nb_kernel331_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,107 行 · 第 1/5 页
S
2,107 行
#### $Id: nb_kernel331_ia32_sse.s,v 1.1.2.2 2007/11/28 09:36:34 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_kernel331_ia32_sse.globl _nb_kernel331_ia32_ssenb_kernel331_ia32_sse: _nb_kernel331_ia32_sse: .set nb331_p_nri, 8.set nb331_iinr, 12.set nb331_jindex, 16.set nb331_jjnr, 20.set nb331_shift, 24.set nb331_shiftvec, 28.set nb331_fshift, 32.set nb331_gid, 36.set nb331_pos, 40.set nb331_faction, 44.set nb331_charge, 48.set nb331_p_facel, 52.set nb331_argkrf, 56.set nb331_argcrf, 60.set nb331_Vc, 64.set nb331_type, 68.set nb331_p_ntype, 72.set nb331_vdwparam, 76.set nb331_Vvdw, 80.set nb331_p_tabscale, 84.set nb331_VFtab, 88.set nb331_invsqrta, 92.set nb331_dvda, 96.set nb331_p_gbtabscale, 100.set nb331_GBtab, 104.set nb331_p_nthreads, 108.set nb331_count, 112.set nb331_mtx, 116.set nb331_outeriter, 120.set nb331_inneriter, 124.set nb331_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb331_ixO, 0.set nb331_iyO, 16.set nb331_izO, 32.set nb331_ixH1, 48.set nb331_iyH1, 64.set nb331_izH1, 80.set nb331_ixH2, 96.set nb331_iyH2, 112.set nb331_izH2, 128.set nb331_iqO, 144.set nb331_iqH, 160.set nb331_dxO, 176.set nb331_dyO, 192.set nb331_dzO, 208.set nb331_dxH1, 224.set nb331_dyH1, 240.set nb331_dzH1, 256.set nb331_dxH2, 272.set nb331_dyH2, 288.set nb331_dzH2, 304.set nb331_qqO, 320.set nb331_qqH, 336.set nb331_rinvO, 352.set nb331_rinvH1, 368.set nb331_rinvH2, 384.set nb331_rO, 400.set nb331_rH1, 416.set nb331_rH2, 432.set nb331_tsc, 448.set nb331_two, 464.set nb331_c6, 480.set nb331_c12, 496.set nb331_vctot, 512.set nb331_Vvdwtot, 528.set nb331_fixO, 544.set nb331_fiyO, 560.set nb331_fizO, 576.set nb331_fixH1, 592.set nb331_fiyH1, 608.set nb331_fizH1, 624.set nb331_fixH2, 640.set nb331_fiyH2, 656.set nb331_fizH2, 672.set nb331_fjx, 688.set nb331_fjy, 704.set nb331_fjz, 720.set nb331_half, 736.set nb331_three, 752.set nb331_is3, 768.set nb331_ii3, 772.set nb331_ntia, 776.set nb331_innerjjnr, 780.set nb331_innerk, 784.set nb331_n, 788.set nb331_nn1, 792.set nb331_nri, 796.set nb331_nouter, 800.set nb331_ninner, 804.set nb331_salign, 808 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $812,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb331_salign(%esp) emms ## Move args passed by reference to stack movl nb331_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb331_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb331_nouter(%esp) movl %eax,nb331_ninner(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb331_half(%esp) movss nb331_half(%esp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## constant 1.0 movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## constant 2.0 addps %xmm2,%xmm3 ## constant 3.0 movaps %xmm1,nb331_half(%esp) movaps %xmm2,nb331_two(%esp) movaps %xmm3,nb331_three(%esp) movl nb331_p_tabscale(%ebp),%eax movss (%eax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb331_tsc(%esp) movl nb331_p_tabscale(%ebp),%eax movss (%eax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb331_tsc(%esp) ## assume we have at least one i particle - start directly movl nb331_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb331_charge(%ebp),%edx movss (%edx,%ebx,4),%xmm3 movss 4(%edx,%ebx,4),%xmm4 movl nb331_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb331_iqO(%esp) movaps %xmm4,nb331_iqH(%esp) movl nb331_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb331_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb331_ntia(%esp)_nb_kernel331_ia32_sse.nb331_threadloop: movl nb331_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel331_ia32_sse.nb331_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_kernel331_ia32_sse.nb331_spinlock ## if(nn1>nri) nn1=nri movl nb331_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,nb331_n(%esp) movl %ebx,nb331_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel331_ia32_sse.nb331_outerstart jmp _nb_kernel331_ia32_sse.nb331_end_nb_kernel331_ia32_sse.nb331_outerstart: ## ebx contains number of outer iterations addl nb331_nouter(%esp),%ebx movl %ebx,nb331_nouter(%esp)_nb_kernel331_ia32_sse.nb331_outer: movl nb331_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,nb331_is3(%esp) ## store is3 movl nb331_shiftvec(%ebp),%eax ## eax = base of shiftvec[] movss (%eax,%ebx,4),%xmm0 movss 4(%eax,%ebx,4),%xmm1 movss 8(%eax,%ebx,4),%xmm2 movl nb331_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb331_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb331_ii3(%esp) addss (%eax,%ebx,4),%xmm3 addss 4(%eax,%ebx,4),%xmm4 addss 8(%eax,%ebx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb331_ixO(%esp) movaps %xmm4,nb331_iyO(%esp) movaps %xmm5,nb331_izO(%esp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%eax,%ebx,4),%xmm0 addss 16(%eax,%ebx,4),%xmm1 addss 20(%eax,%ebx,4),%xmm2 addss 24(%eax,%ebx,4),%xmm3 addss 28(%eax,%ebx,4),%xmm4 addss 32(%eax,%ebx,4),%xmm5 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm0,nb331_ixH1(%esp) movaps %xmm1,nb331_iyH1(%esp) movaps %xmm2,nb331_izH1(%esp) movaps %xmm3,nb331_ixH2(%esp) movaps %xmm4,nb331_iyH2(%esp) movaps %xmm5,nb331_izH2(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb331_vctot(%esp) movaps %xmm4,nb331_Vvdwtot(%esp) movaps %xmm4,nb331_fixO(%esp) movaps %xmm4,nb331_fiyO(%esp) movaps %xmm4,nb331_fizO(%esp) movaps %xmm4,nb331_fixH1(%esp) movaps %xmm4,nb331_fiyH1(%esp) movaps %xmm4,nb331_fizH1(%esp) movaps %xmm4,nb331_fixH2(%esp) movaps %xmm4,nb331_fiyH2(%esp) movaps %xmm4,nb331_fizH2(%esp) movl nb331_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 nb331_pos(%ebp),%esi movl nb331_faction(%ebp),%edi movl nb331_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb331_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb331_ninner(%esp),%ecx movl %ecx,nb331_ninner(%esp) addl $0,%edx movl %edx,nb331_innerk(%esp) ## number of innerloop atoms jge _nb_kernel331_ia32_sse.nb331_unroll_loop jmp _nb_kernel331_ia32_sse.nb331_odd_inner_nb_kernel331_ia32_sse.nb331_unroll_loop: ## quad-unroll innerloop here movl nb331_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx movl 8(%edx),%ecx movl 12(%edx),%edx ## eax-edx=jnr1-4 addl $16,nb331_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb331_charge(%ebp),%esi ## base of charge[] movss (%esi,%eax,4),%xmm3 movss (%esi,%ecx,4),%xmm4 movss (%esi,%ebx,4),%xmm6 movss (%esi,%edx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb331_iqO(%esp),%xmm3 mulps nb331_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb331_qqO(%esp) movaps %xmm4,nb331_qqH(%esp) movl nb331_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb331_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb331_ntia(%esp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%esi,%eax,4),%xmm6 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm6 movhps (%esi,%edx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb331_c6(%esp) movaps %xmm6,nb331_c12(%esp) movl nb331_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx leal (%ecx,%ecx,2),%ecx ## replace jnr with j3 leal (%edx,%edx,2),%edx ## move four coordinates to xmm0-xmm2 movlps (%esi,%eax,4),%xmm4 movlps (%esi,%ecx,4),%xmm5 movss 8(%esi,%eax,4),%xmm2 movss 8(%esi,%ecx,4),%xmm6 movhps (%esi,%ebx,4),%xmm4 movhps (%esi,%edx,4),%xmm5 movss 8(%esi,%ebx,4),%xmm0 movss 8(%esi,%edx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb331_ixO(%esp),%xmm4 movaps nb331_iyO(%esp),%xmm5 movaps nb331_izO(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb331_dxO(%esp) movaps %xmm5,nb331_dyO(%esp) movaps %xmm6,nb331_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?