comparison mp3lib/dct64_k7.c @ 4148:3b29772a4fb2

S->C
author nick
date Mon, 14 Jan 2002 10:34:38 +0000
parents
children 421969d55d5f
comparison
equal deleted inserted replaced
4147:4bbdda22003d 4148:3b29772a4fb2
1 /*
2 * This code was taken from http://www.mpg123.org
3 * See ChangeLog of mpg123-0.59s-pre.1 for detail
4 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
5 * Partial 3dnowex-DSP! optimization by Nick Kurshev
6 *
7 * TODO: optimize scalar 3dnow! code
8 * Warning: Phases 7 & 8 are not tested
9 */
10 #define real float /* ugly - but only way */
11
12 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
13 static float plus_1f = 1.0;
14
15 void __attribute__ (( __stdcall__ )) dct64_MMX_3dnowex(real *a,real *b,real *c)
16 {
17 char tmp[256];
18 __asm __volatile(
19 " movl %2,%%eax\n\t"
20
21 " leal 128+%3,%%edx\n\t"
22 " movl %0,%%esi\n\t"
23 " movl %1,%%edi\n\t"
24 " movl $costab_mmx,%%ebx\n\t"
25 " leal %3,%%ecx\n\t"
26
27 /* Phase 1*/
28 " movq (%%eax), %%mm0\n\t"
29 " movq 8(%%eax), %%mm4\n\t"
30 " movq %%mm0, %%mm3\n\t"
31 " movq %%mm4, %%mm7\n\t"
32 " movq 120(%%eax), %%mm1\n\t"
33 " movq 112(%%eax), %%mm5\n\t"
34 " pswapd %%mm1, %%mm1\n\t"
35 " pswapd %%mm5, %%mm5\n\t"
36 " pfadd %%mm1, %%mm0\n\t"
37 " pfadd %%mm5, %%mm4\n\t"
38 " movq %%mm0, (%%edx)\n\t"
39 " movq %%mm4, 8(%%edx)\n\t"
40 " pfsub %%mm1, %%mm3\n\t"
41 " pfsub %%mm5, %%mm7\n\t"
42 " pfmul (%%ebx), %%mm3\n\t"
43 " pfmul 8(%%ebx), %%mm7\n\t"
44 " pswapd %%mm3, %%mm3\n\t"
45 " pswapd %%mm7, %%mm7\n\t"
46 " movq %%mm3, 120(%%edx)\n\t"
47 " movq %%mm7, 112(%%edx)\n\t"
48
49 " movq 16(%%eax), %%mm0\n\t"
50 " movq 24(%%eax), %%mm4\n\t"
51 " movq %%mm0, %%mm3\n\t"
52 " movq %%mm4, %%mm7\n\t"
53 " movq 104(%%eax), %%mm1\n\t"
54 " movq 96(%%eax), %%mm5\n\t"
55 " pswapd %%mm1, %%mm1\n\t"
56 " pswapd %%mm5, %%mm5\n\t"
57 " pfadd %%mm1, %%mm0\n\t"
58 " pfadd %%mm5, %%mm4\n\t"
59 " movq %%mm0, 16(%%edx)\n\t"
60 " movq %%mm4, 24(%%edx)\n\t"
61 " pfsub %%mm1, %%mm3\n\t"
62 " pfsub %%mm5, %%mm7\n\t"
63 " pfmul 16(%%ebx), %%mm3\n\t"
64 " pfmul 24(%%ebx), %%mm7\n\t"
65 " pswapd %%mm3, %%mm3\n\t"
66 " pswapd %%mm7, %%mm7\n\t"
67 " movq %%mm3, 104(%%edx)\n\t"
68 " movq %%mm7, 96(%%edx)\n\t"
69
70 " movq 32(%%eax), %%mm0\n\t"
71 " movq 40(%%eax), %%mm4\n\t"
72 " movq %%mm0, %%mm3\n\t"
73 " movq %%mm4, %%mm7\n\t"
74 " movq 88(%%eax), %%mm1\n\t"
75 " movq 80(%%eax), %%mm5\n\t"
76 " pswapd %%mm1, %%mm1\n\t"
77 " pswapd %%mm5, %%mm5\n\t"
78 " pfadd %%mm1, %%mm0\n\t"
79 " pfadd %%mm5, %%mm4\n\t"
80 " movq %%mm0, 32(%%edx)\n\t"
81 " movq %%mm4, 40(%%edx)\n\t"
82 " pfsub %%mm1, %%mm3\n\t"
83 " pfsub %%mm5, %%mm7\n\t"
84 " pfmul 32(%%ebx), %%mm3\n\t"
85 " pfmul 40(%%ebx), %%mm7\n\t"
86 " pswapd %%mm3, %%mm3\n\t"
87 " pswapd %%mm7, %%mm7\n\t"
88 " movq %%mm3, 88(%%edx)\n\t"
89 " movq %%mm7, 80(%%edx)\n\t"
90
91 " movq 48(%%eax), %%mm0\n\t"
92 " movq 56(%%eax), %%mm4\n\t"
93 " movq %%mm0, %%mm3\n\t"
94 " movq %%mm4, %%mm7\n\t"
95 " movq 72(%%eax), %%mm1\n\t"
96 " movq 64(%%eax), %%mm5\n\t"
97 " pswapd %%mm1, %%mm1\n\t"
98 " pswapd %%mm5, %%mm5\n\t"
99 " pfadd %%mm1, %%mm0\n\t"
100 " pfadd %%mm5, %%mm4\n\t"
101 " movq %%mm0, 48(%%edx)\n\t"
102 " movq %%mm4, 56(%%edx)\n\t"
103 " pfsub %%mm1, %%mm3\n\t"
104 " pfsub %%mm5, %%mm7\n\t"
105 " pfmul 48(%%ebx), %%mm3\n\t"
106 " pfmul 56(%%ebx), %%mm7\n\t"
107 " pswapd %%mm3, %%mm3\n\t"
108 " pswapd %%mm7, %%mm7\n\t"
109 " movq %%mm3, 72(%%edx)\n\t"
110 " movq %%mm7, 64(%%edx)\n\t"
111
112 /* Phase 2*/
113
114 " movq (%%edx), %%mm0\n\t"
115 " movq 8(%%edx), %%mm4\n\t"
116 " movq %%mm0, %%mm3\n\t"
117 " movq %%mm4, %%mm7\n\t"
118 " movq 56(%%edx), %%mm1\n\t"
119 " movq 48(%%edx), %%mm5\n\t"
120 " pswapd %%mm1, %%mm1\n\t"
121 " pswapd %%mm5, %%mm5\n\t"
122 " pfadd %%mm1, %%mm0\n\t"
123 " pfadd %%mm5, %%mm4\n\t"
124 " movq %%mm0, (%%ecx)\n\t"
125 " movq %%mm4, 8(%%ecx)\n\t"
126 " pfsub %%mm1, %%mm3\n\t"
127 " pfsub %%mm5, %%mm7\n\t"
128 " pfmul 64(%%ebx), %%mm3\n\t"
129 " pfmul 72(%%ebx), %%mm7\n\t"
130 " pswapd %%mm3, %%mm3\n\t"
131 " pswapd %%mm7, %%mm7\n\t"
132 " movq %%mm3, 56(%%ecx)\n\t"
133 " movq %%mm7, 48(%%ecx)\n\t"
134
135 " movq 16(%%edx), %%mm0\n\t"
136 " movq 24(%%edx), %%mm4\n\t"
137 " movq %%mm0, %%mm3\n\t"
138 " movq %%mm4, %%mm7\n\t"
139 " movq 40(%%edx), %%mm1\n\t"
140 " movq 32(%%edx), %%mm5\n\t"
141 " pswapd %%mm1, %%mm1\n\t"
142 " pswapd %%mm5, %%mm5\n\t"
143 " pfadd %%mm1, %%mm0\n\t"
144 " pfadd %%mm5, %%mm4\n\t"
145 " movq %%mm0, 16(%%ecx)\n\t"
146 " movq %%mm4, 24(%%ecx)\n\t"
147 " pfsub %%mm1, %%mm3\n\t"
148 " pfsub %%mm5, %%mm7\n\t"
149 " pfmul 80(%%ebx), %%mm3\n\t"
150 " pfmul 88(%%ebx), %%mm7\n\t"
151 " pswapd %%mm3, %%mm3\n\t"
152 " pswapd %%mm7, %%mm7\n\t"
153 " movq %%mm3, 40(%%ecx)\n\t"
154 " movq %%mm7, 32(%%ecx)\n\t"
155
156 /* Phase 3*/
157
158 " movq 64(%%edx), %%mm0\n\t"
159 " movq 72(%%edx), %%mm4\n\t"
160 " movq %%mm0, %%mm3\n\t"
161 " movq %%mm4, %%mm7\n\t"
162 " movq 120(%%edx), %%mm1\n\t"
163 " movq 112(%%edx), %%mm5\n\t"
164 " pswapd %%mm1, %%mm1\n\t"
165 " pswapd %%mm5, %%mm5\n\t"
166 " pfadd %%mm1, %%mm0\n\t"
167 " pfadd %%mm5, %%mm4\n\t"
168 " movq %%mm0, 64(%%ecx)\n\t"
169 " movq %%mm4, 72(%%ecx)\n\t"
170 " pfsubr %%mm1, %%mm3\n\t"
171 " pfsubr %%mm5, %%mm7\n\t"
172 " pfmul 64(%%ebx), %%mm3\n\t"
173 " pfmul 72(%%ebx), %%mm7\n\t"
174 " pswapd %%mm3, %%mm3\n\t"
175 " pswapd %%mm7, %%mm7\n\t"
176 " movq %%mm3, 120(%%ecx)\n\t"
177 " movq %%mm7, 112(%%ecx)\n\t"
178
179 " movq 80(%%edx), %%mm0\n\t"
180 " movq 88(%%edx), %%mm4\n\t"
181 " movq %%mm0, %%mm3\n\t"
182 " movq %%mm4, %%mm7\n\t"
183 " movq 104(%%edx), %%mm1\n\t"
184 " movq 96(%%edx), %%mm5\n\t"
185 " pswapd %%mm1, %%mm1\n\t"
186 " pswapd %%mm5, %%mm5\n\t"
187 " pfadd %%mm1, %%mm0\n\t"
188 " pfadd %%mm5, %%mm4\n\t"
189 " movq %%mm0, 80(%%ecx)\n\t"
190 " movq %%mm4, 88(%%ecx)\n\t"
191 " pfsubr %%mm1, %%mm3\n\t"
192 " pfsubr %%mm5, %%mm7\n\t"
193 " pfmul 80(%%ebx), %%mm3\n\t"
194 " pfmul 88(%%ebx), %%mm7\n\t"
195 " pswapd %%mm3, %%mm3\n\t"
196 " pswapd %%mm7, %%mm7\n\t"
197 " movq %%mm3, 104(%%ecx)\n\t"
198 " movq %%mm7, 96(%%ecx)\n\t"
199
200 /* Phase 4*/
201
202 " movq 96(%%ebx), %%mm2\n\t"
203 " movq 104(%%ebx), %%mm6\n\t"
204
205 " movq (%%ecx), %%mm0\n\t"
206 " movq 8(%%ecx), %%mm4\n\t"
207 " movq %%mm0, %%mm3\n\t"
208 " movq %%mm4, %%mm7\n\t"
209 " movq 24(%%ecx), %%mm1\n\t"
210 " movq 16(%%ecx), %%mm5\n\t"
211 " pswapd %%mm1, %%mm1\n\t"
212 " pswapd %%mm5, %%mm5\n\t"
213 " pfadd %%mm1, %%mm0\n\t"
214 " pfadd %%mm5, %%mm4\n\t"
215 " movq %%mm0, (%%edx)\n\t"
216 " movq %%mm4, 8(%%edx)\n\t"
217 " pfsub %%mm1, %%mm3\n\t"
218 " pfsub %%mm5, %%mm7\n\t"
219 " pfmul %%mm2, %%mm3\n\t"
220 " pfmul %%mm6, %%mm7\n\t"
221 " pswapd %%mm3, %%mm3\n\t"
222 " pswapd %%mm7, %%mm7\n\t"
223 " movq %%mm3, 24(%%edx)\n\t"
224 " movq %%mm7, 16(%%edx)\n\t"
225
226 " movq 32(%%ecx), %%mm0\n\t"
227 " movq 40(%%ecx), %%mm4\n\t"
228 " movq %%mm0, %%mm3\n\t"
229 " movq %%mm4, %%mm7\n\t"
230 " movq 56(%%ecx), %%mm1\n\t"
231 " movq 48(%%ecx), %%mm5\n\t"
232 " pswapd %%mm1, %%mm1\n\t"
233 " pswapd %%mm5, %%mm5\n\t"
234 " pfadd %%mm1, %%mm0\n\t"
235 " pfadd %%mm5, %%mm4\n\t"
236 " movq %%mm0, 32(%%edx)\n\t"
237 " movq %%mm4, 40(%%edx)\n\t"
238 " pfsubr %%mm1, %%mm3\n\t"
239 " pfsubr %%mm5, %%mm7\n\t"
240 " pfmul %%mm2, %%mm3\n\t"
241 " pfmul %%mm6, %%mm7\n\t"
242 " pswapd %%mm3, %%mm3\n\t"
243 " pswapd %%mm7, %%mm7\n\t"
244 " movq %%mm3, 56(%%edx)\n\t"
245 " movq %%mm7, 48(%%edx)\n\t"
246
247 " movq 64(%%ecx), %%mm0\n\t"
248 " movq 72(%%ecx), %%mm4\n\t"
249 " movq %%mm0, %%mm3\n\t"
250 " movq %%mm4, %%mm7\n\t"
251 " movq 88(%%ecx), %%mm1\n\t"
252 " movq 80(%%ecx), %%mm5\n\t"
253 " pswapd %%mm1, %%mm1\n\t"
254 " pswapd %%mm5, %%mm5\n\t"
255 " pfadd %%mm1, %%mm0\n\t"
256 " pfadd %%mm5, %%mm4\n\t"
257 " movq %%mm0, 64(%%edx)\n\t"
258 " movq %%mm4, 72(%%edx)\n\t"
259 " pfsub %%mm1, %%mm3\n\t"
260 " pfsub %%mm5, %%mm7\n\t"
261 " pfmul %%mm2, %%mm3\n\t"
262 " pfmul %%mm6, %%mm7\n\t"
263 " pswapd %%mm3, %%mm3\n\t"
264 " pswapd %%mm7, %%mm7\n\t"
265 " movq %%mm3, 88(%%edx)\n\t"
266 " movq %%mm7, 80(%%edx)\n\t"
267
268 " movq 96(%%ecx), %%mm0\n\t"
269 " movq 104(%%ecx), %%mm4\n\t"
270 " movq %%mm0, %%mm3\n\t"
271 " movq %%mm4, %%mm7\n\t"
272 " movq 120(%%ecx), %%mm1\n\t"
273 " movq 112(%%ecx), %%mm5\n\t"
274 " pswapd %%mm1, %%mm1\n\t"
275 " pswapd %%mm5, %%mm5\n\t"
276 " pfadd %%mm1, %%mm0\n\t"
277 " pfadd %%mm5, %%mm4\n\t"
278 " movq %%mm0, 96(%%edx)\n\t"
279 " movq %%mm4, 104(%%edx)\n\t"
280 " pfsubr %%mm1, %%mm3\n\t"
281 " pfsubr %%mm5, %%mm7\n\t"
282 " pfmul %%mm2, %%mm3\n\t"
283 " pfmul %%mm6, %%mm7\n\t"
284 " pswapd %%mm3, %%mm3\n\t"
285 " pswapd %%mm7, %%mm7\n\t"
286 " movq %%mm3, 120(%%edx)\n\t"
287 " movq %%mm7, 112(%%edx)\n\t"
288
289 /* Phase 5 */
290
291 " movq 112(%%ebx), %%mm2\n\t"
292
293 " movq (%%edx), %%mm0\n\t"
294 " movq 16(%%edx), %%mm4\n\t"
295 " movq %%mm0, %%mm3\n\t"
296 " movq %%mm4, %%mm7\n\t"
297 " movq 8(%%edx), %%mm1\n\t"
298 " movq 24(%%edx), %%mm5\n\t"
299 " pswapd %%mm1, %%mm1\n\t"
300 " pswapd %%mm5, %%mm5\n\t"
301 " pfadd %%mm1, %%mm0\n\t"
302 " pfadd %%mm5, %%mm4\n\t"
303 " movq %%mm0, (%%ecx)\n\t"
304 " movq %%mm4, 16(%%ecx)\n\t"
305 " pfsub %%mm1, %%mm3\n\t"
306 " pfsubr %%mm5, %%mm7\n\t"
307 " pfmul %%mm2, %%mm3\n\t"
308 " pfmul %%mm2, %%mm7\n\t"
309 " pswapd %%mm3, %%mm3\n\t"
310 " pswapd %%mm7, %%mm7\n\t"
311 " movq %%mm3, 8(%%ecx)\n\t"
312 " movq %%mm7, 24(%%ecx)\n\t"
313
314 " movq 32(%%edx), %%mm0\n\t"
315 " movq 48(%%edx), %%mm4\n\t"
316 " movq %%mm0, %%mm3\n\t"
317 " movq %%mm4, %%mm7\n\t"
318 " movq 40(%%edx), %%mm1\n\t"
319 " movq 56(%%edx), %%mm5\n\t"
320 " pswapd %%mm1, %%mm1\n\t"
321 " pswapd %%mm5, %%mm5\n\t"
322 " pfadd %%mm1, %%mm0\n\t"
323 " pfadd %%mm5, %%mm4\n\t"
324 " movq %%mm0, 32(%%ecx)\n\t"
325 " movq %%mm4, 48(%%ecx)\n\t"
326 " pfsub %%mm1, %%mm3\n\t"
327 " pfsubr %%mm5, %%mm7\n\t"
328 " pfmul %%mm2, %%mm3\n\t"
329 " pfmul %%mm2, %%mm7\n\t"
330 " pswapd %%mm3, %%mm3\n\t"
331 " pswapd %%mm7, %%mm7\n\t"
332 " movq %%mm3, 40(%%ecx)\n\t"
333 " movq %%mm7, 56(%%ecx)\n\t"
334
335 " movq 64(%%edx), %%mm0\n\t"
336 " movq 80(%%edx), %%mm4\n\t"
337 " movq %%mm0, %%mm3\n\t"
338 " movq %%mm4, %%mm7\n\t"
339 " movq 72(%%edx), %%mm1\n\t"
340 " movq 88(%%edx), %%mm5\n\t"
341 " pswapd %%mm1, %%mm1\n\t"
342 " pswapd %%mm5, %%mm5\n\t"
343 " pfadd %%mm1, %%mm0\n\t"
344 " pfadd %%mm5, %%mm4\n\t"
345 " movq %%mm0, 64(%%ecx)\n\t"
346 " movq %%mm4, 80(%%ecx)\n\t"
347 " pfsub %%mm1, %%mm3\n\t"
348 " pfsubr %%mm5, %%mm7\n\t"
349 " pfmul %%mm2, %%mm3\n\t"
350 " pfmul %%mm2, %%mm7\n\t"
351 " pswapd %%mm3, %%mm3\n\t"
352 " pswapd %%mm7, %%mm7\n\t"
353 " movq %%mm3, 72(%%ecx)\n\t"
354 " movq %%mm7, 88(%%ecx)\n\t"
355
356 " movq 96(%%edx), %%mm0\n\t"
357 " movq 112(%%edx), %%mm4\n\t"
358 " movq %%mm0, %%mm3\n\t"
359 " movq %%mm4, %%mm7\n\t"
360 " movq 104(%%edx), %%mm1\n\t"
361 " movq 120(%%edx), %%mm5\n\t"
362 " pswapd %%mm1, %%mm1\n\t"
363 " pswapd %%mm5, %%mm5\n\t"
364 " pfadd %%mm1, %%mm0\n\t"
365 " pfadd %%mm5, %%mm4\n\t"
366 " movq %%mm0, 96(%%ecx)\n\t"
367 " movq %%mm4, 112(%%ecx)\n\t"
368 " pfsub %%mm1, %%mm3\n\t"
369 " pfsubr %%mm5, %%mm7\n\t"
370 " pfmul %%mm2, %%mm3\n\t"
371 " pfmul %%mm2, %%mm7\n\t"
372 " pswapd %%mm3, %%mm3\n\t"
373 " pswapd %%mm7, %%mm7\n\t"
374 " movq %%mm3, 104(%%ecx)\n\t"
375 " movq %%mm7, 120(%%ecx)\n\t"
376
377
378 /* Phase 6. This is the end of easy road. */
379 /* Code below is coded in scalar mode. Should be optimized */
380
381 " movd plus_1f, %%mm6\n\t"
382 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/
383 " movq x_plus_minus_3dnow, %%mm7\n\t" /* mm7 = +1 | -1 */
384
385 " movq 32(%%ecx), %%mm0\n\t"
386 " movq 64(%%ecx), %%mm2\n\t"
387 " movq %%mm0, %%mm1\n\t"
388 " movq %%mm2, %%mm3\n\t"
389 " pxor %%mm7, %%mm1\n\t"
390 " pxor %%mm7, %%mm3\n\t"
391 " pfacc %%mm1, %%mm0\n\t"
392 " pfacc %%mm3, %%mm2\n\t"
393 " pfmul %%mm6, %%mm0\n\t"
394 " pfmul %%mm6, %%mm2\n\t"
395 " movq %%mm0, 32(%%edx)\n\t"
396 " movq %%mm2, 64(%%edx)\n\t"
397
398 " movd 44(%%ecx), %%mm0\n\t"
399 " movd 40(%%ecx), %%mm2\n\t"
400 " movd 120(%%ebx), %%mm3\n\t"
401 " punpckldq 76(%%ecx), %%mm0\n\t"
402 " punpckldq 72(%%ecx), %%mm2\n\t"
403 " punpckldq %%mm3, %%mm3\n\t"
404 " movq %%mm0, %%mm4\n\t"
405 " movq %%mm2, %%mm5\n\t"
406 " pfsub %%mm2, %%mm0\n\t"
407 " pfmul %%mm3, %%mm0\n\t"
408 " movq %%mm0, %%mm1\n\t"
409 " pfadd %%mm5, %%mm0\n\t"
410 " pfadd %%mm4, %%mm0\n\t"
411 " movq %%mm0, %%mm2\n\t"
412 " punpckldq %%mm1, %%mm0\n\t"
413 " punpckhdq %%mm1, %%mm2\n\t"
414 " movq %%mm0, 40(%%edx)\n\t"
415 " movq %%mm2, 72(%%edx)\n\t"
416
417 " movd 48(%%ecx), %%mm3\n\t"
418 " movd 60(%%ecx), %%mm2\n\t"
419 " pfsub 52(%%ecx), %%mm3\n\t"
420 " pfsub 56(%%ecx), %%mm2\n\t"
421 " pfmul 120(%%ebx), %%mm3\n\t"
422 " pfmul 120(%%ebx), %%mm2\n\t"
423 " movq %%mm2, %%mm1\n\t"
424
425 " pfadd 56(%%ecx), %%mm1\n\t"
426 " pfadd 60(%%ecx), %%mm1\n\t"
427 " movq %%mm1, %%mm0\n\t"
428
429 " pfadd 48(%%ecx), %%mm0\n\t"
430 " pfadd 52(%%ecx), %%mm0\n\t"
431 " pfadd %%mm3, %%mm1\n\t"
432 " punpckldq %%mm2, %%mm1\n\t"
433 " pfadd %%mm3, %%mm2\n\t"
434 " punpckldq %%mm2, %%mm0\n\t"
435 " movq %%mm1, 56(%%edx)\n\t"
436 " movq %%mm0, 48(%%edx)\n\t"
437
438 /*---*/
439
440 " movd 92(%%ecx), %%mm1\n\t"
441 " pfsub 88(%%ecx), %%mm1\n\t"
442 " pfmul 120(%%ebx), %%mm1\n\t"
443 " movd %%mm1, 92(%%edx)\n\t"
444 " pfadd 92(%%ecx), %%mm1\n\t"
445 " pfadd 88(%%ecx), %%mm1\n\t"
446 " movq %%mm1, %%mm0\n\t"
447
448 " pfadd 80(%%ecx), %%mm0\n\t"
449 " pfadd 84(%%ecx), %%mm0\n\t"
450 " movd %%mm0, 80(%%edx)\n\t"
451
452 " movd 80(%%ecx), %%mm0\n\t"
453 " pfsub 84(%%ecx), %%mm0\n\t"
454 " pfmul 120(%%ebx), %%mm0\n\t"
455 " pfadd %%mm0, %%mm1\n\t"
456 " pfadd 92(%%edx), %%mm0\n\t"
457 " punpckldq %%mm1, %%mm0\n\t"
458 " movq %%mm0, 84(%%edx)\n\t"
459
460 " movq 96(%%ecx), %%mm0\n\t"
461 " movq %%mm0, %%mm1\n\t"
462 " pxor %%mm7, %%mm1\n\t"
463 " pfacc %%mm1, %%mm0\n\t"
464 " pfmul %%mm6, %%mm0\n\t"
465 " movq %%mm0, 96(%%edx)\n\t"
466
467 " movd 108(%%ecx), %%mm0\n\t"
468 " pfsub 104(%%ecx), %%mm0\n\t"
469 " pfmul 120(%%ebx), %%mm0\n\t"
470 " movd %%mm0, 108(%%edx)\n\t"
471 " pfadd 104(%%ecx), %%mm0\n\t"
472 " pfadd 108(%%ecx), %%mm0\n\t"
473 " movd %%mm0, 104(%%edx)\n\t"
474
475 " movd 124(%%ecx), %%mm1\n\t"
476 " pfsub 120(%%ecx), %%mm1\n\t"
477 " pfmul 120(%%ebx), %%mm1\n\t"
478 " movd %%mm1, 124(%%edx)\n\t"
479 " pfadd 120(%%ecx), %%mm1\n\t"
480 " pfadd 124(%%ecx), %%mm1\n\t"
481 " movq %%mm1, %%mm0\n\t"
482
483 " pfadd 112(%%ecx), %%mm0\n\t"
484 " pfadd 116(%%ecx), %%mm0\n\t"
485 " movd %%mm0, 112(%%edx)\n\t"
486
487 " movd 112(%%ecx), %%mm0\n\t"
488 " pfsub 116(%%ecx), %%mm0\n\t"
489 " pfmul 120(%%ebx), %%mm0\n\t"
490 " pfadd %%mm0,%%mm1\n\t"
491 " pfadd 124(%%edx), %%mm0\n\t"
492 " punpckldq %%mm1, %%mm0\n\t"
493 " movq %%mm0, 116(%%edx)\n\t"
494
495 " jnz .L01\n\t"
496
497 /* Phase 7*/
498 /* Code below is coded in scalar mode. Should be optimized */
499
500 " movd (%%ecx), %%mm0\n\t"
501 " pfadd 4(%%ecx), %%mm0\n\t"
502 " movd %%mm0, 1024(%%esi)\n\t"
503
504 " movd (%%ecx), %%mm0\n\t"
505 " pfsub 4(%%ecx), %%mm0\n\t"
506 " pfmul 120(%%ebx), %%mm0\n\t"
507 " movd %%mm0, (%%esi)\n\t"
508 " movd %%mm0, (%%edi)\n\t"
509
510 " movd 12(%%ecx), %%mm0\n\t"
511 " pfsub 8(%%ecx), %%mm0\n\t"
512 " pfmul 120(%%ebx), %%mm0\n\t"
513 " movd %%mm0, 512(%%edi)\n\t"
514 " pfadd 12(%%ecx), %%mm0\n\t"
515 " pfadd 8(%%ecx), %%mm0\n\t"
516 " movd %%mm0, 512(%%esi)\n\t"
517
518 " movd 16(%%ecx), %%mm0\n\t"
519 " pfsub 20(%%ecx), %%mm0\n\t"
520 " pfmul 120(%%ebx), %%mm0\n\t"
521 " movq %%mm0, %%mm3\n\t"
522
523 " movd 28(%%ecx), %%mm0\n\t"
524 " pfsub 24(%%ecx), %%mm0\n\t"
525 " pfmul 120(%%ebx), %%mm0\n\t"
526 " movd %%mm0, 768(%%edi)\n\t"
527 " movq %%mm0, %%mm2\n\t"
528
529 " pfadd 24(%%ecx), %%mm0\n\t"
530 " pfadd 28(%%ecx), %%mm0\n\t"
531 " movq %%mm0, %%mm1\n\t"
532
533 " pfadd 16(%%ecx), %%mm0\n\t"
534 " pfadd 20(%%ecx), %%mm0\n\t"
535 " movd %%mm0, 768(%%esi)\n\t"
536 " pfadd %%mm3, %%mm1\n\t"
537 " movd %%mm1, 256(%%esi)\n\t"
538 " pfadd %%mm3, %%mm2\n\t"
539 " movd %%mm2, 256(%%edi)\n\t"
540
541 /* Phase 8*/
542
543 " movq 32(%%edx), %%mm0\n\t"
544 " movq 48(%%edx), %%mm1\n\t"
545 " pfadd 48(%%edx), %%mm0\n\t"
546 " pfadd 40(%%edx), %%mm1\n\t"
547 " movd %%mm0, 896(%%esi)\n\t"
548 " movd %%mm1, 640(%%esi)\n\t"
549 " psrlq $32, %%mm0\n\t"
550 " psrlq $32, %%mm1\n\t"
551 " movd %%mm0, 128(%%edi)\n\t"
552 " movd %%mm1, 384(%%edi)\n\t"
553
554 " movd 40(%%edx), %%mm0\n\t"
555 " pfadd 56(%%edx), %%mm0\n\t"
556 " movd %%mm0, 384(%%esi)\n\t"
557
558 " movd 56(%%edx), %%mm0\n\t"
559 " pfadd 36(%%edx), %%mm0\n\t"
560 " movd %%mm0, 128(%%esi)\n\t"
561
562 " movd 60(%%edx), %%mm0\n\t"
563 " movd %%mm0, 896(%%edi)\n\t"
564 " pfadd 44(%%edx), %%mm0\n\t"
565 " movd %%mm0, 640(%%edi)\n\t"
566
567 " movq 96(%%edx), %%mm0\n\t"
568 " movq 112(%%edx), %%mm2\n\t"
569 " movq 104(%%edx), %%mm4\n\t"
570 " pfadd 112(%%edx), %%mm0\n\t"
571 " pfadd 104(%%edx), %%mm2\n\t"
572 " pfadd 120(%%edx), %%mm4\n\t"
573 " movq %%mm0, %%mm1\n\t"
574 " movq %%mm2, %%mm3\n\t"
575 " movq %%mm4, %%mm5\n\t"
576 " pfadd 64(%%edx), %%mm0\n\t"
577 " pfadd 80(%%edx), %%mm2\n\t"
578 " pfadd 72(%%edx), %%mm4\n\t"
579 " movd %%mm0, 960(%%esi)\n\t"
580 " movd %%mm2, 704(%%esi)\n\t"
581 " movd %%mm4, 448(%%esi)\n\t"
582 " psrlq $32, %%mm0\n\t"
583 " psrlq $32, %%mm2\n\t"
584 " psrlq $32, %%mm4\n\t"
585 " movd %%mm0, 64(%%edi)\n\t"
586 " movd %%mm2, 320(%%edi)\n\t"
587 " movd %%mm4, 576(%%edi)\n\t"
588 " pfadd 80(%%edx), %%mm1\n\t"
589 " pfadd 72(%%edx), %%mm3\n\t"
590 " pfadd 88(%%edx), %%mm5\n\t"
591 " movd %%mm1, 832(%%esi)\n\t"
592 " movd %%mm3, 576(%%esi)\n\t"
593 " movd %%mm5, 320(%%esi)\n\t"
594 " psrlq $32, %%mm1\n\t"
595 " psrlq $32, %%mm3\n\t"
596 " psrlq $32, %%mm5\n\t"
597 " movd %%mm1, 192(%%edi)\n\t"
598 " movd %%mm3, 448(%%edi)\n\t"
599 " movd %%mm5, 704(%%edi)\n\t"
600
601 " movd 120(%%edx), %%mm0\n\t"
602 " pfadd 100(%%edx), %%mm0\n\t"
603 " movq %%mm0, %%mm1\n\t"
604 " pfadd 88(%%edx), %%mm0\n\t"
605 " movd %%mm0, 192(%%esi)\n\t"
606 " pfadd 68(%%edx), %%mm1\n\t"
607 " movd %%mm1, 64(%%esi)\n\t"
608
609 " movd 124(%%edx), %%mm0\n\t"
610 " movd %%mm0, 960(%%edi)\n\t"
611 " pfadd 92(%%edx), %%mm0\n\t"
612 " movd %%mm0, 832(%%edi)\n\t"
613
614 " jmp .L_bye\n\t"
615 ".L01: \n\t"
616 /* Phase 9*/
617
618 " movq (%%ecx), %%mm0\n\t"
619 " movq %%mm0, %%mm1\n\t"
620 " pxor %%mm7, %%mm1\n\t"
621 " pfacc %%mm1, %%mm0\n\t"
622 " pfmul %%mm6, %%mm0\n\t"
623 " pf2id %%mm0, %%mm0\n\t"
624 " movd %%mm0, %%eax\n\t"
625 " movw %%ax, 512(%%esi)\n\t"
626 " psrlq $32, %%mm0\n\t"
627 " movd %%mm0, %%eax\n\t"
628 " movw %%ax, (%%esi)\n\t"
629
630 " movd 12(%%ecx), %%mm0\n\t"
631 " pfsub 8(%%ecx), %%mm0\n\t"
632 " pfmul 120(%%ebx), %%mm0\n\t"
633 " pf2id %%mm0, %%mm7\n\t"
634 " movd %%mm7, %%eax\n\t"
635 " movw %%ax, 256(%%edi)\n\t"
636 " pfadd 12(%%ecx), %%mm0\n\t"
637 " pfadd 8(%%ecx), %%mm0\n\t"
638 " pf2id %%mm0, %%mm0\n\t"
639 " movd %%mm0, %%eax\n\t"
640 " movw %%ax, 256(%%esi)\n\t"
641
642 " movd 16(%%ecx), %%mm3\n\t"
643 " pfsub 20(%%ecx), %%mm3\n\t"
644 " pfmul 120(%%ebx), %%mm3\n\t"
645 " movq %%mm3, %%mm2\n\t"
646
647 " movd 28(%%ecx), %%mm2\n\t"
648 " pfsub 24(%%ecx), %%mm2\n\t"
649 " pfmul 120(%%ebx), %%mm2\n\t"
650 " movq %%mm2, %%mm1\n\t"
651
652 " pf2id %%mm2, %%mm7\n\t"
653 " movd %%mm7, %%eax\n\t"
654 " movw %%ax, 384(%%edi)\n\t"
655
656 " pfadd 24(%%ecx), %%mm1\n\t"
657 " pfadd 28(%%ecx), %%mm1\n\t"
658 " movq %%mm1, %%mm0\n\t"
659
660 " pfadd 16(%%ecx), %%mm0\n\t"
661 " pfadd 20(%%ecx), %%mm0\n\t"
662 " pf2id %%mm0, %%mm0\n\t"
663 " movd %%mm0, %%eax\n\t"
664 " movw %%ax, 384(%%esi)\n\t"
665 " pfadd %%mm3, %%mm1\n\t"
666 " pf2id %%mm1, %%mm1\n\t"
667 " movd %%mm1, %%eax\n\t"
668 " movw %%ax, 128(%%esi)\n\t"
669 " pfadd %%mm3, %%mm2\n\t"
670 " pf2id %%mm2, %%mm2\n\t"
671 " movd %%mm2, %%eax\n\t"
672 " movw %%ax, 128(%%edi)\n\t"
673
674 /* Phase 10*/
675
676 " movq 32(%%edx), %%mm0\n\t"
677 " movq 48(%%edx), %%mm1\n\t"
678 " pfadd 48(%%edx), %%mm0\n\t"
679 " pfadd 40(%%edx), %%mm1\n\t"
680 " pf2id %%mm0, %%mm0\n\t"
681 " pf2id %%mm1, %%mm1\n\t"
682 " movd %%mm0, %%eax\n\t"
683 " movd %%mm1, %%ecx\n\t"
684 " movw %%ax, 448(%%esi)\n\t"
685 " movw %%cx, 320(%%esi)\n\t"
686 " psrlq $32, %%mm0\n\t"
687 " psrlq $32, %%mm1\n\t"
688 " movd %%mm0, %%eax\n\t"
689 " movd %%mm1, %%ecx\n\t"
690 " movw %%ax, 64(%%edi)\n\t"
691 " movw %%cx, 192(%%edi)\n\t"
692
693 " movd 40(%%edx), %%mm3\n\t"
694 " movd 56(%%edx), %%mm4\n\t"
695 " movd 60(%%edx), %%mm0\n\t"
696 " movd 44(%%edx), %%mm2\n\t"
697 " movd 120(%%edx), %%mm5\n\t"
698 " punpckldq %%mm4, %%mm3\n\t"
699 " punpckldq 124(%%edx), %%mm0\n\t"
700 " pfadd 100(%%edx), %%mm5\n\t"
701 " punpckldq 36(%%edx), %%mm4\n\t"
702 " punpckldq 92(%%edx), %%mm2\n\t"
703 " movq %%mm5, %%mm6\n\t"
704 " pfadd %%mm4, %%mm3\n\t"
705 " pf2id %%mm0, %%mm1\n\t"
706 " pf2id %%mm3, %%mm3\n\t"
707 " pfadd 88(%%edx), %%mm5\n\t"
708 " movd %%mm1, %%eax\n\t"
709 " movd %%mm3, %%ecx\n\t"
710 " movw %%ax, 448(%%edi)\n\t"
711 " movw %%cx, 192(%%esi)\n\t"
712 " pf2id %%mm5, %%mm5\n\t"
713 " psrlq $32, %%mm1\n\t"
714 " psrlq $32, %%mm3\n\t"
715 " movd %%mm5, %%ebx\n\t"
716 " movd %%mm1, %%eax\n\t"
717 " movd %%mm3, %%ecx\n\t"
718 " movw %%bx, 96(%%esi)\n\t"
719 " movw %%ax, 480(%%edi)\n\t"
720 " movw %%cx, 64(%%esi)\n\t"
721 " pfadd %%mm2, %%mm0\n\t"
722 " pf2id %%mm0, %%mm0\n\t"
723 " movd %%mm0, %%eax\n\t"
724 " pfadd 68(%%edx), %%mm6\n\t"
725 " movw %%ax, 320(%%edi)\n\t"
726 " psrlq $32, %%mm0\n\t"
727 " pf2id %%mm6, %%mm6\n\t"
728 " movd %%mm0, %%eax\n\t"
729 " movd %%mm6, %%ebx\n\t"
730 " movw %%ax, 416(%%edi)\n\t"
731 " movw %%bx, 32(%%esi)\n\t"
732
733 " movq 96(%%edx), %%mm0\n\t"
734 " movq 112(%%edx), %%mm2\n\t"
735 " movq 104(%%edx), %%mm4\n\t"
736 " pfadd %%mm2, %%mm0\n\t"
737 " pfadd %%mm4, %%mm2\n\t"
738 " pfadd 120(%%edx), %%mm4\n\t"
739 " movq %%mm0, %%mm1\n\t"
740 " movq %%mm2, %%mm3\n\t"
741 " movq %%mm4, %%mm5\n\t"
742 " pfadd 64(%%edx), %%mm0\n\t"
743 " pfadd 80(%%edx), %%mm2\n\t"
744 " pfadd 72(%%edx), %%mm4\n\t"
745 " pf2id %%mm0, %%mm0\n\t"
746 " pf2id %%mm2, %%mm2\n\t"
747 " pf2id %%mm4, %%mm4\n\t"
748 " movd %%mm0, %%eax\n\t"
749 " movd %%mm2, %%ecx\n\t"
750 " movd %%mm4, %%ebx\n\t"
751 " movw %%ax, 480(%%esi)\n\t"
752 " movw %%cx, 352(%%esi)\n\t"
753 " movw %%bx, 224(%%esi)\n\t"
754 " psrlq $32, %%mm0\n\t"
755 " psrlq $32, %%mm2\n\t"
756 " psrlq $32, %%mm4\n\t"
757 " movd %%mm0, %%eax\n\t"
758 " movd %%mm2, %%ecx\n\t"
759 " movd %%mm4, %%ebx\n\t"
760 " movw %%ax, 32(%%edi)\n\t"
761 " movw %%cx, 160(%%edi)\n\t"
762 " movw %%bx, 288(%%edi)\n\t"
763 " pfadd 80(%%edx), %%mm1\n\t"
764 " pfadd 72(%%edx), %%mm3\n\t"
765 " pfadd 88(%%edx), %%mm5\n\t"
766 " pf2id %%mm1, %%mm1\n\t"
767 " pf2id %%mm3, %%mm3\n\t"
768 " pf2id %%mm5, %%mm5\n\t"
769 " movd %%mm1, %%eax\n\t"
770 " movd %%mm3, %%ecx\n\t"
771 " movd %%mm5, %%ebx\n\t"
772 " movw %%ax, 416(%%esi)\n\t"
773 " movw %%cx, 288(%%esi)\n\t"
774 " movw %%bx, 160(%%esi)\n\t"
775 " psrlq $32, %%mm1\n\t"
776 " psrlq $32, %%mm3\n\t"
777 " psrlq $32, %%mm5\n\t"
778 " movd %%mm1, %%eax\n\t"
779 " movd %%mm3, %%ecx\n\t"
780 " movd %%mm5, %%ebx\n\t"
781 " movw %%ax, 96(%%edi)\n\t"
782 " movw %%cx, 224(%%edi)\n\t"
783 " movw %%bx, 352(%%edi)\n\t"
784
785 " movsw\n\t"
786
787 ".L_bye:\n\t"
788 " femms\n\t"
789 :
790 :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
791 :"memory","%ebx","%esi","%edi");
792 }