📄 nb_kernel302_ia32_3dnow.s
字号:
pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movd %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqOO(%esp),%mm5 ## vcoul=qq*VV ## update vctot directly, use mm3 for fscal sum. pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) ## time for hydrogens! movq nb302nf_tmprsqH(%esp),%mm0 pfrsqrt %mm0,%mm1 pswapd %mm0,%mm0 pfrsqrt %mm0,%mm2 pswapd %mm0,%mm0 punpckldq %mm2,%mm1 ## seeds are in mm1 now, and rsq in mm0. movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=r pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 movl nb302nf_n1+4(%esp),%ecx shll $2,%ecx punpckldq (%edx,%ecx,4),%mm4 punpckldq 4(%edx,%ecx,4),%mm5 punpckldq 8(%edx,%ecx,4),%mm6 punpckldq 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqOH(%esp),%mm5 ## vcoul=qq*VV ## update vctot pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) ## interactions with j H1 movq 12(%esi,%eax,4),%mm0 movd 20(%esi,%eax,4),%mm1 ## copy & expand to mm2-mm4 for the H interactions movq %mm0,%mm2 movq %mm0,%mm3 movq %mm1,%mm4 punpckldq %mm2,%mm2 punpckhdq %mm3,%mm3 punpckldq %mm4,%mm4 pfsubr nb302nf_ixO(%esp),%mm0 pfsubr nb302nf_izO(%esp),%mm1 pfmul %mm0,%mm0 pfmul %mm1,%mm1 pfacc %mm1,%mm0 pfadd %mm1,%mm0 ## mm0=rsqO punpckldq %mm2,%mm2 punpckldq %mm3,%mm3 punpckldq %mm4,%mm4 ## mm2-mm4 is jx-jz pfsubr nb302nf_ixH(%esp),%mm2 pfsubr nb302nf_iyH(%esp),%mm3 pfsubr nb302nf_izH(%esp),%mm4 ## mm2-mm4 is dxH-dzH pfmul %mm2,%mm2 pfmul %mm3,%mm3 pfmul %mm4,%mm4 pfadd %mm2,%mm3 pfadd %mm4,%mm3 ## mm3=rsqH movq %mm3,nb302nf_tmprsqH(%esp) pfrsqrt %mm0,%mm1 movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=rsq pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movd %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqOH(%esp),%mm5 ## vcoul=qq*VV ## update vctot directly, force is moved to mm3 pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) movq nb302nf_tmprsqH(%esp),%mm0 pfrsqrt %mm0,%mm1 pswapd %mm0,%mm0 pfrsqrt %mm0,%mm2 pswapd %mm0,%mm0 punpckldq %mm2,%mm1 ## seeds are in mm1 now, and rsq in mm0. movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=r pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 movl nb302nf_n1+4(%esp),%ecx shll $2,%ecx punpckldq (%edx,%ecx,4),%mm4 punpckldq 4(%edx,%ecx,4),%mm5 punpckldq 8(%edx,%ecx,4),%mm6 punpckldq 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqHH(%esp),%mm5 ## vcoul=qq*VV ## update vctot pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) ## interactions with j H2 movq 24(%esi,%eax,4),%mm0 movd 32(%esi,%eax,4),%mm1 ## copy & expand to mm2-mm4 for the H interactions movq %mm0,%mm2 movq %mm0,%mm3 movq %mm1,%mm4 punpckldq %mm2,%mm2 punpckhdq %mm3,%mm3 punpckldq %mm4,%mm4 pfsubr nb302nf_ixO(%esp),%mm0 pfsubr nb302nf_izO(%esp),%mm1 pfmul %mm0,%mm0 pfmul %mm1,%mm1 pfacc %mm1,%mm0 pfadd %mm1,%mm0 ## mm0=rsqO punpckldq %mm2,%mm2 punpckldq %mm3,%mm3 punpckldq %mm4,%mm4 ## mm2-mm4 is jx-jz pfsubr nb302nf_ixH(%esp),%mm2 pfsubr nb302nf_iyH(%esp),%mm3 pfsubr nb302nf_izH(%esp),%mm4 ## mm2-mm4 is dxH-dzH pfmul %mm2,%mm2 pfmul %mm3,%mm3 pfmul %mm4,%mm4 pfadd %mm2,%mm3 pfadd %mm4,%mm3 ## mm3=rsqH movq %mm3,nb302nf_tmprsqH(%esp) pfrsqrt %mm0,%mm1 movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movd %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqOH(%esp),%mm5 ## vcoul=qq*VV ## update vctot directly, use mm3 for fscal sum. pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) movq nb302nf_tmprsqH(%esp),%mm0 pfrsqrt %mm0,%mm1 pswapd %mm0,%mm0 pfrsqrt %mm0,%mm2 pswapd %mm0,%mm0 punpckldq %mm2,%mm1 ## seeds are in mm1 now, and rsq in mm0. movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=r pfmul nb302nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb302nf_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb302nf_VFtab(%ebp),%edx movl nb302nf_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 movl nb302nf_n1+4(%esp),%ecx shll $2,%ecx punpckldq (%edx,%ecx,4),%mm4 punpckldq 4(%edx,%ecx,4),%mm5 punpckldq 8(%edx,%ecx,4),%mm6 punpckldq 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb302nf_qqHH(%esp),%mm5 ## vcoul=qq*VV ## update vctot pfadd nb302nf_vctot(%esp),%mm5 movq %mm5,nb302nf_vctot(%esp) ## done - one more? decl nb302nf_innerk(%esp) jz _nb_kernel302nf_ia32_3dnow.nb302nf_updateouterdata jmp _nb_kernel302nf_ia32_3dnow.nb302nf_inner_loop_nb_kernel302nf_ia32_3dnow.nb302nf_updateouterdata: ## get n from stack movl nb302nf_n(%esp),%esi ## get group index for i particle movl nb302nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] movq nb302nf_vctot(%esp),%mm7 pfacc %mm7,%mm7 ## get and sum the two parts of total potential movl nb302nf_Vc(%ebp),%eax movd (%eax,%edx,4),%mm6 pfadd %mm7,%mm6 movd %mm6,(%eax,%edx,4) ## increment vc[gid] ## finish if last movl nb302nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel302nf_ia32_3dnow.nb302nf_outerend ## not last, iterate outer loop once more! movl %esi,nb302nf_n(%esp) jmp _nb_kernel302nf_ia32_3dnow.nb302nf_outer_nb_kernel302nf_ia32_3dnow.nb302nf_outerend: ## check if more outer neighborlists remain movl nb302nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel302nf_ia32_3dnow.nb302nf_end ## non-zero, do one more workunit jmp _nb_kernel302nf_ia32_3dnow.nb302nf_threadloop_nb_kernel302nf_ia32_3dnow.nb302nf_end: femms movl nb302nf_nouter(%esp),%eax movl nb302nf_ninner(%esp),%ebx movl nb302nf_outeriter(%ebp),%ecx movl nb302nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) addl $128,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -