1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * mmx 8x8 block-based halfpel interpolation |
; * - mmx 8x8 block-based halfpel interpolation - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
|
; * 2002 Michael Militzer <isibaar@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 |
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 |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
22 |
; * |
; * |
23 |
; * 06.07.2002 mmx cleanup - Isibaar |
; ****************************************************************************/ |
|
; * 22.12.2001 inital version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
24 |
|
|
25 |
bits 32 |
BITS 32 |
26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
33 |
%endif |
%endif |
34 |
%endmacro |
%endmacro |
35 |
|
|
36 |
section .data |
;============================================================================= |
37 |
|
; Read only data |
38 |
|
;============================================================================= |
39 |
|
|
40 |
align 16 |
%ifdef FORMAT_COFF |
41 |
|
SECTION .rodata data |
42 |
|
%else |
43 |
|
SECTION .rodata data align=16 |
44 |
|
%endif |
45 |
|
|
46 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
47 |
|
; (16 - r) rounding table |
48 |
|
;----------------------------------------------------------------------------- |
49 |
|
|
50 |
|
ALIGN 16 |
51 |
|
rounding_lowpass_mmx: |
52 |
|
times 4 dw 16 |
53 |
|
times 4 dw 15 |
54 |
|
|
55 |
|
;----------------------------------------------------------------------------- |
56 |
; (1 - r) rounding table |
; (1 - r) rounding table |
57 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
58 |
|
|
59 |
rounding1_mmx |
rounding1_mmx: |
60 |
times 4 dw 1 |
times 4 dw 1 |
61 |
times 4 dw 0 |
times 4 dw 0 |
62 |
|
|
63 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
64 |
; (2 - r) rounding table |
; (2 - r) rounding table |
65 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
66 |
|
|
67 |
rounding2_mmx |
rounding2_mmx: |
68 |
times 4 dw 2 |
times 4 dw 2 |
69 |
times 4 dw 1 |
times 4 dw 1 |
70 |
|
|
71 |
mmx_one |
mmx_one: |
72 |
times 8 db 1 |
times 8 db 1 |
73 |
|
|
74 |
section .text |
mmx_two: |
75 |
|
times 8 db 2 |
76 |
|
|
77 |
|
mmx_three: |
78 |
|
times 8 db 3 |
79 |
|
|
80 |
|
mmx_five: |
81 |
|
times 4 dw 5 |
82 |
|
|
83 |
|
mmx_mask: |
84 |
|
times 8 db 254 |
85 |
|
|
86 |
|
mmx_mask2: |
87 |
|
times 8 db 252 |
88 |
|
|
89 |
|
;============================================================================= |
90 |
|
; Code |
91 |
|
;============================================================================= |
92 |
|
|
93 |
|
SECTION .text |
94 |
|
|
95 |
|
cglobal interpolate8x8_halfpel_h_mmx |
96 |
|
cglobal interpolate8x8_halfpel_v_mmx |
97 |
|
cglobal interpolate8x8_halfpel_hv_mmx |
98 |
|
cglobal interpolate8x8_avg4_mmx |
99 |
|
cglobal interpolate8x8_avg2_mmx |
100 |
|
cglobal interpolate8x8_6tap_lowpass_h_mmx |
101 |
|
cglobal interpolate8x8_6tap_lowpass_v_mmx |
102 |
|
|
103 |
%macro CALC_AVG 6 |
%macro CALC_AVG 6 |
104 |
punpcklbw %3, %6 |
punpcklbw %3, %6 |
111 |
|
|
112 |
psrlw %1, 1 ; mm01 >>= 1 |
psrlw %1, 1 ; mm01 >>= 1 |
113 |
psrlw %2, 1 |
psrlw %2, 1 |
|
|
|
114 |
%endmacro |
%endmacro |
115 |
|
|
116 |
|
|
117 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
118 |
; |
; |
119 |
; void interpolate8x8_halfpel_h_mmx(uint8_t * const dst, |
; void interpolate8x8_halfpel_h_mmx(uint8_t * const dst, |
120 |
; const uint8_t * const src, |
; const uint8_t * const src, |
121 |
; const uint32_t stride, |
; const uint32_t stride, |
122 |
; const uint32_t rounding); |
; const uint32_t rounding); |
123 |
; |
; |
124 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
125 |
|
|
126 |
%macro COPY_H_MMX 0 |
%macro COPY_H_MMX 0 |
127 |
movq mm0, [esi] |
movq mm0, [esi] |
141 |
add edi, edx ; dst += stride |
add edi, edx ; dst += stride |
142 |
%endmacro |
%endmacro |
143 |
|
|
144 |
align 16 |
ALIGN 16 |
145 |
cglobal interpolate8x8_halfpel_h_mmx |
interpolate8x8_halfpel_h_mmx: |
|
interpolate8x8_halfpel_h_mmx |
|
146 |
|
|
147 |
push esi |
push esi |
148 |
push edi |
push edi |
|
|
|
149 |
mov eax, [esp + 8 + 16] ; rounding |
mov eax, [esp + 8 + 16] ; rounding |
150 |
|
|
|
interpolate8x8_halfpel_h_mmx.start |
|
151 |
movq mm7, [rounding1_mmx + eax * 8] |
movq mm7, [rounding1_mmx + eax * 8] |
152 |
|
|
153 |
mov edi, [esp + 8 + 4] ; dst |
mov edi, [esp + 8 + 4] ; dst |
171 |
ret |
ret |
172 |
|
|
173 |
|
|
174 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
175 |
; |
; |
176 |
; void interpolate8x8_halfpel_v_mmx(uint8_t * const dst, |
; void interpolate8x8_halfpel_v_mmx(uint8_t * const dst, |
177 |
; const uint8_t * const src, |
; const uint8_t * const src, |
178 |
; const uint32_t stride, |
; const uint32_t stride, |
179 |
; const uint32_t rounding); |
; const uint32_t rounding); |
180 |
; |
; |
181 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
182 |
|
|
183 |
%macro COPY_V_MMX 0 |
%macro COPY_V_MMX 0 |
184 |
movq mm0, [esi] |
movq mm0, [esi] |
198 |
add edi, edx ; dst += stride |
add edi, edx ; dst += stride |
199 |
%endmacro |
%endmacro |
200 |
|
|
201 |
align 16 |
ALIGN 16 |
202 |
cglobal interpolate8x8_halfpel_v_mmx |
interpolate8x8_halfpel_v_mmx: |
|
interpolate8x8_halfpel_v_mmx |
|
203 |
|
|
204 |
push esi |
push esi |
205 |
push edi |
push edi |
206 |
|
|
207 |
mov eax, [esp + 8 + 16] ; rounding |
mov eax, [esp + 8 + 16] ; rounding |
208 |
|
|
|
interpolate8x8_halfpel_v_mmx.start |
|
209 |
movq mm7, [rounding1_mmx + eax * 8] |
movq mm7, [rounding1_mmx + eax * 8] |
210 |
|
|
211 |
mov edi, [esp + 8 + 4] ; dst |
mov edi, [esp + 8 + 4] ; dst |
230 |
ret |
ret |
231 |
|
|
232 |
|
|
233 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
234 |
; |
; |
235 |
; void interpolate8x8_halfpel_hv_mmx(uint8_t * const dst, |
; void interpolate8x8_halfpel_hv_mmx(uint8_t * const dst, |
236 |
; const uint8_t * const src, |
; const uint8_t * const src, |
238 |
; const uint32_t rounding); |
; const uint32_t rounding); |
239 |
; |
; |
240 |
; |
; |
241 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
242 |
|
|
243 |
%macro COPY_HV_MMX 0 |
%macro COPY_HV_MMX 0 |
244 |
; current row |
; current row |
|
|
|
245 |
movq mm0, [esi] |
movq mm0, [esi] |
246 |
movq mm2, [esi + 1] |
movq mm2, [esi + 1] |
247 |
|
|
257 |
paddusw mm1, mm3 |
paddusw mm1, mm3 |
258 |
|
|
259 |
; next row |
; next row |
|
|
|
260 |
movq mm4, [esi + edx] |
movq mm4, [esi + edx] |
261 |
movq mm2, [esi + edx + 1] |
movq mm2, [esi + edx + 1] |
262 |
|
|
272 |
paddusw mm5, mm3 |
paddusw mm5, mm3 |
273 |
|
|
274 |
; add current + next row |
; add current + next row |
|
|
|
275 |
paddusw mm0, mm4 ; mm01 += mm45 |
paddusw mm0, mm4 ; mm01 += mm45 |
276 |
paddusw mm1, mm5 |
paddusw mm1, mm5 |
277 |
paddusw mm0, mm7 ; mm01 += rounding2 |
paddusw mm0, mm7 ; mm01 += rounding2 |
287 |
add edi, edx ; dst += stride |
add edi, edx ; dst += stride |
288 |
%endmacro |
%endmacro |
289 |
|
|
290 |
align 16 |
ALIGN 16 |
291 |
cglobal interpolate8x8_halfpel_hv_mmx |
interpolate8x8_halfpel_hv_mmx: |
|
interpolate8x8_halfpel_hv_mmx |
|
292 |
|
|
293 |
push esi |
push esi |
294 |
push edi |
push edi |
295 |
|
|
296 |
mov eax, [esp + 8 + 16] ; rounding |
mov eax, [esp + 8 + 16] ; rounding |
|
interpolate8x8_halfpel_hv_mmx.start |
|
297 |
|
|
298 |
movq mm7, [rounding2_mmx + eax * 8] |
movq mm7, [rounding2_mmx + eax * 8] |
299 |
|
|
319 |
pop esi |
pop esi |
320 |
|
|
321 |
ret |
ret |
322 |
|
|
323 |
|
;----------------------------------------------------------------------------- |
324 |
|
; |
325 |
|
; void interpolate8x8_avg2_mmx(uint8_t const *dst, |
326 |
|
; const uint8_t * const src1, |
327 |
|
; const uint8_t * const src2, |
328 |
|
; const uint32_t stride, |
329 |
|
; const uint32_t rounding, |
330 |
|
; const uint32_t height); |
331 |
|
; |
332 |
|
;----------------------------------------------------------------------------- |
333 |
|
|
334 |
|
%macro AVG2_MMX_RND0 0 |
335 |
|
movq mm0, [eax] ; src1 -> mm0 |
336 |
|
movq mm1, [ebx] ; src2 -> mm1 |
337 |
|
|
338 |
|
movq mm4, [eax+edx] |
339 |
|
movq mm5, [ebx+edx] |
340 |
|
|
341 |
|
movq mm2, mm0 ; src1 -> mm2 |
342 |
|
movq mm3, mm1 ; src2 -> mm3 |
343 |
|
|
344 |
|
pand mm2, mm7 ; isolate the lsb |
345 |
|
pand mm3, mm7 ; isolate the lsb |
346 |
|
|
347 |
|
por mm2, mm3 ; ODD(src1) OR ODD(src2) -> mm2 |
348 |
|
|
349 |
|
movq mm3, mm4 |
350 |
|
movq mm6, mm5 |
351 |
|
|
352 |
|
pand mm3, mm7 |
353 |
|
pand mm6, mm7 |
354 |
|
|
355 |
|
por mm3, mm6 |
356 |
|
|
357 |
|
pand mm0, [mmx_mask] |
358 |
|
pand mm1, [mmx_mask] |
359 |
|
pand mm4, [mmx_mask] |
360 |
|
pand mm5, [mmx_mask] |
361 |
|
|
362 |
|
psrlq mm0, 1 ; src1 / 2 |
363 |
|
psrlq mm1, 1 ; src2 / 2 |
364 |
|
|
365 |
|
psrlq mm4, 1 |
366 |
|
psrlq mm5, 1 |
367 |
|
|
368 |
|
paddb mm0, mm1 ; src1/2 + src2/2 -> mm0 |
369 |
|
paddb mm0, mm2 ; correct rounding error |
370 |
|
|
371 |
|
paddb mm4, mm5 |
372 |
|
paddb mm4, mm3 |
373 |
|
|
374 |
|
lea eax, [eax+2*edx] |
375 |
|
lea ebx, [ebx+2*edx] |
376 |
|
|
377 |
|
movq [ecx], mm0 ; (src1 + src2 + 1) / 2 -> dst |
378 |
|
movq [ecx+edx], mm4 |
379 |
|
%endmacro |
380 |
|
|
381 |
|
%macro AVG2_MMX_RND1 0 |
382 |
|
movq mm0, [eax] ; src1 -> mm0 |
383 |
|
movq mm1, [ebx] ; src2 -> mm1 |
384 |
|
|
385 |
|
movq mm4, [eax+edx] |
386 |
|
movq mm5, [ebx+edx] |
387 |
|
|
388 |
|
movq mm2, mm0 ; src1 -> mm2 |
389 |
|
movq mm3, mm1 ; src2 -> mm3 |
390 |
|
|
391 |
|
pand mm2, mm7 ; isolate the lsb |
392 |
|
pand mm3, mm7 ; isolate the lsb |
393 |
|
|
394 |
|
pand mm2, mm3 ; ODD(src1) AND ODD(src2) -> mm2 |
395 |
|
|
396 |
|
movq mm3, mm4 |
397 |
|
movq mm6, mm5 |
398 |
|
|
399 |
|
pand mm3, mm7 |
400 |
|
pand mm6, mm7 |
401 |
|
|
402 |
|
pand mm3, mm6 |
403 |
|
|
404 |
|
pand mm0, [mmx_mask] |
405 |
|
pand mm1, [mmx_mask] |
406 |
|
pand mm4, [mmx_mask] |
407 |
|
pand mm5, [mmx_mask] |
408 |
|
|
409 |
|
psrlq mm0, 1 ; src1 / 2 |
410 |
|
psrlq mm1, 1 ; src2 / 2 |
411 |
|
|
412 |
|
psrlq mm4, 1 |
413 |
|
psrlq mm5, 1 |
414 |
|
|
415 |
|
paddb mm0, mm1 ; src1/2 + src2/2 -> mm0 |
416 |
|
paddb mm0, mm2 ; correct rounding error |
417 |
|
|
418 |
|
paddb mm4, mm5 |
419 |
|
paddb mm4, mm3 |
420 |
|
|
421 |
|
lea eax, [eax+2*edx] |
422 |
|
lea ebx, [ebx+2*edx] |
423 |
|
|
424 |
|
movq [ecx], mm0 ; (src1 + src2 + 1) / 2 -> dst |
425 |
|
movq [ecx+edx], mm4 |
426 |
|
%endmacro |
427 |
|
|
428 |
|
ALIGN 16 |
429 |
|
interpolate8x8_avg2_mmx: |
430 |
|
|
431 |
|
push ebx |
432 |
|
|
433 |
|
mov eax, [esp + 4 + 20] ; rounding |
434 |
|
test eax, eax |
435 |
|
|
436 |
|
jnz near .rounding1 |
437 |
|
|
438 |
|
mov eax, [esp + 4 + 24] ; height -> eax |
439 |
|
sub eax, 8 |
440 |
|
test eax, eax |
441 |
|
|
442 |
|
mov ecx, [esp + 4 + 4] ; dst -> edi |
443 |
|
mov eax, [esp + 4 + 8] ; src1 -> esi |
444 |
|
mov ebx, [esp + 4 + 12] ; src2 -> eax |
445 |
|
mov edx, [esp + 4 + 16] ; stride -> edx |
446 |
|
|
447 |
|
movq mm7, [mmx_one] |
448 |
|
|
449 |
|
jz near .start0 |
450 |
|
|
451 |
|
AVG2_MMX_RND0 |
452 |
|
lea ecx, [ecx+2*edx] |
453 |
|
|
454 |
|
.start0 |
455 |
|
|
456 |
|
AVG2_MMX_RND0 |
457 |
|
lea ecx, [ecx+2*edx] |
458 |
|
AVG2_MMX_RND0 |
459 |
|
lea ecx, [ecx+2*edx] |
460 |
|
AVG2_MMX_RND0 |
461 |
|
lea ecx, [ecx+2*edx] |
462 |
|
AVG2_MMX_RND0 |
463 |
|
|
464 |
|
pop ebx |
465 |
|
ret |
466 |
|
|
467 |
|
.rounding1 |
468 |
|
mov eax, [esp + 4 + 24] ; height -> eax |
469 |
|
sub eax, 8 |
470 |
|
test eax, eax |
471 |
|
|
472 |
|
mov ecx, [esp + 4 + 4] ; dst -> edi |
473 |
|
mov eax, [esp + 4 + 8] ; src1 -> esi |
474 |
|
mov ebx, [esp + 4 + 12] ; src2 -> eax |
475 |
|
mov edx, [esp + 4 + 16] ; stride -> edx |
476 |
|
|
477 |
|
movq mm7, [mmx_one] |
478 |
|
|
479 |
|
jz near .start1 |
480 |
|
|
481 |
|
AVG2_MMX_RND1 |
482 |
|
lea ecx, [ecx+2*edx] |
483 |
|
|
484 |
|
.start1 |
485 |
|
|
486 |
|
AVG2_MMX_RND1 |
487 |
|
lea ecx, [ecx+2*edx] |
488 |
|
AVG2_MMX_RND1 |
489 |
|
lea ecx, [ecx+2*edx] |
490 |
|
AVG2_MMX_RND1 |
491 |
|
lea ecx, [ecx+2*edx] |
492 |
|
AVG2_MMX_RND1 |
493 |
|
|
494 |
|
pop ebx |
495 |
|
ret |
496 |
|
|
497 |
|
|
498 |
|
;----------------------------------------------------------------------------- |
499 |
|
; |
500 |
|
; void interpolate8x8_avg4_mmx(uint8_t const *dst, |
501 |
|
; const uint8_t * const src1, |
502 |
|
; const uint8_t * const src2, |
503 |
|
; const uint8_t * const src3, |
504 |
|
; const uint8_t * const src4, |
505 |
|
; const uint32_t stride, |
506 |
|
; const uint32_t rounding); |
507 |
|
; |
508 |
|
;----------------------------------------------------------------------------- |
509 |
|
|
510 |
|
%macro AVG4_MMX_RND0 0 |
511 |
|
movq mm0, [eax] ; src1 -> mm0 |
512 |
|
movq mm1, [ebx] ; src2 -> mm1 |
513 |
|
|
514 |
|
movq mm2, mm0 |
515 |
|
movq mm3, mm1 |
516 |
|
|
517 |
|
pand mm2, [mmx_three] |
518 |
|
pand mm3, [mmx_three] |
519 |
|
|
520 |
|
pand mm0, [mmx_mask2] |
521 |
|
pand mm1, [mmx_mask2] |
522 |
|
|
523 |
|
psrlq mm0, 2 |
524 |
|
psrlq mm1, 2 |
525 |
|
|
526 |
|
lea eax, [eax+edx] |
527 |
|
lea ebx, [ebx+edx] |
528 |
|
|
529 |
|
paddb mm0, mm1 |
530 |
|
paddb mm2, mm3 |
531 |
|
|
532 |
|
movq mm4, [esi] ; src3 -> mm0 |
533 |
|
movq mm5, [edi] ; src4 -> mm1 |
534 |
|
|
535 |
|
movq mm1, mm4 |
536 |
|
movq mm3, mm5 |
537 |
|
|
538 |
|
pand mm1, [mmx_three] |
539 |
|
pand mm3, [mmx_three] |
540 |
|
|
541 |
|
pand mm4, [mmx_mask2] |
542 |
|
pand mm5, [mmx_mask2] |
543 |
|
|
544 |
|
psrlq mm4, 2 |
545 |
|
psrlq mm5, 2 |
546 |
|
|
547 |
|
paddb mm4, mm5 |
548 |
|
paddb mm0, mm4 |
549 |
|
|
550 |
|
paddb mm1, mm3 |
551 |
|
paddb mm2, mm1 |
552 |
|
|
553 |
|
paddb mm2, [mmx_two] |
554 |
|
pand mm2, [mmx_mask2] |
555 |
|
|
556 |
|
psrlq mm2, 2 |
557 |
|
paddb mm0, mm2 |
558 |
|
|
559 |
|
lea esi, [esi+edx] |
560 |
|
lea edi, [edi+edx] |
561 |
|
|
562 |
|
movq [ecx], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
563 |
|
%endmacro |
564 |
|
|
565 |
|
%macro AVG4_MMX_RND1 0 |
566 |
|
movq mm0, [eax] ; src1 -> mm0 |
567 |
|
movq mm1, [ebx] ; src2 -> mm1 |
568 |
|
|
569 |
|
movq mm2, mm0 |
570 |
|
movq mm3, mm1 |
571 |
|
|
572 |
|
pand mm2, [mmx_three] |
573 |
|
pand mm3, [mmx_three] |
574 |
|
|
575 |
|
pand mm0, [mmx_mask2] |
576 |
|
pand mm1, [mmx_mask2] |
577 |
|
|
578 |
|
psrlq mm0, 2 |
579 |
|
psrlq mm1, 2 |
580 |
|
|
581 |
|
lea eax,[eax+edx] |
582 |
|
lea ebx,[ebx+edx] |
583 |
|
|
584 |
|
paddb mm0, mm1 |
585 |
|
paddb mm2, mm3 |
586 |
|
|
587 |
|
movq mm4, [esi] ; src3 -> mm0 |
588 |
|
movq mm5, [edi] ; src4 -> mm1 |
589 |
|
|
590 |
|
movq mm1, mm4 |
591 |
|
movq mm3, mm5 |
592 |
|
|
593 |
|
pand mm1, [mmx_three] |
594 |
|
pand mm3, [mmx_three] |
595 |
|
|
596 |
|
pand mm4, [mmx_mask2] |
597 |
|
pand mm5, [mmx_mask2] |
598 |
|
|
599 |
|
psrlq mm4, 2 |
600 |
|
psrlq mm5, 2 |
601 |
|
|
602 |
|
paddb mm4, mm5 |
603 |
|
paddb mm0, mm4 |
604 |
|
|
605 |
|
paddb mm1, mm3 |
606 |
|
paddb mm2, mm1 |
607 |
|
|
608 |
|
paddb mm2, [mmx_one] |
609 |
|
pand mm2, [mmx_mask2] |
610 |
|
|
611 |
|
psrlq mm2, 2 |
612 |
|
paddb mm0, mm2 |
613 |
|
|
614 |
|
lea esi,[esi+edx] |
615 |
|
lea edi,[edi+edx] |
616 |
|
|
617 |
|
movq [ecx], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
618 |
|
%endmacro |
619 |
|
|
620 |
|
ALIGN 16 |
621 |
|
interpolate8x8_avg4_mmx: |
622 |
|
|
623 |
|
push ebx |
624 |
|
push edi |
625 |
|
push esi |
626 |
|
|
627 |
|
mov eax, [esp + 12 + 28] ; rounding |
628 |
|
|
629 |
|
test eax, eax |
630 |
|
|
631 |
|
mov ecx, [esp + 12 + 4] ; dst -> edi |
632 |
|
mov eax, [esp + 12 + 8] ; src1 -> esi |
633 |
|
mov ebx, [esp + 12 + 12] ; src2 -> eax |
634 |
|
mov esi, [esp + 12 + 16] ; src3 -> esi |
635 |
|
mov edi, [esp + 12 + 20] ; src4 -> edi |
636 |
|
mov edx, [esp + 12 + 24] ; stride -> edx |
637 |
|
|
638 |
|
movq mm7, [mmx_one] |
639 |
|
|
640 |
|
jnz near .rounding1 |
641 |
|
|
642 |
|
AVG4_MMX_RND0 |
643 |
|
lea ecx, [ecx+edx] |
644 |
|
AVG4_MMX_RND0 |
645 |
|
lea ecx, [ecx+edx] |
646 |
|
AVG4_MMX_RND0 |
647 |
|
lea ecx, [ecx+edx] |
648 |
|
AVG4_MMX_RND0 |
649 |
|
lea ecx, [ecx+edx] |
650 |
|
AVG4_MMX_RND0 |
651 |
|
lea ecx, [ecx+edx] |
652 |
|
AVG4_MMX_RND0 |
653 |
|
lea ecx, [ecx+edx] |
654 |
|
AVG4_MMX_RND0 |
655 |
|
lea ecx, [ecx+edx] |
656 |
|
AVG4_MMX_RND0 |
657 |
|
|
658 |
|
pop esi |
659 |
|
pop edi |
660 |
|
pop ebx |
661 |
|
ret |
662 |
|
|
663 |
|
.rounding1 |
664 |
|
AVG4_MMX_RND1 |
665 |
|
lea ecx, [ecx+edx] |
666 |
|
AVG4_MMX_RND1 |
667 |
|
lea ecx, [ecx+edx] |
668 |
|
AVG4_MMX_RND1 |
669 |
|
lea ecx, [ecx+edx] |
670 |
|
AVG4_MMX_RND1 |
671 |
|
lea ecx, [ecx+edx] |
672 |
|
AVG4_MMX_RND1 |
673 |
|
lea ecx, [ecx+edx] |
674 |
|
AVG4_MMX_RND1 |
675 |
|
lea ecx, [ecx+edx] |
676 |
|
AVG4_MMX_RND1 |
677 |
|
lea ecx, [ecx+edx] |
678 |
|
AVG4_MMX_RND1 |
679 |
|
|
680 |
|
pop esi |
681 |
|
pop edi |
682 |
|
pop ebx |
683 |
|
ret |
684 |
|
|
685 |
|
|
686 |
|
;----------------------------------------------------------------------------- |
687 |
|
; |
688 |
|
; void interpolate8x8_6tap_lowpass_h_mmx(uint8_t const *dst, |
689 |
|
; const uint8_t * const src, |
690 |
|
; const uint32_t stride, |
691 |
|
; const uint32_t rounding); |
692 |
|
; |
693 |
|
;----------------------------------------------------------------------------- |
694 |
|
|
695 |
|
%macro LOWPASS_6TAP_H_MMX 0 |
696 |
|
movq mm0, [eax] |
697 |
|
movq mm2, [eax+1] |
698 |
|
|
699 |
|
movq mm1, mm0 |
700 |
|
movq mm3, mm2 |
701 |
|
|
702 |
|
punpcklbw mm0, mm7 |
703 |
|
punpcklbw mm2, mm7 |
704 |
|
|
705 |
|
punpckhbw mm1, mm7 |
706 |
|
punpckhbw mm3, mm7 |
707 |
|
|
708 |
|
paddw mm0, mm2 |
709 |
|
paddw mm1, mm3 |
710 |
|
|
711 |
|
psllw mm0, 2 |
712 |
|
psllw mm1, 2 |
713 |
|
|
714 |
|
movq mm2, [eax-1] |
715 |
|
movq mm4, [eax+2] |
716 |
|
|
717 |
|
movq mm3, mm2 |
718 |
|
movq mm5, mm4 |
719 |
|
|
720 |
|
punpcklbw mm2, mm7 |
721 |
|
punpcklbw mm4, mm7 |
722 |
|
|
723 |
|
punpckhbw mm3, mm7 |
724 |
|
punpckhbw mm5, mm7 |
725 |
|
|
726 |
|
paddw mm2, mm4 |
727 |
|
paddw mm3, mm5 |
728 |
|
|
729 |
|
psubsw mm0, mm2 |
730 |
|
psubsw mm1, mm3 |
731 |
|
|
732 |
|
pmullw mm0, [mmx_five] |
733 |
|
pmullw mm1, [mmx_five] |
734 |
|
|
735 |
|
movq mm2, [eax-2] |
736 |
|
movq mm4, [eax+3] |
737 |
|
|
738 |
|
movq mm3, mm2 |
739 |
|
movq mm5, mm4 |
740 |
|
|
741 |
|
punpcklbw mm2, mm7 |
742 |
|
punpcklbw mm4, mm7 |
743 |
|
|
744 |
|
punpckhbw mm3, mm7 |
745 |
|
punpckhbw mm5, mm7 |
746 |
|
|
747 |
|
paddw mm2, mm4 |
748 |
|
paddw mm3, mm5 |
749 |
|
|
750 |
|
paddsw mm0, mm2 |
751 |
|
paddsw mm1, mm3 |
752 |
|
|
753 |
|
paddsw mm0, mm6 |
754 |
|
paddsw mm1, mm6 |
755 |
|
|
756 |
|
psraw mm0, 5 |
757 |
|
psraw mm1, 5 |
758 |
|
|
759 |
|
lea eax, [eax+edx] |
760 |
|
packuswb mm0, mm1 |
761 |
|
movq [ecx], mm0 |
762 |
|
%endmacro |
763 |
|
|
764 |
|
ALIGN 16 |
765 |
|
interpolate8x8_6tap_lowpass_h_mmx: |
766 |
|
|
767 |
|
mov eax, [esp + 16] ; rounding |
768 |
|
|
769 |
|
movq mm6, [rounding_lowpass_mmx + eax * 8] |
770 |
|
|
771 |
|
mov ecx, [esp + 4] ; dst -> edi |
772 |
|
mov eax, [esp + 8] ; src -> esi |
773 |
|
mov edx, [esp + 12] ; stride -> edx |
774 |
|
|
775 |
|
pxor mm7, mm7 |
776 |
|
|
777 |
|
LOWPASS_6TAP_H_MMX |
778 |
|
lea ecx, [ecx+edx] |
779 |
|
LOWPASS_6TAP_H_MMX |
780 |
|
lea ecx, [ecx+edx] |
781 |
|
LOWPASS_6TAP_H_MMX |
782 |
|
lea ecx, [ecx+edx] |
783 |
|
LOWPASS_6TAP_H_MMX |
784 |
|
lea ecx, [ecx+edx] |
785 |
|
LOWPASS_6TAP_H_MMX |
786 |
|
lea ecx, [ecx+edx] |
787 |
|
LOWPASS_6TAP_H_MMX |
788 |
|
lea ecx, [ecx+edx] |
789 |
|
LOWPASS_6TAP_H_MMX |
790 |
|
lea ecx, [ecx+edx] |
791 |
|
LOWPASS_6TAP_H_MMX |
792 |
|
|
793 |
|
ret |
794 |
|
|
795 |
|
;----------------------------------------------------------------------------- |
796 |
|
; |
797 |
|
; void interpolate8x8_6tap_lowpass_v_mmx(uint8_t const *dst, |
798 |
|
; const uint8_t * const src, |
799 |
|
; const uint32_t stride, |
800 |
|
; const uint32_t rounding); |
801 |
|
; |
802 |
|
;----------------------------------------------------------------------------- |
803 |
|
|
804 |
|
%macro LOWPASS_6TAP_V_MMX 0 |
805 |
|
movq mm0, [eax] |
806 |
|
movq mm2, [eax+edx] |
807 |
|
|
808 |
|
movq mm1, mm0 |
809 |
|
movq mm3, mm2 |
810 |
|
|
811 |
|
punpcklbw mm0, mm7 |
812 |
|
punpcklbw mm2, mm7 |
813 |
|
|
814 |
|
punpckhbw mm1, mm7 |
815 |
|
punpckhbw mm3, mm7 |
816 |
|
|
817 |
|
paddw mm0, mm2 |
818 |
|
paddw mm1, mm3 |
819 |
|
|
820 |
|
psllw mm0, 2 |
821 |
|
psllw mm1, 2 |
822 |
|
|
823 |
|
movq mm4, [eax+2*edx] |
824 |
|
sub eax, ebx |
825 |
|
movq mm2, [eax+2*edx] |
826 |
|
|
827 |
|
movq mm3, mm2 |
828 |
|
movq mm5, mm4 |
829 |
|
|
830 |
|
punpcklbw mm2, mm7 |
831 |
|
punpcklbw mm4, mm7 |
832 |
|
|
833 |
|
punpckhbw mm3, mm7 |
834 |
|
punpckhbw mm5, mm7 |
835 |
|
|
836 |
|
paddw mm2, mm4 |
837 |
|
paddw mm3, mm5 |
838 |
|
|
839 |
|
psubsw mm0, mm2 |
840 |
|
psubsw mm1, mm3 |
841 |
|
|
842 |
|
pmullw mm0, [mmx_five] |
843 |
|
pmullw mm1, [mmx_five] |
844 |
|
|
845 |
|
movq mm2, [eax+edx] |
846 |
|
movq mm4, [eax+2*ebx] |
847 |
|
|
848 |
|
movq mm3, mm2 |
849 |
|
movq mm5, mm4 |
850 |
|
|
851 |
|
punpcklbw mm2, mm7 |
852 |
|
punpcklbw mm4, mm7 |
853 |
|
|
854 |
|
punpckhbw mm3, mm7 |
855 |
|
punpckhbw mm5, mm7 |
856 |
|
|
857 |
|
paddw mm2, mm4 |
858 |
|
paddw mm3, mm5 |
859 |
|
|
860 |
|
paddsw mm0, mm2 |
861 |
|
paddsw mm1, mm3 |
862 |
|
|
863 |
|
paddsw mm0, mm6 |
864 |
|
paddsw mm1, mm6 |
865 |
|
|
866 |
|
psraw mm0, 5 |
867 |
|
psraw mm1, 5 |
868 |
|
|
869 |
|
lea eax, [eax+4*edx] |
870 |
|
packuswb mm0, mm1 |
871 |
|
movq [ecx], mm0 |
872 |
|
%endmacro |
873 |
|
|
874 |
|
ALIGN 16 |
875 |
|
interpolate8x8_6tap_lowpass_v_mmx: |
876 |
|
|
877 |
|
push ebx |
878 |
|
|
879 |
|
mov eax, [esp + 4 + 16] ; rounding |
880 |
|
|
881 |
|
movq mm6, [rounding_lowpass_mmx + eax * 8] |
882 |
|
|
883 |
|
mov ecx, [esp + 4 + 4] ; dst -> edi |
884 |
|
mov eax, [esp + 4 + 8] ; src -> esi |
885 |
|
mov edx, [esp + 4 + 12] ; stride -> edx |
886 |
|
|
887 |
|
mov ebx, edx |
888 |
|
shl ebx, 1 |
889 |
|
add ebx, edx |
890 |
|
|
891 |
|
pxor mm7, mm7 |
892 |
|
|
893 |
|
LOWPASS_6TAP_V_MMX |
894 |
|
lea ecx, [ecx+edx] |
895 |
|
LOWPASS_6TAP_V_MMX |
896 |
|
lea ecx, [ecx+edx] |
897 |
|
LOWPASS_6TAP_V_MMX |
898 |
|
lea ecx, [ecx+edx] |
899 |
|
LOWPASS_6TAP_V_MMX |
900 |
|
lea ecx, [ecx+edx] |
901 |
|
LOWPASS_6TAP_V_MMX |
902 |
|
lea ecx, [ecx+edx] |
903 |
|
LOWPASS_6TAP_V_MMX |
904 |
|
lea ecx, [ecx+edx] |
905 |
|
LOWPASS_6TAP_V_MMX |
906 |
|
lea ecx, [ecx+edx] |
907 |
|
LOWPASS_6TAP_V_MMX |
908 |
|
|
909 |
|
pop ebx |
910 |
|
ret |