nb_kernel110_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,505 行 · 第 1/4 页
S
1,505 行
movhps (%esi,%ebx,4),%xmm6 movhps (%esi,%edx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb110nf_c6(%esp) movaps %xmm6,nb110nf_c12(%esp) movl nb110nf_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx mulps %xmm2,%xmm3 leal (%ecx,%ecx,2),%ecx ## replace jnr with j3 leal (%edx,%edx,2),%edx ## move four coordinates to xmm0-xmm2 movlps (%esi,%eax,4),%xmm4 movlps (%esi,%ecx,4),%xmm5 movss 8(%esi,%eax,4),%xmm2 movss 8(%esi,%ecx,4),%xmm6 movhps (%esi,%ebx,4),%xmm4 movhps (%esi,%edx,4),%xmm5 movss 8(%esi,%ebx,4),%xmm0 movss 8(%esi,%edx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ix-iz to xmm4-xmm6 movaps nb110nf_ix(%esp),%xmm4 movaps nb110nf_iy(%esp),%xmm5 movaps nb110nf_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%esp),%xmm1 mulps nb110nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb110nf_Vvdwtot(%esp),%xmm5 addps nb110nf_vctot(%esp),%xmm3 movaps %xmm3,nb110nf_vctot(%esp) movaps %xmm5,nb110nf_Vvdwtot(%esp) ## should we do one more iteration? subl $4,nb110nf_innerk(%esp) jl _nb_kernel110nf_ia32_sse.nb110nf_finish_inner jmp _nb_kernel110nf_ia32_sse.nb110nf_unroll_loop_nb_kernel110nf_ia32_sse.nb110nf_finish_inner: ## check if at least two particles remain addl $4,nb110nf_innerk(%esp) movl nb110nf_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel110nf_ia32_sse.nb110nf_dopair jmp _nb_kernel110nf_ia32_sse.nb110nf_checksingle_nb_kernel110nf_ia32_sse.nb110nf_dopair: movl nb110nf_charge(%ebp),%esi movl nb110nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb110nf_innerjjnr(%esp) xorps %xmm3,%xmm3 movss (%esi,%eax,4),%xmm3 movss (%esi,%ebx,4),%xmm6 shufps $12,%xmm6,%xmm3 ## constant 00001100 shufps $88,%xmm3,%xmm3 ## constant 01011000 ;# xmm3(0,1) has the charges movl nb110nf_type(%ebp),%esi movl %eax,%ecx movl %ebx,%edx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb110nf_vdwparam(%ebp),%esi shll %ecx shll %edx movl nb110nf_ntia(%esp),%edi addl %edi,%ecx addl %edi,%edx movlps (%esi,%ecx,4),%xmm6 movhps (%esi,%edx,4),%xmm6 movl nb110nf_pos(%ebp),%edi xorps %xmm7,%xmm7 movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## constant 00001000 shufps $13,%xmm6,%xmm6 ## constant 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb110nf_c6(%esp) movaps %xmm6,nb110nf_c12(%esp) leal (%eax,%eax,2),%eax leal (%ebx,%ebx,2),%ebx ## move coordinates to xmm0-xmm2 movlps (%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 movhps (%edi,%ebx,4),%xmm1 movss 8(%edi,%ebx,4),%xmm0 mulps nb110nf_iq(%esp),%xmm3 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## constant 10001000 shufps $136,%xmm0,%xmm0 ## constant 10001000 shufps $221,%xmm1,%xmm1 ## constant 11011101 ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb110nf_ix(%esp),%xmm4 movaps nb110nf_iy(%esp),%xmm5 movaps nb110nf_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%esp),%xmm1 mulps nb110nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb110nf_Vvdwtot(%esp),%xmm5 addps nb110nf_vctot(%esp),%xmm3 movaps %xmm3,nb110nf_vctot(%esp) movaps %xmm5,nb110nf_Vvdwtot(%esp)_nb_kernel110nf_ia32_sse.nb110nf_checksingle: movl nb110nf_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel110nf_ia32_sse.nb110nf_dosingle jmp _nb_kernel110nf_ia32_sse.nb110nf_updateouterdata_nb_kernel110nf_ia32_sse.nb110nf_dosingle: movl nb110nf_charge(%ebp),%esi movl nb110nf_pos(%ebp),%edi movl nb110nf_innerjjnr(%esp),%ecx xorps %xmm3,%xmm3 movl (%ecx),%eax movss (%esi,%eax,4),%xmm3 ## xmm3(0) has the charge movl nb110nf_type(%ebp),%esi movl %eax,%ecx movl (%esi,%ecx,4),%ecx movl nb110nf_vdwparam(%ebp),%esi shll %ecx addl nb110nf_ntia(%esp),%ecx xorps %xmm6,%xmm6 movlps (%esi,%ecx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## constant 11111100 shufps $253,%xmm6,%xmm6 ## constant 11111101 movaps %xmm4,nb110nf_c6(%esp) movaps %xmm6,nb110nf_c12(%esp) leal (%eax,%eax,2),%eax ## move coordinates to xmm0-xmm2 movss (%edi,%eax,4),%xmm0 movss 4(%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 mulps nb110nf_iq(%esp),%xmm3 xorps %xmm7,%xmm7 movaps nb110nf_ix(%esp),%xmm4 movaps nb110nf_iy(%esp),%xmm5 movaps nb110nf_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%esp),%xmm1 mulps nb110nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addss nb110nf_Vvdwtot(%esp),%xmm5 addss nb110nf_vctot(%esp),%xmm3 movss %xmm3,nb110nf_vctot(%esp) movss %xmm5,nb110nf_Vvdwtot(%esp)_nb_kernel110nf_ia32_sse.nb110nf_updateouterdata: ## get n from stack movl nb110nf_n(%esp),%esi ## get group index for i particle movl nb110nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb110nf_vctot(%esp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addps %xmm6,%xmm7 ## pos 0-1 in xmm7 have the sum now movaps %xmm7,%xmm6 shufps $1,%xmm6,%xmm6 addss %xmm6,%xmm7 ## add earlier value from mem movl nb110nf_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## accumulate total lj energy and update it movaps nb110nf_Vvdwtot(%esp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addps %xmm6,%xmm7 ## pos 0-1 in xmm7 have the sum now movaps %xmm7,%xmm6 shufps $1,%xmm6,%xmm6 addss %xmm6,%xmm7 ## add earlier value from mem movl nb110nf_Vvdw(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb110nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel110nf_ia32_sse.nb110nf_outerend ## not last, iterate outer loop once more! movl %esi,nb110nf_n(%esp) jmp _nb_kernel110nf_ia32_sse.nb110nf_outer_nb_kernel110nf_ia32_sse.nb110nf_outerend: ## check if more outer neighborlists remain movl nb110nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel110nf_ia32_sse.nb110nf_end ## non-zero, do one more workunit jmp _nb_kernel110nf_ia32_sse.nb110nf_threadloop_nb_kernel110nf_ia32_sse.nb110nf_end: emms movl nb110nf_nouter(%esp),%eax movl nb110nf_ninner(%esp),%ebx movl nb110nf_outeriter(%ebp),%ecx movl nb110nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb110nf_salign(%esp),%eax addl %eax,%esp addl $216,%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
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?