nb_kernel331_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,277 行 · 第 1/5 页
S
2,277 行
subps xmm3, xmm0 subps xmm4, xmm1 ;# unpack them back for storing movaps xmm7, xmm3 unpcklps xmm7, xmm4 unpckhps xmm3, xmm4 movlps [edi + eax*4], xmm7 movlps [edi + ecx*4], xmm3 movhps [edi + ebx*4], xmm7 movhps [edi + edx*4], xmm3 ;# finally z forces movss xmm0, [edi + eax*4 + 8] movss xmm1, [edi + ebx*4 + 8] movss xmm3, [edi + ecx*4 + 8] movss xmm4, [edi + edx*4 + 8] subss xmm0, xmm2 shufps xmm2, xmm2, 229 ;# constant 11100101 subss xmm1, xmm2 shufps xmm2, xmm2, 234 ;# constant 11101010 subss xmm3, xmm2 shufps xmm2, xmm2, 255 ;# constant 11111111 subss xmm4, xmm2 movss [edi + eax*4 + 8], xmm0 movss [edi + ebx*4 + 8], xmm1 movss [edi + ecx*4 + 8], xmm3 movss [edi + edx*4 + 8], xmm4 ;# should we do one more iteration? sub dword ptr [esp + nb331_innerk], 4 jl .nb331_odd_inner jmp .nb331_unroll_loop.nb331_odd_inner: add dword ptr [esp + nb331_innerk], 4 jnz .nb331_odd_loop jmp .nb331_updateouterdata.nb331_odd_loop: mov edx, [esp + nb331_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb331_innerjjnr], 4 xorps xmm4, xmm4 movss xmm4, [esp + nb331_iqO] mov esi, [ebp + nb331_charge] movhps xmm4, [esp + nb331_iqH] movss xmm3, [esi + eax*4] ;# charge in xmm3 shufps xmm3, xmm3, 0 mulps xmm3, xmm4 movaps [esp + nb331_qqO], xmm3 ;# use oxygen qq for storage xorps xmm6, xmm6 mov esi, [ebp + nb331_type] mov ebx, [esi + eax*4] mov esi, [ebp + nb331_vdwparam] shl ebx, 1 add ebx, [esp + nb331_ntia] movlps xmm6, [esi + ebx*4] movaps xmm7, xmm6 shufps xmm6, xmm6, 252 ;# constant 11111100 shufps xmm7, xmm7, 253 ;# constant 11111101 movaps [esp + nb331_c6], xmm6 movaps [esp + nb331_c12], xmm7 mov esi, [ebp + nb331_pos] lea eax, [eax + eax*2] ;# move j coords to xmm0-xmm2 movss xmm0, [esi + eax*4] movss xmm1, [esi + eax*4 + 4] movss xmm2, [esi + eax*4 + 8] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movss xmm3, [esp + nb331_ixO] movss xmm4, [esp + nb331_iyO] movss xmm5, [esp + nb331_izO] movlps xmm6, [esp + nb331_ixH1] movlps xmm7, [esp + nb331_ixH2] unpcklps xmm6, xmm7 movlhps xmm3, xmm6 movlps xmm6, [esp + nb331_iyH1] movlps xmm7, [esp + nb331_iyH2] unpcklps xmm6, xmm7 movlhps xmm4, xmm6 movlps xmm6, [esp + nb331_izH1] movlps xmm7, [esp + nb331_izH2] unpcklps xmm6, xmm7 movlhps xmm5, xmm6 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb331_dxO], xmm3 movaps [esp + nb331_dyO], xmm4 movaps [esp + nb331_dzO], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm4, xmm3 addps xmm4, xmm5 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [esp + nb331_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb331_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv ;# a little trick to avoid NaNs: ;# positions 0,2,and 3 are valid, but not 1. ;# If it contains NaN it doesnt help to mult by 0, ;# So we shuffle it and copy pos 0 to pos1! shufps xmm0, xmm0, 224 ;# constant 11100000 mulps xmm4, xmm0 ;# xmm4=r movaps [esp + nb331_rinvO], xmm0 mulps xmm4, [esp + nb331_tsc] movhlps xmm7, xmm4 cvttps2pi mm6, xmm4 cvttps2pi mm7, xmm7 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm7, mm7 movlhps xmm3, xmm7 subps xmm4, xmm3 movaps xmm1, xmm4 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd mm0, eax movd mm1, ecx movd mm2, edx mov esi, [ebp + nb331_VFtab] movd eax, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 lea eax, [eax + eax*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*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 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 + nb331_two] ;# two*Heps2 movaps xmm0, [esp + nb331_qqO] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm0 ;# vcoul=qq*VV mulps xmm0, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and xmm0 fijC ;# increment vcoul - then we can get rid of mm5 addps xmm5, [esp + nb331_vctot] movaps [esp + nb331_vctot], xmm5 ;# dispersion movlps xmm5, [esi + eax*4 + 16] ;# half table movaps xmm4, xmm5 shufps xmm4, xmm4, 252 ;# constant 11111100 shufps xmm5, xmm5, 253 ;# constant 11111101 movlps xmm7, [esi + eax*4 + 24] ;# other half of dispersion table movaps xmm6, xmm7 shufps xmm6, xmm6, 252 ;# constant 11111100 shufps xmm7, xmm7, 253 ;# constant 11111101 ;# dispersion table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 ;# Update Vvdwtot directly addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [esp + nb331_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movaps xmm4, [esp + nb331_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 addps xmm0, xmm7 ;# add to fscal ;# Update Vvdwtot directly addps xmm5, [esp + nb331_Vvdwtot] movaps [esp + nb331_Vvdwtot], xmm5 ;# repulsion movlps xmm5, [esi + eax*4 + 32] ;# got half repulsion table movaps xmm4, xmm5 shufps xmm4, xmm4, 136 ;# constant 10001000 shufps xmm5, xmm5, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 40] ;# other half of repulsion table movaps xmm6, xmm7 shufps xmm6, xmm6, 136 ;# constant 10001000 shufps xmm7, xmm7, 221 ;# constant 11011101 ;# repulsion 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, [esp + nb331_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movaps xmm4, [esp + nb331_c12] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, xmm0 ;# add to fscal addps xmm5, [esp + nb331_Vvdwtot] ;# total nonbonded potential in xmm5 xorps xmm4, xmm4 movd eax, mm0 movd ecx, mm1 movd edx, mm2 mulps xmm7, [esp + nb331_rinvO] ;# total fscal now in xmm7 movaps [esp + nb331_Vvdwtot], xmm5 mulps xmm7, [esp + nb331_tsc] subps xmm4, xmm7 movaps xmm0, [esp + nb331_dxO] movaps xmm1, [esp + nb331_dyO] movaps xmm2, [esp + nb331_dzO] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 now contains tx-tz (partial force) movss xmm3, [esp + nb331_fixO] movss xmm4, [esp + nb331_fiyO] movss xmm5, [esp + nb331_fizO] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [esp + nb331_fixO], xmm3 movss [esp + nb331_fiyO], xmm4 movss [esp + nb331_fizO], xmm5 ;# updated the O force now do the H's movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 230 ;# constant 11100110 ;# shift right shufps xmm4, xmm4, 230 ;# constant 11100110 shufps xmm5, xmm5, 230 ;# constant 11100110 addss xmm3, [esp + nb331_fixH1] addss xmm4, [esp + nb331_fiyH1] addss xmm5, [esp + nb331_fizH1] movss [esp + nb331_fixH1], xmm3 movss [esp + nb331_fiyH1], xmm4 movss [esp + nb331_fizH1], xmm5 ;# updated the H1 force mov edi, [ebp + nb331_faction] shufps xmm3, xmm3, 231 ;# constant 11100111 ;# shift right shufps xmm4, xmm4, 231 ;# constant 11100111 shufps xmm5, xmm5, 231 ;# constant 11100111 addss xmm3, [esp + nb331_fixH2] addss xmm4, [esp + nb331_fiyH2] addss xmm5, [esp + nb331_fizH2] movss [esp + nb331_fixH2], xmm3 movss [esp + nb331_fiyH2], xmm4 movss [esp + nb331_fizH2], xmm5 ;# updated the H2 force ;# the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps xmm5, xmm5 movaps xmm3, xmm0 movlps xmm6, [edi + eax*4] movss xmm7, [edi + eax*4 + 8] unpcklps xmm3, xmm1 movlhps xmm3, xmm5 unpckhps xmm0, xmm1 addps xmm0, xmm3 movhlps xmm3, xmm0 addps xmm0, xmm3 ;# x,y sum in xmm0 movhlps xmm1, xmm2 addss xmm2, xmm1 shufps xmm1, xmm1, 1 addss xmm2, xmm1 ;# z sum in xmm2 subps xmm6, xmm0 subss xmm7, xmm2 movlps [edi + eax*4], xmm6 movss [edi + eax*4 + 8], xmm7 dec dword ptr [esp + nb331_innerk] jz .nb331_updateouterdata jmp .nb331_odd_loop.nb331_updateouterdata: mov ecx, [esp + nb331_ii3] mov edi, [ebp + nb331_faction] mov esi, [ebp + nb331_fshift] mov edx, [esp + nb331_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb331_fixO] movaps xmm1, [esp + nb331_fiyO] movaps xmm2, [esp + nb331_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, [edi + ecx*4] movss xmm4, [edi + ecx*4 + 4] movss xmm5, [edi + ecx*4 + 8] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4], xmm3 movss [edi + ecx*4 + 4], xmm4 movss [edi + ecx*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, [esp + nb331_fixH1] movaps xmm1, [esp + nb331_fiyH1] movaps xmm2, [esp + nb331_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 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, [edi + ecx*4 + 12] movss xmm4, [edi + ecx*4 + 16] movss xmm5, [edi + ecx*4 + 20] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4 + 12], xmm3 movss [edi + ecx*4 + 16], xmm4 movss [edi + ecx*4 + 20], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb331_fixH2] movaps xmm1, [esp + nb331_fiyH2] movaps xmm2, [esp + nb331_fizH2] 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, [edi + ecx*4 + 24] movss xmm4, [edi + ecx*4 + 28] movss xmm5, [edi + ecx*4 + 32] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4 + 24], xmm3 movss [edi + ecx*4 + 28], xmm4 movss [edi + ecx*4 + 32], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# increment fshift force movlps xmm3, [esi + edx*4] movss xmm4, [esi + edx*4 + 8] addps xmm3, xmm6 addss xmm4, xmm7 movlps [esi + edx*4], xmm3 movss [esi + edx*4 + 8], xmm4 ;# get n from stack mov esi, [esp + nb331_n] ;# get group index for i particle ⌨️ 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?