nb_kernel311_ia32_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,033 行 · 第 1/5 页
S
2,033 行
#### $Id: nb_kernel311_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_kernel311_ia32_sse2.globl _nb_kernel311_ia32_sse2nb_kernel311_ia32_sse2: _nb_kernel311_ia32_sse2: .set nb311_p_nri, 8.set nb311_iinr, 12.set nb311_jindex, 16.set nb311_jjnr, 20.set nb311_shift, 24.set nb311_shiftvec, 28.set nb311_fshift, 32.set nb311_gid, 36.set nb311_pos, 40.set nb311_faction, 44.set nb311_charge, 48.set nb311_p_facel, 52.set nb311_argkrf, 56.set nb311_argcrf, 60.set nb311_Vc, 64.set nb311_type, 68.set nb311_p_ntype, 72.set nb311_vdwparam, 76.set nb311_Vvdw, 80.set nb311_p_tabscale, 84.set nb311_VFtab, 88.set nb311_invsqrta, 92.set nb311_dvda, 96.set nb311_p_gbtabscale, 100.set nb311_GBtab, 104.set nb311_p_nthreads, 108.set nb311_count, 112.set nb311_mtx, 116.set nb311_outeriter, 120.set nb311_inneriter, 124.set nb311_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb311_ixO, 0.set nb311_iyO, 16.set nb311_izO, 32.set nb311_ixH1, 48.set nb311_iyH1, 64.set nb311_izH1, 80.set nb311_ixH2, 96.set nb311_iyH2, 112.set nb311_izH2, 128.set nb311_iqO, 144.set nb311_iqH, 160.set nb311_dxO, 176.set nb311_dyO, 192.set nb311_dzO, 208.set nb311_dxH1, 224.set nb311_dyH1, 240.set nb311_dzH1, 256.set nb311_dxH2, 272.set nb311_dyH2, 288.set nb311_dzH2, 304.set nb311_qqO, 320.set nb311_qqH, 336.set nb311_rinvO, 352.set nb311_rinvH1, 368.set nb311_rinvH2, 384.set nb311_rO, 400.set nb311_rH1, 416.set nb311_rH2, 432.set nb311_tsc, 448.set nb311_two, 464.set nb311_c6, 480.set nb311_c12, 496.set nb311_six, 512.set nb311_twelve, 528.set nb311_vctot, 544.set nb311_Vvdwtot, 560.set nb311_fixO, 576.set nb311_fiyO, 592.set nb311_fizO, 608.set nb311_fixH1, 624.set nb311_fiyH1, 640.set nb311_fizH1, 656.set nb311_fixH2, 672.set nb311_fiyH2, 688.set nb311_fizH2, 704.set nb311_fjx, 720.set nb311_fjy, 736.set nb311_fjz, 752.set nb311_half, 768.set nb311_three, 784.set nb311_is3, 800.set nb311_ii3, 804.set nb311_ntia, 808.set nb311_innerjjnr, 812.set nb311_innerk, 816.set nb311_n, 820.set nb311_nn1, 824.set nb311_nri, 828.set nb311_nouter, 832.set nb311_ninner, 836.set nb311_salign, 840 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $844,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb311_salign(%esp) emms ## Move args passed by reference to stack movl nb311_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb311_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb311_nouter(%esp) movl %eax,nb311_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,nb311_half(%esp) movl %ebx,nb311_half+4(%esp) movsd nb311_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,nb311_half(%esp) movapd %xmm2,nb311_two(%esp) movapd %xmm3,nb311_three(%esp) movapd %xmm4,nb311_six(%esp) movapd %xmm5,nb311_twelve(%esp) movl nb311_p_tabscale(%ebp),%eax movsd (%eax),%xmm5 shufpd $0,%xmm5,%xmm5 movapd %xmm5,nb311_tsc(%esp) ## assume we have at least one i particle - start directly movl nb311_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb311_charge(%ebp),%edx movsd (%edx,%ebx,8),%xmm3 movsd 8(%edx,%ebx,8),%xmm4 movl nb311_p_facel(%ebp),%esi movsd (%esi),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb311_iqO(%esp) movapd %xmm4,nb311_iqH(%esp) movl nb311_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb311_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb311_ntia(%esp)_nb_kernel311_ia32_sse2.nb311_threadloop: movl nb311_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel311_ia32_sse2.nb311_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_kernel311_ia32_sse2.nb311_spinlock ## if(nn1>nri) nn1=nri movl nb311_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,nb311_n(%esp) movl %ebx,nb311_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel311_ia32_sse2.nb311_outerstart jmp _nb_kernel311_ia32_sse2.nb311_end_nb_kernel311_ia32_sse2.nb311_outerstart: ## ebx contains number of outer iterations addl nb311_nouter(%esp),%ebx movl %ebx,nb311_nouter(%esp)_nb_kernel311_ia32_sse2.nb311_outer: movl nb311_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,nb311_is3(%esp) ## store is3 movl nb311_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 nb311_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb311_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb311_ii3(%esp) addsd (%eax,%ebx,8),%xmm3 addsd 8(%eax,%ebx,8),%xmm4 addsd 16(%eax,%ebx,8),%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb311_ixO(%esp) movapd %xmm4,nb311_iyO(%esp) movapd %xmm5,nb311_izO(%esp) movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 24(%eax,%ebx,8),%xmm0 addsd 32(%eax,%ebx,8),%xmm1 addsd 40(%eax,%ebx,8),%xmm2 addsd 48(%eax,%ebx,8),%xmm3 addsd 56(%eax,%ebx,8),%xmm4 addsd 64(%eax,%ebx,8),%xmm5 shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm0,nb311_ixH1(%esp) movapd %xmm1,nb311_iyH1(%esp) movapd %xmm2,nb311_izH1(%esp) movapd %xmm3,nb311_ixH2(%esp) movapd %xmm4,nb311_iyH2(%esp) movapd %xmm5,nb311_izH2(%esp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb311_vctot(%esp) movapd %xmm4,nb311_Vvdwtot(%esp) movapd %xmm4,nb311_fixO(%esp) movapd %xmm4,nb311_fiyO(%esp) movapd %xmm4,nb311_fizO(%esp) movapd %xmm4,nb311_fixH1(%esp) movapd %xmm4,nb311_fiyH1(%esp) movapd %xmm4,nb311_fizH1(%esp) movapd %xmm4,nb311_fixH2(%esp) movapd %xmm4,nb311_fiyH2(%esp) movapd %xmm4,nb311_fizH2(%esp) movl nb311_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 nb311_pos(%ebp),%esi movl nb311_faction(%ebp),%edi movl nb311_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb311_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb311_ninner(%esp),%ecx movl %ecx,nb311_ninner(%esp) addl $0,%edx movl %edx,nb311_innerk(%esp) ## number of innerloop atoms jge _nb_kernel311_ia32_sse2.nb311_unroll_loop jmp _nb_kernel311_ia32_sse2.nb311_checksingle_nb_kernel311_ia32_sse2.nb311_unroll_loop: ## twice unrolled innerloop here movl nb311_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx addl $8,nb311_innerjjnr(%esp) ## advance pointer (unrolled 2) movl nb311_charge(%ebp),%esi ## base of charge[] movlpd (%esi,%eax,8),%xmm3 movhpd (%esi,%ebx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb311_iqO(%esp),%xmm3 mulpd nb311_iqH(%esp),%xmm4 movapd %xmm3,nb311_qqO(%esp) movapd %xmm4,nb311_qqH(%esp) movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movl nb311_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl nb311_vdwparam(%ebp),%esi shll %eax shll %ebx movl nb311_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,nb311_c6(%esp) movapd %xmm6,nb311_c12(%esp) movl nb311_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 ixO-izO to xmm4-xmm6 movapd nb311_ixO(%esp),%xmm4 movapd nb311_iyO(%esp),%xmm5 movapd nb311_izO(%esp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## store dr movapd %xmm4,nb311_dxO(%esp) movapd %xmm5,nb311_dyO(%esp) movapd %xmm6,nb311_dzO(%esp) ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movapd nb311_ixH1(%esp),%xmm4 movapd nb311_iyH1(%esp),%xmm5 movapd nb311_izH1(%esp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## store dr ⌨️ 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?