<div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><DIV>>+cglobal pixel_satd_32x16, 4,8,8 ;if WIN64 && notcpuflag(avx)<BR>>+ SATD_START_SSE2 m10, m7<BR>>+ mov r6, r0<BR>>+ mov r7, r2<BR>>+%if vertical<BR>>+ mova m7, [pw_00ff]<BR>>+%endif<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ lea r0, [r6 + 16]<BR>>+ lea r2, [r7 + 16]<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ call pixel_satd_16x4_internal<BR>>+ SATD_END_SSE2 m10<BR>>+<BR>> %else<BR>>+cglobal pixel_satd_32x16, 4,8,8<BR>you can't use 8-registers in x86</DIV>
<DIV> </DIV>
<DIV>>+%if WIN64 ;if WIN64 && cpuflag(avx)<BR>>+ SATD_START_SSE2 m6, m7<BR>>+ mov r6, r0<BR>>+ mov r7, r2<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ lea r0, [r6 + 8]<BR>>+ lea r2, [r7 + 8]<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ lea r0, [r6 + 16]<BR>>+ lea r2, [r7 + 16]<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ lea r0, [r6 + 24]<BR>>+ lea r2, [r7 + 24]<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ SATD_END_SSE2 m6<BR>>+%else ;if !WIN64<BR>>+ SATD_START_SSE2 m6, m7<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ mov r0, r0mp<BR>>+ mov r2, r2mp<BR>>+ add r0, 8<BR>>+ add r2, 8<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ mov r0, r0mp<BR>>+ mov r2, r2mp<BR>>+ add r0, 16<BR>>+ add r2, 16<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ mov r0, r0mp<BR>>+ mov r2, r2mp<BR>>+ add r0, 24<BR>>+ add r2, 24<BR>>+ call pixel_satd_8x8_internal<BR>>+ call pixel_satd_8x8_internal<BR>>+ SATD_END_SSE2 m6<BR>>+%endif<BR></DIV></div>