = -(j + 1) * ((i + 1) >> 1);
449 f = pow(2.0, e / 4.0);
451 is_table_lsf[j][k ^ 1][i] = FIXR(f);
452 is_table_lsf[j][k][i] = FIXR(1.0);
453 dprintf("is_table_lsf %d %d: %x %x\n",
454 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
461 cs = 1.0 / sqrt(1.0 + ci * ci);
463 csa_table[i][0] = FIX(cs);
464 csa_table[i][1] = FIX(ca);
467 /* compute mdct windows */
470 v = FIXR(sin(M_PI * (i + 0.5) / 36.0));
476 mdct_win[1][18 + i] = FIXR(1.0);
477 mdct_win[1][24 + i] = FIXR(sin(M_PI * ((i + 6) + 0.5) / 12.0));
478 mdct_win[1][30 + i] = FIXR(0.0);
480 mdct_win[3][i] = FIXR(0.0);
481 mdct_win[3][6 + i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
482 mdct_win[3][12 + i] = FIXR(1.0);
486 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
488 /* NOTE: we do frequency inversion adter the MDCT by changing
489 the sign of the right window coefs */
492 mdct_win[j + 4][i] = mdct_win[j][i];
493 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
499 printf("win%d=\n", j);
501 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
509 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
510 s->inbuf_ptr = s->inbuf;
517 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
521 #define COS0_0 FIXR(0.50060299823519630134)
522 #define COS0_1 FIXR(0.50547095989754365998)
523 #define COS0_2 FIXR(0.51544730992262454697)
524 #define COS0_3 FIXR(0.53104259108978417447)
525 #define COS0_4 FIXR(0.55310389603444452782)
526 #define COS0_5 FIXR(0.58293496820613387367)
527 #define COS0_6 FIXR(0.62250412303566481615)
528 #define COS0_7 FIXR(0.67480834145500574602)
529 #define COS0_8 FIXR(0.74453627100229844977)
530 #define COS0_9 FIXR(0.83934964541552703873)
531 #define COS0_10 FIXR(0.97256823786196069369)
532 #define COS0_11 FIXR(1.16943993343288495515)
533 #define COS0_12 FIXR(1.48416461631416627724)
534 #define COS0_13 FIXR(2.05778100995341155085)
535 #define COS0_14 FIXR(3.40760841846871878570)
536 #define COS0_15 FIXR(10.19000812354805681150)
538 #define COS1_0 FIXR(0.50241928618815570551)
539 #define COS1_1 FIXR(0.52249861493968888062)
540 #define COS1_2 FIXR(0.56694403481635770368)
541 #define COS1_3 FIXR(0.64682178335999012954)
542 #define COS1_4 FIXR(0.78815462345125022473)
543 #define COS1_5 FIXR(1.06067768599034747134)
544 #define COS1_6 FIXR(1.72244709823833392782)
545 #define COS1_7 FIXR(5.10114861868916385802)
547 #define COS2_0 FIXR(0.50979557910415916894)
548 #define COS2_1 FIXR(0.60134488693504528054)
549 #define COS2_2 FIXR(0.89997622313641570463)
550 #define COS2_3 FIXR(2.56291544774150617881)
552 #define COS3_0 FIXR(0.54119610014619698439)
553 #define COS3_1 FIXR(1.30656296487637652785)
555 #define COS4_0 FIXR(0.70710678118654752439)
557 /* butterfly operator */
560 tmp0 = tab[a] + tab[b];\
561 tmp1 = tab[a] - tab[b];\
563 tab[b] = MULL(tmp1, c);\
566 #define BF1(a, b, c, d)\
573 #define BF2(a, b, c, d)\
583 #define ADD(a, b) tab[a] += tab[b]
585 /* DCT32 without 1/sqrt(2) coef zero scaling. */
586 static void dct32(int32_t *out, int32_t *tab)
718 out[ 1] = tab[16] + tab[24];
719 out[17] = tab[17] + tab[25];
720 out[ 9] = tab[18] + tab[26];
721 out[25] = tab[19] + tab[27];
722 out[ 5] = tab[20] + tab[28];
723 out[21] = tab[21] + tab[29];
724 out[13] = tab[22] + tab[30];
725 out[29] = tab[23] + tab[31];
726 out[ 3] = tab[24] + tab[20];
727 out[19] = tab[25] + tab[21];
728 out[11] = tab[26] + tab[22];
729 out[27] = tab[27] + tab[23];
730 out[ 7] = tab[28] + tab[18];
731 out[23] = tab[29] + tab[19];
732 out[15] = tab[30] + tab[17];
736 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
740 #define OUT_SAMPLE(sum)\
743 sum1 = (sum + (1 << (OUT_SHIFT - 1))) >> OUT_SHIFT;\
746 else if (sum1 > 32767)\
752 #define SUM8(off, op) \
754 sum op w[0 * 64 + off] * p[0 * 64];\
755 sum op w[1 * 64 + off] * p[1 * 64];\
756 sum op w[2 * 64 + off] * p[2 * 64];\
757 sum op w[3 * 64 + off] * p[3 * 64];\
758 sum op w[4 * 64 + off] * p[4 * 64];\
759 sum op w[5 * 64 + off] * p[5 * 64];\
760 sum op w[6 * 64 + off] * p[6 * 64];\
761 sum op w[7 * 64 + off] * p[7 * 64];\
766 #define OUT_SAMPLE(sum)\
769 sum1 = (int)((sum + (int64_t_C(1) << (OUT_SHIFT - 1))) >> OUT_SHIFT);\
772 else if (sum1 > 32767)\
778 #define SUM8(off, op) \
780 sum op MUL64(w[0 * 64 + off], p[0 * 64]);\
781 sum op MUL64(w[1 * 64 + off], p[1 * 64]);\
782 sum op MUL64(w[2 * 64 + off], p[2 * 64]);\
783 sum op MUL64(w[3 * 64 + off], p[3 * 64]);\
784 sum op MUL64(w[4 * 64 + off], p[4 * 64]);\
785 sum op MUL64(w[5 * 64 + off], p[5 * 64]);\
786 sum op MUL64(w[6 * 64 + off], p[6 * 64]);\
787 sum op MUL64(w[7 * 64 + off], p[7 * 64]);\
792 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
794 /* XXX: optimize by avoiding ring buffer usage */
795 static void synth_filter(MPADecodeContext *s1,
796 int ch, int16_t *samples, int incr,
797 int32_t sb_samples[SBLIMIT])
800 register MPA_INT *synth_buf, *p;
809 dct32(tmp, sb_samples);
811 offset = s1->synth_buf_offset[ch];
812 synth_buf = s1->synth_buf[ch] + offset;
817 /* NOTE: can cause a loss in precision if very high amplitude
826 /* copy to avoid wrap */
827 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
832 p = synth_buf + 16 + j; /* 0-15 */
834 p = synth_buf + 48 - j; /* 32-47 */
840 p = synth_buf + 32; /* 48 */
848 p = synth_buf + 48 - j; /* 17-31 */
850 p = synth_buf + 16 + j; /* 49-63 */
855 offset = (offset - 32) & 511;
856 s1->synth_buf_offset[ch] = offset;
860 #define C1 FIXR(0.99144486137381041114)
861 #define C3 FIXR(0.92387953251128675612)
862 #define C5 FIXR(0.79335334029123516458)
863 #define C7 FIXR(0.60876142900872063941)
864 #define C9 FIXR(0.38268343236508977173)
865 #define C11 FIXR(0.13052619222005159154)
867 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
869 static void imdct12(int *out, int *in)
872 int64_t in1_3, in1_9, in4_3, in4_9;
874 in1_3 = MUL64(in[1], C3);
875 in1_9 = MUL64(in[1], C9);
876 in4_3 = MUL64(in[4], C3);
877 in4_9 = MUL64(in[4], C9);
879 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
880 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
883 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
884 MUL64(in[2] + in[5], C3) - in4_9);
887 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
888 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
891 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
892 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
895 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
896 MUL64(in[2] + in[5], C9) + in4_3);
899 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
900 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
913 #define C1 FIXR(0.98480775301220805936)
914 #define C2 FIXR(0.93969262078590838405)
915 #define C3 FIXR(0.86602540378443864676)
916 #define C4 FIXR(0.76604444311897803520)
917 #define C5 FIXR(0.64278760968653932632)
919 #define C7 FIXR(0.34202014332566873304)
920 #define C8 FIXR(0.17364817766693034885)
922 /* 0.5 / cos(pi*(2*i+1)/36) */
923 static const int icos36[9] = {
924 FIXR(0.50190991877167369479),
925 FIXR(0.51763809020504152469),
926 FIXR(0.55168895948124587824),
927 FIXR(0.61038729438072803416),
928 FIXR(0.70710678118654752439),
929 FIXR(0.87172339781054900991),
930 FIXR(1.18310079157624925896),
931 FIXR(1.93185165257813657349),
932 FIXR(5.73685662283492756461),
935 static const int icos72[18] = {
936 /* 0.5 / cos(pi*(2*i+19)/72) */
937 FIXR(0.74009361646113053152),
938 FIXR(0.82133981585229078570),
939 FIXR(0.93057949835178895673),
940 FIXR(1.08284028510010010928),
941 FIXR(1.30656296487637652785),
942 FIXR(1.66275476171152078719),
943 FIXR(2.31011315767264929558),
944 FIXR(3.83064878777019433457),
945 FIXR(11.46279281302667383546),
947 /* 0.5 / cos(pi*(2*(i + 18) +19)/72) */
948 FIXR(-0.67817085245462840086),
949 FIXR(-0.63023620700513223342),
950 FIXR(-0.59284452371708034528),
951 FIXR(-0.56369097343317117734),
952 FIXR(-0.54119610014619698439),
953 FIXR(-0.52426456257040533932),
954 FIXR(-0.51213975715725461845),
955 FIXR(-0.50431448029007636036),
956 FIXR(-0.50047634258165998492),
959 /* using Lee like decomposition followed by hand coded 9 points DCT */
960 static void imdct36(int *out, int *in)
962 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
963 int tmp[18], *tmp1, *in1;
964 int64_t in3_3, in6_6;
975 in3_3 = MUL64(in1[2*3], C3);
976 in6_6 = MUL64(in1[2*6], C6);
978 tmp1[0] = FRAC_RND(MUL64(in1[2*1], C1) + in3_3 +
979 MUL64(in1[2*5], C5) + MUL64(in1[2*7], C7));
980 tmp1[2] = in1[2*0] + FRAC_RND(MUL64(in1[2*2], C2) +
981 MUL64(in1[2*4], C4) + in6_6 +
982 MUL64(in1[2*8], C8));
983 tmp1[4] = FRAC_RND(MUL64(in1[2*1] - in1[2*5] - in1[2*7], C3));
984 tmp1[6] = FRAC_RND(MUL64(in1[2*2] - in1[2*4] - in1[2*8], C6)) -
986 tmp1[8] = FRAC_RND(MUL64(in1[2*1], C5) - in3_3 -
987 MUL64(in1[2*5], C7) + MUL64(in1[2*7], C1));
988 tmp1[10] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C8) -
989 MUL64(in1[2*4], C2) + in6_6 +
990 MUL64(in1[2*8], C4));
991 tmp1[12] = FRAC_RND(MUL64(in1[2*1], C7) - in3_3 +
992 MUL64(in1[2*5], C1) -
993 MUL64(in1[2*7], C5));
994 tmp1[14] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C4) +
995 MUL64(in1[2*4], C8) + in6_6 -
996 MUL64(in1[2*8], C2));
997 tmp1[16] = in1[2*0] - in1[2*2] + in1[2*4] - in1[2*6] + in1[2*8];
1009 s1 = MULL(t3 + t2, icos36[j]);
1010 s3 = MULL(t3 - t2, icos36[8 - j]);
1012 t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
1013 t1 = MULL(s0 - s1, icos72[8 - j]);
1014 out[18 + 9 + j] = t0;
1015 out[18 + 8 - j] = t0;
1019 t0 = MULL(s2 + s3, icos72[9+j]);
1020 t1 = MULL(s2 - s3, icos72[j]);
1021 out[18 + 9 + (8 - j)] = t0;
1023 out[9 + (8 - j)] = -t1;
1029 s1 = MULL(tmp[17], icos36[4]);
1030 t0 = MULL(s0 + s1, icos72[9 + 4]);
1031 t1 = MULL(s0 - s1, icos72[4]);
1032 out[18 + 9 + 4] = t0;
1033 out[18 + 8 - 4] = t0;
1038 /* fast header check for resync */
1039 static int check_header(uint32_t header)
1042 if ((header & 0xffe00000) != 0xffe00000)
1045 if (((header >> 17) & 3) == 0)
1048 if (((header >> 12) & 0xf) == 0xf)
1051 if (((header >> 10) & 3) == 3)
1056 /* header + layer + bitrate + freq + lsf/mpeg25 */
1057 #define SAME_HEADER_MASK \
1058 (0xffe00000 | (3 << 17) | (0xf << 12) | (3 << 10) | (3 << 19))
1060 /* header decoding. MUST check the header before because no
1061 consistency check is done there. Return 1 if free format found and
1062 that the frame size must be computed externally */
1063 static int decode_header(MPADecodeContext *s, uint32_t header)
1065 int sample_rate, frame_size, mpeg25, padding;
1066 int sample_rate_index, bitrate_index;
1067 if (header & (1<<20)) {
1068 s->lsf = (header & (1<<19)) ? 0 : 1;
1075 s->layer = 4 - ((header >> 17) & 3);
1076 /* extract frequency */
1077 sample_rate_index = (header >> 10) & 3;
1078 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1079 sample_rate_index += 3 * (s->lsf + mpeg25);
1080 s->sample_rate_index = sample_rate_index;
1081 s->error_protection = ((header >> 16) & 1) ^ 1;
1082 s->sample_rate = sample_rate;
1084 bitrate_index = (header >> 12) & 0xf;
1085 padding = (header >> 9) & 1;
1086 //extension = (header >> 8) & 1;
1087 s->mode = (header >> 6) & 3;
1088 s->mode_ext = (header >> 4) & 3;
1089 //copyright = (header >> 3) & 1;
1090 //original = (header >> 2) & 1;
1091 //emphasis = header & 3;
1093 if (s->mode == MPA_MONO)
1098 if (bitrate_index != 0) {
1099 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1100 s->bit_rate = frame_size * 1000;
1103 frame_size = (frame_size * 12000) / sample_rate;
1104 frame_size = (frame_size + padding) * 4;
1107 frame_size = (frame_size * 144000) / sample_rate;
1108 frame_size += padding;
1112 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1113 frame_size += padding;
1116 s->frame_size = frame_size;
1118 /* if no frame size computed, signal it */
1119 if (!s->free_format_frame_size)
1121 /* free format: compute bitrate and real frame size from the
1122 frame size we extracted by reading the bitstream */
1123 s->frame_size = s->free_format_frame_size;
1126 s->frame_size += padding * 4;
1127 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1130 s->frame_size += padding;
1131 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1135 s->frame_size += padding;
1136 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1142 printf("layer%d, %d Hz, %d kbits/s, ",
1143 s->layer, s->sample_rate, s->bit_rate);
1144 if (s->nb_channels == 2) {
1145 if (s->layer == 3) {
1146 if (s->mode_ext & MODE_EXT_MS_STEREO)
1148 if (s->mode_ext & MODE_EXT_I_STEREO)
1160 /* return the number of decoded frames */
1161 static int mp_decode_layer1(MPADecodeContext *s)
1163 int bound, i, v, n, ch, j, mant;
1164 uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
1165 uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1167 if (s->mode == MPA_JSTEREO)
1168 bound = (s->mode_ext + 1) * 4;
1172 /* allocation bits */
1173 for(i=0;i<bound;i++) {
1174 for(ch=0;ch<s->nb_channels;ch++) {
1175 allocation[ch][i] = get_bits(&s->gb, 4);
1178 for(i=bound;i<SBLIMIT;i++) {
1179 allocation[0][i] = get_bits(&s->gb, 4);
1183 for(i=0;i<bound;i++) {
1184 for(ch=0;ch<s->nb_channels;ch++) {
1185 if (allocation[ch][i])
1186 scale_factors[ch][i] = get_bits(&s->gb, 6);
1189 for(i=bound;i<SBLIMIT;i++) {
1190 if (allocation[0][i]) {
1191 scale_factors[0][i] = get_bits(&s->gb, 6);
1192 scale_factors[1][i] = get_bits(&s->gb, 6);
1196 /* compute samples */
1198 for(i=0;i<bound;i++) {
1199 for(ch=0;ch<s->nb_channels;ch++) {
1200 n = allocation[ch][i];
1202 mant = get_bits(&s->gb, n + 1);
1203 v = l1_unscale(n, mant, scale_factors[ch][i]);
1207 s->sb_samples[ch][j][i] = v;
1210 for(i=bound;i<SBLIMIT;i++) {
1211 n = allocation[0][i];
1213 mant = get_bits(&s->gb, n + 1);
1214 v = l1_unscale(n, mant, scale_factors[0][i]);
1215 s->sb_samples[0][j][i] = v;
1216 v = l1_unscale(n, mant, scale_factors[1][i]);
1217 s->sb_samples[1][j][i] = v;
1219 s->sb_samples[0][j][i] = 0;
1220 s->sb_samples[1][j][i] = 0;
1227 /* bitrate is in kb/s */
1228 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1230 int ch_bitrate, table;
1232 ch_bitrate = bitrate / nb_channels;
1234 if ((freq == 48000 && ch_bitrate >= 56) ||
1235 (ch_bitrate >= 56 && ch_bitrate <= 80))
1237 else if (freq != 48000 && ch_bitrate >= 96)
1239 else if (freq != 32000 && ch_bitrate <= 48)
1249 static int mp_decode_layer2(MPADecodeContext *s)
1251 int sblimit; /* number of used subbands */
1252 const unsigned char *alloc_table;
1253 int table, bit_alloc_bits, i, j, ch, bound, v;
1254 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1255 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1256 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1257 int scale, qindex, bits, steps, k, l, m, b;
1259 /* select decoding table */
1260 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1261 s->sample_rate, s->lsf);
1262 sblimit = sblimit_table[table];
1263 alloc_table = alloc_tables[table];
1265 if (s->mode == MPA_JSTEREO)
1266 bound = (s->mode_ext + 1) * 4;
1270 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1271 /* parse bit allocation */
1273 for(i=0;i<bound;i++) {
1274 bit_alloc_bits = alloc_table[j];
1275 for(ch=0;ch<s->nb_channels;ch++) {
1276 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1278 j += 1 << bit_alloc_bits;
1280 for(i=bound;i<sblimit;i++) {
1281 bit_alloc_bits = alloc_table[j];
1282 v = get_bits(&s->gb, bit_alloc_bits);
1283 bit_alloc[0][i] = v;
1284 bit_alloc[1][i] = v;
1285 j += 1 << bit_alloc_bits;
1290 for(ch=0;ch<s->nb_channels;ch++) {
1291 for(i=0;i<sblimit;i++)
1292 printf(" %d", bit_alloc[ch][i]);
1299 for(i=0;i<sblimit;i++) {
1300 for(ch=0;ch<s->nb_channels;ch++) {
1301 if (bit_alloc[ch][i])
1302 scale_code[ch][i] = get_bits(&s->gb, 2);
1307 for(i=0;i<sblimit;i++) {
1308 for(ch=0;ch<s->nb_channels;ch++) {
1309 if (bit_alloc[ch][i]) {
1310 sf = scale_factors[ch][i];
1311 switch(scale_code[ch][i]) {
1314 sf[0] = get_bits(&s->gb, 6);
1315 sf[1] = get_bits(&s->gb, 6);
1316 sf[2] = get_bits(&s->gb, 6);
1319 sf[0] = get_bits(&s->gb, 6);
1324 sf[0] = get_bits(&s->gb, 6);
1325 sf[2] = get_bits(&s->gb, 6);
1329 sf[0] = get_bits(&s->gb, 6);
1330 sf[2] = get_bits(&s->gb, 6);
1339 for(ch=0;ch<s->nb_channels;ch++) {
1340 for(i=0;i<sblimit;i++) {
1341 if (bit_alloc[ch][i]) {
1342 sf = scale_factors[ch][i];
1343 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1354 for(l=0;l<12;l+=3) {
1356 for(i=0;i<bound;i++) {
1357 bit_alloc_bits = alloc_table[j];
1358 for(ch=0;ch<s->nb_channels;ch++) {
1359 b = bit_alloc[ch][i];
1361 scale = scale_factors[ch][i][k];
1362 qindex = alloc_table[j+b];
1363 bits = quant_bits[qindex];
1365 /* 3 values at the same time */
1366 v = get_bits(&s->gb, -bits);
1367 steps = quant_steps[qindex];
1368 s->sb_samples[ch][k * 12 + l + 0][i] =
1369 l2_unscale_group(steps, v % steps, scale);
1371 s->sb_samples[ch][k * 12 + l + 1][i] =
1372 l2_unscale_group(steps, v % steps, scale);
1374 s->sb_samples[ch][k * 12 + l + 2][i] =
1375 l2_unscale_group(steps, v, scale);
1378 v = get_bits(&s->gb, bits);
1379 v = l1_unscale(bits - 1, v, scale);
1380 s->sb_samples[ch][k * 12 + l + m][i] = v;
1384 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1385 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1386 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1389 /* next subband in alloc table */
1390 j += 1 << bit_alloc_bits;
1392 /* XXX: find a way to avoid this duplication of code */
1393 for(i=bound;i<sblimit;i++) {
1394 bit_alloc_bits = alloc_table[j];
1395 b = bit_alloc[0][i];
1397 int mant, scale0, scale1;
1398 scale0 = scale_factors[0][i][k];
1399 scale1 = scale_factors[1][i][k];
1400 qindex = alloc_table[j+b];
1401 bits = quant_bits[qindex];
1403 /* 3 values at the same time */
1404 v = get_bits(&s->gb, -bits);
1405 steps = quant_steps[qindex];
1408 s->sb_samples[0][k * 12 + l + 0][i] =
1409 l2_unscale_group(steps, mant, scale0);
1410 s->sb_samples[1][k * 12 + l + 0][i] =
1411 l2_unscale_group(steps, mant, scale1);
1414 s->sb_samples[0][k * 12 + l + 1][i] =
1415 l2_unscale_group(steps, mant, scale0);
1416 s->sb_samples[1][k * 12 + l + 1][i] =
1417 l2_unscale_group(steps, mant, scale1);
1418 s->sb_samples[0][k * 12 + l + 2][i] =
1419 l2_unscale_group(steps, v, scale0);
1420 s->sb_samples[1][k * 12 + l + 2][i] =
1421 l2_unscale_group(steps, v, scale1);
1424 mant = get_bits(&s->gb, bits);
1425 s->sb_samples[0][k * 12 + l + m][i] =
1426 l1_unscale(bits - 1, mant, scale0);
1427 s->sb_samples[1][k * 12 + l + m][i] =
1428 l1_unscale(bits - 1, mant, scale1);
1432 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1433 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1434 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1435 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1436 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1437 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1439 /* next subband in alloc table */
1440 j += 1 << bit_alloc_bits;
1442 /* fill remaining samples to zero */
1443 for(i=sblimit;i<SBLIMIT;i++) {
1444 for(ch=0;ch<s->nb_channels;ch++) {
1445 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1446 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1447 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1456 * Seek back in the stream for backstep bytes (at most 511 bytes)
1458 static void seek_to_maindata(MPADecodeContext *s, unsigned int backstep)
1462 /* compute current position in stream */
1463 ptr = s->gb.buffer + (get_bits_count(&s->gb)>>3);
1465 /* copy old data before current one */
1467 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1468 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1469 /* init get bits again */
1470 init_get_bits(&s->gb, ptr, (s->frame_size + backstep)*8);
1472 /* prepare next buffer */
1473 s->inbuf_index ^= 1;
1474 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1475 s->old_frame_size = s->frame_size;
1478 static inline void lsf_sf_expand(int *slen,
1479 int sf, int n1, int n2, int n3)
1498 static void exponents_from_scale_factors(MPADecodeContext *s,
1502 const uint8_t *bstab, *pretab;
1503 int len, i, j, k, l, v0, shift, gain, gains[3];
1506 exp_ptr = exponents;
1507 gain = g->global_gain - 210;
1508 shift = g->scalefac_scale + 1;
1510 bstab = band_size_long[s->sample_rate_index];
1511 pretab = mpa_pretab[g->preflag];
1512 for(i=0;i<g->long_end;i++) {
1513 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1519 if (g->short_start < 13) {
1520 bstab = band_size_short[s->sample_rate_index];
1521 gains[0] = gain - (g->subblock_gain[0] << 3);
1522 gains[1] = gain - (g->subblock_gain[1] << 3);
1523 gains[2] = gain - (g->subblock_gain[2] << 3);
1525 for(i=g->short_start;i<13;i++) {
1528 v0 = gains[l] - (g->scale_factors[k++] << shift);
1536 /* handle n = 0 too */
1537 static inline int get_bitsz(GetBitContext *s, int n)
1542 return get_bits(s, n);
1545 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1546 int16_t *exponents, int end_pos)
1549 int linbits, code, x, y, l, v, i, j, k, pos;
1550 GetBitContext last_gb;
1552 uint8_t *code_table;
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_gb.buffer = NULL;
1610 while (s_index <= 572) {
1611 pos = get_bits_count(&s->gb);
1612 if (pos >= end_pos) {
1613 if (pos > end_pos && last_gb.buffer != NULL) {
1614 /* some encoders generate an incorrect size for this
1615 part. We must go back into the data */
1623 code = get_vlc(&s->gb, vlc);
1624 dprintf("t=%d code=%d\n", g->count1table_select, code);
1628 if (code & (8 >> i)) {
1629 /* non zero value. Could use a hand coded function for
1631 v = l3_unscale(1, exponents[s_index]);
1632 if(get_bits1(&s->gb))
1637 g->sb_hybrid[s_index++] = v;
1640 while (s_index < 576)
1641 g->sb_hybrid[s_index++] = 0;
1645 /* Reorder short blocks from bitstream order to interleaved order. It
1646 would be faster to do it in parsing, but the code would be far more
1648 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1651 int32_t *ptr, *dst, *ptr1;
1654 if (g->block_type != 2)
1657 if (g->switch_point) {
1658 if (s->sample_rate_index != 8) {
1659 ptr = g->sb_hybrid + 36;
1661 ptr = g->sb_hybrid + 48;
1667 for(i=g->short_start;i<13;i++) {
1668 len = band_size_short[s->sample_rate_index][i];
1672 for(j=len;j>0;j--) {
1677 memcpy(ptr1, tmp, len * 3 * sizeof(int32_t));
1681 #define ISQRT2 FIXR(0.70710678118654752440)
1683 static void compute_stereo(MPADecodeContext *s,
1684 GranuleDef *g0, GranuleDef *g1)
1688 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1689 int32_t (*is_tab)[16];
1690 int32_t *tab0, *tab1;
1691 int non_zero_found_short[3];
1693 /* intensity stereo */
1694 if (s->mode_ext & MODE_EXT_I_STEREO) {
1699 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1703 tab0 = g0->sb_hybrid + 576;
1704 tab1 = g1->sb_hybrid + 576;
1706 non_zero_found_short[0] = 0;
1707 non_zero_found_short[1] = 0;
1708 non_zero_found_short[2] = 0;
1709 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1710 for(i = 12;i >= g1->short_start;i--) {
1711 /* for last band, use previous scale factor */
1714 len = band_size_short[s->sample_rate_index][i];
1718 if (!non_zero_found_short[l]) {
1719 /* test if non zero band. if so, stop doing i-stereo */
1720 for(j=0;j<len;j++) {
1722 non_zero_found_short[l] = 1;
1726 sf = g1->scale_factors[k + l];
1732 for(j=0;j<len;j++) {
1734 tab0[j] = MULL(tmp0, v1);
1735 tab1[j] = MULL(tmp0, v2);
1739 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1740 /* lower part of the spectrum : do ms stereo
1742 for(j=0;j<len;j++) {
1745 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1746 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1753 non_zero_found = non_zero_found_short[0] |
1754 non_zero_found_short[1] |
1755 non_zero_found_short[2];
1757 for(i = g1->long_end - 1;i >= 0;i--) {
1758 len = band_size_long[s->sample_rate_index][i];
1761 /* test if non zero band. if so, stop doing i-stereo */
1762 if (!non_zero_found) {
1763 for(j=0;j<len;j++) {
1769 /* for last band, use previous scale factor */
1770 k = (i == 21) ? 20 : i;
1771 sf = g1->scale_factors[k];
1776 for(j=0;j<len;j++) {
1778 tab0[j] = MULL(tmp0, v1);
1779 tab1[j] = MULL(tmp0, v2);
1783 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1784 /* lower part of the spectrum : do ms stereo
1786 for(j=0;j<len;j++) {
1789 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1790 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1795 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1796 /* ms stereo ONLY */
1797 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1799 tab0 = g0->sb_hybrid;
1800 tab1 = g1->sb_hybrid;
1801 for(i=0;i<576;i++) {
1804 tab0[i] = tmp0 + tmp1;
1805 tab1[i] = tmp0 - tmp1;
1810 static void compute_antialias(MPADecodeContext *s,
1813 int32_t *ptr, *p0, *p1, *csa;
1814 int n, tmp0, tmp1, i, j;
1816 /* we antialias only "long" bands */
1817 if (g->block_type == 2) {
1818 if (!g->switch_point)
1820 /* XXX: check this for 8000Hz case */
1826 ptr = g->sb_hybrid + 18;
1827 for(i = n;i > 0;i--) {
1830 csa = &csa_table[0][0];
1834 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1835 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1844 static void compute_imdct(MPADecodeContext *s,
1846 int32_t *sb_samples,
1849 int32_t *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
1853 int i, j, k, mdct_long_end, v, sblimit;
1855 /* find last non zero block */
1856 ptr = g->sb_hybrid + 576;
1857 ptr1 = g->sb_hybrid + 2 * 18;
1858 while (ptr >= ptr1) {
1860 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
1864 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
1866 if (g->block_type == 2) {
1867 /* XXX: check for 8000 Hz */
1868 if (g->switch_point)
1873 mdct_long_end = sblimit;
1878 for(j=0;j<mdct_long_end;j++) {
1880 /* apply window & overlap with previous buffer */
1881 out_ptr = sb_samples + j;
1883 if (g->switch_point && j < 2)
1886 win1 = mdct_win[g->block_type];
1887 /* select frequency inversion */
1888 win = win1 + ((4 * 36) & -(j & 1));
1890 *out_ptr = MULL(out[i], win[i]) + buf[i];
1891 buf[i] = MULL(out[i + 18], win[i + 18]);
1897 for(j=mdct_long_end;j<sblimit;j++) {
1903 /* select frequency inversion */
1904 win = mdct_win[2] + ((4 * 36) & -(j & 1));
1907 /* reorder input for short mdct */
1914 /* apply 12 point window and do small overlap */
1916 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
1917 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
1922 out_ptr = sb_samples + j;
1924 *out_ptr = out[i] + buf[i];
1925 buf[i] = out[i + 18];
1932 for(j=sblimit;j<SBLIMIT;j++) {
1934 out_ptr = sb_samples + j;
1945 void sample_dump(int fnum, int32_t *tab, int n)
1947 static FILE *files[16], *f;
1954 sprintf(buf, "/tmp/out%d.%s.pcm",
1956 #ifdef USE_HIGHPRECISION
1962 f = fopen(buf, "w");
1970 printf("pos=%d\n", pos);
1972 printf(" %0.4f", (double)tab[i] / FRAC_ONE);
1979 /* normalize to 23 frac bits */
1980 v = tab[i] << (23 - FRAC_BITS);
1981 fwrite(&v, 1, sizeof(int32_t), f);
1987 /* main layer3 decoding function */
1988 static int mp_decode_layer3(MPADecodeContext *s)
1990 int nb_granules, main_data_begin, private_bits;
1991 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
1992 GranuleDef granules[2][2], *g;
1993 int16_t exponents[576];
1995 /* read side info */
1997 main_data_begin = get_bits(&s->gb, 8);
1998 if (s->nb_channels == 2)
1999 private_bits = get_bits(&s->gb, 2);
2001 private_bits = get_bits(&s->gb, 1);
2004 main_data_begin = get_bits(&s->gb, 9);
2005 if (s->nb_channels == 2)
2006 private_bits = get_bits(&s->gb, 3);
2008 private_bits = get_bits(&s->gb, 5);
2010 for(ch=0;ch<s->nb_channels;ch++) {
2011 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2012 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2016 for(gr=0;gr<nb_granules;gr++) {
2017 for(ch=0;ch<s->nb_channels;ch++) {
2018 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2019 g = &granules[ch][gr];
2020 g->part2_3_length = get_bits(&s->gb, 12);
2021 g->big_values = get_bits(&s->gb, 9);
2022 g->global_gain = get_bits(&s->gb, 8);
2023 /* if MS stereo only is selected, we precompute the
2024 1/sqrt(2) renormalization factor */
2025 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2027 g->global_gain -= 2;
2029 g->scalefac_compress = get_bits(&s->gb, 9);
2031 g->scalefac_compress = get_bits(&s->gb, 4);
2032 blocksplit_flag = get_bits(&s->gb, 1);
2033 if (blocksplit_flag) {
2034 g->block_type = get_bits(&s->gb, 2);
2035 if (g->block_type == 0)
2037 g->switch_point = get_bits(&s->gb, 1);
2039 g->table_select[i] = get_bits(&s->gb, 5);
2041 g->subblock_gain[i] = get_bits(&s->gb, 3);
2042 /* compute huffman coded region sizes */
2043 if (g->block_type == 2)
2044 g->region_size[0] = (36 / 2);
2046 if (s->sample_rate_index <= 2)
2047 g->region_size[0] = (36 / 2);
2048 else if (s->sample_rate_index != 8)
2049 g->region_size[0] = (54 / 2);
2051 g->region_size[0] = (108 / 2);
2053 g->region_size[1] = (576 / 2);
2055 int region_address1, region_address2, l;
2057 g->switch_point = 0;
2059 g->table_select[i] = get_bits(&s->gb, 5);
2060 /* compute huffman coded region sizes */
2061 region_address1 = get_bits(&s->gb, 4);
2062 region_address2 = get_bits(&s->gb, 3);
2063 dprintf("region1=%d region2=%d\n",
2064 region_address1, region_address2);
2066 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2067 l = region_address1 + region_address2 + 2;
2068 /* should not overflow */
2072 band_index_long[s->sample_rate_index][l] >> 1;
2074 /* convert region offsets to region sizes and truncate
2075 size to big_values */
2076 g->region_size[2] = (576 / 2);
2079 k = g->region_size[i];
2080 if (k > g->big_values)
2082 g->region_size[i] = k - j;
2086 /* compute band indexes */
2087 if (g->block_type == 2) {
2088 if (g->switch_point) {
2089 /* if switched mode, we handle the 36 first samples as
2090 long blocks. For 8000Hz, we handle the 48 first
2091 exponents as long blocks (XXX: check this!) */
2092 if (s->sample_rate_index <= 2)
2094 else if (s->sample_rate_index != 8)
2097 g->long_end = 4; /* 8000 Hz */
2099 if (s->sample_rate_index != 8)
2108 g->short_start = 13;
2114 g->preflag = get_bits(&s->gb, 1);
2115 g->scalefac_scale = get_bits(&s->gb, 1);
2116 g->count1table_select = get_bits(&s->gb, 1);
2117 dprintf("block_type=%d switch_point=%d\n",
2118 g->block_type, g->switch_point);
2122 /* now we get bits from the main_data_begin offset */
2123 dprintf("seekback: %d\n", main_data_begin);
2124 seek_to_maindata(s, main_data_begin);
2126 for(gr=0;gr<nb_granules;gr++) {
2127 for(ch=0;ch<s->nb_channels;ch++) {
2128 g = &granules[ch][gr];
2130 bits_pos = get_bits_count(&s->gb);
2134 int slen, slen1, slen2;
2136 /* MPEG1 scale factors */
2137 slen1 = slen_table[0][g->scalefac_compress];
2138 slen2 = slen_table[1][g->scalefac_compress];
2139 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2140 if (g->block_type == 2) {
2141 n = g->switch_point ? 17 : 18;
2144 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2146 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2148 g->scale_factors[j++] = 0;
2150 sc = granules[ch][0].scale_factors;
2153 n = (k == 0 ? 6 : 5);
2154 if ((g->scfsi & (0x8 >> k)) == 0) {
2155 slen = (k < 2) ? slen1 : slen2;
2157 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2159 /* simply copy from last granule */
2161 g->scale_factors[j] = sc[j];
2166 g->scale_factors[j++] = 0;
2170 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2173 printf(" %d", g->scale_factors[i]);
2178 int tindex, tindex2, slen[4], sl, sf;
2180 /* LSF scale factors */
2181 if (g->block_type == 2) {
2182 tindex = g->switch_point ? 2 : 1;
2186 sf = g->scalefac_compress;
2187 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2188 /* intensity stereo case */
2191 lsf_sf_expand(slen, sf, 6, 6, 0);
2193 } else if (sf < 244) {
2194 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2197 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2203 lsf_sf_expand(slen, sf, 5, 4, 4);
2205 } else if (sf < 500) {
2206 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2209 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2217 n = lsf_nsf_table[tindex2][tindex][k];
2220 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2222 /* XXX: should compute exact size */
2224 g->scale_factors[j] = 0;
2227 printf("gr=%d ch=%d scale_factors:\n",
2230 printf(" %d", g->scale_factors[i]);
2236 exponents_from_scale_factors(s, g, exponents);
2238 /* read Huffman coded residue */
2239 if (huffman_decode(s, g, exponents,
2240 bits_pos + g->part2_3_length) < 0)
2243 sample_dump(0, g->sb_hybrid, 576);
2246 /* skip extension bits */
2247 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2248 if (bits_left < 0) {
2249 dprintf("bits_left=%d\n", bits_left);
2252 while (bits_left >= 16) {
2253 skip_bits(&s->gb, 16);
2257 skip_bits(&s->gb, bits_left);
2260 if (s->nb_channels == 2)
2261 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2263 for(ch=0;ch<s->nb_channels;ch++) {
2264 g = &granules[ch][gr];
2266 reorder_block(s, g);
2268 sample_dump(0, g->sb_hybrid, 576);
2270 compute_antialias(s, g);
2272 sample_dump(1, g->sb_hybrid, 576);
2274 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2276 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2280 return nb_granules * 18;
2283 static int mp_decode_frame(MPADecodeContext *s,
2286 int i, nb_frames, ch;
2289 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2290 (s->inbuf_ptr - s->inbuf - HEADER_SIZE)*8);
2292 /* skip error protection field */
2293 if (s->error_protection)
2294 get_bits(&s->gb, 16);
2296 dprintf("frame %d:\n", s->frame_count);
2299 nb_frames = mp_decode_layer1(s);
2302 nb_frames = mp_decode_layer2(s);
2306 nb_frames = mp_decode_layer3(s);
2310 for(i=0;i<nb_frames;i++) {
2311 for(ch=0;ch<s->nb_channels;ch++) {
2313 printf("%d-%d:", i, ch);
2314 for(j=0;j<SBLIMIT;j++)
2315 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2320 /* apply the synthesis filter */
2321 for(ch=0;ch<s->nb_channels;ch++) {
2322 samples_ptr = samples + ch;
2323 for(i=0;i<nb_frames;i++) {
2324 synth_filter(s, ch, samples_ptr, s->nb_channels,
2325 s->sb_samples[ch][i]);
2326 samples_ptr += 32 * s->nb_channels;
2332 return nb_frames * 32 * sizeof(short) * s->nb_channels;
2335 static int decode_frame(AVCodecContext * avctx,
2336 void *data, int *data_size,
2337 uint8_t * buf, int buf_size)
2339 MPADecodeContext *s = avctx->priv_data;
2343 short *out_samples = data;
2347 while (buf_size > 0) {
2348 len = s->inbuf_ptr - s->inbuf;
2349 if (s->frame_size == 0) {
2350 /* special case for next header for first frame in free
2351 format case (XXX: find a simpler method) */
2352 if (s->free_format_next_header != 0) {
2353 s->inbuf[0] = s->free_format_next_header >> 24;
2354 s->inbuf[1] = s->free_format_next_header >> 16;
2355 s->inbuf[2] = s->free_format_next_header >> 8;
2356 s->inbuf[3] = s->free_format_next_header;
2357 s->inbuf_ptr = s->inbuf + 4;
2358 s->free_format_next_header = 0;
2361 /* no header seen : find one. We need at least HEADER_SIZE
2362 bytes to parse it */
2363 len = HEADER_SIZE - len;
2367 memcpy(s->inbuf_ptr, buf_ptr, len);
2370 s->inbuf_ptr += len;
2372 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2374 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2375 (s->inbuf[2] << 8) | s->inbuf[3];
2377 if (check_header(header) < 0) {
2378 /* no sync found : move by one byte (inefficient, but simple!) */
2379 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2381 dprintf("skip %x\n", header);
2382 /* reset free format frame size to give a chance
2383 to get a new bitrate */
2384 s->free_format_frame_size = 0;
2386 if (decode_header(s, header) == 1) {
2387 /* free format: prepare to compute frame size */
2390 /* update codec info */
2391 avctx->sample_rate = s->sample_rate;
2392 avctx->channels = s->nb_channels;
2393 avctx->bit_rate = s->bit_rate;
2394 avctx->frame_size = s->frame_size;
2397 } else if (s->frame_size == -1) {
2398 /* free format : find next sync to compute frame size */
2399 len = MPA_MAX_CODED_FRAME_SIZE - len;
2403 /* frame too long: resync */
2405 memcpy(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2412 memcpy(s->inbuf_ptr, buf_ptr, len);
2413 /* check for header */
2414 p = s->inbuf_ptr - 3;
2415 pend = s->inbuf_ptr + len - 4;
2417 header = (p[0] << 24) | (p[1] << 16) |
2419 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2420 (s->inbuf[2] << 8) | s->inbuf[3];
2421 /* check with high probability that we have a
2423 if ((header & SAME_HEADER_MASK) ==
2424 (header1 & SAME_HEADER_MASK)) {
2425 /* header found: update pointers */
2426 len = (p + 4) - s->inbuf_ptr;
2430 /* compute frame size */
2431 s->free_format_next_header = header;
2432 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2433 padding = (header1 >> 9) & 1;
2435 s->free_format_frame_size -= padding * 4;
2437 s->free_format_frame_size -= padding;
2438 dprintf("free frame size=%d padding=%d\n",
2439 s->free_format_frame_size, padding);
2440 decode_header(s, header1);
2445 /* not found: simply increase pointers */
2447 s->inbuf_ptr += len;
2450 } else if (len < s->frame_size) {
2451 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2452 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2453 len = s->frame_size - len;
2456 memcpy(s->inbuf_ptr, buf_ptr, len);
2458 s->inbuf_ptr += len;
2461 out_size = mp_decode_frame(s, out_samples);
2462 s->inbuf_ptr = s->inbuf;
2464 *data_size = out_size;
2470 return buf_ptr - buf;
2473 AVCodec mp2_decoder =
2478 sizeof(MPADecodeContext),
2485 AVCodec mp3_decoder =
2490 sizeof(MPADecodeContext),