📄 nb_kernel312_ia32_3dnow.s
字号:
#### $Id: nb_kernel312_ia32_3dnow.s,v 1.1.2.2 2007/11/28 09:34:39 spoel Exp $#### Gromacs 4.0 Copyright (c) 1991-2003 ## David van der Spoel, Erik Lindahl#### This program is free software; you can redistribute it and/or## modify it under the terms of the GNU General Public License## as published by the Free Software Foundation; either version 2## of the License, or (at your option) any later version.#### To help us fund GROMACS development, we humbly ask that you cite## the research papers on the package. Check out http://www.gromacs.org## ## And Hey:## Gnomes, ROck Monsters And Chili Sauce##.globl nb_kernel312_ia32_3dnow.globl _nb_kernel312_ia32_3dnownb_kernel312_ia32_3dnow: _nb_kernel312_ia32_3dnow: .set nb312_p_nri, 8.set nb312_iinr, 12.set nb312_jindex, 16.set nb312_jjnr, 20.set nb312_shift, 24.set nb312_shiftvec, 28.set nb312_fshift, 32.set nb312_gid, 36.set nb312_pos, 40.set nb312_faction, 44.set nb312_charge, 48.set nb312_p_facel, 52.set nb312_p_krf, 56.set nb312_p_crf, 60.set nb312_Vc, 64.set nb312_type, 68.set nb312_p_ntype, 72.set nb312_vdwparam, 76.set nb312_Vvdw, 80.set nb312_p_tabscale, 84.set nb312_VFtab, 88.set nb312_invsqrta, 92.set nb312_dvda, 96.set nb312_p_gbtabscale, 100.set nb312_GBtab, 104.set nb312_p_nthreads, 108.set nb312_count, 112.set nb312_mtx, 116.set nb312_outeriter, 120.set nb312_inneriter, 124.set nb312_work, 128 ## stack offsets for local variables .set nb312_is3, 0.set nb312_ii3, 4.set nb312_ixO, 8.set nb312_iyO, 12.set nb312_izO, 16.set nb312_ixH, 20.set nb312_iyH, 28.set nb312_izH, 36.set nb312_qqOO, 44.set nb312_qqOH, 52.set nb312_qqHH, 60.set nb312_c6, 68.set nb312_c12, 76.set nb312_six, 84.set nb312_twelve, 92.set nb312_two, 100.set nb312_n1, 108.set nb312_tsc, 116.set nb312_vctot, 124.set nb312_Vvdwtot, 132.set nb312_innerjjnr, 140.set nb312_innerk, 144.set nb312_fixO, 148.set nb312_fiyO, 152.set nb312_fizO, 156.set nb312_fixH, 160.set nb312_fiyH, 168.set nb312_fizH, 176.set nb312_dxO, 184.set nb312_dyO, 188.set nb312_dzO, 192.set nb312_dxH, 200.set nb312_dyH, 208.set nb312_dzH, 216.set nb312_tmprsqH, 224.set nb312_n, 232 ## idx for outer loop.set nb312_nn1, 236 ## number of outer iterations.set nb312_nri, 240.set nb312_nouter, 244.set nb312_ninner, 248 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $252,%esp ## local stack space femms movl nb312_p_nri(%ebp),%ecx movl nb312_p_facel(%ebp),%esi movl nb312_p_tabscale(%ebp),%edi movl (%ecx),%ecx movl %ecx,nb312_nri(%esp) movd (%esi),%mm1 ## facel movl nb312_p_ntype(%ebp),%esi ## zero iteration counters movl $0,%eax movl %eax,nb312_nouter(%esp) movl %eax,nb312_ninner(%esp) ## assume we have at least one i particle - start directly movl nb312_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx=ii movl nb312_charge(%ebp),%edx movd (%edx,%ebx,4),%mm2 ## mm2=charge[ii0] (O) movd 4(%edx,%ebx,4),%mm3 ## mm2=charge[ii0+1] (H) movq %mm2,%mm4 pfmul %mm1,%mm4 movq %mm3,%mm6 pfmul %mm1,%mm6 movq %mm4,%mm5 pfmul %mm2,%mm4 ## mm4=qqOO*facel pfmul %mm3,%mm5 ## mm5=qqOH*facel pfmul %mm3,%mm6 ## mm6=qqHH*facel punpckldq %mm5,%mm5 ## spread to both halves punpckldq %mm6,%mm6 ## spread to both halves movq %mm4,nb312_qqOO(%esp) movq %mm5,nb312_qqOH(%esp) movq %mm6,nb312_qqHH(%esp) movl nb312_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl %ecx,%edx imull (%esi),%ecx addl %ecx,%edx movl nb312_vdwparam(%ebp),%eax movd (%eax,%edx,4),%mm0 movd 4(%eax,%edx,4),%mm1 movq %mm0,nb312_c6(%esp) movq %mm1,nb312_c12(%esp) movd (%edi),%mm5 punpckldq %mm5,%mm5 movq %mm5,nb312_tsc(%esp) movl $0x40000000,%eax ## 2.0 movl %eax,nb312_two(%esp) movl %eax,nb312_two+4(%esp) movl $0x40c00000,%ebx ## 6.0 movl %ebx,nb312_six(%esp) movl %ebx,nb312_six+4(%esp) movl $0x41400000,%ecx ## 12.0 movl %ecx,nb312_twelve(%esp) movl %ecx,nb312_twelve+4(%esp)_nb_kernel312_ia32_3dnow.nb312_threadloop: movl nb312_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel312_ia32_3dnow.nb312_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_kernel312_ia32_3dnow.nb312_spinlock ## if(nn1>nri) nn1=nri movl nb312_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,nb312_n(%esp) movl %ebx,nb312_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel312_ia32_3dnow.nb312_outerstart jmp _nb_kernel312_ia32_3dnow.nb312_end_nb_kernel312_ia32_3dnow.nb312_outerstart: ## ebx contains number of outer iterations addl nb312_nouter(%esp),%ebx movl %ebx,nb312_nouter(%esp)_nb_kernel312_ia32_3dnow.nb312_outer: movl nb312_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,nb312_is3(%esp) ## store is3 movl nb312_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 nb312_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx=ii leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb312_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,nb312_ii3(%esp) ## (use mm7 as temp. storage for iz.) pfadd %mm7,%mm6 movq %mm5,nb312_ixO(%esp) movq %mm6,nb312_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,nb312_ixH(%esp) movq %mm1,nb312_iyH(%esp) movq %mm2,nb312_izH(%esp) ## clear vctot and i forces pxor %mm7,%mm7 movq %mm7,nb312_vctot(%esp) movq %mm7,nb312_Vvdwtot(%esp) movq %mm7,nb312_fixO(%esp) movq %mm7,nb312_fizO(%esp) movq %mm7,nb312_fixH(%esp) movq %mm7,nb312_fiyH(%esp) movq %mm7,nb312_fizH(%esp) movl nb312_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,nb312_innerk(%esp) ## number of innerloop atoms addl nb312_ninner(%esp),%edx movl %edx,nb312_ninner(%esp) movl nb312_pos(%ebp),%esi movl nb312_faction(%ebp),%edi movl nb312_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb312_innerjjnr(%esp) ## pointer to jjnr[nj0] _nb_kernel312_ia32_3dnow.nb312_inner_loop: ## a single j particle iteration here - compare with the unrolled code for comments. movl nb312_innerjjnr(%esp),%eax movl (%eax),%eax ## eax=jnr offset addl $4,nb312_innerjjnr(%esp) ## advance pointer 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 nb312_ixO(%esp),%mm0 pfsubr nb312_izO(%esp),%mm1 movq %mm0,nb312_dxO(%esp) pfmul %mm0,%mm0 movd %mm1,nb312_dzO(%esp) pfmul %mm1,%mm1 pfacc %mm0,%mm0 pfadd %mm1,%mm0 ## mm0=rsqO punpckldq %mm2,%mm2 punpckldq %mm3,%mm3 punpckldq %mm4,%mm4 ## mm2-mm4 is jx-jz pfsubr nb312_ixH(%esp),%mm2 pfsubr nb312_iyH(%esp),%mm3 pfsubr nb312_izH(%esp),%mm4 ## mm2-mm4 is dxH-dzH movq %mm2,nb312_dxH(%esp) movq %mm3,nb312_dyH(%esp) movq %mm4,nb312_dzH(%esp) pfmul %mm2,%mm2 pfmul %mm3,%mm3 pfmul %mm4,%mm4 pfadd %mm2,%mm3 pfadd %mm4,%mm3 ## mm3=rsqH movq %mm3,nb312_tmprsqH(%esp) pfrsqrt %mm0,%mm1 movq %mm1,%mm2 pfmul %mm1,%mm1 pfrsqit1 %mm0,%mm1 pfrcpit2 %mm2,%mm1 ## mm1=invsqrt pfmul %mm1,%mm0 ## mm0=rsq pfmul nb312_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movd %mm4,nb312_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 nb312_VFtab(%ebp),%edx movl nb312_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 nb312_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 nb312_qqOO(%esp),%mm5 ## vcoul=qq*VV pfmul nb312_qqOO(%esp),%mm7 ## fijC=qq*FF ## update vctot directly, use mm3 for fscal sum. pfadd nb312_vctot(%esp),%mm5 movq %mm5,nb312_vctot(%esp) movq %mm7,%mm3 pfmul nb312_tsc(%esp),%mm3 movq %mm1,%mm5 pfmul %mm5,%mm5 movq %mm5,%mm4 pfmul %mm5,%mm4 pfmul %mm5,%mm4 movq %mm4,%mm5 pfmul %mm5,%mm5 ## mm4=rinvsix, mm5=rinvtwelve pfmul nb312_c6(%esp),%mm4 pfmul nb312_c12(%esp),%mm5 movq %mm5,%mm6 pfsub %mm4,%mm6 pfmul nb312_six(%esp),%mm4 pfmul nb312_twelve(%esp),%mm5 pfsub %mm4,%mm5 pfmul %mm1,%mm5 pfsubr %mm5,%mm3 pfmul %mm1,%mm3 ## mm3 is total fscal (for the oxygen) now ## update Vvdwtot pfadd nb312_Vvdwtot(%esp),%mm6 ## add the earlier value movq %mm6,nb312_Vvdwtot(%esp) ## store the sum ## Ready with the oxygen - potential is updated, fscal is in mm3. ## time for hydrogens! movq nb312_tmprsqH(%esp),%mm0 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 nb312_tsc(%esp),%mm0 pf2iw %mm0,%mm4 movq %mm4,nb312_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 nb312_VFtab(%ebp),%edx movl nb312_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 nb312_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⌨️ 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 + -