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 #define FRAC_ONE (1 << FRAC_BITS)
53 #define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
54 #define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
55 #define FIX(a) ((int)((a) * FRAC_ONE))
56 /* WARNING: only correct for posititive numbers */
57 #define FIXR(a) ((int)((a) * FRAC_ONE + 0.5))
58 #define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)
61 typedef int16_t MPA_INT;
63 typedef int32_t MPA_INT;
69 #define BACKSTEP_SIZE 512
73 typedef struct MPADecodeContext {
74 uint8_t inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE]; /* input buffer */
76 uint8_t *inbuf_ptr, *inbuf;
78 int free_format_frame_size; /* frame size in case of free format
79 (zero if currently unknown) */
80 /* next header (used in free format parsing) */
81 uint32_t free_format_next_header;
85 int sample_rate_index; /* between 0 and 8 */
93 MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
94 int synth_buf_offset[MPA_MAX_CHANNELS];
95 int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
96 int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
100 void (*compute_antialias)(struct MPADecodeContext *s, struct GranuleDef *g);
101 int adu_mode; ///< 0 for standard mp3, 1 for adu formatted mp3
102 unsigned int dither_state;
105 /* layer 3 "granule" */
106 typedef struct GranuleDef {
111 int scalefac_compress;
113 uint8_t switch_point;
115 int subblock_gain[3];
116 uint8_t scalefac_scale;
117 uint8_t count1table_select;
118 int region_size[3]; /* number of huffman codes in each region */
120 int short_start, long_end; /* long/short band indexes */
121 uint8_t scale_factors[40];
122 int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
125 #define MODE_EXT_MS_STEREO 2
126 #define MODE_EXT_I_STEREO 1
128 /* layer 3 huffman tables */
129 typedef struct HuffTable {
132 const uint16_t *codes;
135 #include "mpegaudiodectab.h"
137 static void compute_antialias_integer(MPADecodeContext *s, GranuleDef *g);
138 static void compute_antialias_float(MPADecodeContext *s, GranuleDef *g);
140 /* vlc structure for decoding layer 3 huffman tables */
141 static VLC huff_vlc[16];
142 static uint8_t *huff_code_table[16];
143 static VLC huff_quad_vlc[2];
144 /* computed from band_size_long */
145 static uint16_t band_index_long[9][23];
146 /* XXX: free when all decoders are closed */
147 #define TABLE_4_3_SIZE (8191 + 16)
148 static int8_t *table_4_3_exp;
150 static uint16_t *table_4_3_value;
152 static uint32_t *table_4_3_value;
154 /* intensity stereo coef table */
155 static int32_t is_table[2][16];
156 static int32_t is_table_lsf[2][2][16];
157 static int32_t csa_table[8][4];
158 static float csa_table_float[8][4];
159 static int32_t mdct_win[8][36];
161 /* lower 2 bits: modulo 3, higher bits: shift */
162 static uint16_t scale_factor_modshift[64];
163 /* [i][j]: 2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
164 static int32_t scale_factor_mult[15][3];
165 /* mult table for layer 2 group quantization */
167 #define SCALE_GEN(v) \
168 { FIXR(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(0.6299605249 * (v)) }
170 static int32_t scale_factor_mult2[3][3] = {
171 SCALE_GEN(4.0 / 3.0), /* 3 steps */
172 SCALE_GEN(4.0 / 5.0), /* 5 steps */
173 SCALE_GEN(4.0 / 9.0), /* 9 steps */
177 static uint32_t scale_factor_mult3[4] = {
179 FIXR(1.18920711500272106671),
180 FIXR(1.41421356237309504880),
181 FIXR(1.68179283050742908605),
184 void ff_mpa_synth_init(MPA_INT *window);
185 static MPA_INT window[512] __attribute__((aligned(16)));
187 /* layer 1 unscaling */
188 /* n = number of bits of the mantissa minus 1 */
189 static inline int l1_unscale(int n, int mant, int scale_factor)
194 shift = scale_factor_modshift[scale_factor];
197 val = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
199 /* NOTE: at this point, 1 <= shift >= 21 + 15 */
200 return (int)((val + (1LL << (shift - 1))) >> shift);
203 static inline int l2_unscale_group(int steps, int mant, int scale_factor)
207 shift = scale_factor_modshift[scale_factor];
211 val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
212 /* NOTE: at this point, 0 <= shift <= 21 */
214 val = (val + (1 << (shift - 1))) >> shift;
218 /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
219 static inline int l3_unscale(int value, int exponent)
228 e = table_4_3_exp[value];
229 e += (exponent >> 2);
235 m = table_4_3_value[value];
237 m = (m * scale_factor_mult3[exponent & 3]);
238 m = (m + (1 << (e-1))) >> e;
241 m = MUL64(m, scale_factor_mult3[exponent & 3]);
242 m = (m + (uint64_t_C(1) << (e-1))) >> e;
247 /* all integer n^(4/3) computation code */
250 #define POW_FRAC_BITS 24
251 #define POW_FRAC_ONE (1 << POW_FRAC_BITS)
252 #define POW_FIX(a) ((int)((a) * POW_FRAC_ONE))
253 #define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)
255 static int dev_4_3_coefs[DEV_ORDER];
257 static int pow_mult3[3] = {
259 POW_FIX(1.25992104989487316476),
260 POW_FIX(1.58740105196819947474),
263 static void int_pow_init(void)
268 for(i=0;i<DEV_ORDER;i++) {
269 a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
270 dev_4_3_coefs[i] = a;
274 /* return the mantissa and the binary exponent */
275 static int int_pow(int i, int *exp_ptr)
283 while (a < (1 << (POW_FRAC_BITS - 1))) {
287 a -= (1 << POW_FRAC_BITS);
289 for(j = DEV_ORDER - 1; j >= 0; j--)
290 a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
291 a = (1 << POW_FRAC_BITS) + a1;
292 /* exponent compute (exact) */
296 a = POW_MULL(a, pow_mult3[er]);
297 while (a >= 2 * POW_FRAC_ONE) {
301 /* convert to float */
302 while (a < POW_FRAC_ONE) {
306 /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
307 #if POW_FRAC_BITS > FRAC_BITS
308 a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
309 /* correct overflow */
310 if (a >= 2 * (1 << FRAC_BITS)) {
319 static int decode_init(AVCodecContext * avctx)
321 MPADecodeContext *s = avctx->priv_data;
325 if(avctx->antialias_algo == FF_AA_INT)
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++) {
410 f = pow((double)i, 4.0 / 3.0);
414 if ((unsigned short)m1 != m1) {
420 if (m != m1 || e != e1) {
421 printf("%4d: m=%x m1=%x e=%d e1=%d\n",
426 /* normalized to FRAC_BITS */
427 table_4_3_value[i] = m;
428 table_4_3_exp[i] = e;
435 f = tan((double)i * M_PI / 12.0);
436 v = FIXR(f / (1.0 + f));
441 is_table[1][6 - i] = v;
445 is_table[0][i] = is_table[1][i] = 0.0;
452 e = -(j + 1) * ((i + 1) >> 1);
453 f = pow(2.0, e / 4.0);
455 is_table_lsf[j][k ^ 1][i] = FIXR(f);
456 is_table_lsf[j][k][i] = FIXR(1.0);
457 dprintf("is_table_lsf %d %d: %x %x\n",
458 i, j, is_table_lsf[j][0][i], is_table_lsf[j][1][i]);
465 cs = 1.0 / sqrt(1.0 + ci * ci);
467 csa_table[i][0] = FIX(cs);
468 csa_table[i][1] = FIX(ca);
469 csa_table[i][2] = FIX(ca) + FIX(cs);
470 csa_table[i][3] = FIX(ca) - FIX(cs);
471 csa_table_float[i][0] = cs;
472 csa_table_float[i][1] = ca;
473 csa_table_float[i][2] = ca + cs;
474 csa_table_float[i][3] = ca - cs;
475 // printf("%d %d %d %d\n", FIX(cs), FIX(cs-1), FIX(ca), FIX(cs)-FIX(ca));
478 /* compute mdct windows */
481 v = FIXR(sin(M_PI * (i + 0.5) / 36.0));
487 mdct_win[1][18 + i] = FIXR(1.0);
488 mdct_win[1][24 + i] = FIXR(sin(M_PI * ((i + 6) + 0.5) / 12.0));
489 mdct_win[1][30 + i] = FIXR(0.0);
491 mdct_win[3][i] = FIXR(0.0);
492 mdct_win[3][6 + i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
493 mdct_win[3][12 + i] = FIXR(1.0);
497 mdct_win[2][i] = FIXR(sin(M_PI * (i + 0.5) / 12.0));
499 /* NOTE: we do frequency inversion adter the MDCT by changing
500 the sign of the right window coefs */
503 mdct_win[j + 4][i] = mdct_win[j][i];
504 mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
510 printf("win%d=\n", j);
512 printf("%f, ", (double)mdct_win[j][i] / FRAC_ONE);
520 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
521 s->inbuf_ptr = s->inbuf;
525 if (avctx->codec_id == CODEC_ID_MP3ADU)
530 /* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */
534 #define COS0_0 FIXR(0.50060299823519630134)
535 #define COS0_1 FIXR(0.50547095989754365998)
536 #define COS0_2 FIXR(0.51544730992262454697)
537 #define COS0_3 FIXR(0.53104259108978417447)
538 #define COS0_4 FIXR(0.55310389603444452782)
539 #define COS0_5 FIXR(0.58293496820613387367)
540 #define COS0_6 FIXR(0.62250412303566481615)
541 #define COS0_7 FIXR(0.67480834145500574602)
542 #define COS0_8 FIXR(0.74453627100229844977)
543 #define COS0_9 FIXR(0.83934964541552703873)
544 #define COS0_10 FIXR(0.97256823786196069369)
545 #define COS0_11 FIXR(1.16943993343288495515)
546 #define COS0_12 FIXR(1.48416461631416627724)
547 #define COS0_13 FIXR(2.05778100995341155085)
548 #define COS0_14 FIXR(3.40760841846871878570)
549 #define COS0_15 FIXR(10.19000812354805681150)
551 #define COS1_0 FIXR(0.50241928618815570551)
552 #define COS1_1 FIXR(0.52249861493968888062)
553 #define COS1_2 FIXR(0.56694403481635770368)
554 #define COS1_3 FIXR(0.64682178335999012954)
555 #define COS1_4 FIXR(0.78815462345125022473)
556 #define COS1_5 FIXR(1.06067768599034747134)
557 #define COS1_6 FIXR(1.72244709823833392782)
558 #define COS1_7 FIXR(5.10114861868916385802)
560 #define COS2_0 FIXR(0.50979557910415916894)
561 #define COS2_1 FIXR(0.60134488693504528054)
562 #define COS2_2 FIXR(0.89997622313641570463)
563 #define COS2_3 FIXR(2.56291544774150617881)
565 #define COS3_0 FIXR(0.54119610014619698439)
566 #define COS3_1 FIXR(1.30656296487637652785)
568 #define COS4_0 FIXR(0.70710678118654752439)
570 /* butterfly operator */
573 tmp0 = tab[a] + tab[b];\
574 tmp1 = tab[a] - tab[b];\
576 tab[b] = MULL(tmp1, c);\
579 #define BF1(a, b, c, d)\
586 #define BF2(a, b, c, d)\
596 #define ADD(a, b) tab[a] += tab[b]
598 /* DCT32 without 1/sqrt(2) coef zero scaling. */
599 static void dct32(int32_t *out, int32_t *tab)
731 out[ 1] = tab[16] + tab[24];
732 out[17] = tab[17] + tab[25];
733 out[ 9] = tab[18] + tab[26];
734 out[25] = tab[19] + tab[27];
735 out[ 5] = tab[20] + tab[28];
736 out[21] = tab[21] + tab[29];
737 out[13] = tab[22] + tab[30];
738 out[29] = tab[23] + tab[31];
739 out[ 3] = tab[24] + tab[20];
740 out[19] = tab[25] + tab[21];
741 out[11] = tab[26] + tab[22];
742 out[27] = tab[27] + tab[23];
743 out[ 7] = tab[28] + tab[18];
744 out[23] = tab[29] + tab[19];
745 out[15] = tab[30] + tab[17];
749 #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15)
753 static inline int round_sample(int *sum)
756 sum1 = (*sum) >> OUT_SHIFT;
757 *sum &= (1<<OUT_SHIFT)-1;
760 else if (sum1 > 32767)
765 #if defined(ARCH_POWERPC_405)
767 /* signed 16x16 -> 32 multiply add accumulate */
768 #define MACS(rt, ra, rb) \
769 asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
771 /* signed 16x16 -> 32 multiply */
772 #define MULS(ra, rb) \
773 ({ int __rt; asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
777 /* signed 16x16 -> 32 multiply add accumulate */
778 #define MACS(rt, ra, rb) rt += (ra) * (rb)
780 /* signed 16x16 -> 32 multiply */
781 #define MULS(ra, rb) ((ra) * (rb))
787 static inline int round_sample(int64_t *sum)
790 sum1 = (int)((*sum) >> OUT_SHIFT);
791 *sum &= (1<<OUT_SHIFT)-1;
794 else if (sum1 > 32767)
799 #define MULS(ra, rb) MUL64(ra, rb)
803 #define SUM8(sum, op, w, p) \
805 sum op MULS((w)[0 * 64], p[0 * 64]);\
806 sum op MULS((w)[1 * 64], p[1 * 64]);\
807 sum op MULS((w)[2 * 64], p[2 * 64]);\
808 sum op MULS((w)[3 * 64], p[3 * 64]);\
809 sum op MULS((w)[4 * 64], p[4 * 64]);\
810 sum op MULS((w)[5 * 64], p[5 * 64]);\
811 sum op MULS((w)[6 * 64], p[6 * 64]);\
812 sum op MULS((w)[7 * 64], p[7 * 64]);\
815 #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
819 sum1 op1 MULS((w1)[0 * 64], tmp);\
820 sum2 op2 MULS((w2)[0 * 64], tmp);\
822 sum1 op1 MULS((w1)[1 * 64], tmp);\
823 sum2 op2 MULS((w2)[1 * 64], tmp);\
825 sum1 op1 MULS((w1)[2 * 64], tmp);\
826 sum2 op2 MULS((w2)[2 * 64], tmp);\
828 sum1 op1 MULS((w1)[3 * 64], tmp);\
829 sum2 op2 MULS((w2)[3 * 64], tmp);\
831 sum1 op1 MULS((w1)[4 * 64], tmp);\
832 sum2 op2 MULS((w2)[4 * 64], tmp);\
834 sum1 op1 MULS((w1)[5 * 64], tmp);\
835 sum2 op2 MULS((w2)[5 * 64], tmp);\
837 sum1 op1 MULS((w1)[6 * 64], tmp);\
838 sum2 op2 MULS((w2)[6 * 64], tmp);\
840 sum1 op1 MULS((w1)[7 * 64], tmp);\
841 sum2 op2 MULS((w2)[7 * 64], tmp);\
844 void ff_mpa_synth_init(MPA_INT *window)
848 /* max = 18760, max sum over all 16 coefs : 44736 */
853 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
863 /* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
865 /* XXX: optimize by avoiding ring buffer usage */
866 void ff_mpa_synth_filter(MPA_INT *synth_buf_ptr, int *synth_buf_offset,
867 MPA_INT *window, int *dither_state,
868 int16_t *samples, int incr,
869 int32_t sb_samples[SBLIMIT])
872 register MPA_INT *synth_buf;
873 register const MPA_INT *w, *w2, *p;
882 dct32(tmp, sb_samples);
884 offset = *synth_buf_offset;
885 synth_buf = synth_buf_ptr + offset;
890 /* NOTE: can cause a loss in precision if very high amplitude
899 /* copy to avoid wrap */
900 memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));
902 samples2 = samples + 31 * incr;
910 SUM8(sum, -=, w + 32, p);
911 *samples = round_sample(&sum);
915 /* we calculate two samples at the same time to avoid one memory
916 access per two sample */
919 p = synth_buf + 16 + j;
920 SUM8P2(sum, +=, sum2, -=, w, w2, p);
921 p = synth_buf + 48 - j;
922 SUM8P2(sum, -=, sum2, -=, w + 32, w2 + 32, p);
924 *samples = round_sample(&sum);
927 *samples2 = round_sample(&sum);
934 SUM8(sum, -=, w + 32, p);
935 *samples = round_sample(&sum);
938 offset = (offset - 32) & 511;
939 *synth_buf_offset = offset;
943 #define C1 FIXR(0.99144486137381041114)
944 #define C3 FIXR(0.92387953251128675612)
945 #define C5 FIXR(0.79335334029123516458)
946 #define C7 FIXR(0.60876142900872063941)
947 #define C9 FIXR(0.38268343236508977173)
948 #define C11 FIXR(0.13052619222005159154)
950 /* 12 points IMDCT. We compute it "by hand" by factorizing obvious
952 static void imdct12(int *out, int *in)
955 int64_t in1_3, in1_9, in4_3, in4_9;
957 in1_3 = MUL64(in[1], C3);
958 in1_9 = MUL64(in[1], C9);
959 in4_3 = MUL64(in[4], C3);
960 in4_9 = MUL64(in[4], C9);
962 tmp = FRAC_RND(MUL64(in[0], C7) - in1_3 - MUL64(in[2], C11) +
963 MUL64(in[3], C1) - in4_9 - MUL64(in[5], C5));
966 tmp = FRAC_RND(MUL64(in[0] - in[3], C9) - in1_3 +
967 MUL64(in[2] + in[5], C3) - in4_9);
970 tmp = FRAC_RND(MUL64(in[0], C11) - in1_9 + MUL64(in[2], C7) -
971 MUL64(in[3], C5) + in4_3 - MUL64(in[5], C1));
974 tmp = FRAC_RND(MUL64(-in[0], C5) + in1_9 + MUL64(in[2], C1) +
975 MUL64(in[3], C11) - in4_3 - MUL64(in[5], C7));
978 tmp = FRAC_RND(MUL64(-in[0] + in[3], C3) - in1_9 +
979 MUL64(in[2] + in[5], C9) + in4_3);
982 tmp = FRAC_RND(-MUL64(in[0], C1) - in1_3 - MUL64(in[2], C5) -
983 MUL64(in[3], C7) - in4_9 - MUL64(in[5], C11));
996 #define C1 FIXR(0.98480775301220805936)
997 #define C2 FIXR(0.93969262078590838405)
998 #define C3 FIXR(0.86602540378443864676)
999 #define C4 FIXR(0.76604444311897803520)
1000 #define C5 FIXR(0.64278760968653932632)
1001 #define C6 FIXR(0.5)
1002 #define C7 FIXR(0.34202014332566873304)
1003 #define C8 FIXR(0.17364817766693034885)
1005 /* 0.5 / cos(pi*(2*i+1)/36) */
1006 static const int icos36[9] = {
1007 FIXR(0.50190991877167369479),
1008 FIXR(0.51763809020504152469),
1009 FIXR(0.55168895948124587824),
1010 FIXR(0.61038729438072803416),
1011 FIXR(0.70710678118654752439),
1012 FIXR(0.87172339781054900991),
1013 FIXR(1.18310079157624925896),
1014 FIXR(1.93185165257813657349),
1015 FIXR(5.73685662283492756461),
1018 static const int icos72[18] = {
1019 /* 0.5 / cos(pi*(2*i+19)/72) */
1020 FIXR(0.74009361646113053152),
1021 FIXR(0.82133981585229078570),
1022 FIXR(0.93057949835178895673),
1023 FIXR(1.08284028510010010928),
1024 FIXR(1.30656296487637652785),
1025 FIXR(1.66275476171152078719),
1026 FIXR(2.31011315767264929558),
1027 FIXR(3.83064878777019433457),
1028 FIXR(11.46279281302667383546),
1030 /* 0.5 / cos(pi*(2*(i + 18) +19)/72) */
1031 FIXR(-0.67817085245462840086),
1032 FIXR(-0.63023620700513223342),
1033 FIXR(-0.59284452371708034528),
1034 FIXR(-0.56369097343317117734),
1035 FIXR(-0.54119610014619698439),
1036 FIXR(-0.52426456257040533932),
1037 FIXR(-0.51213975715725461845),
1038 FIXR(-0.50431448029007636036),
1039 FIXR(-0.50047634258165998492),
1042 /* using Lee like decomposition followed by hand coded 9 points DCT */
1043 static void imdct36(int *out, int *in)
1045 int i, j, t0, t1, t2, t3, s0, s1, s2, s3;
1046 int tmp[18], *tmp1, *in1;
1047 int64_t in3_3, in6_6;
1058 in3_3 = MUL64(in1[2*3], C3);
1059 in6_6 = MUL64(in1[2*6], C6);
1061 tmp1[0] = FRAC_RND(MUL64(in1[2*1], C1) + in3_3 +
1062 MUL64(in1[2*5], C5) + MUL64(in1[2*7], C7));
1063 tmp1[2] = in1[2*0] + FRAC_RND(MUL64(in1[2*2], C2) +
1064 MUL64(in1[2*4], C4) + in6_6 +
1065 MUL64(in1[2*8], C8));
1066 tmp1[4] = FRAC_RND(MUL64(in1[2*1] - in1[2*5] - in1[2*7], C3));
1067 tmp1[6] = FRAC_RND(MUL64(in1[2*2] - in1[2*4] - in1[2*8], C6)) -
1068 in1[2*6] + in1[2*0];
1069 tmp1[8] = FRAC_RND(MUL64(in1[2*1], C5) - in3_3 -
1070 MUL64(in1[2*5], C7) + MUL64(in1[2*7], C1));
1071 tmp1[10] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C8) -
1072 MUL64(in1[2*4], C2) + in6_6 +
1073 MUL64(in1[2*8], C4));
1074 tmp1[12] = FRAC_RND(MUL64(in1[2*1], C7) - in3_3 +
1075 MUL64(in1[2*5], C1) -
1076 MUL64(in1[2*7], C5));
1077 tmp1[14] = in1[2*0] + FRAC_RND(MUL64(-in1[2*2], C4) +
1078 MUL64(in1[2*4], C8) + in6_6 -
1079 MUL64(in1[2*8], C2));
1080 tmp1[16] = in1[2*0] - in1[2*2] + in1[2*4] - in1[2*6] + in1[2*8];
1092 s1 = MULL(t3 + t2, icos36[j]);
1093 s3 = MULL(t3 - t2, icos36[8 - j]);
1095 t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
1096 t1 = MULL(s0 - s1, icos72[8 - j]);
1097 out[18 + 9 + j] = t0;
1098 out[18 + 8 - j] = t0;
1102 t0 = MULL(s2 + s3, icos72[9+j]);
1103 t1 = MULL(s2 - s3, icos72[j]);
1104 out[18 + 9 + (8 - j)] = t0;
1106 out[9 + (8 - j)] = -t1;
1112 s1 = MULL(tmp[17], icos36[4]);
1113 t0 = MULL(s0 + s1, icos72[9 + 4]);
1114 t1 = MULL(s0 - s1, icos72[4]);
1115 out[18 + 9 + 4] = t0;
1116 out[18 + 8 - 4] = t0;
1121 /* header decoding. MUST check the header before because no
1122 consistency check is done there. Return 1 if free format found and
1123 that the frame size must be computed externally */
1124 static int decode_header(MPADecodeContext *s, uint32_t header)
1126 int sample_rate, frame_size, mpeg25, padding;
1127 int sample_rate_index, bitrate_index;
1128 if (header & (1<<20)) {
1129 s->lsf = (header & (1<<19)) ? 0 : 1;
1136 s->layer = 4 - ((header >> 17) & 3);
1137 /* extract frequency */
1138 sample_rate_index = (header >> 10) & 3;
1139 sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
1140 sample_rate_index += 3 * (s->lsf + mpeg25);
1141 s->sample_rate_index = sample_rate_index;
1142 s->error_protection = ((header >> 16) & 1) ^ 1;
1143 s->sample_rate = sample_rate;
1145 bitrate_index = (header >> 12) & 0xf;
1146 padding = (header >> 9) & 1;
1147 //extension = (header >> 8) & 1;
1148 s->mode = (header >> 6) & 3;
1149 s->mode_ext = (header >> 4) & 3;
1150 //copyright = (header >> 3) & 1;
1151 //original = (header >> 2) & 1;
1152 //emphasis = header & 3;
1154 if (s->mode == MPA_MONO)
1159 if (bitrate_index != 0) {
1160 frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
1161 s->bit_rate = frame_size * 1000;
1164 frame_size = (frame_size * 12000) / sample_rate;
1165 frame_size = (frame_size + padding) * 4;
1168 frame_size = (frame_size * 144000) / sample_rate;
1169 frame_size += padding;
1173 frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
1174 frame_size += padding;
1177 s->frame_size = frame_size;
1179 /* if no frame size computed, signal it */
1180 if (!s->free_format_frame_size)
1182 /* free format: compute bitrate and real frame size from the
1183 frame size we extracted by reading the bitstream */
1184 s->frame_size = s->free_format_frame_size;
1187 s->frame_size += padding * 4;
1188 s->bit_rate = (s->frame_size * sample_rate) / 48000;
1191 s->frame_size += padding;
1192 s->bit_rate = (s->frame_size * sample_rate) / 144000;
1196 s->frame_size += padding;
1197 s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
1203 printf("layer%d, %d Hz, %d kbits/s, ",
1204 s->layer, s->sample_rate, s->bit_rate);
1205 if (s->nb_channels == 2) {
1206 if (s->layer == 3) {
1207 if (s->mode_ext & MODE_EXT_MS_STEREO)
1209 if (s->mode_ext & MODE_EXT_I_STEREO)
1221 /* useful helper to get mpeg audio stream infos. Return -1 if error in
1222 header, otherwise the coded frame size in bytes */
1223 int mpa_decode_header(AVCodecContext *avctx, uint32_t head)
1225 MPADecodeContext s1, *s = &s1;
1226 memset( s, 0, sizeof(MPADecodeContext) );
1228 if (ff_mpa_check_header(head) != 0)
1231 if (decode_header(s, head) != 0) {
1237 avctx->frame_size = 384;
1240 avctx->frame_size = 1152;
1245 avctx->frame_size = 576;
1247 avctx->frame_size = 1152;
1251 avctx->sample_rate = s->sample_rate;
1252 avctx->channels = s->nb_channels;
1253 avctx->bit_rate = s->bit_rate;
1254 avctx->sub_id = s->layer;
1255 return s->frame_size;
1258 /* return the number of decoded frames */
1259 static int mp_decode_layer1(MPADecodeContext *s)
1261 int bound, i, v, n, ch, j, mant;
1262 uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
1263 uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];
1265 if (s->mode == MPA_JSTEREO)
1266 bound = (s->mode_ext + 1) * 4;
1270 /* allocation bits */
1271 for(i=0;i<bound;i++) {
1272 for(ch=0;ch<s->nb_channels;ch++) {
1273 allocation[ch][i] = get_bits(&s->gb, 4);
1276 for(i=bound;i<SBLIMIT;i++) {
1277 allocation[0][i] = get_bits(&s->gb, 4);
1281 for(i=0;i<bound;i++) {
1282 for(ch=0;ch<s->nb_channels;ch++) {
1283 if (allocation[ch][i])
1284 scale_factors[ch][i] = get_bits(&s->gb, 6);
1287 for(i=bound;i<SBLIMIT;i++) {
1288 if (allocation[0][i]) {
1289 scale_factors[0][i] = get_bits(&s->gb, 6);
1290 scale_factors[1][i] = get_bits(&s->gb, 6);
1294 /* compute samples */
1296 for(i=0;i<bound;i++) {
1297 for(ch=0;ch<s->nb_channels;ch++) {
1298 n = allocation[ch][i];
1300 mant = get_bits(&s->gb, n + 1);
1301 v = l1_unscale(n, mant, scale_factors[ch][i]);
1305 s->sb_samples[ch][j][i] = v;
1308 for(i=bound;i<SBLIMIT;i++) {
1309 n = allocation[0][i];
1311 mant = get_bits(&s->gb, n + 1);
1312 v = l1_unscale(n, mant, scale_factors[0][i]);
1313 s->sb_samples[0][j][i] = v;
1314 v = l1_unscale(n, mant, scale_factors[1][i]);
1315 s->sb_samples[1][j][i] = v;
1317 s->sb_samples[0][j][i] = 0;
1318 s->sb_samples[1][j][i] = 0;
1325 /* bitrate is in kb/s */
1326 int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
1328 int ch_bitrate, table;
1330 ch_bitrate = bitrate / nb_channels;
1332 if ((freq == 48000 && ch_bitrate >= 56) ||
1333 (ch_bitrate >= 56 && ch_bitrate <= 80))
1335 else if (freq != 48000 && ch_bitrate >= 96)
1337 else if (freq != 32000 && ch_bitrate <= 48)
1347 static int mp_decode_layer2(MPADecodeContext *s)
1349 int sblimit; /* number of used subbands */
1350 const unsigned char *alloc_table;
1351 int table, bit_alloc_bits, i, j, ch, bound, v;
1352 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
1353 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
1354 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
1355 int scale, qindex, bits, steps, k, l, m, b;
1357 /* select decoding table */
1358 table = l2_select_table(s->bit_rate / 1000, s->nb_channels,
1359 s->sample_rate, s->lsf);
1360 sblimit = sblimit_table[table];
1361 alloc_table = alloc_tables[table];
1363 if (s->mode == MPA_JSTEREO)
1364 bound = (s->mode_ext + 1) * 4;
1368 dprintf("bound=%d sblimit=%d\n", bound, sblimit);
1371 if( bound > sblimit ) bound = sblimit;
1373 /* parse bit allocation */
1375 for(i=0;i<bound;i++) {
1376 bit_alloc_bits = alloc_table[j];
1377 for(ch=0;ch<s->nb_channels;ch++) {
1378 bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
1380 j += 1 << bit_alloc_bits;
1382 for(i=bound;i<sblimit;i++) {
1383 bit_alloc_bits = alloc_table[j];
1384 v = get_bits(&s->gb, bit_alloc_bits);
1385 bit_alloc[0][i] = v;
1386 bit_alloc[1][i] = v;
1387 j += 1 << bit_alloc_bits;
1392 for(ch=0;ch<s->nb_channels;ch++) {
1393 for(i=0;i<sblimit;i++)
1394 printf(" %d", bit_alloc[ch][i]);
1401 for(i=0;i<sblimit;i++) {
1402 for(ch=0;ch<s->nb_channels;ch++) {
1403 if (bit_alloc[ch][i])
1404 scale_code[ch][i] = get_bits(&s->gb, 2);
1409 for(i=0;i<sblimit;i++) {
1410 for(ch=0;ch<s->nb_channels;ch++) {
1411 if (bit_alloc[ch][i]) {
1412 sf = scale_factors[ch][i];
1413 switch(scale_code[ch][i]) {
1416 sf[0] = get_bits(&s->gb, 6);
1417 sf[1] = get_bits(&s->gb, 6);
1418 sf[2] = get_bits(&s->gb, 6);
1421 sf[0] = get_bits(&s->gb, 6);
1426 sf[0] = get_bits(&s->gb, 6);
1427 sf[2] = get_bits(&s->gb, 6);
1431 sf[0] = get_bits(&s->gb, 6);
1432 sf[2] = get_bits(&s->gb, 6);
1441 for(ch=0;ch<s->nb_channels;ch++) {
1442 for(i=0;i<sblimit;i++) {
1443 if (bit_alloc[ch][i]) {
1444 sf = scale_factors[ch][i];
1445 printf(" %d %d %d", sf[0], sf[1], sf[2]);
1456 for(l=0;l<12;l+=3) {
1458 for(i=0;i<bound;i++) {
1459 bit_alloc_bits = alloc_table[j];
1460 for(ch=0;ch<s->nb_channels;ch++) {
1461 b = bit_alloc[ch][i];
1463 scale = scale_factors[ch][i][k];
1464 qindex = alloc_table[j+b];
1465 bits = quant_bits[qindex];
1467 /* 3 values at the same time */
1468 v = get_bits(&s->gb, -bits);
1469 steps = quant_steps[qindex];
1470 s->sb_samples[ch][k * 12 + l + 0][i] =
1471 l2_unscale_group(steps, v % steps, scale);
1473 s->sb_samples[ch][k * 12 + l + 1][i] =
1474 l2_unscale_group(steps, v % steps, scale);
1476 s->sb_samples[ch][k * 12 + l + 2][i] =
1477 l2_unscale_group(steps, v, scale);
1480 v = get_bits(&s->gb, bits);
1481 v = l1_unscale(bits - 1, v, scale);
1482 s->sb_samples[ch][k * 12 + l + m][i] = v;
1486 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1487 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1488 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1491 /* next subband in alloc table */
1492 j += 1 << bit_alloc_bits;
1494 /* XXX: find a way to avoid this duplication of code */
1495 for(i=bound;i<sblimit;i++) {
1496 bit_alloc_bits = alloc_table[j];
1497 b = bit_alloc[0][i];
1499 int mant, scale0, scale1;
1500 scale0 = scale_factors[0][i][k];
1501 scale1 = scale_factors[1][i][k];
1502 qindex = alloc_table[j+b];
1503 bits = quant_bits[qindex];
1505 /* 3 values at the same time */
1506 v = get_bits(&s->gb, -bits);
1507 steps = quant_steps[qindex];
1510 s->sb_samples[0][k * 12 + l + 0][i] =
1511 l2_unscale_group(steps, mant, scale0);
1512 s->sb_samples[1][k * 12 + l + 0][i] =
1513 l2_unscale_group(steps, mant, scale1);
1516 s->sb_samples[0][k * 12 + l + 1][i] =
1517 l2_unscale_group(steps, mant, scale0);
1518 s->sb_samples[1][k * 12 + l + 1][i] =
1519 l2_unscale_group(steps, mant, scale1);
1520 s->sb_samples[0][k * 12 + l + 2][i] =
1521 l2_unscale_group(steps, v, scale0);
1522 s->sb_samples[1][k * 12 + l + 2][i] =
1523 l2_unscale_group(steps, v, scale1);
1526 mant = get_bits(&s->gb, bits);
1527 s->sb_samples[0][k * 12 + l + m][i] =
1528 l1_unscale(bits - 1, mant, scale0);
1529 s->sb_samples[1][k * 12 + l + m][i] =
1530 l1_unscale(bits - 1, mant, scale1);
1534 s->sb_samples[0][k * 12 + l + 0][i] = 0;
1535 s->sb_samples[0][k * 12 + l + 1][i] = 0;
1536 s->sb_samples[0][k * 12 + l + 2][i] = 0;
1537 s->sb_samples[1][k * 12 + l + 0][i] = 0;
1538 s->sb_samples[1][k * 12 + l + 1][i] = 0;
1539 s->sb_samples[1][k * 12 + l + 2][i] = 0;
1541 /* next subband in alloc table */
1542 j += 1 << bit_alloc_bits;
1544 /* fill remaining samples to zero */
1545 for(i=sblimit;i<SBLIMIT;i++) {
1546 for(ch=0;ch<s->nb_channels;ch++) {
1547 s->sb_samples[ch][k * 12 + l + 0][i] = 0;
1548 s->sb_samples[ch][k * 12 + l + 1][i] = 0;
1549 s->sb_samples[ch][k * 12 + l + 2][i] = 0;
1558 * Seek back in the stream for backstep bytes (at most 511 bytes)
1560 static void seek_to_maindata(MPADecodeContext *s, unsigned int backstep)
1564 /* compute current position in stream */
1565 ptr = (uint8_t *)(s->gb.buffer + (get_bits_count(&s->gb)>>3));
1567 /* copy old data before current one */
1569 memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] +
1570 BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
1571 /* init get bits again */
1572 init_get_bits(&s->gb, ptr, (s->frame_size + backstep)*8);
1574 /* prepare next buffer */
1575 s->inbuf_index ^= 1;
1576 s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
1577 s->old_frame_size = s->frame_size;
1580 static inline void lsf_sf_expand(int *slen,
1581 int sf, int n1, int n2, int n3)
1600 static void exponents_from_scale_factors(MPADecodeContext *s,
1604 const uint8_t *bstab, *pretab;
1605 int len, i, j, k, l, v0, shift, gain, gains[3];
1608 exp_ptr = exponents;
1609 gain = g->global_gain - 210;
1610 shift = g->scalefac_scale + 1;
1612 bstab = band_size_long[s->sample_rate_index];
1613 pretab = mpa_pretab[g->preflag];
1614 for(i=0;i<g->long_end;i++) {
1615 v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift);
1621 if (g->short_start < 13) {
1622 bstab = band_size_short[s->sample_rate_index];
1623 gains[0] = gain - (g->subblock_gain[0] << 3);
1624 gains[1] = gain - (g->subblock_gain[1] << 3);
1625 gains[2] = gain - (g->subblock_gain[2] << 3);
1627 for(i=g->short_start;i<13;i++) {
1630 v0 = gains[l] - (g->scale_factors[k++] << shift);
1638 /* handle n = 0 too */
1639 static inline int get_bitsz(GetBitContext *s, int n)
1644 return get_bits(s, n);
1647 static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
1648 int16_t *exponents, int end_pos)
1651 int linbits, code, x, y, l, v, i, j, k, pos;
1652 GetBitContext last_gb;
1654 uint8_t *code_table;
1656 /* low frequencies (called big values) */
1659 j = g->region_size[i];
1662 /* select vlc table */
1663 k = g->table_select[i];
1664 l = mpa_huff_data[k][0];
1665 linbits = mpa_huff_data[k][1];
1667 code_table = huff_code_table[l];
1669 /* read huffcode and compute each couple */
1671 if (get_bits_count(&s->gb) >= end_pos)
1674 code = get_vlc(&s->gb, vlc);
1677 y = code_table[code];
1684 dprintf("region=%d n=%d x=%d y=%d exp=%d\n",
1685 i, g->region_size[i] - j, x, y, exponents[s_index]);
1688 x += get_bitsz(&s->gb, linbits);
1689 v = l3_unscale(x, exponents[s_index]);
1690 if (get_bits1(&s->gb))
1695 g->sb_hybrid[s_index++] = v;
1698 y += get_bitsz(&s->gb, linbits);
1699 v = l3_unscale(y, exponents[s_index]);
1700 if (get_bits1(&s->gb))
1705 g->sb_hybrid[s_index++] = v;
1709 /* high frequencies */
1710 vlc = &huff_quad_vlc[g->count1table_select];
1711 last_gb.buffer = NULL;
1712 while (s_index <= 572) {
1713 pos = get_bits_count(&s->gb);
1714 if (pos >= end_pos) {
1715 if (pos > end_pos && last_gb.buffer != NULL) {
1716 /* some encoders generate an incorrect size for this
1717 part. We must go back into the data */
1725 code = get_vlc(&s->gb, vlc);
1726 dprintf("t=%d code=%d\n", g->count1table_select, code);
1730 if (code & (8 >> i)) {
1731 /* non zero value. Could use a hand coded function for
1733 v = l3_unscale(1, exponents[s_index]);
1734 if(get_bits1(&s->gb))
1739 g->sb_hybrid[s_index++] = v;
1742 while (s_index < 576)
1743 g->sb_hybrid[s_index++] = 0;
1747 /* Reorder short blocks from bitstream order to interleaved order. It
1748 would be faster to do it in parsing, but the code would be far more
1750 static void reorder_block(MPADecodeContext *s, GranuleDef *g)
1753 int32_t *ptr, *dst, *ptr1;
1756 if (g->block_type != 2)
1759 if (g->switch_point) {
1760 if (s->sample_rate_index != 8) {
1761 ptr = g->sb_hybrid + 36;
1763 ptr = g->sb_hybrid + 48;
1769 for(i=g->short_start;i<13;i++) {
1770 len = band_size_short[s->sample_rate_index][i];
1774 for(j=len;j>0;j--) {
1779 memcpy(ptr1, tmp, len * 3 * sizeof(int32_t));
1783 #define ISQRT2 FIXR(0.70710678118654752440)
1785 static void compute_stereo(MPADecodeContext *s,
1786 GranuleDef *g0, GranuleDef *g1)
1790 int sf_max, tmp0, tmp1, sf, len, non_zero_found;
1791 int32_t (*is_tab)[16];
1792 int32_t *tab0, *tab1;
1793 int non_zero_found_short[3];
1795 /* intensity stereo */
1796 if (s->mode_ext & MODE_EXT_I_STEREO) {
1801 is_tab = is_table_lsf[g1->scalefac_compress & 1];
1805 tab0 = g0->sb_hybrid + 576;
1806 tab1 = g1->sb_hybrid + 576;
1808 non_zero_found_short[0] = 0;
1809 non_zero_found_short[1] = 0;
1810 non_zero_found_short[2] = 0;
1811 k = (13 - g1->short_start) * 3 + g1->long_end - 3;
1812 for(i = 12;i >= g1->short_start;i--) {
1813 /* for last band, use previous scale factor */
1816 len = band_size_short[s->sample_rate_index][i];
1820 if (!non_zero_found_short[l]) {
1821 /* test if non zero band. if so, stop doing i-stereo */
1822 for(j=0;j<len;j++) {
1824 non_zero_found_short[l] = 1;
1828 sf = g1->scale_factors[k + l];
1834 for(j=0;j<len;j++) {
1836 tab0[j] = MULL(tmp0, v1);
1837 tab1[j] = MULL(tmp0, v2);
1841 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1842 /* lower part of the spectrum : do ms stereo
1844 for(j=0;j<len;j++) {
1847 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1848 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1855 non_zero_found = non_zero_found_short[0] |
1856 non_zero_found_short[1] |
1857 non_zero_found_short[2];
1859 for(i = g1->long_end - 1;i >= 0;i--) {
1860 len = band_size_long[s->sample_rate_index][i];
1863 /* test if non zero band. if so, stop doing i-stereo */
1864 if (!non_zero_found) {
1865 for(j=0;j<len;j++) {
1871 /* for last band, use previous scale factor */
1872 k = (i == 21) ? 20 : i;
1873 sf = g1->scale_factors[k];
1878 for(j=0;j<len;j++) {
1880 tab0[j] = MULL(tmp0, v1);
1881 tab1[j] = MULL(tmp0, v2);
1885 if (s->mode_ext & MODE_EXT_MS_STEREO) {
1886 /* lower part of the spectrum : do ms stereo
1888 for(j=0;j<len;j++) {
1891 tab0[j] = MULL(tmp0 + tmp1, ISQRT2);
1892 tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
1897 } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
1898 /* ms stereo ONLY */
1899 /* NOTE: the 1/sqrt(2) normalization factor is included in the
1901 tab0 = g0->sb_hybrid;
1902 tab1 = g1->sb_hybrid;
1903 for(i=0;i<576;i++) {
1906 tab0[i] = tmp0 + tmp1;
1907 tab1[i] = tmp0 - tmp1;
1912 static void compute_antialias_integer(MPADecodeContext *s,
1915 int32_t *ptr, *p0, *p1, *csa;
1918 /* we antialias only "long" bands */
1919 if (g->block_type == 2) {
1920 if (!g->switch_point)
1922 /* XXX: check this for 8000Hz case */
1928 ptr = g->sb_hybrid + 18;
1929 for(i = n;i > 0;i--) {
1932 csa = &csa_table[0][0];
1937 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1938 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1940 int64_t tmp2= MUL64(tmp0 + tmp1, csa[0]);
1941 *p0 = FRAC_RND(tmp2 - MUL64(tmp1, csa[2]));
1942 *p1 = FRAC_RND(tmp2 + MUL64(tmp0, csa[3]));
1949 *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
1950 *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
1952 tmp2= MUL64(tmp0 + tmp1, csa[0]);
1953 *p0 = FRAC_RND(tmp2 - MUL64(tmp1, csa[2]));
1954 *p1 = FRAC_RND(tmp2 + MUL64(tmp0, csa[3]));
1963 static void compute_antialias_float(MPADecodeContext *s,
1966 int32_t *ptr, *p0, *p1;
1969 /* we antialias only "long" bands */
1970 if (g->block_type == 2) {
1971 if (!g->switch_point)
1973 /* XXX: check this for 8000Hz case */
1979 ptr = g->sb_hybrid + 18;
1980 for(i = n;i > 0;i--) {
1981 float *csa = &csa_table_float[0][0];
1988 *p0 = lrintf(tmp0 * csa[0] - tmp1 * csa[1]);
1989 *p1 = lrintf(tmp0 * csa[1] + tmp1 * csa[0]);
1991 float tmp2= (tmp0 + tmp1) * csa[0];
1992 *p0 = lrintf(tmp2 - tmp1 * csa[2]);
1993 *p1 = lrintf(tmp2 + tmp0 * csa[3]);
2000 *p0 = lrintf(tmp0 * csa[0] - tmp1 * csa[1]);
2001 *p1 = lrintf(tmp0 * csa[1] + tmp1 * csa[0]);
2003 tmp2= (tmp0 + tmp1) * csa[0];
2004 *p0 = lrintf(tmp2 - tmp1 * csa[2]);
2005 *p1 = lrintf(tmp2 + tmp0 * csa[3]);
2014 static void compute_imdct(MPADecodeContext *s,
2016 int32_t *sb_samples,
2019 int32_t *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
2023 int i, j, k, mdct_long_end, v, sblimit;
2025 /* find last non zero block */
2026 ptr = g->sb_hybrid + 576;
2027 ptr1 = g->sb_hybrid + 2 * 18;
2028 while (ptr >= ptr1) {
2030 v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
2034 sblimit = ((ptr - g->sb_hybrid) / 18) + 1;
2036 if (g->block_type == 2) {
2037 /* XXX: check for 8000 Hz */
2038 if (g->switch_point)
2043 mdct_long_end = sblimit;
2048 for(j=0;j<mdct_long_end;j++) {
2050 /* apply window & overlap with previous buffer */
2051 out_ptr = sb_samples + j;
2053 if (g->switch_point && j < 2)
2056 win1 = mdct_win[g->block_type];
2057 /* select frequency inversion */
2058 win = win1 + ((4 * 36) & -(j & 1));
2060 *out_ptr = MULL(out[i], win[i]) + buf[i];
2061 buf[i] = MULL(out[i + 18], win[i + 18]);
2067 for(j=mdct_long_end;j<sblimit;j++) {
2073 /* select frequency inversion */
2074 win = mdct_win[2] + ((4 * 36) & -(j & 1));
2077 /* reorder input for short mdct */
2084 /* apply 12 point window and do small overlap */
2086 buf2[i] = MULL(out2[i], win[i]) + buf2[i];
2087 buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
2092 out_ptr = sb_samples + j;
2094 *out_ptr = out[i] + buf[i];
2095 buf[i] = out[i + 18];
2102 for(j=sblimit;j<SBLIMIT;j++) {
2104 out_ptr = sb_samples + j;
2115 void sample_dump(int fnum, int32_t *tab, int n)
2117 static FILE *files[16], *f;
2124 snprintf(buf, sizeof(buf), "/tmp/out%d.%s.pcm",
2126 #ifdef USE_HIGHPRECISION
2132 f = fopen(buf, "w");
2140 printf("pos=%d\n", pos);
2142 printf(" %0.4f", (double)tab[i] / FRAC_ONE);
2149 /* normalize to 23 frac bits */
2150 v = tab[i] << (23 - FRAC_BITS);
2151 fwrite(&v, 1, sizeof(int32_t), f);
2157 /* main layer3 decoding function */
2158 static int mp_decode_layer3(MPADecodeContext *s)
2160 int nb_granules, main_data_begin, private_bits;
2161 int gr, ch, blocksplit_flag, i, j, k, n, bits_pos, bits_left;
2162 GranuleDef granules[2][2], *g;
2163 int16_t exponents[576];
2165 /* read side info */
2167 main_data_begin = get_bits(&s->gb, 8);
2168 if (s->nb_channels == 2)
2169 private_bits = get_bits(&s->gb, 2);
2171 private_bits = get_bits(&s->gb, 1);
2174 main_data_begin = get_bits(&s->gb, 9);
2175 if (s->nb_channels == 2)
2176 private_bits = get_bits(&s->gb, 3);
2178 private_bits = get_bits(&s->gb, 5);
2180 for(ch=0;ch<s->nb_channels;ch++) {
2181 granules[ch][0].scfsi = 0; /* all scale factors are transmitted */
2182 granules[ch][1].scfsi = get_bits(&s->gb, 4);
2186 for(gr=0;gr<nb_granules;gr++) {
2187 for(ch=0;ch<s->nb_channels;ch++) {
2188 dprintf("gr=%d ch=%d: side_info\n", gr, ch);
2189 g = &granules[ch][gr];
2190 g->part2_3_length = get_bits(&s->gb, 12);
2191 g->big_values = get_bits(&s->gb, 9);
2192 g->global_gain = get_bits(&s->gb, 8);
2193 /* if MS stereo only is selected, we precompute the
2194 1/sqrt(2) renormalization factor */
2195 if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
2197 g->global_gain -= 2;
2199 g->scalefac_compress = get_bits(&s->gb, 9);
2201 g->scalefac_compress = get_bits(&s->gb, 4);
2202 blocksplit_flag = get_bits(&s->gb, 1);
2203 if (blocksplit_flag) {
2204 g->block_type = get_bits(&s->gb, 2);
2205 if (g->block_type == 0)
2207 g->switch_point = get_bits(&s->gb, 1);
2209 g->table_select[i] = get_bits(&s->gb, 5);
2211 g->subblock_gain[i] = get_bits(&s->gb, 3);
2212 /* compute huffman coded region sizes */
2213 if (g->block_type == 2)
2214 g->region_size[0] = (36 / 2);
2216 if (s->sample_rate_index <= 2)
2217 g->region_size[0] = (36 / 2);
2218 else if (s->sample_rate_index != 8)
2219 g->region_size[0] = (54 / 2);
2221 g->region_size[0] = (108 / 2);
2223 g->region_size[1] = (576 / 2);
2225 int region_address1, region_address2, l;
2227 g->switch_point = 0;
2229 g->table_select[i] = get_bits(&s->gb, 5);
2230 /* compute huffman coded region sizes */
2231 region_address1 = get_bits(&s->gb, 4);
2232 region_address2 = get_bits(&s->gb, 3);
2233 dprintf("region1=%d region2=%d\n",
2234 region_address1, region_address2);
2236 band_index_long[s->sample_rate_index][region_address1 + 1] >> 1;
2237 l = region_address1 + region_address2 + 2;
2238 /* should not overflow */
2242 band_index_long[s->sample_rate_index][l] >> 1;
2244 /* convert region offsets to region sizes and truncate
2245 size to big_values */
2246 g->region_size[2] = (576 / 2);
2249 k = g->region_size[i];
2250 if (k > g->big_values)
2252 g->region_size[i] = k - j;
2256 /* compute band indexes */
2257 if (g->block_type == 2) {
2258 if (g->switch_point) {
2259 /* if switched mode, we handle the 36 first samples as
2260 long blocks. For 8000Hz, we handle the 48 first
2261 exponents as long blocks (XXX: check this!) */
2262 if (s->sample_rate_index <= 2)
2264 else if (s->sample_rate_index != 8)
2267 g->long_end = 4; /* 8000 Hz */
2269 if (s->sample_rate_index != 8)
2278 g->short_start = 13;
2284 g->preflag = get_bits(&s->gb, 1);
2285 g->scalefac_scale = get_bits(&s->gb, 1);
2286 g->count1table_select = get_bits(&s->gb, 1);
2287 dprintf("block_type=%d switch_point=%d\n",
2288 g->block_type, g->switch_point);
2293 /* now we get bits from the main_data_begin offset */
2294 dprintf("seekback: %d\n", main_data_begin);
2295 seek_to_maindata(s, main_data_begin);
2298 for(gr=0;gr<nb_granules;gr++) {
2299 for(ch=0;ch<s->nb_channels;ch++) {
2300 g = &granules[ch][gr];
2302 bits_pos = get_bits_count(&s->gb);
2306 int slen, slen1, slen2;
2308 /* MPEG1 scale factors */
2309 slen1 = slen_table[0][g->scalefac_compress];
2310 slen2 = slen_table[1][g->scalefac_compress];
2311 dprintf("slen1=%d slen2=%d\n", slen1, slen2);
2312 if (g->block_type == 2) {
2313 n = g->switch_point ? 17 : 18;
2316 g->scale_factors[j++] = get_bitsz(&s->gb, slen1);
2318 g->scale_factors[j++] = get_bitsz(&s->gb, slen2);
2320 g->scale_factors[j++] = 0;
2322 sc = granules[ch][0].scale_factors;
2325 n = (k == 0 ? 6 : 5);
2326 if ((g->scfsi & (0x8 >> k)) == 0) {
2327 slen = (k < 2) ? slen1 : slen2;
2329 g->scale_factors[j++] = get_bitsz(&s->gb, slen);
2331 /* simply copy from last granule */
2333 g->scale_factors[j] = sc[j];
2338 g->scale_factors[j++] = 0;
2342 printf("scfsi=%x gr=%d ch=%d scale_factors:\n",
2345 printf(" %d", g->scale_factors[i]);
2350 int tindex, tindex2, slen[4], sl, sf;
2352 /* LSF scale factors */
2353 if (g->block_type == 2) {
2354 tindex = g->switch_point ? 2 : 1;
2358 sf = g->scalefac_compress;
2359 if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
2360 /* intensity stereo case */
2363 lsf_sf_expand(slen, sf, 6, 6, 0);
2365 } else if (sf < 244) {
2366 lsf_sf_expand(slen, sf - 180, 4, 4, 0);
2369 lsf_sf_expand(slen, sf - 244, 3, 0, 0);
2375 lsf_sf_expand(slen, sf, 5, 4, 4);
2377 } else if (sf < 500) {
2378 lsf_sf_expand(slen, sf - 400, 5, 4, 0);
2381 lsf_sf_expand(slen, sf - 500, 3, 0, 0);
2389 n = lsf_nsf_table[tindex2][tindex][k];
2392 g->scale_factors[j++] = get_bitsz(&s->gb, sl);
2394 /* XXX: should compute exact size */
2396 g->scale_factors[j] = 0;
2399 printf("gr=%d ch=%d scale_factors:\n",
2402 printf(" %d", g->scale_factors[i]);
2408 exponents_from_scale_factors(s, g, exponents);
2410 /* read Huffman coded residue */
2411 if (huffman_decode(s, g, exponents,
2412 bits_pos + g->part2_3_length) < 0)
2415 sample_dump(0, g->sb_hybrid, 576);
2418 /* skip extension bits */
2419 bits_left = g->part2_3_length - (get_bits_count(&s->gb) - bits_pos);
2420 if (bits_left < 0) {
2421 dprintf("bits_left=%d\n", bits_left);
2424 while (bits_left >= 16) {
2425 skip_bits(&s->gb, 16);
2429 skip_bits(&s->gb, bits_left);
2432 if (s->nb_channels == 2)
2433 compute_stereo(s, &granules[0][gr], &granules[1][gr]);
2435 for(ch=0;ch<s->nb_channels;ch++) {
2436 g = &granules[ch][gr];
2438 reorder_block(s, g);
2440 sample_dump(0, g->sb_hybrid, 576);
2442 s->compute_antialias(s, g);
2444 sample_dump(1, g->sb_hybrid, 576);
2446 compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
2448 sample_dump(2, &s->sb_samples[ch][18 * gr][0], 576);
2452 return nb_granules * 18;
2455 static int mp_decode_frame(MPADecodeContext *s,
2458 int i, nb_frames, ch;
2461 init_get_bits(&s->gb, s->inbuf + HEADER_SIZE,
2462 (s->inbuf_ptr - s->inbuf - HEADER_SIZE)*8);
2464 /* skip error protection field */
2465 if (s->error_protection)
2466 get_bits(&s->gb, 16);
2468 dprintf("frame %d:\n", s->frame_count);
2471 nb_frames = mp_decode_layer1(s);
2474 nb_frames = mp_decode_layer2(s);
2478 nb_frames = mp_decode_layer3(s);
2482 for(i=0;i<nb_frames;i++) {
2483 for(ch=0;ch<s->nb_channels;ch++) {
2485 printf("%d-%d:", i, ch);
2486 for(j=0;j<SBLIMIT;j++)
2487 printf(" %0.6f", (double)s->sb_samples[ch][i][j] / FRAC_ONE);
2492 /* apply the synthesis filter */
2493 for(ch=0;ch<s->nb_channels;ch++) {
2494 samples_ptr = samples + ch;
2495 for(i=0;i<nb_frames;i++) {
2496 ff_mpa_synth_filter(s->synth_buf[ch], &(s->synth_buf_offset[ch]),
2497 window, &s->dither_state,
2498 samples_ptr, s->nb_channels,
2499 s->sb_samples[ch][i]);
2500 samples_ptr += 32 * s->nb_channels;
2506 return nb_frames * 32 * sizeof(short) * s->nb_channels;
2509 static int decode_frame(AVCodecContext * avctx,
2510 void *data, int *data_size,
2511 uint8_t * buf, int buf_size)
2513 MPADecodeContext *s = avctx->priv_data;
2517 short *out_samples = data;
2520 while (buf_size > 0) {
2521 len = s->inbuf_ptr - s->inbuf;
2522 if (s->frame_size == 0) {
2523 /* special case for next header for first frame in free
2524 format case (XXX: find a simpler method) */
2525 if (s->free_format_next_header != 0) {
2526 s->inbuf[0] = s->free_format_next_header >> 24;
2527 s->inbuf[1] = s->free_format_next_header >> 16;
2528 s->inbuf[2] = s->free_format_next_header >> 8;
2529 s->inbuf[3] = s->free_format_next_header;
2530 s->inbuf_ptr = s->inbuf + 4;
2531 s->free_format_next_header = 0;
2534 /* no header seen : find one. We need at least HEADER_SIZE
2535 bytes to parse it */
2536 len = HEADER_SIZE - len;
2540 memcpy(s->inbuf_ptr, buf_ptr, len);
2543 s->inbuf_ptr += len;
2545 if ((s->inbuf_ptr - s->inbuf) >= HEADER_SIZE) {
2547 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2548 (s->inbuf[2] << 8) | s->inbuf[3];
2550 if (ff_mpa_check_header(header) < 0) {
2551 /* no sync found : move by one byte (inefficient, but simple!) */
2552 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2554 dprintf("skip %x\n", header);
2555 /* reset free format frame size to give a chance
2556 to get a new bitrate */
2557 s->free_format_frame_size = 0;
2559 if (decode_header(s, header) == 1) {
2560 /* free format: prepare to compute frame size */
2563 /* update codec info */
2564 avctx->sample_rate = s->sample_rate;
2565 avctx->channels = s->nb_channels;
2566 avctx->bit_rate = s->bit_rate;
2567 avctx->sub_id = s->layer;
2570 avctx->frame_size = 384;
2573 avctx->frame_size = 1152;
2577 avctx->frame_size = 576;
2579 avctx->frame_size = 1152;
2584 } else if (s->frame_size == -1) {
2585 /* free format : find next sync to compute frame size */
2586 len = MPA_MAX_CODED_FRAME_SIZE - len;
2590 /* frame too long: resync */
2592 memmove(s->inbuf, s->inbuf + 1, s->inbuf_ptr - s->inbuf - 1);
2599 memcpy(s->inbuf_ptr, buf_ptr, len);
2600 /* check for header */
2601 p = s->inbuf_ptr - 3;
2602 pend = s->inbuf_ptr + len - 4;
2604 header = (p[0] << 24) | (p[1] << 16) |
2606 header1 = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2607 (s->inbuf[2] << 8) | s->inbuf[3];
2608 /* check with high probability that we have a
2610 if ((header & SAME_HEADER_MASK) ==
2611 (header1 & SAME_HEADER_MASK)) {
2612 /* header found: update pointers */
2613 len = (p + 4) - s->inbuf_ptr;
2617 /* compute frame size */
2618 s->free_format_next_header = header;
2619 s->free_format_frame_size = s->inbuf_ptr - s->inbuf;
2620 padding = (header1 >> 9) & 1;
2622 s->free_format_frame_size -= padding * 4;
2624 s->free_format_frame_size -= padding;
2625 dprintf("free frame size=%d padding=%d\n",
2626 s->free_format_frame_size, padding);
2627 decode_header(s, header1);
2632 /* not found: simply increase pointers */
2634 s->inbuf_ptr += len;
2637 } else if (len < s->frame_size) {
2638 if (s->frame_size > MPA_MAX_CODED_FRAME_SIZE)
2639 s->frame_size = MPA_MAX_CODED_FRAME_SIZE;
2640 len = s->frame_size - len;
2643 memcpy(s->inbuf_ptr, buf_ptr, len);
2645 s->inbuf_ptr += len;
2649 if (s->frame_size > 0 &&
2650 (s->inbuf_ptr - s->inbuf) >= s->frame_size) {
2651 if (avctx->parse_only) {
2652 /* simply return the frame data */
2653 *(uint8_t **)data = s->inbuf;
2654 out_size = s->inbuf_ptr - s->inbuf;
2656 out_size = mp_decode_frame(s, out_samples);
2658 s->inbuf_ptr = s->inbuf;
2660 *data_size = out_size;
2664 return buf_ptr - buf;
2668 static int decode_frame_adu(AVCodecContext * avctx,
2669 void *data, int *data_size,
2670 uint8_t * buf, int buf_size)
2672 MPADecodeContext *s = avctx->priv_data;
2675 short *out_samples = data;
2679 // Discard too short frames
2680 if (buf_size < HEADER_SIZE) {
2686 if (len > MPA_MAX_CODED_FRAME_SIZE)
2687 len = MPA_MAX_CODED_FRAME_SIZE;
2689 memcpy(s->inbuf, buf, len);
2690 s->inbuf_ptr = s->inbuf + len;
2692 // Get header and restore sync word
2693 header = (s->inbuf[0] << 24) | (s->inbuf[1] << 16) |
2694 (s->inbuf[2] << 8) | s->inbuf[3] | 0xffe00000;
2696 if (ff_mpa_check_header(header) < 0) { // Bad header, discard frame
2701 decode_header(s, header);
2702 /* update codec info */
2703 avctx->sample_rate = s->sample_rate;
2704 avctx->channels = s->nb_channels;
2705 avctx->bit_rate = s->bit_rate;
2706 avctx->sub_id = s->layer;
2708 avctx->frame_size=s->frame_size = len;
2710 if (avctx->parse_only) {
2711 /* simply return the frame data */
2712 *(uint8_t **)data = s->inbuf;
2713 out_size = s->inbuf_ptr - s->inbuf;
2715 out_size = mp_decode_frame(s, out_samples);
2718 *data_size = out_size;
2723 AVCodec mp2_decoder =
2728 sizeof(MPADecodeContext),
2733 CODEC_CAP_PARSE_ONLY,
2736 AVCodec mp3_decoder =
2741 sizeof(MPADecodeContext),
2746 CODEC_CAP_PARSE_ONLY,
2749 AVCodec mp3adu_decoder =
2754 sizeof(MPADecodeContext),
2759 CODEC_CAP_PARSE_ONLY,