Parent Directory | Revision Log
Revision 1.2.2.1 - (view) (download)
1 : | edgomez | 1.2.2.1 | ;/**************************************************************************** |
2 : | edgomez | 1.2 | ; * |
3 : | edgomez | 1.2.2.1 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - 8<->16 bit transfer functions - | ||
5 : | edgomez | 1.2 | ; * |
6 : | edgomez | 1.2.2.1 | ; * Copyright (C) 2002 Jaan Kalda |
7 : | edgomez | 1.2 | ; * |
8 : | edgomez | 1.2.2.1 | ; * 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 | 1.2.2.1 | ; * 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 | 1.2.2.1 | ; * 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 | 1.2.2.1 | ; * $Id$ |
23 : | ; * | ||
24 : | ; ***************************************************************************/ | ||
25 : | edgomez | 1.2 | |
26 : | edgomez | 1.2.2.1 | ; these 3dne functions are compatible with iSSE, but are optimized specifically |
27 : | ; for K7 pipelines | ||
28 : | edgomez | 1.2 | |
29 : | edgomez | 1.2.2.1 | BITS 32 |
30 : | edgomez | 1.2 | |
31 : | edgomez | 1.2.2.1 | %macro cglobal 1 |
32 : | edgomez | 1.2 | %ifdef PREFIX |
33 : | edgomez | 1.2.2.1 | global _%1 |
34 : | edgomez | 1.2 | %define %1 _%1 |
35 : | %else | ||
36 : | global %1 | ||
37 : | %endif | ||
38 : | %endmacro | ||
39 : | edgomez | 1.2.2.1 | |
40 : | ;============================================================================= | ||
41 : | ; Read only data | ||
42 : | ;============================================================================= | ||
43 : | |||
44 : | SECTION .rodata | ||
45 : | |||
46 : | ALIGN 8 | ||
47 : | mm_zero: | ||
48 : | dd 0,0 | ||
49 : | ;============================================================================= | ||
50 : | ; Macros | ||
51 : | ;============================================================================= | ||
52 : | |||
53 : | edgomez | 1.2 | %macro nop4 0 |
54 : | edgomez | 1.2.2.1 | db 08Dh, 074h, 026h, 0 |
55 : | edgomez | 1.2 | %endmacro |
56 : | |||
57 : | edgomez | 1.2.2.1 | ;============================================================================= |
58 : | ; Code | ||
59 : | ;============================================================================= | ||
60 : | |||
61 : | SECTION .text | ||
62 : | edgomez | 1.2 | |
63 : | cglobal transfer_8to16copy_3dne | ||
64 : | cglobal transfer_16to8copy_3dne | ||
65 : | cglobal transfer_8to16sub_3dne | ||
66 : | cglobal transfer_8to16subro_3dne | ||
67 : | cglobal transfer_8to16sub2_3dne | ||
68 : | cglobal transfer_16to8add_3dne | ||
69 : | cglobal transfer8x8_copy_3dne | ||
70 : | |||
71 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
72 : | edgomez | 1.2 | ; |
73 : | ; void transfer_8to16copy_3dne(int16_t * const dst, | ||
74 : | ; const uint8_t * const src, | ||
75 : | ; uint32_t stride); | ||
76 : | ; | ||
77 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
78 : | edgomez | 1.2 | |
79 : | edgomez | 1.2.2.1 | ALIGN 16 |
80 : | edgomez | 1.2 | transfer_8to16copy_3dne: |
81 : | |||
82 : | mov eax, [esp+ 8] ; Src | ||
83 : | mov edx, [esp+12] ; Stride | ||
84 : | mov ecx, [esp+ 4] ; Dst | ||
85 : | punpcklbw mm0, [byte eax] | ||
86 : | punpcklbw mm1, [eax+4] | ||
87 : | edgomez | 1.2.2.1 | movq mm2, [eax+edx] |
88 : | movq mm3, [eax+edx] | ||
89 : | pxor mm7, mm7 | ||
90 : | lea eax, [eax+2*edx] | ||
91 : | punpcklbw mm2, mm7 | ||
92 : | punpckhbw mm3, mm7 | ||
93 : | psrlw mm0, 8 | ||
94 : | psrlw mm1, 8 | ||
95 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
96 : | punpcklbw mm5, [eax+edx+4] | ||
97 : | movq [byte ecx+0*64], mm0 | ||
98 : | movq [ecx+0*64+8], mm1 | ||
99 : | punpcklbw mm6, [eax+edx] | ||
100 : | punpcklbw mm7, [eax+4] | ||
101 : | edgomez | 1.2.2.1 | lea eax, [byte eax+2*edx] |
102 : | psrlw mm4, 8 | ||
103 : | psrlw mm5, 8 | ||
104 : | edgomez | 1.2 | punpcklbw mm0, [eax] |
105 : | punpcklbw mm1, [eax+edx+4] | ||
106 : | movq [ecx+0*64+16], mm2 | ||
107 : | movq [ecx+0*64+24], mm3 | ||
108 : | edgomez | 1.2.2.1 | psrlw mm6, 8 |
109 : | psrlw mm7, 8 | ||
110 : | edgomez | 1.2 | punpcklbw mm2, [eax+edx] |
111 : | punpcklbw mm3, [eax+4] | ||
112 : | edgomez | 1.2.2.1 | lea eax, [byte eax+2*edx] |
113 : | edgomez | 1.2 | movq [byte ecx+0*64+32], mm4 |
114 : | movq [ecx+0*64+56], mm5 | ||
115 : | edgomez | 1.2.2.1 | psrlw mm0, 8 |
116 : | psrlw mm1, 8 | ||
117 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
118 : | punpcklbw mm5, [eax+edx+4] | ||
119 : | movq [byte ecx+0*64+48], mm6 | ||
120 : | movq [ecx+0*64+40], mm7 | ||
121 : | edgomez | 1.2.2.1 | psrlw mm2, 8 |
122 : | psrlw mm3, 8 | ||
123 : | edgomez | 1.2 | punpcklbw mm6, [eax+edx] |
124 : | punpcklbw mm7, [eax+4] | ||
125 : | movq [byte ecx+1*64], mm0 | ||
126 : | movq [ecx+1*64+24], mm1 | ||
127 : | edgomez | 1.2.2.1 | psrlw mm4, 8 |
128 : | psrlw mm5, 8 | ||
129 : | edgomez | 1.2 | movq [ecx+1*64+16], mm2 |
130 : | movq [ecx+1*64+8], mm3 | ||
131 : | edgomez | 1.2.2.1 | psrlw mm6, 8 |
132 : | psrlw mm7, 8 | ||
133 : | edgomez | 1.2 | movq [byte ecx+1*64+32], mm4 |
134 : | edgomez | 1.2.2.1 | movq [ecx+1*64+56], mm5 |
135 : | edgomez | 1.2 | movq [byte ecx+1*64+48], mm6 |
136 : | movq [ecx+1*64+40], mm7 | ||
137 : | edgomez | 1.2.2.1 | ret |
138 : | edgomez | 1.2 | |
139 : | edgomez | 1.2.2.1 | |
140 : | |||
141 : | ;----------------------------------------------------------------------------- | ||
142 : | edgomez | 1.2 | ; |
143 : | ; void transfer_16to8copy_3dne(uint8_t * const dst, | ||
144 : | ; const int16_t * const src, | ||
145 : | ; uint32_t stride); | ||
146 : | ; | ||
147 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
148 : | edgomez | 1.2 | |
149 : | edgomez | 1.2.2.1 | ALIGN 16 |
150 : | edgomez | 1.2 | transfer_16to8copy_3dne: |
151 : | |||
152 : | mov eax, [esp+ 8] ; Src | ||
153 : | mov ecx, [esp+ 4] ; Dst | ||
154 : | mov edx, [esp+12] ; Stride | ||
155 : | |||
156 : | movq mm0, [byte eax+0*32] | ||
157 : | edgomez | 1.2.2.1 | packuswb mm0, [eax+0*32+8] |
158 : | edgomez | 1.2 | movq mm1, [eax+0*32+16] |
159 : | edgomez | 1.2.2.1 | packuswb mm1, [eax+0*32+24] |
160 : | edgomez | 1.2 | movq mm5, [eax+2*32+16] |
161 : | movq mm2, [eax+1*32] | ||
162 : | edgomez | 1.2.2.1 | packuswb mm2, [eax+1*32+8] |
163 : | edgomez | 1.2 | movq mm3, [eax+1*32+16] |
164 : | packuswb mm3, [eax+1*32+24] | ||
165 : | edgomez | 1.2.2.1 | movq mm6, [eax+3*32] |
166 : | edgomez | 1.2 | movq mm4, [eax+2*32] |
167 : | edgomez | 1.2.2.1 | packuswb mm4, [eax+2*32+8] |
168 : | edgomez | 1.2 | packuswb mm5, [eax+2*32+24] |
169 : | movq mm7, [eax+3*32+16] | ||
170 : | packuswb mm7, [eax+3*32+24] | ||
171 : | edgomez | 1.2.2.1 | packuswb mm6, [eax+3*32+8] |
172 : | edgomez | 1.2 | movq [ecx], mm0 |
173 : | edgomez | 1.2.2.1 | lea eax, [3*edx] |
174 : | add eax, ecx | ||
175 : | edgomez | 1.2 | movq [ecx+edx], mm1 |
176 : | movq [ecx+2*edx], mm2 | ||
177 : | movq [byte eax], mm3 | ||
178 : | movq [ecx+4*edx], mm4 | ||
179 : | edgomez | 1.2.2.1 | lea ecx, [byte ecx+4*edx] |
180 : | edgomez | 1.2 | movq [eax+2*edx], mm5 |
181 : | movq [eax+4*edx], mm7 | ||
182 : | movq [ecx+2*edx], mm6 | ||
183 : | ret | ||
184 : | |||
185 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
186 : | edgomez | 1.2 | ; |
187 : | ; void transfer_8to16sub_3dne(int16_t * const dct, | ||
188 : | ; uint8_t * const cur, | ||
189 : | ; const uint8_t * const ref, | ||
190 : | ; const uint32_t stride); | ||
191 : | ; | ||
192 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
193 : | edgomez | 1.2 | |
194 : | ; when second argument == 1, reference (ebx) block is to current (eax) | ||
195 : | %macro COPY_8_TO_16_SUB 2 | ||
196 : | movq mm1, [eax] ; cur | ||
197 : | movq mm0, mm1 | ||
198 : | movq mm4, [ecx] ; ref | ||
199 : | movq mm6, mm4 | ||
200 : | %if %2 == 1 | ||
201 : | movq [eax], mm4 | ||
202 : | %endif | ||
203 : | punpckhbw mm1, mm7 | ||
204 : | punpckhbw mm6, mm7 | ||
205 : | punpcklbw mm4, mm7 | ||
206 : | edgomez | 1.2.2.1 | ALIGN 8 |
207 : | edgomez | 1.2 | movq mm2, [byte eax+edx] |
208 : | punpcklbw mm0, mm7 | ||
209 : | movq mm3, [byte eax+edx] | ||
210 : | punpcklbw mm2, mm7 | ||
211 : | movq mm5, [byte ecx+edx] ; ref | ||
212 : | punpckhbw mm3, mm7 | ||
213 : | %if %2 == 1 | ||
214 : | movq [byte eax+edx], mm5 | ||
215 : | %endif | ||
216 : | psubsw mm1, mm6 | ||
217 : | |||
218 : | movq mm6, mm5 | ||
219 : | psubsw mm0, mm4 | ||
220 : | edgomez | 1.2.2.1 | %if (%1 < 3) |
221 : | edgomez | 1.2 | lea eax,[eax+2*edx] |
222 : | lea ecx,[ecx+2*edx] | ||
223 : | %else | ||
224 : | mov ecx,[esp] | ||
225 : | add esp,byte 4 | ||
226 : | %endif | ||
227 : | movq [edi+%1*32+ 8], mm1 | ||
228 : | movq [byte edi+%1*32+ 0], mm0 ; dst | ||
229 : | punpcklbw mm5, mm7 | ||
230 : | punpckhbw mm6, mm7 | ||
231 : | psubsw mm2, mm5 | ||
232 : | psubsw mm3, mm6 | ||
233 : | movq [edi+%1*32+16], mm2 | ||
234 : | movq [edi+%1*32+24], mm3 | ||
235 : | %endmacro | ||
236 : | |||
237 : | edgomez | 1.2.2.1 | ALIGN 16 |
238 : | edgomez | 1.2 | transfer_8to16sub_3dne: |
239 : | mov eax, [esp + 8] ; Cur | ||
240 : | mov ecx, [esp +12] ; Ref | ||
241 : | push edi | ||
242 : | mov edx, [dword esp+4+16] ; Stride | ||
243 : | mov edi, [esp+4+ 4] ; Dst | ||
244 : | pxor mm7, mm7 | ||
245 : | nop | ||
246 : | edgomez | 1.2.2.1 | ALIGN 4 |
247 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 1 |
248 : | COPY_8_TO_16_SUB 1, 1 | ||
249 : | COPY_8_TO_16_SUB 2, 1 | ||
250 : | COPY_8_TO_16_SUB 3, 1 | ||
251 : | edgomez | 1.2.2.1 | mov edi, ecx |
252 : | edgomez | 1.2 | ret |
253 : | |||
254 : | edgomez | 1.2.2.1 | ALIGN 16 |
255 : | edgomez | 1.2 | transfer_8to16subro_3dne: |
256 : | mov eax, [esp + 8] ; Cur | ||
257 : | mov ecx, [esp +12] ; Ref | ||
258 : | push edi | ||
259 : | mov edx, [dword esp+4+16] ; Stride | ||
260 : | mov edi, [esp+4+ 4] ; Dst | ||
261 : | pxor mm7, mm7 | ||
262 : | nop | ||
263 : | edgomez | 1.2.2.1 | ALIGN 4 |
264 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 0 |
265 : | COPY_8_TO_16_SUB 1, 0 | ||
266 : | COPY_8_TO_16_SUB 2, 0 | ||
267 : | COPY_8_TO_16_SUB 3, 0 | ||
268 : | edgomez | 1.2.2.1 | mov edi, ecx |
269 : | edgomez | 1.2 | ret |
270 : | |||
271 : | |||
272 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
273 : | edgomez | 1.2 | ; |
274 : | ; void transfer_8to16sub2_3dne(int16_t * const dct, | ||
275 : | ; uint8_t * const cur, | ||
276 : | ; const uint8_t * ref1, | ||
277 : | ; const uint8_t * ref2, | ||
278 : | ; const uint32_t stride) | ||
279 : | ; | ||
280 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
281 : | edgomez | 1.2 | |
282 : | %macro COPY_8_TO_16_SUB2_SSE 1 | ||
283 : | edgomez | 1.2.2.1 | db 0Fh, 6Fh, 44h, 20h, 00 ;movq mm0, [byte eax] ; cur |
284 : | edgomez | 1.2 | punpcklbw mm0, mm7 |
285 : | movq mm2, [byte eax+edx] | ||
286 : | punpcklbw mm2, mm7 | ||
287 : | db 0Fh, 6Fh, 4ch, 20h, 00 ;movq mm1, [byte eax] | ||
288 : | punpckhbw mm1, mm7 | ||
289 : | movq mm3, [byte eax+edx] | ||
290 : | punpckhbw mm3, mm7 | ||
291 : | edgomez | 1.2.2.1 | |
292 : | edgomez | 1.2 | movq mm4, [byte ebx] ; ref1 |
293 : | pavgb mm4, [byte esi] ; ref2 | ||
294 : | movq mm5, [ebx+edx] ; ref | ||
295 : | pavgb mm5, [esi+edx] ; ref2 | ||
296 : | movq mm6, mm4 | ||
297 : | punpcklbw mm4, mm7 | ||
298 : | punpckhbw mm6, mm7 | ||
299 : | %if (%1 < 3) | ||
300 : | lea esi,[esi+2*edx] | ||
301 : | lea ebx,[byte ebx+2*edx] | ||
302 : | lea eax,[eax+2*edx] | ||
303 : | %else | ||
304 : | mov esi,[esp] | ||
305 : | mov ebx,[esp+4] | ||
306 : | add esp,byte 8 | ||
307 : | edgomez | 1.2.2.1 | %endif |
308 : | edgomez | 1.2 | psubsw mm0, mm4 |
309 : | psubsw mm1, mm6 | ||
310 : | movq mm6, mm5 | ||
311 : | punpcklbw mm5, mm7 | ||
312 : | punpckhbw mm6, mm7 | ||
313 : | psubsw mm2, mm5 | ||
314 : | psubsw mm3, mm6 | ||
315 : | movq [byte ecx+%1*32+ 0], mm0 ; dst | ||
316 : | movq [ecx+%1*32+ 8], mm1 | ||
317 : | movq [ecx+%1*32+16], mm2 | ||
318 : | movq [ecx+%1*32+24], mm3 | ||
319 : | %endmacro | ||
320 : | |||
321 : | edgomez | 1.2.2.1 | ALIGN 16 |
322 : | edgomez | 1.2 | transfer_8to16sub2_3dne: |
323 : | mov edx, [esp +20] ; Stride | ||
324 : | mov ecx, [esp + 4] ; Dst | ||
325 : | mov eax, [esp + 8] ; Cur | ||
326 : | push ebx | ||
327 : | lea ebp,[byte ebp] | ||
328 : | mov ebx, [esp+4+12] ; Ref1 | ||
329 : | push esi | ||
330 : | pxor mm7, mm7 | ||
331 : | mov esi, [esp+8+16] ; Ref2 | ||
332 : | edgomez | 1.2.2.1 | nop4 |
333 : | edgomez | 1.2 | COPY_8_TO_16_SUB2_SSE 0 |
334 : | COPY_8_TO_16_SUB2_SSE 1 | ||
335 : | COPY_8_TO_16_SUB2_SSE 2 | ||
336 : | COPY_8_TO_16_SUB2_SSE 3 | ||
337 : | |||
338 : | ret | ||
339 : | |||
340 : | |||
341 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
342 : | edgomez | 1.2 | ; |
343 : | ; void transfer_16to8add_3dne(uint8_t * const dst, | ||
344 : | ; const int16_t * const src, | ||
345 : | ; uint32_t stride); | ||
346 : | ; | ||
347 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
348 : | edgomez | 1.2 | |
349 : | %macro COPY_16_TO_8_ADD 1 | ||
350 : | db 0Fh, 6Fh, 44h, 21h, 00 ;movq mm0, [byte ecx] | ||
351 : | punpcklbw mm0, mm7 | ||
352 : | movq mm2, [byte ecx+edx] | ||
353 : | punpcklbw mm2, mm7 | ||
354 : | db 0Fh, 6Fh, 4ch, 21h, 00 ;movq mm1, [byte ecx] | ||
355 : | punpckhbw mm1, mm7 | ||
356 : | movq mm3, [byte ecx+edx] | ||
357 : | punpckhbw mm3, mm7 | ||
358 : | paddsw mm0, [byte eax+%1*32+ 0] | ||
359 : | paddsw mm1, [eax+%1*32+ 8] | ||
360 : | paddsw mm2, [eax+%1*32+16] | ||
361 : | paddsw mm3, [eax+%1*32+24] | ||
362 : | packuswb mm0, mm1 | ||
363 : | packuswb mm2, mm3 | ||
364 : | edgomez | 1.2.2.1 | mov esp,esp |
365 : | edgomez | 1.2 | movq [byte ecx], mm0 |
366 : | movq [ecx+edx], mm2 | ||
367 : | %endmacro | ||
368 : | |||
369 : | |||
370 : | edgomez | 1.2.2.1 | ALIGN 16 |
371 : | edgomez | 1.2 | transfer_16to8add_3dne: |
372 : | mov ecx, [esp+ 4] ; Dst | ||
373 : | mov edx, [esp+12] ; Stride | ||
374 : | mov eax, [esp+ 8] ; Src | ||
375 : | pxor mm7, mm7 | ||
376 : | nop | ||
377 : | |||
378 : | COPY_16_TO_8_ADD 0 | ||
379 : | lea ecx,[byte ecx+2*edx] | ||
380 : | COPY_16_TO_8_ADD 1 | ||
381 : | lea ecx,[byte ecx+2*edx] | ||
382 : | COPY_16_TO_8_ADD 2 | ||
383 : | lea ecx,[byte ecx+2*edx] | ||
384 : | COPY_16_TO_8_ADD 3 | ||
385 : | ret | ||
386 : | |||
387 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
388 : | edgomez | 1.2 | ; |
389 : | ; void transfer8x8_copy_3dne(uint8_t * const dst, | ||
390 : | ; const uint8_t * const src, | ||
391 : | ; const uint32_t stride); | ||
392 : | ; | ||
393 : | ; | ||
394 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
395 : | edgomez | 1.2 | |
396 : | %macro COPY_8_TO_8 0 | ||
397 : | movq mm0, [byte eax] | ||
398 : | movq mm1, [eax+edx] | ||
399 : | movq [byte ecx], mm0 | ||
400 : | lea eax,[byte eax+2*edx] | ||
401 : | movq [ecx+edx], mm1 | ||
402 : | %endmacro | ||
403 : | |||
404 : | edgomez | 1.2.2.1 | ALIGN 16 |
405 : | edgomez | 1.2 | transfer8x8_copy_3dne: |
406 : | mov eax, [esp+ 8] ; Src | ||
407 : | mov edx, [esp+12] ; Stride | ||
408 : | mov ecx, [esp+ 4] ; Dst | ||
409 : | |||
410 : | COPY_8_TO_8 | ||
411 : | lea ecx,[byte ecx+2*edx] | ||
412 : | COPY_8_TO_8 | ||
413 : | lea ecx,[byte ecx+2*edx] | ||
414 : | COPY_8_TO_8 | ||
415 : | lea ecx,[byte ecx+2*edx] | ||
416 : | COPY_8_TO_8 | ||
417 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |