nb_kernel104_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,968 行 · 第 1/5 页
S
1,968 行
#### $Id: nb_kernel104_ia32_sse.s,v 1.1.2.2 2007/11/28 09:36:33 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_kernel104_ia32_sse.globl _nb_kernel104_ia32_ssenb_kernel104_ia32_sse: _nb_kernel104_ia32_sse: .set nb104_p_nri, 8.set nb104_iinr, 12.set nb104_jindex, 16.set nb104_jjnr, 20.set nb104_shift, 24.set nb104_shiftvec, 28.set nb104_fshift, 32.set nb104_gid, 36.set nb104_pos, 40.set nb104_faction, 44.set nb104_charge, 48.set nb104_p_facel, 52.set nb104_p_krf, 56.set nb104_p_crf, 60.set nb104_Vc, 64.set nb104_type, 68.set nb104_p_ntype, 72.set nb104_vdwparam, 76.set nb104_Vvdw, 80.set nb104_p_tabscale, 84.set nb104_VFtab, 88.set nb104_invsqrta, 92.set nb104_dvda, 96.set nb104_p_gbtabscale, 100.set nb104_GBtab, 104.set nb104_p_nthreads, 108.set nb104_count, 112.set nb104_mtx, 116.set nb104_outeriter, 120.set nb104_inneriter, 124.set nb104_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb104_ixH1, 0.set nb104_iyH1, 16.set nb104_izH1, 32.set nb104_ixH2, 48.set nb104_iyH2, 64.set nb104_izH2, 80.set nb104_ixM, 96.set nb104_iyM, 112.set nb104_izM, 128.set nb104_jxH1, 144.set nb104_jyH1, 160.set nb104_jzH1, 176.set nb104_jxH2, 192.set nb104_jyH2, 208.set nb104_jzH2, 224.set nb104_jxM, 240.set nb104_jyM, 256.set nb104_jzM, 272.set nb104_dxH1H1, 288.set nb104_dyH1H1, 304.set nb104_dzH1H1, 320.set nb104_dxH1H2, 336.set nb104_dyH1H2, 352.set nb104_dzH1H2, 368.set nb104_dxH1M, 384.set nb104_dyH1M, 400.set nb104_dzH1M, 416.set nb104_dxH2H1, 432.set nb104_dyH2H1, 448.set nb104_dzH2H1, 464.set nb104_dxH2H2, 480.set nb104_dyH2H2, 496.set nb104_dzH2H2, 512.set nb104_dxH2M, 528.set nb104_dyH2M, 544.set nb104_dzH2M, 560.set nb104_dxMH1, 576.set nb104_dyMH1, 592.set nb104_dzMH1, 608.set nb104_dxMH2, 624.set nb104_dyMH2, 640.set nb104_dzMH2, 656.set nb104_dxMM, 672.set nb104_dyMM, 688.set nb104_dzMM, 704.set nb104_qqHH, 720.set nb104_qqMH, 736.set nb104_qqMM, 752.set nb104_vctot, 768.set nb104_fixH1, 784.set nb104_fiyH1, 800.set nb104_fizH1, 816.set nb104_fixH2, 832.set nb104_fiyH2, 848.set nb104_fizH2, 864.set nb104_fixM, 880.set nb104_fiyM, 896.set nb104_fizM, 912.set nb104_fjxH1, 928.set nb104_fjyH1, 944.set nb104_fjzH1, 960.set nb104_fjxH2, 976.set nb104_fjyH2, 992.set nb104_fjzH2, 1008.set nb104_fjxM, 1024.set nb104_fjyM, 1040.set nb104_fjzM, 1056.set nb104_fjzMb, 1060.set nb104_fjzMc, 1064.set nb104_fjzMd, 1068.set nb104_half, 1072.set nb104_three, 1088.set nb104_rsqH1H1, 1104.set nb104_rsqH1H2, 1120.set nb104_rsqH1M, 1136.set nb104_rsqH2H1, 1152.set nb104_rsqH2H2, 1168.set nb104_rsqH2M, 1184.set nb104_rsqMH1, 1200.set nb104_rsqMH2, 1216.set nb104_rsqMM, 1232.set nb104_rinvH1H1, 1248.set nb104_rinvH1H2, 1264.set nb104_rinvH1M, 1280.set nb104_rinvH2H1, 1296.set nb104_rinvH2H2, 1312.set nb104_rinvH2M, 1328.set nb104_rinvMH1, 1344.set nb104_rinvMH2, 1360.set nb104_rinvMM, 1376.set nb104_is3, 1392.set nb104_ii3, 1396.set nb104_innerjjnr, 1400.set nb104_innerk, 1404.set nb104_n, 1408.set nb104_nn1, 1412.set nb104_nri, 1416.set nb104_nouter, 1420.set nb104_ninner, 1424.set nb104_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,nb104_salign(%esp) emms ## Move args passed by reference to stack movl nb104_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb104_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb104_nouter(%esp) movl %eax,nb104_ninner(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb104_half(%esp) movss nb104_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,nb104_half(%esp) movaps %xmm3,nb104_three(%esp) ## assume we have at least one i particle - start directly movl nb104_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb104_charge(%ebp),%edx movss 4(%edx,%ebx,4),%xmm3 movss %xmm3,%xmm4 movss 12(%edx,%ebx,4),%xmm5 movl nb104_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,nb104_qqHH(%esp) movaps %xmm4,nb104_qqMH(%esp) movaps %xmm5,nb104_qqMM(%esp)_nb_kernel104_ia32_sse.nb104_threadloop: movl nb104_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel104_ia32_sse.nb104_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_kernel104_ia32_sse.nb104_spinlock ## if(nn1>nri) nn1=nri movl nb104_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,nb104_n(%esp) movl %ebx,nb104_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel104_ia32_sse.nb104_outerstart jmp _nb_kernel104_ia32_sse.nb104_end_nb_kernel104_ia32_sse.nb104_outerstart: ## ebx contains number of outer iterations addl nb104_nouter(%esp),%ebx movl %ebx,nb104_nouter(%esp)_nb_kernel104_ia32_sse.nb104_outer: movl nb104_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,nb104_is3(%esp) ## store is3 movl nb104_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 nb104_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb104_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb104_ii3(%esp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 addss 12(%eax,%ebx,4),%xmm3 addss 16(%eax,%ebx,4),%xmm4 addss 20(%eax,%ebx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb104_ixH1(%esp) movaps %xmm4,nb104_iyH1(%esp) movaps %xmm5,nb104_izH1(%esp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 24(%eax,%ebx,4),%xmm0 addss 28(%eax,%ebx,4),%xmm1 addss 32(%eax,%ebx,4),%xmm2 addss 36(%eax,%ebx,4),%xmm3 addss 40(%eax,%ebx,4),%xmm4 addss 44(%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,nb104_ixH2(%esp) movaps %xmm1,nb104_iyH2(%esp) movaps %xmm2,nb104_izH2(%esp) movaps %xmm3,nb104_ixM(%esp) movaps %xmm4,nb104_iyM(%esp) movaps %xmm5,nb104_izM(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb104_vctot(%esp) movaps %xmm4,nb104_fixH1(%esp) movaps %xmm4,nb104_fiyH1(%esp) movaps %xmm4,nb104_fizH1(%esp) movaps %xmm4,nb104_fixH2(%esp) movaps %xmm4,nb104_fiyH2(%esp) movaps %xmm4,nb104_fizH2(%esp) movaps %xmm4,nb104_fixM(%esp) movaps %xmm4,nb104_fiyM(%esp) movaps %xmm4,nb104_fizM(%esp) movl nb104_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 nb104_pos(%ebp),%esi movl nb104_faction(%ebp),%edi movl nb104_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb104_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb104_ninner(%esp),%ecx movl %ecx,nb104_ninner(%esp) addl $0,%edx movl %edx,nb104_innerk(%esp) ## number of innerloop atoms jge _nb_kernel104_ia32_sse.nb104_unroll_loop jmp _nb_kernel104_ia32_sse.nb104_single_check_nb_kernel104_ia32_sse.nb104_unroll_loop: ## quad-unroll innerloop here movl nb104_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,nb104_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb104_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 12(%esi,%eax,4),%xmm2 movlps 24(%esi,%eax,4),%xmm3 movlps 36(%esi,%eax,4),%xmm4 movlps 12(%esi,%ebx,4),%xmm5 movlps 24(%esi,%ebx,4),%xmm6 movlps 36(%esi,%ebx,4),%xmm7 movhps 12(%esi,%ecx,4),%xmm2 movhps 24(%esi,%ecx,4),%xmm3 movhps 36(%esi,%ecx,4),%xmm4 movhps 12(%esi,%edx,4),%xmm5 movhps 24(%esi,%edx,4),%xmm6 movhps 36(%esi,%edx,4),%xmm7 ## current state: ## xmm2= jxh1a jyH1a jxH1c jyH1c ## xmm3= jxH2a jyH2a jxH2c jyH2c ## xmm4= jxMa jyMa jxMc jyMc ## xmm5= jxH1b jyH1b jxH1d jyH1d ## xmm6= jxH2b jyH2b jxH2d jyH2d ## xmm7= jxMb jyMb jxMd jyMd movaps %xmm2,%xmm0 movaps %xmm3,%xmm1 unpcklps %xmm5,%xmm0 ## xmm0= jxH1a jxH1b jyH1a jyH1b unpcklps %xmm6,%xmm1 ## xmm1= jxH2a jxH2b jyH2a jyH2b unpckhps %xmm5,%xmm2 ## xmm2= jxH1c jxH1d jyH1c jyH1d unpckhps %xmm6,%xmm3 ## xmm3= jxH2c jxH2d jyH2c jyH2d movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 ## xmm4= jxMa jxMb jyMa jyMb unpckhps %xmm7,%xmm5 ## xmm5= jxMc jxMd jyMc jyMd movaps %xmm1,%xmm7⌨️ 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?