nb_kernel301_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,967 行 · 第 1/5 页
S
1,967 行
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 nb301_ixH2(%esp),%xmm3 movaps nb301_iyH2(%esp),%xmm4 movaps nb301_izH2(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## store dr movaps %xmm3,nb301_dxH2(%esp) movaps %xmm4,nb301_dyH2(%esp) movaps %xmm5,nb301_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 ## start with rsqO - seed to xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb301_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 nb301_half(%esp),%xmm4 movaps %xmm4,nb301_rinvO(%esp) ## rinvO in xmm4 mulps %xmm4,%xmm7 movaps %xmm7,nb301_rO(%esp) ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb301_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 nb301_half(%esp),%xmm4 movaps %xmm4,nb301_rinvH1(%esp) ## rinvH1 in xmm4 mulps %xmm4,%xmm6 movaps %xmm6,nb301_rH1(%esp) ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb301_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 nb301_half(%esp),%xmm4 movaps %xmm4,nb301_rinvH2(%esp) ## rinvH2 in xmm4 mulps %xmm4,%xmm5 movaps %xmm5,nb301_rH2(%esp) ## do O interactions ## rO is still in xmm7 mulps nb301_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb301_VFtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm5 movhps (%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%ebx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## 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 nb301_two(%esp),%xmm7 ## two*Heps2 movaps nb301_qqO(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm0 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm0 fijC ## increment vcoul - then we can get rid of mm5 addps nb301_vctot(%esp),%xmm5 movaps %xmm5,nb301_vctot(%esp) xorps %xmm4,%xmm4 mulps nb301_tsc(%esp),%xmm0 mulps nb301_rinvO(%esp),%xmm0 subps %xmm0,%xmm4 movaps nb301_dxO(%esp),%xmm0 movaps nb301_dyO(%esp),%xmm1 movaps nb301_dzO(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## tx in xmm0-xmm2 ## update O forces movaps nb301_fixO(%esp),%xmm3 movaps nb301_fiyO(%esp),%xmm4 movaps nb301_fizO(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb301_fixO(%esp) movaps %xmm4,nb301_fiyO(%esp) movaps %xmm7,nb301_fizO(%esp) ## update j forces with water O movaps %xmm0,nb301_fjx(%esp) movaps %xmm1,nb301_fjy(%esp) movaps %xmm2,nb301_fjz(%esp) ## Done with O interactions - now H1! movaps nb301_rH1(%esp),%xmm7 mulps nb301_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm5 movhps (%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%ebx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## 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 nb301_two(%esp),%xmm7 ## two*Heps2 movaps nb301_qqH(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm0,%xmm7 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm7 fijC ## increment vcoul xorps %xmm4,%xmm4 addps nb301_vctot(%esp),%xmm5 mulps nb301_rinvH1(%esp),%xmm7 movaps %xmm5,nb301_vctot(%esp) mulps nb301_tsc(%esp),%xmm7 subps %xmm7,%xmm4 movaps nb301_dxH1(%esp),%xmm0 movaps nb301_dyH1(%esp),%xmm1 movaps nb301_dzH1(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H1 forces movaps nb301_fixH1(%esp),%xmm3 movaps nb301_fiyH1(%esp),%xmm4 movaps nb301_fizH1(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb301_fixH1(%esp) movaps %xmm4,nb301_fiyH1(%esp) movaps %xmm7,nb301_fizH1(%esp) ## update j forces with water H1 addps nb301_fjx(%esp),%xmm0 addps nb301_fjy(%esp),%xmm1 addps nb301_fjz(%esp),%xmm2 movaps %xmm0,nb301_fjx(%esp) movaps %xmm1,nb301_fjy(%esp) movaps %xmm2,nb301_fjz(%esp) ## Done with H1, finally we do H2 interactions movaps nb301_rH2(%esp),%xmm7 mulps nb301_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm5 movhps (%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%ebx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## 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 nb301_two(%esp),%xmm7 ## two*Heps2 movaps nb301_qqH(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm0,%xmm7 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm0 fijC ## increment vcoul xorps %xmm4,%xmm4 addps nb301_vctot(%esp),%xmm5 mulps nb301_rinvH2(%esp),%xmm7 movaps %xmm5,nb301_vctot(%esp) mulps nb301_tsc(%esp),%xmm7 subps %xmm7,%xmm4 movaps nb301_dxH2(%esp),%xmm0 movaps nb301_dyH2(%esp),%xmm1 movaps nb301_dzH2(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx ## update H2 forces movaps nb301_fixH2(%esp),%xmm3 movaps nb301_fiyH2(%esp),%xmm4 movaps nb301_fizH2(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb301_fixH2(%esp) movaps %xmm4,nb301_fiyH2(%esp) movaps %xmm7,nb301_fizH2(%esp) movl nb301_faction(%ebp),%edi ## update j forces addps nb301_fjx(%esp),%xmm0 addps nb301_fjy(%esp),%xmm1 addps nb301_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,nb301_innerk(%esp) jl _nb_kernel301_ia32_sse.nb301_odd_inner jmp _nb_kernel301_ia32_sse.nb301_unroll_loop_nb_kernel301_ia32_sse.nb301_odd_inner: addl $4,nb301_innerk(%esp) jnz _nb_kernel301_ia32_sse.nb301_odd_loop jmp _nb_kernel301_ia32_sse.nb301_updateouterdata_nb_kernel301_ia32_sse.nb301_odd_loop: movl nb301_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb301_innerjjnr(%esp)⌨️ 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?