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
;# Step2 - transpose fjyH1, fjzH1 and fjxH2, fjyH2 movapd xmm0, [esp + nb334_fjyH1] movapd xmm1, [esp + nb334_fjzH1] movapd xmm2, [esp + nb334_fjxH2] movapd xmm3, [esp + nb334_fjyH2] movapd xmm4, xmm0 movapd xmm5, xmm2 unpcklpd xmm0, xmm1 ;# fjOxA fjOyA unpckhpd xmm4, xmm1 ;# fjOxB fjOyB unpcklpd xmm2, xmm3 ;# fjOzA fjH1xA unpckhpd xmm5, xmm3 ;# fjOzB fjH1xB movlpd xmm1, [edi + eax*8 + 32] movlpd xmm3, [edi + ebx*8 + 32] movhpd xmm1, [edi + eax*8 + 40] movhpd xmm3, [edi + ebx*8 + 40] movlpd xmm6, [edi + eax*8 + 48] movlpd xmm7, [edi + ebx*8 + 48] movhpd xmm6, [edi + eax*8 + 56] movhpd xmm7, [edi + ebx*8 + 56] addpd xmm1, xmm0 addpd xmm3, xmm4 addpd xmm6, xmm2 addpd xmm7, xmm5 movlpd [edi + eax*8 + 32], xmm1 movlpd [edi + ebx*8 + 32], xmm3 movhpd [edi + eax*8 + 40], xmm1 movhpd [edi + ebx*8 + 40], xmm3 movlpd [edi + eax*8 + 48], xmm6 movlpd [edi + ebx*8 + 48], xmm7 movhpd [edi + eax*8 + 56], xmm6 movhpd [edi + ebx*8 + 56], xmm7 ;# Step3 - transpose fjzH2, fjxM and fjyM, fjzM movapd xmm0, [esp + nb334_fjzH2] movapd xmm1, [esp + nb334_fjxM] movapd xmm2, [esp + nb334_fjyM] movapd xmm3, [esp + nb334_fjzM] movapd xmm4, xmm0 movapd xmm5, xmm2 unpcklpd xmm0, xmm1 ;# fjOxA fjOyA unpckhpd xmm4, xmm1 ;# fjOxB fjOyB unpcklpd xmm2, xmm3 ;# fjOzA fjH1xA unpckhpd xmm5, xmm3 ;# fjOzB fjH1xB movlpd xmm1, [edi + eax*8 + 64] movlpd xmm3, [edi + ebx*8 + 64] movhpd xmm1, [edi + eax*8 + 72] movhpd xmm3, [edi + ebx*8 + 72] movlpd xmm6, [edi + eax*8 + 80] movlpd xmm7, [edi + ebx*8 + 80] movhpd xmm6, [edi + eax*8 + 88] movhpd xmm7, [edi + ebx*8 + 88] addpd xmm1, xmm0 addpd xmm3, xmm4 addpd xmm6, xmm2 addpd xmm7, xmm5 movlpd [edi + eax*8 + 64], xmm1 movlpd [edi + ebx*8 + 64], xmm3 movhpd [edi + eax*8 + 72], xmm1 movhpd [edi + ebx*8 + 72], xmm3 movlpd [edi + eax*8 + 80], xmm6 movlpd [edi + ebx*8 + 80], xmm7 movhpd [edi + eax*8 + 88], xmm6 movhpd [edi + ebx*8 + 88], xmm7 ;# should we do one more iteration? sub dword ptr [esp + nb334_innerk], 2 jl .nb334_checksingle jmp .nb334_unroll_loop.nb334_checksingle: mov edx, [esp + nb334_innerk] and edx, 1 jnz .nb334_dosingle jmp .nb334_updateouterdata.nb334_dosingle: mov edx, [esp + nb334_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov esi, [ebp + nb334_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 ;# move j coordinates to local temp variables ;# load ox, oy, oz, h1x movlpd xmm0, [esi + eax*8] movhpd xmm0, [esi + eax*8 + 8] movlpd xmm1, [esi + eax*8 + 16] movhpd xmm1, [esi + eax*8 + 24] movlpd xmm2, [esi + eax*8 + 32] movhpd xmm2, [esi + eax*8 + 40] movlpd xmm3, [esi + eax*8 + 48] movhpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movhpd xmm4, [esi + eax*8 + 72] movlpd xmm5, [esi + eax*8 + 80] movhpd xmm5, [esi + eax*8 + 88] movsd [esp + nb334_jxO], xmm0 movsd [esp + nb334_jzO], xmm1 movsd [esp + nb334_jyH1], xmm2 movsd [esp + nb334_jxH2], xmm3 movsd [esp + nb334_jzH2], xmm4 movsd [esp + nb334_jyM], xmm5 unpckhpd xmm0, xmm0 unpckhpd xmm1, xmm1 unpckhpd xmm2, xmm2 unpckhpd xmm3, xmm3 unpckhpd xmm4, xmm4 unpckhpd xmm5, xmm5 movsd [esp + nb334_jyO], xmm0 movsd [esp + nb334_jxH1], xmm1 movsd [esp + nb334_jzH1], xmm2 movsd [esp + nb334_jyH2], xmm3 movsd [esp + nb334_jxM], xmm4 movsd [esp + nb334_jzM], xmm5 ;# 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] subsd xmm0, [esp + nb334_jxO] subsd xmm1, [esp + nb334_jyO] subsd xmm2, [esp + nb334_jzO] subsd xmm3, [esp + nb334_jxH1] subsd xmm4, [esp + nb334_jyH1] subsd xmm5, [esp + nb334_jzH1] movapd [esp + nb334_dxOO], xmm0 movapd [esp + nb334_dyOO], xmm1 movapd [esp + nb334_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb334_dxH1H1], xmm3 movapd [esp + nb334_dyH1H1], xmm4 movapd [esp + nb334_dzH1H1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd 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 subsd xmm0, [esp + nb334_jxH2] subsd xmm1, [esp + nb334_jyH2] subsd xmm2, [esp + nb334_jzH2] subsd xmm3, [esp + nb334_jxM] subsd xmm4, [esp + nb334_jyM] subsd xmm5, [esp + nb334_jzM] movapd [esp + nb334_dxH1H2], xmm0 movapd [esp + nb334_dyH1H2], xmm1 movapd [esp + nb334_dzH1H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb334_dxH1M], xmm3 movapd [esp + nb334_dyH1M], xmm4 movapd [esp + nb334_dzH1M], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd 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 subsd xmm0, [esp + nb334_jxH1] subsd xmm1, [esp + nb334_jyH1] subsd xmm2, [esp + nb334_jzH1] subsd xmm3, [esp + nb334_jxH2] subsd xmm4, [esp + nb334_jyH2] subsd xmm5, [esp + nb334_jzH2] movapd [esp + nb334_dxH2H1], xmm0 movapd [esp + nb334_dyH2H1], xmm1 movapd [esp + nb334_dzH2H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb334_dxH2H2], xmm3 movapd [esp + nb334_dyH2H2], xmm4 movapd [esp + nb334_dzH2H2], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd 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] subsd xmm0, [esp + nb334_jxM] subsd xmm1, [esp + nb334_jyM] subsd xmm2, [esp + nb334_jzM] subsd xmm3, [esp + nb334_jxH1] subsd xmm4, [esp + nb334_jyH1] subsd xmm5, [esp + nb334_jzH1] movapd [esp + nb334_dxH2M], xmm0 movapd [esp + nb334_dyH2M], xmm1 movapd [esp + nb334_dzH2M], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb334_dxMH1], xmm3 movapd [esp + nb334_dyMH1], xmm4 movapd [esp + nb334_dzMH1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm4, xmm3 addsd 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 subsd xmm0, [esp + nb334_jxH2] subsd xmm1, [esp + nb334_jyH2] subsd xmm2, [esp + nb334_jzH2] subsd xmm3, [esp + nb334_jxM] subsd xmm4, [esp + nb334_jyM] subsd xmm5, [esp + nb334_jzM] movapd [esp + nb334_dxMH2], xmm0 movapd [esp + nb334_dyMH2], xmm1 movapd [esp + nb334_dzMH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb334_dxMM], xmm3 movapd [esp + nb334_dyMM], xmm4 movapd [esp + nb334_dzMM], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm4, xmm3 addsd 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) cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 ;# luA cvtss2sd xmm5, xmm5 ;# luB movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulsd xmm1, xmm1 ;# luA*luA mulsd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA mulsd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subsd xmm3, xmm1 ;# 3-rsqA*luA*luA subsd xmm7, xmm5 ;# 3-rsqB*luB*luB mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm3, [esp + nb334_half] ;# iter1 mulsd xmm7, [esp + nb334_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulsd xmm3, xmm3 ;# luA*luA mulsd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA mulsd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subsd xmm1, xmm3 ;# 3-rsqA*luA*luA subsd xmm5, xmm7 ;# 3-rsqB*luB*luB mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm1, [esp + nb334_half] ;# rinv mulsd 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] cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 cvtss2sd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulsd xmm1, xmm1 ;# luA*luA mulsd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA mulsd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subsd xmm3, xmm1 ;# 3-rsqA*luA*luA subsd xmm7, xmm5 ;# 3-rsqB*luB*luB mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm3, [esp + nb334_half] ;# iter1 of mulsd xmm7, [esp + nb334_half] ;# iter1 of movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulsd xmm3, xmm3 ;# luA*luA mulsd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA mulsd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subsd xmm1, xmm3 ;# 3-rsqA*luA*luA subsd xmm5, xmm7 ;# 3-rsqB*luB*luB mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm1, [esp + nb334_half] ;# rinv mulsd 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] cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 cvtss2sd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulsd xmm1, xmm1 ;# luA*luA mulsd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA mulsd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subsd xmm3, xmm1 ;# 3-rsqA*luA*luA subsd xmm7, xmm5 ;# 3-rsqB*luB*luB mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm3, [esp + nb334_half] ;# iter1 mulsd xmm7, [esp + nb334_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulsd xmm3, xmm3 ;# luA*luA mulsd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA mulsd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subsd xmm1, xmm3 ;# 3-rsqA*luA*luA subsd xmm5, xmm7 ;# 3-rsqB*luB*luB mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm1, [esp + nb334_half] ;# rinv mulsd 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] cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 cvtss2sd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulsd xmm1, xmm1 ;# luA*luA mulsd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb334_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA mulsd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subsd xmm3, xmm1 ;# 3-rsqA*luA*luA subsd xmm7, xmm5 ;# 3-rsqB*luB*luB mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm3, [esp + nb334_half] ;# iter1a mulsd xmm7, [esp + nb334_half] ;# iter1b movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulsd xmm3, xmm3 ;# luA*luA mulsd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb334_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA mulsd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subsd xmm1, xmm3 ;# 3-rsqA*luA*luA subsd xmm5, xmm7 ;# 3-rsqB*luB*luB mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm1, [esp + nb334_half] ;# rinv mulsd xmm5, [esp + nb334_half] ;# rinv movapd [esp + n
⌨️ 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?