126 |
%endmacro |
%endmacro |
127 |
|
|
128 |
|
|
129 |
SECTION .rotext align=SECTION_ALIGN |
TEXT |
130 |
|
|
131 |
cglobal lum_8x8_mmx |
cglobal lum_8x8_mmx |
132 |
cglobal consim_sse2 |
cglobal consim_sse2 |
159 |
|
|
160 |
ALIGN SECTION_ALIGN |
ALIGN SECTION_ALIGN |
161 |
consim_sse2: |
consim_sse2: |
162 |
|
PUSH_XMM6_XMM7 |
163 |
mov TMP0,prm1 ;ptro |
mov TMP0,prm1 ;ptro |
164 |
mov TMP1,prm2 ;ptrc |
mov TMP1,prm2 ;ptrc |
165 |
mov _EAX, prm3 ;stride |
mov _EAX, prm3 ;stride |
169 |
pxor xmm6,xmm6;devc |
pxor xmm6,xmm6;devc |
170 |
pxor xmm7,xmm7;corr |
pxor xmm7,xmm7;corr |
171 |
|
|
|
;broadcast lumo/c |
|
|
punpcklbw xmm6,xmm6 |
|
|
punpcklwd xmm6,xmm6 |
|
|
pshufd xmm6,xmm6,00000000b;or shufps |
|
|
punpcklbw xmm7,xmm7 |
|
|
punpcklwd xmm7,xmm7 |
|
|
pshufd xmm7,xmm7,00000000b |
|
|
|
|
172 |
CONSIM_1x8_SSE2 |
CONSIM_1x8_SSE2 |
173 |
add TMP0,_EAX |
add TMP0,_EAX |
174 |
add TMP1,_EAX |
add TMP1,_EAX |
209 |
paddd xmm7, xmm2 |
paddd xmm7, xmm2 |
210 |
|
|
211 |
CONSIM_WRITEOUT xmm5,xmm6,xmm7 |
CONSIM_WRITEOUT xmm5,xmm6,xmm7 |
212 |
|
|
213 |
|
POP_XMM6_XMM7 |
214 |
ret |
ret |
215 |
ENDFUNC |
ENDFUNC |
216 |
|
|
265 |
ret |
ret |
266 |
ENDFUNC |
ENDFUNC |
267 |
|
|
268 |
%ifidn __OUTPUT_FORMAT__,elf |
NON_EXEC_STACK |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
|
|
%endif |
|