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 #define real float /* ugly - but only way */
  13
  14 static unsigned long long attribute_used __attribute__((aligned(8))) null_one = 0x0000ffff0000ffffULL;
  15 static unsigned long long attribute_used __attribute__((aligned(8))) one_null = 0xffff0000ffff0000ULL;
  16 unsigned long __attribute__((aligned(8))) 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 static int temp; // buggy gcc 3.x fails if this is moved into the function :(
  52 void synth_1to1_MMX_s(real *bandPtr, int channel, short *samples,
  53                       short *buffs, int *bo)
  54 {
  55
  56 __asm __volatile(
  57         "movl %1,%%ecx\n\t"
  58         "movl %2,%%edi\n\t"
  59         "movl $15,%%ebx\n\t"
  60         "movl %4,%%edx\n\t"
  61         "leal (%%edi,%%ecx,2),%%edi\n\t"
  62         "decl %%ecx\n\t"
  63         "movl %3,%%esi\n\t"
  64         "movl (%%edx),%%eax\n\t"
  65         "jecxz .L01\n\t"
  66         "decl %%eax\n\t"
  67         "andl %%ebx,%%eax\n\t"
  68         "leal 1088(%%esi),%%esi\n\t"
  69         "movl %%eax,(%%edx)\n\t"
  70 ".L01:\n\t"
  71         "leal (%%esi,%%eax,2),%%edx\n\t"
  72         "movl %%eax,%5\n\t"
  73         "incl %%eax\n\t"
  74         "andl %%ebx,%%eax\n\t"
  75         "leal 544(%%esi,%%eax,2),%%ecx\n\t"
  76         "incl %%ebx\n\t"
  77         "testl $1, %%eax\n\t"
  78         "jnz .L02\n\t"
  79         "xchgl %%edx,%%ecx\n\t"
  80         "incl %5\n\t"
  81         "leal 544(%%esi),%%esi\n\t"
  82 ".L02:\n\t"
  83         "emms\n\t"
  84         "pushl %0\n\t"
  85         "pushl %%edx\n\t"
  86         "pushl %%ecx\n\t"
  87         "call *"MANGLE(dct64_MMX_func)"\n\t"
  88         "addl $12, %%esp\n\t"
  89         "leal 1(%%ebx), %%ecx\n\t"
  90         "subl %5,%%ebx\n\t"
  91         "pushl %%ecx\n\t"
  92         "leal "MANGLE(mp3lib_decwins)"(%%ebx,%%ebx,1), %%edx\n\t"
  93         "shrl $1, %%ecx\n\t"
  94 ".balign 16\n\t"
  95 ".L03:\n\t"
  96         "movq  (%%edx),%%mm0\n\t"
  97         "movq  64(%%edx),%%mm4\n\t"
  98         "pmaddwd (%%esi),%%mm0\n\t"
  99         "pmaddwd 32(%%esi),%%mm4\n\t"
 100         "movq  8(%%edx),%%mm1\n\t"
 101         "movq  72(%%edx),%%mm5\n\t"
 102         "pmaddwd 8(%%esi),%%mm1\n\t"
 103         "pmaddwd 40(%%esi),%%mm5\n\t"
 104         "movq  16(%%edx),%%mm2\n\t"
 105         "movq  80(%%edx),%%mm6\n\t"
 106         "pmaddwd 16(%%esi),%%mm2\n\t"
 107         "pmaddwd 48(%%esi),%%mm6\n\t"
 108         "movq  24(%%edx),%%mm3\n\t"
 109         "movq  88(%%edx),%%mm7\n\t"
 110         "pmaddwd 24(%%esi),%%mm3\n\t"
 111         "pmaddwd 56(%%esi),%%mm7\n\t"
 112         "paddd %%mm1,%%mm0\n\t"
 113         "paddd %%mm5,%%mm4\n\t"
 114         "paddd %%mm2,%%mm0\n\t"
 115         "paddd %%mm6,%%mm4\n\t"
 116         "paddd %%mm3,%%mm0\n\t"
 117         "paddd %%mm7,%%mm4\n\t"
 118         "movq  %%mm0,%%mm1\n\t"
 119         "movq  %%mm4,%%mm5\n\t"
 120         "psrlq $32,%%mm1\n\t"
 121         "psrlq $32,%%mm5\n\t"
 122         "paddd %%mm1,%%mm0\n\t"
 123         "paddd %%mm5,%%mm4\n\t"
 124         "psrad $13,%%mm0\n\t"
 125         "psrad $13,%%mm4\n\t"
 126         "packssdw %%mm0,%%mm0\n\t"
 127         "packssdw %%mm4,%%mm4\n\t"
 128
 129         "movq   (%%edi), %%mm1\n\t"
 130         "punpckldq %%mm4, %%mm0\n\t"
 131         "pand   "MANGLE(one_null)", %%mm1\n\t"
 132         "pand   "MANGLE(null_one)", %%mm0\n\t"
 133         "por    %%mm0, %%mm1\n\t"
 134         "movq   %%mm1,(%%edi)\n\t"
 135
 136         "leal 64(%%esi),%%esi\n\t"
 137         "leal 128(%%edx),%%edx\n\t"
 138         "leal 8(%%edi),%%edi\n\t"
 139
 140         "decl %%ecx\n\t"
 141         "jnz  .L03\n\t"
 142
 143         "popl %%ecx\n\t"
 144         "andl $1, %%ecx\n\t"
 145         "jecxz .next_loop\n\t"
 146
 147         "movq  (%%edx),%%mm0\n\t"
 148         "pmaddwd (%%esi),%%mm0\n\t"
 149         "movq  8(%%edx),%%mm1\n\t"
 150         "pmaddwd 8(%%esi),%%mm1\n\t"
 151         "movq  16(%%edx),%%mm2\n\t"
 152         "pmaddwd 16(%%esi),%%mm2\n\t"
 153         "movq  24(%%edx),%%mm3\n\t"
 154         "pmaddwd 24(%%esi),%%mm3\n\t"
 155         "paddd %%mm1,%%mm0\n\t"
 156         "paddd %%mm2,%%mm0\n\t"
 157         "paddd %%mm3,%%mm0\n\t"
 158         "movq  %%mm0,%%mm1\n\t"
 159         "psrlq $32,%%mm1\n\t"
 160         "paddd %%mm1,%%mm0\n\t"
 161         "psrad $13,%%mm0\n\t"
 162         "packssdw %%mm0,%%mm0\n\t"
 163         "movd %%mm0,%%eax\n\t"
 164         "movw %%ax, (%%edi)\n\t"
 165         "leal 32(%%esi),%%esi\n\t"
 166         "leal 64(%%edx),%%edx\n\t"
 167         "leal 4(%%edi),%%edi\n\t"
 168
 169 ".next_loop:\n\t"
 170         "subl $64,%%esi\n\t"
 171         "movl $7,%%ecx\n\t"
 172 ".balign 16\n\t"
 173 ".L04:\n\t"
 174         "movq  (%%edx),%%mm0\n\t"
 175         "movq  64(%%edx),%%mm4\n\t"
 176         "pmaddwd (%%esi),%%mm0\n\t"
 177         "pmaddwd -32(%%esi),%%mm4\n\t"
 178         "movq  8(%%edx),%%mm1\n\t"
 179         "movq  72(%%edx),%%mm5\n\t"
 180         "pmaddwd 8(%%esi),%%mm1\n\t"
 181         "pmaddwd -24(%%esi),%%mm5\n\t"
 182         "movq  16(%%edx),%%mm2\n\t"
 183         "movq  80(%%edx),%%mm6\n\t"
 184         "pmaddwd 16(%%esi),%%mm2\n\t"
 185         "pmaddwd -16(%%esi),%%mm6\n\t"
 186         "movq  24(%%edx),%%mm3\n\t"
 187         "movq  88(%%edx),%%mm7\n\t"
 188         "pmaddwd 24(%%esi),%%mm3\n\t"
 189         "pmaddwd -8(%%esi),%%mm7\n\t"
 190         "paddd %%mm1,%%mm0\n\t"
 191         "paddd %%mm5,%%mm4\n\t"
 192         "paddd %%mm2,%%mm0\n\t"
 193         "paddd %%mm6,%%mm4\n\t"
 194         "paddd %%mm3,%%mm0\n\t"
 195         "paddd %%mm7,%%mm4\n\t"
 196         "movq  %%mm0,%%mm1\n\t"
 197         "movq  %%mm4,%%mm5\n\t"
 198         "psrlq $32,%%mm1\n\t"
 199         "psrlq $32,%%mm5\n\t"
 200         "paddd %%mm0,%%mm1\n\t"
 201         "paddd %%mm4,%%mm5\n\t"
 202         "psrad $13,%%mm1\n\t"
 203         "psrad $13,%%mm5\n\t"
 204         "packssdw %%mm1,%%mm1\n\t"
 205         "packssdw %%mm5,%%mm5\n\t"
 206         "psubd %%mm0,%%mm0\n\t"
 207         "psubd %%mm4,%%mm4\n\t"
 208         "psubsw %%mm1,%%mm0\n\t"
 209         "psubsw %%mm5,%%mm4\n\t"
 210
 211         "movq   (%%edi), %%mm1\n\t"
 212         "punpckldq %%mm4, %%mm0\n\t"
 213         "pand   "MANGLE(one_null)", %%mm1\n\t"
 214         "pand   "MANGLE(null_one)", %%mm0\n\t"
 215         "por    %%mm0, %%mm1\n\t"
 216         "movq   %%mm1,(%%edi)\n\t"
 217
 218         "subl $64,%%esi\n\t"
 219         "addl $128,%%edx\n\t"
 220         "leal 8(%%edi),%%edi\n\t"
 221         "decl %%ecx\n\t"
 222         "jnz  .L04\n\t"
 223
 224         "movq  (%%edx),%%mm0\n\t"
 225         "pmaddwd (%%esi),%%mm0\n\t"
 226         "movq  8(%%edx),%%mm1\n\t"
 227         "pmaddwd 8(%%esi),%%mm1\n\t"
 228         "movq  16(%%edx),%%mm2\n\t"
 229         "pmaddwd 16(%%esi),%%mm2\n\t"
 230         "movq  24(%%edx),%%mm3\n\t"
 231         "pmaddwd 24(%%esi),%%mm3\n\t"
 232         "paddd %%mm1,%%mm0\n\t"
 233         "paddd %%mm2,%%mm0\n\t"
 234         "paddd %%mm3,%%mm0\n\t"
 235         "movq  %%mm0,%%mm1\n\t"
 236         "psrlq $32,%%mm1\n\t"
 237         "paddd %%mm0,%%mm1\n\t"
 238         "psrad $13,%%mm1\n\t"
 239         "packssdw %%mm1,%%mm1\n\t"
 240         "psubd %%mm0,%%mm0\n\t"
 241         "psubsw %%mm1,%%mm0\n\t"
 242         "movd %%mm0,%%eax\n\t"
 243         "movw %%ax,(%%edi)\n\t"
 244         "emms\n\t"
 245         :
 246         :"m"(bandPtr),"m"(channel),"m"(samples),"m"(buffs),"m"(bo), "m"(temp)
 247         :"memory","%edi","%esi","%ebx","%esp");
 248 }