Parent Directory | Revision Log
Revision 1.7 - (view) (download)
1 : | edgomez | 1.5 | ;/***************************************************************************** |
2 : | Isibaar | 1.1 | ; * |
3 : | edgomez | 1.5 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - mmx 8x8 block-based halfpel interpolation - | ||
5 : | Isibaar | 1.1 | ; * |
6 : | edgomez | 1.5 | ; * Copyright(C) 2002 Michael Militzer <isibaar@xvid.org> |
7 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
8 : | Isibaar | 1.1 | ; * |
9 : | edgomez | 1.5 | ; * 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 : | edgomez | 1.3 | ; * |
14 : | edgomez | 1.5 | ; * 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 | 1.1 | ; * |
19 : | edgomez | 1.5 | ; * 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 : | edgomez | 1.4 | ; * |
23 : | edgomez | 1.5 | ; ****************************************************************************/ |
24 : | edgomez | 1.4 | |
25 : | edgomez | 1.5 | BITS 32 |
26 : | Isibaar | 1.1 | |
27 : | edgomez | 1.5 | %macro cglobal 1 |
28 : | Isibaar | 1.1 | %ifdef PREFIX |
29 : | edgomez | 1.5 | global _%1 |
30 : | Isibaar | 1.1 | %define %1 _%1 |
31 : | %else | ||
32 : | global %1 | ||
33 : | %endif | ||
34 : | %endmacro | ||
35 : | |||
36 : | edgomez | 1.5 | ;============================================================================= |
37 : | ; Read only data | ||
38 : | ;============================================================================= | ||
39 : | |||
40 : | %ifdef FORMAT_COFF | ||
41 : | edgomez | 1.6 | SECTION .rodata |
42 : | edgomez | 1.5 | %else |
43 : | edgomez | 1.6 | SECTION .rodata align=16 |
44 : | edgomez | 1.5 | %endif |
45 : | |||
46 : | ALIGN 16 | ||
47 : | mmx_one: | ||
48 : | times 8 db 1 | ||
49 : | Isibaar | 1.1 | |
50 : | edgomez | 1.5 | SECTION .text |
51 : | Isibaar | 1.1 | |
52 : | cglobal interpolate8x8_halfpel_h_xmm | ||
53 : | cglobal interpolate8x8_halfpel_v_xmm | ||
54 : | cglobal interpolate8x8_halfpel_hv_xmm | ||
55 : | |||
56 : | edgomez | 1.7 | cglobal interpolate8x8_halfpel_add_xmm |
57 : | cglobal interpolate8x8_halfpel_h_add_xmm | ||
58 : | cglobal interpolate8x8_halfpel_v_add_xmm | ||
59 : | cglobal interpolate8x8_halfpel_hv_add_xmm | ||
60 : | |||
61 : | Isibaar | 1.1 | ;=========================================================================== |
62 : | ; | ||
63 : | ; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, | ||
64 : | ; const uint8_t * const src, | ||
65 : | ; const uint32_t stride, | ||
66 : | ; const uint32_t rounding); | ||
67 : | ; | ||
68 : | ;=========================================================================== | ||
69 : | |||
70 : | %macro COPY_H_SSE_RND0 0 | ||
71 : | movq mm0, [eax] | ||
72 : | pavgb mm0, [eax+1] | ||
73 : | movq mm1, [eax+edx] | ||
74 : | pavgb mm1, [eax+edx+1] | ||
75 : | lea eax,[eax+2*edx] | ||
76 : | movq [ecx],mm0 | ||
77 : | movq [ecx+edx],mm1 | ||
78 : | %endmacro | ||
79 : | |||
80 : | %macro COPY_H_SSE_RND1 0 | ||
81 : | movq mm0, [eax] | ||
82 : | movq mm1, [eax+edx] | ||
83 : | movq mm4, mm0 | ||
84 : | movq mm5, mm1 | ||
85 : | edgomez | 1.5 | movq mm2, [eax+1] |
86 : | Isibaar | 1.1 | movq mm3, [eax+edx+1] |
87 : | pavgb mm0, mm2 | ||
88 : | pxor mm2, mm4 | ||
89 : | pavgb mm1, mm3 | ||
90 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
91 : | Isibaar | 1.1 | pxor mm3, mm5 |
92 : | pand mm2, mm7 | ||
93 : | pand mm3, mm7 | ||
94 : | psubb mm0, mm2 | ||
95 : | movq [ecx], mm0 | ||
96 : | psubb mm1, mm3 | ||
97 : | edgomez | 1.5 | movq [ecx+edx], mm1 |
98 : | Isibaar | 1.1 | %endmacro |
99 : | |||
100 : | edgomez | 1.5 | ALIGN 16 |
101 : | Isibaar | 1.1 | interpolate8x8_halfpel_h_xmm: |
102 : | |||
103 : | edgomez | 1.5 | mov eax, [esp+16] ; rounding |
104 : | mov ecx, [esp+ 4] ; Dst | ||
105 : | Isibaar | 1.1 | test eax,eax |
106 : | edgomez | 1.5 | mov eax, [esp+ 8] ; Src |
107 : | mov edx, [esp+12] ; stride | ||
108 : | Isibaar | 1.1 | |
109 : | jnz near .rounding1 | ||
110 : | |||
111 : | COPY_H_SSE_RND0 | ||
112 : | lea ecx,[ecx+2*edx] | ||
113 : | COPY_H_SSE_RND0 | ||
114 : | lea ecx,[ecx+2*edx] | ||
115 : | COPY_H_SSE_RND0 | ||
116 : | lea ecx,[ecx+2*edx] | ||
117 : | COPY_H_SSE_RND0 | ||
118 : | ret | ||
119 : | |||
120 : | .rounding1 | ||
121 : | edgomez | 1.5 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
122 : | Isibaar | 1.1 | movq mm7, [mmx_one] |
123 : | COPY_H_SSE_RND1 | ||
124 : | lea ecx, [ecx+2*edx] | ||
125 : | COPY_H_SSE_RND1 | ||
126 : | lea ecx,[ecx+2*edx] | ||
127 : | COPY_H_SSE_RND1 | ||
128 : | lea ecx,[ecx+2*edx] | ||
129 : | COPY_H_SSE_RND1 | ||
130 : | ret | ||
131 : | |||
132 : | ;=========================================================================== | ||
133 : | ; | ||
134 : | ; void interpolate8x8_halfpel_v_xmm(uint8_t * const dst, | ||
135 : | edgomez | 1.5 | ; const uint8_t * const src, |
136 : | ; const uint32_t stride, | ||
137 : | ; const uint32_t rounding); | ||
138 : | Isibaar | 1.1 | ; |
139 : | ;=========================================================================== | ||
140 : | |||
141 : | %macro COPY_V_SSE_RND0 0 | ||
142 : | edgomez | 1.5 | movq mm0, [eax] |
143 : | movq mm1, [eax+edx] | ||
144 : | Isibaar | 1.1 | pavgb mm0, mm1 |
145 : | pavgb mm1, [eax+2*edx] | ||
146 : | edgomez | 1.5 | lea eax, [eax+2*edx] |
147 : | movq [ecx], mm0 | ||
148 : | Isibaar | 1.1 | movq [ecx+edx],mm1 |
149 : | %endmacro | ||
150 : | |||
151 : | %macro COPY_V_SSE_RND1 0 | ||
152 : | movq mm0, mm2 | ||
153 : | movq mm1, [eax] | ||
154 : | movq mm2, [eax+edx] | ||
155 : | lea eax,[eax+2*edx] | ||
156 : | movq mm4, mm0 | ||
157 : | movq mm5, mm1 | ||
158 : | pavgb mm0, mm1 | ||
159 : | edgomez | 1.5 | pxor mm4, mm1 |
160 : | Isibaar | 1.1 | pavgb mm1, mm2 |
161 : | pxor mm5, mm2 | ||
162 : | edgomez | 1.5 | pand mm4, mm7 ; lsb's of (i^j)... |
163 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
164 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
165 : | Isibaar | 1.1 | movq [ecx], mm0 |
166 : | edgomez | 1.5 | psubb mm1, mm5 ; ...are substracted from result of pavgb |
167 : | Isibaar | 1.1 | movq [ecx+edx], mm1 |
168 : | %endmacro | ||
169 : | |||
170 : | edgomez | 1.5 | ALIGN 16 |
171 : | Isibaar | 1.1 | interpolate8x8_halfpel_v_xmm: |
172 : | |||
173 : | mov eax, [esp+16]; rounding | ||
174 : | edgomez | 1.5 | mov ecx, [esp+ 4] ; Dst |
175 : | Isibaar | 1.1 | test eax,eax |
176 : | edgomez | 1.5 | mov eax, [esp+ 8] ; Src |
177 : | mov edx, [esp+12] ; stride | ||
178 : | Isibaar | 1.1 | |
179 : | edgomez | 1.5 | ; we process 2 line at a time |
180 : | Isibaar | 1.1 | jnz near .rounding1 |
181 : | |||
182 : | COPY_V_SSE_RND0 | ||
183 : | lea ecx, [ecx+2*edx] | ||
184 : | COPY_V_SSE_RND0 | ||
185 : | lea ecx, [ecx+2*edx] | ||
186 : | COPY_V_SSE_RND0 | ||
187 : | lea ecx, [ecx+2*edx] | ||
188 : | COPY_V_SSE_RND0 | ||
189 : | ret | ||
190 : | |||
191 : | .rounding1 | ||
192 : | edgomez | 1.5 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
193 : | Isibaar | 1.1 | movq mm7, [mmx_one] |
194 : | edgomez | 1.5 | movq mm2, [eax] ; loop invariant |
195 : | Isibaar | 1.1 | add eax, edx |
196 : | |||
197 : | COPY_V_SSE_RND1 | ||
198 : | lea ecx,[ecx+2*edx] | ||
199 : | COPY_V_SSE_RND1 | ||
200 : | lea ecx,[ecx+2*edx] | ||
201 : | COPY_V_SSE_RND1 | ||
202 : | lea ecx,[ecx+2*edx] | ||
203 : | COPY_V_SSE_RND1 | ||
204 : | ret | ||
205 : | |||
206 : | ;=========================================================================== | ||
207 : | ; | ||
208 : | ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, | ||
209 : | edgomez | 1.5 | ; const uint8_t * const src, |
210 : | ; const uint32_t stride, | ||
211 : | ; const uint32_t rounding); | ||
212 : | Isibaar | 1.1 | ; |
213 : | ; | ||
214 : | ;=========================================================================== | ||
215 : | |||
216 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
217 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
218 : | ; The boolean relations are: | ||
219 : | edgomez | 1.5 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
220 : | Isibaar | 1.1 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
221 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
222 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
223 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
224 : | |||
225 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
226 : | |||
227 : | %macro COPY_HV_SSE_RND0 0 | ||
228 : | edgomez | 1.5 | lea eax, [eax+edx] |
229 : | Isibaar | 1.1 | |
230 : | edgomez | 1.5 | movq mm0, [eax] |
231 : | movq mm1, [eax+1] | ||
232 : | Isibaar | 1.1 | |
233 : | edgomez | 1.5 | movq mm6, mm0 |
234 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
235 : | lea eax, [eax+edx] | ||
236 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
237 : | Isibaar | 1.1 | |
238 : | edgomez | 1.5 | por mm3, mm1 ; ij |= jk |
239 : | movq mm6, mm2 | ||
240 : | pxor mm6, mm0 ; mm6 = s^t | ||
241 : | pand mm3, mm6 ; (ij|jk) &= st | ||
242 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
243 : | pand mm3, mm7 ; mask lsb | ||
244 : | psubb mm2, mm3 ; apply. | ||
245 : | |||
246 : | movq [ecx], mm2 | ||
247 : | Isibaar | 1.1 | |
248 : | edgomez | 1.5 | movq mm2, [eax] |
249 : | movq mm3, [eax+1] | ||
250 : | movq mm6, mm2 | ||
251 : | pavgb mm2, mm3 ; preserved for next iteration | ||
252 : | lea ecx,[ecx+edx] | ||
253 : | pxor mm3, mm6 ; preserved for next iteration | ||
254 : | |||
255 : | por mm1, mm3 | ||
256 : | movq mm6, mm0 | ||
257 : | pxor mm6, mm2 | ||
258 : | pand mm1, mm6 | ||
259 : | pavgb mm0, mm2 | ||
260 : | |||
261 : | pand mm1, mm7 | ||
262 : | psubb mm0, mm1 | ||
263 : | |||
264 : | movq [ecx], mm0 | ||
265 : | Isibaar | 1.1 | %endmacro |
266 : | |||
267 : | %macro COPY_HV_SSE_RND1 0 | ||
268 : | edgomez | 1.5 | lea eax, [eax+edx] |
269 : | |||
270 : | movq mm0, [eax] | ||
271 : | movq mm1, [eax+1] | ||
272 : | Isibaar | 1.1 | |
273 : | edgomez | 1.5 | movq mm6, mm0 |
274 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
275 : | lea eax, [eax+edx] | ||
276 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
277 : | Isibaar | 1.1 | |
278 : | edgomez | 1.5 | pand mm3, mm1 |
279 : | movq mm6, mm2 | ||
280 : | pxor mm6, mm0 | ||
281 : | por mm3, mm6 | ||
282 : | pavgb mm2, mm0 | ||
283 : | pand mm3, mm7 | ||
284 : | psubb mm2, mm3 | ||
285 : | |||
286 : | movq [ecx], mm2 | ||
287 : | Isibaar | 1.1 | |
288 : | edgomez | 1.5 | movq mm2, [eax] |
289 : | movq mm3, [eax+1] | ||
290 : | movq mm6, mm2 | ||
291 : | pavgb mm2, mm3 ; preserved for next iteration | ||
292 : | lea ecx,[ecx+edx] | ||
293 : | pxor mm3, mm6 ; preserved for next iteration | ||
294 : | |||
295 : | pand mm1, mm3 | ||
296 : | movq mm6, mm0 | ||
297 : | pxor mm6, mm2 | ||
298 : | por mm1, mm6 | ||
299 : | pavgb mm0, mm2 | ||
300 : | pand mm1, mm7 | ||
301 : | psubb mm0, mm1 | ||
302 : | |||
303 : | movq [ecx], mm0 | ||
304 : | Isibaar | 1.1 | %endmacro |
305 : | |||
306 : | edgomez | 1.5 | ALIGN 16 |
307 : | Isibaar | 1.1 | interpolate8x8_halfpel_hv_xmm: |
308 : | edgomez | 1.5 | mov eax, [esp+16] ; rounding |
309 : | mov ecx, [esp+ 4] ; Dst | ||
310 : | test eax, eax | ||
311 : | mov eax, [esp+ 8] ; Src | ||
312 : | mov edx, [esp+12] ; stride | ||
313 : | Isibaar | 1.1 | |
314 : | movq mm7, [mmx_one] | ||
315 : | |||
316 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
317 : | movq mm2, [eax] | ||
318 : | movq mm3, [eax+1] | ||
319 : | movq mm6, mm2 | ||
320 : | pavgb mm2, mm3 | ||
321 : | edgomez | 1.5 | pxor mm3, mm6 ; mm2/mm3 ready |
322 : | Isibaar | 1.1 | |
323 : | jnz near .rounding1 | ||
324 : | |||
325 : | COPY_HV_SSE_RND0 | ||
326 : | add ecx, edx | ||
327 : | COPY_HV_SSE_RND0 | ||
328 : | add ecx, edx | ||
329 : | COPY_HV_SSE_RND0 | ||
330 : | add ecx, edx | ||
331 : | COPY_HV_SSE_RND0 | ||
332 : | ret | ||
333 : | |||
334 : | .rounding1 | ||
335 : | COPY_HV_SSE_RND1 | ||
336 : | add ecx, edx | ||
337 : | COPY_HV_SSE_RND1 | ||
338 : | add ecx, edx | ||
339 : | COPY_HV_SSE_RND1 | ||
340 : | add ecx, edx | ||
341 : | COPY_HV_SSE_RND1 | ||
342 : | edgomez | 1.5 | ret |
343 : | edgomez | 1.7 | |
344 : | ;=========================================================================== | ||
345 : | ; | ||
346 : | ; The next functions combine both source halfpel interpolation step and the | ||
347 : | ; averaging (with rouding) step to avoid wasting memory bandwidth computing | ||
348 : | ; intermediate halfpel images and then averaging them. | ||
349 : | ; | ||
350 : | ;=========================================================================== | ||
351 : | |||
352 : | %macro PROLOG0 0 | ||
353 : | mov ecx, [esp+ 4] ; Dst | ||
354 : | mov eax, [esp+ 8] ; Src | ||
355 : | mov edx, [esp+12] ; BpS | ||
356 : | %endmacro | ||
357 : | %macro PROLOG1 0 | ||
358 : | PROLOG0 | ||
359 : | test dword [esp+16], 1; Rounding? | ||
360 : | %endmacro | ||
361 : | %macro EPILOG 0 | ||
362 : | ret | ||
363 : | %endmacro | ||
364 : | |||
365 : | ;=========================================================================== | ||
366 : | ; | ||
367 : | ; void interpolate8x8_halfpel_add_xmm(uint8_t * const dst, | ||
368 : | ; const uint8_t * const src, | ||
369 : | ; const uint32_t stride, | ||
370 : | ; const uint32_t rounding); | ||
371 : | ; | ||
372 : | ; | ||
373 : | ;=========================================================================== | ||
374 : | |||
375 : | %macro ADD_FF 2 | ||
376 : | movq mm0, [eax+%1] | ||
377 : | movq mm1, [eax+%2] | ||
378 : | ;;--- | ||
379 : | ;; movq mm2, mm0 | ||
380 : | ;; movq mm3, mm1 | ||
381 : | ;;--- | ||
382 : | pavgb mm0, [ecx+%1] | ||
383 : | pavgb mm1, [ecx+%2] | ||
384 : | ;;-- | ||
385 : | ;; por mm2, [ecx+%1] | ||
386 : | ;; por mm3, [ecx+%2] | ||
387 : | ;; pand mm2, [mmx_one] | ||
388 : | ;; pand mm3, [mmx_one] | ||
389 : | ;; psubsb mm0, mm2 | ||
390 : | ;; psubsb mm1, mm3 | ||
391 : | ;;-- | ||
392 : | movq [ecx+%1], mm0 | ||
393 : | movq [ecx+%2], mm1 | ||
394 : | %endmacro | ||
395 : | |||
396 : | ALIGN 16 | ||
397 : | interpolate8x8_halfpel_add_xmm: ; 23c | ||
398 : | PROLOG1 | ||
399 : | ADD_FF 0, edx | ||
400 : | lea eax,[eax+2*edx] | ||
401 : | lea ecx,[ecx+2*edx] | ||
402 : | ADD_FF 0, edx | ||
403 : | lea eax,[eax+2*edx] | ||
404 : | lea ecx,[ecx+2*edx] | ||
405 : | ADD_FF 0, edx | ||
406 : | lea eax,[eax+2*edx] | ||
407 : | lea ecx,[ecx+2*edx] | ||
408 : | ADD_FF 0, edx | ||
409 : | EPILOG | ||
410 : | |||
411 : | ;=========================================================================== | ||
412 : | ; | ||
413 : | ; void interpolate8x8_halfpel_h_add_xmm(uint8_t * const dst, | ||
414 : | ; const uint8_t * const src, | ||
415 : | ; const uint32_t stride, | ||
416 : | ; const uint32_t rounding); | ||
417 : | ; | ||
418 : | ; | ||
419 : | ;=========================================================================== | ||
420 : | |||
421 : | |||
422 : | %macro ADD_FH_RND0 2 | ||
423 : | movq mm0, [eax+%1] | ||
424 : | movq mm1, [eax+%2] | ||
425 : | pavgb mm0, [eax+%1+1] | ||
426 : | pavgb mm1, [eax+%2+1] | ||
427 : | pavgb mm0, [ecx+%1] | ||
428 : | pavgb mm1, [ecx+%2] | ||
429 : | movq [ecx+%1],mm0 | ||
430 : | movq [ecx+%2],mm1 | ||
431 : | %endmacro | ||
432 : | |||
433 : | %macro ADD_FH_RND1 2 | ||
434 : | movq mm0, [eax+%1] | ||
435 : | movq mm1, [eax+%2] | ||
436 : | movq mm4, mm0 | ||
437 : | movq mm5, mm1 | ||
438 : | movq mm2, [eax+%1+1] | ||
439 : | movq mm3, [eax+%2+1] | ||
440 : | pavgb mm0, mm2 | ||
441 : | ; lea ?? | ||
442 : | pxor mm2, mm4 | ||
443 : | pavgb mm1, mm3 | ||
444 : | pxor mm3, mm5 | ||
445 : | pand mm2, [mmx_one] | ||
446 : | pand mm3, [mmx_one] | ||
447 : | psubb mm0, mm2 | ||
448 : | psubb mm1, mm3 | ||
449 : | pavgb mm0, [ecx+%1] | ||
450 : | pavgb mm1, [ecx+%2] | ||
451 : | movq [ecx+%1],mm0 | ||
452 : | movq [ecx+%2],mm1 | ||
453 : | %endmacro | ||
454 : | |||
455 : | ALIGN 16 | ||
456 : | interpolate8x8_halfpel_h_add_xmm: ; 32c | ||
457 : | PROLOG1 | ||
458 : | jnz near .Loop1 | ||
459 : | ADD_FH_RND0 0, edx | ||
460 : | lea eax,[eax+2*edx] | ||
461 : | lea ecx,[ecx+2*edx] | ||
462 : | ADD_FH_RND0 0, edx | ||
463 : | lea eax,[eax+2*edx] | ||
464 : | lea ecx,[ecx+2*edx] | ||
465 : | ADD_FH_RND0 0, edx | ||
466 : | lea eax,[eax+2*edx] | ||
467 : | lea ecx,[ecx+2*edx] | ||
468 : | ADD_FH_RND0 0, edx | ||
469 : | EPILOG | ||
470 : | |||
471 : | .Loop1 | ||
472 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
473 : | ; movq mm7, [mmx_one] | ||
474 : | ADD_FH_RND1 0, edx | ||
475 : | lea eax,[eax+2*edx] | ||
476 : | lea ecx,[ecx+2*edx] | ||
477 : | ADD_FH_RND1 0, edx | ||
478 : | lea eax,[eax+2*edx] | ||
479 : | lea ecx,[ecx+2*edx] | ||
480 : | ADD_FH_RND1 0, edx | ||
481 : | lea eax,[eax+2*edx] | ||
482 : | lea ecx,[ecx+2*edx] | ||
483 : | ADD_FH_RND1 0, edx | ||
484 : | EPILOG | ||
485 : | |||
486 : | |||
487 : | ;=========================================================================== | ||
488 : | ; | ||
489 : | ; void interpolate8x8_halfpel_v_add_xmm(uint8_t * const dst, | ||
490 : | ; const uint8_t * const src, | ||
491 : | ; const uint32_t stride, | ||
492 : | ; const uint32_t rounding); | ||
493 : | ; | ||
494 : | ; | ||
495 : | ;=========================================================================== | ||
496 : | |||
497 : | %macro ADD_8_HF_RND0 0 | ||
498 : | movq mm0, [eax] | ||
499 : | movq mm1, [eax+edx] | ||
500 : | pavgb mm0, mm1 | ||
501 : | pavgb mm1, [eax+2*edx] | ||
502 : | lea eax,[eax+2*edx] | ||
503 : | pavgb mm0, [ecx] | ||
504 : | pavgb mm1, [ecx+edx] | ||
505 : | movq [ecx],mm0 | ||
506 : | movq [ecx+edx],mm1 | ||
507 : | %endmacro | ||
508 : | |||
509 : | %macro ADD_8_HF_RND1 0 | ||
510 : | movq mm1, [eax+edx] | ||
511 : | movq mm2, [eax+2*edx] | ||
512 : | lea eax,[eax+2*edx] | ||
513 : | movq mm4, mm0 | ||
514 : | movq mm5, mm1 | ||
515 : | pavgb mm0, mm1 | ||
516 : | pxor mm4, mm1 | ||
517 : | pavgb mm1, mm2 | ||
518 : | pxor mm5, mm2 | ||
519 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
520 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
521 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
522 : | pavgb mm0, [ecx] | ||
523 : | movq [ecx], mm0 | ||
524 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
525 : | pavgb mm1, [ecx+edx] | ||
526 : | movq [ecx+edx], mm1 | ||
527 : | %endmacro | ||
528 : | |||
529 : | ALIGN 16 | ||
530 : | interpolate8x8_halfpel_v_add_xmm: | ||
531 : | PROLOG1 | ||
532 : | |||
533 : | jnz near .Loop1 | ||
534 : | pxor mm7, mm7 ; this is a NOP | ||
535 : | |||
536 : | ADD_8_HF_RND0 | ||
537 : | lea ecx,[ecx+2*edx] | ||
538 : | ADD_8_HF_RND0 | ||
539 : | lea ecx,[ecx+2*edx] | ||
540 : | ADD_8_HF_RND0 | ||
541 : | lea ecx,[ecx+2*edx] | ||
542 : | ADD_8_HF_RND0 | ||
543 : | EPILOG | ||
544 : | |||
545 : | .Loop1 | ||
546 : | movq mm0, [eax] ; loop invariant | ||
547 : | movq mm7, [mmx_one] | ||
548 : | |||
549 : | ADD_8_HF_RND1 | ||
550 : | movq mm0, mm2 | ||
551 : | lea ecx,[ecx+2*edx] | ||
552 : | ADD_8_HF_RND1 | ||
553 : | movq mm0, mm2 | ||
554 : | lea ecx,[ecx+2*edx] | ||
555 : | ADD_8_HF_RND1 | ||
556 : | movq mm0, mm2 | ||
557 : | lea ecx,[ecx+2*edx] | ||
558 : | ADD_8_HF_RND1 | ||
559 : | EPILOG | ||
560 : | |||
561 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
562 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
563 : | ; The boolean relations are: | ||
564 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
565 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
566 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
567 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
568 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
569 : | |||
570 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
571 : | |||
572 : | ;=========================================================================== | ||
573 : | ; | ||
574 : | ; void interpolate8x8_halfpel_hv_add_xmm(uint8_t * const dst, | ||
575 : | ; const uint8_t * const src, | ||
576 : | ; const uint32_t stride, | ||
577 : | ; const uint32_t rounding); | ||
578 : | ; | ||
579 : | ; | ||
580 : | ;=========================================================================== | ||
581 : | |||
582 : | %macro ADD_HH_RND0 0 | ||
583 : | lea eax,[eax+edx] | ||
584 : | |||
585 : | movq mm0, [eax] | ||
586 : | movq mm1, [eax+1] | ||
587 : | |||
588 : | movq mm6, mm0 | ||
589 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
590 : | lea eax,[eax+edx] | ||
591 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
592 : | |||
593 : | por mm3, mm1 ; ij |= jk | ||
594 : | movq mm6, mm2 | ||
595 : | pxor mm6, mm0 ; mm6 = s^t | ||
596 : | pand mm3, mm6 ; (ij|jk) &= st | ||
597 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
598 : | pand mm3, mm7 ; mask lsb | ||
599 : | psubb mm2, mm3 ; apply. | ||
600 : | |||
601 : | pavgb mm2, [ecx] | ||
602 : | movq [ecx], mm2 | ||
603 : | |||
604 : | movq mm2, [eax] | ||
605 : | movq mm3, [eax+1] | ||
606 : | movq mm6, mm2 | ||
607 : | pavgb mm2, mm3 ; preserved for next iteration | ||
608 : | lea ecx,[ecx+edx] | ||
609 : | pxor mm3, mm6 ; preserved for next iteration | ||
610 : | |||
611 : | por mm1, mm3 | ||
612 : | movq mm6, mm0 | ||
613 : | pxor mm6, mm2 | ||
614 : | pand mm1, mm6 | ||
615 : | pavgb mm0, mm2 | ||
616 : | |||
617 : | pand mm1, mm7 | ||
618 : | psubb mm0, mm1 | ||
619 : | |||
620 : | pavgb mm0, [ecx] | ||
621 : | movq [ecx], mm0 | ||
622 : | %endmacro | ||
623 : | |||
624 : | %macro ADD_HH_RND1 0 | ||
625 : | lea eax,[eax+edx] | ||
626 : | |||
627 : | movq mm0, [eax] | ||
628 : | movq mm1, [eax+1] | ||
629 : | |||
630 : | movq mm6, mm0 | ||
631 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
632 : | lea eax,[eax+edx] | ||
633 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
634 : | |||
635 : | pand mm3, mm1 | ||
636 : | movq mm6, mm2 | ||
637 : | pxor mm6, mm0 | ||
638 : | por mm3, mm6 | ||
639 : | pavgb mm2, mm0 | ||
640 : | pand mm3, mm7 | ||
641 : | psubb mm2, mm3 | ||
642 : | |||
643 : | pavgb mm2, [ecx] | ||
644 : | movq [ecx], mm2 | ||
645 : | |||
646 : | movq mm2, [eax] | ||
647 : | movq mm3, [eax+1] | ||
648 : | movq mm6, mm2 | ||
649 : | pavgb mm2, mm3 ; preserved for next iteration | ||
650 : | lea ecx,[ecx+edx] | ||
651 : | pxor mm3, mm6 ; preserved for next iteration | ||
652 : | |||
653 : | pand mm1, mm3 | ||
654 : | movq mm6, mm0 | ||
655 : | pxor mm6, mm2 | ||
656 : | por mm1, mm6 | ||
657 : | pavgb mm0, mm2 | ||
658 : | pand mm1, mm7 | ||
659 : | psubb mm0, mm1 | ||
660 : | |||
661 : | pavgb mm0, [ecx] | ||
662 : | movq [ecx], mm0 | ||
663 : | %endmacro | ||
664 : | |||
665 : | ALIGN 16 | ||
666 : | interpolate8x8_halfpel_hv_add_xmm: | ||
667 : | PROLOG1 | ||
668 : | |||
669 : | movq mm7, [mmx_one] | ||
670 : | |||
671 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
672 : | movq mm2, [eax] | ||
673 : | movq mm3, [eax+1] | ||
674 : | movq mm6, mm2 | ||
675 : | pavgb mm2, mm3 | ||
676 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
677 : | |||
678 : | jnz near .Loop1 | ||
679 : | |||
680 : | ADD_HH_RND0 | ||
681 : | add ecx, edx | ||
682 : | ADD_HH_RND0 | ||
683 : | add ecx, edx | ||
684 : | ADD_HH_RND0 | ||
685 : | add ecx, edx | ||
686 : | ADD_HH_RND0 | ||
687 : | EPILOG | ||
688 : | |||
689 : | .Loop1 | ||
690 : | ADD_HH_RND1 | ||
691 : | add ecx, edx | ||
692 : | ADD_HH_RND1 | ||
693 : | add ecx, edx | ||
694 : | ADD_HH_RND1 | ||
695 : | add ecx, edx | ||
696 : | ADD_HH_RND1 | ||
697 : | |||
698 : | EPILOG |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |