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"
32 * - in low precision mode, use more 16 bit multiplies in synth filter
33 * - test lsf / mpeg25 extensively.
36 /* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
38 #ifdef CONFIG_MPEGAUDIO_HP
39 #define USE_HIGHPRECISION
42 #ifdef USE_HIGHPRECISION
43 #define FRAC_BITS 23 /* fractional bits for sb_samples and dct */
44 #define WFRAC_BITS 16 /* fractional bits for window */
46 #define FRAC_BITS 15 /* fractional bits for sb_samples and dct */
47 #define WFRAC_BITS 14 /* fractional bits for window */
50 #define FRAC_ONE (1 << FRAC_BITS)
52 #define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
53 #define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
54 #define FIX(a) ((int)((a) * FRAC_ONE))
55 /* WARNING: only correct for posititive numbers */
56 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
57 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
60 typedef int16_t MPA_INT;
62 typedef int32_t MPA_INT;
68 #define BACKSTEP_SIZE 512
72 typedef struct MPADecodeContext {
73 uint8_t inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
75 uint8_t *inbuf_ptr, *inbuf;
77 int free_format_frame_size; /* frame size in case of free format
78 (zero if currently unknown) */
79 /* next header (used in free format parsing) */
80 uint32_t free_format_next_header;
84 int sample_rate_index; /* between 0 and 8 */
92 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
93 int synth_buf_offset[MPA_MAX_CHANNELS];
94 int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
95 int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
99 void (*compute_antialias)(struct MPADecodeContext *s, struct GranuleDef *g);
102 /* layer 3 "granule" */
103 typedef struct GranuleDef {
108 int scalefac_compress;
110 uint8_t switch_point;
112 int subblock_gain[3];
113 uint8_t scalefac_scale;
114 uint8_t count1table_select;
115 int region_size[3]; /* number of huffman codes in each region */
117 int short_start, long_end; /* long/short band indexes */
118 uint8_t scale_factors[40];
119 int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
122 #define MODE_EXT_MS_STEREO 2
123 #define MODE_EXT_I_STEREO 1
125 /* layer 3 huffman tables */
126 typedef struct HuffTable {
129 const uint16_t *codes;
132 #include "mpegaudiodectab.h"
134 static void compute_antialias_integer(MPADecodeContext *s, GranuleDef *g);
135 static void compute_antialias_float(MPADecodeContext *s, GranuleDef *g);
137 /* vlc structure for decoding layer 3 huffman tables */
138 static VLC huff_vlc[16];
139 static uint8_t *huff_code_table[16];
140 static VLC huff_quad_vlc[2];
141 /* computed from band_size_long */
142 static uint16_t band_index_long[9][23];
143 /* XXX: free when all decoders are closed */
144 #define TABLE_4_3_SIZE (8191 + 16)
145 static int8_t *table_4_3_exp;
147 static uint16_t *table_4_3_value;
149 static uint32_t *table_4_3_value;
151 /* intensity stereo coef table */
152 static int32_t is_table[2][16];
153 static int32_t is_table_lsf[2][2][16];
154 static int32_t csa_table[8][4];
155 static float csa_table_float[8][4];
156 static int32_t mdct_win[8][36];
158 /* lower 2 bits: modulo 3, higher bits: shift */
159 static uint16_t scale_factor_modshift[64];
160 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
161 static int32_t scale_factor_mult[15][3];
162 /* mult table for layer 2 group quantization */
164 #define SCALE_GEN(v) \
165 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
167 static int32_t scale_factor_mult2[3][3] = {
168 SCALE_GEN(4.0 / 3.0), /* 3 steps */
169 SCALE_GEN(4.0 / 5.0), /* 5 steps */
170 SCALE_GEN(4.0 / 9.0), /* 9 steps */
174 static uint32_t scale_factor_mult3[4] = {
176 FIXR(1.18920711500272106671),
177 FIXR(1.41421356237309504880),
178 FIXR(1.68179283050742908605),
181 static MPA_INT window[512] __attribute__((aligned(16)));
183 /* layer 1 unscaling */
184 /* n = number of bits of the mantissa minus 1 */
185 static inline int l1_unscale(int n, int mant, int scale_factor)
190 shift = scale_factor_modshift[scale_factor];
193 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
195 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
196 return (int)((val + (1LL << (shift - 1))) >> shift);
199 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
203 shift = scale_factor_modshift[scale_factor];
207 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
208 /* NOTE: at this point, 0 <= shift <= 21 */
210 val = (val + (1 << (shift - 1))) >> shift;
214 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
215 static inline int l3_unscale(int value, int exponent)
224 e = table_4_3_exp[value];
225 e += (exponent >> 2);
231 m = table_4_3_value[value];
233 m = (m * scale_factor_mult3[exponent & 3]);
234 m = (m + (1 << (e-1))) >> e;
237 m = MUL64(m, scale_factor_mult3[exponent & 3]);
238 m = (m + (uint64_t_C(1) << (e-1))) >> e;
243 /* all integer n^(4/3) computation code */
246 #define POW_FRAC_BITS 24
247 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
248 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
249 #define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)
251 static int dev_4_3_coefs[DEV_ORDER];
253 static int pow_mult3[3] = {
255 POW_FIX(1.25992104989487316476),
256 POW_FIX(1.58740105196819947474),
259 static void int_pow_init(void)
264 for(i=0;i<DEV_ORDER;i++) {
265 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
266 dev_4_3_coefs[i] = a;
270 /* return the mantissa and the binary exponent */
271 static int int_pow(int i, int *exp_ptr)
279 while (a < (1 << (POW_FRAC_BITS - 1))) {
283 a -= (1 << POW_FRAC_BITS);
285 for(j = DEV_ORDER - 1; j >= 0; j--)
286 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
287 a = (1 << POW_FRAC_BITS) + a1;
288 /* exponent compute (exact) */
292 a = POW_MULL(a, pow_mult3[er]);
293 while (a >= 2 * POW_FRAC_ONE) {
297 /* convert to float */
298 while (a < POW_FRAC_ONE) {
302 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
303 #if POW_FRAC_BITS > FRAC_BITS
304 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
305 /* correct overflow */
306 if (a >= 2 * (1 << FRAC_BITS)) {
315 static int decode_init(AVCodecContext * avctx)
317 MPADecodeContext *s = avctx->priv_data;
321 if(avctx->antialias_algo == FF_AA_INT)
322 s->compute_antialias= compute_antialias_integer;
324 s->compute_antialias= compute_antialias_float;
326 if (!init && !avctx->parse_only) {
327 /* scale factors table for layer 1/2 */
330 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
333 scale_factor_modshift[i] = mod | (shift << 2);
336 /* scale factor multiply for layer 1 */
340 norm = ((int64_t_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
341 scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
342 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
343 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
344 dprintf("%d: norm=%x s=%x %x %x\n",
346 scale_factor_mult[i][0],
347 scale_factor_mult[i][1],
348 scale_factor_mult[i][2]);
352 /* max = 18760, max sum over all 16 coefs : 44736 */
357 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
366 /* huffman decode tables */
367 huff_code_table[0] = NULL;
369 const HuffTable *h = &mpa_huff_tables[i];
377 init_vlc(&huff_vlc[i], 8, n,
378 h->bits, 1, 1, h->codes, 2, 2);
380 code_table = av_mallocz(n);
382 for(x=0;x<xsize;x++) {
384 code_table[j++] = (x << 4) | y;
386 huff_code_table[i] = code_table;
389 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
390 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1);
396 band_index_long[i][j] = k;
397 k += band_size_long[i][j];
399 band_index_long[i][22] = k;
402 /* compute n ^ (4/3) and store it in mantissa/exp format */
403 table_4_3_exp= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_exp[0]));
406 table_4_3_value= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_value[0]));
411 for(i=1;i<TABLE_4_3_SIZE;i++) {
419 f = pow((double)i, 4.0 / 3.0);
423 if ((unsigned short)m1 != m1) {
429 if (m != m1 || e != e1) {
430 printf("%4d: m=%x m1=%x e=%d e1=%d\n",
435 /* normalized to FRAC_BITS */
436 table_4_3_value[i] = m;
437 table_4_3_exp[i] = e;
444 f = tan((double)i * M_PI / 12.0);
445 v = FIXR(f / (1.0 + f));
450 is_table[1][6 - i] = v;
454 is_table[0][i] = is_table[1][i] = 0.0;
461 e = -(j + 1) * ((i + 1) >> 1);
462 f = pow(2.0, e / 4.0);
464 is_table_lsf[j][k ^ 1][i] = FIXR(f);
465 is_table_lsf[j][k][i] = FIXR(1.0);
466 dprintf("is_table_lsf %d %d: %x %x\n",
467 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
474 cs = 1.0 / sqrt(1.0 + ci * ci);
476 csa_table[i][0] = FIX(cs);
477 csa_table[i][1] = FIX(ca);
478 csa_table[i][2] = FIX(ca) + FIX(cs);
479 csa_table[i][3] = FIX(ca) - FIX(cs);
480 csa_table_float[i][0] = cs;
481 csa_table_float[i][1] = ca;
482 csa_table_float[i][2] = ca + cs;
483 csa_table_float[i][3] = ca - cs;
484 // printf("%d %d %d %d\n", FIX(cs), FIX(cs-1), FIX(ca), FIX(cs)-FIX(ca));
487 /* compute mdct windows */
490 v = FIXR(sin(M_PI * (i + 0.5) / 36.0));
496 mdct_win[1][18 + i] = FIXR(1.0);
497 mdct_win[1][24 + i] = FIXR(sin(M_PI * ((i + 6) + 0.5) / 12.0));
498 mdct_win[1][30 + i] = FIXR(0.0);
500 mdct_win[3][i] = FIXR(0.0);
501 mdct_win[3][6 + i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
502 mdct_win[3][12 + i] = FIXR(1.0);
506 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
508 /* NOTE: we do frequency inversion adter the MDCT by changing
509 the sign of the right window coefs */
512 mdct_win[j + 4][i] = mdct_win[j][i];
513 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
519 printf("win%d=\n", j);
521 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
529 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
530 s->inbuf_ptr = s->inbuf;
537 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
541 #define COS0_0 FIXR(0.50060299823519630134)
542 #define COS0_1 FIXR(0.50547095989754365998)
543 #define COS0_2 FIXR(0.51544730992262454697)
544 #define COS0_3 FIXR(0.53104259108978417447)
545 #define COS0_4 FIXR(0.55310389603444452782)
546 #define COS0_5 FIXR(0.58293496820613387367)
547 #define COS0_6 FIXR(0.62250412303566481615)
548 #define COS0_7 FIXR(0.67480834145500574602)
549 #define COS0_8 FIXR(0.74453627100229844977)
550 #define COS0_9 FIXR(0.83934964541552703873)
551 #define COS0_10 FIXR(0.97256823786196069369)
552 #define COS0_11 FIXR(1.16943993343288495515)
553 #define COS0_12 FIXR(1.48416461631416627724)
554 #define COS0_13 FIXR(2.05778100995341155085)
555 #define COS0_14 FIXR(3.40760841846871878570)
556 #define COS0_15 FIXR(10.19000812354805681150)
558 #define COS1_0 FIXR(0.50241928618815570551)
559 #define COS1_1 FIXR(0.52249861493968888062)
560 #define COS1_2 FIXR(0.56694403481635770368)
561 #define COS1_3 FIXR(0.64682178335999012954)
562 #define COS1_4 FIXR(0.78815462345125022473)
563 #define COS1_5 FIXR(1.06067768599034747134)
564 #define COS1_6 FIXR(1.72244709823833392782)
565 #define COS1_7 FIXR(5.10114861868916385802)
567 #define COS2_0 FIXR(0.50979557910415916894)
568 #define COS2_1 FIXR(0.60134488693504528054)
569 #define COS2_2 FIXR(0.89997622313641570463)
570 #define COS2_3 FIXR(2.56291544774150617881)
572 #define COS3_0 FIXR(0.54119610014619698439)
573 #define COS3_1 FIXR(1.30656296487637652785)
575 #define COS4_0 FIXR(0.70710678118654752439)
577 /* butterfly operator */
580 tmp0 = tab[a] + tab[b];\
581 tmp1 = tab[a] - tab[b];\
583 tab[b] = MULL(tmp1, c);\
586 #define BF1(a, b, c, d)\
593 #define BF2(a, b, c, d)\
603 #define ADD(a, b) tab[a] += tab[b]
605 /* DCT32 without 1/sqrt(2) coef zero scaling. */
606 static void dct32(int32_t *out, int32_t *tab)
738 out[ 1] = tab[16] + tab[24];
739 out[17] = tab[17] + tab[25];
740 out[ 9] = tab[18] + tab[26];
741 out[25] = tab[19] + tab[27];
742 out[ 5] = tab[20] + tab[28];
743 out[21] = tab[21] + tab[29];
744 out[13] = tab[22] + tab[30];
745 out[29] = tab[23] + tab[31];
746 out[ 3] = tab[24] + tab[20];
747 out[19] = tab[25] + tab[21];
748 out[11] = tab[26] + tab[22];
749 out[27] = tab[27] + tab[23];
750 out[ 7] = tab[28] + tab[18];
751 out[23] = tab[29] + tab[19];
752 out[15] = tab[30] + tab[17];
756 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
760 static inline int round_sample(int sum)
763 sum1 = (sum + (1 << (OUT_SHIFT - 1))) >> OUT_SHIFT;
766 else if (sum1 > 32767)
771 #if defined(ARCH_POWERPC_405)
773 /* signed 16x16 -> 32 multiply add accumulate */
774 #define MACS(rt, ra, rb) \
775 asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
777 /* signed 16x16 -> 32 multiply */
778 #define MULS(ra, rb) \
779 ({ int __rt; asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
783 /* signed 16x16 -> 32 multiply add accumulate */
784 #define MACS(rt, ra, rb) rt += (ra) * (rb)
786 /* signed 16x16 -> 32 multiply */
787 #define MULS(ra, rb) ((ra) * (rb))
793 static inline int round_sample(int64_t sum)
796 sum1 = (int)((sum + (int64_t_C(1) << (OUT_SHIFT - 1))) >> OUT_SHIFT);
799 else if (sum1 > 32767)
804 #define MULS(ra, rb) MUL64(ra, rb)
808 #define SUM8(sum, op, w, p) \
810 sum op MULS((w)[0 * 64], p[0 * 64]);\
811 sum op MULS((w)[1 * 64], p[1 * 64]);\
812 sum op MULS((w)[2 * 64], p[2 * 64]);\
813 sum op MULS((w)[3 * 64], p[3 * 64]);\
814 sum op MULS((w)[4 * 64], p[4 * 64]);\
815 sum op MULS((w)[5 * 64], p[5 * 64]);\
816 sum op MULS((w)[6 * 64], p[6 * 64]);\
817 sum op MULS((w)[7 * 64], p[7 * 64]);\
820 #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
824 sum1 op1 MULS((w1)[0 * 64], tmp);\
825 sum2 op2 MULS((w2)[0 * 64], tmp);\
827 sum1 op1 MULS((w1)[1 * 64], tmp);\
828 sum2 op2 MULS((w2)[1 * 64], tmp);\
830 sum1 op1 MULS((w1)[2 * 64], tmp);\
831 sum2 op2 MULS((w2)[2 * 64], tmp);\
833 sum1 op1 MULS((w1)[3 * 64], tmp);\
834 sum2 op2 MULS((w2)[3 * 64], tmp);\
836 sum1 op1 MULS((w1)[4 * 64], tmp);\
837 sum2 op2 MULS((w2)[4 * 64], tmp);\
839 sum1 op1 MULS((w1)[5 * 64], tmp);\
840 sum2 op2 MULS((w2)[5 * 64], tmp);\
842 sum1 op1 MULS((w1)[6 * 64], tmp);\
843 sum2 op2 MULS((w2)[6 * 64], tmp);\
845 sum1 op1 MULS((w1)[7 * 64], tmp);\
846 sum2 op2 MULS((w2)[7 * 64], tmp);\
850 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
852 /* XXX: optimize by avoiding ring buffer usage */
853 static void synth_filter(MPADecodeContext *s1,
854 int ch, int16_t *samples, int incr,
855 int32_t sb_samples[SBLIMIT])
858 register MPA_INT *synth_buf;
859 const register MPA_INT *w, *w2, *p;
868 dct32(tmp, sb_samples);
870 offset = s1->synth_buf_offset[ch];
871 synth_buf = s1->synth_buf[ch] + offset;
876 /* NOTE: can cause a loss in precision if very high amplitude
885 /* copy to avoid wrap */
886 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
888 samples2 = samples + 31 * incr;
896 SUM8(sum, -=, w + 32, p);
897 *samples = round_sample(sum);
901 /* we calculate two samples at the same time to avoid one memory
902 access per two sample */
906 p = synth_buf + 16 + j;
907 SUM8P2(sum, +=, sum2, -=, w, w2, p);
908 p = synth_buf + 48 - j;
909 SUM8P2(sum, -=, sum2, -=, w + 32, w2 + 32, p);
911 *samples = round_sample(sum);
913 *samples2 = round_sample(sum2);
921 SUM8(sum, -=, w + 32, p);
922 *samples = round_sample(sum);
924 offset = (offset - 32) & 511;
925 s1->synth_buf_offset[ch] = offset;
929 #define C1 FIXR(0.99144486137381041114)
930 #define C3 FIXR(0.92387953251128675612)
931 #define C5 FIXR(0.79335334029123516458)
932 #define C7 FIXR(0.60876142900872063941)
933 #define C9 FIXR(0.38268343236508977173)
934 #define C11 FIXR(0.13052619222005159154)
936 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
938 static void imdct12(int *out, int *in)
941 int64_t in1_3, in1_9, in4_3, in4_9;
943 in1_3 = MUL64(in[1], C3);
944 in1_9 = MUL64(in[1], C9);
945 in4_3 = MUL64(in[4], C3);
946 in4_9 = MUL64(in[4], C9);
948 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
949 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
952 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
953 MUL64(in[2] + in[5], C3) - in4_9);
956 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
957 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
960 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
961 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
964 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
965 MUL64(in[2] + in[5], C9) + in4_3);
968 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
969 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
982 #define C1 FIXR(0.98480775301220805936)
983 #define C2 FIXR(0.93969262078590838405)
984 #define C3 FIXR(0.86602540378443864676)
985 #define C4 FIXR(0.76604444311897803520)
986 #define C5 FIXR(0.64278760968653932632)
988 #define C7 FIXR(0.34202014332566873304)
989 #define C8 FIXR(0.17364817766693034885)
991 /* 0.5 / cos(pi*(2*i+1)/36) */
992 static const int icos36[9] = {
993 FIXR(0.50190991877167369479),
994 FIXR(0.51763809020504152469),
995 FIXR(0.55168895948124587824),
996 FIXR(0.61038729438072803416),
997 FIXR(0.70710678118654752439),
998 FIXR(0.87172339781054900991),
999 FIXR(1.18310079157624925896),
1000 FIXR(1.93185165257813657349),
1001 FIXR(5.73685662283492756461),
1004 static const int icos72[18] = {
1005 /* 0.5 / cos(pi*(2*i+19)/72) */
1006 FIXR(0.74009361646113053152),
1007 FIXR(0.82133981585229078570),
1008 FIXR(0.93057949835178895673),
1009 FIXR(1.08284028510010010928),
1010 FIXR(1.30656296487637652785),
1011 FIXR(1.66275476171152078719),
1012 FIXR(2.31011315767264929558),
1013 FIXR(3.83064878777019433457),
1014 FIXR(11.46279281302667383546),
1016 /* 0.5 / cos(pi*(2*(i + 18) +19)/72) */
1017 FIXR(-0.67817085245462840086),
1018 FIXR(-0.63023620700513223342),
1019 FIXR(-0.59284452371708034528),
1020 FIXR(-0.56369097343317117734),
1021 FIXR(-0.54119610014619698439),
1022 FIXR(-0.52426456257040533932),
1023 FIXR(-0.51213975715725461845),
1024 FIXR(-0.50431448029007636036),
1025 FIXR(-0.50047634258165998492),
1028 /* using Lee like decomposition followed by hand coded 9 points DCT */
1029 static void imdct36(int *out, int *in)
1031 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
1032 int tmp[18], *tmp1, *in1;
1033 int64_t in3_3, in6_6;
1044 in3_3 = MUL64(in1[2*3], C3);
1045 in6_6 = MUL64(in1[2*6], C6);
1047 tmp1[0] = FRAC_RND(MUL64(in1[2*1], C1) + in3_3 +
1048 MUL64(in1[2*5], C5) + MUL64(in1[2*7], C7));
1049 tmp1[2] = in1[2*0] + FRAC_RND(MUL64(in1[2*2], C2) +
1050 MUL64(in1[2*4], C4) + in6_6 +
1051 MUL64(in1[2*8], C8));
1052 tmp1[4] = FRAC_RND(MUL64(in1[2*1] - in1[2*5] - in1[2*7], C3));
1053 tmp1[6] = FRAC_RND(MUL64(in1[2*2] - in1[2*4] - in1[2*8], C6)) -
1054 in1[2*6] + in1[2*0];
1055 tmp1[8] = FRAC_RND(MUL64(in1[2*1], C5) - in3_3 -
1056 MUL64(in1[2*5], C7) + MUL64(in1[2*7], C1));
1057 tmp1[10] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C8) -
1058 MUL64(in1[2*4], C2) + in6_6 +
1059 MUL64(in1[2*8], C4));
1060 tmp1[12] = FRAC_RND(MUL64(in1[2*1], C7) - in3_3 +
1061 MUL64(in1[2*5], C1) -
1062 MUL64(in1[2*7], C5));
1063 tmp1[14] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C4) +
1064 MUL64(in1[2*4], C8) + in6_6 -
1065 MUL64(in1[2*8], C2));
1066 tmp1[16] = in1[2*0] - in1[2*2] + in1[2*4] - in1[2*6] + in1[2*8];
1078 s1 = MULL(t3 + t2, icos36[j]);
1079 s3 = MULL(t3 - t2, icos36[8 - j]);
1081 t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
1082 t1 = MULL(s0 - s1, icos72[8 - j]);
1083 out[18 + 9 + j] = t0;
1084 out[18 + 8 - j] = t0;
1088 t0 = MULL(s2 + s3, icos72[9+j]);
1089 t1 = MULL(s2 - s3, icos72[j]);
1090 out[18 + 9 + (8 - j)] = t0;
1092 out[9 + (8 - j)] = -t1;
1098 s1 = MULL(tmp[17], icos36[4]);
1099 t0 = MULL(s0 + s1, icos72[9 + 4]);
1100 t1 = MULL(s0 - s1, icos72[4]);
1101 out[18 + 9 + 4] = t0;
1102 out[18 + 8 - 4] = t0;
1107 /* fast header check for resync */
1108 static int check_header(uint32_t header)
1111 if ((header & 0xffe00000) != 0xffe00000)
1114 if (((header >> 17) & 3) == 0)
1117 if (((header >> 12) & 0xf) == 0xf)
1120 if (((header >> 10) & 3) == 3)
1125 /* header + layer + bitrate + freq + lsf/mpeg25 */
1126 #define SAME_HEADER_MASK \
1127 (0xffe00000 | (3 << 17) | (0xf << 12) | (3 << 10) | (3 << 19))
1129 /* header decoding. MUST check the header before because no
1130 consistency check is done there. Return 1 if free format found and
1131 that the frame size must be computed externally */
1132 static int decode_header(MPADecodeContext *s, uint32_t header)
1134 int sample_rate, frame_size, mpeg25, padding;
1135 int sample_rate_index, bitrate_index;
1136 if (header & (1<<20)) {
1137 s->lsf = (header & (1<<19)) ? 0 : 1;
1144 s->layer = 4 - ((header >> 17) & 3);
1145 /* extract frequency */
1146 sample_rate_index = (header >> 10) & 3;
1147 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1148 sample_rate_index += 3 * (s->lsf + mpeg25);
1149 s->sample_rate_index = sample_rate_index;
1150 s->error_protection = ((header >> 16) & 1) ^ 1;
1151 s->sample_rate = sample_rate;
1153 bitrate_index = (header >> 12) & 0xf;
1154 padding = (header >> 9) & 1;
1155 //extension = (header >> 8) & 1;
1156 s->mode = (header >> 6) & 3;
1157 s->mode_ext = (header >> 4) & 3;
1158 //copyright = (header >> 3) & 1;
1159 //original = (header >> 2) & 1;
1160 //emphasis = header & 3;
1162 if (s->mode == MPA_MONO)
1167 if (bitrate_index != 0) {
1168 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1169 s->bit_rate = frame_size * 1000;
1172 frame_size = (frame_size * 12000) / sample_rate;
1173 frame_size = (frame_size + padding) * 4;
1176 frame_size = (frame_size * 144000) / sample_rate;
1177 frame_size += padding;
1181 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1182 frame_size += padding;
1185 s->frame_size = frame_size;
1187 /* if no frame size computed, signal it */
1188 if (!s->free_format_frame_size)
1190 /* free format: compute bitrate and real frame size from the
1191 frame size we extracted by reading the bitstream */
1192 s->frame_size = s->free_format_frame_size;
1195 s->frame_size += padding * 4;
1196 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1199 s->frame_size += padding;
1200 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1204 s->frame_size += padding;
1205 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1211 printf("layer%d, %d Hz, %d kbits/s, ",
1212 s->layer, s->sample_rate, s->bit_rate);
1213 if (s->nb_channels == 2) {
1214 if (s->layer == 3) {
1215 if (s->mode_ext & MODE_EXT_MS_STEREO)
1217 if (s->mode_ext & MODE_EXT_I_STEREO)
1229 /* useful helper to get mpeg audio stream infos. Return -1 if error in
1230 header, otherwise the coded frame size in bytes */
1231 int mpa_decode_header(AVCodecContext *avctx, uint32_t head)
1233 MPADecodeContext s1, *s = &s1;
1234 memset( s, 0, sizeof(MPADecodeContext) );
1236 if (check_header(head) != 0)
1239 if (decode_header(s, head) != 0) {
1245 avctx->frame_size = 384;
1248 avctx->frame_size = 1152;
1253 avctx->frame_size = 576;
1255 avctx->frame_size = 1152;
1259 avctx->sample_rate = s->sample_rate;
1260 avctx->channels = s->nb_channels;
1261 avctx->bit_rate = s->bit_rate;
1262 avctx->sub_id = s->layer;
1263 return s->frame_size;
1266 /* return the number of decoded frames */
1267 static int mp_decode_layer1(MPADecodeContext *s)
1269 int bound, i, v, n, ch, j, mant;
1270 uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
1271 uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1273 if (s->mode == MPA_JSTEREO)
1274 bound = (s->mode_ext + 1) * 4;
1278 /* allocation bits */
1279 for(i=0;i<bound;i++) {
1280 for(ch=0;ch<s->nb_channels;ch++) {
1281 allocation[ch][i] = get_bits(&s->gb, 4);
1284 for(i=bound;i<SBLIMIT;i++) {
1285 allocation[0][i] = get_bits(&s->gb, 4);
1289 for(i=0;i<bound;i++) {
1290 for(ch=0;ch<s->nb_channels;ch++) {
1291 if (allocation[ch][i])
1292 scale_factors[ch][i] = get_bits(&s->gb, 6);
1295 for(i=bound;i<SBLIMIT;i++) {
1296 if (allocation[0][i]) {
1297 scale_factors[0][i] = get_bits(&s->gb, 6);
1298 scale_factors[1][i] = get_bits(&s->gb, 6);
1302 /* compute samples */
1304 for(i=0;i<bound;i++) {
1305 for(ch=0;ch<s->nb_channels;ch++) {
1306 n = allocation[ch][i];
1308 mant = get_bits(&s->gb, n + 1);
1309 v = l1_unscale(n, mant, scale_factors[ch][i]);
1313 s->sb_samples[ch][j][i] = v;
1316 for(i=bound;i<SBLIMIT;i++) {
1317 n = allocation[0][i];
1319 mant = get_bits(&s->gb, n + 1);
1320 v = l1_unscale(n, mant, scale_factors[0][i]);
1321 s->sb_samples[0][j][i] = v;
1322 v = l1_unscale(n, mant, scale_factors[1][i]);
1323 s->sb_samples[1][j][i] = v;
1325 s->sb_samples[0][j][i] = 0;
1326 s->sb_samples[1][j][i] = 0;
1333 /* bitrate is in kb/s */
1334 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1336 int ch_bitrate, table;
1338 ch_bitrate = bitrate / nb_channels;
1340 if ((freq == 48000 && ch_bitrate >= 56) ||
1341 (ch_bitrate >= 56 && ch_bitrate <= 80))
1343 else if (freq != 48000 && ch_bitrate >= 96)
1345 else if (freq != 32000 && ch_bitrate <= 48)
1355 static int mp_decode_layer2(MPADecodeContext *s)
1357 int sblimit; /* number of used subbands */
1358 const unsigned char *alloc_table;
1359 int table, bit_alloc_bits, i, j, ch, bound, v;
1360 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1361 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1362 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1363 int scale, qindex, bits, steps, k, l, m, b;
1365 /* select decoding table */
1366 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1367 s->sample_rate, s->lsf);
1368 sblimit = sblimit_table[table];
1369 alloc_table = alloc_tables[table];
1371 if (s->mode == MPA_JSTEREO)
1372 bound = (s->mode_ext + 1) * 4;
1376 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1379 if( bound > sblimit ) bound = sblimit;
1381 /* parse bit allocation */
1383 for(i=0;i<bound;i++) {
1384 bit_alloc_bits = alloc_table[j];
1385 for(ch=0;ch<s->nb_channels;ch++) {
1386 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1388 j += 1 << bit_alloc_bits;
1390 for(i=bound;i<sblimit;i++) {
1391 bit_alloc_bits = alloc_table[j];
1392 v = get_bits(&s->gb, bit_alloc_bits);
1393 bit_alloc[0][i] = v;
1394 bit_alloc[1][i] = v;
1395 j += 1 << bit_alloc_bits;
1400 for(ch=0;ch<s->nb_channels;ch++) {
1401 for(i=0;i<sblimit;i++)
1402 printf(" %d", bit_alloc[ch][i]);
1409 for(i=0;i<sblimit;i++) {
1410 for(ch=0;ch<s->nb_channels;ch++) {
1411 if (bit_alloc[ch][i])
1412 scale_code[ch][i] = get_bits(&s->gb, 2);
1417 for(i=0;i<sblimit;i++) {
1418 for(ch=0;ch<s->nb_channels;ch++) {
1419 if (bit_alloc[ch][i]) {
1420 sf = scale_factors[ch][i];
1421 switch(scale_code[ch][i]) {
1424 sf[0] = get_bits(&s->gb, 6);
1425 sf[1] = get_bits(&s->gb, 6);
1426 sf[2] = get_bits(&s->gb, 6);
1429 sf[0] = get_bits(&s->gb, 6);
1434 sf[0] = get_bits(&s->gb, 6);
1435 sf[2] = get_bits(&s->gb, 6);
1439 sf[0] = get_bits(&s->gb, 6);
1440 sf[2] = get_bits(&s->gb, 6);
1449 for(ch=0;ch<s->nb_channels;ch++) {
1450 for(i=0;i<sblimit;i++) {
1451 if (bit_alloc[ch][i]) {
1452 sf = scale_factors[ch][i];
1453 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1464 for(l=0;l<12;l+=3) {
1466 for(i=0;i<bound;i++) {
1467 bit_alloc_bits = alloc_table[j];
1468 for(ch=0;ch<s->nb_channels;ch++) {
1469 b = bit_alloc[ch][i];
1471 scale = scale_factors[ch][i][k];
1472 qindex = alloc_table[j+b];
1473 bits = quant_bits[qindex];
1475 /* 3 values at the same time */
1476 v = get_bits(&s->gb, -bits);
1477 steps = quant_steps[qindex];
1478 s->sb_samples[ch][k * 12 + l + 0][i] =
1479 l2_unscale_group(steps, v % steps, scale);
1481 s->sb_samples[ch][k * 12 + l + 1][i] =
1482 l2_unscale_group(steps, v % steps, scale);
1484 s->sb_samples[ch][k * 12 + l + 2][i] =
1485 l2_unscale_group(steps, v, scale);
1488 v = get_bits(&s->gb, bits);
1489 v = l1_unscale(bits - 1, v, scale);
1490 s->sb_samples[ch][k * 12 + l + m][i] = v;
1494 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1495 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1496 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1499 /* next subband in alloc table */
1500 j += 1 << bit_alloc_bits;
1502 /* XXX: find a way to avoid this duplication of code */
1503 for(i=bound;i<sblimit;i++) {
1504 bit_alloc_bits = alloc_table[j];
1505 b = bit_alloc[0][i];
1507 int mant, scale0, scale1;
1508 scale0 = scale_factors[0][i][k];
1509 scale1 = scale_factors[1][i][k];
1510 qindex = alloc_table[j+b];
1511 bits = quant_bits[qindex];
1513 /* 3 values at the same time */
1514 v = get_bits(&s->gb, -bits);
1515 steps = quant_steps[qindex];
1518 s->sb_samples[0][k * 12 + l + 0][i] =
1519 l2_unscale_group(steps, mant, scale0);
1520 s->sb_samples[1][k * 12 + l + 0][i] =
1521 l2_unscale_group(steps, mant, scale1);
1524 s->sb_samples[0][k * 12 + l + 1][i] =
1525 l2_unscale_group(steps, mant, scale0);
1526 s->sb_samples[1][k * 12 + l + 1][i] =
1527 l2_unscale_group(steps, mant, scale1);
1528 s->sb_samples[0][k * 12 + l + 2][i] =
1529 l2_unscale_group(steps, v, scale0);
1530 s->sb_samples[1][k * 12 + l + 2][i] =
1531 l2_unscale_group(steps, v, scale1);
1534 mant = get_bits(&s->gb, bits);
1535 s->sb_samples[0][k * 12 + l + m][i] =
1536 l1_unscale(bits - 1, mant, scale0);
1537 s->sb_samples[1][k * 12 + l + m][i] =
1538 l1_unscale(bits - 1, mant, scale1);
1542 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1543 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1544 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1545 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1546 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1547 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1549 /* next subband in alloc table */
1550 j += 1 << bit_alloc_bits;
1552 /* fill remaining samples to zero */
1553 for(i=sblimit;i<SBLIMIT;i++) {
1554 for(ch=0;ch<s->nb_channels;ch++) {
1555 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1556 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1557 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1566 * Seek back in the stream for backstep bytes (at most 511 bytes)
1568 static void seek_to_maindata(MPADecodeContext *s, unsigned int backstep)
1572 /* compute current position in stream */
1573 ptr = (uint8_t *)(s->gb.buffer + (get_bits_count(&s->gb)>>3));
1575 /* copy old data before current one */
1577 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1578 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1579 /* init get bits again */
1580 init_get_bits(&s->gb, ptr, (s->frame_size + backstep)*8);
1582 /* prepare next buffer */
1583 s->inbuf_index ^= 1;
1584 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1585 s->old_frame_size = s->frame_size;
1588 static inline void lsf_sf_expand(int *slen,
1589 int sf, int n1, int n2, int n3)
1608 static void exponents_from_scale_factors(MPADecodeContext *s,
1612 const uint8_t *bstab, *pretab;
1613 int len, i, j, k, l, v0, shift, gain, gains[3];
1616 exp_ptr = exponents;
1617 gain = g->global_gain - 210;
1618 shift = g->scalefac_scale + 1;
1620 bstab = band_size_long[s->sample_rate_index];
1621 pretab = mpa_pretab[g->preflag];
1622 for(i=0;i<g->long_end;i++) {
1623 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1629 if (g->short_start < 13) {
1630 bstab = band_size_short[s->sample_rate_index];
1631 gains[0] = gain - (g->subblock_gain[0] << 3);
1632 gains[1] = gain - (g->subblock_gain[1] << 3);
1633 gains[2] = gain - (g->subblock_gain[2] << 3);
1635 for(i=g->short_start;i<13;i++) {
1638 v0 = gains[l] - (g->scale_factors[k++] << shift);
1646 /* handle n = 0 too */
1647 static inline int get_bitsz(GetBitContext *s, int n)
1652 return get_bits(s, n);
1655 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1656 int16_t *exponents, int end_pos)
1659 int linbits, code, x, y, l, v, i, j, k, pos;
1660 GetBitContext last_gb;
1662 uint8_t *code_table;
1664 /* low frequencies (called big values) */
1667 j = g->region_size[i];
1670 /* select vlc table */
1671 k = g->table_select[i];
1672 l = mpa_huff_data[k][0];
1673 linbits = mpa_huff_data[k][1];
1675 code_table = huff_code_table[l];
1677 /* read huffcode and compute each couple */
1679 if (get_bits_count(&s->gb) >= end_pos)
1682 code = get_vlc(&s->gb, vlc);
1685 y = code_table[code];
1692 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1693 i, g->region_size[i] - j, x, y, exponents[s_index]);
1696 x += get_bitsz(&s->gb, linbits);
1697 v = l3_unscale(x, exponents[s_index]);
1698 if (get_bits1(&s->gb))
1703 g->sb_hybrid[s_index++] = v;
1706 y += get_bitsz(&s->gb, linbits);
1707 v = l3_unscale(y, exponents[s_index]);
1708 if (get_bits1(&s->gb))
1713 g->sb_hybrid[s_index++] = v;
1717 /* high frequencies */
1718 vlc = &huff_quad_vlc[g->count1table_select];
1719 last_gb.buffer = NULL;
1720 while (s_index <= 572) {
1721 pos = get_bits_count(&s->gb);
1722 if (pos >= end_pos) {
1723 if (pos > end_pos && last_gb.buffer != NULL) {
1724 /* some encoders generate an incorrect size for this
1725 part. We must go back into the data */
1733 code = get_vlc(&s->gb, vlc);
1734 dprintf("t=%d code=%d\n", g->count1table_select, code);
1738 if (code & (8 >> i)) {
1739 /* non zero value. Could use a hand coded function for
1741 v = l3_unscale(1, exponents[s_index]);
1742 if(get_bits1(&s->gb))
1747 g->sb_hybrid[s_index++] = v;
1750 while (s_index < 576)
1751 g->sb_hybrid[s_index++] = 0;
1755 /* Reorder short blocks from bitstream order to interleaved order. It
1756 would be faster to do it in parsing, but the code would be far more
1758 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1761 int32_t *ptr, *dst, *ptr1;
1764 if (g->block_type != 2)
1767 if (g->switch_point) {
1768 if (s->sample_rate_index != 8) {
1769 ptr = g->sb_hybrid + 36;
1771 ptr = g->sb_hybrid + 48;
1777 for(i=g->short_start;i<13;i++) {
1778 len = band_size_short[s->sample_rate_index][i];
1782 for(j=len;j>0;j--) {
1787 memcpy(ptr1, tmp, len * 3 * sizeof(int32_t));
1791 #define ISQRT2 FIXR(0.70710678118654752440)
1793 static void compute_stereo(MPADecodeContext *s,
1794 GranuleDef *g0, GranuleDef *g1)
1798 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1799 int32_t (*is_tab)[16];
1800 int32_t *tab0, *tab1;
1801 int non_zero_found_short[3];
1803 /* intensity stereo */
1804 if (s->mode_ext & MODE_EXT_I_STEREO) {
1809 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1813 tab0 = g0->sb_hybrid + 576;
1814 tab1 = g1->sb_hybrid + 576;
1816 non_zero_found_short[0] = 0;
1817 non_zero_found_short[1] = 0;
1818 non_zero_found_short[2] = 0;
1819 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1820 for(i = 12;i >= g1->short_start;i--) {
1821 /* for last band, use previous scale factor */
1824 len = band_size_short[s->sample_rate_index][i];
1828 if (!non_zero_found_short[l]) {
1829 /* test if non zero band. if so, stop doing i-stereo */
1830 for(j=0;j<len;j++) {
1832 non_zero_found_short[l] = 1;
1836 sf = g1->scale_factors[k + l];
1842 for(j=0;j<len;j++) {
1844 tab0[j] = MULL(tmp0, v1);
1845 tab1[j] = MULL(tmp0, v2);
1849 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1850 /* lower part of the spectrum : do ms stereo
1852 for(j=0;j<len;j++) {
1855 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1856 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1863 non_zero_found = non_zero_found_short[0] |
1864 non_zero_found_short[1] |
1865 non_zero_found_short[2];
1867 for(i = g1->long_end - 1;i >= 0;i--) {
1868 len = band_size_long[s->sample_rate_index][i];
1871 /* test if non zero band. if so, stop doing i-stereo */
1872 if (!non_zero_found) {
1873 for(j=0;j<len;j++) {
1879 /* for last band, use previous scale factor */
1880 k = (i == 21) ? 20 : i;
1881 sf = g1->scale_factors[k];
1886 for(j=0;j<len;j++) {
1888 tab0[j] = MULL(tmp0, v1);
1889 tab1[j] = MULL(tmp0, v2);
1893 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1894 /* lower part of the spectrum : do ms stereo
1896 for(j=0;j<len;j++) {
1899 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1900 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1905 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1906 /* ms stereo ONLY */
1907 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1909 tab0 = g0->sb_hybrid;
1910 tab1 = g1->sb_hybrid;
1911 for(i=0;i<576;i++) {
1914 tab0[i] = tmp0 + tmp1;
1915 tab1[i] = tmp0 - tmp1;
1920 static void compute_antialias_integer(MPADecodeContext *s,
1923 int32_t *ptr, *p0, *p1, *csa;
1926 /* we antialias only "long" bands */
1927 if (g->block_type == 2) {
1928 if (!g->switch_point)
1930 /* XXX: check this for 8000Hz case */
1936 ptr = g->sb_hybrid + 18;
1937 for(i = n;i > 0;i--) {
1940 csa = &csa_table[0][0];
1945 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1946 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1948 int64_t tmp2= MUL64(tmp0 + tmp1, csa[0]);
1949 *p0 = FRAC_RND(tmp2 - MUL64(tmp1, csa[2]));
1950 *p1 = FRAC_RND(tmp2 + MUL64(tmp0, csa[3]));
1957 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1958 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1960 tmp2= MUL64(tmp0 + tmp1, csa[0]);
1961 *p0 = FRAC_RND(tmp2 - MUL64(tmp1, csa[2]));
1962 *p1 = FRAC_RND(tmp2 + MUL64(tmp0, csa[3]));
1971 static void compute_antialias_float(MPADecodeContext *s,
1974 int32_t *ptr, *p0, *p1;
1977 /* we antialias only "long" bands */
1978 if (g->block_type == 2) {
1979 if (!g->switch_point)
1981 /* XXX: check this for 8000Hz case */
1987 ptr = g->sb_hybrid + 18;
1988 for(i = n;i > 0;i--) {
1989 float *csa = &csa_table_float[0][0];
1996 *p0 = lrintf(tmp0 * csa[0] - tmp1 * csa[1]);
1997 *p1 = lrintf(tmp0 * csa[1] + tmp1 * csa[0]);
1999 float tmp2= (tmp0 + tmp1) * csa[0];
2000 *p0 = lrintf(tmp2 - tmp1 * csa[2]);
2001 *p1 = lrintf(tmp2 + tmp0 * csa[3]);
2008 *p0 = lrintf(tmp0 * csa[0] - tmp1 * csa[1]);
2009 *p1 = lrintf(tmp0 * csa[1] + tmp1 * csa[0]);
2011 tmp2= (tmp0 + tmp1) * csa[0];
2012 *p0 = lrintf(tmp2 - tmp1 * csa[2]);
2013 *p1 = lrintf(tmp2 + tmp0 * csa[3]);
2022 static void compute_imdct(MPADecodeContext *s,
2024 int32_t *sb_samples,
2027 int32_t *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
2031 int i, j, k, mdct_long_end, v, sblimit;
2033 /* find last non zero block */
2034 ptr = g->sb_hybrid + 576;
2035 ptr1 = g->sb_hybrid + 2 * 18;
2036 while (ptr >= ptr1) {
2038 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
2042 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
2044 if (g->block_type == 2) {
2045 /* XXX: check for 8000 Hz */
2046 if (g->switch_point)
2051 mdct_long_end = sblimit;
2056 for(j=0;j<mdct_long_end;j++) {
2058 /* apply window & overlap with previous buffer */
2059 out_ptr = sb_samples + j;
2061 if (g->switch_point && j < 2)
2064 win1 = mdct_win[g->block_type];
2065 /* select frequency inversion */
2066 win = win1 + ((4 * 36) & -(j & 1));
2068 *out_ptr = MULL(out[i], win[i]) + buf[i];
2069 buf[i] = MULL(out[i + 18], win[i + 18]);
2075 for(j=mdct_long_end;j<sblimit;j++) {
2081 /* select frequency inversion */
2082 win = mdct_win[2] + ((4 * 36) & -(j & 1));
2085 /* reorder input for short mdct */
2092 /* apply 12 point window and do small overlap */
2094 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
2095 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
2100 out_ptr = sb_samples + j;
2102 *out_ptr = out[i] + buf[i];
2103 buf[i] = out[i + 18];
2110 for(j=sblimit;j<SBLIMIT;j++) {
2112 out_ptr = sb_samples + j;
2123 void sample_dump(int fnum, int32_t *tab, int n)
2125 static FILE *files[16], *f;
2132 sprintf(buf, "/tmp/out%d.%s.pcm",
2134 #ifdef USE_HIGHPRECISION
2140 f = fopen(buf, "w");
2148 printf("pos=%d\n", pos);
2150 printf(" %0.4f", (double)tab[i] / FRAC_ONE);
2157 /* normalize to 23 frac bits */
2158 v = tab[i] << (23 - FRAC_BITS);
2159 fwrite(&v, 1, sizeof(int32_t), f);
2165 /* main layer3 decoding function */
2166 static int mp_decode_layer3(MPADecodeContext *s)
2168 int nb_granules, main_data_begin, private_bits;
2169 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
2170 GranuleDef granules[2][2], *g;
2171 int16_t exponents[576];
2173 /* read side info */
2175 main_data_begin = get_bits(&s->gb, 8);
2176 if (s->nb_channels == 2)
2177 private_bits = get_bits(&s->gb, 2);
2179 private_bits = get_bits(&s->gb, 1);
2182 main_data_begin = get_bits(&s->gb, 9);
2183 if (s->nb_channels == 2)
2184 private_bits = get_bits(&s->gb, 3);
2186 private_bits = get_bits(&s->gb, 5);
2188 for(ch=0;ch<s->nb_channels;ch++) {
2189 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2190 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2194 for(gr=0;gr<nb_granules;gr++) {
2195 for(ch=0;ch<s->nb_channels;ch++) {
2196 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2197 g = &granules[ch][gr];
2198 g->part2_3_length = get_bits(&s->gb, 12);
2199 g->big_values = get_bits(&s->gb, 9);
2200 g->global_gain = get_bits(&s->gb, 8);
2201 /* if MS stereo only is selected, we precompute the
2202 1/sqrt(2) renormalization factor */
2203 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2205 g->global_gain -= 2;
2207 g->scalefac_compress = get_bits(&s->gb, 9);
2209 g->scalefac_compress = get_bits(&s->gb, 4);
2210 blocksplit_flag = get_bits(&s->gb, 1);
2211 if (blocksplit_flag) {
2212 g->block_type = get_bits(&s->gb, 2);
2213 if (g->block_type == 0)
2215 g->switch_point = get_bits(&s->gb, 1);
2217 g->table_select[i] = get_bits(&s->gb, 5);
2219 g->subblock_gain[i] = get_bits(&s->gb, 3);
2220 /* compute huffman coded region sizes */
2221 if (g->block_type == 2)
2222 g->region_size[0] = (36 / 2);
2224 if (s->sample_rate_index <= 2)
2225 g->region_size[0] = (36 / 2);
2226 else if (s->sample_rate_index != 8)
2227 g->region_size[0] = (54 / 2);
2229 g->region_size[0] = (108 / 2);
2231 g->region_size[1] = (576 / 2);
2233 int region_address1, region_address2, l;
2235 g->switch_point = 0;
2237 g->table_select[i] = get_bits(&s->gb, 5);
2238 /* compute huffman coded region sizes */
2239 region_address1 = get_bits(&s->gb, 4);
2240 region_address2 = get_bits(&s->gb, 3);
2241 dprintf("region1=%d region2=%d\n",
2242 region_address1, region_address2);
2244 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2245 l = region_address1 + region_address2 + 2;
2246 /* should not overflow */
2250 band_index_long[s->sample_rate_index][l] >> 1;
2252 /* convert region offsets to region sizes and truncate
2253 size to big_values */
2254 g->region_size[2] = (576 / 2);
2257 k = g->region_size[i];
2258 if (k > g->big_values)
2260 g->region_size[i] = k - j;
2264 /* compute band indexes */
2265 if (g->block_type == 2) {
2266 if (g->switch_point) {
2267 /* if switched mode, we handle the 36 first samples as
2268 long blocks. For 8000Hz, we handle the 48 first
2269 exponents as long blocks (XXX: check this!) */
2270 if (s->sample_rate_index <= 2)
2272 else if (s->sample_rate_index != 8)
2275 g->long_end = 4; /* 8000 Hz */
2277 if (s->sample_rate_index != 8)
2286 g->short_start = 13;
2292 g->preflag = get_bits(&s->gb, 1);
2293 g->scalefac_scale = get_bits(&s->gb, 1);
2294 g->count1table_select = get_bits(&s->gb, 1);
2295 dprintf("block_type=%d switch_point=%d\n",
2296 g->block_type, g->switch_point);
2300 /* now we get bits from the main_data_begin offset */
2301 dprintf("seekback: %d\n", main_data_begin);
2302 seek_to_maindata(s, main_data_begin);
2304 for(gr=0;gr<nb_granules;gr++) {
2305 for(ch=0;ch<s->nb_channels;ch++) {
2306 g = &granules[ch][gr];
2308 bits_pos = get_bits_count(&s->gb);
2312 int slen, slen1, slen2;
2314 /* MPEG1 scale factors */
2315 slen1 = slen_table[0][g->scalefac_compress];
2316 slen2 = slen_table[1][g->scalefac_compress];
2317 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2318 if (g->block_type == 2) {
2319 n = g->switch_point ? 17 : 18;
2322 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2324 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2326 g->scale_factors[j++] = 0;
2328 sc = granules[ch][0].scale_factors;
2331 n = (k == 0 ? 6 : 5);
2332 if ((g->scfsi & (0x8 >> k)) == 0) {
2333 slen = (k < 2) ? slen1 : slen2;
2335 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2337 /* simply copy from last granule */
2339 g->scale_factors[j] = sc[j];
2344 g->scale_factors[j++] = 0;
2348 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2351 printf(" %d", g->scale_factors[i]);
2356 int tindex, tindex2, slen[4], sl, sf;
2358 /* LSF scale factors */
2359 if (g->block_type == 2) {
2360 tindex = g->switch_point ? 2 : 1;
2364 sf = g->scalefac_compress;
2365 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2366 /* intensity stereo case */
2369 lsf_sf_expand(slen, sf, 6, 6, 0);
2371 } else if (sf < 244) {
2372 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2375 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2381 lsf_sf_expand(slen, sf, 5, 4, 4);
2383 } else if (sf < 500) {
2384 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2387 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2395 n = lsf_nsf_table[tindex2][tindex][k];
2398 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2400 /* XXX: should compute exact size */
2402 g->scale_factors[j] = 0;
2405 printf("gr=%d ch=%d scale_factors:\n",
2408 printf(" %d", g->scale_factors[i]);
2414 exponents_from_scale_factors(s, g, exponents);
2416 /* read Huffman coded residue */
2417 if (huffman_decode(s, g, exponents,
2418 bits_pos + g->part2_3_length) < 0)
2421 sample_dump(0, g->sb_hybrid, 576);
2424 /* skip extension bits */
2425 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2426 if (bits_left < 0) {
2427 dprintf("bits_left=%d\n", bits_left);
2430 while (bits_left >= 16) {
2431 skip_bits(&s->gb, 16);
2435 skip_bits(&s->gb, bits_left);
2438 if (s->nb_channels == 2)
2439 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2441 for(ch=0;ch<s->nb_channels;ch++) {
2442 g = &granules[ch][gr];
2444 reorder_block(s, g);
2446 sample_dump(0, g->sb_hybrid, 576);
2448 s->compute_antialias(s, g);
2450 sample_dump(1, g->sb_hybrid, 576);
2452 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2454 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2458 return nb_granules * 18;
2461 static int mp_decode_frame(MPADecodeContext *s,
2464 int i, nb_frames, ch;
2467 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2468 (s->inbuf_ptr - s->inbuf - HEADER_SIZE)*8);
2470 /* skip error protection field */
2471 if (s->error_protection)
2472 get_bits(&s->gb, 16);
2474 dprintf("frame %d:\n", s->frame_count);
2477 nb_frames = mp_decode_layer1(s);
2480 nb_frames = mp_decode_layer2(s);
2484 nb_frames = mp_decode_layer3(s);
2488 for(i=0;i<nb_frames;i++) {
2489 for(ch=0;ch<s->nb_channels;ch++) {
2491 printf("%d-%d:", i, ch);
2492 for(j=0;j<SBLIMIT;j++)
2493 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2498 /* apply the synthesis filter */
2499 for(ch=0;ch<s->nb_channels;ch++) {
2500 samples_ptr = samples + ch;
2501 for(i=0;i<nb_frames;i++) {
2502 synth_filter(s, ch, samples_ptr, s->nb_channels,
2503 s->sb_samples[ch][i]);
2504 samples_ptr += 32 * s->nb_channels;
2510 return nb_frames * 32 * sizeof(short) * s->nb_channels;
2513 static int decode_frame(AVCodecContext * avctx,
2514 void *data, int *data_size,
2515 uint8_t * buf, int buf_size)
2517 MPADecodeContext *s = avctx->priv_data;
2521 short *out_samples = data;
2524 while (buf_size > 0) {
2525 len = s->inbuf_ptr - s->inbuf;
2526 if (s->frame_size == 0) {
2527 /* special case for next header for first frame in free
2528 format case (XXX: find a simpler method) */
2529 if (s->free_format_next_header != 0) {
2530 s->inbuf[0] = s->free_format_next_header >> 24;
2531 s->inbuf[1] = s->free_format_next_header >> 16;
2532 s->inbuf[2] = s->free_format_next_header >> 8;
2533 s->inbuf[3] = s->free_format_next_header;
2534 s->inbuf_ptr = s->inbuf + 4;
2535 s->free_format_next_header = 0;
2538 /* no header seen : find one. We need at least HEADER_SIZE
2539 bytes to parse it */
2540 len = HEADER_SIZE - len;
2544 memcpy(s->inbuf_ptr, buf_ptr, len);
2547 s->inbuf_ptr += len;
2549 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2551 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2552 (s->inbuf[2] << 8) | s->inbuf[3];
2554 if (check_header(header) < 0) {
2555 /* no sync found : move by one byte (inefficient, but simple!) */
2556 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2558 dprintf("skip %x\n", header);
2559 /* reset free format frame size to give a chance
2560 to get a new bitrate */
2561 s->free_format_frame_size = 0;
2563 if (decode_header(s, header) == 1) {
2564 /* free format: prepare to compute frame size */
2567 /* update codec info */
2568 avctx->sample_rate = s->sample_rate;
2569 avctx->channels = s->nb_channels;
2570 avctx->bit_rate = s->bit_rate;
2571 avctx->sub_id = s->layer;
2574 avctx->frame_size = 384;
2577 avctx->frame_size = 1152;
2581 avctx->frame_size = 576;
2583 avctx->frame_size = 1152;
2588 } else if (s->frame_size == -1) {
2589 /* free format : find next sync to compute frame size */
2590 len = MPA_MAX_CODED_FRAME_SIZE - len;
2594 /* frame too long: resync */
2596 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2603 memcpy(s->inbuf_ptr, buf_ptr, len);
2604 /* check for header */
2605 p = s->inbuf_ptr - 3;
2606 pend = s->inbuf_ptr + len - 4;
2608 header = (p[0] << 24) | (p[1] << 16) |
2610 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2611 (s->inbuf[2] << 8) | s->inbuf[3];
2612 /* check with high probability that we have a
2614 if ((header & SAME_HEADER_MASK) ==
2615 (header1 & SAME_HEADER_MASK)) {
2616 /* header found: update pointers */
2617 len = (p + 4) - s->inbuf_ptr;
2621 /* compute frame size */
2622 s->free_format_next_header = header;
2623 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2624 padding = (header1 >> 9) & 1;
2626 s->free_format_frame_size -= padding * 4;
2628 s->free_format_frame_size -= padding;
2629 dprintf("free frame size=%d padding=%d\n",
2630 s->free_format_frame_size, padding);
2631 decode_header(s, header1);
2636 /* not found: simply increase pointers */
2638 s->inbuf_ptr += len;
2641 } else if (len < s->frame_size) {
2642 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2643 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2644 len = s->frame_size - len;
2647 memcpy(s->inbuf_ptr, buf_ptr, len);
2649 s->inbuf_ptr += len;
2653 if (s->frame_size > 0 &&
2654 (s->inbuf_ptr - s->inbuf) >= s->frame_size) {
2655 if (avctx->parse_only) {
2656 /* simply return the frame data */
2657 *(uint8_t **)data = s->inbuf;
2658 out_size = s->inbuf_ptr - s->inbuf;
2660 out_size = mp_decode_frame(s, out_samples);
2662 s->inbuf_ptr = s->inbuf;
2664 *data_size = out_size;
2668 return buf_ptr - buf;
2671 AVCodec mp2_decoder =
2676 sizeof(MPADecodeContext),
2681 CODEC_CAP_PARSE_ONLY,
2684 AVCodec mp3_decoder =
2689 sizeof(MPADecodeContext),
2694 CODEC_CAP_PARSE_ONLY,