annotate mp3lib/dct64_3dnow.c @ 4294:21dbbbbd5479

a few filters (should be removed/merged when arpis videofilter stuff is finished) -ssf cgb=<float> -> chrominance gaussian blur (nice to improve picture quality) -ssf lgb=<float> -> luminance gaussian blur (dunno what its good for) -ssf chs=<int> -> shift chrominance horizontally -ssf cvs=<int> -> shift chrominance vertically these will only work if the swscaler is used at all
author michael
date Mon, 21 Jan 2002 15:22:28 +0000
parents 3b29772a4fb2
children 421969d55d5f
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 3dnow! 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
nick
parents:
diff changeset
12 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
nick
parents:
diff changeset
13 static float plus_1f = 1.0;
nick
parents:
diff changeset
14
nick
parents:
diff changeset
15 void __attribute__ (( __stdcall__ )) dct64_MMX_3dnow(real *a,real *b,real *c)
nick
parents:
diff changeset
16 {
nick
parents:
diff changeset
17 char tmp[256];
nick
parents:
diff changeset
18 __asm __volatile(
nick
parents:
diff changeset
19 " movl %2,%%eax\n\t"
nick
parents:
diff changeset
20
nick
parents:
diff changeset
21 " leal 128+%3,%%edx\n\t"
nick
parents:
diff changeset
22 " movl %0,%%esi\n\t"
nick
parents:
diff changeset
23 " movl %1,%%edi\n\t"
nick
parents:
diff changeset
24 " movl $costab_mmx,%%ebx\n\t"
nick
parents:
diff changeset
25 " leal %3,%%ecx\n\t"
nick
parents:
diff changeset
26
nick
parents:
diff changeset
27 /* Phase 1*/
nick
parents:
diff changeset
28 " movq (%%eax), %%mm0\n\t"
nick
parents:
diff changeset
29 " movq 8(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
30 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
31 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
32 " movq 120(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
33 " movq 112(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
34 /* n.b.: pswapd*/
nick
parents:
diff changeset
35 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
36 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
37 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
38 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
39 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
40 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
41 /**/
nick
parents:
diff changeset
42 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
43 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
44 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
45 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
46 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
47 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
48 " pfmul (%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
49 " pfmul 8(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
50 " movd %%mm3, 124(%%edx)\n\t"
nick
parents:
diff changeset
51 " movd %%mm7, 116(%%edx)\n\t"
nick
parents:
diff changeset
52 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
53 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
54 " movd %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
55 " movd %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
56
nick
parents:
diff changeset
57 " movq 16(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
58 " movq 24(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
59 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
60 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
61 " movq 104(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
62 " movq 96(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
63 /* n.b.: pswapd*/
nick
parents:
diff changeset
64 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
65 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
66 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
67 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
68 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
69 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
70 /**/
nick
parents:
diff changeset
71 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
72 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
73 " movq %%mm0, 16(%%edx)\n\t"
nick
parents:
diff changeset
74 " movq %%mm4, 24(%%edx)\n\t"
nick
parents:
diff changeset
75 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
76 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
77 " pfmul 16(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
78 " pfmul 24(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
79 " movd %%mm3, 108(%%edx)\n\t"
nick
parents:
diff changeset
80 " movd %%mm7, 100(%%edx)\n\t"
nick
parents:
diff changeset
81 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
82 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
83 " movd %%mm3, 104(%%edx)\n\t"
nick
parents:
diff changeset
84 " movd %%mm7, 96(%%edx)\n\t"
nick
parents:
diff changeset
85
nick
parents:
diff changeset
86 " movq 32(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
87 " movq 40(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
88 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
89 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
90 " movq 88(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
91 " movq 80(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
92 /* n.b.: pswapd*/
nick
parents:
diff changeset
93 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
94 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
95 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
96 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
97 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
98 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
99 /**/
nick
parents:
diff changeset
100 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
101 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
102 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
103 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
104 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
105 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
106 " pfmul 32(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
107 " pfmul 40(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
108 " movd %%mm3, 92(%%edx)\n\t"
nick
parents:
diff changeset
109 " movd %%mm7, 84(%%edx)\n\t"
nick
parents:
diff changeset
110 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
111 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
112 " movd %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
113 " movd %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
114
nick
parents:
diff changeset
115 " movq 48(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
116 " movq 56(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
117 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
118 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
119 " movq 72(%%eax), %%mm1\n\t"
nick
parents:
diff changeset
120 " movq 64(%%eax), %%mm5\n\t"
nick
parents:
diff changeset
121 /* n.b.: pswapd*/
nick
parents:
diff changeset
122 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
123 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
124 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
125 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
126 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
127 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
128 /**/
nick
parents:
diff changeset
129 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
130 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
131 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
132 " movq %%mm4, 56(%%edx)\n\t"
nick
parents:
diff changeset
133 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
134 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
135 " pfmul 48(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
136 " pfmul 56(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
137 " movd %%mm3, 76(%%edx)\n\t"
nick
parents:
diff changeset
138 " movd %%mm7, 68(%%edx)\n\t"
nick
parents:
diff changeset
139 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
140 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
141 " movd %%mm3, 72(%%edx)\n\t"
nick
parents:
diff changeset
142 " movd %%mm7, 64(%%edx)\n\t"
nick
parents:
diff changeset
143
nick
parents:
diff changeset
144 /* Phase 2*/
nick
parents:
diff changeset
145
nick
parents:
diff changeset
146 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
147 " movq 8(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
148 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
149 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
150 " movq 56(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
151 " movq 48(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
152 /* n.b.: pswapd*/
nick
parents:
diff changeset
153 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
154 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
155 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
156 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
157 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
158 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
159 /**/
nick
parents:
diff changeset
160 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
161 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
162 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
163 " movq %%mm4, 8(%%ecx)\n\t"
nick
parents:
diff changeset
164 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
165 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
166 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
167 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
168 " movd %%mm3, 60(%%ecx)\n\t"
nick
parents:
diff changeset
169 " movd %%mm7, 52(%%ecx)\n\t"
nick
parents:
diff changeset
170 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
171 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
172 " movd %%mm3, 56(%%ecx)\n\t"
nick
parents:
diff changeset
173 " movd %%mm7, 48(%%ecx)\n\t"
nick
parents:
diff changeset
174
nick
parents:
diff changeset
175 " movq 16(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
176 " movq 24(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
177 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
178 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
179 " movq 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
180 " movq 32(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
181 /* n.b.: pswapd*/
nick
parents:
diff changeset
182 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
183 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
184 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
185 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
186 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
187 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
188 /**/
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, 16(%%ecx)\n\t"
nick
parents:
diff changeset
192 " movq %%mm4, 24(%%ecx)\n\t"
nick
parents:
diff changeset
193 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
194 " pfsub %%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 " movd %%mm3, 44(%%ecx)\n\t"
nick
parents:
diff changeset
198 " movd %%mm7, 36(%%ecx)\n\t"
nick
parents:
diff changeset
199 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
200 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
201 " movd %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
202 " movd %%mm7, 32(%%ecx)\n\t"
nick
parents:
diff changeset
203
nick
parents:
diff changeset
204 /* Phase 3*/
nick
parents:
diff changeset
205
nick
parents:
diff changeset
206 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
207 " movq 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
208 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
209 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
210 " movq 120(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
211 " movq 112(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
212 /* n.b.: pswapd*/
nick
parents:
diff changeset
213 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
214 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
215 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
216 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
217 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
218 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
219 /**/
nick
parents:
diff changeset
220 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
221 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
222 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
223 " movq %%mm4, 72(%%ecx)\n\t"
nick
parents:
diff changeset
224 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
225 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
226 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
227 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
228 " movd %%mm3, 124(%%ecx)\n\t"
nick
parents:
diff changeset
229 " movd %%mm7, 116(%%ecx)\n\t"
nick
parents:
diff changeset
230 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
231 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
232 " movd %%mm3, 120(%%ecx)\n\t"
nick
parents:
diff changeset
233 " movd %%mm7, 112(%%ecx)\n\t"
nick
parents:
diff changeset
234
nick
parents:
diff changeset
235 " movq 80(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
236 " movq 88(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
237 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
238 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
239 " movq 104(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
240 " movq 96(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
241 /* n.b.: pswapd*/
nick
parents:
diff changeset
242 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
243 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
244 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
245 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
246 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
247 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
248 /**/
nick
parents:
diff changeset
249 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
250 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
251 " movq %%mm0, 80(%%ecx)\n\t"
nick
parents:
diff changeset
252 " movq %%mm4, 88(%%ecx)\n\t"
nick
parents:
diff changeset
253 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
254 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
255 " pfmul 80(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
256 " pfmul 88(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
257 " movd %%mm3, 108(%%ecx)\n\t"
nick
parents:
diff changeset
258 " movd %%mm7, 100(%%ecx)\n\t"
nick
parents:
diff changeset
259 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
260 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
261 " movd %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
262 " movd %%mm7, 96(%%ecx)\n\t"
nick
parents:
diff changeset
263
nick
parents:
diff changeset
264 /* Phase 4*/
nick
parents:
diff changeset
265
nick
parents:
diff changeset
266 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
267 " movq 8(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
268 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
269 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
270 " movq 24(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
271 " movq 16(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
272 /* n.b.: pswapd*/
nick
parents:
diff changeset
273 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
274 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
275 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
276 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
277 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
278 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
279 /**/
nick
parents:
diff changeset
280 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
281 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
282 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
283 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
284 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
285 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
286 " pfmul 96(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
287 " pfmul 104(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
288 " movd %%mm3, 28(%%edx)\n\t"
nick
parents:
diff changeset
289 " movd %%mm7, 20(%%edx)\n\t"
nick
parents:
diff changeset
290 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
291 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
292 " movd %%mm3, 24(%%edx)\n\t"
nick
parents:
diff changeset
293 " movd %%mm7, 16(%%edx)\n\t"
nick
parents:
diff changeset
294
nick
parents:
diff changeset
295 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
296 " movq 40(%%ecx), %%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 56(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
300 " movq 48(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
301 /* n.b.: pswapd*/
nick
parents:
diff changeset
302 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
303 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
304 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
305 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
306 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
307 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
308 /**/
nick
parents:
diff changeset
309 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
310 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
311 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
312 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
313 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
314 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
315 " pfmul 96(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
316 " pfmul 104(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
317 " movd %%mm3, 60(%%edx)\n\t"
nick
parents:
diff changeset
318 " movd %%mm7, 52(%%edx)\n\t"
nick
parents:
diff changeset
319 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
320 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
321 " movd %%mm3, 56(%%edx)\n\t"
nick
parents:
diff changeset
322 " movd %%mm7, 48(%%edx)\n\t"
nick
parents:
diff changeset
323
nick
parents:
diff changeset
324 " movq 64(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
325 " movq 72(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
326 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
327 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
328 " movq 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
329 " movq 80(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
330 /* n.b.: pswapd*/
nick
parents:
diff changeset
331 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
332 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
333 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
334 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
335 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
336 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
337 /**/
nick
parents:
diff changeset
338 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
339 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
340 " movq %%mm0, 64(%%edx)\n\t"
nick
parents:
diff changeset
341 " movq %%mm4, 72(%%edx)\n\t"
nick
parents:
diff changeset
342 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
343 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
344 " pfmul 96(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
345 " pfmul 104(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
346 " movd %%mm3, 92(%%edx)\n\t"
nick
parents:
diff changeset
347 " movd %%mm7, 84(%%edx)\n\t"
nick
parents:
diff changeset
348 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
349 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
350 " movd %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
351 " movd %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
352
nick
parents:
diff changeset
353 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
354 " movq 104(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
355 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
356 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
357 " movq 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
358 " movq 112(%%ecx), %%mm5\n\t"
nick
parents:
diff changeset
359 /* n.b.: pswapd*/
nick
parents:
diff changeset
360 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
361 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
362 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
363 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
364 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
365 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
366 /**/
nick
parents:
diff changeset
367 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
368 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
369 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
370 " movq %%mm4, 104(%%edx)\n\t"
nick
parents:
diff changeset
371 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
372 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
373 " pfmul 96(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
374 " pfmul 104(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
375 " movd %%mm3, 124(%%edx)\n\t"
nick
parents:
diff changeset
376 " movd %%mm7, 116(%%edx)\n\t"
nick
parents:
diff changeset
377 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
378 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
379 " movd %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
380 " movd %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
381
nick
parents:
diff changeset
382 /* Phase 5 */
nick
parents:
diff changeset
383
nick
parents:
diff changeset
384 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
385 " movq 16(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
386 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
387 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
388 " movq 8(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
389 " movq 24(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
390 /* n.b.: pswapd*/
nick
parents:
diff changeset
391 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
392 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
393 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
394 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
395 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
396 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
397 /**/
nick
parents:
diff changeset
398 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
399 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
400 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
401 " movq %%mm4, 16(%%ecx)\n\t"
nick
parents:
diff changeset
402 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
403 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
404 " pfmul 112(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
405 " pfmul 112(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
406 " movd %%mm3, 12(%%ecx)\n\t"
nick
parents:
diff changeset
407 " movd %%mm7, 28(%%ecx)\n\t"
nick
parents:
diff changeset
408 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
409 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
410 " movd %%mm3, 8(%%ecx)\n\t"
nick
parents:
diff changeset
411 " movd %%mm7, 24(%%ecx)\n\t"
nick
parents:
diff changeset
412
nick
parents:
diff changeset
413 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
414 " movq 48(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
415 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
416 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
417 " movq 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
418 " movq 56(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
419 /* n.b.: pswapd*/
nick
parents:
diff changeset
420 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
421 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
422 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
423 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
424 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
425 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
426 /**/
nick
parents:
diff changeset
427 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
428 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
429 " movq %%mm0, 32(%%ecx)\n\t"
nick
parents:
diff changeset
430 " movq %%mm4, 48(%%ecx)\n\t"
nick
parents:
diff changeset
431 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
432 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
433 " pfmul 112(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
434 " pfmul 112(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
435 " movd %%mm3, 44(%%ecx)\n\t"
nick
parents:
diff changeset
436 " movd %%mm7, 60(%%ecx)\n\t"
nick
parents:
diff changeset
437 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
438 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
439 " movd %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
440 " movd %%mm7, 56(%%ecx)\n\t"
nick
parents:
diff changeset
441
nick
parents:
diff changeset
442 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
443 " movq 80(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
444 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
445 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
446 " movq 72(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
447 " movq 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
448 /* n.b.: pswapd*/
nick
parents:
diff changeset
449 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
450 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
451 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
452 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
453 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
454 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
455 /**/
nick
parents:
diff changeset
456 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
457 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
458 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
459 " movq %%mm4, 80(%%ecx)\n\t"
nick
parents:
diff changeset
460 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
461 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
462 " pfmul 112(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
463 " pfmul 112(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
464 " movd %%mm3, 76(%%ecx)\n\t"
nick
parents:
diff changeset
465 " movd %%mm7, 92(%%ecx)\n\t"
nick
parents:
diff changeset
466 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
467 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
468 " movd %%mm3, 72(%%ecx)\n\t"
nick
parents:
diff changeset
469 " movd %%mm7, 88(%%ecx)\n\t"
nick
parents:
diff changeset
470
nick
parents:
diff changeset
471 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
472 " movq 112(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
473 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
474 " movq %%mm4, %%mm7\n\t"
nick
parents:
diff changeset
475 " movq 104(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
476 " movq 120(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
477 /* n.b.: pswapd*/
nick
parents:
diff changeset
478 " movq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
479 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
480 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
481 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
482 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
483 " punpckldq %%mm6, %%mm5\n\t"
nick
parents:
diff changeset
484 /**/
nick
parents:
diff changeset
485 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
486 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
487 " movq %%mm0, 96(%%ecx)\n\t"
nick
parents:
diff changeset
488 " movq %%mm4, 112(%%ecx)\n\t"
nick
parents:
diff changeset
489 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
490 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
491 " pfmul 112(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
492 " pfmul 112(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
493 " movd %%mm3, 108(%%ecx)\n\t"
nick
parents:
diff changeset
494 " movd %%mm7, 124(%%ecx)\n\t"
nick
parents:
diff changeset
495 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
496 " psrlq $32, %%mm7\n\t"
nick
parents:
diff changeset
497 " movd %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
498 " movd %%mm7, 120(%%ecx)\n\t"
nick
parents:
diff changeset
499
nick
parents:
diff changeset
500 /* Phase 6. This is the end of easy road. */
nick
parents:
diff changeset
501 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
502
nick
parents:
diff changeset
503 " movd plus_1f, %%mm6\n\t"
nick
parents:
diff changeset
504 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/
nick
parents:
diff changeset
505 " movq x_plus_minus_3dnow, %%mm7\n\t" /* mm7 = +1 | -1 */
nick
parents:
diff changeset
506
nick
parents:
diff changeset
507 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
508 " movq 64(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
509 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
510 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
511 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
512 " pxor %%mm7, %%mm3\n\t"
nick
parents:
diff changeset
513 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
514 " pfacc %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
515 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
516 " pfmul %%mm6, %%mm2\n\t"
nick
parents:
diff changeset
517 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
518 " movq %%mm2, 64(%%edx)\n\t"
nick
parents:
diff changeset
519
nick
parents:
diff changeset
520 " movd 44(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
521 " movd 40(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
522 " movd 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
523 " punpckldq 76(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
524 " punpckldq 72(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
525 " punpckldq %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
526 " movq %%mm0, %%mm4\n\t"
nick
parents:
diff changeset
527 " movq %%mm2, %%mm5\n\t"
nick
parents:
diff changeset
528 " pfsub %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
529 " pfmul %%mm3, %%mm0\n\t"
nick
parents:
diff changeset
530 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
531 " pfadd %%mm5, %%mm0\n\t"
nick
parents:
diff changeset
532 " pfadd %%mm4, %%mm0\n\t"
nick
parents:
diff changeset
533 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
534 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
535 " punpckhdq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
536 " movq %%mm0, 40(%%edx)\n\t"
nick
parents:
diff changeset
537 " movq %%mm2, 72(%%edx)\n\t"
nick
parents:
diff changeset
538
nick
parents:
diff changeset
539 " movd 48(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
540 " movd 60(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
541 " pfsub 52(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
542 " pfsub 56(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
543 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
544 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
545 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
546
nick
parents:
diff changeset
547 " pfadd 56(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
548 " pfadd 60(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
549 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
550
nick
parents:
diff changeset
551 " pfadd 48(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
552 " pfadd 52(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
553 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
554 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
555 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
556 " punpckldq %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
557 " movq %%mm1, 56(%%edx)\n\t"
nick
parents:
diff changeset
558 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
559
nick
parents:
diff changeset
560 /*---*/
nick
parents:
diff changeset
561
nick
parents:
diff changeset
562 " movd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
563 " pfsub 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
564 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
565 " movd %%mm1, 92(%%edx)\n\t"
nick
parents:
diff changeset
566 " pfadd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
567 " pfadd 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
568 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
569
nick
parents:
diff changeset
570 " pfadd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
571 " pfadd 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
572 " movd %%mm0, 80(%%edx)\n\t"
nick
parents:
diff changeset
573
nick
parents:
diff changeset
574 " movd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
575 " pfsub 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
576 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
577 " pfadd %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
578 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
579 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
580 " movq %%mm0, 84(%%edx)\n\t"
nick
parents:
diff changeset
581
nick
parents:
diff changeset
582 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
583 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
584 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
585 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
586 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
587 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
588
nick
parents:
diff changeset
589 " movd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
590 " pfsub 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
591 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
592 " movd %%mm0, 108(%%edx)\n\t"
nick
parents:
diff changeset
593 " pfadd 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
594 " pfadd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
595 " movd %%mm0, 104(%%edx)\n\t"
nick
parents:
diff changeset
596
nick
parents:
diff changeset
597 " movd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
598 " pfsub 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
599 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
600 " movd %%mm1, 124(%%edx)\n\t"
nick
parents:
diff changeset
601 " pfadd 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
602 " pfadd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
603 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
604
nick
parents:
diff changeset
605 " pfadd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
606 " pfadd 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
607 " movd %%mm0, 112(%%edx)\n\t"
nick
parents:
diff changeset
608
nick
parents:
diff changeset
609 " movd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
610 " pfsub 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
611 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
612 " pfadd %%mm0,%%mm1\n\t"
nick
parents:
diff changeset
613 " pfadd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
614 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
615 " movq %%mm0, 116(%%edx)\n\t"
nick
parents:
diff changeset
616
nick
parents:
diff changeset
617 " jnz .L01\n\t"
nick
parents:
diff changeset
618
nick
parents:
diff changeset
619 /* Phase 7*/
nick
parents:
diff changeset
620 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
621
nick
parents:
diff changeset
622 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
623 " pfadd 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
624 " movd %%mm0, 1024(%%esi)\n\t"
nick
parents:
diff changeset
625
nick
parents:
diff changeset
626 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
627 " pfsub 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
628 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
629 " movd %%mm0, (%%esi)\n\t"
nick
parents:
diff changeset
630 " movd %%mm0, (%%edi)\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 " movd %%mm0, 512(%%edi)\n\t"
nick
parents:
diff changeset
636 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
637 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
638 " movd %%mm0, 512(%%esi)\n\t"
nick
parents:
diff changeset
639
nick
parents:
diff changeset
640 " movd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
641 " pfsub 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
642 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
643 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
644
nick
parents:
diff changeset
645 " movd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
646 " pfsub 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
647 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
648 " movd %%mm0, 768(%%edi)\n\t"
nick
parents:
diff changeset
649 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
650
nick
parents:
diff changeset
651 " pfadd 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
652 " pfadd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
653 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
654
nick
parents:
diff changeset
655 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
656 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
657 " movd %%mm0, 768(%%esi)\n\t"
nick
parents:
diff changeset
658 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
659 " movd %%mm1, 256(%%esi)\n\t"
nick
parents:
diff changeset
660 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
661 " movd %%mm2, 256(%%edi)\n\t"
nick
parents:
diff changeset
662
nick
parents:
diff changeset
663 /* Phase 8*/
nick
parents:
diff changeset
664
nick
parents:
diff changeset
665 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
666 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
667 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
668 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
669 " movd %%mm0, 896(%%esi)\n\t"
nick
parents:
diff changeset
670 " movd %%mm1, 640(%%esi)\n\t"
nick
parents:
diff changeset
671 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
672 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
673 " movd %%mm0, 128(%%edi)\n\t"
nick
parents:
diff changeset
674 " movd %%mm1, 384(%%edi)\n\t"
nick
parents:
diff changeset
675
nick
parents:
diff changeset
676 " movd 40(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
677 " pfadd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
678 " movd %%mm0, 384(%%esi)\n\t"
nick
parents:
diff changeset
679
nick
parents:
diff changeset
680 " movd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
681 " pfadd 36(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
682 " movd %%mm0, 128(%%esi)\n\t"
nick
parents:
diff changeset
683
nick
parents:
diff changeset
684 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
685 " movd %%mm0, 896(%%edi)\n\t"
nick
parents:
diff changeset
686 " pfadd 44(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
687 " movd %%mm0, 640(%%edi)\n\t"
nick
parents:
diff changeset
688
nick
parents:
diff changeset
689 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
690 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
691 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
692 " pfadd 112(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
693 " pfadd 104(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
694 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
695 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
696 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
697 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
698 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
699 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
700 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
701 " movd %%mm0, 960(%%esi)\n\t"
nick
parents:
diff changeset
702 " movd %%mm2, 704(%%esi)\n\t"
nick
parents:
diff changeset
703 " movd %%mm4, 448(%%esi)\n\t"
nick
parents:
diff changeset
704 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
705 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
706 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
707 " movd %%mm0, 64(%%edi)\n\t"
nick
parents:
diff changeset
708 " movd %%mm2, 320(%%edi)\n\t"
nick
parents:
diff changeset
709 " movd %%mm4, 576(%%edi)\n\t"
nick
parents:
diff changeset
710 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
711 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
712 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
713 " movd %%mm1, 832(%%esi)\n\t"
nick
parents:
diff changeset
714 " movd %%mm3, 576(%%esi)\n\t"
nick
parents:
diff changeset
715 " movd %%mm5, 320(%%esi)\n\t"
nick
parents:
diff changeset
716 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
717 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
718 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
719 " movd %%mm1, 192(%%edi)\n\t"
nick
parents:
diff changeset
720 " movd %%mm3, 448(%%edi)\n\t"
nick
parents:
diff changeset
721 " movd %%mm5, 704(%%edi)\n\t"
nick
parents:
diff changeset
722
nick
parents:
diff changeset
723 " movd 120(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
724 " pfadd 100(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
725 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
726 " pfadd 88(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
727 " movd %%mm0, 192(%%esi)\n\t"
nick
parents:
diff changeset
728 " pfadd 68(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
729 " movd %%mm1, 64(%%esi)\n\t"
nick
parents:
diff changeset
730
nick
parents:
diff changeset
731 " movd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
732 " movd %%mm0, 960(%%edi)\n\t"
nick
parents:
diff changeset
733 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
734 " movd %%mm0, 832(%%edi)\n\t"
nick
parents:
diff changeset
735
nick
parents:
diff changeset
736 " jmp .L_bye\n\t"
nick
parents:
diff changeset
737 ".L01:\n\t"
nick
parents:
diff changeset
738 /* Phase 9*/
nick
parents:
diff changeset
739
nick
parents:
diff changeset
740 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
741 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
742 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
743 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
744 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
745 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
746 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
747 " movw %%ax, 512(%%esi)\n\t"
nick
parents:
diff changeset
748 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
749 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
750 " movw %%ax, (%%esi)\n\t"
nick
parents:
diff changeset
751
nick
parents:
diff changeset
752 " movd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
753 " pfsub 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
754 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
755 " pf2id %%mm0, %%mm7\n\t"
nick
parents:
diff changeset
756 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
757 " movw %%ax, 256(%%edi)\n\t"
nick
parents:
diff changeset
758 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
759 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
760 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
761 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
762 " movw %%ax, 256(%%esi)\n\t"
nick
parents:
diff changeset
763
nick
parents:
diff changeset
764 " movd 16(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
765 " pfsub 20(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
766 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
767 " movq %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
768
nick
parents:
diff changeset
769 " movd 28(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
770 " pfsub 24(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
771 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
772 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
773
nick
parents:
diff changeset
774 " pf2id %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
775 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
776 " movw %%ax, 384(%%edi)\n\t"
nick
parents:
diff changeset
777
nick
parents:
diff changeset
778 " pfadd 24(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
779 " pfadd 28(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
780 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
781
nick
parents:
diff changeset
782 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
783 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
784 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
785 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
786 " movw %%ax, 384(%%esi)\n\t"
nick
parents:
diff changeset
787 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
788 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
789 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
790 " movw %%ax, 128(%%esi)\n\t"
nick
parents:
diff changeset
791 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
792 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
793 " movd %%mm2, %%eax\n\t"
nick
parents:
diff changeset
794 " movw %%ax, 128(%%edi)\n\t"
nick
parents:
diff changeset
795
nick
parents:
diff changeset
796 /* Phase 10*/
nick
parents:
diff changeset
797
nick
parents:
diff changeset
798 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
799 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
800 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
801 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
802 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
803 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
804 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
805 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
806 " movw %%ax, 448(%%esi)\n\t"
nick
parents:
diff changeset
807 " movw %%cx, 320(%%esi)\n\t"
nick
parents:
diff changeset
808 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
809 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
810 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
811 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
812 " movw %%ax, 64(%%edi)\n\t"
nick
parents:
diff changeset
813 " movw %%cx, 192(%%edi)\n\t"
nick
parents:
diff changeset
814
nick
parents:
diff changeset
815 " movd 40(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
816 " movd 56(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
817 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
818 " movd 44(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
819 " movd 120(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
820 " punpckldq %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
821 " punpckldq 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
822 " pfadd 100(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
823 " punpckldq 36(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
824 " punpckldq 92(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
825 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
826 " pfadd %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
827 " pf2id %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
828 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
829 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
830 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
831 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
832 " movw %%ax, 448(%%edi)\n\t"
nick
parents:
diff changeset
833 " movw %%cx, 192(%%esi)\n\t"
nick
parents:
diff changeset
834 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
835 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
836 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
837 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
838 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
839 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
840 " movw %%bx, 96(%%esi)\n\t"
nick
parents:
diff changeset
841 " movw %%ax, 480(%%edi)\n\t"
nick
parents:
diff changeset
842 " movw %%cx, 64(%%esi)\n\t"
nick
parents:
diff changeset
843 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
844 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
845 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
846 " pfadd 68(%%edx), %%mm6\n\t"
nick
parents:
diff changeset
847 " movw %%ax, 320(%%edi)\n\t"
nick
parents:
diff changeset
848 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
849 " pf2id %%mm6, %%mm6\n\t"
nick
parents:
diff changeset
850 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
851 " movd %%mm6, %%ebx\n\t"
nick
parents:
diff changeset
852 " movw %%ax, 416(%%edi)\n\t"
nick
parents:
diff changeset
853 " movw %%bx, 32(%%esi)\n\t"
nick
parents:
diff changeset
854
nick
parents:
diff changeset
855 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
856 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
857 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
858 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
859 " pfadd %%mm4, %%mm2\n\t"
nick
parents:
diff changeset
860 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
861 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
862 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
863 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
864 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
865 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
866 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
867 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
868 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
869 " pf2id %%mm4, %%mm4\n\t"
nick
parents:
diff changeset
870 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
871 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
872 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
873 " movw %%ax, 480(%%esi)\n\t"
nick
parents:
diff changeset
874 " movw %%cx, 352(%%esi)\n\t"
nick
parents:
diff changeset
875 " movw %%bx, 224(%%esi)\n\t"
nick
parents:
diff changeset
876 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
877 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
878 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
879 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
880 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
881 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
882 " movw %%ax, 32(%%edi)\n\t"
nick
parents:
diff changeset
883 " movw %%cx, 160(%%edi)\n\t"
nick
parents:
diff changeset
884 " movw %%bx, 288(%%edi)\n\t"
nick
parents:
diff changeset
885 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
886 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
887 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
888 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
889 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
890 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
891 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
892 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
893 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
894 " movw %%ax, 416(%%esi)\n\t"
nick
parents:
diff changeset
895 " movw %%cx, 288(%%esi)\n\t"
nick
parents:
diff changeset
896 " movw %%bx, 160(%%esi)\n\t"
nick
parents:
diff changeset
897 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
898 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
899 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
900 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
901 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
902 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
903 " movw %%ax, 96(%%edi)\n\t"
nick
parents:
diff changeset
904 " movw %%cx, 224(%%edi)\n\t"
nick
parents:
diff changeset
905 " movw %%bx, 352(%%edi)\n\t"
nick
parents:
diff changeset
906
nick
parents:
diff changeset
907 " movsw\n\t"
nick
parents:
diff changeset
908
nick
parents:
diff changeset
909 ".L_bye:\n\t"
nick
parents:
diff changeset
910 " femms\n\t"
nick
parents:
diff changeset
911 :
nick
parents:
diff changeset
912 :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
nick
parents:
diff changeset
913 :"memory","%ebx","%esi","%edi");
nick
parents:
diff changeset
914 }