Parent Directory | Revision Log
Revision 1.8 - (view) (download)
1 : | edgomez | 1.5 | ;/***************************************************************************** |
2 : | Isibaar | 1.1 | ; * |
3 : | edgomez | 1.5 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - 3dnow 8x8 block-based halfpel interpolation - | ||
5 : | Isibaar | 1.1 | ; * |
6 : | edgomez | 1.5 | ; * 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 | 1.1 | ; * |
10 : | edgomez | 1.5 | ; * 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 : | edgomez | 1.3 | ; * |
15 : | edgomez | 1.5 | ; * 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 | 1.1 | ; * |
20 : | edgomez | 1.5 | ; * 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 : | edgomez | 1.4 | ; * |
24 : | edgomez | 1.5 | ; ****************************************************************************/ |
25 : | edgomez | 1.4 | |
26 : | edgomez | 1.5 | BITS 32 |
27 : | Isibaar | 1.1 | |
28 : | edgomez | 1.5 | %macro cglobal 1 |
29 : | Isibaar | 1.1 | %ifdef PREFIX |
30 : | edgomez | 1.7 | %ifdef MARK_FUNCS |
31 : | edgomez | 1.8 | global _%1:function %1.endfunc-%1 |
32 : | %define %1 _%1:function %1.endfunc-%1 | ||
33 : | edgomez | 1.7 | %else |
34 : | global _%1 | ||
35 : | %define %1 _%1 | ||
36 : | %endif | ||
37 : | Isibaar | 1.1 | %else |
38 : | edgomez | 1.7 | %ifdef MARK_FUNCS |
39 : | edgomez | 1.8 | global %1:function %1.endfunc-%1 |
40 : | edgomez | 1.7 | %else |
41 : | global %1 | ||
42 : | %endif | ||
43 : | Isibaar | 1.1 | %endif |
44 : | %endmacro | ||
45 : | |||
46 : | edgomez | 1.5 | ;============================================================================= |
47 : | ; Read Only data | ||
48 : | ;============================================================================= | ||
49 : | |||
50 : | %ifdef FORMAT_COFF | ||
51 : | edgomez | 1.6 | SECTION .rodata |
52 : | edgomez | 1.5 | %else |
53 : | edgomez | 1.6 | SECTION .rodata align=16 |
54 : | edgomez | 1.5 | %endif |
55 : | |||
56 : | ALIGN 16 | ||
57 : | mmx_one: | ||
58 : | times 8 db 1 | ||
59 : | |||
60 : | ;============================================================================= | ||
61 : | ; Code | ||
62 : | ;============================================================================= | ||
63 : | Isibaar | 1.1 | |
64 : | edgomez | 1.5 | SECTION .text |
65 : | Isibaar | 1.1 | |
66 : | cglobal interpolate8x8_halfpel_h_3dn | ||
67 : | cglobal interpolate8x8_halfpel_v_3dn | ||
68 : | cglobal interpolate8x8_halfpel_hv_3dn | ||
69 : | |||
70 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
71 : | Isibaar | 1.1 | ; |
72 : | ; void interpolate8x8_halfpel_h_3dn(uint8_t * const dst, | ||
73 : | edgomez | 1.5 | ; const uint8_t * const src, |
74 : | ; const uint32_t stride, | ||
75 : | ; const uint32_t rounding); | ||
76 : | Isibaar | 1.1 | ; |
77 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
78 : | Isibaar | 1.1 | |
79 : | %macro COPY_H_3DN_RND0 0 | ||
80 : | edgomez | 1.5 | movq mm0, [eax] |
81 : | Isibaar | 1.1 | pavgusb mm0, [eax+1] |
82 : | edgomez | 1.5 | movq mm1, [eax+edx] |
83 : | Isibaar | 1.1 | pavgusb mm1, [eax+edx+1] |
84 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
85 : | movq [ecx], mm0 | ||
86 : | movq [ecx+edx], mm1 | ||
87 : | Isibaar | 1.1 | %endmacro |
88 : | |||
89 : | %macro COPY_H_3DN_RND1 0 | ||
90 : | movq mm0, [eax] | ||
91 : | movq mm1, [eax+edx] | ||
92 : | movq mm4, mm0 | ||
93 : | movq mm5, mm1 | ||
94 : | edgomez | 1.5 | movq mm2, [eax+1] |
95 : | Isibaar | 1.1 | movq mm3, [eax+edx+1] |
96 : | pavgusb mm0, mm2 | ||
97 : | pxor mm2, mm4 | ||
98 : | pavgusb mm1, mm3 | ||
99 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
100 : | Isibaar | 1.1 | pxor mm3, mm5 |
101 : | pand mm2, mm7 | ||
102 : | pand mm3, mm7 | ||
103 : | psubb mm0, mm2 | ||
104 : | movq [ecx], mm0 | ||
105 : | psubb mm1, mm3 | ||
106 : | movq [ecx+edx], mm1 | ||
107 : | %endmacro | ||
108 : | |||
109 : | edgomez | 1.5 | ALIGN 16 |
110 : | Isibaar | 1.1 | interpolate8x8_halfpel_h_3dn: |
111 : | |||
112 : | mov eax, [esp+16] ; rounding | ||
113 : | mov ecx, [esp+ 4] ; Dst | ||
114 : | edgomez | 1.5 | test eax, eax |
115 : | Isibaar | 1.1 | mov eax, [esp+ 8] ; Src |
116 : | mov edx, [esp+12] ; stride | ||
117 : | |||
118 : | jnz near .rounding1 | ||
119 : | |||
120 : | COPY_H_3DN_RND0 | ||
121 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
122 : | Isibaar | 1.1 | COPY_H_3DN_RND0 |
123 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
124 : | Isibaar | 1.1 | COPY_H_3DN_RND0 |
125 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
126 : | Isibaar | 1.1 | COPY_H_3DN_RND0 |
127 : | ret | ||
128 : | |||
129 : | .rounding1 | ||
130 : | edgomez | 1.5 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
131 : | Isibaar | 1.1 | movq mm7, [mmx_one] |
132 : | COPY_H_3DN_RND1 | ||
133 : | lea ecx, [ecx+2*edx] | ||
134 : | COPY_H_3DN_RND1 | ||
135 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
136 : | Isibaar | 1.1 | COPY_H_3DN_RND1 |
137 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
138 : | Isibaar | 1.1 | COPY_H_3DN_RND1 |
139 : | ret | ||
140 : | edgomez | 1.8 | .endfunc |
141 : | Isibaar | 1.1 | |
142 : | |||
143 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
144 : | Isibaar | 1.1 | ; |
145 : | ; void interpolate8x8_halfpel_v_3dn(uint8_t * const dst, | ||
146 : | edgomez | 1.5 | ; const uint8_t * const src, |
147 : | ; const uint32_t stride, | ||
148 : | ; const uint32_t rounding); | ||
149 : | Isibaar | 1.1 | ; |
150 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
151 : | Isibaar | 1.1 | |
152 : | %macro COPY_V_3DN_RND0 0 | ||
153 : | edgomez | 1.5 | movq mm0, [eax] |
154 : | movq mm1, [eax+edx] | ||
155 : | Isibaar | 1.1 | pavgusb mm0, mm1 |
156 : | pavgusb mm1, [eax+2*edx] | ||
157 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
158 : | movq [ecx], mm0 | ||
159 : | movq [ecx+edx], mm1 | ||
160 : | Isibaar | 1.1 | %endmacro |
161 : | |||
162 : | %macro COPY_V_3DN_RND1 0 | ||
163 : | movq mm0, mm2 | ||
164 : | movq mm1, [eax] | ||
165 : | movq mm2, [eax+edx] | ||
166 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
167 : | Isibaar | 1.1 | movq mm4, mm0 |
168 : | movq mm5, mm1 | ||
169 : | pavgusb mm0, mm1 | ||
170 : | edgomez | 1.5 | pxor mm4, mm1 |
171 : | Isibaar | 1.1 | pavgusb mm1, mm2 |
172 : | pxor mm5, mm2 | ||
173 : | edgomez | 1.5 | pand mm4, mm7 ; lsb's of (i^j)... |
174 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
175 : | psubb mm0, mm4 ; ...are substracted from result of pavgusb | ||
176 : | Isibaar | 1.1 | movq [ecx], mm0 |
177 : | edgomez | 1.5 | psubb mm1, mm5 ; ...are substracted from result of pavgusb |
178 : | Isibaar | 1.1 | movq [ecx+edx], mm1 |
179 : | %endmacro | ||
180 : | |||
181 : | edgomez | 1.5 | ALIGN 16 |
182 : | Isibaar | 1.1 | interpolate8x8_halfpel_v_3dn: |
183 : | |||
184 : | mov eax, [esp+16] ; rounding | ||
185 : | mov ecx, [esp+ 4] ; Dst | ||
186 : | test eax,eax | ||
187 : | mov eax, [esp+ 8] ; Src | ||
188 : | mov edx, [esp+12] ; stride | ||
189 : | |||
190 : | ; we process 2 line at a time | ||
191 : | |||
192 : | jnz near .rounding1 | ||
193 : | |||
194 : | COPY_V_3DN_RND0 | ||
195 : | lea ecx, [ecx+2*edx] | ||
196 : | COPY_V_3DN_RND0 | ||
197 : | lea ecx, [ecx+2*edx] | ||
198 : | COPY_V_3DN_RND0 | ||
199 : | lea ecx, [ecx+2*edx] | ||
200 : | COPY_V_3DN_RND0 | ||
201 : | ret | ||
202 : | |||
203 : | .rounding1 | ||
204 : | edgomez | 1.5 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
205 : | Isibaar | 1.1 | movq mm7, [mmx_one] |
206 : | edgomez | 1.5 | movq mm2, [eax] ; loop invariant |
207 : | Isibaar | 1.1 | add eax, edx |
208 : | |||
209 : | COPY_V_3DN_RND1 | ||
210 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
211 : | Isibaar | 1.1 | COPY_V_3DN_RND1 |
212 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
213 : | Isibaar | 1.1 | COPY_V_3DN_RND1 |
214 : | edgomez | 1.5 | lea ecx, [ecx+2*edx] |
215 : | Isibaar | 1.1 | COPY_V_3DN_RND1 |
216 : | ret | ||
217 : | edgomez | 1.8 | .endfunc |
218 : | Isibaar | 1.1 | |
219 : | |||
220 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
221 : | Isibaar | 1.1 | ; |
222 : | ; void interpolate8x8_halfpel_hv_3dn(uint8_t * const dst, | ||
223 : | edgomez | 1.5 | ; const uint8_t * const src, |
224 : | ; const uint32_t stride, | ||
225 : | ; const uint32_t rounding); | ||
226 : | Isibaar | 1.1 | ; |
227 : | ; | ||
228 : | edgomez | 1.5 | ;----------------------------------------------------------------------------- |
229 : | Isibaar | 1.1 | |
230 : | ; The trick is to correct the result of 'pavgusb' with some combination of the | ||
231 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). | ||
232 : | ; The boolean relations are: | ||
233 : | edgomez | 1.5 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
234 : | Isibaar | 1.1 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
235 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
236 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
237 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
238 : | |||
239 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
240 : | |||
241 : | %macro COPY_HV_3DN_RND0 0 | ||
242 : | edgomez | 1.5 | lea eax, [eax+edx] |
243 : | |||
244 : | movq mm0, [eax] | ||
245 : | movq mm1, [eax+1] | ||
246 : | |||
247 : | movq mm6, mm0 | ||
248 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
249 : | lea eax, [eax+edx] | ||
250 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
251 : | |||
252 : | por mm3, mm1 ; ij |= jk | ||
253 : | movq mm6, mm2 | ||
254 : | pxor mm6, mm0 ; mm6 = s^t | ||
255 : | pand mm3, mm6 ; (ij|jk) &= st | ||
256 : | pavgusb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
257 : | pand mm3, mm7 ; mask lsb | ||
258 : | psubb mm2, mm3 ; apply. | ||
259 : | Isibaar | 1.1 | |
260 : | edgomez | 1.5 | movq [ecx], mm2 |
261 : | Isibaar | 1.1 | |
262 : | edgomez | 1.5 | movq mm2, [eax] |
263 : | movq mm3, [eax+1] | ||
264 : | movq mm6, mm2 | ||
265 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
266 : | lea ecx, [ecx+edx] | ||
267 : | pxor mm3, mm6 ; preserved for next iteration | ||
268 : | |||
269 : | por mm1, mm3 | ||
270 : | movq mm6, mm0 | ||
271 : | pxor mm6, mm2 | ||
272 : | pand mm1, mm6 | ||
273 : | pavgusb mm0, mm2 | ||
274 : | Isibaar | 1.1 | |
275 : | edgomez | 1.5 | pand mm1, mm7 |
276 : | psubb mm0, mm1 | ||
277 : | Isibaar | 1.1 | |
278 : | edgomez | 1.5 | movq [ecx], mm0 |
279 : | Isibaar | 1.1 | %endmacro |
280 : | |||
281 : | %macro COPY_HV_3DN_RND1 0 | ||
282 : | edgomez | 1.5 | lea eax,[eax+edx] |
283 : | |||
284 : | movq mm0, [eax] | ||
285 : | movq mm1, [eax+1] | ||
286 : | |||
287 : | movq mm6, mm0 | ||
288 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
289 : | lea eax, [eax+edx] | ||
290 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
291 : | |||
292 : | pand mm3, mm1 | ||
293 : | movq mm6, mm2 | ||
294 : | pxor mm6, mm0 | ||
295 : | por mm3, mm6 | ||
296 : | pavgusb mm2, mm0 | ||
297 : | pand mm3, mm7 | ||
298 : | psubb mm2, mm3 | ||
299 : | Isibaar | 1.1 | |
300 : | edgomez | 1.5 | movq [ecx], mm2 |
301 : | Isibaar | 1.1 | |
302 : | edgomez | 1.5 | movq mm2, [eax] |
303 : | movq mm3, [eax+1] | ||
304 : | movq mm6, mm2 | ||
305 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
306 : | lea ecx, [ecx+edx] | ||
307 : | pxor mm3, mm6 ; preserved for next iteration | ||
308 : | |||
309 : | pand mm1, mm3 | ||
310 : | movq mm6, mm0 | ||
311 : | pxor mm6, mm2 | ||
312 : | por mm1, mm6 | ||
313 : | pavgusb mm0, mm2 | ||
314 : | pand mm1, mm7 | ||
315 : | psubb mm0, mm1 | ||
316 : | Isibaar | 1.1 | |
317 : | edgomez | 1.5 | movq [ecx], mm0 |
318 : | Isibaar | 1.1 | %endmacro |
319 : | |||
320 : | edgomez | 1.5 | ALIGN 16 |
321 : | Isibaar | 1.1 | interpolate8x8_halfpel_hv_3dn |
322 : | mov eax, [esp+16] ; rounding | ||
323 : | mov ecx, [esp+ 4] ; Dst | ||
324 : | edgomez | 1.5 | test eax, eax |
325 : | Isibaar | 1.1 | mov eax, [esp+ 8] ; Src |
326 : | mov edx, [esp+12] ; stride | ||
327 : | |||
328 : | movq mm7, [mmx_one] | ||
329 : | |||
330 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
331 : | movq mm2, [eax] | ||
332 : | movq mm3, [eax+1] | ||
333 : | movq mm6, mm2 | ||
334 : | pavgusb mm2, mm3 | ||
335 : | edgomez | 1.5 | pxor mm3, mm6 ; mm2/mm3 ready |
336 : | Isibaar | 1.1 | |
337 : | jnz near .rounding1 | ||
338 : | |||
339 : | COPY_HV_3DN_RND0 | ||
340 : | add ecx, edx | ||
341 : | COPY_HV_3DN_RND0 | ||
342 : | add ecx, edx | ||
343 : | COPY_HV_3DN_RND0 | ||
344 : | add ecx, edx | ||
345 : | COPY_HV_3DN_RND0 | ||
346 : | ret | ||
347 : | |||
348 : | .rounding1 | ||
349 : | COPY_HV_3DN_RND1 | ||
350 : | add ecx, edx | ||
351 : | COPY_HV_3DN_RND1 | ||
352 : | add ecx, edx | ||
353 : | COPY_HV_3DN_RND1 | ||
354 : | add ecx, edx | ||
355 : | COPY_HV_3DN_RND1 | ||
356 : | edgomez | 1.5 | ret |
357 : | edgomez | 1.8 | .endfunc |
358 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |