mp3lib/dct36_3dnow.c

   1 /*
   2  * dct36_3dnow.c - 3DNow! optimized dct36()
   3  *
   4  * This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
   5  * <squash@mb.kcom.ne.jp>, only two types of changes have been made:
   6  *
   7  * - removed PREFETCH instruction for speedup
   8  * - changed function name for support 3DNow! automatic detection
   9  *
  10  * You can find Kashiyama's original 3dnow! support patch
  11  * (for mpg123-0.59o) at
  12  * http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
  13  *
  14  * by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
  15  *                    <kim@comtec.co.jp>               - after  1.Apr.1999
  16  *
  17  * Modified for use with MPlayer, for details see the changelog at
  18  * http://svn.mplayerhq.hu/mplayer/trunk/
  19  * $Id$
  20  *
  21  * Original disclaimer:
  22  *  The author of this program disclaim whole expressed or implied
  23  *  warranties with regard to this program, and in no event shall the
  24  *  author of this program liable to whatever resulted from the use of
  25  *  this program. Use it at your own risk.
  26  *
  27  * 2003/06/21: Moved to GCC inline assembly - Alex Beregszaszi
  28  */
  29
  30 #define real float /* ugly - but only way */
  31
  32 #include "mangle.h"
  33
  34 #ifdef DCT36_OPTIMIZE_FOR_K7
  35 void dct36_3dnowex(real *inbuf, real *o1,
  36     real *o2, real *wintab, real *tsbuf)
  37 #else
  38 void dct36_3dnow(real *inbuf, real *o1,
  39     real *o2, real *wintab, real *tsbuf)
  40 #endif
  41 {
  42     __asm__ __volatile__(
  43         "movq (%%eax),%%mm0\n\t"
  44         "movq 4(%%eax),%%mm1\n\t"
  45         "pfadd %%mm1,%%mm0\n\t"
  46         "movq %%mm0,4(%%eax)\n\t"
  47         "psrlq $32,%%mm1\n\t"
  48         "movq 12(%%eax),%%mm2\n\t"
  49         "punpckldq %%mm2,%%mm1\n\t"
  50         "pfadd %%mm2,%%mm1\n\t"
  51         "movq %%mm1,12(%%eax)\n\t"
  52         "psrlq $32,%%mm2\n\t"
  53         "movq 20(%%eax),%%mm3\n\t"
  54         "punpckldq %%mm3,%%mm2\n\t"
  55         "pfadd %%mm3,%%mm2\n\t"
  56         "movq %%mm2,20(%%eax)\n\t"
  57         "psrlq $32,%%mm3\n\t"
  58         "movq 28(%%eax),%%mm4\n\t"
  59         "punpckldq %%mm4,%%mm3\n\t"
  60         "pfadd %%mm4,%%mm3\n\t"
  61         "movq %%mm3,28(%%eax)\n\t"
  62         "psrlq $32,%%mm4\n\t"
  63         "movq 36(%%eax),%%mm5\n\t"
  64         "punpckldq %%mm5,%%mm4\n\t"
  65         "pfadd %%mm5,%%mm4\n\t"
  66         "movq %%mm4,36(%%eax)\n\t"
  67         "psrlq $32,%%mm5\n\t"
  68         "movq 44(%%eax),%%mm6\n\t"
  69         "punpckldq %%mm6,%%mm5\n\t"
  70         "pfadd %%mm6,%%mm5\n\t"
  71         "movq %%mm5,44(%%eax)\n\t"
  72         "psrlq $32,%%mm6\n\t"
  73         "movq 52(%%eax),%%mm7\n\t"
  74         "punpckldq %%mm7,%%mm6\n\t"
  75         "pfadd %%mm7,%%mm6\n\t"
  76         "movq %%mm6,52(%%eax)\n\t"
  77         "psrlq $32,%%mm7\n\t"
  78         "movq 60(%%eax),%%mm0\n\t"
  79         "punpckldq %%mm0,%%mm7\n\t"
  80         "pfadd %%mm0,%%mm7\n\t"
  81         "movq %%mm7,60(%%eax)\n\t"
  82         "psrlq $32,%%mm0\n\t"
  83         "movd 68(%%eax),%%mm1\n\t"
  84         "pfadd %%mm1,%%mm0\n\t"
  85         "movd %%mm0,68(%%eax)\n\t"
  86         "movd 4(%%eax),%%mm0\n\t"
  87         "movd 12(%%eax),%%mm1\n\t"
  88         "punpckldq %%mm1,%%mm0\n\t"
  89         "punpckldq 20(%%eax),%%mm1\n\t"
  90         "pfadd %%mm1,%%mm0\n\t"
  91         "movd %%mm0,12(%%eax)\n\t"
  92         "psrlq $32,%%mm0\n\t"
  93         "movd %%mm0,20(%%eax)\n\t"
  94         "psrlq $32,%%mm1\n\t"
  95         "movd 28(%%eax),%%mm2\n\t"
  96         "punpckldq %%mm2,%%mm1\n\t"
  97         "punpckldq 36(%%eax),%%mm2\n\t"
  98         "pfadd %%mm2,%%mm1\n\t"
  99         "movd %%mm1,28(%%eax)\n\t"
 100         "psrlq $32,%%mm1\n\t"
 101         "movd %%mm1,36(%%eax)\n\t"
 102         "psrlq $32,%%mm2\n\t"
 103         "movd 44(%%eax),%%mm3\n\t"
 104         "punpckldq %%mm3,%%mm2\n\t"
 105         "punpckldq 52(%%eax),%%mm3\n\t"
 106         "pfadd %%mm3,%%mm2\n\t"
 107         "movd %%mm2,44(%%eax)\n\t"
 108         "psrlq $32,%%mm2\n\t"
 109         "movd %%mm2,52(%%eax)\n\t"
 110         "psrlq $32,%%mm3\n\t"
 111         "movd 60(%%eax),%%mm4\n\t"
 112         "punpckldq %%mm4,%%mm3\n\t"
 113         "punpckldq 68(%%eax),%%mm4\n\t"
 114         "pfadd %%mm4,%%mm3\n\t"
 115         "movd %%mm3,60(%%eax)\n\t"
 116         "psrlq $32,%%mm3\n\t"
 117         "movd %%mm3,68(%%eax)\n\t"
 118
 119         "movq 24(%%eax),%%mm0\n\t"
 120         "movq 48(%%eax),%%mm1\n\t"
 121         "movd "MANGLE(COS9)"+12,%%mm2\n\t"
 122         "punpckldq %%mm2,%%mm2\n\t"
 123         "movd "MANGLE(COS9)"+24,%%mm3\n\t"
 124         "punpckldq %%mm3,%%mm3\n\t"
 125         "pfmul %%mm2,%%mm0\n\t"
 126         "pfmul %%mm3,%%mm1\n\t"
 127         "pushl %%eax\n\t"
 128         "movl $1,%%eax\n\t"
 129         "movd %%eax,%%mm7\n\t"
 130         "pi2fd %%mm7,%%mm7\n\t"
 131         "popl %%eax\n\t"
 132         "movq 8(%%eax),%%mm2\n\t"
 133         "movd "MANGLE(COS9)"+4,%%mm3\n\t"
 134         "punpckldq %%mm3,%%mm3\n\t"
 135         "pfmul %%mm3,%%mm2\n\t"
 136         "pfadd %%mm0,%%mm2\n\t"
 137         "movq 40(%%eax),%%mm3\n\t"
 138         "movd "MANGLE(COS9)"+20,%%mm4\n\t"
 139         "punpckldq %%mm4,%%mm4\n\t"
 140         "pfmul %%mm4,%%mm3\n\t"
 141         "pfadd %%mm3,%%mm2\n\t"
 142         "movq 56(%%eax),%%mm3\n\t"
 143         "movd "MANGLE(COS9)"+28,%%mm4\n\t"
 144         "punpckldq %%mm4,%%mm4\n\t"
 145         "pfmul %%mm4,%%mm3\n\t"
 146         "pfadd %%mm3,%%mm2\n\t"
 147         "movq (%%eax),%%mm3\n\t"
 148         "movq 16(%%eax),%%mm4\n\t"
 149         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 150         "punpckldq %%mm5,%%mm5\n\t"
 151         "pfmul %%mm5,%%mm4\n\t"
 152         "pfadd %%mm4,%%mm3\n\t"
 153         "movq 32(%%eax),%%mm4\n\t"
 154         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 155         "punpckldq %%mm5,%%mm5\n\t"
 156         "pfmul %%mm5,%%mm4\n\t"
 157         "pfadd %%mm4,%%mm3\n\t"
 158         "pfadd %%mm1,%%mm3\n\t"
 159         "movq 64(%%eax),%%mm4\n\t"
 160         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 161         "punpckldq %%mm5,%%mm5\n\t"
 162         "pfmul %%mm5,%%mm4\n\t"
 163         "pfadd %%mm4,%%mm3\n\t"
 164         "movq %%mm2,%%mm4\n\t"
 165         "pfadd %%mm3,%%mm4\n\t"
 166         "movq %%mm7,%%mm5\n\t"
 167         "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t"
 168         "pfmul %%mm5,%%mm4\n\t"
 169         "movq %%mm4,%%mm5\n\t"
 170         "pfacc %%mm5,%%mm5\n\t"
 171         "movd 108(%%edx),%%mm6\n\t"
 172         "punpckldq 104(%%edx),%%mm6\n\t"
 173         "pfmul %%mm6,%%mm5\n\t"
 174 #ifdef DCT36_OPTIMIZE_FOR_K7
 175         "pswapd %%mm5,%%mm5\n\t"
 176         "movq %%mm5,32(%%ecx)\n\t"
 177 #else
 178         "movd %%mm5,36(%%ecx)\n\t"
 179         "psrlq $32,%%mm5\n\t"
 180         "movd %%mm5,32(%%ecx)\n\t"
 181 #endif
 182         "movq %%mm4,%%mm6\n\t"
 183         "punpckldq %%mm6,%%mm5\n\t"
 184         "pfsub %%mm6,%%mm5\n\t"
 185         "punpckhdq %%mm5,%%mm5\n\t"
 186         "movd 32(%%edx),%%mm6\n\t"
 187         "punpckldq 36(%%edx),%%mm6\n\t"
 188         "pfmul %%mm6,%%mm5\n\t"
 189         "movd 32(%%esi),%%mm6\n\t"
 190         "punpckldq 36(%%esi),%%mm6\n\t"
 191         "pfadd %%mm6,%%mm5\n\t"
 192         "movd %%mm5,1024(%%ebx)\n\t"
 193         "psrlq $32,%%mm5\n\t"
 194         "movd %%mm5,1152(%%ebx)\n\t"
 195         "movq %%mm3,%%mm4\n\t"
 196         "pfsub %%mm2,%%mm4\n\t"
 197         "movq %%mm7,%%mm5\n\t"
 198         "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t"
 199         "pfmul %%mm5,%%mm4\n\t"
 200         "movq %%mm4,%%mm5\n\t"
 201         "pfacc %%mm5,%%mm5\n\t"
 202         "movd 140(%%edx),%%mm6\n\t"
 203         "punpckldq 72(%%edx),%%mm6\n\t"
 204         "pfmul %%mm6,%%mm5\n\t"
 205         "movd %%mm5,68(%%ecx)\n\t"
 206         "psrlq $32,%%mm5\n\t"
 207         "movd %%mm5,0(%%ecx)\n\t"
 208         "movq %%mm4,%%mm6\n\t"
 209         "punpckldq %%mm6,%%mm5\n\t"
 210         "pfsub %%mm6,%%mm5\n\t"
 211         "punpckhdq %%mm5,%%mm5\n\t"
 212         "movd 0(%%edx),%%mm6\n\t"
 213         "punpckldq 68(%%edx),%%mm6\n\t"
 214         "pfmul %%mm6,%%mm5\n\t"
 215         "movd 0(%%esi),%%mm6\n\t"
 216         "punpckldq 68(%%esi),%%mm6\n\t"
 217         "pfadd %%mm6,%%mm5\n\t"
 218         "movd %%mm5,0(%%ebx)\n\t"
 219         "psrlq $32,%%mm5\n\t"
 220         "movd %%mm5,2176(%%ebx)\n\t"
 221         "movq 8(%%eax),%%mm2\n\t"
 222         "movq 40(%%eax),%%mm3\n\t"
 223         "pfsub %%mm3,%%mm2\n\t"
 224         "movq 56(%%eax),%%mm3\n\t"
 225         "pfsub %%mm3,%%mm2\n\t"
 226         "movd "MANGLE(COS9)"+12,%%mm3\n\t"
 227         "punpckldq %%mm3,%%mm3\n\t"
 228         "pfmul %%mm3,%%mm2\n\t"
 229         "movq 16(%%eax),%%mm3\n\t"
 230         "movq 32(%%eax),%%mm4\n\t"
 231         "pfsub %%mm4,%%mm3\n\t"
 232         "movq 64(%%eax),%%mm4\n\t"
 233         "pfsub %%mm4,%%mm3\n\t"
 234         "movd "MANGLE(COS9)"+24,%%mm4\n\t"
 235         "punpckldq %%mm4,%%mm4\n\t"
 236         "pfmul %%mm4,%%mm3\n\t"
 237         "movq 48(%%eax),%%mm4\n\t"
 238         "pfsub %%mm4,%%mm3\n\t"
 239         "movq (%%eax),%%mm4\n\t"
 240         "pfadd %%mm4,%%mm3\n\t"
 241         "movq %%mm2,%%mm4\n\t"
 242         "pfadd %%mm3,%%mm4\n\t"
 243         "movq %%mm7,%%mm5\n\t"
 244         "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t"
 245         "pfmul %%mm5,%%mm4\n\t"
 246         "movq %%mm4,%%mm5\n\t"
 247         "pfacc %%mm5,%%mm5\n\t"
 248         "movd 112(%%edx),%%mm6\n\t"
 249         "punpckldq 100(%%edx),%%mm6\n\t"
 250         "pfmul %%mm6,%%mm5\n\t"
 251         "movd %%mm5,40(%%ecx)\n\t"
 252         "psrlq $32,%%mm5\n\t"
 253         "movd %%mm5,28(%%ecx)\n\t"
 254         "movq %%mm4,%%mm6\n\t"
 255         "punpckldq %%mm6,%%mm5\n\t"
 256         "pfsub %%mm6,%%mm5\n\t"
 257         "punpckhdq %%mm5,%%mm5\n\t"
 258         "movd 28(%%edx),%%mm6\n\t"
 259         "punpckldq 40(%%edx),%%mm6\n\t"
 260         "pfmul %%mm6,%%mm5\n\t"
 261         "movd 28(%%esi),%%mm6\n\t"
 262         "punpckldq 40(%%esi),%%mm6\n\t"
 263         "pfadd %%mm6,%%mm5\n\t"
 264         "movd %%mm5,896(%%ebx)\n\t"
 265         "psrlq $32,%%mm5\n\t"
 266         "movd %%mm5,1280(%%ebx)\n\t"
 267         "movq %%mm3,%%mm4\n\t"
 268         "pfsub %%mm2,%%mm4\n\t"
 269         "movq %%mm7,%%mm5\n\t"
 270         "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t"
 271         "pfmul %%mm5,%%mm4\n\t"
 272         "movq %%mm4,%%mm5\n\t"
 273         "pfacc %%mm5,%%mm5\n\t"
 274         "movd 136(%%edx),%%mm6\n\t"
 275         "punpckldq 76(%%edx),%%mm6\n\t"
 276         "pfmul %%mm6,%%mm5\n\t"
 277         "movd %%mm5,64(%%ecx)\n\t"
 278         "psrlq $32,%%mm5\n\t"
 279         "movd %%mm5,4(%%ecx)\n\t"
 280         "movq %%mm4,%%mm6\n\t"
 281         "punpckldq %%mm6,%%mm5\n\t"
 282         "pfsub %%mm6,%%mm5\n\t"
 283         "punpckhdq %%mm5,%%mm5\n\t"
 284         "movd 4(%%edx),%%mm6\n\t"
 285         "punpckldq 64(%%edx),%%mm6\n\t"
 286         "pfmul %%mm6,%%mm5\n\t"
 287         "movd 4(%%esi),%%mm6\n\t"
 288         "punpckldq 64(%%esi),%%mm6\n\t"
 289         "pfadd %%mm6,%%mm5\n\t"
 290         "movd %%mm5,128(%%ebx)\n\t"
 291         "psrlq $32,%%mm5\n\t"
 292         "movd %%mm5,2048(%%ebx)\n\t"
 293
 294         "movq 8(%%eax),%%mm2\n\t"
 295         "movd "MANGLE(COS9)"+20,%%mm3\n\t"
 296         "punpckldq %%mm3,%%mm3\n\t"
 297         "pfmul %%mm3,%%mm2\n\t"
 298         "pfsub %%mm0,%%mm2\n\t"
 299         "movq 40(%%eax),%%mm3\n\t"
 300         "movd "MANGLE(COS9)"+28,%%mm4\n\t"
 301         "punpckldq %%mm4,%%mm4\n\t"
 302         "pfmul %%mm4,%%mm3\n\t"
 303         "pfsub %%mm3,%%mm2\n\t"
 304         "movq 56(%%eax),%%mm3\n\t"
 305         "movd "MANGLE(COS9)"+4,%%mm4\n\t"
 306         "punpckldq %%mm4,%%mm4\n\t"
 307         "pfmul %%mm4,%%mm3\n\t"
 308         "pfadd %%mm3,%%mm2\n\t"
 309         "movq (%%eax),%%mm3\n\t"
 310         "movq 16(%%eax),%%mm4\n\t"
 311         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 312         "punpckldq %%mm5,%%mm5\n\t"
 313         "pfmul %%mm5,%%mm4\n\t"
 314         "pfsub %%mm4,%%mm3\n\t"
 315         "movq 32(%%eax),%%mm4\n\t"
 316         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 317         "punpckldq %%mm5,%%mm5\n\t"
 318         "pfmul %%mm5,%%mm4\n\t"
 319         "pfsub %%mm4,%%mm3\n\t"
 320         "pfadd %%mm1,%%mm3\n\t"
 321         "movq 64(%%eax),%%mm4\n\t"
 322         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 323         "punpckldq %%mm5,%%mm5\n\t"
 324         "pfmul %%mm5,%%mm4\n\t"
 325         "pfadd %%mm4,%%mm3\n\t"
 326         "movq %%mm2,%%mm4\n\t"
 327         "pfadd %%mm3,%%mm4\n\t"
 328         "movq %%mm7,%%mm5\n\t"
 329         "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t"
 330         "pfmul %%mm5,%%mm4\n\t"
 331         "movq %%mm4,%%mm5\n\t"
 332         "pfacc %%mm5,%%mm5\n\t"
 333         "movd 116(%%edx),%%mm6\n\t"
 334         "punpckldq 96(%%edx),%%mm6\n\t"
 335         "pfmul %%mm6,%%mm5\n\t"
 336         "movd %%mm5,44(%%ecx)\n\t"
 337         "psrlq $32,%%mm5\n\t"
 338         "movd %%mm5,24(%%ecx)\n\t"
 339         "movq %%mm4,%%mm6\n\t"
 340         "punpckldq %%mm6,%%mm5\n\t"
 341         "pfsub %%mm6,%%mm5\n\t"
 342         "punpckhdq %%mm5,%%mm5\n\t"
 343         "movd 24(%%edx),%%mm6\n\t"
 344         "punpckldq 44(%%edx),%%mm6\n\t"
 345         "pfmul %%mm6,%%mm5\n\t"
 346         "movd 24(%%esi),%%mm6\n\t"
 347         "punpckldq 44(%%esi),%%mm6\n\t"
 348         "pfadd %%mm6,%%mm5\n\t"
 349         "movd %%mm5,768(%%ebx)\n\t"
 350         "psrlq $32,%%mm5\n\t"
 351         "movd %%mm5,1408(%%ebx)\n\t"
 352         "movq %%mm3,%%mm4\n\t"
 353         "pfsub %%mm2,%%mm4\n\t"
 354         "movq %%mm7,%%mm5\n\t"
 355         "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t"
 356         "pfmul %%mm5,%%mm4\n\t"
 357         "movq %%mm4,%%mm5\n\t"
 358         "pfacc %%mm5,%%mm5\n\t"
 359         "movd 132(%%edx),%%mm6\n\t"
 360         "punpckldq 80(%%edx),%%mm6\n\t"
 361         "pfmul %%mm6,%%mm5\n\t"
 362         "movd %%mm5,60(%%ecx)\n\t"
 363         "psrlq $32,%%mm5\n\t"
 364         "movd %%mm5,8(%%ecx)\n\t"
 365         "movq %%mm4,%%mm6\n\t"
 366         "punpckldq %%mm6,%%mm5\n\t"
 367         "pfsub %%mm6,%%mm5\n\t"
 368         "punpckhdq %%mm5,%%mm5\n\t"
 369         "movd 8(%%edx),%%mm6\n\t"
 370         "punpckldq 60(%%edx),%%mm6\n\t"
 371         "pfmul %%mm6,%%mm5\n\t"
 372         "movd 8(%%esi),%%mm6\n\t"
 373         "punpckldq 60(%%esi),%%mm6\n\t"
 374         "pfadd %%mm6,%%mm5\n\t"
 375         "movd %%mm5,256(%%ebx)\n\t"
 376         "psrlq $32,%%mm5\n\t"
 377         "movd %%mm5,1920(%%ebx)\n\t"
 378         "movq 8(%%eax),%%mm2\n\t"
 379         "movd "MANGLE(COS9)"+28,%%mm3\n\t"
 380         "punpckldq %%mm3,%%mm3\n\t"
 381         "pfmul %%mm3,%%mm2\n\t"
 382         "pfsub %%mm0,%%mm2\n\t"
 383         "movq 40(%%eax),%%mm3\n\t"
 384         "movd "MANGLE(COS9)"+4,%%mm4\n\t"
 385         "punpckldq %%mm4,%%mm4\n\t"
 386         "pfmul %%mm4,%%mm3\n\t"
 387         "pfadd %%mm3,%%mm2\n\t"
 388         "movq 56(%%eax),%%mm3\n\t"
 389         "movd "MANGLE(COS9)"+20,%%mm4\n\t"
 390         "punpckldq %%mm4,%%mm4\n\t"
 391         "pfmul %%mm4,%%mm3\n\t"
 392         "pfsub %%mm3,%%mm2\n\t"
 393         "movq (%%eax),%%mm3\n\t"
 394         "movq 16(%%eax),%%mm4\n\t"
 395         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 396         "punpckldq %%mm5,%%mm5\n\t"
 397         "pfmul %%mm5,%%mm4\n\t"
 398         "pfsub %%mm4,%%mm3\n\t"
 399         "movq 32(%%eax),%%mm4\n\t"
 400         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 401         "punpckldq %%mm5,%%mm5\n\t"
 402         "pfmul %%mm5,%%mm4\n\t"
 403         "pfadd %%mm4,%%mm3\n\t"
 404         "pfadd %%mm1,%%mm3\n\t"
 405         "movq 64(%%eax),%%mm4\n\t"
 406         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 407         "punpckldq %%mm5,%%mm5\n\t"
 408         "pfmul %%mm5,%%mm4\n\t"
 409         "pfsub %%mm4,%%mm3\n\t"
 410         "movq %%mm2,%%mm4\n\t"
 411         "pfadd %%mm3,%%mm4\n\t"
 412         "movq %%mm7,%%mm5\n\t"
 413         "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t"
 414         "pfmul %%mm5,%%mm4\n\t"
 415         "movq %%mm4,%%mm5\n\t"
 416         "pfacc %%mm5,%%mm5\n\t"
 417         "movd 120(%%edx),%%mm6\n\t"
 418         "punpckldq 92(%%edx),%%mm6\n\t"
 419         "pfmul %%mm6,%%mm5\n\t"
 420         "movd %%mm5,48(%%ecx)\n\t"
 421         "psrlq $32,%%mm5\n\t"
 422         "movd %%mm5,20(%%ecx)\n\t"
 423         "movq %%mm4,%%mm6\n\t"
 424         "punpckldq %%mm6,%%mm5\n\t"
 425         "pfsub %%mm6,%%mm5\n\t"
 426         "punpckhdq %%mm5,%%mm5\n\t"
 427         "movd 20(%%edx),%%mm6\n\t"
 428         "punpckldq 48(%%edx),%%mm6\n\t"
 429         "pfmul %%mm6,%%mm5\n\t"
 430         "movd 20(%%esi),%%mm6\n\t"
 431         "punpckldq 48(%%esi),%%mm6\n\t"
 432         "pfadd %%mm6,%%mm5\n\t"
 433         "movd %%mm5,640(%%ebx)\n\t"
 434         "psrlq $32,%%mm5\n\t"
 435         "movd %%mm5,1536(%%ebx)\n\t"
 436         "movq %%mm3,%%mm4\n\t"
 437         "pfsub %%mm2,%%mm4\n\t"
 438         "movq %%mm7,%%mm5\n\t"
 439         "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t"
 440         "pfmul %%mm5,%%mm4\n\t"
 441         "movq %%mm4,%%mm5\n\t"
 442         "pfacc %%mm5,%%mm5\n\t"
 443         "movd 128(%%edx),%%mm6\n\t"
 444         "punpckldq 84(%%edx),%%mm6\n\t"
 445         "pfmul %%mm6,%%mm5\n\t"
 446         "movd %%mm5,56(%%ecx)\n\t"
 447         "psrlq $32,%%mm5\n\t"
 448         "movd %%mm5,12(%%ecx)\n\t"
 449         "movq %%mm4,%%mm6\n\t"
 450         "punpckldq %%mm6,%%mm5\n\t"
 451         "pfsub %%mm6,%%mm5\n\t"
 452         "punpckhdq %%mm5,%%mm5\n\t"
 453         "movd 12(%%edx),%%mm6\n\t"
 454         "punpckldq 56(%%edx),%%mm6\n\t"
 455         "pfmul %%mm6,%%mm5\n\t"
 456         "movd 12(%%esi),%%mm6\n\t"
 457         "punpckldq 56(%%esi),%%mm6\n\t"
 458         "pfadd %%mm6,%%mm5\n\t"
 459         "movd %%mm5,384(%%ebx)\n\t"
 460         "psrlq $32,%%mm5\n\t"
 461         "movd %%mm5,1792(%%ebx)\n\t"
 462
 463         "movq (%%eax),%%mm4\n\t"
 464         "movq 16(%%eax),%%mm3\n\t"
 465         "pfsub %%mm3,%%mm4\n\t"
 466         "movq 32(%%eax),%%mm3\n\t"
 467         "pfadd %%mm3,%%mm4\n\t"
 468         "movq 48(%%eax),%%mm3\n\t"
 469         "pfsub %%mm3,%%mm4\n\t"
 470         "movq 64(%%eax),%%mm3\n\t"
 471         "pfadd %%mm3,%%mm4\n\t"
 472         "movq %%mm7,%%mm5\n\t"
 473         "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t"
 474         "pfmul %%mm5,%%mm4\n\t"
 475         "movq %%mm4,%%mm5\n\t"
 476         "pfacc %%mm5,%%mm5\n\t"
 477         "movd 124(%%edx),%%mm6\n\t"
 478         "punpckldq 88(%%edx),%%mm6\n\t"
 479         "pfmul %%mm6,%%mm5\n\t"
 480         "movd %%mm5,52(%%ecx)\n\t"
 481         "psrlq $32,%%mm5\n\t"
 482         "movd %%mm5,16(%%ecx)\n\t"
 483         "movq %%mm4,%%mm6\n\t"
 484         "punpckldq %%mm6,%%mm5\n\t"
 485         "pfsub %%mm6,%%mm5\n\t"
 486         "punpckhdq %%mm5,%%mm5\n\t"
 487         "movd 16(%%edx),%%mm6\n\t"
 488         "punpckldq 52(%%edx),%%mm6\n\t"
 489         "pfmul %%mm6,%%mm5\n\t"
 490         "movd 16(%%esi),%%mm6\n\t"
 491         "punpckldq 52(%%esi),%%mm6\n\t"
 492         "pfadd %%mm6,%%mm5\n\t"
 493         "movd %%mm5,512(%%ebx)\n\t"
 494         "psrlq $32,%%mm5\n\t"
 495         "movd %%mm5,1664(%%ebx)\n\t"
 496
 497         "femms\n\t"
 498         :
 499         : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf)
 500         : "memory");
 501 }