nb_kernel310_ia32_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,395 行 · 第 1/3 页
S
1,395 行
#### $Id: nb_kernel310_ia32_sse2.s,v 1.1.2.2 2007/11/28 18:35:36 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_kernel310_ia32_sse2.globl _nb_kernel310_ia32_sse2nb_kernel310_ia32_sse2: _nb_kernel310_ia32_sse2: .set nb310_p_nri, 8.set nb310_iinr, 12.set nb310_jindex, 16.set nb310_jjnr, 20.set nb310_shift, 24.set nb310_shiftvec, 28.set nb310_fshift, 32.set nb310_gid, 36.set nb310_pos, 40.set nb310_faction, 44.set nb310_charge, 48.set nb310_p_facel, 52.set nb310_argkrf, 56.set nb310_argcrf, 60.set nb310_Vc, 64.set nb310_type, 68.set nb310_p_ntype, 72.set nb310_vdwparam, 76.set nb310_Vvdw, 80.set nb310_p_tabscale, 84.set nb310_VFtab, 88.set nb310_invsqrta, 92.set nb310_dvda, 96.set nb310_p_gbtabscale, 100.set nb310_GBtab, 104.set nb310_p_nthreads, 108.set nb310_count, 112.set nb310_mtx, 116.set nb310_outeriter, 120.set nb310_inneriter, 124.set nb310_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb310_ix, 0.set nb310_iy, 16.set nb310_iz, 32.set nb310_iq, 48.set nb310_dx, 64.set nb310_dy, 80.set nb310_dz, 96.set nb310_two, 112.set nb310_six, 128.set nb310_twelve, 144.set nb310_tsc, 160.set nb310_qq, 176.set nb310_c6, 192.set nb310_c12, 208.set nb310_fscal, 224.set nb310_vctot, 240.set nb310_Vvdwtot, 256.set nb310_fix, 272.set nb310_fiy, 288.set nb310_fiz, 304.set nb310_half, 320.set nb310_three, 336.set nb310_is3, 352.set nb310_ii3, 356.set nb310_ntia, 360.set nb310_innerjjnr, 364.set nb310_innerk, 368.set nb310_n, 372.set nb310_nn1, 376.set nb310_nri, 380.set nb310_facel, 384 ## uses 8 bytes.set nb310_ntype, 392.set nb310_nouter, 396.set nb310_ninner, 400.set nb310_salign, 404 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $408,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb310_salign(%esp) emms ## Move args passed by reference to stack movl nb310_p_nri(%ebp),%ecx movl nb310_p_facel(%ebp),%esi movl nb310_p_ntype(%ebp),%edi movl (%ecx),%ecx movsd (%esi),%xmm7 movl (%edi),%edi movl %ecx,nb310_nri(%esp) movsd %xmm7,nb310_facel(%esp) movl %edi,nb310_ntype(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb310_nouter(%esp) movl %eax,nb310_ninner(%esp) movl nb310_p_tabscale(%ebp),%eax movsd (%eax),%xmm5 shufpd $0,%xmm5,%xmm5 movapd %xmm5,nb310_tsc(%esp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double 0.5 IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb310_half(%esp) movl %ebx,nb310_half+4(%esp) movsd nb310_half(%esp),%xmm1 shufpd $0,%xmm1,%xmm1 ## splat to all elements movapd %xmm1,%xmm3 addpd %xmm3,%xmm3 ## 1.0 movapd %xmm3,%xmm2 addpd %xmm2,%xmm2 ## 2.0 addpd %xmm2,%xmm3 ## 3.0 movapd %xmm3,%xmm4 addpd %xmm4,%xmm4 ## 6.0 movapd %xmm4,%xmm5 addpd %xmm5,%xmm5 ## 12.0 movapd %xmm1,nb310_half(%esp) movapd %xmm2,nb310_two(%esp) movapd %xmm3,nb310_three(%esp) movapd %xmm4,nb310_six(%esp) movapd %xmm5,nb310_twelve(%esp)_nb_kernel310_ia32_sse2.nb310_threadloop: movl nb310_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel310_ia32_sse2.nb310_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_kernel310_ia32_sse2.nb310_spinlock ## if(nn1>nri) nn1=nri movl nb310_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,nb310_n(%esp) movl %ebx,nb310_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel310_ia32_sse2.nb310_outerstart jmp _nb_kernel310_ia32_sse2.nb310_end_nb_kernel310_ia32_sse2.nb310_outerstart: ## ebx contains number of outer iterations addl nb310_nouter(%esp),%ebx movl %ebx,nb310_nouter(%esp)_nb_kernel310_ia32_sse2.nb310_outer: movl nb310_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,nb310_is3(%esp) ## store is3 movl nb310_shiftvec(%ebp),%eax ## eax = base of shiftvec[] movsd (%eax,%ebx,8),%xmm0 movsd 8(%eax,%ebx,8),%xmm1 movsd 16(%eax,%ebx,8),%xmm2 movl nb310_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movl nb310_charge(%ebp),%edx movsd (%edx,%ebx,8),%xmm3 mulsd nb310_facel(%esp),%xmm3 shufpd $0,%xmm3,%xmm3 movl nb310_type(%ebp),%edx movl (%edx,%ebx,4),%edx imull nb310_ntype(%esp),%edx shll %edx movl %edx,nb310_ntia(%esp) leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb310_pos(%ebp),%eax ## eax = base of pos[] addsd (%eax,%ebx,8),%xmm0 addsd 8(%eax,%ebx,8),%xmm1 addsd 16(%eax,%ebx,8),%xmm2 movapd %xmm3,nb310_iq(%esp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb310_ix(%esp) movapd %xmm1,nb310_iy(%esp) movapd %xmm2,nb310_iz(%esp) movl %ebx,nb310_ii3(%esp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb310_vctot(%esp) movapd %xmm4,nb310_Vvdwtot(%esp) movapd %xmm4,nb310_fix(%esp) movapd %xmm4,nb310_fiy(%esp) movapd %xmm4,nb310_fiz(%esp) movl nb310_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 nb310_pos(%ebp),%esi movl nb310_faction(%ebp),%edi movl nb310_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb310_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb310_ninner(%esp),%ecx movl %ecx,nb310_ninner(%esp) addl $0,%edx movl %edx,nb310_innerk(%esp) ## number of innerloop atoms jge _nb_kernel310_ia32_sse2.nb310_unroll_loop jmp _nb_kernel310_ia32_sse2.nb310_checksingle_nb_kernel310_ia32_sse2.nb310_unroll_loop: ## twice unrolled innerloop here movl nb310_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx addl $8,nb310_innerjjnr(%esp) ## advance pointer (unrolled 2) movl nb310_charge(%ebp),%esi ## base of charge[] movlpd (%esi,%eax,8),%xmm3 movhpd (%esi,%ebx,8),%xmm3 movapd nb310_iq(%esp),%xmm2 mulpd %xmm2,%xmm3 movapd %xmm3,nb310_qq(%esp) movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movl nb310_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl nb310_vdwparam(%ebp),%esi shll %eax shll %ebx movl nb310_ntia(%esp),%edi addl %edi,%eax addl %edi,%ebx movlpd (%esi,%eax,8),%xmm6 ## c6a movlpd (%esi,%ebx,8),%xmm7 ## c6b movhpd 8(%esi,%eax,8),%xmm6 ## c6a c12a movhpd 8(%esi,%ebx,8),%xmm7 ## c6b c12b movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movd %mm0,%eax movd %mm1,%ebx movapd %xmm4,nb310_c6(%esp) movapd %xmm6,nb310_c12(%esp) movl nb310_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx ## move two coordinates to xmm0-xmm2 movlpd (%esi,%eax,8),%xmm0 movlpd 8(%esi,%eax,8),%xmm1 movlpd 16(%esi,%eax,8),%xmm2 movhpd (%esi,%ebx,8),%xmm0 movhpd 8(%esi,%ebx,8),%xmm1 movhpd 16(%esi,%ebx,8),%xmm2 ## move ix-iz to xmm4-xmm6 movapd nb310_ix(%esp),%xmm4 movapd nb310_iy(%esp),%xmm5 movapd nb310_iz(%esp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## store dr movapd %xmm4,nb310_dx(%esp) movapd %xmm5,nb310_dy(%esp) movapd %xmm6,nb310_dz(%esp) ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 ## rsq in xmm4 cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb310_three(%esp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb310_half(%esp),%xmm0 subpd %xmm2,%xmm1 ## 30-rsq*lu*lu mulpd %xmm5,%xmm1 mulpd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movapd %xmm1,%xmm5 ## copy of lu mulpd %xmm1,%xmm1 ## lu*lu movapd nb310_three(%esp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb310_half(%esp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=rinv mulpd %xmm0,%xmm4 ## xmm4=r mulpd nb310_tsc(%esp),%xmm4 cvttpd2pi %xmm4,%mm6 ## mm6 = lu idx cvtpi2pd %mm6,%xmm5 subpd %xmm5,%xmm4 movapd %xmm4,%xmm1 ## xmm1=eps movapd %xmm1,%xmm2 mulpd %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 ## idx *= 4 movd %eax,%mm0 movd %ebx,%mm1 movl nb310_VFtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm6,%ebx ## indices in eax/ebx movlpd (%esi,%eax,8),%xmm4 ## Y1 movlpd (%esi,%ebx,8),%xmm3 ## Y2 movhpd 8(%esi,%eax,8),%xmm4 ## Y1 F1 movhpd 8(%esi,%ebx,8),%xmm3 ## Y2 F2 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movlpd 16(%esi,%eax,8),%xmm6 ## G1 movlpd 16(%esi,%ebx,8),%xmm3 ## G2 movhpd 24(%esi,%eax,8),%xmm6 ## G1 H1 movhpd 24(%esi,%ebx,8),%xmm3 ## G2 H2 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## coulomb table ready, in xmm4-xmm7 mulpd %xmm1,%xmm6 ## xmm6=Geps mulpd %xmm2,%xmm7 ## xmm7=Heps2 addpd %xmm6,%xmm5 addpd %xmm7,%xmm5 ## xmm5=Fp mulpd nb310_two(%esp),%xmm7 ## two*Heps2 movapd nb310_qq(%esp),%xmm3 addpd %xmm6,%xmm7 addpd %xmm5,%xmm7 ## xmm7=FF mulpd %xmm1,%xmm5 ## xmm5=eps*Fp addpd %xmm4,%xmm5 ## xmm5=VV mulpd %xmm3,%xmm5 ## vcoul=qq*VV mulpd %xmm7,%xmm3 ## fijC=FF*qq ## at this point mm5 contains vcoul and mm3 fijC ## L-J movapd %xmm0,%xmm4 mulpd %xmm0,%xmm4 ## xmm4=rinvsq ## increment vcoul - then we can get rid of mm5 ## update vctot addpd nb310_vctot(%esp),%xmm5 movapd %xmm4,%xmm6 mulpd %xmm4,%xmm6 movapd %xmm5,nb310_vctot(%esp) mulpd %xmm4,%xmm6 ## xmm6=rinvsix movapd %xmm6,%xmm4 mulpd %xmm4,%xmm4 ## xmm4=rinvtwelve mulpd nb310_c6(%esp),%xmm6 mulpd nb310_c12(%esp),%xmm4 movapd nb310_Vvdwtot(%esp),%xmm7 addpd %xmm4,%xmm7 mulpd nb310_twelve(%esp),%xmm4 subpd %xmm6,%xmm7 mulpd nb310_tsc(%esp),%xmm3 mulpd nb310_six(%esp),%xmm6 movapd %xmm7,nb310_Vvdwtot(%esp) subpd %xmm6,%xmm4 mulpd %xmm0,%xmm4 subpd %xmm3,%xmm4 mulpd %xmm0,%xmm4 movapd nb310_dx(%esp),%xmm0 movapd nb310_dy(%esp),%xmm1 movapd nb310_dz(%esp),%xmm2 movd %mm0,%eax movd %mm1,%ebx movl nb310_faction(%ebp),%edi mulpd %xmm4,%xmm0 mulpd %xmm4,%xmm1 mulpd %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movapd nb310_fix(%esp),%xmm3 movapd nb310_fiy(%esp),%xmm4 movapd nb310_fiz(%esp),%xmm5 addpd %xmm0,%xmm3 addpd %xmm1,%xmm4 addpd %xmm2,%xmm5 movapd %xmm3,nb310_fix(%esp) movapd %xmm4,nb310_fiy(%esp) movapd %xmm5,nb310_fiz(%esp) ## the fj's - start by accumulating forces from memory movlpd (%edi,%eax,8),%xmm3 movlpd 8(%edi,%eax,8),%xmm4 movlpd 16(%edi,%eax,8),%xmm5 movhpd (%edi,%ebx,8),%xmm3 movhpd 8(%edi,%ebx,8),%xmm4 movhpd 16(%edi,%ebx,8),%xmm5⌨️ 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?