📄 nb_kernel301_ia32_3dnow.s
字号:
pfmul nb301_two(%esp),%mm7 ## two*Heps2 pfadd %mm6,%mm7 pfadd %mm5,%mm7 ## mm7=FF pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb301_qqO(%esp),%mm5 ## vcoul=qq*VV pfmul nb301_qqO(%esp),%mm7 ## fijC=qq*FF ## update vctot directly, use mm3 for fscal sum. pfadd nb301_vctot(%esp),%mm5 movq %mm5,nb301_vctot(%esp) movq %mm7,%mm3 ## change sign of fscal and multiply with rinv pxor %mm0,%mm0 pfsubr %mm0,%mm3 pfmul nb301_tsc(%esp),%mm3 pfmul %mm1,%mm3 ## mm3 is total fscal (for the oxygen) now ## Ready with the oxygen - potential is updated, fscal is in mm3. ## now do the two hydrogens. movq nb301_tmprsqH(%esp),%mm0 ## mm0=rsqH pfrsqrt %mm0,%mm1 pswapd %mm0,%mm0 pfrsqrt %mm0,%mm2 pswapd %mm0,%mm0 punpckldq %mm2,%mm1 ## seeds are in mm1 now, and rsq in mm0. movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=r pfmul nb301_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb301_n1(%esp) pi2fd %mm4,%mm4 pfsub %mm4,%mm0 ## now mm0 is eps and mm4 n0 movq %mm0,%mm2 pfmul %mm2,%mm2 ## mm0 is eps, mm2 eps2 ## coulomb table movl nb301_VFtab(%ebp),%edx movl nb301_n1(%esp),%ecx shll $2,%ecx ## load all values we need movd (%edx,%ecx,4),%mm4 movd 4(%edx,%ecx,4),%mm5 movd 8(%edx,%ecx,4),%mm6 movd 12(%edx,%ecx,4),%mm7 movl nb301_n1+4(%esp),%ecx shll $2,%ecx punpckldq (%edx,%ecx,4),%mm4 punpckldq 4(%edx,%ecx,4),%mm5 punpckldq 8(%edx,%ecx,4),%mm6 punpckldq 12(%edx,%ecx,4),%mm7 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul nb301_two(%esp),%mm7 ## two*Heps2 pfadd %mm6,%mm7 pfadd %mm5,%mm7 ## mm7=FF pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb301_qqH(%esp),%mm5 ## vcoul=qq*VV pfmul nb301_qqH(%esp),%mm7 ## fijC=qq*FF ## update vctot pfadd nb301_vctot(%esp),%mm5 movq %mm5,nb301_vctot(%esp) ## change sign of fijC and multiply by rinv pxor %mm4,%mm4 pfsub %mm7,%mm4 pfmul nb301_tsc(%esp),%mm4 pfmul %mm1,%mm4 ## mm4 is total fscal (for the hydrogens) now ## spread oxygen fscalar to both positions punpckldq %mm3,%mm3 ## calc vectorial force for O prefetchw (%edi,%eax,4) ## prefetch faction to cache movq nb301_dxO(%esp),%mm0 movd nb301_dzO(%esp),%mm1 pfmul %mm3,%mm0 pfmul %mm3,%mm1 ## calc vectorial force for H's movq nb301_dxH(%esp),%mm5 movq nb301_dyH(%esp),%mm6 movq nb301_dzH(%esp),%mm7 pfmul %mm4,%mm5 pfmul %mm4,%mm6 pfmul %mm4,%mm7 ## update iO particle force movq nb301_fixO(%esp),%mm2 movd nb301_fizO(%esp),%mm3 pfadd %mm0,%mm2 pfadd %mm1,%mm3 movq %mm2,nb301_fixO(%esp) movd %mm3,nb301_fizO(%esp) ## update iH forces movq nb301_fixH(%esp),%mm2 movq nb301_fiyH(%esp),%mm3 movq nb301_fizH(%esp),%mm4 pfadd %mm5,%mm2 pfadd %mm6,%mm3 pfadd %mm7,%mm4 movq %mm2,nb301_fixH(%esp) movq %mm3,nb301_fiyH(%esp) movq %mm4,nb301_fizH(%esp) ## pack j forces from H in the same form as the oxygen force. pfacc %mm6,%mm5 ## mm5(l)=fjx(H1+ h2) mm5(h)=fjy(H1+ h2) pfacc %mm7,%mm7 ## mm7(l)=fjz(H1+ h2) pfadd %mm5,%mm0 ## add up total force on j particle. pfadd %mm7,%mm1 ## update j particle force movq (%edi,%eax,4),%mm2 movd 8(%edi,%eax,4),%mm3 pfsub %mm0,%mm2 pfsub %mm1,%mm3 movq %mm2,(%edi,%eax,4) movd %mm3,8(%edi,%eax,4) ## done - one more? decl nb301_innerk(%esp) jz _nb_kernel301_ia32_3dnow.nb301_updateouterdata jmp _nb_kernel301_ia32_3dnow.nb301_inner_loop_nb_kernel301_ia32_3dnow.nb301_updateouterdata: movl nb301_ii3(%esp),%ecx movq (%edi,%ecx,4),%mm6 ## increment iO force movd 8(%edi,%ecx,4),%mm7 pfadd nb301_fixO(%esp),%mm6 pfadd nb301_fizO(%esp),%mm7 movq %mm6,(%edi,%ecx,4) movd %mm7,8(%edi,%ecx,4) movq nb301_fixH(%esp),%mm0 movq nb301_fiyH(%esp),%mm3 movq nb301_fizH(%esp),%mm1 movq %mm0,%mm2 punpckldq %mm3,%mm0 ## mm0(l)=fxH1, mm0(h)=fyH1 punpckhdq %mm3,%mm2 ## mm2(l)=fxH2, mm2(h)=fyH2 movq %mm1,%mm3 pswapd %mm3,%mm3 ## mm1 is fzH1 ## mm3 is fzH2 movq 12(%edi,%ecx,4),%mm6 ## increment iH1 force movd 20(%edi,%ecx,4),%mm7 pfadd %mm0,%mm6 pfadd %mm1,%mm7 movq %mm6,12(%edi,%ecx,4) movd %mm7,20(%edi,%ecx,4) movq 24(%edi,%ecx,4),%mm6 ## increment iH2 force movd 32(%edi,%ecx,4),%mm7 pfadd %mm2,%mm6 pfadd %mm3,%mm7 movq %mm6,24(%edi,%ecx,4) movd %mm7,32(%edi,%ecx,4) movl nb301_fshift(%ebp),%ebx ## increment fshift force movl nb301_is3(%esp),%edx movq (%ebx,%edx,4),%mm6 movd 8(%ebx,%edx,4),%mm7 pfadd nb301_fixO(%esp),%mm6 pfadd nb301_fizO(%esp),%mm7 pfadd %mm0,%mm6 pfadd %mm1,%mm7 pfadd %mm2,%mm6 pfadd %mm3,%mm7 movq %mm6,(%ebx,%edx,4) movd %mm7,8(%ebx,%edx,4) ## get n from stack movl nb301_n(%esp),%esi ## get group index for i particle movl nb301_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] movq nb301_vctot(%esp),%mm7 pfacc %mm7,%mm7 ## get and sum the two parts of total potential movl nb301_Vc(%ebp),%eax movd (%eax,%edx,4),%mm6 pfadd %mm7,%mm6 movd %mm6,(%eax,%edx,4) ## increment vc[gid] ## finish if last movl nb301_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel301_ia32_3dnow.nb301_outerend ## not last, iterate outer loop once more! movl %esi,nb301_n(%esp) jmp _nb_kernel301_ia32_3dnow.nb301_outer_nb_kernel301_ia32_3dnow.nb301_outerend: ## check if more outer neighborlists remain movl nb301_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel301_ia32_3dnow.nb301_end ## non-zero, do one more workunit jmp _nb_kernel301_ia32_3dnow.nb301_threadloop_nb_kernel301_ia32_3dnow.nb301_end: femms movl nb301_nouter(%esp),%eax movl nb301_ninner(%esp),%ebx movl nb301_outeriter(%ebp),%ecx movl nb301_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) addl $216,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret.globl nb_kernel301nf_ia32_3dnow.globl _nb_kernel301nf_ia32_3dnownb_kernel301nf_ia32_3dnow: _nb_kernel301nf_ia32_3dnow: .set nb301nf_p_nri, 8.set nb301nf_iinr, 12.set nb301nf_jindex, 16.set nb301nf_jjnr, 20.set nb301nf_shift, 24.set nb301nf_shiftvec, 28.set nb301nf_fshift, 32.set nb301nf_gid, 36.set nb301nf_pos, 40.set nb301nf_faction, 44.set nb301nf_charge, 48.set nb301nf_p_facel, 52.set nb301nf_p_krf, 56.set nb301nf_p_crf, 60.set nb301nf_Vc, 64.set nb301nf_type, 68.set nb301nf_p_ntype, 72.set nb301nf_vdwparam, 76.set nb301nf_Vvdw, 80.set nb301nf_p_tabscale, 84.set nb301nf_VFtab, 88.set nb301nf_invsqrta, 92.set nb301nf_dvda, 96.set nb301nf_p_gbtabscale, 100.set nb301nf_GBtab, 104.set nb301nf_p_nthreads, 108.set nb301nf_count, 112.set nb301nf_mtx, 116.set nb301nf_outeriter, 120.set nb301nf_inneriter, 124.set nb301nf_work, 128 ## stack offsets for local variables .set nb301nf_is3, 0.set nb301nf_ii3, 4.set nb301nf_ixO, 8.set nb301nf_iyO, 12.set nb301nf_izO, 16.set nb301nf_ixH, 20.set nb301nf_iyH, 28.set nb301nf_izH, 36.set nb301nf_iqO, 44.set nb301nf_iqH, 52.set nb301nf_qqO, 60.set nb301nf_qqH, 68.set nb301nf_vctot, 76.set nb301nf_n1, 84.set nb301nf_tsc, 92.set nb301nf_innerjjnr, 100.set nb301nf_innerk, 104.set nb301nf_tmprsqH, 108.set nb301nf_n, 116 ## idx for outer loop.set nb301nf_nn1, 120 ## number of outer iterations.set nb301nf_nri, 124.set nb301nf_nouter, 128.set nb301nf_ninner, 132 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $136,%esp ## local stack space femms movl nb301nf_p_nri(%ebp),%ecx movl nb301nf_p_facel(%ebp),%esi movl nb301nf_p_tabscale(%ebp),%edi movl (%ecx),%ecx movl %ecx,nb301nf_nri(%esp) ## zero iteration counters
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -