nb_kernel133_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,066 行 · 第 1/5 页
S
2,066 行
movss %xmm4,8(%edi,%edx,4) ## should we do one more iteration? subl $4,nb133_innerk(%esp) jl _nb_kernel133_ia32_sse.nb133_odd_inner jmp _nb_kernel133_ia32_sse.nb133_unroll_loop_nb_kernel133_ia32_sse.nb133_odd_inner: addl $4,nb133_innerk(%esp) jnz _nb_kernel133_ia32_sse.nb133_odd_loop jmp _nb_kernel133_ia32_sse.nb133_updateouterdata_nb_kernel133_ia32_sse.nb133_odd_loop: movl nb133_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb133_innerjjnr(%esp) xorps %xmm4,%xmm4 ## clear reg. movss nb133_iqM(%esp),%xmm4 movl nb133_charge(%ebp),%esi movhps nb133_iqH(%esp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb133_qqM(%esp) ## use dummy qq for storage xorps %xmm6,%xmm6 movl nb133_type(%ebp),%esi movl (%esi,%eax,4),%ebx movl nb133_vdwparam(%ebp),%esi shll %ebx addl nb133_ntia(%esp),%ebx movlps (%esi,%ebx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb133_c6(%esp) movaps %xmm7,nb133_c12(%esp) movl nb133_pos(%ebp),%esi leal (%eax,%eax,2),%eax movss nb133_ixO(%esp),%xmm3 movss nb133_iyO(%esp),%xmm4 movss nb133_izO(%esp),%xmm5 movss nb133_ixH1(%esp),%xmm0 movss nb133_iyH1(%esp),%xmm1 movss nb133_izH1(%esp),%xmm2 unpcklps nb133_ixH2(%esp),%xmm3 ## ixO ixH2 - - unpcklps nb133_iyH2(%esp),%xmm4 ## iyO iyH2 - - unpcklps nb133_izH2(%esp),%xmm5 ## izO izH2 - - unpcklps nb133_ixM(%esp),%xmm0 ## ixH1 ixM - - unpcklps nb133_iyM(%esp),%xmm1 ## iyH1 iyM - - unpcklps nb133_izM(%esp),%xmm2 ## izH1 izM - - unpcklps %xmm0,%xmm3 ## ixO ixH1 ixH2 ixM unpcklps %xmm1,%xmm4 ## same for y unpcklps %xmm2,%xmm5 ## same for z ## 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 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use O distances for storage movaps %xmm3,nb133_dxO(%esp) movaps %xmm4,nb133_dyO(%esp) movaps %xmm5,nb133_dzO(%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 nb133_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb133_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv, xmm4=rsq mulps %xmm0,%xmm4 mulps nb133_tsc(%esp),%xmm4 ## rtab cvttps2pi %xmm4,%mm6 cvtpi2ps %mm6,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 movd %eax,%mm0 movl nb133_VFtab(%ebp),%esi movd %mm6,%eax ## dispersion movlps (%esi,%eax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb133_two(%esp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb133_c6(%esp),%xmm4 mulss %xmm4,%xmm7 ## fijD mulss %xmm4,%xmm5 ## Vvdw6 mulss nb133_tsc(%esp),%xmm7 ## put scalar force on stack Update Vvdwtot directly addss nb133_Vvdwtot(%esp),%xmm5 movss %xmm7,nb133_fstmp(%esp) movss %xmm5,nb133_Vvdwtot(%esp) ## repulsion movlps 16(%esi,%eax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%esi,%eax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb133_two(%esp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb133_c12(%esp),%xmm4 mulss %xmm4,%xmm7 ## fijR mulss %xmm4,%xmm5 ## Vvdw12 mulss nb133_tsc(%esp),%xmm7 addss nb133_fstmp(%esp),%xmm7 movss %xmm7,nb133_fstmp(%esp) addss nb133_Vvdwtot(%esp),%xmm5 movss %xmm5,nb133_Vvdwtot(%esp) movd %mm0,%eax movaps %xmm0,%xmm4 mulps nb133_qqM(%esp),%xmm4 movaps %xmm4,%xmm2 mulps %xmm0,%xmm4 subss nb133_fstmp(%esp),%xmm4 mulps %xmm0,%xmm4 addps nb133_vctot(%esp),%xmm2 movaps %xmm2,nb133_vctot(%esp) movaps nb133_dxO(%esp),%xmm0 movaps nb133_dyO(%esp),%xmm1 movaps nb133_dzO(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb133_fixO(%esp),%xmm3 movss nb133_fiyO(%esp),%xmm4 movss nb133_fizO(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb133_fixO(%esp) movss %xmm4,nb133_fiyO(%esp) movss %xmm5,nb133_fizO(%esp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb133_fixH1(%esp),%xmm3 addss nb133_fiyH1(%esp),%xmm4 addss nb133_fizH1(%esp),%xmm5 movss %xmm3,nb133_fixH1(%esp) movss %xmm4,nb133_fiyH1(%esp) movss %xmm5,nb133_fizH1(%esp) ## updated the H1 force shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb133_fixH2(%esp),%xmm3 addss nb133_fiyH2(%esp),%xmm4 addss nb133_fizH2(%esp),%xmm5 movss %xmm3,nb133_fixH2(%esp) movss %xmm4,nb133_fiyH2(%esp) movss %xmm5,nb133_fizH2(%esp) ## updated the H2 force movl nb133_faction(%ebp),%edi shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb133_fixM(%esp),%xmm3 addss nb133_fiyM(%esp),%xmm4 addss nb133_fizM(%esp),%xmm5 movss %xmm3,nb133_fixM(%esp) movss %xmm4,nb133_fiyM(%esp) movss %xmm5,nb133_fizM(%esp) ## updated the M force ## the fj's - move in from mem start by acc. tx/ty/tz in xmm0, xmm1 movlps (%edi,%eax,4),%xmm6 movss 8(%edi,%eax,4),%xmm7 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,%xmm0 movaps %xmm4,%xmm1 movaps %xmm5,%xmm2 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 unpcklps %xmm1,%xmm0 ## x,y sum in xmm0, z sum in xmm2 subps %xmm0,%xmm6 subss %xmm2,%xmm7 movlps %xmm6,(%edi,%eax,4) movss %xmm7,8(%edi,%eax,4) decl nb133_innerk(%esp) jz _nb_kernel133_ia32_sse.nb133_updateouterdata jmp _nb_kernel133_ia32_sse.nb133_odd_loop_nb_kernel133_ia32_sse.nb133_updateouterdata: movl nb133_ii3(%esp),%ecx movl nb133_faction(%ebp),%edi movl nb133_fshift(%ebp),%esi movl nb133_is3(%esp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb133_fixO(%esp),%xmm0 movaps nb133_fiyO(%esp),%xmm1 movaps nb133_fizO(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss (%edi,%ecx,4),%xmm3 movss 4(%edi,%ecx,4),%xmm4 movss 8(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,(%edi,%ecx,4) movss %xmm4,4(%edi,%ecx,4) movss %xmm5,8(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## constant 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb133_fixH1(%esp),%xmm0 movaps nb133_fiyH1(%esp),%xmm1 movaps nb133_fizH1(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 12(%edi,%ecx,4),%xmm3 movss 16(%edi,%ecx,4),%xmm4 movss 20(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,12(%edi,%ecx,4) movss %xmm4,16(%edi,%ecx,4) movss %xmm5,20(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb133_fixH2(%esp),%xmm0 movaps nb133_fiyH2(%esp),%xmm1 movaps nb133_fizH2(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 24(%edi,%ecx,4),%xmm3 movss 28(%edi,%ecx,4),%xmm4 movss 32(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,24(%edi,%ecx,4) movss %xmm4,28(%edi,%ecx,4) movss %xmm5,32(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate Mi forces in xmm0, xmm1, xmm2 movaps nb133_fixM(%esp),%xmm0 movaps nb133_fiyM(%esp),%xmm1 movaps nb133_fizM(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ⌨️ 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?