Parent Directory | Revision Log
Revision 1793 - (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 | 1535 | %ifdef MARK_FUNCS |
31 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
32 : | %define %1 _%1:function %1.endfunc-%1 | ||
33 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
34 : | edgomez | 1535 | %else |
35 : | global _%1 | ||
36 : | %define %1 _%1 | ||
37 : | Isibaar | 1793 | %define ENDFUNC |
38 : | edgomez | 1535 | %endif |
39 : | Isibaar | 262 | %else |
40 : | edgomez | 1535 | %ifdef MARK_FUNCS |
41 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
42 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
43 : | edgomez | 1535 | %else |
44 : | global %1 | ||
45 : | Isibaar | 1793 | %define ENDFUNC |
46 : | edgomez | 1535 | %endif |
47 : | Isibaar | 262 | %endif |
48 : | %endmacro | ||
49 : | |||
50 : | edgomez | 1382 | ;============================================================================= |
51 : | ; Read Only data | ||
52 : | ;============================================================================= | ||
53 : | Isibaar | 262 | |
54 : | edgomez | 1382 | %ifdef FORMAT_COFF |
55 : | edgomez | 1519 | SECTION .rodata |
56 : | edgomez | 1382 | %else |
57 : | edgomez | 1519 | SECTION .rodata align=16 |
58 : | edgomez | 1382 | %endif |
59 : | Isibaar | 262 | |
60 : | edgomez | 1382 | ALIGN 16 |
61 : | mmx_one: | ||
62 : | times 8 db 1 | ||
63 : | Isibaar | 262 | |
64 : | edgomez | 1382 | ;============================================================================= |
65 : | ; Code | ||
66 : | ;============================================================================= | ||
67 : | Isibaar | 262 | |
68 : | edgomez | 1382 | SECTION .text |
69 : | Isibaar | 262 | |
70 : | cglobal interpolate8x8_halfpel_h_3dn | ||
71 : | cglobal interpolate8x8_halfpel_v_3dn | ||
72 : | cglobal interpolate8x8_halfpel_hv_3dn | ||
73 : | |||
74 : | suxen_drol | 1632 | cglobal interpolate8x4_halfpel_h_3dn |
75 : | cglobal interpolate8x4_halfpel_v_3dn | ||
76 : | cglobal interpolate8x4_halfpel_hv_3dn | ||
77 : | |||
78 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
79 : | Isibaar | 262 | ; |
80 : | ; void interpolate8x8_halfpel_h_3dn(uint8_t * const dst, | ||
81 : | edgomez | 1382 | ; const uint8_t * const src, |
82 : | ; const uint32_t stride, | ||
83 : | ; const uint32_t rounding); | ||
84 : | Isibaar | 262 | ; |
85 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
86 : | Isibaar | 262 | |
87 : | %macro COPY_H_3DN_RND0 0 | ||
88 : | edgomez | 1382 | movq mm0, [eax] |
89 : | Isibaar | 262 | pavgusb mm0, [eax+1] |
90 : | edgomez | 1382 | movq mm1, [eax+edx] |
91 : | Isibaar | 262 | pavgusb mm1, [eax+edx+1] |
92 : | edgomez | 1382 | lea eax, [eax+2*edx] |
93 : | movq [ecx], mm0 | ||
94 : | movq [ecx+edx], mm1 | ||
95 : | Isibaar | 262 | %endmacro |
96 : | |||
97 : | %macro COPY_H_3DN_RND1 0 | ||
98 : | movq mm0, [eax] | ||
99 : | movq mm1, [eax+edx] | ||
100 : | movq mm4, mm0 | ||
101 : | movq mm5, mm1 | ||
102 : | edgomez | 1382 | movq mm2, [eax+1] |
103 : | Isibaar | 262 | movq mm3, [eax+edx+1] |
104 : | pavgusb mm0, mm2 | ||
105 : | pxor mm2, mm4 | ||
106 : | pavgusb mm1, mm3 | ||
107 : | edgomez | 1382 | lea eax, [eax+2*edx] |
108 : | Isibaar | 262 | pxor mm3, mm5 |
109 : | pand mm2, mm7 | ||
110 : | pand mm3, mm7 | ||
111 : | psubb mm0, mm2 | ||
112 : | movq [ecx], mm0 | ||
113 : | psubb mm1, mm3 | ||
114 : | movq [ecx+edx], mm1 | ||
115 : | %endmacro | ||
116 : | |||
117 : | edgomez | 1382 | ALIGN 16 |
118 : | Isibaar | 262 | interpolate8x8_halfpel_h_3dn: |
119 : | |||
120 : | mov eax, [esp+16] ; rounding | ||
121 : | mov ecx, [esp+ 4] ; Dst | ||
122 : | edgomez | 1382 | test eax, eax |
123 : | Isibaar | 262 | mov eax, [esp+ 8] ; Src |
124 : | mov edx, [esp+12] ; stride | ||
125 : | |||
126 : | jnz near .rounding1 | ||
127 : | |||
128 : | COPY_H_3DN_RND0 | ||
129 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
130 : | Isibaar | 262 | COPY_H_3DN_RND0 |
131 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
132 : | Isibaar | 262 | COPY_H_3DN_RND0 |
133 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
134 : | Isibaar | 262 | COPY_H_3DN_RND0 |
135 : | ret | ||
136 : | |||
137 : | Isibaar | 1793 | .rounding1: |
138 : | edgomez | 1382 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
139 : | Isibaar | 262 | movq mm7, [mmx_one] |
140 : | COPY_H_3DN_RND1 | ||
141 : | lea ecx, [ecx+2*edx] | ||
142 : | COPY_H_3DN_RND1 | ||
143 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
144 : | Isibaar | 262 | COPY_H_3DN_RND1 |
145 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
146 : | Isibaar | 262 | COPY_H_3DN_RND1 |
147 : | ret | ||
148 : | Isibaar | 1793 | ENDFUNC |
149 : | Isibaar | 262 | |
150 : | |||
151 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
152 : | Isibaar | 262 | ; |
153 : | ; void interpolate8x8_halfpel_v_3dn(uint8_t * const dst, | ||
154 : | edgomez | 1382 | ; const uint8_t * const src, |
155 : | ; const uint32_t stride, | ||
156 : | ; const uint32_t rounding); | ||
157 : | Isibaar | 262 | ; |
158 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
159 : | Isibaar | 262 | |
160 : | %macro COPY_V_3DN_RND0 0 | ||
161 : | edgomez | 1382 | movq mm0, [eax] |
162 : | movq mm1, [eax+edx] | ||
163 : | Isibaar | 262 | pavgusb mm0, mm1 |
164 : | pavgusb mm1, [eax+2*edx] | ||
165 : | edgomez | 1382 | lea eax, [eax+2*edx] |
166 : | movq [ecx], mm0 | ||
167 : | movq [ecx+edx], mm1 | ||
168 : | Isibaar | 262 | %endmacro |
169 : | |||
170 : | %macro COPY_V_3DN_RND1 0 | ||
171 : | movq mm0, mm2 | ||
172 : | movq mm1, [eax] | ||
173 : | movq mm2, [eax+edx] | ||
174 : | edgomez | 1382 | lea eax, [eax+2*edx] |
175 : | Isibaar | 262 | movq mm4, mm0 |
176 : | movq mm5, mm1 | ||
177 : | pavgusb mm0, mm1 | ||
178 : | edgomez | 1382 | pxor mm4, mm1 |
179 : | Isibaar | 262 | pavgusb mm1, mm2 |
180 : | pxor mm5, mm2 | ||
181 : | edgomez | 1382 | pand mm4, mm7 ; lsb's of (i^j)... |
182 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
183 : | psubb mm0, mm4 ; ...are substracted from result of pavgusb | ||
184 : | Isibaar | 262 | movq [ecx], mm0 |
185 : | edgomez | 1382 | psubb mm1, mm5 ; ...are substracted from result of pavgusb |
186 : | Isibaar | 262 | movq [ecx+edx], mm1 |
187 : | %endmacro | ||
188 : | |||
189 : | edgomez | 1382 | ALIGN 16 |
190 : | Isibaar | 262 | interpolate8x8_halfpel_v_3dn: |
191 : | |||
192 : | mov eax, [esp+16] ; rounding | ||
193 : | mov ecx, [esp+ 4] ; Dst | ||
194 : | test eax,eax | ||
195 : | mov eax, [esp+ 8] ; Src | ||
196 : | mov edx, [esp+12] ; stride | ||
197 : | |||
198 : | ; we process 2 line at a time | ||
199 : | |||
200 : | jnz near .rounding1 | ||
201 : | |||
202 : | COPY_V_3DN_RND0 | ||
203 : | lea ecx, [ecx+2*edx] | ||
204 : | COPY_V_3DN_RND0 | ||
205 : | lea ecx, [ecx+2*edx] | ||
206 : | COPY_V_3DN_RND0 | ||
207 : | lea ecx, [ecx+2*edx] | ||
208 : | COPY_V_3DN_RND0 | ||
209 : | ret | ||
210 : | |||
211 : | Isibaar | 1793 | .rounding1: |
212 : | edgomez | 1382 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
213 : | Isibaar | 262 | movq mm7, [mmx_one] |
214 : | edgomez | 1382 | movq mm2, [eax] ; loop invariant |
215 : | Isibaar | 262 | add eax, edx |
216 : | |||
217 : | COPY_V_3DN_RND1 | ||
218 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
219 : | Isibaar | 262 | COPY_V_3DN_RND1 |
220 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
221 : | Isibaar | 262 | COPY_V_3DN_RND1 |
222 : | edgomez | 1382 | lea ecx, [ecx+2*edx] |
223 : | Isibaar | 262 | COPY_V_3DN_RND1 |
224 : | ret | ||
225 : | Isibaar | 1793 | ENDFUNC |
226 : | Isibaar | 262 | |
227 : | |||
228 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
229 : | Isibaar | 262 | ; |
230 : | ; void interpolate8x8_halfpel_hv_3dn(uint8_t * const dst, | ||
231 : | edgomez | 1382 | ; const uint8_t * const src, |
232 : | ; const uint32_t stride, | ||
233 : | ; const uint32_t rounding); | ||
234 : | Isibaar | 262 | ; |
235 : | ; | ||
236 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
237 : | Isibaar | 262 | |
238 : | ; The trick is to correct the result of 'pavgusb' with some combination of the | ||
239 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). | ||
240 : | ; The boolean relations are: | ||
241 : | edgomez | 1382 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
242 : | Isibaar | 262 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
243 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
244 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
245 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
246 : | |||
247 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
248 : | |||
249 : | %macro COPY_HV_3DN_RND0 0 | ||
250 : | edgomez | 1382 | lea eax, [eax+edx] |
251 : | Isibaar | 262 | |
252 : | edgomez | 1382 | movq mm0, [eax] |
253 : | movq mm1, [eax+1] | ||
254 : | Isibaar | 262 | |
255 : | edgomez | 1382 | movq mm6, mm0 |
256 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
257 : | lea eax, [eax+edx] | ||
258 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
259 : | Isibaar | 262 | |
260 : | edgomez | 1382 | por mm3, mm1 ; ij |= jk |
261 : | movq mm6, mm2 | ||
262 : | pxor mm6, mm0 ; mm6 = s^t | ||
263 : | pand mm3, mm6 ; (ij|jk) &= st | ||
264 : | pavgusb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
265 : | pand mm3, mm7 ; mask lsb | ||
266 : | psubb mm2, mm3 ; apply. | ||
267 : | Isibaar | 262 | |
268 : | edgomez | 1382 | movq [ecx], mm2 |
269 : | Isibaar | 262 | |
270 : | edgomez | 1382 | movq mm2, [eax] |
271 : | movq mm3, [eax+1] | ||
272 : | movq mm6, mm2 | ||
273 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
274 : | lea ecx, [ecx+edx] | ||
275 : | pxor mm3, mm6 ; preserved for next iteration | ||
276 : | Isibaar | 262 | |
277 : | edgomez | 1382 | por mm1, mm3 |
278 : | movq mm6, mm0 | ||
279 : | pxor mm6, mm2 | ||
280 : | pand mm1, mm6 | ||
281 : | pavgusb mm0, mm2 | ||
282 : | Isibaar | 262 | |
283 : | edgomez | 1382 | pand mm1, mm7 |
284 : | psubb mm0, mm1 | ||
285 : | Isibaar | 262 | |
286 : | edgomez | 1382 | movq [ecx], mm0 |
287 : | Isibaar | 262 | %endmacro |
288 : | |||
289 : | %macro COPY_HV_3DN_RND1 0 | ||
290 : | edgomez | 1382 | lea eax,[eax+edx] |
291 : | Isibaar | 262 | |
292 : | edgomez | 1382 | movq mm0, [eax] |
293 : | movq mm1, [eax+1] | ||
294 : | Isibaar | 262 | |
295 : | edgomez | 1382 | movq mm6, mm0 |
296 : | pavgusb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
297 : | lea eax, [eax+edx] | ||
298 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
299 : | Isibaar | 262 | |
300 : | edgomez | 1382 | pand mm3, mm1 |
301 : | movq mm6, mm2 | ||
302 : | pxor mm6, mm0 | ||
303 : | por mm3, mm6 | ||
304 : | pavgusb mm2, mm0 | ||
305 : | pand mm3, mm7 | ||
306 : | psubb mm2, mm3 | ||
307 : | Isibaar | 262 | |
308 : | edgomez | 1382 | movq [ecx], mm2 |
309 : | Isibaar | 262 | |
310 : | edgomez | 1382 | movq mm2, [eax] |
311 : | movq mm3, [eax+1] | ||
312 : | movq mm6, mm2 | ||
313 : | pavgusb mm2, mm3 ; preserved for next iteration | ||
314 : | lea ecx, [ecx+edx] | ||
315 : | pxor mm3, mm6 ; preserved for next iteration | ||
316 : | Isibaar | 262 | |
317 : | edgomez | 1382 | pand mm1, mm3 |
318 : | movq mm6, mm0 | ||
319 : | pxor mm6, mm2 | ||
320 : | por mm1, mm6 | ||
321 : | pavgusb mm0, mm2 | ||
322 : | pand mm1, mm7 | ||
323 : | psubb mm0, mm1 | ||
324 : | Isibaar | 262 | |
325 : | edgomez | 1382 | movq [ecx], mm0 |
326 : | Isibaar | 262 | %endmacro |
327 : | |||
328 : | edgomez | 1382 | ALIGN 16 |
329 : | Isibaar | 1793 | interpolate8x8_halfpel_hv_3dn: |
330 : | Isibaar | 262 | mov eax, [esp+16] ; rounding |
331 : | mov ecx, [esp+ 4] ; Dst | ||
332 : | edgomez | 1382 | test eax, eax |
333 : | Isibaar | 262 | mov eax, [esp+ 8] ; Src |
334 : | mov edx, [esp+12] ; stride | ||
335 : | |||
336 : | movq mm7, [mmx_one] | ||
337 : | |||
338 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
339 : | movq mm2, [eax] | ||
340 : | movq mm3, [eax+1] | ||
341 : | movq mm6, mm2 | ||
342 : | pavgusb mm2, mm3 | ||
343 : | edgomez | 1382 | pxor mm3, mm6 ; mm2/mm3 ready |
344 : | Isibaar | 262 | |
345 : | jnz near .rounding1 | ||
346 : | |||
347 : | COPY_HV_3DN_RND0 | ||
348 : | add ecx, edx | ||
349 : | COPY_HV_3DN_RND0 | ||
350 : | add ecx, edx | ||
351 : | COPY_HV_3DN_RND0 | ||
352 : | add ecx, edx | ||
353 : | COPY_HV_3DN_RND0 | ||
354 : | ret | ||
355 : | |||
356 : | Isibaar | 1793 | .rounding1: |
357 : | Isibaar | 262 | COPY_HV_3DN_RND1 |
358 : | add ecx, edx | ||
359 : | COPY_HV_3DN_RND1 | ||
360 : | add ecx, edx | ||
361 : | COPY_HV_3DN_RND1 | ||
362 : | add ecx, edx | ||
363 : | COPY_HV_3DN_RND1 | ||
364 : | edgomez | 1382 | ret |
365 : | Isibaar | 1793 | ENDFUNC |
366 : | edgomez | 1540 | |
367 : | suxen_drol | 1632 | ;----------------------------------------------------------------------------- |
368 : | ; | ||
369 : | ; void interpolate8x4_halfpel_h_3dn(uint8_t * const dst, | ||
370 : | ; const uint8_t * const src, | ||
371 : | ; const uint32_t stride, | ||
372 : | ; const uint32_t rounding); | ||
373 : | ; | ||
374 : | ;----------------------------------------------------------------------------- | ||
375 : | |||
376 : | ALIGN 16 | ||
377 : | interpolate8x4_halfpel_h_3dn: | ||
378 : | |||
379 : | mov eax, [esp+16] ; rounding | ||
380 : | mov ecx, [esp+ 4] ; Dst | ||
381 : | test eax, eax | ||
382 : | mov eax, [esp+ 8] ; Src | ||
383 : | mov edx, [esp+12] ; stride | ||
384 : | |||
385 : | jnz near .rounding1 | ||
386 : | |||
387 : | COPY_H_3DN_RND0 | ||
388 : | lea ecx, [ecx+2*edx] | ||
389 : | COPY_H_3DN_RND0 | ||
390 : | ret | ||
391 : | |||
392 : | Isibaar | 1793 | .rounding1: |
393 : | suxen_drol | 1632 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
394 : | movq mm7, [mmx_one] | ||
395 : | COPY_H_3DN_RND1 | ||
396 : | lea ecx, [ecx+2*edx] | ||
397 : | COPY_H_3DN_RND1 | ||
398 : | ret | ||
399 : | Isibaar | 1793 | ENDFUNC |
400 : | suxen_drol | 1632 | |
401 : | |||
402 : | ;----------------------------------------------------------------------------- | ||
403 : | ; | ||
404 : | ; void interpolate8x4_halfpel_v_3dn(uint8_t * const dst, | ||
405 : | ; const uint8_t * const src, | ||
406 : | ; const uint32_t stride, | ||
407 : | ; const uint32_t rounding); | ||
408 : | ; | ||
409 : | ;----------------------------------------------------------------------------- | ||
410 : | |||
411 : | ALIGN 16 | ||
412 : | interpolate8x4_halfpel_v_3dn: | ||
413 : | |||
414 : | mov eax, [esp+16] ; rounding | ||
415 : | mov ecx, [esp+ 4] ; Dst | ||
416 : | test eax,eax | ||
417 : | mov eax, [esp+ 8] ; Src | ||
418 : | mov edx, [esp+12] ; stride | ||
419 : | |||
420 : | ; we process 2 line at a time | ||
421 : | |||
422 : | jnz near .rounding1 | ||
423 : | |||
424 : | COPY_V_3DN_RND0 | ||
425 : | lea ecx, [ecx+2*edx] | ||
426 : | COPY_V_3DN_RND0 | ||
427 : | ret | ||
428 : | |||
429 : | Isibaar | 1793 | .rounding1: |
430 : | suxen_drol | 1632 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
431 : | movq mm7, [mmx_one] | ||
432 : | movq mm2, [eax] ; loop invariant | ||
433 : | add eax, edx | ||
434 : | |||
435 : | COPY_V_3DN_RND1 | ||
436 : | lea ecx, [ecx+2*edx] | ||
437 : | COPY_V_3DN_RND1 | ||
438 : | ret | ||
439 : | Isibaar | 1793 | ENDFUNC |
440 : | suxen_drol | 1632 | |
441 : | |||
442 : | ;----------------------------------------------------------------------------- | ||
443 : | ; | ||
444 : | ; void interpolate8x4_halfpel_hv_3dn(uint8_t * const dst, | ||
445 : | ; const uint8_t * const src, | ||
446 : | ; const uint32_t stride, | ||
447 : | ; const uint32_t rounding); | ||
448 : | ; | ||
449 : | ; | ||
450 : | ;----------------------------------------------------------------------------- | ||
451 : | |||
452 : | ; The trick is to correct the result of 'pavgusb' with some combination of the | ||
453 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). | ||
454 : | ; The boolean relations are: | ||
455 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
456 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
457 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
458 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
459 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
460 : | |||
461 : | ALIGN 16 | ||
462 : | Isibaar | 1793 | interpolate8x4_halfpel_hv_3dn: |
463 : | suxen_drol | 1632 | mov eax, [esp+16] ; rounding |
464 : | mov ecx, [esp+ 4] ; Dst | ||
465 : | test eax, eax | ||
466 : | mov eax, [esp+ 8] ; Src | ||
467 : | mov edx, [esp+12] ; stride | ||
468 : | |||
469 : | movq mm7, [mmx_one] | ||
470 : | |||
471 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
472 : | movq mm2, [eax] | ||
473 : | movq mm3, [eax+1] | ||
474 : | movq mm6, mm2 | ||
475 : | pavgusb mm2, mm3 | ||
476 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
477 : | |||
478 : | jnz near .rounding1 | ||
479 : | |||
480 : | COPY_HV_3DN_RND0 | ||
481 : | add ecx, edx | ||
482 : | COPY_HV_3DN_RND0 | ||
483 : | ret | ||
484 : | |||
485 : | Isibaar | 1793 | .rounding1: |
486 : | suxen_drol | 1632 | COPY_HV_3DN_RND1 |
487 : | add ecx, edx | ||
488 : | COPY_HV_3DN_RND1 | ||
489 : | ret | ||
490 : | Isibaar | 1793 | ENDFUNC |
491 : | suxen_drol | 1632 | |
492 : | Isibaar | 1790 | |
493 : | %ifidn __OUTPUT_FORMAT__,elf | ||
494 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
495 : | %endif | ||
496 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |