nb_kernel304_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,258 行 · 第 1/5 页
S
2,258 行
mulps xmm3, [esp + nb304_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb304_fjxH2] movaps xmm4, [esp + nb304_fjyH2] movaps xmm5, [esp + nb304_fjzH2] mulps xmm0, [esp + nb304_dxMH2] mulps xmm1, [esp + nb304_dyMH2] mulps xmm2, [esp + nb304_dzMH2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb304_fixM] addps xmm1, [esp + nb304_fiyM] addps xmm2, [esp + nb304_fizM] movaps [esp + nb304_fjxH2], xmm3 movaps [esp + nb304_fjyH2], xmm4 movaps [esp + nb304_fjzH2], xmm5 movaps [esp + nb304_fixM], xmm0 movaps [esp + nb304_fiyM], xmm1 movaps [esp + nb304_fizM], xmm2 ;# M-M interaction movaps xmm0, [esp + nb304_rinvMM] movaps xmm1, xmm0 mulps xmm1, [esp + nb304_rsqMM] ;# xmm1=r mulps xmm1, [esp + nb304_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 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 + nb304_two] ;# two*Heps2 movaps xmm3, [esp + nb304_qqMM] 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 + nb304_vctot] movaps [esp + nb304_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb304_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb304_fjxM] movaps xmm4, [esp + nb304_fjyM] movaps xmm5, [esp + nb304_fjzM] mulps xmm0, [esp + nb304_dxMM] mulps xmm1, [esp + nb304_dyMM] mulps xmm2, [esp + nb304_dzMM] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb304_fixM] addps xmm1, [esp + nb304_fiyM] addps xmm2, [esp + nb304_fizM] movaps [esp + nb304_fjxM], xmm3 movaps [esp + nb304_fjyM], xmm4 movaps [esp + nb304_fjzM], xmm5 movaps [esp + nb304_fixM], xmm0 movaps [esp + nb304_fiyM], xmm1 movaps [esp + nb304_fizM], xmm2 mov edi, [ebp + nb304_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: ;# fjxH1, fjyH1, fjzH1, ... , fjzM. ;# 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 1st Hydrogen X & Y forces for 4 j particles movaps xmm0, [esp + nb304_fjxH1] ;# xmm0= fjxH1a fjxH1b fjxH1c fjxH1d movaps xmm2, [esp + nb304_fjyH1] ;# xmm1= fjyH1a fjyH1b fjyH1c fjyH1d movlps xmm3, [edi + eax*4 + 12] movlps xmm4, [edi + ecx*4 + 12] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjxH1a fjyH1a fjxH1b fjyH1b unpckhps xmm1, xmm2 ;# xmm1= fjxH1c fjyH1c fjxH1d fjyH1d movhps xmm3, [edi + ebx*4 + 12] movhps xmm4, [edi + edx*4 + 12] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 12], xmm3 movlps [edi + ecx*4 + 12], xmm4 movhps [edi + ebx*4 + 12], xmm3 movhps [edi + edx*4 + 12], xmm4 ;# constant 1st Hydrogen Z & 2nd hydrogen X forces for 4 j particles movaps xmm0, [esp + nb304_fjzH1] ;# xmm0= fjzH1a fjzH1b fjzH1c fjzH1d movaps xmm2, [esp + nb304_fjxH2] ;# xmm1= fjxH2a fjxH2b fjxH2c fjxH2d movlps xmm3, [edi + eax*4 + 20] movlps xmm4, [edi + ecx*4 + 20] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjzH1a fjxH2a fjzH1b fjxH2b unpckhps xmm1, xmm2 ;# xmm1= fjzH1c fjxH2c fjzH1d fjxH2d movhps xmm3, [edi + ebx*4 + 20] movhps xmm4, [edi + edx*4 + 20] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 20], xmm3 movlps [edi + ecx*4 + 20], xmm4 movhps [edi + ebx*4 + 20], xmm3 movhps [edi + edx*4 + 20], xmm4 ;# constant 2nd hydrogen Y & Z forces for 4 j particles movaps xmm0, [esp + nb304_fjyH2] ;# xmm0= fjyH2a fjyH2b fjyH2c fjyH2d movaps xmm2, [esp + nb304_fjzH2] ;# xmm1= fjzH2a fjzH2b fjzH2c fjzH2d movlps xmm3, [edi + eax*4 + 28] movlps xmm4, [edi + ecx*4 + 28] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjyH2a fjzH2a fjyH2b fjzH2b unpckhps xmm1, xmm2 ;# xmm1= fjyH2c fjzH2c fjyH2d fjzH2d movhps xmm3, [edi + ebx*4 + 28] movhps xmm4, [edi + edx*4 + 28] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 28], xmm3 movlps [edi + ecx*4 + 28], xmm4 movhps [edi + ebx*4 + 28], xmm3 movhps [edi + edx*4 + 28], xmm4 ;# Dummy (M) X & Y forces for 4 j particles movaps xmm0, [esp + nb304_fjxM] ;# xmm0= fjxMa fjxMb fjxMc fjxMd movaps xmm2, [esp + nb304_fjyM] ;# xmm1= fjyMa fjyMb fjyMc fjyMd movlps xmm3, [edi + eax*4 + 36] movlps xmm4, [edi + ecx*4 + 36] movaps xmm1, xmm0 unpcklps xmm0, xmm2 ;# xmm0= fjxMa fjyMa fjxMb fjyMb unpckhps xmm1, xmm2 ;# xmm1= fjxMc fjyMc fjxMd fjyMd movhps xmm3, [edi + ebx*4 + 36] movhps xmm4, [edi + edx*4 + 36] addps xmm3, xmm0 addps xmm4, xmm1 movlps [edi + eax*4 + 36], xmm3 movlps [edi + ecx*4 + 36], xmm4 movhps [edi + ebx*4 + 36], xmm3 movhps [edi + edx*4 + 36], xmm4 ;# Dummy (M) Z forces for 4 j particles ;# Just load the four Z coords into one reg. each movss xmm4, [edi + eax*4 + 44] movss xmm5, [edi + ebx*4 + 44] movss xmm6, [edi + ecx*4 + 44] movss xmm7, [edi + edx*4 + 44] ;# add what we have on the stack addss xmm4, [esp + nb304_fjzM] addss xmm5, [esp + nb304_fjzMb] addss xmm6, [esp + nb304_fjzMc] addss xmm7, [esp + nb304_fjzMd] ;# store back movss [edi + eax*4 + 44], xmm4 movss [edi + ebx*4 + 44], xmm5 movss [edi + ecx*4 + 44], xmm6 movss [edi + edx*4 + 44], xmm7 ;# should we do one more iteration? sub dword ptr [esp + nb304_innerk], 4 jl .nb304_single_check jmp .nb304_unroll_loop.nb304_single_check: add dword ptr [esp + nb304_innerk], 4 jnz .nb304_single_loop jmp .nb304_updateouterdata.nb304_single_loop: mov edx, [esp + nb304_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb304_innerjjnr], 4 mov esi, [ebp + nb304_pos] lea eax, [eax + eax*2] ;# fetch j coordinates xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 movss xmm3, [esi + eax*4 + 36] ;# jxM - - - movss xmm4, [esi + eax*4 + 40] ;# jyM - - - movss xmm5, [esi + eax*4 + 44] ;# jzM - - - 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 + nb304_ixM] movaps xmm1, [esp + nb304_iyM] movaps xmm2, [esp + nb304_izM] movlhps xmm3, xmm6 ;# xmm3 = jxM 0 jxH1 jxH2 shufps xmm4, xmm6, 228 ;# constant 11100100 ;# xmm4 = jyM 0 jyH1 jyH2 shufps xmm5, xmm7, 68 ;# constant 01000100 ;# xmm5 = jzM 0 jzH1 jzH2 ;# store all j coordinates in jM movaps [esp + nb304_jxM], xmm3 movaps [esp + nb304_jyM], xmm4 movaps [esp + nb304_jzM], xmm5 subps xmm0, xmm3 subps xmm1, xmm4 subps xmm2, xmm5 movaps [esp + nb304_dxMM], xmm0 movaps [esp + nb304_dyMM], xmm1 movaps [esp + nb304_dzMM], 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 + nb304_three] mulps xmm1, xmm0 subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb304_half] ;# rinv iO - j water movaps xmm1, xmm3 mulps xmm1, xmm0 ;# xmm1=r movaps xmm0, xmm3 ;# xmm0=rinv mulps xmm1, [esp + nb304_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 ebx, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 ;# table indices in ebx,ecx,edx mov esi, [ebp + nb304_VFtab] movlps xmm5, [esi + ebx*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 + ebx*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 + nb304_two] ;# two*Heps2 xorps xmm3, xmm3 ;# fetch charges to xmm3 (temporary) movss xmm3, [esp + nb304_qqMM] movhps xmm3, [esp + nb304_qqMH] 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 xmm5 contains vcoul and xmm3 fijC addps xmm5, [esp + nb304_vctot] movaps [esp + nb304_vctot], xmm5 xorps xmm2, xmm2 mulps xmm3, [esp + nb304_tsc] subps xmm2, xmm3 mulps xmm0, xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [esp + nb304_dxMM] mulps xmm1, [esp + nb304_dyMM] mulps xmm2, [esp + nb304_dzMM] ;# 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 + nb304_fjxM], xmm3 movaps [esp + nb304_fjyM], xmm4 movaps [esp + nb304_fjzM], xmm5 addps xmm0, [esp + nb304_fixM] addps xmm1, [esp + nb304_fiyM] addps xmm2, [esp + nb304_fizM] movaps [esp + nb304_fixM], xmm0 movaps [esp + nb304_fiyM], xmm1 movaps [esp + nb304_fizM], xmm2 ;# done with i M Now do i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [esp + nb304_ixH1] movaps xmm1, [esp + nb304_iyH1] movaps xmm2, [esp + nb304_izH1] movaps xmm3, [esp + nb304_ixH2] movaps xmm4, [esp + nb304_iyH2] movaps xmm5, [esp + nb304_izH2] subps xmm0, [esp + nb304_jxM] subps xmm1, [esp + nb304_jyM] subps xmm2, [esp + nb304_jzM] subps xmm3, [esp + nb304_jxM] subps xmm4, [esp + nb304_jyM] subps xmm5, [esp + nb304_jzM] movaps [esp + nb304_dxH1M], xmm0 movaps [esp + nb304_dyH1M], xmm1 movaps [esp + nb304_dzH1M], xmm2 movaps [esp + nb304_dxH2M], xmm3 movaps [esp + nb304_dyH2M], xmm4 movaps [esp + nb304_dzH2M], 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 ;# start with H1, save H2 data movaps [esp + nb304_rsqH2M], xmm4 ;# do invsqrt rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb304_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 + nb304_half] ;# rinv H1 - j water mulps xmm7, [esp + nb304_half] ;# rinv H2 - j water ;# start with H1, save H2 data movaps [esp + nb304_rinvH2M], xmm7 movaps xmm1, xmm3 mulps xmm1, xmm0 ;# xmm1=r movaps xmm0, xmm3 ;# xmm0=rinv mulps xmm1, [esp + nb304_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 ebx, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 ;# table indices in ebx,ecx,edx movlps xmm5, [esi + ebx*4]⌨️ 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?