Parent Directory | Revision Log
Revision 1.4 - (view) (download)
1 : | edgomez | 1.3 | ;/**************************************************************************** |
2 : | edgomez | 1.2 | ; * |
3 : | edgomez | 1.3 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - 8<->16 bit transfer functions - | ||
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 | ; * $Id$ |
23 : | ; * | ||
24 : | ; ***************************************************************************/ | ||
25 : | |||
26 : | ; these 3dne functions are compatible with iSSE, but are optimized specifically | ||
27 : | ; for K7 pipelines | ||
28 : | |||
29 : | BITS 32 | ||
30 : | edgomez | 1.2 | |
31 : | edgomez | 1.3 | %macro cglobal 1 |
32 : | %ifdef PREFIX | ||
33 : | global _%1 | ||
34 : | %define %1 _%1 | ||
35 : | %else | ||
36 : | global %1 | ||
37 : | %endif | ||
38 : | %endmacro | ||
39 : | edgomez | 1.2 | |
40 : | edgomez | 1.3 | ;============================================================================= |
41 : | ; Read only data | ||
42 : | ;============================================================================= | ||
43 : | edgomez | 1.2 | |
44 : | %ifdef FORMAT_COFF | ||
45 : | edgomez | 1.4 | SECTION .rodata |
46 : | edgomez | 1.2 | %else |
47 : | edgomez | 1.4 | SECTION .rodata align=16 |
48 : | edgomez | 1.2 | %endif |
49 : | |||
50 : | edgomez | 1.3 | ALIGN 8 |
51 : | edgomez | 1.2 | mm_zero: |
52 : | edgomez | 1.3 | dd 0,0 |
53 : | ;============================================================================= | ||
54 : | ; Macros | ||
55 : | ;============================================================================= | ||
56 : | edgomez | 1.2 | |
57 : | %macro nop4 0 | ||
58 : | edgomez | 1.3 | db 08Dh, 074h, 026h, 0 |
59 : | edgomez | 1.2 | %endmacro |
60 : | |||
61 : | edgomez | 1.3 | ;============================================================================= |
62 : | ; Code | ||
63 : | ;============================================================================= | ||
64 : | |||
65 : | SECTION .text | ||
66 : | edgomez | 1.2 | |
67 : | cglobal transfer_8to16copy_3dne | ||
68 : | cglobal transfer_16to8copy_3dne | ||
69 : | cglobal transfer_8to16sub_3dne | ||
70 : | cglobal transfer_8to16subro_3dne | ||
71 : | cglobal transfer_8to16sub2_3dne | ||
72 : | cglobal transfer_16to8add_3dne | ||
73 : | cglobal transfer8x8_copy_3dne | ||
74 : | |||
75 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
76 : | edgomez | 1.2 | ; |
77 : | ; void transfer_8to16copy_3dne(int16_t * const dst, | ||
78 : | ; const uint8_t * const src, | ||
79 : | ; uint32_t stride); | ||
80 : | ; | ||
81 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
82 : | edgomez | 1.2 | |
83 : | edgomez | 1.3 | ALIGN 16 |
84 : | edgomez | 1.2 | transfer_8to16copy_3dne: |
85 : | |||
86 : | mov eax, [esp+ 8] ; Src | ||
87 : | mov edx, [esp+12] ; Stride | ||
88 : | mov ecx, [esp+ 4] ; Dst | ||
89 : | punpcklbw mm0, [byte eax] | ||
90 : | punpcklbw mm1, [eax+4] | ||
91 : | edgomez | 1.3 | movq mm2, [eax+edx] |
92 : | movq mm3, [eax+edx] | ||
93 : | pxor mm7, mm7 | ||
94 : | lea eax, [eax+2*edx] | ||
95 : | punpcklbw mm2, mm7 | ||
96 : | punpckhbw mm3, mm7 | ||
97 : | psrlw mm0, 8 | ||
98 : | psrlw mm1, 8 | ||
99 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
100 : | punpcklbw mm5, [eax+edx+4] | ||
101 : | movq [byte ecx+0*64], mm0 | ||
102 : | movq [ecx+0*64+8], mm1 | ||
103 : | punpcklbw mm6, [eax+edx] | ||
104 : | punpcklbw mm7, [eax+4] | ||
105 : | edgomez | 1.3 | lea eax, [byte eax+2*edx] |
106 : | psrlw mm4, 8 | ||
107 : | psrlw mm5, 8 | ||
108 : | edgomez | 1.2 | punpcklbw mm0, [eax] |
109 : | punpcklbw mm1, [eax+edx+4] | ||
110 : | movq [ecx+0*64+16], mm2 | ||
111 : | movq [ecx+0*64+24], mm3 | ||
112 : | edgomez | 1.3 | psrlw mm6, 8 |
113 : | psrlw mm7, 8 | ||
114 : | edgomez | 1.2 | punpcklbw mm2, [eax+edx] |
115 : | punpcklbw mm3, [eax+4] | ||
116 : | edgomez | 1.3 | lea eax, [byte eax+2*edx] |
117 : | edgomez | 1.2 | movq [byte ecx+0*64+32], mm4 |
118 : | movq [ecx+0*64+56], mm5 | ||
119 : | edgomez | 1.3 | psrlw mm0, 8 |
120 : | psrlw mm1, 8 | ||
121 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
122 : | punpcklbw mm5, [eax+edx+4] | ||
123 : | movq [byte ecx+0*64+48], mm6 | ||
124 : | movq [ecx+0*64+40], mm7 | ||
125 : | edgomez | 1.3 | psrlw mm2, 8 |
126 : | psrlw mm3, 8 | ||
127 : | edgomez | 1.2 | punpcklbw mm6, [eax+edx] |
128 : | punpcklbw mm7, [eax+4] | ||
129 : | movq [byte ecx+1*64], mm0 | ||
130 : | movq [ecx+1*64+24], mm1 | ||
131 : | edgomez | 1.3 | psrlw mm4, 8 |
132 : | psrlw mm5, 8 | ||
133 : | edgomez | 1.2 | movq [ecx+1*64+16], mm2 |
134 : | movq [ecx+1*64+8], mm3 | ||
135 : | edgomez | 1.3 | psrlw mm6, 8 |
136 : | psrlw mm7, 8 | ||
137 : | edgomez | 1.2 | movq [byte ecx+1*64+32], mm4 |
138 : | edgomez | 1.3 | movq [ecx+1*64+56], mm5 |
139 : | edgomez | 1.2 | movq [byte ecx+1*64+48], mm6 |
140 : | movq [ecx+1*64+40], mm7 | ||
141 : | edgomez | 1.3 | ret |
142 : | |||
143 : | edgomez | 1.2 | |
144 : | edgomez | 1.3 | |
145 : | ;----------------------------------------------------------------------------- | ||
146 : | edgomez | 1.2 | ; |
147 : | ; void transfer_16to8copy_3dne(uint8_t * const dst, | ||
148 : | ; const int16_t * const src, | ||
149 : | ; uint32_t stride); | ||
150 : | ; | ||
151 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
152 : | edgomez | 1.2 | |
153 : | edgomez | 1.3 | ALIGN 16 |
154 : | edgomez | 1.2 | transfer_16to8copy_3dne: |
155 : | |||
156 : | mov eax, [esp+ 8] ; Src | ||
157 : | mov ecx, [esp+ 4] ; Dst | ||
158 : | mov edx, [esp+12] ; Stride | ||
159 : | |||
160 : | movq mm0, [byte eax+0*32] | ||
161 : | edgomez | 1.3 | packuswb mm0, [eax+0*32+8] |
162 : | edgomez | 1.2 | movq mm1, [eax+0*32+16] |
163 : | edgomez | 1.3 | packuswb mm1, [eax+0*32+24] |
164 : | edgomez | 1.2 | movq mm5, [eax+2*32+16] |
165 : | movq mm2, [eax+1*32] | ||
166 : | edgomez | 1.3 | packuswb mm2, [eax+1*32+8] |
167 : | edgomez | 1.2 | movq mm3, [eax+1*32+16] |
168 : | packuswb mm3, [eax+1*32+24] | ||
169 : | edgomez | 1.3 | movq mm6, [eax+3*32] |
170 : | edgomez | 1.2 | movq mm4, [eax+2*32] |
171 : | edgomez | 1.3 | packuswb mm4, [eax+2*32+8] |
172 : | edgomez | 1.2 | packuswb mm5, [eax+2*32+24] |
173 : | movq mm7, [eax+3*32+16] | ||
174 : | packuswb mm7, [eax+3*32+24] | ||
175 : | edgomez | 1.3 | packuswb mm6, [eax+3*32+8] |
176 : | edgomez | 1.2 | movq [ecx], mm0 |
177 : | edgomez | 1.3 | lea eax, [3*edx] |
178 : | add eax, ecx | ||
179 : | edgomez | 1.2 | movq [ecx+edx], mm1 |
180 : | movq [ecx+2*edx], mm2 | ||
181 : | movq [byte eax], mm3 | ||
182 : | movq [ecx+4*edx], mm4 | ||
183 : | edgomez | 1.3 | lea ecx, [byte ecx+4*edx] |
184 : | edgomez | 1.2 | movq [eax+2*edx], mm5 |
185 : | movq [eax+4*edx], mm7 | ||
186 : | movq [ecx+2*edx], mm6 | ||
187 : | ret | ||
188 : | |||
189 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
190 : | edgomez | 1.2 | ; |
191 : | ; void transfer_8to16sub_3dne(int16_t * const dct, | ||
192 : | ; uint8_t * const cur, | ||
193 : | ; const uint8_t * const ref, | ||
194 : | ; const uint32_t stride); | ||
195 : | ; | ||
196 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
197 : | edgomez | 1.2 | |
198 : | ; when second argument == 1, reference (ebx) block is to current (eax) | ||
199 : | %macro COPY_8_TO_16_SUB 2 | ||
200 : | movq mm1, [eax] ; cur | ||
201 : | movq mm0, mm1 | ||
202 : | movq mm4, [ecx] ; ref | ||
203 : | movq mm6, mm4 | ||
204 : | %if %2 == 1 | ||
205 : | movq [eax], mm4 | ||
206 : | %endif | ||
207 : | punpckhbw mm1, mm7 | ||
208 : | punpckhbw mm6, mm7 | ||
209 : | punpcklbw mm4, mm7 | ||
210 : | edgomez | 1.3 | ALIGN 8 |
211 : | edgomez | 1.2 | movq mm2, [byte eax+edx] |
212 : | punpcklbw mm0, mm7 | ||
213 : | movq mm3, [byte eax+edx] | ||
214 : | punpcklbw mm2, mm7 | ||
215 : | movq mm5, [byte ecx+edx] ; ref | ||
216 : | punpckhbw mm3, mm7 | ||
217 : | %if %2 == 1 | ||
218 : | movq [byte eax+edx], mm5 | ||
219 : | %endif | ||
220 : | psubsw mm1, mm6 | ||
221 : | |||
222 : | movq mm6, mm5 | ||
223 : | psubsw mm0, mm4 | ||
224 : | edgomez | 1.3 | %if (%1 < 3) |
225 : | edgomez | 1.2 | lea eax,[eax+2*edx] |
226 : | lea ecx,[ecx+2*edx] | ||
227 : | %else | ||
228 : | mov ecx,[esp] | ||
229 : | add esp,byte 4 | ||
230 : | %endif | ||
231 : | movq [edi+%1*32+ 8], mm1 | ||
232 : | movq [byte edi+%1*32+ 0], mm0 ; dst | ||
233 : | punpcklbw mm5, mm7 | ||
234 : | punpckhbw mm6, mm7 | ||
235 : | psubsw mm2, mm5 | ||
236 : | psubsw mm3, mm6 | ||
237 : | movq [edi+%1*32+16], mm2 | ||
238 : | movq [edi+%1*32+24], mm3 | ||
239 : | %endmacro | ||
240 : | |||
241 : | edgomez | 1.3 | ALIGN 16 |
242 : | edgomez | 1.2 | transfer_8to16sub_3dne: |
243 : | mov eax, [esp + 8] ; Cur | ||
244 : | mov ecx, [esp +12] ; Ref | ||
245 : | push edi | ||
246 : | mov edx, [dword esp+4+16] ; Stride | ||
247 : | mov edi, [esp+4+ 4] ; Dst | ||
248 : | pxor mm7, mm7 | ||
249 : | nop | ||
250 : | edgomez | 1.3 | ALIGN 4 |
251 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 1 |
252 : | COPY_8_TO_16_SUB 1, 1 | ||
253 : | COPY_8_TO_16_SUB 2, 1 | ||
254 : | COPY_8_TO_16_SUB 3, 1 | ||
255 : | edgomez | 1.3 | mov edi, ecx |
256 : | edgomez | 1.2 | ret |
257 : | |||
258 : | edgomez | 1.3 | ALIGN 16 |
259 : | edgomez | 1.2 | transfer_8to16subro_3dne: |
260 : | mov eax, [esp + 8] ; Cur | ||
261 : | mov ecx, [esp +12] ; Ref | ||
262 : | push edi | ||
263 : | mov edx, [dword esp+4+16] ; Stride | ||
264 : | mov edi, [esp+4+ 4] ; Dst | ||
265 : | pxor mm7, mm7 | ||
266 : | nop | ||
267 : | edgomez | 1.3 | ALIGN 4 |
268 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 0 |
269 : | COPY_8_TO_16_SUB 1, 0 | ||
270 : | COPY_8_TO_16_SUB 2, 0 | ||
271 : | COPY_8_TO_16_SUB 3, 0 | ||
272 : | edgomez | 1.3 | mov edi, ecx |
273 : | edgomez | 1.2 | ret |
274 : | |||
275 : | |||
276 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
277 : | edgomez | 1.2 | ; |
278 : | ; void transfer_8to16sub2_3dne(int16_t * const dct, | ||
279 : | ; uint8_t * const cur, | ||
280 : | ; const uint8_t * ref1, | ||
281 : | ; const uint8_t * ref2, | ||
282 : | ; const uint32_t stride) | ||
283 : | ; | ||
284 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
285 : | edgomez | 1.2 | |
286 : | %macro COPY_8_TO_16_SUB2_SSE 1 | ||
287 : | edgomez | 1.3 | db 0Fh, 6Fh, 44h, 20h, 00 ;movq mm0, [byte eax] ; cur |
288 : | edgomez | 1.2 | punpcklbw mm0, mm7 |
289 : | movq mm2, [byte eax+edx] | ||
290 : | punpcklbw mm2, mm7 | ||
291 : | db 0Fh, 6Fh, 4ch, 20h, 00 ;movq mm1, [byte eax] | ||
292 : | punpckhbw mm1, mm7 | ||
293 : | movq mm3, [byte eax+edx] | ||
294 : | punpckhbw mm3, mm7 | ||
295 : | edgomez | 1.3 | |
296 : | edgomez | 1.2 | movq mm4, [byte ebx] ; ref1 |
297 : | pavgb mm4, [byte esi] ; ref2 | ||
298 : | edgomez | 1.3 | movq [eax], mm4 |
299 : | edgomez | 1.2 | movq mm5, [ebx+edx] ; ref |
300 : | pavgb mm5, [esi+edx] ; ref2 | ||
301 : | edgomez | 1.3 | movq [eax+edx], mm5 |
302 : | edgomez | 1.2 | movq mm6, mm4 |
303 : | punpcklbw mm4, mm7 | ||
304 : | punpckhbw mm6, mm7 | ||
305 : | %if (%1 < 3) | ||
306 : | lea esi,[esi+2*edx] | ||
307 : | lea ebx,[byte ebx+2*edx] | ||
308 : | lea eax,[eax+2*edx] | ||
309 : | %else | ||
310 : | mov esi,[esp] | ||
311 : | mov ebx,[esp+4] | ||
312 : | add esp,byte 8 | ||
313 : | edgomez | 1.3 | %endif |
314 : | edgomez | 1.2 | psubsw mm0, mm4 |
315 : | psubsw mm1, mm6 | ||
316 : | movq mm6, mm5 | ||
317 : | punpcklbw mm5, mm7 | ||
318 : | punpckhbw mm6, mm7 | ||
319 : | psubsw mm2, mm5 | ||
320 : | psubsw mm3, mm6 | ||
321 : | movq [byte ecx+%1*32+ 0], mm0 ; dst | ||
322 : | movq [ecx+%1*32+ 8], mm1 | ||
323 : | movq [ecx+%1*32+16], mm2 | ||
324 : | movq [ecx+%1*32+24], mm3 | ||
325 : | %endmacro | ||
326 : | |||
327 : | edgomez | 1.3 | ALIGN 16 |
328 : | edgomez | 1.2 | transfer_8to16sub2_3dne: |
329 : | mov edx, [esp +20] ; Stride | ||
330 : | mov ecx, [esp + 4] ; Dst | ||
331 : | mov eax, [esp + 8] ; Cur | ||
332 : | push ebx | ||
333 : | lea ebp,[byte ebp] | ||
334 : | mov ebx, [esp+4+12] ; Ref1 | ||
335 : | push esi | ||
336 : | pxor mm7, mm7 | ||
337 : | mov esi, [esp+8+16] ; Ref2 | ||
338 : | edgomez | 1.3 | nop4 |
339 : | edgomez | 1.2 | COPY_8_TO_16_SUB2_SSE 0 |
340 : | COPY_8_TO_16_SUB2_SSE 1 | ||
341 : | COPY_8_TO_16_SUB2_SSE 2 | ||
342 : | COPY_8_TO_16_SUB2_SSE 3 | ||
343 : | |||
344 : | ret | ||
345 : | |||
346 : | |||
347 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
348 : | edgomez | 1.2 | ; |
349 : | ; void transfer_16to8add_3dne(uint8_t * const dst, | ||
350 : | ; const int16_t * const src, | ||
351 : | ; uint32_t stride); | ||
352 : | ; | ||
353 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
354 : | edgomez | 1.2 | |
355 : | %macro COPY_16_TO_8_ADD 1 | ||
356 : | db 0Fh, 6Fh, 44h, 21h, 00 ;movq mm0, [byte ecx] | ||
357 : | punpcklbw mm0, mm7 | ||
358 : | movq mm2, [byte ecx+edx] | ||
359 : | punpcklbw mm2, mm7 | ||
360 : | db 0Fh, 6Fh, 4ch, 21h, 00 ;movq mm1, [byte ecx] | ||
361 : | punpckhbw mm1, mm7 | ||
362 : | movq mm3, [byte ecx+edx] | ||
363 : | punpckhbw mm3, mm7 | ||
364 : | paddsw mm0, [byte eax+%1*32+ 0] | ||
365 : | paddsw mm1, [eax+%1*32+ 8] | ||
366 : | paddsw mm2, [eax+%1*32+16] | ||
367 : | paddsw mm3, [eax+%1*32+24] | ||
368 : | packuswb mm0, mm1 | ||
369 : | packuswb mm2, mm3 | ||
370 : | edgomez | 1.3 | mov esp,esp |
371 : | edgomez | 1.2 | movq [byte ecx], mm0 |
372 : | movq [ecx+edx], mm2 | ||
373 : | %endmacro | ||
374 : | |||
375 : | |||
376 : | edgomez | 1.3 | ALIGN 16 |
377 : | edgomez | 1.2 | transfer_16to8add_3dne: |
378 : | mov ecx, [esp+ 4] ; Dst | ||
379 : | mov edx, [esp+12] ; Stride | ||
380 : | mov eax, [esp+ 8] ; Src | ||
381 : | pxor mm7, mm7 | ||
382 : | nop | ||
383 : | |||
384 : | COPY_16_TO_8_ADD 0 | ||
385 : | lea ecx,[byte ecx+2*edx] | ||
386 : | COPY_16_TO_8_ADD 1 | ||
387 : | lea ecx,[byte ecx+2*edx] | ||
388 : | COPY_16_TO_8_ADD 2 | ||
389 : | lea ecx,[byte ecx+2*edx] | ||
390 : | COPY_16_TO_8_ADD 3 | ||
391 : | ret | ||
392 : | |||
393 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
394 : | edgomez | 1.2 | ; |
395 : | ; void transfer8x8_copy_3dne(uint8_t * const dst, | ||
396 : | ; const uint8_t * const src, | ||
397 : | ; const uint32_t stride); | ||
398 : | ; | ||
399 : | ; | ||
400 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
401 : | edgomez | 1.2 | |
402 : | %macro COPY_8_TO_8 0 | ||
403 : | movq mm0, [byte eax] | ||
404 : | movq mm1, [eax+edx] | ||
405 : | movq [byte ecx], mm0 | ||
406 : | lea eax,[byte eax+2*edx] | ||
407 : | movq [ecx+edx], mm1 | ||
408 : | %endmacro | ||
409 : | |||
410 : | edgomez | 1.3 | ALIGN 16 |
411 : | edgomez | 1.2 | transfer8x8_copy_3dne: |
412 : | mov eax, [esp+ 8] ; Src | ||
413 : | mov edx, [esp+12] ; Stride | ||
414 : | mov ecx, [esp+ 4] ; Dst | ||
415 : | |||
416 : | COPY_8_TO_8 | ||
417 : | lea ecx,[byte ecx+2*edx] | ||
418 : | COPY_8_TO_8 | ||
419 : | lea ecx,[byte ecx+2*edx] | ||
420 : | COPY_8_TO_8 | ||
421 : | lea ecx,[byte ecx+2*edx] | ||
422 : | COPY_8_TO_8 | ||
423 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |