Parent Directory | Revision Log
Revision 1192 - (view) (download)
1 : | edgomez | 1192 | ;/***************************************************************************** |
2 : | Isibaar | 262 | ; * |
3 : | edgomez | 1192 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - mmx 8x8 block-based halfpel interpolation - | ||
5 : | Isibaar | 262 | ; * |
6 : | edgomez | 1192 | ; * Copyright(C) 2002 Michael Militzer <isibaar@xvid.org> |
7 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
8 : | Isibaar | 262 | ; * |
9 : | edgomez | 1192 | ; * This program is free software ; you can redistribute it and/or modify |
10 : | ; * it under the terms of the GNU General Public License as published by | ||
11 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
12 : | ; * (at your option) any later version. | ||
13 : | Isibaar | 262 | ; * |
14 : | edgomez | 1192 | ; * This program is distributed in the hope that it will be useful, |
15 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
16 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
17 : | ; * GNU General Public License for more details. | ||
18 : | Isibaar | 262 | ; * |
19 : | edgomez | 1192 | ; * You should have received a copy of the GNU General Public License |
20 : | ; * along with this program ; if not, write to the Free Software | ||
21 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
22 : | Isibaar | 262 | ; * |
23 : | edgomez | 1192 | ; ****************************************************************************/ |
24 : | Isibaar | 262 | |
25 : | edgomez | 1192 | BITS 32 |
26 : | edgomez | 851 | |
27 : | edgomez | 1192 | %macro cglobal 1 |
28 : | Isibaar | 262 | %ifdef PREFIX |
29 : | edgomez | 1192 | global _%1 |
30 : | Isibaar | 262 | %define %1 _%1 |
31 : | %else | ||
32 : | global %1 | ||
33 : | %endif | ||
34 : | %endmacro | ||
35 : | |||
36 : | edgomez | 1192 | ;============================================================================= |
37 : | ; Read only data | ||
38 : | ;============================================================================= | ||
39 : | Isibaar | 262 | |
40 : | edgomez | 1192 | SECTION .rodata |
41 : | Isibaar | 262 | |
42 : | edgomez | 1192 | ALIGN 16 |
43 : | mmx_one: | ||
44 : | times 8 db 1 | ||
45 : | Isibaar | 262 | |
46 : | edgomez | 1192 | SECTION .text |
47 : | Isibaar | 262 | |
48 : | cglobal interpolate8x8_halfpel_h_xmm | ||
49 : | cglobal interpolate8x8_halfpel_v_xmm | ||
50 : | cglobal interpolate8x8_halfpel_hv_xmm | ||
51 : | |||
52 : | ;=========================================================================== | ||
53 : | ; | ||
54 : | ; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, | ||
55 : | ; const uint8_t * const src, | ||
56 : | ; const uint32_t stride, | ||
57 : | ; const uint32_t rounding); | ||
58 : | ; | ||
59 : | ;=========================================================================== | ||
60 : | |||
61 : | %macro COPY_H_SSE_RND0 0 | ||
62 : | movq mm0, [eax] | ||
63 : | pavgb mm0, [eax+1] | ||
64 : | movq mm1, [eax+edx] | ||
65 : | pavgb mm1, [eax+edx+1] | ||
66 : | lea eax,[eax+2*edx] | ||
67 : | movq [ecx],mm0 | ||
68 : | movq [ecx+edx],mm1 | ||
69 : | %endmacro | ||
70 : | |||
71 : | %macro COPY_H_SSE_RND1 0 | ||
72 : | movq mm0, [eax] | ||
73 : | movq mm1, [eax+edx] | ||
74 : | movq mm4, mm0 | ||
75 : | movq mm5, mm1 | ||
76 : | edgomez | 1192 | movq mm2, [eax+1] |
77 : | Isibaar | 262 | movq mm3, [eax+edx+1] |
78 : | pavgb mm0, mm2 | ||
79 : | pxor mm2, mm4 | ||
80 : | pavgb mm1, mm3 | ||
81 : | edgomez | 1192 | lea eax, [eax+2*edx] |
82 : | Isibaar | 262 | pxor mm3, mm5 |
83 : | pand mm2, mm7 | ||
84 : | pand mm3, mm7 | ||
85 : | psubb mm0, mm2 | ||
86 : | movq [ecx], mm0 | ||
87 : | psubb mm1, mm3 | ||
88 : | edgomez | 1192 | movq [ecx+edx], mm1 |
89 : | Isibaar | 262 | %endmacro |
90 : | |||
91 : | edgomez | 1192 | ALIGN 16 |
92 : | Isibaar | 262 | interpolate8x8_halfpel_h_xmm: |
93 : | |||
94 : | edgomez | 1192 | mov eax, [esp+16] ; rounding |
95 : | mov ecx, [esp+ 4] ; Dst | ||
96 : | Isibaar | 262 | test eax,eax |
97 : | edgomez | 1192 | mov eax, [esp+ 8] ; Src |
98 : | mov edx, [esp+12] ; stride | ||
99 : | Isibaar | 262 | |
100 : | jnz near .rounding1 | ||
101 : | |||
102 : | COPY_H_SSE_RND0 | ||
103 : | lea ecx,[ecx+2*edx] | ||
104 : | COPY_H_SSE_RND0 | ||
105 : | lea ecx,[ecx+2*edx] | ||
106 : | COPY_H_SSE_RND0 | ||
107 : | lea ecx,[ecx+2*edx] | ||
108 : | COPY_H_SSE_RND0 | ||
109 : | ret | ||
110 : | |||
111 : | .rounding1 | ||
112 : | edgomez | 1192 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
113 : | Isibaar | 262 | movq mm7, [mmx_one] |
114 : | COPY_H_SSE_RND1 | ||
115 : | lea ecx, [ecx+2*edx] | ||
116 : | COPY_H_SSE_RND1 | ||
117 : | lea ecx,[ecx+2*edx] | ||
118 : | COPY_H_SSE_RND1 | ||
119 : | lea ecx,[ecx+2*edx] | ||
120 : | COPY_H_SSE_RND1 | ||
121 : | ret | ||
122 : | |||
123 : | ;=========================================================================== | ||
124 : | ; | ||
125 : | ; void interpolate8x8_halfpel_v_xmm(uint8_t * const dst, | ||
126 : | edgomez | 1192 | ; const uint8_t * const src, |
127 : | ; const uint32_t stride, | ||
128 : | ; const uint32_t rounding); | ||
129 : | Isibaar | 262 | ; |
130 : | ;=========================================================================== | ||
131 : | |||
132 : | %macro COPY_V_SSE_RND0 0 | ||
133 : | edgomez | 1192 | movq mm0, [eax] |
134 : | movq mm1, [eax+edx] | ||
135 : | Isibaar | 262 | pavgb mm0, mm1 |
136 : | pavgb mm1, [eax+2*edx] | ||
137 : | edgomez | 1192 | lea eax, [eax+2*edx] |
138 : | movq [ecx], mm0 | ||
139 : | Isibaar | 262 | movq [ecx+edx],mm1 |
140 : | %endmacro | ||
141 : | |||
142 : | %macro COPY_V_SSE_RND1 0 | ||
143 : | movq mm0, mm2 | ||
144 : | movq mm1, [eax] | ||
145 : | movq mm2, [eax+edx] | ||
146 : | lea eax,[eax+2*edx] | ||
147 : | movq mm4, mm0 | ||
148 : | movq mm5, mm1 | ||
149 : | pavgb mm0, mm1 | ||
150 : | edgomez | 1192 | pxor mm4, mm1 |
151 : | Isibaar | 262 | pavgb mm1, mm2 |
152 : | pxor mm5, mm2 | ||
153 : | edgomez | 1192 | pand mm4, mm7 ; lsb's of (i^j)... |
154 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
155 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
156 : | Isibaar | 262 | movq [ecx], mm0 |
157 : | edgomez | 1192 | psubb mm1, mm5 ; ...are substracted from result of pavgb |
158 : | Isibaar | 262 | movq [ecx+edx], mm1 |
159 : | %endmacro | ||
160 : | |||
161 : | edgomez | 1192 | ALIGN 16 |
162 : | Isibaar | 262 | interpolate8x8_halfpel_v_xmm: |
163 : | |||
164 : | mov eax, [esp+16]; rounding | ||
165 : | edgomez | 1192 | mov ecx, [esp+ 4] ; Dst |
166 : | Isibaar | 262 | test eax,eax |
167 : | edgomez | 1192 | mov eax, [esp+ 8] ; Src |
168 : | mov edx, [esp+12] ; stride | ||
169 : | Isibaar | 262 | |
170 : | edgomez | 1192 | ; we process 2 line at a time |
171 : | Isibaar | 262 | jnz near .rounding1 |
172 : | |||
173 : | COPY_V_SSE_RND0 | ||
174 : | lea ecx, [ecx+2*edx] | ||
175 : | COPY_V_SSE_RND0 | ||
176 : | lea ecx, [ecx+2*edx] | ||
177 : | COPY_V_SSE_RND0 | ||
178 : | lea ecx, [ecx+2*edx] | ||
179 : | COPY_V_SSE_RND0 | ||
180 : | ret | ||
181 : | |||
182 : | .rounding1 | ||
183 : | edgomez | 1192 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
184 : | Isibaar | 262 | movq mm7, [mmx_one] |
185 : | edgomez | 1192 | movq mm2, [eax] ; loop invariant |
186 : | Isibaar | 262 | add eax, edx |
187 : | |||
188 : | COPY_V_SSE_RND1 | ||
189 : | lea ecx,[ecx+2*edx] | ||
190 : | COPY_V_SSE_RND1 | ||
191 : | lea ecx,[ecx+2*edx] | ||
192 : | COPY_V_SSE_RND1 | ||
193 : | lea ecx,[ecx+2*edx] | ||
194 : | COPY_V_SSE_RND1 | ||
195 : | ret | ||
196 : | |||
197 : | ;=========================================================================== | ||
198 : | ; | ||
199 : | ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, | ||
200 : | edgomez | 1192 | ; const uint8_t * const src, |
201 : | ; const uint32_t stride, | ||
202 : | ; const uint32_t rounding); | ||
203 : | Isibaar | 262 | ; |
204 : | ; | ||
205 : | ;=========================================================================== | ||
206 : | |||
207 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
208 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
209 : | ; The boolean relations are: | ||
210 : | edgomez | 1192 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
211 : | Isibaar | 262 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
212 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
213 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
214 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
215 : | |||
216 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
217 : | |||
218 : | %macro COPY_HV_SSE_RND0 0 | ||
219 : | edgomez | 1192 | lea eax, [eax+edx] |
220 : | Isibaar | 262 | |
221 : | edgomez | 1192 | movq mm0, [eax] |
222 : | movq mm1, [eax+1] | ||
223 : | Isibaar | 262 | |
224 : | edgomez | 1192 | movq mm6, mm0 |
225 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
226 : | lea eax, [eax+edx] | ||
227 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
228 : | Isibaar | 262 | |
229 : | edgomez | 1192 | por mm3, mm1 ; ij |= jk |
230 : | movq mm6, mm2 | ||
231 : | pxor mm6, mm0 ; mm6 = s^t | ||
232 : | pand mm3, mm6 ; (ij|jk) &= st | ||
233 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
234 : | pand mm3, mm7 ; mask lsb | ||
235 : | psubb mm2, mm3 ; apply. | ||
236 : | Isibaar | 262 | |
237 : | edgomez | 1192 | movq [ecx], mm2 |
238 : | Isibaar | 262 | |
239 : | edgomez | 1192 | movq mm2, [eax] |
240 : | movq mm3, [eax+1] | ||
241 : | movq mm6, mm2 | ||
242 : | pavgb mm2, mm3 ; preserved for next iteration | ||
243 : | lea ecx,[ecx+edx] | ||
244 : | pxor mm3, mm6 ; preserved for next iteration | ||
245 : | Isibaar | 262 | |
246 : | edgomez | 1192 | por mm1, mm3 |
247 : | movq mm6, mm0 | ||
248 : | pxor mm6, mm2 | ||
249 : | pand mm1, mm6 | ||
250 : | pavgb mm0, mm2 | ||
251 : | Isibaar | 262 | |
252 : | edgomez | 1192 | pand mm1, mm7 |
253 : | psubb mm0, mm1 | ||
254 : | Isibaar | 262 | |
255 : | edgomez | 1192 | movq [ecx], mm0 |
256 : | Isibaar | 262 | %endmacro |
257 : | |||
258 : | %macro COPY_HV_SSE_RND1 0 | ||
259 : | edgomez | 1192 | lea eax, [eax+edx] |
260 : | Isibaar | 262 | |
261 : | edgomez | 1192 | movq mm0, [eax] |
262 : | movq mm1, [eax+1] | ||
263 : | Isibaar | 262 | |
264 : | edgomez | 1192 | movq mm6, mm0 |
265 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
266 : | lea eax, [eax+edx] | ||
267 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
268 : | Isibaar | 262 | |
269 : | edgomez | 1192 | pand mm3, mm1 |
270 : | movq mm6, mm2 | ||
271 : | pxor mm6, mm0 | ||
272 : | por mm3, mm6 | ||
273 : | pavgb mm2, mm0 | ||
274 : | pand mm3, mm7 | ||
275 : | psubb mm2, mm3 | ||
276 : | Isibaar | 262 | |
277 : | edgomez | 1192 | movq [ecx], mm2 |
278 : | Isibaar | 262 | |
279 : | edgomez | 1192 | movq mm2, [eax] |
280 : | movq mm3, [eax+1] | ||
281 : | movq mm6, mm2 | ||
282 : | pavgb mm2, mm3 ; preserved for next iteration | ||
283 : | lea ecx,[ecx+edx] | ||
284 : | pxor mm3, mm6 ; preserved for next iteration | ||
285 : | Isibaar | 262 | |
286 : | edgomez | 1192 | pand mm1, mm3 |
287 : | movq mm6, mm0 | ||
288 : | pxor mm6, mm2 | ||
289 : | por mm1, mm6 | ||
290 : | pavgb mm0, mm2 | ||
291 : | pand mm1, mm7 | ||
292 : | psubb mm0, mm1 | ||
293 : | Isibaar | 262 | |
294 : | edgomez | 1192 | movq [ecx], mm0 |
295 : | Isibaar | 262 | %endmacro |
296 : | |||
297 : | edgomez | 1192 | ALIGN 16 |
298 : | Isibaar | 262 | interpolate8x8_halfpel_hv_xmm: |
299 : | edgomez | 1192 | mov eax, [esp+16] ; rounding |
300 : | mov ecx, [esp+ 4] ; Dst | ||
301 : | test eax, eax | ||
302 : | mov eax, [esp+ 8] ; Src | ||
303 : | mov edx, [esp+12] ; stride | ||
304 : | Isibaar | 262 | |
305 : | movq mm7, [mmx_one] | ||
306 : | |||
307 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
308 : | movq mm2, [eax] | ||
309 : | movq mm3, [eax+1] | ||
310 : | movq mm6, mm2 | ||
311 : | pavgb mm2, mm3 | ||
312 : | edgomez | 1192 | pxor mm3, mm6 ; mm2/mm3 ready |
313 : | Isibaar | 262 | |
314 : | jnz near .rounding1 | ||
315 : | |||
316 : | COPY_HV_SSE_RND0 | ||
317 : | add ecx, edx | ||
318 : | COPY_HV_SSE_RND0 | ||
319 : | add ecx, edx | ||
320 : | COPY_HV_SSE_RND0 | ||
321 : | add ecx, edx | ||
322 : | COPY_HV_SSE_RND0 | ||
323 : | ret | ||
324 : | |||
325 : | .rounding1 | ||
326 : | COPY_HV_SSE_RND1 | ||
327 : | add ecx, edx | ||
328 : | COPY_HV_SSE_RND1 | ||
329 : | add ecx, edx | ||
330 : | COPY_HV_SSE_RND1 | ||
331 : | add ecx, edx | ||
332 : | COPY_HV_SSE_RND1 | ||
333 : | edgomez | 1192 | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |