nb_kernel134_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,215 行 · 第 1/5 页
S
2,215 行
shufps xmm3, xmm2, 136 ;# (10001000) Oy H1y H2y My shufps xmm1, xmm5, 200 ;# (11001000) Oz H1z H2z Mz ;# store all j coordinates in jO movaps [rsp + nb134_jxO], xmm6 movaps [rsp + nb134_jyO], xmm3 movaps [rsp + nb134_jzO], xmm1 movaps xmm0, xmm6 ;# jxO movaps xmm2, xmm1 ;# jzO movaps xmm1, xmm3 ;# jyO movaps xmm4, xmm3 ;# jyO movaps xmm3, xmm6 ;# jxO movaps xmm5, xmm2 ;# jzO ;# do O and M in parallel subps xmm0, [rsp + nb134_ixO] subps xmm1, [rsp + nb134_iyO] subps xmm2, [rsp + nb134_izO] subps xmm3, [rsp + nb134_ixM] subps xmm4, [rsp + nb134_iyM] subps xmm5, [rsp + nb134_izM] movaps [rsp + nb134_dxOO], xmm0 movaps [rsp + nb134_dyOO], xmm1 movaps [rsp + nb134_dzOO], xmm2 movaps [rsp + nb134_dxMM], xmm3 movaps [rsp + nb134_dyMM], xmm4 movaps [rsp + nb134_dzMM], xmm5 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 ;# have rsq in xmm0 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm4, xmm3 addps xmm4, xmm5 ;# have rsq in xmm4 ;# Save data movaps [rsp + nb134_rsqOO], xmm0 movaps [rsp + nb134_rsqMM], xmm4 ;# do 1/x for O rsqrtps xmm1, xmm0 movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [rsp + nb134_three] mulps xmm1, xmm0 ;# rsq*lu*lu subps xmm3, xmm1 ;# constant 30-rsq*lu*lu mulps xmm3, xmm2 ;# lu*(3-rsq*lu*lu) mulps xmm3, [rsp + nb134_half] movaps [rsp + nb134_rinvOO], xmm3 ;# rinvH2 ;# 1/sqrt(x) for M rsqrtps xmm5, xmm4 movaps xmm6, xmm5 mulps xmm5, xmm5 movaps xmm7, [rsp + nb134_three] mulps xmm5, xmm4 subps xmm7, xmm5 mulps xmm7, xmm6 mulps xmm7, [rsp + nb134_half] ;# rinv iH1 - j water movaps [rsp + nb134_rinvMM], xmm7 ;# LJ table interaction movaps xmm0, [rsp + nb134_rinvOO] movss xmm1, xmm0 mulss xmm1, [rsp + nb134_rsqOO] ;# xmm1=r mulss xmm1, [rsp + nb134_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 + nb134_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 + nb134_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 + nb134_c6] mulss xmm7, xmm4 ;# fijD mulss xmm5, xmm4 ;# Vvdw6 xorps xmm3, xmm3 mulps xmm7, [rsp + nb134_tsc] subss xmm3, xmm7 movss [rsp + nb134_fstmp], xmm3 addss xmm5, [rsp + nb134_Vvdwtot] movss [rsp + nb134_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 + nb134_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 + nb134_c12] mulss xmm7, xmm4 ;# fijR mulss xmm5, xmm4 ;# Vvdw12 movaps xmm3, [rsp + nb134_fstmp] mulss xmm7, [rsp + nb134_tsc] subss xmm3, xmm7 addss xmm5, [rsp + nb134_Vvdwtot] movss [rsp + nb134_Vvdwtot], xmm5 mulss xmm0, xmm3 movaps xmm1, xmm0 movaps xmm2, xmm0 mulss xmm0, [rsp + nb134_dxOO] mulss xmm1, [rsp + nb134_dyOO] mulss xmm2, [rsp + nb134_dzOO] xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movaps [rsp + nb134_fjxO], xmm3 movaps [rsp + nb134_fjyO], xmm4 movaps [rsp + nb134_fjzO], xmm5 addss xmm0, [rsp + nb134_fixO] addss xmm1, [rsp + nb134_fiyO] addss xmm2, [rsp + nb134_fizO] movss [rsp + nb134_fixO], xmm0 movss [rsp + nb134_fiyO], xmm1 movss [rsp + nb134_fizO], xmm2 ;# do M coulomb interaction movaps xmm0, [rsp + nb134_rinvMM] movaps xmm7, xmm0 ;# xmm7=rinv mulps xmm0, xmm0 ;# xmm0=rinvsq ;# fetch charges to xmm3 (temporary) xorps xmm3, xmm3 movss xmm3, [rsp + nb134_qqMH] movhps xmm3, [rsp + nb134_qqMM] shufps xmm3, xmm3, 193 ;# constant 11000001 mulps xmm7, xmm3 ;# vcoul=rinv*qq movaps xmm6, xmm7 addps xmm7, [rsp + nb134_vctot] movaps [rsp + nb134_vctot], xmm7 mulps xmm0, xmm6 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [rsp + nb134_dxMM] mulps xmm1, [rsp + nb134_dyMM] mulps xmm2, [rsp + nb134_dzMM] ;# update forces M - j water movaps xmm3, [rsp + nb134_fjxO] movaps xmm4, [rsp + nb134_fjyO] movaps xmm5, [rsp + nb134_fjzO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [rsp + nb134_fjxO], xmm3 movaps [rsp + nb134_fjyO], xmm4 movaps [rsp + nb134_fjzO], xmm5 addps xmm0, [rsp + nb134_fixM] addps xmm1, [rsp + nb134_fiyM] addps xmm2, [rsp + nb134_fizM] movaps [rsp + nb134_fixM], xmm0 movaps [rsp + nb134_fiyM], xmm1 movaps [rsp + nb134_fizM], xmm2 ;# i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [rsp + nb134_jxO] movaps xmm1, [rsp + nb134_jyO] movaps xmm2, [rsp + nb134_jzO] movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 subps xmm0, [rsp + nb134_ixH1] subps xmm1, [rsp + nb134_iyH1] subps xmm2, [rsp + nb134_izH1] subps xmm3, [rsp + nb134_ixH2] subps xmm4, [rsp + nb134_iyH2] subps xmm5, [rsp + nb134_izH2] movaps [rsp + nb134_dxH1H1], xmm0 movaps [rsp + nb134_dyH1H1], xmm1 movaps [rsp + nb134_dzH1H1], xmm2 movaps [rsp + nb134_dxH2H2], xmm3 movaps [rsp + nb134_dyH2H2], xmm4 movaps [rsp + nb134_dzH2H2], 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 movaps [rsp + nb134_rsqH1H1], xmm0 movaps [rsp + nb134_rsqH2H2], xmm4 ;# start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb134_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 + nb134_half] ;# rinvH1H1 mulps xmm7, [rsp + nb134_half] ;# rinvH2H2 movaps [rsp + nb134_rinvH1H1], xmm3 movaps [rsp + nb134_rinvH2H2], xmm7 ;# Do H1 coulomb interaction movaps xmm0, xmm3 movaps xmm7, xmm0 ;# xmm7=rinv mulps xmm0, xmm0 ;# xmm0=rinvsq ;# fetch charges to xmm3 (temporary) xorps xmm3, xmm3 movss xmm3, [rsp + nb134_qqHH] movhps xmm3, [rsp + nb134_qqMH] shufps xmm3, xmm3, 193 ;# constant 11000001 mulps xmm7, xmm3 ;#vcoul mulps xmm0, xmm7 addps xmm7, [rsp + nb134_vctot] movaps [rsp + nb134_vctot], xmm7 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [rsp + nb134_dxH1H1] mulps xmm1, [rsp + nb134_dyH1H1] mulps xmm2, [rsp + nb134_dzH1H1] ;# update forces H1 - j water movaps xmm3, [rsp + nb134_fjxO] movaps xmm4, [rsp + nb134_fjyO] movaps xmm5, [rsp + nb134_fjzO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [rsp + nb134_fjxO], xmm3 movaps [rsp + nb134_fjyO], xmm4 movaps [rsp + nb134_fjzO], xmm5 addps xmm0, [rsp + nb134_fixH1] addps xmm1, [rsp + nb134_fiyH1] addps xmm2, [rsp + nb134_fizH1] movaps [rsp + nb134_fixH1], xmm0 movaps [rsp + nb134_fiyH1], xmm1 movaps [rsp + nb134_fizH1], xmm2 ;# H2 Coulomb movaps xmm0, [rsp + nb134_rinvH2H2] movaps xmm7, xmm0 ;# xmm7=rinv mulps xmm0, xmm0 ;# xmm0=rinvsq ;# fetch charges to xmm3 (temporary) xorps xmm3, xmm3 movss xmm3, [rsp + nb134_qqHH] movhps xmm3, [rsp + nb134_qqMH] shufps xmm3, xmm3, 193 ;# constant 11000001 mulps xmm7, xmm3 mulps xmm0, xmm7 addps xmm7, [rsp + nb134_vctot] ;# local vctot summation variable movaps [rsp + nb134_vctot], xmm7 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [rsp + nb134_dxH2H2] mulps xmm1, [rsp + nb134_dyH2H2] mulps xmm2, [rsp + nb134_dzH2H2] ;# update forces H2 - j water movaps xmm3, [rsp + nb134_fjxO] movaps xmm4, [rsp + nb134_fjyO] movaps xmm5, [rsp + nb134_fjzO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [rsp + nb134_fjxO], xmm3 movaps [rsp + nb134_fjyO], xmm4 movaps [rsp + nb134_fjzO], xmm5 addps xmm0, [rsp + nb134_fixH2] addps xmm1, [rsp + nb134_fiyH2] addps xmm2, [rsp + nb134_fizH2] movaps [rsp + nb134_fixH2], xmm0 movaps [rsp + nb134_fiyH2], xmm1 movaps [rsp + nb134_fizH2], xmm2 mov rsi, [rbp + nb134_faction] ;# update j water forces from local variables. ;# transpose back first movaps xmm0, [rsp + nb134_fjxO] ;# Ox H1x H2x Mx movaps xmm1, [rsp + nb134_fjyO] ;# Oy H1y H2y My movaps xmm2, [rsp + nb134_fjzO] ;# Oz H1z H2z Mz movaps xmm3, xmm0 movaps xmm4, xmm0 unpcklps xmm3, xmm1 ;# Ox Oy - - shufps xmm4, xmm2, 0x1 ;# h1x - Oz - movaps xmm5, xmm1 movaps xmm6, xmm0 unpcklps xmm5, xmm2 ;# - - H1y H1z unpckhps xmm6, xmm1 ;# h2x h2y - - unpckhps xmm1, xmm2 ;# - - My Mz shufps xmm2, xmm0, 50 ;# (00110010) h2z - Mx - shufps xmm3, xmm4, 36 ;# constant 00100100 ;# Ox Oy Oz H1x shufps xmm5, xmm6, 78 ;# constant 01001110 ;# h1y h1z h2x h2y shufps xmm2, xmm1, 232 ;# constant 11101000 ;# h2z mx my mz movlps xmm0, [rsi + rax*4] movlps xmm1, [rsi + rax*4 + 16] movlps xmm4, [rsi + rax*4 + 32] movhps xmm0, [rsi + rax*4 + 8] movhps xmm1, [rsi + rax*4 + 24] movhps xmm4, [rsi + rax*4 + 40] addps xmm0, xmm3 addps xmm1, xmm5 addps xmm4, xmm2 movlps [rsi + rax*4], xmm0 movlps [rsi + rax*4 + 16], xmm1 movlps [rsi + rax*4 + 32], xmm4 movhps [rsi + rax*4 + 8], xmm0 movhps [rsi + rax*4 + 24], xmm1 movhps [rsi + rax*4 + 40], xmm4 dec dword ptr [rsp + nb134_innerk] jz .nb134_updateouterdata jmp .nb134_single_loop.nb134_updateouterdata: mov ecx, [rsp + nb134_ii3] mov rdi, [rbp + nb134_faction] mov rsi, [rbp + nb134_fshift] mov edx, [rsp + nb134_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb134_fixO] movaps xmm1, [rsp + nb134_fiyO] movaps xmm2, [rsp + nb134_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 + nb134_fixH1] movaps xmm1, [rsp + nb134_fiyH1] movaps xmm2, [rsp + nb134_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 ⌨️ 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?