nb_kernel233_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,059 行 · 第 1/5 页
S
2,059 行
shufps $0,%xmm5,%xmm5 movaps %xmm6,nb233nf_izH1(%esp) movaps %xmm0,nb233nf_ixH2(%esp) movaps %xmm1,nb233nf_iyH2(%esp) movaps %xmm2,nb233nf_izH2(%esp) movaps %xmm3,nb233nf_ixM(%esp) movaps %xmm4,nb233nf_iyM(%esp) movaps %xmm5,nb233nf_izM(%esp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb233nf_vctot(%esp) movaps %xmm4,nb233nf_Vvdwtot(%esp) movl nb233nf_jindex(%ebp),%eax movl (%eax,%esi,4),%ecx ## jindex[n] movl 4(%eax,%esi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movl nb233nf_pos(%ebp),%esi movl nb233nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb233nf_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb233nf_ninner(%esp),%ecx movl %ecx,nb233nf_ninner(%esp) addl $0,%edx movl %edx,nb233nf_innerk(%esp) ## number of innerloop atoms jge _nb_kernel233nf_ia32_sse.nb233nf_unroll_loop jmp _nb_kernel233nf_ia32_sse.nb233nf_odd_inner_nb_kernel233nf_ia32_sse.nb233nf_unroll_loop: ## quad-unroll innerloop here movl nb233nf_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx movl 8(%edx),%ecx movl 12(%edx),%edx ## eax-edx=jnr1-4 addl $16,nb233nf_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb233nf_charge(%ebp),%esi ## base of charge[] movss (%esi,%eax,4),%xmm3 movss (%esi,%ecx,4),%xmm4 movss (%esi,%ebx,4),%xmm6 movss (%esi,%edx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb233nf_iqM(%esp),%xmm3 mulps nb233nf_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb233nf_qqM(%esp) movaps %xmm4,nb233nf_qqH(%esp) movl nb233nf_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb233nf_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb233nf_ntia(%esp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%esi,%eax,4),%xmm6 movlps (%esi,%ecx,4),%xmm7 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,nb233nf_c6(%esp) movaps %xmm6,nb233nf_c12(%esp) movl nb233nf_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx 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 ixO-izO to xmm4-xmm6 movaps nb233nf_ixO(%esp),%xmm4 movaps nb233nf_iyO(%esp),%xmm5 movaps nb233nf_izO(%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 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb233nf_ixH1(%esp),%xmm4 movaps nb233nf_iyH1(%esp),%xmm5 movaps nb233nf_izH1(%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,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb233nf_ixH2(%esp),%xmm3 movaps nb233nf_iyH2(%esp),%xmm4 movaps nb233nf_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 ## move ixM-izM to xmm2-xmm4 movaps nb233nf_iyM(%esp),%xmm3 movaps nb233nf_izM(%esp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb233nf_ixM(%esp),%xmm2 subps %xmm0,%xmm2 ## square it mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 addps %xmm3,%xmm4 addps %xmm2,%xmm4 ## rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movaps %xmm4,%xmm0 movaps %xmm5,%xmm1 movaps %xmm6,%xmm2 mulps nb233nf_krf(%esp),%xmm0 mulps nb233nf_krf(%esp),%xmm1 mulps nb233nf_krf(%esp),%xmm2 movaps %xmm0,nb233nf_krsqM(%esp) movaps %xmm1,nb233nf_krsqH2(%esp) movaps %xmm2,nb233nf_krsqH1(%esp) ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb233nf_three(%esp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb233nf_half(%esp),%xmm0 movaps %xmm0,nb233nf_rinvH1(%esp) ## rinvH1 ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb233nf_three(%esp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb233nf_half(%esp),%xmm0 movaps %xmm0,nb233nf_rinvH2(%esp) ## rinvH2 ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb233nf_three(%esp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb233nf_half(%esp),%xmm0 movaps %xmm0,nb233nf_rinvM(%esp) ## Do the O LJ-only interaction directly. ## rsqO is in xmm7 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb233nf_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 nb233nf_half(%esp),%xmm4 movaps %xmm4,%xmm0 ## xmm0=rinvO mulps %xmm0,%xmm7 mulps nb233nf_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 nb233nf_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 nb233nf_c6(%esp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw6 addps nb233nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb233nf_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 nb233nf_c12(%esp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw12 addps nb233nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb233nf_Vvdwtot(%esp) ## Do H1 interaction movaps nb233nf_rinvH1(%esp),%xmm7 movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq movaps %xmm7,%xmm0 movaps nb233nf_krsqH1(%esp),%xmm1 addps %xmm1,%xmm0 subps nb233nf_crf(%esp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb233nf_qqH(%esp),%xmm0 addps nb233nf_vctot(%esp),%xmm0 movaps %xmm0,nb233nf_vctot(%esp) ## Done with H1, do H2 interactions movaps nb233nf_rinvH2(%esp),%xmm7 movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq movaps %xmm7,%xmm0 movaps nb233nf_krsqH2(%esp),%xmm1 addps %xmm1,%xmm0 subps nb233nf_crf(%esp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb233nf_qqH(%esp),%xmm0 addps nb233nf_vctot(%esp),%xmm0 movaps %xmm0,nb233nf_vctot(%esp) ## Done with H2, do M interactions movaps nb233nf_rinvM(%esp),%xmm7 movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq movaps %xmm7,%xmm0 movaps nb233nf_krsqM(%esp),%xmm1 addps %xmm1,%xmm0 subps nb233nf_crf(%esp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb233nf_qqM(%esp),%xmm0 addps nb233nf_vctot(%esp),%xmm0 movaps %xmm0,nb233nf_vctot(%esp) ## should we do one more iteration? subl $4,nb233nf_innerk(%esp) jl _nb_kernel233nf_ia32_sse.nb233nf_odd_inner jmp _nb_kernel233nf_ia32_sse.nb233nf_unroll_loop_nb_kernel233nf_ia32_sse.nb233nf_odd_inner: addl $4,nb233nf_innerk(%esp) jnz _nb_kernel233nf_ia32_sse.nb233nf_odd_loop jmp _nb_kernel233nf_ia32_sse.nb233nf_updateouterdata_nb_kernel233nf_ia32_sse.nb233nf_odd_loop: movl nb233nf_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb233nf_innerjjnr(%esp) xorps %xmm4,%xmm4 ## clear reg. movss nb233nf_iqM(%esp),%xmm4 movl nb233nf_charge(%ebp),%esi movhps nb233nf_iqH(%esp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb233nf_qqM(%esp) ## use dummy qq for storage xorps %xmm6,%xmm6 movl nb233nf_type(%ebp),%esi movl (%esi,%eax,4),%ebx movl nb233nf_vdwparam(%ebp),%esi shll %ebx addl nb233nf_ntia(%esp),%ebx movlps (%esi,%ebx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps⌨️ 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?