nb_kernel203_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,563 行 · 第 1/4 页
S
1,563 行
#### $Id: nb_kernel203_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_kernel203_ia32_sse.globl _nb_kernel203_ia32_ssenb_kernel203_ia32_sse: _nb_kernel203_ia32_sse: .set nb203_p_nri, 8.set nb203_iinr, 12.set nb203_jindex, 16.set nb203_jjnr, 20.set nb203_shift, 24.set nb203_shiftvec, 28.set nb203_fshift, 32.set nb203_gid, 36.set nb203_pos, 40.set nb203_faction, 44.set nb203_charge, 48.set nb203_p_facel, 52.set nb203_argkrf, 56.set nb203_argcrf, 60.set nb203_Vc, 64.set nb203_type, 68.set nb203_p_ntype, 72.set nb203_vdwparam, 76.set nb203_Vvdw, 80.set nb203_p_tabscale, 84.set nb203_VFtab, 88.set nb203_invsqrta, 92.set nb203_dvda, 96.set nb203_p_gbtabscale, 100.set nb203_GBtab, 104.set nb203_p_nthreads, 108.set nb203_count, 112.set nb203_mtx, 116.set nb203_outeriter, 120.set nb203_inneriter, 124.set nb203_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb203_ixH1, 0.set nb203_iyH1, 16.set nb203_izH1, 32.set nb203_ixH2, 48.set nb203_iyH2, 64.set nb203_izH2, 80.set nb203_ixM, 96.set nb203_iyM, 112.set nb203_izM, 128.set nb203_iqH, 144.set nb203_iqM, 160.set nb203_dxH1, 176.set nb203_dyH1, 192.set nb203_dzH1, 208.set nb203_dxH2, 224.set nb203_dyH2, 240.set nb203_dzH2, 256.set nb203_dxM, 272.set nb203_dyM, 288.set nb203_dzM, 304.set nb203_qqH, 320.set nb203_qqM, 336.set nb203_vctot, 352.set nb203_fixH1, 384.set nb203_fiyH1, 400.set nb203_fizH1, 416.set nb203_fixH2, 432.set nb203_fiyH2, 448.set nb203_fizH2, 464.set nb203_fixM, 480.set nb203_fiyM, 496.set nb203_fizM, 512.set nb203_fjx, 528.set nb203_fjy, 544.set nb203_fjz, 560.set nb203_half, 576.set nb203_three, 592.set nb203_two, 608.set nb203_krf, 624.set nb203_crf, 640.set nb203_krsqH1, 656.set nb203_krsqH2, 672.set nb203_krsqM, 688.set nb203_is3, 704.set nb203_ii3, 708.set nb203_innerjjnr, 712.set nb203_innerk, 716.set nb203_n, 720.set nb203_nn1, 724.set nb203_nri, 728.set nb203_nouter, 732.set nb203_ninner, 736.set nb203_salign, 740 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $744,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb203_salign(%esp) emms ## Move args passed by reference to stack movl nb203_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb203_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb203_nouter(%esp) movl %eax,nb203_ninner(%esp) movl nb203_argkrf(%ebp),%esi movl nb203_argcrf(%ebp),%edi movss (%esi),%xmm5 movss (%edi),%xmm6 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 movaps %xmm5,nb203_krf(%esp) movaps %xmm6,nb203_crf(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb203_half(%esp) movss nb203_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,nb203_half(%esp) movaps %xmm2,nb203_two(%esp) movaps %xmm3,nb203_three(%esp) ## assume we have at least one i particle - start directly movl nb203_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb203_charge(%ebp),%edx movss 4(%edx,%ebx,4),%xmm3 movss 12(%edx,%ebx,4),%xmm4 movl nb203_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb203_iqH(%esp) movaps %xmm4,nb203_iqM(%esp)_nb_kernel203_ia32_sse.nb203_threadloop: movl nb203_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel203_ia32_sse.nb203_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_kernel203_ia32_sse.nb203_spinlock ## if(nn1>nri) nn1=nri movl nb203_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,nb203_n(%esp) movl %ebx,nb203_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel203_ia32_sse.nb203_outerstart jmp _nb_kernel203_ia32_sse.nb203_end_nb_kernel203_ia32_sse.nb203_outerstart: ## ebx contains number of outer iterations addl nb203_nouter(%esp),%ebx movl %ebx,nb203_nouter(%esp)_nb_kernel203_ia32_sse.nb203_outer: movl nb203_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,nb203_is3(%esp) ## store is3 movl nb203_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 nb203_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 nb203_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb203_ii3(%esp) 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,nb203_ixH1(%esp) movaps %xmm4,nb203_iyH1(%esp) movaps %xmm5,nb203_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,nb203_ixH2(%esp) movaps %xmm1,nb203_iyH2(%esp) movaps %xmm2,nb203_izH2(%esp) movaps %xmm3,nb203_ixM(%esp) movaps %xmm4,nb203_iyM(%esp) movaps %xmm5,nb203_izM(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb203_vctot(%esp) movaps %xmm4,nb203_fixH1(%esp) movaps %xmm4,nb203_fiyH1(%esp) movaps %xmm4,nb203_fizH1(%esp) movaps %xmm4,nb203_fixH2(%esp) movaps %xmm4,nb203_fiyH2(%esp) movaps %xmm4,nb203_fizH2(%esp) movaps %xmm4,nb203_fixM(%esp) movaps %xmm4,nb203_fiyM(%esp) movaps %xmm4,nb203_fizM(%esp) movl nb203_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 nb203_pos(%ebp),%esi movl nb203_faction(%ebp),%edi movl nb203_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb203_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb203_ninner(%esp),%ecx movl %ecx,nb203_ninner(%esp) addl $0,%edx movl %edx,nb203_innerk(%esp) ## number of innerloop atoms jge _nb_kernel203_ia32_sse.nb203_unroll_loop jmp _nb_kernel203_ia32_sse.nb203_odd_inner_nb_kernel203_ia32_sse.nb203_unroll_loop: ## quad-unroll innerloop here movl nb203_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,nb203_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb203_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 nb203_iqH(%esp),%xmm3 mulps nb203_iqM(%esp),%xmm4 movaps %xmm3,nb203_qqH(%esp) movaps %xmm4,nb203_qqM(%esp) movl nb203_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 ixH1-izH1 to xmm4-xmm6 movaps nb203_ixH1(%esp),%xmm4 movaps nb203_iyH1(%esp),%xmm5 movaps nb203_izH1(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb203_dxH1(%esp) movaps %xmm5,nb203_dyH1(%esp) movaps %xmm6,nb203_dzH1(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqH1 in xmm7 ## move ixH2-izH2 to xmm4-xmm6 movaps nb203_ixH2(%esp),%xmm4 movaps nb203_iyH2(%esp),%xmm5 movaps nb203_izH2(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb203_dxH2(%esp) movaps %xmm5,nb203_dyH2(%esp) movaps %xmm6,nb203_dzH2(%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?