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 "bitstream.h"
28 #include "mpegaudio.h"
33 * - in low precision mode, use more 16 bit multiplies in synth filter
34 * - test lsf / mpeg25 extensively.
37 /* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
39 #ifdef CONFIG_MPEGAUDIO_HP
40 #define USE_HIGHPRECISION
43 #ifdef USE_HIGHPRECISION
44 #define FRAC_BITS 23 /* fractional bits for sb_samples and dct */
45 #define WFRAC_BITS 16 /* fractional bits for window */
47 #define FRAC_BITS 15 /* fractional bits for sb_samples and dct */
48 #define WFRAC_BITS 14 /* fractional bits for window */
51 #if defined(USE_HIGHPRECISION) && defined(CONFIG_AUDIO_NONSHORT)
52 typedef int32_t OUT_INT;
53 #define OUT_MAX INT32_MAX
54 #define OUT_MIN INT32_MIN
55 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 31)
57 typedef int16_t OUT_INT;
58 #define OUT_MAX INT16_MAX
59 #define OUT_MIN INT16_MIN
60 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
63 #define FRAC_ONE (1 << FRAC_BITS)
65 #define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
66 #define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
67 #define FIX(a) ((int)((a) * FRAC_ONE))
68 /* WARNING: only correct for posititive numbers */
69 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
70 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
72 #define FIXHR(a) ((int)((a) * (1LL<<32) + 0.5))
73 //#define MULH(a,b) (((int64_t)(a) * (int64_t)(b))>>32) //gcc 3.4 creates an incredibly bloated mess out of this
74 static always_inline int MULH(int a, int b){
75 return ((int64_t)(a) * (int64_t)(b))>>32;
79 typedef int16_t MPA_INT;
81 typedef int32_t MPA_INT;
87 #define BACKSTEP_SIZE 512
91 typedef struct MPADecodeContext {
92 uint8_t inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
94 uint8_t *inbuf_ptr, *inbuf;
96 int free_format_frame_size; /* frame size in case of free format
97 (zero if currently unknown) */
98 /* next header (used in free format parsing) */
99 uint32_t free_format_next_header;
100 int error_protection;
103 int sample_rate_index; /* between 0 and 8 */
111 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
112 int synth_buf_offset[MPA_MAX_CHANNELS];
113 int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
114 int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
118 void (*compute_antialias)(struct MPADecodeContext *s, struct GranuleDef *g);
119 int adu_mode; ///< 0 for standard mp3, 1 for adu formatted mp3
120 unsigned int dither_state;
123 /* layer 3 "granule" */
124 typedef struct GranuleDef {
129 int scalefac_compress;
131 uint8_t switch_point;
133 int subblock_gain[3];
134 uint8_t scalefac_scale;
135 uint8_t count1table_select;
136 int region_size[3]; /* number of huffman codes in each region */
138 int short_start, long_end; /* long/short band indexes */
139 uint8_t scale_factors[40];
140 int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
143 #define MODE_EXT_MS_STEREO 2
144 #define MODE_EXT_I_STEREO 1
146 /* layer 3 huffman tables */
147 typedef struct HuffTable {
150 const uint16_t *codes;
153 #include "mpegaudiodectab.h"
155 static void compute_antialias_integer(MPADecodeContext *s, GranuleDef *g);
156 static void compute_antialias_float(MPADecodeContext *s, GranuleDef *g);
158 /* vlc structure for decoding layer 3 huffman tables */
159 static VLC huff_vlc[16];
160 static uint8_t *huff_code_table[16];
161 static VLC huff_quad_vlc[2];
162 /* computed from band_size_long */
163 static uint16_t band_index_long[9][23];
164 /* XXX: free when all decoders are closed */
165 #define TABLE_4_3_SIZE (8191 + 16)*4
166 static int8_t *table_4_3_exp;
167 static uint32_t *table_4_3_value;
168 /* intensity stereo coef table */
169 static int32_t is_table[2][16];
170 static int32_t is_table_lsf[2][2][16];
171 static int32_t csa_table[8][4];
172 static float csa_table_float[8][4];
173 static int32_t mdct_win[8][36];
175 /* lower 2 bits: modulo 3, higher bits: shift */
176 static uint16_t scale_factor_modshift[64];
177 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
178 static int32_t scale_factor_mult[15][3];
179 /* mult table for layer 2 group quantization */
181 #define SCALE_GEN(v) \
182 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
184 static int32_t scale_factor_mult2[3][3] = {
185 SCALE_GEN(4.0 / 3.0), /* 3 steps */
186 SCALE_GEN(4.0 / 5.0), /* 5 steps */
187 SCALE_GEN(4.0 / 9.0), /* 9 steps */
190 void ff_mpa_synth_init(MPA_INT *window);
191 static MPA_INT window[512] __attribute__((aligned(16)));
193 /* layer 1 unscaling */
194 /* n = number of bits of the mantissa minus 1 */
195 static inline int l1_unscale(int n, int mant, int scale_factor)
200 shift = scale_factor_modshift[scale_factor];
203 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
205 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
206 return (int)((val + (1LL << (shift - 1))) >> shift);
209 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
213 shift = scale_factor_modshift[scale_factor];
217 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
218 /* NOTE: at this point, 0 <= shift <= 21 */
220 val = (val + (1 << (shift - 1))) >> shift;
224 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
225 static inline int l3_unscale(int value, int exponent)
230 e = table_4_3_exp [4*value + (exponent&3)];
231 m = table_4_3_value[4*value + (exponent&3)];
232 e -= (exponent >> 2);
236 m = (m + (1 << (e-1))) >> e;
241 /* all integer n^(4/3) computation code */
244 #define POW_FRAC_BITS 24
245 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
246 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
247 #define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)
249 static int dev_4_3_coefs[DEV_ORDER];
251 static int pow_mult3[3] = {
253 POW_FIX(1.25992104989487316476),
254 POW_FIX(1.58740105196819947474),
257 static void int_pow_init(void)
262 for(i=0;i<DEV_ORDER;i++) {
263 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
264 dev_4_3_coefs[i] = a;
268 /* return the mantissa and the binary exponent */
269 static int int_pow(int i, int *exp_ptr)
277 while (a < (1 << (POW_FRAC_BITS - 1))) {
281 a -= (1 << POW_FRAC_BITS);
283 for(j = DEV_ORDER - 1; j >= 0; j--)
284 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
285 a = (1 << POW_FRAC_BITS) + a1;
286 /* exponent compute (exact) */
290 a = POW_MULL(a, pow_mult3[er]);
291 while (a >= 2 * POW_FRAC_ONE) {
295 /* convert to float */
296 while (a < POW_FRAC_ONE) {
300 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
301 #if POW_FRAC_BITS > FRAC_BITS
302 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
303 /* correct overflow */
304 if (a >= 2 * (1 << FRAC_BITS)) {
313 static int decode_init(AVCodecContext * avctx)
315 MPADecodeContext *s = avctx->priv_data;
319 #if defined(USE_HIGHPRECISION) && defined(CONFIG_AUDIO_NONSHORT)
320 avctx->sample_fmt= SAMPLE_FMT_S32;
322 avctx->sample_fmt= SAMPLE_FMT_S16;
325 if(avctx->antialias_algo != FF_AA_FLOAT)
326 s->compute_antialias= compute_antialias_integer;
328 s->compute_antialias= compute_antialias_float;
330 if (!init && !avctx->parse_only) {
331 /* scale factors table for layer 1/2 */
334 /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
337 scale_factor_modshift[i] = mod | (shift << 2);
340 /* scale factor multiply for layer 1 */
344 norm = ((int64_t_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
345 scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
346 scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
347 scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
348 dprintf("%d: norm=%x s=%x %x %x\n",
350 scale_factor_mult[i][0],
351 scale_factor_mult[i][1],
352 scale_factor_mult[i][2]);
355 ff_mpa_synth_init(window);
357 /* huffman decode tables */
358 huff_code_table[0] = NULL;
360 const HuffTable *h = &mpa_huff_tables[i];
368 init_vlc(&huff_vlc[i], 8, n,
369 h->bits, 1, 1, h->codes, 2, 2, 1);
371 code_table = av_mallocz(n);
373 for(x=0;x<xsize;x++) {
375 code_table[j++] = (x << 4) | y;
377 huff_code_table[i] = code_table;
380 init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
381 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1, 1);
387 band_index_long[i][j] = k;
388 k += band_size_long[i][j];
390 band_index_long[i][22] = k;
393 /* compute n ^ (4/3) and store it in mantissa/exp format */
394 table_4_3_exp= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_exp[0]));
397 table_4_3_value= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_value[0]));
402 for(i=1;i<TABLE_4_3_SIZE;i++) {
405 f = pow((double)(i/4), 4.0 / 3.0) * pow(2, (i&3)*0.25);
410 /* normalized to FRAC_BITS */
411 table_4_3_value[i] = m;
412 // av_log(NULL, AV_LOG_DEBUG, "%d %d %f\n", i, m, pow((double)i, 4.0 / 3.0));
413 table_4_3_exp[i] = -e;
420 f = tan((double)i * M_PI / 12.0);
421 v = FIXR(f / (1.0 + f));
426 is_table[1][6 - i] = v;
430 is_table[0][i] = is_table[1][i] = 0.0;
437 e = -(j + 1) * ((i + 1) >> 1);
438 f = pow(2.0, e / 4.0);
440 is_table_lsf[j][k ^ 1][i] = FIXR(f);
441 is_table_lsf[j][k][i] = FIXR(1.0);
442 dprintf("is_table_lsf %d %d: %x %x\n",
443 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
450 cs = 1.0 / sqrt(1.0 + ci * ci);
452 csa_table[i][0] = FIXHR(cs/4);
453 csa_table[i][1] = FIXHR(ca/4);
454 csa_table[i][2] = FIXHR(ca/4) + FIXHR(cs/4);
455 csa_table[i][3] = FIXHR(ca/4) - FIXHR(cs/4);
456 csa_table_float[i][0] = cs;
457 csa_table_float[i][1] = ca;
458 csa_table_float[i][2] = ca + cs;
459 csa_table_float[i][3] = ca - cs;
460 // printf("%d %d %d %d\n", FIX(cs), FIX(cs-1), FIX(ca), FIX(cs)-FIX(ca));
461 // av_log(NULL, AV_LOG_DEBUG,"%f %f %f %f\n", cs, ca, ca+cs, ca-cs);
464 /* compute mdct windows */
470 d= sin(M_PI * (i + 0.5) / 36.0);
473 else if(i>=24) d= sin(M_PI * (i - 18 + 0.5) / 12.0);
477 else if(i< 12) d= sin(M_PI * (i - 6 + 0.5) / 12.0);
480 //merge last stage of imdct into the window coefficients
481 if (i/9 == 0) d*= 0.5 / cos(M_PI*(2*( i) +19)/72);
482 else if(i/9 == 1) d*= 0.5 / cos(M_PI*(2*(17 - i) +19)/72);
483 else if(i/9 == 2) d*= 0.5 / cos(M_PI*(2*( i) +19)/72);
484 else d*=-0.5 / cos(M_PI*(2*(17 - i) +19)/72);
485 mdct_win[j][i] = FIXHR((d / (1<<5)));
486 // av_log(NULL, AV_LOG_DEBUG, "%2d %d %f\n", i,j,d / (1<<5));
491 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
493 /* NOTE: we do frequency inversion adter the MDCT by changing
494 the sign of the right window coefs */
497 mdct_win[j + 4][i] = mdct_win[j][i];
498 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
504 printf("win%d=\n", j);
506 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
514 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
515 s->inbuf_ptr = s->inbuf;
519 if (avctx->codec_id == CODEC_ID_MP3ADU)
524 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
528 #define COS0_0 FIXR(0.50060299823519630134)
529 #define COS0_1 FIXR(0.50547095989754365998)
530 #define COS0_2 FIXR(0.51544730992262454697)
531 #define COS0_3 FIXR(0.53104259108978417447)
532 #define COS0_4 FIXR(0.55310389603444452782)
533 #define COS0_5 FIXR(0.58293496820613387367)
534 #define COS0_6 FIXR(0.62250412303566481615)
535 #define COS0_7 FIXR(0.67480834145500574602)
536 #define COS0_8 FIXR(0.74453627100229844977)
537 #define COS0_9 FIXR(0.83934964541552703873)
538 #define COS0_10 FIXR(0.97256823786196069369)
539 #define COS0_11 FIXR(1.16943993343288495515)
540 #define COS0_12 FIXR(1.48416461631416627724)
541 #define COS0_13 FIXR(2.05778100995341155085)
542 #define COS0_14 FIXR(3.40760841846871878570)
543 #define COS0_15 FIXR(10.19000812354805681150)
545 #define COS1_0 FIXR(0.50241928618815570551)
546 #define COS1_1 FIXR(0.52249861493968888062)
547 #define COS1_2 FIXR(0.56694403481635770368)
548 #define COS1_3 FIXR(0.64682178335999012954)
549 #define COS1_4 FIXR(0.78815462345125022473)
550 #define COS1_5 FIXR(1.06067768599034747134)
551 #define COS1_6 FIXR(1.72244709823833392782)
552 #define COS1_7 FIXR(5.10114861868916385802)
554 #define COS2_0 FIXR(0.50979557910415916894)
555 #define COS2_1 FIXR(0.60134488693504528054)
556 #define COS2_2 FIXR(0.89997622313641570463)
557 #define COS2_3 FIXR(2.56291544774150617881)
559 #define COS3_0 FIXR(0.54119610014619698439)
560 #define COS3_1 FIXR(1.30656296487637652785)
562 #define COS4_0 FIXR(0.70710678118654752439)
564 /* butterfly operator */
567 tmp0 = tab[a] + tab[b];\
568 tmp1 = tab[a] - tab[b];\
570 tab[b] = MULL(tmp1, c);\
573 #define BF1(a, b, c, d)\
580 #define BF2(a, b, c, d)\
590 #define ADD(a, b) tab[a] += tab[b]
592 /* DCT32 without 1/sqrt(2) coef zero scaling. */
593 static void dct32(int32_t *out, int32_t *tab)
725 out[ 1] = tab[16] + tab[24];
726 out[17] = tab[17] + tab[25];
727 out[ 9] = tab[18] + tab[26];
728 out[25] = tab[19] + tab[27];
729 out[ 5] = tab[20] + tab[28];
730 out[21] = tab[21] + tab[29];
731 out[13] = tab[22] + tab[30];
732 out[29] = tab[23] + tab[31];
733 out[ 3] = tab[24] + tab[20];
734 out[19] = tab[25] + tab[21];
735 out[11] = tab[26] + tab[22];
736 out[27] = tab[27] + tab[23];
737 out[ 7] = tab[28] + tab[18];
738 out[23] = tab[29] + tab[19];
739 out[15] = tab[30] + tab[17];
745 static inline int round_sample(int *sum)
748 sum1 = (*sum) >> OUT_SHIFT;
749 *sum &= (1<<OUT_SHIFT)-1;
752 else if (sum1 > OUT_MAX)
757 #if defined(ARCH_POWERPC_405)
759 /* signed 16x16 -> 32 multiply add accumulate */
760 #define MACS(rt, ra, rb) \
761 asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
763 /* signed 16x16 -> 32 multiply */
764 #define MULS(ra, rb) \
765 ({ int __rt; asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
769 /* signed 16x16 -> 32 multiply add accumulate */
770 #define MACS(rt, ra, rb) rt += (ra) * (rb)
772 /* signed 16x16 -> 32 multiply */
773 #define MULS(ra, rb) ((ra) * (rb))
779 static inline int round_sample(int64_t *sum)
782 sum1 = (int)((*sum) >> OUT_SHIFT);
783 *sum &= (1<<OUT_SHIFT)-1;
786 else if (sum1 > OUT_MAX)
791 #define MULS(ra, rb) MUL64(ra, rb)
795 #define SUM8(sum, op, w, p) \
797 sum op MULS((w)[0 * 64], p[0 * 64]);\
798 sum op MULS((w)[1 * 64], p[1 * 64]);\
799 sum op MULS((w)[2 * 64], p[2 * 64]);\
800 sum op MULS((w)[3 * 64], p[3 * 64]);\
801 sum op MULS((w)[4 * 64], p[4 * 64]);\
802 sum op MULS((w)[5 * 64], p[5 * 64]);\
803 sum op MULS((w)[6 * 64], p[6 * 64]);\
804 sum op MULS((w)[7 * 64], p[7 * 64]);\
807 #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
811 sum1 op1 MULS((w1)[0 * 64], tmp);\
812 sum2 op2 MULS((w2)[0 * 64], tmp);\
814 sum1 op1 MULS((w1)[1 * 64], tmp);\
815 sum2 op2 MULS((w2)[1 * 64], tmp);\
817 sum1 op1 MULS((w1)[2 * 64], tmp);\
818 sum2 op2 MULS((w2)[2 * 64], tmp);\
820 sum1 op1 MULS((w1)[3 * 64], tmp);\
821 sum2 op2 MULS((w2)[3 * 64], tmp);\
823 sum1 op1 MULS((w1)[4 * 64], tmp);\
824 sum2 op2 MULS((w2)[4 * 64], tmp);\
826 sum1 op1 MULS((w1)[5 * 64], tmp);\
827 sum2 op2 MULS((w2)[5 * 64], tmp);\
829 sum1 op1 MULS((w1)[6 * 64], tmp);\
830 sum2 op2 MULS((w2)[6 * 64], tmp);\
832 sum1 op1 MULS((w1)[7 * 64], tmp);\
833 sum2 op2 MULS((w2)[7 * 64], tmp);\
836 void ff_mpa_synth_init(MPA_INT *window)
840 /* max = 18760, max sum over all 16 coefs : 44736 */
845 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
855 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
857 /* XXX: optimize by avoiding ring buffer usage */
858 void ff_mpa_synth_filter(MPA_INT *synth_buf_ptr, int *synth_buf_offset,
859 MPA_INT *window, int *dither_state,
860 OUT_INT *samples, int incr,
861 int32_t sb_samples[SBLIMIT])
864 register MPA_INT *synth_buf;
865 register const MPA_INT *w, *w2, *p;
874 dct32(tmp, sb_samples);
876 offset = *synth_buf_offset;
877 synth_buf = synth_buf_ptr + offset;
882 /* NOTE: can cause a loss in precision if very high amplitude
891 /* copy to avoid wrap */
892 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
894 samples2 = samples + 31 * incr;
902 SUM8(sum, -=, w + 32, p);
903 *samples = round_sample(&sum);
907 /* we calculate two samples at the same time to avoid one memory
908 access per two sample */
911 p = synth_buf + 16 + j;
912 SUM8P2(sum, +=, sum2, -=, w, w2, p);
913 p = synth_buf + 48 - j;
914 SUM8P2(sum, -=, sum2, -=, w + 32, w2 + 32, p);
916 *samples = round_sample(&sum);
919 *samples2 = round_sample(&sum);
926 SUM8(sum, -=, w + 32, p);
927 *samples = round_sample(&sum);
930 offset = (offset - 32) & 511;
931 *synth_buf_offset = offset;
935 #define C1 FIXR(0.99144486137381041114)
936 #define C3 FIXR(0.92387953251128675612)
937 #define C5 FIXR(0.79335334029123516458)
938 #define C7 FIXR(0.60876142900872063941)
939 #define C9 FIXR(0.38268343236508977173)
940 #define C11 FIXR(0.13052619222005159154)
942 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
944 static void imdct12(int *out, int *in)
947 int64_t in1_3, in1_9, in4_3, in4_9;
949 in1_3 = MUL64(in[1], C3);
950 in1_9 = MUL64(in[1], C9);
951 in4_3 = MUL64(in[4], C3);
952 in4_9 = MUL64(in[4], C9);
954 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
955 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
958 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
959 MUL64(in[2] + in[5], C3) - in4_9);
962 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
963 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
966 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
967 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
970 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
971 MUL64(in[2] + in[5], C9) + in4_3);
974 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
975 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
988 #define C1 FIXHR(0.98480775301220805936/2)
989 #define C2 FIXHR(0.93969262078590838405/2)
990 #define C3 FIXHR(0.86602540378443864676/2)
991 #define C4 FIXHR(0.76604444311897803520/2)
992 #define C5 FIXHR(0.64278760968653932632/2)
993 #define C6 FIXHR(0.5/2)
994 #define C7 FIXHR(0.34202014332566873304/2)
995 #define C8 FIXHR(0.17364817766693034885/2)
998 /* 0.5 / cos(pi*(2*i+1)/36) */
999 static const int icos36[9] = {
1000 FIXR(0.50190991877167369479),
1001 FIXR(0.51763809020504152469),
1002 FIXR(0.55168895948124587824),
1003 FIXR(0.61038729438072803416),
1004 FIXR(0.70710678118654752439),
1005 FIXR(0.87172339781054900991),
1006 FIXR(1.18310079157624925896),
1007 FIXR(1.93185165257813657349),
1008 FIXR(5.73685662283492756461),
1010 /* using Lee like decomposition followed by hand coded 9 points DCT */
1011 static void imdct36(int *out, int *buf, int *in, int *win)
1013 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
1014 int tmp[18], *tmp1, *in1;
1025 //more accurate but slower
1026 int64_t t0, t1, t2, t3;
1027 t2 = in1[2*4] + in1[2*8] - in1[2*2];
1029 t3 = (in1[2*0] + (int64_t)(in1[2*6]>>1))<<32;
1030 t1 = in1[2*0] - in1[2*6];
1031 tmp1[ 6] = t1 - (t2>>1);
1034 t0 = MUL64(2*(in1[2*2] + in1[2*4]), C2);
1035 t1 = MUL64( in1[2*4] - in1[2*8] , -2*C8);
1036 t2 = MUL64(2*(in1[2*2] + in1[2*8]), -C4);
1038 tmp1[10] = (t3 - t0 - t2) >> 32;
1039 tmp1[ 2] = (t3 + t0 + t1) >> 32;
1040 tmp1[14] = (t3 + t2 - t1) >> 32;
1042 tmp1[ 4] = MULH(2*(in1[2*5] + in1[2*7] - in1[2*1]), -C3);
1043 t2 = MUL64(2*(in1[2*1] + in1[2*5]), C1);
1044 t3 = MUL64( in1[2*5] - in1[2*7] , -2*C7);
1045 t0 = MUL64(2*in1[2*3], C3);
1047 t1 = MUL64(2*(in1[2*1] + in1[2*7]), -C5);
1049 tmp1[ 0] = (t2 + t3 + t0) >> 32;
1050 tmp1[12] = (t2 + t1 - t0) >> 32;
1051 tmp1[ 8] = (t3 - t1 - t0) >> 32;
1053 t2 = in1[2*4] + in1[2*8] - in1[2*2];
1055 t3 = in1[2*0] + (in1[2*6]>>1);
1056 t1 = in1[2*0] - in1[2*6];
1057 tmp1[ 6] = t1 - (t2>>1);
1060 t0 = MULH(2*(in1[2*2] + in1[2*4]), C2);
1061 t1 = MULH( in1[2*4] - in1[2*8] , -2*C8);
1062 t2 = MULH(2*(in1[2*2] + in1[2*8]), -C4);
1064 tmp1[10] = t3 - t0 - t2;
1065 tmp1[ 2] = t3 + t0 + t1;
1066 tmp1[14] = t3 + t2 - t1;
1068 tmp1[ 4] = MULH(2*(in1[2*5] + in1[2*7] - in1[2*1]), -C3);
1069 t2 = MULH(2*(in1[2*1] + in1[2*5]), C1);
1070 t3 = MULH( in1[2*5] - in1[2*7] , -2*C7);
1071 t0 = MULH(2*in1[2*3], C3);
1073 t1 = MULH(2*(in1[2*1] + in1[2*7]), -C5);
1075 tmp1[ 0] = t2 + t3 + t0;
1076 tmp1[12] = t2 + t1 - t0;
1077 tmp1[ 8] = t3 - t1 - t0;
1090 s1 = MULL(t3 + t2, icos36[j]);
1091 s3 = MULL(t3 - t2, icos36[8 - j]);
1093 t0 = (s0 + s1) << 5;
1094 t1 = (s0 - s1) << 5;
1095 out[(9 + j)*SBLIMIT] = -MULH(t1, win[9 + j]) + buf[9 + j];
1096 out[(8 - j)*SBLIMIT] = MULH(t1, win[8 - j]) + buf[8 - j];
1097 buf[9 + j] = MULH(t0, win[18 + 9 + j]);
1098 buf[8 - j] = MULH(t0, win[18 + 8 - j]);
1100 t0 = (s2 + s3) << 5;
1101 t1 = (s2 - s3) << 5;
1102 out[(9 + 8 - j)*SBLIMIT] = -MULH(t1, win[9 + 8 - j]) + buf[9 + 8 - j];
1103 out[( j)*SBLIMIT] = MULH(t1, win[ j]) + buf[ j];
1104 buf[9 + 8 - j] = MULH(t0, win[18 + 9 + 8 - j]);
1105 buf[ + j] = MULH(t0, win[18 + j]);
1110 s1 = MULL(tmp[17], icos36[4]);
1111 t0 = (s0 + s1) << 5;
1112 t1 = (s0 - s1) << 5;
1113 out[(9 + 4)*SBLIMIT] = -MULH(t1, win[9 + 4]) + buf[9 + 4];
1114 out[(8 - 4)*SBLIMIT] = MULH(t1, win[8 - 4]) + buf[8 - 4];
1115 buf[9 + 4] = MULH(t0, win[18 + 9 + 4]);
1116 buf[8 - 4] = MULH(t0, win[18 + 8 - 4]);
1119 /* header decoding. MUST check the header before because no
1120 consistency check is done there. Return 1 if free format found and
1121 that the frame size must be computed externally */
1122 static int decode_header(MPADecodeContext *s, uint32_t header)
1124 int sample_rate, frame_size, mpeg25, padding;
1125 int sample_rate_index, bitrate_index;
1126 if (header & (1<<20)) {
1127 s->lsf = (header & (1<<19)) ? 0 : 1;
1134 s->layer = 4 - ((header >> 17) & 3);
1135 /* extract frequency */
1136 sample_rate_index = (header >> 10) & 3;
1137 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1138 sample_rate_index += 3 * (s->lsf + mpeg25);
1139 s->sample_rate_index = sample_rate_index;
1140 s->error_protection = ((header >> 16) & 1) ^ 1;
1141 s->sample_rate = sample_rate;
1143 bitrate_index = (header >> 12) & 0xf;
1144 padding = (header >> 9) & 1;
1145 //extension = (header >> 8) & 1;
1146 s->mode = (header >> 6) & 3;
1147 s->mode_ext = (header >> 4) & 3;
1148 //copyright = (header >> 3) & 1;
1149 //original = (header >> 2) & 1;
1150 //emphasis = header & 3;
1152 if (s->mode == MPA_MONO)
1157 if (bitrate_index != 0) {
1158 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1159 s->bit_rate = frame_size * 1000;
1162 frame_size = (frame_size * 12000) / sample_rate;
1163 frame_size = (frame_size + padding) * 4;
1166 frame_size = (frame_size * 144000) / sample_rate;
1167 frame_size += padding;
1171 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1172 frame_size += padding;
1175 s->frame_size = frame_size;
1177 /* if no frame size computed, signal it */
1178 if (!s->free_format_frame_size)
1180 /* free format: compute bitrate and real frame size from the
1181 frame size we extracted by reading the bitstream */
1182 s->frame_size = s->free_format_frame_size;
1185 s->frame_size += padding * 4;
1186 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1189 s->frame_size += padding;
1190 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1194 s->frame_size += padding;
1195 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1201 printf("layer%d, %d Hz, %d kbits/s, ",
1202 s->layer, s->sample_rate, s->bit_rate);
1203 if (s->nb_channels == 2) {
1204 if (s->layer == 3) {
1205 if (s->mode_ext & MODE_EXT_MS_STEREO)
1207 if (s->mode_ext & MODE_EXT_I_STEREO)
1219 /* useful helper to get mpeg audio stream infos. Return -1 if error in
1220 header, otherwise the coded frame size in bytes */
1221 int mpa_decode_header(AVCodecContext *avctx, uint32_t head)
1223 MPADecodeContext s1, *s = &s1;
1224 memset( s, 0, sizeof(MPADecodeContext) );
1226 if (ff_mpa_check_header(head) != 0)
1229 if (decode_header(s, head) != 0) {
1235 avctx->frame_size = 384;
1238 avctx->frame_size = 1152;
1243 avctx->frame_size = 576;
1245 avctx->frame_size = 1152;
1249 avctx->sample_rate = s->sample_rate;
1250 avctx->channels = s->nb_channels;
1251 avctx->bit_rate = s->bit_rate;
1252 avctx->sub_id = s->layer;
1253 return s->frame_size;
1256 /* return the number of decoded frames */
1257 static int mp_decode_layer1(MPADecodeContext *s)
1259 int bound, i, v, n, ch, j, mant;
1260 uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
1261 uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1263 if (s->mode == MPA_JSTEREO)
1264 bound = (s->mode_ext + 1) * 4;
1268 /* allocation bits */
1269 for(i=0;i<bound;i++) {
1270 for(ch=0;ch<s->nb_channels;ch++) {
1271 allocation[ch][i] = get_bits(&s->gb, 4);
1274 for(i=bound;i<SBLIMIT;i++) {
1275 allocation[0][i] = get_bits(&s->gb, 4);
1279 for(i=0;i<bound;i++) {
1280 for(ch=0;ch<s->nb_channels;ch++) {
1281 if (allocation[ch][i])
1282 scale_factors[ch][i] = get_bits(&s->gb, 6);
1285 for(i=bound;i<SBLIMIT;i++) {
1286 if (allocation[0][i]) {
1287 scale_factors[0][i] = get_bits(&s->gb, 6);
1288 scale_factors[1][i] = get_bits(&s->gb, 6);
1292 /* compute samples */
1294 for(i=0;i<bound;i++) {
1295 for(ch=0;ch<s->nb_channels;ch++) {
1296 n = allocation[ch][i];
1298 mant = get_bits(&s->gb, n + 1);
1299 v = l1_unscale(n, mant, scale_factors[ch][i]);
1303 s->sb_samples[ch][j][i] = v;
1306 for(i=bound;i<SBLIMIT;i++) {
1307 n = allocation[0][i];
1309 mant = get_bits(&s->gb, n + 1);
1310 v = l1_unscale(n, mant, scale_factors[0][i]);
1311 s->sb_samples[0][j][i] = v;
1312 v = l1_unscale(n, mant, scale_factors[1][i]);
1313 s->sb_samples[1][j][i] = v;
1315 s->sb_samples[0][j][i] = 0;
1316 s->sb_samples[1][j][i] = 0;
1323 /* bitrate is in kb/s */
1324 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1326 int ch_bitrate, table;
1328 ch_bitrate = bitrate / nb_channels;
1330 if ((freq == 48000 && ch_bitrate >= 56) ||
1331 (ch_bitrate >= 56 && ch_bitrate <= 80))
1333 else if (freq != 48000 && ch_bitrate >= 96)
1335 else if (freq != 32000 && ch_bitrate <= 48)
1345 static int mp_decode_layer2(MPADecodeContext *s)
1347 int sblimit; /* number of used subbands */
1348 const unsigned char *alloc_table;
1349 int table, bit_alloc_bits, i, j, ch, bound, v;
1350 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1351 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1352 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1353 int scale, qindex, bits, steps, k, l, m, b;
1355 /* select decoding table */
1356 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1357 s->sample_rate, s->lsf);
1358 sblimit = sblimit_table[table];
1359 alloc_table = alloc_tables[table];
1361 if (s->mode == MPA_JSTEREO)
1362 bound = (s->mode_ext + 1) * 4;
1366 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1369 if( bound > sblimit ) bound = sblimit;
1371 /* parse bit allocation */
1373 for(i=0;i<bound;i++) {
1374 bit_alloc_bits = alloc_table[j];
1375 for(ch=0;ch<s->nb_channels;ch++) {
1376 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1378 j += 1 << bit_alloc_bits;
1380 for(i=bound;i<sblimit;i++) {
1381 bit_alloc_bits = alloc_table[j];
1382 v = get_bits(&s->gb, bit_alloc_bits);
1383 bit_alloc[0][i] = v;
1384 bit_alloc[1][i] = v;
1385 j += 1 << bit_alloc_bits;
1390 for(ch=0;ch<s->nb_channels;ch++) {
1391 for(i=0;i<sblimit;i++)
1392 printf(" %d", bit_alloc[ch][i]);
1399 for(i=0;i<sblimit;i++) {
1400 for(ch=0;ch<s->nb_channels;ch++) {
1401 if (bit_alloc[ch][i])
1402 scale_code[ch][i] = get_bits(&s->gb, 2);
1407 for(i=0;i<sblimit;i++) {
1408 for(ch=0;ch<s->nb_channels;ch++) {
1409 if (bit_alloc[ch][i]) {
1410 sf = scale_factors[ch][i];
1411 switch(scale_code[ch][i]) {
1414 sf[0] = get_bits(&s->gb, 6);
1415 sf[1] = get_bits(&s->gb, 6);
1416 sf[2] = get_bits(&s->gb, 6);
1419 sf[0] = get_bits(&s->gb, 6);
1424 sf[0] = get_bits(&s->gb, 6);
1425 sf[2] = get_bits(&s->gb, 6);
1429 sf[0] = get_bits(&s->gb, 6);
1430 sf[2] = get_bits(&s->gb, 6);
1439 for(ch=0;ch<s->nb_channels;ch++) {
1440 for(i=0;i<sblimit;i++) {
1441 if (bit_alloc[ch][i]) {
1442 sf = scale_factors[ch][i];
1443 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1454 for(l=0;l<12;l+=3) {
1456 for(i=0;i<bound;i++) {
1457 bit_alloc_bits = alloc_table[j];
1458 for(ch=0;ch<s->nb_channels;ch++) {
1459 b = bit_alloc[ch][i];
1461 scale = scale_factors[ch][i][k];
1462 qindex = alloc_table[j+b];
1463 bits = quant_bits[qindex];
1465 /* 3 values at the same time */
1466 v = get_bits(&s->gb, -bits);
1467 steps = quant_steps[qindex];
1468 s->sb_samples[ch][k * 12 + l + 0][i] =
1469 l2_unscale_group(steps, v % steps, scale);
1471 s->sb_samples[ch][k * 12 + l + 1][i] =
1472 l2_unscale_group(steps, v % steps, scale);
1474 s->sb_samples[ch][k * 12 + l + 2][i] =
1475 l2_unscale_group(steps, v, scale);
1478 v = get_bits(&s->gb, bits);
1479 v = l1_unscale(bits - 1, v, scale);
1480 s->sb_samples[ch][k * 12 + l + m][i] = v;
1484 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1485 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1486 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1489 /* next subband in alloc table */
1490 j += 1 << bit_alloc_bits;
1492 /* XXX: find a way to avoid this duplication of code */
1493 for(i=bound;i<sblimit;i++) {
1494 bit_alloc_bits = alloc_table[j];
1495 b = bit_alloc[0][i];
1497 int mant, scale0, scale1;
1498 scale0 = scale_factors[0][i][k];
1499 scale1 = scale_factors[1][i][k];
1500 qindex = alloc_table[j+b];
1501 bits = quant_bits[qindex];
1503 /* 3 values at the same time */
1504 v = get_bits(&s->gb, -bits);
1505 steps = quant_steps[qindex];
1508 s->sb_samples[0][k * 12 + l + 0][i] =
1509 l2_unscale_group(steps, mant, scale0);
1510 s->sb_samples[1][k * 12 + l + 0][i] =
1511 l2_unscale_group(steps, mant, scale1);
1514 s->sb_samples[0][k * 12 + l + 1][i] =
1515 l2_unscale_group(steps, mant, scale0);
1516 s->sb_samples[1][k * 12 + l + 1][i] =
1517 l2_unscale_group(steps, mant, scale1);
1518 s->sb_samples[0][k * 12 + l + 2][i] =
1519 l2_unscale_group(steps, v, scale0);
1520 s->sb_samples[1][k * 12 + l + 2][i] =
1521 l2_unscale_group(steps, v, scale1);
1524 mant = get_bits(&s->gb, bits);
1525 s->sb_samples[0][k * 12 + l + m][i] =
1526 l1_unscale(bits - 1, mant, scale0);
1527 s->sb_samples[1][k * 12 + l + m][i] =
1528 l1_unscale(bits - 1, mant, scale1);
1532 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1533 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1534 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1535 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1536 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1537 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1539 /* next subband in alloc table */
1540 j += 1 << bit_alloc_bits;
1542 /* fill remaining samples to zero */
1543 for(i=sblimit;i<SBLIMIT;i++) {
1544 for(ch=0;ch<s->nb_channels;ch++) {
1545 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1546 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1547 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1556 * Seek back in the stream for backstep bytes (at most 511 bytes)
1558 static void seek_to_maindata(MPADecodeContext *s, unsigned int backstep)
1562 /* compute current position in stream */
1563 ptr = (uint8_t *)(s->gb.buffer + (get_bits_count(&s->gb)>>3));
1565 /* copy old data before current one */
1567 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1568 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1569 /* init get bits again */
1570 init_get_bits(&s->gb, ptr, (s->frame_size + backstep)*8);
1572 /* prepare next buffer */
1573 s->inbuf_index ^= 1;
1574 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1575 s->old_frame_size = s->frame_size;
1578 static inline void lsf_sf_expand(int *slen,
1579 int sf, int n1, int n2, int n3)
1598 static void exponents_from_scale_factors(MPADecodeContext *s,
1602 const uint8_t *bstab, *pretab;
1603 int len, i, j, k, l, v0, shift, gain, gains[3];
1606 exp_ptr = exponents;
1607 gain = g->global_gain - 210;
1608 shift = g->scalefac_scale + 1;
1610 bstab = band_size_long[s->sample_rate_index];
1611 pretab = mpa_pretab[g->preflag];
1612 for(i=0;i<g->long_end;i++) {
1613 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1619 if (g->short_start < 13) {
1620 bstab = band_size_short[s->sample_rate_index];
1621 gains[0] = gain - (g->subblock_gain[0] << 3);
1622 gains[1] = gain - (g->subblock_gain[1] << 3);
1623 gains[2] = gain - (g->subblock_gain[2] << 3);
1625 for(i=g->short_start;i<13;i++) {
1628 v0 = gains[l] - (g->scale_factors[k++] << shift);
1636 /* handle n = 0 too */
1637 static inline int get_bitsz(GetBitContext *s, int n)
1642 return get_bits(s, n);
1645 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1646 int16_t *exponents, int end_pos)
1649 int linbits, code, x, y, l, v, i, j, k, pos;
1650 GetBitContext last_gb;
1652 uint8_t *code_table;
1654 /* low frequencies (called big values) */
1657 j = g->region_size[i];
1660 /* select vlc table */
1661 k = g->table_select[i];
1662 l = mpa_huff_data[k][0];
1663 linbits = mpa_huff_data[k][1];
1665 code_table = huff_code_table[l];
1667 /* read huffcode and compute each couple */
1669 if (get_bits_count(&s->gb) >= end_pos)
1672 code = get_vlc(&s->gb, vlc);
1675 y = code_table[code];
1682 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1683 i, g->region_size[i] - j, x, y, exponents[s_index]);
1686 x += get_bitsz(&s->gb, linbits);
1687 v = l3_unscale(x, exponents[s_index]);
1688 if (get_bits1(&s->gb))
1693 g->sb_hybrid[s_index++] = v;
1696 y += get_bitsz(&s->gb, linbits);
1697 v = l3_unscale(y, exponents[s_index]);
1698 if (get_bits1(&s->gb))
1703 g->sb_hybrid[s_index++] = v;
1707 /* high frequencies */
1708 vlc = &huff_quad_vlc[g->count1table_select];
1709 last_gb.buffer = NULL;
1710 while (s_index <= 572) {
1711 pos = get_bits_count(&s->gb);
1712 if (pos >= end_pos) {
1713 if (pos > end_pos && last_gb.buffer != NULL) {
1714 /* some encoders generate an incorrect size for this
1715 part. We must go back into the data */
1723 code = get_vlc(&s->gb, vlc);
1724 dprintf("t=%d code=%d\n", g->count1table_select, code);
1728 if (code & (8 >> i)) {
1729 /* non zero value. Could use a hand coded function for
1731 v = l3_unscale(1, exponents[s_index]);
1732 if(get_bits1(&s->gb))
1737 g->sb_hybrid[s_index++] = v;
1740 while (s_index < 576)
1741 g->sb_hybrid[s_index++] = 0;
1745 /* Reorder short blocks from bitstream order to interleaved order. It
1746 would be faster to do it in parsing, but the code would be far more
1748 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1751 int32_t *ptr, *dst, *ptr1;
1754 if (g->block_type != 2)
1757 if (g->switch_point) {
1758 if (s->sample_rate_index != 8) {
1759 ptr = g->sb_hybrid + 36;
1761 ptr = g->sb_hybrid + 48;
1767 for(i=g->short_start;i<13;i++) {
1768 len = band_size_short[s->sample_rate_index][i];
1772 for(j=len;j>0;j--) {
1777 memcpy(ptr1, tmp, len * 3 * sizeof(int32_t));
1781 #define ISQRT2 FIXR(0.70710678118654752440)
1783 static void compute_stereo(MPADecodeContext *s,
1784 GranuleDef *g0, GranuleDef *g1)
1788 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1789 int32_t (*is_tab)[16];
1790 int32_t *tab0, *tab1;
1791 int non_zero_found_short[3];
1793 /* intensity stereo */
1794 if (s->mode_ext & MODE_EXT_I_STEREO) {
1799 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1803 tab0 = g0->sb_hybrid + 576;
1804 tab1 = g1->sb_hybrid + 576;
1806 non_zero_found_short[0] = 0;
1807 non_zero_found_short[1] = 0;
1808 non_zero_found_short[2] = 0;
1809 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1810 for(i = 12;i >= g1->short_start;i--) {
1811 /* for last band, use previous scale factor */
1814 len = band_size_short[s->sample_rate_index][i];
1818 if (!non_zero_found_short[l]) {
1819 /* test if non zero band. if so, stop doing i-stereo */
1820 for(j=0;j<len;j++) {
1822 non_zero_found_short[l] = 1;
1826 sf = g1->scale_factors[k + l];
1832 for(j=0;j<len;j++) {
1834 tab0[j] = MULL(tmp0, v1);
1835 tab1[j] = MULL(tmp0, v2);
1839 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1840 /* lower part of the spectrum : do ms stereo
1842 for(j=0;j<len;j++) {
1845 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1846 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1853 non_zero_found = non_zero_found_short[0] |
1854 non_zero_found_short[1] |
1855 non_zero_found_short[2];
1857 for(i = g1->long_end - 1;i >= 0;i--) {
1858 len = band_size_long[s->sample_rate_index][i];
1861 /* test if non zero band. if so, stop doing i-stereo */
1862 if (!non_zero_found) {
1863 for(j=0;j<len;j++) {
1869 /* for last band, use previous scale factor */
1870 k = (i == 21) ? 20 : i;
1871 sf = g1->scale_factors[k];
1876 for(j=0;j<len;j++) {
1878 tab0[j] = MULL(tmp0, v1);
1879 tab1[j] = MULL(tmp0, v2);
1883 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1884 /* lower part of the spectrum : do ms stereo
1886 for(j=0;j<len;j++) {
1889 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1890 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1895 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1896 /* ms stereo ONLY */
1897 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1899 tab0 = g0->sb_hybrid;
1900 tab1 = g1->sb_hybrid;
1901 for(i=0;i<576;i++) {
1904 tab0[i] = tmp0 + tmp1;
1905 tab1[i] = tmp0 - tmp1;
1910 static void compute_antialias_integer(MPADecodeContext *s,
1916 /* we antialias only "long" bands */
1917 if (g->block_type == 2) {
1918 if (!g->switch_point)
1920 /* XXX: check this for 8000Hz case */
1926 ptr = g->sb_hybrid + 18;
1927 for(i = n;i > 0;i--) {
1928 int tmp0, tmp1, tmp2;
1929 csa = &csa_table[0][0];
1931 tmp0 = 4*(ptr[-1-j]);\
1932 tmp1 = 4*(ptr[ j]);\
1933 tmp2= MULH(tmp0 + tmp1, csa[0+4*j]);\
1934 ptr[-1-j] = tmp2 - MULH(tmp1, csa[2+4*j]);\
1935 ptr[ j] = tmp2 + MULH(tmp0, csa[3+4*j]);
1950 static void compute_antialias_float(MPADecodeContext *s,
1956 /* we antialias only "long" bands */
1957 if (g->block_type == 2) {
1958 if (!g->switch_point)
1960 /* XXX: check this for 8000Hz case */
1966 ptr = g->sb_hybrid + 18;
1967 for(i = n;i > 0;i--) {
1969 float *csa = &csa_table_float[0][0];
1970 #define FLOAT_AA(j)\
1973 ptr[-1-j] = lrintf(tmp0 * csa[0+4*j] - tmp1 * csa[1+4*j]);\
1974 ptr[ j] = lrintf(tmp0 * csa[1+4*j] + tmp1 * csa[0+4*j]);
1989 static void compute_imdct(MPADecodeContext *s,
1991 int32_t *sb_samples,
1994 int32_t *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
1998 int i, j, k, mdct_long_end, v, sblimit;
2000 /* find last non zero block */
2001 ptr = g->sb_hybrid + 576;
2002 ptr1 = g->sb_hybrid + 2 * 18;
2003 while (ptr >= ptr1) {
2005 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
2009 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
2011 if (g->block_type == 2) {
2012 /* XXX: check for 8000 Hz */
2013 if (g->switch_point)
2018 mdct_long_end = sblimit;
2023 for(j=0;j<mdct_long_end;j++) {
2024 /* apply window & overlap with previous buffer */
2025 out_ptr = sb_samples + j;
2027 if (g->switch_point && j < 2)
2030 win1 = mdct_win[g->block_type];
2031 /* select frequency inversion */
2032 win = win1 + ((4 * 36) & -(j & 1));
2033 imdct36(out_ptr, buf, ptr, win);
2034 out_ptr += 18*SBLIMIT;
2038 for(j=mdct_long_end;j<sblimit;j++) {
2044 /* select frequency inversion */
2045 win = mdct_win[2] + ((4 * 36) & -(j & 1));
2048 /* reorder input for short mdct */
2055 /* apply 12 point window and do small overlap */
2057 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
2058 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
2063 out_ptr = sb_samples + j;
2065 *out_ptr = out[i] + buf[i];
2066 buf[i] = out[i + 18];
2073 for(j=sblimit;j<SBLIMIT;j++) {
2075 out_ptr = sb_samples + j;
2086 void sample_dump(int fnum, int32_t *tab, int n)
2088 static FILE *files[16], *f;
2095 snprintf(buf, sizeof(buf), "/tmp/out%d.%s.pcm",
2097 #ifdef USE_HIGHPRECISION
2103 f = fopen(buf, "w");
2111 av_log(NULL, AV_LOG_DEBUG, "pos=%d\n", pos);
2113 av_log(NULL, AV_LOG_DEBUG, " %0.4f", (double)tab[i] / FRAC_ONE);
2115 av_log(NULL, AV_LOG_DEBUG, "\n");
2120 /* normalize to 23 frac bits */
2121 v = tab[i] << (23 - FRAC_BITS);
2122 fwrite(&v, 1, sizeof(int32_t), f);
2128 /* main layer3 decoding function */
2129 static int mp_decode_layer3(MPADecodeContext *s)
2131 int nb_granules, main_data_begin, private_bits;
2132 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
2133 GranuleDef granules[2][2], *g;
2134 int16_t exponents[576];
2136 /* read side info */
2138 main_data_begin = get_bits(&s->gb, 8);
2139 if (s->nb_channels == 2)
2140 private_bits = get_bits(&s->gb, 2);
2142 private_bits = get_bits(&s->gb, 1);
2145 main_data_begin = get_bits(&s->gb, 9);
2146 if (s->nb_channels == 2)
2147 private_bits = get_bits(&s->gb, 3);
2149 private_bits = get_bits(&s->gb, 5);
2151 for(ch=0;ch<s->nb_channels;ch++) {
2152 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2153 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2157 for(gr=0;gr<nb_granules;gr++) {
2158 for(ch=0;ch<s->nb_channels;ch++) {
2159 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2160 g = &granules[ch][gr];
2161 g->part2_3_length = get_bits(&s->gb, 12);
2162 g->big_values = get_bits(&s->gb, 9);
2163 g->global_gain = get_bits(&s->gb, 8);
2164 /* if MS stereo only is selected, we precompute the
2165 1/sqrt(2) renormalization factor */
2166 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2168 g->global_gain -= 2;
2170 g->scalefac_compress = get_bits(&s->gb, 9);
2172 g->scalefac_compress = get_bits(&s->gb, 4);
2173 blocksplit_flag = get_bits(&s->gb, 1);
2174 if (blocksplit_flag) {
2175 g->block_type = get_bits(&s->gb, 2);
2176 if (g->block_type == 0)
2178 g->switch_point = get_bits(&s->gb, 1);
2180 g->table_select[i] = get_bits(&s->gb, 5);
2182 g->subblock_gain[i] = get_bits(&s->gb, 3);
2183 /* compute huffman coded region sizes */
2184 if (g->block_type == 2)
2185 g->region_size[0] = (36 / 2);
2187 if (s->sample_rate_index <= 2)
2188 g->region_size[0] = (36 / 2);
2189 else if (s->sample_rate_index != 8)
2190 g->region_size[0] = (54 / 2);
2192 g->region_size[0] = (108 / 2);
2194 g->region_size[1] = (576 / 2);
2196 int region_address1, region_address2, l;
2198 g->switch_point = 0;
2200 g->table_select[i] = get_bits(&s->gb, 5);
2201 /* compute huffman coded region sizes */
2202 region_address1 = get_bits(&s->gb, 4);
2203 region_address2 = get_bits(&s->gb, 3);
2204 dprintf("region1=%d region2=%d\n",
2205 region_address1, region_address2);
2207 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2208 l = region_address1 + region_address2 + 2;
2209 /* should not overflow */
2213 band_index_long[s->sample_rate_index][l] >> 1;
2215 /* convert region offsets to region sizes and truncate
2216 size to big_values */
2217 g->region_size[2] = (576 / 2);
2220 k = g->region_size[i];
2221 if (k > g->big_values)
2223 g->region_size[i] = k - j;
2227 /* compute band indexes */
2228 if (g->block_type == 2) {
2229 if (g->switch_point) {
2230 /* if switched mode, we handle the 36 first samples as
2231 long blocks. For 8000Hz, we handle the 48 first
2232 exponents as long blocks (XXX: check this!) */
2233 if (s->sample_rate_index <= 2)
2235 else if (s->sample_rate_index != 8)
2238 g->long_end = 4; /* 8000 Hz */
2240 if (s->sample_rate_index != 8)
2249 g->short_start = 13;
2255 g->preflag = get_bits(&s->gb, 1);
2256 g->scalefac_scale = get_bits(&s->gb, 1);
2257 g->count1table_select = get_bits(&s->gb, 1);
2258 dprintf("block_type=%d switch_point=%d\n",
2259 g->block_type, g->switch_point);
2264 /* now we get bits from the main_data_begin offset */
2265 dprintf("seekback: %d\n", main_data_begin);
2266 seek_to_maindata(s, main_data_begin);
2269 for(gr=0;gr<nb_granules;gr++) {
2270 for(ch=0;ch<s->nb_channels;ch++) {
2271 g = &granules[ch][gr];
2273 bits_pos = get_bits_count(&s->gb);
2277 int slen, slen1, slen2;
2279 /* MPEG1 scale factors */
2280 slen1 = slen_table[0][g->scalefac_compress];
2281 slen2 = slen_table[1][g->scalefac_compress];
2282 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2283 if (g->block_type == 2) {
2284 n = g->switch_point ? 17 : 18;
2287 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2289 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2291 g->scale_factors[j++] = 0;
2293 sc = granules[ch][0].scale_factors;
2296 n = (k == 0 ? 6 : 5);
2297 if ((g->scfsi & (0x8 >> k)) == 0) {
2298 slen = (k < 2) ? slen1 : slen2;
2300 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2302 /* simply copy from last granule */
2304 g->scale_factors[j] = sc[j];
2309 g->scale_factors[j++] = 0;
2313 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2316 printf(" %d", g->scale_factors[i]);
2321 int tindex, tindex2, slen[4], sl, sf;
2323 /* LSF scale factors */
2324 if (g->block_type == 2) {
2325 tindex = g->switch_point ? 2 : 1;
2329 sf = g->scalefac_compress;
2330 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2331 /* intensity stereo case */
2334 lsf_sf_expand(slen, sf, 6, 6, 0);
2336 } else if (sf < 244) {
2337 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2340 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2346 lsf_sf_expand(slen, sf, 5, 4, 4);
2348 } else if (sf < 500) {
2349 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2352 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2360 n = lsf_nsf_table[tindex2][tindex][k];
2363 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2365 /* XXX: should compute exact size */
2367 g->scale_factors[j] = 0;
2370 printf("gr=%d ch=%d scale_factors:\n",
2373 printf(" %d", g->scale_factors[i]);
2379 exponents_from_scale_factors(s, g, exponents);
2381 /* read Huffman coded residue */
2382 if (huffman_decode(s, g, exponents,
2383 bits_pos + g->part2_3_length) < 0)
2386 sample_dump(0, g->sb_hybrid, 576);
2389 /* skip extension bits */
2390 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2391 if (bits_left < 0) {
2392 dprintf("bits_left=%d\n", bits_left);
2395 while (bits_left >= 16) {
2396 skip_bits(&s->gb, 16);
2400 skip_bits(&s->gb, bits_left);
2403 if (s->nb_channels == 2)
2404 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2406 for(ch=0;ch<s->nb_channels;ch++) {
2407 g = &granules[ch][gr];
2409 reorder_block(s, g);
2411 sample_dump(0, g->sb_hybrid, 576);
2413 s->compute_antialias(s, g);
2415 sample_dump(1, g->sb_hybrid, 576);
2417 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2419 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2423 return nb_granules * 18;
2426 static int mp_decode_frame(MPADecodeContext *s,
2429 int i, nb_frames, ch;
2430 OUT_INT *samples_ptr;
2432 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2433 (s->inbuf_ptr - s->inbuf - HEADER_SIZE)*8);
2435 /* skip error protection field */
2436 if (s->error_protection)
2437 get_bits(&s->gb, 16);
2439 dprintf("frame %d:\n", s->frame_count);
2442 nb_frames = mp_decode_layer1(s);
2445 nb_frames = mp_decode_layer2(s);
2449 nb_frames = mp_decode_layer3(s);
2453 for(i=0;i<nb_frames;i++) {
2454 for(ch=0;ch<s->nb_channels;ch++) {
2456 printf("%d-%d:", i, ch);
2457 for(j=0;j<SBLIMIT;j++)
2458 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2463 /* apply the synthesis filter */
2464 for(ch=0;ch<s->nb_channels;ch++) {
2465 samples_ptr = samples + ch;
2466 for(i=0;i<nb_frames;i++) {
2467 ff_mpa_synth_filter(s->synth_buf[ch], &(s->synth_buf_offset[ch]),
2468 window, &s->dither_state,
2469 samples_ptr, s->nb_channels,
2470 s->sb_samples[ch][i]);
2471 samples_ptr += 32 * s->nb_channels;
2477 return nb_frames * 32 * sizeof(OUT_INT) * s->nb_channels;
2480 static int decode_frame(AVCodecContext * avctx,
2481 void *data, int *data_size,
2482 uint8_t * buf, int buf_size)
2484 MPADecodeContext *s = avctx->priv_data;
2488 OUT_INT *out_samples = data;
2491 while (buf_size > 0) {
2492 len = s->inbuf_ptr - s->inbuf;
2493 if (s->frame_size == 0) {
2494 /* special case for next header for first frame in free
2495 format case (XXX: find a simpler method) */
2496 if (s->free_format_next_header != 0) {
2497 s->inbuf[0] = s->free_format_next_header >> 24;
2498 s->inbuf[1] = s->free_format_next_header >> 16;
2499 s->inbuf[2] = s->free_format_next_header >> 8;
2500 s->inbuf[3] = s->free_format_next_header;
2501 s->inbuf_ptr = s->inbuf + 4;
2502 s->free_format_next_header = 0;
2505 /* no header seen : find one. We need at least HEADER_SIZE
2506 bytes to parse it */
2507 len = HEADER_SIZE - len;
2511 memcpy(s->inbuf_ptr, buf_ptr, len);
2514 s->inbuf_ptr += len;
2516 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2518 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2519 (s->inbuf[2] << 8) | s->inbuf[3];
2521 if (ff_mpa_check_header(header) < 0) {
2522 /* no sync found : move by one byte (inefficient, but simple!) */
2523 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2525 dprintf("skip %x\n", header);
2526 /* reset free format frame size to give a chance
2527 to get a new bitrate */
2528 s->free_format_frame_size = 0;
2530 if (decode_header(s, header) == 1) {
2531 /* free format: prepare to compute frame size */
2534 /* update codec info */
2535 avctx->sample_rate = s->sample_rate;
2536 avctx->channels = s->nb_channels;
2537 avctx->bit_rate = s->bit_rate;
2538 avctx->sub_id = s->layer;
2541 avctx->frame_size = 384;
2544 avctx->frame_size = 1152;
2548 avctx->frame_size = 576;
2550 avctx->frame_size = 1152;
2555 } else if (s->frame_size == -1) {
2556 /* free format : find next sync to compute frame size */
2557 len = MPA_MAX_CODED_FRAME_SIZE - len;
2561 /* frame too long: resync */
2563 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2570 memcpy(s->inbuf_ptr, buf_ptr, len);
2571 /* check for header */
2572 p = s->inbuf_ptr - 3;
2573 pend = s->inbuf_ptr + len - 4;
2575 header = (p[0] << 24) | (p[1] << 16) |
2577 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2578 (s->inbuf[2] << 8) | s->inbuf[3];
2579 /* check with high probability that we have a
2581 if ((header & SAME_HEADER_MASK) ==
2582 (header1 & SAME_HEADER_MASK)) {
2583 /* header found: update pointers */
2584 len = (p + 4) - s->inbuf_ptr;
2588 /* compute frame size */
2589 s->free_format_next_header = header;
2590 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2591 padding = (header1 >> 9) & 1;
2593 s->free_format_frame_size -= padding * 4;
2595 s->free_format_frame_size -= padding;
2596 dprintf("free frame size=%d padding=%d\n",
2597 s->free_format_frame_size, padding);
2598 decode_header(s, header1);
2603 /* not found: simply increase pointers */
2605 s->inbuf_ptr += len;
2608 } else if (len < s->frame_size) {
2609 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2610 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2611 len = s->frame_size - len;
2614 memcpy(s->inbuf_ptr, buf_ptr, len);
2616 s->inbuf_ptr += len;
2620 if (s->frame_size > 0 &&
2621 (s->inbuf_ptr - s->inbuf) >= s->frame_size) {
2622 if (avctx->parse_only) {
2623 /* simply return the frame data */
2624 *(uint8_t **)data = s->inbuf;
2625 out_size = s->inbuf_ptr - s->inbuf;
2627 out_size = mp_decode_frame(s, out_samples);
2629 s->inbuf_ptr = s->inbuf;
2631 *data_size = out_size;
2635 return buf_ptr - buf;
2639 static int decode_frame_adu(AVCodecContext * avctx,
2640 void *data, int *data_size,
2641 uint8_t * buf, int buf_size)
2643 MPADecodeContext *s = avctx->priv_data;
2646 OUT_INT *out_samples = data;
2650 // Discard too short frames
2651 if (buf_size < HEADER_SIZE) {
2657 if (len > MPA_MAX_CODED_FRAME_SIZE)
2658 len = MPA_MAX_CODED_FRAME_SIZE;
2660 memcpy(s->inbuf, buf, len);
2661 s->inbuf_ptr = s->inbuf + len;
2663 // Get header and restore sync word
2664 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2665 (s->inbuf[2] << 8) | s->inbuf[3] | 0xffe00000;
2667 if (ff_mpa_check_header(header) < 0) { // Bad header, discard frame
2672 decode_header(s, header);
2673 /* update codec info */
2674 avctx->sample_rate = s->sample_rate;
2675 avctx->channels = s->nb_channels;
2676 avctx->bit_rate = s->bit_rate;
2677 avctx->sub_id = s->layer;
2679 avctx->frame_size=s->frame_size = len;
2681 if (avctx->parse_only) {
2682 /* simply return the frame data */
2683 *(uint8_t **)data = s->inbuf;
2684 out_size = s->inbuf_ptr - s->inbuf;
2686 out_size = mp_decode_frame(s, out_samples);
2689 *data_size = out_size;
2694 AVCodec mp2_decoder =
2699 sizeof(MPADecodeContext),
2704 CODEC_CAP_PARSE_ONLY,
2707 AVCodec mp3_decoder =
2712 sizeof(MPADecodeContext),
2717 CODEC_CAP_PARSE_ONLY,
2720 AVCodec mp3adu_decoder =
2725 sizeof(MPADecodeContext),
2730 CODEC_CAP_PARSE_ONLY,