nb_kernel102_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,986 行 · 第 1/5 页
S
1,986 行
#### $Id: nb_kernel102_ia32_sse.s,v 1.1.2.2 2007/11/28 09:36:32 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_kernel102_ia32_sse.globl _nb_kernel102_ia32_ssenb_kernel102_ia32_sse: _nb_kernel102_ia32_sse: .set nb102_p_nri, 8.set nb102_iinr, 12.set nb102_jindex, 16.set nb102_jjnr, 20.set nb102_shift, 24.set nb102_shiftvec, 28.set nb102_fshift, 32.set nb102_gid, 36.set nb102_pos, 40.set nb102_faction, 44.set nb102_charge, 48.set nb102_p_facel, 52.set nb102_p_krf, 56.set nb102_p_crf, 60.set nb102_Vc, 64.set nb102_type, 68.set nb102_p_ntype, 72.set nb102_vdwparam, 76.set nb102_Vvdw, 80.set nb102_p_tabscale, 84.set nb102_VFtab, 88.set nb102_invsqrta, 92.set nb102_dvda, 96.set nb102_p_gbtabscale, 100.set nb102_GBtab, 104.set nb102_p_nthreads, 108.set nb102_count, 112.set nb102_mtx, 116.set nb102_outeriter, 120.set nb102_inneriter, 124.set nb102_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb102_ixO, 0.set nb102_iyO, 16.set nb102_izO, 32.set nb102_ixH1, 48.set nb102_iyH1, 64.set nb102_izH1, 80.set nb102_ixH2, 96.set nb102_iyH2, 112.set nb102_izH2, 128.set nb102_jxO, 144.set nb102_jyO, 160.set nb102_jzO, 176.set nb102_jxH1, 192.set nb102_jyH1, 208.set nb102_jzH1, 224.set nb102_jxH2, 240.set nb102_jyH2, 256.set nb102_jzH2, 272.set nb102_dxOO, 288.set nb102_dyOO, 304.set nb102_dzOO, 320.set nb102_dxOH1, 336.set nb102_dyOH1, 352.set nb102_dzOH1, 368.set nb102_dxOH2, 384.set nb102_dyOH2, 400.set nb102_dzOH2, 416.set nb102_dxH1O, 432.set nb102_dyH1O, 448.set nb102_dzH1O, 464.set nb102_dxH1H1, 480.set nb102_dyH1H1, 496.set nb102_dzH1H1, 512.set nb102_dxH1H2, 528.set nb102_dyH1H2, 544.set nb102_dzH1H2, 560.set nb102_dxH2O, 576.set nb102_dyH2O, 592.set nb102_dzH2O, 608.set nb102_dxH2H1, 624.set nb102_dyH2H1, 640.set nb102_dzH2H1, 656.set nb102_dxH2H2, 672.set nb102_dyH2H2, 688.set nb102_dzH2H2, 704.set nb102_qqOO, 720.set nb102_qqOH, 736.set nb102_qqHH, 752.set nb102_vctot, 768.set nb102_fixO, 784.set nb102_fiyO, 800.set nb102_fizO, 816.set nb102_fixH1, 832.set nb102_fiyH1, 848.set nb102_fizH1, 864.set nb102_fixH2, 880.set nb102_fiyH2, 896.set nb102_fizH2, 912.set nb102_fjxO, 928.set nb102_fjyO, 944.set nb102_fjzO, 960.set nb102_fjxH1, 976.set nb102_fjyH1, 992.set nb102_fjzH1, 1008.set nb102_fjxH2, 1024.set nb102_fjyH2, 1040.set nb102_fjzH2, 1056.set nb102_fjzH2b, 1060.set nb102_fjzH2c, 1064.set nb102_fjzH2d, 1068.set nb102_half, 1072.set nb102_three, 1088.set nb102_rsqOO, 1104.set nb102_rsqOH1, 1120.set nb102_rsqOH2, 1136.set nb102_rsqH1O, 1152.set nb102_rsqH1H1, 1168.set nb102_rsqH1H2, 1184.set nb102_rsqH2O, 1200.set nb102_rsqH2H1, 1216.set nb102_rsqH2H2, 1232.set nb102_rinvOO, 1248.set nb102_rinvOH1, 1264.set nb102_rinvOH2, 1280.set nb102_rinvH1O, 1296.set nb102_rinvH1H1, 1312.set nb102_rinvH1H2, 1328.set nb102_rinvH2O, 1344.set nb102_rinvH2H1, 1360.set nb102_rinvH2H2, 1376.set nb102_is3, 1392.set nb102_ii3, 1396.set nb102_innerjjnr, 1400.set nb102_innerk, 1404.set nb102_n, 1408.set nb102_nn1, 1412.set nb102_nri, 1416.set nb102_nouter, 1420.set nb102_ninner, 1424.set nb102_salign, 1428 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $1432,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb102_salign(%esp) emms ## Move args passed by reference to stack movl nb102_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb102_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb102_nouter(%esp) movl %eax,nb102_ninner(%esp) ## assume we have at least one i particle - start directly movl nb102_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb102_charge(%ebp),%edx movss (%edx,%ebx,4),%xmm3 movss %xmm3,%xmm4 movss 4(%edx,%ebx,4),%xmm5 movl nb102_p_facel(%ebp),%esi movss (%esi),%xmm6 mulss %xmm3,%xmm3 mulss %xmm5,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm3 mulss %xmm6,%xmm4 mulss %xmm6,%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb102_qqOO(%esp) movaps %xmm4,nb102_qqOH(%esp) movaps %xmm5,nb102_qqHH(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb102_half(%esp) movss nb102_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,nb102_half(%esp) movaps %xmm3,nb102_three(%esp)_nb_kernel102_ia32_sse.nb102_threadloop: movl nb102_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel102_ia32_sse.nb102_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_kernel102_ia32_sse.nb102_spinlock ## if(nn1>nri) nn1=nri movl nb102_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,nb102_n(%esp) movl %ebx,nb102_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel102_ia32_sse.nb102_outerstart jmp _nb_kernel102_ia32_sse.nb102_end_nb_kernel102_ia32_sse.nb102_outerstart: ## ebx contains number of outer iterations addl nb102_nouter(%esp),%ebx movl %ebx,nb102_nouter(%esp)_nb_kernel102_ia32_sse.nb102_outer: movl nb102_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,nb102_is3(%esp) ## store is3 movl nb102_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 nb102_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb102_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb102_ii3(%esp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 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,nb102_ixO(%esp) movaps %xmm4,nb102_iyO(%esp) movaps %xmm5,nb102_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,nb102_ixH1(%esp) movaps %xmm1,nb102_iyH1(%esp) movaps %xmm2,nb102_izH1(%esp) movaps %xmm3,nb102_ixH2(%esp) movaps %xmm4,nb102_iyH2(%esp) movaps %xmm5,nb102_izH2(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb102_vctot(%esp) movaps %xmm4,nb102_fixO(%esp) movaps %xmm4,nb102_fiyO(%esp) movaps %xmm4,nb102_fizO(%esp) movaps %xmm4,nb102_fixH1(%esp) movaps %xmm4,nb102_fiyH1(%esp) movaps %xmm4,nb102_fizH1(%esp) movaps %xmm4,nb102_fixH2(%esp) movaps %xmm4,nb102_fiyH2(%esp) movaps %xmm4,nb102_fizH2(%esp) movl nb102_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 nb102_pos(%ebp),%esi movl nb102_faction(%ebp),%edi movl nb102_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb102_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb102_ninner(%esp),%ecx movl %ecx,nb102_ninner(%esp) addl $0,%edx movl %edx,nb102_innerk(%esp) ## number of innerloop atoms jge _nb_kernel102_ia32_sse.nb102_unroll_loop jmp _nb_kernel102_ia32_sse.nb102_single_check_nb_kernel102_ia32_sse.nb102_unroll_loop: ## quad-unroll innerloop here movl nb102_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,nb102_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb102_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 j coordinates to local temp variables movlps (%esi,%eax,4),%xmm2 movlps 12(%esi,%eax,4),%xmm3 movlps 24(%esi,%eax,4),%xmm4 movlps (%esi,%ebx,4),%xmm5 movlps 12(%esi,%ebx,4),%xmm6 movlps 24(%esi,%ebx,4),%xmm7 movhps (%esi,%ecx,4),%xmm2 movhps 12(%esi,%ecx,4),%xmm3 movhps 24(%esi,%ecx,4),%xmm4 movhps (%esi,%edx,4),%xmm5 movhps 12(%esi,%edx,4),%xmm6 movhps 24(%esi,%edx,4),%xmm7 ## current state: ## xmm2= jxOa jyOa jxOc jyOc ## xmm3= jxH1a jyH1a jxH1c jyH1c ## xmm4= jxH2a jyH2a jxH2c jyH2c ## xmm5= jxOb jyOb jxOd jyOd ## xmm6= jxH1b jyH1b jxH1d jyH1d ## xmm7= jxH2b jyH2b jxH2d jyH2d movaps %xmm2,%xmm0 movaps %xmm3,%xmm1 unpcklps %xmm5,%xmm0 ## xmm0= jxOa jxOb jyOa jyOb unpcklps %xmm6,%xmm1 ## xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps %xmm5,%xmm2 ## xmm2= jxOc jxOd jyOc jyOd unpckhps %xmm6,%xmm3 ## xmm3= jxH1c jxH1d jyH1c jyH1d movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 ## xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps %xmm7,%xmm5 ## xmm5= jxH2c jxH2d jyH2c jyH2d movaps %xmm1,%xmm7 movlhps %xmm2,%xmm0 ## xmm0= jxOa jxOb jxOc jxOd movaps %xmm0,nb102_jxO(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb102_jyO(%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?