📄 nb_kernel311_ia32_3dnow.s
字号:
movl nb311nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx=ii movl nb311nf_charge(%ebp),%edx movd (%edx,%ebx,4),%mm2 ## mm2=charge[ii0] pfmul %mm1,%mm2 movq %mm2,nb311nf_iqO(%esp) ## iqO = facel*charge[ii] movd 4(%edx,%ebx,4),%mm2 ## mm2=charge[ii0+1] pfmul %mm1,%mm2 punpckldq %mm2,%mm2 ## spread to both halves movq %mm2,nb311nf_iqH(%esp) ## iqH = facel*charge[ii0+1] movl nb311nf_type(%ebp),%edx movl (%edx,%ebx,4),%edx shll %edx movl %edx,%ecx imull (%esi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb311nf_ntia(%esp) movq (%edi),%mm6 punpckldq %mm6,%mm6 ## spread to both halves movq %mm6,nb311nf_tsc(%esp)_nb_kernel311nf_ia32_3dnow.nb311nf_threadloop: movl nb311nf_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel311nf_ia32_3dnow.nb311nf_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%esi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel311nf_ia32_3dnow.nb311nf_spinlock ## if(nn1>nri) nn1=nri movl nb311nf_nri(%esp),%ecx movl %ecx,%edx subl %ebx,%ecx cmovlel %edx,%ebx ## if(nn1>nri) nn1=nri ## Cleared the spinlock if we got here. ## eax contains nn0, ebx contains nn1. movl %eax,nb311nf_n(%esp) movl %ebx,nb311nf_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel311nf_ia32_3dnow.nb311nf_outerstart jmp _nb_kernel311nf_ia32_3dnow.nb311nf_end_nb_kernel311nf_ia32_3dnow.nb311nf_outerstart: ## ebx contains number of outer iterations addl nb311nf_nouter(%esp),%ebx movl %ebx,nb311nf_nouter(%esp)_nb_kernel311nf_ia32_3dnow.nb311nf_outer: movl nb311nf_shift(%ebp),%eax ## eax = pointer into shift[] movl (%eax,%esi,4),%ebx ## ebx=shift[n] leal (%ebx,%ebx,2),%ebx ## ebx=3*is movl %ebx,nb311nf_is3(%esp) ## store is3 movl nb311nf_shiftvec(%ebp),%eax ## eax = base of shiftvec[] movq (%eax,%ebx,4),%mm5 ## move shX/shY to mm5 and shZ to mm6. movd 8(%eax,%ebx,4),%mm6 movq %mm5,%mm0 movq %mm5,%mm1 movq %mm6,%mm2 punpckldq %mm0,%mm0 ## also expand shX,Y,Z in mm0--mm2. punpckhdq %mm1,%mm1 punpckldq %mm2,%mm2 movl nb311nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx=ii leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb311nf_pos(%ebp),%eax ## eax = base of pos[] pfadd (%eax,%ebx,4),%mm5 ## ix = shX + posX (and iy too) movd 8(%eax,%ebx,4),%mm7 ## cant use direct memory add for 4 bytes (iz) movl %ebx,nb311nf_ii3(%esp) ## (use mm7 as temp. storage for iz.) pfadd %mm7,%mm6 movq %mm5,nb311nf_ixO(%esp) movq %mm6,nb311nf_izO(%esp) movd 12(%eax,%ebx,4),%mm3 movd 16(%eax,%ebx,4),%mm4 movd 20(%eax,%ebx,4),%mm5 punpckldq 24(%eax,%ebx,4),%mm3 punpckldq 28(%eax,%ebx,4),%mm4 punpckldq 32(%eax,%ebx,4),%mm5 ## coords of H1 in low mm3-mm5, H2 in high pfadd %mm3,%mm0 pfadd %mm4,%mm1 pfadd %mm5,%mm2 movq %mm0,nb311nf_ixH(%esp) movq %mm1,nb311nf_iyH(%esp) movq %mm2,nb311nf_izH(%esp) ## clear vctot and i forces pxor %mm7,%mm7 movq %mm7,nb311nf_vctot(%esp) movq %mm7,nb311nf_Vvdwtot(%esp) movl nb311nf_jindex(%ebp),%eax movl (%eax,%esi,4),%ecx ## jindex[n] movl 4(%eax,%esi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movl %edx,nb311nf_innerk(%esp) addl nb311nf_ninner(%esp),%edx movl %edx,nb311nf_ninner(%esp) movl nb311nf_pos(%ebp),%esi movl nb311nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb311nf_innerjjnr(%esp) ## pointer to jjnr[nj0] _nb_kernel311nf_ia32_3dnow.nb311nf_inner_loop: ## a single j particle iteration here - compare with the unrolled code for comments. movl nb311nf_innerjjnr(%esp),%eax movl (%eax),%eax ## eax=jnr offset addl $4,nb311nf_innerjjnr(%esp) ## advance pointer prefetch 16(%ecx) ## prefetch data - trial and error says 16 is best movl nb311nf_charge(%ebp),%ecx movd (%ecx,%eax,4),%mm7 punpckldq %mm7,%mm7 movq %mm7,%mm6 pfmul nb311nf_iqO(%esp),%mm6 pfmul nb311nf_iqH(%esp),%mm7 ## mm6=qqO, mm7=qqH movd %mm6,nb311nf_qqO(%esp) movq %mm7,nb311nf_qqH(%esp) movl nb311nf_type(%ebp),%ecx movl (%ecx,%eax,4),%edx ## type [jnr] movl nb311nf_vdwparam(%ebp),%ecx shll %edx addl nb311nf_ntia(%esp),%edx ## tja = ntia + 2*type movd (%ecx,%edx,4),%mm5 ## mm5 = 1st c6 movq %mm5,nb311nf_c6(%esp) movd 4(%ecx,%edx,4),%mm5 ## mm5 = 1st c12 movq %mm5,nb311nf_c12(%esp) leal (%eax,%eax,2),%eax movq (%esi,%eax,4),%mm0 movd 8(%esi,%eax,4),%mm1 ## copy & expand to mm2-mm4 for the H interactions movq %mm0,%mm2 movq %mm0,%mm3 movq %mm1,%mm4 punpckldq %mm2,%mm2 punpckhdq %mm3,%mm3 punpckldq %mm4,%mm4 pfsubr nb311nf_ixO(%esp),%mm0 pfsubr nb311nf_izO(%esp),%mm1 pfmul %mm0,%mm0 pfmul %mm1,%mm1 pfacc %mm1,%mm0 pfadd %mm1,%mm0 ## mm0=rsqO punpckldq %mm2,%mm2 punpckldq %mm3,%mm3 punpckldq %mm4,%mm4 ## mm2-mm4 is jx-jz pfsubr nb311nf_ixH(%esp),%mm2 pfsubr nb311nf_iyH(%esp),%mm3 pfsubr nb311nf_izH(%esp),%mm4 ## mm2-mm4 is dxH-dzH pfmul %mm2,%mm2 pfmul %mm3,%mm3 pfmul %mm4,%mm4 pfadd %mm2,%mm3 pfadd %mm4,%mm3 ## mm3=rsqH movq %mm3,nb311nf_tmprsqH(%esp) pfrsqrt %mm0,%mm1 movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=r pfmul nb311nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movd %mm4,nb311nf_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 nb311nf_VFtab(%ebp),%edx movl nb311nf_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 pfmul %mm0,%mm6 ## mm6 = Geps pfmul %mm2,%mm7 ## mm7 = Heps2 pfadd %mm6,%mm5 pfadd %mm7,%mm5 ## mm5 = Fp pfmul %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb311nf_qqO(%esp),%mm5 ## vcoul=qq*VV ## update vctot directly pfadd nb311nf_vctot(%esp),%mm5 movq %mm5,nb311nf_vctot(%esp) ## nontabulated LJ - mm1 is invsqrt. - keep mm1! movq %mm1,%mm0 pfmul %mm0,%mm0 ## mm0 is invsq movq %mm0,%mm2 pfmul %mm0,%mm2 pfmul %mm0,%mm2 ## mm2 = rinvsix movq %mm2,%mm4 pfmul %mm4,%mm4 ## mm4=rinvtwelve pfmul nb311nf_c12(%esp),%mm4 pfmul nb311nf_c6(%esp),%mm2 pfsub %mm2,%mm4 ## mm4=Vvdw12-Vvdw6 ## update Vvdwtot pfadd nb311nf_Vvdwtot(%esp),%mm4 ## add the earlier value movq %mm4,nb311nf_Vvdwtot(%esp) ## store the sum ## now do the two hydrogens. movq nb311nf_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 nb311nf_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb311nf_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 nb311nf_VFtab(%ebp),%edx movl nb311nf_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 nb311nf_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 %mm0,%mm5 ## mm5=eps*Fp pfadd %mm4,%mm5 ## mm5= VV pfmul nb311nf_qqH(%esp),%mm5 ## vcoul=qq*VV ## update vctot pfadd nb311nf_vctot(%esp),%mm5 movq %mm5,nb311nf_vctot(%esp) ## done - one more? decl nb311nf_innerk(%esp) jz _nb_kernel311nf_ia32_3dnow.nb311nf_updateouterdata jmp _nb_kernel311nf_ia32_3dnow.nb311nf_inner_loop_nb_kernel311nf_ia32_3dnow.nb311nf_updateouterdata: ## get n from stack movl nb311nf_n(%esp),%esi ## get group index for i particle movl nb311nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] movq nb311nf_vctot(%esp),%mm7 pfacc %mm7,%mm7 ## get and sum the two parts of total potential movl nb311nf_Vc(%ebp),%eax movd (%eax,%edx,4),%mm6 pfadd %mm7,%mm6 movd %mm6,(%eax,%edx,4) ## increment vc[gid] movq nb311nf_Vvdwtot(%esp),%mm7 pfacc %mm7,%mm7 ## same for Vvdw movl nb311nf_Vvdw(%ebp),%eax movd (%eax,%edx,4),%mm6 pfadd %mm7,%mm6 movd %mm6,(%eax,%edx,4) ## increment Vvdw[gid] ## finish if last movl nb311nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel311nf_ia32_3dnow.nb311nf_outerend ## not last, iterate outer loop once more! movl %esi,nb311nf_n(%esp) jmp _nb_kernel311nf_ia32_3dnow.nb311nf_outer_nb_kernel311nf_ia32_3dnow.nb311nf_outerend: ## check if more outer neighborlists remain movl nb311nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel311nf_ia32_3dnow.nb311nf_end ## non-zero, do one more workunit jmp _nb_kernel311nf_ia32_3dnow.nb311nf_threadloop_nb_kernel311nf_ia32_3dnow.nb311nf_end: femms movl nb311nf_nouter(%esp),%eax movl nb311nf_ninner(%esp),%ebx movl nb311nf_outeriter(%ebp),%ecx movl nb311nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) addl $164,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret⌨️ 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 + -