nb_kernel130_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,392 行 · 第 1/3 页
S
1,392 行
movq nb130nf_pos(%rbp),%rax ## eax = base of pos[] addsd (%rax,%rbx,8),%xmm0 addsd 8(%rax,%rbx,8),%xmm1 addsd 16(%rax,%rbx,8),%xmm2 movapd %xmm3,nb130nf_iq(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb130nf_ix(%rsp) movapd %xmm1,nb130nf_iy(%rsp) movapd %xmm2,nb130nf_iz(%rsp) movl %ebx,nb130nf_ii3(%rsp) ## clear vctot xorpd %xmm4,%xmm4 movapd %xmm4,nb130nf_vctot(%rsp) movapd %xmm4,nb130nf_Vvdwtot(%rsp) movq nb130nf_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 nb130nf_pos(%rbp),%rsi movq nb130nf_faction(%rbp),%rdi movq nb130nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb130nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb130nf_ninner(%rsp),%ecx movl %ecx,nb130nf_ninner(%rsp) addl $0,%edx movl %edx,nb130nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel130nf_x86_64_sse2.nb130nf_unroll_loop jmp _nb_kernel130nf_x86_64_sse2.nb130nf_checksingle_nb_kernel130nf_x86_64_sse2.nb130nf_unroll_loop: ## twice unrolled innerloop here movq nb130nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb130nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb130nf_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd nb130nf_iq(%rsp),%xmm5 mulpd %xmm5,%xmm3 ## qq movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movq nb130nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movq nb130nf_vdwparam(%rbp),%rsi shll %eax shll %ebx movl nb130nf_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,nb130nf_c6(%rsp) movapd %xmm6,nb130nf_c12(%rsp) movq nb130nf_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 nb130nf_ix(%rsp),%xmm4 movapd nb130nf_iy(%rsp),%xmm5 movapd nb130nf_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 nb130nf_three(%rsp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb130nf_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 nb130nf_three(%rsp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb130nf_half(%rsp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=rinv movapd %xmm0,%xmm6 mulpd %xmm3,%xmm6 ## vcoul addpd nb130nf_vctot(%rsp),%xmm6 movapd %xmm6,nb130nf_vctot(%rsp) ## LJ table interaction. xmm0=rinv, xmm4=rsq mulpd %xmm0,%xmm4 ## xmm4=r mulpd nb130nf_tsc(%rsp),%xmm4 cvttpd2pi %xmm4,%mm6 ## mm6 = lu idx cvtpi2pd %mm6,%xmm5 subpd %xmm5,%xmm4 movapd %xmm4,%xmm1 ## xmm1=eps movapd %xmm1,%xmm2 mulpd %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 ## idx *= 8 movq nb130nf_VFtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm6,%ebx ## dispersion movlpd (%rsi,%rax,8),%xmm4 ## Y1 movlpd (%rsi,%rbx,8),%xmm3 ## Y2 movhpd 8(%rsi,%rax,8),%xmm4 ## Y1 F1 movhpd 8(%rsi,%rbx,8),%xmm3 ## Y2 F2 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movlpd 16(%rsi,%rax,8),%xmm6 ## G1 movlpd 16(%rsi,%rbx,8),%xmm3 ## G2 movhpd 24(%rsi,%rax,8),%xmm6 ## G1 H1 movhpd 24(%rsi,%rbx,8),%xmm3 ## G2 H2 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## dispersion table ready, in xmm4-xmm7 mulpd %xmm1,%xmm6 ## xmm6=Geps mulpd %xmm2,%xmm7 ## xmm7=Heps2 addpd %xmm6,%xmm5 addpd %xmm7,%xmm5 ## xmm5=Fp mulpd %xmm1,%xmm5 ## xmm5=eps*Fp addpd %xmm4,%xmm5 ## xmm5=VV movapd nb130nf_c6(%rsp),%xmm4 mulpd %xmm4,%xmm5 ## Vvdw6 ##Update Vvdwtot directly addpd nb130nf_Vvdwtot(%rsp),%xmm5 movapd %xmm5,nb130nf_Vvdwtot(%rsp) ## repulsion movlpd 32(%rsi,%rax,8),%xmm4 ## Y1 movlpd 32(%rsi,%rbx,8),%xmm3 ## Y2 movhpd 40(%rsi,%rax,8),%xmm4 ## Y1 F1 movhpd 40(%rsi,%rbx,8),%xmm3 ## Y2 F2 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movlpd 48(%rsi,%rax,8),%xmm6 ## G1 movlpd 48(%rsi,%rbx,8),%xmm3 ## G2 movhpd 56(%rsi,%rax,8),%xmm6 ## G1 H1 movhpd 56(%rsi,%rbx,8),%xmm3 ## G2 H2 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## table ready, in xmm4-xmm7 mulpd %xmm1,%xmm6 ## xmm6=Geps mulpd %xmm2,%xmm7 ## xmm7=Heps2 addpd %xmm6,%xmm5 addpd %xmm7,%xmm5 ## xmm5=Fp mulpd %xmm1,%xmm5 ## xmm5=eps*Fp addpd %xmm4,%xmm5 ## xmm5=VV movapd nb130nf_c12(%rsp),%xmm4 mulpd %xmm4,%xmm5 addpd nb130nf_Vvdwtot(%rsp),%xmm5 movapd %xmm5,nb130nf_Vvdwtot(%rsp) ## should we do one more iteration? subl $2,nb130nf_innerk(%rsp) jl _nb_kernel130nf_x86_64_sse2.nb130nf_checksingle jmp _nb_kernel130nf_x86_64_sse2.nb130nf_unroll_loop_nb_kernel130nf_x86_64_sse2.nb130nf_checksingle: movl nb130nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel130nf_x86_64_sse2.nb130nf_dosingle jmp _nb_kernel130nf_x86_64_sse2.nb130nf_updateouterdata_nb_kernel130nf_x86_64_sse2.nb130nf_dosingle: movq nb130nf_charge(%rbp),%rsi movq nb130nf_pos(%rbp),%rdi movq nb130nf_innerjjnr(%rsp),%rcx xorpd %xmm3,%xmm3 movl (%rcx),%eax movlpd (%rsi,%rax,8),%xmm3 movapd nb130nf_iq(%rsp),%xmm5 mulpd %xmm5,%xmm3 ## qq movd %eax,%mm0 ## use mmx registers as temp storage movq nb130nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movq nb130nf_vdwparam(%rbp),%rsi shll %eax movl nb130nf_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,nb130nf_c6(%rsp) movapd %xmm6,nb130nf_c12(%rsp) movq nb130nf_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 nb130nf_ix(%rsp),%xmm4 movapd nb130nf_iy(%rsp),%xmm5 movapd nb130nf_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 nb130nf_three(%rsp),%xmm1 mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb130nf_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 nb130nf_three(%rsp),%xmm2 mulsd %xmm4,%xmm1 ## rsq*lu*lu movapd nb130nf_half(%rsp),%xmm0 subsd %xmm1,%xmm2 ## 30-rsq*lu*lu mulsd %xmm5,%xmm2 mulsd %xmm2,%xmm0 ## xmm0=rinv movapd %xmm0,%xmm6 movapd %xmm4,%xmm1 mulsd %xmm3,%xmm6 ## xmm6=vcoul=qq*rinv addsd nb130nf_vctot(%rsp),%xmm6 movsd %xmm6,nb130nf_vctot(%rsp) ## LJ table interaction. xmm0=rinv, cmm4=rsq mulsd %xmm0,%xmm4 ## xmm4=r mulsd nb130nf_tsc(%rsp),%xmm4 cvttsd2si %xmm4,%ebx ## mm6 = lu idx cvtsi2sd %ebx,%xmm5 subsd %xmm5,%xmm4 movsd %xmm4,%xmm1 ## xmm1=eps movsd %xmm1,%xmm2 mulsd %xmm2,%xmm2 ## xmm2=eps2 shll $3,%ebx movq nb130nf_VFtab(%rbp),%rsi ## dispersion movlpd (%rsi,%rbx,8),%xmm4 ## Y1 movhpd 8(%rsi,%rbx,8),%xmm4 ## Y1 F1 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movlpd 16(%rsi,%rbx,8),%xmm6 ## G1 movhpd 24(%rsi,%rbx,8),%xmm6 ## G1 H1 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## dispersion table ready, in xmm4-xmm7 mulsd %xmm1,%xmm6 ## xmm6=Geps mulsd %xmm2,%xmm7 ## xmm7=Heps2 addsd %xmm6,%xmm5 addsd %xmm7,%xmm5 ## xmm5=Fp mulsd %xmm1,%xmm5 ## xmm5=eps*Fp addsd %xmm4,%xmm5 ## xmm5=VV movsd nb130nf_c6(%rsp),%xmm4 mulsd %xmm4,%xmm5 ## Vvdw6 ## put scalar force on stack Update Vvdwtot directly addsd nb130nf_Vvdwtot(%rsp),%xmm5 movsd %xmm5,nb130nf_Vvdwtot(%rsp) ## repulsion movlpd 32(%rsi,%rbx,8),%xmm4 ## Y1 movhpd 40(%rsi,%rbx,8),%xmm4 ## Y1 F1 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movlpd 48(%rsi,%rbx,8),%xmm6 ## G1 movhpd 56(%rsi,%rbx,8),%xmm6 ## G1 H1 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## table ready, in xmm4-xmm7 mulsd %xmm1,%xmm6 ## xmm6=Geps mulsd %xmm2,%xmm7 ## xmm7=Heps2 addsd %xmm6,%xmm5 addsd %xmm7,%xmm5 ## xmm5=Fp mulsd %xmm1,%xmm5 ## xmm5=eps*Fp addsd %xmm4,%xmm5 ## xmm5=VV movsd nb130nf_c12(%rsp),%xmm4 mulsd %xmm4,%xmm5 addsd nb130nf_Vvdwtot(%rsp),%xmm5 movsd %xmm5,nb130nf_Vvdwtot(%rsp)_nb_kernel130nf_x86_64_sse2.nb130nf_updateouterdata: ## get n from stack movl nb130nf_n(%rsp),%esi ## get group index for i particle movq nb130nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movapd nb130nf_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb130nf_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 nb130nf_Vvdwtot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb130nf_Vvdw(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## finish if last movl nb130nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel130nf_x86_64_sse2.nb130nf_outerend ## not last, iterate outer loop once more! movl %esi,nb130nf_n(%rsp) jmp _nb_kernel130nf_x86_64_sse2.nb130nf_outer_nb_kernel130nf_x86_64_sse2.nb130nf_outerend: ## check if more outer neighborlists remain movl nb130nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel130nf_x86_64_sse2.nb130nf_end ## non-zero, do one more workunit jmp _nb_kernel130nf_x86_64_sse2.nb130nf_threadloop_nb_kernel130nf_x86_64_sse2.nb130nf_end: movl nb130nf_nouter(%rsp),%eax movl nb130nf_ninner(%rsp),%ebx movq nb130nf_outeriter(%rbp),%rcx movq nb130nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $328,%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?