32 |
bits 32 |
bits 32 |
33 |
|
|
34 |
%macro cglobal 1 |
%macro cglobal 1 |
35 |
|
%ifdef PREFIX |
36 |
global _%1 |
global _%1 |
37 |
%define %1 _%1 |
%define %1 _%1 |
38 |
%else |
%else |
48 |
section .text |
section .text |
49 |
|
|
50 |
cglobal sad16_mmx |
cglobal sad16_mmx |
51 |
|
cglobal sad16v_mmx |
52 |
cglobal sad8_mmx |
cglobal sad8_mmx |
53 |
cglobal sad16bi_mmx |
cglobal sad16bi_mmx |
54 |
cglobal sad8bi_mmx |
cglobal sad8bi_mmx |
59 |
; uint32_t sad16_mmx(const uint8_t * const cur, |
; uint32_t sad16_mmx(const uint8_t * const cur, |
60 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
61 |
; const uint32_t stride, |
; const uint32_t stride, |
62 |
|
; const uint32_t best_sad); |
63 |
|
; |
64 |
; (early termination ignore; slows this down) |
; (early termination ignore; slows this down) |
65 |
; |
; |
66 |
;=========================================================================== |
;=========================================================================== |
203 |
ret |
ret |
204 |
|
|
205 |
|
|
206 |
|
;=========================================================================== |
207 |
|
; |
208 |
|
; uint32_t sad16v_mmx(const uint8_t * const cur, |
209 |
|
; const uint8_t * const ref, |
210 |
|
; const uint32_t stride, |
211 |
|
; int32_t *sad); |
212 |
|
; |
213 |
|
;=========================================================================== |
214 |
|
|
215 |
|
%macro SADV_16x16_MMX 0 |
216 |
|
movq mm0, [eax] |
217 |
|
movq mm1, [edx] |
218 |
|
|
219 |
|
movq mm2, [eax+8] |
220 |
|
movq mm3, [edx+8] |
221 |
|
|
222 |
|
movq mm4, mm0 |
223 |
|
psubusb mm0, mm1 |
224 |
|
|
225 |
|
psubusb mm1, mm4 |
226 |
|
por mm0, mm1 |
227 |
|
lea eax,[eax+ecx] |
228 |
|
|
229 |
|
movq mm4, mm2 |
230 |
|
psubusb mm2, mm3 |
231 |
|
|
232 |
|
psubusb mm3, mm4 |
233 |
|
por mm2, mm3 |
234 |
|
lea edx,[edx+ecx] |
235 |
|
|
236 |
|
movq mm1,mm0 |
237 |
|
movq mm3,mm2 |
238 |
|
|
239 |
|
punpcklbw mm0,mm7 |
240 |
|
punpckhbw mm1,mm7 |
241 |
|
punpcklbw mm2,mm7 |
242 |
|
punpckhbw mm3,mm7 |
243 |
|
|
244 |
|
paddusw mm0,mm1 |
245 |
|
paddusw mm2,mm3 |
246 |
|
|
247 |
|
paddusw mm5, mm0 |
248 |
|
paddusw mm6, mm2 |
249 |
|
%endmacro |
250 |
|
|
251 |
|
align 16 |
252 |
|
sad16v_mmx: |
253 |
|
|
254 |
|
push ebx |
255 |
|
push edi |
256 |
|
|
257 |
|
mov eax, [esp + 8 + 4] ; Src1 |
258 |
|
mov edx, [esp + 8 + 8] ; Src2 |
259 |
|
mov ecx, [esp + 8 + 12] ; Stride |
260 |
|
mov ebx, [esp + 8 + 16] ; sad ptr |
261 |
|
|
262 |
|
pxor mm5, mm5 ; accum |
263 |
|
pxor mm6, mm6 ; accum |
264 |
|
pxor mm7, mm7 ; zero |
265 |
|
|
266 |
|
SADV_16x16_MMX |
267 |
|
SADV_16x16_MMX |
268 |
|
SADV_16x16_MMX |
269 |
|
SADV_16x16_MMX |
270 |
|
SADV_16x16_MMX |
271 |
|
SADV_16x16_MMX |
272 |
|
SADV_16x16_MMX |
273 |
|
SADV_16x16_MMX |
274 |
|
|
275 |
|
pmaddwd mm5, [mmx_one] ; collapse |
276 |
|
pmaddwd mm6, [mmx_one] ; collapse |
277 |
|
|
278 |
|
movq mm2, mm5 |
279 |
|
movq mm3, mm6 |
280 |
|
|
281 |
|
psrlq mm2, 32 |
282 |
|
psrlq mm3, 32 |
283 |
|
|
284 |
|
paddd mm5, mm2 |
285 |
|
paddd mm6, mm3 |
286 |
|
|
287 |
|
movd [ebx], mm5 |
288 |
|
movd [ebx + 4], mm6 |
289 |
|
|
290 |
|
paddd mm5, mm6 |
291 |
|
|
292 |
|
movd edi, mm5 |
293 |
|
|
294 |
|
pxor mm5, mm5 |
295 |
|
pxor mm6, mm6 |
296 |
|
|
297 |
|
SADV_16x16_MMX |
298 |
|
SADV_16x16_MMX |
299 |
|
SADV_16x16_MMX |
300 |
|
SADV_16x16_MMX |
301 |
|
SADV_16x16_MMX |
302 |
|
SADV_16x16_MMX |
303 |
|
SADV_16x16_MMX |
304 |
|
SADV_16x16_MMX |
305 |
|
|
306 |
|
pmaddwd mm5, [mmx_one] ; collapse |
307 |
|
pmaddwd mm6, [mmx_one] ; collapse |
308 |
|
|
309 |
|
movq mm2, mm5 |
310 |
|
movq mm3, mm6 |
311 |
|
|
312 |
|
psrlq mm2, 32 |
313 |
|
psrlq mm3, 32 |
314 |
|
|
315 |
|
paddd mm5, mm2 |
316 |
|
paddd mm6, mm3 |
317 |
|
|
318 |
|
movd [ebx + 8], mm5 |
319 |
|
movd [ebx + 12], mm6 |
320 |
|
|
321 |
|
paddd mm5, mm6 |
322 |
|
|
323 |
|
movd eax, mm5 |
324 |
|
|
325 |
|
add eax, edi |
326 |
|
|
327 |
|
pop edi |
328 |
|
pop ebx |
329 |
|
ret |
330 |
|
|
331 |
|
|
332 |
|
|
334 |
; |
; |
335 |
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
336 |
; const uint8_t * const ref1, |
; const uint8_t * const ref1, |
337 |
|
; const uint8_t * const ref2, |
338 |
|
; const uint32_t stride); |
339 |
|
; |
340 |
|
;=========================================================================== |
341 |
|
%macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
342 |
|
|
343 |
|
movq mm0, [edx+%1] |
344 |
|
movq mm2, [ebx+%1] |
345 |
|
movq mm1, mm0 |
346 |
|
movq mm3, mm2 |
347 |
|
|
348 |
|
%if %2 != 0 |
349 |
|
add edx, ecx |
350 |
|
%endif |
351 |
|
|
352 |
|
punpcklbw mm0, mm7 |
353 |
|
punpckhbw mm1, mm7 |
354 |
|
punpcklbw mm2, mm7 |
355 |
|
punpckhbw mm3, mm7 |
356 |
|
|
357 |
|
%if %2 != 0 |
358 |
|
add ebx, ecx |
359 |
|
%endif |
360 |
|
|
361 |
|
paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
362 |
|
paddusw mm1, mm3 |
363 |
|
paddusw mm0, [mmx_one] ; mm01 += 1 |
364 |
|
paddusw mm1, [mmx_one] |
365 |
|
psrlw mm0, 1 ; mm01 >>= 1 |
366 |
|
psrlw mm1, 1 |
367 |
|
|
368 |
|
movq mm2, [eax+%1] |
369 |
|
movq mm3, mm2 |
370 |
|
punpcklbw mm2, mm7 ; mm23 = src |
371 |
|
punpckhbw mm3, mm7 |
372 |
|
|
373 |
|
%if %2 != 0 |
374 |
|
add eax, ecx |
375 |
|
%endif |
376 |
|
|
377 |
|
movq mm4, mm0 |
378 |
|
movq mm5, mm1 |
379 |
|
psubusw mm0, mm2 |
380 |
|
psubusw mm1, mm3 |
381 |
|
psubusw mm2, mm4 |
382 |
|
psubusw mm3, mm5 |
383 |
|
por mm0, mm2 ; mm01 = ABS(mm01 - mm23) |
384 |
|
por mm1, mm3 |
385 |
|
|
386 |
|
paddusw mm6,mm0 ; mm6 += mm01 |
387 |
|
paddusw mm6,mm1 |
388 |
|
|
389 |
|
%endmacro |
390 |
|
|
391 |
|
align 16 |
392 |
|
sad16bi_mmx: |
393 |
|
push ebx |
394 |
|
mov eax, [esp+4+ 4] ; Src |
395 |
|
mov edx, [esp+4+ 8] ; Ref1 |
396 |
|
mov ebx, [esp+4+12] ; Ref2 |
397 |
|
mov ecx, [esp+4+16] ; Stride |
398 |
|
|
399 |
|
pxor mm6, mm6 ; accum2 |
400 |
|
pxor mm7, mm7 |
401 |
|
.Loop |
402 |
|
SADBI_16x16_MMX 0, 0 |
403 |
|
SADBI_16x16_MMX 8, 1 |
404 |
|
SADBI_16x16_MMX 0, 0 |
405 |
|
SADBI_16x16_MMX 8, 1 |
406 |
|
SADBI_16x16_MMX 0, 0 |
407 |
|
SADBI_16x16_MMX 8, 1 |
408 |
|
SADBI_16x16_MMX 0, 0 |
409 |
|
SADBI_16x16_MMX 8, 1 |
410 |
|
SADBI_16x16_MMX 0, 0 |
411 |
|
SADBI_16x16_MMX 8, 1 |
412 |
|
SADBI_16x16_MMX 0, 0 |
413 |
|
SADBI_16x16_MMX 8, 1 |
414 |
|
SADBI_16x16_MMX 0, 0 |
415 |
|
SADBI_16x16_MMX 8, 1 |
416 |
|
SADBI_16x16_MMX 0, 0 |
417 |
|
SADBI_16x16_MMX 8, 1 |
418 |
|
|
419 |
|
SADBI_16x16_MMX 0, 0 |
420 |
|
SADBI_16x16_MMX 8, 1 |
421 |
|
SADBI_16x16_MMX 0, 0 |
422 |
|
SADBI_16x16_MMX 8, 1 |
423 |
|
SADBI_16x16_MMX 0, 0 |
424 |
|
SADBI_16x16_MMX 8, 1 |
425 |
|
SADBI_16x16_MMX 0, 0 |
426 |
|
SADBI_16x16_MMX 8, 1 |
427 |
|
SADBI_16x16_MMX 0, 0 |
428 |
|
SADBI_16x16_MMX 8, 1 |
429 |
|
SADBI_16x16_MMX 0, 0 |
430 |
|
SADBI_16x16_MMX 8, 1 |
431 |
|
SADBI_16x16_MMX 0, 0 |
432 |
|
SADBI_16x16_MMX 8, 1 |
433 |
|
SADBI_16x16_MMX 0, 0 |
434 |
|
SADBI_16x16_MMX 8, 1 |
435 |
|
|
436 |
|
pmaddwd mm6, [mmx_one] ; collapse |
437 |
|
movq mm7, mm6 |
438 |
|
psrlq mm7, 32 |
439 |
|
paddd mm6, mm7 |
440 |
|
|
441 |
|
movd eax, mm6 |
442 |
|
pop ebx |
443 |
|
ret |
444 |
|
|
445 |
|
;=========================================================================== |
446 |
|
; |
447 |
|
; uint32_t sad8bi_mmx(const uint8_t * const cur, |
448 |
|
; const uint8_t * const ref1, |
449 |
|
; const uint8_t * const ref2, |
450 |
|
; const uint32_t stride); |
451 |
|
; |
452 |
|
;=========================================================================== |
453 |
|
align 16 |
454 |
|
sad8bi_mmx: |
455 |
|
push ebx |
456 |
|
mov eax, [esp+4+ 4] ; Src |
457 |
|
mov edx, [esp+4+ 8] ; Ref1 |
458 |
|
mov ebx, [esp+4+12] ; Ref2 |
459 |
|
mov ecx, [esp+4+16] ; Stride |
460 |
|
|
461 |
|
pxor mm6, mm6 ; accum2 |
462 |
|
pxor mm7, mm7 |
463 |
|
.Loop |
464 |
|
SADBI_16x16_MMX 0, 1 |
465 |
|
SADBI_16x16_MMX 0, 1 |
466 |
|
SADBI_16x16_MMX 0, 1 |
467 |
|
SADBI_16x16_MMX 0, 1 |
468 |
|
SADBI_16x16_MMX 0, 1 |
469 |
|
SADBI_16x16_MMX 0, 1 |
470 |
|
SADBI_16x16_MMX 0, 1 |
471 |
|
SADBI_16x16_MMX 0, 1 |
472 |
|
|
473 |
|
pmaddwd mm6, [mmx_one] ; collapse |
474 |
|
movq mm7, mm6 |
475 |
|
psrlq mm7, 32 |
476 |
|
paddd mm6, mm7 |
477 |
|
|
478 |
|
movd eax, mm6 |
479 |
|
pop ebx |
480 |
|
ret |
481 |
|
|
482 |
|
|
483 |
|
|
484 |
|
|
485 |
|
;=========================================================================== |
486 |
|
; |
487 |
|
; uint32_t dev16_mmx(const uint8_t * const cur, |
488 |
|
; const uint32_t stride); |
489 |
|
; |
490 |
|
;=========================================================================== |
491 |
|
|
492 |
|
%macro MEAN_16x16_MMX 0 |
493 |
movq mm0, [eax] |
movq mm0, [eax] |
494 |
movq mm2, [eax+8] |
movq mm2, [eax+8] |
495 |
lea eax,[eax+ecx] |
lea eax,[eax+ecx] |