nb_kernel103_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,466 行 · 第 1/4 页
S
1,466 行
subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb103_dxH2(%esp) movaps %xmm5,nb103_dyH2(%esp) movaps %xmm6,nb103_dzH2(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixM-izM to xmm3-xmm5 movaps nb103_ixM(%esp),%xmm3 movaps nb103_iyM(%esp),%xmm4 movaps nb103_izM(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## store dr movaps %xmm3,nb103_dxM(%esp) movaps %xmm4,nb103_dyM(%esp) movaps %xmm5,nb103_dzM(%esp) ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## rsqM in xmm5, rsqH2 in xmm6, rsqH1 in xmm7 ## start with rsqH1 - seed in xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb103_three(%esp),%xmm4 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb103_half(%esp),%xmm4 movaps %xmm4,%xmm7 ## rinvH1 in xmm7 ## rsqH2 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb103_three(%esp),%xmm4 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb103_half(%esp),%xmm4 movaps %xmm4,%xmm6 ## rinvH2 in xmm6 ## rsqM - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb103_three(%esp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb103_half(%esp),%xmm4 movaps %xmm4,%xmm5 ## rinvM in xmm5 ## do H1 interactions movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq mulps nb103_qqH(%esp),%xmm7 ## xmm7=vcoul mulps %xmm7,%xmm4 ## total fsH1 in xmm4 addps nb103_vctot(%esp),%xmm7 movaps %xmm7,nb103_vctot(%esp) movaps nb103_dxH1(%esp),%xmm0 movaps nb103_dyH1(%esp),%xmm1 movaps nb103_dzH1(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H1 forces movaps nb103_fixH1(%esp),%xmm3 movaps nb103_fiyH1(%esp),%xmm4 movaps nb103_fizH1(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb103_fixH1(%esp) movaps %xmm4,nb103_fiyH1(%esp) movaps %xmm7,nb103_fizH1(%esp) ## update j forces with water O movaps %xmm0,nb103_fjx(%esp) movaps %xmm1,nb103_fjy(%esp) movaps %xmm2,nb103_fjz(%esp) ## H2 interactions movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm6=rinv, xmm4=rinvsq mulps nb103_qqH(%esp),%xmm6 ## xmm6=vcoul mulps %xmm6,%xmm4 ## total fsH2 in xmm4 addps nb103_vctot(%esp),%xmm6 movaps nb103_dxH2(%esp),%xmm0 movaps nb103_dyH2(%esp),%xmm1 movaps nb103_dzH2(%esp),%xmm2 movaps %xmm6,nb103_vctot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H2 forces movaps nb103_fixH2(%esp),%xmm3 movaps nb103_fiyH2(%esp),%xmm4 movaps nb103_fizH2(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb103_fixH2(%esp) movaps %xmm4,nb103_fiyH2(%esp) movaps %xmm7,nb103_fizH2(%esp) ## update j forces with water H2 addps nb103_fjx(%esp),%xmm0 addps nb103_fjy(%esp),%xmm1 addps nb103_fjz(%esp),%xmm2 movaps %xmm0,nb103_fjx(%esp) movaps %xmm1,nb103_fjy(%esp) movaps %xmm2,nb103_fjz(%esp) ## M interactions movaps %xmm5,%xmm4 mulps %xmm4,%xmm4 ## xmm5=rinv, xmm4=rinvsq mulps nb103_qqM(%esp),%xmm5 ## xmm5=vcoul mulps %xmm5,%xmm4 ## total fsM in xmm4 addps nb103_vctot(%esp),%xmm5 movaps nb103_dxM(%esp),%xmm0 movaps nb103_dyM(%esp),%xmm1 movaps nb103_dzM(%esp),%xmm2 movaps %xmm5,nb103_vctot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update M forces movaps nb103_fixM(%esp),%xmm3 movaps nb103_fiyM(%esp),%xmm4 movaps nb103_fizM(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb103_fixM(%esp) movaps %xmm4,nb103_fiyM(%esp) movaps %xmm7,nb103_fizM(%esp) movl nb103_faction(%ebp),%edi ## update j forces addps nb103_fjx(%esp),%xmm0 addps nb103_fjy(%esp),%xmm1 addps nb103_fjz(%esp),%xmm2 movlps (%edi,%eax,4),%xmm4 movlps (%edi,%ecx,4),%xmm7 movhps (%edi,%ebx,4),%xmm4 movhps (%edi,%edx,4),%xmm7 movaps %xmm4,%xmm3 shufps $136,%xmm7,%xmm3 ## constant 10001000 shufps $221,%xmm7,%xmm4 ## constant 11011101 ## xmm3 has fjx, xmm4 has fjy subps %xmm0,%xmm3 subps %xmm1,%xmm4 ## unpack them back for storing movaps %xmm3,%xmm7 unpcklps %xmm4,%xmm7 unpckhps %xmm4,%xmm3 movlps %xmm7,(%edi,%eax,4) movlps %xmm3,(%edi,%ecx,4) movhps %xmm7,(%edi,%ebx,4) movhps %xmm3,(%edi,%edx,4) ## finally z forces movss 8(%edi,%eax,4),%xmm0 movss 8(%edi,%ebx,4),%xmm1 movss 8(%edi,%ecx,4),%xmm3 movss 8(%edi,%edx,4),%xmm4 subss %xmm2,%xmm0 shufps $229,%xmm2,%xmm2 ## constant 11100101 subss %xmm2,%xmm1 shufps $234,%xmm2,%xmm2 ## constant 11101010 subss %xmm2,%xmm3 shufps $255,%xmm2,%xmm2 ## constant 11111111 subss %xmm2,%xmm4 movss %xmm0,8(%edi,%eax,4) movss %xmm1,8(%edi,%ebx,4) movss %xmm3,8(%edi,%ecx,4) movss %xmm4,8(%edi,%edx,4) ## should we do one more iteration? subl $4,nb103_innerk(%esp) jl _nb_kernel103_ia32_sse.nb103_odd_inner jmp _nb_kernel103_ia32_sse.nb103_unroll_loop_nb_kernel103_ia32_sse.nb103_odd_inner: addl $4,nb103_innerk(%esp) jnz _nb_kernel103_ia32_sse.nb103_odd_loop jmp _nb_kernel103_ia32_sse.nb103_updateouterdata_nb_kernel103_ia32_sse.nb103_odd_loop: movl nb103_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb103_innerjjnr(%esp) xorps %xmm4,%xmm4 movss nb103_iqM(%esp),%xmm4 movl nb103_charge(%ebp),%esi movhps nb103_iqH(%esp),%xmm4 movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb103_qqM(%esp) ## use dummy qq for storage movl nb103_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## move j coords to xmm0-xmm2 movss (%esi,%eax,4),%xmm0 movss 4(%esi,%eax,4),%xmm1 movss 8(%esi,%eax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movss nb103_ixM(%esp),%xmm3 movss nb103_iyM(%esp),%xmm4 movss nb103_izM(%esp),%xmm5 movlps nb103_ixH1(%esp),%xmm6 movlps nb103_ixH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb103_iyH1(%esp),%xmm6 movlps nb103_iyH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb103_izH1(%esp),%xmm6 movlps nb103_izH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use dummy dx for storage movaps %xmm3,nb103_dxM(%esp) movaps %xmm4,nb103_dyM(%esp) movaps %xmm5,nb103_dzM(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb103_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb103_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## 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 $224,%xmm0,%xmm0 ## constant 11100000 movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps nb103_qqM(%esp),%xmm3 mulps %xmm0,%xmm3 ## xmm3=vcoul mulps %xmm3,%xmm4 ## xmm4=total fscal addps nb103_vctot(%esp),%xmm3 movaps nb103_dxM(%esp),%xmm0 movaps nb103_dyM(%esp),%xmm1 movaps nb103_dzM(%esp),%xmm2 movaps %xmm3,nb103_vctot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb103_fixM(%esp),%xmm3 movss nb103_fiyM(%esp),%xmm4 movss nb103_fizM(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb103_fixM(%esp) movss %xmm4,nb103_fiyM(%esp) movss %xmm5,nb103_fizM(%esp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## constant 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## constant 11100110 shufps $230,%xmm5,%xmm5 ## constant 11100110 addss nb103_fixH1(%esp),%xmm3 addss nb103_fiyH1(%esp),%xmm4 addss nb103_fizH1(%esp),%xmm5 movss %xmm3,nb103_fixH1(%esp) movss %xmm4,nb103_fiyH1(%esp) movss %xmm5,nb103_fizH1(%esp) ## updated the H1 force movl nb103_faction(%ebp),%edi shufps $231,%xmm3,%xmm3 ## constant 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## constant 11100111 shufps $231,%xmm5,%xmm5 ## constant 11100111 addss nb103_fixH2(%esp),%xmm3 addss nb103_fiyH2(%esp),%xmm4 addss nb103_fizH2(%esp),%xmm5 movss %xmm3,nb103_fixH2(%esp) movss %xmm4,nb103_fiyH2(%esp) movss %xmm5,nb103_fizH2(%esp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%edi,%eax,4),%xmm6 movss 8(%edi,%eax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 subps %xmm0,%xmm6 subss %xmm2,%xmm7 movlps %xmm6,(%edi,%eax,4) movss %xmm7,8(%edi,%eax,4) decl nb103_innerk(%esp) jz _nb_kernel103_ia32_sse.nb103_updateouterdata jmp _nb_kernel103_ia32_sse.nb103_odd_loop_nb_kernel103_ia32_sse.nb103_updateouterdata: movl nb103_ii3(%esp),%ecx movl nb103_faction(%ebp),%edi movl nb103_fshift(%ebp),%esi⌨️ 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?