nb_kernel210_ia32_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,212 行 · 第 1/3 页
S
1,212 行
#### $Id: nb_kernel210_ia32_sse2.s,v 1.2.2.2 2007/11/28 18:35:34 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_kernel210_ia32_sse2.globl _nb_kernel210_ia32_sse2nb_kernel210_ia32_sse2: _nb_kernel210_ia32_sse2: .set nb210_p_nri, 8.set nb210_iinr, 12.set nb210_jindex, 16.set nb210_jjnr, 20.set nb210_shift, 24.set nb210_shiftvec, 28.set nb210_fshift, 32.set nb210_gid, 36.set nb210_pos, 40.set nb210_faction, 44.set nb210_charge, 48.set nb210_p_facel, 52.set nb210_argkrf, 56.set nb210_argcrf, 60.set nb210_Vc, 64.set nb210_type, 68.set nb210_p_ntype, 72.set nb210_vdwparam, 76.set nb210_Vvdw, 80.set nb210_p_tabscale, 84.set nb210_VFtab, 88.set nb210_invsqrta, 92.set nb210_dvda, 96.set nb210_p_gbtabscale, 100.set nb210_GBtab, 104.set nb210_p_nthreads, 108.set nb210_count, 112.set nb210_mtx, 116.set nb210_outeriter, 120.set nb210_inneriter, 124.set nb210_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb210_ix, 0.set nb210_iy, 16.set nb210_iz, 32.set nb210_iq, 48.set nb210_dx, 64.set nb210_dy, 80.set nb210_dz, 96.set nb210_c6, 112.set nb210_c12, 128.set nb210_six, 144.set nb210_twelve, 160.set nb210_vctot, 176.set nb210_Vvdwtot, 192.set nb210_fix, 208.set nb210_fiy, 224.set nb210_fiz, 240.set nb210_half, 256.set nb210_three, 272.set nb210_two, 288.set nb210_krf, 304.set nb210_crf, 320.set nb210_facel, 336 ## uses 8 bytes.set nb210_is3, 344.set nb210_ii3, 348.set nb210_ntia, 352.set nb210_innerjjnr, 356.set nb210_innerk, 360.set nb210_n, 364.set nb210_nn1, 368.set nb210_nri, 372.set nb210_ntype, 376.set nb210_nouter, 380.set nb210_ninner, 384.set nb210_salign, 388 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $392,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb210_salign(%esp) emms ## Move args passed by reference to stack movl nb210_p_nri(%ebp),%ecx movl nb210_p_facel(%ebp),%esi movl nb210_p_ntype(%ebp),%edi movl (%ecx),%ecx movsd (%esi),%xmm7 movl (%edi),%edi movl %ecx,nb210_nri(%esp) movsd %xmm7,nb210_facel(%esp) movl %edi,nb210_ntype(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb210_nouter(%esp) movl %eax,nb210_ninner(%esp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double 0.5 IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb210_half(%esp) movl %ebx,nb210_half+4(%esp) movsd nb210_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,nb210_half(%esp) movapd %xmm2,nb210_two(%esp) movapd %xmm3,nb210_three(%esp) movapd %xmm4,nb210_six(%esp) movapd %xmm5,nb210_twelve(%esp) movl nb210_argkrf(%ebp),%esi movl nb210_argcrf(%ebp),%edi movsd (%esi),%xmm5 movsd (%edi),%xmm6 shufpd $0,%xmm5,%xmm5 shufpd $0,%xmm6,%xmm6 movapd %xmm5,nb210_krf(%esp) movapd %xmm6,nb210_crf(%esp)_nb_kernel210_ia32_sse2.nb210_threadloop: movl nb210_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel210_ia32_sse2.nb210_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_kernel210_ia32_sse2.nb210_spinlock ## if(nn1>nri) nn1=nri movl nb210_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,nb210_n(%esp) movl %ebx,nb210_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel210_ia32_sse2.nb210_outerstart jmp _nb_kernel210_ia32_sse2.nb210_end_nb_kernel210_ia32_sse2.nb210_outerstart: ## ebx contains number of outer iterations addl nb210_nouter(%esp),%ebx movl %ebx,nb210_nouter(%esp)_nb_kernel210_ia32_sse2.nb210_outer: movl nb210_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,nb210_is3(%esp) ## store is3 movl nb210_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 nb210_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movl nb210_charge(%ebp),%edx movsd (%edx,%ebx,8),%xmm3 mulsd nb210_facel(%esp),%xmm3 shufpd $0,%xmm3,%xmm3 movl nb210_type(%ebp),%edx movl (%edx,%ebx,4),%edx imull nb210_ntype(%esp),%edx shll %edx movl %edx,nb210_ntia(%esp) leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb210_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,nb210_iq(%esp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb210_ix(%esp) movapd %xmm1,nb210_iy(%esp) movapd %xmm2,nb210_iz(%esp) movl %ebx,nb210_ii3(%esp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb210_vctot(%esp) movapd %xmm4,nb210_Vvdwtot(%esp) movapd %xmm4,nb210_fix(%esp) movapd %xmm4,nb210_fiy(%esp) movapd %xmm4,nb210_fiz(%esp) movl nb210_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 nb210_pos(%ebp),%esi movl nb210_faction(%ebp),%edi movl nb210_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb210_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb210_ninner(%esp),%ecx movl %ecx,nb210_ninner(%esp) addl $0,%edx movl %edx,nb210_innerk(%esp) ## number of innerloop atoms jge _nb_kernel210_ia32_sse2.nb210_unroll_loop jmp _nb_kernel210_ia32_sse2.nb210_checksingle_nb_kernel210_ia32_sse2.nb210_unroll_loop: ## twice unrolled innerloop here movl nb210_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx addl $8,nb210_innerjjnr(%esp) ## advance pointer (unrolled 2) movl nb210_charge(%ebp),%esi ## base of charge[] movlpd (%esi,%eax,8),%xmm3 movhpd (%esi,%ebx,8),%xmm3 movapd nb210_iq(%esp),%xmm5 mulpd %xmm5,%xmm3 ## qq movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movl nb210_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl nb210_vdwparam(%ebp),%esi shll %eax shll %ebx movl nb210_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,nb210_c6(%esp) movapd %xmm6,nb210_c12(%esp) movl nb210_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 nb210_ix(%esp),%xmm4 movapd nb210_iy(%esp),%xmm5 movapd nb210_iz(%esp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## store dr movapd %xmm4,nb210_dx(%esp) movapd %xmm5,nb210_dy(%esp) movapd %xmm6,nb210_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 movapd nb210_krf(%esp),%xmm7 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb210_three(%esp),%xmm1 mulpd %xmm4,%xmm7 ## krsq mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb210_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 nb210_three(%esp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb210_half(%esp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=rinv movapd %xmm0,%xmm4 mulpd %xmm4,%xmm4 ## xmm4=rinvsq movapd %xmm0,%xmm6 addpd %xmm7,%xmm6 ## xmm6=rinv+ krsq movapd %xmm4,%xmm1 subpd nb210_crf(%esp),%xmm6 mulpd %xmm4,%xmm1 mulpd %xmm4,%xmm1 ## xmm1=rinvsix movapd %xmm1,%xmm2 mulpd %xmm2,%xmm2 ## xmm2=rinvtwelve mulpd %xmm3,%xmm6 ## xmm6=vcoul=qq*(rinv+ krsq) mulpd nb210_two(%esp),%xmm7 mulpd nb210_c6(%esp),%xmm1 mulpd nb210_c12(%esp),%xmm2 movapd %xmm2,%xmm5 subpd %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addpd nb210_Vvdwtot(%esp),%xmm5 mulpd nb210_six(%esp),%xmm1 mulpd nb210_twelve(%esp),%xmm2 subpd %xmm1,%xmm2 subpd %xmm7,%xmm0 mulpd %xmm0,%xmm3 addpd %xmm3,%xmm2 mulpd %xmm2,%xmm4 ## xmm4=total fscal addpd nb210_vctot(%esp),%xmm6 movapd nb210_dx(%esp),%xmm0 movapd nb210_dy(%esp),%xmm1 movapd nb210_dz(%esp),%xmm2 movapd %xmm6,nb210_vctot(%esp) movapd %xmm5,nb210_Vvdwtot(%esp) movl nb210_faction(%ebp),%edi mulpd %xmm4,%xmm0 mulpd %xmm4,%xmm1 mulpd %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movapd nb210_fix(%esp),%xmm3 movapd nb210_fiy(%esp),%xmm4 movapd nb210_fiz(%esp),%xmm5 addpd %xmm0,%xmm3⌨️ 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?