nb_kernel131_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,004 行 · 第 1/5 页
S
2,004 行
mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb131nf_ixH2(%esp),%xmm3 movaps nb131nf_iyH2(%esp),%xmm4 movaps nb131nf_izH2(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## start with rsqO (still in xmm7) - seed to xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb131nf_three(%esp),%xmm4 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb131nf_half(%esp),%xmm4 movaps %xmm4,nb131nf_rinvO(%esp) ## rinvO in xmm0, rsqO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb131nf_three(%esp),%xmm4 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb131nf_half(%esp),%xmm4 movaps %xmm4,nb131nf_rinvH1(%esp) ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb131nf_three(%esp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb131nf_half(%esp),%xmm4 movaps %xmm4,nb131nf_rinvH2(%esp) ## rinvH2 in xmm5 ## do O table interactions - rsqO in xmm7. mulps nb131nf_rinvO(%esp),%xmm7 mulps nb131nf_tsc(%esp),%xmm7 ## rtab movhlps %xmm7,%xmm5 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm5,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm6 cvtpi2ps %mm7,%xmm5 movlhps %xmm5,%xmm6 subps %xmm6,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 pslld $3,%mm7 movl nb131nf_VFtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx ## dispersion movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm5 movhps (%esi,%edx,4),%xmm7 ## got half dispersion table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%ebx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm3 ## other half of dispersion table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb131nf_c6(%esp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw6 addps nb131nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb131nf_Vvdwtot(%esp) ## repulsion movlps 16(%esi,%eax,4),%xmm5 movlps 16(%esi,%ecx,4),%xmm7 movhps 16(%esi,%ebx,4),%xmm5 movhps 16(%esi,%edx,4),%xmm7 ## got half repulsion table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%esi,%eax,4),%xmm7 movlps 24(%esi,%ecx,4),%xmm3 movhps 24(%esi,%ebx,4),%xmm7 movhps 24(%esi,%edx,4),%xmm3 ## other half of repulsion table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb131nf_c12(%esp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw12 addps nb131nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb131nf_Vvdwtot(%esp) movaps nb131nf_rinvO(%esp),%xmm2 mulps nb131nf_qqO(%esp),%xmm2 ## H1 & H2 interactions movaps nb131nf_rinvH1(%esp),%xmm3 addps nb131nf_rinvH2(%esp),%xmm3 mulps nb131nf_qqH(%esp),%xmm3 addps %xmm3,%xmm2 addps nb131nf_vctot(%esp),%xmm2 movaps %xmm2,nb131nf_vctot(%esp) ## should we do one more iteration? subl $4,nb131nf_innerk(%esp) jl _nb_kernel131nf_ia32_sse.nb131nf_odd_inner jmp _nb_kernel131nf_ia32_sse.nb131nf_unroll_loop_nb_kernel131nf_ia32_sse.nb131nf_odd_inner: addl $4,nb131nf_innerk(%esp) jnz _nb_kernel131nf_ia32_sse.nb131nf_odd_loop jmp _nb_kernel131nf_ia32_sse.nb131nf_updateouterdata_nb_kernel131nf_ia32_sse.nb131nf_odd_loop: movl nb131nf_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb131nf_innerjjnr(%esp) xorps %xmm4,%xmm4 movss nb131nf_iqO(%esp),%xmm4 movl nb131nf_charge(%ebp),%esi movhps nb131nf_iqH(%esp),%xmm4 movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb131nf_qqO(%esp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movl nb131nf_type(%ebp),%esi movl (%esi,%eax,4),%ebx movl nb131nf_vdwparam(%ebp),%esi shll %ebx addl nb131nf_ntia(%esp),%ebx movlps (%esi,%ebx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb131nf_c6(%esp) movaps %xmm7,nb131nf_c12(%esp) movl nb131nf_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## move j coords to xmm0-xmm2 movss (%esi,%eax,4),%xmm0 movss 4(%esi,%eax,4),%xmm1 movss 8(%esi,%eax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movss nb131nf_ixO(%esp),%xmm3 movss nb131nf_iyO(%esp),%xmm4 movss nb131nf_izO(%esp),%xmm5 movlps nb131nf_ixH1(%esp),%xmm6 movlps nb131nf_ixH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb131nf_iyH1(%esp),%xmm6 movlps nb131nf_iyH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb131nf_izH1(%esp),%xmm6 movlps nb131nf_izH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb131nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb131nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv ## a little trick to avoid NaNs: ## positions 0,2,and 3 are valid, but not 1. ## If it contains NaN it doesnt help to mult by 0, ## So we shuffle it and copy pos 0 to pos1! shufps $224,%xmm0,%xmm0 ## constant 11100000 ## rsq still in xmm4, rinv in xmm0. mulps %xmm0,%xmm4 mulps nb131nf_tsc(%esp),%xmm4 ## rtab cvttps2pi %xmm4,%mm6 cvtpi2ps %mm6,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 movl nb131nf_VFtab(%ebp),%esi movd %mm6,%eax ## dispersion movlps (%esi,%eax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb131nf_c6(%esp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw6 addss nb131nf_Vvdwtot(%esp),%xmm5 movss %xmm5,nb131nf_Vvdwtot(%esp) ## repulsion movlps 16(%esi,%eax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%esi,%eax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb131nf_c12(%esp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw12 addss nb131nf_Vvdwtot(%esp),%xmm5 movss %xmm5,nb131nf_Vvdwtot(%esp) ## xmm0=rinv mulps nb131nf_qqO(%esp),%xmm0 ## xmm0=vcoul addps nb131nf_vctot(%esp),%xmm0 movaps %xmm0,nb131nf_vctot(%esp) decl nb131nf_innerk(%esp) jz _nb_kernel131nf_ia32_sse.nb131nf_updateouterdata jmp _nb_kernel131nf_ia32_sse.nb131nf_odd_loop_nb_kernel131nf_ia32_sse.nb131nf_updateouterdata: ## get n from stack movl nb131nf_n(%esp),%esi ## get group index for i particle movl nb131nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb131nf_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 nb131nf_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 nb131nf_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 nb131nf_Vvdw(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb131nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel131nf_ia32_sse.nb131nf_outerend ## not last, iterate outer loop once more! movl %esi,nb131nf_n(%esp) jmp _nb_kernel131nf_ia32_sse.nb131nf_outer_nb_kernel131nf_ia32_sse.nb131nf_outerend: ## check if more outer neighborlists remain movl nb131nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel131nf_ia32_sse.nb131nf_end ## non-zero, do one more workunit jmp _nb_kernel131nf_ia32_sse.nb131nf_threadloop_nb_kernel131nf_ia32_sse.nb131nf_end: emms movl nb131nf_nouter(%esp),%eax movl nb131nf_ninner(%esp),%ebx movl nb131nf_outeriter(%ebp),%ecx movl nb131nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb131nf_salign(%esp),%eax addl %eax,%esp addl $492,%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?