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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; * $Id: sad_mmx.asm,v 1.12 2004-03-22 22:36:24 edgomez Exp $ |
; * $Id: sad_mmx.asm,v 1.15 2004-08-22 11:46:10 edgomez Exp $ |
24 |
; * |
; * |
25 |
; ***************************************************************************/ |
; ***************************************************************************/ |
26 |
|
|
28 |
|
|
29 |
%macro cglobal 1 |
%macro cglobal 1 |
30 |
%ifdef PREFIX |
%ifdef PREFIX |
31 |
|
%ifdef MARK_FUNCS |
32 |
|
global _%1:function |
33 |
|
%define %1 _%1:function |
34 |
|
%else |
35 |
global _%1 |
global _%1 |
36 |
%define %1 _%1 |
%define %1 _%1 |
37 |
|
%endif |
38 |
|
%else |
39 |
|
%ifdef MARK_FUNCS |
40 |
|
global %1:function |
41 |
%else |
%else |
42 |
global %1 |
global %1 |
43 |
%endif |
%endif |
44 |
|
%endif |
45 |
%endmacro |
%endmacro |
46 |
|
|
47 |
;============================================================================= |
;============================================================================= |
49 |
;============================================================================= |
;============================================================================= |
50 |
|
|
51 |
%ifdef FORMAT_COFF |
%ifdef FORMAT_COFF |
52 |
SECTION .rodata data |
SECTION .rodata |
53 |
%else |
%else |
54 |
SECTION .rodata data align=16 |
SECTION .rodata align=16 |
55 |
%endif |
%endif |
56 |
|
|
57 |
ALIGN 16 |
ALIGN 16 |
279 |
cglobal sad8bi_mmx |
cglobal sad8bi_mmx |
280 |
cglobal dev16_mmx |
cglobal dev16_mmx |
281 |
cglobal sse8_16bit_mmx |
cglobal sse8_16bit_mmx |
282 |
|
cglobal sse8_8bit_mmx |
283 |
|
|
284 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
285 |
; |
; |
640 |
; |
; |
641 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
642 |
|
|
643 |
%macro ROW_SSE_MMX 2 |
%macro ROW_SSE_16bit_MMX 2 |
644 |
movq mm0, [%1] |
movq mm0, [%1] |
645 |
movq mm1, [%1+8] |
movq mm1, [%1+8] |
646 |
psubw mm0, [%2] |
psubw mm0, [%2] |
664 |
pxor mm2, mm2 |
pxor mm2, mm2 |
665 |
|
|
666 |
;; Let's go |
;; Let's go |
667 |
ROW_SSE_MMX esi, edi |
%rep 8 |
668 |
lea esi, [esi+edx] |
ROW_SSE_16bit_MMX esi, edi |
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
|
lea esi, [esi+edx] |
|
|
lea edi, [edi+edx] |
|
|
ROW_SSE_MMX esi, edi |
|
669 |
lea esi, [esi+edx] |
lea esi, [esi+edx] |
670 |
lea edi, [edi+edx] |
lea edi, [edi+edx] |
671 |
|
%endrep |
672 |
|
|
673 |
;; Finish adding each dword of the accumulator |
;; Finish adding each dword of the accumulator |
674 |
movq mm3, mm2 |
movq mm3, mm2 |
680 |
pop edi |
pop edi |
681 |
pop esi |
pop esi |
682 |
ret |
ret |
683 |
|
|
684 |
|
;----------------------------------------------------------------------------- |
685 |
|
; |
686 |
|
; uint32_t sse8_8bit_mmx(const int8_t *b1, |
687 |
|
; const int8_t *b2, |
688 |
|
; const uint32_t stride); |
689 |
|
; |
690 |
|
;----------------------------------------------------------------------------- |
691 |
|
|
692 |
|
%macro ROW_SSE_8bit_MMX 2 |
693 |
|
movq mm0, [%1] ; load a row |
694 |
|
movq mm2, [%2] ; load a row |
695 |
|
|
696 |
|
movq mm1, mm0 ; copy row |
697 |
|
movq mm3, mm2 ; copy row |
698 |
|
|
699 |
|
punpcklbw mm0, mm7 ; turn the 4low elements into 16bit |
700 |
|
punpckhbw mm1, mm7 ; turn the 4high elements into 16bit |
701 |
|
|
702 |
|
punpcklbw mm2, mm7 ; turn the 4low elements into 16bit |
703 |
|
punpckhbw mm3, mm7 ; turn the 4high elements into 16bit |
704 |
|
|
705 |
|
psubw mm0, mm2 ; low part of src-dst |
706 |
|
psubw mm1, mm3 ; high part of src-dst |
707 |
|
|
708 |
|
pmaddwd mm0, mm0 ; compute the square sum |
709 |
|
pmaddwd mm1, mm1 ; compute the square sum |
710 |
|
|
711 |
|
paddd mm6, mm0 ; add to the accumulator |
712 |
|
paddd mm6, mm1 ; add to the accumulator |
713 |
|
%endmacro |
714 |
|
|
715 |
|
sse8_8bit_mmx: |
716 |
|
push esi |
717 |
|
push edi |
718 |
|
|
719 |
|
;; Load the function params |
720 |
|
mov esi, [esp+8+4] |
721 |
|
mov edi, [esp+8+8] |
722 |
|
mov edx, [esp+8+12] |
723 |
|
|
724 |
|
;; Reset the sse accumulator |
725 |
|
pxor mm6, mm6 |
726 |
|
|
727 |
|
;; Used to interleave 8bit data with 0x00 values |
728 |
|
pxor mm7, mm7 |
729 |
|
|
730 |
|
;; Let's go |
731 |
|
%rep 8 |
732 |
|
ROW_SSE_8bit_MMX esi, edi |
733 |
|
lea esi, [esi+edx] |
734 |
|
lea edi, [edi+edx] |
735 |
|
%endrep |
736 |
|
|
737 |
|
;; Finish adding each dword of the accumulator |
738 |
|
movq mm7, mm6 |
739 |
|
psrlq mm6, 32 |
740 |
|
paddd mm6, mm7 |
741 |
|
movd eax, mm6 |
742 |
|
|
743 |
|
;; All done |
744 |
|
pop edi |
745 |
|
pop esi |
746 |
|
ret |