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 "config.h"
  33 #include "mangle.h"
  34
  35 #ifdef DCT36_OPTIMIZE_FOR_K7
  36 void dct36_3dnowex(real *inbuf, real *o1,
  37     real *o2, real *wintab, real *tsbuf)
  38 #else
  39 void dct36_3dnow(real *inbuf, real *o1,
  40     real *o2, real *wintab, real *tsbuf)
  41 #endif
  42 {
  43     __asm__ volatile(
  44         "movq (%%eax),%%mm0\n\t"
  45         "movq 4(%%eax),%%mm1\n\t"
  46         "pfadd %%mm1,%%mm0\n\t"
  47         "movq %%mm0,4(%%eax)\n\t"
  48         "psrlq $32,%%mm1\n\t"
  49         "movq 12(%%eax),%%mm2\n\t"
  50         "punpckldq %%mm2,%%mm1\n\t"
  51         "pfadd %%mm2,%%mm1\n\t"
  52         "movq %%mm1,12(%%eax)\n\t"
  53         "psrlq $32,%%mm2\n\t"
  54         "movq 20(%%eax),%%mm3\n\t"
  55         "punpckldq %%mm3,%%mm2\n\t"
  56         "pfadd %%mm3,%%mm2\n\t"
  57         "movq %%mm2,20(%%eax)\n\t"
  58         "psrlq $32,%%mm3\n\t"
  59         "movq 28(%%eax),%%mm4\n\t"
  60         "punpckldq %%mm4,%%mm3\n\t"
  61         "pfadd %%mm4,%%mm3\n\t"
  62         "movq %%mm3,28(%%eax)\n\t"
  63         "psrlq $32,%%mm4\n\t"
  64         "movq 36(%%eax),%%mm5\n\t"
  65         "punpckldq %%mm5,%%mm4\n\t"
  66         "pfadd %%mm5,%%mm4\n\t"
  67         "movq %%mm4,36(%%eax)\n\t"
  68         "psrlq $32,%%mm5\n\t"
  69         "movq 44(%%eax),%%mm6\n\t"
  70         "punpckldq %%mm6,%%mm5\n\t"
  71         "pfadd %%mm6,%%mm5\n\t"
  72         "movq %%mm5,44(%%eax)\n\t"
  73         "psrlq $32,%%mm6\n\t"
  74         "movq 52(%%eax),%%mm7\n\t"
  75         "punpckldq %%mm7,%%mm6\n\t"
  76         "pfadd %%mm7,%%mm6\n\t"
  77         "movq %%mm6,52(%%eax)\n\t"
  78         "psrlq $32,%%mm7\n\t"
  79         "movq 60(%%eax),%%mm0\n\t"
  80         "punpckldq %%mm0,%%mm7\n\t"
  81         "pfadd %%mm0,%%mm7\n\t"
  82         "movq %%mm7,60(%%eax)\n\t"
  83         "psrlq $32,%%mm0\n\t"
  84         "movd 68(%%eax),%%mm1\n\t"
  85         "pfadd %%mm1,%%mm0\n\t"
  86         "movd %%mm0,68(%%eax)\n\t"
  87         "movd 4(%%eax),%%mm0\n\t"
  88         "movd 12(%%eax),%%mm1\n\t"
  89         "punpckldq %%mm1,%%mm0\n\t"
  90         "punpckldq 20(%%eax),%%mm1\n\t"
  91         "pfadd %%mm1,%%mm0\n\t"
  92         "movd %%mm0,12(%%eax)\n\t"
  93         "psrlq $32,%%mm0\n\t"
  94         "movd %%mm0,20(%%eax)\n\t"
  95         "psrlq $32,%%mm1\n\t"
  96         "movd 28(%%eax),%%mm2\n\t"
  97         "punpckldq %%mm2,%%mm1\n\t"
  98         "punpckldq 36(%%eax),%%mm2\n\t"
  99         "pfadd %%mm2,%%mm1\n\t"
 100         "movd %%mm1,28(%%eax)\n\t"
 101         "psrlq $32,%%mm1\n\t"
 102         "movd %%mm1,36(%%eax)\n\t"
 103         "psrlq $32,%%mm2\n\t"
 104         "movd 44(%%eax),%%mm3\n\t"
 105         "punpckldq %%mm3,%%mm2\n\t"
 106         "punpckldq 52(%%eax),%%mm3\n\t"
 107         "pfadd %%mm3,%%mm2\n\t"
 108         "movd %%mm2,44(%%eax)\n\t"
 109         "psrlq $32,%%mm2\n\t"
 110         "movd %%mm2,52(%%eax)\n\t"
 111         "psrlq $32,%%mm3\n\t"
 112         "movd 60(%%eax),%%mm4\n\t"
 113         "punpckldq %%mm4,%%mm3\n\t"
 114         "punpckldq 68(%%eax),%%mm4\n\t"
 115         "pfadd %%mm4,%%mm3\n\t"
 116         "movd %%mm3,60(%%eax)\n\t"
 117         "psrlq $32,%%mm3\n\t"
 118         "movd %%mm3,68(%%eax)\n\t"
 119
 120         "movq 24(%%eax),%%mm0\n\t"
 121         "movq 48(%%eax),%%mm1\n\t"
 122         "movd "MANGLE(COS9)"+12,%%mm2\n\t"
 123         "punpckldq %%mm2,%%mm2\n\t"
 124         "movd "MANGLE(COS9)"+24,%%mm3\n\t"
 125         "punpckldq %%mm3,%%mm3\n\t"
 126         "pfmul %%mm2,%%mm0\n\t"
 127         "pfmul %%mm3,%%mm1\n\t"
 128         "pushl %%eax\n\t"
 129         "movl $1,%%eax\n\t"
 130         "movd %%eax,%%mm7\n\t"
 131         "pi2fd %%mm7,%%mm7\n\t"
 132         "popl %%eax\n\t"
 133         "movq 8(%%eax),%%mm2\n\t"
 134         "movd "MANGLE(COS9)"+4,%%mm3\n\t"
 135         "punpckldq %%mm3,%%mm3\n\t"
 136         "pfmul %%mm3,%%mm2\n\t"
 137         "pfadd %%mm0,%%mm2\n\t"
 138         "movq 40(%%eax),%%mm3\n\t"
 139         "movd "MANGLE(COS9)"+20,%%mm4\n\t"
 140         "punpckldq %%mm4,%%mm4\n\t"
 141         "pfmul %%mm4,%%mm3\n\t"
 142         "pfadd %%mm3,%%mm2\n\t"
 143         "movq 56(%%eax),%%mm3\n\t"
 144         "movd "MANGLE(COS9)"+28,%%mm4\n\t"
 145         "punpckldq %%mm4,%%mm4\n\t"
 146         "pfmul %%mm4,%%mm3\n\t"
 147         "pfadd %%mm3,%%mm2\n\t"
 148         "movq (%%eax),%%mm3\n\t"
 149         "movq 16(%%eax),%%mm4\n\t"
 150         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 151         "punpckldq %%mm5,%%mm5\n\t"
 152         "pfmul %%mm5,%%mm4\n\t"
 153         "pfadd %%mm4,%%mm3\n\t"
 154         "movq 32(%%eax),%%mm4\n\t"
 155         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 156         "punpckldq %%mm5,%%mm5\n\t"
 157         "pfmul %%mm5,%%mm4\n\t"
 158         "pfadd %%mm4,%%mm3\n\t"
 159         "pfadd %%mm1,%%mm3\n\t"
 160         "movq 64(%%eax),%%mm4\n\t"
 161         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 162         "punpckldq %%mm5,%%mm5\n\t"
 163         "pfmul %%mm5,%%mm4\n\t"
 164         "pfadd %%mm4,%%mm3\n\t"
 165         "movq %%mm2,%%mm4\n\t"
 166         "pfadd %%mm3,%%mm4\n\t"
 167         "movq %%mm7,%%mm5\n\t"
 168         "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t"
 169         "pfmul %%mm5,%%mm4\n\t"
 170         "movq %%mm4,%%mm5\n\t"
 171         "pfacc %%mm5,%%mm5\n\t"
 172         "movd 108(%%edx),%%mm6\n\t"
 173         "punpckldq 104(%%edx),%%mm6\n\t"
 174         "pfmul %%mm6,%%mm5\n\t"
 175 #ifdef DCT36_OPTIMIZE_FOR_K7
 176         "pswapd %%mm5,%%mm5\n\t"
 177         "movq %%mm5,32(%%ecx)\n\t"
 178 #else
 179         "movd %%mm5,36(%%ecx)\n\t"
 180         "psrlq $32,%%mm5\n\t"
 181         "movd %%mm5,32(%%ecx)\n\t"
 182 #endif
 183         "movq %%mm4,%%mm6\n\t"
 184         "punpckldq %%mm6,%%mm5\n\t"
 185         "pfsub %%mm6,%%mm5\n\t"
 186         "punpckhdq %%mm5,%%mm5\n\t"
 187         "movd 32(%%edx),%%mm6\n\t"
 188         "punpckldq 36(%%edx),%%mm6\n\t"
 189         "pfmul %%mm6,%%mm5\n\t"
 190         "movd 32(%%esi),%%mm6\n\t"
 191         "punpckldq 36(%%esi),%%mm6\n\t"
 192         "pfadd %%mm6,%%mm5\n\t"
 193         "movd %%mm5,1024(%%ebx)\n\t"
 194         "psrlq $32,%%mm5\n\t"
 195         "movd %%mm5,1152(%%ebx)\n\t"
 196         "movq %%mm3,%%mm4\n\t"
 197         "pfsub %%mm2,%%mm4\n\t"
 198         "movq %%mm7,%%mm5\n\t"
 199         "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t"
 200         "pfmul %%mm5,%%mm4\n\t"
 201         "movq %%mm4,%%mm5\n\t"
 202         "pfacc %%mm5,%%mm5\n\t"
 203         "movd 140(%%edx),%%mm6\n\t"
 204         "punpckldq 72(%%edx),%%mm6\n\t"
 205         "pfmul %%mm6,%%mm5\n\t"
 206         "movd %%mm5,68(%%ecx)\n\t"
 207         "psrlq $32,%%mm5\n\t"
 208         "movd %%mm5,0(%%ecx)\n\t"
 209         "movq %%mm4,%%mm6\n\t"
 210         "punpckldq %%mm6,%%mm5\n\t"
 211         "pfsub %%mm6,%%mm5\n\t"
 212         "punpckhdq %%mm5,%%mm5\n\t"
 213         "movd 0(%%edx),%%mm6\n\t"
 214         "punpckldq 68(%%edx),%%mm6\n\t"
 215         "pfmul %%mm6,%%mm5\n\t"
 216         "movd 0(%%esi),%%mm6\n\t"
 217         "punpckldq 68(%%esi),%%mm6\n\t"
 218         "pfadd %%mm6,%%mm5\n\t"
 219         "movd %%mm5,0(%%ebx)\n\t"
 220         "psrlq $32,%%mm5\n\t"
 221         "movd %%mm5,2176(%%ebx)\n\t"
 222         "movq 8(%%eax),%%mm2\n\t"
 223         "movq 40(%%eax),%%mm3\n\t"
 224         "pfsub %%mm3,%%mm2\n\t"
 225         "movq 56(%%eax),%%mm3\n\t"
 226         "pfsub %%mm3,%%mm2\n\t"
 227         "movd "MANGLE(COS9)"+12,%%mm3\n\t"
 228         "punpckldq %%mm3,%%mm3\n\t"
 229         "pfmul %%mm3,%%mm2\n\t"
 230         "movq 16(%%eax),%%mm3\n\t"
 231         "movq 32(%%eax),%%mm4\n\t"
 232         "pfsub %%mm4,%%mm3\n\t"
 233         "movq 64(%%eax),%%mm4\n\t"
 234         "pfsub %%mm4,%%mm3\n\t"
 235         "movd "MANGLE(COS9)"+24,%%mm4\n\t"
 236         "punpckldq %%mm4,%%mm4\n\t"
 237         "pfmul %%mm4,%%mm3\n\t"
 238         "movq 48(%%eax),%%mm4\n\t"
 239         "pfsub %%mm4,%%mm3\n\t"
 240         "movq (%%eax),%%mm4\n\t"
 241         "pfadd %%mm4,%%mm3\n\t"
 242         "movq %%mm2,%%mm4\n\t"
 243         "pfadd %%mm3,%%mm4\n\t"
 244         "movq %%mm7,%%mm5\n\t"
 245         "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t"
 246         "pfmul %%mm5,%%mm4\n\t"
 247         "movq %%mm4,%%mm5\n\t"
 248         "pfacc %%mm5,%%mm5\n\t"
 249         "movd 112(%%edx),%%mm6\n\t"
 250         "punpckldq 100(%%edx),%%mm6\n\t"
 251         "pfmul %%mm6,%%mm5\n\t"
 252         "movd %%mm5,40(%%ecx)\n\t"
 253         "psrlq $32,%%mm5\n\t"
 254         "movd %%mm5,28(%%ecx)\n\t"
 255         "movq %%mm4,%%mm6\n\t"
 256         "punpckldq %%mm6,%%mm5\n\t"
 257         "pfsub %%mm6,%%mm5\n\t"
 258         "punpckhdq %%mm5,%%mm5\n\t"
 259         "movd 28(%%edx),%%mm6\n\t"
 260         "punpckldq 40(%%edx),%%mm6\n\t"
 261         "pfmul %%mm6,%%mm5\n\t"
 262         "movd 28(%%esi),%%mm6\n\t"
 263         "punpckldq 40(%%esi),%%mm6\n\t"
 264         "pfadd %%mm6,%%mm5\n\t"
 265         "movd %%mm5,896(%%ebx)\n\t"
 266         "psrlq $32,%%mm5\n\t"
 267         "movd %%mm5,1280(%%ebx)\n\t"
 268         "movq %%mm3,%%mm4\n\t"
 269         "pfsub %%mm2,%%mm4\n\t"
 270         "movq %%mm7,%%mm5\n\t"
 271         "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t"
 272         "pfmul %%mm5,%%mm4\n\t"
 273         "movq %%mm4,%%mm5\n\t"
 274         "pfacc %%mm5,%%mm5\n\t"
 275         "movd 136(%%edx),%%mm6\n\t"
 276         "punpckldq 76(%%edx),%%mm6\n\t"
 277         "pfmul %%mm6,%%mm5\n\t"
 278         "movd %%mm5,64(%%ecx)\n\t"
 279         "psrlq $32,%%mm5\n\t"
 280         "movd %%mm5,4(%%ecx)\n\t"
 281         "movq %%mm4,%%mm6\n\t"
 282         "punpckldq %%mm6,%%mm5\n\t"
 283         "pfsub %%mm6,%%mm5\n\t"
 284         "punpckhdq %%mm5,%%mm5\n\t"
 285         "movd 4(%%edx),%%mm6\n\t"
 286         "punpckldq 64(%%edx),%%mm6\n\t"
 287         "pfmul %%mm6,%%mm5\n\t"
 288         "movd 4(%%esi),%%mm6\n\t"
 289         "punpckldq 64(%%esi),%%mm6\n\t"
 290         "pfadd %%mm6,%%mm5\n\t"
 291         "movd %%mm5,128(%%ebx)\n\t"
 292         "psrlq $32,%%mm5\n\t"
 293         "movd %%mm5,2048(%%ebx)\n\t"
 294
 295         "movq 8(%%eax),%%mm2\n\t"
 296         "movd "MANGLE(COS9)"+20,%%mm3\n\t"
 297         "punpckldq %%mm3,%%mm3\n\t"
 298         "pfmul %%mm3,%%mm2\n\t"
 299         "pfsub %%mm0,%%mm2\n\t"
 300         "movq 40(%%eax),%%mm3\n\t"
 301         "movd "MANGLE(COS9)"+28,%%mm4\n\t"
 302         "punpckldq %%mm4,%%mm4\n\t"
 303         "pfmul %%mm4,%%mm3\n\t"
 304         "pfsub %%mm3,%%mm2\n\t"
 305         "movq 56(%%eax),%%mm3\n\t"
 306         "movd "MANGLE(COS9)"+4,%%mm4\n\t"
 307         "punpckldq %%mm4,%%mm4\n\t"
 308         "pfmul %%mm4,%%mm3\n\t"
 309         "pfadd %%mm3,%%mm2\n\t"
 310         "movq (%%eax),%%mm3\n\t"
 311         "movq 16(%%eax),%%mm4\n\t"
 312         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 313         "punpckldq %%mm5,%%mm5\n\t"
 314         "pfmul %%mm5,%%mm4\n\t"
 315         "pfsub %%mm4,%%mm3\n\t"
 316         "movq 32(%%eax),%%mm4\n\t"
 317         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 318         "punpckldq %%mm5,%%mm5\n\t"
 319         "pfmul %%mm5,%%mm4\n\t"
 320         "pfsub %%mm4,%%mm3\n\t"
 321         "pfadd %%mm1,%%mm3\n\t"
 322         "movq 64(%%eax),%%mm4\n\t"
 323         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 324         "punpckldq %%mm5,%%mm5\n\t"
 325         "pfmul %%mm5,%%mm4\n\t"
 326         "pfadd %%mm4,%%mm3\n\t"
 327         "movq %%mm2,%%mm4\n\t"
 328         "pfadd %%mm3,%%mm4\n\t"
 329         "movq %%mm7,%%mm5\n\t"
 330         "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t"
 331         "pfmul %%mm5,%%mm4\n\t"
 332         "movq %%mm4,%%mm5\n\t"
 333         "pfacc %%mm5,%%mm5\n\t"
 334         "movd 116(%%edx),%%mm6\n\t"
 335         "punpckldq 96(%%edx),%%mm6\n\t"
 336         "pfmul %%mm6,%%mm5\n\t"
 337         "movd %%mm5,44(%%ecx)\n\t"
 338         "psrlq $32,%%mm5\n\t"
 339         "movd %%mm5,24(%%ecx)\n\t"
 340         "movq %%mm4,%%mm6\n\t"
 341         "punpckldq %%mm6,%%mm5\n\t"
 342         "pfsub %%mm6,%%mm5\n\t"
 343         "punpckhdq %%mm5,%%mm5\n\t"
 344         "movd 24(%%edx),%%mm6\n\t"
 345         "punpckldq 44(%%edx),%%mm6\n\t"
 346         "pfmul %%mm6,%%mm5\n\t"
 347         "movd 24(%%esi),%%mm6\n\t"
 348         "punpckldq 44(%%esi),%%mm6\n\t"
 349         "pfadd %%mm6,%%mm5\n\t"
 350         "movd %%mm5,768(%%ebx)\n\t"
 351         "psrlq $32,%%mm5\n\t"
 352         "movd %%mm5,1408(%%ebx)\n\t"
 353         "movq %%mm3,%%mm4\n\t"
 354         "pfsub %%mm2,%%mm4\n\t"
 355         "movq %%mm7,%%mm5\n\t"
 356         "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t"
 357         "pfmul %%mm5,%%mm4\n\t"
 358         "movq %%mm4,%%mm5\n\t"
 359         "pfacc %%mm5,%%mm5\n\t"
 360         "movd 132(%%edx),%%mm6\n\t"
 361         "punpckldq 80(%%edx),%%mm6\n\t"
 362         "pfmul %%mm6,%%mm5\n\t"
 363         "movd %%mm5,60(%%ecx)\n\t"
 364         "psrlq $32,%%mm5\n\t"
 365         "movd %%mm5,8(%%ecx)\n\t"
 366         "movq %%mm4,%%mm6\n\t"
 367         "punpckldq %%mm6,%%mm5\n\t"
 368         "pfsub %%mm6,%%mm5\n\t"
 369         "punpckhdq %%mm5,%%mm5\n\t"
 370         "movd 8(%%edx),%%mm6\n\t"
 371         "punpckldq 60(%%edx),%%mm6\n\t"
 372         "pfmul %%mm6,%%mm5\n\t"
 373         "movd 8(%%esi),%%mm6\n\t"
 374         "punpckldq 60(%%esi),%%mm6\n\t"
 375         "pfadd %%mm6,%%mm5\n\t"
 376         "movd %%mm5,256(%%ebx)\n\t"
 377         "psrlq $32,%%mm5\n\t"
 378         "movd %%mm5,1920(%%ebx)\n\t"
 379         "movq 8(%%eax),%%mm2\n\t"
 380         "movd "MANGLE(COS9)"+28,%%mm3\n\t"
 381         "punpckldq %%mm3,%%mm3\n\t"
 382         "pfmul %%mm3,%%mm2\n\t"
 383         "pfsub %%mm0,%%mm2\n\t"
 384         "movq 40(%%eax),%%mm3\n\t"
 385         "movd "MANGLE(COS9)"+4,%%mm4\n\t"
 386         "punpckldq %%mm4,%%mm4\n\t"
 387         "pfmul %%mm4,%%mm3\n\t"
 388         "pfadd %%mm3,%%mm2\n\t"
 389         "movq 56(%%eax),%%mm3\n\t"
 390         "movd "MANGLE(COS9)"+20,%%mm4\n\t"
 391         "punpckldq %%mm4,%%mm4\n\t"
 392         "pfmul %%mm4,%%mm3\n\t"
 393         "pfsub %%mm3,%%mm2\n\t"
 394         "movq (%%eax),%%mm3\n\t"
 395         "movq 16(%%eax),%%mm4\n\t"
 396         "movd "MANGLE(COS9)"+16,%%mm5\n\t"
 397         "punpckldq %%mm5,%%mm5\n\t"
 398         "pfmul %%mm5,%%mm4\n\t"
 399         "pfsub %%mm4,%%mm3\n\t"
 400         "movq 32(%%eax),%%mm4\n\t"
 401         "movd "MANGLE(COS9)"+32,%%mm5\n\t"
 402         "punpckldq %%mm5,%%mm5\n\t"
 403         "pfmul %%mm5,%%mm4\n\t"
 404         "pfadd %%mm4,%%mm3\n\t"
 405         "pfadd %%mm1,%%mm3\n\t"
 406         "movq 64(%%eax),%%mm4\n\t"
 407         "movd "MANGLE(COS9)"+8,%%mm5\n\t"
 408         "punpckldq %%mm5,%%mm5\n\t"
 409         "pfmul %%mm5,%%mm4\n\t"
 410         "pfsub %%mm4,%%mm3\n\t"
 411         "movq %%mm2,%%mm4\n\t"
 412         "pfadd %%mm3,%%mm4\n\t"
 413         "movq %%mm7,%%mm5\n\t"
 414         "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t"
 415         "pfmul %%mm5,%%mm4\n\t"
 416         "movq %%mm4,%%mm5\n\t"
 417         "pfacc %%mm5,%%mm5\n\t"
 418         "movd 120(%%edx),%%mm6\n\t"
 419         "punpckldq 92(%%edx),%%mm6\n\t"
 420         "pfmul %%mm6,%%mm5\n\t"
 421         "movd %%mm5,48(%%ecx)\n\t"
 422         "psrlq $32,%%mm5\n\t"
 423         "movd %%mm5,20(%%ecx)\n\t"
 424         "movq %%mm4,%%mm6\n\t"
 425         "punpckldq %%mm6,%%mm5\n\t"
 426         "pfsub %%mm6,%%mm5\n\t"
 427         "punpckhdq %%mm5,%%mm5\n\t"
 428         "movd 20(%%edx),%%mm6\n\t"
 429         "punpckldq 48(%%edx),%%mm6\n\t"
 430         "pfmul %%mm6,%%mm5\n\t"
 431         "movd 20(%%esi),%%mm6\n\t"
 432         "punpckldq 48(%%esi),%%mm6\n\t"
 433         "pfadd %%mm6,%%mm5\n\t"
 434         "movd %%mm5,640(%%ebx)\n\t"
 435         "psrlq $32,%%mm5\n\t"
 436         "movd %%mm5,1536(%%ebx)\n\t"
 437         "movq %%mm3,%%mm4\n\t"
 438         "pfsub %%mm2,%%mm4\n\t"
 439         "movq %%mm7,%%mm5\n\t"
 440         "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t"
 441         "pfmul %%mm5,%%mm4\n\t"
 442         "movq %%mm4,%%mm5\n\t"
 443         "pfacc %%mm5,%%mm5\n\t"
 444         "movd 128(%%edx),%%mm6\n\t"
 445         "punpckldq 84(%%edx),%%mm6\n\t"
 446         "pfmul %%mm6,%%mm5\n\t"
 447         "movd %%mm5,56(%%ecx)\n\t"
 448         "psrlq $32,%%mm5\n\t"
 449         "movd %%mm5,12(%%ecx)\n\t"
 450         "movq %%mm4,%%mm6\n\t"
 451         "punpckldq %%mm6,%%mm5\n\t"
 452         "pfsub %%mm6,%%mm5\n\t"
 453         "punpckhdq %%mm5,%%mm5\n\t"
 454         "movd 12(%%edx),%%mm6\n\t"
 455         "punpckldq 56(%%edx),%%mm6\n\t"
 456         "pfmul %%mm6,%%mm5\n\t"
 457         "movd 12(%%esi),%%mm6\n\t"
 458         "punpckldq 56(%%esi),%%mm6\n\t"
 459         "pfadd %%mm6,%%mm5\n\t"
 460         "movd %%mm5,384(%%ebx)\n\t"
 461         "psrlq $32,%%mm5\n\t"
 462         "movd %%mm5,1792(%%ebx)\n\t"
 463
 464         "movq (%%eax),%%mm4\n\t"
 465         "movq 16(%%eax),%%mm3\n\t"
 466         "pfsub %%mm3,%%mm4\n\t"
 467         "movq 32(%%eax),%%mm3\n\t"
 468         "pfadd %%mm3,%%mm4\n\t"
 469         "movq 48(%%eax),%%mm3\n\t"
 470         "pfsub %%mm3,%%mm4\n\t"
 471         "movq 64(%%eax),%%mm3\n\t"
 472         "pfadd %%mm3,%%mm4\n\t"
 473         "movq %%mm7,%%mm5\n\t"
 474         "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t"
 475         "pfmul %%mm5,%%mm4\n\t"
 476         "movq %%mm4,%%mm5\n\t"
 477         "pfacc %%mm5,%%mm5\n\t"
 478         "movd 124(%%edx),%%mm6\n\t"
 479         "punpckldq 88(%%edx),%%mm6\n\t"
 480         "pfmul %%mm6,%%mm5\n\t"
 481         "movd %%mm5,52(%%ecx)\n\t"
 482         "psrlq $32,%%mm5\n\t"
 483         "movd %%mm5,16(%%ecx)\n\t"
 484         "movq %%mm4,%%mm6\n\t"
 485         "punpckldq %%mm6,%%mm5\n\t"
 486         "pfsub %%mm6,%%mm5\n\t"
 487         "punpckhdq %%mm5,%%mm5\n\t"
 488         "movd 16(%%edx),%%mm6\n\t"
 489         "punpckldq 52(%%edx),%%mm6\n\t"
 490         "pfmul %%mm6,%%mm5\n\t"
 491         "movd 16(%%esi),%%mm6\n\t"
 492         "punpckldq 52(%%esi),%%mm6\n\t"
 493         "pfadd %%mm6,%%mm5\n\t"
 494         "movd %%mm5,512(%%ebx)\n\t"
 495         "psrlq $32,%%mm5\n\t"
 496         "movd %%mm5,1664(%%ebx)\n\t"
 497
 498         "femms\n\t"
 499         :
 500         : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf)
 501         : "memory");
 502 }