Parent Directory | Revision Log
Revision 1740 -
(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 : | ;load a dq from mem to a xmm reg | ||
55 : | %macro LOAD_XMM 2 | ||
56 : | movdqu %1,[%2] | ||
57 : | ;movhps %1,[%2+8] | ||
58 : | %endmacro | ||
59 : | |||
60 : | %macro WRITE_XMM 2 | ||
61 : | ;movlps [%1],%2 | ||
62 : | ;movhps [%1+8],%2 | ||
63 : | movdqu [%1],%2 | ||
64 : | %endmacro | ||
65 : | |||
66 : | %macro CONSIM_1x8_SSE2 0 | ||
67 : | LOAD_XMM xmm0,ecx | ||
68 : | LOAD_XMM xmm1,edx | ||
69 : | pxor xmm2,xmm2 | ||
70 : | |||
71 : | ;unpack to words | ||
72 : | punpcklbw xmm0,xmm2 | ||
73 : | punpcklbw xmm1,xmm2 | ||
74 : | |||
75 : | ;devo | ||
76 : | psubw xmm0,xmm6 | ||
77 : | movaps xmm2,xmm0 | ||
78 : | pmaddwd xmm2,xmm0 | ||
79 : | paddd xmm3,xmm2 | ||
80 : | |||
81 : | ;devc | ||
82 : | psubw xmm1,xmm7 | ||
83 : | movaps xmm2,xmm1 | ||
84 : | pmaddwd xmm2,xmm1 | ||
85 : | paddd xmm4,xmm2 | ||
86 : | |||
87 : | ;corr | ||
88 : | pmaddwd xmm1,xmm0 | ||
89 : | paddd xmm5,xmm1 | ||
90 : | %endmacro | ||
91 : | |||
92 : | |||
93 : | %macro CONSIM_1x8_MMX 0 | ||
94 : | movq mm0,[ecx];orig | ||
95 : | movq mm1,[edx];comp | ||
96 : | pxor mm2,mm2;null vector | ||
97 : | |||
98 : | ;unpack low half of qw to words | ||
99 : | punpcklbw mm0,mm2 | ||
100 : | punpcklbw mm1,mm2 | ||
101 : | |||
102 : | ;devo | ||
103 : | psubw mm0,mm6 | ||
104 : | movq mm2,mm0 | ||
105 : | pmaddwd mm2,mm0 | ||
106 : | paddd mm3,mm2; | ||
107 : | |||
108 : | ;devc | ||
109 : | psubw mm1,mm7 | ||
110 : | movq mm2,mm1 | ||
111 : | pmaddwd mm2,mm1 | ||
112 : | paddd mm4,mm2 | ||
113 : | |||
114 : | ;corr | ||
115 : | pmaddwd mm1,mm0 | ||
116 : | paddd mm5,mm1 | ||
117 : | |||
118 : | movq mm0,[ecx] | ||
119 : | movq mm1,[edx] | ||
120 : | pxor mm2,mm2;null vector | ||
121 : | |||
122 : | ;unpack high half of qw to words | ||
123 : | punpckhbw mm0,mm2 | ||
124 : | punpckhbw mm1,mm2 | ||
125 : | |||
126 : | ;devo | ||
127 : | psubw mm0,mm6 | ||
128 : | movq mm2,mm0 | ||
129 : | pmaddwd mm2,mm0 | ||
130 : | paddd mm3,mm2; | ||
131 : | |||
132 : | ;devc | ||
133 : | psubw mm1,mm7 | ||
134 : | movq mm2,mm1 | ||
135 : | pmaddwd mm2,mm1 | ||
136 : | paddd mm4,mm2 | ||
137 : | |||
138 : | ;corr | ||
139 : | pmaddwd mm1,mm0 | ||
140 : | paddd mm5,mm1 | ||
141 : | %endmacro | ||
142 : | |||
143 : | |||
144 : | |||
145 : | SECTION .text | ||
146 : | |||
147 : | cglobal lum_8x8_mmx | ||
148 : | cglobal consim_sse2 | ||
149 : | cglobal consim_mmx | ||
150 : | |||
151 : | ;int lum_8x8_c(uint8_t* ptr, uint32_t stride) | ||
152 : | |||
153 : | ALIGN 16 | ||
154 : | lum_8x8_mmx: | ||
155 : | mov ecx, [esp + 4] ;ptr | ||
156 : | mov edx, [esp + 8];stride | ||
157 : | |||
158 : | pxor mm0,mm0 | ||
159 : | |||
160 : | Skal | 1740 | ACC_ROW mm1, mm2 |
161 : | Skal | 1730 | |
162 : | Skal | 1740 | ACC_ROW mm3, mm4 |
163 : | Skal | 1730 | |
164 : | Skal | 1740 | ACC_ROW mm5, mm6 |
165 : | Skal | 1730 | |
166 : | Skal | 1740 | ACC_ROW mm7, mm4 |
167 : | Skal | 1730 | |
168 : | paddw mm1, mm3 | ||
169 : | paddw mm5, mm7 | ||
170 : | paddw mm1, mm5 | ||
171 : | |||
172 : | movd eax,mm1 | ||
173 : | ret | ||
174 : | .endfunc | ||
175 : | |||
176 : | ALIGN 16 | ||
177 : | consim_mmx: | ||
178 : | mov ecx,[esp+4] ;ptro | ||
179 : | pxor mm6,mm6; | ||
180 : | |||
181 : | mov edx,[esp+8] ;ptrc | ||
182 : | pxor mm3,mm3;devo | ||
183 : | pxor mm4,mm4;devc | ||
184 : | movd mm6,[esp + 16];lumo | ||
185 : | pxor mm7,mm7 | ||
186 : | mov eax,[esp+12];stride | ||
187 : | movd mm7,[esp + 20];lumc | ||
188 : | Skal | 1740 | pshufw mm6,mm6,00000000b ; TODO: remove later! not MMX, but SSE |
189 : | Skal | 1730 | pxor mm5,mm5;corr |
190 : | pshufw mm7,mm7,00000000b | ||
191 : | |||
192 : | CONSIM_1x8_MMX | ||
193 : | add ecx,eax | ||
194 : | add edx,eax | ||
195 : | CONSIM_1x8_MMX | ||
196 : | add ecx,eax | ||
197 : | add edx,eax | ||
198 : | CONSIM_1x8_MMX | ||
199 : | add ecx,eax | ||
200 : | add edx,eax | ||
201 : | CONSIM_1x8_MMX | ||
202 : | add ecx,eax | ||
203 : | add edx,eax | ||
204 : | CONSIM_1x8_MMX | ||
205 : | add ecx,eax | ||
206 : | add edx,eax | ||
207 : | CONSIM_1x8_MMX | ||
208 : | add ecx,eax | ||
209 : | add edx,eax | ||
210 : | CONSIM_1x8_MMX | ||
211 : | add ecx,eax | ||
212 : | add edx,eax | ||
213 : | CONSIM_1x8_MMX | ||
214 : | |||
215 : | pshufw mm0,mm3,01001110b | ||
216 : | paddd mm3,mm0 | ||
217 : | pshufw mm1,mm4,01001110b | ||
218 : | paddd mm4,mm1 | ||
219 : | pshufw mm2,mm5,01001110b | ||
220 : | paddd mm5,mm2 | ||
221 : | |||
222 : | ;load target pointer | ||
223 : | mov ecx,[esp + 24]; pdevo | ||
224 : | movd [ecx],mm3 | ||
225 : | mov edx,[esp + 28]; pdevc | ||
226 : | movd [edx],mm4 | ||
227 : | mov eax,[esp + 32]; corr | ||
228 : | movd [eax],mm5 | ||
229 : | emms | ||
230 : | ret | ||
231 : | .endfunc | ||
232 : | |||
233 : | consim_sse2: | ||
234 : | mov ecx,[esp+4] ;ptro | ||
235 : | pxor xmm6,xmm6; | ||
236 : | mov edx,[esp+8] ;ptrc | ||
237 : | pxor xmm3,xmm3;devo | ||
238 : | pxor xmm4,xmm4;devc | ||
239 : | movd xmm6,[esp + 16];lumo | ||
240 : | pxor xmm7,xmm7 | ||
241 : | mov eax,[esp+12];stride | ||
242 : | movd xmm7,[esp + 20];lumc | ||
243 : | pxor xmm5,xmm5;corr | ||
244 : | |||
245 : | ;broadcast lumo/c | ||
246 : | ;punpcklbw xmm6,xmm6 | ||
247 : | punpcklwd xmm6,xmm6 | ||
248 : | pshufd xmm6,xmm6,00000000b;or shufps | ||
249 : | ;punpcklbw xmm7,xmm7 | ||
250 : | punpcklwd xmm7,xmm7 | ||
251 : | pshufd xmm7,xmm7,00000000b | ||
252 : | |||
253 : | CONSIM_1x8_SSE2 | ||
254 : | add ecx,eax | ||
255 : | add edx,eax | ||
256 : | CONSIM_1x8_SSE2 | ||
257 : | add ecx,eax | ||
258 : | add edx,eax | ||
259 : | CONSIM_1x8_SSE2 | ||
260 : | add ecx,eax | ||
261 : | add edx,eax | ||
262 : | CONSIM_1x8_SSE2 | ||
263 : | add ecx,eax | ||
264 : | add edx,eax | ||
265 : | CONSIM_1x8_SSE2 | ||
266 : | add ecx,eax | ||
267 : | add edx,eax | ||
268 : | CONSIM_1x8_SSE2 | ||
269 : | add ecx,eax | ||
270 : | add edx,eax | ||
271 : | CONSIM_1x8_SSE2 | ||
272 : | add ecx,eax | ||
273 : | add edx,eax | ||
274 : | CONSIM_1x8_SSE2 | ||
275 : | |||
276 : | ;accumulate xmm3-5 | ||
277 : | pshufd xmm0, xmm3, 0EH ; Get bit 64-127 from xmm1 (or use movhlps) | ||
278 : | paddd xmm3, xmm0 ; Sums are in 2 dwords | ||
279 : | pshufd xmm0, xmm3, 01H ; Get bit 32-63 from xmm0 | ||
280 : | paddd xmm3, xmm0 ; Sum is in one dword | ||
281 : | |||
282 : | pshufd xmm1, xmm4, 0EH ; Get bit 64-127 from xmm1 (or use movhlps) | ||
283 : | paddd xmm4, xmm1 ; Sums are in 2 dwords | ||
284 : | pshufd xmm1, xmm4, 01H ; Get bit 32-63 from xmm0 | ||
285 : | paddd xmm4, xmm1 ; Sum is in one dword | ||
286 : | |||
287 : | pshufd xmm2, xmm5, 0EH ; Get bit 64-127 from xmm1 (or use movhlps) | ||
288 : | paddd xmm5, xmm2 ; Sums are in 2 dwords | ||
289 : | pshufd xmm2, xmm5, 01H ; Get bit 32-63 from xmm0 | ||
290 : | paddd xmm5, xmm2 ; Sum is in one dword | ||
291 : | |||
292 : | |||
293 : | ;load target pointer | ||
294 : | mov ecx,[esp + 24]; pdevo | ||
295 : | movd [ecx],xmm3 | ||
296 : | mov edx,[esp + 28]; pdevc | ||
297 : | movd [edx],xmm4 | ||
298 : | mov eax,[esp + 32]; corr | ||
299 : | movd [eax],xmm5 | ||
300 : | ret | ||
301 : | .endfunc |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |