nb_kernel231_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,053 行 · 第 1/5 页
S
2,053 行
mulps nb231nf_krf(%rsp),%xmm0 mulps nb231nf_krf(%rsp),%xmm1 mulps nb231nf_krf(%rsp),%xmm2 movaps %xmm0,nb231nf_krsqH2(%rsp) movaps %xmm1,nb231nf_krsqH1(%rsp) movaps %xmm2,nb231nf_krsqO(%rsp) ## start with rsqO (still in xmm7) - seed to xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb231nf_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 nb231nf_half(%rsp),%xmm4 movaps %xmm4,nb231nf_rinvO(%rsp) ## rinvO in xmm0, rsqO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb231nf_three(%rsp),%xmm4 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb231nf_half(%rsp),%xmm4 movaps %xmm4,nb231nf_rinvH1(%rsp) ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb231nf_three(%rsp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb231nf_half(%rsp),%xmm4 movaps %xmm4,nb231nf_rinvH2(%rsp) ## rinvH2 in xmm5 ## do O table interactions - rsqO in xmm7. mulps nb231nf_rinvO(%rsp),%xmm7 mulps nb231nf_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 nb231nf_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 nb231nf_c6(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw6 addps nb231nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb231nf_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 nb231nf_c12(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw12 addps nb231nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb231nf_Vvdwtot(%rsp) movaps nb231nf_rinvO(%rsp),%xmm2 movaps nb231nf_krsqO(%rsp),%xmm1 addps %xmm1,%xmm2 ## rinv+krsq subps nb231nf_crf(%rsp),%xmm2 ## rinv+krsq-crf mulps nb231nf_qqO(%rsp),%xmm2 addps nb231nf_vctot(%rsp),%xmm2 movaps %xmm2,nb231nf_vctot(%rsp) ## H1 interactions movaps nb231nf_rinvH1(%rsp),%xmm2 movaps nb231nf_krsqH1(%rsp),%xmm1 addps %xmm1,%xmm2 ## rinv+krsq subps nb231nf_crf(%rsp),%xmm2 ## rinv+krsq-crf mulps nb231nf_qqH(%rsp),%xmm2 addps nb231nf_vctot(%rsp),%xmm2 movaps %xmm2,nb231nf_vctot(%rsp) ## H2 interactions movaps nb231nf_rinvH2(%rsp),%xmm2 movaps nb231nf_krsqH2(%rsp),%xmm1 addps %xmm1,%xmm2 ## rinv+krsq subps nb231nf_crf(%rsp),%xmm2 ## rinv+krsq-crf mulps nb231nf_qqH(%rsp),%xmm2 addps nb231nf_vctot(%rsp),%xmm2 movaps %xmm2,nb231nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb231nf_innerk(%rsp) jl _nb_kernel231nf_x86_64_sse.nb231nf_odd_inner jmp _nb_kernel231nf_x86_64_sse.nb231nf_unroll_loop_nb_kernel231nf_x86_64_sse.nb231nf_odd_inner: addl $4,nb231nf_innerk(%rsp) jnz _nb_kernel231nf_x86_64_sse.nb231nf_odd_loop jmp _nb_kernel231nf_x86_64_sse.nb231nf_updateouterdata_nb_kernel231nf_x86_64_sse.nb231nf_odd_loop: movq nb231nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb231nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb231nf_iqO(%rsp),%xmm4 movq nb231nf_charge(%rbp),%rsi movhps nb231nf_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb231nf_qqO(%rsp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movq nb231nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb231nf_vdwparam(%rbp),%rsi shll %ebx addl nb231nf_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,nb231nf_c6(%rsp) movaps %xmm7,nb231nf_c12(%rsp) movq nb231nf_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## 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 movss nb231nf_ixO(%rsp),%xmm3 movss nb231nf_iyO(%rsp),%xmm4 movss nb231nf_izO(%rsp),%xmm5 movlps nb231nf_ixH1(%rsp),%xmm6 movlps nb231nf_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb231nf_iyH1(%rsp),%xmm6 movlps nb231nf_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb231nf_izH1(%rsp),%xmm6 movlps nb231nf_izH2(%rsp),%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 movaps %xmm4,%xmm0 mulps nb231nf_krf(%rsp),%xmm0 movaps %xmm0,nb231nf_krsqO(%rsp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb231nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb231nf_half(%rsp),%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 nb231nf_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 nb231nf_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 nb231nf_c6(%rsp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw6 addss nb231nf_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb231nf_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 nb231nf_c12(%rsp),%xmm4 mulss %xmm4,%xmm5 ## Vvdw12 addss nb231nf_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb231nf_Vvdwtot(%rsp) movaps %xmm0,%xmm1 ## xmm1=rinv movaps %xmm0,%xmm4 movaps nb231nf_krsqO(%rsp),%xmm3 addps %xmm3,%xmm0 ## xmm0=rinv+ krsq subps nb231nf_crf(%rsp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb231nf_qqO(%rsp),%xmm0 ## xmm0=vcoul addps nb231nf_vctot(%rsp),%xmm0 movaps %xmm0,nb231nf_vctot(%rsp) decl nb231nf_innerk(%rsp) jz _nb_kernel231nf_x86_64_sse.nb231nf_updateouterdata jmp _nb_kernel231nf_x86_64_sse.nb231nf_odd_loop_nb_kernel231nf_x86_64_sse.nb231nf_updateouterdata: ## get n from stack movl nb231nf_n(%rsp),%esi ## get group index for i particle movq nb231nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb231nf_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 nb231nf_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 nb231nf_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 nb231nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb231nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel231nf_x86_64_sse.nb231nf_outerend ## not last, iterate outer loop once more! movl %esi,nb231nf_n(%rsp) jmp _nb_kernel231nf_x86_64_sse.nb231nf_outer_nb_kernel231nf_x86_64_sse.nb231nf_outerend: ## check if more outer neighborlists remain movl nb231nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel231nf_x86_64_sse.nb231nf_end ## non-zero, do one more workunit jmp _nb_kernel231nf_x86_64_sse.nb231nf_threadloop_nb_kernel231nf_x86_64_sse.nb231nf_end: movl nb231nf_nouter(%rsp),%eax movl nb231nf_ninner(%rsp),%ebx movq nb231nf_outeriter(%rbp),%rcx movq nb231nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $552,%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?