nb_kernel102_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,198 行 · 第 1/5 页
S
2,198 行
movaps [esp + nb102_fjyO], xmm4 movaps [esp + nb102_fjzO], xmm5 movaps [esp + nb102_fixH2], xmm0 movaps [esp + nb102_fiyH2], xmm1 movaps [esp + nb102_fizH2], xmm2 ;# H2-H1 interaction movaps xmm0, [esp + nb102_rinvH2H1] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb102_qqHH] mulps xmm0, xmm1 ;# fsH2H1 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb102_fjxH1] movaps xmm4, [esp + nb102_fjyH1] movaps xmm5, [esp + nb102_fjzH1] mulps xmm0, [esp + nb102_dxH2H1] mulps xmm1, [esp + nb102_dyH2H1] mulps xmm2, [esp + nb102_dzH2H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb102_fixH2] addps xmm1, [esp + nb102_fiyH2] addps xmm2, [esp + nb102_fizH2] movaps [esp + nb102_fjxH1], xmm3 movaps [esp + nb102_fjyH1], xmm4 movaps [esp + nb102_fjzH1], xmm5 movaps [esp + nb102_fixH2], xmm0 movaps [esp + nb102_fiyH2], xmm1 movaps [esp + nb102_fizH2], xmm2 ;# H2-H2 interaction movaps xmm0, [esp + nb102_rinvH2H2] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb102_qqHH] mulps xmm0, xmm1 ;# fsH2H2 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps [esp + nb102_vctot], xmm7 movaps xmm2, xmm0 movaps xmm3, [esp + nb102_fjxH2] movaps xmm4, [esp + nb102_fjyH2] movaps xmm5, [esp + nb102_fjzH2] mulps xmm0, [esp + nb102_dxH2H2] mulps xmm1, [esp + nb102_dyH2H2] mulps xmm2, [esp + nb102_dzH2H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb102_fixH2] addps xmm1, [esp + nb102_fiyH2] addps xmm2, [esp + nb102_fizH2] movaps [esp + nb102_fjxH2], xmm3 movaps [esp + nb102_fjyH2], xmm4 movaps [esp + nb102_fjzH2], xmm5 movaps [esp + nb102_fixH2], xmm0 movaps [esp + nb102_fiyH2], xmm1 movaps [esp + nb102_fizH2], xmm2 mov edi, [ebp + nb102_faction] ;# Did all interactions - now update j forces ;# At this stage forces are still on the stack, in positions: ;# fjxO, fjyO, fjzO, ... , fjzH2. ;# Each position is a quadruplet of forces for the four ;# corresponding j waters, so we need to transpose them before ;# adding to the memory positions. ;# ;# This _used_ to be a simple transpose, but the resulting high number ;# of unaligned 128-bit load/stores might trigger a possible hardware ;# bug on Athlon and Opteron chips, so I have worked around it ;# to use 64-bit load/stores instead. The performance hit should be ;# very modest, since the 128-bit unaligned memory instructions were ;# slow anyway. ;# 4 j waters with three atoms each - first do Oxygen X & Y forces for 4 j particles movaps xmm0, [esp + nb102_fjxO] ;# xmm0= fjxOa fjxOb fjxOc fjxOd movaps xmm2, [esp + nb102_fjyO] ;# xmm1= fjyOa fjyOb fjyOc fjyOd movlps xmm3, [edi + eax*4] movlps xmm4, [edi + ecx*4] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjxOa fjyOa fjxOb fjyOb unpckhps xmm1, xmm2 ;# xmm1= fjxOc fjyOc fjxOd fjyOd movhps xmm3, [edi + ebx*4] movhps xmm4, [edi + edx*4] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4], xmm3 movlps [edi + ecx*4], xmm4 movhps [edi + ebx*4], xmm3 movhps [edi + edx*4], xmm4 ;# Oxygen Z & first hydrogen X forces for 4 j particles movaps xmm0, [esp + nb102_fjzO] ;# xmm0= fjzOa fjzOb fjzOc fjzOd movaps xmm2, [esp + nb102_fjxH1] ;# xmm1= fjxH1a fjxH1b fjxH1c fjxH1d movlps xmm3, [edi + eax*4 + 8] movlps xmm4, [edi + ecx*4 + 8] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjzOa fjxH1a fjzOb fjxH1b unpckhps xmm1, xmm2 ;# xmm1= fjzOc fjxH1c fjzOd fjxH1d movhps xmm3, [edi + ebx*4 + 8] movhps xmm4, [edi + edx*4 + 8] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 8], xmm3 movlps [edi + ecx*4 + 8], xmm4 movhps [edi + ebx*4 + 8], xmm3 movhps [edi + edx*4 + 8], xmm4 ;# First hydrogen Y & Z forces for 4 j particles movaps xmm0, [esp + nb102_fjyH1] ;# xmm0= fjyH1a fjyH1b fjyH1c fjyH1d movaps xmm2, [esp + nb102_fjzH1] ;# xmm1= fjzH1a fjzH1b fjzH1c fjzH1d movlps xmm3, [edi + eax*4 + 16] movlps xmm4, [edi + ecx*4 + 16] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjyH1a fjzH1a fjyH1b fjzH1b unpckhps xmm1, xmm2 ;# xmm1= fjyH1c fjzH1c fjyH1d fjzH1d movhps xmm3, [edi + ebx*4 + 16] movhps xmm4, [edi + edx*4 + 16] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 16], xmm3 movlps [edi + ecx*4 + 16], xmm4 movhps [edi + ebx*4 + 16], xmm3 movhps [edi + edx*4 + 16], xmm4 ;# Second hydrogen X & Y forces for 4 j particles movaps xmm0, [esp + nb102_fjxH2] ;# xmm0= fjxH2a fjxH2b fjxH2c fjxH2d movaps xmm2, [esp + nb102_fjyH2] ;# xmm1= fjyH2a fjyH2b fjyH2c fjyH2d movlps xmm3, [edi + eax*4 + 24] movlps xmm4, [edi + ecx*4 + 24] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjxH2a fjyH2a fjxH2b fjyH2b unpckhps xmm1, xmm2 ;# xmm1= fjxH2c fjyH2c fjxH2d fjyH2d movhps xmm3, [edi + ebx*4 + 24] movhps xmm4, [edi + edx*4 + 24] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 24], xmm3 movlps [edi + ecx*4 + 24], xmm4 movhps [edi + ebx*4 + 24], xmm3 movhps [edi + edx*4 + 24], xmm4 ;# Second hydrogen Z forces for 4 j particles ;# Just load the four Z coords into one reg. each movss xmm4, [edi + eax*4 + 32] movss xmm5, [edi + ebx*4 + 32] movss xmm6, [edi + ecx*4 + 32] movss xmm7, [edi + edx*4 + 32] ;# add what we have on the stack addss xmm4, [esp + nb102_fjzH2] addss xmm5, [esp + nb102_fjzH2b] addss xmm6, [esp + nb102_fjzH2c] addss xmm7, [esp + nb102_fjzH2d] ;# store back movss [edi + eax*4 + 32], xmm4 movss [edi + ebx*4 + 32], xmm5 movss [edi + ecx*4 + 32], xmm6 movss [edi + edx*4 + 32], xmm7 ;# should we do one more iteration? sub dword ptr [esp + nb102_innerk], 4 jl .nb102_single_check jmp .nb102_unroll_loop.nb102_single_check: add dword ptr [esp + nb102_innerk], 4 jnz .nb102_single_loop jmp .nb102_updateouterdata.nb102_single_loop: mov edx, [esp + nb102_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb102_innerjjnr], 4 mov esi, [ebp + nb102_pos] lea eax, [eax + eax*2] ;# fetch j coordinates xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 movss xmm3, [esi + eax*4] ;# jxO - - - movss xmm4, [esi + eax*4 + 4] ;# jyO - - - movss xmm5, [esi + eax*4 + 8] ;# jzO - - - movlps xmm6, [esi + eax*4 + 12] ;# xmm6 = jxH1 jyH1 - - movss xmm7, [esi + eax*4 + 20] ;# xmm7 = jzH1 - - - movhps xmm6, [esi + eax*4 + 24] ;# xmm6 = jxH1 jyH1 jxH2 jyH2 movss xmm2, [esi + eax*4 + 32] ;# xmm2 = jzH2 - - - ;# have all coords, time for some shuffling. shufps xmm6, xmm6, 216 ;# constant 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps xmm7, xmm2 ;# xmm7 = jzH1 jzH2 - - movaps xmm0, [esp + nb102_ixO] movaps xmm1, [esp + nb102_iyO] movaps xmm2, [esp + nb102_izO] movlhps xmm3, xmm6 ;# xmm3 = jxO 0 jxH1 jxH2 shufps xmm4, xmm6, 228 ;# constant 11100100 ;# xmm4 = jyO 0 jyH1 jyH2 shufps xmm5, xmm7, 68 ;# constant 01000100 ;# xmm5 = jzO 0 jzH1 jzH2 ;# store all j coordinates in jO movaps [esp + nb102_jxO], xmm3 movaps [esp + nb102_jyO], xmm4 movaps [esp + nb102_jzO], xmm5 subps xmm0, xmm3 subps xmm1, xmm4 subps xmm2, xmm5 movaps [esp + nb102_dxOO], xmm0 movaps [esp + nb102_dyOO], xmm1 movaps [esp + nb102_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 ;# have rsq in xmm0 ;# do invsqrt rsqrtps xmm1, xmm0 movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb102_three] mulps xmm1, xmm0 subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb102_half] ;# rinv iO - j water xorps xmm1, xmm1 movaps xmm0, xmm3 xorps xmm4, xmm4 mulps xmm0, xmm0 ;# xmm0=rinvsq ;# fetch charges to xmm4 (temporary) movss xmm4, [esp + nb102_qqOO] movhps xmm4, [esp + nb102_qqOH] mulps xmm3, xmm4 ;# xmm3=vcoul mulps xmm0, xmm3 ;# total fscal addps xmm3, [esp + nb102_vctot] movaps [esp + nb102_vctot], xmm3 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [esp + nb102_dxOO] mulps xmm1, [esp + nb102_dyOO] mulps xmm2, [esp + nb102_dzOO] ;# initial update for j forces xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb102_fjxO], xmm3 movaps [esp + nb102_fjyO], xmm4 movaps [esp + nb102_fjzO], xmm5 addps xmm0, [esp + nb102_fixO] addps xmm1, [esp + nb102_fiyO] addps xmm2, [esp + nb102_fizO] movaps [esp + nb102_fixO], xmm0 movaps [esp + nb102_fiyO], xmm1 movaps [esp + nb102_fizO], xmm2 ;# done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [esp + nb102_ixH1] movaps xmm1, [esp + nb102_iyH1] movaps xmm2, [esp + nb102_izH1] movaps xmm3, [esp + nb102_ixH2] movaps xmm4, [esp + nb102_iyH2] movaps xmm5, [esp + nb102_izH2] subps xmm0, [esp + nb102_jxO] subps xmm1, [esp + nb102_jyO] subps xmm2, [esp + nb102_jzO] subps xmm3, [esp + nb102_jxO] subps xmm4, [esp + nb102_jyO] subps xmm5, [esp + nb102_jzO] movaps [esp + nb102_dxH1O], xmm0 movaps [esp + nb102_dyH1O], xmm1 movaps [esp + nb102_dzH1O], xmm2 movaps [esp + nb102_dxH2O], xmm3 movaps [esp + nb102_dyH2O], xmm4 movaps [esp + nb102_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, [esp + nb102_three] movaps xmm7, xmm3 mulps xmm1, xmm0 mulps xmm5, xmm4 subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb102_half] ;# rinv H1 - j water mulps xmm7, [esp + nb102_half] ;# rinv H2 - j water ;# assemble charges in xmm6 xorps xmm6, xmm6 ;# do coulomb interaction movaps xmm0, xmm3 movss xmm6, [esp + nb102_qqOH] movaps xmm4, xmm7 movhps xmm6, [esp + nb102_qqHH] mulps xmm0, xmm0 ;# rinvsq mulps xmm4, xmm4 ;# rinvsq mulps xmm3, xmm6 ;# vcoul mulps xmm7, xmm6 ;# vcoul movaps xmm2, xmm3 addps xmm2, xmm7 ;# total vcoul mulps xmm0, xmm3 ;# fscal addps xmm2, [esp + nb102_vctot] mulps xmm7, xmm4 ;# fscal movaps [esp + nb102_vctot], xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [esp + nb102_dxH1O] mulps xmm1, [esp + nb102_dyH1O] mulps xmm2, [esp + nb102_dzH1O] ;# update forces H1 - j water movaps xmm3, [esp + nb102_fjxO] movaps xmm4, [esp + nb102_fjyO] movaps xmm5, [esp + nb102_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb102_fjxO], xmm3 movaps [esp + nb102_fjyO], xmm4 movaps [esp + nb102_fjzO], xmm5 addps xmm0, [esp + nb102_fixH1] addps xmm1, [esp + nb102_fiyH1] addps xmm2, [esp + nb102_fizH1] movaps [esp + nb102_fixH1], xmm0 movaps [esp + nb102_fiyH1], xmm1 movaps [esp + nb102_fizH1], xmm2 ;# do forces H2 - j water movaps xmm0, xmm7 movaps xmm1, xmm7 movaps xmm2, xmm7 mulps xmm0, [esp + nb102_dxH2O] mulps xmm1, [esp + nb102_dyH2O] mulps xmm2, [esp + nb102_dzH2O] movaps xmm3, [esp + nb102_fjxO] movaps xmm4, [esp + nb102_fjyO] movaps xmm5, [esp + nb102_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 mov esi, [ebp + nb102_faction] movaps [esp + nb102_fjxO], xmm3 movaps [esp + nb102_fjyO], xmm4 movaps [esp + nb102_fjzO], xmm5 addps xmm0, [esp + nb102_fixH2] addps xmm1, [esp + nb102_fiyH2] addps xmm2, [esp + nb102_fizH2] movaps [esp + nb102_fixH2], xmm0 movaps [esp + nb102_fiyH2], xmm1 movaps [esp + nb102_fizH2], xmm2 ;# update j water forces from local variables movlps xmm0, [esi + eax*4] movlps xmm1, [esi + eax*4 + 12] movhps xmm1, [esi + eax*4 + 24] movaps xmm3, [esp + nb102_fjxO] movaps xmm4, [esp + nb102_fjyO] movaps xmm5, [esp + nb102_fjzO] movaps xmm6, xmm5 movaps xmm7, xmm5 shufps xmm6, xmm6, 2 ;# constant 00000010 shufps xmm7, xmm7, 3 ;# constant 00000011 addss xmm5, [esi + eax*4 + 8] addss xmm6, [esi + eax*4 + 20] addss xmm7, [esi + eax*4 + 32] movss [esi + eax*4 + 8], xmm5 movss [esi + eax*4 + 20], xmm6 movss [esi + eax*4 + 32], xmm7 movaps xmm5, xmm3 unpcklps xmm3, xmm4 unpckhps xmm5, xmm4 addps xmm0, xmm3 addps xmm1, xmm5 movlps [esi + eax*4], xmm0 movlps [esi + eax*4 + 12], xmm1 movhps [esi + eax*4 + 24], xmm1 dec dword ptr [esp + nb102_innerk] jz .nb102_updateouterdata jmp .nb102_single_loop.nb102_updateouterdata: mov ecx, [esp + nb102_ii3] mov edi, [ebp + nb102_faction] mov esi, [ebp + nb102_fshift] mov edx, [esp + nb102_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb102_fixO] movaps xmm1, [esp + nb102_fiyO] movaps xmm2, [esp + nb102_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 ⌨️ 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?