nb_kernel334_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,176 行 · 第 1/5 页
S
2,176 行
mov esi, [ebp + nb334_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 lea ebx, [ebx + ebx*2] ;# move j coordinates to local temp variables ;# load ox, oy, oz, h1x movlpd xmm0, [esi + eax*8] movlpd xmm2, [esi + ebx*8] movhpd xmm0, [esi + eax*8 + 8] movhpd xmm2, [esi + ebx*8 + 8] movlpd xmm3, [esi + eax*8 + 16] movlpd xmm5, [esi + ebx*8 + 16] movhpd xmm3, [esi + eax*8 + 24] movhpd xmm5, [esi + ebx*8 + 24] movapd xmm1, xmm0 movapd xmm4, xmm3 unpcklpd xmm0, xmm2 ;# ox unpckhpd xmm1, xmm2 ;# oy unpcklpd xmm3, xmm5 ;# ox unpckhpd xmm4, xmm5 ;# oy movapd [esp + nb334_jxO], xmm0 movapd [esp + nb334_jyO], xmm1 movapd [esp + nb334_jzO], xmm3 movapd [esp + nb334_jxH1], xmm4 ;# load h1y, h1z, h2x, h2y movlpd xmm0, [esi + eax*8 + 32] movlpd xmm2, [esi + ebx*8 + 32] movhpd xmm0, [esi + eax*8 + 40] movhpd xmm2, [esi + ebx*8 + 40] movlpd xmm3, [esi + eax*8 + 48] movlpd xmm5, [esi + ebx*8 + 48] movhpd xmm3, [esi + eax*8 + 56] movhpd xmm5, [esi + ebx*8 + 56] movapd xmm1, xmm0 movapd xmm4, xmm3 unpcklpd xmm0, xmm2 ;# h1y unpckhpd xmm1, xmm2 ;# h1z unpcklpd xmm3, xmm5 ;# h2x unpckhpd xmm4, xmm5 ;# h2y movapd [esp + nb334_jyH1], xmm0 movapd [esp + nb334_jzH1], xmm1 movapd [esp + nb334_jxH2], xmm3 movapd [esp + nb334_jyH2], xmm4 ;# load h2z, mx, my, mz movlpd xmm0, [esi + eax*8 + 64] movlpd xmm2, [esi + ebx*8 + 64] movhpd xmm0, [esi + eax*8 + 72] movhpd xmm2, [esi + ebx*8 + 72] movlpd xmm3, [esi + eax*8 + 80] movlpd xmm5, [esi + ebx*8 + 80] movhpd xmm3, [esi + eax*8 + 88] movhpd xmm5, [esi + ebx*8 + 88] movapd xmm1, xmm0 movapd xmm4, xmm3 unpcklpd xmm0, xmm2 ;# h2z unpckhpd xmm1, xmm2 ;# mx unpcklpd xmm3, xmm5 ;# my unpckhpd xmm4, xmm5 ;# mz movapd [esp + nb334_jzH2], xmm0 movapd [esp + nb334_jxM], xmm1 movapd [esp + nb334_jyM], xmm3 movapd [esp + nb334_jzM], xmm4 ;# start calculating pairwise distances movapd xmm0, [esp + nb334_ixO] movapd xmm1, [esp + nb334_iyO] movapd xmm2, [esp + nb334_izO] movapd xmm3, [esp + nb334_ixH1] movapd xmm4, [esp + nb334_iyH1] movapd xmm5, [esp + nb334_izH1] subpd xmm0, [esp + nb334_jxO] subpd xmm1, [esp + nb334_jyO] subpd xmm2, [esp + nb334_jzO] subpd xmm3, [esp + nb334_jxH1] subpd xmm4, [esp + nb334_jyH1] subpd xmm5, [esp + nb334_jzH1] movapd [esp + nb334_dxOO], xmm0 movapd [esp + nb334_dyOO], xmm1 movapd [esp + nb334_dzOO], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb334_dxH1H1], xmm3 movapd [esp + nb334_dyH1H1], xmm4 movapd [esp + nb334_dzH1H1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb334_rsqOO], xmm0 movapd [esp + nb334_rsqH1H1], xmm3 movapd xmm0, [esp + nb334_ixH1] movapd xmm1, [esp + nb334_iyH1] movapd xmm2, [esp + nb334_izH1] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subpd xmm0, [esp + nb334_jxH2] subpd xmm1, [esp + nb334_jyH2] subpd xmm2, [esp + nb334_jzH2] subpd xmm3, [esp + nb334_jxM] subpd xmm4, [esp + nb334_jyM] subpd xmm5, [esp + nb334_jzM] movapd [esp + nb334_dxH1H2], xmm0 movapd [esp + nb334_dyH1H2], xmm1 movapd [esp + nb334_dzH1H2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb334_dxH1M], xmm3 movapd [esp + nb334_dyH1M], xmm4 movapd [esp + nb334_dzH1M], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb334_rsqH1H2], xmm0 movapd [esp + nb334_rsqH1M], xmm3 movapd xmm0, [esp + nb334_ixH2] movapd xmm1, [esp + nb334_iyH2] movapd xmm2, [esp + nb334_izH2] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subpd xmm0, [esp + nb334_jxH1] subpd xmm1, [esp + nb334_jyH1] subpd xmm2, [esp + nb334_jzH1] subpd xmm3, [esp + nb334_jxH2] subpd xmm4, [esp + nb334_jyH2] subpd xmm5, [esp + nb334_jzH2] movapd [esp + nb334_dxH2H1], xmm0 movapd [esp + nb334_dyH2H1], xmm1 movapd [esp + nb334_dzH2H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb334_dxH2H2], xmm3 movapd [esp + nb334_dyH2H2], xmm4 movapd [esp + nb334_dzH2H2], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb334_rsqH2H1], xmm0 movapd [esp + nb334_rsqH2H2], xmm3 movapd xmm0, [esp + nb334_ixH2] movapd xmm1, [esp + nb334_iyH2] movapd xmm2, [esp + nb334_izH2] movapd xmm3, [esp + nb334_ixM] movapd xmm4, [esp + nb334_iyM] movapd xmm5, [esp + nb334_izM] subpd xmm0, [esp + nb334_jxM] subpd xmm1, [esp + nb334_jyM] subpd xmm2, [esp + nb334_jzM] subpd xmm3, [esp + nb334_jxH1] subpd xmm4, [esp + nb334_jyH1] subpd xmm5, [esp + nb334_jzH1] movapd [esp + nb334_dxH2M], xmm0 movapd [esp + nb334_dyH2M], xmm1 movapd [esp + nb334_dzH2M], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb334_dxMH1], xmm3 movapd [esp + nb334_dyMH1], xmm4 movapd [esp + nb334_dzMH1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm4, xmm3 addpd xmm4, xmm5 movapd [esp + nb334_rsqH2M], xmm0 movapd [esp + nb334_rsqMH1], xmm4 movapd xmm0, [esp + nb334_ixM] movapd xmm1, [esp + nb334_iyM] movapd xmm2, [esp + nb334_izM] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subpd xmm0, [esp + nb334_jxH2] subpd xmm1, [esp + nb334_jyH2] subpd xmm2, [esp + nb334_jzH2] subpd xmm3, [esp + nb334_jxM] subpd xmm4, [esp + nb334_jyM] subpd xmm5, [esp + nb334_jzM] movapd [esp + nb334_dxMH2], xmm0 movapd [esp + nb334_dyMH2], xmm1 movapd [esp + nb334_dzMH2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb334_dxMM], xmm3 movapd [esp + nb334_dyMM], xmm4 movapd [esp + nb334_dzMM], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm4, xmm3 addpd xmm4, xmm5 movapd [esp + nb334_rsqMH2], xmm0 movapd [esp + nb334_rsqMM], xmm4 ;# Invsqrt form rsq M-H2 (rsq in xmm0) and MM (rsq in xmm4) cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 ;# luA cvtps2pd xmm5, xmm5 ;# luB movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb334_half] ;# iter1 mulpd xmm7, [esp + nb334_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb334_half] ;# rinv mulpd xmm5, [esp + nb334_half] ;# rinv movapd [esp + nb334_rinvMH2], xmm1 movapd [esp + nb334_rinvMM], xmm5 movapd xmm0, [esp + nb334_rsqOO] movapd xmm4, [esp + nb334_rsqH1H1] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb334_half] ;# iter1 of mulpd xmm7, [esp + nb334_half] ;# iter1 of movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb334_half] ;# rinv mulpd xmm5, [esp + nb334_half] ;# rinv movapd [esp + nb334_rinvOO], xmm1 movapd [esp + nb334_rinvH1H1], xmm5 movapd xmm0, [esp + nb334_rsqH1H2] movapd xmm4, [esp + nb334_rsqH1M] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb334_half] ;# iter1 mulpd xmm7, [esp + nb334_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb334_half] ;# rinv mulpd xmm5, [esp + nb334_half] ;# rinv movapd [esp + nb334_rinvH1H2], xmm1 movapd [esp + nb334_rinvH1M], xmm5 movapd xmm0, [esp + nb334_rsqH2H1] movapd xmm4, [esp + nb334_rsqH2H2] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb334_half] ;# iter1a mulpd xmm7, [esp + nb334_half] ;# iter1b movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb334_half] ;# rinv mulpd xmm5, [esp + nb334_half] ;# rinv movapd [esp + nb334_rinvH2H1], xmm1 movapd [esp + nb334_rinvH2H2], xmm5 movapd xmm0, [esp + nb334_rsqMH1] movapd xmm4, [esp + nb334_rsqH2M] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb334_half] ;# iter1a mulpd xmm7, [esp + nb334_half] ;# iter1b movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb334_half] ;# rinv mulpd xmm5, [esp + nb334_half] ;# rinv movapd [esp + nb334_rinvMH1], xmm1 movapd [esp + nb334_rinvH2M], xmm5 ;# start with OO interaction movapd xmm0, [esp + nb334_rinvOO] movapd xmm1, xmm0 mulpd xmm1, [esp + nb334_rsqOO] ;# xmm1=r mulpd xmm1, [esp + nb334_tsc] cvttpd2pi mm6, xmm1 ;# mm6 = lu idx cvtpi2pd xmm6, mm6 subpd xmm1, xmm6 ;# xmm1=eps
⌨️ 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?