|
562
|
1 /*
|
|
|
2 dct64_3dnowext: extended 3DNow optimized DCT64
|
|
|
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 dct64_3dnowext.{S,c}
|
|
|
9
|
|
|
10 MPlayer comment follows.
|
|
|
11 */
|
|
|
12
|
|
|
13 /*
|
|
|
14 * This code was taken from http://www.mpg123.org
|
|
|
15 * See ChangeLog of mpg123-0.59s-pre.1 for detail
|
|
|
16 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
|
|
|
17 * Partial 3dnowex-DSP! optimization by Nick Kurshev
|
|
|
18 *
|
|
|
19 * TODO: optimize scalar 3dnow! code
|
|
|
20 * Warning: Phases 7 & 8 are not tested
|
|
|
21 */
|
|
|
22
|
|
|
23 #include "mangle.h"
|
|
|
24
|
|
|
25 .data
|
|
|
26 ALIGN4
|
|
|
27 /* .type plus_1f, @object
|
|
|
28 .size plus_1f, 4 */
|
|
|
29 plus_1f:
|
|
|
30 .long 1065353216
|
|
|
31 ALIGN8
|
|
|
32 /* .type x_plus_minus_3dnow, @object
|
|
|
33 .size x_plus_minus_3dnow, 8 */
|
|
|
34 x_plus_minus_3dnow:
|
|
|
35 .long 0
|
|
|
36 .long -2147483648
|
|
|
37
|
|
|
38 .text
|
|
|
39 ALIGN32,,31
|
|
|
40 .globl ASM_NAME(dct64_3dnowext)
|
|
|
41 /* .type ASM_NAME(dct64_3dnowext), @function */
|
|
|
42 ASM_NAME(dct64_3dnowext):
|
|
|
43 pushl %ebp
|
|
|
44 movl %esp, %ebp
|
|
|
45 pushl %edi
|
|
|
46 pushl %esi
|
|
|
47 pushl %ebx
|
|
|
48 subl $256, %esp
|
|
|
49 /* APP */
|
|
|
50 movl 16(%ebp),%eax
|
|
|
51 leal 128+-268(%ebp),%edx
|
|
|
52 movl 8(%ebp),%esi
|
|
|
53 movl 12(%ebp),%edi
|
|
|
54 movl $ASM_NAME(costab_mmxsse),%ebx
|
|
|
55 leal -268(%ebp),%ecx
|
|
|
56 movq (%eax), %mm0
|
|
|
57 movq 8(%eax), %mm4
|
|
|
58 movq %mm0, %mm3
|
|
|
59 movq %mm4, %mm7
|
|
|
60 pswapd 120(%eax), %mm1
|
|
|
61 pswapd 112(%eax), %mm5
|
|
|
62 pfadd %mm1, %mm0
|
|
|
63 pfadd %mm5, %mm4
|
|
|
64 movq %mm0, (%edx)
|
|
|
65 movq %mm4, 8(%edx)
|
|
|
66 pfsub %mm1, %mm3
|
|
|
67 pfsub %mm5, %mm7
|
|
|
68 pfmul (%ebx), %mm3
|
|
|
69 pfmul 8(%ebx), %mm7
|
|
|
70 pswapd %mm3, %mm3
|
|
|
71 pswapd %mm7, %mm7
|
|
|
72 movq %mm3, 120(%edx)
|
|
|
73 movq %mm7, 112(%edx)
|
|
|
74 movq 16(%eax), %mm0
|
|
|
75 movq 24(%eax), %mm4
|
|
|
76 movq %mm0, %mm3
|
|
|
77 movq %mm4, %mm7
|
|
|
78 pswapd 104(%eax), %mm1
|
|
|
79 pswapd 96(%eax), %mm5
|
|
|
80 pfadd %mm1, %mm0
|
|
|
81 pfadd %mm5, %mm4
|
|
|
82 movq %mm0, 16(%edx)
|
|
|
83 movq %mm4, 24(%edx)
|
|
|
84 pfsub %mm1, %mm3
|
|
|
85 pfsub %mm5, %mm7
|
|
|
86 pfmul 16(%ebx), %mm3
|
|
|
87 pfmul 24(%ebx), %mm7
|
|
|
88 pswapd %mm3, %mm3
|
|
|
89 pswapd %mm7, %mm7
|
|
|
90 movq %mm3, 104(%edx)
|
|
|
91 movq %mm7, 96(%edx)
|
|
|
92 movq 32(%eax), %mm0
|
|
|
93 movq 40(%eax), %mm4
|
|
|
94 movq %mm0, %mm3
|
|
|
95 movq %mm4, %mm7
|
|
|
96 pswapd 88(%eax), %mm1
|
|
|
97 pswapd 80(%eax), %mm5
|
|
|
98 pfadd %mm1, %mm0
|
|
|
99 pfadd %mm5, %mm4
|
|
|
100 movq %mm0, 32(%edx)
|
|
|
101 movq %mm4, 40(%edx)
|
|
|
102 pfsub %mm1, %mm3
|
|
|
103 pfsub %mm5, %mm7
|
|
|
104 pfmul 32(%ebx), %mm3
|
|
|
105 pfmul 40(%ebx), %mm7
|
|
|
106 pswapd %mm3, %mm3
|
|
|
107 pswapd %mm7, %mm7
|
|
|
108 movq %mm3, 88(%edx)
|
|
|
109 movq %mm7, 80(%edx)
|
|
|
110 movq 48(%eax), %mm0
|
|
|
111 movq 56(%eax), %mm4
|
|
|
112 movq %mm0, %mm3
|
|
|
113 movq %mm4, %mm7
|
|
|
114 pswapd 72(%eax), %mm1
|
|
|
115 pswapd 64(%eax), %mm5
|
|
|
116 pfadd %mm1, %mm0
|
|
|
117 pfadd %mm5, %mm4
|
|
|
118 movq %mm0, 48(%edx)
|
|
|
119 movq %mm4, 56(%edx)
|
|
|
120 pfsub %mm1, %mm3
|
|
|
121 pfsub %mm5, %mm7
|
|
|
122 pfmul 48(%ebx), %mm3
|
|
|
123 pfmul 56(%ebx), %mm7
|
|
|
124 pswapd %mm3, %mm3
|
|
|
125 pswapd %mm7, %mm7
|
|
|
126 movq %mm3, 72(%edx)
|
|
|
127 movq %mm7, 64(%edx)
|
|
|
128 movq (%edx), %mm0
|
|
|
129 movq 8(%edx), %mm4
|
|
|
130 movq %mm0, %mm3
|
|
|
131 movq %mm4, %mm7
|
|
|
132 pswapd 56(%edx), %mm1
|
|
|
133 pswapd 48(%edx), %mm5
|
|
|
134 pfadd %mm1, %mm0
|
|
|
135 pfadd %mm5, %mm4
|
|
|
136 movq %mm0, (%ecx)
|
|
|
137 movq %mm4, 8(%ecx)
|
|
|
138 pfsub %mm1, %mm3
|
|
|
139 pfsub %mm5, %mm7
|
|
|
140 pfmul 64(%ebx), %mm3
|
|
|
141 pfmul 72(%ebx), %mm7
|
|
|
142 pswapd %mm3, %mm3
|
|
|
143 pswapd %mm7, %mm7
|
|
|
144 movq %mm3, 56(%ecx)
|
|
|
145 movq %mm7, 48(%ecx)
|
|
|
146 movq 16(%edx), %mm0
|
|
|
147 movq 24(%edx), %mm4
|
|
|
148 movq %mm0, %mm3
|
|
|
149 movq %mm4, %mm7
|
|
|
150 pswapd 40(%edx), %mm1
|
|
|
151 pswapd 32(%edx), %mm5
|
|
|
152 pfadd %mm1, %mm0
|
|
|
153 pfadd %mm5, %mm4
|
|
|
154 movq %mm0, 16(%ecx)
|
|
|
155 movq %mm4, 24(%ecx)
|
|
|
156 pfsub %mm1, %mm3
|
|
|
157 pfsub %mm5, %mm7
|
|
|
158 pfmul 80(%ebx), %mm3
|
|
|
159 pfmul 88(%ebx), %mm7
|
|
|
160 pswapd %mm3, %mm3
|
|
|
161 pswapd %mm7, %mm7
|
|
|
162 movq %mm3, 40(%ecx)
|
|
|
163 movq %mm7, 32(%ecx)
|
|
|
164 movq 64(%edx), %mm0
|
|
|
165 movq 72(%edx), %mm4
|
|
|
166 movq %mm0, %mm3
|
|
|
167 movq %mm4, %mm7
|
|
|
168 pswapd 120(%edx), %mm1
|
|
|
169 pswapd 112(%edx), %mm5
|
|
|
170 pfadd %mm1, %mm0
|
|
|
171 pfadd %mm5, %mm4
|
|
|
172 movq %mm0, 64(%ecx)
|
|
|
173 movq %mm4, 72(%ecx)
|
|
|
174 pfsubr %mm1, %mm3
|
|
|
175 pfsubr %mm5, %mm7
|
|
|
176 pfmul 64(%ebx), %mm3
|
|
|
177 pfmul 72(%ebx), %mm7
|
|
|
178 pswapd %mm3, %mm3
|
|
|
179 pswapd %mm7, %mm7
|
|
|
180 movq %mm3, 120(%ecx)
|
|
|
181 movq %mm7, 112(%ecx)
|
|
|
182 movq 80(%edx), %mm0
|
|
|
183 movq 88(%edx), %mm4
|
|
|
184 movq %mm0, %mm3
|
|
|
185 movq %mm4, %mm7
|
|
|
186 pswapd 104(%edx), %mm1
|
|
|
187 pswapd 96(%edx), %mm5
|
|
|
188 pfadd %mm1, %mm0
|
|
|
189 pfadd %mm5, %mm4
|
|
|
190 movq %mm0, 80(%ecx)
|
|
|
191 movq %mm4, 88(%ecx)
|
|
|
192 pfsubr %mm1, %mm3
|
|
|
193 pfsubr %mm5, %mm7
|
|
|
194 pfmul 80(%ebx), %mm3
|
|
|
195 pfmul 88(%ebx), %mm7
|
|
|
196 pswapd %mm3, %mm3
|
|
|
197 pswapd %mm7, %mm7
|
|
|
198 movq %mm3, 104(%ecx)
|
|
|
199 movq %mm7, 96(%ecx)
|
|
|
200 movq 96(%ebx), %mm2
|
|
|
201 movq 104(%ebx), %mm6
|
|
|
202 movq (%ecx), %mm0
|
|
|
203 movq 8(%ecx), %mm4
|
|
|
204 movq %mm0, %mm3
|
|
|
205 movq %mm4, %mm7
|
|
|
206 pswapd 24(%ecx), %mm1
|
|
|
207 pswapd 16(%ecx), %mm5
|
|
|
208 pfadd %mm1, %mm0
|
|
|
209 pfadd %mm5, %mm4
|
|
|
210 movq %mm0, (%edx)
|
|
|
211 movq %mm4, 8(%edx)
|
|
|
212 pfsub %mm1, %mm3
|
|
|
213 pfsub %mm5, %mm7
|
|
|
214 pfmul %mm2, %mm3
|
|
|
215 pfmul %mm6, %mm7
|
|
|
216 pswapd %mm3, %mm3
|
|
|
217 pswapd %mm7, %mm7
|
|
|
218 movq %mm3, 24(%edx)
|
|
|
219 movq %mm7, 16(%edx)
|
|
|
220 movq 32(%ecx), %mm0
|
|
|
221 movq 40(%ecx), %mm4
|
|
|
222 movq %mm0, %mm3
|
|
|
223 movq %mm4, %mm7
|
|
|
224 pswapd 56(%ecx), %mm1
|
|
|
225 pswapd 48(%ecx), %mm5
|
|
|
226 pfadd %mm1, %mm0
|
|
|
227 pfadd %mm5, %mm4
|
|
|
228 movq %mm0, 32(%edx)
|
|
|
229 movq %mm4, 40(%edx)
|
|
|
230 pfsubr %mm1, %mm3
|
|
|
231 pfsubr %mm5, %mm7
|
|
|
232 pfmul %mm2, %mm3
|
|
|
233 pfmul %mm6, %mm7
|
|
|
234 pswapd %mm3, %mm3
|
|
|
235 pswapd %mm7, %mm7
|
|
|
236 movq %mm3, 56(%edx)
|
|
|
237 movq %mm7, 48(%edx)
|
|
|
238 movq 64(%ecx), %mm0
|
|
|
239 movq 72(%ecx), %mm4
|
|
|
240 movq %mm0, %mm3
|
|
|
241 movq %mm4, %mm7
|
|
|
242 pswapd 88(%ecx), %mm1
|
|
|
243 pswapd 80(%ecx), %mm5
|
|
|
244 pfadd %mm1, %mm0
|
|
|
245 pfadd %mm5, %mm4
|
|
|
246 movq %mm0, 64(%edx)
|
|
|
247 movq %mm4, 72(%edx)
|
|
|
248 pfsub %mm1, %mm3
|
|
|
249 pfsub %mm5, %mm7
|
|
|
250 pfmul %mm2, %mm3
|
|
|
251 pfmul %mm6, %mm7
|
|
|
252 pswapd %mm3, %mm3
|
|
|
253 pswapd %mm7, %mm7
|
|
|
254 movq %mm3, 88(%edx)
|
|
|
255 movq %mm7, 80(%edx)
|
|
|
256 movq 96(%ecx), %mm0
|
|
|
257 movq 104(%ecx), %mm4
|
|
|
258 movq %mm0, %mm3
|
|
|
259 movq %mm4, %mm7
|
|
|
260 pswapd 120(%ecx), %mm1
|
|
|
261 pswapd 112(%ecx), %mm5
|
|
|
262 pfadd %mm1, %mm0
|
|
|
263 pfadd %mm5, %mm4
|
|
|
264 movq %mm0, 96(%edx)
|
|
|
265 movq %mm4, 104(%edx)
|
|
|
266 pfsubr %mm1, %mm3
|
|
|
267 pfsubr %mm5, %mm7
|
|
|
268 pfmul %mm2, %mm3
|
|
|
269 pfmul %mm6, %mm7
|
|
|
270 pswapd %mm3, %mm3
|
|
|
271 pswapd %mm7, %mm7
|
|
|
272 movq %mm3, 120(%edx)
|
|
|
273 movq %mm7, 112(%edx)
|
|
|
274 movq 112(%ebx), %mm2
|
|
|
275 movq (%edx), %mm0
|
|
|
276 movq 16(%edx), %mm4
|
|
|
277 movq %mm0, %mm3
|
|
|
278 movq %mm4, %mm7
|
|
|
279 pswapd 8(%edx), %mm1
|
|
|
280 pswapd 24(%edx), %mm5
|
|
|
281 pfadd %mm1, %mm0
|
|
|
282 pfadd %mm5, %mm4
|
|
|
283 movq %mm0, (%ecx)
|
|
|
284 movq %mm4, 16(%ecx)
|
|
|
285 pfsub %mm1, %mm3
|
|
|
286 pfsubr %mm5, %mm7
|
|
|
287 pfmul %mm2, %mm3
|
|
|
288 pfmul %mm2, %mm7
|
|
|
289 pswapd %mm3, %mm3
|
|
|
290 pswapd %mm7, %mm7
|
|
|
291 movq %mm3, 8(%ecx)
|
|
|
292 movq %mm7, 24(%ecx)
|
|
|
293 movq 32(%edx), %mm0
|
|
|
294 movq 48(%edx), %mm4
|
|
|
295 movq %mm0, %mm3
|
|
|
296 movq %mm4, %mm7
|
|
|
297 pswapd 40(%edx), %mm1
|
|
|
298 pswapd 56(%edx), %mm5
|
|
|
299 pfadd %mm1, %mm0
|
|
|
300 pfadd %mm5, %mm4
|
|
|
301 movq %mm0, 32(%ecx)
|
|
|
302 movq %mm4, 48(%ecx)
|
|
|
303 pfsub %mm1, %mm3
|
|
|
304 pfsubr %mm5, %mm7
|
|
|
305 pfmul %mm2, %mm3
|
|
|
306 pfmul %mm2, %mm7
|
|
|
307 pswapd %mm3, %mm3
|
|
|
308 pswapd %mm7, %mm7
|
|
|
309 movq %mm3, 40(%ecx)
|
|
|
310 movq %mm7, 56(%ecx)
|
|
|
311 movq 64(%edx), %mm0
|
|
|
312 movq 80(%edx), %mm4
|
|
|
313 movq %mm0, %mm3
|
|
|
314 movq %mm4, %mm7
|
|
|
315 pswapd 72(%edx), %mm1
|
|
|
316 pswapd 88(%edx), %mm5
|
|
|
317 pfadd %mm1, %mm0
|
|
|
318 pfadd %mm5, %mm4
|
|
|
319 movq %mm0, 64(%ecx)
|
|
|
320 movq %mm4, 80(%ecx)
|
|
|
321 pfsub %mm1, %mm3
|
|
|
322 pfsubr %mm5, %mm7
|
|
|
323 pfmul %mm2, %mm3
|
|
|
324 pfmul %mm2, %mm7
|
|
|
325 pswapd %mm3, %mm3
|
|
|
326 pswapd %mm7, %mm7
|
|
|
327 movq %mm3, 72(%ecx)
|
|
|
328 movq %mm7, 88(%ecx)
|
|
|
329 movq 96(%edx), %mm0
|
|
|
330 movq 112(%edx), %mm4
|
|
|
331 movq %mm0, %mm3
|
|
|
332 movq %mm4, %mm7
|
|
|
333 pswapd 104(%edx), %mm1
|
|
|
334 pswapd 120(%edx), %mm5
|
|
|
335 pfadd %mm1, %mm0
|
|
|
336 pfadd %mm5, %mm4
|
|
|
337 movq %mm0, 96(%ecx)
|
|
|
338 movq %mm4, 112(%ecx)
|
|
|
339 pfsub %mm1, %mm3
|
|
|
340 pfsubr %mm5, %mm7
|
|
|
341 pfmul %mm2, %mm3
|
|
|
342 pfmul %mm2, %mm7
|
|
|
343 pswapd %mm3, %mm3
|
|
|
344 pswapd %mm7, %mm7
|
|
|
345 movq %mm3, 104(%ecx)
|
|
|
346 movq %mm7, 120(%ecx)
|
|
|
347 movd plus_1f, %mm6
|
|
|
348 punpckldq 120(%ebx), %mm6
|
|
|
349 movq x_plus_minus_3dnow, %mm7
|
|
|
350 movq 32(%ecx), %mm0
|
|
|
351 movq 64(%ecx), %mm2
|
|
|
352 movq %mm0, %mm1
|
|
|
353 movq %mm2, %mm3
|
|
|
354 pxor %mm7, %mm1
|
|
|
355 pxor %mm7, %mm3
|
|
|
356 pfacc %mm1, %mm0
|
|
|
357 pfacc %mm3, %mm2
|
|
|
358 pfmul %mm6, %mm0
|
|
|
359 pfmul %mm6, %mm2
|
|
|
360 movq %mm0, 32(%edx)
|
|
|
361 movq %mm2, 64(%edx)
|
|
|
362 movd 44(%ecx), %mm0
|
|
|
363 movd 40(%ecx), %mm2
|
|
|
364 movd 120(%ebx), %mm3
|
|
|
365 punpckldq 76(%ecx), %mm0
|
|
|
366 punpckldq 72(%ecx), %mm2
|
|
|
367 punpckldq %mm3, %mm3
|
|
|
368 movq %mm0, %mm4
|
|
|
369 movq %mm2, %mm5
|
|
|
370 pfsub %mm2, %mm0
|
|
|
371 pfmul %mm3, %mm0
|
|
|
372 movq %mm0, %mm1
|
|
|
373 pfadd %mm5, %mm0
|
|
|
374 pfadd %mm4, %mm0
|
|
|
375 movq %mm0, %mm2
|
|
|
376 punpckldq %mm1, %mm0
|
|
|
377 punpckhdq %mm1, %mm2
|
|
|
378 movq %mm0, 40(%edx)
|
|
|
379 movq %mm2, 72(%edx)
|
|
|
380 movd 48(%ecx), %mm3
|
|
|
381 movd 60(%ecx), %mm2
|
|
|
382 pfsub 52(%ecx), %mm3
|
|
|
383 pfsub 56(%ecx), %mm2
|
|
|
384 pfmul 120(%ebx), %mm3
|
|
|
385 pfmul 120(%ebx), %mm2
|
|
|
386 movq %mm2, %mm1
|
|
|
387 pfadd 56(%ecx), %mm1
|
|
|
388 pfadd 60(%ecx), %mm1
|
|
|
389 movq %mm1, %mm0
|
|
|
390 pfadd 48(%ecx), %mm0
|
|
|
391 pfadd 52(%ecx), %mm0
|
|
|
392 pfadd %mm3, %mm1
|
|
|
393 punpckldq %mm2, %mm1
|
|
|
394 pfadd %mm3, %mm2
|
|
|
395 punpckldq %mm2, %mm0
|
|
|
396 movq %mm1, 56(%edx)
|
|
|
397 movq %mm0, 48(%edx)
|
|
|
398 movd 92(%ecx), %mm1
|
|
|
399 pfsub 88(%ecx), %mm1
|
|
|
400 pfmul 120(%ebx), %mm1
|
|
|
401 movd %mm1, 92(%edx)
|
|
|
402 pfadd 92(%ecx), %mm1
|
|
|
403 pfadd 88(%ecx), %mm1
|
|
|
404 movq %mm1, %mm0
|
|
|
405 pfadd 80(%ecx), %mm0
|
|
|
406 pfadd 84(%ecx), %mm0
|
|
|
407 movd %mm0, 80(%edx)
|
|
|
408 movd 80(%ecx), %mm0
|
|
|
409 pfsub 84(%ecx), %mm0
|
|
|
410 pfmul 120(%ebx), %mm0
|
|
|
411 pfadd %mm0, %mm1
|
|
|
412 pfadd 92(%edx), %mm0
|
|
|
413 punpckldq %mm1, %mm0
|
|
|
414 movq %mm0, 84(%edx)
|
|
|
415 movq 96(%ecx), %mm0
|
|
|
416 movq %mm0, %mm1
|
|
|
417 pxor %mm7, %mm1
|
|
|
418 pfacc %mm1, %mm0
|
|
|
419 pfmul %mm6, %mm0
|
|
|
420 movq %mm0, 96(%edx)
|
|
|
421 movd 108(%ecx), %mm0
|
|
|
422 pfsub 104(%ecx), %mm0
|
|
|
423 pfmul 120(%ebx), %mm0
|
|
|
424 movd %mm0, 108(%edx)
|
|
|
425 pfadd 104(%ecx), %mm0
|
|
|
426 pfadd 108(%ecx), %mm0
|
|
|
427 movd %mm0, 104(%edx)
|
|
|
428 movd 124(%ecx), %mm1
|
|
|
429 pfsub 120(%ecx), %mm1
|
|
|
430 pfmul 120(%ebx), %mm1
|
|
|
431 movd %mm1, 124(%edx)
|
|
|
432 pfadd 120(%ecx), %mm1
|
|
|
433 pfadd 124(%ecx), %mm1
|
|
|
434 movq %mm1, %mm0
|
|
|
435 pfadd 112(%ecx), %mm0
|
|
|
436 pfadd 116(%ecx), %mm0
|
|
|
437 movd %mm0, 112(%edx)
|
|
|
438 movd 112(%ecx), %mm0
|
|
|
439 pfsub 116(%ecx), %mm0
|
|
|
440 pfmul 120(%ebx), %mm0
|
|
|
441 pfadd %mm0,%mm1
|
|
|
442 pfadd 124(%edx), %mm0
|
|
|
443 punpckldq %mm1, %mm0
|
|
|
444 movq %mm0, 116(%edx)
|
|
|
445 jnz .L01
|
|
|
446 movd (%ecx), %mm0
|
|
|
447 pfadd 4(%ecx), %mm0
|
|
|
448 movd %mm0, 1024(%esi)
|
|
|
449 movd (%ecx), %mm0
|
|
|
450 pfsub 4(%ecx), %mm0
|
|
|
451 pfmul 120(%ebx), %mm0
|
|
|
452 movd %mm0, (%esi)
|
|
|
453 movd %mm0, (%edi)
|
|
|
454 movd 12(%ecx), %mm0
|
|
|
455 pfsub 8(%ecx), %mm0
|
|
|
456 pfmul 120(%ebx), %mm0
|
|
|
457 movd %mm0, 512(%edi)
|
|
|
458 pfadd 12(%ecx), %mm0
|
|
|
459 pfadd 8(%ecx), %mm0
|
|
|
460 movd %mm0, 512(%esi)
|
|
|
461 movd 16(%ecx), %mm0
|
|
|
462 pfsub 20(%ecx), %mm0
|
|
|
463 pfmul 120(%ebx), %mm0
|
|
|
464 movq %mm0, %mm3
|
|
|
465 movd 28(%ecx), %mm0
|
|
|
466 pfsub 24(%ecx), %mm0
|
|
|
467 pfmul 120(%ebx), %mm0
|
|
|
468 movd %mm0, 768(%edi)
|
|
|
469 movq %mm0, %mm2
|
|
|
470 pfadd 24(%ecx), %mm0
|
|
|
471 pfadd 28(%ecx), %mm0
|
|
|
472 movq %mm0, %mm1
|
|
|
473 pfadd 16(%ecx), %mm0
|
|
|
474 pfadd 20(%ecx), %mm0
|
|
|
475 movd %mm0, 768(%esi)
|
|
|
476 pfadd %mm3, %mm1
|
|
|
477 movd %mm1, 256(%esi)
|
|
|
478 pfadd %mm3, %mm2
|
|
|
479 movd %mm2, 256(%edi)
|
|
|
480 movq 32(%edx), %mm0
|
|
|
481 movq 48(%edx), %mm1
|
|
|
482 pfadd 48(%edx), %mm0
|
|
|
483 pfadd 40(%edx), %mm1
|
|
|
484 movd %mm0, 896(%esi)
|
|
|
485 movd %mm1, 640(%esi)
|
|
|
486 psrlq $32, %mm0
|
|
|
487 psrlq $32, %mm1
|
|
|
488 movd %mm0, 128(%edi)
|
|
|
489 movd %mm1, 384(%edi)
|
|
|
490 movd 40(%edx), %mm0
|
|
|
491 pfadd 56(%edx), %mm0
|
|
|
492 movd %mm0, 384(%esi)
|
|
|
493 movd 56(%edx), %mm0
|
|
|
494 pfadd 36(%edx), %mm0
|
|
|
495 movd %mm0, 128(%esi)
|
|
|
496 movd 60(%edx), %mm0
|
|
|
497 movd %mm0, 896(%edi)
|
|
|
498 pfadd 44(%edx), %mm0
|
|
|
499 movd %mm0, 640(%edi)
|
|
|
500 movq 96(%edx), %mm0
|
|
|
501 movq 112(%edx), %mm2
|
|
|
502 movq 104(%edx), %mm4
|
|
|
503 pfadd 112(%edx), %mm0
|
|
|
504 pfadd 104(%edx), %mm2
|
|
|
505 pfadd 120(%edx), %mm4
|
|
|
506 movq %mm0, %mm1
|
|
|
507 movq %mm2, %mm3
|
|
|
508 movq %mm4, %mm5
|
|
|
509 pfadd 64(%edx), %mm0
|
|
|
510 pfadd 80(%edx), %mm2
|
|
|
511 pfadd 72(%edx), %mm4
|
|
|
512 movd %mm0, 960(%esi)
|
|
|
513 movd %mm2, 704(%esi)
|
|
|
514 movd %mm4, 448(%esi)
|
|
|
515 psrlq $32, %mm0
|
|
|
516 psrlq $32, %mm2
|
|
|
517 psrlq $32, %mm4
|
|
|
518 movd %mm0, 64(%edi)
|
|
|
519 movd %mm2, 320(%edi)
|
|
|
520 movd %mm4, 576(%edi)
|
|
|
521 pfadd 80(%edx), %mm1
|
|
|
522 pfadd 72(%edx), %mm3
|
|
|
523 pfadd 88(%edx), %mm5
|
|
|
524 movd %mm1, 832(%esi)
|
|
|
525 movd %mm3, 576(%esi)
|
|
|
526 movd %mm5, 320(%esi)
|
|
|
527 psrlq $32, %mm1
|
|
|
528 psrlq $32, %mm3
|
|
|
529 psrlq $32, %mm5
|
|
|
530 movd %mm1, 192(%edi)
|
|
|
531 movd %mm3, 448(%edi)
|
|
|
532 movd %mm5, 704(%edi)
|
|
|
533 movd 120(%edx), %mm0
|
|
|
534 pfadd 100(%edx), %mm0
|
|
|
535 movq %mm0, %mm1
|
|
|
536 pfadd 88(%edx), %mm0
|
|
|
537 movd %mm0, 192(%esi)
|
|
|
538 pfadd 68(%edx), %mm1
|
|
|
539 movd %mm1, 64(%esi)
|
|
|
540 movd 124(%edx), %mm0
|
|
|
541 movd %mm0, 960(%edi)
|
|
|
542 pfadd 92(%edx), %mm0
|
|
|
543 movd %mm0, 832(%edi)
|
|
|
544 jmp .L_bye
|
|
|
545 .L01:
|
|
|
546 movq (%ecx), %mm0
|
|
|
547 movq %mm0, %mm1
|
|
|
548 pxor %mm7, %mm1
|
|
|
549 pfacc %mm1, %mm0
|
|
|
550 pfmul %mm6, %mm0
|
|
|
551 pf2iw %mm0, %mm0
|
|
|
552 movd %mm0, %eax
|
|
|
553 movw %ax, 512(%esi)
|
|
|
554 psrlq $32, %mm0
|
|
|
555 movd %mm0, %eax
|
|
|
556 movw %ax, (%esi)
|
|
|
557 movd 12(%ecx), %mm0
|
|
|
558 pfsub 8(%ecx), %mm0
|
|
|
559 pfmul 120(%ebx), %mm0
|
|
|
560 pf2iw %mm0, %mm7
|
|
|
561 movd %mm7, %eax
|
|
|
562 movw %ax, 256(%edi)
|
|
|
563 pfadd 12(%ecx), %mm0
|
|
|
564 pfadd 8(%ecx), %mm0
|
|
|
565 pf2iw %mm0, %mm0
|
|
|
566 movd %mm0, %eax
|
|
|
567 movw %ax, 256(%esi)
|
|
|
568 movd 16(%ecx), %mm3
|
|
|
569 pfsub 20(%ecx), %mm3
|
|
|
570 pfmul 120(%ebx), %mm3
|
|
|
571 movq %mm3, %mm2
|
|
|
572 movd 28(%ecx), %mm2
|
|
|
573 pfsub 24(%ecx), %mm2
|
|
|
574 pfmul 120(%ebx), %mm2
|
|
|
575 movq %mm2, %mm1
|
|
|
576 pf2iw %mm2, %mm7
|
|
|
577 movd %mm7, %eax
|
|
|
578 movw %ax, 384(%edi)
|
|
|
579 pfadd 24(%ecx), %mm1
|
|
|
580 pfadd 28(%ecx), %mm1
|
|
|
581 movq %mm1, %mm0
|
|
|
582 pfadd 16(%ecx), %mm0
|
|
|
583 pfadd 20(%ecx), %mm0
|
|
|
584 pf2iw %mm0, %mm0
|
|
|
585 movd %mm0, %eax
|
|
|
586 movw %ax, 384(%esi)
|
|
|
587 pfadd %mm3, %mm1
|
|
|
588 pf2iw %mm1, %mm1
|
|
|
589 movd %mm1, %eax
|
|
|
590 movw %ax, 128(%esi)
|
|
|
591 pfadd %mm3, %mm2
|
|
|
592 pf2iw %mm2, %mm2
|
|
|
593 movd %mm2, %eax
|
|
|
594 movw %ax, 128(%edi)
|
|
|
595 movq 32(%edx), %mm0
|
|
|
596 movq 48(%edx), %mm1
|
|
|
597 pfadd 48(%edx), %mm0
|
|
|
598 pfadd 40(%edx), %mm1
|
|
|
599 pf2iw %mm0, %mm0
|
|
|
600 pf2iw %mm1, %mm1
|
|
|
601 movd %mm0, %eax
|
|
|
602 movd %mm1, %ecx
|
|
|
603 movw %ax, 448(%esi)
|
|
|
604 movw %cx, 320(%esi)
|
|
|
605 psrlq $32, %mm0
|
|
|
606 psrlq $32, %mm1
|
|
|
607 movd %mm0, %eax
|
|
|
608 movd %mm1, %ecx
|
|
|
609 movw %ax, 64(%edi)
|
|
|
610 movw %cx, 192(%edi)
|
|
|
611 movd 40(%edx), %mm3
|
|
|
612 movd 56(%edx), %mm4
|
|
|
613 movd 60(%edx), %mm0
|
|
|
614 movd 44(%edx), %mm2
|
|
|
615 movd 120(%edx), %mm5
|
|
|
616 punpckldq %mm4, %mm3
|
|
|
617 punpckldq 124(%edx), %mm0
|
|
|
618 pfadd 100(%edx), %mm5
|
|
|
619 punpckldq 36(%edx), %mm4
|
|
|
620 punpckldq 92(%edx), %mm2
|
|
|
621 movq %mm5, %mm6
|
|
|
622 pfadd %mm4, %mm3
|
|
|
623 pf2iw %mm0, %mm1
|
|
|
624 pf2iw %mm3, %mm3
|
|
|
625 pfadd 88(%edx), %mm5
|
|
|
626 movd %mm1, %eax
|
|
|
627 movd %mm3, %ecx
|
|
|
628 movw %ax, 448(%edi)
|
|
|
629 movw %cx, 192(%esi)
|
|
|
630 pf2iw %mm5, %mm5
|
|
|
631 psrlq $32, %mm1
|
|
|
632 psrlq $32, %mm3
|
|
|
633 movd %mm5, %ebx
|
|
|
634 movd %mm1, %eax
|
|
|
635 movd %mm3, %ecx
|
|
|
636 movw %bx, 96(%esi)
|
|
|
637 movw %ax, 480(%edi)
|
|
|
638 movw %cx, 64(%esi)
|
|
|
639 pfadd %mm2, %mm0
|
|
|
640 pf2iw %mm0, %mm0
|
|
|
641 movd %mm0, %eax
|
|
|
642 pfadd 68(%edx), %mm6
|
|
|
643 movw %ax, 320(%edi)
|
|
|
644 psrlq $32, %mm0
|
|
|
645 pf2iw %mm6, %mm6
|
|
|
646 movd %mm0, %eax
|
|
|
647 movd %mm6, %ebx
|
|
|
648 movw %ax, 416(%edi)
|
|
|
649 movw %bx, 32(%esi)
|
|
|
650 movq 96(%edx), %mm0
|
|
|
651 movq 112(%edx), %mm2
|
|
|
652 movq 104(%edx), %mm4
|
|
|
653 pfadd %mm2, %mm0
|
|
|
654 pfadd %mm4, %mm2
|
|
|
655 pfadd 120(%edx), %mm4
|
|
|
656 movq %mm0, %mm1
|
|
|
657 movq %mm2, %mm3
|
|
|
658 movq %mm4, %mm5
|
|
|
659 pfadd 64(%edx), %mm0
|
|
|
660 pfadd 80(%edx), %mm2
|
|
|
661 pfadd 72(%edx), %mm4
|
|
|
662 pf2iw %mm0, %mm0
|
|
|
663 pf2iw %mm2, %mm2
|
|
|
664 pf2iw %mm4, %mm4
|
|
|
665 movd %mm0, %eax
|
|
|
666 movd %mm2, %ecx
|
|
|
667 movd %mm4, %ebx
|
|
|
668 movw %ax, 480(%esi)
|
|
|
669 movw %cx, 352(%esi)
|
|
|
670 movw %bx, 224(%esi)
|
|
|
671 psrlq $32, %mm0
|
|
|
672 psrlq $32, %mm2
|
|
|
673 psrlq $32, %mm4
|
|
|
674 movd %mm0, %eax
|
|
|
675 movd %mm2, %ecx
|
|
|
676 movd %mm4, %ebx
|
|
|
677 movw %ax, 32(%edi)
|
|
|
678 movw %cx, 160(%edi)
|
|
|
679 movw %bx, 288(%edi)
|
|
|
680 pfadd 80(%edx), %mm1
|
|
|
681 pfadd 72(%edx), %mm3
|
|
|
682 pfadd 88(%edx), %mm5
|
|
|
683 pf2iw %mm1, %mm1
|
|
|
684 pf2iw %mm3, %mm3
|
|
|
685 pf2iw %mm5, %mm5
|
|
|
686 movd %mm1, %eax
|
|
|
687 movd %mm3, %ecx
|
|
|
688 movd %mm5, %ebx
|
|
|
689 movw %ax, 416(%esi)
|
|
|
690 movw %cx, 288(%esi)
|
|
|
691 movw %bx, 160(%esi)
|
|
|
692 psrlq $32, %mm1
|
|
|
693 psrlq $32, %mm3
|
|
|
694 psrlq $32, %mm5
|
|
|
695 movd %mm1, %eax
|
|
|
696 movd %mm3, %ecx
|
|
|
697 movd %mm5, %ebx
|
|
|
698 movw %ax, 96(%edi)
|
|
|
699 movw %cx, 224(%edi)
|
|
|
700 movw %bx, 352(%edi)
|
|
|
701 movsw
|
|
|
702 .L_bye:
|
|
|
703 femms
|
|
|
704
|
|
|
705 /* NO_APP */
|
|
|
706 addl $256, %esp
|
|
|
707 popl %ebx
|
|
|
708 popl %esi
|
|
|
709 popl %edi
|
|
|
710 leave
|
|
|
711 ret
|
|
|
712 /* .size ASM_NAME(dct64_3dnowext), .-ASM_NAME(dct64_3dnowext) */
|
|
|
713
|
|
|
714 /* Mark non-executable stack. */
|
|
|
715 #if defined(__linux__) && defined(__ELF__)
|
|
|
716 .section .note.GNU-stack,"",%progbits
|
|
|
717 #endif
|
