4 |
; * - mmx 8x8 block-based halfpel interpolation - |
; * - mmx 8x8 block-based halfpel interpolation - |
5 |
; * |
; * |
6 |
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
; * 2002 Michael Militzer <isibaar@xvid.org> |
; * 2002-2008 Michael Militzer <michael@xvid.org> |
8 |
; * |
; * |
9 |
; * This program is free software ; you can redistribute it and/or modify |
; * This program is free software ; you can redistribute it and/or modify |
10 |
; * it under the terms of the GNU General Public License as published by |
; * it under the terms of the GNU General Public License as published by |
22 |
; * |
; * |
23 |
; ****************************************************************************/ |
; ****************************************************************************/ |
24 |
|
|
25 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
26 |
|
|
27 |
;============================================================================= |
;============================================================================= |
28 |
; Read only data |
; Read only data |
29 |
;============================================================================= |
;============================================================================= |
30 |
|
|
31 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata |
|
|
%else |
|
|
SECTION .rodata align=16 |
|
|
%endif |
|
32 |
|
|
33 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
34 |
; (16 - r) rounding table |
; (16 - r) rounding table |
35 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
36 |
|
|
37 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
38 |
rounding_lowpass_mmx: |
rounding_lowpass_mmx: |
39 |
times 4 dw 16 |
times 4 dw 16 |
40 |
times 4 dw 15 |
times 4 dw 15 |
77 |
; Code |
; Code |
78 |
;============================================================================= |
;============================================================================= |
79 |
|
|
80 |
SECTION .text |
SECTION .rotext align=SECTION_ALIGN |
81 |
|
|
82 |
cglobal interpolate8x8_halfpel_h_mmx |
cglobal interpolate8x8_halfpel_h_mmx |
83 |
cglobal interpolate8x8_halfpel_v_mmx |
cglobal interpolate8x8_halfpel_v_mmx |
84 |
cglobal interpolate8x8_halfpel_hv_mmx |
cglobal interpolate8x8_halfpel_hv_mmx |
85 |
|
|
86 |
|
cglobal interpolate8x4_halfpel_h_mmx |
87 |
|
cglobal interpolate8x4_halfpel_v_mmx |
88 |
|
cglobal interpolate8x4_halfpel_hv_mmx |
89 |
|
|
90 |
cglobal interpolate8x8_avg4_mmx |
cglobal interpolate8x8_avg4_mmx |
91 |
cglobal interpolate8x8_avg2_mmx |
cglobal interpolate8x8_avg2_mmx |
92 |
|
|
122 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
123 |
|
|
124 |
%macro COPY_H_MMX 0 |
%macro COPY_H_MMX 0 |
125 |
movq mm0, [esi] |
movq mm0, [TMP0] |
126 |
movq mm2, [esi + 1] |
movq mm2, [TMP0 + 1] |
127 |
movq mm1, mm0 |
movq mm1, mm0 |
128 |
movq mm3, mm2 |
movq mm3, mm2 |
129 |
|
|
133 |
CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
134 |
|
|
135 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
136 |
movq [edi], mm0 ; [dst] = mm01 |
movq [_EAX], mm0 ; [dst] = mm01 |
137 |
|
|
138 |
add esi, edx ; src += stride |
add TMP0, TMP1 ; src += stride |
139 |
add edi, edx ; dst += stride |
add _EAX, TMP1 ; dst += stride |
140 |
%endmacro |
%endmacro |
141 |
|
|
142 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
143 |
interpolate8x8_halfpel_h_mmx: |
interpolate8x8_halfpel_h_mmx: |
144 |
|
|
145 |
push esi |
mov _EAX, prm4 ; rounding |
146 |
push edi |
lea TMP0, [rounding1_mmx] |
147 |
mov eax, [esp + 8 + 16] ; rounding |
movq mm7, [TMP0 + _EAX * 8] |
148 |
|
|
149 |
movq mm7, [rounding1_mmx + eax * 8] |
mov _EAX, prm1 ; dst |
150 |
|
mov TMP0, prm2 ; src |
151 |
mov edi, [esp + 8 + 4] ; dst |
mov TMP1, prm3 ; stride |
|
mov esi, [esp + 8 + 8] ; src |
|
|
mov edx, [esp + 8 + 12] ; stride |
|
152 |
|
|
153 |
pxor mm6, mm6 ; zero |
pxor mm6, mm6 ; zero |
154 |
|
|
161 |
COPY_H_MMX |
COPY_H_MMX |
162 |
COPY_H_MMX |
COPY_H_MMX |
163 |
|
|
|
pop edi |
|
|
pop esi |
|
|
|
|
164 |
ret |
ret |
165 |
|
ENDFUNC |
166 |
|
|
167 |
|
|
168 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
175 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
176 |
|
|
177 |
%macro COPY_V_MMX 0 |
%macro COPY_V_MMX 0 |
178 |
movq mm0, [esi] |
movq mm0, [TMP0] |
179 |
movq mm2, [esi + edx] |
movq mm2, [TMP0 + TMP1] |
180 |
movq mm1, mm0 |
movq mm1, mm0 |
181 |
movq mm3, mm2 |
movq mm3, mm2 |
182 |
|
|
186 |
CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
187 |
|
|
188 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
189 |
movq [edi], mm0 ; [dst] = mm01 |
movq [_EAX], mm0 ; [dst] = mm01 |
190 |
|
|
191 |
add esi, edx ; src += stride |
add TMP0, TMP1 ; src += stride |
192 |
add edi, edx ; dst += stride |
add _EAX, TMP1 ; dst += stride |
193 |
%endmacro |
%endmacro |
194 |
|
|
195 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
196 |
interpolate8x8_halfpel_v_mmx: |
interpolate8x8_halfpel_v_mmx: |
197 |
|
|
198 |
push esi |
mov _EAX, prm4 ; rounding |
199 |
push edi |
lea TMP0, [rounding1_mmx] |
200 |
|
movq mm7, [TMP0 + _EAX * 8] |
201 |
mov eax, [esp + 8 + 16] ; rounding |
|
202 |
|
mov _EAX, prm1 ; dst |
203 |
movq mm7, [rounding1_mmx + eax * 8] |
mov TMP0, prm2 ; src |
204 |
|
mov TMP1, prm3 ; stride |
|
mov edi, [esp + 8 + 4] ; dst |
|
|
mov esi, [esp + 8 + 8] ; src |
|
|
mov edx, [esp + 8 + 12] ; stride |
|
205 |
|
|
206 |
pxor mm6, mm6 ; zero |
pxor mm6, mm6 ; zero |
207 |
|
|
215 |
COPY_V_MMX |
COPY_V_MMX |
216 |
COPY_V_MMX |
COPY_V_MMX |
217 |
|
|
|
pop edi |
|
|
pop esi |
|
|
|
|
218 |
ret |
ret |
219 |
|
ENDFUNC |
220 |
|
|
221 |
|
|
222 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
231 |
|
|
232 |
%macro COPY_HV_MMX 0 |
%macro COPY_HV_MMX 0 |
233 |
; current row |
; current row |
234 |
movq mm0, [esi] |
movq mm0, [TMP0] |
235 |
movq mm2, [esi + 1] |
movq mm2, [TMP0 + 1] |
236 |
|
|
237 |
movq mm1, mm0 |
movq mm1, mm0 |
238 |
movq mm3, mm2 |
movq mm3, mm2 |
246 |
paddusw mm1, mm3 |
paddusw mm1, mm3 |
247 |
|
|
248 |
; next row |
; next row |
249 |
movq mm4, [esi + edx] |
movq mm4, [TMP0 + TMP1] |
250 |
movq mm2, [esi + edx + 1] |
movq mm2, [TMP0 + TMP1 + 1] |
251 |
|
|
252 |
movq mm5, mm4 |
movq mm5, mm4 |
253 |
movq mm3, mm2 |
movq mm3, mm2 |
270 |
psrlw mm1, 2 |
psrlw mm1, 2 |
271 |
|
|
272 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
273 |
movq [edi], mm0 ; [dst] = mm01 |
movq [_EAX], mm0 ; [dst] = mm01 |
274 |
|
|
275 |
add esi, edx ; src += stride |
add TMP0, TMP1 ; src += stride |
276 |
add edi, edx ; dst += stride |
add _EAX, TMP1 ; dst += stride |
277 |
%endmacro |
%endmacro |
278 |
|
|
279 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
280 |
interpolate8x8_halfpel_hv_mmx: |
interpolate8x8_halfpel_hv_mmx: |
281 |
|
|
282 |
push esi |
mov _EAX, prm4 ; rounding |
283 |
push edi |
lea TMP0, [rounding2_mmx] |
284 |
|
movq mm7, [TMP0 + _EAX * 8] |
285 |
|
|
286 |
mov eax, [esp + 8 + 16] ; rounding |
mov _EAX, prm1 ; dst |
287 |
|
mov TMP0, prm2 ; src |
|
movq mm7, [rounding2_mmx + eax * 8] |
|
|
|
|
|
mov edi, [esp + 8 + 4] ; dst |
|
|
mov esi, [esp + 8 + 8] ; src |
|
|
|
|
|
mov eax, 8 |
|
288 |
|
|
289 |
pxor mm6, mm6 ; zero |
pxor mm6, mm6 ; zero |
290 |
|
|
291 |
mov edx, [esp + 8 + 12] ; stride |
mov TMP1, prm3 ; stride |
292 |
|
|
293 |
COPY_HV_MMX |
COPY_HV_MMX |
294 |
COPY_HV_MMX |
COPY_HV_MMX |
299 |
COPY_HV_MMX |
COPY_HV_MMX |
300 |
COPY_HV_MMX |
COPY_HV_MMX |
301 |
|
|
302 |
pop edi |
ret |
303 |
pop esi |
ENDFUNC |
304 |
|
|
305 |
|
;----------------------------------------------------------------------------- |
306 |
|
; |
307 |
|
; void interpolate8x4_halfpel_h_mmx(uint8_t * const dst, |
308 |
|
; const uint8_t * const src, |
309 |
|
; const uint32_t stride, |
310 |
|
; const uint32_t rounding); |
311 |
|
; |
312 |
|
;----------------------------------------------------------------------------- |
313 |
|
|
314 |
|
ALIGN SECTION_ALIGN |
315 |
|
interpolate8x4_halfpel_h_mmx: |
316 |
|
|
317 |
|
mov _EAX, prm4 ; rounding |
318 |
|
lea TMP0, [rounding1_mmx] |
319 |
|
movq mm7, [TMP0 + _EAX * 8] |
320 |
|
|
321 |
|
mov _EAX, prm1 ; dst |
322 |
|
mov TMP0, prm2 ; src |
323 |
|
mov TMP1, prm3 ; stride |
324 |
|
|
325 |
|
pxor mm6, mm6 ; zero |
326 |
|
|
327 |
|
COPY_H_MMX |
328 |
|
COPY_H_MMX |
329 |
|
COPY_H_MMX |
330 |
|
COPY_H_MMX |
331 |
|
|
332 |
|
ret |
333 |
|
ENDFUNC |
334 |
|
|
335 |
|
|
336 |
|
;----------------------------------------------------------------------------- |
337 |
|
; |
338 |
|
; void interpolate8x4_halfpel_v_mmx(uint8_t * const dst, |
339 |
|
; const uint8_t * const src, |
340 |
|
; const uint32_t stride, |
341 |
|
; const uint32_t rounding); |
342 |
|
; |
343 |
|
;----------------------------------------------------------------------------- |
344 |
|
|
345 |
|
ALIGN SECTION_ALIGN |
346 |
|
interpolate8x4_halfpel_v_mmx: |
347 |
|
|
348 |
|
mov _EAX, prm4 ; rounding |
349 |
|
lea TMP0, [rounding1_mmx] |
350 |
|
movq mm7, [TMP0 + _EAX * 8] |
351 |
|
|
352 |
|
mov _EAX, prm1 ; dst |
353 |
|
mov TMP0, prm2 ; src |
354 |
|
mov TMP1, prm3 ; stride |
355 |
|
|
356 |
|
pxor mm6, mm6 ; zero |
357 |
|
|
358 |
|
|
359 |
|
COPY_V_MMX |
360 |
|
COPY_V_MMX |
361 |
|
COPY_V_MMX |
362 |
|
COPY_V_MMX |
363 |
|
|
364 |
ret |
ret |
365 |
|
ENDFUNC |
366 |
|
|
367 |
|
|
368 |
|
;----------------------------------------------------------------------------- |
369 |
|
; |
370 |
|
; void interpolate8x4_halfpel_hv_mmx(uint8_t * const dst, |
371 |
|
; const uint8_t * const src, |
372 |
|
; const uint32_t stride, |
373 |
|
; const uint32_t rounding); |
374 |
|
; |
375 |
|
; |
376 |
|
;----------------------------------------------------------------------------- |
377 |
|
|
378 |
|
ALIGN SECTION_ALIGN |
379 |
|
interpolate8x4_halfpel_hv_mmx: |
380 |
|
|
381 |
|
mov _EAX, prm4 ; rounding |
382 |
|
lea TMP0, [rounding2_mmx] |
383 |
|
movq mm7, [TMP0 + _EAX * 8] |
384 |
|
|
385 |
|
mov _EAX, prm1 ; dst |
386 |
|
mov TMP0, prm2 ; src |
387 |
|
|
388 |
|
pxor mm6, mm6 ; zero |
389 |
|
|
390 |
|
mov TMP1, prm3 ; stride |
391 |
|
|
392 |
|
COPY_HV_MMX |
393 |
|
COPY_HV_MMX |
394 |
|
COPY_HV_MMX |
395 |
|
COPY_HV_MMX |
396 |
|
|
397 |
|
ret |
398 |
|
ENDFUNC |
399 |
|
|
400 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
401 |
; |
; |
409 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
410 |
|
|
411 |
%macro AVG2_MMX_RND0 0 |
%macro AVG2_MMX_RND0 0 |
412 |
movq mm0, [eax] ; src1 -> mm0 |
movq mm0, [_EAX] ; src1 -> mm0 |
413 |
movq mm1, [ebx] ; src2 -> mm1 |
movq mm1, [_EBX] ; src2 -> mm1 |
414 |
|
|
415 |
movq mm4, [eax+edx] |
movq mm4, [_EAX+TMP1] |
416 |
movq mm5, [ebx+edx] |
movq mm5, [_EBX+TMP1] |
417 |
|
|
418 |
movq mm2, mm0 ; src1 -> mm2 |
movq mm2, mm0 ; src1 -> mm2 |
419 |
movq mm3, mm1 ; src2 -> mm3 |
movq mm3, mm1 ; src2 -> mm3 |
448 |
paddb mm4, mm5 |
paddb mm4, mm5 |
449 |
paddb mm4, mm3 |
paddb mm4, mm3 |
450 |
|
|
451 |
lea eax, [eax+2*edx] |
lea _EAX, [_EAX+2*TMP1] |
452 |
lea ebx, [ebx+2*edx] |
lea _EBX, [_EBX+2*TMP1] |
453 |
|
|
454 |
movq [ecx], mm0 ; (src1 + src2 + 1) / 2 -> dst |
movq [TMP0], mm0 ; (src1 + src2 + 1) / 2 -> dst |
455 |
movq [ecx+edx], mm4 |
movq [TMP0+TMP1], mm4 |
456 |
%endmacro |
%endmacro |
457 |
|
|
458 |
%macro AVG2_MMX_RND1 0 |
%macro AVG2_MMX_RND1 0 |
459 |
movq mm0, [eax] ; src1 -> mm0 |
movq mm0, [_EAX] ; src1 -> mm0 |
460 |
movq mm1, [ebx] ; src2 -> mm1 |
movq mm1, [_EBX] ; src2 -> mm1 |
461 |
|
|
462 |
movq mm4, [eax+edx] |
movq mm4, [_EAX+TMP1] |
463 |
movq mm5, [ebx+edx] |
movq mm5, [_EBX+TMP1] |
464 |
|
|
465 |
movq mm2, mm0 ; src1 -> mm2 |
movq mm2, mm0 ; src1 -> mm2 |
466 |
movq mm3, mm1 ; src2 -> mm3 |
movq mm3, mm1 ; src2 -> mm3 |
495 |
paddb mm4, mm5 |
paddb mm4, mm5 |
496 |
paddb mm4, mm3 |
paddb mm4, mm3 |
497 |
|
|
498 |
lea eax, [eax+2*edx] |
lea _EAX, [_EAX+2*TMP1] |
499 |
lea ebx, [ebx+2*edx] |
lea _EBX, [_EBX+2*TMP1] |
500 |
|
|
501 |
movq [ecx], mm0 ; (src1 + src2 + 1) / 2 -> dst |
movq [TMP0], mm0 ; (src1 + src2 + 1) / 2 -> dst |
502 |
movq [ecx+edx], mm4 |
movq [TMP0+TMP1], mm4 |
503 |
%endmacro |
%endmacro |
504 |
|
|
505 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
506 |
interpolate8x8_avg2_mmx: |
interpolate8x8_avg2_mmx: |
507 |
|
|
508 |
push ebx |
mov eax, prm5d ; rounding |
509 |
|
test _EAX, _EAX |
|
mov eax, [esp + 4 + 20] ; rounding |
|
|
test eax, eax |
|
510 |
|
|
511 |
jnz near .rounding1 |
jnz near .rounding1 |
512 |
|
|
513 |
mov eax, [esp + 4 + 24] ; height -> eax |
mov eax, prm6d ; height -> _EAX |
514 |
sub eax, 8 |
sub _EAX, 8 |
515 |
test eax, eax |
|
516 |
|
mov TMP0, prm1 ; dst -> edi |
517 |
mov ecx, [esp + 4 + 4] ; dst -> edi |
mov _EAX, prm2 ; src1 -> esi |
518 |
mov eax, [esp + 4 + 8] ; src1 -> esi |
mov TMP1, prm4 ; stride -> TMP1 |
519 |
mov ebx, [esp + 4 + 12] ; src2 -> eax |
|
520 |
mov edx, [esp + 4 + 16] ; stride -> edx |
push _EBX |
521 |
|
%ifdef ARCH_IS_X86_64 |
522 |
|
mov _EBX, prm3 |
523 |
|
%else |
524 |
|
mov _EBX, [esp + 4 + 12] ; src2 -> eax |
525 |
|
%endif |
526 |
|
|
527 |
movq mm7, [mmx_one] |
movq mm7, [mmx_one] |
528 |
|
|
529 |
jz near .start0 |
jz near .start0 |
530 |
|
|
531 |
AVG2_MMX_RND0 |
AVG2_MMX_RND0 |
532 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
533 |
|
|
534 |
.start0 |
.start0: |
535 |
|
|
536 |
AVG2_MMX_RND0 |
AVG2_MMX_RND0 |
537 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
538 |
AVG2_MMX_RND0 |
AVG2_MMX_RND0 |
539 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
540 |
AVG2_MMX_RND0 |
AVG2_MMX_RND0 |
541 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
542 |
AVG2_MMX_RND0 |
AVG2_MMX_RND0 |
543 |
|
|
544 |
pop ebx |
pop _EBX |
545 |
ret |
ret |
546 |
|
|
547 |
.rounding1 |
.rounding1: |
548 |
mov eax, [esp + 4 + 24] ; height -> eax |
mov eax, prm6d ; height -> _EAX |
549 |
sub eax, 8 |
sub _EAX, 8 |
550 |
test eax, eax |
|
551 |
|
mov TMP0, prm1 ; dst -> edi |
552 |
mov ecx, [esp + 4 + 4] ; dst -> edi |
mov _EAX, prm2 ; src1 -> esi |
553 |
mov eax, [esp + 4 + 8] ; src1 -> esi |
mov TMP1, prm4 ; stride -> TMP1 |
554 |
mov ebx, [esp + 4 + 12] ; src2 -> eax |
|
555 |
mov edx, [esp + 4 + 16] ; stride -> edx |
push _EBX |
556 |
|
%ifdef ARCH_IS_X86_64 |
557 |
|
mov _EBX, prm3 |
558 |
|
%else |
559 |
|
mov _EBX, [esp + 4 + 12] ; src2 -> eax |
560 |
|
%endif |
561 |
|
|
562 |
movq mm7, [mmx_one] |
movq mm7, [mmx_one] |
563 |
|
|
564 |
jz near .start1 |
jz near .start1 |
565 |
|
|
566 |
AVG2_MMX_RND1 |
AVG2_MMX_RND1 |
567 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
568 |
|
|
569 |
.start1 |
.start1: |
570 |
|
|
571 |
AVG2_MMX_RND1 |
AVG2_MMX_RND1 |
572 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
573 |
AVG2_MMX_RND1 |
AVG2_MMX_RND1 |
574 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
575 |
AVG2_MMX_RND1 |
AVG2_MMX_RND1 |
576 |
lea ecx, [ecx+2*edx] |
lea TMP0, [TMP0+2*TMP1] |
577 |
AVG2_MMX_RND1 |
AVG2_MMX_RND1 |
578 |
|
|
579 |
pop ebx |
pop _EBX |
580 |
ret |
ret |
581 |
|
ENDFUNC |
582 |
|
|
583 |
|
|
584 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
594 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
595 |
|
|
596 |
%macro AVG4_MMX_RND0 0 |
%macro AVG4_MMX_RND0 0 |
597 |
movq mm0, [eax] ; src1 -> mm0 |
movq mm0, [_EAX] ; src1 -> mm0 |
598 |
movq mm1, [ebx] ; src2 -> mm1 |
movq mm1, [_EBX] ; src2 -> mm1 |
599 |
|
|
600 |
movq mm2, mm0 |
movq mm2, mm0 |
601 |
movq mm3, mm1 |
movq mm3, mm1 |
609 |
psrlq mm0, 2 |
psrlq mm0, 2 |
610 |
psrlq mm1, 2 |
psrlq mm1, 2 |
611 |
|
|
612 |
lea eax, [eax+edx] |
lea _EAX, [_EAX+TMP1] |
613 |
lea ebx, [ebx+edx] |
lea _EBX, [_EBX+TMP1] |
614 |
|
|
615 |
paddb mm0, mm1 |
paddb mm0, mm1 |
616 |
paddb mm2, mm3 |
paddb mm2, mm3 |
617 |
|
|
618 |
movq mm4, [esi] ; src3 -> mm0 |
movq mm4, [_ESI] ; src3 -> mm0 |
619 |
movq mm5, [edi] ; src4 -> mm1 |
movq mm5, [_EDI] ; src4 -> mm1 |
620 |
|
|
621 |
movq mm1, mm4 |
movq mm1, mm4 |
622 |
movq mm3, mm5 |
movq mm3, mm5 |
642 |
psrlq mm2, 2 |
psrlq mm2, 2 |
643 |
paddb mm0, mm2 |
paddb mm0, mm2 |
644 |
|
|
645 |
lea esi, [esi+edx] |
lea _ESI, [_ESI+TMP1] |
646 |
lea edi, [edi+edx] |
lea _EDI, [_EDI+TMP1] |
647 |
|
|
648 |
movq [ecx], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
movq [TMP0], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
649 |
%endmacro |
%endmacro |
650 |
|
|
651 |
%macro AVG4_MMX_RND1 0 |
%macro AVG4_MMX_RND1 0 |
652 |
movq mm0, [eax] ; src1 -> mm0 |
movq mm0, [_EAX] ; src1 -> mm0 |
653 |
movq mm1, [ebx] ; src2 -> mm1 |
movq mm1, [_EBX] ; src2 -> mm1 |
654 |
|
|
655 |
movq mm2, mm0 |
movq mm2, mm0 |
656 |
movq mm3, mm1 |
movq mm3, mm1 |
664 |
psrlq mm0, 2 |
psrlq mm0, 2 |
665 |
psrlq mm1, 2 |
psrlq mm1, 2 |
666 |
|
|
667 |
lea eax,[eax+edx] |
lea _EAX,[_EAX+TMP1] |
668 |
lea ebx,[ebx+edx] |
lea _EBX,[_EBX+TMP1] |
669 |
|
|
670 |
paddb mm0, mm1 |
paddb mm0, mm1 |
671 |
paddb mm2, mm3 |
paddb mm2, mm3 |
672 |
|
|
673 |
movq mm4, [esi] ; src3 -> mm0 |
movq mm4, [_ESI] ; src3 -> mm0 |
674 |
movq mm5, [edi] ; src4 -> mm1 |
movq mm5, [_EDI] ; src4 -> mm1 |
675 |
|
|
676 |
movq mm1, mm4 |
movq mm1, mm4 |
677 |
movq mm3, mm5 |
movq mm3, mm5 |
697 |
psrlq mm2, 2 |
psrlq mm2, 2 |
698 |
paddb mm0, mm2 |
paddb mm0, mm2 |
699 |
|
|
700 |
lea esi,[esi+edx] |
lea _ESI,[_ESI+TMP1] |
701 |
lea edi,[edi+edx] |
lea _EDI,[_EDI+TMP1] |
702 |
|
|
703 |
movq [ecx], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
movq [TMP0], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
704 |
%endmacro |
%endmacro |
705 |
|
|
706 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
707 |
interpolate8x8_avg4_mmx: |
interpolate8x8_avg4_mmx: |
708 |
|
|
709 |
push ebx |
mov eax, prm7d ; rounding |
710 |
push edi |
test _EAX, _EAX |
711 |
push esi |
|
712 |
|
mov TMP0, prm1 ; dst -> edi |
713 |
mov eax, [esp + 12 + 28] ; rounding |
mov _EAX, prm5 ; src4 -> edi |
714 |
|
mov TMP1d, prm6d ; stride -> TMP1 |
715 |
test eax, eax |
|
716 |
|
|
717 |
mov ecx, [esp + 12 + 4] ; dst -> edi |
push _EBX |
718 |
mov eax, [esp + 12 + 8] ; src1 -> esi |
push _EDI |
719 |
mov ebx, [esp + 12 + 12] ; src2 -> eax |
push _ESI |
720 |
mov esi, [esp + 12 + 16] ; src3 -> esi |
|
721 |
mov edi, [esp + 12 + 20] ; src4 -> edi |
mov _EDI, _EAX |
722 |
mov edx, [esp + 12 + 24] ; stride -> edx |
|
723 |
|
%ifdef ARCH_IS_X86_64 |
724 |
|
mov _EAX, prm2 |
725 |
|
mov _EBX, prm3 |
726 |
|
mov _ESI, prm4 |
727 |
|
%else |
728 |
|
mov _EAX, [esp + 12 + 8] ; src1 -> esi |
729 |
|
mov _EBX, [esp + 12 + 12] ; src2 -> _EAX |
730 |
|
mov _ESI, [esp + 12 + 16] ; src3 -> esi |
731 |
|
%endif |
732 |
|
|
733 |
movq mm7, [mmx_one] |
movq mm7, [mmx_one] |
734 |
|
|
735 |
jnz near .rounding1 |
jnz near .rounding1 |
736 |
|
|
737 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
738 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
739 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
740 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
741 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
742 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
743 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
744 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
745 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
746 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
747 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
748 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
749 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
750 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
751 |
AVG4_MMX_RND0 |
AVG4_MMX_RND0 |
752 |
|
|
753 |
pop esi |
pop _ESI |
754 |
pop edi |
pop _EDI |
755 |
pop ebx |
pop _EBX |
756 |
ret |
ret |
757 |
|
|
758 |
.rounding1 |
.rounding1: |
759 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
760 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
761 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
762 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
763 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
764 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
765 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
766 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
767 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
768 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
769 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
770 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
771 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
772 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
773 |
AVG4_MMX_RND1 |
AVG4_MMX_RND1 |
774 |
|
|
775 |
pop esi |
pop _ESI |
776 |
pop edi |
pop _EDI |
777 |
pop ebx |
pop _EBX |
778 |
ret |
ret |
779 |
|
ENDFUNC |
780 |
|
|
781 |
|
|
782 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
789 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
790 |
|
|
791 |
%macro LOWPASS_6TAP_H_MMX 0 |
%macro LOWPASS_6TAP_H_MMX 0 |
792 |
movq mm0, [eax] |
movq mm0, [_EAX] |
793 |
movq mm2, [eax+1] |
movq mm2, [_EAX+1] |
794 |
|
|
795 |
movq mm1, mm0 |
movq mm1, mm0 |
796 |
movq mm3, mm2 |
movq mm3, mm2 |
807 |
psllw mm0, 2 |
psllw mm0, 2 |
808 |
psllw mm1, 2 |
psllw mm1, 2 |
809 |
|
|
810 |
movq mm2, [eax-1] |
movq mm2, [_EAX-1] |
811 |
movq mm4, [eax+2] |
movq mm4, [_EAX+2] |
812 |
|
|
813 |
movq mm3, mm2 |
movq mm3, mm2 |
814 |
movq mm5, mm4 |
movq mm5, mm4 |
828 |
pmullw mm0, [mmx_five] |
pmullw mm0, [mmx_five] |
829 |
pmullw mm1, [mmx_five] |
pmullw mm1, [mmx_five] |
830 |
|
|
831 |
movq mm2, [eax-2] |
movq mm2, [_EAX-2] |
832 |
movq mm4, [eax+3] |
movq mm4, [_EAX+3] |
833 |
|
|
834 |
movq mm3, mm2 |
movq mm3, mm2 |
835 |
movq mm5, mm4 |
movq mm5, mm4 |
852 |
psraw mm0, 5 |
psraw mm0, 5 |
853 |
psraw mm1, 5 |
psraw mm1, 5 |
854 |
|
|
855 |
lea eax, [eax+edx] |
lea _EAX, [_EAX+TMP1] |
856 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
857 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
858 |
%endmacro |
%endmacro |
859 |
|
|
860 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
861 |
interpolate8x8_6tap_lowpass_h_mmx: |
interpolate8x8_6tap_lowpass_h_mmx: |
862 |
|
|
863 |
mov eax, [esp + 16] ; rounding |
mov _EAX, prm4 ; rounding |
864 |
|
|
865 |
movq mm6, [rounding_lowpass_mmx + eax * 8] |
lea TMP0, [rounding_lowpass_mmx] |
866 |
|
movq mm6, [TMP0 + _EAX * 8] |
867 |
|
|
868 |
mov ecx, [esp + 4] ; dst -> edi |
mov TMP0, prm1 ; dst -> edi |
869 |
mov eax, [esp + 8] ; src -> esi |
mov _EAX, prm2 ; src -> esi |
870 |
mov edx, [esp + 12] ; stride -> edx |
mov TMP1, prm3 ; stride -> edx |
871 |
|
|
872 |
pxor mm7, mm7 |
pxor mm7, mm7 |
873 |
|
|
874 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
875 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
876 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
877 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
878 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
879 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
880 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
881 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
882 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
883 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
884 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
885 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
886 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
887 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
888 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
889 |
|
|
890 |
ret |
ret |
891 |
|
ENDFUNC |
892 |
|
|
893 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
894 |
; |
; |
900 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
901 |
|
|
902 |
%macro LOWPASS_6TAP_V_MMX 0 |
%macro LOWPASS_6TAP_V_MMX 0 |
903 |
movq mm0, [eax] |
movq mm0, [_EAX] |
904 |
movq mm2, [eax+edx] |
movq mm2, [_EAX+TMP1] |
905 |
|
|
906 |
movq mm1, mm0 |
movq mm1, mm0 |
907 |
movq mm3, mm2 |
movq mm3, mm2 |
918 |
psllw mm0, 2 |
psllw mm0, 2 |
919 |
psllw mm1, 2 |
psllw mm1, 2 |
920 |
|
|
921 |
movq mm4, [eax+2*edx] |
movq mm4, [_EAX+2*TMP1] |
922 |
sub eax, ebx |
sub _EAX, _EBX |
923 |
movq mm2, [eax+2*edx] |
movq mm2, [_EAX+2*TMP1] |
924 |
|
|
925 |
movq mm3, mm2 |
movq mm3, mm2 |
926 |
movq mm5, mm4 |
movq mm5, mm4 |
940 |
pmullw mm0, [mmx_five] |
pmullw mm0, [mmx_five] |
941 |
pmullw mm1, [mmx_five] |
pmullw mm1, [mmx_five] |
942 |
|
|
943 |
movq mm2, [eax+edx] |
movq mm2, [_EAX+TMP1] |
944 |
movq mm4, [eax+2*ebx] |
movq mm4, [_EAX+2*_EBX] |
945 |
|
|
946 |
movq mm3, mm2 |
movq mm3, mm2 |
947 |
movq mm5, mm4 |
movq mm5, mm4 |
964 |
psraw mm0, 5 |
psraw mm0, 5 |
965 |
psraw mm1, 5 |
psraw mm1, 5 |
966 |
|
|
967 |
lea eax, [eax+4*edx] |
lea _EAX, [_EAX+4*TMP1] |
968 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
969 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
970 |
%endmacro |
%endmacro |
971 |
|
|
972 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
973 |
interpolate8x8_6tap_lowpass_v_mmx: |
interpolate8x8_6tap_lowpass_v_mmx: |
974 |
|
|
975 |
push ebx |
mov _EAX, prm4 ; rounding |
|
|
|
|
mov eax, [esp + 4 + 16] ; rounding |
|
976 |
|
|
977 |
movq mm6, [rounding_lowpass_mmx + eax * 8] |
lea TMP0, [rounding_lowpass_mmx] |
978 |
|
movq mm6, [TMP0 + _EAX * 8] |
979 |
|
|
980 |
mov ecx, [esp + 4 + 4] ; dst -> edi |
mov TMP0, prm1 ; dst -> edi |
981 |
mov eax, [esp + 4 + 8] ; src -> esi |
mov _EAX, prm2 ; src -> esi |
982 |
mov edx, [esp + 4 + 12] ; stride -> edx |
mov TMP1, prm3 ; stride -> edx |
983 |
|
|
984 |
mov ebx, edx |
push _EBX |
985 |
shl ebx, 1 |
lea _EBX, [TMP1+TMP1*2] |
|
add ebx, edx |
|
986 |
|
|
987 |
pxor mm7, mm7 |
pxor mm7, mm7 |
988 |
|
|
989 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
990 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
991 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
992 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
993 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
994 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
995 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
996 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
997 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
998 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
999 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
1000 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
1001 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
1002 |
lea ecx, [ecx+edx] |
lea TMP0, [TMP0+TMP1] |
1003 |
LOWPASS_6TAP_V_MMX |
LOWPASS_6TAP_V_MMX |
1004 |
|
|
1005 |
pop ebx |
pop _EBX |
1006 |
ret |
ret |
1007 |
|
ENDFUNC |
1008 |
|
|
1009 |
;=========================================================================== |
;=========================================================================== |
1010 |
; |
; |
1015 |
;=========================================================================== |
;=========================================================================== |
1016 |
|
|
1017 |
%macro PROLOG0 0 |
%macro PROLOG0 0 |
1018 |
mov ecx, [esp+ 4] ; Dst |
mov TMP0, prm1 ; Dst |
1019 |
mov eax, [esp+ 8] ; Src |
mov _EAX, prm2 ; Src |
1020 |
mov edx, [esp+12] ; BpS |
mov TMP1, prm3 ; BpS |
1021 |
%endmacro |
%endmacro |
1022 |
|
|
1023 |
%macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder |
%macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder |
1085 |
;=========================================================================== |
;=========================================================================== |
1086 |
|
|
1087 |
%macro ADD_FF_MMX 1 |
%macro ADD_FF_MMX 1 |
1088 |
movq mm0, [eax] |
movq mm0, [_EAX] |
1089 |
movq mm2, [ecx] |
movq mm2, [TMP0] |
1090 |
movq mm1, mm0 |
movq mm1, mm0 |
1091 |
movq mm3, mm2 |
movq mm3, mm2 |
1092 |
%if (%1!=0) |
%if (%1!=0) |
1093 |
lea eax,[eax+%1*edx] |
lea _EAX,[_EAX+%1*TMP1] |
1094 |
%endif |
%endif |
1095 |
MIX |
MIX |
1096 |
paddusw mm0, mm5 ; rounder |
paddusw mm0, mm5 ; rounder |
1099 |
psrlw mm1, 1 |
psrlw mm1, 1 |
1100 |
|
|
1101 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
1102 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
1103 |
%if (%1!=0) |
%if (%1!=0) |
1104 |
lea ecx,[ecx+%1*edx] |
lea TMP0,[TMP0+%1*TMP1] |
1105 |
%endif |
%endif |
1106 |
%endmacro |
%endmacro |
1107 |
|
|
1108 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
1109 |
interpolate8x8_halfpel_add_mmx: |
interpolate8x8_halfpel_add_mmx: |
1110 |
PROLOG rounding1_mmx, 1 |
PROLOG rounding1_mmx, 1 |
1111 |
ADD_FF_MMX 1 |
ADD_FF_MMX 1 |
1117 |
ADD_FF_MMX 1 |
ADD_FF_MMX 1 |
1118 |
ADD_FF_MMX 0 |
ADD_FF_MMX 0 |
1119 |
ret |
ret |
1120 |
|
ENDFUNC |
1121 |
|
|
1122 |
;=========================================================================== |
;=========================================================================== |
1123 |
; |
; |
1130 |
;=========================================================================== |
;=========================================================================== |
1131 |
|
|
1132 |
%macro ADD_FH_MMX 0 |
%macro ADD_FH_MMX 0 |
1133 |
movq mm0, [eax] |
movq mm0, [_EAX] |
1134 |
movq mm2, [eax+1] |
movq mm2, [_EAX+1] |
1135 |
movq mm1, mm0 |
movq mm1, mm0 |
1136 |
movq mm3, mm2 |
movq mm3, mm2 |
1137 |
|
|
1138 |
lea eax,[eax+edx] |
lea _EAX,[_EAX+TMP1] |
1139 |
|
|
1140 |
MIX |
MIX |
1141 |
movq mm2, [ecx] ; prepare mix with Dst[0] |
movq mm2, [TMP0] ; prepare mix with Dst[0] |
1142 |
MIX_DST |
MIX_DST |
1143 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
1144 |
%endmacro |
%endmacro |
1145 |
|
|
1146 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
1147 |
interpolate8x8_halfpel_h_add_mmx: |
interpolate8x8_halfpel_h_add_mmx: |
1148 |
PROLOG rounding1_mmx, 1 |
PROLOG rounding1_mmx, 1 |
1149 |
|
|
1150 |
ADD_FH_MMX |
ADD_FH_MMX |
1151 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1152 |
ADD_FH_MMX |
ADD_FH_MMX |
1153 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1154 |
ADD_FH_MMX |
ADD_FH_MMX |
1155 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1156 |
ADD_FH_MMX |
ADD_FH_MMX |
1157 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1158 |
ADD_FH_MMX |
ADD_FH_MMX |
1159 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1160 |
ADD_FH_MMX |
ADD_FH_MMX |
1161 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1162 |
ADD_FH_MMX |
ADD_FH_MMX |
1163 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1164 |
ADD_FH_MMX |
ADD_FH_MMX |
1165 |
ret |
ret |
1166 |
|
ENDFUNC |
1167 |
|
|
1168 |
;=========================================================================== |
;=========================================================================== |
1169 |
; |
; |
1176 |
;=========================================================================== |
;=========================================================================== |
1177 |
|
|
1178 |
%macro ADD_HF_MMX 0 |
%macro ADD_HF_MMX 0 |
1179 |
movq mm0, [eax] |
movq mm0, [_EAX] |
1180 |
movq mm2, [eax+edx] |
movq mm2, [_EAX+TMP1] |
1181 |
movq mm1, mm0 |
movq mm1, mm0 |
1182 |
movq mm3, mm2 |
movq mm3, mm2 |
1183 |
|
|
1184 |
lea eax,[eax+edx] |
lea _EAX,[_EAX+TMP1] |
1185 |
|
|
1186 |
MIX |
MIX |
1187 |
movq mm2, [ecx] ; prepare mix with Dst[0] |
movq mm2, [TMP0] ; prepare mix with Dst[0] |
1188 |
MIX_DST |
MIX_DST |
1189 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
1190 |
|
|
1191 |
%endmacro |
%endmacro |
1192 |
|
|
1193 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
1194 |
interpolate8x8_halfpel_v_add_mmx: |
interpolate8x8_halfpel_v_add_mmx: |
1195 |
PROLOG rounding1_mmx, 1 |
PROLOG rounding1_mmx, 1 |
1196 |
|
|
1197 |
ADD_HF_MMX |
ADD_HF_MMX |
1198 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1199 |
ADD_HF_MMX |
ADD_HF_MMX |
1200 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1201 |
ADD_HF_MMX |
ADD_HF_MMX |
1202 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1203 |
ADD_HF_MMX |
ADD_HF_MMX |
1204 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1205 |
ADD_HF_MMX |
ADD_HF_MMX |
1206 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1207 |
ADD_HF_MMX |
ADD_HF_MMX |
1208 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1209 |
ADD_HF_MMX |
ADD_HF_MMX |
1210 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1211 |
ADD_HF_MMX |
ADD_HF_MMX |
1212 |
ret |
ret |
1213 |
|
ENDFUNC |
1214 |
|
|
1215 |
; The trick is to correct the result of 'pavgb' with some combination of the |
; The trick is to correct the result of 'pavgb' with some combination of the |
1216 |
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
1234 |
;=========================================================================== |
;=========================================================================== |
1235 |
|
|
1236 |
%macro ADD_HH_MMX 0 |
%macro ADD_HH_MMX 0 |
1237 |
lea eax,[eax+edx] |
lea _EAX,[_EAX+TMP1] |
1238 |
|
|
1239 |
; transfert prev line to mm0/mm1 |
; transfert prev line to mm0/mm1 |
1240 |
movq mm0, mm2 |
movq mm0, mm2 |
1241 |
movq mm1, mm3 |
movq mm1, mm3 |
1242 |
|
|
1243 |
; load new line in mm2/mm3 |
; load new line in mm2/mm3 |
1244 |
movq mm2, [eax] |
movq mm2, [_EAX] |
1245 |
movq mm4, [eax+1] |
movq mm4, [_EAX+1] |
1246 |
movq mm3, mm2 |
movq mm3, mm2 |
1247 |
movq mm5, mm4 |
movq mm5, mm4 |
1248 |
|
|
1259 |
paddusw mm0, mm2 |
paddusw mm0, mm2 |
1260 |
paddusw mm1, mm3 |
paddusw mm1, mm3 |
1261 |
|
|
1262 |
movq mm4, [ecx] ; prepare mix with Dst[0] |
movq mm4, [TMP0] ; prepare mix with Dst[0] |
1263 |
movq mm5, mm4 |
movq mm5, mm4 |
1264 |
|
|
1265 |
paddusw mm0, mm7 ; finish mixing current line |
paddusw mm0, mm7 ; finish mixing current line |
1282 |
|
|
1283 |
packuswb mm0, mm1 |
packuswb mm0, mm1 |
1284 |
|
|
1285 |
movq [ecx], mm0 |
movq [TMP0], mm0 |
1286 |
%endmacro |
%endmacro |
1287 |
|
|
1288 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
1289 |
interpolate8x8_halfpel_hv_add_mmx: |
interpolate8x8_halfpel_hv_add_mmx: |
1290 |
PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder |
PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder |
1291 |
|
|
1292 |
; preprocess first line |
; preprocess first line |
1293 |
movq mm0, [eax] |
movq mm0, [_EAX] |
1294 |
movq mm2, [eax+1] |
movq mm2, [_EAX+1] |
1295 |
movq mm1, mm0 |
movq mm1, mm0 |
1296 |
movq mm3, mm2 |
movq mm3, mm2 |
1297 |
|
|
1305 |
; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line |
; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line |
1306 |
|
|
1307 |
ADD_HH_MMX |
ADD_HH_MMX |
1308 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1309 |
ADD_HH_MMX |
ADD_HH_MMX |
1310 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1311 |
ADD_HH_MMX |
ADD_HH_MMX |
1312 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1313 |
ADD_HH_MMX |
ADD_HH_MMX |
1314 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1315 |
ADD_HH_MMX |
ADD_HH_MMX |
1316 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1317 |
ADD_HH_MMX |
ADD_HH_MMX |
1318 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1319 |
ADD_HH_MMX |
ADD_HH_MMX |
1320 |
lea ecx,[ecx+edx] |
lea TMP0,[TMP0+TMP1] |
1321 |
ADD_HH_MMX |
ADD_HH_MMX |
1322 |
|
|
1323 |
ret |
ret |
1324 |
|
ENDFUNC |
1325 |
|
|
1326 |
|
|
1327 |
|
%ifidn __OUTPUT_FORMAT__,elf |
1328 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
1329 |
|
%endif |
1330 |
|
|