Rev 281 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
168 | serge | 1 | ; |
2 | ; This file is part of the Infinity sound library. |
||
3 | ; (C) copyright Serge 2006 |
||
4 | ; email: infinity_sound@mail.ru |
||
5 | ; |
||
6 | ; This program is free software; you can redistribute it and/or modify |
||
7 | ; it under the terms of the GNU General Public License as published by |
||
8 | ; the Free Software Foundation; either version 2 of the License, or |
||
9 | ; (at your option) any later version. |
||
10 | ; |
||
11 | ; This program is distributed in the hope that it will be useful, |
||
12 | ; but WITHOUT ANY WARRANTY; without even the implied warranty of |
||
13 | ; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
||
14 | ; GNU General Public License for more details. |
||
15 | |||
16 | align 4 |
||
17 | proc new_mix stdcall, output:dword |
||
281 | serge | 18 | locals |
19 | mixCounter dd ? |
||
20 | mixIndex dd ? |
||
21 | streamIndex dd ? |
||
22 | inputCount dd ? |
||
23 | main_count dd ? |
||
24 | blockCount dd ? |
||
25 | mix_out dd ? |
||
26 | endl |
||
168 | serge | 27 | |
281 | serge | 28 | call prepare_playlist |
168 | serge | 29 | |
281 | serge | 30 | cmp [play_count], 0 |
31 | je .exit |
||
188 | serge | 32 | call FpuSave |
281 | serge | 33 | mov [main_count], 32; |
168 | serge | 34 | .l00: |
281 | serge | 35 | mov [mix_buff_map], 0x0000FFFF; |
36 | xor eax, eax |
||
37 | mov [mixCounter], eax |
||
38 | mov [mixIndex],eax |
||
39 | mov [streamIndex], eax; |
||
40 | mov ebx, [play_count] |
||
41 | mov [inputCount], ebx |
||
168 | serge | 42 | .l0: |
281 | serge | 43 | mov ecx, 4 |
168 | serge | 44 | .l1: |
281 | serge | 45 | mov ebx, [streamIndex] |
46 | mov esi, [play_list+ebx*4] |
||
47 | mov eax, [esi+STREAM.work_read] |
||
48 | add [esi+STREAM.work_read], 512 |
||
168 | serge | 49 | |
281 | serge | 50 | mov ebx, [mixIndex] |
51 | mov [mix_input+ebx*4], eax |
||
52 | inc [mixCounter] |
||
53 | inc [mixIndex] |
||
54 | inc [streamIndex] |
||
55 | dec [inputCount] |
||
56 | jz .m2 |
||
168 | serge | 57 | |
281 | serge | 58 | dec ecx |
59 | jnz .l1 |
||
168 | serge | 60 | |
281 | serge | 61 | cmp [mixCounter], 4 |
62 | jnz .m2 |
||
168 | serge | 63 | |
281 | serge | 64 | stdcall mix_4_1, [mix_input],[mix_input+4],[mix_input+8],[mix_input+12] |
65 | sub [mixIndex],4 |
||
66 | mov ebx, [mixIndex] |
||
67 | mov [mix_input+ebx*4], eax |
||
68 | inc [mixIndex] |
||
69 | mov [mixCounter], 0 |
||
168 | serge | 70 | |
281 | serge | 71 | cmp [inputCount], 0 |
72 | jnz .l0 |
||
168 | serge | 73 | .m2: |
281 | serge | 74 | cmp [mixIndex], 1 |
75 | jne @f |
||
76 | stdcall copy_mem, [output], [mix_input] |
||
77 | jmp .m3 |
||
168 | serge | 78 | @@: |
281 | serge | 79 | cmp [mixIndex], 2 |
80 | jne @f |
||
81 | stdcall mix_2_1, [output], [mix_input], [mix_input+4] |
||
82 | jmp .m3 |
||
168 | serge | 83 | @@: |
281 | serge | 84 | cmp [mixIndex], 3 |
85 | jne @f |
||
86 | stdcall mix_3_1, [output],[mix_input],[mix_input+4],[mix_input+8] |
||
87 | jmp .m3 |
||
168 | serge | 88 | @@: |
281 | serge | 89 | stdcall final_mix, [output],[mix_input],[mix_input+4],[mix_input+8], [mix_input+12] |
168 | serge | 90 | .m3: |
281 | serge | 91 | add [output],512 |
168 | serge | 92 | |
281 | serge | 93 | sub [main_count], 1 |
94 | jnz .l00 |
||
168 | serge | 95 | |
281 | serge | 96 | call update_stream |
203 | serge | 97 | emms |
188 | serge | 98 | call FpuRestore |
281 | serge | 99 | ret |
168 | serge | 100 | .exit: |
281 | serge | 101 | mov edi, [output] |
102 | mov ecx, 0x1000 |
||
103 | xor eax, eax |
||
104 | cld |
||
105 | rep stosd |
||
106 | ret |
||
168 | serge | 107 | endp |
108 | |||
109 | align 4 |
||
110 | proc update_stream |
||
281 | serge | 111 | locals |
170 | serge | 112 | stream_index dd ? |
227 | serge | 113 | ev_code dd ? ;EVENT |
114 | ev_offs dd ? |
||
115 | rd 4 |
||
281 | serge | 116 | endl |
168 | serge | 117 | |
281 | serge | 118 | mov [stream_index], 0 |
168 | serge | 119 | .l1: |
281 | serge | 120 | mov edx, [stream_index] |
121 | mov esi, [play_list+edx*4] |
||
168 | serge | 122 | |
281 | serge | 123 | mov eax, [esi+STREAM.work_read] |
124 | cmp eax, [esi+STREAM.work_top] |
||
125 | jb @f |
||
126 | mov eax, [esi+STREAM.work_buff] |
||
168 | serge | 127 | @@: |
281 | serge | 128 | mov [esi+STREAM.work_read], eax |
168 | serge | 129 | |
281 | serge | 130 | cmp [esi+STREAM.format], PCM_2_16_48 |
131 | je .copy |
||
168 | serge | 132 | |
281 | serge | 133 | sub [esi+STREAM.work_count], 16384 |
168 | serge | 134 | |
281 | serge | 135 | cmp [esi+STREAM.work_count], 32768 |
136 | ja @f |
||
168 | serge | 137 | |
281 | serge | 138 | stdcall refill, esi |
168 | serge | 139 | @@: |
281 | serge | 140 | inc [stream_index] |
141 | dec [play_count] |
||
142 | jnz .l1 |
||
143 | ret |
||
168 | serge | 144 | .copy: |
281 | serge | 145 | mov ebx, esi |
146 | mov edi, [ebx+STREAM.work_write] |
||
147 | cmp edi, [ebx+STREAM.work_top] |
||
148 | jb @f |
||
149 | mov edi, [ebx+STREAM.work_buff] |
||
150 | mov [ebx+STREAM.work_write], edi |
||
168 | serge | 151 | @@: |
281 | serge | 152 | mov esi, [ebx+STREAM.curr_seg] |
153 | mov ecx, 16384/4 |
||
154 | cld |
||
155 | rep movsd |
||
168 | serge | 156 | |
157 | mov [ebx+STREAM.work_write], edi |
||
158 | |||
281 | serge | 159 | cmp esi, [ebx+STREAM.lim_0] |
160 | jb @f |
||
168 | serge | 161 | |
281 | serge | 162 | mov esi, [ebx+STREAM.seg_0] |
163 | mov eax, [ebx+STREAM.lim_0] |
||
164 | xchg esi, [ebx+STREAM.seg_1] |
||
165 | xchg eax, [ebx+STREAM.lim_1] |
||
166 | mov [ebx+STREAM.seg_0], esi |
||
167 | mov [ebx+STREAM.lim_0], eax |
||
168 | serge | 168 | @@: |
281 | serge | 169 | mov [ebx+STREAM.curr_seg], esi |
168 | serge | 170 | |
281 | serge | 171 | xor ecx, ecx |
172 | cmp esi, [ebx+STREAM.notify_off2] |
||
173 | je @f |
||
168 | serge | 174 | |
281 | serge | 175 | mov ecx,0x8000 |
176 | cmp esi, [ebx+STREAM.notify_off1] |
||
177 | je @f |
||
168 | serge | 178 | |
281 | serge | 179 | inc [stream_index] |
180 | dec [play_count] |
||
181 | jnz .l1 |
||
182 | ret |
||
168 | serge | 183 | @@: |
227 | serge | 184 | mov [ev_code], 0xFF000001 |
185 | mov [ev_offs], ecx |
||
168 | serge | 186 | mov eax, [ebx+STREAM.notify_task] |
227 | serge | 187 | |
188 | lea edx, [ev_code] |
||
189 | push ebx |
||
190 | stdcall SendEvent, eax, edx |
||
191 | pop ebx |
||
168 | serge | 192 | test eax, eax |
227 | serge | 193 | jnz .l_end |
194 | |||
168 | serge | 195 | not eax |
196 | mov [ebx+STREAM.notify_task], eax ;-1 |
||
197 | .l_end: |
||
281 | serge | 198 | inc [stream_index] |
199 | dec [play_count] |
||
200 | jnz .l1 |
||
201 | ret |
||
168 | serge | 202 | endp |
203 | |||
204 | align 4 |
||
205 | proc refill stdcall, str:dword |
||
281 | serge | 206 | locals |
227 | serge | 207 | ev_code dd ? ;EVENT |
208 | ev_offs dd ? |
||
209 | rd 4 |
||
281 | serge | 210 | endl |
168 | serge | 211 | |
281 | serge | 212 | mov ebx, [str] |
213 | mov ecx, [ebx+STREAM.work_write] |
||
214 | cmp ecx, [ebx+STREAM.work_top] |
||
215 | jbe .m2 |
||
216 | mov esi, [ebx+STREAM.work_top] |
||
217 | sub ecx, esi |
||
218 | mov edi, [ebx+STREAM.work_buff] |
||
219 | shr ecx, 2 |
||
220 | rep movsd ;call memcpy |
||
168 | serge | 221 | |
281 | serge | 222 | mov [ebx+STREAM.work_write], edi |
168 | serge | 223 | .m2: |
281 | serge | 224 | mov esi, [ebx+STREAM.curr_seg] |
225 | mov edi, [ebx+STREAM.work_write] |
||
168 | serge | 226 | |
285 | serge | 227 | stdcall [ebx+STREAM.resample], edi, esi, \ |
281 | serge | 228 | [ebx+STREAM.r_dt],[ebx+STREAM.r_size],[ebx+STREAM.r_end] |
168 | serge | 229 | |
281 | serge | 230 | mov ebx, [str] |
168 | serge | 231 | |
281 | serge | 232 | add [ebx+STREAM.work_count], eax; |
233 | add [ebx+STREAM.work_write], eax; |
||
168 | serge | 234 | |
281 | serge | 235 | mov eax, [ebx+STREAM.curr_seg] |
236 | add eax, [ebx+STREAM.r_size] |
||
237 | cmp eax, [ebx+STREAM.lim_0] |
||
238 | jb @f |
||
239 | |||
285 | serge | 240 | mov esi, [ebx+STREAM.seg_0] |
241 | lea edi, [esi-128] |
||
242 | add esi, 0x7F80 |
||
243 | mov ecx, 128/4 |
||
244 | cld |
||
245 | rep movsd |
||
246 | |||
281 | serge | 247 | mov eax, [ebx+STREAM.seg_0] |
248 | mov ecx, [ebx+STREAM.lim_0] |
||
249 | xchg eax, [ebx+STREAM.seg_1] |
||
250 | xchg ecx, [ebx+STREAM.lim_1] |
||
251 | mov [ebx+STREAM.seg_0], eax |
||
252 | mov [ebx+STREAM.lim_0], ecx |
||
168 | serge | 253 | @@: |
281 | serge | 254 | mov [ebx+STREAM.curr_seg], eax |
168 | serge | 255 | |
281 | serge | 256 | xor ecx, ecx |
257 | cmp eax, [ebx+STREAM.notify_off2] |
||
258 | je @f |
||
168 | serge | 259 | |
281 | serge | 260 | mov ecx,0x8000 |
261 | cmp eax, [ebx+STREAM.notify_off1] |
||
262 | je @f |
||
263 | ret |
||
168 | serge | 264 | @@: |
227 | serge | 265 | mov [ev_code], 0xFF000001 |
266 | mov [ev_offs], ecx |
||
168 | serge | 267 | mov eax, [ebx+STREAM.notify_task] |
227 | serge | 268 | |
269 | lea edx, [ev_code] |
||
270 | push ebx |
||
271 | stdcall SendEvent, eax, edx |
||
272 | pop ebx |
||
168 | serge | 273 | test eax, eax |
227 | serge | 274 | jnz @F |
168 | serge | 275 | not eax |
276 | mov [ebx+STREAM.notify_task], eax ;-1 |
||
277 | @@: |
||
278 | ret |
||
279 | endp |
||
280 | |||
281 | align 4 |
||
285 | serge | 282 | proc resample_1 stdcall, dest:dword,src:dword,\ |
168 | serge | 283 | r_dt:dword, r_size:dword,r_end:dword |
284 | |||
285 | serge | 285 | ; dest equ esp+8 |
286 | ; src equ esp+12 |
||
287 | ; r_dt equ esp+16 |
||
288 | ; r_size equ esp+20 |
||
289 | ;r_end equ esp+24 |
||
168 | serge | 290 | |
285 | serge | 291 | mov edi, [dest] |
292 | mov edx, [src] |
||
293 | sub edx, 32*2 |
||
294 | mov eax, 16 |
||
168 | serge | 295 | |
296 | align 16 |
||
297 | .l1: |
||
298 | mov ecx, eax |
||
299 | mov esi, eax |
||
300 | and ecx, 0x7FFF |
||
301 | shr esi, 15 |
||
302 | lea esi, [edx+esi*2] |
||
303 | |||
304 | movsx ebp, word [esi] |
||
305 | movsx esi, word [esi+2] |
||
306 | mov ebx, 32768 |
||
307 | imul esi, ecx |
||
308 | sub ebx, ecx |
||
309 | imul ebx, ebp |
||
310 | lea ecx, [ebx+esi+16384] |
||
311 | sar ecx, 15 |
||
312 | cmp ecx, 32767 ; 00007fffH |
||
313 | jle @f |
||
314 | mov ecx, 32767 ; 00007fffH |
||
315 | jmp .write |
||
316 | @@: |
||
317 | cmp ecx, -32768 ; ffff8000H |
||
318 | jge .write |
||
319 | mov ecx, -32768 ; ffff8000H |
||
320 | .write: |
||
321 | mov ebx, ecx |
||
322 | shl ebx, 16 |
||
323 | mov bx, cx |
||
324 | mov [edi], ebx |
||
325 | add edi, 4 |
||
326 | |||
285 | serge | 327 | add eax, [esp+16] |
328 | cmp eax, [esp+24] |
||
168 | serge | 329 | jb .l1 |
330 | |||
331 | mov ebp, esp |
||
332 | |||
333 | sub edi, [dest] |
||
334 | mov eax, edi |
||
335 | ret |
||
336 | endp |
||
337 | |||
338 | align 4 |
||
285 | serge | 339 | proc resample_18 stdcall, dest:dword,src:dword,\ |
168 | serge | 340 | r_dt:dword, r_size:dword,r_end:dword |
341 | |||
342 | |||
343 | mov edi, [dest] |
||
285 | serge | 344 | mov edx, [src] |
345 | sub edx, 32 |
||
346 | |||
168 | serge | 347 | mov esi, 16 |
348 | |||
349 | align 16 |
||
350 | .l1: |
||
351 | mov ecx, esi |
||
352 | mov eax, esi |
||
353 | and ecx, 0x7FFF |
||
354 | shr eax, 15 |
||
355 | lea eax, [edx+eax] |
||
356 | |||
357 | mov bx, word [eax] |
||
358 | sub bh, 0x80 |
||
359 | sub bl, 0x80 |
||
360 | movsx eax, bh |
||
361 | shl eax,8 |
||
362 | movsx ebp, bl |
||
363 | shl ebp,8 |
||
364 | mov ebx, 32768 |
||
365 | imul eax, ecx |
||
366 | sub ebx, ecx |
||
367 | imul ebx, ebp |
||
368 | lea ecx, [ebx+eax+16384] |
||
369 | sar ecx, 15 |
||
370 | cmp ecx, 32767 ; 00007fffH |
||
371 | jle @f |
||
372 | mov ecx, 32767 ; 00007fffH |
||
373 | jmp .write |
||
374 | @@: |
||
375 | cmp ecx, -32768 ; ffff8000H |
||
376 | jge .write |
||
377 | mov ecx, -32768 ; ffff8000H |
||
378 | .write: |
||
379 | mov ebx, ecx |
||
380 | shl ebx, 16 |
||
381 | mov bx, cx |
||
382 | mov [edi], ebx |
||
383 | add edi, 4 |
||
384 | |||
285 | serge | 385 | add esi, [esp+16] |
386 | cmp esi, [esp+24] |
||
168 | serge | 387 | jb .l1 |
388 | |||
389 | mov ebp, esp |
||
390 | sub edi, [dest] |
||
391 | mov eax, edi |
||
392 | ret |
||
393 | endp |
||
394 | |||
395 | align 4 |
||
285 | serge | 396 | proc copy_stream stdcall, dest:dword,src:dword,\ |
168 | serge | 397 | r_dt:dword, r_size:dword,r_end:dword |
398 | |||
281 | serge | 399 | mov ecx, [r_size] |
400 | mov eax, ecx |
||
401 | shr ecx, 2 |
||
402 | mov esi, [src] |
||
403 | mov edi, [dest] |
||
404 | rep movsd |
||
405 | mov eax, 16384 |
||
406 | ret |
||
168 | serge | 407 | endp |
408 | |||
409 | align 4 |
||
285 | serge | 410 | proc resample_2 stdcall, dest:dword,src:dword,\ |
168 | serge | 411 | r_dt:dword, r_size:dword,r_end:dword |
412 | |||
285 | serge | 413 | mov edx, [src] |
414 | sub edx, 32*4 |
||
415 | mov edi, [dest] |
||
416 | mov ebx, [r_dt] |
||
417 | mov eax, 16 |
||
418 | emms |
||
168 | serge | 419 | |
420 | align 16 |
||
421 | .l1: |
||
285 | serge | 422 | mov ecx, eax |
423 | mov esi, eax |
||
424 | and ecx, 0x7FFF |
||
425 | shr esi, 15 |
||
426 | lea esi, [edx+esi*4] |
||
168 | serge | 427 | |
285 | serge | 428 | movq mm0, [esi] |
429 | movq mm1, mm0 |
||
168 | serge | 430 | |
285 | serge | 431 | movd mm2, ecx |
432 | punpcklwd mm2, mm2 |
||
433 | movq mm3, qword [m7] ;0x8000 |
||
168 | serge | 434 | |
285 | serge | 435 | psubw mm3, mm2 ; ;0x8000 - iconst |
436 | punpckldq mm3, mm2 |
||
168 | serge | 437 | |
285 | serge | 438 | pmulhw mm0, mm3 |
439 | pmullw mm1, mm3 |
||
168 | serge | 440 | |
285 | serge | 441 | movq mm4, mm1 |
442 | punpcklwd mm1, mm0 |
||
443 | punpckhwd mm4, mm0 |
||
444 | paddd mm1, mm4 |
||
445 | psrad mm1, 15 |
||
446 | packssdw mm1, mm1 |
||
447 | movd [edi], mm1 |
||
448 | add edi, 4 |
||
168 | serge | 449 | |
285 | serge | 450 | add eax, ebx |
451 | cmp eax, [r_end] |
||
452 | jb .l1 |
||
453 | emms |
||
168 | serge | 454 | |
285 | serge | 455 | sub edi, [dest] |
456 | mov eax, edi |
||
457 | ret |
||
168 | serge | 458 | endp |
459 | |||
460 | align 4 |
||
285 | serge | 461 | proc resample_28 stdcall, dest:dword,src:dword,\ |
168 | serge | 462 | r_dt:dword, r_size:dword,r_end:dword |
463 | |||
285 | serge | 464 | mov edx, [src] |
465 | sub edx, 32*2 |
||
466 | mov edi, [dest] |
||
467 | mov ebx, [r_dt] |
||
468 | mov eax, 16 |
||
469 | emms |
||
470 | movq mm7,[mm80] |
||
471 | movq mm6,[mm_mask] |
||
168 | serge | 472 | |
473 | align 16 |
||
474 | .l1: |
||
475 | mov ecx, eax |
||
476 | mov esi, eax |
||
477 | and ecx, 0x7FFF |
||
478 | shr esi, 15 |
||
479 | lea esi, [edx+esi*2] |
||
480 | |||
481 | movq mm0, [esi] |
||
482 | psubb mm0,mm7 |
||
483 | punpcklbw mm0,mm0 |
||
484 | pand mm0,mm6 |
||
485 | |||
486 | movq mm1, mm0 |
||
487 | |||
488 | movd mm2, ecx |
||
489 | punpcklwd mm2, mm2 |
||
490 | movq mm3, qword [m7] ; // 0x8000 |
||
491 | |||
492 | psubw mm3, mm2 ; // 0x8000 - iconst |
||
493 | punpckldq mm3, mm2 |
||
494 | |||
495 | pmulhw mm0, mm3 |
||
496 | pmullw mm1, mm3 |
||
497 | |||
498 | movq mm4, mm1 |
||
499 | punpcklwd mm1, mm0 |
||
500 | punpckhwd mm4, mm0 |
||
501 | paddd mm1, mm4 |
||
502 | psrad mm1, 15 |
||
503 | packssdw mm1, mm1 |
||
504 | movd [edi], mm1 |
||
505 | add edi, 4 |
||
506 | |||
507 | add eax, ebx |
||
508 | cmp eax, [r_end] |
||
509 | jb .l1 |
||
510 | emms |
||
511 | |||
512 | |||
513 | sub edi, [dest] |
||
514 | mov eax, edi |
||
515 | ret |
||
516 | endp |
||
517 | |||
518 | |||
285 | serge | 519 | proc m16_stereo stdcall, dest:dword,src:dword,\ |
168 | serge | 520 | r_dt:dword, r_size:dword,r_end:dword |
521 | |||
522 | mov esi, [src] |
||
523 | mov edi, [dest] |
||
524 | mov ecx, [r_size] |
||
525 | shr ecx,8 |
||
526 | @@: |
||
527 | call m16_s_mmx |
||
528 | add edi, 128 |
||
529 | add esi, 64 |
||
530 | call m16_s_mmx |
||
531 | add edi, 128 |
||
532 | add esi, 64 |
||
533 | call m16_s_mmx |
||
534 | add edi, 128 |
||
535 | add esi, 64 |
||
536 | call m16_s_mmx |
||
537 | add edi, 128 |
||
538 | add esi, 64 |
||
539 | dec ecx |
||
540 | jnz @b |
||
541 | |||
542 | mov eax, [r_size] |
||
543 | add eax, eax |
||
544 | ret |
||
545 | endp |
||
546 | |||
547 | align 4 |
||
285 | serge | 548 | proc s8_stereo stdcall, dest:dword,src:dword,\ |
168 | serge | 549 | r_dt:dword, r_size:dword,r_end:dword |
550 | |||
281 | serge | 551 | mov esi, [src] |
552 | mov edi, [dest] |
||
553 | mov ecx, [r_size] |
||
554 | shr ecx, 7 |
||
168 | serge | 555 | |
281 | serge | 556 | movq mm7, [mm80] |
557 | movq mm6, [mm_mask] |
||
168 | serge | 558 | @@: |
281 | serge | 559 | call s8_s_mmx |
560 | add edi, 64 |
||
561 | add esi, 32 |
||
562 | call s8_s_mmx |
||
563 | add edi, 64 |
||
564 | add esi, 32 |
||
565 | call s8_s_mmx |
||
566 | add edi, 64 |
||
567 | add esi, 32 |
||
568 | call s8_s_mmx |
||
569 | add edi, 64 |
||
570 | add esi, 32 |
||
168 | serge | 571 | dec ecx |
281 | serge | 572 | jnz @b |
168 | serge | 573 | |
281 | serge | 574 | mov eax, [r_size] |
575 | add eax, eax |
||
576 | ret |
||
168 | serge | 577 | endp |
578 | |||
285 | serge | 579 | proc m8_stereo stdcall, dest:dword,src:dword,\ |
168 | serge | 580 | r_dt:dword, r_size:dword,r_end:dword |
581 | |||
281 | serge | 582 | mov esi, [src] |
583 | mov edi, [dest] |
||
584 | mov ecx, [r_size] |
||
585 | shr ecx, 6 |
||
168 | serge | 586 | |
281 | serge | 587 | movq mm7, [mm80] |
588 | movq mm6, [mm_mask] |
||
168 | serge | 589 | @@: |
281 | serge | 590 | call m8_s_mmx |
591 | add edi, 64 |
||
592 | add esi, 16 |
||
593 | call m8_s_mmx |
||
594 | add edi, 64 |
||
595 | add esi, 16 |
||
596 | call m8_s_mmx |
||
597 | add edi, 64 |
||
598 | add esi, 16 |
||
599 | call m8_s_mmx |
||
600 | add edi, 64 |
||
601 | add esi, 16 |
||
602 | dec ecx |
||
603 | jnz @b |
||
168 | serge | 604 | |
281 | serge | 605 | mov eax, [r_size] |
606 | add eax, eax |
||
607 | add eax, eax |
||
608 | ret |
||
168 | serge | 609 | endp |
610 | |||
611 | align 4 |
||
612 | proc alloc_mix_buff |
||
613 | |||
281 | serge | 614 | bsf eax, [mix_buff_map] |
615 | jnz .find |
||
616 | xor eax, eax |
||
617 | ret |
||
168 | serge | 618 | .find: |
281 | serge | 619 | btr [mix_buff_map], eax |
620 | shl eax, 9 |
||
621 | add eax, [mix_buff] |
||
622 | ret |
||
168 | serge | 623 | endp |
624 | |||
625 | proc m16_s_mmx |
||
626 | |||
627 | movq mm0, [esi] |
||
628 | movq mm1, mm0 |
||
629 | punpcklwd mm0, mm0 |
||
630 | punpckhwd mm1, mm1 |
||
631 | movq [edi], mm0 |
||
632 | movq [edi+8], mm1 |
||
633 | |||
634 | movq mm0, [esi+8] |
||
635 | movq mm1, mm0 |
||
636 | punpcklwd mm0, mm0 |
||
637 | punpckhwd mm1, mm1 |
||
638 | movq [edi+16], mm0 |
||
639 | movq [edi+24], mm1 |
||
640 | |||
641 | movq mm0, [esi+16] |
||
642 | movq mm1, mm0 |
||
643 | punpcklwd mm0, mm0 |
||
644 | punpckhwd mm1, mm1 |
||
645 | movq [edi+32], mm0 |
||
646 | movq [edi+40], mm1 |
||
647 | |||
648 | movq mm0, [esi+24] |
||
649 | movq mm1, mm0 |
||
650 | punpcklwd mm0, mm0 |
||
651 | punpckhwd mm1, mm1 |
||
652 | movq [edi+48], mm0 |
||
653 | movq [edi+56], mm1 |
||
654 | |||
655 | movq mm0, [esi+32] |
||
656 | movq mm1, mm0 |
||
657 | punpcklwd mm0, mm0 |
||
658 | punpckhwd mm1, mm1 |
||
659 | movq [edi+64], mm0 |
||
660 | movq [edi+72], mm1 |
||
661 | |||
662 | movq mm0, [esi+40] |
||
663 | movq mm1, mm0 |
||
664 | punpcklwd mm0, mm0 |
||
665 | punpckhwd mm1, mm1 |
||
666 | movq [edi+80], mm0 |
||
667 | movq [edi+88], mm1 |
||
668 | |||
669 | |||
670 | movq mm0, [esi+48] |
||
671 | movq mm1, mm0 |
||
672 | punpcklwd mm0, mm0 |
||
673 | punpckhwd mm1, mm1 |
||
674 | movq [edi+96], mm0 |
||
675 | movq [edi+104], mm1 |
||
676 | |||
677 | movq mm0, [esi+56] |
||
678 | movq mm1, mm0 |
||
679 | punpcklwd mm0, mm0 |
||
680 | punpckhwd mm1, mm1 |
||
681 | movq [edi+112], mm0 |
||
682 | movq [edi+120], mm1 |
||
683 | |||
684 | ret |
||
685 | endp |
||
686 | |||
687 | align 4 |
||
688 | proc s8_s_mmx |
||
689 | |||
281 | serge | 690 | movq mm0, [esi] |
691 | psubb mm0, mm7 |
||
692 | movq mm1, mm0 |
||
693 | punpcklbw mm0, mm0 |
||
694 | pand mm0, mm6 |
||
695 | punpckhbw mm1, mm1 |
||
696 | pand mm1, mm6 |
||
697 | movq [edi], mm0 |
||
698 | movq [edi+8], mm1 |
||
168 | serge | 699 | |
281 | serge | 700 | movq mm0, [esi+8] |
701 | psubb mm0, mm7 |
||
702 | movq mm1, mm0 |
||
703 | punpcklbw mm0, mm0 |
||
704 | pand mm0, mm6 |
||
705 | punpckhbw mm1, mm1 |
||
706 | pand mm1, mm6 |
||
707 | movq [edi+16], mm0 |
||
708 | movq [edi+24], mm1 |
||
168 | serge | 709 | |
281 | serge | 710 | movq mm0, [esi+16] |
711 | psubb mm0, mm7 |
||
712 | movq mm1, mm0 |
||
713 | punpcklbw mm0, mm0 |
||
714 | pand mm0, mm6 |
||
715 | punpckhbw mm1, mm1 |
||
716 | pand mm1, mm6 |
||
717 | movq [edi+32], mm0 |
||
718 | movq [edi+40], mm1 |
||
168 | serge | 719 | |
281 | serge | 720 | movq mm0, [esi+24] |
721 | psubb mm0, mm7 |
||
722 | movq mm1, mm0 |
||
723 | punpcklbw mm0, mm0 |
||
724 | pand mm0, mm6 |
||
725 | punpckhbw mm1, mm1 |
||
726 | pand mm1, mm6 |
||
727 | movq [edi+48], mm0 |
||
728 | movq [edi+56], mm1 |
||
168 | serge | 729 | |
281 | serge | 730 | ret |
168 | serge | 731 | |
732 | endp |
||
733 | |||
734 | align 4 |
||
735 | proc m8_s_mmx |
||
736 | |||
281 | serge | 737 | movq mm0, [esi] |
738 | psubb mm0, mm7 |
||
739 | movq mm1, mm0 |
||
740 | punpcklbw mm0, mm0 |
||
741 | pand mm0, mm6 |
||
742 | punpckhbw mm1, mm1 |
||
743 | pand mm1, mm6 |
||
744 | movq mm2, mm0 |
||
745 | punpcklwd mm0, mm0 |
||
746 | punpckhwd mm2, mm2 |
||
168 | serge | 747 | |
281 | serge | 748 | movq mm3, mm1 |
749 | punpcklwd mm1, mm1 |
||
750 | punpckhwd mm3, mm3 |
||
168 | serge | 751 | |
281 | serge | 752 | movq [edi], mm0 |
753 | movq [edi+8], mm2 |
||
754 | movq [edi+16], mm1 |
||
755 | movq [edi+24], mm3 |
||
168 | serge | 756 | |
281 | serge | 757 | movq mm0, [esi+8] |
758 | psubb mm0, mm7 |
||
759 | movq mm1, mm0 |
||
760 | punpcklbw mm0, mm0 |
||
761 | pand mm0, mm6 |
||
762 | punpckhbw mm1, mm1 |
||
763 | pand mm1, mm6 |
||
764 | movq mm2, mm0 |
||
765 | punpcklwd mm0, mm0 |
||
766 | punpckhwd mm2, mm2 |
||
168 | serge | 767 | |
281 | serge | 768 | movq mm3, mm1 |
769 | punpcklwd mm1, mm1 |
||
770 | punpckhwd mm3, mm3 |
||
168 | serge | 771 | |
281 | serge | 772 | movq [edi+32], mm0 |
773 | movq [edi+40], mm2 |
||
774 | movq [edi+48], mm1 |
||
775 | movq [edi+56], mm3 |
||
168 | serge | 776 | |
281 | serge | 777 | ret |
168 | serge | 778 | endp |
779 | |||
780 | |||
781 | align 4 |
||
782 | proc mix_2_1 stdcall, output:dword, str0:dword, str1:dword |
||
783 | |||
281 | serge | 784 | mov edi, [output] |
168 | serge | 785 | |
281 | serge | 786 | stdcall mix_2_1_mmx, edi, [str0],[str1] |
787 | add edi, 128 |
||
788 | add [str0], 128 |
||
789 | add [str1], 128 |
||
790 | stdcall mix_2_1_mmx, edi, [str0],[str1] |
||
791 | add edi, 128 |
||
792 | add [str0], 128 |
||
793 | add [str1], 128 |
||
794 | stdcall mix_2_1_mmx, edi, [str0],[str1] |
||
795 | add edi, 128 |
||
796 | add [str0], 128 |
||
797 | add [str1], 128 |
||
798 | stdcall mix_2_1_mmx, edi, [str0],[str1] |
||
168 | serge | 799 | |
281 | serge | 800 | ret |
168 | serge | 801 | endp |
802 | |||
803 | |||
804 | align 4 |
||
805 | proc mix_3_1 stdcall, output:dword, str0:dword, str1:dword, str2:dword |
||
806 | |||
807 | mov edi, [output] |
||
808 | |||
809 | stdcall mix_3_1_mmx, edi, [str0],[str1],[str2] |
||
810 | add edi, 128 |
||
811 | add [str0], 128 |
||
812 | add [str1], 128 |
||
813 | add [str2], 128 |
||
814 | stdcall mix_3_1_mmx, edi, [str0],[str1],[str2] |
||
815 | add edi, 128 |
||
816 | add [str0], 128 |
||
817 | add [str1], 128 |
||
818 | add [str2], 128 |
||
819 | stdcall mix_3_1_mmx, edi, [str0],[str1],[str2] |
||
820 | add edi, 128 |
||
821 | add [str0], 128 |
||
822 | add [str1], 128 |
||
823 | add [str2], 128 |
||
824 | stdcall mix_3_1_mmx, edi, [str0],[str1],[str2] |
||
825 | |||
826 | ret |
||
827 | endp |
||
828 | |||
829 | align 4 |
||
830 | proc mix_4_1 stdcall, str0:dword, str1:dword,\ |
||
831 | str2:dword, str3:dword |
||
832 | |||
833 | local output:DWORD |
||
834 | |||
835 | call alloc_mix_buff |
||
836 | and eax, eax |
||
837 | jz .err |
||
838 | mov [output], eax |
||
839 | |||
840 | mov edi, eax |
||
841 | |||
842 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
843 | add edi, 128 |
||
844 | add [str0], 128 |
||
845 | add [str1], 128 |
||
846 | add [str2], 128 |
||
847 | add [str3], 128 |
||
848 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
849 | add edi, 128 |
||
850 | add [str0], 128 |
||
851 | add [str1], 128 |
||
852 | add [str2], 128 |
||
853 | add [str3], 128 |
||
854 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
855 | add edi, 128 |
||
856 | add [str0], 128 |
||
857 | add [str1], 128 |
||
858 | add [str2], 128 |
||
859 | add [str3], 128 |
||
860 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
861 | mov eax, [output] |
||
862 | ret |
||
863 | .err: |
||
864 | xor eax, eax |
||
865 | ret |
||
866 | endp |
||
867 | |||
868 | |||
869 | align 4 |
||
870 | proc final_mix stdcall, output:dword, str0:dword, str1:dword,\ |
||
871 | str2:dword, str3:dword |
||
872 | |||
873 | mov edi, [output] |
||
874 | |||
875 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
876 | add edi, 128 |
||
877 | add [str0], 128 |
||
878 | add [str1], 128 |
||
879 | add [str2], 128 |
||
880 | add [str3], 128 |
||
881 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
882 | add edi, 128 |
||
883 | add [str0], 128 |
||
884 | add [str1], 128 |
||
885 | add [str2], 128 |
||
886 | add [str3], 128 |
||
887 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
888 | add edi, 128 |
||
889 | add [str0], 128 |
||
890 | add [str1], 128 |
||
891 | add [str2], 128 |
||
892 | add [str3], 128 |
||
893 | stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3] |
||
894 | |||
895 | ret |
||
896 | endp |
||
897 | |||
898 | align 4 |
||
899 | proc mix_2_1_mmx stdcall, output:dword, str0:dword, str1:dword |
||
900 | |||
281 | serge | 901 | mov edx, [output] |
902 | mov eax, [str0] |
||
903 | mov ecx, [str1] |
||
168 | serge | 904 | |
281 | serge | 905 | movq mm0, [eax] |
906 | paddsw mm0, [ecx] |
||
907 | ; psraw mm0, 1 |
||
908 | movq [edx], mm0 |
||
168 | serge | 909 | |
281 | serge | 910 | movq mm1, [eax+8] |
911 | paddsw mm1,[ecx+8] |
||
912 | ; psraw mm1, 1 |
||
913 | movq [edx+8], mm1 |
||
168 | serge | 914 | |
281 | serge | 915 | movq mm2, [eax+16] |
916 | paddsw mm2, [ecx+16] |
||
917 | ; psraw mm2, 1 |
||
918 | movq [edx+16], mm2 |
||
168 | serge | 919 | |
281 | serge | 920 | movq mm3, [eax+24] |
921 | paddsw mm3, [ecx+24] |
||
922 | ; psraw mm3, 1 |
||
923 | movq [edx+24], mm3 |
||
168 | serge | 924 | |
281 | serge | 925 | movq mm0, [eax+32] |
926 | paddsw mm0, [ecx+32] |
||
927 | ; psraw mm0, 1 |
||
928 | movq [edx+32], mm0 |
||
168 | serge | 929 | |
281 | serge | 930 | movq mm1, [eax+40] |
931 | paddsw mm1, [ecx+40] |
||
932 | ; psraw mm1, 1 |
||
933 | movq [edx+40], mm1 |
||
168 | serge | 934 | |
281 | serge | 935 | movq mm2, [eax+48] |
936 | paddsw mm2, [ecx+48] |
||
937 | ; psraw mm2, 1 |
||
938 | movq [edx+48], mm2 |
||
168 | serge | 939 | |
281 | serge | 940 | movq mm3, [eax+56] |
941 | paddsw mm3, [ecx+56] |
||
942 | ; psraw mm3, 1 |
||
943 | movq [edx+56], mm3 |
||
168 | serge | 944 | |
281 | serge | 945 | movq mm0, [eax+64] |
946 | paddsw mm0, [ecx+64] |
||
947 | ; psraw mm0, 1 |
||
948 | movq [edx+64], mm0 |
||
168 | serge | 949 | |
281 | serge | 950 | movq mm1, [eax+72] |
951 | paddsw mm1, [ecx+72] |
||
952 | ; psraw mm1, 1 |
||
953 | movq [edx+72], mm1 |
||
168 | serge | 954 | |
281 | serge | 955 | movq mm2, [eax+80] |
956 | paddsw mm2, [ecx+80] |
||
957 | ; psraw mm2, 1 |
||
958 | movq [edx+80], mm2 |
||
168 | serge | 959 | |
281 | serge | 960 | movq mm3, [eax+88] |
961 | paddsw mm3, [ecx+88] |
||
962 | ; psraw mm3, 1 |
||
963 | movq [edx+88], mm3 |
||
168 | serge | 964 | |
281 | serge | 965 | movq mm0, [eax+96] |
966 | paddsw mm0, [ecx+96] |
||
967 | ; psraw mm0, 1 |
||
968 | movq [edx+96], mm0 |
||
168 | serge | 969 | |
281 | serge | 970 | movq mm1, [eax+104] |
971 | paddsw mm1, [ecx+104] |
||
972 | ; psraw mm1, 1 |
||
973 | movq [edx+104], mm1 |
||
168 | serge | 974 | |
281 | serge | 975 | movq mm2, [eax+112] |
976 | paddsw mm2, [ecx+112] |
||
977 | ; psraw mm2, 1 |
||
978 | movq [edx+112], mm2 |
||
168 | serge | 979 | |
281 | serge | 980 | movq mm3, [eax+120] |
981 | paddsw mm3, [ecx+120] |
||
982 | ; psraw mm3, 1 |
||
983 | movq [edx+120], mm3 |
||
168 | serge | 984 | |
281 | serge | 985 | ret |
168 | serge | 986 | endp |
987 | |||
988 | align 4 |
||
989 | proc mix_3_1_mmx stdcall, output:dword, str0:dword, str1:dword, str2:dword |
||
990 | |||
281 | serge | 991 | mov edx, [output] |
992 | mov eax, [str0] |
||
993 | mov ebx, [str1] |
||
994 | mov ecx, [str2] |
||
168 | serge | 995 | |
281 | serge | 996 | movq mm0, [eax] |
997 | paddsw mm0, [ebx] |
||
998 | paddsw mm0, [ecx] |
||
999 | movq [edx], mm0 |
||
168 | serge | 1000 | |
281 | serge | 1001 | movq mm1, [eax+8] |
1002 | paddsw mm1,[ebx+8] |
||
1003 | paddsw mm1,[ecx+8] |
||
1004 | movq [edx+8], mm1 |
||
168 | serge | 1005 | |
281 | serge | 1006 | movq mm2, [eax+16] |
1007 | paddsw mm2, [ebx+16] |
||
1008 | paddsw mm2, [ecx+16] |
||
1009 | movq [edx+16], mm2 |
||
168 | serge | 1010 | |
281 | serge | 1011 | movq mm3, [eax+24] |
1012 | paddsw mm3, [ebx+24] |
||
1013 | paddsw mm3, [ecx+24] |
||
1014 | movq [edx+24], mm3 |
||
168 | serge | 1015 | |
281 | serge | 1016 | movq mm0, [eax+32] |
1017 | paddsw mm0, [ebx+32] |
||
1018 | paddsw mm0, [ecx+32] |
||
1019 | movq [edx+32], mm0 |
||
168 | serge | 1020 | |
281 | serge | 1021 | movq mm1, [eax+40] |
1022 | paddsw mm1, [ebx+40] |
||
1023 | paddsw mm1, [ecx+40] |
||
1024 | movq [edx+40], mm1 |
||
168 | serge | 1025 | |
281 | serge | 1026 | movq mm2, [eax+48] |
1027 | paddsw mm2, [ebx+48] |
||
1028 | paddsw mm2, [ecx+48] |
||
1029 | movq [edx+48], mm2 |
||
168 | serge | 1030 | |
281 | serge | 1031 | movq mm3, [eax+56] |
1032 | paddsw mm3, [ebx+56] |
||
1033 | paddsw mm3, [ecx+56] |
||
1034 | movq [edx+56], mm3 |
||
168 | serge | 1035 | |
281 | serge | 1036 | movq mm0, [eax+64] |
1037 | paddsw mm0, [ebx+64] |
||
1038 | paddsw mm0, [ecx+64] |
||
1039 | movq [edx+64], mm0 |
||
168 | serge | 1040 | |
281 | serge | 1041 | movq mm1, [eax+72] |
1042 | paddsw mm1, [ebx+72] |
||
1043 | paddsw mm1, [ecx+72] |
||
1044 | movq [edx+72], mm1 |
||
168 | serge | 1045 | |
281 | serge | 1046 | movq mm2, [eax+80] |
1047 | paddsw mm2, [ebx+80] |
||
1048 | paddsw mm2, [ecx+80] |
||
1049 | movq [edx+80], mm2 |
||
168 | serge | 1050 | |
281 | serge | 1051 | movq mm3, [eax+88] |
1052 | paddsw mm3, [ebx+88] |
||
1053 | paddsw mm3, [ecx+88] |
||
1054 | movq [edx+88], mm3 |
||
168 | serge | 1055 | |
281 | serge | 1056 | movq mm0, [eax+96] |
1057 | paddsw mm0, [ebx+96] |
||
1058 | paddsw mm0, [ecx+96] |
||
1059 | movq [edx+96], mm0 |
||
168 | serge | 1060 | |
281 | serge | 1061 | movq mm1, [eax+104] |
1062 | paddsw mm1, [ebx+104] |
||
1063 | paddsw mm1, [ecx+104] |
||
1064 | movq [edx+104], mm1 |
||
168 | serge | 1065 | |
281 | serge | 1066 | movq mm2, [eax+112] |
1067 | paddsw mm2, [ebx+112] |
||
1068 | paddsw mm2, [ecx+112] |
||
1069 | movq [edx+112], mm2 |
||
168 | serge | 1070 | |
281 | serge | 1071 | movq mm3, [eax+120] |
1072 | paddsw mm3, [ebx+120] |
||
1073 | paddsw mm3, [ecx+120] |
||
1074 | movq [edx+120], mm3 |
||
168 | serge | 1075 | |
281 | serge | 1076 | ret |
168 | serge | 1077 | endp |
1078 | |||
1079 | align 4 |
||
1080 | proc mix_4_1_mmx stdcall, output:dword, str0:dword, str1:dword,\ |
||
281 | serge | 1081 | str2:dword, str3:dword |
168 | serge | 1082 | |
281 | serge | 1083 | mov edx, [output] |
1084 | mov esi, [str0] |
||
1085 | mov eax, [str1] |
||
1086 | mov ebx, [str2] |
||
1087 | mov ecx, [str3] |
||
168 | serge | 1088 | |
281 | serge | 1089 | movq mm0, [esi] |
1090 | movq mm1, [eax] |
||
1091 | paddsw mm0, [ebx] |
||
1092 | paddsw mm1, [ecx] |
||
1093 | paddsw mm0, mm1 |
||
1094 | movq [edx], mm0 |
||
168 | serge | 1095 | |
281 | serge | 1096 | movq mm2, [esi+8] |
1097 | movq mm3, [eax+8] |
||
1098 | paddsw mm2, [ebx+8] |
||
1099 | paddsw mm3, [ecx+8] |
||
1100 | paddsw mm2, mm3 |
||
1101 | movq [edx+8], mm2 |
||
168 | serge | 1102 | |
281 | serge | 1103 | movq mm0, [esi+16] |
1104 | movq mm1, [eax+16] |
||
1105 | paddsw mm0, [ebx+16] |
||
1106 | paddsw mm1, [ecx+16] |
||
1107 | paddsw mm0, mm1 |
||
1108 | movq [edx+16], mm0 |
||
168 | serge | 1109 | |
281 | serge | 1110 | movq mm2, [esi+24] |
1111 | movq mm3, [eax+24] |
||
1112 | paddsw mm2, [ebx+24] |
||
1113 | paddsw mm3, [ecx+24] |
||
1114 | paddsw mm2, mm3 |
||
1115 | movq [edx+24], mm2 |
||
168 | serge | 1116 | |
281 | serge | 1117 | movq mm0, [esi+32] |
1118 | movq mm1, [eax+32] |
||
1119 | paddsw mm0, [ebx+32] |
||
1120 | paddsw mm1, [ecx+32] |
||
1121 | paddsw mm0, mm1 |
||
1122 | movq [edx+32], mm0 |
||
168 | serge | 1123 | |
281 | serge | 1124 | movq mm2, [esi+40] |
1125 | movq mm3, [eax+40] |
||
1126 | paddsw mm2, [ebx+40] |
||
1127 | paddsw mm3, [ecx+40] |
||
1128 | paddsw mm2, mm3 |
||
1129 | movq [edx+40], mm2 |
||
168 | serge | 1130 | |
281 | serge | 1131 | movq mm0, [esi+48] |
1132 | movq mm1, [eax+48] |
||
1133 | paddsw mm0, [ebx+48] |
||
1134 | paddsw mm1, [ecx+48] |
||
1135 | paddsw mm0, mm1 |
||
1136 | movq [edx+48], mm0 |
||
168 | serge | 1137 | |
281 | serge | 1138 | movq mm2, [esi+56] |
1139 | movq mm3, [eax+56] |
||
1140 | paddsw mm2, [ebx+56] |
||
1141 | paddsw mm3, [ecx+56] |
||
1142 | paddsw mm2, mm3 |
||
1143 | movq [edx+56], mm2 |
||
168 | serge | 1144 | |
281 | serge | 1145 | movq mm0, [esi+64] |
1146 | movq mm1, [eax+64] |
||
1147 | paddsw mm0, [ebx+64] |
||
1148 | paddsw mm1, [ecx+64] |
||
1149 | paddsw mm0, mm1 |
||
1150 | movq [edx+64], mm0 |
||
168 | serge | 1151 | |
281 | serge | 1152 | movq mm2, [esi+72] |
1153 | movq mm3, [eax+72] |
||
1154 | paddsw mm2, [ebx+72] |
||
1155 | paddsw mm3, [ecx+72] |
||
1156 | paddsw mm2, mm3 |
||
1157 | movq [edx+72], mm2 |
||
168 | serge | 1158 | |
281 | serge | 1159 | movq mm2, [esi+80] |
1160 | movq mm3, [eax+80] |
||
1161 | paddsw mm2, [ebx+80] |
||
1162 | paddsw mm3, [ecx+80] |
||
1163 | paddsw mm2, mm3 |
||
1164 | movq [edx+80], mm2 |
||
168 | serge | 1165 | |
281 | serge | 1166 | movq mm2, [esi+88] |
1167 | movq mm3, [eax+88] |
||
1168 | paddsw mm2, [ebx+88] |
||
1169 | paddsw mm3, [ecx+88] |
||
1170 | paddsw mm2, mm3 |
||
1171 | movq [edx+88], mm2 |
||
168 | serge | 1172 | |
281 | serge | 1173 | movq mm2, [esi+96] |
1174 | movq mm3, [eax+96] |
||
1175 | paddsw mm2, [ebx+96] |
||
1176 | paddsw mm3, [ecx+96] |
||
1177 | paddsw mm2, mm3 |
||
1178 | movq [edx+96], mm2 |
||
168 | serge | 1179 | |
281 | serge | 1180 | movq mm2, [esi+104] |
1181 | movq mm3, [eax+104] |
||
1182 | paddsw mm2, [ebx+104] |
||
1183 | paddsw mm3, [ecx+104] |
||
1184 | paddsw mm2, mm3 |
||
1185 | movq [edx+104], mm2 |
||
168 | serge | 1186 | |
281 | serge | 1187 | movq mm2, [esi+112] |
1188 | movq mm3, [eax+112] |
||
1189 | paddsw mm2, [ebx+112] |
||
1190 | paddsw mm3, [ecx+112] |
||
1191 | paddsw mm2, mm3 |
||
1192 | movq [edx+112], mm2 |
||
168 | serge | 1193 | |
281 | serge | 1194 | movq mm2, [esi+120] |
1195 | movq mm3, [eax+120] |
||
1196 | paddsw mm2, [ebx+120] |
||
1197 | paddsw mm3, [ecx+120] |
||
1198 | paddsw mm2, mm3 |
||
1199 | movq [edx+120], mm2 |
||
168 | serge | 1200 | |
281 | serge | 1201 | ret |
168 | serge | 1202 | endp |
1203 | |||
1204 | align 4 |
||
1205 | proc copy_mem stdcall, output:dword, input:dword |
||
1206 | |||
1207 | mov edi, [output] |
||
1208 | mov esi, [input] |
||
1209 | mov ecx, 0x80 |
||
1210 | .l1: |
||
1211 | mov eax, [esi] |
||
1212 | mov [edi], eax |
||
1213 | add esi, 4 |
||
1214 | add edi, 4 |
||
1215 | loop .l1 |
||
1216 | |||
1217 | ret |
||
1218 | endp |
||
1219 | |||
1220 | proc memcpy |
||
1221 | @@: |
||
1222 | mov eax, [esi] |
||
1223 | mov [edi], eax |
||
1224 | add esi, 4 |
||
1225 | add edi, 4 |
||
1226 | dec ecx |
||
1227 | jnz @B |
||
1228 | ret |
||
1229 | endp |
||
1230 |