nb_kernel111_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,633 行 · 第 1/4 页
S
1,633 行
movq nb111nf_pos(%rbp),%rsi movq nb111nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb111nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb111nf_ninner(%rsp),%ecx movl %ecx,nb111nf_ninner(%rsp) addl $0,%edx movl %edx,nb111nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel111nf_x86_64_sse.nb111nf_unroll_loop jmp _nb_kernel111nf_x86_64_sse.nb111nf_odd_inner_nb_kernel111nf_x86_64_sse.nb111nf_unroll_loop: ## quad-unroll innerloop here movq nb111nf_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,nb111nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb111nf_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 nb111nf_iqO(%rsp),%xmm3 mulps nb111nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb111nf_qqO(%rsp) movaps %xmm4,nb111nf_qqH(%rsp) movq nb111nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb111nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb111nf_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,nb111nf_c6(%rsp) movaps %xmm6,nb111nf_c12(%rsp) movq nb111nf_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 nb111nf_ixO(%rsp),%xmm4 movaps nb111nf_iyO(%rsp),%xmm5 movaps nb111nf_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 nb111nf_ixH1(%rsp),%xmm4 movaps nb111nf_iyH1(%rsp),%xmm5 movaps nb111nf_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 nb111nf_ixH2(%rsp),%xmm3 movaps nb111nf_iyH2(%rsp),%xmm4 movaps nb111nf_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 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## start with rsqO - seed in xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb111nf_three(%rsp),%xmm4 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb111nf_half(%rsp),%xmm4 movaps %xmm4,%xmm7 ## rinvO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb111nf_three(%rsp),%xmm4 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb111nf_half(%rsp),%xmm4 movaps %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb111nf_three(%rsp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb111nf_half(%rsp),%xmm4 movaps %xmm4,%xmm5 ## rinvH2 in xmm5 ## do O interactions movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb111nf_qqO(%rsp),%xmm7 ## xmm7=vcoul mulps nb111nf_c6(%rsp),%xmm1 mulps nb111nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm3 subps %xmm1,%xmm3 ## Vvdw=Vvdw12-Vvdw6 addps nb111nf_Vvdwtot(%rsp),%xmm3 addps nb111nf_vctot(%rsp),%xmm7 movaps %xmm3,nb111nf_Vvdwtot(%rsp) movaps %xmm7,nb111nf_vctot(%rsp) ## H1 & H2 interactions addps %xmm5,%xmm6 ## add H2 rinv mulps nb111nf_qqH(%rsp),%xmm6 ## xmm6=vcoul addps nb111nf_vctot(%rsp),%xmm6 movaps %xmm6,nb111nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb111nf_innerk(%rsp) jl _nb_kernel111nf_x86_64_sse.nb111nf_odd_inner jmp _nb_kernel111nf_x86_64_sse.nb111nf_unroll_loop_nb_kernel111nf_x86_64_sse.nb111nf_odd_inner: addl $4,nb111nf_innerk(%rsp) jnz _nb_kernel111nf_x86_64_sse.nb111nf_odd_loop jmp _nb_kernel111nf_x86_64_sse.nb111nf_updateouterdata_nb_kernel111nf_x86_64_sse.nb111nf_odd_loop: movq nb111nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb111nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb111nf_iqO(%rsp),%xmm4 movq nb111nf_charge(%rbp),%rsi movhps nb111nf_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb111nf_qqO(%rsp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movq nb111nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb111nf_vdwparam(%rbp),%rsi shll %ebx addl nb111nf_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb111nf_c6(%rsp) movaps %xmm7,nb111nf_c12(%rsp) movq nb111nf_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 nb111nf_ixO(%rsp),%xmm3 movss nb111nf_iyO(%rsp),%xmm4 movss nb111nf_izO(%rsp),%xmm5 movlps nb111nf_ixH1(%rsp),%xmm6 movlps nb111nf_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb111nf_iyH1(%rsp),%xmm6 movlps nb111nf_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb111nf_izH1(%rsp),%xmm6 movlps nb111nf_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 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb111nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb111nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 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 ## 11100000 movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulss %xmm4,%xmm1 movaps nb111nf_qqO(%rsp),%xmm3 mulss %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb111nf_c6(%rsp),%xmm1 mulps nb111nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subss %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb111nf_Vvdwtot(%rsp),%xmm5 addps nb111nf_vctot(%rsp),%xmm3 movaps %xmm3,nb111nf_vctot(%rsp) movaps %xmm5,nb111nf_Vvdwtot(%rsp) decl nb111nf_innerk(%rsp) jz _nb_kernel111nf_x86_64_sse.nb111nf_updateouterdata jmp _nb_kernel111nf_x86_64_sse.nb111nf_odd_loop_nb_kernel111nf_x86_64_sse.nb111nf_updateouterdata: ## get n from stack movl nb111nf_n(%rsp),%esi ## get group index for i particle movq nb111nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] movaps nb111nf_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 nb111nf_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 nb111nf_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 nb111nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb111nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel111nf_x86_64_sse.nb111nf_outerend ## not last, iterate outer loop once more! movl %esi,nb111nf_n(%rsp) jmp _nb_kernel111nf_x86_64_sse.nb111nf_outer_nb_kernel111nf_x86_64_sse.nb111nf_outerend: ## check if more outer neighborlists remain movl nb111nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel111nf_x86_64_sse.nb111nf_end ## non-zero, do one more workunit jmp _nb_kernel111nf_x86_64_sse.nb111nf_threadloop_nb_kernel111nf_x86_64_sse.nb111nf_end: movl nb111nf_nouter(%rsp),%eax movl nb111nf_ninner(%rsp),%ebx movq nb111nf_outeriter(%rbp),%rcx movq nb111nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $424,%rsp emms 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?