Parent Directory | Revision Log
Revision 1.2.2.2 - (view) (download)
1 : | edgomez | 1.2 | ;/***************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | edgomez | 1.2.2.1 | ; * - Reduced-Resolution utilities - |
5 : | edgomez | 1.2 | ; * |
6 : | ; * Copyright(C) 2002 Pascal Massimino <skal@planet-d.net> | ||
7 : | ; * | ||
8 : | ; * XviD is free software; you can redistribute it and/or modify it | ||
9 : | ; * under the terms of the GNU General Public License as published by | ||
10 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
11 : | ; * (at your option) any later version. | ||
12 : | ; * | ||
13 : | ; * This program is distributed in the hope that it will be useful, | ||
14 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
16 : | ; * GNU General Public License for more details. | ||
17 : | ; * | ||
18 : | ; * You should have received a copy of the GNU General Public License | ||
19 : | ; * along with this program; if not, write to the Free Software | ||
20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | ; * | ||
22 : | edgomez | 1.2.2.1 | ; * $Id$ |
23 : | edgomez | 1.2 | ; * |
24 : | ; *************************************************************************/ | ||
25 : | |||
26 : | edgomez | 1.2.2.1 | BITS 32 |
27 : | edgomez | 1.2 | |
28 : | edgomez | 1.2.2.1 | %macro cglobal 1 |
29 : | edgomez | 1.2 | %ifdef PREFIX |
30 : | edgomez | 1.2.2.1 | global _%1 |
31 : | edgomez | 1.2 | %define %1 _%1 |
32 : | %else | ||
33 : | global %1 | ||
34 : | %endif | ||
35 : | %endmacro | ||
36 : | |||
37 : | ;=========================================================================== | ||
38 : | |||
39 : | edgomez | 1.2.2.2 | %ifdef FORMAT_COFF |
40 : | SECTION .rodata data | ||
41 : | %else | ||
42 : | SECTION .rodata data align=16 | ||
43 : | %endif | ||
44 : | edgomez | 1.2 | |
45 : | align 16 | ||
46 : | Up31 dw 3, 1, 3, 1 | ||
47 : | Up13 dw 1, 3, 1, 3 | ||
48 : | Up93 dw 9, 3, 9, 3 | ||
49 : | Up39 dw 3, 9, 3, 9 | ||
50 : | Cst0 dw 0, 0, 0, 0 | ||
51 : | Cst2 dw 2, 2, 2, 2 | ||
52 : | Cst3 dw 3, 3, 3, 3 | ||
53 : | Cst32 dw 32,32,32,32 | ||
54 : | Cst2000 dw 2, 0, 0, 0 | ||
55 : | Cst0002 dw 0, 0, 0, 2 | ||
56 : | |||
57 : | Mask_ff dw 0xff,0xff,0xff,0xff | ||
58 : | |||
59 : | ;=========================================================================== | ||
60 : | |||
61 : | edgomez | 1.2.2.1 | SECTION .text |
62 : | edgomez | 1.2 | |
63 : | cglobal xvid_Copy_Upsampled_8x8_16To8_mmx | ||
64 : | cglobal xvid_Add_Upsampled_8x8_16To8_mmx | ||
65 : | cglobal xvid_Copy_Upsampled_8x8_16To8_xmm | ||
66 : | cglobal xvid_Add_Upsampled_8x8_16To8_xmm | ||
67 : | |||
68 : | cglobal xvid_HFilter_31_mmx | ||
69 : | cglobal xvid_VFilter_31_x86 | ||
70 : | cglobal xvid_HFilter_31_x86 | ||
71 : | |||
72 : | cglobal xvid_Filter_18x18_To_8x8_mmx | ||
73 : | cglobal xvid_Filter_Diff_18x18_To_8x8_mmx | ||
74 : | |||
75 : | |||
76 : | ;////////////////////////////////////////////////////////////////////// | ||
77 : | ;// 8x8 -> 16x16 upsampling (16b) | ||
78 : | ;////////////////////////////////////////////////////////////////////// | ||
79 : | |||
80 : | %macro MUL_PACK 4 ; %1/%2: regs %3/%4/%5: Up13/Up31 | ||
81 : | pmullw %1, %3 ; [Up13] | ||
82 : | pmullw mm4, %4 ; [Up31] | ||
83 : | pmullw %2, %3 ; [Up13] | ||
84 : | pmullw mm5, %4 ; [Up31] | ||
85 : | paddsw %1, [Cst2] | ||
86 : | paddsw %2, [Cst2] | ||
87 : | paddsw %1, mm4 | ||
88 : | paddsw %2, mm5 | ||
89 : | %endmacro | ||
90 : | |||
91 : | ; MMX-way of reordering columns... | ||
92 : | |||
93 : | %macro COL03 3 ;%1/%2: regs, %3: row -output: mm4/mm5 | ||
94 : | movq %1, [edx+%3*16+0*2] ; %1 = 0|1|2|3 | ||
95 : | movq %2,[edx+%3*16+1*2] ; %2 = 1|2|3|4 | ||
96 : | movq mm5, %1 ; mm5 = 0|1|2|3 | ||
97 : | movq mm4, %1 ; mm4 = 0|1|2|3 | ||
98 : | punpckhwd mm5,%2 ; mm5 = 2|3|3|4 | ||
99 : | punpcklwd mm4,%2 ; mm4 = 0|1|1|2 | ||
100 : | punpcklwd %1,%1 ; %1 = 0|0|1|1 | ||
101 : | punpcklwd %2, mm5 ; %2 = 1|2|2|3 | ||
102 : | punpcklwd %1, mm4 ; %1 = 0|0|0|1 | ||
103 : | %endmacro | ||
104 : | |||
105 : | %macro COL47 3 ;%1-%2: regs, %3: row -output: mm4/mm5 | ||
106 : | movq mm5, [edx+%3*16+4*2] ; mm5 = 4|5|6|7 | ||
107 : | movq %1, [edx+%3*16+3*2] ; %1 = 3|4|5|6 | ||
108 : | movq %2, mm5 ; %2 = 4|5|6|7 | ||
109 : | movq mm4, mm5 ; mm4 = 4|5|6|7 | ||
110 : | punpckhwd %2, %2 ; %2 = 6|6|7|7 | ||
111 : | punpckhwd mm5, %2 ; mm5 = 6|7|7|7 | ||
112 : | movq %2, %1 ; %2 = 3|4|5|6 | ||
113 : | punpcklwd %1, mm4 ; %1 = 3|4|4|5 | ||
114 : | punpckhwd %2, mm4 ; %2 = 5|6|6|7 | ||
115 : | punpcklwd mm4, %2 ; mm4 = 4|5|5|6 | ||
116 : | %endmacro | ||
117 : | |||
118 : | %macro MIX_ROWS 4 ; %1/%2:prev %3/4:cur (preserved) mm4/mm5: output | ||
119 : | ; we need to perform: (%1,%3) -> (%1 = 3*%1+%3, mm4 = 3*%3+%1), %3 preserved. | ||
120 : | movq mm4, [Cst3] | ||
121 : | movq mm5, [Cst3] | ||
122 : | pmullw mm4, %3 | ||
123 : | pmullw mm5, %4 | ||
124 : | paddsw mm4, %1 | ||
125 : | paddsw mm5, %2 | ||
126 : | pmullw %1, [Cst3] | ||
127 : | pmullw %2, [Cst3] | ||
128 : | paddsw %1, %3 | ||
129 : | paddsw %2, %4 | ||
130 : | %endmacro | ||
131 : | |||
132 : | ;=========================================================================== | ||
133 : | ; | ||
134 : | edgomez | 1.2.2.1 | ; void xvid_Copy_Upsampled_8x8_16To8_mmx(uint8_t *Dst, |
135 : | edgomez | 1.2 | ; const int16_t *Src, const int BpS); |
136 : | ; | ||
137 : | ;=========================================================================== | ||
138 : | |||
139 : | edgomez | 1.2.2.1 | ; Note: we can use ">>2" instead of "/4" here, since we |
140 : | edgomez | 1.2 | ; are (supposed to be) averaging positive values |
141 : | |||
142 : | %macro STORE_1 2 | ||
143 : | psraw %1, 2 | ||
144 : | psraw %2, 2 | ||
145 : | edgomez | 1.2.2.1 | packuswb %1,%2 |
146 : | edgomez | 1.2 | movq [ecx], %1 |
147 : | %endmacro | ||
148 : | |||
149 : | %macro STORE_2 2 ; pack and store (%1,%2) + (mm4,mm5) | ||
150 : | psraw %1, 4 | ||
151 : | psraw %2, 4 | ||
152 : | psraw mm4, 4 | ||
153 : | psraw mm5, 4 | ||
154 : | packuswb %1,%2 | ||
155 : | packuswb mm4, mm5 | ||
156 : | movq [ecx], %1 | ||
157 : | movq [ecx+eax], mm4 | ||
158 : | lea ecx, [ecx+2*eax] | ||
159 : | %endmacro | ||
160 : | |||
161 : | ;////////////////////////////////////////////////////////////////////// | ||
162 : | |||
163 : | align 16 | ||
164 : | xvid_Copy_Upsampled_8x8_16To8_mmx: ; 344c | ||
165 : | |||
166 : | mov ecx, [esp+4] ; Dst | ||
167 : | mov edx, [esp+8] ; Src | ||
168 : | mov eax, [esp+12] ; BpS | ||
169 : | |||
170 : | movq mm6, [Up13] | ||
171 : | movq mm7, [Up31] | ||
172 : | |||
173 : | COL03 mm0, mm1, 0 | ||
174 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
175 : | movq mm4, mm0 | ||
176 : | movq mm5, mm1 | ||
177 : | STORE_1 mm4, mm5 | ||
178 : | add ecx, eax | ||
179 : | |||
180 : | COL03 mm2, mm3, 1 | ||
181 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
182 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
183 : | STORE_2 mm0, mm1 | ||
184 : | |||
185 : | COL03 mm0, mm1, 2 | ||
186 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
187 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
188 : | STORE_2 mm2, mm3 | ||
189 : | |||
190 : | COL03 mm2, mm3, 3 | ||
191 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
192 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
193 : | STORE_2 mm0, mm1 | ||
194 : | |||
195 : | COL03 mm0, mm1, 4 | ||
196 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
197 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
198 : | STORE_2 mm2, mm3 | ||
199 : | |||
200 : | COL03 mm2, mm3, 5 | ||
201 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
202 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
203 : | STORE_2 mm0, mm1 | ||
204 : | |||
205 : | COL03 mm0, mm1, 6 | ||
206 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
207 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
208 : | STORE_2 mm2, mm3 | ||
209 : | |||
210 : | COL03 mm2, mm3, 7 | ||
211 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
212 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
213 : | STORE_2 mm0, mm1 | ||
214 : | |||
215 : | STORE_1 mm2, mm3 | ||
216 : | |||
217 : | mov ecx, [esp+4] | ||
218 : | add ecx, 8 | ||
219 : | |||
220 : | COL47 mm0, mm1, 0 | ||
221 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
222 : | movq mm4, mm0 | ||
223 : | movq mm5, mm1 | ||
224 : | STORE_1 mm4, mm5 | ||
225 : | add ecx, eax | ||
226 : | |||
227 : | COL47 mm2, mm3, 1 | ||
228 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
229 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
230 : | STORE_2 mm0, mm1 | ||
231 : | |||
232 : | COL47 mm0, mm1, 2 | ||
233 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
234 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
235 : | STORE_2 mm2, mm3 | ||
236 : | |||
237 : | COL47 mm2, mm3, 3 | ||
238 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
239 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
240 : | STORE_2 mm0, mm1 | ||
241 : | |||
242 : | COL47 mm0, mm1, 4 | ||
243 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
244 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
245 : | STORE_2 mm2, mm3 | ||
246 : | |||
247 : | COL47 mm2, mm3, 5 | ||
248 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
249 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
250 : | STORE_2 mm0, mm1 | ||
251 : | |||
252 : | COL47 mm0, mm1, 6 | ||
253 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
254 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
255 : | STORE_2 mm2, mm3 | ||
256 : | |||
257 : | COL47 mm2, mm3, 7 | ||
258 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
259 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
260 : | STORE_2 mm0, mm1 | ||
261 : | |||
262 : | STORE_1 mm2, mm3 | ||
263 : | |||
264 : | ret | ||
265 : | |||
266 : | ;=========================================================================== | ||
267 : | ; | ||
268 : | edgomez | 1.2.2.1 | ; void xvid_Add_Upsampled_8x8_16To8_mmx(uint8_t *Dst, |
269 : | edgomez | 1.2 | ; const int16_t *Src, const int BpS); |
270 : | ; | ||
271 : | ;=========================================================================== | ||
272 : | |||
273 : | ; Note: grrr... the 'pcmpgtw' stuff are the "/4" and "/16" operators | ||
274 : | edgomez | 1.2.2.1 | ; implemented with ">>2" and ">>4" using: |
275 : | edgomez | 1.2 | ; x/4 = ( (x-(x<0))>>2 ) + (x<0) |
276 : | ; x/16 = ( (x-(x<0))>>4 ) + (x<0) | ||
277 : | |||
278 : | %macro STORE_ADD_1 2 | ||
279 : | ; We substract the rounder '2' for corner pixels, | ||
280 : | ; since when 'x' is negative, (x*4 + 2)/4 is *not* | ||
281 : | ; equal to 'x'. In fact, the correct relation is: | ||
282 : | ; (x*4 + 2)/4 = x - (x<0) | ||
283 : | ; So, better revert to (x*4)/4 = x. | ||
284 : | |||
285 : | psubsw %1, [Cst2000] | ||
286 : | psubsw %2, [Cst0002] | ||
287 : | pxor mm6, mm6 | ||
288 : | pxor mm7, mm7 | ||
289 : | pcmpgtw mm6, %1 | ||
290 : | pcmpgtw mm7, %2 | ||
291 : | paddsw %1, mm6 | ||
292 : | paddsw %2, mm7 | ||
293 : | psraw %1, 2 | ||
294 : | psraw %2, 2 | ||
295 : | psubsw %1, mm6 | ||
296 : | psubsw %2, mm7 | ||
297 : | |||
298 : | ; mix with destination [ecx] | ||
299 : | movq mm6, [ecx] | ||
300 : | movq mm7, [ecx] | ||
301 : | punpcklbw mm6, [Cst0] | ||
302 : | punpckhbw mm7, [Cst0] | ||
303 : | paddsw %1, mm6 | ||
304 : | paddsw %2, mm7 | ||
305 : | packuswb %1,%2 | ||
306 : | movq [ecx], %1 | ||
307 : | %endmacro | ||
308 : | |||
309 : | %macro STORE_ADD_2 2 | ||
310 : | pxor mm6, mm6 | ||
311 : | pxor mm7, mm7 | ||
312 : | pcmpgtw mm6, %1 | ||
313 : | pcmpgtw mm7, %2 | ||
314 : | paddsw %1, mm6 | ||
315 : | paddsw %2, mm7 | ||
316 : | psraw %1, 4 | ||
317 : | psraw %2, 4 | ||
318 : | psubsw %1, mm6 | ||
319 : | psubsw %2, mm7 | ||
320 : | |||
321 : | pxor mm6, mm6 | ||
322 : | pxor mm7, mm7 | ||
323 : | pcmpgtw mm6, mm4 | ||
324 : | pcmpgtw mm7, mm5 | ||
325 : | paddsw mm4, mm6 | ||
326 : | paddsw mm5, mm7 | ||
327 : | psraw mm4, 4 | ||
328 : | psraw mm5, 4 | ||
329 : | psubsw mm4, mm6 | ||
330 : | psubsw mm5, mm7 | ||
331 : | |||
332 : | ; mix with destination | ||
333 : | movq mm6, [ecx] | ||
334 : | movq mm7, [ecx] | ||
335 : | punpcklbw mm6, [Cst0] | ||
336 : | punpckhbw mm7, [Cst0] | ||
337 : | paddsw %1, mm6 | ||
338 : | paddsw %2, mm7 | ||
339 : | |||
340 : | movq mm6, [ecx+eax] | ||
341 : | movq mm7, [ecx+eax] | ||
342 : | |||
343 : | punpcklbw mm6, [Cst0] | ||
344 : | punpckhbw mm7, [Cst0] | ||
345 : | paddsw mm4, mm6 | ||
346 : | paddsw mm5, mm7 | ||
347 : | |||
348 : | packuswb %1,%2 | ||
349 : | packuswb mm4, mm5 | ||
350 : | |||
351 : | movq [ecx], %1 | ||
352 : | movq [ecx+eax], mm4 | ||
353 : | |||
354 : | lea ecx, [ecx+2*eax] | ||
355 : | %endmacro | ||
356 : | |||
357 : | ;////////////////////////////////////////////////////////////////////// | ||
358 : | |||
359 : | align 16 | ||
360 : | xvid_Add_Upsampled_8x8_16To8_mmx: ; 579c | ||
361 : | |||
362 : | mov ecx, [esp+4] ; Dst | ||
363 : | mov edx, [esp+8] ; Src | ||
364 : | mov eax, [esp+12] ; BpS | ||
365 : | |||
366 : | COL03 mm0, mm1, 0 | ||
367 : | edgomez | 1.2.2.1 | MUL_PACK mm0,mm1, [Up13], [Up31] |
368 : | edgomez | 1.2 | movq mm4, mm0 |
369 : | movq mm5, mm1 | ||
370 : | STORE_ADD_1 mm4, mm5 | ||
371 : | add ecx, eax | ||
372 : | |||
373 : | COL03 mm2, mm3, 1 | ||
374 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
375 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
376 : | STORE_ADD_2 mm0, mm1 | ||
377 : | |||
378 : | COL03 mm0, mm1, 2 | ||
379 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
380 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
381 : | STORE_ADD_2 mm2, mm3 | ||
382 : | |||
383 : | COL03 mm2, mm3, 3 | ||
384 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
385 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
386 : | STORE_ADD_2 mm0, mm1 | ||
387 : | |||
388 : | COL03 mm0, mm1, 4 | ||
389 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
390 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
391 : | STORE_ADD_2 mm2, mm3 | ||
392 : | |||
393 : | COL03 mm2, mm3, 5 | ||
394 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
395 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
396 : | STORE_ADD_2 mm0, mm1 | ||
397 : | |||
398 : | COL03 mm0, mm1, 6 | ||
399 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
400 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
401 : | STORE_ADD_2 mm2, mm3 | ||
402 : | |||
403 : | COL03 mm2, mm3, 7 | ||
404 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
405 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
406 : | STORE_ADD_2 mm0, mm1 | ||
407 : | |||
408 : | STORE_ADD_1 mm2, mm3 | ||
409 : | |||
410 : | |||
411 : | mov ecx, [esp+4] | ||
412 : | add ecx, 8 | ||
413 : | |||
414 : | COL47 mm0, mm1, 0 | ||
415 : | edgomez | 1.2.2.1 | MUL_PACK mm0,mm1, [Up13], [Up31] |
416 : | edgomez | 1.2 | movq mm4, mm0 |
417 : | edgomez | 1.2.2.1 | movq mm5, mm1 |
418 : | edgomez | 1.2 | STORE_ADD_1 mm4, mm5 |
419 : | add ecx, eax | ||
420 : | |||
421 : | COL47 mm2, mm3, 1 | ||
422 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
423 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
424 : | STORE_ADD_2 mm0, mm1 | ||
425 : | |||
426 : | COL47 mm0, mm1, 2 | ||
427 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
428 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
429 : | STORE_ADD_2 mm2, mm3 | ||
430 : | |||
431 : | COL47 mm2, mm3, 3 | ||
432 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
433 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
434 : | STORE_ADD_2 mm0, mm1 | ||
435 : | |||
436 : | COL47 mm0, mm1, 4 | ||
437 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
438 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
439 : | STORE_ADD_2 mm2, mm3 | ||
440 : | |||
441 : | COL47 mm2, mm3, 5 | ||
442 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
443 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
444 : | STORE_ADD_2 mm0, mm1 | ||
445 : | |||
446 : | COL47 mm0, mm1, 6 | ||
447 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
448 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
449 : | STORE_ADD_2 mm2, mm3 | ||
450 : | |||
451 : | COL47 mm2, mm3, 7 | ||
452 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
453 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
454 : | STORE_ADD_2 mm0, mm1 | ||
455 : | |||
456 : | STORE_ADD_1 mm2, mm3 | ||
457 : | |||
458 : | ret | ||
459 : | |||
460 : | ;=========================================================================== | ||
461 : | ; | ||
462 : | edgomez | 1.2.2.1 | ; void xvid_Copy_Upsampled_8x8_16To8_xmm(uint8_t *Dst, |
463 : | edgomez | 1.2 | ; const int16_t *Src, const int BpS); |
464 : | ; | ||
465 : | ;=========================================================================== | ||
466 : | |||
467 : | ; xmm version can take (little) advantage of 'pshufw' | ||
468 : | |||
469 : | %macro COL03_SSE 3 ;%1/%2: regs, %3: row -trashes mm4/mm5 | ||
470 : | movq %2, [edx+%3*16+0*2] ; <- 0|1|2|3 | ||
471 : | pshufw %1, %2, (0+0*4+0*16+1*64) ; %1 = 0|0|0|1 | ||
472 : | pshufw mm4, %2, (0+1*4+1*16+2*64) ; mm4= 0|1|1|2 | ||
473 : | pshufw %2, %2, (1+2*4+2*16+3*64) ; %2 = 1|2|2|3 | ||
474 : | pshufw mm5, [edx+%3*16+2*2], (0+1*4+1*16+2*64) ; mm5 = 2|3|3|4 | ||
475 : | %endmacro | ||
476 : | |||
477 : | %macro COL47_SSE 3 ;%1-%2: regs, %3: row -trashes mm4/mm5 | ||
478 : | pshufw %1, [edx+%3*16+2*2], (1+2*4+2*16+3*64) ; 3|4|4|5 | ||
479 : | movq mm5, [edx+%3*16+2*4] ; <- 4|5|6|7 | ||
480 : | pshufw mm4, mm5, (0+1*4+1*16+2*64) ; 4|5|5|6 | ||
481 : | pshufw %2, mm5, (1+2*4+2*16+3*64) ; 5|6|6|7 | ||
482 : | pshufw mm5, mm5, (2+3*4+3*16+3*64) ; 6|7|7|7 | ||
483 : | %endmacro | ||
484 : | |||
485 : | |||
486 : | ;////////////////////////////////////////////////////////////////////// | ||
487 : | |||
488 : | align 16 | ||
489 : | xvid_Copy_Upsampled_8x8_16To8_xmm: ; 315c | ||
490 : | |||
491 : | mov ecx, [esp+4] ; Dst | ||
492 : | mov edx, [esp+8] ; Src | ||
493 : | mov eax, [esp+12] ; BpS | ||
494 : | |||
495 : | movq mm6, [Up13] | ||
496 : | movq mm7, [Up31] | ||
497 : | |||
498 : | COL03_SSE mm0, mm1, 0 | ||
499 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
500 : | movq mm4, mm0 | ||
501 : | movq mm5, mm1 | ||
502 : | STORE_1 mm4, mm5 | ||
503 : | add ecx, eax | ||
504 : | |||
505 : | COL03_SSE mm2, mm3, 1 | ||
506 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
507 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
508 : | STORE_2 mm0, mm1 | ||
509 : | |||
510 : | COL03_SSE mm0, mm1, 2 | ||
511 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
512 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
513 : | STORE_2 mm2, mm3 | ||
514 : | |||
515 : | COL03_SSE mm2, mm3, 3 | ||
516 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
517 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
518 : | STORE_2 mm0, mm1 | ||
519 : | |||
520 : | COL03_SSE mm0, mm1, 4 | ||
521 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
522 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
523 : | STORE_2 mm2, mm3 | ||
524 : | |||
525 : | COL03_SSE mm2, mm3, 5 | ||
526 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
527 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
528 : | STORE_2 mm0, mm1 | ||
529 : | |||
530 : | COL03_SSE mm0, mm1, 6 | ||
531 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
532 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
533 : | STORE_2 mm2, mm3 | ||
534 : | |||
535 : | COL03_SSE mm2, mm3, 7 | ||
536 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
537 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
538 : | STORE_2 mm0, mm1 | ||
539 : | |||
540 : | STORE_1 mm2, mm3 | ||
541 : | |||
542 : | mov ecx, [esp+4] | ||
543 : | add ecx, 8 | ||
544 : | |||
545 : | COL47_SSE mm0, mm1, 0 | ||
546 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
547 : | movq mm4, mm0 | ||
548 : | movq mm5, mm1 | ||
549 : | STORE_1 mm4, mm5 | ||
550 : | add ecx, eax | ||
551 : | |||
552 : | COL47_SSE mm2, mm3, 1 | ||
553 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
554 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
555 : | STORE_2 mm0, mm1 | ||
556 : | |||
557 : | COL47_SSE mm0, mm1, 2 | ||
558 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
559 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
560 : | STORE_2 mm2, mm3 | ||
561 : | |||
562 : | COL47_SSE mm2, mm3, 3 | ||
563 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
564 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
565 : | STORE_2 mm0, mm1 | ||
566 : | |||
567 : | COL47_SSE mm0, mm1, 4 | ||
568 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
569 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
570 : | STORE_2 mm2, mm3 | ||
571 : | |||
572 : | COL47_SSE mm2, mm3, 5 | ||
573 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
574 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
575 : | STORE_2 mm0, mm1 | ||
576 : | |||
577 : | COL47_SSE mm0, mm1, 6 | ||
578 : | MUL_PACK mm0,mm1, mm6, mm7 | ||
579 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
580 : | STORE_2 mm2, mm3 | ||
581 : | |||
582 : | COL47_SSE mm2, mm3, 7 | ||
583 : | MUL_PACK mm2,mm3, mm6, mm7 | ||
584 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
585 : | STORE_2 mm0, mm1 | ||
586 : | |||
587 : | STORE_1 mm2, mm3 | ||
588 : | |||
589 : | ret | ||
590 : | |||
591 : | ;=========================================================================== | ||
592 : | ; | ||
593 : | edgomez | 1.2.2.1 | ; void xvid_Add_Upsampled_8x8_16To8_xmm(uint8_t *Dst, |
594 : | edgomez | 1.2 | ; const int16_t *Src, const int BpS); |
595 : | ; | ||
596 : | ;=========================================================================== | ||
597 : | |||
598 : | align 16 | ||
599 : | xvid_Add_Upsampled_8x8_16To8_xmm: ; 549c | ||
600 : | |||
601 : | mov ecx, [esp+4] ; Dst | ||
602 : | mov edx, [esp+8] ; Src | ||
603 : | mov eax, [esp+12] ; BpS | ||
604 : | |||
605 : | COL03_SSE mm0, mm1, 0 | ||
606 : | edgomez | 1.2.2.1 | MUL_PACK mm0,mm1, [Up13], [Up31] |
607 : | edgomez | 1.2 | movq mm4, mm0 |
608 : | movq mm5, mm1 | ||
609 : | STORE_ADD_1 mm4, mm5 | ||
610 : | add ecx, eax | ||
611 : | |||
612 : | COL03_SSE mm2, mm3, 1 | ||
613 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
614 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
615 : | STORE_ADD_2 mm0, mm1 | ||
616 : | |||
617 : | COL03_SSE mm0, mm1, 2 | ||
618 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
619 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
620 : | STORE_ADD_2 mm2, mm3 | ||
621 : | |||
622 : | COL03_SSE mm2, mm3, 3 | ||
623 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
624 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
625 : | STORE_ADD_2 mm0, mm1 | ||
626 : | |||
627 : | COL03_SSE mm0, mm1, 4 | ||
628 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
629 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
630 : | STORE_ADD_2 mm2, mm3 | ||
631 : | |||
632 : | COL03_SSE mm2, mm3, 5 | ||
633 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
634 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
635 : | STORE_ADD_2 mm0, mm1 | ||
636 : | |||
637 : | COL03_SSE mm0, mm1, 6 | ||
638 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
639 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
640 : | STORE_ADD_2 mm2, mm3 | ||
641 : | |||
642 : | COL03_SSE mm2, mm3, 7 | ||
643 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
644 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
645 : | STORE_ADD_2 mm0, mm1 | ||
646 : | |||
647 : | STORE_ADD_1 mm2, mm3 | ||
648 : | |||
649 : | |||
650 : | mov ecx, [esp+4] | ||
651 : | add ecx, 8 | ||
652 : | |||
653 : | COL47_SSE mm0, mm1, 0 | ||
654 : | edgomez | 1.2.2.1 | MUL_PACK mm0,mm1, [Up13], [Up31] |
655 : | edgomez | 1.2 | movq mm4, mm0 |
656 : | edgomez | 1.2.2.1 | movq mm5, mm1 |
657 : | edgomez | 1.2 | STORE_ADD_1 mm4, mm5 |
658 : | add ecx, eax | ||
659 : | |||
660 : | COL47_SSE mm2, mm3, 1 | ||
661 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
662 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
663 : | STORE_ADD_2 mm0, mm1 | ||
664 : | |||
665 : | COL47_SSE mm0, mm1, 2 | ||
666 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
667 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
668 : | STORE_ADD_2 mm2, mm3 | ||
669 : | |||
670 : | COL47_SSE mm2, mm3, 3 | ||
671 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
672 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
673 : | STORE_ADD_2 mm0, mm1 | ||
674 : | |||
675 : | COL47_SSE mm0, mm1, 4 | ||
676 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
677 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
678 : | STORE_ADD_2 mm2, mm3 | ||
679 : | |||
680 : | COL47_SSE mm2, mm3, 5 | ||
681 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
682 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
683 : | STORE_ADD_2 mm0, mm1 | ||
684 : | |||
685 : | COL47_SSE mm0, mm1, 6 | ||
686 : | MUL_PACK mm0,mm1, [Up13], [Up31] | ||
687 : | MIX_ROWS mm2, mm3, mm0, mm1 | ||
688 : | STORE_ADD_2 mm2, mm3 | ||
689 : | |||
690 : | COL47_SSE mm2, mm3, 7 | ||
691 : | MUL_PACK mm2,mm3, [Up13], [Up31] | ||
692 : | MIX_ROWS mm0, mm1, mm2, mm3 | ||
693 : | STORE_ADD_2 mm0, mm1 | ||
694 : | |||
695 : | STORE_ADD_1 mm2, mm3 | ||
696 : | |||
697 : | ret | ||
698 : | |||
699 : | |||
700 : | ;=========================================================================== | ||
701 : | ; | ||
702 : | ; void xvid_HFilter_31_mmx(uint8_t *Src1, uint8_t *Src2, int Nb_Blks); | ||
703 : | ; void xvid_VFilter_31_x86(uint8_t *Src1, uint8_t *Src2, const int BpS, int Nb_Blks); | ||
704 : | ; void xvid_HFilter_31_x86(uint8_t *Src1, uint8_t *Src2, int Nb_Blks); | ||
705 : | ; | ||
706 : | ;=========================================================================== | ||
707 : | |||
708 : | ;////////////////////////////////////////////////////////////////////// | ||
709 : | ;// horizontal/vertical filtering: [x,y] -> [ (3x+y+2)>>2, (x+3y+2)>>2 ] | ||
710 : | ;// | ||
711 : | ;// We use the trick: tmp = (x+y+2) -> [x = (tmp+2x)>>2, y = (tmp+2y)>>2] | ||
712 : | ;////////////////////////////////////////////////////////////////////// | ||
713 : | |||
714 : | align 16 | ||
715 : | xvid_HFilter_31_mmx: | ||
716 : | push esi | ||
717 : | push edi | ||
718 : | mov esi, [esp+4 +8] ; Src1 | ||
719 : | mov edi, [esp+8 +8] ; Src2 | ||
720 : | mov eax, [esp+12 +8] ; Nb_Blks | ||
721 : | lea eax,[eax*2] | ||
722 : | movq mm5, [Cst2] | ||
723 : | pxor mm7, mm7 | ||
724 : | |||
725 : | lea esi, [esi+eax*4] | ||
726 : | lea edi, [edi+eax*4] | ||
727 : | |||
728 : | neg eax | ||
729 : | |||
730 : | .Loop: ;12c | ||
731 : | edgomez | 1.2.2.1 | movd mm0, [esi+eax*4] |
732 : | edgomez | 1.2 | movd mm1, [edi+eax*4] |
733 : | edgomez | 1.2.2.1 | movq mm2, mm5 |
734 : | edgomez | 1.2 | punpcklbw mm0, mm7 |
735 : | punpcklbw mm1, mm7 | ||
736 : | paddsw mm2, mm0 | ||
737 : | paddsw mm0, mm0 | ||
738 : | paddsw mm2, mm1 | ||
739 : | paddsw mm1, mm1 | ||
740 : | paddsw mm0, mm2 | ||
741 : | paddsw mm1, mm2 | ||
742 : | psraw mm0, 2 | ||
743 : | psraw mm1, 2 | ||
744 : | packuswb mm0, mm7 | ||
745 : | packuswb mm1, mm7 | ||
746 : | movd [esi+eax*4], mm0 | ||
747 : | movd [edi+eax*4], mm1 | ||
748 : | add eax,1 | ||
749 : | jl .Loop | ||
750 : | |||
751 : | pop edi | ||
752 : | pop esi | ||
753 : | ret | ||
754 : | |||
755 : | ; mmx is of no use here. Better use plain ASM. Moreover, | ||
756 : | ; this is for the fun of ASM coding, coz' every modern compiler can | ||
757 : | ; end up with a code that looks very much like this one... | ||
758 : | |||
759 : | align 16 | ||
760 : | xvid_VFilter_31_x86: | ||
761 : | push esi | ||
762 : | push edi | ||
763 : | push ebx | ||
764 : | push ebp | ||
765 : | mov esi, [esp+4 +16] ; Src1 | ||
766 : | mov edi, [esp+8 +16] ; Src2 | ||
767 : | mov ebp, [esp+12 +16] ; BpS | ||
768 : | mov eax, [esp+16 +16] ; Nb_Blks | ||
769 : | lea eax,[eax*8] | ||
770 : | |||
771 : | .Loop: ;7c | ||
772 : | movzx ecx, byte [esi] | ||
773 : | movzx edx, byte [edi] | ||
774 : | |||
775 : | lea ebx, [ecx+edx+2] | ||
776 : | lea ecx,[ebx+2*ecx] | ||
777 : | lea edx,[ebx+2*edx] | ||
778 : | |||
779 : | shr ecx,2 | ||
780 : | shr edx,2 | ||
781 : | mov [esi], cl | ||
782 : | mov [edi], dl | ||
783 : | lea esi, [esi+ebp] | ||
784 : | lea edi, [edi+ebp] | ||
785 : | dec eax | ||
786 : | jg .Loop | ||
787 : | |||
788 : | pop ebp | ||
789 : | pop ebx | ||
790 : | pop edi | ||
791 : | pop esi | ||
792 : | ret | ||
793 : | |||
794 : | ; this one's just a little faster than gcc's code. Very little. | ||
795 : | |||
796 : | align 16 | ||
797 : | xvid_HFilter_31_x86: | ||
798 : | push esi | ||
799 : | push edi | ||
800 : | push ebx | ||
801 : | mov esi, [esp+4 +12] ; Src1 | ||
802 : | mov edi, [esp+8 +12] ; Src2 | ||
803 : | mov eax, [esp+12 +12] ; Nb_Blks | ||
804 : | |||
805 : | lea eax,[eax*8] | ||
806 : | lea esi, [esi+eax] | ||
807 : | lea edi, [esi+eax] | ||
808 : | neg eax | ||
809 : | |||
810 : | .Loop: ; 6c | ||
811 : | movzx ecx, byte [esi+eax] | ||
812 : | movzx edx, byte [edi+eax] | ||
813 : | |||
814 : | lea ebx, [ecx+edx+2] | ||
815 : | lea ecx,[ebx+2*ecx] | ||
816 : | lea edx,[ebx+2*edx] | ||
817 : | shr ecx,2 | ||
818 : | shr edx,2 | ||
819 : | mov [esi+eax], cl | ||
820 : | mov [edi+eax], dl | ||
821 : | inc eax | ||
822 : | |||
823 : | jl .Loop | ||
824 : | |||
825 : | pop ebx | ||
826 : | pop edi | ||
827 : | pop esi | ||
828 : | ret | ||
829 : | |||
830 : | ;////////////////////////////////////////////////////////////////////// | ||
831 : | ;// 16b downsampling 16x16 -> 8x8 | ||
832 : | ;////////////////////////////////////////////////////////////////////// | ||
833 : | |||
834 : | %macro HFILTER_1331 2 ;%1:src %2:dst reg. -trashes mm0/mm1/mm2 | ||
835 : | movq mm2, [Mask_ff] | ||
836 : | movq %2, [%1-1] ;-10123456 | ||
837 : | movq mm0, [%1] ; 01234567 | ||
838 : | movq mm1, [%1+1] ; 12345678 | ||
839 : | pand %2, mm2 ;-1|1|3|5 | ||
840 : | pand mm0, mm2 ; 0|2|4|6 | ||
841 : | pand mm1, mm2 ; 1|3|5|7 | ||
842 : | pand mm2, [%1+2] ; 2|4|6|8 | ||
843 : | paddusw mm0, mm1 | ||
844 : | paddusw %2, mm2 | ||
845 : | pmullw mm0, mm7 | ||
846 : | paddusw %2, mm0 | ||
847 : | %endmacro | ||
848 : | |||
849 : | %macro VFILTER_1331 4 ; %1-4: regs %1-%2: trashed | ||
850 : | paddsw %1, [Cst32] | ||
851 : | edgomez | 1.2.2.1 | paddsw %2, %3 |
852 : | edgomez | 1.2 | pmullw %2, mm7 |
853 : | paddsw %1,%4 | ||
854 : | paddsw %1, %2 | ||
855 : | psraw %1, 6 | ||
856 : | %endmacro | ||
857 : | |||
858 : | ;=========================================================================== | ||
859 : | ; | ||
860 : | ; void xvid_Filter_18x18_To_8x8_mmx(int16_t *Dst, | ||
861 : | ; const uint8_t *Src, const int BpS); | ||
862 : | ; | ||
863 : | ;=========================================================================== | ||
864 : | |||
865 : | %macro COPY_TWO_LINES_1331 1 ; %1: dst | ||
866 : | HFILTER_1331 edx , mm5 | ||
867 : | HFILTER_1331 edx+eax, mm6 | ||
868 : | lea edx, [edx+2*eax] | ||
869 : | VFILTER_1331 mm3,mm4,mm5, mm6 | ||
870 : | movq [%1], mm3 | ||
871 : | |||
872 : | HFILTER_1331 edx , mm3 | ||
873 : | HFILTER_1331 edx+eax, mm4 | ||
874 : | lea edx, [edx+2*eax] | ||
875 : | VFILTER_1331 mm5,mm6,mm3,mm4 | ||
876 : | movq [%1+16], mm5 | ||
877 : | %endmacro | ||
878 : | |||
879 : | align 16 | ||
880 : | xvid_Filter_18x18_To_8x8_mmx: ; 283c (~4.4c per output pixel) | ||
881 : | |||
882 : | mov ecx, [esp+4] ; Dst | ||
883 : | mov edx, [esp+8] ; Src | ||
884 : | mov eax, [esp+12] ; BpS | ||
885 : | |||
886 : | movq mm7, [Cst3] | ||
887 : | sub edx, eax | ||
888 : | |||
889 : | ; mm3/mm4/mm5/mm6 is used as a 4-samples delay line. | ||
890 : | |||
891 : | ; process columns 0-3 | ||
892 : | |||
893 : | HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 | ||
894 : | HFILTER_1331 edx+eax, mm4 | ||
895 : | lea edx, [edx+2*eax] | ||
896 : | |||
897 : | COPY_TWO_LINES_1331 ecx + 0*16 | ||
898 : | COPY_TWO_LINES_1331 ecx + 2*16 | ||
899 : | COPY_TWO_LINES_1331 ecx + 4*16 | ||
900 : | COPY_TWO_LINES_1331 ecx + 6*16 | ||
901 : | |||
902 : | ; process columns 4-7 | ||
903 : | |||
904 : | mov edx, [esp+8] | ||
905 : | sub edx, eax | ||
906 : | add edx, 8 | ||
907 : | |||
908 : | HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 | ||
909 : | HFILTER_1331 edx+eax, mm4 | ||
910 : | lea edx, [edx+2*eax] | ||
911 : | |||
912 : | COPY_TWO_LINES_1331 ecx + 0*16 +8 | ||
913 : | COPY_TWO_LINES_1331 ecx + 2*16 +8 | ||
914 : | COPY_TWO_LINES_1331 ecx + 4*16 +8 | ||
915 : | COPY_TWO_LINES_1331 ecx + 6*16 +8 | ||
916 : | |||
917 : | ret | ||
918 : | |||
919 : | ;=========================================================================== | ||
920 : | ; | ||
921 : | ; void xvid_Filter_Diff_18x18_To_8x8_mmx(int16_t *Dst, | ||
922 : | ; const uint8_t *Src, const int BpS); | ||
923 : | ; | ||
924 : | ;=========================================================================== | ||
925 : | |||
926 : | %macro DIFF_TWO_LINES_1331 1 ; %1: dst | ||
927 : | HFILTER_1331 edx , mm5 | ||
928 : | HFILTER_1331 edx+eax, mm6 | ||
929 : | lea edx, [edx+2*eax] | ||
930 : | movq mm2, [%1] | ||
931 : | VFILTER_1331 mm3,mm4,mm5, mm6 | ||
932 : | psubsw mm2, mm3 | ||
933 : | movq [%1], mm2 | ||
934 : | |||
935 : | HFILTER_1331 edx , mm3 | ||
936 : | HFILTER_1331 edx+eax, mm4 | ||
937 : | lea edx, [edx+2*eax] | ||
938 : | movq mm2, [%1+16] | ||
939 : | VFILTER_1331 mm5,mm6,mm3,mm4 | ||
940 : | psubsw mm2, mm5 | ||
941 : | movq [%1+16], mm2 | ||
942 : | %endmacro | ||
943 : | |||
944 : | align 16 | ||
945 : | xvid_Filter_Diff_18x18_To_8x8_mmx: ; 302c | ||
946 : | |||
947 : | mov ecx, [esp+4] ; Dst | ||
948 : | mov edx, [esp+8] ; Src | ||
949 : | mov eax, [esp+12] ; BpS | ||
950 : | |||
951 : | movq mm7, [Cst3] | ||
952 : | sub edx, eax | ||
953 : | |||
954 : | ; mm3/mm4/mm5/mm6 is used as a 4-samples delay line. | ||
955 : | |||
956 : | ; process columns 0-3 | ||
957 : | |||
958 : | HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 | ||
959 : | HFILTER_1331 edx+eax, mm4 | ||
960 : | lea edx, [edx+2*eax] | ||
961 : | |||
962 : | DIFF_TWO_LINES_1331 ecx + 0*16 | ||
963 : | DIFF_TWO_LINES_1331 ecx + 2*16 | ||
964 : | DIFF_TWO_LINES_1331 ecx + 4*16 | ||
965 : | DIFF_TWO_LINES_1331 ecx + 6*16 | ||
966 : | |||
967 : | ; process columns 4-7 | ||
968 : | mov edx, [esp+8] | ||
969 : | sub edx, eax | ||
970 : | add edx, 8 | ||
971 : | |||
972 : | HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 | ||
973 : | HFILTER_1331 edx+eax, mm4 | ||
974 : | lea edx, [edx+2*eax] | ||
975 : | |||
976 : | DIFF_TWO_LINES_1331 ecx + 0*16 +8 | ||
977 : | DIFF_TWO_LINES_1331 ecx + 2*16 +8 | ||
978 : | DIFF_TWO_LINES_1331 ecx + 4*16 +8 | ||
979 : | DIFF_TWO_LINES_1331 ecx + 6*16 +8 | ||
980 : | |||
981 : | ret | ||
982 : | |||
983 : | ;////////////////////////////////////////////////////////////////////// | ||
984 : | |||
985 : | ; pfeewwww... Never Do That On Stage Again. :) | ||
986 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |