Parent Directory | Revision Log
Revision 1519 - (view) (download)
1 : | edgomez | 1382 | ;/***************************************************************************** |
2 : | Isibaar | 262 | ; * |
3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - 3dnow 8x8 block-based halfpel interpolation - | ||
5 : | Isibaar | 262 | ; * |
6 : | edgomez | 1382 | ; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 : | ; * 2002 Michael Militzer <isibaar@xvid.org> | ||
8 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
9 : | Isibaar | 262 | ; * |
10 : | edgomez | 1382 | ; * This program is free software ; you can redistribute it and/or modify |
11 : | ; * it under the terms of the GNU General Public License as published by | ||
12 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
13 : | ; * (at your option) any later version. | ||
14 : | Isibaar | 262 | ; * |
15 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
16 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
17 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
18 : | ; * GNU General Public License for more details. | ||
19 : | Isibaar | 262 | ; * |
20 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
21 : | ; * along with this program ; if not, write to the Free Software | ||
22 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
23 : | Isibaar | 262 | ; * |
24 : | edgomez | 1382 | ; ****************************************************************************/ |
25 : | Isibaar | 262 | |
26 : | edgomez | 1382 | BITS 32 |
27 : | edgomez | 851 | |
28 : | edgomez | 1382 | %macro cglobal 1 |
29 : | Isibaar | 262 | %ifdef PREFIX |
30 : | edgomez | 1382 | global _%1 |
31 : | Isibaar | 262 | %define %1 _%1 |
32 : | %else | ||
33 : | global %1 | ||
34 : | %endif | ||
35 : | %endmacro | ||
36 : | |||
37 : | edgomez | 1382 | ;============================================================================= |
38 : | ; Read Only data | ||
39 : | ;============================================================================= | ||
40 : | Isibaar | 262 | |
41 : | edgomez | 1382 | %ifdef FORMAT_COFF |
42 : | edgomez | 1519 | SECTION .rodata |
43 : | edgomez | 1382 | %else |
44 : | edgomez | 1519 | SECTION .rodata align=16 |
45 : | edgomez | 1382 | %endif |
46 : | Isibaar | 262 | |
47 : | edgomez | 1382 | ALIGN 16 |
48 : | mmx_one: | ||
49 : | times 8 db 1 | ||
50 : | Isibaar | 262 | |
51 : | edgomez | 1382 | ;============================================================================= |
52 : | ; Code | ||
53 : | ;============================================================================= | ||
54 : | Isibaar | 262 | |
55 : | edgomez | 1382 | SECTION .text |
56 : | Isibaar | 262 | |
57 : | cglobal interpolate8x8_halfpel_h_3dn | ||
58 : | cglobal interpolate8x8_halfpel_v_3dn | ||
59 : | cglobal interpolate8x8_halfpel_hv_3dn | ||
60 : | |||
61 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
62 : | Isibaar | 262 | ; |
63 : | ; void interpolate8x8_halfpel_h_3dn(uint8_t * const dst, | ||
64 : | edgomez | 1382 | ; const uint8_t * const src, |
65 : | ; const uint32_t stride, | ||
66 : | ; const uint32_t rounding); | ||
67 : | Isibaar | 262 | ; |
68 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
69 : | Isibaar | 262 | |
70 : | %macro COPY_H_3DN_RND0 0 | ||
71 : | edgomez | 1382 | movq mm0, [eax] |
72 : | Isibaar | 262 | pavgusb mm0, [eax+1] |
73 : | edgomez | 1382 | movq mm1, [eax+edx] |
74 : | Isibaar | 262 | pavgusb mm1, [eax+edx+1] |
75 : | edgomez | 1382 | lea eax, [eax+2*edx] |
76 : | movq [ecx], mm0 | ||
77 : | movq [ecx+edx], mm1 | ||
78 : | Isibaar | 262 | %endmacro |
79 : | |||
80 : | %macro COPY_H_3DN_RND1 0 | ||
81 : | movq mm0, [eax] | ||
82 : | movq mm1, [eax+edx] | ||
83 : | movq mm4, mm0 | ||
84 : | movq mm5, mm1 | ||
85 : | edgomez | 1382 | movq mm2, [eax+1] |
86 : | Isibaar | 262 | movq mm3, [eax+edx+1] |
87 : | pavgusb mm0, mm2 | ||
88 : | pxor mm2, mm4 | ||
89 : | pavgusb mm1, mm3 | ||
90 : | edgomez | 1382 | lea eax, [eax+2*edx] |
91 : | Isibaar | 262 | pxor mm3, mm5 |
92 : | pand mm2, mm7 | ||
93 : | pand mm3, mm7 | ||
94 : | psubb mm0, mm2 | ||
95 : | movq [ecx], mm0 | ||
96 : | psubb mm1, mm3 | ||
97 : | movq [ecx+edx], mm1 | ||
98 : | %endmacro | ||
99 : | |||
100 : | edgomez | 1382 | ALIGN 16 |
101 : | Isibaar | 262 | interpolate8x8_halfpel_h_3dn: |
102 : | |||
103 : | mov eax, [esp+16] ; rounding | ||
104 : | mov ecx, [esp+ 4] ; Dst | ||
105 : | edgomez | 1382 | test eax, eax |
106 : | Isibaar | 262 | mov eax, [esp+ 8] ; Src |
107 : | mov edx, [esp+12] ; stride | ||
108 : | |||
109 : | jnz near .rounding1 | ||
110 : | |||
111 : | COPY_H_3DN_RND0 | ||
112 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
113 : | Isibaar | 262 | COPY_H_3DN_RND0 |
114 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
115 : | Isibaar | 262 | COPY_H_3DN_RND0 |
116 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
117 : | Isibaar | 262 | COPY_H_3DN_RND0 |
118 : | ret | ||
119 : | |||
120 : | .rounding1 | ||
121 : | edgomez | 1382 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
122 : | Isibaar | 262 | movq mm7, [mmx_one] |
123 : | COPY_H_3DN_RND1 | ||
124 : | lea ecx, [ecx+2*edx] | ||
125 : | COPY_H_3DN_RND1 | ||
126 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
127 : | Isibaar | 262 | COPY_H_3DN_RND1 |
128 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
129 : | Isibaar | 262 | COPY_H_3DN_RND1 |
130 : | ret | ||
131 : | |||
132 : | |||
133 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
134 : | Isibaar | 262 | ; |
135 : | ; void interpolate8x8_halfpel_v_3dn(uint8_t * const dst, | ||
136 : | edgomez | 1382 | ; const uint8_t * const src, |
137 : | ; const uint32_t stride, | ||
138 : | ; const uint32_t rounding); | ||
139 : | Isibaar | 262 | ; |
140 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
141 : | Isibaar | 262 | |
142 : | %macro COPY_V_3DN_RND0 0 | ||
143 : | edgomez | 1382 | movq mm0, [eax] |
144 : | movq mm1, [eax+edx] | ||
145 : | Isibaar | 262 | pavgusb mm0, mm1 |
146 : | pavgusb mm1, [eax+2*edx] | ||
147 : | edgomez | 1382 | lea eax, [eax+2*edx] |
148 : | movq [ecx], mm0 | ||
149 : | movq [ecx+edx], mm1 | ||
150 : | Isibaar | 262 | %endmacro |
151 : | |||
152 : | %macro COPY_V_3DN_RND1 0 | ||
153 : | movq mm0, mm2 | ||
154 : | movq mm1, [eax] | ||
155 : | movq mm2, [eax+edx] | ||
156 : | edgomez | 1382 | lea eax, [eax+2*edx] |
157 : | Isibaar | 262 | movq mm4, mm0 |
158 : | movq mm5, mm1 | ||
159 : | pavgusb mm0, mm1 | ||
160 : | edgomez | 1382 | pxor mm4, mm1 |
161 : | Isibaar | 262 | pavgusb mm1, mm2 |
162 : | pxor mm5, mm2 | ||
163 : | edgomez | 1382 | pand mm4, mm7 ; lsb's of (i^j)... |
164 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
165 : | psubb mm0, mm4 ; ...are substracted from result of pavgusb | ||
166 : | Isibaar | 262 | movq [ecx], mm0 |
167 : | edgomez | 1382 | psubb mm1, mm5 ; ...are substracted from result of pavgusb |
168 : | Isibaar | 262 | movq [ecx+edx], mm1 |
169 : | %endmacro | ||
170 : | |||
171 : | edgomez | 1382 | ALIGN 16 |
172 : | Isibaar | 262 | interpolate8x8_halfpel_v_3dn: |
173 : | |||
174 : | mov eax, [esp+16] ; rounding | ||
175 : | mov ecx, [esp+ 4] ; Dst | ||
176 : | test eax,eax | ||
177 : | mov eax, [esp+ 8] ; Src | ||
178 : | mov edx, [esp+12] ; stride | ||
179 : | |||
180 : | ; we process 2 line at a time | ||
181 : | |||
182 : | jnz near .rounding1 | ||
183 : | |||
184 : | COPY_V_3DN_RND0 | ||
185 : | lea ecx, [ecx+2*edx] | ||
186 : | COPY_V_3DN_RND0 | ||
187 : | lea ecx, [ecx+2*edx] | ||
188 : | COPY_V_3DN_RND0 | ||
189 : | lea ecx, [ecx+2*edx] | ||
190 : | COPY_V_3DN_RND0 | ||
191 : | ret | ||
192 : | |||
193 : | .rounding1 | ||
194 : | edgomez | 1382 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
195 : | Isibaar | 262 | movq mm7, [mmx_one] |
196 : | edgomez | 1382 | movq mm2, [eax] ; loop invariant |
197 : | Isibaar | 262 | add eax, edx |
198 : | |||
199 : | COPY_V_3DN_RND1 | ||
200 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
201 : | Isibaar | 262 | COPY_V_3DN_RND1 |
202 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
203 : | Isibaar | 262 | COPY_V_3DN_RND1 |
204 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
205 : | Isibaar | 262 | COPY_V_3DN_RND1 |
206 : | ret | ||
207 : | |||
208 : | |||
209 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
210 : | Isibaar | 262 | ; |
211 : | ; void interpolate8x8_halfpel_hv_3dn(uint8_t * const dst, | ||
212 : | edgomez | 1382 | ; const uint8_t * const src, |
213 : | ; const uint32_t stride, | ||
214 : | ; const uint32_t rounding); | ||
215 : | Isibaar | 262 | ; |
216 : | ; | ||
217 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
218 : | Isibaar | 262 | |
219 : | ; The trick is to correct the result of 'pavgusb' with some combination of the | ||
220 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). | ||
221 : | ; The boolean relations are: | ||
222 : | edgomez | 1382 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
223 : | Isibaar | 262 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
224 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
225 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
226 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
227 : | |||
228 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
229 : | |||
230 : | %macro COPY_HV_3DN_RND0 0 | ||
231 : | edgomez | 1382 | lea eax, [eax+edx] |
232 : | Isibaar | 262 | |
233 : | edgomez | 1382 | movq mm0, [eax] |
234 : | movq mm1, [eax+1] | ||
235 : | Isibaar | 262 | |
236 : | edgomez | 1382 | movq mm6, mm0 |
237 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
238 : | lea eax, [eax+edx] | ||
239 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
240 : | Isibaar | 262 | |
241 : | edgomez | 1382 | por mm3, mm1 ; ij |= jk |
242 : | movq mm6, mm2 | ||
243 : | pxor mm6, mm0 ; mm6 = s^t | ||
244 : | pand mm3, mm6 ; (ij|jk) &= st | ||
245 : | pavgusb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
246 : | pand mm3, mm7 ; mask lsb | ||
247 : | psubb mm2, mm3 ; apply. | ||
248 : | Isibaar | 262 | |
249 : | edgomez | 1382 | movq [ecx], mm2 |
250 : | Isibaar | 262 | |
251 : | edgomez | 1382 | movq mm2, [eax] |
252 : | movq mm3, [eax+1] | ||
253 : | movq mm6, mm2 | ||
254 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
255 : | lea ecx, [ecx+edx] | ||
256 : | pxor mm3, mm6 ; preserved for next iteration | ||
257 : | Isibaar | 262 | |
258 : | edgomez | 1382 | por mm1, mm3 |
259 : | movq mm6, mm0 | ||
260 : | pxor mm6, mm2 | ||
261 : | pand mm1, mm6 | ||
262 : | pavgusb mm0, mm2 | ||
263 : | Isibaar | 262 | |
264 : | edgomez | 1382 | pand mm1, mm7 |
265 : | psubb mm0, mm1 | ||
266 : | Isibaar | 262 | |
267 : | edgomez | 1382 | movq [ecx], mm0 |
268 : | Isibaar | 262 | %endmacro |
269 : | |||
270 : | %macro COPY_HV_3DN_RND1 0 | ||
271 : | edgomez | 1382 | lea eax,[eax+edx] |
272 : | Isibaar | 262 | |
273 : | edgomez | 1382 | movq mm0, [eax] |
274 : | movq mm1, [eax+1] | ||
275 : | Isibaar | 262 | |
276 : | edgomez | 1382 | movq mm6, mm0 |
277 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
278 : | lea eax, [eax+edx] | ||
279 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
280 : | Isibaar | 262 | |
281 : | edgomez | 1382 | pand mm3, mm1 |
282 : | movq mm6, mm2 | ||
283 : | pxor mm6, mm0 | ||
284 : | por mm3, mm6 | ||
285 : | pavgusb mm2, mm0 | ||
286 : | pand mm3, mm7 | ||
287 : | psubb mm2, mm3 | ||
288 : | Isibaar | 262 | |
289 : | edgomez | 1382 | movq [ecx], mm2 |
290 : | Isibaar | 262 | |
291 : | edgomez | 1382 | movq mm2, [eax] |
292 : | movq mm3, [eax+1] | ||
293 : | movq mm6, mm2 | ||
294 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
295 : | lea ecx, [ecx+edx] | ||
296 : | pxor mm3, mm6 ; preserved for next iteration | ||
297 : | Isibaar | 262 | |
298 : | edgomez | 1382 | pand mm1, mm3 |
299 : | movq mm6, mm0 | ||
300 : | pxor mm6, mm2 | ||
301 : | por mm1, mm6 | ||
302 : | pavgusb mm0, mm2 | ||
303 : | pand mm1, mm7 | ||
304 : | psubb mm0, mm1 | ||
305 : | Isibaar | 262 | |
306 : | edgomez | 1382 | movq [ecx], mm0 |
307 : | Isibaar | 262 | %endmacro |
308 : | |||
309 : | edgomez | 1382 | ALIGN 16 |
310 : | Isibaar | 262 | interpolate8x8_halfpel_hv_3dn |
311 : | mov eax, [esp+16] ; rounding | ||
312 : | mov ecx, [esp+ 4] ; Dst | ||
313 : | edgomez | 1382 | test eax, eax |
314 : | Isibaar | 262 | mov eax, [esp+ 8] ; Src |
315 : | mov edx, [esp+12] ; stride | ||
316 : | |||
317 : | movq mm7, [mmx_one] | ||
318 : | |||
319 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
320 : | movq mm2, [eax] | ||
321 : | movq mm3, [eax+1] | ||
322 : | movq mm6, mm2 | ||
323 : | pavgusb mm2, mm3 | ||
324 : | edgomez | 1382 | pxor mm3, mm6 ; mm2/mm3 ready |
325 : | Isibaar | 262 | |
326 : | jnz near .rounding1 | ||
327 : | |||
328 : | COPY_HV_3DN_RND0 | ||
329 : | add ecx, edx | ||
330 : | COPY_HV_3DN_RND0 | ||
331 : | add ecx, edx | ||
332 : | COPY_HV_3DN_RND0 | ||
333 : | add ecx, edx | ||
334 : | COPY_HV_3DN_RND0 | ||
335 : | ret | ||
336 : | |||
337 : | .rounding1 | ||
338 : | COPY_HV_3DN_RND1 | ||
339 : | add ecx, edx | ||
340 : | COPY_HV_3DN_RND1 | ||
341 : | add ecx, edx | ||
342 : | COPY_HV_3DN_RND1 | ||
343 : | add ecx, edx | ||
344 : | COPY_HV_3DN_RND1 | ||
345 : | edgomez | 1382 | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |