Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
3960 | Serge | 1 | /* |
2 | dct36_3dnowext: extended 3DNow optimized DCT36 |
||
3 | |||
4 | copyright ?-2007 by the mpg123 project - free software under the terms of the LGPL 2.1 |
||
5 | see COPYING and AUTHORS files in distribution or http://mpg123.org |
||
6 | |||
7 | Transformed back into standalone asm, with help of |
||
8 | gcc -S -DHAVE_CONFIG_H -I. -march=k6-3 -O3 -Wall -pedantic -fno-strict-aliasing -DREAL_IS_FLOAT -c -o dct36_3dnowext.{S,c} |
||
9 | |||
10 | MPlayer comment follows. |
||
11 | */ |
||
12 | |||
13 | /* |
||
14 | * dct36_3dnow.c - 3DNow! optimized dct36() |
||
15 | * |
||
16 | * This code based 'dct36_3dnow.s' by Syuuhei Kashiyama |
||
17 | * |
||
18 | * |
||
19 | * - removed PREFETCH instruction for speedup |
||
20 | * - changed function name for support 3DNow! automatic detection |
||
21 | * |
||
22 | * You can find Kashiyama's original 3dnow! support patch |
||
23 | * (for mpg123-0.59o) at |
||
24 | * http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese). |
||
25 | * |
||
26 | * by KIMURA Takuhiro |
||
27 | * |
||
28 | * |
||
29 | * Modified for use with MPlayer, for details see the changelog at |
||
30 | * http://svn.mplayerhq.hu/mplayer/trunk/ |
||
31 | * $Id: dct36_3dnow.c 18786 2006-06-22 13:34:00Z diego $ |
||
32 | * |
||
33 | * Original disclaimer: |
||
34 | * The author of this program disclaim whole expressed or implied |
||
35 | * warranties with regard to this program, and in no event shall the |
||
36 | * author of this program liable to whatever resulted from the use of |
||
37 | * this program. Use it at your own risk. |
||
38 | * |
||
39 | * 2003/06/21: Moved to GCC inline assembly - Alex Beregszaszi |
||
40 | */ |
||
41 | |||
42 | #include "mangle.h" |
||
43 | |||
44 | .text |
||
45 | ALIGN32 |
||
46 | .globl ASM_NAME(dct36_3dnowext) |
||
47 | /* .type ASM_NAME(dct36_3dnowext), @function */ |
||
48 | ASM_NAME(dct36_3dnowext): |
||
49 | pushl %ebp |
||
50 | movl %esp, %ebp |
||
51 | pushl %esi |
||
52 | pushl %ebx |
||
53 | movl 8(%ebp), %eax |
||
54 | movl 12(%ebp), %esi |
||
55 | movl 16(%ebp), %ecx |
||
56 | movl 20(%ebp), %edx |
||
57 | movl 24(%ebp), %ebx |
||
58 | /* APP */ |
||
59 | movq (%eax),%mm0 |
||
60 | movq 4(%eax),%mm1 |
||
61 | pfadd %mm1,%mm0 |
||
62 | movq %mm0,4(%eax) |
||
63 | psrlq $32,%mm1 |
||
64 | movq 12(%eax),%mm2 |
||
65 | punpckldq %mm2,%mm1 |
||
66 | pfadd %mm2,%mm1 |
||
67 | movq %mm1,12(%eax) |
||
68 | psrlq $32,%mm2 |
||
69 | movq 20(%eax),%mm3 |
||
70 | punpckldq %mm3,%mm2 |
||
71 | pfadd %mm3,%mm2 |
||
72 | movq %mm2,20(%eax) |
||
73 | psrlq $32,%mm3 |
||
74 | movq 28(%eax),%mm4 |
||
75 | punpckldq %mm4,%mm3 |
||
76 | pfadd %mm4,%mm3 |
||
77 | movq %mm3,28(%eax) |
||
78 | psrlq $32,%mm4 |
||
79 | movq 36(%eax),%mm5 |
||
80 | punpckldq %mm5,%mm4 |
||
81 | pfadd %mm5,%mm4 |
||
82 | movq %mm4,36(%eax) |
||
83 | psrlq $32,%mm5 |
||
84 | movq 44(%eax),%mm6 |
||
85 | punpckldq %mm6,%mm5 |
||
86 | pfadd %mm6,%mm5 |
||
87 | movq %mm5,44(%eax) |
||
88 | psrlq $32,%mm6 |
||
89 | movq 52(%eax),%mm7 |
||
90 | punpckldq %mm7,%mm6 |
||
91 | pfadd %mm7,%mm6 |
||
92 | movq %mm6,52(%eax) |
||
93 | psrlq $32,%mm7 |
||
94 | movq 60(%eax),%mm0 |
||
95 | punpckldq %mm0,%mm7 |
||
96 | pfadd %mm0,%mm7 |
||
97 | movq %mm7,60(%eax) |
||
98 | psrlq $32,%mm0 |
||
99 | movd 68(%eax),%mm1 |
||
100 | pfadd %mm1,%mm0 |
||
101 | movd %mm0,68(%eax) |
||
102 | movd 4(%eax),%mm0 |
||
103 | movd 12(%eax),%mm1 |
||
104 | punpckldq %mm1,%mm0 |
||
105 | punpckldq 20(%eax),%mm1 |
||
106 | pfadd %mm1,%mm0 |
||
107 | movd %mm0,12(%eax) |
||
108 | psrlq $32,%mm0 |
||
109 | movd %mm0,20(%eax) |
||
110 | psrlq $32,%mm1 |
||
111 | movd 28(%eax),%mm2 |
||
112 | punpckldq %mm2,%mm1 |
||
113 | punpckldq 36(%eax),%mm2 |
||
114 | pfadd %mm2,%mm1 |
||
115 | movd %mm1,28(%eax) |
||
116 | psrlq $32,%mm1 |
||
117 | movd %mm1,36(%eax) |
||
118 | psrlq $32,%mm2 |
||
119 | movd 44(%eax),%mm3 |
||
120 | punpckldq %mm3,%mm2 |
||
121 | punpckldq 52(%eax),%mm3 |
||
122 | pfadd %mm3,%mm2 |
||
123 | movd %mm2,44(%eax) |
||
124 | psrlq $32,%mm2 |
||
125 | movd %mm2,52(%eax) |
||
126 | psrlq $32,%mm3 |
||
127 | movd 60(%eax),%mm4 |
||
128 | punpckldq %mm4,%mm3 |
||
129 | punpckldq 68(%eax),%mm4 |
||
130 | pfadd %mm4,%mm3 |
||
131 | movd %mm3,60(%eax) |
||
132 | psrlq $32,%mm3 |
||
133 | movd %mm3,68(%eax) |
||
134 | movq 24(%eax),%mm0 |
||
135 | movq 48(%eax),%mm1 |
||
136 | movd ASM_NAME(COS9)+12,%mm2 |
||
137 | punpckldq %mm2,%mm2 |
||
138 | movd ASM_NAME(COS9)+24,%mm3 |
||
139 | punpckldq %mm3,%mm3 |
||
140 | pfmul %mm2,%mm0 |
||
141 | pfmul %mm3,%mm1 |
||
142 | pushl %eax |
||
143 | movl $1,%eax |
||
144 | movd %eax,%mm7 |
||
145 | pi2fd %mm7,%mm7 |
||
146 | popl %eax |
||
147 | movq 8(%eax),%mm2 |
||
148 | movd ASM_NAME(COS9)+4,%mm3 |
||
149 | punpckldq %mm3,%mm3 |
||
150 | pfmul %mm3,%mm2 |
||
151 | pfadd %mm0,%mm2 |
||
152 | movq 40(%eax),%mm3 |
||
153 | movd ASM_NAME(COS9)+20,%mm4 |
||
154 | punpckldq %mm4,%mm4 |
||
155 | pfmul %mm4,%mm3 |
||
156 | pfadd %mm3,%mm2 |
||
157 | movq 56(%eax),%mm3 |
||
158 | movd ASM_NAME(COS9)+28,%mm4 |
||
159 | punpckldq %mm4,%mm4 |
||
160 | pfmul %mm4,%mm3 |
||
161 | pfadd %mm3,%mm2 |
||
162 | movq (%eax),%mm3 |
||
163 | movq 16(%eax),%mm4 |
||
164 | movd ASM_NAME(COS9)+8,%mm5 |
||
165 | punpckldq %mm5,%mm5 |
||
166 | pfmul %mm5,%mm4 |
||
167 | pfadd %mm4,%mm3 |
||
168 | movq 32(%eax),%mm4 |
||
169 | movd ASM_NAME(COS9)+16,%mm5 |
||
170 | punpckldq %mm5,%mm5 |
||
171 | pfmul %mm5,%mm4 |
||
172 | pfadd %mm4,%mm3 |
||
173 | pfadd %mm1,%mm3 |
||
174 | movq 64(%eax),%mm4 |
||
175 | movd ASM_NAME(COS9)+32,%mm5 |
||
176 | punpckldq %mm5,%mm5 |
||
177 | pfmul %mm5,%mm4 |
||
178 | pfadd %mm4,%mm3 |
||
179 | movq %mm2,%mm4 |
||
180 | pfadd %mm3,%mm4 |
||
181 | movq %mm7,%mm5 |
||
182 | punpckldq ASM_NAME(tfcos36)+0,%mm5 |
||
183 | pfmul %mm5,%mm4 |
||
184 | movq %mm4,%mm5 |
||
185 | pfacc %mm5,%mm5 |
||
186 | movd 108(%edx),%mm6 |
||
187 | punpckldq 104(%edx),%mm6 |
||
188 | pfmul %mm6,%mm5 |
||
189 | pswapd %mm5,%mm5 |
||
190 | movq %mm5,32(%ecx) |
||
191 | movq %mm4,%mm6 |
||
192 | punpckldq %mm6,%mm5 |
||
193 | pfsub %mm6,%mm5 |
||
194 | punpckhdq %mm5,%mm5 |
||
195 | movd 32(%edx),%mm6 |
||
196 | punpckldq 36(%edx),%mm6 |
||
197 | pfmul %mm6,%mm5 |
||
198 | movd 32(%esi),%mm6 |
||
199 | punpckldq 36(%esi),%mm6 |
||
200 | pfadd %mm6,%mm5 |
||
201 | movd %mm5,1024(%ebx) |
||
202 | psrlq $32,%mm5 |
||
203 | movd %mm5,1152(%ebx) |
||
204 | movq %mm3,%mm4 |
||
205 | pfsub %mm2,%mm4 |
||
206 | movq %mm7,%mm5 |
||
207 | punpckldq ASM_NAME(tfcos36)+32,%mm5 |
||
208 | pfmul %mm5,%mm4 |
||
209 | movq %mm4,%mm5 |
||
210 | pfacc %mm5,%mm5 |
||
211 | movd 140(%edx),%mm6 |
||
212 | punpckldq 72(%edx),%mm6 |
||
213 | pfmul %mm6,%mm5 |
||
214 | movd %mm5,68(%ecx) |
||
215 | psrlq $32,%mm5 |
||
216 | movd %mm5,0(%ecx) |
||
217 | movq %mm4,%mm6 |
||
218 | punpckldq %mm6,%mm5 |
||
219 | pfsub %mm6,%mm5 |
||
220 | punpckhdq %mm5,%mm5 |
||
221 | movd 0(%edx),%mm6 |
||
222 | punpckldq 68(%edx),%mm6 |
||
223 | pfmul %mm6,%mm5 |
||
224 | movd 0(%esi),%mm6 |
||
225 | punpckldq 68(%esi),%mm6 |
||
226 | pfadd %mm6,%mm5 |
||
227 | movd %mm5,0(%ebx) |
||
228 | psrlq $32,%mm5 |
||
229 | movd %mm5,2176(%ebx) |
||
230 | movq 8(%eax),%mm2 |
||
231 | movq 40(%eax),%mm3 |
||
232 | pfsub %mm3,%mm2 |
||
233 | movq 56(%eax),%mm3 |
||
234 | pfsub %mm3,%mm2 |
||
235 | movd ASM_NAME(COS9)+12,%mm3 |
||
236 | punpckldq %mm3,%mm3 |
||
237 | pfmul %mm3,%mm2 |
||
238 | movq 16(%eax),%mm3 |
||
239 | movq 32(%eax),%mm4 |
||
240 | pfsub %mm4,%mm3 |
||
241 | movq 64(%eax),%mm4 |
||
242 | pfsub %mm4,%mm3 |
||
243 | movd ASM_NAME(COS9)+24,%mm4 |
||
244 | punpckldq %mm4,%mm4 |
||
245 | pfmul %mm4,%mm3 |
||
246 | movq 48(%eax),%mm4 |
||
247 | pfsub %mm4,%mm3 |
||
248 | movq (%eax),%mm4 |
||
249 | pfadd %mm4,%mm3 |
||
250 | movq %mm2,%mm4 |
||
251 | pfadd %mm3,%mm4 |
||
252 | movq %mm7,%mm5 |
||
253 | punpckldq ASM_NAME(tfcos36)+4,%mm5 |
||
254 | pfmul %mm5,%mm4 |
||
255 | movq %mm4,%mm5 |
||
256 | pfacc %mm5,%mm5 |
||
257 | movd 112(%edx),%mm6 |
||
258 | punpckldq 100(%edx),%mm6 |
||
259 | pfmul %mm6,%mm5 |
||
260 | movd %mm5,40(%ecx) |
||
261 | psrlq $32,%mm5 |
||
262 | movd %mm5,28(%ecx) |
||
263 | movq %mm4,%mm6 |
||
264 | punpckldq %mm6,%mm5 |
||
265 | pfsub %mm6,%mm5 |
||
266 | punpckhdq %mm5,%mm5 |
||
267 | movd 28(%edx),%mm6 |
||
268 | punpckldq 40(%edx),%mm6 |
||
269 | pfmul %mm6,%mm5 |
||
270 | movd 28(%esi),%mm6 |
||
271 | punpckldq 40(%esi),%mm6 |
||
272 | pfadd %mm6,%mm5 |
||
273 | movd %mm5,896(%ebx) |
||
274 | psrlq $32,%mm5 |
||
275 | movd %mm5,1280(%ebx) |
||
276 | movq %mm3,%mm4 |
||
277 | pfsub %mm2,%mm4 |
||
278 | movq %mm7,%mm5 |
||
279 | punpckldq ASM_NAME(tfcos36)+28,%mm5 |
||
280 | pfmul %mm5,%mm4 |
||
281 | movq %mm4,%mm5 |
||
282 | pfacc %mm5,%mm5 |
||
283 | movd 136(%edx),%mm6 |
||
284 | punpckldq 76(%edx),%mm6 |
||
285 | pfmul %mm6,%mm5 |
||
286 | movd %mm5,64(%ecx) |
||
287 | psrlq $32,%mm5 |
||
288 | movd %mm5,4(%ecx) |
||
289 | movq %mm4,%mm6 |
||
290 | punpckldq %mm6,%mm5 |
||
291 | pfsub %mm6,%mm5 |
||
292 | punpckhdq %mm5,%mm5 |
||
293 | movd 4(%edx),%mm6 |
||
294 | punpckldq 64(%edx),%mm6 |
||
295 | pfmul %mm6,%mm5 |
||
296 | movd 4(%esi),%mm6 |
||
297 | punpckldq 64(%esi),%mm6 |
||
298 | pfadd %mm6,%mm5 |
||
299 | movd %mm5,128(%ebx) |
||
300 | psrlq $32,%mm5 |
||
301 | movd %mm5,2048(%ebx) |
||
302 | movq 8(%eax),%mm2 |
||
303 | movd ASM_NAME(COS9)+20,%mm3 |
||
304 | punpckldq %mm3,%mm3 |
||
305 | pfmul %mm3,%mm2 |
||
306 | pfsub %mm0,%mm2 |
||
307 | movq 40(%eax),%mm3 |
||
308 | movd ASM_NAME(COS9)+28,%mm4 |
||
309 | punpckldq %mm4,%mm4 |
||
310 | pfmul %mm4,%mm3 |
||
311 | pfsub %mm3,%mm2 |
||
312 | movq 56(%eax),%mm3 |
||
313 | movd ASM_NAME(COS9)+4,%mm4 |
||
314 | punpckldq %mm4,%mm4 |
||
315 | pfmul %mm4,%mm3 |
||
316 | pfadd %mm3,%mm2 |
||
317 | movq (%eax),%mm3 |
||
318 | movq 16(%eax),%mm4 |
||
319 | movd ASM_NAME(COS9)+32,%mm5 |
||
320 | punpckldq %mm5,%mm5 |
||
321 | pfmul %mm5,%mm4 |
||
322 | pfsub %mm4,%mm3 |
||
323 | movq 32(%eax),%mm4 |
||
324 | movd ASM_NAME(COS9)+8,%mm5 |
||
325 | punpckldq %mm5,%mm5 |
||
326 | pfmul %mm5,%mm4 |
||
327 | pfsub %mm4,%mm3 |
||
328 | pfadd %mm1,%mm3 |
||
329 | movq 64(%eax),%mm4 |
||
330 | movd ASM_NAME(COS9)+16,%mm5 |
||
331 | punpckldq %mm5,%mm5 |
||
332 | pfmul %mm5,%mm4 |
||
333 | pfadd %mm4,%mm3 |
||
334 | movq %mm2,%mm4 |
||
335 | pfadd %mm3,%mm4 |
||
336 | movq %mm7,%mm5 |
||
337 | punpckldq ASM_NAME(tfcos36)+8,%mm5 |
||
338 | pfmul %mm5,%mm4 |
||
339 | movq %mm4,%mm5 |
||
340 | pfacc %mm5,%mm5 |
||
341 | movd 116(%edx),%mm6 |
||
342 | punpckldq 96(%edx),%mm6 |
||
343 | pfmul %mm6,%mm5 |
||
344 | movd %mm5,44(%ecx) |
||
345 | psrlq $32,%mm5 |
||
346 | movd %mm5,24(%ecx) |
||
347 | movq %mm4,%mm6 |
||
348 | punpckldq %mm6,%mm5 |
||
349 | pfsub %mm6,%mm5 |
||
350 | punpckhdq %mm5,%mm5 |
||
351 | movd 24(%edx),%mm6 |
||
352 | punpckldq 44(%edx),%mm6 |
||
353 | pfmul %mm6,%mm5 |
||
354 | movd 24(%esi),%mm6 |
||
355 | punpckldq 44(%esi),%mm6 |
||
356 | pfadd %mm6,%mm5 |
||
357 | movd %mm5,768(%ebx) |
||
358 | psrlq $32,%mm5 |
||
359 | movd %mm5,1408(%ebx) |
||
360 | movq %mm3,%mm4 |
||
361 | pfsub %mm2,%mm4 |
||
362 | movq %mm7,%mm5 |
||
363 | punpckldq ASM_NAME(tfcos36)+24,%mm5 |
||
364 | pfmul %mm5,%mm4 |
||
365 | movq %mm4,%mm5 |
||
366 | pfacc %mm5,%mm5 |
||
367 | movd 132(%edx),%mm6 |
||
368 | punpckldq 80(%edx),%mm6 |
||
369 | pfmul %mm6,%mm5 |
||
370 | movd %mm5,60(%ecx) |
||
371 | psrlq $32,%mm5 |
||
372 | movd %mm5,8(%ecx) |
||
373 | movq %mm4,%mm6 |
||
374 | punpckldq %mm6,%mm5 |
||
375 | pfsub %mm6,%mm5 |
||
376 | punpckhdq %mm5,%mm5 |
||
377 | movd 8(%edx),%mm6 |
||
378 | punpckldq 60(%edx),%mm6 |
||
379 | pfmul %mm6,%mm5 |
||
380 | movd 8(%esi),%mm6 |
||
381 | punpckldq 60(%esi),%mm6 |
||
382 | pfadd %mm6,%mm5 |
||
383 | movd %mm5,256(%ebx) |
||
384 | psrlq $32,%mm5 |
||
385 | movd %mm5,1920(%ebx) |
||
386 | movq 8(%eax),%mm2 |
||
387 | movd ASM_NAME(COS9)+28,%mm3 |
||
388 | punpckldq %mm3,%mm3 |
||
389 | pfmul %mm3,%mm2 |
||
390 | pfsub %mm0,%mm2 |
||
391 | movq 40(%eax),%mm3 |
||
392 | movd ASM_NAME(COS9)+4,%mm4 |
||
393 | punpckldq %mm4,%mm4 |
||
394 | pfmul %mm4,%mm3 |
||
395 | pfadd %mm3,%mm2 |
||
396 | movq 56(%eax),%mm3 |
||
397 | movd ASM_NAME(COS9)+20,%mm4 |
||
398 | punpckldq %mm4,%mm4 |
||
399 | pfmul %mm4,%mm3 |
||
400 | pfsub %mm3,%mm2 |
||
401 | movq (%eax),%mm3 |
||
402 | movq 16(%eax),%mm4 |
||
403 | movd ASM_NAME(COS9)+16,%mm5 |
||
404 | punpckldq %mm5,%mm5 |
||
405 | pfmul %mm5,%mm4 |
||
406 | pfsub %mm4,%mm3 |
||
407 | movq 32(%eax),%mm4 |
||
408 | movd ASM_NAME(COS9)+32,%mm5 |
||
409 | punpckldq %mm5,%mm5 |
||
410 | pfmul %mm5,%mm4 |
||
411 | pfadd %mm4,%mm3 |
||
412 | pfadd %mm1,%mm3 |
||
413 | movq 64(%eax),%mm4 |
||
414 | movd ASM_NAME(COS9)+8,%mm5 |
||
415 | punpckldq %mm5,%mm5 |
||
416 | pfmul %mm5,%mm4 |
||
417 | pfsub %mm4,%mm3 |
||
418 | movq %mm2,%mm4 |
||
419 | pfadd %mm3,%mm4 |
||
420 | movq %mm7,%mm5 |
||
421 | punpckldq ASM_NAME(tfcos36)+12,%mm5 |
||
422 | pfmul %mm5,%mm4 |
||
423 | movq %mm4,%mm5 |
||
424 | pfacc %mm5,%mm5 |
||
425 | movd 120(%edx),%mm6 |
||
426 | punpckldq 92(%edx),%mm6 |
||
427 | pfmul %mm6,%mm5 |
||
428 | movd %mm5,48(%ecx) |
||
429 | psrlq $32,%mm5 |
||
430 | movd %mm5,20(%ecx) |
||
431 | movq %mm4,%mm6 |
||
432 | punpckldq %mm6,%mm5 |
||
433 | pfsub %mm6,%mm5 |
||
434 | punpckhdq %mm5,%mm5 |
||
435 | movd 20(%edx),%mm6 |
||
436 | punpckldq 48(%edx),%mm6 |
||
437 | pfmul %mm6,%mm5 |
||
438 | movd 20(%esi),%mm6 |
||
439 | punpckldq 48(%esi),%mm6 |
||
440 | pfadd %mm6,%mm5 |
||
441 | movd %mm5,640(%ebx) |
||
442 | psrlq $32,%mm5 |
||
443 | movd %mm5,1536(%ebx) |
||
444 | movq %mm3,%mm4 |
||
445 | pfsub %mm2,%mm4 |
||
446 | movq %mm7,%mm5 |
||
447 | punpckldq ASM_NAME(tfcos36)+20,%mm5 |
||
448 | pfmul %mm5,%mm4 |
||
449 | movq %mm4,%mm5 |
||
450 | pfacc %mm5,%mm5 |
||
451 | movd 128(%edx),%mm6 |
||
452 | punpckldq 84(%edx),%mm6 |
||
453 | pfmul %mm6,%mm5 |
||
454 | movd %mm5,56(%ecx) |
||
455 | psrlq $32,%mm5 |
||
456 | movd %mm5,12(%ecx) |
||
457 | movq %mm4,%mm6 |
||
458 | punpckldq %mm6,%mm5 |
||
459 | pfsub %mm6,%mm5 |
||
460 | punpckhdq %mm5,%mm5 |
||
461 | movd 12(%edx),%mm6 |
||
462 | punpckldq 56(%edx),%mm6 |
||
463 | pfmul %mm6,%mm5 |
||
464 | movd 12(%esi),%mm6 |
||
465 | punpckldq 56(%esi),%mm6 |
||
466 | pfadd %mm6,%mm5 |
||
467 | movd %mm5,384(%ebx) |
||
468 | psrlq $32,%mm5 |
||
469 | movd %mm5,1792(%ebx) |
||
470 | movq (%eax),%mm4 |
||
471 | movq 16(%eax),%mm3 |
||
472 | pfsub %mm3,%mm4 |
||
473 | movq 32(%eax),%mm3 |
||
474 | pfadd %mm3,%mm4 |
||
475 | movq 48(%eax),%mm3 |
||
476 | pfsub %mm3,%mm4 |
||
477 | movq 64(%eax),%mm3 |
||
478 | pfadd %mm3,%mm4 |
||
479 | movq %mm7,%mm5 |
||
480 | punpckldq ASM_NAME(tfcos36)+16,%mm5 |
||
481 | pfmul %mm5,%mm4 |
||
482 | movq %mm4,%mm5 |
||
483 | pfacc %mm5,%mm5 |
||
484 | movd 124(%edx),%mm6 |
||
485 | punpckldq 88(%edx),%mm6 |
||
486 | pfmul %mm6,%mm5 |
||
487 | movd %mm5,52(%ecx) |
||
488 | psrlq $32,%mm5 |
||
489 | movd %mm5,16(%ecx) |
||
490 | movq %mm4,%mm6 |
||
491 | punpckldq %mm6,%mm5 |
||
492 | pfsub %mm6,%mm5 |
||
493 | punpckhdq %mm5,%mm5 |
||
494 | movd 16(%edx),%mm6 |
||
495 | punpckldq 52(%edx),%mm6 |
||
496 | pfmul %mm6,%mm5 |
||
497 | movd 16(%esi),%mm6 |
||
498 | punpckldq 52(%esi),%mm6 |
||
499 | pfadd %mm6,%mm5 |
||
500 | movd %mm5,512(%ebx) |
||
501 | psrlq $32,%mm5 |
||
502 | movd %mm5,1664(%ebx) |
||
503 | femms |
||
504 | |||
505 | /* NO_APP */ |
||
506 | popl %ebx |
||
507 | popl %esi |
||
508 | leave |
||
509 | ret |
||
510 | /* .size ASM_NAME(dct36_3dnowext), .-ASM_NAME(dct36_3dnowext) */ |
||
511 | |||
512 | NONEXEC_STACK |