nb_kernel313_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,088 行 · 第 1/5 页
S
2,088 行
#### $Id: nb_kernel313_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_kernel313_ia32_sse.globl _nb_kernel313_ia32_ssenb_kernel313_ia32_sse: _nb_kernel313_ia32_sse: .set nb313_p_nri, 8.set nb313_iinr, 12.set nb313_jindex, 16.set nb313_jjnr, 20.set nb313_shift, 24.set nb313_shiftvec, 28.set nb313_fshift, 32.set nb313_gid, 36.set nb313_pos, 40.set nb313_faction, 44.set nb313_charge, 48.set nb313_p_facel, 52.set nb313_argkrf, 56.set nb313_argcrf, 60.set nb313_Vc, 64.set nb313_type, 68.set nb313_p_ntype, 72.set nb313_vdwparam, 76.set nb313_Vvdw, 80.set nb313_p_tabscale, 84.set nb313_VFtab, 88.set nb313_invsqrta, 92.set nb313_dvda, 96.set nb313_p_gbtabscale, 100.set nb313_GBtab, 104.set nb313_p_nthreads, 108.set nb313_count, 112.set nb313_mtx, 116.set nb313_outeriter, 120.set nb313_inneriter, 124.set nb313_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb313_ixO, 0.set nb313_iyO, 16.set nb313_izO, 32.set nb313_ixH1, 48.set nb313_iyH1, 64.set nb313_izH1, 80.set nb313_ixH2, 96.set nb313_iyH2, 112.set nb313_izH2, 128.set nb313_ixM, 144.set nb313_iyM, 160.set nb313_izM, 176.set nb313_iqM, 192.set nb313_iqH, 208.set nb313_dxO, 224.set nb313_dyO, 240.set nb313_dzO, 256.set nb313_dxH1, 272.set nb313_dyH1, 288.set nb313_dzH1, 304.set nb313_dxH2, 320.set nb313_dyH2, 336.set nb313_dzH2, 352.set nb313_dxM, 368.set nb313_dyM, 384.set nb313_dzM, 400.set nb313_qqM, 416.set nb313_qqH, 432.set nb313_rinvH1, 448.set nb313_rinvH2, 464.set nb313_rinvM, 480.set nb313_rH1, 496.set nb313_rH2, 512.set nb313_rM, 528.set nb313_tsc, 544.set nb313_two, 560.set nb313_c6, 576.set nb313_c12, 592.set nb313_six, 608.set nb313_twelve, 624.set nb313_vctot, 640.set nb313_Vvdwtot, 656.set nb313_fixO, 672.set nb313_fiyO, 688.set nb313_fizO, 704.set nb313_fixH1, 720.set nb313_fiyH1, 736.set nb313_fizH1, 752.set nb313_fixH2, 768.set nb313_fiyH2, 784.set nb313_fizH2, 800.set nb313_fixM, 816.set nb313_fiyM, 832.set nb313_fizM, 848.set nb313_fjx, 864.set nb313_fjy, 880.set nb313_fjz, 896.set nb313_half, 912.set nb313_three, 928.set nb313_is3, 944.set nb313_ii3, 948.set nb313_ntia, 952.set nb313_innerjjnr, 956.set nb313_innerk, 960.set nb313_n, 964.set nb313_nn1, 968.set nb313_nri, 972.set nb313_nouter, 976.set nb313_ninner, 980.set nb313_salign, 984 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $988,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb313_salign(%esp) emms ## Move args passed by reference to stack movl nb313_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb313_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb313_nouter(%esp) movl %eax,nb313_ninner(%esp) movl nb313_p_tabscale(%ebp),%eax movss (%eax),%xmm5 shufps $0,%xmm5,%xmm5 movaps %xmm5,nb313_tsc(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb313_half(%esp) movss nb313_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,nb313_half(%esp) movaps %xmm2,nb313_two(%esp) movaps %xmm3,nb313_three(%esp) movaps %xmm4,nb313_six(%esp) movaps %xmm5,nb313_twelve(%esp) ## assume we have at least one i particle - start directly movl nb313_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb313_charge(%ebp),%edx movss 4(%edx,%ebx,4),%xmm4 movss 12(%edx,%ebx,4),%xmm3 movl nb313_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb313_iqM(%esp) movaps %xmm4,nb313_iqH(%esp) movl nb313_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb313_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb313_ntia(%esp)_nb_kernel313_ia32_sse.nb313_threadloop: movl nb313_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel313_ia32_sse.nb313_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_kernel313_ia32_sse.nb313_spinlock ## if(nn1>nri) nn1=nri movl nb313_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,nb313_n(%esp) movl %ebx,nb313_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel313_ia32_sse.nb313_outerstart jmp _nb_kernel313_ia32_sse.nb313_end_nb_kernel313_ia32_sse.nb313_outerstart: ## ebx contains number of outer iterations addl nb313_nouter(%esp),%ebx movl %ebx,nb313_nouter(%esp)_nb_kernel313_ia32_sse.nb313_outer: movl nb313_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,nb313_is3(%esp) ## store is3 movl nb313_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 nb313_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb313_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb313_ii3(%esp) addss (%eax,%ebx,4),%xmm3 ## ox addss 4(%eax,%ebx,4),%xmm4 ## oy addss 8(%eax,%ebx,4),%xmm5 ## oz addss 12(%eax,%ebx,4),%xmm6 ## h1x addss 16(%eax,%ebx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb313_ixO(%esp) movaps %xmm4,nb313_iyO(%esp) movaps %xmm5,nb313_izO(%esp) movaps %xmm6,nb313_ixH1(%esp) movaps %xmm7,nb313_iyH1(%esp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%eax,%ebx,4),%xmm6 ## h1z addss 24(%eax,%ebx,4),%xmm0 ## h2x addss 28(%eax,%ebx,4),%xmm1 ## h2y addss 32(%eax,%ebx,4),%xmm2 ## h2z addss 36(%eax,%ebx,4),%xmm3 ## mx addss 40(%eax,%ebx,4),%xmm4 ## my addss 44(%eax,%ebx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 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 %xmm6,nb313_izH1(%esp) movaps %xmm0,nb313_ixH2(%esp) movaps %xmm1,nb313_iyH2(%esp) movaps %xmm2,nb313_izH2(%esp) movaps %xmm3,nb313_ixM(%esp) movaps %xmm4,nb313_iyM(%esp) movaps %xmm5,nb313_izM(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb313_vctot(%esp) movaps %xmm4,nb313_Vvdwtot(%esp) movaps %xmm4,nb313_fixO(%esp) movaps %xmm4,nb313_fiyO(%esp) movaps %xmm4,nb313_fizO(%esp) movaps %xmm4,nb313_fixH1(%esp) movaps %xmm4,nb313_fiyH1(%esp) movaps %xmm4,nb313_fizH1(%esp) movaps %xmm4,nb313_fixH2(%esp) movaps %xmm4,nb313_fiyH2(%esp) movaps %xmm4,nb313_fizH2(%esp) movaps %xmm4,nb313_fixM(%esp) movaps %xmm4,nb313_fiyM(%esp) movaps %xmm4,nb313_fizM(%esp) movl nb313_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 nb313_pos(%ebp),%esi movl nb313_faction(%ebp),%edi movl nb313_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb313_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb313_ninner(%esp),%ecx movl %ecx,nb313_ninner(%esp) addl $0,%edx movl %edx,nb313_innerk(%esp) ## number of innerloop atoms jge _nb_kernel313_ia32_sse.nb313_unroll_loop jmp _nb_kernel313_ia32_sse.nb313_odd_inner_nb_kernel313_ia32_sse.nb313_unroll_loop: ## quad-unroll innerloop here movl nb313_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,nb313_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb313_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 nb313_iqM(%esp),%xmm3 mulps nb313_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb313_qqM(%esp) movaps %xmm4,nb313_qqH(%esp) movl nb313_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb313_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb313_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,nb313_c6(%esp) movaps %xmm6,nb313_c12(%esp) movl nb313_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⌨️ 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?