nb_kernel311_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,112 行 · 第 1/5 页
S
2,112 行
#### $Id: nb_kernel311_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_kernel311_ia32_sse.globl _nb_kernel311_ia32_ssenb_kernel311_ia32_sse: _nb_kernel311_ia32_sse: .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 sse 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) movl nb311_p_tabscale(%ebp),%eax movss (%eax),%xmm5 shufps $0,%xmm5,%xmm5 movaps %xmm5,nb311_tsc(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb311_half(%esp) movss nb311_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 %xmm3,%xmm4 addps %xmm4,%xmm4 ## 6.0 movaps %xmm4,%xmm5 addps %xmm5,%xmm5 ## constant 12.0 movaps %xmm1,nb311_half(%esp) movaps %xmm2,nb311_two(%esp) movaps %xmm3,nb311_three(%esp) movaps %xmm4,nb311_six(%esp) movaps %xmm5,nb311_twelve(%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 movss (%edx,%ebx,4),%xmm3 movss 4(%edx,%ebx,4),%xmm4 movl nb311_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb311_iqO(%esp) movaps %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_sse.nb311_threadloop: movl nb311_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel311_ia32_sse.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_sse.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_sse.nb311_outerstart jmp _nb_kernel311_ia32_sse.nb311_end_nb_kernel311_ia32_sse.nb311_outerstart: ## ebx contains number of outer iterations addl nb311_nouter(%esp),%ebx movl %ebx,nb311_nouter(%esp)_nb_kernel311_ia32_sse.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[] movss (%eax,%ebx,4),%xmm0 movss 4(%eax,%ebx,4),%xmm1 movss 8(%eax,%ebx,4),%xmm2 movl nb311_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 nb311_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb311_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,nb311_ixO(%esp) movaps %xmm4,nb311_iyO(%esp) movaps %xmm5,nb311_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,nb311_ixH1(%esp) movaps %xmm1,nb311_iyH1(%esp) movaps %xmm2,nb311_izH1(%esp) movaps %xmm3,nb311_ixH2(%esp) movaps %xmm4,nb311_iyH2(%esp) movaps %xmm5,nb311_izH2(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb311_vctot(%esp) movaps %xmm4,nb311_Vvdwtot(%esp) movaps %xmm4,nb311_fixO(%esp) movaps %xmm4,nb311_fiyO(%esp) movaps %xmm4,nb311_fizO(%esp) movaps %xmm4,nb311_fixH1(%esp) movaps %xmm4,nb311_fiyH1(%esp) movaps %xmm4,nb311_fizH1(%esp) movaps %xmm4,nb311_fixH2(%esp) movaps %xmm4,nb311_fiyH2(%esp) movaps %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 $4,%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_sse.nb311_unroll_loop jmp _nb_kernel311_ia32_sse.nb311_odd_inner_nb_kernel311_ia32_sse.nb311_unroll_loop: ## quad-unroll innerloop here movl nb311_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,nb311_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb311_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 nb311_iqO(%esp),%xmm3 mulps nb311_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb311_qqO(%esp) movaps %xmm4,nb311_qqH(%esp) movl nb311_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb311_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb311_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,nb311_c6(%esp) movaps %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 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 nb311_ixO(%esp),%xmm4 movaps nb311_iyO(%esp),%xmm5 movaps nb311_izO(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb311_dxO(%esp) movaps %xmm5,nb311_dyO(%esp) movaps %xmm6,nb311_dzO(%esp)⌨️ 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?