nb_kernel231_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,053 行 · 第 1/5 页
S
2,053 行
addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 movaps %xmm4,%xmm0 movaps %xmm0,nb231_rsqO(%rsp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb231_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb231_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## constant 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 ## constant 11100000 ## rsq still in xmm4, rinv in xmm0. mulps %xmm0,%xmm4 mulps nb231_tsc(%rsp),%xmm4 ## rtab cvttps2pi %xmm4,%mm6 cvtpi2ps %mm6,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 movq nb231_VFtab(%rbp),%rsi movd %mm6,%r8d ## dispersion movlps (%rsi,%r8,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%rsi,%r8,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb231_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb231_c6(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijD mulss %xmm4,%xmm5 ## Vvdw6 mulss nb231_tsc(%rsp),%xmm7 ## put scalar force on stack Update Vvdwtot directly addss nb231_Vvdwtot(%rsp),%xmm5 movss %xmm7,nb231_fstmp(%rsp) movss %xmm5,nb231_Vvdwtot(%rsp) ## repulsion movlps 16(%rsi,%r8,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%rsi,%r8,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb231_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb231_c12(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijR mulss %xmm4,%xmm5 ## Vvdw12 mulss nb231_tsc(%rsp),%xmm7 addss nb231_fstmp(%rsp),%xmm7 movss %xmm7,nb231_fstmp(%rsp) addss nb231_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb231_Vvdwtot(%rsp) movaps nb231_rsqO(%rsp),%xmm3 movaps %xmm0,%xmm1 ## xmm1=rinv movaps %xmm0,%xmm4 mulps nb231_krf(%rsp),%xmm3 addps %xmm3,%xmm0 ## xmm0=rinv+ krsq mulps nb231_two(%rsp),%xmm3 subps nb231_crf(%rsp),%xmm0 ## xmm0=rinv+ krsq-crf subps %xmm3,%xmm1 ## xmm1=rinv-2*krsq mulps nb231_qqO(%rsp),%xmm0 ## xmm0=vcoul mulps nb231_qqO(%rsp),%xmm1 ## xmm1=coul part of fs mulps %xmm4,%xmm1 subss nb231_fstmp(%rsp),%xmm1 mulps %xmm1,%xmm4 addps nb231_vctot(%rsp),%xmm0 movaps %xmm0,nb231_vctot(%rsp) movaps nb231_dxO(%rsp),%xmm0 movaps nb231_dyO(%rsp),%xmm1 movaps nb231_dzO(%rsp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb231_fixO(%rsp),%xmm3 movss nb231_fiyO(%rsp),%xmm4 movss nb231_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb231_fixO(%rsp) movss %xmm4,nb231_fiyO(%rsp) movss %xmm5,nb231_fizO(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## constant 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## constant 11100110 shufps $230,%xmm5,%xmm5 ## constant 11100110 addss nb231_fixH1(%rsp),%xmm3 addss nb231_fiyH1(%rsp),%xmm4 addss nb231_fizH1(%rsp),%xmm5 movss %xmm3,nb231_fixH1(%rsp) movss %xmm4,nb231_fiyH1(%rsp) movss %xmm5,nb231_fizH1(%rsp) ## updated the H1 force movq nb231_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## constant 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## constant 11100111 shufps $231,%xmm5,%xmm5 ## constant 11100111 addss nb231_fixH2(%rsp),%xmm3 addss nb231_fiyH2(%rsp),%xmm4 addss nb231_fizH2(%rsp),%xmm5 movss %xmm3,nb231_fixH2(%rsp) movss %xmm4,nb231_fiyH2(%rsp) movss %xmm5,nb231_fizH2(%rsp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb231_innerk(%rsp) jz _nb_kernel231_x86_64_sse.nb231_updateouterdata jmp _nb_kernel231_x86_64_sse.nb231_odd_loop_nb_kernel231_x86_64_sse.nb231_updateouterdata: movl nb231_ii3(%rsp),%ecx movq nb231_faction(%rbp),%rdi movq nb231_fshift(%rbp),%rsi movl nb231_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb231_fixO(%rsp),%xmm0 movaps nb231_fiyO(%rsp),%xmm1 movaps nb231_fizO(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss (%rdi,%rcx,4),%xmm3 movss 4(%rdi,%rcx,4),%xmm4 movss 8(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%rdi,%rcx,4) movss %xmm4,4(%rdi,%rcx,4) movss %xmm5,8(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## constant 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb231_fixH1(%rsp),%xmm0 movaps nb231_fiyH1(%rsp),%xmm1 movaps nb231_fizH1(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 12(%rdi,%rcx,4),%xmm3 movss 16(%rdi,%rcx,4),%xmm4 movss 20(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,12(%rdi,%rcx,4) movss %xmm4,16(%rdi,%rcx,4) movss %xmm5,20(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb231_fixH2(%rsp),%xmm0 movaps nb231_fiyH2(%rsp),%xmm1 movaps nb231_fizH2(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 24(%rdi,%rcx,4),%xmm3 movss 28(%rdi,%rcx,4),%xmm4 movss 32(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,24(%rdi,%rcx,4) movss %xmm4,28(%rdi,%rcx,4) movss %xmm5,32(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## increment fshift force movlps (%rsi,%rdx,4),%xmm3 movss 8(%rsi,%rdx,4),%xmm4 subps %xmm6,%xmm3 subss %xmm7,%xmm4 movlps %xmm3,(%rsi,%rdx,4) movss %xmm4,8(%rsi,%rdx,4) ## get n from stack movl nb231_n(%rsp),%esi ## get group index for i particle movq nb231_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb231_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 nb231_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 nb231_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 nb231_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb231_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel231_x86_64_sse.nb231_outerend ## not last, iterate outer loop once more! movl %esi,nb231_n(%rsp) jmp _nb_kernel231_x86_64_sse.nb231_outer_nb_kernel231_x86_64_sse.nb231_outerend: ## check if more outer neighborlists remain movl nb231_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel231_x86_64_sse.nb231_end ## non-zero, do one more workunit jmp _nb_kernel231_x86_64_sse.nb231_threadloop_nb_kernel231_x86_64_sse.nb231_end: movl nb231_nouter(%rsp),%eax movl nb231_ninner(%rsp),%ebx movq nb231_outeriter(%rbp),%rcx movq nb231_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $920,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel231nf_x86_64_sse.globl _nb_kernel231nf_x86_64_ssenb_kernel231nf_x86_64_sse: _nb_kernel231nf_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb231nf_fshift, 16.set nb231nf_gid, 24.set nb231nf_pos, 32.set nb231nf_faction, 40.set nb231nf_charge, 48.set nb231nf_p_facel, 56.set nb231nf_argkrf, 64.set nb231nf_argcrf, 72⌨️ 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?