annotate mp3lib/dct64_k7.c @ 7179:a258b9b7669e

dct_algo
author michael
date Fri, 30 Aug 2002 21:30:59 +0000
parents 82a03c17f314
children 7d20830716a1
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
4148
nick
parents:
diff changeset
1 /*
nick
parents:
diff changeset
2 * This code was taken from http://www.mpg123.org
nick
parents:
diff changeset
3 * See ChangeLog of mpg123-0.59s-pre.1 for detail
nick
parents:
diff changeset
4 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
nick
parents:
diff changeset
5 * Partial 3dnowex-DSP! optimization by Nick Kurshev
nick
parents:
diff changeset
6 *
nick
parents:
diff changeset
7 * TODO: optimize scalar 3dnow! code
nick
parents:
diff changeset
8 * Warning: Phases 7 & 8 are not tested
nick
parents:
diff changeset
9 */
nick
parents:
diff changeset
10 #define real float /* ugly - but only way */
nick
parents:
diff changeset
11
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
12 #include "../mangle.h"
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
13
4148
nick
parents:
diff changeset
14 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
nick
parents:
diff changeset
15 static float plus_1f = 1.0;
nick
parents:
diff changeset
16
nick
parents:
diff changeset
17 void __attribute__ (( __stdcall__ )) dct64_MMX_3dnowex(real *a,real *b,real *c)
nick
parents:
diff changeset
18 {
nick
parents:
diff changeset
19 char tmp[256];
nick
parents:
diff changeset
20 __asm __volatile(
nick
parents:
diff changeset
21 " movl %2,%%eax\n\t"
nick
parents:
diff changeset
22
nick
parents:
diff changeset
23 " leal 128+%3,%%edx\n\t"
nick
parents:
diff changeset
24 " movl %0,%%esi\n\t"
nick
parents:
diff changeset
25 " movl %1,%%edi\n\t"
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
26 " movl $"MANGLE(costab_mmx)",%%ebx\n\t"
4148
nick
parents:
diff changeset
27 " leal %3,%%ecx\n\t"
nick
parents:
diff changeset
28
nick
parents:
diff changeset
29 /* Phase 1*/
nick
parents:
diff changeset
30 " movq (%%eax), %%mm0\n\t"
nick
parents:
diff changeset
31 " movq 8(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
32 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
33 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
34 " movq 120(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
35 " movq 112(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
36 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
37 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
38 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
39 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
40 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
41 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
42 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
43 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
44 " pfmul (%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
45 " pfmul 8(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
46 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
47 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
48 " movq %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
49 " movq %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
50
nick
parents:
diff changeset
51 " movq 16(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
52 " movq 24(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
53 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
54 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
55 " movq 104(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
56 " movq 96(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
57 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
58 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
59 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
60 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
61 " movq %%mm0, 16(%%edx)\n\t"
nick
parents:
diff changeset
62 " movq %%mm4, 24(%%edx)\n\t"
nick
parents:
diff changeset
63 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
64 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
65 " pfmul 16(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
66 " pfmul 24(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
67 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
68 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
69 " movq %%mm3, 104(%%edx)\n\t"
nick
parents:
diff changeset
70 " movq %%mm7, 96(%%edx)\n\t"
nick
parents:
diff changeset
71
nick
parents:
diff changeset
72 " movq 32(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
73 " movq 40(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
74 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
75 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
76 " movq 88(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
77 " movq 80(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
78 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
79 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
80 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
81 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
82 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
83 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
84 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
85 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
86 " pfmul 32(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
87 " pfmul 40(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
88 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
89 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
90 " movq %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
91 " movq %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
92
nick
parents:
diff changeset
93 " movq 48(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
94 " movq 56(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
95 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
96 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
97 " movq 72(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
98 " movq 64(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
99 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
100 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
101 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
102 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
103 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
104 " movq %%mm4, 56(%%edx)\n\t"
nick
parents:
diff changeset
105 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
106 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
107 " pfmul 48(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
108 " pfmul 56(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
109 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
110 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
111 " movq %%mm3, 72(%%edx)\n\t"
nick
parents:
diff changeset
112 " movq %%mm7, 64(%%edx)\n\t"
nick
parents:
diff changeset
113
nick
parents:
diff changeset
114 /* Phase 2*/
nick
parents:
diff changeset
115
nick
parents:
diff changeset
116 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
117 " movq 8(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
118 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
119 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
120 " movq 56(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
121 " movq 48(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
122 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
123 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
124 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
125 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
126 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
127 " movq %%mm4, 8(%%ecx)\n\t"
nick
parents:
diff changeset
128 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
129 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
130 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
131 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
132 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
133 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
134 " movq %%mm3, 56(%%ecx)\n\t"
nick
parents:
diff changeset
135 " movq %%mm7, 48(%%ecx)\n\t"
nick
parents:
diff changeset
136
nick
parents:
diff changeset
137 " movq 16(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
138 " movq 24(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
139 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
140 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
141 " movq 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
142 " movq 32(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
143 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
144 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
145 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
146 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
147 " movq %%mm0, 16(%%ecx)\n\t"
nick
parents:
diff changeset
148 " movq %%mm4, 24(%%ecx)\n\t"
nick
parents:
diff changeset
149 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
150 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
151 " pfmul 80(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
152 " pfmul 88(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
153 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
154 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
155 " movq %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
156 " movq %%mm7, 32(%%ecx)\n\t"
nick
parents:
diff changeset
157
nick
parents:
diff changeset
158 /* Phase 3*/
nick
parents:
diff changeset
159
nick
parents:
diff changeset
160 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
161 " movq 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
162 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
163 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
164 " movq 120(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
165 " movq 112(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
166 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
167 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
168 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
169 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
170 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
171 " movq %%mm4, 72(%%ecx)\n\t"
nick
parents:
diff changeset
172 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
173 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
174 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
175 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
176 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
177 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
178 " movq %%mm3, 120(%%ecx)\n\t"
nick
parents:
diff changeset
179 " movq %%mm7, 112(%%ecx)\n\t"
nick
parents:
diff changeset
180
nick
parents:
diff changeset
181 " movq 80(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
182 " movq 88(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
183 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
184 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
185 " movq 104(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
186 " movq 96(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
187 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
188 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
189 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
190 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
191 " movq %%mm0, 80(%%ecx)\n\t"
nick
parents:
diff changeset
192 " movq %%mm4, 88(%%ecx)\n\t"
nick
parents:
diff changeset
193 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
194 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
195 " pfmul 80(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
196 " pfmul 88(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
197 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
198 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
199 " movq %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
200 " movq %%mm7, 96(%%ecx)\n\t"
nick
parents:
diff changeset
201
nick
parents:
diff changeset
202 /* Phase 4*/
nick
parents:
diff changeset
203
nick
parents:
diff changeset
204 " movq 96(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
205 " movq 104(%%ebx), %%mm6\n\t"
nick
parents:
diff changeset
206
nick
parents:
diff changeset
207 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
208 " movq 8(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
209 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
210 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
211 " movq 24(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
212 " movq 16(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
213 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
214 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
215 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
216 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
217 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
218 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
219 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
220 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
221 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
222 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
223 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
224 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
225 " movq %%mm3, 24(%%edx)\n\t"
nick
parents:
diff changeset
226 " movq %%mm7, 16(%%edx)\n\t"
nick
parents:
diff changeset
227
nick
parents:
diff changeset
228 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
229 " movq 40(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
230 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
231 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
232 " movq 56(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
233 " movq 48(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
234 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
235 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
236 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
237 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
238 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
239 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
240 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
241 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
242 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
243 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
244 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
245 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
246 " movq %%mm3, 56(%%edx)\n\t"
nick
parents:
diff changeset
247 " movq %%mm7, 48(%%edx)\n\t"
nick
parents:
diff changeset
248
nick
parents:
diff changeset
249 " movq 64(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
250 " movq 72(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
251 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
252 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
253 " movq 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
254 " movq 80(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
255 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
256 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
257 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
258 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
259 " movq %%mm0, 64(%%edx)\n\t"
nick
parents:
diff changeset
260 " movq %%mm4, 72(%%edx)\n\t"
nick
parents:
diff changeset
261 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
262 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
263 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
264 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
265 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
266 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
267 " movq %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
268 " movq %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
269
nick
parents:
diff changeset
270 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
271 " movq 104(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
272 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
273 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
274 " movq 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
275 " movq 112(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
276 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
277 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
278 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
279 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
280 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
281 " movq %%mm4, 104(%%edx)\n\t"
nick
parents:
diff changeset
282 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
283 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
284 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
285 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
286 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
287 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
288 " movq %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
289 " movq %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
290
nick
parents:
diff changeset
291 /* Phase 5 */
nick
parents:
diff changeset
292
nick
parents:
diff changeset
293 " movq 112(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
294
nick
parents:
diff changeset
295 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
296 " movq 16(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
297 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
298 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
299 " movq 8(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
300 " movq 24(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
301 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
302 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
303 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
304 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
305 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
306 " movq %%mm4, 16(%%ecx)\n\t"
nick
parents:
diff changeset
307 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
308 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
309 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
310 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
311 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
312 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
313 " movq %%mm3, 8(%%ecx)\n\t"
nick
parents:
diff changeset
314 " movq %%mm7, 24(%%ecx)\n\t"
nick
parents:
diff changeset
315
nick
parents:
diff changeset
316 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
317 " movq 48(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
318 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
319 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
320 " movq 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
321 " movq 56(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
322 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
323 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
324 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
325 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
326 " movq %%mm0, 32(%%ecx)\n\t"
nick
parents:
diff changeset
327 " movq %%mm4, 48(%%ecx)\n\t"
nick
parents:
diff changeset
328 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
329 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
330 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
331 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
332 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
333 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
334 " movq %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
335 " movq %%mm7, 56(%%ecx)\n\t"
nick
parents:
diff changeset
336
nick
parents:
diff changeset
337 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
338 " movq 80(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
339 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
340 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
341 " movq 72(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
342 " movq 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
343 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
344 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
345 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
346 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
347 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
348 " movq %%mm4, 80(%%ecx)\n\t"
nick
parents:
diff changeset
349 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
350 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
351 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
352 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
353 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
354 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
355 " movq %%mm3, 72(%%ecx)\n\t"
nick
parents:
diff changeset
356 " movq %%mm7, 88(%%ecx)\n\t"
nick
parents:
diff changeset
357
nick
parents:
diff changeset
358 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
359 " movq 112(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
360 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
361 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
362 " movq 104(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
363 " movq 120(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
364 " pswapd %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
365 " pswapd %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
366 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
367 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
368 " movq %%mm0, 96(%%ecx)\n\t"
nick
parents:
diff changeset
369 " movq %%mm4, 112(%%ecx)\n\t"
nick
parents:
diff changeset
370 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
371 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
372 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
373 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
374 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
375 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
376 " movq %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
377 " movq %%mm7, 120(%%ecx)\n\t"
nick
parents:
diff changeset
378
nick
parents:
diff changeset
379
nick
parents:
diff changeset
380 /* Phase 6. This is the end of easy road. */
nick
parents:
diff changeset
381 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
382
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
383 " movd "MANGLE(plus_1f)", %%mm6\n\t"
4148
nick
parents:
diff changeset
384 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/
nick
parents:
diff changeset
385 " movq x_plus_minus_3dnow, %%mm7\n\t" /* mm7 = +1 | -1 */
nick
parents:
diff changeset
386
nick
parents:
diff changeset
387 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
388 " movq 64(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
389 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
390 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
391 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
392 " pxor %%mm7, %%mm3\n\t"
nick
parents:
diff changeset
393 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
394 " pfacc %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
395 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
396 " pfmul %%mm6, %%mm2\n\t"
nick
parents:
diff changeset
397 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
398 " movq %%mm2, 64(%%edx)\n\t"
nick
parents:
diff changeset
399
nick
parents:
diff changeset
400 " movd 44(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
401 " movd 40(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
402 " movd 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
403 " punpckldq 76(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
404 " punpckldq 72(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
405 " punpckldq %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
406 " movq %%mm0, %%mm4\n\t"
nick
parents:
diff changeset
407 " movq %%mm2, %%mm5\n\t"
nick
parents:
diff changeset
408 " pfsub %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
409 " pfmul %%mm3, %%mm0\n\t"
nick
parents:
diff changeset
410 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
411 " pfadd %%mm5, %%mm0\n\t"
nick
parents:
diff changeset
412 " pfadd %%mm4, %%mm0\n\t"
nick
parents:
diff changeset
413 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
414 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
415 " punpckhdq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
416 " movq %%mm0, 40(%%edx)\n\t"
nick
parents:
diff changeset
417 " movq %%mm2, 72(%%edx)\n\t"
nick
parents:
diff changeset
418
nick
parents:
diff changeset
419 " movd 48(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
420 " movd 60(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
421 " pfsub 52(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
422 " pfsub 56(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
423 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
424 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
425 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
426
nick
parents:
diff changeset
427 " pfadd 56(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
428 " pfadd 60(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
429 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
430
nick
parents:
diff changeset
431 " pfadd 48(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
432 " pfadd 52(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
433 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
434 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
435 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
436 " punpckldq %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
437 " movq %%mm1, 56(%%edx)\n\t"
nick
parents:
diff changeset
438 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
439
nick
parents:
diff changeset
440 /*---*/
nick
parents:
diff changeset
441
nick
parents:
diff changeset
442 " movd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
443 " pfsub 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
444 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
445 " movd %%mm1, 92(%%edx)\n\t"
nick
parents:
diff changeset
446 " pfadd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
447 " pfadd 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
448 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
449
nick
parents:
diff changeset
450 " pfadd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
451 " pfadd 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
452 " movd %%mm0, 80(%%edx)\n\t"
nick
parents:
diff changeset
453
nick
parents:
diff changeset
454 " movd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
455 " pfsub 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
456 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
457 " pfadd %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
458 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
459 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
460 " movq %%mm0, 84(%%edx)\n\t"
nick
parents:
diff changeset
461
nick
parents:
diff changeset
462 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
463 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
464 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
465 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
466 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
467 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
468
nick
parents:
diff changeset
469 " movd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
470 " pfsub 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
471 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
472 " movd %%mm0, 108(%%edx)\n\t"
nick
parents:
diff changeset
473 " pfadd 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
474 " pfadd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
475 " movd %%mm0, 104(%%edx)\n\t"
nick
parents:
diff changeset
476
nick
parents:
diff changeset
477 " movd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
478 " pfsub 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
479 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
480 " movd %%mm1, 124(%%edx)\n\t"
nick
parents:
diff changeset
481 " pfadd 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
482 " pfadd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
483 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
484
nick
parents:
diff changeset
485 " pfadd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
486 " pfadd 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
487 " movd %%mm0, 112(%%edx)\n\t"
nick
parents:
diff changeset
488
nick
parents:
diff changeset
489 " movd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
490 " pfsub 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
491 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
492 " pfadd %%mm0,%%mm1\n\t"
nick
parents:
diff changeset
493 " pfadd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
494 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
495 " movq %%mm0, 116(%%edx)\n\t"
nick
parents:
diff changeset
496
nick
parents:
diff changeset
497 " jnz .L01\n\t"
nick
parents:
diff changeset
498
nick
parents:
diff changeset
499 /* Phase 7*/
nick
parents:
diff changeset
500 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
501
nick
parents:
diff changeset
502 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
503 " pfadd 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
504 " movd %%mm0, 1024(%%esi)\n\t"
nick
parents:
diff changeset
505
nick
parents:
diff changeset
506 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
507 " pfsub 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
508 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
509 " movd %%mm0, (%%esi)\n\t"
nick
parents:
diff changeset
510 " movd %%mm0, (%%edi)\n\t"
nick
parents:
diff changeset
511
nick
parents:
diff changeset
512 " movd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
513 " pfsub 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
514 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
515 " movd %%mm0, 512(%%edi)\n\t"
nick
parents:
diff changeset
516 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
517 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
518 " movd %%mm0, 512(%%esi)\n\t"
nick
parents:
diff changeset
519
nick
parents:
diff changeset
520 " movd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
521 " pfsub 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
522 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
523 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
524
nick
parents:
diff changeset
525 " movd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
526 " pfsub 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
527 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
528 " movd %%mm0, 768(%%edi)\n\t"
nick
parents:
diff changeset
529 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
530
nick
parents:
diff changeset
531 " pfadd 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
532 " pfadd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
533 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
534
nick
parents:
diff changeset
535 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
536 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
537 " movd %%mm0, 768(%%esi)\n\t"
nick
parents:
diff changeset
538 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
539 " movd %%mm1, 256(%%esi)\n\t"
nick
parents:
diff changeset
540 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
541 " movd %%mm2, 256(%%edi)\n\t"
nick
parents:
diff changeset
542
nick
parents:
diff changeset
543 /* Phase 8*/
nick
parents:
diff changeset
544
nick
parents:
diff changeset
545 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
546 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
547 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
548 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
549 " movd %%mm0, 896(%%esi)\n\t"
nick
parents:
diff changeset
550 " movd %%mm1, 640(%%esi)\n\t"
nick
parents:
diff changeset
551 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
552 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
553 " movd %%mm0, 128(%%edi)\n\t"
nick
parents:
diff changeset
554 " movd %%mm1, 384(%%edi)\n\t"
nick
parents:
diff changeset
555
nick
parents:
diff changeset
556 " movd 40(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
557 " pfadd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
558 " movd %%mm0, 384(%%esi)\n\t"
nick
parents:
diff changeset
559
nick
parents:
diff changeset
560 " movd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
561 " pfadd 36(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
562 " movd %%mm0, 128(%%esi)\n\t"
nick
parents:
diff changeset
563
nick
parents:
diff changeset
564 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
565 " movd %%mm0, 896(%%edi)\n\t"
nick
parents:
diff changeset
566 " pfadd 44(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
567 " movd %%mm0, 640(%%edi)\n\t"
nick
parents:
diff changeset
568
nick
parents:
diff changeset
569 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
570 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
571 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
572 " pfadd 112(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
573 " pfadd 104(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
574 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
575 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
576 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
577 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
578 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
579 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
580 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
581 " movd %%mm0, 960(%%esi)\n\t"
nick
parents:
diff changeset
582 " movd %%mm2, 704(%%esi)\n\t"
nick
parents:
diff changeset
583 " movd %%mm4, 448(%%esi)\n\t"
nick
parents:
diff changeset
584 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
585 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
586 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
587 " movd %%mm0, 64(%%edi)\n\t"
nick
parents:
diff changeset
588 " movd %%mm2, 320(%%edi)\n\t"
nick
parents:
diff changeset
589 " movd %%mm4, 576(%%edi)\n\t"
nick
parents:
diff changeset
590 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
591 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
592 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
593 " movd %%mm1, 832(%%esi)\n\t"
nick
parents:
diff changeset
594 " movd %%mm3, 576(%%esi)\n\t"
nick
parents:
diff changeset
595 " movd %%mm5, 320(%%esi)\n\t"
nick
parents:
diff changeset
596 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
597 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
598 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
599 " movd %%mm1, 192(%%edi)\n\t"
nick
parents:
diff changeset
600 " movd %%mm3, 448(%%edi)\n\t"
nick
parents:
diff changeset
601 " movd %%mm5, 704(%%edi)\n\t"
nick
parents:
diff changeset
602
nick
parents:
diff changeset
603 " movd 120(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
604 " pfadd 100(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
605 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
606 " pfadd 88(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
607 " movd %%mm0, 192(%%esi)\n\t"
nick
parents:
diff changeset
608 " pfadd 68(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
609 " movd %%mm1, 64(%%esi)\n\t"
nick
parents:
diff changeset
610
nick
parents:
diff changeset
611 " movd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
612 " movd %%mm0, 960(%%edi)\n\t"
nick
parents:
diff changeset
613 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
614 " movd %%mm0, 832(%%edi)\n\t"
nick
parents:
diff changeset
615
nick
parents:
diff changeset
616 " jmp .L_bye\n\t"
nick
parents:
diff changeset
617 ".L01: \n\t"
nick
parents:
diff changeset
618 /* Phase 9*/
nick
parents:
diff changeset
619
nick
parents:
diff changeset
620 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
621 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
622 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
623 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
624 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
625 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
626 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
627 " movw %%ax, 512(%%esi)\n\t"
nick
parents:
diff changeset
628 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
629 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
630 " movw %%ax, (%%esi)\n\t"
nick
parents:
diff changeset
631
nick
parents:
diff changeset
632 " movd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
633 " pfsub 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
634 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
635 " pf2id %%mm0, %%mm7\n\t"
nick
parents:
diff changeset
636 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
637 " movw %%ax, 256(%%edi)\n\t"
nick
parents:
diff changeset
638 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
639 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
640 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
641 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
642 " movw %%ax, 256(%%esi)\n\t"
nick
parents:
diff changeset
643
nick
parents:
diff changeset
644 " movd 16(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
645 " pfsub 20(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
646 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
647 " movq %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
648
nick
parents:
diff changeset
649 " movd 28(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
650 " pfsub 24(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
651 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
652 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
653
nick
parents:
diff changeset
654 " pf2id %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
655 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
656 " movw %%ax, 384(%%edi)\n\t"
nick
parents:
diff changeset
657
nick
parents:
diff changeset
658 " pfadd 24(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
659 " pfadd 28(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
660 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
661
nick
parents:
diff changeset
662 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
663 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
664 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
665 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
666 " movw %%ax, 384(%%esi)\n\t"
nick
parents:
diff changeset
667 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
668 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
669 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
670 " movw %%ax, 128(%%esi)\n\t"
nick
parents:
diff changeset
671 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
672 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
673 " movd %%mm2, %%eax\n\t"
nick
parents:
diff changeset
674 " movw %%ax, 128(%%edi)\n\t"
nick
parents:
diff changeset
675
nick
parents:
diff changeset
676 /* Phase 10*/
nick
parents:
diff changeset
677
nick
parents:
diff changeset
678 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
679 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
680 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
681 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
682 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
683 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
684 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
685 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
686 " movw %%ax, 448(%%esi)\n\t"
nick
parents:
diff changeset
687 " movw %%cx, 320(%%esi)\n\t"
nick
parents:
diff changeset
688 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
689 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
690 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
691 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
692 " movw %%ax, 64(%%edi)\n\t"
nick
parents:
diff changeset
693 " movw %%cx, 192(%%edi)\n\t"
nick
parents:
diff changeset
694
nick
parents:
diff changeset
695 " movd 40(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
696 " movd 56(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
697 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
698 " movd 44(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
699 " movd 120(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
700 " punpckldq %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
701 " punpckldq 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
702 " pfadd 100(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
703 " punpckldq 36(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
704 " punpckldq 92(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
705 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
706 " pfadd %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
707 " pf2id %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
708 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
709 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
710 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
711 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
712 " movw %%ax, 448(%%edi)\n\t"
nick
parents:
diff changeset
713 " movw %%cx, 192(%%esi)\n\t"
nick
parents:
diff changeset
714 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
715 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
716 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
717 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
718 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
719 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
720 " movw %%bx, 96(%%esi)\n\t"
nick
parents:
diff changeset
721 " movw %%ax, 480(%%edi)\n\t"
nick
parents:
diff changeset
722 " movw %%cx, 64(%%esi)\n\t"
nick
parents:
diff changeset
723 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
724 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
725 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
726 " pfadd 68(%%edx), %%mm6\n\t"
nick
parents:
diff changeset
727 " movw %%ax, 320(%%edi)\n\t"
nick
parents:
diff changeset
728 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
729 " pf2id %%mm6, %%mm6\n\t"
nick
parents:
diff changeset
730 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
731 " movd %%mm6, %%ebx\n\t"
nick
parents:
diff changeset
732 " movw %%ax, 416(%%edi)\n\t"
nick
parents:
diff changeset
733 " movw %%bx, 32(%%esi)\n\t"
nick
parents:
diff changeset
734
nick
parents:
diff changeset
735 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
736 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
737 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
738 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
739 " pfadd %%mm4, %%mm2\n\t"
nick
parents:
diff changeset
740 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
741 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
742 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
743 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
744 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
745 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
746 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
747 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
748 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
749 " pf2id %%mm4, %%mm4\n\t"
nick
parents:
diff changeset
750 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
751 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
752 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
753 " movw %%ax, 480(%%esi)\n\t"
nick
parents:
diff changeset
754 " movw %%cx, 352(%%esi)\n\t"
nick
parents:
diff changeset
755 " movw %%bx, 224(%%esi)\n\t"
nick
parents:
diff changeset
756 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
757 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
758 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
759 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
760 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
761 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
762 " movw %%ax, 32(%%edi)\n\t"
nick
parents:
diff changeset
763 " movw %%cx, 160(%%edi)\n\t"
nick
parents:
diff changeset
764 " movw %%bx, 288(%%edi)\n\t"
nick
parents:
diff changeset
765 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
766 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
767 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
768 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
769 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
770 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
771 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
772 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
773 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
774 " movw %%ax, 416(%%esi)\n\t"
nick
parents:
diff changeset
775 " movw %%cx, 288(%%esi)\n\t"
nick
parents:
diff changeset
776 " movw %%bx, 160(%%esi)\n\t"
nick
parents:
diff changeset
777 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
778 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
779 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
780 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
781 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
782 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
783 " movw %%ax, 96(%%edi)\n\t"
nick
parents:
diff changeset
784 " movw %%cx, 224(%%edi)\n\t"
nick
parents:
diff changeset
785 " movw %%bx, 352(%%edi)\n\t"
nick
parents:
diff changeset
786
nick
parents:
diff changeset
787 " movsw\n\t"
nick
parents:
diff changeset
788
nick
parents:
diff changeset
789 ".L_bye:\n\t"
nick
parents:
diff changeset
790 " femms\n\t"
nick
parents:
diff changeset
791 :
nick
parents:
diff changeset
792 :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
nick
parents:
diff changeset
793 :"memory","%ebx","%esi","%edi");
5842
d6eab895c742 Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents: 5291
diff changeset
794 }