nb_kernel333_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,182 行 · 第 1/5 页
S
2,182 行
movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb333nf_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb333nf_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 ## ox addss 4(%rax,%rbx,4),%xmm4 ## oy addss 8(%rax,%rbx,4),%xmm5 ## oz addss 12(%rax,%rbx,4),%xmm6 ## h1x addss 16(%rax,%rbx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb333nf_ixO(%rsp) movaps %xmm4,nb333nf_iyO(%rsp) movaps %xmm5,nb333nf_izO(%rsp) movaps %xmm6,nb333nf_ixH1(%rsp) movaps %xmm7,nb333nf_iyH1(%rsp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%rax,%rbx,4),%xmm6 ## h1z addss 24(%rax,%rbx,4),%xmm0 ## h2x addss 28(%rax,%rbx,4),%xmm1 ## h2y addss 32(%rax,%rbx,4),%xmm2 ## h2z addss 36(%rax,%rbx,4),%xmm3 ## mx addss 40(%rax,%rbx,4),%xmm4 ## my addss 44(%rax,%rbx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm6,nb333nf_izH1(%rsp) movaps %xmm0,nb333nf_ixH2(%rsp) movaps %xmm1,nb333nf_iyH2(%rsp) movaps %xmm2,nb333nf_izH2(%rsp) movaps %xmm3,nb333nf_ixM(%rsp) movaps %xmm4,nb333nf_iyM(%rsp) movaps %xmm5,nb333nf_izM(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb333nf_vctot(%rsp) movaps %xmm4,nb333nf_Vvdwtot(%rsp) movq nb333nf_jindex(%rsp),%rax movl (%rax,%rsi,4),%ecx ## jindex[n] movl 4(%rax,%rsi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movq nb333nf_pos(%rbp),%rsi movq nb333nf_faction(%rbp),%rdi movq nb333nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb333nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb333nf_ninner(%rsp),%ecx movl %ecx,nb333nf_ninner(%rsp) addl $0,%edx movl %edx,nb333nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel333nf_x86_64_sse.nb333nf_unroll_loop jmp _nb_kernel333nf_x86_64_sse.nb333nf_odd_inner_nb_kernel333nf_x86_64_sse.nb333nf_unroll_loop: ## quad-unroll innerloop here movq nb333nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb333nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb333nf_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb333nf_iqM(%rsp),%xmm3 mulps nb333nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb333nf_qqM(%rsp) movaps %xmm4,nb333nf_qqH(%rsp) movq nb333nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb333nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb333nf_ntia(%rsp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rax,4),%xmm6 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm6 ## 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb333nf_c6(%rsp) movaps %xmm6,nb333nf_c12(%rsp) movq nb333nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx ## replace jnr with j3 lea (%rdx,%rdx,2),%rdx ## move four coordinates to xmm0-xmm2 movlps (%rsi,%rax,4),%xmm4 movlps (%rsi,%rcx,4),%xmm5 movss 8(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm6 movhps (%rsi,%rbx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movss 8(%rsi,%rbx,4),%xmm0 movss 8(%rsi,%rdx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb333nf_ixO(%rsp),%xmm4 movaps nb333nf_iyO(%rsp),%xmm5 movaps nb333nf_izO(%rsp),%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 nb333nf_ixH1(%rsp),%xmm4 movaps nb333nf_iyH1(%rsp),%xmm5 movaps nb333nf_izH1(%rsp),%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 nb333nf_ixH2(%rsp),%xmm3 movaps nb333nf_iyH2(%rsp),%xmm4 movaps nb333nf_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 nb333nf_iyM(%rsp),%xmm3 movaps nb333nf_izM(%rsp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb333nf_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 nb333nf_three(%rsp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%rsp),%xmm0 movaps %xmm0,nb333nf_rinvH1(%rsp) ## rinvH1 mulps %xmm0,%xmm6 movaps %xmm6,nb333nf_rH1(%rsp) ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%rsp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%rsp),%xmm0 movaps %xmm0,nb333nf_rinvH2(%rsp) ## rinvH2 mulps %xmm0,%xmm5 movaps %xmm5,nb333nf_rH2(%rsp) ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%rsp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%rsp),%xmm0 movaps %xmm0,nb333nf_rinvM(%rsp) ## rinvM mulps %xmm0,%xmm4 movaps %xmm4,nb333nf_rM(%rsp) ## Do the O LJ table interaction directly. rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%rsp),%xmm0 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%rsp),%xmm0 ## rinv movaps %xmm0,%xmm1 mulps %xmm7,%xmm1 ## xmm1=r mulps nb333nf_tsc(%rsp),%xmm1 ## r*tabscale movhlps %xmm1,%xmm2 cvttps2pi %xmm1,%mm6 cvttps2pi %xmm2,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm2 movlhps %xmm2,%xmm3 subps %xmm3,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movq nb333nf_VFtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx lea (%rax,%rax,2),%rax lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx ## load dispersion table data into xmm4-xmm7 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 coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## dispersion table YFGH 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 nb333nf_c6(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw6 ## Update Vvdwtot directly addps nb333nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb333nf_Vvdwtot(%rsp) ## load repulsion table data into xmm4-xmm7 movlps 32(%rsi,%rax,4),%xmm5 movlps 32(%rsi,%rcx,4),%xmm7 movhps 32(%rsi,%rbx,4),%xmm5 movhps 32(%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 40(%rsi,%rax,4),%xmm7 movlps 40(%rsi,%rcx,4),%xmm3 movhps 40(%rsi,%rbx,4),%xmm7 movhps 40(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## repulsion table YFGH 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 nb333nf_c12(%rsp),%xmm4 mulps %xmm4,%xmm5 ## Vvdw12 addps nb333nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb333nf_Vvdwtot(%rsp) ## Do H1 interaction movq nb333nf_VFtab(%rbp),%rsi movaps nb333nf_rH1(%rsp),%xmm7 mulps nb333nf_tsc(%rsp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx lea (%rax,%rax,2),%rax lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps ⌨️ 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?