nb_kernel133_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,130 行 · 第 1/5 页
S
2,130 行
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 nb133nf_ixH2(%rsp),%xmm3 movaps nb133nf_iyH2(%rsp),%xmm4 movaps nb133nf_izH2(%rsp),%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 nb133nf_iyM(%rsp),%xmm3 movaps nb133nf_izM(%rsp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb133nf_ixM(%rsp),%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 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb133nf_three(%rsp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb133nf_half(%rsp),%xmm0 movaps %xmm0,nb133nf_rinvH1(%rsp) ## rinvH1 ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb133nf_three(%rsp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb133nf_half(%rsp),%xmm0 movaps %xmm0,nb133nf_rinvH2(%rsp) ## rinvH2 ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb133nf_three(%rsp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb133nf_half(%rsp),%xmm0 movaps %xmm0,nb133nf_rinvM(%rsp) ## Do the O LJ-only interaction directly. ## rsqO is in xmm7 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb133nf_three(%rsp),%xmm4 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb133nf_half(%rsp),%xmm4 movaps %xmm4,%xmm0 ## xmm0=rinvO mulps %xmm0,%xmm7 mulps nb133nf_tsc(%rsp),%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 movq nb133nf_VFtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx ## dispersion movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half dispersion table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 movhps 8(%rsi,%rdx,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 nb133nf_c6(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw6 addps nb133nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb133nf_Vvdwtot(%rsp) ## repulsion movlps 16(%rsi,%rax,4),%xmm5 movlps 16(%rsi,%rcx,4),%xmm7 movhps 16(%rsi,%rbx,4),%xmm5 movhps 16(%rsi,%rdx,4),%xmm7 ## got half repulsion table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%rsi,%rax,4),%xmm7 movlps 24(%rsi,%rcx,4),%xmm3 movhps 24(%rsi,%rbx,4),%xmm7 movhps 24(%rsi,%rdx,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 nb133nf_c12(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw12 addps nb133nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb133nf_Vvdwtot(%rsp) ## Do H1-H2-M interactions movaps nb133nf_rinvH1(%rsp),%xmm7 addps nb133nf_rinvH2(%rsp),%xmm7 movaps nb133nf_rinvM(%rsp),%xmm6 mulps nb133nf_qqH(%rsp),%xmm7 mulps nb133nf_qqM(%rsp),%xmm6 addps %xmm6,%xmm7 addps nb133nf_vctot(%rsp),%xmm7 movaps %xmm7,nb133nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb133nf_innerk(%rsp) jl _nb_kernel133nf_x86_64_sse.nb133nf_odd_inner jmp _nb_kernel133nf_x86_64_sse.nb133nf_unroll_loop_nb_kernel133nf_x86_64_sse.nb133nf_odd_inner: addl $4,nb133nf_innerk(%rsp) jnz _nb_kernel133nf_x86_64_sse.nb133nf_odd_loop jmp _nb_kernel133nf_x86_64_sse.nb133nf_updateouterdata_nb_kernel133nf_x86_64_sse.nb133nf_odd_loop: movq nb133nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb133nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 ## clear reg. movss nb133nf_iqM(%rsp),%xmm4 movq nb133nf_charge(%rbp),%rsi movhps nb133nf_iqH(%rsp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb133nf_qqM(%rsp) ## use dummy qq for storage xorps %xmm6,%xmm6 movq nb133nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb133nf_vdwparam(%rbp),%rsi shll %ebx addl nb133nf_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb133nf_c6(%rsp) movaps %xmm7,nb133nf_c12(%rsp) movq nb133nf_pos(%rbp),%rsi lea (%rax,%rax,2),%rax movss nb133nf_ixO(%rsp),%xmm3 movss nb133nf_iyO(%rsp),%xmm4 movss nb133nf_izO(%rsp),%xmm5 movss nb133nf_ixH1(%rsp),%xmm0 movss nb133nf_iyH1(%rsp),%xmm1 movss nb133nf_izH1(%rsp),%xmm2 unpcklps nb133nf_ixH2(%rsp),%xmm3 ## ixO ixH2 - - unpcklps nb133nf_iyH2(%rsp),%xmm4 ## iyO iyH2 - - unpcklps nb133nf_izH2(%rsp),%xmm5 ## izO izH2 - - unpcklps nb133nf_ixM(%rsp),%xmm0 ## ixH1 ixM - - unpcklps nb133nf_iyM(%rsp),%xmm1 ## iyH1 iyM - - unpcklps nb133nf_izM(%rsp),%xmm2 ## izH1 izM - - unpcklps %xmm0,%xmm3 ## ixO ixH1 ixH2 ixM unpcklps %xmm1,%xmm4 ## same for y unpcklps %xmm2,%xmm5 ## same for z ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm0 movss 4(%rsi,%rax,4),%xmm1 movss 8(%rsi,%rax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 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 nb133nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb133nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv, xmm4=rsq mulps %xmm0,%xmm4 mulps nb133nf_tsc(%rsp),%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 movq nb133nf_VFtab(%rbp),%rsi movd %mm6,%eax ## dispersion movlps (%rsi,%rax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%rsi,%rax,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 nb133nf_c6(%rsp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw6 ## put scalar force on stack Update Vvdwtot directly addss nb133nf_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb133nf_Vvdwtot(%rsp) ## repulsion movlps 16(%rsi,%rax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%rsi,%rax,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 nb133nf_c12(%rsp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw12 addss nb133nf_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb133nf_Vvdwtot(%rsp) mulps nb133nf_qqM(%rsp),%xmm0 ## xmm0=vcoul addps nb133nf_vctot(%rsp),%xmm0 movaps %xmm0,nb133nf_vctot(%rsp) decl nb133nf_innerk(%rsp) jz _nb_kernel133nf_x86_64_sse.nb133nf_updateouterdata jmp _nb_kernel133nf_x86_64_sse.nb133nf_odd_loop_nb_kernel133nf_x86_64_sse.nb133nf_updateouterdata: ## get n from stack movl nb133nf_n(%rsp),%esi ## get group index for i particle movq nb133nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb133nf_vctot(%rsp),%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 movq nb133nf_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## accumulate total lj energy and update it movaps nb133nf_Vvdwtot(%rsp),%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 movq nb133nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb133nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel133nf_x86_64_sse.nb133nf_outerend ## not last, iterate outer loop once more! movl %esi,nb133nf_n(%rsp) jmp _nb_kernel133nf_x86_64_sse.nb133nf_outer_nb_kernel133nf_x86_64_sse.nb133nf_outerend: ## check if more outer neighborlists remain movl nb133nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel133nf_x86_64_sse.nb133nf_end ## non-zero, do one more workunit jmp _nb_kernel133nf_x86_64_sse.nb133nf_threadloop_nb_kernel133nf_x86_64_sse.nb133nf_end: movl nb133nf_nouter(%rsp),%eax movl nb133nf_ninner(%rsp),%ebx movq nb133nf_outeriter(%rbp),%rcx movq nb133nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $600,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp 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?