Parent Directory | Revision Log
Revision 1793 - (view) (download)
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 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
33 : | Skal | 1730 | %else |
34 : | global _%1 | ||
35 : | %define %1 _%1 | ||
36 : | Isibaar | 1793 | %define ENDFUNC |
37 : | Skal | 1730 | %endif |
38 : | %else | ||
39 : | %ifdef MARK_FUNCS | ||
40 : | global %1:function %1.endfunc-%1 | ||
41 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
42 : | Skal | 1730 | %else |
43 : | global %1 | ||
44 : | Isibaar | 1793 | %define ENDFUNC |
45 : | Skal | 1730 | %endif |
46 : | %endif | ||
47 : | %endmacro | ||
48 : | |||
49 : | Skal | 1740 | %macro ACC_ROW 2 |
50 : | movq %1,[ ecx] | ||
51 : | movq %2,[ecx+edx] | ||
52 : | Skal | 1730 | psadbw %1,mm0 |
53 : | Skal | 1740 | psadbw %2,mm0 |
54 : | lea ecx, [ecx+2*edx] | ||
55 : | paddw %1, %2 | ||
56 : | Skal | 1730 | %endmacro |
57 : | |||
58 : | %macro CONSIM_1x8_SSE2 0 | ||
59 : | Skal | 1743 | movdqu xmm0,[ecx] |
60 : | movdqu xmm1,[edx] | ||
61 : | Skal | 1730 | |
62 : | ;unpack to words | ||
63 : | punpcklbw xmm0,xmm2 | ||
64 : | punpcklbw xmm1,xmm2 | ||
65 : | |||
66 : | Skal | 1743 | movaps xmm3,xmm0 |
67 : | movaps xmm4,xmm1 | ||
68 : | Skal | 1730 | |
69 : | Skal | 1743 | pmaddwd xmm0,xmm0;orig |
70 : | pmaddwd xmm1,xmm1;comp | ||
71 : | pmaddwd xmm3,xmm4;corr | ||
72 : | Skal | 1730 | |
73 : | Skal | 1743 | paddd xmm5,xmm0 |
74 : | paddd xmm6,xmm1 | ||
75 : | paddd xmm7,xmm3 | ||
76 : | Skal | 1730 | %endmacro |
77 : | |||
78 : | %macro CONSIM_1x8_MMX 0 | ||
79 : | Skal | 1743 | movq mm0,[ecx];orig |
80 : | movq mm1,[edx];comp | ||
81 : | Skal | 1730 | |
82 : | Skal | 1743 | ;unpack low half of qw to words |
83 : | punpcklbw mm0,mm2 | ||
84 : | punpcklbw mm1,mm2 | ||
85 : | Skal | 1730 | |
86 : | Skal | 1743 | movq mm3,mm0 |
87 : | pmaddwd mm3,mm0 | ||
88 : | paddd mm5,mm3; | ||
89 : | Skal | 1730 | |
90 : | Skal | 1743 | movq mm4,mm1 |
91 : | pmaddwd mm4,mm1 | ||
92 : | paddd mm6,mm4; | ||
93 : | |||
94 : | Skal | 1730 | pmaddwd mm1,mm0 |
95 : | Skal | 1743 | paddd mm7,mm1 |
96 : | Skal | 1730 | |
97 : | Skal | 1743 | movq mm0,[ecx];orig |
98 : | movq mm1,[edx];comp | ||
99 : | Skal | 1730 | |
100 : | Skal | 1743 | ;unpack high half of qw to words |
101 : | punpckhbw mm0,mm2 | ||
102 : | punpckhbw mm1,mm2 | ||
103 : | Skal | 1730 | |
104 : | Skal | 1743 | movq mm3,mm0 |
105 : | pmaddwd mm3,mm0 | ||
106 : | paddd mm5,mm3; | ||
107 : | Skal | 1730 | |
108 : | Skal | 1743 | movq mm4,mm1 |
109 : | pmaddwd mm4,mm1 | ||
110 : | paddd mm6,mm4; | ||
111 : | |||
112 : | Skal | 1730 | pmaddwd mm1,mm0 |
113 : | Skal | 1743 | paddd mm7,mm1 |
114 : | Skal | 1730 | %endmacro |
115 : | |||
116 : | Skal | 1743 | %macro CONSIM_WRITEOUT 3 |
117 : | mov eax,[esp + 16];lumo | ||
118 : | mul eax; lumo^2 | ||
119 : | Skal | 1744 | add eax, 32 |
120 : | Skal | 1743 | shr eax,6; 64*lum0^2 |
121 : | movd ecx,%1 | ||
122 : | sub ecx,eax | ||
123 : | Skal | 1730 | |
124 : | Skal | 1743 | mov edx,[esp + 24]; pdevo |
125 : | mov [edx],ecx | ||
126 : | Skal | 1730 | |
127 : | Skal | 1743 | mov eax,[esp + 20];lumc |
128 : | mul eax; lumc^2 | ||
129 : | Skal | 1744 | add eax, 32 |
130 : | Skal | 1743 | shr eax,6; 64*lumc^2 |
131 : | movd ecx,%2 | ||
132 : | sub ecx,eax | ||
133 : | |||
134 : | mov edx,[esp + 28]; pdevc | ||
135 : | mov [edx],ecx | ||
136 : | |||
137 : | mov eax,[esp + 16];lumo | ||
138 : | mul dword [esp + 20]; lumo*lumc, should fit in eax | ||
139 : | Skal | 1744 | add eax, 32 |
140 : | Skal | 1743 | shr eax,6; 64*lumo*lumc |
141 : | movd ecx,%3 | ||
142 : | sub ecx,eax | ||
143 : | |||
144 : | mov edx,[esp + 32]; pcorr | ||
145 : | mov [edx],ecx | ||
146 : | %endmacro | ||
147 : | |||
148 : | |||
149 : | Skal | 1730 | SECTION .text |
150 : | |||
151 : | cglobal lum_8x8_mmx | ||
152 : | cglobal consim_sse2 | ||
153 : | cglobal consim_mmx | ||
154 : | |||
155 : | ;int lum_8x8_c(uint8_t* ptr, uint32_t stride) | ||
156 : | |||
157 : | ALIGN 16 | ||
158 : | lum_8x8_mmx: | ||
159 : | mov ecx, [esp + 4] ;ptr | ||
160 : | mov edx, [esp + 8];stride | ||
161 : | |||
162 : | pxor mm0,mm0 | ||
163 : | |||
164 : | Skal | 1740 | ACC_ROW mm1, mm2 |
165 : | Skal | 1730 | |
166 : | Skal | 1740 | ACC_ROW mm3, mm4 |
167 : | Skal | 1730 | |
168 : | Skal | 1740 | ACC_ROW mm5, mm6 |
169 : | Skal | 1730 | |
170 : | Skal | 1740 | ACC_ROW mm7, mm4 |
171 : | Skal | 1730 | |
172 : | paddw mm1, mm3 | ||
173 : | paddw mm5, mm7 | ||
174 : | paddw mm1, mm5 | ||
175 : | |||
176 : | movd eax,mm1 | ||
177 : | ret | ||
178 : | Isibaar | 1793 | ENDFUNC |
179 : | Skal | 1730 | |
180 : | ALIGN 16 | ||
181 : | Skal | 1743 | consim_sse2: |
182 : | Skal | 1730 | mov ecx,[esp+4] ;ptro |
183 : | mov edx,[esp+8] ;ptrc | ||
184 : | mov eax,[esp+12];stride | ||
185 : | |||
186 : | Skal | 1743 | pxor xmm2,xmm2;null vektor |
187 : | pxor xmm5,xmm5;devo | ||
188 : | pxor xmm6,xmm6;devc | ||
189 : | pxor xmm7,xmm7;corr | ||
190 : | |||
191 : | ;broadcast lumo/c | ||
192 : | punpcklbw xmm6,xmm6 | ||
193 : | punpcklwd xmm6,xmm6 | ||
194 : | pshufd xmm6,xmm6,00000000b;or shufps | ||
195 : | punpcklbw xmm7,xmm7 | ||
196 : | punpcklwd xmm7,xmm7 | ||
197 : | pshufd xmm7,xmm7,00000000b | ||
198 : | |||
199 : | CONSIM_1x8_SSE2 | ||
200 : | Skal | 1730 | add ecx,eax |
201 : | add edx,eax | ||
202 : | Skal | 1743 | CONSIM_1x8_SSE2 |
203 : | Skal | 1730 | add ecx,eax |
204 : | add edx,eax | ||
205 : | Skal | 1743 | CONSIM_1x8_SSE2 |
206 : | Skal | 1730 | add ecx,eax |
207 : | add edx,eax | ||
208 : | Skal | 1743 | CONSIM_1x8_SSE2 |
209 : | Skal | 1730 | add ecx,eax |
210 : | add edx,eax | ||
211 : | Skal | 1743 | CONSIM_1x8_SSE2 |
212 : | Skal | 1730 | add ecx,eax |
213 : | add edx,eax | ||
214 : | Skal | 1743 | CONSIM_1x8_SSE2 |
215 : | Skal | 1730 | add ecx,eax |
216 : | add edx,eax | ||
217 : | Skal | 1743 | CONSIM_1x8_SSE2 |
218 : | Skal | 1730 | add ecx,eax |
219 : | add edx,eax | ||
220 : | Skal | 1743 | CONSIM_1x8_SSE2 |
221 : | Skal | 1730 | |
222 : | Skal | 1743 | ;accumulate xmm5-7 |
223 : | chl | 1745 | pshufd xmm0, xmm5, 0x0E |
224 : | Skal | 1743 | paddd xmm5, xmm0 |
225 : | chl | 1745 | pshufd xmm0, xmm5, 0x01 |
226 : | Skal | 1743 | paddd xmm5, xmm0 |
227 : | Skal | 1730 | |
228 : | chl | 1745 | pshufd xmm1, xmm6, 0x0E |
229 : | Skal | 1743 | paddd xmm6, xmm1 |
230 : | chl | 1745 | pshufd xmm1, xmm6, 0x01 |
231 : | Skal | 1743 | paddd xmm6, xmm1 |
232 : | |||
233 : | chl | 1745 | pshufd xmm2, xmm7, 0x0E |
234 : | Skal | 1743 | paddd xmm7, xmm2 |
235 : | chl | 1745 | pshufd xmm2, xmm7, 0x01 |
236 : | Skal | 1743 | paddd xmm7, xmm2 |
237 : | |||
238 : | CONSIM_WRITEOUT xmm5,xmm6,xmm7 | ||
239 : | Skal | 1730 | ret |
240 : | Isibaar | 1793 | ENDFUNC |
241 : | Skal | 1730 | |
242 : | Skal | 1743 | |
243 : | |||
244 : | |||
245 : | |||
246 : | ALIGN 16 | ||
247 : | consim_mmx: | ||
248 : | Skal | 1730 | mov ecx,[esp+4] ;ptro |
249 : | mov edx,[esp+8] ;ptrc | ||
250 : | mov eax,[esp+12];stride | ||
251 : | Skal | 1743 | pxor mm2,mm2;null |
252 : | pxor mm5,mm5;devo | ||
253 : | pxor mm6,mm6;devc | ||
254 : | pxor mm7,mm7;corr | ||
255 : | Skal | 1730 | |
256 : | Skal | 1743 | CONSIM_1x8_MMX |
257 : | Skal | 1730 | add ecx,eax |
258 : | add edx,eax | ||
259 : | Skal | 1743 | CONSIM_1x8_MMX |
260 : | Skal | 1730 | add ecx,eax |
261 : | add edx,eax | ||
262 : | Skal | 1743 | CONSIM_1x8_MMX |
263 : | Skal | 1730 | add ecx,eax |
264 : | add edx,eax | ||
265 : | Skal | 1743 | CONSIM_1x8_MMX |
266 : | Skal | 1730 | add ecx,eax |
267 : | add edx,eax | ||
268 : | Skal | 1743 | CONSIM_1x8_MMX |
269 : | Skal | 1730 | add ecx,eax |
270 : | add edx,eax | ||
271 : | Skal | 1743 | CONSIM_1x8_MMX |
272 : | Skal | 1730 | add ecx,eax |
273 : | add edx,eax | ||
274 : | Skal | 1743 | CONSIM_1x8_MMX |
275 : | Skal | 1730 | add ecx,eax |
276 : | add edx,eax | ||
277 : | Skal | 1743 | CONSIM_1x8_MMX |
278 : | Skal | 1730 | |
279 : | Skal | 1743 | movq mm0,mm5 |
280 : | psrlq mm0,32 | ||
281 : | paddd mm5,mm0 | ||
282 : | movq mm1,mm6 | ||
283 : | psrlq mm1,32 | ||
284 : | paddd mm6,mm1 | ||
285 : | movq mm2,mm7 | ||
286 : | psrlq mm2,32 | ||
287 : | paddd mm7,mm2 | ||
288 : | Skal | 1730 | |
289 : | Skal | 1743 | CONSIM_WRITEOUT mm5,mm6,mm7 |
290 : | Skal | 1730 | ret |
291 : | Isibaar | 1793 | ENDFUNC |
292 : | Isibaar | 1790 | |
293 : | %ifidn __OUTPUT_FORMAT__,elf | ||
294 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
295 : | %endif |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |