28 |
|
|
29 |
%macro cglobal 1 |
%macro cglobal 1 |
30 |
%ifdef PREFIX |
%ifdef PREFIX |
31 |
|
%ifdef MARK_FUNCS |
32 |
|
global _%1:function %1.endfunc-%1 |
33 |
|
%define %1 _%1:function %1.endfunc-%1 |
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 %1.endfunc-%1 |
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 |
74 |
lea eax, [eax+ecx] |
lea eax, [eax+ecx] |
75 |
movq mm5, mm2 |
movq mm5, mm2 |
76 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
|
lea edx, [edx+ecx] |
|
77 |
|
|
78 |
psubusb mm1, mm4 |
psubusb mm1, mm4 |
|
por mm0, mm1 |
|
79 |
psubusb mm3, mm5 |
psubusb mm3, mm5 |
80 |
|
por mm0, mm1 |
81 |
por mm2, mm3 |
por mm2, mm3 |
82 |
|
|
83 |
movq mm1, mm0 |
movq mm1, mm0 |
|
movq mm3, mm2 |
|
|
|
|
84 |
punpcklbw mm0,mm7 |
punpcklbw mm0,mm7 |
85 |
|
movq mm3, mm2 |
86 |
punpckhbw mm1,mm7 |
punpckhbw mm1,mm7 |
87 |
|
lea edx, [edx+ecx] |
88 |
punpcklbw mm2,mm7 |
punpcklbw mm2,mm7 |
|
punpckhbw mm3,mm7 |
|
|
|
|
89 |
paddusw mm0, mm1 |
paddusw mm0, mm1 |
90 |
|
punpckhbw mm3,mm7 |
91 |
paddusw mm6, mm0 |
paddusw mm6, mm0 |
92 |
paddusw mm2, mm3 |
paddusw mm2, mm3 |
93 |
paddusw mm6, mm2 |
paddusw mm6, mm2 |
94 |
|
|
95 |
%endmacro |
%endmacro |
96 |
|
|
97 |
%macro SAD_8x8_MMX 0 |
%macro SAD_8x8_MMX 0 |
110 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
111 |
|
|
112 |
psubusb mm1, mm4 |
psubusb mm1, mm4 |
|
por mm0, mm1 |
|
113 |
psubusb mm3, mm5 |
psubusb mm3, mm5 |
114 |
|
por mm0, mm1 |
115 |
por mm2, mm3 |
por mm2, mm3 |
116 |
|
|
117 |
movq mm1,mm0 |
movq mm1,mm0 |
|
movq mm3,mm2 |
|
|
|
|
118 |
punpcklbw mm0,mm7 |
punpcklbw mm0,mm7 |
119 |
|
movq mm3,mm2 |
120 |
punpckhbw mm1,mm7 |
punpckhbw mm1,mm7 |
121 |
punpcklbw mm2,mm7 |
punpcklbw mm2,mm7 |
|
punpckhbw mm3,mm7 |
|
|
|
|
122 |
paddusw mm0,mm1 |
paddusw mm0,mm1 |
123 |
|
punpckhbw mm3,mm7 |
124 |
paddusw mm6,mm0 |
paddusw mm6,mm0 |
125 |
paddusw mm2,mm3 |
paddusw mm2,mm3 |
126 |
paddusw mm6,mm2 |
paddusw mm6,mm2 |
127 |
%endmacro |
%endmacro |
128 |
|
|
129 |
|
|
130 |
%macro SADV_16x16_MMX 0 |
%macro SADV_16x16_MMX 0 |
131 |
movq mm0, [eax] |
movq mm0, [eax] |
132 |
movq mm1, [edx] |
movq mm1, [edx] |
133 |
|
|
134 |
movq mm2, [eax+8] |
movq mm2, [eax+8] |
|
movq mm3, [edx+8] |
|
|
|
|
135 |
movq mm4, mm0 |
movq mm4, mm0 |
136 |
|
movq mm3, [edx+8] |
137 |
psubusb mm0, mm1 |
psubusb mm0, mm1 |
138 |
|
|
139 |
psubusb mm1, mm4 |
psubusb mm1, mm4 |
|
por mm0, mm1 |
|
140 |
lea eax,[eax+ecx] |
lea eax,[eax+ecx] |
141 |
|
por mm0, mm1 |
142 |
|
|
143 |
movq mm4, mm2 |
movq mm4, mm2 |
144 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
145 |
|
|
146 |
psubusb mm3, mm4 |
psubusb mm3, mm4 |
147 |
por mm2, mm3 |
por mm2, mm3 |
|
lea edx,[edx+ecx] |
|
148 |
|
|
149 |
movq mm1,mm0 |
movq mm1,mm0 |
|
movq mm3,mm2 |
|
|
|
|
150 |
punpcklbw mm0,mm7 |
punpcklbw mm0,mm7 |
151 |
|
movq mm3,mm2 |
152 |
punpckhbw mm1,mm7 |
punpckhbw mm1,mm7 |
153 |
punpcklbw mm2,mm7 |
punpcklbw mm2,mm7 |
|
punpckhbw mm3,mm7 |
|
|
|
|
154 |
paddusw mm0,mm1 |
paddusw mm0,mm1 |
155 |
paddusw mm2,mm3 |
punpckhbw mm3,mm7 |
|
|
|
156 |
paddusw mm5, mm0 |
paddusw mm5, mm0 |
157 |
|
paddusw mm2,mm3 |
158 |
|
lea edx,[edx+ecx] |
159 |
paddusw mm6, mm2 |
paddusw mm6, mm2 |
160 |
%endmacro |
%endmacro |
161 |
|
|
214 |
movq mm2, [eax+8] |
movq mm2, [eax+8] |
215 |
lea eax, [eax+ecx] |
lea eax, [eax+ecx] |
216 |
movq mm1, mm0 |
movq mm1, mm0 |
|
movq mm3, mm2 |
|
217 |
punpcklbw mm0, mm7 |
punpcklbw mm0, mm7 |
218 |
punpcklbw mm2, mm7 |
movq mm3, mm2 |
219 |
punpckhbw mm1, mm7 |
punpckhbw mm1, mm7 |
|
punpckhbw mm3, mm7 |
|
220 |
paddw mm5, mm0 |
paddw mm5, mm0 |
221 |
|
punpcklbw mm2, mm7 |
222 |
paddw mm6, mm1 |
paddw mm6, mm1 |
223 |
|
punpckhbw mm3, mm7 |
224 |
paddw mm5, mm2 |
paddw mm5, mm2 |
225 |
paddw mm6, mm3 |
paddw mm6, mm3 |
226 |
%endmacro |
%endmacro |
273 |
cglobal sad8bi_mmx |
cglobal sad8bi_mmx |
274 |
cglobal dev16_mmx |
cglobal dev16_mmx |
275 |
cglobal sse8_16bit_mmx |
cglobal sse8_16bit_mmx |
276 |
|
cglobal sse8_8bit_mmx |
277 |
|
|
278 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
279 |
; |
; |
322 |
movd eax, mm6 |
movd eax, mm6 |
323 |
|
|
324 |
ret |
ret |
325 |
|
.endfunc |
326 |
|
|
327 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
328 |
; |
; |
355 |
movd eax, mm6 |
movd eax, mm6 |
356 |
|
|
357 |
ret |
ret |
358 |
|
.endfunc |
359 |
|
|
360 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
361 |
; |
; |
446 |
pop ebx |
pop ebx |
447 |
|
|
448 |
ret |
ret |
449 |
|
.endfunc |
450 |
|
|
451 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
452 |
; |
; |
511 |
pop ebx |
pop ebx |
512 |
|
|
513 |
ret |
ret |
514 |
|
.endfunc |
515 |
|
|
516 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
517 |
; |
; |
550 |
movd eax, mm6 |
movd eax, mm6 |
551 |
pop ebx |
pop ebx |
552 |
ret |
ret |
553 |
|
.endfunc |
554 |
|
|
555 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
556 |
; |
; |
630 |
movd eax, mm6 |
movd eax, mm6 |
631 |
|
|
632 |
ret |
ret |
633 |
|
.endfunc |
634 |
|
|
635 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
636 |
; |
; |
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 |
|
.endfunc |
684 |
|
|
685 |
|
;----------------------------------------------------------------------------- |
686 |
|
; |
687 |
|
; uint32_t sse8_8bit_mmx(const int8_t *b1, |
688 |
|
; const int8_t *b2, |
689 |
|
; const uint32_t stride); |
690 |
|
; |
691 |
|
;----------------------------------------------------------------------------- |
692 |
|
|
693 |
|
%macro ROW_SSE_8bit_MMX 2 |
694 |
|
movq mm0, [%1] ; load a row |
695 |
|
movq mm2, [%2] ; load a row |
696 |
|
|
697 |
|
movq mm1, mm0 ; copy row |
698 |
|
movq mm3, mm2 ; copy row |
699 |
|
|
700 |
|
punpcklbw mm0, mm7 ; turn the 4low elements into 16bit |
701 |
|
punpckhbw mm1, mm7 ; turn the 4high elements into 16bit |
702 |
|
|
703 |
|
punpcklbw mm2, mm7 ; turn the 4low elements into 16bit |
704 |
|
punpckhbw mm3, mm7 ; turn the 4high elements into 16bit |
705 |
|
|
706 |
|
psubw mm0, mm2 ; low part of src-dst |
707 |
|
psubw mm1, mm3 ; high part of src-dst |
708 |
|
|
709 |
|
pmaddwd mm0, mm0 ; compute the square sum |
710 |
|
pmaddwd mm1, mm1 ; compute the square sum |
711 |
|
|
712 |
|
paddd mm6, mm0 ; add to the accumulator |
713 |
|
paddd mm6, mm1 ; add to the accumulator |
714 |
|
%endmacro |
715 |
|
|
716 |
|
sse8_8bit_mmx: |
717 |
|
push esi |
718 |
|
push edi |
719 |
|
|
720 |
|
;; Load the function params |
721 |
|
mov esi, [esp+8+4] |
722 |
|
mov edi, [esp+8+8] |
723 |
|
mov edx, [esp+8+12] |
724 |
|
|
725 |
|
;; Reset the sse accumulator |
726 |
|
pxor mm6, mm6 |
727 |
|
|
728 |
|
;; Used to interleave 8bit data with 0x00 values |
729 |
|
pxor mm7, mm7 |
730 |
|
|
731 |
|
;; Let's go |
732 |
|
%rep 8 |
733 |
|
ROW_SSE_8bit_MMX esi, edi |
734 |
|
lea esi, [esi+edx] |
735 |
|
lea edi, [edi+edx] |
736 |
|
%endrep |
737 |
|
|
738 |
|
;; Finish adding each dword of the accumulator |
739 |
|
movq mm7, mm6 |
740 |
|
psrlq mm6, 32 |
741 |
|
paddd mm6, mm7 |
742 |
|
movd eax, mm6 |
743 |
|
|
744 |
|
;; All done |
745 |
|
pop edi |
746 |
|
pop esi |
747 |
|
ret |
748 |
|
.endfunc |
749 |
|
|