1 |
;/************************************************************************** |
;/**************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * mmx sum of absolute difference |
; * - K7 optimized SAD operators - |
5 |
; * |
; * |
6 |
; * This program is free software; you can redistribute it and/or modify |
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
; * it under the terms of the GNU General Public License as published by |
; * 2002 Pascal Massimino <skal@planet-d.net> |
8 |
|
; * |
9 |
|
; * This program is free software; you can redistribute it and/or modify it |
10 |
|
; * under the terms of the GNU General Public License as published by |
11 |
; * the Free Software Foundation; either version 2 of the License, or |
; * the Free Software Foundation; either version 2 of the License, or |
12 |
; * (at your option) any later version. |
; * (at your option) any later version. |
13 |
; * |
; * |
18 |
; * |
; * |
19 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
20 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
21 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
22 |
; * |
; * |
23 |
; * History: |
; * $Id: sad_mmx.asm,v 1.11.2.2 2003-11-03 15:51:50 edgomez Exp $ |
24 |
; * |
; * |
25 |
; * 23.07.2002 sad[16,8]bi_xmm; <pross@xvid.org> |
; ***************************************************************************/ |
|
; * 04.06.2002 cleanup -Skal- |
|
|
; * 12.11.2001 inital version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
26 |
|
|
27 |
bits 32 |
BITS 32 |
28 |
|
|
29 |
%macro cglobal 1 |
%macro cglobal 1 |
30 |
|
%ifdef PREFIX |
31 |
global _%1 |
global _%1 |
32 |
%define %1 _%1 |
%define %1 _%1 |
33 |
%else |
%else |
35 |
%endif |
%endif |
36 |
%endmacro |
%endmacro |
37 |
|
|
38 |
section .data |
;============================================================================= |
39 |
|
; Read only data |
40 |
align 16 |
;============================================================================= |
|
mmx_one times 4 dw 1 |
|
41 |
|
|
42 |
section .text |
%ifdef FORMAT_COFF |
43 |
|
SECTION .rodata data |
44 |
cglobal sad16_mmx |
%else |
45 |
cglobal sad8_mmx |
SECTION .rodata data align=16 |
46 |
cglobal sad16bi_mmx |
%endif |
|
cglobal sad8bi_mmx |
|
|
cglobal dev16_mmx |
|
47 |
|
|
48 |
;=========================================================================== |
ALIGN 16 |
49 |
; |
mmx_one: |
50 |
; uint32_t sad16_mmx(const uint8_t * const cur, |
times 4 dw 1 |
51 |
; const uint8_t * const ref, |
|
52 |
; const uint32_t stride, |
;============================================================================= |
53 |
; (early termination ignore; slows this down) |
; Helper macros |
54 |
; |
;============================================================================= |
|
;=========================================================================== |
|
55 |
|
|
56 |
%macro SAD_16x16_MMX 0 |
%macro SAD_16x16_MMX 0 |
57 |
movq mm0, [eax] |
movq mm0, [eax] |
86 |
paddusw mm6,mm2 |
paddusw mm6,mm2 |
87 |
%endmacro |
%endmacro |
88 |
|
|
|
align 16 |
|
|
sad16_mmx: |
|
|
|
|
|
mov eax, [esp+ 4] ; Src1 |
|
|
mov edx, [esp+ 8] ; Src2 |
|
|
mov ecx, [esp+12] ; Stride |
|
|
|
|
|
pxor mm6, mm6 ; accum |
|
|
pxor mm7, mm7 ; zero |
|
|
|
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
|
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
SAD_16x16_MMX |
|
|
|
|
|
pmaddwd mm6, [mmx_one] ; collapse |
|
|
movq mm7, mm6 |
|
|
psrlq mm7, 32 |
|
|
paddd mm6, mm7 |
|
|
|
|
|
movd eax, mm6 |
|
|
|
|
|
ret |
|
|
|
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
|
; uint32_t sad8_mmx(const uint8_t * const cur, |
|
|
; const uint8_t * const ref, |
|
|
; const uint32_t stride); |
|
|
; |
|
|
;=========================================================================== |
|
|
|
|
89 |
%macro SAD_8x8_MMX 0 |
%macro SAD_8x8_MMX 0 |
90 |
movq mm0, [eax] |
movq mm0, [eax] |
91 |
movq mm1, [edx] |
movq mm1, [edx] |
120 |
paddusw mm6,mm2 |
paddusw mm6,mm2 |
121 |
%endmacro |
%endmacro |
122 |
|
|
123 |
align 16 |
%macro SADV_16x16_MMX 0 |
124 |
sad8_mmx: |
movq mm0, [eax] |
125 |
|
movq mm1, [edx] |
126 |
|
|
127 |
mov eax, [esp+ 4] ; Src1 |
movq mm2, [eax+8] |
128 |
mov edx, [esp+ 8] ; Src2 |
movq mm3, [edx+8] |
|
mov ecx, [esp+12] ; Stride |
|
129 |
|
|
130 |
pxor mm6, mm6 ; accum |
movq mm4, mm0 |
131 |
pxor mm7, mm7 ; zero |
psubusb mm0, mm1 |
132 |
|
|
133 |
SAD_8x8_MMX |
psubusb mm1, mm4 |
134 |
SAD_8x8_MMX |
por mm0, mm1 |
135 |
SAD_8x8_MMX |
lea eax,[eax+ecx] |
|
SAD_8x8_MMX |
|
136 |
|
|
137 |
pmaddwd mm6, [mmx_one] ; collapse |
movq mm4, mm2 |
138 |
movq mm7, mm6 |
psubusb mm2, mm3 |
|
psrlq mm7, 32 |
|
|
paddd mm6, mm7 |
|
139 |
|
|
140 |
movd eax, mm6 |
psubusb mm3, mm4 |
141 |
|
por mm2, mm3 |
142 |
|
lea edx,[edx+ecx] |
143 |
|
|
144 |
ret |
movq mm1,mm0 |
145 |
|
movq mm3,mm2 |
146 |
|
|
147 |
|
punpcklbw mm0,mm7 |
148 |
|
punpckhbw mm1,mm7 |
149 |
|
punpcklbw mm2,mm7 |
150 |
|
punpckhbw mm3,mm7 |
151 |
|
|
152 |
|
paddusw mm0,mm1 |
153 |
|
paddusw mm2,mm3 |
154 |
|
|
155 |
|
paddusw mm5, mm0 |
156 |
|
paddusw mm6, mm2 |
157 |
|
%endmacro |
158 |
|
|
159 |
|
%macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
160 |
|
|
161 |
|
movq mm0, [edx+%1] |
162 |
|
movq mm2, [ebx+%1] |
163 |
|
movq mm1, mm0 |
164 |
|
movq mm3, mm2 |
165 |
|
|
166 |
;=========================================================================== |
%if %2 != 0 |
167 |
; |
add edx, ecx |
168 |
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
%endif |
169 |
; const uint8_t * const ref1, |
|
170 |
|
punpcklbw mm0, mm7 |
171 |
|
punpckhbw mm1, mm7 |
172 |
|
punpcklbw mm2, mm7 |
173 |
|
punpckhbw mm3, mm7 |
174 |
|
|
175 |
|
%if %2 != 0 |
176 |
|
add ebx, ecx |
177 |
|
%endif |
178 |
|
|
179 |
|
paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
180 |
|
paddusw mm1, mm3 |
181 |
|
paddusw mm0, [mmx_one] ; mm01 += 1 |
182 |
|
paddusw mm1, [mmx_one] |
183 |
|
psrlw mm0, 1 ; mm01 >>= 1 |
184 |
|
psrlw mm1, 1 |
185 |
|
|
186 |
|
movq mm2, [eax+%1] |
187 |
|
movq mm3, mm2 |
188 |
|
punpcklbw mm2, mm7 ; mm23 = src |
189 |
|
punpckhbw mm3, mm7 |
190 |
|
|
191 |
|
%if %2 != 0 |
192 |
|
add eax, ecx |
193 |
|
%endif |
194 |
|
|
195 |
|
movq mm4, mm0 |
196 |
|
movq mm5, mm1 |
197 |
|
psubusw mm0, mm2 |
198 |
|
psubusw mm1, mm3 |
199 |
|
psubusw mm2, mm4 |
200 |
|
psubusw mm3, mm5 |
201 |
|
por mm0, mm2 ; mm01 = ABS(mm01 - mm23) |
202 |
|
por mm1, mm3 |
203 |
|
|
204 |
|
paddusw mm6, mm0 ; mm6 += mm01 |
205 |
|
paddusw mm6, mm1 |
206 |
|
|
207 |
|
%endmacro |
208 |
|
|
209 |
|
%macro MEAN_16x16_MMX 0 |
210 |
movq mm0, [eax] |
movq mm0, [eax] |
211 |
movq mm2, [eax+8] |
movq mm2, [eax+8] |
212 |
lea eax,[eax+ecx] |
lea eax,[eax+ecx] |
257 |
paddw mm5, mm2 |
paddw mm5, mm2 |
258 |
%endmacro |
%endmacro |
259 |
|
|
260 |
align 16 |
;============================================================================= |
261 |
|
; Code |
262 |
|
;============================================================================= |
263 |
|
|
264 |
|
SECTION .text |
265 |
|
|
266 |
|
cglobal sad16_mmx |
267 |
|
cglobal sad16v_mmx |
268 |
|
cglobal sad8_mmx |
269 |
|
cglobal sad16bi_mmx |
270 |
|
cglobal sad8bi_mmx |
271 |
|
cglobal dev16_mmx |
272 |
|
|
273 |
|
;----------------------------------------------------------------------------- |
274 |
|
; |
275 |
|
; uint32_t sad16_mmx(const uint8_t * const cur, |
276 |
|
; const uint8_t * const ref, |
277 |
|
; const uint32_t stride, |
278 |
|
; const uint32_t best_sad); |
279 |
|
; |
280 |
|
; (early termination ignore; slows this down) |
281 |
|
; |
282 |
|
;----------------------------------------------------------------------------- |
283 |
|
|
284 |
|
ALIGN 16 |
285 |
|
sad16_mmx: |
286 |
|
|
287 |
|
mov eax, [esp+ 4] ; Src1 |
288 |
|
mov edx, [esp+ 8] ; Src2 |
289 |
|
mov ecx, [esp+12] ; Stride |
290 |
|
|
291 |
|
pxor mm6, mm6 ; accum |
292 |
|
pxor mm7, mm7 ; zero |
293 |
|
|
294 |
|
SAD_16x16_MMX |
295 |
|
SAD_16x16_MMX |
296 |
|
SAD_16x16_MMX |
297 |
|
SAD_16x16_MMX |
298 |
|
SAD_16x16_MMX |
299 |
|
SAD_16x16_MMX |
300 |
|
SAD_16x16_MMX |
301 |
|
SAD_16x16_MMX |
302 |
|
|
303 |
|
SAD_16x16_MMX |
304 |
|
SAD_16x16_MMX |
305 |
|
SAD_16x16_MMX |
306 |
|
SAD_16x16_MMX |
307 |
|
SAD_16x16_MMX |
308 |
|
SAD_16x16_MMX |
309 |
|
SAD_16x16_MMX |
310 |
|
SAD_16x16_MMX |
311 |
|
|
312 |
|
pmaddwd mm6, [mmx_one] ; collapse |
313 |
|
movq mm7, mm6 |
314 |
|
psrlq mm7, 32 |
315 |
|
paddd mm6, mm7 |
316 |
|
|
317 |
|
movd eax, mm6 |
318 |
|
|
319 |
|
ret |
320 |
|
|
321 |
|
;----------------------------------------------------------------------------- |
322 |
|
; |
323 |
|
; uint32_t sad8_mmx(const uint8_t * const cur, |
324 |
|
; const uint8_t * const ref, |
325 |
|
; const uint32_t stride); |
326 |
|
; |
327 |
|
;----------------------------------------------------------------------------- |
328 |
|
|
329 |
|
ALIGN 16 |
330 |
|
sad8_mmx: |
331 |
|
|
332 |
|
mov eax, [esp+ 4] ; Src1 |
333 |
|
mov edx, [esp+ 8] ; Src2 |
334 |
|
mov ecx, [esp+12] ; Stride |
335 |
|
|
336 |
|
pxor mm6, mm6 ; accum |
337 |
|
pxor mm7, mm7 ; zero |
338 |
|
|
339 |
|
SAD_8x8_MMX |
340 |
|
SAD_8x8_MMX |
341 |
|
SAD_8x8_MMX |
342 |
|
SAD_8x8_MMX |
343 |
|
|
344 |
|
pmaddwd mm6, [mmx_one] ; collapse |
345 |
|
movq mm7, mm6 |
346 |
|
psrlq mm7, 32 |
347 |
|
paddd mm6, mm7 |
348 |
|
|
349 |
|
movd eax, mm6 |
350 |
|
|
351 |
|
ret |
352 |
|
|
353 |
|
;----------------------------------------------------------------------------- |
354 |
|
; |
355 |
|
; uint32_t sad16v_mmx(const uint8_t * const cur, |
356 |
|
; const uint8_t * const ref, |
357 |
|
; const uint32_t stride, |
358 |
|
; int32_t *sad); |
359 |
|
; |
360 |
|
;----------------------------------------------------------------------------- |
361 |
|
|
362 |
|
ALIGN 16 |
363 |
|
sad16v_mmx: |
364 |
|
|
365 |
|
push ebx |
366 |
|
push edi |
367 |
|
|
368 |
|
mov eax, [esp + 8 + 4] ; Src1 |
369 |
|
mov edx, [esp + 8 + 8] ; Src2 |
370 |
|
mov ecx, [esp + 8 + 12] ; Stride |
371 |
|
mov ebx, [esp + 8 + 16] ; sad ptr |
372 |
|
|
373 |
|
pxor mm5, mm5 ; accum |
374 |
|
pxor mm6, mm6 ; accum |
375 |
|
pxor mm7, mm7 ; zero |
376 |
|
|
377 |
|
SADV_16x16_MMX |
378 |
|
SADV_16x16_MMX |
379 |
|
SADV_16x16_MMX |
380 |
|
SADV_16x16_MMX |
381 |
|
SADV_16x16_MMX |
382 |
|
SADV_16x16_MMX |
383 |
|
SADV_16x16_MMX |
384 |
|
SADV_16x16_MMX |
385 |
|
|
386 |
|
pmaddwd mm5, [mmx_one] ; collapse |
387 |
|
pmaddwd mm6, [mmx_one] ; collapse |
388 |
|
|
389 |
|
movq mm2, mm5 |
390 |
|
movq mm3, mm6 |
391 |
|
|
392 |
|
psrlq mm2, 32 |
393 |
|
psrlq mm3, 32 |
394 |
|
|
395 |
|
paddd mm5, mm2 |
396 |
|
paddd mm6, mm3 |
397 |
|
|
398 |
|
movd [ebx], mm5 |
399 |
|
movd [ebx + 4], mm6 |
400 |
|
|
401 |
|
paddd mm5, mm6 |
402 |
|
|
403 |
|
movd edi, mm5 |
404 |
|
|
405 |
|
pxor mm5, mm5 |
406 |
|
pxor mm6, mm6 |
407 |
|
|
408 |
|
SADV_16x16_MMX |
409 |
|
SADV_16x16_MMX |
410 |
|
SADV_16x16_MMX |
411 |
|
SADV_16x16_MMX |
412 |
|
SADV_16x16_MMX |
413 |
|
SADV_16x16_MMX |
414 |
|
SADV_16x16_MMX |
415 |
|
SADV_16x16_MMX |
416 |
|
|
417 |
|
pmaddwd mm5, [mmx_one] ; collapse |
418 |
|
pmaddwd mm6, [mmx_one] ; collapse |
419 |
|
|
420 |
|
movq mm2, mm5 |
421 |
|
movq mm3, mm6 |
422 |
|
|
423 |
|
psrlq mm2, 32 |
424 |
|
psrlq mm3, 32 |
425 |
|
|
426 |
|
paddd mm5, mm2 |
427 |
|
paddd mm6, mm3 |
428 |
|
|
429 |
|
movd [ebx + 8], mm5 |
430 |
|
movd [ebx + 12], mm6 |
431 |
|
|
432 |
|
paddd mm5, mm6 |
433 |
|
|
434 |
|
movd eax, mm5 |
435 |
|
|
436 |
|
add eax, edi |
437 |
|
|
438 |
|
pop edi |
439 |
|
pop ebx |
440 |
|
|
441 |
|
ret |
442 |
|
|
443 |
|
;----------------------------------------------------------------------------- |
444 |
|
; |
445 |
|
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
446 |
|
; const uint8_t * const ref1, |
447 |
|
; const uint8_t * const ref2, |
448 |
|
; const uint32_t stride); |
449 |
|
; |
450 |
|
;----------------------------------------------------------------------------- |
451 |
|
|
452 |
|
ALIGN 16 |
453 |
|
sad16bi_mmx: |
454 |
|
push ebx |
455 |
|
mov eax, [esp+4+ 4] ; Src |
456 |
|
mov edx, [esp+4+ 8] ; Ref1 |
457 |
|
mov ebx, [esp+4+12] ; Ref2 |
458 |
|
mov ecx, [esp+4+16] ; Stride |
459 |
|
|
460 |
|
pxor mm6, mm6 ; accum2 |
461 |
|
pxor mm7, mm7 |
462 |
|
.Loop |
463 |
|
SADBI_16x16_MMX 0, 0 |
464 |
|
SADBI_16x16_MMX 8, 1 |
465 |
|
SADBI_16x16_MMX 0, 0 |
466 |
|
SADBI_16x16_MMX 8, 1 |
467 |
|
SADBI_16x16_MMX 0, 0 |
468 |
|
SADBI_16x16_MMX 8, 1 |
469 |
|
SADBI_16x16_MMX 0, 0 |
470 |
|
SADBI_16x16_MMX 8, 1 |
471 |
|
SADBI_16x16_MMX 0, 0 |
472 |
|
SADBI_16x16_MMX 8, 1 |
473 |
|
SADBI_16x16_MMX 0, 0 |
474 |
|
SADBI_16x16_MMX 8, 1 |
475 |
|
SADBI_16x16_MMX 0, 0 |
476 |
|
SADBI_16x16_MMX 8, 1 |
477 |
|
SADBI_16x16_MMX 0, 0 |
478 |
|
SADBI_16x16_MMX 8, 1 |
479 |
|
|
480 |
|
SADBI_16x16_MMX 0, 0 |
481 |
|
SADBI_16x16_MMX 8, 1 |
482 |
|
SADBI_16x16_MMX 0, 0 |
483 |
|
SADBI_16x16_MMX 8, 1 |
484 |
|
SADBI_16x16_MMX 0, 0 |
485 |
|
SADBI_16x16_MMX 8, 1 |
486 |
|
SADBI_16x16_MMX 0, 0 |
487 |
|
SADBI_16x16_MMX 8, 1 |
488 |
|
SADBI_16x16_MMX 0, 0 |
489 |
|
SADBI_16x16_MMX 8, 1 |
490 |
|
SADBI_16x16_MMX 0, 0 |
491 |
|
SADBI_16x16_MMX 8, 1 |
492 |
|
SADBI_16x16_MMX 0, 0 |
493 |
|
SADBI_16x16_MMX 8, 1 |
494 |
|
SADBI_16x16_MMX 0, 0 |
495 |
|
SADBI_16x16_MMX 8, 1 |
496 |
|
|
497 |
|
pmaddwd mm6, [mmx_one] ; collapse |
498 |
|
movq mm7, mm6 |
499 |
|
psrlq mm7, 32 |
500 |
|
paddd mm6, mm7 |
501 |
|
|
502 |
|
movd eax, mm6 |
503 |
|
pop ebx |
504 |
|
|
505 |
|
ret |
506 |
|
|
507 |
|
;----------------------------------------------------------------------------- |
508 |
|
; |
509 |
|
; uint32_t sad8bi_mmx(const uint8_t * const cur, |
510 |
|
; const uint8_t * const ref1, |
511 |
|
; const uint8_t * const ref2, |
512 |
|
; const uint32_t stride); |
513 |
|
; |
514 |
|
;----------------------------------------------------------------------------- |
515 |
|
|
516 |
|
ALIGN 16 |
517 |
|
sad8bi_mmx: |
518 |
|
push ebx |
519 |
|
mov eax, [esp+4+ 4] ; Src |
520 |
|
mov edx, [esp+4+ 8] ; Ref1 |
521 |
|
mov ebx, [esp+4+12] ; Ref2 |
522 |
|
mov ecx, [esp+4+16] ; Stride |
523 |
|
|
524 |
|
pxor mm6, mm6 ; accum2 |
525 |
|
pxor mm7, mm7 |
526 |
|
.Loop |
527 |
|
SADBI_16x16_MMX 0, 1 |
528 |
|
SADBI_16x16_MMX 0, 1 |
529 |
|
SADBI_16x16_MMX 0, 1 |
530 |
|
SADBI_16x16_MMX 0, 1 |
531 |
|
SADBI_16x16_MMX 0, 1 |
532 |
|
SADBI_16x16_MMX 0, 1 |
533 |
|
SADBI_16x16_MMX 0, 1 |
534 |
|
SADBI_16x16_MMX 0, 1 |
535 |
|
|
536 |
|
pmaddwd mm6, [mmx_one] ; collapse |
537 |
|
movq mm7, mm6 |
538 |
|
psrlq mm7, 32 |
539 |
|
paddd mm6, mm7 |
540 |
|
|
541 |
|
movd eax, mm6 |
542 |
|
pop ebx |
543 |
|
ret |
544 |
|
|
545 |
|
;----------------------------------------------------------------------------- |
546 |
|
; |
547 |
|
; uint32_t dev16_mmx(const uint8_t * const cur, |
548 |
|
; const uint32_t stride); |
549 |
|
; |
550 |
|
;----------------------------------------------------------------------------- |
551 |
|
|
552 |
|
ALIGN 16 |
553 |
dev16_mmx: |
dev16_mmx: |
554 |
mov eax, [esp+ 4] ; Src |
mov eax, [esp+ 4] ; Src |
555 |
mov ecx, [esp+ 8] ; Stride |
mov ecx, [esp+ 8] ; Stride |
618 |
paddd mm6, mm5 |
paddd mm6, mm5 |
619 |
|
|
620 |
movd eax, mm6 |
movd eax, mm6 |
621 |
|
|
622 |
ret |
ret |