nb_kernel332_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,165 行 · 第 1/5 页
S
2,165 行
mulps xmm7, [esp + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_qqHH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxH2] movaps xmm4, [esp + nb332_fjyH2] movaps xmm5, [esp + nb332_fjzH2] mulps xmm0, [esp + nb332_dxH1H2] mulps xmm1, [esp + nb332_dyH1H2] mulps xmm2, [esp + nb332_dzH1H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH1] addps xmm1, [esp + nb332_fiyH1] addps xmm2, [esp + nb332_fizH1] movaps [esp + nb332_fjxH2], xmm3 movaps [esp + nb332_fjyH2], xmm4 movaps [esp + nb332_fjzH2], xmm5 movaps [esp + nb332_fixH1], xmm0 movaps [esp + nb332_fiyH1], xmm1 movaps [esp + nb332_fizH1], xmm2 ;# H2-O interaction movaps xmm0, [esp + nb332_rinvH2O] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH2O] ;# xmm1=r mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm7, [esi + ebx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_qqOH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxO] movaps xmm4, [esp + nb332_fjyO] movaps xmm5, [esp + nb332_fjzO] mulps xmm0, [esp + nb332_dxH2O] mulps xmm1, [esp + nb332_dyH2O] mulps xmm2, [esp + nb332_dzH2O] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH2] addps xmm1, [esp + nb332_fiyH2] addps xmm2, [esp + nb332_fizH2] movaps [esp + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 movaps [esp + nb332_fixH2], xmm0 movaps [esp + nb332_fiyH2], xmm1 movaps [esp + nb332_fizH2], xmm2 ;# H2-H1 interaction movaps xmm0, [esp + nb332_rinvH2H1] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH2H1] ;# xmm1=r mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm7, [esi + ebx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_qqHH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxH1] movaps xmm4, [esp + nb332_fjyH1] movaps xmm5, [esp + nb332_fjzH1] mulps xmm0, [esp + nb332_dxH2H1] mulps xmm1, [esp + nb332_dyH2H1] mulps xmm2, [esp + nb332_dzH2H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH2] addps xmm1, [esp + nb332_fiyH2] addps xmm2, [esp + nb332_fizH2] movaps [esp + nb332_fjxH1], xmm3 movaps [esp + nb332_fjyH1], xmm4 movaps [esp + nb332_fjzH1], xmm5 movaps [esp + nb332_fixH2], xmm0 movaps [esp + nb332_fiyH2], xmm1 movaps [esp + nb332_fizH2], xmm2 ;# H2-H2 interaction movaps xmm0, [esp + nb332_rinvH2H2] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH2H2] ;# xmm1=r mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm7, [esi + ebx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_qqHH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxH2] movaps xmm4, [esp + nb332_fjyH2] movaps xmm5, [esp + nb332_fjzH2] mulps xmm0, [esp + nb332_dxH2H2] mulps xmm1, [esp + nb332_dyH2H2] mulps xmm2, [esp + nb332_dzH2H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH2] addps xmm1, [esp + nb332_fiyH2] addps xmm2, [esp + nb332_fizH2] movaps [esp + nb332_fjxH2], xmm3 movaps [esp + nb332_fjyH2], xmm4 movaps [esp + nb332_fjzH2], xmm5 movaps [esp + nb332_fixH2], xmm0 movaps [esp + nb332_fiyH2], xmm1 movaps [esp + nb332_fizH2], xmm2 mov edi, [ebp + nb332_faction] movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 ;# 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 + nb332_fjxO] ;# xmm0= fjxOa fjxOb fjxOc fjxOd movaps xmm2, [esp + nb332_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 + nb332_fjzO] ;# xmm0= fjzOa fjzOb fjzOc fjzOd movaps xmm2, [esp + nb332_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 + nb332_fjyH1] ;# xmm0= fjyH1a fjyH1b fjyH1c fjyH1d movaps xmm2, [esp + nb332_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 + nb332_fjxH2] ;# xmm0= fjxH2a fjxH2b fjxH2c fjxH2d movaps xmm2, [esp + nb332_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 + nb332_fjzH2] addss xmm5, [esp + nb332_fjzH2b] addss xmm6, [esp + nb332_fjzH2c] addss xmm7, [esp + nb332_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 + nb332_innerk], 4 jl .nb332_single_check jmp .nb332_unroll_loop.nb332_single_check: add dword ptr [esp + nb332_innerk], 4 jnz .nb332_single_loop jmp .nb332_updateouterdata.nb332_single_loop: mov edx, [esp + nb332_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb332_innerjjnr], 4 mov esi, [ebp + nb332_pos] lea eax, [eax + eax*2] ⌨️ 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?