mp3lib/decode_mmx.c

   1 /*
   2  * this code comes under GPL
   3  * This code was taken from http://www.mpg123.org
   4  * See ChangeLog of mpg123-0.59s-pre.1 for detail
   5  * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
   6  *
   7  * Local ChangeLog:
   8  * - Partial loops unrolling and removing MOVW insn from loops
   9 */
  10 #include "config.h"
  11 #include "mangle.h"
  12 #include "mpg123.h"
  13
  14 static const unsigned long long attribute_used __attribute__((aligned(8))) null_one = 0x0000ffff0000ffffULL;
  15 static const unsigned long long attribute_used __attribute__((aligned(8))) one_null = 0xffff0000ffff0000ULL;
  16 const unsigned int __attribute__((aligned(16))) costab_mmx[] =
  17 {
  18     1056974725,
  19     1057056395,
  20     1057223771,
  21     1057485416,
  22     1057855544,
  23     1058356026,
  24     1059019886,
  25     1059897405,
  26     1061067246,
  27     1062657950,
  28     1064892987,
  29     1066774581,
  30     1069414683,
  31     1073984175,
  32     1079645762,
  33     1092815430,
  34     1057005197,
  35     1057342072,
  36     1058087743,
  37     1059427869,
  38     1061799040,
  39     1065862217,
  40     1071413542,
  41     1084439708,
  42     1057128951,
  43     1058664893,
  44     1063675095,
  45     1076102863,
  46     1057655764,
  47     1067924853,
  48     1060439283,
  49 };
  50
  51 /**
  52   This array of magic numbers were calculated by the pure function
  53   make_decode_tables_MMX(32768), which had been implemented in (deleted since
  54   r23383) tabinit_MMX.c.
  55   */
  56 static const short __attribute__((aligned(8))) mp3lib_decwins[] =
  57 {
  58         0,      7,     54,    114,    510,   1288,   1644,   9372,
  59     18760,  -9373,   1644,  -1289,    510,   -115,     54,     -8,
  60         0,      7,     54,    114,    510,   1288,   1644,   9372,
  61     18760,  -9373,   1644,  -1289,    510,   -115,     54,     -8,
  62         0,      7,     55,    129,    500,   1379,   1490,   9834,
  63     18748,  -8910,   1784,  -1197,    516,   -101,     52,     -7,
  64         0,      7,     55,    129,    500,   1379,   1490,   9834,
  65     18748,  -8910,   1784,  -1197,    516,   -101,     52,     -7,
  66         0,      8,     56,    145,    488,   1469,   1322,  10294,
  67     18714,  -8448,   1910,  -1107,    520,    -87,     51,     -6,
  68         0,      8,     56,    145,    488,   1469,   1322,  10294,
  69     18714,  -8448,   1910,  -1107,    520,    -87,     51,     -6,
  70         0,      9,     57,    161,    474,   1559,   1141,  10751,
  71     18658,  -7987,   2023,  -1016,    522,    -74,     49,     -6,
  72         0,      9,     57,    161,    474,   1559,   1141,  10751,
  73     18658,  -7987,   2023,  -1016,    522,    -74,     49,     -6,
  74         0,     10,     57,    177,    456,   1647,    944,  11205,
  75     18579,  -7528,   2123,   -927,    522,    -61,     48,     -5,
  76         0,     10,     57,    177,    456,   1647,    944,  11205,
  77     18579,  -7528,   2123,   -927,    522,    -61,     48,     -5,
  78         0,     11,     57,    194,    435,   1733,    734,  11654,
  79     18477,  -7073,   2210,   -838,    519,    -50,     46,     -5,
  80         0,     11,     57,    194,    435,   1733,    734,  11654,
  81     18477,  -7073,   2210,   -838,    519,    -50,     46,     -5,
  82         0,     12,     57,    212,    411,   1817,    510,  12097,
  83     18354,  -6621,   2285,   -751,    515,    -39,     44,     -4,
  84         0,     12,     57,    212,    411,   1817,    510,  12097,
  85     18354,  -6621,   2285,   -751,    515,    -39,     44,     -4,
  86         0,     13,     57,    229,    384,   1899,    271,  12534,
  87     18209,  -6174,   2348,   -666,    508,    -28,     43,     -4,
  88         0,     13,     57,    229,    384,   1899,    271,  12534,
  89     18209,  -6174,   2348,   -666,    508,    -28,     43,     -4,
  90         0,     14,     56,    247,    354,   1977,     18,  12963,
  91     18043,  -5733,   2398,   -583,    501,    -18,     41,     -4,
  92         0,     14,     56,    247,    354,   1977,     18,  12963,
  93     18043,  -5733,   2398,   -583,    501,    -18,     41,     -4,
  94         0,     15,     56,    266,    320,   2052,   -249,  13383,
  95     17855,  -5298,   2438,   -502,    491,     -9,     39,     -3,
  96         0,     15,     56,    266,    320,   2052,   -249,  13383,
  97     17855,  -5298,   2438,   -502,    491,     -9,     39,     -3,
  98         0,     17,     54,    284,    283,   2122,   -530,  13794,
  99     17648,  -4870,   2466,   -423,    480,     -1,     37,     -3,
 100         0,     17,     54,    284,    283,   2122,   -530,  13794,
 101     17648,  -4870,   2466,   -423,    480,     -1,     37,     -3,
 102         0,     18,     52,    302,    243,   2188,   -825,  14194,
 103     17420,  -4450,   2484,   -347,    468,      7,     35,     -3,
 104         0,     18,     52,    302,    243,   2188,   -825,  14194,
 105     17420,  -4450,   2484,   -347,    468,      7,     35,     -3,
 106         0,     19,     50,    320,    199,   2249,  -1133,  14583,
 107     17173,  -4039,   2492,   -274,    455,     14,     33,     -2,
 108         0,     19,     50,    320,    199,   2249,  -1133,  14583,
 109     17173,  -4039,   2492,   -274,    455,     14,     33,     -2,
 110        -1,     21,     48,    339,    152,   2304,  -1454,  14959,
 111     16908,  -3637,   2490,   -204,    440,     20,     32,     -2,
 112        -1,     21,     48,    339,    152,   2304,  -1454,  14959,
 113     16908,  -3637,   2490,   -204,    440,     20,     32,     -2,
 114        -1,     22,     45,    357,    101,   2354,  -1788,  15322,
 115     16624,  -3245,   2479,   -137,    425,     26,     30,     -2,
 116        -1,     22,     45,    357,    101,   2354,  -1788,  15322,
 117     16624,  -3245,   2479,   -137,    425,     26,     30,     -2,
 118        -1,     24,     41,    374,     47,   2396,  -2135,  15671,
 119     16323,  -2864,   2460,    -72,    409,     31,     28,     -2,
 120        -1,     24,     41,    374,     47,   2396,  -2135,  15671,
 121     16323,  -2864,   2460,    -72,    409,     31,     28,     -2,
 122        -1,     26,     37,    391,    -11,   2431,  -2493,  16004,
 123     16005,  -2494,   2432,    -12,    392,     36,     26,     -2,
 124        -1,     26,     37,    391,    -11,   2431,  -2493,  16004,
 125     16005,  -2494,   2432,    -12,    392,     36,     26,     -2,
 126        -2,    -28,     31,   -409,    -72,  -2460,  -2864, -16323,
 127     15671,   2135,   2396,    -47,    374,    -41,     24,      1,
 128        -2,    -28,     31,   -409,    -72,  -2460,  -2864, -16323,
 129     15671,   2135,   2396,    -47,    374,    -41,     24,      1,
 130        -2,    -30,     26,   -425,   -137,  -2479,  -3245, -16624,
 131     15322,   1788,   2354,   -101,    357,    -45,     22,      1,
 132        -2,    -30,     26,   -425,   -137,  -2479,  -3245, -16624,
 133     15322,   1788,   2354,   -101,    357,    -45,     22,      1,
 134        -2,    -32,     20,   -440,   -204,  -2490,  -3637, -16908,
 135     14959,   1454,   2304,   -152,    339,    -48,     21,      1,
 136        -2,    -32,     20,   -440,   -204,  -2490,  -3637, -16908,
 137     14959,   1454,   2304,   -152,    339,    -48,     21,      1,
 138        -2,    -33,     14,   -455,   -274,  -2492,  -4039, -17173,
 139     14583,   1133,   2249,   -199,    320,    -50,     19,      0,
 140        -2,    -33,     14,   -455,   -274,  -2492,  -4039, -17173,
 141     14583,   1133,   2249,   -199,    320,    -50,     19,      0,
 142        -3,    -35,      7,   -468,   -347,  -2484,  -4450, -17420,
 143     14194,    825,   2188,   -243,    302,    -52,     18,      0,
 144        -3,    -35,      7,   -468,   -347,  -2484,  -4450, -17420,
 145     14194,    825,   2188,   -243,    302,    -52,     18,      0,
 146        -3,    -37,     -1,   -480,   -423,  -2466,  -4870, -17648,
 147     13794,    530,   2122,   -283,    284,    -54,     17,      0,
 148        -3,    -37,     -1,   -480,   -423,  -2466,  -4870, -17648,
 149     13794,    530,   2122,   -283,    284,    -54,     17,      0,
 150        -3,    -39,     -9,   -491,   -502,  -2438,  -5298, -17855,
 151     13383,    249,   2052,   -320,    266,    -56,     15,      0,
 152        -3,    -39,     -9,   -491,   -502,  -2438,  -5298, -17855,
 153     13383,    249,   2052,   -320,    266,    -56,     15,      0,
 154        -4,    -41,    -18,   -501,   -583,  -2398,  -5733, -18043,
 155     12963,    -18,   1977,   -354,    247,    -56,     14,      0,
 156        -4,    -41,    -18,   -501,   -583,  -2398,  -5733, -18043,
 157     12963,    -18,   1977,   -354,    247,    -56,     14,      0,
 158        -4,    -43,    -28,   -508,   -666,  -2348,  -6174, -18209,
 159     12534,   -271,   1899,   -384,    229,    -57,     13,      0,
 160        -4,    -43,    -28,   -508,   -666,  -2348,  -6174, -18209,
 161     12534,   -271,   1899,   -384,    229,    -57,     13,      0,
 162        -4,    -44,    -39,   -515,   -751,  -2285,  -6621, -18354,
 163     12097,   -510,   1817,   -411,    212,    -57,     12,      0,
 164        -4,    -44,    -39,   -515,   -751,  -2285,  -6621, -18354,
 165     12097,   -510,   1817,   -411,    212,    -57,     12,      0,
 166        -5,    -46,    -50,   -519,   -838,  -2210,  -7073, -18477,
 167     11654,   -734,   1733,   -435,    194,    -57,     11,      0,
 168        -5,    -46,    -50,   -519,   -838,  -2210,  -7073, -18477,
 169     11654,   -734,   1733,   -435,    194,    -57,     11,      0,
 170        -5,    -48,    -61,   -522,   -927,  -2123,  -7528, -18579,
 171     11205,   -944,   1647,   -456,    177,    -57,     10,      0,
 172        -5,    -48,    -61,   -522,   -927,  -2123,  -7528, -18579,
 173     11205,   -944,   1647,   -456,    177,    -57,     10,      0,
 174        -6,    -49,    -74,   -522,  -1016,  -2023,  -7987, -18658,
 175     10751,  -1141,   1559,   -474,    161,    -57,      9,      0,
 176        -6,    -49,    -74,   -522,  -1016,  -2023,  -7987, -18658,
 177     10751,  -1141,   1559,   -474,    161,    -57,      9,      0,
 178        -6,    -51,    -87,   -520,  -1107,  -1910,  -8448, -18714,
 179     10294,  -1322,   1469,   -488,    145,    -56,      8,      0,
 180        -6,    -51,    -87,   -520,  -1107,  -1910,  -8448, -18714,
 181     10294,  -1322,   1469,   -488,    145,    -56,      8,      0,
 182        -7,    -52,   -101,   -516,  -1197,  -1784,  -8910, -18748,
 183      9834,  -1490,   1379,   -500,    129,    -55,      7,      0,
 184        -7,    -52,   -101,   -516,  -1197,  -1784,  -8910, -18748,
 185      9834,  -1490,   1379,   -500,    129,    -55,      7,      0,
 186 };
 187
 188 int synth_1to1_MMX(real *bandPtr, int channel, short *samples)
 189 {
 190     static short buffs[2][2][0x110] __attribute__((aligned(8)));
 191     static int bo = 1;
 192     short *b0, (*buf)[0x110], *a, *b;
 193     const short* window;
 194     int bo1, i = 8;
 195
 196     if (channel == 0) {
 197         bo = (bo - 1) & 0xf;
 198         buf = buffs[1];
 199     } else {
 200         samples++;
 201         buf = buffs[0];
 202     }
 203
 204     if (bo & 1) {
 205         b0 = buf[1];
 206         bo1 = bo + 1;
 207         a = buf[0] + bo;
 208         b = buf[1] + ((bo + 1) & 0xf);
 209     } else {
 210         b0 = buf[0];
 211         bo1 = bo;
 212         b = buf[0] + bo;
 213         a = buf[1] + ((bo + 1) & 0xf);
 214     }
 215
 216     dct64_MMX_func(a, b, bandPtr);
 217     window = mp3lib_decwins + 16 - bo1;
 218     //printf("DEBUG: channel %d, bo %d, off %d\n", channel, bo, 16 - bo1);
 219 __asm__ volatile(
 220 ASMALIGN(4)
 221 "0:\n\t"
 222         "movq  (%1),%%mm0\n\t"
 223         "movq  64(%1),%%mm4\n\t"
 224         "pmaddwd (%2),%%mm0\n\t"
 225         "pmaddwd 32(%2),%%mm4\n\t"
 226         "movq  8(%1),%%mm1\n\t"
 227         "movq  72(%1),%%mm5\n\t"
 228         "pmaddwd 8(%2),%%mm1\n\t"
 229         "pmaddwd 40(%2),%%mm5\n\t"
 230         "movq  16(%1),%%mm2\n\t"
 231         "movq  80(%1),%%mm6\n\t"
 232         "pmaddwd 16(%2),%%mm2\n\t"
 233         "pmaddwd 48(%2),%%mm6\n\t"
 234         "movq  24(%1),%%mm3\n\t"
 235         "movq  88(%1),%%mm7\n\t"
 236         "pmaddwd 24(%2),%%mm3\n\t"
 237         "pmaddwd 56(%2),%%mm7\n\t"
 238         "paddd %%mm1,%%mm0\n\t"
 239         "paddd %%mm5,%%mm4\n\t"
 240         "paddd %%mm2,%%mm0\n\t"
 241         "paddd %%mm6,%%mm4\n\t"
 242         "paddd %%mm3,%%mm0\n\t"
 243         "paddd %%mm7,%%mm4\n\t"
 244         "movq  %%mm0,%%mm1\n\t"
 245         "movq  %%mm4,%%mm5\n\t"
 246         "psrlq $32,%%mm1\n\t"
 247         "psrlq $32,%%mm5\n\t"
 248         "paddd %%mm1,%%mm0\n\t"
 249         "paddd %%mm5,%%mm4\n\t"
 250         "psrad $13,%%mm0\n\t"
 251         "psrad $13,%%mm4\n\t"
 252         "packssdw %%mm0,%%mm0\n\t"
 253         "packssdw %%mm4,%%mm4\n\t"
 254
 255         "movq   (%3), %%mm1\n\t"
 256         "punpckldq %%mm4, %%mm0\n\t"
 257         "pand   "MANGLE(one_null)", %%mm1\n\t"
 258         "pand   "MANGLE(null_one)", %%mm0\n\t"
 259         "por    %%mm0, %%mm1\n\t"
 260         "movq   %%mm1,(%3)\n\t"
 261
 262         "add $64,%2\n\t"
 263         "add $128,%1\n\t"
 264         "add $8,%3\n\t"
 265
 266         "decl %0\n\t"
 267         "jnz  0b\n\t"
 268
 269         "movq  (%1),%%mm0\n\t"
 270         "pmaddwd (%2),%%mm0\n\t"
 271         "movq  8(%1),%%mm1\n\t"
 272         "pmaddwd 8(%2),%%mm1\n\t"
 273         "movq  16(%1),%%mm2\n\t"
 274         "pmaddwd 16(%2),%%mm2\n\t"
 275         "movq  24(%1),%%mm3\n\t"
 276         "pmaddwd 24(%2),%%mm3\n\t"
 277         "paddd %%mm1,%%mm0\n\t"
 278         "paddd %%mm2,%%mm0\n\t"
 279         "paddd %%mm3,%%mm0\n\t"
 280         "movq  %%mm0,%%mm1\n\t"
 281         "psrlq $32,%%mm1\n\t"
 282         "paddd %%mm1,%%mm0\n\t"
 283         "psrad $13,%%mm0\n\t"
 284         "packssdw %%mm0,%%mm0\n\t"
 285         "movd %%mm0,%%eax\n\t"
 286         "movw %%ax, (%3)\n\t"
 287         "sub $32,%2\n\t"
 288         "add $64,%1\n\t"
 289         "add $4,%3\n\t"
 290
 291         "movl $7,%0\n\t"
 292 ASMALIGN(4)
 293 "1:\n\t"
 294         "movq  (%1),%%mm0\n\t"
 295         "movq  64(%1),%%mm4\n\t"
 296         "pmaddwd (%2),%%mm0\n\t"
 297         "pmaddwd -32(%2),%%mm4\n\t"
 298         "movq  8(%1),%%mm1\n\t"
 299         "movq  72(%1),%%mm5\n\t"
 300         "pmaddwd 8(%2),%%mm1\n\t"
 301         "pmaddwd -24(%2),%%mm5\n\t"
 302         "movq  16(%1),%%mm2\n\t"
 303         "movq  80(%1),%%mm6\n\t"
 304         "pmaddwd 16(%2),%%mm2\n\t"
 305         "pmaddwd -16(%2),%%mm6\n\t"
 306         "movq  24(%1),%%mm3\n\t"
 307         "movq  88(%1),%%mm7\n\t"
 308         "pmaddwd 24(%2),%%mm3\n\t"
 309         "pmaddwd -8(%2),%%mm7\n\t"
 310         "paddd %%mm1,%%mm0\n\t"
 311         "paddd %%mm5,%%mm4\n\t"
 312         "paddd %%mm2,%%mm0\n\t"
 313         "paddd %%mm6,%%mm4\n\t"
 314         "paddd %%mm3,%%mm0\n\t"
 315         "paddd %%mm7,%%mm4\n\t"
 316         "movq  %%mm0,%%mm1\n\t"
 317         "movq  %%mm4,%%mm5\n\t"
 318         "psrlq $32,%%mm1\n\t"
 319         "psrlq $32,%%mm5\n\t"
 320         "paddd %%mm0,%%mm1\n\t"
 321         "paddd %%mm4,%%mm5\n\t"
 322         "psrad $13,%%mm1\n\t"
 323         "psrad $13,%%mm5\n\t"
 324         "packssdw %%mm1,%%mm1\n\t"
 325         "packssdw %%mm5,%%mm5\n\t"
 326         "psubd %%mm0,%%mm0\n\t"
 327         "psubd %%mm4,%%mm4\n\t"
 328         "psubsw %%mm1,%%mm0\n\t"
 329         "psubsw %%mm5,%%mm4\n\t"
 330
 331         "movq   (%3), %%mm1\n\t"
 332         "punpckldq %%mm4, %%mm0\n\t"
 333         "pand   "MANGLE(one_null)", %%mm1\n\t"
 334         "pand   "MANGLE(null_one)", %%mm0\n\t"
 335         "por    %%mm0, %%mm1\n\t"
 336         "movq   %%mm1,(%3)\n\t"
 337
 338         "sub $64,%2\n\t"
 339         "add $128,%1\n\t"
 340         "add $8,%3\n\t"
 341         "decl %0\n\t"
 342         "jnz  1b\n\t"
 343
 344         "movq  (%1),%%mm0\n\t"
 345         "pmaddwd (%2),%%mm0\n\t"
 346         "movq  8(%1),%%mm1\n\t"
 347         "pmaddwd 8(%2),%%mm1\n\t"
 348         "movq  16(%1),%%mm2\n\t"
 349         "pmaddwd 16(%2),%%mm2\n\t"
 350         "movq  24(%1),%%mm3\n\t"
 351         "pmaddwd 24(%2),%%mm3\n\t"
 352         "paddd %%mm1,%%mm0\n\t"
 353         "paddd %%mm2,%%mm0\n\t"
 354         "paddd %%mm3,%%mm0\n\t"
 355         "movq  %%mm0,%%mm1\n\t"
 356         "psrlq $32,%%mm1\n\t"
 357         "paddd %%mm0,%%mm1\n\t"
 358         "psrad $13,%%mm1\n\t"
 359         "packssdw %%mm1,%%mm1\n\t"
 360         "psubd %%mm0,%%mm0\n\t"
 361         "psubsw %%mm1,%%mm0\n\t"
 362         "movd %%mm0,%%eax\n\t"
 363         "movw %%ax,(%3)\n\t"
 364         "emms\n\t"
 365         :"+r"(i), "+r"(window), "+r"(b0), "+r"(samples)
 366         :
 367         :"memory", "%eax");
 368     return 0;
 369 }