nb_kernel110_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,440 行 · 第 1/3 页
S
1,440 行
movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb110nf_is3(%rsp) ## store is3 movq nb110nf_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb110nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb110nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb110nf_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 movq nb110nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb110nf_ntype(%rsp),%edx shll %edx movl %edx,nb110nf_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb110nf_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 movaps %xmm3,nb110nf_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb110nf_ix(%rsp) movaps %xmm1,nb110nf_iy(%rsp) movaps %xmm2,nb110nf_iz(%rsp) movl %ebx,nb110nf_ii3(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb110nf_vctot(%rsp) movaps %xmm4,nb110nf_Vvdwtot(%rsp) movq nb110nf_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 nb110nf_pos(%rbp),%rsi movq nb110nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb110nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb110nf_ninner(%rsp),%ecx movl %ecx,nb110nf_ninner(%rsp) addl $0,%edx movl %edx,nb110nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel110nf_x86_64_sse.nb110nf_unroll_loop jmp _nb_kernel110nf_x86_64_sse.nb110nf_finish_inner_nb_kernel110nf_x86_64_sse.nb110nf_unroll_loop: ## quad-unroll innerloop here movq nb110nf_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,nb110nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb110nf_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 movaps nb110nf_iq(%rsp),%xmm2 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movq nb110nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb110nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb110nf_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,nb110nf_c6(%rsp) movaps %xmm6,nb110nf_c12(%rsp) movq nb110nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx mulps %xmm2,%xmm3 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 ix-iz to xmm4-xmm6 movaps nb110nf_ix(%rsp),%xmm4 movaps nb110nf_iy(%rsp),%xmm5 movaps nb110nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%rsp),%xmm1 mulps nb110nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb110nf_Vvdwtot(%rsp),%xmm5 addps nb110nf_vctot(%rsp),%xmm3 movaps %xmm3,nb110nf_vctot(%rsp) movaps %xmm5,nb110nf_Vvdwtot(%rsp) ## should we do one more iteration? subl $4,nb110nf_innerk(%rsp) jl _nb_kernel110nf_x86_64_sse.nb110nf_finish_inner jmp _nb_kernel110nf_x86_64_sse.nb110nf_unroll_loop_nb_kernel110nf_x86_64_sse.nb110nf_finish_inner: ## check if at least two particles remain addl $4,nb110nf_innerk(%rsp) movl nb110nf_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel110nf_x86_64_sse.nb110nf_dopair jmp _nb_kernel110nf_x86_64_sse.nb110nf_checksingle_nb_kernel110nf_x86_64_sse.nb110nf_dopair: movq nb110nf_charge(%rbp),%rsi movq nb110nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movl 4(%rcx),%ebx addq $8,nb110nf_innerjjnr(%rsp) xorps %xmm3,%xmm3 movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rbx,4),%xmm6 shufps $12,%xmm6,%xmm3 ## 00001100 shufps $88,%xmm3,%xmm3 ## 01011000 ;# xmm3(0,1) has the charges movq nb110nf_type(%rbp),%rsi movl %eax,%ecx movl %ebx,%edx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb110nf_vdwparam(%rbp),%rsi shll %ecx shll %edx movl nb110nf_ntia(%rsp),%edi addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rcx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm6 movq nb110nf_pos(%rbp),%rdi xorps %xmm7,%xmm7 movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## 00001000 shufps $13,%xmm6,%xmm6 ## 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb110nf_c6(%rsp) movaps %xmm6,nb110nf_c12(%rsp) lea (%rax,%rax,2),%rax lea (%rbx,%rbx,2),%rbx ## move coordinates to xmm0-xmm2 movlps (%rdi,%rax,4),%xmm1 movss 8(%rdi,%rax,4),%xmm2 movhps (%rdi,%rbx,4),%xmm1 movss 8(%rdi,%rbx,4),%xmm0 mulps nb110nf_iq(%rsp),%xmm3 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## 10001000 shufps $136,%xmm0,%xmm0 ## 10001000 shufps $221,%xmm1,%xmm1 ## 11011101 ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb110nf_ix(%rsp),%xmm4 movaps nb110nf_iy(%rsp),%xmm5 movaps nb110nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%rsp),%xmm1 mulps nb110nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb110nf_Vvdwtot(%rsp),%xmm5 addps nb110nf_vctot(%rsp),%xmm3 movaps %xmm3,nb110nf_vctot(%rsp) movaps %xmm5,nb110nf_Vvdwtot(%rsp)_nb_kernel110nf_x86_64_sse.nb110nf_checksingle: movl nb110nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel110nf_x86_64_sse.nb110nf_dosingle jmp _nb_kernel110nf_x86_64_sse.nb110nf_updateouterdata_nb_kernel110nf_x86_64_sse.nb110nf_dosingle: movq nb110nf_charge(%rbp),%rsi movq nb110nf_pos(%rbp),%rdi movq nb110nf_innerjjnr(%rsp),%rcx xorps %xmm3,%xmm3 movl (%rcx),%eax movss (%rsi,%rax,4),%xmm3 ## xmm3(0) has the charge movq nb110nf_type(%rbp),%rsi movl %eax,%ecx movl (%rsi,%rcx,4),%ecx movq nb110nf_vdwparam(%rbp),%rsi shll %ecx addl nb110nf_ntia(%rsp),%ecx xorps %xmm6,%xmm6 movlps (%rsi,%rcx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## 11111100 shufps $253,%xmm6,%xmm6 ## 11111101 movaps %xmm4,nb110nf_c6(%rsp) movaps %xmm6,nb110nf_c12(%rsp) lea (%rax,%rax,2),%rax ## move coordinates to xmm0-xmm2 movss (%rdi,%rax,4),%xmm0 movss 4(%rdi,%rax,4),%xmm1 movss 8(%rdi,%rax,4),%xmm2 mulps nb110nf_iq(%rsp),%xmm3 xorps %xmm7,%xmm7 movaps nb110nf_ix(%rsp),%xmm4 movaps nb110nf_iy(%rsp),%xmm5 movaps nb110nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb110nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb110nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulps %xmm4,%xmm1 mulps %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb110nf_c6(%rsp),%xmm1 mulps nb110nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addss nb110nf_Vvdwtot(%rsp),%xmm5 addss nb110nf_vctot(%rsp),%xmm3 movss %xmm3,nb110nf_vctot(%rsp) movss %xmm5,nb110nf_Vvdwtot(%rsp)_nb_kernel110nf_x86_64_sse.nb110nf_updateouterdata: ## get n from stack movl nb110nf_n(%rsp),%esi ## get group index for i particle movq nb110nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb110nf_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 nb110nf_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 nb110nf_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 nb110nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb110nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel110nf_x86_64_sse.nb110nf_outerend ## not last, iterate outer loop once more! movl %esi,nb110nf_n(%rsp) jmp _nb_kernel110nf_x86_64_sse.nb110nf_outer_nb_kernel110nf_x86_64_sse.nb110nf_outerend: ## check if more outer neighborlists remain movl nb110nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel110nf_x86_64_sse.nb110nf_end ## non-zero, do one more workunit jmp _nb_kernel110nf_x86_64_sse.nb110nf_threadloop_nb_kernel110nf_x86_64_sse.nb110nf_end: movl nb110nf_nouter(%rsp),%eax movl nb110nf_ninner(%rsp),%ebx movq nb110nf_outeriter(%rbp),%rcx movq nb110nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $280,%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?