Parent Directory | Revision Log
Revision 1.8 - (view) (download)
1 : | edgomez | 1.2 | ;/***************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - Quarter-pixel interpolation - | ||
5 : | ; * Copyright(C) 2002 Pascal Massimino <skal@planet-d.net> | ||
6 : | ; * | ||
7 : | ; * This file is part of XviD, a free MPEG-4 video encoder/decoder | ||
8 : | ; * | ||
9 : | Isibaar | 1.8 | ; * XviD is free software; you can rDST_PTRstribute it and/or modify it |
10 : | edgomez | 1.2 | ; * 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 : | ; * | ||
14 : | ; * 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 : | ; * | ||
19 : | ; * 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 : | ; * | ||
23 : | Isibaar | 1.8 | ; * $Id: qpel_mmx.asm,v 1.7 2008/11/11 20:46:24 Isibaar Exp $ |
24 : | edgomez | 1.2 | ; * |
25 : | ; *************************************************************************/ | ||
26 : | |||
27 : | ;/************************************************************************** | ||
28 : | ; * | ||
29 : | ; * History: | ||
30 : | ; * | ||
31 : | ; * 22.10.2002 initial coding. unoptimized 'proof of concept', | ||
32 : | ; * just to heft the qpel filtering. - Skal - | ||
33 : | ; * | ||
34 : | ; *************************************************************************/ | ||
35 : | |||
36 : | |||
37 : | %define USE_TABLES ; in order to use xvid_FIR_x_x_x_x tables | ||
38 : | ; instead of xvid_Expand_mmx... | ||
39 : | |||
40 : | |||
41 : | Isibaar | 1.8 | %include "nasm.inc" |
42 : | edgomez | 1.2 | |
43 : | ;////////////////////////////////////////////////////////////////////// | ||
44 : | ;// Declarations | ||
45 : | ;// all signatures are: | ||
46 : | ;// void XXX(uint8_t *dst, const uint8_t *src, | ||
47 : | ;// int32_t length, int32_t stride, int32_t rounding) | ||
48 : | ;////////////////////////////////////////////////////////////////////// | ||
49 : | |||
50 : | cglobal xvid_H_Pass_16_mmx | ||
51 : | cglobal xvid_H_Pass_Avrg_16_mmx | ||
52 : | cglobal xvid_H_Pass_Avrg_Up_16_mmx | ||
53 : | cglobal xvid_V_Pass_16_mmx | ||
54 : | cglobal xvid_V_Pass_Avrg_16_mmx | ||
55 : | cglobal xvid_V_Pass_Avrg_Up_16_mmx | ||
56 : | cglobal xvid_H_Pass_8_mmx | ||
57 : | cglobal xvid_H_Pass_Avrg_8_mmx | ||
58 : | cglobal xvid_H_Pass_Avrg_Up_8_mmx | ||
59 : | cglobal xvid_V_Pass_8_mmx | ||
60 : | cglobal xvid_V_Pass_Avrg_8_mmx | ||
61 : | cglobal xvid_V_Pass_Avrg_Up_8_mmx | ||
62 : | |||
63 : | cglobal xvid_H_Pass_Add_16_mmx | ||
64 : | cglobal xvid_H_Pass_Avrg_Add_16_mmx | ||
65 : | cglobal xvid_H_Pass_Avrg_Up_Add_16_mmx | ||
66 : | cglobal xvid_V_Pass_Add_16_mmx | ||
67 : | cglobal xvid_V_Pass_Avrg_Add_16_mmx | ||
68 : | cglobal xvid_V_Pass_Avrg_Up_Add_16_mmx | ||
69 : | cglobal xvid_H_Pass_8_Add_mmx | ||
70 : | cglobal xvid_H_Pass_Avrg_8_Add_mmx | ||
71 : | cglobal xvid_H_Pass_Avrg_Up_8_Add_mmx | ||
72 : | cglobal xvid_V_Pass_8_Add_mmx | ||
73 : | cglobal xvid_V_Pass_Avrg_8_Add_mmx | ||
74 : | cglobal xvid_V_Pass_Avrg_Up_8_Add_mmx | ||
75 : | |||
76 : | Isibaar | 1.8 | cglobal xvid_Expand_mmx |
77 : | |||
78 : | cglobal xvid_FIR_1_0_0_0 | ||
79 : | cglobal xvid_FIR_3_1_0_0 | ||
80 : | cglobal xvid_FIR_6_3_1_0 | ||
81 : | cglobal xvid_FIR_14_3_2_1 | ||
82 : | cglobal xvid_FIR_20_6_3_1 | ||
83 : | cglobal xvid_FIR_20_20_6_3 | ||
84 : | cglobal xvid_FIR_23_19_6_3 | ||
85 : | cglobal xvid_FIR_7_20_20_6 | ||
86 : | cglobal xvid_FIR_6_20_20_6 | ||
87 : | cglobal xvid_FIR_6_20_20_7 | ||
88 : | cglobal xvid_FIR_3_6_20_20 | ||
89 : | cglobal xvid_FIR_3_6_19_23 | ||
90 : | cglobal xvid_FIR_1_3_6_20 | ||
91 : | cglobal xvid_FIR_1_2_3_14 | ||
92 : | cglobal xvid_FIR_0_1_3_6 | ||
93 : | cglobal xvid_FIR_0_0_1_3 | ||
94 : | cglobal xvid_FIR_0_0_0_1 | ||
95 : | |||
96 : | SECTION .data align=SECTION_ALIGN | ||
97 : | |||
98 : | align SECTION_ALIGN | ||
99 : | xvid_Expand_mmx: | ||
100 : | times 256*4 dw 0 ; uint16_t xvid_Expand_mmx[256][4] | ||
101 : | ENDFUNC | ||
102 : | |||
103 : | xvid_FIR_1_0_0_0: | ||
104 : | times 256*4 dw 0 | ||
105 : | ENDFUNC | ||
106 : | |||
107 : | xvid_FIR_3_1_0_0: | ||
108 : | times 256*4 dw 0 | ||
109 : | ENDFUNC | ||
110 : | |||
111 : | xvid_FIR_6_3_1_0: | ||
112 : | times 256*4 dw 0 | ||
113 : | ENDFUNC | ||
114 : | |||
115 : | xvid_FIR_14_3_2_1: | ||
116 : | times 256*4 dw 0 | ||
117 : | ENDFUNC | ||
118 : | edgomez | 1.2 | |
119 : | Isibaar | 1.8 | xvid_FIR_20_6_3_1: |
120 : | times 256*4 dw 0 | ||
121 : | ENDFUNC | ||
122 : | edgomez | 1.2 | |
123 : | Isibaar | 1.8 | xvid_FIR_20_20_6_3: |
124 : | times 256*4 dw 0 | ||
125 : | ENDFUNC | ||
126 : | edgomez | 1.2 | |
127 : | Isibaar | 1.8 | xvid_FIR_23_19_6_3: |
128 : | times 256*4 dw 0 | ||
129 : | ENDFUNC | ||
130 : | |||
131 : | xvid_FIR_7_20_20_6: | ||
132 : | times 256*4 dw 0 | ||
133 : | ENDFUNC | ||
134 : | |||
135 : | xvid_FIR_6_20_20_6: | ||
136 : | times 256*4 dw 0 | ||
137 : | ENDFUNC | ||
138 : | |||
139 : | xvid_FIR_6_20_20_7: | ||
140 : | times 256*4 dw 0 | ||
141 : | ENDFUNC | ||
142 : | |||
143 : | xvid_FIR_3_6_20_20: | ||
144 : | times 256*4 dw 0 | ||
145 : | ENDFUNC | ||
146 : | |||
147 : | xvid_FIR_3_6_19_23: | ||
148 : | times 256*4 dw 0 | ||
149 : | ENDFUNC | ||
150 : | |||
151 : | xvid_FIR_1_3_6_20: | ||
152 : | times 256*4 dw 0 | ||
153 : | ENDFUNC | ||
154 : | |||
155 : | xvid_FIR_1_2_3_14: | ||
156 : | times 256*4 dw 0 | ||
157 : | ENDFUNC | ||
158 : | |||
159 : | xvid_FIR_0_1_3_6: | ||
160 : | times 256*4 dw 0 | ||
161 : | ENDFUNC | ||
162 : | |||
163 : | xvid_FIR_0_0_1_3: | ||
164 : | times 256*4 dw 0 | ||
165 : | ENDFUNC | ||
166 : | |||
167 : | xvid_FIR_0_0_0_1: | ||
168 : | times 256*4 dw 0 | ||
169 : | ENDFUNC | ||
170 : | edgomez | 1.2 | |
171 : | ;////////////////////////////////////////////////////////////////////// | ||
172 : | |||
173 : | Isibaar | 1.8 | DATA |
174 : | edgomez | 1.2 | |
175 : | Isibaar | 1.8 | align SECTION_ALIGN |
176 : | edgomez | 1.2 | Rounder1_MMX: |
177 : | times 4 dw 1 | ||
178 : | Rounder0_MMX: | ||
179 : | times 4 dw 0 | ||
180 : | |||
181 : | Isibaar | 1.8 | align SECTION_ALIGN |
182 : | Isibaar | 1.7 | Rounder_QP_MMX: |
183 : | edgomez | 1.2 | times 4 dw 16 |
184 : | times 4 dw 15 | ||
185 : | |||
186 : | %ifndef USE_TABLES | ||
187 : | |||
188 : | Isibaar | 1.8 | align SECTION_ALIGN |
189 : | edgomez | 1.2 | |
190 : | ; H-Pass table shared by 16x? and 8x? filters | ||
191 : | |||
192 : | FIR_R0: dw 14, -3, 2, -1 | ||
193 : | Isibaar | 1.8 | align SECTION_ALIGN |
194 : | edgomez | 1.2 | FIR_R1: dw 23, 19, -6, 3, -1, 0, 0, 0 |
195 : | |||
196 : | FIR_R2: dw -7, 20, 20, -6, 3, -1, 0, 0 | ||
197 : | |||
198 : | FIR_R3: dw 3, -6, 20, 20, -6, 3, -1, 0 | ||
199 : | |||
200 : | FIR_R4: dw -1, 3, -6, 20, 20, -6, 3, -1 | ||
201 : | |||
202 : | FIR_R5: dw 0, -1, 3, -6, 20, 20, -6, 3, -1, 0, 0, 0 | ||
203 : | Isibaar | 1.8 | align SECTION_ALIGN |
204 : | edgomez | 1.2 | FIR_R6: dw 0, 0, -1, 3, -6, 20, 20, -6, 3, -1, 0, 0 |
205 : | Isibaar | 1.8 | align SECTION_ALIGN |
206 : | edgomez | 1.2 | FIR_R7: dw 0, 0, 0, -1, 3, -6, 20, 20, -6, 3, -1, 0 |
207 : | Isibaar | 1.8 | align SECTION_ALIGN |
208 : | edgomez | 1.2 | FIR_R8: dw -1, 3, -6, 20, 20, -6, 3, -1 |
209 : | |||
210 : | FIR_R9: dw 0, -1, 3, -6, 20, 20, -6, 3, -1, 0, 0, 0 | ||
211 : | Isibaar | 1.8 | align SECTION_ALIGN |
212 : | edgomez | 1.2 | FIR_R10: dw 0, 0, -1, 3, -6, 20, 20, -6, 3, -1, 0, 0 |
213 : | Isibaar | 1.8 | align SECTION_ALIGN |
214 : | edgomez | 1.2 | FIR_R11: dw 0, 0, 0, -1, 3, -6, 20, 20, -6, 3, -1, 0 |
215 : | Isibaar | 1.8 | align SECTION_ALIGN |
216 : | edgomez | 1.2 | FIR_R12: dw -1, 3, -6, 20, 20, -6, 3, -1 |
217 : | |||
218 : | FIR_R13: dw 0, -1, 3, -6, 20, 20, -6, 3 | ||
219 : | |||
220 : | FIR_R14: dw 0, 0, -1, 3, -6, 20, 20, -7 | ||
221 : | |||
222 : | FIR_R15: dw 0, 0, 0, -1, 3, -6, 19, 23 | ||
223 : | |||
224 : | FIR_R16: dw -1, 2, -3, 14 | ||
225 : | |||
226 : | %endif ; !USE_TABLES | ||
227 : | |||
228 : | ; V-Pass taps | ||
229 : | |||
230 : | Isibaar | 1.8 | align SECTION_ALIGN |
231 : | edgomez | 1.2 | FIR_Cm7: times 4 dw -7 |
232 : | FIR_Cm6: times 4 dw -6 | ||
233 : | FIR_Cm3: times 4 dw -3 | ||
234 : | FIR_Cm1: times 4 dw -1 | ||
235 : | FIR_C2: times 4 dw 2 | ||
236 : | FIR_C3: times 4 dw 3 | ||
237 : | FIR_C14: times 4 dw 14 | ||
238 : | FIR_C19: times 4 dw 19 | ||
239 : | FIR_C20: times 4 dw 20 | ||
240 : | FIR_C23: times 4 dw 23 | ||
241 : | |||
242 : | Isibaar | 1.8 | SECTION .rotext align=SECTION_ALIGN |
243 : | edgomez | 1.2 | |
244 : | ;////////////////////////////////////////////////////////////////////// | ||
245 : | ;// Here we go with the Q-Pel mess. | ||
246 : | ;// For horizontal passes, we process 4 *output* pixel in parallel | ||
247 : | ;// For vertical ones, we process 4 *input* pixel in parallel. | ||
248 : | ;////////////////////////////////////////////////////////////////////// | ||
249 : | |||
250 : | Isibaar | 1.8 | %ifdef ARCH_IS_X86_64 |
251 : | %macro XVID_MOVQ 3 | ||
252 : | lea r9, [%2] | ||
253 : | movq %1, [r9 + %3] | ||
254 : | %endmacro | ||
255 : | %macro XVID_PADDW 3 | ||
256 : | lea r9, [%2] | ||
257 : | paddw %1, [r9 + %3] | ||
258 : | %endmacro | ||
259 : | %ifdef WINDOWS | ||
260 : | %define SRC_PTR _EDX | ||
261 : | %define DST_PTR _ECX | ||
262 : | %else | ||
263 : | %define SRC_PTR _ESI | ||
264 : | %define DST_PTR _EDI | ||
265 : | %endif | ||
266 : | %else | ||
267 : | %macro XVID_MOVQ 3 | ||
268 : | movq %1, [%2 + %3] | ||
269 : | %endmacro | ||
270 : | %macro XVID_PADDW 3 | ||
271 : | paddw %1, [%2 + %3] | ||
272 : | %endmacro | ||
273 : | %define SRC_PTR _ESI | ||
274 : | %define DST_PTR _EDI | ||
275 : | %endif | ||
276 : | |||
277 : | edgomez | 1.2 | %macro PROLOG_NO_AVRG 0 |
278 : | Isibaar | 1.8 | mov TMP0, prm3 ; Size |
279 : | mov TMP1, prm4 ; BpS | ||
280 : | mov eax, prm5d ; Rnd | ||
281 : | |||
282 : | %ifndef ARCH_IS_X86_64 | ||
283 : | push SRC_PTR | ||
284 : | push DST_PTR | ||
285 : | %endif | ||
286 : | push _EBP | ||
287 : | mov _EBP, TMP1 | ||
288 : | |||
289 : | %ifndef ARCH_IS_X86_64 | ||
290 : | mov DST_PTR, [_ESP+16 + 0*4] ; Dst | ||
291 : | mov SRC_PTR, [_ESP+16 + 1*4] ; Src | ||
292 : | %endif | ||
293 : | |||
294 : | and _EAX, 1 | ||
295 : | lea TMP1, [Rounder_QP_MMX] | ||
296 : | movq mm7, [TMP1+_EAX*8] ; rounder | ||
297 : | edgomez | 1.2 | %endmacro |
298 : | |||
299 : | %macro EPILOG_NO_AVRG 0 | ||
300 : | Isibaar | 1.8 | pop _EBP |
301 : | %ifndef ARCH_IS_X86_64 | ||
302 : | pop DST_PTR | ||
303 : | pop SRC_PTR | ||
304 : | %endif | ||
305 : | edgomez | 1.2 | ret |
306 : | %endmacro | ||
307 : | |||
308 : | %macro PROLOG_AVRG 0 | ||
309 : | Isibaar | 1.8 | mov TMP0, prm3 ; Size |
310 : | mov TMP1, prm4 ; BpS | ||
311 : | mov eax, prm5d ; Rnd | ||
312 : | |||
313 : | push _EBX | ||
314 : | push _EBP | ||
315 : | %ifndef ARCH_IS_X86_64 | ||
316 : | push SRC_PTR | ||
317 : | push DST_PTR | ||
318 : | %endif | ||
319 : | mov _EBP, TMP1 | ||
320 : | |||
321 : | %ifndef ARCH_IS_X86_64 | ||
322 : | mov DST_PTR, [_ESP+20 + 0*4] ; Dst | ||
323 : | mov SRC_PTR, [_ESP+20 + 1*4] ; Src | ||
324 : | %endif | ||
325 : | |||
326 : | and _EAX, 1 | ||
327 : | lea TMP1, [Rounder_QP_MMX] | ||
328 : | movq mm7, [TMP1+_EAX*8] ; rounder | ||
329 : | lea TMP1, [Rounder1_MMX] | ||
330 : | lea _EBX, [TMP1+_EAX*8] ; *Rounder2 | ||
331 : | edgomez | 1.2 | %endmacro |
332 : | |||
333 : | %macro EPILOG_AVRG 0 | ||
334 : | Isibaar | 1.8 | %ifndef ARCH_IS_X86_64 |
335 : | pop DST_PTR | ||
336 : | pop SRC_PTR | ||
337 : | %endif | ||
338 : | pop _EBP | ||
339 : | pop _EBX | ||
340 : | edgomez | 1.2 | ret |
341 : | %endmacro | ||
342 : | |||
343 : | ;////////////////////////////////////////////////////////////////////// | ||
344 : | ;// | ||
345 : | ;// All horizontal passes | ||
346 : | ;// | ||
347 : | ;////////////////////////////////////////////////////////////////////// | ||
348 : | |||
349 : | ; macros for USE_TABLES | ||
350 : | |||
351 : | %macro TLOAD 2 ; %1,%2: src pixels | ||
352 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
353 : | movzx TMP1, byte [SRC_PTR+%2] | ||
354 : | XVID_MOVQ mm0, xvid_FIR_14_3_2_1, _EAX*8 | ||
355 : | XVID_MOVQ mm3, xvid_FIR_1_2_3_14, TMP1*8 | ||
356 : | edgomez | 1.2 | paddw mm0, mm7 |
357 : | paddw mm3, mm7 | ||
358 : | %endmacro | ||
359 : | |||
360 : | %macro TACCUM2 5 ;%1:src pixel/%2-%3:Taps tables/ %4-%5:dst regs | ||
361 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
362 : | XVID_PADDW %4, %2, _EAX*8 | ||
363 : | XVID_PADDW %5, %3, _EAX*8 | ||
364 : | edgomez | 1.2 | %endmacro |
365 : | |||
366 : | %macro TACCUM3 7 ;%1:src pixel/%2-%4:Taps tables/%5-%7:dst regs | ||
367 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
368 : | XVID_PADDW %5, %2, _EAX*8 | ||
369 : | XVID_PADDW %6, %3, _EAX*8 | ||
370 : | XVID_PADDW %7, %4, _EAX*8 | ||
371 : | edgomez | 1.2 | %endmacro |
372 : | |||
373 : | ;////////////////////////////////////////////////////////////////////// | ||
374 : | |||
375 : | ; macros without USE_TABLES | ||
376 : | |||
377 : | %macro LOAD 2 ; %1,%2: src pixels | ||
378 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
379 : | movzx TMP1, byte [SRC_PTR+%2] | ||
380 : | XVID_MOVQ mm0, xvid_Expand_mmx, _EAX*8 | ||
381 : | XVID_MOVQ mm3, xvid_Expand_mmx, TMP1*8 | ||
382 : | edgomez | 1.2 | pmullw mm0, [FIR_R0 ] |
383 : | pmullw mm3, [FIR_R16] | ||
384 : | paddw mm0, mm7 | ||
385 : | paddw mm3, mm7 | ||
386 : | %endmacro | ||
387 : | |||
388 : | %macro ACCUM2 4 ;src pixel/Taps/dst regs #1-#2 | ||
389 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
390 : | XVID_MOVQ mm4, xvid_Expand_mmx, _EAX*8 | ||
391 : | edgomez | 1.2 | movq mm5, mm4 |
392 : | pmullw mm4, [%2] | ||
393 : | pmullw mm5, [%2+8] | ||
394 : | paddw %3, mm4 | ||
395 : | paddw %4, mm5 | ||
396 : | %endmacro | ||
397 : | |||
398 : | %macro ACCUM3 5 ;src pixel/Taps/dst regs #1-#2-#3 | ||
399 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR+%1] |
400 : | XVID_MOVQ mm4, xvid_Expand_mmx, _EAX*8 | ||
401 : | edgomez | 1.2 | movq mm5, mm4 |
402 : | movq mm6, mm5 | ||
403 : | pmullw mm4, [%2 ] | ||
404 : | pmullw mm5, [%2+ 8] | ||
405 : | pmullw mm6, [%2+16] | ||
406 : | paddw %3, mm4 | ||
407 : | paddw %4, mm5 | ||
408 : | paddw %5, mm6 | ||
409 : | %endmacro | ||
410 : | |||
411 : | ;////////////////////////////////////////////////////////////////////// | ||
412 : | |||
413 : | %macro MIX 3 ; %1:reg, %2:src, %3:rounder | ||
414 : | pxor mm6, mm6 | ||
415 : | movq mm4, [%2] | ||
416 : | movq mm1, %1 | ||
417 : | movq mm5, mm4 | ||
418 : | punpcklbw %1, mm6 | ||
419 : | punpcklbw mm4, mm6 | ||
420 : | punpckhbw mm1, mm6 | ||
421 : | punpckhbw mm5, mm6 | ||
422 : | movq mm6, [%3] ; rounder #2 | ||
423 : | paddusw %1, mm4 | ||
424 : | paddusw mm1, mm5 | ||
425 : | paddusw %1, mm6 | ||
426 : | paddusw mm1, mm6 | ||
427 : | psrlw %1, 1 | ||
428 : | psrlw mm1, 1 | ||
429 : | packuswb %1, mm1 | ||
430 : | %endmacro | ||
431 : | |||
432 : | ;////////////////////////////////////////////////////////////////////// | ||
433 : | |||
434 : | %macro H_PASS_16 2 ; %1:src-op (0=NONE,1=AVRG,2=AVRG-UP), %2:dst-op (NONE/AVRG) | ||
435 : | |||
436 : | %if (%2==0) && (%1==0) | ||
437 : | PROLOG_NO_AVRG | ||
438 : | %else | ||
439 : | PROLOG_AVRG | ||
440 : | %endif | ||
441 : | |||
442 : | Isibaar | 1.7 | .Loop: |
443 : | edgomez | 1.2 | |
444 : | ; mm0..mm3 serves as a 4x4 delay line | ||
445 : | |||
446 : | %ifndef USE_TABLES | ||
447 : | |||
448 : | LOAD 0, 16 ; special case for 1rst/last pixel | ||
449 : | movq mm1, mm7 | ||
450 : | movq mm2, mm7 | ||
451 : | |||
452 : | ACCUM2 1, FIR_R1, mm0, mm1 | ||
453 : | ACCUM2 2, FIR_R2, mm0, mm1 | ||
454 : | ACCUM2 3, FIR_R3, mm0, mm1 | ||
455 : | ACCUM2 4, FIR_R4, mm0, mm1 | ||
456 : | |||
457 : | ACCUM3 5, FIR_R5, mm0, mm1, mm2 | ||
458 : | ACCUM3 6, FIR_R6, mm0, mm1, mm2 | ||
459 : | ACCUM3 7, FIR_R7, mm0, mm1, mm2 | ||
460 : | ACCUM2 8, FIR_R8, mm1, mm2 | ||
461 : | ACCUM3 9, FIR_R9, mm1, mm2, mm3 | ||
462 : | ACCUM3 10, FIR_R10,mm1, mm2, mm3 | ||
463 : | ACCUM3 11, FIR_R11,mm1, mm2, mm3 | ||
464 : | |||
465 : | ACCUM2 12, FIR_R12, mm2, mm3 | ||
466 : | ACCUM2 13, FIR_R13, mm2, mm3 | ||
467 : | ACCUM2 14, FIR_R14, mm2, mm3 | ||
468 : | ACCUM2 15, FIR_R15, mm2, mm3 | ||
469 : | |||
470 : | %else | ||
471 : | |||
472 : | TLOAD 0, 16 ; special case for 1rst/last pixel | ||
473 : | movq mm1, mm7 | ||
474 : | movq mm2, mm7 | ||
475 : | |||
476 : | TACCUM2 1, xvid_FIR_23_19_6_3, xvid_FIR_1_0_0_0 , mm0, mm1 | ||
477 : | TACCUM2 2, xvid_FIR_7_20_20_6, xvid_FIR_3_1_0_0 , mm0, mm1 | ||
478 : | TACCUM2 3, xvid_FIR_3_6_20_20, xvid_FIR_6_3_1_0 , mm0, mm1 | ||
479 : | TACCUM2 4, xvid_FIR_1_3_6_20 , xvid_FIR_20_6_3_1, mm0, mm1 | ||
480 : | |||
481 : | TACCUM3 5, xvid_FIR_0_1_3_6 , xvid_FIR_20_20_6_3, xvid_FIR_1_0_0_0 , mm0, mm1, mm2 | ||
482 : | TACCUM3 6, xvid_FIR_0_0_1_3 , xvid_FIR_6_20_20_6, xvid_FIR_3_1_0_0 , mm0, mm1, mm2 | ||
483 : | TACCUM3 7, xvid_FIR_0_0_0_1 , xvid_FIR_3_6_20_20, xvid_FIR_6_3_1_0 , mm0, mm1, mm2 | ||
484 : | |||
485 : | TACCUM2 8, xvid_FIR_1_3_6_20 , xvid_FIR_20_6_3_1 , mm1, mm2 | ||
486 : | |||
487 : | TACCUM3 9, xvid_FIR_0_1_3_6 , xvid_FIR_20_20_6_3, xvid_FIR_1_0_0_0, mm1, mm2, mm3 | ||
488 : | TACCUM3 10, xvid_FIR_0_0_1_3 , xvid_FIR_6_20_20_6, xvid_FIR_3_1_0_0, mm1, mm2, mm3 | ||
489 : | TACCUM3 11, xvid_FIR_0_0_0_1 , xvid_FIR_3_6_20_20, xvid_FIR_6_3_1_0, mm1, mm2, mm3 | ||
490 : | |||
491 : | TACCUM2 12, xvid_FIR_1_3_6_20, xvid_FIR_20_6_3_1 , mm2, mm3 | ||
492 : | TACCUM2 13, xvid_FIR_0_1_3_6 , xvid_FIR_20_20_6_3, mm2, mm3 | ||
493 : | TACCUM2 14, xvid_FIR_0_0_1_3 , xvid_FIR_6_20_20_7, mm2, mm3 | ||
494 : | TACCUM2 15, xvid_FIR_0_0_0_1 , xvid_FIR_3_6_19_23, mm2, mm3 | ||
495 : | |||
496 : | %endif | ||
497 : | |||
498 : | psraw mm0, 5 | ||
499 : | psraw mm1, 5 | ||
500 : | psraw mm2, 5 | ||
501 : | psraw mm3, 5 | ||
502 : | packuswb mm0, mm1 | ||
503 : | packuswb mm2, mm3 | ||
504 : | |||
505 : | %if (%1==1) | ||
506 : | Isibaar | 1.8 | MIX mm0, SRC_PTR, _EBX |
507 : | edgomez | 1.2 | %elif (%1==2) |
508 : | Isibaar | 1.8 | MIX mm0, SRC_PTR+1, _EBX |
509 : | edgomez | 1.2 | %endif |
510 : | %if (%2==1) | ||
511 : | Isibaar | 1.8 | MIX mm0, DST_PTR, Rounder1_MMX |
512 : | edgomez | 1.2 | %endif |
513 : | |||
514 : | %if (%1==1) | ||
515 : | Isibaar | 1.8 | MIX mm2, SRC_PTR+8, _EBX |
516 : | edgomez | 1.2 | %elif (%1==2) |
517 : | Isibaar | 1.8 | MIX mm2, SRC_PTR+9, _EBX |
518 : | edgomez | 1.2 | %endif |
519 : | %if (%2==1) | ||
520 : | Isibaar | 1.8 | MIX mm2, DST_PTR+8, Rounder1_MMX |
521 : | edgomez | 1.2 | %endif |
522 : | |||
523 : | Isibaar | 1.8 | lea SRC_PTR, [SRC_PTR+_EBP] |
524 : | edgomez | 1.2 | |
525 : | Isibaar | 1.8 | movq [DST_PTR+0], mm0 |
526 : | movq [DST_PTR+8], mm2 | ||
527 : | edgomez | 1.2 | |
528 : | Isibaar | 1.8 | add DST_PTR, _EBP |
529 : | dec TMP0 | ||
530 : | edgomez | 1.2 | jg .Loop |
531 : | |||
532 : | %if (%2==0) && (%1==0) | ||
533 : | EPILOG_NO_AVRG | ||
534 : | %else | ||
535 : | EPILOG_AVRG | ||
536 : | %endif | ||
537 : | |||
538 : | %endmacro | ||
539 : | |||
540 : | |||
541 : | ;////////////////////////////////////////////////////////////////////// | ||
542 : | |||
543 : | %macro H_PASS_8 2 ; %1:src-op (0=NONE,1=AVRG,2=AVRG-UP), %2:dst-op (NONE/AVRG) | ||
544 : | |||
545 : | %if (%2==0) && (%1==0) | ||
546 : | PROLOG_NO_AVRG | ||
547 : | %else | ||
548 : | PROLOG_AVRG | ||
549 : | %endif | ||
550 : | |||
551 : | Isibaar | 1.7 | .Loop: |
552 : | edgomez | 1.2 | ; mm0..mm3 serves as a 4x4 delay line |
553 : | |||
554 : | %ifndef USE_TABLES | ||
555 : | |||
556 : | LOAD 0, 8 ; special case for 1rst/last pixel | ||
557 : | ACCUM2 1, FIR_R1, mm0, mm3 | ||
558 : | ACCUM2 2, FIR_R2, mm0, mm3 | ||
559 : | ACCUM2 3, FIR_R3, mm0, mm3 | ||
560 : | ACCUM2 4, FIR_R4, mm0, mm3 | ||
561 : | |||
562 : | ACCUM2 5, FIR_R13, mm0, mm3 | ||
563 : | ACCUM2 6, FIR_R14, mm0, mm3 | ||
564 : | ACCUM2 7, FIR_R15, mm0, mm3 | ||
565 : | |||
566 : | %else | ||
567 : | |||
568 : | %if 0 ; test with no unrolling | ||
569 : | |||
570 : | TLOAD 0, 8 ; special case for 1rst/last pixel | ||
571 : | TACCUM2 1, xvid_FIR_23_19_6_3, xvid_FIR_1_0_0_0 , mm0, mm3 | ||
572 : | TACCUM2 2, xvid_FIR_7_20_20_6, xvid_FIR_3_1_0_0 , mm0, mm3 | ||
573 : | TACCUM2 3, xvid_FIR_3_6_20_20, xvid_FIR_6_3_1_0 , mm0, mm3 | ||
574 : | TACCUM2 4, xvid_FIR_1_3_6_20 , xvid_FIR_20_6_3_1 , mm0, mm3 | ||
575 : | TACCUM2 5, xvid_FIR_0_1_3_6 , xvid_FIR_20_20_6_3, mm0, mm3 | ||
576 : | TACCUM2 6, xvid_FIR_0_0_1_3 , xvid_FIR_6_20_20_7, mm0, mm3 | ||
577 : | TACCUM2 7, xvid_FIR_0_0_0_1 , xvid_FIR_3_6_19_23, mm0, mm3 | ||
578 : | |||
579 : | %else ; test with unrolling (little faster, but not much) | ||
580 : | |||
581 : | Isibaar | 1.8 | movzx _EAX, byte [SRC_PTR] |
582 : | movzx TMP1, byte [SRC_PTR+8] | ||
583 : | XVID_MOVQ mm0, xvid_FIR_14_3_2_1, _EAX*8 | ||
584 : | movzx _EAX, byte [SRC_PTR+1] | ||
585 : | XVID_MOVQ mm3, xvid_FIR_1_2_3_14, TMP1*8 | ||
586 : | edgomez | 1.2 | paddw mm0, mm7 |
587 : | paddw mm3, mm7 | ||
588 : | |||
589 : | Isibaar | 1.8 | movzx TMP1, byte [SRC_PTR+2] |
590 : | XVID_PADDW mm0, xvid_FIR_23_19_6_3, _EAX*8 | ||
591 : | XVID_PADDW mm3, xvid_FIR_1_0_0_0, _EAX*8 | ||
592 : | |||
593 : | movzx _EAX, byte [SRC_PTR+3] | ||
594 : | XVID_PADDW mm0, xvid_FIR_7_20_20_6, TMP1*8 | ||
595 : | XVID_PADDW mm3, xvid_FIR_3_1_0_0, TMP1*8 | ||
596 : | |||
597 : | movzx TMP1, byte [SRC_PTR+4] | ||
598 : | XVID_PADDW mm0, xvid_FIR_3_6_20_20, _EAX*8 | ||
599 : | XVID_PADDW mm3, xvid_FIR_6_3_1_0, _EAX*8 | ||
600 : | |||
601 : | movzx _EAX, byte [SRC_PTR+5] | ||
602 : | XVID_PADDW mm0, xvid_FIR_1_3_6_20, TMP1*8 | ||
603 : | XVID_PADDW mm3, xvid_FIR_20_6_3_1, TMP1*8 | ||
604 : | |||
605 : | movzx TMP1, byte [SRC_PTR+6] | ||
606 : | XVID_PADDW mm0, xvid_FIR_0_1_3_6, _EAX*8 | ||
607 : | XVID_PADDW mm3, xvid_FIR_20_20_6_3, _EAX*8 | ||
608 : | |||
609 : | movzx _EAX, byte [SRC_PTR+7] | ||
610 : | XVID_PADDW mm0, xvid_FIR_0_0_1_3, TMP1*8 | ||
611 : | XVID_PADDW mm3, xvid_FIR_6_20_20_7, TMP1*8 | ||
612 : | edgomez | 1.2 | |
613 : | Isibaar | 1.8 | XVID_PADDW mm0, xvid_FIR_0_0_0_1, _EAX*8 |
614 : | XVID_PADDW mm3, xvid_FIR_3_6_19_23, _EAX*8 | ||
615 : | edgomez | 1.2 | |
616 : | %endif | ||
617 : | |||
618 : | %endif ; !USE_TABLES | ||
619 : | |||
620 : | psraw mm0, 5 | ||
621 : | psraw mm3, 5 | ||
622 : | packuswb mm0, mm3 | ||
623 : | |||
624 : | %if (%1==1) | ||
625 : | Isibaar | 1.8 | MIX mm0, SRC_PTR, _EBX |
626 : | edgomez | 1.2 | %elif (%1==2) |
627 : | Isibaar | 1.8 | MIX mm0, SRC_PTR+1, _EBX |
628 : | edgomez | 1.2 | %endif |
629 : | %if (%2==1) | ||
630 : | Isibaar | 1.8 | MIX mm0, DST_PTR, Rounder1_MMX |
631 : | edgomez | 1.2 | %endif |
632 : | |||
633 : | Isibaar | 1.8 | movq [DST_PTR], mm0 |
634 : | edgomez | 1.2 | |
635 : | Isibaar | 1.8 | add DST_PTR, _EBP |
636 : | add SRC_PTR, _EBP | ||
637 : | dec TMP0 | ||
638 : | edgomez | 1.2 | jg .Loop |
639 : | |||
640 : | %if (%2==0) && (%1==0) | ||
641 : | EPILOG_NO_AVRG | ||
642 : | %else | ||
643 : | EPILOG_AVRG | ||
644 : | %endif | ||
645 : | |||
646 : | %endmacro | ||
647 : | |||
648 : | ;////////////////////////////////////////////////////////////////////// | ||
649 : | ;// 16x? copy Functions | ||
650 : | |||
651 : | xvid_H_Pass_16_mmx: | ||
652 : | H_PASS_16 0, 0 | ||
653 : | Isibaar | 1.7 | ENDFUNC |
654 : | edgomez | 1.2 | xvid_H_Pass_Avrg_16_mmx: |
655 : | H_PASS_16 1, 0 | ||
656 : | Isibaar | 1.7 | ENDFUNC |
657 : | edgomez | 1.2 | xvid_H_Pass_Avrg_Up_16_mmx: |
658 : | H_PASS_16 2, 0 | ||
659 : | Isibaar | 1.7 | ENDFUNC |
660 : | edgomez | 1.2 | |
661 : | ;////////////////////////////////////////////////////////////////////// | ||
662 : | ;// 8x? copy Functions | ||
663 : | |||
664 : | xvid_H_Pass_8_mmx: | ||
665 : | H_PASS_8 0, 0 | ||
666 : | Isibaar | 1.7 | ENDFUNC |
667 : | edgomez | 1.2 | xvid_H_Pass_Avrg_8_mmx: |
668 : | H_PASS_8 1, 0 | ||
669 : | Isibaar | 1.7 | ENDFUNC |
670 : | edgomez | 1.2 | xvid_H_Pass_Avrg_Up_8_mmx: |
671 : | H_PASS_8 2, 0 | ||
672 : | Isibaar | 1.7 | ENDFUNC |
673 : | edgomez | 1.2 | |
674 : | ;////////////////////////////////////////////////////////////////////// | ||
675 : | ;// 16x? avrg Functions | ||
676 : | |||
677 : | xvid_H_Pass_Add_16_mmx: | ||
678 : | H_PASS_16 0, 1 | ||
679 : | Isibaar | 1.7 | ENDFUNC |
680 : | edgomez | 1.2 | xvid_H_Pass_Avrg_Add_16_mmx: |
681 : | H_PASS_16 1, 1 | ||
682 : | Isibaar | 1.7 | ENDFUNC |
683 : | edgomez | 1.2 | xvid_H_Pass_Avrg_Up_Add_16_mmx: |
684 : | H_PASS_16 2, 1 | ||
685 : | Isibaar | 1.7 | ENDFUNC |
686 : | edgomez | 1.2 | |
687 : | ;////////////////////////////////////////////////////////////////////// | ||
688 : | ;// 8x? avrg Functions | ||
689 : | |||
690 : | xvid_H_Pass_8_Add_mmx: | ||
691 : | H_PASS_8 0, 1 | ||
692 : | Isibaar | 1.7 | ENDFUNC |
693 : | edgomez | 1.2 | xvid_H_Pass_Avrg_8_Add_mmx: |
694 : | H_PASS_8 1, 1 | ||
695 : | Isibaar | 1.7 | ENDFUNC |
696 : | edgomez | 1.2 | xvid_H_Pass_Avrg_Up_8_Add_mmx: |
697 : | H_PASS_8 2, 1 | ||
698 : | Isibaar | 1.7 | ENDFUNC |
699 : | edgomez | 1.2 | |
700 : | |||
701 : | |||
702 : | ;////////////////////////////////////////////////////////////////////// | ||
703 : | ;// | ||
704 : | ;// All vertical passes | ||
705 : | ;// | ||
706 : | ;////////////////////////////////////////////////////////////////////// | ||
707 : | |||
708 : | %macro V_LOAD 1 ; %1=Last? | ||
709 : | |||
710 : | Isibaar | 1.8 | movd mm4, dword [TMP1] |
711 : | edgomez | 1.2 | pxor mm6, mm6 |
712 : | %if (%1==0) | ||
713 : | Isibaar | 1.8 | add TMP1, _EBP |
714 : | edgomez | 1.2 | %endif |
715 : | punpcklbw mm4, mm6 | ||
716 : | |||
717 : | %endmacro | ||
718 : | |||
719 : | %macro V_ACC1 2 ; %1:reg; 2:tap | ||
720 : | pmullw mm4, [%2] | ||
721 : | paddw %1, mm4 | ||
722 : | %endmacro | ||
723 : | |||
724 : | %macro V_ACC2 4 ; %1-%2: regs, %3-%4: taps | ||
725 : | movq mm5, mm4 | ||
726 : | movq mm6, mm4 | ||
727 : | pmullw mm5, [%3] | ||
728 : | pmullw mm6, [%4] | ||
729 : | paddw %1, mm5 | ||
730 : | paddw %2, mm6 | ||
731 : | %endmacro | ||
732 : | |||
733 : | %macro V_ACC2l 4 ; %1-%2: regs, %3-%4: taps | ||
734 : | movq mm5, mm4 | ||
735 : | pmullw mm5, [%3] | ||
736 : | pmullw mm4, [%4] | ||
737 : | paddw %1, mm5 | ||
738 : | paddw %2, mm4 | ||
739 : | %endmacro | ||
740 : | |||
741 : | %macro V_ACC4 8 ; %1-%4: regs, %5-%8: taps | ||
742 : | V_ACC2 %1,%2, %5,%6 | ||
743 : | V_ACC2l %3,%4, %7,%8 | ||
744 : | %endmacro | ||
745 : | |||
746 : | |||
747 : | %macro V_MIX 3 ; %1:dst-reg, %2:src, %3: rounder | ||
748 : | pxor mm6, mm6 | ||
749 : | movq mm4, [%2] | ||
750 : | punpcklbw %1, mm6 | ||
751 : | punpcklbw mm4, mm6 | ||
752 : | paddusw %1, mm4 | ||
753 : | paddusw %1, [%3] | ||
754 : | psrlw %1, 1 | ||
755 : | packuswb %1, %1 | ||
756 : | %endmacro | ||
757 : | |||
758 : | %macro V_STORE 4 ; %1-%2: mix ops, %3: reg, %4:last? | ||
759 : | |||
760 : | psraw %3, 5 | ||
761 : | packuswb %3, %3 | ||
762 : | |||
763 : | %if (%1==1) | ||
764 : | Isibaar | 1.8 | V_MIX %3, SRC_PTR, _EBX |
765 : | add SRC_PTR, _EBP | ||
766 : | edgomez | 1.2 | %elif (%1==2) |
767 : | Isibaar | 1.8 | add SRC_PTR, _EBP |
768 : | V_MIX %3, SRC_PTR, _EBX | ||
769 : | edgomez | 1.2 | %endif |
770 : | %if (%2==1) | ||
771 : | Isibaar | 1.8 | V_MIX %3, DST_PTR, Rounder1_MMX |
772 : | edgomez | 1.2 | %endif |
773 : | |||
774 : | movd eax, %3 | ||
775 : | Isibaar | 1.8 | mov [DST_PTR], _EAX |
776 : | edgomez | 1.2 | |
777 : | %if (%4==0) | ||
778 : | Isibaar | 1.8 | add DST_PTR, _EBP |
779 : | edgomez | 1.2 | %endif |
780 : | |||
781 : | %endmacro | ||
782 : | |||
783 : | ;////////////////////////////////////////////////////////////////////// | ||
784 : | |||
785 : | %macro V_PASS_16 2 ; %1:src-op (0=NONE,1=AVRG,2=AVRG-UP), %2:dst-op (NONE/AVRG) | ||
786 : | |||
787 : | %if (%2==0) && (%1==0) | ||
788 : | PROLOG_NO_AVRG | ||
789 : | %else | ||
790 : | PROLOG_AVRG | ||
791 : | %endif | ||
792 : | |||
793 : | ; we process one stripe of 4x16 pixel each time. | ||
794 : | ; the size (3rd argument) is meant to be a multiple of 4 | ||
795 : | ; mm0..mm3 serves as a 4x4 delay line | ||
796 : | |||
797 : | Isibaar | 1.7 | .Loop: |
798 : | edgomez | 1.2 | |
799 : | Isibaar | 1.8 | push DST_PTR |
800 : | push SRC_PTR ; SRC_PTR is preserved for src-mixing | ||
801 : | mov TMP1, SRC_PTR | ||
802 : | edgomez | 1.2 | |
803 : | ; ouput rows [0..3], from input rows [0..8] | ||
804 : | |||
805 : | movq mm0, mm7 | ||
806 : | movq mm1, mm7 | ||
807 : | movq mm2, mm7 | ||
808 : | movq mm3, mm7 | ||
809 : | |||
810 : | V_LOAD 0 | ||
811 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C14, FIR_Cm3, FIR_C2, FIR_Cm1 | ||
812 : | V_LOAD 0 | ||
813 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C23, FIR_C19, FIR_Cm6, FIR_C3 | ||
814 : | V_LOAD 0 | ||
815 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm7, FIR_C20, FIR_C20, FIR_Cm6 | ||
816 : | V_LOAD 0 | ||
817 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
818 : | V_LOAD 0 | ||
819 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
820 : | V_STORE %1, %2, mm0, 0 | ||
821 : | |||
822 : | V_LOAD 0 | ||
823 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
824 : | V_ACC1 mm3, FIR_Cm6 | ||
825 : | V_STORE %1, %2, mm1, 0 | ||
826 : | |||
827 : | V_LOAD 0 | ||
828 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
829 : | V_STORE %1, %2, mm2, 0 | ||
830 : | |||
831 : | V_LOAD 1 | ||
832 : | V_ACC1 mm3, FIR_Cm1 | ||
833 : | V_STORE %1, %2, mm3, 0 | ||
834 : | |||
835 : | ; ouput rows [4..7], from input rows [1..11] (!!) | ||
836 : | |||
837 : | Isibaar | 1.8 | mov SRC_PTR, [_ESP] |
838 : | lea TMP1, [SRC_PTR+_EBP] | ||
839 : | edgomez | 1.2 | |
840 : | Isibaar | 1.8 | lea SRC_PTR, [SRC_PTR+4*_EBP] ; for src-mixing |
841 : | push SRC_PTR ; this will be the new value for next round | ||
842 : | edgomez | 1.2 | |
843 : | movq mm0, mm7 | ||
844 : | movq mm1, mm7 | ||
845 : | movq mm2, mm7 | ||
846 : | movq mm3, mm7 | ||
847 : | |||
848 : | V_LOAD 0 | ||
849 : | V_ACC1 mm0, FIR_Cm1 | ||
850 : | |||
851 : | V_LOAD 0 | ||
852 : | V_ACC2l mm0, mm1, FIR_C3, FIR_Cm1 | ||
853 : | |||
854 : | V_LOAD 0 | ||
855 : | V_ACC2 mm0, mm1, FIR_Cm6, FIR_C3 | ||
856 : | V_ACC1 mm2, FIR_Cm1 | ||
857 : | |||
858 : | V_LOAD 0 | ||
859 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C20, FIR_Cm6, FIR_C3, FIR_Cm1 | ||
860 : | V_LOAD 0 | ||
861 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C20, FIR_C20, FIR_Cm6, FIR_C3 | ||
862 : | V_LOAD 0 | ||
863 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm6, FIR_C20, FIR_C20, FIR_Cm6 | ||
864 : | V_LOAD 0 | ||
865 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
866 : | V_LOAD 0 | ||
867 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
868 : | V_STORE %1, %2, mm0, 0 | ||
869 : | |||
870 : | V_LOAD 0 | ||
871 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
872 : | V_ACC1 mm3, FIR_Cm6 | ||
873 : | V_STORE %1, %2, mm1, 0 | ||
874 : | |||
875 : | V_LOAD 0 | ||
876 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
877 : | V_STORE %1, %2, mm2, 0 | ||
878 : | |||
879 : | V_LOAD 1 | ||
880 : | V_ACC1 mm3, FIR_Cm1 | ||
881 : | V_STORE %1, %2, mm3, 0 | ||
882 : | |||
883 : | ; ouput rows [8..11], from input rows [5..15] | ||
884 : | |||
885 : | Isibaar | 1.8 | pop SRC_PTR |
886 : | lea TMP1, [SRC_PTR+_EBP] | ||
887 : | edgomez | 1.2 | |
888 : | Isibaar | 1.8 | lea SRC_PTR, [SRC_PTR+4*_EBP] ; for src-mixing |
889 : | push SRC_PTR ; this will be the new value for next round | ||
890 : | edgomez | 1.2 | |
891 : | movq mm0, mm7 | ||
892 : | movq mm1, mm7 | ||
893 : | movq mm2, mm7 | ||
894 : | movq mm3, mm7 | ||
895 : | |||
896 : | V_LOAD 0 | ||
897 : | V_ACC1 mm0, FIR_Cm1 | ||
898 : | |||
899 : | V_LOAD 0 | ||
900 : | V_ACC2l mm0, mm1, FIR_C3, FIR_Cm1 | ||
901 : | |||
902 : | V_LOAD 0 | ||
903 : | V_ACC2 mm0, mm1, FIR_Cm6, FIR_C3 | ||
904 : | V_ACC1 mm2, FIR_Cm1 | ||
905 : | |||
906 : | V_LOAD 0 | ||
907 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C20, FIR_Cm6, FIR_C3, FIR_Cm1 | ||
908 : | V_LOAD 0 | ||
909 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C20, FIR_C20, FIR_Cm6, FIR_C3 | ||
910 : | V_LOAD 0 | ||
911 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm6, FIR_C20, FIR_C20, FIR_Cm6 | ||
912 : | V_LOAD 0 | ||
913 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
914 : | V_LOAD 0 | ||
915 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
916 : | |||
917 : | V_STORE %1, %2, mm0, 0 | ||
918 : | |||
919 : | V_LOAD 0 | ||
920 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
921 : | V_ACC1 mm3, FIR_Cm6 | ||
922 : | V_STORE %1, %2, mm1, 0 | ||
923 : | |||
924 : | V_LOAD 0 | ||
925 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
926 : | V_STORE %1, %2, mm2, 0 | ||
927 : | |||
928 : | V_LOAD 1 | ||
929 : | V_ACC1 mm3, FIR_Cm1 | ||
930 : | V_STORE %1, %2, mm3, 0 | ||
931 : | |||
932 : | |||
933 : | ; ouput rows [12..15], from input rows [9.16] | ||
934 : | |||
935 : | Isibaar | 1.8 | pop SRC_PTR |
936 : | lea TMP1, [SRC_PTR+_EBP] | ||
937 : | edgomez | 1.2 | |
938 : | %if (%1!=0) | ||
939 : | Isibaar | 1.8 | lea SRC_PTR, [SRC_PTR+4*_EBP] ; for src-mixing |
940 : | edgomez | 1.2 | %endif |
941 : | |||
942 : | movq mm0, mm7 | ||
943 : | movq mm1, mm7 | ||
944 : | movq mm2, mm7 | ||
945 : | movq mm3, mm7 | ||
946 : | |||
947 : | V_LOAD 0 | ||
948 : | V_ACC1 mm3, FIR_Cm1 | ||
949 : | |||
950 : | V_LOAD 0 | ||
951 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
952 : | |||
953 : | V_LOAD 0 | ||
954 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
955 : | V_ACC1 mm3, FIR_Cm6 | ||
956 : | |||
957 : | V_LOAD 0 | ||
958 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
959 : | V_LOAD 0 | ||
960 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
961 : | V_LOAD 0 | ||
962 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm7, FIR_C20, FIR_C20, FIR_Cm6 | ||
963 : | V_LOAD 0 | ||
964 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C23, FIR_C19, FIR_Cm6, FIR_C3 | ||
965 : | V_LOAD 1 | ||
966 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C14, FIR_Cm3, FIR_C2, FIR_Cm1 | ||
967 : | |||
968 : | V_STORE %1, %2, mm3, 0 | ||
969 : | V_STORE %1, %2, mm2, 0 | ||
970 : | V_STORE %1, %2, mm1, 0 | ||
971 : | V_STORE %1, %2, mm0, 1 | ||
972 : | |||
973 : | ; ... next 4 columns | ||
974 : | |||
975 : | Isibaar | 1.8 | pop SRC_PTR |
976 : | pop DST_PTR | ||
977 : | add SRC_PTR, 4 | ||
978 : | add DST_PTR, 4 | ||
979 : | sub TMP0, 4 | ||
980 : | edgomez | 1.2 | jg .Loop |
981 : | |||
982 : | %if (%2==0) && (%1==0) | ||
983 : | EPILOG_NO_AVRG | ||
984 : | %else | ||
985 : | EPILOG_AVRG | ||
986 : | %endif | ||
987 : | |||
988 : | %endmacro | ||
989 : | |||
990 : | ;////////////////////////////////////////////////////////////////////// | ||
991 : | |||
992 : | %macro V_PASS_8 2 ; %1:src-op (0=NONE,1=AVRG,2=AVRG-UP), %2:dst-op (NONE/AVRG) | ||
993 : | |||
994 : | %if (%2==0) && (%1==0) | ||
995 : | PROLOG_NO_AVRG | ||
996 : | %else | ||
997 : | PROLOG_AVRG | ||
998 : | %endif | ||
999 : | |||
1000 : | ; we process one stripe of 4x8 pixel each time | ||
1001 : | ; the size (3rd argument) is meant to be a multiple of 4 | ||
1002 : | ; mm0..mm3 serves as a 4x4 delay line | ||
1003 : | Isibaar | 1.7 | .Loop: |
1004 : | edgomez | 1.2 | |
1005 : | Isibaar | 1.8 | push DST_PTR |
1006 : | push SRC_PTR ; SRC_PTR is preserved for src-mixing | ||
1007 : | mov TMP1, SRC_PTR | ||
1008 : | edgomez | 1.2 | |
1009 : | ; ouput rows [0..3], from input rows [0..8] | ||
1010 : | |||
1011 : | movq mm0, mm7 | ||
1012 : | movq mm1, mm7 | ||
1013 : | movq mm2, mm7 | ||
1014 : | movq mm3, mm7 | ||
1015 : | |||
1016 : | V_LOAD 0 | ||
1017 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C14, FIR_Cm3, FIR_C2, FIR_Cm1 | ||
1018 : | V_LOAD 0 | ||
1019 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C23, FIR_C19, FIR_Cm6, FIR_C3 | ||
1020 : | V_LOAD 0 | ||
1021 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm7, FIR_C20, FIR_C20, FIR_Cm6 | ||
1022 : | V_LOAD 0 | ||
1023 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
1024 : | V_LOAD 0 | ||
1025 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
1026 : | V_STORE %1, %2, mm0, 0 | ||
1027 : | |||
1028 : | V_LOAD 0 | ||
1029 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
1030 : | V_ACC1 mm3, FIR_Cm6 | ||
1031 : | |||
1032 : | V_STORE %1, %2, mm1, 0 | ||
1033 : | |||
1034 : | V_LOAD 0 | ||
1035 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
1036 : | V_STORE %1, %2, mm2, 0 | ||
1037 : | |||
1038 : | V_LOAD 1 | ||
1039 : | V_ACC1 mm3, FIR_Cm1 | ||
1040 : | V_STORE %1, %2, mm3, 0 | ||
1041 : | |||
1042 : | ; ouput rows [4..7], from input rows [1..9] | ||
1043 : | |||
1044 : | Isibaar | 1.8 | mov SRC_PTR, [_ESP] |
1045 : | lea TMP1, [SRC_PTR+_EBP] | ||
1046 : | edgomez | 1.2 | |
1047 : | %if (%1!=0) | ||
1048 : | Isibaar | 1.8 | lea SRC_PTR, [SRC_PTR+4*_EBP] ; for src-mixing |
1049 : | edgomez | 1.2 | %endif |
1050 : | |||
1051 : | movq mm0, mm7 | ||
1052 : | movq mm1, mm7 | ||
1053 : | movq mm2, mm7 | ||
1054 : | movq mm3, mm7 | ||
1055 : | |||
1056 : | V_LOAD 0 | ||
1057 : | V_ACC1 mm3, FIR_Cm1 | ||
1058 : | |||
1059 : | V_LOAD 0 | ||
1060 : | V_ACC2l mm2, mm3, FIR_Cm1, FIR_C3 | ||
1061 : | |||
1062 : | V_LOAD 0 | ||
1063 : | V_ACC2 mm1, mm2, FIR_Cm1, FIR_C3 | ||
1064 : | V_ACC1 mm3, FIR_Cm6 | ||
1065 : | |||
1066 : | V_LOAD 0 | ||
1067 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm1, FIR_C3, FIR_Cm6, FIR_C20 | ||
1068 : | V_LOAD 0 | ||
1069 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C3, FIR_Cm6, FIR_C20, FIR_C20 | ||
1070 : | V_LOAD 0 | ||
1071 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_Cm7, FIR_C20, FIR_C20, FIR_Cm6 | ||
1072 : | V_LOAD 0 | ||
1073 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C23, FIR_C19, FIR_Cm6, FIR_C3 | ||
1074 : | V_LOAD 1 | ||
1075 : | V_ACC4 mm0, mm1, mm2, mm3, FIR_C14, FIR_Cm3, FIR_C2, FIR_Cm1 | ||
1076 : | |||
1077 : | V_STORE %1, %2, mm3, 0 | ||
1078 : | V_STORE %1, %2, mm2, 0 | ||
1079 : | V_STORE %1, %2, mm1, 0 | ||
1080 : | V_STORE %1, %2, mm0, 1 | ||
1081 : | |||
1082 : | ; ... next 4 columns | ||
1083 : | |||
1084 : | Isibaar | 1.8 | pop SRC_PTR |
1085 : | pop DST_PTR | ||
1086 : | add SRC_PTR, 4 | ||
1087 : | add DST_PTR, 4 | ||
1088 : | sub TMP0, 4 | ||
1089 : | edgomez | 1.2 | jg .Loop |
1090 : | |||
1091 : | %if (%2==0) && (%1==0) | ||
1092 : | EPILOG_NO_AVRG | ||
1093 : | %else | ||
1094 : | EPILOG_AVRG | ||
1095 : | %endif | ||
1096 : | |||
1097 : | %endmacro | ||
1098 : | |||
1099 : | |||
1100 : | ;////////////////////////////////////////////////////////////////////// | ||
1101 : | ;// 16x? copy Functions | ||
1102 : | |||
1103 : | xvid_V_Pass_16_mmx: | ||
1104 : | V_PASS_16 0, 0 | ||
1105 : | Isibaar | 1.7 | ENDFUNC |
1106 : | edgomez | 1.2 | xvid_V_Pass_Avrg_16_mmx: |
1107 : | V_PASS_16 1, 0 | ||
1108 : | Isibaar | 1.7 | ENDFUNC |
1109 : | edgomez | 1.2 | xvid_V_Pass_Avrg_Up_16_mmx: |
1110 : | V_PASS_16 2, 0 | ||
1111 : | Isibaar | 1.7 | ENDFUNC |
1112 : | edgomez | 1.2 | |
1113 : | ;////////////////////////////////////////////////////////////////////// | ||
1114 : | ;// 8x? copy Functions | ||
1115 : | |||
1116 : | xvid_V_Pass_8_mmx: | ||
1117 : | V_PASS_8 0, 0 | ||
1118 : | Isibaar | 1.7 | ENDFUNC |
1119 : | edgomez | 1.2 | xvid_V_Pass_Avrg_8_mmx: |
1120 : | V_PASS_8 1, 0 | ||
1121 : | Isibaar | 1.7 | ENDFUNC |
1122 : | edgomez | 1.2 | xvid_V_Pass_Avrg_Up_8_mmx: |
1123 : | V_PASS_8 2, 0 | ||
1124 : | Isibaar | 1.7 | ENDFUNC |
1125 : | edgomez | 1.2 | |
1126 : | ;////////////////////////////////////////////////////////////////////// | ||
1127 : | ;// 16x? avrg Functions | ||
1128 : | |||
1129 : | xvid_V_Pass_Add_16_mmx: | ||
1130 : | V_PASS_16 0, 1 | ||
1131 : | Isibaar | 1.7 | ENDFUNC |
1132 : | edgomez | 1.2 | xvid_V_Pass_Avrg_Add_16_mmx: |
1133 : | V_PASS_16 1, 1 | ||
1134 : | Isibaar | 1.7 | ENDFUNC |
1135 : | edgomez | 1.2 | xvid_V_Pass_Avrg_Up_Add_16_mmx: |
1136 : | V_PASS_16 2, 1 | ||
1137 : | Isibaar | 1.7 | ENDFUNC |
1138 : | edgomez | 1.2 | |
1139 : | ;////////////////////////////////////////////////////////////////////// | ||
1140 : | ;// 8x? avrg Functions | ||
1141 : | |||
1142 : | xvid_V_Pass_8_Add_mmx: | ||
1143 : | V_PASS_8 0, 1 | ||
1144 : | Isibaar | 1.7 | ENDFUNC |
1145 : | edgomez | 1.2 | xvid_V_Pass_Avrg_8_Add_mmx: |
1146 : | V_PASS_8 1, 1 | ||
1147 : | Isibaar | 1.7 | ENDFUNC |
1148 : | edgomez | 1.2 | xvid_V_Pass_Avrg_Up_8_Add_mmx: |
1149 : | V_PASS_8 2, 1 | ||
1150 : | Isibaar | 1.7 | ENDFUNC |
1151 : | edgomez | 1.2 | |
1152 : | ;////////////////////////////////////////////////////////////////////// | ||
1153 : | Isibaar | 1.6 | |
1154 : | Isibaar | 1.8 | %undef SRC_PTR |
1155 : | %undef DST_PTR | ||
1156 : | |||
1157 : | Isibaar | 1.6 | %ifidn __OUTPUT_FORMAT__,elf |
1158 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
1159 : | %endif | ||
1160 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |