nb_kernel213_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,930 行 · 第 1/4 页
S
1,930 行
shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm6,nb213nf_izH1(%rsp) movaps %xmm0,nb213nf_ixH2(%rsp) movaps %xmm1,nb213nf_iyH2(%rsp) movaps %xmm2,nb213nf_izH2(%rsp) movaps %xmm3,nb213nf_ixM(%rsp) movaps %xmm4,nb213nf_iyM(%rsp) movaps %xmm5,nb213nf_izM(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb213nf_vctot(%rsp) movaps %xmm4,nb213nf_Vvdwtot(%rsp) movq nb213nf_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 nb213nf_pos(%rbp),%rsi movq nb213nf_faction(%rbp),%rdi movq nb213nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb213nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb213nf_ninner(%rsp),%ecx movl %ecx,nb213nf_ninner(%rsp) addl $0,%edx movl %edx,nb213nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel213nf_x86_64_sse.nb213nf_unroll_loop jmp _nb_kernel213nf_x86_64_sse.nb213nf_odd_inner_nb_kernel213nf_x86_64_sse.nb213nf_unroll_loop: ## quad-unroll innerloop here movq nb213nf_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,nb213nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb213nf_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 nb213nf_iqM(%rsp),%xmm3 mulps nb213nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb213nf_qqM(%rsp) movaps %xmm4,nb213nf_qqH(%rsp) movq nb213nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb213nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb213nf_ntia(%rsp),%edi addq %rdi,%rax addq %rdi,%rbx addq %rdi,%rcx addq %rdi,%rdx 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,nb213nf_c6(%rsp) movaps %xmm6,nb213nf_c12(%rsp) movq nb213nf_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 nb213nf_ixO(%rsp),%xmm4 movaps nb213nf_iyO(%rsp),%xmm5 movaps nb213nf_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 nb213nf_ixH1(%rsp),%xmm4 movaps nb213nf_iyH1(%rsp),%xmm5 movaps nb213nf_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 nb213nf_ixH2(%rsp),%xmm3 movaps nb213nf_iyH2(%rsp),%xmm4 movaps nb213nf_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 nb213nf_iyM(%rsp),%xmm3 movaps nb213nf_izM(%rsp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb213nf_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 movaps %xmm4,%xmm0 movaps %xmm5,%xmm1 movaps %xmm6,%xmm2 mulps nb213nf_krf(%rsp),%xmm0 mulps nb213nf_krf(%rsp),%xmm1 mulps nb213nf_krf(%rsp),%xmm2 movaps %xmm0,nb213nf_krsqM(%rsp) movaps %xmm1,nb213nf_krsqH2(%rsp) movaps %xmm2,nb213nf_krsqH1(%rsp) ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb213nf_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 nb213nf_half(%rsp),%xmm0 movaps %xmm0,nb213nf_rinvH1(%rsp) ## rinvH1 ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb213nf_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 nb213nf_half(%rsp),%xmm0 movaps %xmm0,nb213nf_rinvH2(%rsp) ## rinvH2 ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb213nf_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 nb213nf_half(%rsp),%xmm0 movaps %xmm0,nb213nf_rinvM(%rsp) ## Do the O LJ-only interaction directly. rcpps %xmm7,%xmm2 movaps nb213nf_two(%rsp),%xmm1 mulps %xmm2,%xmm7 subps %xmm7,%xmm1 mulps %xmm1,%xmm2 ## rinvsq movaps %xmm2,%xmm0 mulps %xmm2,%xmm0 ## r4 mulps %xmm2,%xmm0 ## r6 movaps %xmm0,%xmm1 mulps %xmm1,%xmm1 ## r12 mulps nb213nf_c6(%rsp),%xmm0 mulps nb213nf_c12(%rsp),%xmm1 movaps %xmm1,%xmm3 subps %xmm0,%xmm3 ## Vvdw12-Vvdw6 addps nb213nf_Vvdwtot(%rsp),%xmm3 movaps %xmm3,nb213nf_Vvdwtot(%rsp) ## do H1 interactions movaps nb213nf_rinvH1(%rsp),%xmm7 addps nb213nf_krsqH1(%rsp),%xmm7 subps nb213nf_crf(%rsp),%xmm7 ## xmm7=rinv+ krsq-crf mulps nb213nf_qqH(%rsp),%xmm7 addps nb213nf_vctot(%rsp),%xmm7 ## H2 interactions movaps nb213nf_rinvH2(%rsp),%xmm6 addps nb213nf_krsqH2(%rsp),%xmm6 subps nb213nf_crf(%rsp),%xmm6 ## xmm6=rinv+ krsq-crf mulps nb213nf_qqH(%rsp),%xmm6 addps %xmm7,%xmm6 ## M interactions movaps nb213nf_rinvM(%rsp),%xmm5 addps nb213nf_krsqM(%rsp),%xmm5 subps nb213nf_crf(%rsp),%xmm5 ## xmm5=rinv+ krsq-crf mulps nb213nf_qqM(%rsp),%xmm5 addps %xmm6,%xmm5 movaps %xmm5,nb213nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb213nf_innerk(%rsp) jl _nb_kernel213nf_x86_64_sse.nb213nf_odd_inner jmp _nb_kernel213nf_x86_64_sse.nb213nf_unroll_loop_nb_kernel213nf_x86_64_sse.nb213nf_odd_inner: addl $4,nb213nf_innerk(%rsp) jnz _nb_kernel213nf_x86_64_sse.nb213nf_odd_loop jmp _nb_kernel213nf_x86_64_sse.nb213nf_updateouterdata_nb_kernel213nf_x86_64_sse.nb213nf_odd_loop: movq nb213nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb213nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 ## clear reg. movss nb213nf_iqM(%rsp),%xmm4 movq nb213nf_charge(%rbp),%rsi movhps nb213nf_iqH(%rsp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb213nf_qqM(%rsp) ## use dummy qq for storage xorps %xmm6,%xmm6 movq nb213nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb213nf_vdwparam(%rbp),%rsi shll %ebx addl nb213nf_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb213nf_c6(%rsp) movaps %xmm7,nb213nf_c12(%rsp) movq nb213nf_pos(%rbp),%rsi lea (%rax,%rax,2),%rax movss nb213nf_ixO(%rsp),%xmm3 movss nb213nf_iyO(%rsp),%xmm4 movss nb213nf_izO(%rsp),%xmm5 movss nb213nf_ixH1(%rsp),%xmm0 movss nb213nf_iyH1(%rsp),%xmm1 movss nb213nf_izH1(%rsp),%xmm2 unpcklps nb213nf_ixH2(%rsp),%xmm3 ## ixO ixH2 - - unpcklps nb213nf_iyH2(%rsp),%xmm4 ## iyO iyH2 - - unpcklps nb213nf_izH2(%rsp),%xmm5 ## izO izH2 - - unpcklps nb213nf_ixM(%rsp),%xmm0 ## ixH1 ixM - - unpcklps nb213nf_iyM(%rsp),%xmm1 ## iyH1 iyM - - unpcklps nb213nf_izM(%rsp),%xmm2 ## izH1 izM - - unpcklps %xmm0,%xmm3 ## ixO ixH1 ixH2 ixM unpcklps %xmm1,%xmm4 ## same for y unpcklps %xmm2,%xmm5 ## same for z ## 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 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 movaps %xmm4,%xmm0 mulps nb213nf_krf(%rsp),%xmm0 movaps %xmm0,nb213nf_krsqM(%rsp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb213nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb213nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm1 ## xmm1=rinv addps nb213nf_krsqM(%rsp),%xmm1 subps nb213nf_crf(%rsp),%xmm1 ## xmm0=rinv+ krsq-crf mulps nb213nf_qqM(%rsp),%xmm1 addps nb213nf_vctot(%rsp),%xmm1 movaps %xmm1,nb213nf_vctot(%rsp) movaps %xmm0,%xmm1 mulps %xmm1,%xmm1 movaps %xmm1,%xmm2 mulss %xmm1,%xmm1 mulss %xmm2,%xmm1 ## xmm1=rinvsix xorps %xmm4,%xmm4 movss %xmm1,%xmm4 mulss %xmm4,%xmm4 ## xmm4=rinvtwelve mulss nb213nf_c6(%rsp),%xmm1 mulss nb213nf_c12(%rsp),%xmm4 movaps %xmm4,%xmm3 subss %xmm1,%xmm3 ## xmm3=Vvdw12-Vvdw6 addss nb213nf_Vvdwtot(%rsp),%xmm3 movss %xmm3,nb213nf_Vvdwtot(%rsp) decl nb213nf_innerk(%rsp) jz _nb_kernel213nf_x86_64_sse.nb213nf_updateouterdata jmp _nb_kernel213nf_x86_64_sse.nb213nf_odd_loop_nb_kernel213nf_x86_64_sse.nb213nf_updateouterdata: ## get n from stack movl nb213nf_n(%rsp),%esi ## get group index for i particle movq nb213nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb213nf_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 nb213nf_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 nb213nf_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 nb213nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb213nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel213nf_x86_64_sse.nb213nf_outerend ## not last, iterate outer loop once more! movl %esi,nb213nf_n(%rsp) jmp _nb_kernel213nf_x86_64_sse.nb213nf_outer_nb_kernel213nf_x86_64_sse.nb213nf_outerend: ## check if more outer neighborlists remain movl nb213nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel213nf_x86_64_sse.nb213nf_end ## non-zero, do one more workunit jmp _nb_kernel213nf_x86_64_sse.nb213nf_threadloop_nb_kernel213nf_x86_64_sse.nb213nf_end: movl nb213nf_nouter(%rsp),%eax movl nb213nf_ninner(%rsp),%ebx movq nb213nf_outeriter(%rbp),%rcx movq nb213nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $600,%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?