nb_kernel232_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,208 行 · 第 1/5 页
S
2,208 行
movaps xmm0, xmm3 ;# rinv subps xmm3, xmm7 ;# xmm3=rinv-2*krsq xorps xmm4, xmm4 ;# fetch charges to xmm4 (temporary) movss xmm4, [rsp + nb232_qqOO] movhps xmm4, [rsp + nb232_qqOH] mulps xmm6, xmm4 ;# vcoul mulps xmm3, xmm4 ;# coul part of fs mulps xmm3, xmm0 movaps [rsp + nb232_fstmp], xmm3 ;# save it addps xmm6, [rsp + nb232_vctot] movaps [rsp + nb232_vctot], xmm6 movaps xmm0, [rsp + nb232_rinvOO] movss xmm1, xmm0 mulss xmm1, [rsp + nb232_rsqOO] ;# xmm1=r mulss xmm1, [rsp + nb232_tsc] cvttps2pi mm6, xmm1 cvtpi2ps xmm3, mm6 subss xmm1, xmm3 ;# xmm1=eps movss xmm2, xmm1 mulss xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 3 mov rsi, [rbp + nb232_VFtab] movd r8d, mm6 ;# dispersion movlps xmm5, [rsi + r8*4] movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [rsi + r8*4 + 8] movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# dispersion table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [rsp + nb232_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movss xmm4, [rsp + nb232_c6] mulss xmm7, xmm4 ;# fijD mulss xmm5, xmm4 ;# Vvdw6 movss xmm3, [rsp + nb232_fstmp] mulps xmm7, [rsp + nb232_tsc] subss xmm3, xmm7 movss [rsp + nb232_fstmp], xmm3 addss xmm5, [rsp + nb232_Vvdwtot] movss [rsp + nb232_Vvdwtot], xmm5 ;# repulsion movlps xmm5, [rsi + r8*4 + 16] movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [rsi + r8*4 + 24] movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [rsp + nb232_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movss xmm4, [rsp + nb232_c12] mulss xmm7, xmm4 ;# fijR mulss xmm5, xmm4 ;# Vvdw12 movaps xmm3, [rsp + nb232_fstmp] mulss xmm7, [rsp + nb232_tsc] subss xmm3, xmm7 addss xmm5, [rsp + nb232_Vvdwtot] movss [rsp + nb232_Vvdwtot], xmm5 mulps xmm0, xmm3 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [rsp + nb232_dxOO] mulps xmm1, [rsp + nb232_dyOO] mulps xmm2, [rsp + nb232_dzOO] ;# initial update for j forces xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [rsp + nb232_fjxO], xmm3 movaps [rsp + nb232_fjyO], xmm4 movaps [rsp + nb232_fjzO], xmm5 addps xmm0, [rsp + nb232_fixO] addps xmm1, [rsp + nb232_fiyO] addps xmm2, [rsp + nb232_fizO] movaps [rsp + nb232_fixO], xmm0 movaps [rsp + nb232_fiyO], xmm1 movaps [rsp + nb232_fizO], xmm2 ;# done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [rsp + nb232_jxO] movaps xmm1, [rsp + nb232_jyO] movaps xmm2, [rsp + nb232_jzO] movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 subps xmm0, [rsp + nb232_ixH1] subps xmm1, [rsp + nb232_iyH1] subps xmm2, [rsp + nb232_izH1] subps xmm3, [rsp + nb232_ixH2] subps xmm4, [rsp + nb232_iyH2] subps xmm5, [rsp + nb232_izH2] movaps [rsp + nb232_dxH1O], xmm0 movaps [rsp + nb232_dyH1O], xmm1 movaps [rsp + nb232_dzH1O], xmm2 movaps [rsp + nb232_dxH2O], xmm3 movaps [rsp + nb232_dyH2O], xmm4 movaps [rsp + nb232_dzH2O], xmm5 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm4, xmm3 addps xmm0, xmm2 ;# have rsqH1 in xmm0 addps xmm4, xmm5 ;# have rsqH2 in xmm4 ;# do invsqrt rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb232_three] movaps xmm7, xmm3 mulps xmm1, xmm0 mulps xmm5, xmm4 subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [rsp + nb232_half] ;# rinv H1 - j water mulps xmm7, [rsp + nb232_half] ;# rinv H2 - j water mulps xmm0, [rsp + nb232_krf] ;# krsq mulps xmm4, [rsp + nb232_krf] ;# krsq ;# assemble charges in xmm6 xorps xmm6, xmm6 movss xmm6, [rsp + nb232_qqOH] movhps xmm6, [rsp + nb232_qqHH] movaps xmm1, xmm0 movaps xmm5, xmm4 addps xmm0, xmm3 ;# krsq+ rinv addps xmm4, xmm7 ;# krsq+ rinv subps xmm0, [rsp + nb232_crf] subps xmm4, [rsp + nb232_crf] mulps xmm1, [rsp + nb232_two] mulps xmm5, [rsp + nb232_two] mulps xmm0, xmm6 ;# vcoul mulps xmm4, xmm6 ;# vcoul addps xmm4, xmm0 addps xmm4, [rsp + nb232_vctot] movaps [rsp + nb232_vctot], xmm4 movaps xmm0, xmm3 movaps xmm4, xmm7 mulps xmm3, xmm3 mulps xmm7, xmm7 subps xmm0, xmm1 subps xmm4, xmm5 mulps xmm0, xmm6 mulps xmm4, xmm6 mulps xmm0, xmm3 ;# fscal mulps xmm7, xmm4 ;# fscal movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [rsp + nb232_dxH1O] mulps xmm1, [rsp + nb232_dyH1O] mulps xmm2, [rsp + nb232_dzH1O] ;# update forces H1 - j water movaps xmm3, [rsp + nb232_fjxO] movaps xmm4, [rsp + nb232_fjyO] movaps xmm5, [rsp + nb232_fjzO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [rsp + nb232_fjxO], xmm3 movaps [rsp + nb232_fjyO], xmm4 movaps [rsp + nb232_fjzO], xmm5 addps xmm0, [rsp + nb232_fixH1] addps xmm1, [rsp + nb232_fiyH1] addps xmm2, [rsp + nb232_fizH1] movaps [rsp + nb232_fixH1], xmm0 movaps [rsp + nb232_fiyH1], xmm1 movaps [rsp + nb232_fizH1], xmm2 ;# do forces H2 - j water movaps xmm0, xmm7 movaps xmm1, xmm7 movaps xmm2, xmm7 mulps xmm0, [rsp + nb232_dxH2O] mulps xmm1, [rsp + nb232_dyH2O] mulps xmm2, [rsp + nb232_dzH2O] movaps xmm3, [rsp + nb232_fjxO] movaps xmm4, [rsp + nb232_fjyO] movaps xmm5, [rsp + nb232_fjzO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 mov rsi, [rbp + nb232_faction] movaps [rsp + nb232_fjxO], xmm3 movaps [rsp + nb232_fjyO], xmm4 movaps [rsp + nb232_fjzO], xmm5 addps xmm0, [rsp + nb232_fixH2] addps xmm1, [rsp + nb232_fiyH2] addps xmm2, [rsp + nb232_fizH2] movaps [rsp + nb232_fixH2], xmm0 movaps [rsp + nb232_fiyH2], xmm1 movaps [rsp + nb232_fizH2], xmm2 ;# update j water forces from local variables movlps xmm0, [rsi + rax*4] movlps xmm1, [rsi + rax*4 + 12] movhps xmm1, [rsi + rax*4 + 24] movaps xmm3, [rsp + nb232_fjxO] movaps xmm4, [rsp + nb232_fjyO] movaps xmm5, [rsp + nb232_fjzO] movaps xmm6, xmm5 movaps xmm7, xmm5 shufps xmm6, xmm6, 2 ;# constant 00000010 shufps xmm7, xmm7, 3 ;# constant 00000011 addss xmm5, [rsi + rax*4 + 8] addss xmm6, [rsi + rax*4 + 20] addss xmm7, [rsi + rax*4 + 32] movss [rsi + rax*4 + 8], xmm5 movss [rsi + rax*4 + 20], xmm6 movss [rsi + rax*4 + 32], xmm7 movaps xmm5, xmm3 unpcklps xmm3, xmm4 unpckhps xmm5, xmm4 addps xmm0, xmm3 addps xmm1, xmm5 movlps [rsi + rax*4], xmm0 movlps [rsi + rax*4 + 12], xmm1 movhps [rsi + rax*4 + 24], xmm1 dec dword ptr [rsp + nb232_innerk] jz .nb232_updateouterdata jmp .nb232_single_loop.nb232_updateouterdata: mov ecx, [rsp + nb232_ii3] mov rdi, [rbp + nb232_faction] mov rsi, [rbp + nb232_fshift] mov edx, [rsp + nb232_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb232_fixO] movaps xmm1, [rsp + nb232_fiyO] movaps xmm2, [rsp + nb232_fizO] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4] movss xmm4, [rdi + rcx*4 + 4] movss xmm5, [rdi + rcx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4], xmm3 movss [rdi + rcx*4 + 4], xmm4 movss [rdi + rcx*4 + 8], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movaps xmm6, xmm0 movss xmm7, xmm2 movlhps xmm6, xmm1 shufps xmm6, xmm6, 8 ;# constant 00001000 ;# accumulate H1i forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb232_fixH1] movaps xmm1, [rsp + nb232_fiyH1] movaps xmm2, [rsp + nb232_fizH1] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4 + 12] movss xmm4, [rdi + rcx*4 + 16] movss xmm5, [rdi + rcx*4 + 20] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4 + 12], xmm3 movss [rdi + rcx*4 + 16], xmm4 movss [rdi + rcx*4 + 20], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb232_fixH2] movaps xmm1, [rsp + nb232_fiyH2] movaps xmm2, [rsp + nb232_fizH2] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4 + 24] movss xmm4, [rdi + rcx*4 + 28] movss xmm5, [rdi + rcx*4 + 32] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4 + 24], xmm3 movss [rdi + rcx*4 + 28], xmm4 movss [rdi + rcx*4 + 32], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# increment fshift force movlps xmm3, [rsi + rdx*4] movss xmm4, [rsi + rdx*4 + 8] subps xmm3, xmm6 subss xmm4, xmm7 movlps [rsi + rdx*4], xmm3 movss [rsi + rdx*4 + 8], xmm4 ;# get n from stack mov esi, [rsp + nb232_n] ;# get group index for i particle mov rdx, [rbp + nb232_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [rsp + nb232_vctot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb232_Vc] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# accumulate total lj energy and update it movaps xmm7, [rsp + nb232_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb232_Vvdw] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ⌨️ 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?