nb_kernel213_ia32_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,087 行 · 第 1/5 页
S
2,087 行
#### $Id: nb_kernel213_ia32_sse2.s,v 1.1.2.2 2007/11/28 18:35:35 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_kernel213_ia32_sse2.globl _nb_kernel213_ia32_sse2nb_kernel213_ia32_sse2: _nb_kernel213_ia32_sse2: .set nb213_p_nri, 8.set nb213_iinr, 12.set nb213_jindex, 16.set nb213_jjnr, 20.set nb213_shift, 24.set nb213_shiftvec, 28.set nb213_fshift, 32.set nb213_gid, 36.set nb213_pos, 40.set nb213_faction, 44.set nb213_charge, 48.set nb213_p_facel, 52.set nb213_argkrf, 56.set nb213_argcrf, 60.set nb213_Vc, 64.set nb213_type, 68.set nb213_p_ntype, 72.set nb213_vdwparam, 76.set nb213_Vvdw, 80.set nb213_p_tabscale, 84.set nb213_VFtab, 88.set nb213_invsqrta, 92.set nb213_dvda, 96.set nb213_p_gbtabscale, 100.set nb213_GBtab, 104.set nb213_p_nthreads, 108.set nb213_count, 112.set nb213_mtx, 116.set nb213_outeriter, 120.set nb213_inneriter, 124.set nb213_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb213_ixO, 0.set nb213_iyO, 16.set nb213_izO, 32.set nb213_ixH1, 48.set nb213_iyH1, 64.set nb213_izH1, 80.set nb213_ixH2, 96.set nb213_iyH2, 112.set nb213_izH2, 128.set nb213_ixM, 144.set nb213_iyM, 160.set nb213_izM, 176.set nb213_iqH, 192.set nb213_iqM, 208.set nb213_dxO, 224.set nb213_dyO, 240.set nb213_dzO, 256.set nb213_dxH1, 272.set nb213_dyH1, 288.set nb213_dzH1, 304.set nb213_dxH2, 320.set nb213_dyH2, 336.set nb213_dzH2, 352.set nb213_dxM, 368.set nb213_dyM, 384.set nb213_dzM, 400.set nb213_qqH, 416.set nb213_qqM, 432.set nb213_c6, 448.set nb213_c12, 464.set nb213_six, 480.set nb213_twelve, 496.set nb213_vctot, 512.set nb213_Vvdwtot, 528.set nb213_fixO, 544.set nb213_fiyO, 560.set nb213_fizO, 576.set nb213_fixH1, 592.set nb213_fiyH1, 608.set nb213_fizH1, 624.set nb213_fixH2, 640.set nb213_fiyH2, 656.set nb213_fizH2, 672.set nb213_fixM, 688.set nb213_fiyM, 704.set nb213_fizM, 720.set nb213_fjx, 736.set nb213_fjy, 752.set nb213_fjz, 768.set nb213_half, 784.set nb213_three, 800.set nb213_two, 816.set nb213_rinvH1, 832.set nb213_rinvH2, 848.set nb213_rinvM, 864.set nb213_krsqH1, 880.set nb213_krsqH2, 896.set nb213_krsqM, 912.set nb213_krf, 928.set nb213_crf, 944.set nb213_is3, 960.set nb213_ii3, 964.set nb213_ntia, 968.set nb213_innerjjnr, 972.set nb213_innerk, 976.set nb213_n, 980.set nb213_nn1, 984.set nb213_nri, 988.set nb213_nouter, 992.set nb213_ninner, 996.set nb213_salign, 1000 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $1004,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb213_salign(%esp) emms ## Move args passed by reference to stack movl nb213_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb213_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb213_nouter(%esp) movl %eax,nb213_ninner(%esp) movl nb213_argkrf(%ebp),%esi movl nb213_argcrf(%ebp),%edi movsd (%esi),%xmm5 movsd (%edi),%xmm6 shufpd $0,%xmm5,%xmm5 shufpd $0,%xmm6,%xmm6 movapd %xmm5,nb213_krf(%esp) movapd %xmm6,nb213_crf(%esp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double 0.5 IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb213_half(%esp) movl %ebx,nb213_half+4(%esp) movsd nb213_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,nb213_half(%esp) movapd %xmm2,nb213_two(%esp) movapd %xmm3,nb213_three(%esp) movapd %xmm4,nb213_six(%esp) movapd %xmm5,nb213_twelve(%esp) ## assume we have at least one i particle - start directly movl nb213_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb213_charge(%ebp),%edx movsd 8(%edx,%ebx,8),%xmm3 movsd 24(%edx,%ebx,8),%xmm4 movl nb213_p_facel(%ebp),%esi movsd (%esi),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb213_iqH(%esp) movapd %xmm4,nb213_iqM(%esp) movl nb213_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb213_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb213_ntia(%esp)_nb_kernel213_ia32_sse2.nb213_threadloop: movl nb213_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel213_ia32_sse2.nb213_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_kernel213_ia32_sse2.nb213_spinlock ## if(nn1>nri) nn1=nri movl nb213_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,nb213_n(%esp) movl %ebx,nb213_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel213_ia32_sse2.nb213_outerstart jmp _nb_kernel213_ia32_sse2.nb213_end_nb_kernel213_ia32_sse2.nb213_outerstart: ## ebx contains number of outer iterations addl nb213_nouter(%esp),%ebx movl %ebx,nb213_nouter(%esp)_nb_kernel213_ia32_sse2.nb213_outer: movl nb213_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,nb213_is3(%esp) ## store is3 movl nb213_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 nb213_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb213_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb213_ii3(%esp) addsd (%eax,%ebx,8),%xmm3 ## ox addsd 8(%eax,%ebx,8),%xmm4 ## oy addsd 16(%eax,%ebx,8),%xmm5 ## oz addsd 24(%eax,%ebx,8),%xmm6 ## h1x addsd 32(%eax,%ebx,8),%xmm7 ## h1y shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 shufpd $0,%xmm6,%xmm6 shufpd $0,%xmm7,%xmm7 movapd %xmm3,nb213_ixO(%esp) movapd %xmm4,nb213_iyO(%esp) movapd %xmm5,nb213_izO(%esp) movapd %xmm6,nb213_ixH1(%esp) movapd %xmm7,nb213_iyH1(%esp) movsd %xmm2,%xmm6 movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 40(%eax,%ebx,8),%xmm6 ## h1z addsd 48(%eax,%ebx,8),%xmm0 ## h2x addsd 56(%eax,%ebx,8),%xmm1 ## h2y addsd 64(%eax,%ebx,8),%xmm2 ## h2z addsd 72(%eax,%ebx,8),%xmm3 ## mx addsd 80(%eax,%ebx,8),%xmm4 ## my addsd 88(%eax,%ebx,8),%xmm5 ## mz shufpd $0,%xmm6,%xmm6 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 %xmm6,nb213_izH1(%esp) movapd %xmm0,nb213_ixH2(%esp) movapd %xmm1,nb213_iyH2(%esp) movapd %xmm2,nb213_izH2(%esp) movapd %xmm3,nb213_ixM(%esp) movapd %xmm4,nb213_iyM(%esp) movapd %xmm5,nb213_izM(%esp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb213_vctot(%esp) movapd %xmm4,nb213_Vvdwtot(%esp) movapd %xmm4,nb213_fixO(%esp) movapd %xmm4,nb213_fiyO(%esp) movapd %xmm4,nb213_fizO(%esp) movapd %xmm4,nb213_fixH1(%esp) movapd %xmm4,nb213_fiyH1(%esp) movapd %xmm4,nb213_fizH1(%esp) movapd %xmm4,nb213_fixH2(%esp) movapd %xmm4,nb213_fiyH2(%esp) movapd %xmm4,nb213_fizH2(%esp) movapd %xmm4,nb213_fixM(%esp) movapd %xmm4,nb213_fiyM(%esp) movapd %xmm4,nb213_fizM(%esp) movl nb213_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 nb213_pos(%ebp),%esi movl nb213_faction(%ebp),%edi movl nb213_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb213_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb213_ninner(%esp),%ecx movl %ecx,nb213_ninner(%esp) addl $0,%edx movl %edx,nb213_innerk(%esp) ## number of innerloop atoms jge _nb_kernel213_ia32_sse2.nb213_unroll_loop jmp _nb_kernel213_ia32_sse2.nb213_checksingle_nb_kernel213_ia32_sse2.nb213_unroll_loop: ## twice unrolled innerloop here movl nb213_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx addl $8,nb213_innerjjnr(%esp) ## advance pointer (unrolled 2) movl nb213_charge(%ebp),%esi ## base of charge[] movlpd (%esi,%eax,8),%xmm3 movhpd (%esi,%ebx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb213_iqM(%esp),%xmm3 mulpd nb213_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movapd %xmm3,nb213_qqM(%esp) movapd %xmm4,nb213_qqH(%esp) movl nb213_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl nb213_vdwparam(%ebp),%esi shll %eax shll %ebx movl nb213_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,nb213_c6(%esp) movapd %xmm6,nb213_c12(%esp) movl nb213_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 nb213_ixO(%esp),%xmm4 movapd nb213_iyO(%esp),%xmm5 movapd nb213_izO(%esp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## store dr movapd %xmm4,nb213_dxO(%esp) movapd %xmm5,nb213_dyO(%esp) movapd %xmm6,nb213_dzO(%esp) ## square it ⌨️ 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?