Parent Directory | Revision Log
Revision 1.8 - (view) (download)
1 : | edgomez | 1.3 | ;/***************************************************************************** |
2 : | edgomez | 1.2 | ; * |
3 : | edgomez | 1.3 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - 3dne pipeline optimized 8x8 block-based halfpel interpolation - | ||
5 : | edgomez | 1.2 | ; * |
6 : | edgomez | 1.3 | ; * Copyright(C) 2002 Jaan Kalda |
7 : | edgomez | 1.2 | ; * |
8 : | edgomez | 1.3 | ; * This program is free software ; you can redistribute it and/or modify |
9 : | ; * it 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 : | edgomez | 1.2 | ; * |
13 : | edgomez | 1.3 | ; * 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 : | edgomez | 1.2 | ; * |
18 : | edgomez | 1.3 | ; * 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 : | edgomez | 1.2 | ; * |
22 : | edgomez | 1.3 | ; ****************************************************************************/ |
23 : | edgomez | 1.2 | |
24 : | edgomez | 1.3 | ; these 3dne functions are compatible with iSSE, but are optimized specifically |
25 : | ; for K7 pipelines | ||
26 : | edgomez | 1.2 | |
27 : | edgomez | 1.3 | BITS 32 |
28 : | edgomez | 1.2 | |
29 : | edgomez | 1.3 | %macro cglobal 1 |
30 : | edgomez | 1.2 | %ifdef PREFIX |
31 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
32 : | edgomez | 1.6 | global _%1:function %1.endfunc-%1 |
33 : | %define %1 _%1:function %1.endfunc-%1 | ||
34 : | edgomez | 1.5 | %else |
35 : | global _%1 | ||
36 : | %define %1 _%1 | ||
37 : | %endif | ||
38 : | edgomez | 1.2 | %else |
39 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
40 : | edgomez | 1.6 | global %1:function %1.endfunc-%1 |
41 : | edgomez | 1.5 | %else |
42 : | global %1 | ||
43 : | %endif | ||
44 : | edgomez | 1.2 | %endif |
45 : | %endmacro | ||
46 : | edgomez | 1.3 | |
47 : | ;============================================================================= | ||
48 : | ; Read only data | ||
49 : | ;============================================================================= | ||
50 : | |||
51 : | edgomez | 1.2 | %ifdef FORMAT_COFF |
52 : | edgomez | 1.4 | SECTION .rodata |
53 : | edgomez | 1.2 | %else |
54 : | edgomez | 1.4 | SECTION .rodata align=16 |
55 : | edgomez | 1.2 | %endif |
56 : | |||
57 : | edgomez | 1.3 | ALIGN 16 |
58 : | mmx_one: | ||
59 : | times 8 db 1 | ||
60 : | |||
61 : | ALIGN 8 | ||
62 : | mm_minusone: | ||
63 : | dd -1,-1 | ||
64 : | edgomez | 1.2 | |
65 : | edgomez | 1.3 | ;============================================================================= |
66 : | ; Macros | ||
67 : | ;============================================================================= | ||
68 : | edgomez | 1.2 | |
69 : | edgomez | 1.3 | %macro nop4 0 |
70 : | DB 08Dh,074h,026h,0 | ||
71 : | %endmacro | ||
72 : | |||
73 : | ;============================================================================= | ||
74 : | ; Macros | ||
75 : | ;============================================================================= | ||
76 : | edgomez | 1.2 | |
77 : | edgomez | 1.3 | SECTION .text |
78 : | edgomez | 1.2 | |
79 : | cglobal interpolate8x8_halfpel_h_3dne | ||
80 : | cglobal interpolate8x8_halfpel_v_3dne | ||
81 : | cglobal interpolate8x8_halfpel_hv_3dne | ||
82 : | |||
83 : | suxen_drol | 1.7 | cglobal interpolate8x4_halfpel_h_3dne |
84 : | cglobal interpolate8x4_halfpel_v_3dne | ||
85 : | cglobal interpolate8x4_halfpel_hv_3dne | ||
86 : | |||
87 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
88 : | edgomez | 1.2 | ; |
89 : | ; void interpolate8x8_halfpel_h_3dne(uint8_t * const dst, | ||
90 : | edgomez | 1.3 | ; const uint8_t * const src, |
91 : | ; const uint32_t stride, | ||
92 : | ; const uint32_t rounding); | ||
93 : | edgomez | 1.2 | ; |
94 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
95 : | edgomez | 1.2 | |
96 : | %macro COPY_H_SSE_RND0 1 | ||
97 : | %if (%1) | ||
98 : | edgomez | 1.3 | movq mm0, [eax] |
99 : | edgomez | 1.2 | %else |
100 : | suxen_drol | 1.8 | movq mm0, [eax+0] |
101 : | ; --- | ||
102 : | ; nasm >0.99.x rejects the original statement: | ||
103 : | ; movq mm0, [dword eax] | ||
104 : | ; as it is ambiguous. for this statement nasm <0.99.x would | ||
105 : | ; generate "movq mm0,[eax+0]" | ||
106 : | ; --- | ||
107 : | edgomez | 1.2 | %endif |
108 : | pavgb mm0, [eax+1] | ||
109 : | edgomez | 1.3 | movq mm1, [eax+edx] |
110 : | edgomez | 1.2 | pavgb mm1, [eax+edx+1] |
111 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
112 : | movq [ecx], mm0 | ||
113 : | movq [ecx+edx], mm1 | ||
114 : | edgomez | 1.2 | %endmacro |
115 : | |||
116 : | %macro COPY_H_SSE_RND1 0 | ||
117 : | movq mm0, [eax] | ||
118 : | movq mm1, [eax+edx] | ||
119 : | movq mm4, mm0 | ||
120 : | movq mm5, mm1 | ||
121 : | edgomez | 1.3 | movq mm2, [eax+1] |
122 : | edgomez | 1.2 | movq mm3, [eax+edx+1] |
123 : | pavgb mm0, mm2 | ||
124 : | pxor mm2, mm4 | ||
125 : | pavgb mm1, mm3 | ||
126 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
127 : | edgomez | 1.2 | pxor mm3, mm5 |
128 : | pand mm2, mm7 | ||
129 : | pand mm3, mm7 | ||
130 : | psubb mm0, mm2 | ||
131 : | movq [ecx], mm0 | ||
132 : | psubb mm1, mm3 | ||
133 : | edgomez | 1.3 | movq [ecx+edx], mm1 |
134 : | edgomez | 1.2 | %endmacro |
135 : | |||
136 : | edgomez | 1.3 | ALIGN 16 |
137 : | edgomez | 1.2 | interpolate8x8_halfpel_h_3dne: |
138 : | |||
139 : | mov eax, [esp+ 8] ; Src | ||
140 : | mov edx, [esp+12] ; stride | ||
141 : | dec dword [esp+16]; rounding | ||
142 : | |||
143 : | jz .rounding1 | ||
144 : | mov ecx, [esp+ 4] ; Dst | ||
145 : | |||
146 : | COPY_H_SSE_RND0 0 | ||
147 : | lea ecx,[ecx+2*edx] | ||
148 : | COPY_H_SSE_RND0 1 | ||
149 : | lea ecx,[ecx+2*edx] | ||
150 : | COPY_H_SSE_RND0 1 | ||
151 : | lea ecx,[ecx+2*edx] | ||
152 : | COPY_H_SSE_RND0 1 | ||
153 : | ret | ||
154 : | |||
155 : | .rounding1 | ||
156 : | edgomez | 1.3 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
157 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
158 : | movq mm7, [mmx_one] | ||
159 : | COPY_H_SSE_RND1 | ||
160 : | lea ecx, [ecx+2*edx] | ||
161 : | COPY_H_SSE_RND1 | ||
162 : | lea ecx,[ecx+2*edx] | ||
163 : | COPY_H_SSE_RND1 | ||
164 : | lea ecx,[ecx+2*edx] | ||
165 : | COPY_H_SSE_RND1 | ||
166 : | ret | ||
167 : | edgomez | 1.6 | .endfunc |
168 : | edgomez | 1.2 | |
169 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
170 : | edgomez | 1.2 | ; |
171 : | ; void interpolate8x8_halfpel_v_3dne(uint8_t * const dst, | ||
172 : | edgomez | 1.3 | ; const uint8_t * const src, |
173 : | ; const uint32_t stride, | ||
174 : | ; const uint32_t rounding); | ||
175 : | edgomez | 1.2 | ; |
176 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
177 : | edgomez | 1.2 | |
178 : | edgomez | 1.3 | ALIGN 16 |
179 : | edgomez | 1.2 | interpolate8x8_halfpel_v_3dne: |
180 : | |||
181 : | mov eax, [esp+ 8] ; Src | ||
182 : | mov edx, [esp+12] ; stride | ||
183 : | dec dword [esp+16]; rounding | ||
184 : | |||
185 : | ; we process 2 line at a time | ||
186 : | |||
187 : | jz .rounding1 | ||
188 : | pxor mm2,mm2 | ||
189 : | edgomez | 1.3 | movq mm0, [eax] |
190 : | movq mm1, [eax+edx] | ||
191 : | por mm2, [eax+2*edx] | ||
192 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
193 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
194 : | pxor mm4, mm4 | ||
195 : | edgomez | 1.2 | pavgb mm0, mm1 |
196 : | edgomez | 1.3 | pavgb mm1, mm2 |
197 : | movq [byte ecx], mm0 | ||
198 : | movq [ecx+edx], mm1 | ||
199 : | pxor mm6, mm6 | ||
200 : | add eax, edx | ||
201 : | lea ecx, [ecx+2*edx] | ||
202 : | movq mm3, [byte eax] | ||
203 : | por mm4, [eax+edx] | ||
204 : | lea eax, [eax+2*edx] | ||
205 : | edgomez | 1.2 | pavgb mm2, mm3 |
206 : | pavgb mm3, mm4 | ||
207 : | edgomez | 1.3 | movq [ecx], mm2 |
208 : | movq [ecx+edx], mm3 | ||
209 : | lea ecx, [byte ecx+2*edx] | ||
210 : | movq mm5, [byte eax] | ||
211 : | por mm6, [eax+edx] | ||
212 : | lea eax, [eax+2*edx] | ||
213 : | edgomez | 1.2 | pavgb mm4, mm5 |
214 : | pavgb mm5, mm6 | ||
215 : | edgomez | 1.3 | movq [ecx], mm4 |
216 : | movq [ecx+edx], mm5 | ||
217 : | lea ecx, [ecx+2*edx] | ||
218 : | movq mm7, [eax] | ||
219 : | movq mm0, [eax+edx] | ||
220 : | edgomez | 1.2 | pavgb mm6, mm7 |
221 : | pavgb mm7, mm0 | ||
222 : | edgomez | 1.3 | movq [ecx], mm6 |
223 : | movq [ecx+edx], mm7 | ||
224 : | edgomez | 1.2 | ret |
225 : | |||
226 : | edgomez | 1.3 | ALIGN 8 |
227 : | edgomez | 1.2 | .rounding1 |
228 : | edgomez | 1.3 | pcmpeqb mm0, mm0 |
229 : | psubusb mm0, [eax] | ||
230 : | add eax, edx | ||
231 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
232 : | push esi | ||
233 : | edgomez | 1.3 | pcmpeqb mm1, mm1 |
234 : | pcmpeqb mm2, mm2 | ||
235 : | mov esi, mm_minusone | ||
236 : | psubusb mm1, [byte eax] | ||
237 : | psubusb mm2, [eax+edx] | ||
238 : | lea eax, [eax+2*edx] | ||
239 : | edgomez | 1.2 | movq mm6, [esi] |
240 : | movq mm7, [esi] | ||
241 : | pavgb mm0, mm1 | ||
242 : | pavgb mm1, mm2 | ||
243 : | edgomez | 1.3 | psubusb mm6, mm0 |
244 : | psubusb mm7, mm1 | ||
245 : | edgomez | 1.2 | movq [ecx], mm6 |
246 : | movq [ecx+edx], mm7 | ||
247 : | edgomez | 1.3 | lea ecx, [ecx+2*edx] |
248 : | pcmpeqb mm3, mm3 | ||
249 : | pcmpeqb mm4, mm4 | ||
250 : | psubusb mm3, [eax] | ||
251 : | psubusb mm4, [eax+edx] | ||
252 : | lea eax, [eax+2*edx] | ||
253 : | edgomez | 1.2 | pavgb mm2, mm3 |
254 : | pavgb mm3, mm4 | ||
255 : | movq mm0, [esi] | ||
256 : | movq mm1, [esi] | ||
257 : | edgomez | 1.3 | psubusb mm0, mm2 |
258 : | psubusb mm1, mm3 | ||
259 : | edgomez | 1.2 | movq [ecx], mm0 |
260 : | movq [ecx+edx], mm1 | ||
261 : | lea ecx,[ecx+2*edx] | ||
262 : | |||
263 : | edgomez | 1.3 | pcmpeqb mm5, mm5 |
264 : | pcmpeqb mm6, mm6 | ||
265 : | psubusb mm5, [eax] | ||
266 : | psubusb mm6, [eax+edx] | ||
267 : | lea eax, [eax+2*edx] | ||
268 : | edgomez | 1.2 | pavgb mm4, mm5 |
269 : | pavgb mm5, mm6 | ||
270 : | movq mm2, [esi] | ||
271 : | movq mm3, [esi] | ||
272 : | edgomez | 1.3 | psubusb mm2, mm4 |
273 : | psubusb mm3, mm5 | ||
274 : | edgomez | 1.2 | movq [ecx], mm2 |
275 : | movq [ecx+edx], mm3 | ||
276 : | edgomez | 1.3 | lea ecx, [ecx+2*edx] |
277 : | pcmpeqb mm7, mm7 | ||
278 : | pcmpeqb mm0, mm0 | ||
279 : | psubusb mm7, [eax] | ||
280 : | psubusb mm0, [eax+edx] | ||
281 : | edgomez | 1.2 | pavgb mm6, mm7 |
282 : | pavgb mm7, mm0 | ||
283 : | movq mm4, [esi] | ||
284 : | movq mm5, [esi] | ||
285 : | edgomez | 1.3 | psubusb mm4, mm6 |
286 : | edgomez | 1.2 | pop esi |
287 : | edgomez | 1.3 | psubusb mm5, mm7 |
288 : | edgomez | 1.2 | movq [ecx], mm4 |
289 : | movq [ecx+edx], mm5 | ||
290 : | ret | ||
291 : | edgomez | 1.6 | .endfunc |
292 : | edgomez | 1.3 | |
293 : | ;----------------------------------------------------------------------------- | ||
294 : | edgomez | 1.2 | ; |
295 : | ; void interpolate8x8_halfpel_hv_3dne(uint8_t * const dst, | ||
296 : | edgomez | 1.3 | ; const uint8_t * const src, |
297 : | ; const uint32_t stride, | ||
298 : | ; const uint32_t rounding); | ||
299 : | edgomez | 1.2 | ; |
300 : | ; | ||
301 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
302 : | edgomez | 1.2 | |
303 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
304 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
305 : | ; The boolean relations are: | ||
306 : | edgomez | 1.3 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
307 : | edgomez | 1.2 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
308 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
309 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
310 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
311 : | |||
312 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
313 : | |||
314 : | %macro COPY_HV_SSE_RND0 0 | ||
315 : | |||
316 : | edgomez | 1.3 | movq mm0, [eax+edx] |
317 : | movq mm1, [eax+edx+1] | ||
318 : | edgomez | 1.2 | |
319 : | edgomez | 1.3 | movq mm6, mm0 |
320 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
321 : | lea eax, [eax+2*edx] | ||
322 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
323 : | edgomez | 1.2 | |
324 : | edgomez | 1.3 | por mm3, mm1 ; ij |= jk |
325 : | movq mm6, mm2 | ||
326 : | pxor mm6, mm0 ; mm6 = s^t | ||
327 : | pand mm3, mm6 ; (ij|jk) &= st | ||
328 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
329 : | movq mm6, [eax] | ||
330 : | pand mm3, mm7 ; mask lsb | ||
331 : | psubb mm2, mm3 ; apply. | ||
332 : | |||
333 : | movq [ecx], mm2 | ||
334 : | edgomez | 1.2 | |
335 : | edgomez | 1.3 | movq mm2, [eax] |
336 : | movq mm3, [eax+1] | ||
337 : | pavgb mm2, mm3 ; preserved for next iteration | ||
338 : | pxor mm3, mm6 ; preserved for next iteration | ||
339 : | |||
340 : | por mm1, mm3 | ||
341 : | movq mm6, mm0 | ||
342 : | pxor mm6, mm2 | ||
343 : | pand mm1, mm6 | ||
344 : | pavgb mm0, mm2 | ||
345 : | |||
346 : | pand mm1, mm7 | ||
347 : | psubb mm0, mm1 | ||
348 : | |||
349 : | movq [ecx+edx], mm0 | ||
350 : | edgomez | 1.2 | %endmacro |
351 : | |||
352 : | %macro COPY_HV_SSE_RND1 0 | ||
353 : | edgomez | 1.3 | movq mm0, [eax+edx] |
354 : | movq mm1, [eax+edx+1] | ||
355 : | |||
356 : | movq mm6, mm0 | ||
357 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
358 : | lea eax,[eax+2*edx] | ||
359 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
360 : | edgomez | 1.2 | |
361 : | edgomez | 1.3 | pand mm3, mm1 |
362 : | movq mm6, mm2 | ||
363 : | pxor mm6, mm0 | ||
364 : | por mm3, mm6 | ||
365 : | pavgb mm2, mm0 | ||
366 : | movq mm6, [eax] | ||
367 : | pand mm3, mm7 | ||
368 : | psubb mm2, mm3 | ||
369 : | |||
370 : | movq [ecx], mm2 | ||
371 : | |||
372 : | movq mm2, [eax] | ||
373 : | movq mm3, [eax+1] | ||
374 : | pavgb mm2, mm3 ; preserved for next iteration | ||
375 : | pxor mm3, mm6 ; preserved for next iteration | ||
376 : | |||
377 : | pand mm1, mm3 | ||
378 : | movq mm6, mm0 | ||
379 : | pxor mm6, mm2 | ||
380 : | por mm1, mm6 | ||
381 : | pavgb mm0, mm2 | ||
382 : | pand mm1, mm7 | ||
383 : | psubb mm0, mm1 | ||
384 : | movq [ecx+edx], mm0 | ||
385 : | edgomez | 1.2 | %endmacro |
386 : | |||
387 : | edgomez | 1.3 | ALIGN 16 |
388 : | edgomez | 1.2 | interpolate8x8_halfpel_hv_3dne: |
389 : | edgomez | 1.3 | mov eax, [esp+ 8] ; Src |
390 : | mov edx, [esp+12] ; stride | ||
391 : | dec dword [esp+16] ; rounding | ||
392 : | edgomez | 1.2 | |
393 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
394 : | movq mm2, [eax] | ||
395 : | movq mm3, [eax+1] | ||
396 : | movq mm6, mm2 | ||
397 : | pavgb mm2, mm3 | ||
398 : | edgomez | 1.3 | pxor mm3, mm6 ; mm2/mm3 ready |
399 : | mov ecx, [esp+ 4] ; Dst | ||
400 : | edgomez | 1.2 | movq mm7, [mmx_one] |
401 : | |||
402 : | jz near .rounding1 | ||
403 : | edgomez | 1.3 | lea ebp,[byte ebp] |
404 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
405 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
406 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
407 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
408 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
409 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
410 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
411 : | ret | ||
412 : | |||
413 : | edgomez | 1.3 | ALIGN 16 |
414 : | edgomez | 1.2 | .rounding1 |
415 : | COPY_HV_SSE_RND1 | ||
416 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
417 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
418 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
419 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
420 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
421 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
422 : | edgomez | 1.3 | ret |
423 : | edgomez | 1.6 | .endfunc |
424 : | |||
425 : | suxen_drol | 1.7 | ;----------------------------------------------------------------------------- |
426 : | ; | ||
427 : | ; void interpolate8x4_halfpel_h_3dne(uint8_t * const dst, | ||
428 : | ; const uint8_t * const src, | ||
429 : | ; const uint32_t stride, | ||
430 : | ; const uint32_t rounding); | ||
431 : | ; | ||
432 : | ;----------------------------------------------------------------------------- | ||
433 : | |||
434 : | ALIGN 16 | ||
435 : | interpolate8x4_halfpel_h_3dne: | ||
436 : | |||
437 : | mov eax, [esp+ 8] ; Src | ||
438 : | mov edx, [esp+12] ; stride | ||
439 : | dec dword [esp+16]; rounding | ||
440 : | |||
441 : | jz .rounding1 | ||
442 : | mov ecx, [esp+ 4] ; Dst | ||
443 : | |||
444 : | COPY_H_SSE_RND0 0 | ||
445 : | lea ecx,[ecx+2*edx] | ||
446 : | COPY_H_SSE_RND0 1 | ||
447 : | ret | ||
448 : | |||
449 : | .rounding1 | ||
450 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
451 : | mov ecx, [esp+ 4] ; Dst | ||
452 : | movq mm7, [mmx_one] | ||
453 : | COPY_H_SSE_RND1 | ||
454 : | lea ecx, [ecx+2*edx] | ||
455 : | COPY_H_SSE_RND1 | ||
456 : | ret | ||
457 : | .endfunc | ||
458 : | |||
459 : | ;----------------------------------------------------------------------------- | ||
460 : | ; | ||
461 : | ; void interpolate8x4_halfpel_v_3dne(uint8_t * const dst, | ||
462 : | ; const uint8_t * const src, | ||
463 : | ; const uint32_t stride, | ||
464 : | ; const uint32_t rounding); | ||
465 : | ; | ||
466 : | ;----------------------------------------------------------------------------- | ||
467 : | |||
468 : | ALIGN 16 | ||
469 : | interpolate8x4_halfpel_v_3dne: | ||
470 : | |||
471 : | mov eax, [esp+ 8] ; Src | ||
472 : | mov edx, [esp+12] ; stride | ||
473 : | dec dword [esp+16]; rounding | ||
474 : | |||
475 : | ; we process 2 line at a time | ||
476 : | |||
477 : | jz .rounding1 | ||
478 : | pxor mm2,mm2 | ||
479 : | movq mm0, [eax] | ||
480 : | movq mm1, [eax+edx] | ||
481 : | por mm2, [eax+2*edx] ; Something like preload (pipelining) | ||
482 : | mov ecx, [esp+ 4] ; Dst | ||
483 : | lea eax, [eax+2*edx] | ||
484 : | pxor mm4, mm4 | ||
485 : | pavgb mm0, mm1 | ||
486 : | pavgb mm1, mm2 | ||
487 : | movq [byte ecx], mm0 | ||
488 : | movq [ecx+edx], mm1 | ||
489 : | |||
490 : | pxor mm6, mm6 | ||
491 : | add eax, edx | ||
492 : | lea ecx, [ecx+2*edx] | ||
493 : | movq mm3, [byte eax] | ||
494 : | por mm4, [eax+edx] | ||
495 : | lea eax, [eax+2*edx] | ||
496 : | pavgb mm2, mm3 | ||
497 : | pavgb mm3, mm4 | ||
498 : | movq [ecx], mm2 | ||
499 : | movq [ecx+edx], mm3 | ||
500 : | |||
501 : | ret | ||
502 : | |||
503 : | ALIGN 8 | ||
504 : | .rounding1 | ||
505 : | pcmpeqb mm0, mm0 | ||
506 : | psubusb mm0, [eax] ; eax==line0 | ||
507 : | add eax, edx ; eax==line1 | ||
508 : | mov ecx, [esp+ 4] ; Dst | ||
509 : | |||
510 : | push esi | ||
511 : | |||
512 : | pcmpeqb mm1, mm1 | ||
513 : | pcmpeqb mm2, mm2 | ||
514 : | mov esi, mm_minusone | ||
515 : | psubusb mm1, [byte eax] ; line1 | ||
516 : | psubusb mm2, [eax+edx] ; line2 | ||
517 : | lea eax, [eax+2*edx] ; eax==line3 | ||
518 : | movq mm6, [esi] | ||
519 : | movq mm7, [esi] | ||
520 : | pavgb mm0, mm1 | ||
521 : | pavgb mm1, mm2 | ||
522 : | psubusb mm6, mm0 | ||
523 : | psubusb mm7, mm1 | ||
524 : | movq [ecx], mm6 ; store line0 | ||
525 : | movq [ecx+edx], mm7 ; store line1 | ||
526 : | |||
527 : | lea ecx, [ecx+2*edx] | ||
528 : | pcmpeqb mm3, mm3 | ||
529 : | pcmpeqb mm4, mm4 | ||
530 : | psubusb mm3, [eax] ; line3 | ||
531 : | psubusb mm4, [eax+edx] ; line4 | ||
532 : | lea eax, [eax+2*edx] ; eax==line 5 | ||
533 : | pavgb mm2, mm3 | ||
534 : | pavgb mm3, mm4 | ||
535 : | movq mm0, [esi] | ||
536 : | movq mm1, [esi] | ||
537 : | psubusb mm0, mm2 | ||
538 : | psubusb mm1, mm3 | ||
539 : | movq [ecx], mm0 | ||
540 : | movq [ecx+edx], mm1 | ||
541 : | |||
542 : | pop esi | ||
543 : | |||
544 : | ret | ||
545 : | |||
546 : | .endfunc | ||
547 : | |||
548 : | ;----------------------------------------------------------------------------- | ||
549 : | ; | ||
550 : | ; void interpolate8x4_halfpel_hv_3dne(uint8_t * const dst, | ||
551 : | ; const uint8_t * const src, | ||
552 : | ; const uint32_t stride, | ||
553 : | ; const uint32_t rounding); | ||
554 : | ; | ||
555 : | ; | ||
556 : | ;----------------------------------------------------------------------------- | ||
557 : | |||
558 : | ALIGN 16 | ||
559 : | interpolate8x4_halfpel_hv_3dne: | ||
560 : | mov eax, [esp+ 8] ; Src | ||
561 : | mov edx, [esp+12] ; stride | ||
562 : | dec dword [esp+16] ; rounding | ||
563 : | |||
564 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
565 : | movq mm2, [eax] | ||
566 : | movq mm3, [eax+1] | ||
567 : | movq mm6, mm2 | ||
568 : | pavgb mm2, mm3 | ||
569 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
570 : | mov ecx, [esp+ 4] ; Dst | ||
571 : | movq mm7, [mmx_one] | ||
572 : | |||
573 : | jz near .rounding1 | ||
574 : | lea ebp,[byte ebp] | ||
575 : | COPY_HV_SSE_RND0 | ||
576 : | lea ecx,[ecx+2*edx] | ||
577 : | COPY_HV_SSE_RND0 | ||
578 : | ret | ||
579 : | |||
580 : | ALIGN 16 | ||
581 : | .rounding1 | ||
582 : | COPY_HV_SSE_RND1 | ||
583 : | lea ecx,[ecx+2*edx] | ||
584 : | COPY_HV_SSE_RND1 | ||
585 : | ret | ||
586 : | .endfunc | ||
587 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |