nb_kernel400_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,703 行 · 第 1/4 页
S
1,703 行
movl nb400_jnrd(%esp),%edx movl nb400_dvda(%ebp),%esi ## Calculate dVda xorps %xmm7,%xmm7 mulps nb400_gbscale(%esp),%xmm3 movaps %xmm3,%xmm6 mulps nb400_r(%esp),%xmm6 addps %xmm5,%xmm6 addps nb400_vctot(%esp),%xmm5 movaps %xmm5,nb400_vctot(%esp) ## xmm6=(vcoul+fijC*r) subps %xmm6,%xmm7 movaps %xmm7,%xmm6 ## update dvdasum addps nb400_dvdasum(%esp),%xmm7 movaps %xmm7,nb400_dvdasum(%esp) ## update j atoms dvdaj movhlps %xmm6,%xmm7 movaps %xmm6,%xmm5 movaps %xmm7,%xmm4 shufps $0x1,%xmm5,%xmm5 shufps $0x1,%xmm4,%xmm4 ## xmm6=dvdaj1 xmm5=dvdaj2 xmm7=dvdaj3 xmm4=dvdaj4 addss (%esi,%eax,4),%xmm6 addss (%esi,%ebx,4),%xmm5 addss (%esi,%ecx,4),%xmm7 addss (%esi,%edx,4),%xmm4 movss %xmm6,(%esi,%eax,4) movss %xmm5,(%esi,%ebx,4) movss %xmm7,(%esi,%ecx,4) movss %xmm4,(%esi,%edx,4) xorps %xmm4,%xmm4 mulps %xmm0,%xmm3 subps %xmm3,%xmm4 movaps nb400_dx(%esp),%xmm0 movaps nb400_dy(%esp),%xmm1 movaps nb400_dz(%esp),%xmm2 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movl nb400_faction(%ebp),%edi mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movaps nb400_fix(%esp),%xmm3 movaps nb400_fiy(%esp),%xmm4 movaps nb400_fiz(%esp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb400_fix(%esp) movaps %xmm4,nb400_fiy(%esp) movaps %xmm5,nb400_fiz(%esp) ## the fj's - start by accumulating x & y forces from memory movlps (%edi,%eax,4),%xmm4 movlps (%edi,%ecx,4),%xmm6 movhps (%edi,%ebx,4),%xmm4 movhps (%edi,%edx,4),%xmm6 movaps %xmm4,%xmm3 shufps $136,%xmm6,%xmm3 ## constant 10001000 shufps $221,%xmm6,%xmm4 ## constant 11011101 ## now xmm3-xmm5 contains fjx, fjy, fjz subps %xmm0,%xmm3 subps %xmm1,%xmm4 ## unpack them back so we can store them - first x & y in xmm3/xmm4 movaps %xmm3,%xmm6 unpcklps %xmm4,%xmm6 unpckhps %xmm4,%xmm3 ## xmm6(l)=x & y for j1, (h) for j2 ## xmm3(l)=x & y for j3, (h) for j4 movlps %xmm6,(%edi,%eax,4) movlps %xmm3,(%edi,%ecx,4) movhps %xmm6,(%edi,%ebx,4) movhps %xmm3,(%edi,%edx,4) ## and the z forces movss 8(%edi,%eax,4),%xmm4 movss 8(%edi,%ebx,4),%xmm5 movss 8(%edi,%ecx,4),%xmm6 movss 8(%edi,%edx,4),%xmm7 subss %xmm2,%xmm4 shufps $229,%xmm2,%xmm2 ## constant 11100101 subss %xmm2,%xmm5 shufps $234,%xmm2,%xmm2 ## constant 11101010 subss %xmm2,%xmm6 shufps $255,%xmm2,%xmm2 ## constant 11111111 subss %xmm2,%xmm7 movss %xmm4,8(%edi,%eax,4) movss %xmm5,8(%edi,%ebx,4) movss %xmm6,8(%edi,%ecx,4) movss %xmm7,8(%edi,%edx,4) ## should we do one more iteration? subl $4,nb400_innerk(%esp) jl _nb_kernel400_ia32_sse.nb400_finish_inner jmp _nb_kernel400_ia32_sse.nb400_unroll_loop_nb_kernel400_ia32_sse.nb400_finish_inner: ## check if at least two particles remain addl $4,nb400_innerk(%esp) movl nb400_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel400_ia32_sse.nb400_dopair jmp _nb_kernel400_ia32_sse.nb400_checksingle_nb_kernel400_ia32_sse.nb400_dopair: movl nb400_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb400_innerjjnr(%esp) xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 ## load isaj movl nb400_invsqrta(%ebp),%esi movss (%esi,%eax,4),%xmm2 movss (%esi,%ebx,4),%xmm3 unpcklps %xmm3,%xmm2 ## isaj in xmm2(0,1) mulps nb400_isai(%esp),%xmm2 movaps %xmm2,nb400_isaprod(%esp) movaps %xmm2,%xmm1 mulps nb400_gbtsc(%esp),%xmm1 movaps %xmm1,nb400_gbscale(%esp) movl nb400_charge(%ebp),%esi ## base of charge[] movss (%esi,%eax,4),%xmm3 movss (%esi,%ebx,4),%xmm6 unpcklps %xmm6,%xmm3 ## constant 00001000 ;# xmm3(0,1) has the charges mulps nb400_iq(%esp),%xmm2 mulps %xmm2,%xmm3 movaps %xmm3,nb400_qq(%esp) movl nb400_pos(%ebp),%edi movd %eax,%mm0 ## copy jnr to mm0/mm1 movd %ebx,%mm1 leal (%eax,%eax,2),%eax leal (%ebx,%ebx,2),%ebx ## move coordinates to xmm0-xmm2 movlps (%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 movhps (%edi,%ebx,4),%xmm1 movss 8(%edi,%ebx,4),%xmm0 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## constant 10001000 shufps $136,%xmm0,%xmm0 ## constant 10001000 shufps $221,%xmm1,%xmm1 ## constant 11011101 movl nb400_faction(%ebp),%edi ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb400_ix(%esp),%xmm4 movaps nb400_iy(%esp),%xmm5 movaps nb400_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb400_dx(%esp) movaps %xmm5,nb400_dy(%esp) movaps %xmm6,nb400_dz(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb400_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb400_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm4,nb400_r(%esp) mulps nb400_gbscale(%esp),%xmm4 cvttps2pi %xmm4,%mm6 ## mm6 contain lu indices cvtpi2ps %mm6,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 movl nb400_GBtab(%ebp),%esi movd %mm6,%ecx psrlq $32,%mm6 movd %mm6,%edx ## load coulomb table movaps (%esi,%ecx,4),%xmm4 movaps (%esi,%edx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm4,%xmm6 unpcklps %xmm7,%xmm4 ## Y1 Y2 F1 F2 unpckhps %xmm7,%xmm6 ## G1 G2 H1 H2 movhlps %xmm4,%xmm5 ## F1 F2 movhlps %xmm6,%xmm7 ## H1 H2 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb400_two(%esp),%xmm7 ## two*Heps2 movaps nb400_qq(%esp),%xmm3 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm3 ## fijC=FF*qq ## at this point mm5 contains vcoul and mm3 fijC ## get jnr from mm0/mm1 movd %mm0,%ecx movd %mm1,%edx movl nb400_dvda(%ebp),%esi ## Calculate dVda xorps %xmm7,%xmm7 mulps nb400_gbscale(%esp),%xmm3 movaps %xmm3,%xmm6 mulps nb400_r(%esp),%xmm6 addps %xmm5,%xmm6 addps nb400_vctot(%esp),%xmm5 movaps %xmm5,nb400_vctot(%esp) ## xmm6=(vcoul+fijC*r) subps %xmm6,%xmm7 movaps %xmm7,%xmm6 ## update dvdasum addps nb400_dvdasum(%esp),%xmm7 movaps %xmm7,nb400_dvdasum(%esp) ## update j atoms dvdaj movaps %xmm6,%xmm7 shufps $0x1,%xmm7,%xmm7 addss (%esi,%ecx,4),%xmm6 addss (%esi,%edx,4),%xmm7 movss %xmm6,(%esi,%ecx,4) movss %xmm7,(%esi,%edx,4) xorps %xmm4,%xmm4 mulps %xmm0,%xmm3 subps %xmm3,%xmm4 movaps nb400_dx(%esp),%xmm0 movaps nb400_dy(%esp),%xmm1 movaps nb400_dz(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movaps nb400_fix(%esp),%xmm3 movaps nb400_fiy(%esp),%xmm4 movaps nb400_fiz(%esp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb400_fix(%esp) movaps %xmm4,nb400_fiy(%esp) movaps %xmm5,nb400_fiz(%esp) ## update the fj's movss (%edi,%eax,4),%xmm3 movss 4(%edi,%eax,4),%xmm4 movss 8(%edi,%eax,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%edi,%eax,4) movss %xmm4,4(%edi,%eax,4) movss %xmm5,8(%edi,%eax,4) shufps $225,%xmm0,%xmm0 ## constant 11100001 shufps $225,%xmm1,%xmm1 ## constant 11100001 shufps $225,%xmm2,%xmm2 ## constant 11100001 movss (%edi,%ebx,4),%xmm3 movss 4(%edi,%ebx,4),%xmm4 movss 8(%edi,%ebx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%edi,%ebx,4) movss %xmm4,4(%edi,%ebx,4) movss %xmm5,8(%edi,%ebx,4)_nb_kernel400_ia32_sse.nb400_checksingle: movl nb400_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel400_ia32_sse.nb400_dosingle jmp _nb_kernel400_ia32_sse.nb400_updateouterdata_nb_kernel400_ia32_sse.nb400_dosingle: movl nb400_charge(%ebp),%esi movl nb400_invsqrta(%ebp),%edx movl nb400_pos(%ebp),%edi movl nb400_innerjjnr(%esp),%ecx movl (%ecx),%eax xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 movss (%edx,%eax,4),%xmm2 ## isaj mulss nb400_isai(%esp),%xmm2 movss %xmm2,nb400_isaprod(%esp) movss %xmm2,%xmm1 mulss nb400_gbtsc(%esp),%xmm1 movss %xmm1,nb400_gbscale(%esp) mulss nb400_iq(%esp),%xmm2 movss (%esi,%eax,4),%xmm6 ## xmm6(0) has the charge mulss %xmm2,%xmm6 movss %xmm6,nb400_qq(%esp) movd %eax,%mm0 leal (%eax,%eax,2),%eax ## move coordinates to xmm0-xmm2 movss (%edi,%eax,4),%xmm0 movss 4(%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 movss nb400_ix(%esp),%xmm4 movss nb400_iy(%esp),%xmm5 movss nb400_iz(%esp),%xmm6 ## calc dr subss %xmm0,%xmm4 subss %xmm1,%xmm5 subss %xmm2,%xmm6 ## store dr movss %xmm4,nb400_dx(%esp) movss %xmm5,nb400_dy(%esp) movss %xmm6,nb400_dz(%esp) ## square it mulss %xmm4,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm6 addss %xmm5,%xmm4 addss %xmm6,%xmm4 ## rsq in xmm4 rsqrtss %xmm4,%xmm5 ## lookup seed in xmm5 movss %xmm5,%xmm2 mulss %xmm5,%xmm5 movss nb400_three(%esp),%xmm1 mulss %xmm4,%xmm5 ## rsq*lu*lu movss nb400_half(%esp),%xmm0 subss %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulss %xmm2,%xmm1 mulss %xmm1,%xmm0 ## xmm0=rinv mulss %xmm0,%xmm4 ## xmm4=r movss %xmm4,nb400_r(%esp) mulss nb400_gbscale(%esp),%xmm4 cvttss2si %xmm4,%ebx ## mm6 contain lu indices cvtsi2ss %ebx,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 shll $2,%ebx movl nb400_GBtab(%ebp),%esi movaps (%esi,%ebx,4),%xmm4 movhlps %xmm4,%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $1,%xmm5,%xmm5 shufps $1,%xmm7,%xmm7 ## table ready in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb400_two(%esp),%xmm7 ## two*Heps2 movss nb400_qq(%esp),%xmm3 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF ⌨️ 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?