nb_kernel110_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,111 行 · 第 1/3 页
S
1,111 行
jg _nb_kernel110nf_x86_64_sse2.nb110nf_outerstart jmp _nb_kernel110nf_x86_64_sse2.nb110nf_end_nb_kernel110nf_x86_64_sse2.nb110nf_outerstart: ## ebx contains number of outer iterations addl nb110nf_nouter(%rsp),%ebx movl %ebx,nb110nf_nouter(%rsp)_nb_kernel110nf_x86_64_sse2.nb110nf_outer: movq nb110nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## rbx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movq nb110nf_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%xmm2 movq nb110nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb110nf_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 mulsd nb110nf_facel(%rsp),%xmm3 shufpd $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[] addsd (%rax,%rbx,8),%xmm0 addsd 8(%rax,%rbx,8),%xmm1 addsd 16(%rax,%rbx,8),%xmm2 movapd %xmm3,nb110nf_iq(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb110nf_ix(%rsp) movapd %xmm1,nb110nf_iy(%rsp) movapd %xmm2,nb110nf_iz(%rsp) movl %ebx,nb110nf_ii3(%rsp) ## clear vctot xorpd %xmm4,%xmm4 movapd %xmm4,nb110nf_vctot(%rsp) movapd %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 $2,%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_sse2.nb110nf_unroll_loop jmp _nb_kernel110nf_x86_64_sse2.nb110nf_checksingle_nb_kernel110nf_x86_64_sse2.nb110nf_unroll_loop: ## twice unrolled innerloop here movq nb110nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb110nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb110nf_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd nb110nf_iq(%rsp),%xmm5 mulpd %xmm5,%xmm3 ## qq movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movq nb110nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movq nb110nf_vdwparam(%rbp),%rsi shll %eax shll %ebx movl nb110nf_ntia(%rsp),%edi addl %edi,%eax addl %edi,%ebx movlpd (%rsi,%rax,8),%xmm6 ## c6a movlpd (%rsi,%rbx,8),%xmm7 ## c6b movhpd 8(%rsi,%rax,8),%xmm6 ## c6a c12a movhpd 8(%rsi,%rbx,8),%xmm7 ## c6b c12b movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movd %mm0,%eax movd %mm1,%ebx movapd %xmm4,nb110nf_c6(%rsp) movapd %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 ## move two coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm0 movhpd 8(%rsi,%rbx,8),%xmm1 movhpd 16(%rsi,%rbx,8),%xmm2 ## move ix-iz to xmm4-xmm6 movapd nb110nf_ix(%rsp),%xmm4 movapd nb110nf_iy(%rsp),%xmm5 movapd nb110nf_iz(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 ## rsq in xmm4 cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb110nf_three(%rsp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb110nf_half(%rsp),%xmm0 subpd %xmm2,%xmm1 ## 30-rsq*lu*lu mulpd %xmm5,%xmm1 mulpd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movapd %xmm1,%xmm5 ## copy of lu mulpd %xmm1,%xmm1 ## lu*lu movapd nb110nf_three(%rsp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb110nf_half(%rsp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=rinv movapd %xmm0,%xmm4 mulpd %xmm4,%xmm4 ## xmm4=rinvsq movapd %xmm4,%xmm1 mulpd %xmm4,%xmm1 mulpd %xmm4,%xmm1 ## xmm1=rinvsix movapd %xmm1,%xmm2 mulpd %xmm2,%xmm2 ## xmm2=rinvtwelve mulpd %xmm0,%xmm3 ## xmm3=vcoul mulpd nb110nf_c6(%rsp),%xmm1 mulpd nb110nf_c12(%rsp),%xmm2 movapd %xmm2,%xmm5 subpd %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addpd nb110nf_Vvdwtot(%rsp),%xmm5 addpd nb110nf_vctot(%rsp),%xmm3 movapd %xmm3,nb110nf_vctot(%rsp) movapd %xmm5,nb110nf_Vvdwtot(%rsp) ## should we do one more iteration? subl $2,nb110nf_innerk(%rsp) jl _nb_kernel110nf_x86_64_sse2.nb110nf_checksingle jmp _nb_kernel110nf_x86_64_sse2.nb110nf_unroll_loop_nb_kernel110nf_x86_64_sse2.nb110nf_checksingle: movl nb110nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel110nf_x86_64_sse2.nb110nf_dosingle jmp _nb_kernel110nf_x86_64_sse2.nb110nf_updateouterdata_nb_kernel110nf_x86_64_sse2.nb110nf_dosingle: movq nb110nf_charge(%rbp),%rsi movq nb110nf_pos(%rbp),%rdi movq nb110nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax xorpd %xmm3,%xmm3 movlpd (%rsi,%rax,8),%xmm3 movapd nb110nf_iq(%rsp),%xmm5 mulsd %xmm5,%xmm3 ## qq movd %eax,%mm0 ## use mmx registers as temp storage movq nb110nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movq nb110nf_vdwparam(%rbp),%rsi shll %eax movl nb110nf_ntia(%rsp),%edi addl %edi,%eax movlpd (%rsi,%rax,8),%xmm6 ## c6a movhpd 8(%rsi,%rax,8),%xmm6 ## c6a c12a xorpd %xmm7,%xmm7 movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movd %mm0,%eax movapd %xmm4,nb110nf_c6(%rsp) movapd %xmm6,nb110nf_c12(%rsp) movq nb110nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move two coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 ## move ix-iz to xmm4-xmm6 movapd nb110nf_ix(%rsp),%xmm4 movapd nb110nf_iy(%rsp),%xmm5 movapd nb110nf_iz(%rsp),%xmm6 ## calc dr subsd %xmm0,%xmm4 subsd %xmm1,%xmm5 subsd %xmm2,%xmm6 ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 ## rsq in xmm4 cvtsd2ss %xmm4,%xmm5 rsqrtss %xmm5,%xmm5 cvtss2sd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulsd %xmm2,%xmm2 ## lu*lu movapd nb110nf_three(%rsp),%xmm1 mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb110nf_half(%rsp),%xmm0 subsd %xmm2,%xmm1 ## 30-rsq*lu*lu mulsd %xmm5,%xmm1 mulsd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movapd %xmm1,%xmm5 ## copy of lu mulsd %xmm1,%xmm1 ## lu*lu movapd nb110nf_three(%rsp),%xmm2 mulsd %xmm4,%xmm1 ## rsq*lu*lu movapd nb110nf_half(%rsp),%xmm0 subsd %xmm1,%xmm2 ## 30-rsq*lu*lu mulsd %xmm5,%xmm2 mulsd %xmm2,%xmm0 ## xmm0=rinv movapd %xmm0,%xmm4 mulsd %xmm4,%xmm4 ## xmm4=rinvsq movapd %xmm4,%xmm1 mulsd %xmm4,%xmm1 mulsd %xmm4,%xmm1 ## xmm1=rinvsix movapd %xmm1,%xmm2 mulsd %xmm2,%xmm2 ## xmm2=rinvtwelve mulsd %xmm0,%xmm3 ## xmm3=vcoul mulsd nb110nf_c6(%rsp),%xmm1 mulsd nb110nf_c12(%rsp),%xmm2 movapd %xmm2,%xmm5 subsd %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addsd nb110nf_Vvdwtot(%rsp),%xmm5 addsd nb110nf_vctot(%rsp),%xmm3 movlpd %xmm3,nb110nf_vctot(%rsp) movlpd %xmm5,nb110nf_Vvdwtot(%rsp)_nb_kernel110nf_x86_64_sse2.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 movapd nb110nf_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb110nf_Vc(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## accumulate total lj energy and update it movapd nb110nf_Vvdwtot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb110nf_Vvdw(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## finish if last movl nb110nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel110nf_x86_64_sse2.nb110nf_outerend ## not last, iterate outer loop once more! movl %esi,nb110nf_n(%rsp) jmp _nb_kernel110nf_x86_64_sse2.nb110nf_outer_nb_kernel110nf_x86_64_sse2.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_sse2.nb110nf_end ## non-zero, do one more workunit jmp _nb_kernel110nf_x86_64_sse2.nb110nf_threadloop_nb_kernel110nf_x86_64_sse2.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 %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?