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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 * @file mpegaudiodec.c
27 #include "bitstream.h"
30 /* Assume that all Intel XScale processors support armv5 edsp instructions */
31 #if defined(ARCH_ARMV4L) && defined (HAVE_IWMMXT)
37 * - in low precision mode, use more 16 bit multiplies in synth filter
38 * - test lsf / mpeg25 extensively.
41 /* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
43 #ifdef CONFIG_MPEGAUDIO_HP
44 # define USE_HIGHPRECISION
47 #include "mpegaudio.h"
49 #define FRAC_ONE (1 << FRAC_BITS)
52 # define MULL(ra, rb) \
53 ({ int rt, dummy; asm (\
55 "shrdl %4, %%edx, %%eax \n\t"\
56 : "=a"(rt), "=d"(dummy)\
57 : "a" (ra), "rm" (rb), "i"(FRAC_BITS));\
59 # define MUL64(ra, rb) \
60 ({ int64_t rt; asm ("imull %2\n\t" : "=A"(rt) : "a" (ra), "g" (rb)); rt; })
61 # define MULH(ra, rb) \
62 ({ int rt, dummy; asm ("imull %3\n\t" : "=d"(rt), "=a"(dummy): "a" (ra), "rm" (rb)); rt; })
63 #elif defined(ARCH_ARMV4L)
66 asm("smull %0, %1, %2, %3 \n\t"\
67 "mov %0, %0, lsr %4\n\t"\
68 "add %1, %0, %1, lsl %5\n\t"\
69 : "=&r"(lo), "=&r"(hi)\
70 : "r"(b), "r"(a), "i"(FRAC_BITS), "i"(32-FRAC_BITS));\
72 # define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
73 # define MULH(a, b) ({ int lo, hi; asm ("smull %0, %1, %2, %3" : "=&r"(lo), "=&r"(hi) : "r"(b), "r"(a)); hi; })
75 # define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
76 # define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
77 //#define MULH(a,b) (((int64_t)(a) * (int64_t)(b))>>32) //gcc 3.4 creates an incredibly bloated mess out of this
78 static always_inline int MULH(int a, int b){
79 return ((int64_t)(a) * (int64_t)(b))>>32;
82 #define FIX(a) ((int)((a) * FRAC_ONE))
83 /* WARNING: only correct for posititive numbers */
84 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
85 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
87 #define FIXHR(a) ((int)((a) * (1LL<<32) + 0.5))
92 #define BACKSTEP_SIZE 512
97 typedef struct MPADecodeContext {
98 DECLARE_ALIGNED_8(uint8_t, last_buf[2*BACKSTEP_SIZE + EXTRABYTES]);
101 int free_format_frame_size; /* frame size in case of free format
102 (zero if currently unknown) */
103 /* next header (used in free format parsing) */
104 uint32_t free_format_next_header;
105 int error_protection;
108 int sample_rate_index; /* between 0 and 8 */
116 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
117 int synth_buf_offset[MPA_MAX_CHANNELS];
118 int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
119 int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
123 void (*compute_antialias)(struct MPADecodeContext *s, struct GranuleDef *g);
124 int adu_mode; ///< 0 for standard mp3, 1 for adu formatted mp3
125 unsigned int dither_state;
129 * Context for MP3On4 decoder
131 typedef struct MP3On4DecodeContext {
132 int frames; ///< number of mp3 frames per block (number of mp3 decoder instances)
133 int chan_cfg; ///< channel config number
134 MPADecodeContext *mp3decctx[5]; ///< MPADecodeContext for every decoder instance
135 } MP3On4DecodeContext;
137 /* layer 3 "granule" */
138 typedef struct GranuleDef {
143 int scalefac_compress;
145 uint8_t switch_point;
147 int subblock_gain[3];
148 uint8_t scalefac_scale;
149 uint8_t count1table_select;
150 int region_size[3]; /* number of huffman codes in each region */
152 int short_start, long_end; /* long/short band indexes */
153 uint8_t scale_factors[40];
154 int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
157 #define MODE_EXT_MS_STEREO 2
158 #define MODE_EXT_I_STEREO 1
160 /* layer 3 huffman tables */
161 typedef struct HuffTable {
164 const uint16_t *codes;
167 #include "mpegaudiodectab.h"
169 static void compute_antialias_integer(MPADecodeContext *s, GranuleDef *g);
170 static void compute_antialias_float(MPADecodeContext *s, GranuleDef *g);
172 /* vlc structure for decoding layer 3 huffman tables */
173 static VLC huff_vlc[16];
174 static VLC huff_quad_vlc[2];
175 /* computed from band_size_long */
176 static uint16_t band_index_long[9][23];
177 /* XXX: free when all decoders are closed */
178 #define TABLE_4_3_SIZE (8191 + 16)*4
179 static int8_t *table_4_3_exp;
180 static uint32_t *table_4_3_value;
181 static uint32_t exp_table[512];
182 static uint32_t expval_table[512][16];
183 /* intensity stereo coef table */
184 static int32_t is_table[2][16];
185 static int32_t is_table_lsf[2][2][16];
186 static int32_t csa_table[8][4];
187 static float csa_table_float[8][4];
188 static int32_t mdct_win[8][36];
190 /* lower 2 bits: modulo 3, higher bits: shift */
191 static uint16_t scale_factor_modshift[64];
192 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
193 static int32_t scale_factor_mult[15][3];
194 /* mult table for layer 2 group quantization */
196 #define SCALE_GEN(v) \
197 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
199 static const int32_t scale_factor_mult2[3][3] = {
200 SCALE_GEN(4.0 / 3.0), /* 3 steps */
201 SCALE_GEN(4.0 / 5.0), /* 5 steps */
202 SCALE_GEN(4.0 / 9.0), /* 9 steps */
205 static MPA_INT window[512] __attribute__((aligned(16)));
207 /* layer 1 unscaling */
208 /* n = number of bits of the mantissa minus 1 */
209 static inline int l1_unscale(int n, int mant, int scale_factor)
214 shift = scale_factor_modshift[scale_factor];
217 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
219 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
220 return (int)((val + (1LL << (shift - 1))) >> shift);
223 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
227 shift = scale_factor_modshift[scale_factor];
231 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
232 /* NOTE: at this point, 0 <= shift <= 21 */
234 val = (val + (1 << (shift - 1))) >> shift;
238 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
239 static inline int l3_unscale(int value, int exponent)
244 e = table_4_3_exp [4*value + (exponent&3)];
245 m = table_4_3_value[4*value + (exponent&3)];
246 e -= (exponent >> 2);
250 m = (m + (1 << (e-1))) >> e;
255 /* all integer n^(4/3) computation code */
258 #define POW_FRAC_BITS 24
259 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
260 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
261 #define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)
263 static int dev_4_3_coefs[DEV_ORDER];
266 static int pow_mult3[3] = {
268 POW_FIX(1.25992104989487316476),
269 POW_FIX(1.58740105196819947474),
273 static void int_pow_init(void)
278 for(i=0;i<DEV_ORDER;i++) {
279 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
280 dev_4_3_coefs[i] = a;
284 #if 0 /* unused, remove? */
285 /* return the mantissa and the binary exponent */
286 static int int_pow(int i, int *exp_ptr)
294 while (a < (1 << (POW_FRAC_BITS - 1))) {
298 a -= (1 << POW_FRAC_BITS);
300 for(j = DEV_ORDER - 1; j >= 0; j--)
301 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
302 a = (1 << POW_FRAC_BITS) + a1;
303 /* exponent compute (exact) */
307 a = POW_MULL(a, pow_mult3[er]);
308 while (a >= 2 * POW_FRAC_ONE) {
312 /* convert to float */
313 while (a < POW_FRAC_ONE) {
317 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
318 #if POW_FRAC_BITS > FRAC_BITS
319 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
320 /* correct overflow */
321 if (a >= 2 * (1 << FRAC_BITS)) {
331 static int decode_init(AVCodecContext * avctx)
333 MPADecodeContext *s = avctx->priv_data;
337 #if defined(USE_HIGHPRECISION) && defined(CONFIG_AUDIO_NONSHORT)
338 avctx->sample_fmt= SAMPLE_FMT_S32;
340 avctx->sample_fmt= SAMPLE_FMT_S16;
343 if(avctx->antialias_algo != FF_AA_FLOAT)
344 s->compute_antialias= compute_antialias_integer;
346 s->compute_antialias= compute_antialias_float;
348 if (!init && !avctx->parse_only) {
349 /* scale factors table for layer 1/2 */
352 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
355 scale_factor_modshift[i] = mod | (shift << 2);
358 /* scale factor multiply for layer 1 */
362 norm = ((int64_t_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
363 scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
364 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
365 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
366 dprintf("%d: norm=%x s=%x %x %x\n",
368 scale_factor_mult[i][0],
369 scale_factor_mult[i][1],
370 scale_factor_mult[i][2]);
373 ff_mpa_synth_init(window);
375 /* huffman decode tables */
377 const HuffTable *h = &mpa_huff_tables[i];
380 uint8_t tmp_bits [512];
381 uint16_t tmp_codes[512];
383 memset(tmp_bits , 0, sizeof(tmp_bits ));
384 memset(tmp_codes, 0, sizeof(tmp_codes));
390 for(x=0;x<xsize;x++) {
391 for(y=0;y<xsize;y++){
392 tmp_bits [(x << 5) | y | ((x&&y)<<4)]= h->bits [j ];
393 tmp_codes[(x << 5) | y | ((x&&y)<<4)]= h->codes[j++];
398 init_vlc(&huff_vlc[i], 7, 512,
399 tmp_bits, 1, 1, tmp_codes, 2, 2, 1);
402 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
403 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1, 1);
409 band_index_long[i][j] = k;
410 k += band_size_long[i][j];
412 band_index_long[i][22] = k;
415 /* compute n ^ (4/3) and store it in mantissa/exp format */
416 table_4_3_exp= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_exp[0]));
419 table_4_3_value= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_value[0]));
424 for(i=1;i<TABLE_4_3_SIZE;i++) {
427 f = pow((double)(i/4), 4.0 / 3.0) * pow(2, (i&3)*0.25);
429 m = (uint32_t)(fm*(1LL<<31) + 0.5);
430 e+= FRAC_BITS - 31 + 5 - 100;
432 /* normalized to FRAC_BITS */
433 table_4_3_value[i] = m;
434 // av_log(NULL, AV_LOG_DEBUG, "%d %d %f\n", i, m, pow((double)i, 4.0 / 3.0));
435 table_4_3_exp[i] = -e;
437 for(i=0; i<512*16; i++){
438 int exponent= (i>>4);
439 double f= pow(i&15, 4.0 / 3.0) * pow(2, (exponent-400)*0.25 + FRAC_BITS + 5);
440 expval_table[exponent][i&15]= lrintf(f);
442 exp_table[exponent]= lrintf(f);
449 f = tan((double)i * M_PI / 12.0);
450 v = FIXR(f / (1.0 + f));
455 is_table[1][6 - i] = v;
459 is_table[0][i] = is_table[1][i] = 0.0;
466 e = -(j + 1) * ((i + 1) >> 1);
467 f = pow(2.0, e / 4.0);
469 is_table_lsf[j][k ^ 1][i] = FIXR(f);
470 is_table_lsf[j][k][i] = FIXR(1.0);
471 dprintf("is_table_lsf %d %d: %x %x\n",
472 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
479 cs = 1.0 / sqrt(1.0 + ci * ci);
481 csa_table[i][0] = FIXHR(cs/4);
482 csa_table[i][1] = FIXHR(ca/4);
483 csa_table[i][2] = FIXHR(ca/4) + FIXHR(cs/4);
484 csa_table[i][3] = FIXHR(ca/4) - FIXHR(cs/4);
485 csa_table_float[i][0] = cs;
486 csa_table_float[i][1] = ca;
487 csa_table_float[i][2] = ca + cs;
488 csa_table_float[i][3] = ca - cs;
489 // printf("%d %d %d %d\n", FIX(cs), FIX(cs-1), FIX(ca), FIX(cs)-FIX(ca));
490 // av_log(NULL, AV_LOG_DEBUG,"%f %f %f %f\n", cs, ca, ca+cs, ca-cs);
493 /* compute mdct windows */
501 d= sin(M_PI * (i + 0.5) / 36.0);
504 else if(i>=24) d= sin(M_PI * (i - 18 + 0.5) / 12.0);
508 else if(i< 12) d= sin(M_PI * (i - 6 + 0.5) / 12.0);
511 //merge last stage of imdct into the window coefficients
512 d*= 0.5 / cos(M_PI*(2*i + 19)/72);
515 mdct_win[j][i/3] = FIXHR((d / (1<<5)));
517 mdct_win[j][i ] = FIXHR((d / (1<<5)));
518 // av_log(NULL, AV_LOG_DEBUG, "%2d %d %f\n", i,j,d / (1<<5));
522 /* NOTE: we do frequency inversion adter the MDCT by changing
523 the sign of the right window coefs */
526 mdct_win[j + 4][i] = mdct_win[j][i];
527 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
533 av_log(avctx, AV_LOG_DEBUG, "win%d=\n", j);
535 av_log(avctx, AV_LOG_DEBUG, "%f, ", (double)mdct_win[j][i] / FRAC_ONE);
536 av_log(avctx, AV_LOG_DEBUG, "\n");
545 if (avctx->codec_id == CODEC_ID_MP3ADU)
550 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
554 #define COS0_0 FIXHR(0.50060299823519630134/2)
555 #define COS0_1 FIXHR(0.50547095989754365998/2)
556 #define COS0_2 FIXHR(0.51544730992262454697/2)
557 #define COS0_3 FIXHR(0.53104259108978417447/2)
558 #define COS0_4 FIXHR(0.55310389603444452782/2)
559 #define COS0_5 FIXHR(0.58293496820613387367/2)
560 #define COS0_6 FIXHR(0.62250412303566481615/2)
561 #define COS0_7 FIXHR(0.67480834145500574602/2)
562 #define COS0_8 FIXHR(0.74453627100229844977/2)
563 #define COS0_9 FIXHR(0.83934964541552703873/2)
564 #define COS0_10 FIXHR(0.97256823786196069369/2)
565 #define COS0_11 FIXHR(1.16943993343288495515/4)
566 #define COS0_12 FIXHR(1.48416461631416627724/4)
567 #define COS0_13 FIXHR(2.05778100995341155085/8)
568 #define COS0_14 FIXHR(3.40760841846871878570/8)
569 #define COS0_15 FIXHR(10.19000812354805681150/32)
571 #define COS1_0 FIXHR(0.50241928618815570551/2)
572 #define COS1_1 FIXHR(0.52249861493968888062/2)
573 #define COS1_2 FIXHR(0.56694403481635770368/2)
574 #define COS1_3 FIXHR(0.64682178335999012954/2)
575 #define COS1_4 FIXHR(0.78815462345125022473/2)
576 #define COS1_5 FIXHR(1.06067768599034747134/4)
577 #define COS1_6 FIXHR(1.72244709823833392782/4)
578 #define COS1_7 FIXHR(5.10114861868916385802/16)
580 #define COS2_0 FIXHR(0.50979557910415916894/2)
581 #define COS2_1 FIXHR(0.60134488693504528054/2)
582 #define COS2_2 FIXHR(0.89997622313641570463/2)
583 #define COS2_3 FIXHR(2.56291544774150617881/8)
585 #define COS3_0 FIXHR(0.54119610014619698439/2)
586 #define COS3_1 FIXHR(1.30656296487637652785/4)
588 #define COS4_0 FIXHR(0.70710678118654752439/2)
590 /* butterfly operator */
591 #define BF(a, b, c, s)\
593 tmp0 = tab[a] + tab[b];\
594 tmp1 = tab[a] - tab[b];\
596 tab[b] = MULH(tmp1<<(s), c);\
599 #define BF1(a, b, c, d)\
601 BF(a, b, COS4_0, 1);\
602 BF(c, d,-COS4_0, 1);\
606 #define BF2(a, b, c, d)\
608 BF(a, b, COS4_0, 1);\
609 BF(c, d,-COS4_0, 1);\
616 #define ADD(a, b) tab[a] += tab[b]
618 /* DCT32 without 1/sqrt(2) coef zero scaling. */
619 static void dct32(int32_t *out, int32_t *tab)
624 BF( 0, 31, COS0_0 , 1);
625 BF(15, 16, COS0_15, 5);
627 BF( 0, 15, COS1_0 , 1);
628 BF(16, 31,-COS1_0 , 1);
630 BF( 7, 24, COS0_7 , 1);
631 BF( 8, 23, COS0_8 , 1);
633 BF( 7, 8, COS1_7 , 4);
634 BF(23, 24,-COS1_7 , 4);
636 BF( 0, 7, COS2_0 , 1);
637 BF( 8, 15,-COS2_0 , 1);
638 BF(16, 23, COS2_0 , 1);
639 BF(24, 31,-COS2_0 , 1);
641 BF( 3, 28, COS0_3 , 1);
642 BF(12, 19, COS0_12, 2);
644 BF( 3, 12, COS1_3 , 1);
645 BF(19, 28,-COS1_3 , 1);
647 BF( 4, 27, COS0_4 , 1);
648 BF(11, 20, COS0_11, 2);
650 BF( 4, 11, COS1_4 , 1);
651 BF(20, 27,-COS1_4 , 1);
653 BF( 3, 4, COS2_3 , 3);
654 BF(11, 12,-COS2_3 , 3);
655 BF(19, 20, COS2_3 , 3);
656 BF(27, 28,-COS2_3 , 3);
658 BF( 0, 3, COS3_0 , 1);
659 BF( 4, 7,-COS3_0 , 1);
660 BF( 8, 11, COS3_0 , 1);
661 BF(12, 15,-COS3_0 , 1);
662 BF(16, 19, COS3_0 , 1);
663 BF(20, 23,-COS3_0 , 1);
664 BF(24, 27, COS3_0 , 1);
665 BF(28, 31,-COS3_0 , 1);
670 BF( 1, 30, COS0_1 , 1);
671 BF(14, 17, COS0_14, 3);
673 BF( 1, 14, COS1_1 , 1);
674 BF(17, 30,-COS1_1 , 1);
676 BF( 6, 25, COS0_6 , 1);
677 BF( 9, 22, COS0_9 , 1);
679 BF( 6, 9, COS1_6 , 2);
680 BF(22, 25,-COS1_6 , 2);
682 BF( 1, 6, COS2_1 , 1);
683 BF( 9, 14,-COS2_1 , 1);
684 BF(17, 22, COS2_1 , 1);
685 BF(25, 30,-COS2_1 , 1);
688 BF( 2, 29, COS0_2 , 1);
689 BF(13, 18, COS0_13, 3);
691 BF( 2, 13, COS1_2 , 1);
692 BF(18, 29,-COS1_2 , 1);
694 BF( 5, 26, COS0_5 , 1);
695 BF(10, 21, COS0_10, 1);
697 BF( 5, 10, COS1_5 , 2);
698 BF(21, 26,-COS1_5 , 2);
700 BF( 2, 5, COS2_2 , 1);
701 BF(10, 13,-COS2_2 , 1);
702 BF(18, 21, COS2_2 , 1);
703 BF(26, 29,-COS2_2 , 1);
705 BF( 1, 2, COS3_1 , 2);
706 BF( 5, 6,-COS3_1 , 2);
707 BF( 9, 10, COS3_1 , 2);
708 BF(13, 14,-COS3_1 , 2);
709 BF(17, 18, COS3_1 , 2);
710 BF(21, 22,-COS3_1 , 2);
711 BF(25, 26, COS3_1 , 2);
712 BF(29, 30,-COS3_1 , 2);
759 out[ 1] = tab[16] + tab[24];
760 out[17] = tab[17] + tab[25];
761 out[ 9] = tab[18] + tab[26];
762 out[25] = tab[19] + tab[27];
763 out[ 5] = tab[20] + tab[28];
764 out[21] = tab[21] + tab[29];
765 out[13] = tab[22] + tab[30];
766 out[29] = tab[23] + tab[31];
767 out[ 3] = tab[24] + tab[20];
768 out[19] = tab[25] + tab[21];
769 out[11] = tab[26] + tab[22];
770 out[27] = tab[27] + tab[23];
771 out[ 7] = tab[28] + tab[18];
772 out[23] = tab[29] + tab[19];
773 out[15] = tab[30] + tab[17];
779 static inline int round_sample(int *sum)
782 sum1 = (*sum) >> OUT_SHIFT;
783 *sum &= (1<<OUT_SHIFT)-1;
786 else if (sum1 > OUT_MAX)
791 # if defined(ARCH_POWERPC_405)
792 /* signed 16x16 -> 32 multiply add accumulate */
793 # define MACS(rt, ra, rb) \
794 asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
796 /* signed 16x16 -> 32 multiply */
797 # define MULS(ra, rb) \
798 ({ int __rt; asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
800 # elif defined(ARCH_ARM5E)
802 /* signed 16x16 -> 32 multiply add accumulate */
803 # define MACS(rt, ra, rb) \
804 asm ("smlabb %0, %2, %3, %0" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
806 /* signed 16x16 -> 32 multiply */
807 # define MULS(ra, rb) \
808 ({ int __rt; asm ("smulbb %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
811 /* signed 16x16 -> 32 multiply add accumulate */
812 # define MACS(rt, ra, rb) rt += (ra) * (rb)
814 /* signed 16x16 -> 32 multiply */
815 # define MULS(ra, rb) ((ra) * (rb))
819 static inline int round_sample(int64_t *sum)
822 sum1 = (int)((*sum) >> OUT_SHIFT);
823 *sum &= (1<<OUT_SHIFT)-1;
826 else if (sum1 > OUT_MAX)
831 # define MULS(ra, rb) MUL64(ra, rb)
834 #define SUM8(sum, op, w, p) \
836 sum op MULS((w)[0 * 64], p[0 * 64]);\
837 sum op MULS((w)[1 * 64], p[1 * 64]);\
838 sum op MULS((w)[2 * 64], p[2 * 64]);\
839 sum op MULS((w)[3 * 64], p[3 * 64]);\
840 sum op MULS((w)[4 * 64], p[4 * 64]);\
841 sum op MULS((w)[5 * 64], p[5 * 64]);\
842 sum op MULS((w)[6 * 64], p[6 * 64]);\
843 sum op MULS((w)[7 * 64], p[7 * 64]);\
846 #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
850 sum1 op1 MULS((w1)[0 * 64], tmp);\
851 sum2 op2 MULS((w2)[0 * 64], tmp);\
853 sum1 op1 MULS((w1)[1 * 64], tmp);\
854 sum2 op2 MULS((w2)[1 * 64], tmp);\
856 sum1 op1 MULS((w1)[2 * 64], tmp);\
857 sum2 op2 MULS((w2)[2 * 64], tmp);\
859 sum1 op1 MULS((w1)[3 * 64], tmp);\
860 sum2 op2 MULS((w2)[3 * 64], tmp);\
862 sum1 op1 MULS((w1)[4 * 64], tmp);\
863 sum2 op2 MULS((w2)[4 * 64], tmp);\
865 sum1 op1 MULS((w1)[5 * 64], tmp);\
866 sum2 op2 MULS((w2)[5 * 64], tmp);\
868 sum1 op1 MULS((w1)[6 * 64], tmp);\
869 sum2 op2 MULS((w2)[6 * 64], tmp);\
871 sum1 op1 MULS((w1)[7 * 64], tmp);\
872 sum2 op2 MULS((w2)[7 * 64], tmp);\
875 void ff_mpa_synth_init(MPA_INT *window)
879 /* max = 18760, max sum over all 16 coefs : 44736 */
884 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
894 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
896 /* XXX: optimize by avoiding ring buffer usage */
897 void ff_mpa_synth_filter(MPA_INT *synth_buf_ptr, int *synth_buf_offset,
898 MPA_INT *window, int *dither_state,
899 OUT_INT *samples, int incr,
900 int32_t sb_samples[SBLIMIT])
903 register MPA_INT *synth_buf;
904 register const MPA_INT *w, *w2, *p;
913 dct32(tmp, sb_samples);
915 offset = *synth_buf_offset;
916 synth_buf = synth_buf_ptr + offset;
921 /* NOTE: can cause a loss in precision if very high amplitude
930 /* copy to avoid wrap */
931 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
933 samples2 = samples + 31 * incr;
941 SUM8(sum, -=, w + 32, p);
942 *samples = round_sample(&sum);
946 /* we calculate two samples at the same time to avoid one memory
947 access per two sample */
950 p = synth_buf + 16 + j;
951 SUM8P2(sum, +=, sum2, -=, w, w2, p);
952 p = synth_buf + 48 - j;
953 SUM8P2(sum, -=, sum2, -=, w + 32, w2 + 32, p);
955 *samples = round_sample(&sum);
958 *samples2 = round_sample(&sum);
965 SUM8(sum, -=, w + 32, p);
966 *samples = round_sample(&sum);
969 offset = (offset - 32) & 511;
970 *synth_buf_offset = offset;
973 #define C3 FIXHR(0.86602540378443864676/2)
975 /* 0.5 / cos(pi*(2*i+1)/36) */
976 static const int icos36[9] = {
977 FIXR(0.50190991877167369479),
978 FIXR(0.51763809020504152469), //0
979 FIXR(0.55168895948124587824),
980 FIXR(0.61038729438072803416),
981 FIXR(0.70710678118654752439), //1
982 FIXR(0.87172339781054900991),
983 FIXR(1.18310079157624925896),
984 FIXR(1.93185165257813657349), //2
985 FIXR(5.73685662283492756461),
988 /* 0.5 / cos(pi*(2*i+1)/36) */
989 static const int icos36h[9] = {
990 FIXHR(0.50190991877167369479/2),
991 FIXHR(0.51763809020504152469/2), //0
992 FIXHR(0.55168895948124587824/2),
993 FIXHR(0.61038729438072803416/2),
994 FIXHR(0.70710678118654752439/2), //1
995 FIXHR(0.87172339781054900991/2),
996 FIXHR(1.18310079157624925896/4),
997 FIXHR(1.93185165257813657349/4), //2
998 // FIXHR(5.73685662283492756461),
1001 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
1003 static void imdct12(int *out, int *in)
1005 int in0, in1, in2, in3, in4, in5, t1, t2;
1008 in1= in[1*3] + in[0*3];
1009 in2= in[2*3] + in[1*3];
1010 in3= in[3*3] + in[2*3];
1011 in4= in[4*3] + in[3*3];
1012 in5= in[5*3] + in[4*3];
1016 in2= MULH(2*in2, C3);
1017 in3= MULH(4*in3, C3);
1020 t2 = MULH(2*(in1 - in5), icos36h[4]);
1030 in1 = MULH(in5 + in3, icos36h[1]);
1037 in5 = MULH(2*(in5 - in3), icos36h[7]);
1045 #define C1 FIXHR(0.98480775301220805936/2)
1046 #define C2 FIXHR(0.93969262078590838405/2)
1047 #define C3 FIXHR(0.86602540378443864676/2)
1048 #define C4 FIXHR(0.76604444311897803520/2)
1049 #define C5 FIXHR(0.64278760968653932632/2)
1050 #define C6 FIXHR(0.5/2)
1051 #define C7 FIXHR(0.34202014332566873304/2)
1052 #define C8 FIXHR(0.17364817766693034885/2)
1055 /* using Lee like decomposition followed by hand coded 9 points DCT */
1056 static void imdct36(int *out, int *buf, int *in, int *win)
1058 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
1059 int tmp[18], *tmp1, *in1;
1070 //more accurate but slower
1071 int64_t t0, t1, t2, t3;
1072 t2 = in1[2*4] + in1[2*8] - in1[2*2];
1074 t3 = (in1[2*0] + (int64_t)(in1[2*6]>>1))<<32;
1075 t1 = in1[2*0] - in1[2*6];
1076 tmp1[ 6] = t1 - (t2>>1);
1079 t0 = MUL64(2*(in1[2*2] + in1[2*4]), C2);
1080 t1 = MUL64( in1[2*4] - in1[2*8] , -2*C8);
1081 t2 = MUL64(2*(in1[2*2] + in1[2*8]), -C4);
1083 tmp1[10] = (t3 - t0 - t2) >> 32;
1084 tmp1[ 2] = (t3 + t0 + t1) >> 32;
1085 tmp1[14] = (t3 + t2 - t1) >> 32;
1087 tmp1[ 4] = MULH(2*(in1[2*5] + in1[2*7] - in1[2*1]), -C3);
1088 t2 = MUL64(2*(in1[2*1] + in1[2*5]), C1);
1089 t3 = MUL64( in1[2*5] - in1[2*7] , -2*C7);
1090 t0 = MUL64(2*in1[2*3], C3);
1092 t1 = MUL64(2*(in1[2*1] + in1[2*7]), -C5);
1094 tmp1[ 0] = (t2 + t3 + t0) >> 32;
1095 tmp1[12] = (t2 + t1 - t0) >> 32;
1096 tmp1[ 8] = (t3 - t1 - t0) >> 32;
1098 t2 = in1[2*4] + in1[2*8] - in1[2*2];
1100 t3 = in1[2*0] + (in1[2*6]>>1);
1101 t1 = in1[2*0] - in1[2*6];
1102 tmp1[ 6] = t1 - (t2>>1);
1105 t0 = MULH(2*(in1[2*2] + in1[2*4]), C2);
1106 t1 = MULH( in1[2*4] - in1[2*8] , -2*C8);
1107 t2 = MULH(2*(in1[2*2] + in1[2*8]), -C4);
1109 tmp1[10] = t3 - t0 - t2;
1110 tmp1[ 2] = t3 + t0 + t1;
1111 tmp1[14] = t3 + t2 - t1;
1113 tmp1[ 4] = MULH(2*(in1[2*5] + in1[2*7] - in1[2*1]), -C3);
1114 t2 = MULH(2*(in1[2*1] + in1[2*5]), C1);
1115 t3 = MULH( in1[2*5] - in1[2*7] , -2*C7);
1116 t0 = MULH(2*in1[2*3], C3);
1118 t1 = MULH(2*(in1[2*1] + in1[2*7]), -C5);
1120 tmp1[ 0] = t2 + t3 + t0;
1121 tmp1[12] = t2 + t1 - t0;
1122 tmp1[ 8] = t3 - t1 - t0;
1135 s1 = MULH(2*(t3 + t2), icos36h[j]);
1136 s3 = MULL(t3 - t2, icos36[8 - j]);
1140 out[(9 + j)*SBLIMIT] = MULH(t1, win[9 + j]) + buf[9 + j];
1141 out[(8 - j)*SBLIMIT] = MULH(t1, win[8 - j]) + buf[8 - j];
1142 buf[9 + j] = MULH(t0, win[18 + 9 + j]);
1143 buf[8 - j] = MULH(t0, win[18 + 8 - j]);
1147 out[(9 + 8 - j)*SBLIMIT] = MULH(t1, win[9 + 8 - j]) + buf[9 + 8 - j];
1148 out[( j)*SBLIMIT] = MULH(t1, win[ j]) + buf[ j];
1149 buf[9 + 8 - j] = MULH(t0, win[18 + 9 + 8 - j]);
1150 buf[ + j] = MULH(t0, win[18 + j]);
1155 s1 = MULH(2*tmp[17], icos36h[4]);
1158 out[(9 + 4)*SBLIMIT] = MULH(t1, win[9 + 4]) + buf[9 + 4];
1159 out[(8 - 4)*SBLIMIT] = MULH(t1, win[8 - 4]) + buf[8 - 4];
1160 buf[9 + 4] = MULH(t0, win[18 + 9 + 4]);
1161 buf[8 - 4] = MULH(t0, win[18 + 8 - 4]);
1164 /* header decoding. MUST check the header before because no
1165 consistency check is done there. Return 1 if free format found and
1166 that the frame size must be computed externally */
1167 static int decode_header(MPADecodeContext *s, uint32_t header)
1169 int sample_rate, frame_size, mpeg25, padding;
1170 int sample_rate_index, bitrate_index;
1171 if (header & (1<<20)) {
1172 s->lsf = (header & (1<<19)) ? 0 : 1;
1179 s->layer = 4 - ((header >> 17) & 3);
1180 /* extract frequency */
1181 sample_rate_index = (header >> 10) & 3;
1182 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1183 sample_rate_index += 3 * (s->lsf + mpeg25);
1184 s->sample_rate_index = sample_rate_index;
1185 s->error_protection = ((header >> 16) & 1) ^ 1;
1186 s->sample_rate = sample_rate;
1188 bitrate_index = (header >> 12) & 0xf;
1189 padding = (header >> 9) & 1;
1190 //extension = (header >> 8) & 1;
1191 s->mode = (header >> 6) & 3;
1192 s->mode_ext = (header >> 4) & 3;
1193 //copyright = (header >> 3) & 1;
1194 //original = (header >> 2) & 1;
1195 //emphasis = header & 3;
1197 if (s->mode == MPA_MONO)
1202 if (bitrate_index != 0) {
1203 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1204 s->bit_rate = frame_size * 1000;
1207 frame_size = (frame_size * 12000) / sample_rate;
1208 frame_size = (frame_size + padding) * 4;
1211 frame_size = (frame_size * 144000) / sample_rate;
1212 frame_size += padding;
1216 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1217 frame_size += padding;
1220 s->frame_size = frame_size;
1222 /* if no frame size computed, signal it */
1223 if (!s->free_format_frame_size)
1225 /* free format: compute bitrate and real frame size from the
1226 frame size we extracted by reading the bitstream */
1227 s->frame_size = s->free_format_frame_size;
1230 s->frame_size += padding * 4;
1231 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1234 s->frame_size += padding;
1235 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1239 s->frame_size += padding;
1240 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1246 dprintf("layer%d, %d Hz, %d kbits/s, ",
1247 s->layer, s->sample_rate, s->bit_rate);
1248 if (s->nb_channels == 2) {
1249 if (s->layer == 3) {
1250 if (s->mode_ext & MODE_EXT_MS_STEREO)
1252 if (s->mode_ext & MODE_EXT_I_STEREO)
1264 /* useful helper to get mpeg audio stream infos. Return -1 if error in
1265 header, otherwise the coded frame size in bytes */
1266 int mpa_decode_header(AVCodecContext *avctx, uint32_t head)
1268 MPADecodeContext s1, *s = &s1;
1270 if (ff_mpa_check_header(head) != 0)
1273 if (decode_header(s, head) != 0) {
1279 avctx->frame_size = 384;
1282 avctx->frame_size = 1152;
1287 avctx->frame_size = 576;
1289 avctx->frame_size = 1152;
1293 avctx->sample_rate = s->sample_rate;
1294 avctx->channels = s->nb_channels;
1295 avctx->bit_rate = s->bit_rate;
1296 avctx->sub_id = s->layer;
1297 return s->frame_size;
1300 /* return the number of decoded frames */
1301 static int mp_decode_layer1(MPADecodeContext *s)
1303 int bound, i, v, n, ch, j, mant;
1304 uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
1305 uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1307 if (s->mode == MPA_JSTEREO)
1308 bound = (s->mode_ext + 1) * 4;
1312 /* allocation bits */
1313 for(i=0;i<bound;i++) {
1314 for(ch=0;ch<s->nb_channels;ch++) {
1315 allocation[ch][i] = get_bits(&s->gb, 4);
1318 for(i=bound;i<SBLIMIT;i++) {
1319 allocation[0][i] = get_bits(&s->gb, 4);
1323 for(i=0;i<bound;i++) {
1324 for(ch=0;ch<s->nb_channels;ch++) {
1325 if (allocation[ch][i])
1326 scale_factors[ch][i] = get_bits(&s->gb, 6);
1329 for(i=bound;i<SBLIMIT;i++) {
1330 if (allocation[0][i]) {
1331 scale_factors[0][i] = get_bits(&s->gb, 6);
1332 scale_factors[1][i] = get_bits(&s->gb, 6);
1336 /* compute samples */
1338 for(i=0;i<bound;i++) {
1339 for(ch=0;ch<s->nb_channels;ch++) {
1340 n = allocation[ch][i];
1342 mant = get_bits(&s->gb, n + 1);
1343 v = l1_unscale(n, mant, scale_factors[ch][i]);
1347 s->sb_samples[ch][j][i] = v;
1350 for(i=bound;i<SBLIMIT;i++) {
1351 n = allocation[0][i];
1353 mant = get_bits(&s->gb, n + 1);
1354 v = l1_unscale(n, mant, scale_factors[0][i]);
1355 s->sb_samples[0][j][i] = v;
1356 v = l1_unscale(n, mant, scale_factors[1][i]);
1357 s->sb_samples[1][j][i] = v;
1359 s->sb_samples[0][j][i] = 0;
1360 s->sb_samples[1][j][i] = 0;
1367 /* bitrate is in kb/s */
1368 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1370 int ch_bitrate, table;
1372 ch_bitrate = bitrate / nb_channels;
1374 if ((freq == 48000 && ch_bitrate >= 56) ||
1375 (ch_bitrate >= 56 && ch_bitrate <= 80))
1377 else if (freq != 48000 && ch_bitrate >= 96)
1379 else if (freq != 32000 && ch_bitrate <= 48)
1389 static int mp_decode_layer2(MPADecodeContext *s)
1391 int sblimit; /* number of used subbands */
1392 const unsigned char *alloc_table;
1393 int table, bit_alloc_bits, i, j, ch, bound, v;
1394 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1395 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1396 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1397 int scale, qindex, bits, steps, k, l, m, b;
1399 /* select decoding table */
1400 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1401 s->sample_rate, s->lsf);
1402 sblimit = sblimit_table[table];
1403 alloc_table = alloc_tables[table];
1405 if (s->mode == MPA_JSTEREO)
1406 bound = (s->mode_ext + 1) * 4;
1410 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1413 if( bound > sblimit ) bound = sblimit;
1415 /* parse bit allocation */
1417 for(i=0;i<bound;i++) {
1418 bit_alloc_bits = alloc_table[j];
1419 for(ch=0;ch<s->nb_channels;ch++) {
1420 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1422 j += 1 << bit_alloc_bits;
1424 for(i=bound;i<sblimit;i++) {
1425 bit_alloc_bits = alloc_table[j];
1426 v = get_bits(&s->gb, bit_alloc_bits);
1427 bit_alloc[0][i] = v;
1428 bit_alloc[1][i] = v;
1429 j += 1 << bit_alloc_bits;
1434 for(ch=0;ch<s->nb_channels;ch++) {
1435 for(i=0;i<sblimit;i++)
1436 dprintf(" %d", bit_alloc[ch][i]);
1443 for(i=0;i<sblimit;i++) {
1444 for(ch=0;ch<s->nb_channels;ch++) {
1445 if (bit_alloc[ch][i])
1446 scale_code[ch][i] = get_bits(&s->gb, 2);
1451 for(i=0;i<sblimit;i++) {
1452 for(ch=0;ch<s->nb_channels;ch++) {
1453 if (bit_alloc[ch][i]) {
1454 sf = scale_factors[ch][i];
1455 switch(scale_code[ch][i]) {
1458 sf[0] = get_bits(&s->gb, 6);
1459 sf[1] = get_bits(&s->gb, 6);
1460 sf[2] = get_bits(&s->gb, 6);
1463 sf[0] = get_bits(&s->gb, 6);
1468 sf[0] = get_bits(&s->gb, 6);
1469 sf[2] = get_bits(&s->gb, 6);
1473 sf[0] = get_bits(&s->gb, 6);
1474 sf[2] = get_bits(&s->gb, 6);
1483 for(ch=0;ch<s->nb_channels;ch++) {
1484 for(i=0;i<sblimit;i++) {
1485 if (bit_alloc[ch][i]) {
1486 sf = scale_factors[ch][i];
1487 dprintf(" %d %d %d", sf[0], sf[1], sf[2]);
1498 for(l=0;l<12;l+=3) {
1500 for(i=0;i<bound;i++) {
1501 bit_alloc_bits = alloc_table[j];
1502 for(ch=0;ch<s->nb_channels;ch++) {
1503 b = bit_alloc[ch][i];
1505 scale = scale_factors[ch][i][k];
1506 qindex = alloc_table[j+b];
1507 bits = quant_bits[qindex];
1509 /* 3 values at the same time */
1510 v = get_bits(&s->gb, -bits);
1511 steps = quant_steps[qindex];
1512 s->sb_samples[ch][k * 12 + l + 0][i] =
1513 l2_unscale_group(steps, v % steps, scale);
1515 s->sb_samples[ch][k * 12 + l + 1][i] =
1516 l2_unscale_group(steps, v % steps, scale);
1518 s->sb_samples[ch][k * 12 + l + 2][i] =
1519 l2_unscale_group(steps, v, scale);
1522 v = get_bits(&s->gb, bits);
1523 v = l1_unscale(bits - 1, v, scale);
1524 s->sb_samples[ch][k * 12 + l + m][i] = v;
1528 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1529 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1530 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1533 /* next subband in alloc table */
1534 j += 1 << bit_alloc_bits;
1536 /* XXX: find a way to avoid this duplication of code */
1537 for(i=bound;i<sblimit;i++) {
1538 bit_alloc_bits = alloc_table[j];
1539 b = bit_alloc[0][i];
1541 int mant, scale0, scale1;
1542 scale0 = scale_factors[0][i][k];
1543 scale1 = scale_factors[1][i][k];
1544 qindex = alloc_table[j+b];
1545 bits = quant_bits[qindex];
1547 /* 3 values at the same time */
1548 v = get_bits(&s->gb, -bits);
1549 steps = quant_steps[qindex];
1552 s->sb_samples[0][k * 12 + l + 0][i] =
1553 l2_unscale_group(steps, mant, scale0);
1554 s->sb_samples[1][k * 12 + l + 0][i] =
1555 l2_unscale_group(steps, mant, scale1);
1558 s->sb_samples[0][k * 12 + l + 1][i] =
1559 l2_unscale_group(steps, mant, scale0);
1560 s->sb_samples[1][k * 12 + l + 1][i] =
1561 l2_unscale_group(steps, mant, scale1);
1562 s->sb_samples[0][k * 12 + l + 2][i] =
1563 l2_unscale_group(steps, v, scale0);
1564 s->sb_samples[1][k * 12 + l + 2][i] =
1565 l2_unscale_group(steps, v, scale1);
1568 mant = get_bits(&s->gb, bits);
1569 s->sb_samples[0][k * 12 + l + m][i] =
1570 l1_unscale(bits - 1, mant, scale0);
1571 s->sb_samples[1][k * 12 + l + m][i] =
1572 l1_unscale(bits - 1, mant, scale1);
1576 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1577 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1578 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1579 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1580 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1581 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1583 /* next subband in alloc table */
1584 j += 1 << bit_alloc_bits;
1586 /* fill remaining samples to zero */
1587 for(i=sblimit;i<SBLIMIT;i++) {
1588 for(ch=0;ch<s->nb_channels;ch++) {
1589 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1590 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1591 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1599 static inline void lsf_sf_expand(int *slen,
1600 int sf, int n1, int n2, int n3)
1619 static void exponents_from_scale_factors(MPADecodeContext *s,
1623 const uint8_t *bstab, *pretab;
1624 int len, i, j, k, l, v0, shift, gain, gains[3];
1627 exp_ptr = exponents;
1628 gain = g->global_gain - 210;
1629 shift = g->scalefac_scale + 1;
1631 bstab = band_size_long[s->sample_rate_index];
1632 pretab = mpa_pretab[g->preflag];
1633 for(i=0;i<g->long_end;i++) {
1634 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift) + 400;
1640 if (g->short_start < 13) {
1641 bstab = band_size_short[s->sample_rate_index];
1642 gains[0] = gain - (g->subblock_gain[0] << 3);
1643 gains[1] = gain - (g->subblock_gain[1] << 3);
1644 gains[2] = gain - (g->subblock_gain[2] << 3);
1646 for(i=g->short_start;i<13;i++) {
1649 v0 = gains[l] - (g->scale_factors[k++] << shift) + 400;
1657 /* handle n = 0 too */
1658 static inline int get_bitsz(GetBitContext *s, int n)
1663 return get_bits(s, n);
1666 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1667 int16_t *exponents, int end_pos2)
1671 int last_pos, bits_left;
1673 int end_pos= FFMIN(end_pos2, s->gb.size_in_bits);
1675 /* low frequencies (called big values) */
1678 int j, k, l, linbits;
1679 j = g->region_size[i];
1682 /* select vlc table */
1683 k = g->table_select[i];
1684 l = mpa_huff_data[k][0];
1685 linbits = mpa_huff_data[k][1];
1689 memset(&g->sb_hybrid[s_index], 0, sizeof(*g->sb_hybrid)*2*j);
1694 /* read huffcode and compute each couple */
1696 int exponent, x, y, v;
1697 int pos= get_bits_count(&s->gb);
1699 if (pos >= end_pos){
1700 // av_log(NULL, AV_LOG_ERROR, "pos: %d %d %d %d\n", pos, end_pos, end_pos2, s_index);
1701 if(s->in_gb.buffer && pos >= s->gb.size_in_bits){
1703 s->in_gb.buffer=NULL;
1704 assert((get_bits_count(&s->gb) & 7) == 0);
1705 skip_bits_long(&s->gb, pos - end_pos);
1707 end_pos= end_pos2 + get_bits_count(&s->gb) - pos;
1708 pos= get_bits_count(&s->gb);
1710 // av_log(NULL, AV_LOG_ERROR, "new pos: %d %d\n", pos, end_pos);
1714 y = get_vlc2(&s->gb, vlc->table, 7, 3);
1717 g->sb_hybrid[s_index ] =
1718 g->sb_hybrid[s_index+1] = 0;
1723 exponent= exponents[s_index];
1725 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1726 i, g->region_size[i] - j, x, y, exponent);
1731 v = expval_table[ exponent ][ x ];
1732 // v = expval_table[ (exponent&3) ][ x ] >> FFMIN(0 - (exponent>>2), 31);
1734 x += get_bitsz(&s->gb, linbits);
1735 v = l3_unscale(x, exponent);
1737 if (get_bits1(&s->gb))
1739 g->sb_hybrid[s_index] = v;
1741 v = expval_table[ exponent ][ y ];
1743 y += get_bitsz(&s->gb, linbits);
1744 v = l3_unscale(y, exponent);
1746 if (get_bits1(&s->gb))
1748 g->sb_hybrid[s_index+1] = v;
1754 v = expval_table[ exponent ][ x ];
1756 x += get_bitsz(&s->gb, linbits);
1757 v = l3_unscale(x, exponent);
1759 if (get_bits1(&s->gb))
1761 g->sb_hybrid[s_index+!!y] = v;
1762 g->sb_hybrid[s_index+ !y] = 0;
1768 /* high frequencies */
1769 vlc = &huff_quad_vlc[g->count1table_select];
1771 while (s_index <= 572) {
1773 pos = get_bits_count(&s->gb);
1774 if (pos >= end_pos) {
1775 if (pos > end_pos2 && last_pos){
1776 /* some encoders generate an incorrect size for this
1777 part. We must go back into the data */
1779 skip_bits_long(&s->gb, last_pos - pos);
1780 av_log(NULL, AV_LOG_INFO, "overread, skip %d enddists: %d %d\n", last_pos - pos, end_pos-pos, end_pos2-pos);
1783 // av_log(NULL, AV_LOG_ERROR, "pos2: %d %d %d %d\n", pos, end_pos, end_pos2, s_index);
1784 if(s->in_gb.buffer && pos >= s->gb.size_in_bits){
1786 s->in_gb.buffer=NULL;
1787 assert((get_bits_count(&s->gb) & 7) == 0);
1788 skip_bits_long(&s->gb, pos - end_pos);
1790 end_pos= end_pos2 + get_bits_count(&s->gb) - pos;
1791 pos= get_bits_count(&s->gb);
1793 // av_log(NULL, AV_LOG_ERROR, "new pos2: %d %d %d\n", pos, end_pos, s_index);
1799 code = get_vlc2(&s->gb, vlc->table, vlc->bits, 1);
1800 dprintf("t=%d code=%d\n", g->count1table_select, code);
1801 g->sb_hybrid[s_index+0]=
1802 g->sb_hybrid[s_index+1]=
1803 g->sb_hybrid[s_index+2]=
1804 g->sb_hybrid[s_index+3]= 0;
1806 const static int idxtab[16]={3,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0};
1808 int pos= s_index+idxtab[code];
1809 code ^= 8>>idxtab[code];
1810 v = exp_table[ exponents[pos] ];
1811 // v = exp_table[ (exponents[pos]&3) ] >> FFMIN(0 - (exponents[pos]>>2), 31);
1812 if(get_bits1(&s->gb))
1814 g->sb_hybrid[pos] = v;
1818 memset(&g->sb_hybrid[s_index], 0, sizeof(*g->sb_hybrid)*(576 - s_index));
1820 /* skip extension bits */
1821 bits_left = end_pos - get_bits_count(&s->gb);
1822 //av_log(NULL, AV_LOG_ERROR, "left:%d buf:%p\n", bits_left, s->in_gb.buffer);
1823 if (bits_left < 0) {
1824 dprintf("bits_left=%d\n", bits_left);
1827 skip_bits_long(&s->gb, bits_left);
1832 /* Reorder short blocks from bitstream order to interleaved order. It
1833 would be faster to do it in parsing, but the code would be far more
1835 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1838 int32_t *ptr, *dst, *ptr1;
1841 if (g->block_type != 2)
1844 if (g->switch_point) {
1845 if (s->sample_rate_index != 8) {
1846 ptr = g->sb_hybrid + 36;
1848 ptr = g->sb_hybrid + 48;
1854 for(i=g->short_start;i<13;i++) {
1855 len = band_size_short[s->sample_rate_index][i];
1858 for(j=len;j>0;j--) {
1859 *dst++ = ptr[0*len];
1860 *dst++ = ptr[1*len];
1861 *dst++ = ptr[2*len];
1865 memcpy(ptr1, tmp, len * 3 * sizeof(*ptr1));
1869 #define ISQRT2 FIXR(0.70710678118654752440)
1871 static void compute_stereo(MPADecodeContext *s,
1872 GranuleDef *g0, GranuleDef *g1)
1876 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1877 int32_t (*is_tab)[16];
1878 int32_t *tab0, *tab1;
1879 int non_zero_found_short[3];
1881 /* intensity stereo */
1882 if (s->mode_ext & MODE_EXT_I_STEREO) {
1887 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1891 tab0 = g0->sb_hybrid + 576;
1892 tab1 = g1->sb_hybrid + 576;
1894 non_zero_found_short[0] = 0;
1895 non_zero_found_short[1] = 0;
1896 non_zero_found_short[2] = 0;
1897 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1898 for(i = 12;i >= g1->short_start;i--) {
1899 /* for last band, use previous scale factor */
1902 len = band_size_short[s->sample_rate_index][i];
1906 if (!non_zero_found_short[l]) {
1907 /* test if non zero band. if so, stop doing i-stereo */
1908 for(j=0;j<len;j++) {
1910 non_zero_found_short[l] = 1;
1914 sf = g1->scale_factors[k + l];
1920 for(j=0;j<len;j++) {
1922 tab0[j] = MULL(tmp0, v1);
1923 tab1[j] = MULL(tmp0, v2);
1927 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1928 /* lower part of the spectrum : do ms stereo
1930 for(j=0;j<len;j++) {
1933 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1934 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1941 non_zero_found = non_zero_found_short[0] |
1942 non_zero_found_short[1] |
1943 non_zero_found_short[2];
1945 for(i = g1->long_end - 1;i >= 0;i--) {
1946 len = band_size_long[s->sample_rate_index][i];
1949 /* test if non zero band. if so, stop doing i-stereo */
1950 if (!non_zero_found) {
1951 for(j=0;j<len;j++) {
1957 /* for last band, use previous scale factor */
1958 k = (i == 21) ? 20 : i;
1959 sf = g1->scale_factors[k];
1964 for(j=0;j<len;j++) {
1966 tab0[j] = MULL(tmp0, v1);
1967 tab1[j] = MULL(tmp0, v2);
1971 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1972 /* lower part of the spectrum : do ms stereo
1974 for(j=0;j<len;j++) {
1977 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1978 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1983 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1984 /* ms stereo ONLY */
1985 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1987 tab0 = g0->sb_hybrid;
1988 tab1 = g1->sb_hybrid;
1989 for(i=0;i<576;i++) {
1992 tab0[i] = tmp0 + tmp1;
1993 tab1[i] = tmp0 - tmp1;
1998 static void compute_antialias_integer(MPADecodeContext *s,
2004 /* we antialias only "long" bands */
2005 if (g->block_type == 2) {
2006 if (!g->switch_point)
2008 /* XXX: check this for 8000Hz case */
2014 ptr = g->sb_hybrid + 18;
2015 for(i = n;i > 0;i--) {
2016 int tmp0, tmp1, tmp2;
2017 csa = &csa_table[0][0];
2021 tmp2= MULH(tmp0 + tmp1, csa[0+4*j]);\
2022 ptr[-1-j] = 4*(tmp2 - MULH(tmp1, csa[2+4*j]));\
2023 ptr[ j] = 4*(tmp2 + MULH(tmp0, csa[3+4*j]));
2038 static void compute_antialias_float(MPADecodeContext *s,
2044 /* we antialias only "long" bands */
2045 if (g->block_type == 2) {
2046 if (!g->switch_point)
2048 /* XXX: check this for 8000Hz case */
2054 ptr = g->sb_hybrid + 18;
2055 for(i = n;i > 0;i--) {
2057 float *csa = &csa_table_float[0][0];
2058 #define FLOAT_AA(j)\
2061 ptr[-1-j] = lrintf(tmp0 * csa[0+4*j] - tmp1 * csa[1+4*j]);\
2062 ptr[ j] = lrintf(tmp0 * csa[1+4*j] + tmp1 * csa[0+4*j]);
2077 static void compute_imdct(MPADecodeContext *s,
2079 int32_t *sb_samples,
2082 int32_t *ptr, *win, *win1, *buf, *out_ptr, *ptr1;
2084 int i, j, mdct_long_end, v, sblimit;
2086 /* find last non zero block */
2087 ptr = g->sb_hybrid + 576;
2088 ptr1 = g->sb_hybrid + 2 * 18;
2089 while (ptr >= ptr1) {
2091 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
2095 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
2097 if (g->block_type == 2) {
2098 /* XXX: check for 8000 Hz */
2099 if (g->switch_point)
2104 mdct_long_end = sblimit;
2109 for(j=0;j<mdct_long_end;j++) {
2110 /* apply window & overlap with previous buffer */
2111 out_ptr = sb_samples + j;
2113 if (g->switch_point && j < 2)
2116 win1 = mdct_win[g->block_type];
2117 /* select frequency inversion */
2118 win = win1 + ((4 * 36) & -(j & 1));
2119 imdct36(out_ptr, buf, ptr, win);
2120 out_ptr += 18*SBLIMIT;
2124 for(j=mdct_long_end;j<sblimit;j++) {
2125 /* select frequency inversion */
2126 win = mdct_win[2] + ((4 * 36) & -(j & 1));
2127 out_ptr = sb_samples + j;
2133 imdct12(out2, ptr + 0);
2135 *out_ptr = MULH(out2[i], win[i]) + buf[i + 6*1];
2136 buf[i + 6*2] = MULH(out2[i + 6], win[i + 6]);
2139 imdct12(out2, ptr + 1);
2141 *out_ptr = MULH(out2[i], win[i]) + buf[i + 6*2];
2142 buf[i + 6*0] = MULH(out2[i + 6], win[i + 6]);
2145 imdct12(out2, ptr + 2);
2147 buf[i + 6*0] = MULH(out2[i], win[i]) + buf[i + 6*0];
2148 buf[i + 6*1] = MULH(out2[i + 6], win[i + 6]);
2155 for(j=sblimit;j<SBLIMIT;j++) {
2157 out_ptr = sb_samples + j;
2168 void sample_dump(int fnum, int32_t *tab, int n)
2170 static FILE *files[16], *f;
2177 snprintf(buf, sizeof(buf), "/tmp/out%d.%s.pcm",
2179 #ifdef USE_HIGHPRECISION
2185 f = fopen(buf, "w");
2193 av_log(NULL, AV_LOG_DEBUG, "pos=%d\n", pos);
2195 av_log(NULL, AV_LOG_DEBUG, " %0.4f", (double)tab[i] / FRAC_ONE);
2197 av_log(NULL, AV_LOG_DEBUG, "\n");
2202 /* normalize to 23 frac bits */
2203 v = tab[i] << (23 - FRAC_BITS);
2204 fwrite(&v, 1, sizeof(int32_t), f);
2210 /* main layer3 decoding function */
2211 static int mp_decode_layer3(MPADecodeContext *s)
2213 int nb_granules, main_data_begin, private_bits;
2214 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos;
2215 GranuleDef granules[2][2], *g;
2216 int16_t exponents[576];
2218 /* read side info */
2220 main_data_begin = get_bits(&s->gb, 8);
2221 private_bits = get_bits(&s->gb, s->nb_channels);
2224 main_data_begin = get_bits(&s->gb, 9);
2225 if (s->nb_channels == 2)
2226 private_bits = get_bits(&s->gb, 3);
2228 private_bits = get_bits(&s->gb, 5);
2230 for(ch=0;ch<s->nb_channels;ch++) {
2231 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2232 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2236 for(gr=0;gr<nb_granules;gr++) {
2237 for(ch=0;ch<s->nb_channels;ch++) {
2238 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2239 g = &granules[ch][gr];
2240 g->part2_3_length = get_bits(&s->gb, 12);
2241 g->big_values = get_bits(&s->gb, 9);
2242 g->global_gain = get_bits(&s->gb, 8);
2243 /* if MS stereo only is selected, we precompute the
2244 1/sqrt(2) renormalization factor */
2245 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2247 g->global_gain -= 2;
2249 g->scalefac_compress = get_bits(&s->gb, 9);
2251 g->scalefac_compress = get_bits(&s->gb, 4);
2252 blocksplit_flag = get_bits(&s->gb, 1);
2253 if (blocksplit_flag) {
2254 g->block_type = get_bits(&s->gb, 2);
2255 if (g->block_type == 0)
2257 g->switch_point = get_bits(&s->gb, 1);
2259 g->table_select[i] = get_bits(&s->gb, 5);
2261 g->subblock_gain[i] = get_bits(&s->gb, 3);
2262 /* compute huffman coded region sizes */
2263 if (g->block_type == 2)
2264 g->region_size[0] = (36 / 2);
2266 if (s->sample_rate_index <= 2)
2267 g->region_size[0] = (36 / 2);
2268 else if (s->sample_rate_index != 8)
2269 g->region_size[0] = (54 / 2);
2271 g->region_size[0] = (108 / 2);
2273 g->region_size[1] = (576 / 2);
2275 int region_address1, region_address2, l;
2277 g->switch_point = 0;
2279 g->table_select[i] = get_bits(&s->gb, 5);
2280 /* compute huffman coded region sizes */
2281 region_address1 = get_bits(&s->gb, 4);
2282 region_address2 = get_bits(&s->gb, 3);
2283 dprintf("region1=%d region2=%d\n",
2284 region_address1, region_address2);
2286 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2287 l = region_address1 + region_address2 + 2;
2288 /* should not overflow */
2292 band_index_long[s->sample_rate_index][l] >> 1;
2294 /* convert region offsets to region sizes and truncate
2295 size to big_values */
2296 g->region_size[2] = (576 / 2);
2299 k = FFMIN(g->region_size[i], g->big_values);
2300 g->region_size[i] = k - j;
2304 /* compute band indexes */
2305 if (g->block_type == 2) {
2306 if (g->switch_point) {
2307 /* if switched mode, we handle the 36 first samples as
2308 long blocks. For 8000Hz, we handle the 48 first
2309 exponents as long blocks (XXX: check this!) */
2310 if (s->sample_rate_index <= 2)
2312 else if (s->sample_rate_index != 8)
2315 g->long_end = 4; /* 8000 Hz */
2317 g->short_start = 2 + (s->sample_rate_index != 8);
2323 g->short_start = 13;
2329 g->preflag = get_bits(&s->gb, 1);
2330 g->scalefac_scale = get_bits(&s->gb, 1);
2331 g->count1table_select = get_bits(&s->gb, 1);
2332 dprintf("block_type=%d switch_point=%d\n",
2333 g->block_type, g->switch_point);
2338 const uint8_t *ptr = s->gb.buffer + (get_bits_count(&s->gb)>>3);
2339 assert((get_bits_count(&s->gb) & 7) == 0);
2340 /* now we get bits from the main_data_begin offset */
2341 dprintf("seekback: %d\n", main_data_begin);
2342 //av_log(NULL, AV_LOG_ERROR, "backstep:%d, lastbuf:%d\n", main_data_begin, s->last_buf_size);
2343 if(main_data_begin > s->last_buf_size){
2344 av_log(NULL, AV_LOG_ERROR, "backstep:%d, lastbuf:%d\n", main_data_begin, s->last_buf_size);
2345 s->last_buf_size= main_data_begin;
2348 memcpy(s->last_buf + s->last_buf_size, ptr, EXTRABYTES);
2350 init_get_bits(&s->gb, s->last_buf + s->last_buf_size - main_data_begin, main_data_begin*8);
2353 for(gr=0;gr<nb_granules;gr++) {
2354 for(ch=0;ch<s->nb_channels;ch++) {
2355 g = &granules[ch][gr];
2357 bits_pos = get_bits_count(&s->gb);
2361 int slen, slen1, slen2;
2363 /* MPEG1 scale factors */
2364 slen1 = slen_table[0][g->scalefac_compress];
2365 slen2 = slen_table[1][g->scalefac_compress];
2366 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2367 if (g->block_type == 2) {
2368 n = g->switch_point ? 17 : 18;
2372 g->scale_factors[j++] = get_bits(&s->gb, slen1);
2375 g->scale_factors[j++] = 0;
2379 g->scale_factors[j++] = get_bits(&s->gb, slen2);
2381 g->scale_factors[j++] = 0;
2384 g->scale_factors[j++] = 0;
2387 sc = granules[ch][0].scale_factors;
2390 n = (k == 0 ? 6 : 5);
2391 if ((g->scfsi & (0x8 >> k)) == 0) {
2392 slen = (k < 2) ? slen1 : slen2;
2395 g->scale_factors[j++] = get_bits(&s->gb, slen);
2398 g->scale_factors[j++] = 0;
2401 /* simply copy from last granule */
2403 g->scale_factors[j] = sc[j];
2408 g->scale_factors[j++] = 0;
2412 dprintf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2415 dprintf(" %d", g->scale_factors[i]);
2420 int tindex, tindex2, slen[4], sl, sf;
2422 /* LSF scale factors */
2423 if (g->block_type == 2) {
2424 tindex = g->switch_point ? 2 : 1;
2428 sf = g->scalefac_compress;
2429 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2430 /* intensity stereo case */
2433 lsf_sf_expand(slen, sf, 6, 6, 0);
2435 } else if (sf < 244) {
2436 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2439 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2445 lsf_sf_expand(slen, sf, 5, 4, 4);
2447 } else if (sf < 500) {
2448 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2451 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2459 n = lsf_nsf_table[tindex2][tindex][k];
2463 g->scale_factors[j++] = get_bits(&s->gb, sl);
2466 g->scale_factors[j++] = 0;
2469 /* XXX: should compute exact size */
2471 g->scale_factors[j] = 0;
2474 dprintf("gr=%d ch=%d scale_factors:\n",
2477 dprintf(" %d", g->scale_factors[i]);
2483 exponents_from_scale_factors(s, g, exponents);
2485 /* read Huffman coded residue */
2486 if (huffman_decode(s, g, exponents,
2487 bits_pos + g->part2_3_length) < 0)
2490 sample_dump(0, g->sb_hybrid, 576);
2494 if (s->nb_channels == 2)
2495 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2497 for(ch=0;ch<s->nb_channels;ch++) {
2498 g = &granules[ch][gr];
2500 reorder_block(s, g);
2502 sample_dump(0, g->sb_hybrid, 576);
2504 s->compute_antialias(s, g);
2506 sample_dump(1, g->sb_hybrid, 576);
2508 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2510 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2514 return nb_granules * 18;
2517 static int mp_decode_frame(MPADecodeContext *s,
2518 OUT_INT *samples, const uint8_t *buf, int buf_size)
2520 int i, nb_frames, ch;
2521 OUT_INT *samples_ptr;
2523 init_get_bits(&s->gb, buf + HEADER_SIZE, (buf_size - HEADER_SIZE)*8);
2525 /* skip error protection field */
2526 if (s->error_protection)
2527 get_bits(&s->gb, 16);
2529 dprintf("frame %d:\n", s->frame_count);
2532 nb_frames = mp_decode_layer1(s);
2535 nb_frames = mp_decode_layer2(s);
2539 nb_frames = mp_decode_layer3(s);
2542 if(s->in_gb.buffer){
2543 align_get_bits(&s->gb);
2544 i= (s->gb.size_in_bits - get_bits_count(&s->gb))>>3;
2545 if(i >= 0 && i <= BACKSTEP_SIZE){
2546 memmove(s->last_buf, s->gb.buffer + (get_bits_count(&s->gb)>>3), i);
2549 av_log(NULL, AV_LOG_ERROR, "invalid old backstep %d\n", i);
2553 align_get_bits(&s->gb);
2554 assert((get_bits_count(&s->gb) & 7) == 0);
2555 i= (s->gb.size_in_bits - get_bits_count(&s->gb))>>3;
2557 if(i<0 || i > BACKSTEP_SIZE || nb_frames<0){
2558 av_log(NULL, AV_LOG_ERROR, "invalid new backstep %d\n", i);
2559 i= FFMIN(BACKSTEP_SIZE, buf_size - HEADER_SIZE);
2561 assert(i <= buf_size - HEADER_SIZE && i>= 0);
2562 memcpy(s->last_buf + s->last_buf_size, s->gb.buffer + buf_size - HEADER_SIZE - i, i);
2563 s->last_buf_size += i;
2568 for(i=0;i<nb_frames;i++) {
2569 for(ch=0;ch<s->nb_channels;ch++) {
2571 dprintf("%d-%d:", i, ch);
2572 for(j=0;j<SBLIMIT;j++)
2573 dprintf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2578 /* apply the synthesis filter */
2579 for(ch=0;ch<s->nb_channels;ch++) {
2580 samples_ptr = samples + ch;
2581 for(i=0;i<nb_frames;i++) {
2582 ff_mpa_synth_filter(s->synth_buf[ch], &(s->synth_buf_offset[ch]),
2583 window, &s->dither_state,
2584 samples_ptr, s->nb_channels,
2585 s->sb_samples[ch][i]);
2586 samples_ptr += 32 * s->nb_channels;
2592 return nb_frames * 32 * sizeof(OUT_INT) * s->nb_channels;
2595 static int decode_frame(AVCodecContext * avctx,
2596 void *data, int *data_size,
2597 uint8_t * buf, int buf_size)
2599 MPADecodeContext *s = avctx->priv_data;
2602 OUT_INT *out_samples = data;
2605 if(buf_size < HEADER_SIZE)
2608 header = (buf[0] << 24) | (buf[1] << 16) | (buf[2] << 8) | buf[3];
2609 if(ff_mpa_check_header(header) < 0){
2612 av_log(avctx, AV_LOG_ERROR, "header missing skiping one byte\n");
2616 if (decode_header(s, header) == 1) {
2617 /* free format: prepare to compute frame size */
2621 /* update codec info */
2622 avctx->sample_rate = s->sample_rate;
2623 avctx->channels = s->nb_channels;
2624 avctx->bit_rate = s->bit_rate;
2625 avctx->sub_id = s->layer;
2628 avctx->frame_size = 384;
2631 avctx->frame_size = 1152;
2635 avctx->frame_size = 576;
2637 avctx->frame_size = 1152;
2641 if(s->frame_size<=0 || s->frame_size > buf_size){
2642 av_log(avctx, AV_LOG_ERROR, "incomplete frame\n");
2644 }else if(s->frame_size < buf_size){
2645 av_log(avctx, AV_LOG_ERROR, "incorrect frame size\n");
2648 out_size = mp_decode_frame(s, out_samples, buf, buf_size);
2650 *data_size = out_size;
2652 av_log(avctx, AV_LOG_DEBUG, "Error while decoding mpeg audio frame\n"); //FIXME return -1 / but also return the number of bytes consumed
2658 static int decode_frame_adu(AVCodecContext * avctx,
2659 void *data, int *data_size,
2660 uint8_t * buf, int buf_size)
2662 MPADecodeContext *s = avctx->priv_data;
2665 OUT_INT *out_samples = data;
2669 // Discard too short frames
2670 if (buf_size < HEADER_SIZE) {
2676 if (len > MPA_MAX_CODED_FRAME_SIZE)
2677 len = MPA_MAX_CODED_FRAME_SIZE;
2679 // Get header and restore sync word
2680 header = (buf[0] << 24) | (buf[1] << 16) | (buf[2] << 8) | buf[3] | 0xffe00000;
2682 if (ff_mpa_check_header(header) < 0) { // Bad header, discard frame
2687 decode_header(s, header);
2688 /* update codec info */
2689 avctx->sample_rate = s->sample_rate;
2690 avctx->channels = s->nb_channels;
2691 avctx->bit_rate = s->bit_rate;
2692 avctx->sub_id = s->layer;
2694 avctx->frame_size=s->frame_size = len;
2696 if (avctx->parse_only) {
2697 out_size = buf_size;
2699 out_size = mp_decode_frame(s, out_samples, buf, buf_size);
2702 *data_size = out_size;
2707 /* Next 3 arrays are indexed by channel config number (passed via codecdata) */
2708 static int mp3Frames[16] = {0,1,1,2,3,3,4,5,2}; /* number of mp3 decoder instances */
2709 static int mp3Channels[16] = {0,1,2,3,4,5,6,8,4}; /* total output channels */
2710 /* offsets into output buffer, assume output order is FL FR BL BR C LFE */
2711 static int chan_offset[9][5] = {
2716 {2,0,3}, // C FLR BS
2717 {4,0,2}, // C FLR BLRS
2718 {4,0,2,5}, // C FLR BLRS LFE
2719 {4,0,2,6,5}, // C FLR BLRS BLR LFE
2724 static int decode_init_mp3on4(AVCodecContext * avctx)
2726 MP3On4DecodeContext *s = avctx->priv_data;
2729 if ((avctx->extradata_size < 2) || (avctx->extradata == NULL)) {
2730 av_log(avctx, AV_LOG_ERROR, "Codec extradata missing or too short.\n");
2734 s->chan_cfg = (((unsigned char *)avctx->extradata)[1] >> 3) & 0x0f;
2735 s->frames = mp3Frames[s->chan_cfg];
2737 av_log(avctx, AV_LOG_ERROR, "Invalid channel config number.\n");
2740 avctx->channels = mp3Channels[s->chan_cfg];
2742 /* Init the first mp3 decoder in standard way, so that all tables get builded
2743 * We replace avctx->priv_data with the context of the first decoder so that
2744 * decode_init() does not have to be changed.
2745 * Other decoders will be inited here copying data from the first context
2747 // Allocate zeroed memory for the first decoder context
2748 s->mp3decctx[0] = av_mallocz(sizeof(MPADecodeContext));
2749 // Put decoder context in place to make init_decode() happy
2750 avctx->priv_data = s->mp3decctx[0];
2752 // Restore mp3on4 context pointer
2753 avctx->priv_data = s;
2754 s->mp3decctx[0]->adu_mode = 1; // Set adu mode
2756 /* Create a separate codec/context for each frame (first is already ok).
2757 * Each frame is 1 or 2 channels - up to 5 frames allowed
2759 for (i = 1; i < s->frames; i++) {
2760 s->mp3decctx[i] = av_mallocz(sizeof(MPADecodeContext));
2761 s->mp3decctx[i]->compute_antialias = s->mp3decctx[0]->compute_antialias;
2762 s->mp3decctx[i]->adu_mode = 1;
2769 static int decode_close_mp3on4(AVCodecContext * avctx)
2771 MP3On4DecodeContext *s = avctx->priv_data;
2774 for (i = 0; i < s->frames; i++)
2775 if (s->mp3decctx[i])
2776 av_free(s->mp3decctx[i]);
2782 static int decode_frame_mp3on4(AVCodecContext * avctx,
2783 void *data, int *data_size,
2784 uint8_t * buf, int buf_size)
2786 MP3On4DecodeContext *s = avctx->priv_data;
2787 MPADecodeContext *m;
2788 int len, out_size = 0;
2790 OUT_INT *out_samples = data;
2791 OUT_INT decoded_buf[MPA_FRAME_SIZE * MPA_MAX_CHANNELS];
2792 OUT_INT *outptr, *bp;
2794 unsigned char *start2 = buf, *start;
2796 int off = avctx->channels;
2797 int *coff = chan_offset[s->chan_cfg];
2801 // Discard too short frames
2802 if (buf_size < HEADER_SIZE) {
2807 // If only one decoder interleave is not needed
2808 outptr = s->frames == 1 ? out_samples : decoded_buf;
2810 for (fr = 0; fr < s->frames; fr++) {
2812 fsize = (start[0] << 4) | (start[1] >> 4);
2817 if (fsize > MPA_MAX_CODED_FRAME_SIZE)
2818 fsize = MPA_MAX_CODED_FRAME_SIZE;
2819 m = s->mp3decctx[fr];
2823 header = (start[0] << 24) | (start[1] << 16) | (start[2] << 8) | start[3] | 0xfff00000;
2825 if (ff_mpa_check_header(header) < 0) { // Bad header, discard block
2830 decode_header(m, header);
2831 mp_decode_frame(m, decoded_buf, start, fsize);
2833 n = MPA_FRAME_SIZE * m->nb_channels;
2834 out_size += n * sizeof(OUT_INT);
2836 /* interleave output data */
2837 bp = out_samples + coff[fr];
2838 if(m->nb_channels == 1) {
2839 for(j = 0; j < n; j++) {
2840 *bp = decoded_buf[j];
2844 for(j = 0; j < n; j++) {
2845 bp[0] = decoded_buf[j++];
2846 bp[1] = decoded_buf[j];
2853 /* update codec info */
2854 avctx->sample_rate = s->mp3decctx[0]->sample_rate;
2855 avctx->frame_size= buf_size;
2856 avctx->bit_rate = 0;
2857 for (i = 0; i < s->frames; i++)
2858 avctx->bit_rate += s->mp3decctx[i]->bit_rate;
2860 *data_size = out_size;
2865 AVCodec mp2_decoder =
2870 sizeof(MPADecodeContext),
2875 CODEC_CAP_PARSE_ONLY,
2878 AVCodec mp3_decoder =
2883 sizeof(MPADecodeContext),
2888 CODEC_CAP_PARSE_ONLY,
2891 AVCodec mp3adu_decoder =
2896 sizeof(MPADecodeContext),
2901 CODEC_CAP_PARSE_ONLY,
2904 AVCodec mp3on4_decoder =
2909 sizeof(MP3On4DecodeContext),
2912 decode_close_mp3on4,
2913 decode_frame_mp3on4,