nb_kernel201_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,574 行 · 第 1/4 页
S
1,574 行
## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb201_ixH2(%esp),%xmm3 movaps nb201_iyH2(%esp),%xmm4 movaps nb201_izH2(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## store dr movaps %xmm3,nb201_dxH2(%esp) movaps %xmm4,nb201_dyH2(%esp) movaps %xmm5,nb201_dzH2(%esp) ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movaps %xmm5,%xmm0 movaps %xmm6,%xmm1 movaps %xmm7,%xmm2 mulps nb201_krf(%esp),%xmm0 mulps nb201_krf(%esp),%xmm1 mulps nb201_krf(%esp),%xmm2 movaps %xmm0,nb201_krsqH2(%esp) movaps %xmm1,nb201_krsqH1(%esp) movaps %xmm2,nb201_krsqO(%esp) ## start with rsqO - seed in xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb201_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 nb201_half(%esp),%xmm4 movaps %xmm4,%xmm7 ## rinvO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb201_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 nb201_half(%esp),%xmm4 movaps %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb201_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 nb201_half(%esp),%xmm4 movaps %xmm4,%xmm5 ## rinvH2 in xmm5 ## do O interactions movaps %xmm7,%xmm4 mulps %xmm4,%xmm4 ## xmm7=rinv, xmm4=rinvsq movaps %xmm7,%xmm0 movaps nb201_krsqO(%esp),%xmm1 addps %xmm1,%xmm0 subps nb201_crf(%esp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb201_two(%esp),%xmm1 subps %xmm1,%xmm7 mulps nb201_qqO(%esp),%xmm0 mulps nb201_qqO(%esp),%xmm7 mulps %xmm7,%xmm4 ## total fsO in xmm4 addps nb201_vctot(%esp),%xmm0 movaps %xmm0,nb201_vctot(%esp) movaps nb201_dxO(%esp),%xmm0 movaps nb201_dyO(%esp),%xmm1 movaps nb201_dzO(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update O forces movaps nb201_fixO(%esp),%xmm3 movaps nb201_fiyO(%esp),%xmm4 movaps nb201_fizO(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb201_fixO(%esp) movaps %xmm4,nb201_fiyO(%esp) movaps %xmm7,nb201_fizO(%esp) ## update j forces with water O movaps %xmm0,nb201_fjx(%esp) movaps %xmm1,nb201_fjy(%esp) movaps %xmm2,nb201_fjz(%esp) ## H1 interactions movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm6=rinv, xmm4=rinvsq movaps %xmm6,%xmm7 movaps nb201_krsqH1(%esp),%xmm0 addps %xmm0,%xmm6 ## xmm6=rinv+ krsq subps nb201_crf(%esp),%xmm6 ## xmm6=rinv+ krsq-crf mulps nb201_two(%esp),%xmm0 subps %xmm0,%xmm7 ## xmm7=rinv-2*krsq mulps nb201_qqH(%esp),%xmm6 ## vcoul mulps nb201_qqH(%esp),%xmm7 mulps %xmm7,%xmm4 ## total fsH1 in xmm4 addps nb201_vctot(%esp),%xmm6 movaps nb201_dxH1(%esp),%xmm0 movaps nb201_dyH1(%esp),%xmm1 movaps nb201_dzH1(%esp),%xmm2 movaps %xmm6,nb201_vctot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H1 forces movaps nb201_fixH1(%esp),%xmm3 movaps nb201_fiyH1(%esp),%xmm4 movaps nb201_fizH1(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb201_fixH1(%esp) movaps %xmm4,nb201_fiyH1(%esp) movaps %xmm7,nb201_fizH1(%esp) ## update j forces with water H1 addps nb201_fjx(%esp),%xmm0 addps nb201_fjy(%esp),%xmm1 addps nb201_fjz(%esp),%xmm2 movaps %xmm0,nb201_fjx(%esp) movaps %xmm1,nb201_fjy(%esp) movaps %xmm2,nb201_fjz(%esp) ## H2 interactions movaps %xmm5,%xmm4 mulps %xmm4,%xmm4 ## xmm5=rinv, xmm4=rinvsq movaps %xmm5,%xmm7 movaps nb201_krsqH2(%esp),%xmm0 addps %xmm0,%xmm5 ## xmm6=rinv+ krsq subps nb201_crf(%esp),%xmm5 ## xmm5=rinv+ krsq-crf mulps nb201_two(%esp),%xmm0 subps %xmm0,%xmm7 ## xmm7=rinv-2*krsq mulps nb201_qqH(%esp),%xmm5 ## vcoul mulps nb201_qqH(%esp),%xmm7 mulps %xmm7,%xmm4 ## total fsH2 in xmm4 addps nb201_vctot(%esp),%xmm5 movaps nb201_dxH2(%esp),%xmm0 movaps nb201_dyH2(%esp),%xmm1 movaps nb201_dzH2(%esp),%xmm2 movaps %xmm5,nb201_vctot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H2 forces movaps nb201_fixH2(%esp),%xmm3 movaps nb201_fiyH2(%esp),%xmm4 movaps nb201_fizH2(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb201_fixH2(%esp) movaps %xmm4,nb201_fiyH2(%esp) movaps %xmm7,nb201_fizH2(%esp) movl nb201_faction(%ebp),%edi ## update j forces addps nb201_fjx(%esp),%xmm0 addps nb201_fjy(%esp),%xmm1 addps nb201_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,nb201_innerk(%esp) jl _nb_kernel201_ia32_sse.nb201_odd_inner jmp _nb_kernel201_ia32_sse.nb201_unroll_loop_nb_kernel201_ia32_sse.nb201_odd_inner: addl $4,nb201_innerk(%esp) jnz _nb_kernel201_ia32_sse.nb201_odd_loop jmp _nb_kernel201_ia32_sse.nb201_updateouterdata_nb_kernel201_ia32_sse.nb201_odd_loop: movl nb201_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb201_innerjjnr(%esp) xorps %xmm4,%xmm4 movss nb201_iqO(%esp),%xmm4 movl nb201_charge(%ebp),%esi movhps nb201_iqH(%esp),%xmm4 movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb201_qqO(%esp) ## use oxygen qq for storage movl nb201_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 nb201_ixO(%esp),%xmm3 movss nb201_iyO(%esp),%xmm4 movss nb201_izO(%esp),%xmm5 movlps nb201_ixH1(%esp),%xmm6 movlps nb201_ixH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb201_iyH1(%esp),%xmm6 movlps nb201_iyH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb201_izH1(%esp),%xmm6 movlps nb201_izH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb201_dxO(%esp) movaps %xmm4,nb201_dyO(%esp) movaps %xmm5,nb201_dzO(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 movaps %xmm4,%xmm0 mulps nb201_krf(%esp),%xmm0 movaps %xmm0,nb201_krsqO(%esp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb201_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb201_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 %xmm0,%xmm1 ## xmm1=rinv movaps nb201_krsqO(%esp),%xmm3 addps %xmm3,%xmm0 ## xmm0=rinv+ krsq subps nb201_crf(%esp),%xmm0 ## xmm0=rinv+ krsq-crf mulps nb201_two(%esp),%xmm3 subps %xmm3,%xmm1 ## xmm1=rinv-2*krsq mulps nb201_qqO(%esp),%xmm0 ## xmm0=vcoul mulps nb201_qqO(%esp),%xmm1 ## xmm1=coul part of fs mulps %xmm1,%xmm4 ## xmm4=total fscal addps nb201_vctot(%esp),%xmm0 movaps %xmm0,nb201_vctot(%esp) movaps nb201_dxO(%esp),%xmm0 movaps nb201_dyO(%esp),%xmm1 movaps nb201_dzO(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb201_fixO(%esp),%xmm3 movss nb201_fiyO(%esp),%xmm4 movss nb201_fizO(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb201_fixO(%esp) movss %xmm4,nb201_fiyO(%esp) movss %xmm5,nb201_fizO(%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 nb201_fixH1(%esp),%xmm3 addss nb201_fiyH1(%esp),%xmm4 addss nb201_fizH1(%esp),%xmm5 movss %xmm3,nb201_fixH1(%esp) movss %xmm4,nb201_fiyH1(%esp) movss %xmm5,nb201_fizH1(%esp) ## updated the H1 force movl nb201_faction(%ebp),%edi shufps $231,%xmm3,%xmm3 ## constant 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## constant 11100111 shufps $231,%xmm5,%xmm5 ## constant 11100111 addss nb201_fixH2(%esp),%xmm3 addss nb201_fiyH2(%esp),%xmm4 addss nb201_fizH2(%esp),%xmm5 movss %xmm3,nb201_fixH2(%esp) movss %xmm4,nb201_fiyH2(%esp) movss %xmm5,nb201_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)⌨️ 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?