29 |
|
|
30 |
%macro cglobal 1 |
%macro cglobal 1 |
31 |
%ifdef PREFIX |
%ifdef PREFIX |
32 |
|
%ifdef MARK_FUNCS |
33 |
|
global _%1:function %1.endfunc-%1 |
34 |
|
%define %1 _%1:function %1.endfunc-%1 |
35 |
|
%define ENDFUNC .endfunc |
36 |
|
%else |
37 |
global _%1 |
global _%1 |
38 |
%define %1 _%1 |
%define %1 _%1 |
39 |
|
%define ENDFUNC |
40 |
|
%endif |
41 |
|
%else |
42 |
|
%ifdef MARK_FUNCS |
43 |
|
global %1:function %1.endfunc-%1 |
44 |
|
%define ENDFUNC .endfunc |
45 |
%else |
%else |
46 |
global %1 |
global %1 |
47 |
|
%define ENDFUNC |
48 |
|
%endif |
49 |
%endif |
%endif |
50 |
%endmacro |
%endmacro |
51 |
|
|
54 |
;============================================================================= |
;============================================================================= |
55 |
|
|
56 |
%ifdef FORMAT_COFF |
%ifdef FORMAT_COFF |
57 |
SECTION .rodata data |
SECTION .rodata |
58 |
%else |
%else |
59 |
SECTION .rodata data align=16 |
SECTION .rodata align=16 |
60 |
%endif |
%endif |
61 |
|
|
62 |
ALIGN 16 |
ALIGN 16 |
75 |
cglobal transfer_8to16subro_mmx |
cglobal transfer_8to16subro_mmx |
76 |
cglobal transfer_8to16sub2_mmx |
cglobal transfer_8to16sub2_mmx |
77 |
cglobal transfer_8to16sub2_xmm |
cglobal transfer_8to16sub2_xmm |
78 |
|
cglobal transfer_8to16sub2ro_xmm |
79 |
cglobal transfer_16to8add_mmx |
cglobal transfer_16to8add_mmx |
80 |
cglobal transfer8x8_copy_mmx |
cglobal transfer8x8_copy_mmx |
81 |
|
cglobal transfer8x4_copy_mmx |
82 |
|
|
83 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
84 |
; |
; |
117 |
COPY_8_TO_16 2 |
COPY_8_TO_16 2 |
118 |
COPY_8_TO_16 3 |
COPY_8_TO_16 3 |
119 |
ret |
ret |
120 |
|
ENDFUNC |
121 |
|
|
122 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
123 |
; |
; |
153 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
154 |
COPY_16_TO_8 3 |
COPY_16_TO_8 3 |
155 |
ret |
ret |
156 |
|
ENDFUNC |
157 |
|
|
158 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
159 |
; |
; |
217 |
|
|
218 |
pop ebx |
pop ebx |
219 |
ret |
ret |
220 |
|
ENDFUNC |
221 |
|
|
222 |
|
|
223 |
ALIGN 16 |
ALIGN 16 |
236 |
|
|
237 |
pop ebx |
pop ebx |
238 |
ret |
ret |
239 |
|
ENDFUNC |
240 |
|
|
241 |
|
|
242 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
335 |
pop esi |
pop esi |
336 |
pop ebx |
pop ebx |
337 |
ret |
ret |
338 |
|
ENDFUNC |
339 |
|
|
340 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
341 |
; |
; |
403 |
pop esi |
pop esi |
404 |
pop ebx |
pop ebx |
405 |
ret |
ret |
406 |
|
ENDFUNC |
407 |
|
|
408 |
|
|
409 |
|
;----------------------------------------------------------------------------- |
410 |
|
; |
411 |
|
; void transfer_8to16sub2ro_xmm(int16_t * const dct, |
412 |
|
; const uint8_t * const cur, |
413 |
|
; const uint8_t * ref1, |
414 |
|
; const uint8_t * ref2, |
415 |
|
; const uint32_t stride) |
416 |
|
; |
417 |
|
;----------------------------------------------------------------------------- |
418 |
|
|
419 |
|
%macro COPY_8_TO_16_SUB2RO_SSE 1 |
420 |
|
movq mm0, [eax] ; cur |
421 |
|
movq mm2, [eax+edx] |
422 |
|
movq mm1, mm0 |
423 |
|
movq mm3, mm2 |
424 |
|
|
425 |
|
punpcklbw mm0, mm7 |
426 |
|
punpcklbw mm2, mm7 |
427 |
|
movq mm4, [ebx] ; ref1 |
428 |
|
pavgb mm4, [esi] ; ref2 |
429 |
|
punpckhbw mm1, mm7 |
430 |
|
punpckhbw mm3, mm7 |
431 |
|
movq mm5, [ebx+edx] ; ref |
432 |
|
pavgb mm5, [esi+edx] ; ref2 |
433 |
|
|
434 |
|
movq mm6, mm4 |
435 |
|
punpcklbw mm4, mm7 |
436 |
|
punpckhbw mm6, mm7 |
437 |
|
psubsw mm0, mm4 |
438 |
|
psubsw mm1, mm6 |
439 |
|
lea esi, [esi+2*edx] |
440 |
|
movq mm6, mm5 |
441 |
|
punpcklbw mm5, mm7 |
442 |
|
punpckhbw mm6, mm7 |
443 |
|
psubsw mm2, mm5 |
444 |
|
lea eax, [eax+2*edx] |
445 |
|
psubsw mm3, mm6 |
446 |
|
lea ebx, [ebx+2*edx] |
447 |
|
|
448 |
|
movq [ecx+%1*32+ 0], mm0 ; dst |
449 |
|
movq [ecx+%1*32+ 8], mm1 |
450 |
|
movq [ecx+%1*32+16], mm2 |
451 |
|
movq [ecx+%1*32+24], mm3 |
452 |
|
%endmacro |
453 |
|
|
454 |
|
ALIGN 16 |
455 |
|
transfer_8to16sub2ro_xmm: |
456 |
|
pxor mm7, mm7 |
457 |
|
mov ecx, [esp + 4] ; Dst |
458 |
|
mov eax, [esp + 8] ; Cur |
459 |
|
push ebx |
460 |
|
mov ebx, [esp+4+12] ; Ref1 |
461 |
|
push esi |
462 |
|
mov esi, [esp+8+16] ; Ref2 |
463 |
|
mov edx, [esp+8+20] ; Stride |
464 |
|
|
465 |
|
COPY_8_TO_16_SUB2RO_SSE 0 |
466 |
|
COPY_8_TO_16_SUB2RO_SSE 1 |
467 |
|
COPY_8_TO_16_SUB2RO_SSE 2 |
468 |
|
COPY_8_TO_16_SUB2RO_SSE 3 |
469 |
|
|
470 |
|
pop esi |
471 |
|
pop ebx |
472 |
|
ret |
473 |
|
ENDFUNC |
474 |
|
|
475 |
|
|
476 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
477 |
; |
; |
516 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
517 |
COPY_16_TO_8_ADD 3 |
COPY_16_TO_8_ADD 3 |
518 |
ret |
ret |
519 |
|
ENDFUNC |
520 |
|
|
521 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
522 |
; |
; |
549 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
550 |
COPY_8_TO_8 |
COPY_8_TO_8 |
551 |
ret |
ret |
552 |
|
ENDFUNC |
553 |
|
|
554 |
|
;----------------------------------------------------------------------------- |
555 |
|
; |
556 |
|
; void transfer8x4_copy_mmx(uint8_t * const dst, |
557 |
|
; const uint8_t * const src, |
558 |
|
; const uint32_t stride); |
559 |
|
; |
560 |
|
; |
561 |
|
;----------------------------------------------------------------------------- |
562 |
|
|
563 |
|
ALIGN 16 |
564 |
|
transfer8x4_copy_mmx: |
565 |
|
mov ecx, [esp+ 4] ; Dst |
566 |
|
mov eax, [esp+ 8] ; Src |
567 |
|
mov edx, [esp+12] ; Stride |
568 |
|
|
569 |
|
COPY_8_TO_8 |
570 |
|
lea ecx,[ecx+2*edx] |
571 |
|
COPY_8_TO_8 |
572 |
|
ret |
573 |
|
ENDFUNC |
574 |
|
|
575 |
|
|
576 |
|
%ifidn __OUTPUT_FORMAT__,elf |
577 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
578 |
|
%endif |
579 |
|
|