Parent Directory | Revision Log
Revision 1743 -
(view)
(download)
Original Path: trunk/xvidcore/src/plugins/x86_asm/plugin_ssim-a.asm
1 : | Skal | 1730 | ;/**************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - optimized SSIM routines - | ||
5 : | ; * | ||
6 : | ; * Copyright(C) 2006 Johannes Reinhardt <johannes.reinhardt@gmx.de> | ||
7 : | ; * | ||
8 : | ; * This program is free software; you can redistribute it and/or modify it | ||
9 : | ; * under the terms of the GNU General Public License as published by | ||
10 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
11 : | ; * (at your option) any later version. | ||
12 : | ; * | ||
13 : | ; * This program is distributed in the hope that it will be useful, | ||
14 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
16 : | ; * GNU General Public License for more details. | ||
17 : | ; * | ||
18 : | ; * You should have received a copy of the GNU General Public License | ||
19 : | ; * along with this program; if not, write to the Free Software | ||
20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | ; * | ||
22 : | ; * | ||
23 : | ; ***************************************************************************/ | ||
24 : | |||
25 : | BITS 32 | ||
26 : | |||
27 : | %macro cglobal 1 | ||
28 : | %ifdef PREFIX | ||
29 : | %ifdef MARK_FUNCS | ||
30 : | global _%1:function %1.endfunc-%1 | ||
31 : | %define %1 _%1:function %1.endfunc-%1 | ||
32 : | %else | ||
33 : | global _%1 | ||
34 : | %define %1 _%1 | ||
35 : | %endif | ||
36 : | %else | ||
37 : | %ifdef MARK_FUNCS | ||
38 : | global %1:function %1.endfunc-%1 | ||
39 : | %else | ||
40 : | global %1 | ||
41 : | %endif | ||
42 : | %endif | ||
43 : | %endmacro | ||
44 : | |||
45 : | Skal | 1740 | %macro ACC_ROW 2 |
46 : | movq %1,[ ecx] | ||
47 : | movq %2,[ecx+edx] | ||
48 : | Skal | 1730 | psadbw %1,mm0 |
49 : | Skal | 1740 | psadbw %2,mm0 |
50 : | lea ecx, [ecx+2*edx] | ||
51 : | paddw %1, %2 | ||
52 : | Skal | 1730 | %endmacro |
53 : | |||
54 : | %macro CONSIM_1x8_SSE2 0 | ||
55 : | Skal | 1743 | movdqu xmm0,[ecx] |
56 : | movdqu xmm1,[edx] | ||
57 : | Skal | 1730 | |
58 : | ;unpack to words | ||
59 : | punpcklbw xmm0,xmm2 | ||
60 : | punpcklbw xmm1,xmm2 | ||
61 : | |||
62 : | Skal | 1743 | movaps xmm3,xmm0 |
63 : | movaps xmm4,xmm1 | ||
64 : | Skal | 1730 | |
65 : | Skal | 1743 | pmaddwd xmm0,xmm0;orig |
66 : | pmaddwd xmm1,xmm1;comp | ||
67 : | pmaddwd xmm3,xmm4;corr | ||
68 : | Skal | 1730 | |
69 : | Skal | 1743 | paddd xmm5,xmm0 |
70 : | paddd xmm6,xmm1 | ||
71 : | paddd xmm7,xmm3 | ||
72 : | Skal | 1730 | %endmacro |
73 : | |||
74 : | %macro CONSIM_1x8_MMX 0 | ||
75 : | Skal | 1743 | movq mm0,[ecx];orig |
76 : | movq mm1,[edx];comp | ||
77 : | Skal | 1730 | |
78 : | Skal | 1743 | ;unpack low half of qw to words |
79 : | punpcklbw mm0,mm2 | ||
80 : | punpcklbw mm1,mm2 | ||
81 : | Skal | 1730 | |
82 : | Skal | 1743 | movq mm3,mm0 |
83 : | pmaddwd mm3,mm0 | ||
84 : | paddd mm5,mm3; | ||
85 : | Skal | 1730 | |
86 : | Skal | 1743 | movq mm4,mm1 |
87 : | pmaddwd mm4,mm1 | ||
88 : | paddd mm6,mm4; | ||
89 : | |||
90 : | Skal | 1730 | pmaddwd mm1,mm0 |
91 : | Skal | 1743 | paddd mm7,mm1 |
92 : | Skal | 1730 | |
93 : | Skal | 1743 | movq mm0,[ecx];orig |
94 : | movq mm1,[edx];comp | ||
95 : | Skal | 1730 | |
96 : | Skal | 1743 | ;unpack high half of qw to words |
97 : | punpckhbw mm0,mm2 | ||
98 : | punpckhbw mm1,mm2 | ||
99 : | Skal | 1730 | |
100 : | Skal | 1743 | movq mm3,mm0 |
101 : | pmaddwd mm3,mm0 | ||
102 : | paddd mm5,mm3; | ||
103 : | Skal | 1730 | |
104 : | Skal | 1743 | movq mm4,mm1 |
105 : | pmaddwd mm4,mm1 | ||
106 : | paddd mm6,mm4; | ||
107 : | |||
108 : | Skal | 1730 | pmaddwd mm1,mm0 |
109 : | Skal | 1743 | paddd mm7,mm1 |
110 : | Skal | 1730 | %endmacro |
111 : | |||
112 : | Skal | 1743 | %macro CONSIM_WRITEOUT 3 |
113 : | mov eax,[esp + 16];lumo | ||
114 : | mul eax; lumo^2 | ||
115 : | shr eax,6; 64*lum0^2 | ||
116 : | movd ecx,%1 | ||
117 : | sub ecx,eax | ||
118 : | Skal | 1730 | |
119 : | Skal | 1743 | mov edx,[esp + 24]; pdevo |
120 : | mov [edx],ecx | ||
121 : | Skal | 1730 | |
122 : | Skal | 1743 | mov eax,[esp + 20];lumc |
123 : | mul eax; lumc^2 | ||
124 : | shr eax,6; 64*lumc^2 | ||
125 : | movd ecx,%2 | ||
126 : | sub ecx,eax | ||
127 : | |||
128 : | mov edx,[esp + 28]; pdevc | ||
129 : | mov [edx],ecx | ||
130 : | |||
131 : | mov eax,[esp + 16];lumo | ||
132 : | mul dword [esp + 20]; lumo*lumc, should fit in eax | ||
133 : | shr eax,6; 64*lumo*lumc | ||
134 : | movd ecx,%3 | ||
135 : | sub ecx,eax | ||
136 : | |||
137 : | mov edx,[esp + 32]; pcorr | ||
138 : | mov [edx],ecx | ||
139 : | %endmacro | ||
140 : | |||
141 : | |||
142 : | Skal | 1730 | SECTION .text |
143 : | |||
144 : | cglobal lum_8x8_mmx | ||
145 : | cglobal consim_sse2 | ||
146 : | cglobal consim_mmx | ||
147 : | |||
148 : | ;int lum_8x8_c(uint8_t* ptr, uint32_t stride) | ||
149 : | |||
150 : | ALIGN 16 | ||
151 : | lum_8x8_mmx: | ||
152 : | mov ecx, [esp + 4] ;ptr | ||
153 : | mov edx, [esp + 8];stride | ||
154 : | |||
155 : | pxor mm0,mm0 | ||
156 : | |||
157 : | Skal | 1740 | ACC_ROW mm1, mm2 |
158 : | Skal | 1730 | |
159 : | Skal | 1740 | ACC_ROW mm3, mm4 |
160 : | Skal | 1730 | |
161 : | Skal | 1740 | ACC_ROW mm5, mm6 |
162 : | Skal | 1730 | |
163 : | Skal | 1740 | ACC_ROW mm7, mm4 |
164 : | Skal | 1730 | |
165 : | paddw mm1, mm3 | ||
166 : | paddw mm5, mm7 | ||
167 : | paddw mm1, mm5 | ||
168 : | |||
169 : | movd eax,mm1 | ||
170 : | ret | ||
171 : | .endfunc | ||
172 : | |||
173 : | ALIGN 16 | ||
174 : | Skal | 1743 | consim_sse2: |
175 : | Skal | 1730 | mov ecx,[esp+4] ;ptro |
176 : | mov edx,[esp+8] ;ptrc | ||
177 : | mov eax,[esp+12];stride | ||
178 : | |||
179 : | Skal | 1743 | pxor xmm2,xmm2;null vektor |
180 : | pxor xmm5,xmm5;devo | ||
181 : | pxor xmm6,xmm6;devc | ||
182 : | pxor xmm7,xmm7;corr | ||
183 : | |||
184 : | ;broadcast lumo/c | ||
185 : | punpcklbw xmm6,xmm6 | ||
186 : | punpcklwd xmm6,xmm6 | ||
187 : | pshufd xmm6,xmm6,00000000b;or shufps | ||
188 : | punpcklbw xmm7,xmm7 | ||
189 : | punpcklwd xmm7,xmm7 | ||
190 : | pshufd xmm7,xmm7,00000000b | ||
191 : | |||
192 : | CONSIM_1x8_SSE2 | ||
193 : | Skal | 1730 | add ecx,eax |
194 : | add edx,eax | ||
195 : | Skal | 1743 | CONSIM_1x8_SSE2 |
196 : | Skal | 1730 | add ecx,eax |
197 : | add edx,eax | ||
198 : | Skal | 1743 | CONSIM_1x8_SSE2 |
199 : | Skal | 1730 | add ecx,eax |
200 : | add edx,eax | ||
201 : | Skal | 1743 | CONSIM_1x8_SSE2 |
202 : | Skal | 1730 | add ecx,eax |
203 : | add edx,eax | ||
204 : | Skal | 1743 | CONSIM_1x8_SSE2 |
205 : | Skal | 1730 | add ecx,eax |
206 : | add edx,eax | ||
207 : | Skal | 1743 | CONSIM_1x8_SSE2 |
208 : | Skal | 1730 | add ecx,eax |
209 : | add edx,eax | ||
210 : | Skal | 1743 | CONSIM_1x8_SSE2 |
211 : | Skal | 1730 | add ecx,eax |
212 : | add edx,eax | ||
213 : | Skal | 1743 | CONSIM_1x8_SSE2 |
214 : | Skal | 1730 | |
215 : | Skal | 1743 | ;accumulate xmm5-7 |
216 : | pshufd xmm0, xmm5, 0EH | ||
217 : | paddd xmm5, xmm0 | ||
218 : | pshufd xmm0, xmm5, 01H | ||
219 : | paddd xmm5, xmm0 | ||
220 : | Skal | 1730 | |
221 : | Skal | 1743 | pshufd xmm1, xmm6, 0EH |
222 : | paddd xmm6, xmm1 | ||
223 : | pshufd xmm1, xmm6, 01H | ||
224 : | paddd xmm6, xmm1 | ||
225 : | |||
226 : | pshufd xmm2, xmm7, 0EH | ||
227 : | paddd xmm7, xmm2 | ||
228 : | pshufd xmm2, xmm7, 01H | ||
229 : | paddd xmm7, xmm2 | ||
230 : | |||
231 : | CONSIM_WRITEOUT xmm5,xmm6,xmm7 | ||
232 : | Skal | 1730 | ret |
233 : | .endfunc | ||
234 : | |||
235 : | Skal | 1743 | |
236 : | |||
237 : | |||
238 : | |||
239 : | ALIGN 16 | ||
240 : | consim_mmx: | ||
241 : | Skal | 1730 | mov ecx,[esp+4] ;ptro |
242 : | mov edx,[esp+8] ;ptrc | ||
243 : | mov eax,[esp+12];stride | ||
244 : | Skal | 1743 | pxor mm2,mm2;null |
245 : | pxor mm5,mm5;devo | ||
246 : | pxor mm6,mm6;devc | ||
247 : | pxor mm7,mm7;corr | ||
248 : | Skal | 1730 | |
249 : | Skal | 1743 | CONSIM_1x8_MMX |
250 : | Skal | 1730 | add ecx,eax |
251 : | add edx,eax | ||
252 : | Skal | 1743 | CONSIM_1x8_MMX |
253 : | Skal | 1730 | add ecx,eax |
254 : | add edx,eax | ||
255 : | Skal | 1743 | CONSIM_1x8_MMX |
256 : | Skal | 1730 | add ecx,eax |
257 : | add edx,eax | ||
258 : | Skal | 1743 | CONSIM_1x8_MMX |
259 : | Skal | 1730 | add ecx,eax |
260 : | add edx,eax | ||
261 : | Skal | 1743 | CONSIM_1x8_MMX |
262 : | Skal | 1730 | add ecx,eax |
263 : | add edx,eax | ||
264 : | Skal | 1743 | CONSIM_1x8_MMX |
265 : | Skal | 1730 | add ecx,eax |
266 : | add edx,eax | ||
267 : | Skal | 1743 | CONSIM_1x8_MMX |
268 : | Skal | 1730 | add ecx,eax |
269 : | add edx,eax | ||
270 : | Skal | 1743 | CONSIM_1x8_MMX |
271 : | Skal | 1730 | |
272 : | Skal | 1743 | movq mm0,mm5 |
273 : | psrlq mm0,32 | ||
274 : | paddd mm5,mm0 | ||
275 : | movq mm1,mm6 | ||
276 : | psrlq mm1,32 | ||
277 : | paddd mm6,mm1 | ||
278 : | movq mm2,mm7 | ||
279 : | psrlq mm2,32 | ||
280 : | paddd mm7,mm2 | ||
281 : | Skal | 1730 | |
282 : | Skal | 1743 | CONSIM_WRITEOUT mm5,mm6,mm7 |
283 : | Skal | 1730 | ret |
284 : | Skal | 1743 | .endfunc |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |