nb_kernel113_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,895 行 · 第 1/4 页
S
1,895 行
;# rsqO in xmm7 ;# move ixH1-izH1 to xmm4-xmm6 movaps xmm4, [esp + nb113_ixH1] movaps xmm5, [esp + nb113_iyH1] movaps xmm6, [esp + nb113_izH1] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb113_dxH1], xmm4 movaps [esp + nb113_dyH1], xmm5 movaps [esp + nb113_dzH1], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm6, xmm5 addps xmm6, xmm4 ;# rsqH1 in xmm6 ;# move ixH2-izH2 to xmm3-xmm5 movaps xmm3, [esp + nb113_ixH2] movaps xmm4, [esp + nb113_iyH2] movaps xmm5, [esp + nb113_izH2] ;# calc dr subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# store dr movaps [esp + nb113_dxH2], xmm3 movaps [esp + nb113_dyH2], xmm4 movaps [esp + nb113_dzH2], xmm5 ;# square it mulps xmm3,xmm3 mulps xmm4,xmm4 mulps xmm5,xmm5 addps xmm5, xmm4 addps xmm5, xmm3 ;# move ixM-izM to xmm2-xmm4 movaps xmm3, [esp + nb113_iyM] movaps xmm4, [esp + nb113_izM] subps xmm3, xmm1 subps xmm4, xmm2 movaps xmm2, [esp + nb113_ixM] subps xmm2, xmm0 ;# store dr movaps [esp + nb113_dxM], xmm2 movaps [esp + nb113_dyM], xmm3 movaps [esp + nb113_dzM], xmm4 ;# square it mulps xmm2,xmm2 mulps xmm3,xmm3 mulps xmm4,xmm4 addps xmm4, xmm3 addps xmm4, xmm2 ;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb113_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb113_half] movaps [esp + nb113_rinvH1], xmm0 ;# rinvH1 ;# rsqH2 - seed to xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb113_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb113_half] movaps [esp + nb113_rinvH2], xmm0 ;# rinvH2 ;# rsqM - seed to xmm2 rsqrtps xmm2, xmm4 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb113_three] mulps xmm2, xmm4 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb113_half] movaps [esp + nb113_rinvM], xmm0 ;# Do the O LJ-only interaction directly. rcpps xmm2, xmm7 movaps xmm1, [esp + nb113_two] mulps xmm7, xmm2 subps xmm1, xmm7 mulps xmm2, xmm1 ;# rinvsq movaps xmm0, xmm2 mulps xmm0, xmm2 ;# r4 mulps xmm0, xmm2 ;# r6 movaps xmm1, xmm0 mulps xmm1, xmm1 ;# r12 mulps xmm0, [esp + nb113_c6] mulps xmm1, [esp + nb113_c12] movaps xmm3, xmm1 subps xmm3, xmm0 ;# Vvdw12-Vvdw6 addps xmm3, [esp + nb113_Vvdwtot] movaps [esp + nb113_Vvdwtot], xmm3 mulps xmm0, [esp + nb113_six] mulps xmm1, [esp + nb113_twelve] subps xmm1, xmm0 mulps xmm1, xmm2 ;# fscal movaps xmm3, [esp + nb113_dxO] movaps xmm4, [esp + nb113_dyO] movaps xmm5, [esp + nb113_dzO] mulps xmm3, xmm1 mulps xmm4, xmm1 mulps xmm5, xmm1 ;# tx in xmm3-xmm5 ;# update O forces movaps xmm0, [esp + nb113_fixO] movaps xmm1, [esp + nb113_fiyO] movaps xmm2, [esp + nb113_fizO] addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 movaps [esp + nb113_fixO], xmm0 movaps [esp + nb113_fiyO], xmm1 movaps [esp + nb113_fizO], xmm2 ;# update j forces with water O movaps [esp + nb113_fjx], xmm3 movaps [esp + nb113_fjy], xmm4 movaps [esp + nb113_fjz], xmm5 ;# Do H1 interaction movaps xmm7, [esp + nb113_rinvH1] movaps xmm4, xmm7 mulps xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq mulps xmm7, [esp + nb113_qqH] ;# xmm7=vcoul mulps xmm4, xmm7 ;# total fsH1 in xmm4 addps xmm7, [esp + nb113_vctot] movaps [esp + nb113_vctot], xmm7 movaps xmm0, [esp + nb113_dxH1] movaps xmm1, [esp + nb113_dyH1] movaps xmm2, [esp + nb113_dzH1] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H1 forces movaps xmm3, [esp + nb113_fixH1] movaps xmm4, [esp + nb113_fiyH1] movaps xmm7, [esp + nb113_fizH1] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb113_fixH1], xmm3 movaps [esp + nb113_fiyH1], xmm4 movaps [esp + nb113_fizH1], xmm7 ;# update j forces with water H1 addps xmm0, [esp + nb113_fjx] addps xmm1, [esp + nb113_fjy] addps xmm2, [esp + nb113_fjz] movaps [esp + nb113_fjx], xmm0 movaps [esp + nb113_fjy], xmm1 movaps [esp + nb113_fjz], xmm2 ;# Done with H1, do H2 interactions movaps xmm7, [esp + nb113_rinvH2] movaps xmm4, xmm7 mulps xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq mulps xmm7, [esp + nb113_qqH] ;# xmm7=vcoul mulps xmm4, xmm7 ;# total fsH2 in xmm4 addps xmm7, [esp + nb113_vctot] movaps [esp + nb113_vctot], xmm7 movaps xmm0, [esp + nb113_dxH2] movaps xmm1, [esp + nb113_dyH2] movaps xmm2, [esp + nb113_dzH2] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H2 forces movaps xmm3, [esp + nb113_fixH2] movaps xmm4, [esp + nb113_fiyH2] movaps xmm7, [esp + nb113_fizH2] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb113_fixH2], xmm3 movaps [esp + nb113_fiyH2], xmm4 movaps [esp + nb113_fizH2], xmm7 addps xmm0, [esp + nb113_fjx] addps xmm1, [esp + nb113_fjy] addps xmm2, [esp + nb113_fjz] movaps [esp + nb113_fjx], xmm0 movaps [esp + nb113_fjy], xmm1 movaps [esp + nb113_fjz], xmm2 ;# Done with H2, do M interactions movaps xmm7, [esp + nb113_rinvM] movaps xmm4, xmm7 mulps xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq mulps xmm7, [esp + nb113_qqM] ;# xmm7=vcoul mulps xmm4, xmm7 ;# total fsM in xmm4 addps xmm7, [esp + nb113_vctot] movaps [esp + nb113_vctot], xmm7 movaps xmm0, [esp + nb113_dxM] movaps xmm1, [esp + nb113_dyM] movaps xmm2, [esp + nb113_dzM] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update M forces movaps xmm3, [esp + nb113_fixM] movaps xmm4, [esp + nb113_fiyM] movaps xmm7, [esp + nb113_fizM] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb113_fixM], xmm3 movaps [esp + nb113_fiyM], xmm4 movaps [esp + nb113_fizM], xmm7 mov edi, [ebp + nb113_faction] ;# update j forces from stored values addps xmm0, [esp + nb113_fjx] addps xmm1, [esp + nb113_fjy] addps xmm2, [esp + nb113_fjz] movlps xmm4, [edi + eax*4] movlps xmm7, [edi + ecx*4] movhps xmm4, [edi + ebx*4] movhps xmm7, [edi + edx*4] movaps xmm3, xmm4 shufps xmm3, xmm7, 136 ;# constant 10001000 shufps xmm4, xmm7, 221 ;# constant 11011101 ;# xmm3 has fjx, xmm4 has fjy 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 + nb113_innerk], 4 jl .nb113_odd_inner jmp .nb113_unroll_loop.nb113_odd_inner: add dword ptr [esp + nb113_innerk], 4 jnz .nb113_odd_loop jmp .nb113_updateouterdata.nb113_odd_loop: mov edx, [esp + nb113_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb113_innerjjnr], 4 xorps xmm4, xmm4 ;# clear reg. movss xmm4, [esp + nb113_iqM] mov esi, [ebp + nb113_charge] movhps xmm4, [esp + nb113_iqH] ;# [qM 0 qH qH] shufps xmm4, xmm4, 41 ;# [0 qH qH qM] movss xmm3, [esi + eax*4] ;# charge in xmm3 shufps xmm3, xmm3, 0 mulps xmm3, xmm4 movaps [esp + nb113_qqM], xmm3 ;# use dummy qq for storage xorps xmm6, xmm6 mov esi, [ebp + nb113_type] mov ebx, [esi + eax*4] mov esi, [ebp + nb113_vdwparam] shl ebx, 1 add ebx, [esp + nb113_ntia] movlps xmm6, [esi + ebx*4] movaps xmm7, xmm6 shufps xmm6, xmm6, 252 ;# constant 11111100 shufps xmm7, xmm7, 253 ;# constant 11111101 movaps [esp + nb113_c6], xmm6 movaps [esp + nb113_c12], xmm7 mov esi, [ebp + nb113_pos] lea eax, [eax + eax*2] movss xmm3, [esp + nb113_ixO] movss xmm4, [esp + nb113_iyO] movss xmm5, [esp + nb113_izO] movss xmm0, [esp + nb113_ixH1] movss xmm1, [esp + nb113_iyH1] movss xmm2, [esp + nb113_izH1] unpcklps xmm3, [esp + nb113_ixH2] ;# ixO ixH2 - - unpcklps xmm4, [esp + nb113_iyH2] ;# iyO iyH2 - - unpcklps xmm5, [esp + nb113_izH2] ;# izO izH2 - - unpcklps xmm0, [esp + nb113_ixM] ;# ixH1 ixM - - unpcklps xmm1, [esp + nb113_iyM] ;# iyH1 iyM - - unpcklps xmm2, [esp + nb113_izM] ;# izH1 izM - - unpcklps xmm3, xmm0 ;# ixO ixH1 ixH2 ixM unpcklps xmm4, xmm1 ;# same for y unpcklps xmm5, xmm2 ;# same for z ;# 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 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# use O distances for storage movaps [esp + nb113_dxO], xmm3 movaps [esp + nb113_dyO], xmm4 movaps [esp + nb113_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 + nb113_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb113_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv movaps xmm1, xmm0 mulps xmm1, xmm1 ;# rinvsq movaps xmm7, xmm0 mulps xmm7, [esp + nb113_qqM] ;# vcoul movaps xmm6, xmm7 addps xmm7, [esp + nb113_vctot] movaps [esp + nb113_vctot], xmm7 movaps xmm2, xmm1 mulss xmm1, xmm1 mulss xmm1, xmm2 ;# xmm1=rinvsix xorps xmm4, xmm4 movss xmm4, xmm1 mulss xmm4, xmm4 ;# xmm4=rinvtwelve mulss xmm1, [esp + nb113_c6] mulss xmm4, [esp + nb113_c12] movaps xmm3, xmm4 subss xmm3, xmm1 ;# xmm3=Vvdw12-Vvdw6 mulss xmm1, [esp + nb113_six] mulss xmm4, [esp + nb113_twelve] subss xmm4, xmm1 addss xmm3, [esp + nb113_Vvdwtot] movss [esp + nb113_Vvdwtot], xmm3 addps xmm4, xmm6 mulps xmm4, xmm0 mulps xmm4, xmm0 ;# fscal movaps xmm0, [esp + nb113_dxO] movaps xmm1, [esp + nb113_dyO] movaps xmm2, [esp + nb113_dzO] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 now contains tx-tz (partial force) movss xmm3, [esp + nb113_fixO] movss xmm4, [esp + nb113_fiyO] movss xmm5, [esp + nb113_fizO] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [esp + nb113_fixO], xmm3 movss [esp + nb113_fiyO], xmm4 movss [esp + nb113_fizO], xmm5 ;# updated the O force now do the H's movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 0x39 ;# shift right shufps xmm4, xmm4, 0x39 shufps xmm5, xmm5, 0x39 addss xmm3, [esp + nb113_fixH1] addss xmm4, [esp + nb113_fiyH1] addss xmm5, [esp + nb113_fizH1] movss [esp + nb113_fixH1], xmm3 movss [esp + nb113_fiyH1], xmm4 movss [esp + nb113_fizH1], xmm5 ;# updated the H1 force shufps xmm3, xmm3, 0x39 shufps xmm4, xmm4, 0x39 shufps xmm5, xmm5, 0x39 addss xmm3, [esp + nb113_fixH2] addss xmm4, [esp + nb113_fiyH2] addss xmm5, [esp + nb113_fizH2] movss [esp + nb113_fixH2], xmm3 movss [esp + nb113_fiyH2], xmm4 movss [esp + nb113_fizH2], xmm5 ;# updated the H2 force mov edi, [ebp + nb113_faction] shufps xmm3, xmm3, 0x39 shufps xmm4, xmm4, 0x39 shufps xmm5, xmm5, 0x39 addss xmm3, [esp + nb113_fixM] addss xmm4, [esp + nb113_fiyM] addss xmm5, [esp + nb113_fizM] movss [esp + nb113_fixM], xmm3 movss [esp + nb113_fiyM], xmm4 movss [esp + nb113_fizM], xmm5 ;# updated the M force ;# the fj's - move in from mem start by acc. tx/ty/tz in xmm0, xmm1 movlps xmm6, [edi + eax*4] movss xmm7, [edi + eax*4 + 8] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps xmm0, xmm3 movaps xmm1, xmm4 movaps xmm2, xmm5 shufps xmm3, xmm3, 0x39 ;# shift right shufps xmm4, xmm4, 0x39 shufps xmm5, xmm5, 0x39 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 unpcklps xmm0, xmm1 ;# x,y sum in xmm0, z sum in xmm2⌨️ 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?