nb_kernel430_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,138 行 · 第 1/5 页
S
2,138 行
#### $Id: nb_kernel430_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_kernel430_ia32_sse.globl _nb_kernel430_ia32_ssenb_kernel430_ia32_sse: _nb_kernel430_ia32_sse: .set nb430_p_nri, 8.set nb430_iinr, 12.set nb430_jindex, 16.set nb430_jjnr, 20.set nb430_shift, 24.set nb430_shiftvec, 28.set nb430_fshift, 32.set nb430_gid, 36.set nb430_pos, 40.set nb430_faction, 44.set nb430_charge, 48.set nb430_p_facel, 52.set nb430_argkrf, 56.set nb430_argcrf, 60.set nb430_Vc, 64.set nb430_type, 68.set nb430_p_ntype, 72.set nb430_vdwparam, 76.set nb430_Vvdw, 80.set nb430_p_tabscale, 84.set nb430_VFtab, 88.set nb430_invsqrta, 92.set nb430_dvda, 96.set nb430_p_gbtabscale, 100.set nb430_GBtab, 104.set nb430_p_nthreads, 108.set nb430_count, 112.set nb430_mtx, 116.set nb430_outeriter, 120.set nb430_inneriter, 124.set nb430_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb430_ix, 0.set nb430_iy, 16.set nb430_iz, 32.set nb430_iq, 48.set nb430_dx, 64.set nb430_dy, 80.set nb430_dz, 96.set nb430_two, 112.set nb430_gbtsc, 128.set nb430_tsc, 144.set nb430_qq, 160.set nb430_c6, 176.set nb430_c12, 192.set nb430_fscal, 208.set nb430_vctot, 224.set nb430_Vvdwtot, 240.set nb430_fix, 256.set nb430_fiy, 272.set nb430_fiz, 288.set nb430_half, 304.set nb430_three, 320.set nb430_r, 336.set nb430_isai, 352.set nb430_isaprod, 368.set nb430_dvdasum, 384.set nb430_gbscale, 400.set nb430_ii, 416.set nb430_is3, 420.set nb430_ii3, 424.set nb430_ntia, 428.set nb430_innerjjnr, 432.set nb430_innerk, 436.set nb430_n, 440.set nb430_nn1, 444.set nb430_jnra, 448.set nb430_jnrb, 452.set nb430_jnrc, 456.set nb430_jnrd, 460.set nb430_nri, 464.set nb430_facel, 468.set nb430_ntype, 472.set nb430_nouter, 476.set nb430_ninner, 480.set nb430_salign, 484 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $488,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb430_salign(%esp) emms ## Move args passed by reference to stack movl nb430_p_nri(%ebp),%ecx movl nb430_p_facel(%ebp),%esi movl nb430_p_ntype(%ebp),%edi movl (%ecx),%ecx movl (%esi),%esi movl (%edi),%edi movl %ecx,nb430_nri(%esp) movl %esi,nb430_facel(%esp) movl %edi,nb430_ntype(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb430_nouter(%esp) movl %eax,nb430_ninner(%esp) movl nb430_p_gbtabscale(%ebp),%eax movss (%eax),%xmm3 movl nb430_p_tabscale(%ebp),%eax movss (%eax),%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb430_gbtsc(%esp) movaps %xmm4,nb430_tsc(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb430_half(%esp) movss nb430_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,nb430_half(%esp) movaps %xmm2,nb430_two(%esp) movaps %xmm3,nb430_three(%esp)_nb_kernel430_ia32_sse.nb430_threadloop: movl nb430_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel430_ia32_sse.nb430_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_kernel430_ia32_sse.nb430_spinlock ## if(nn1>nri) nn1=nri movl nb430_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,nb430_n(%esp) movl %ebx,nb430_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel430_ia32_sse.nb430_outerstart jmp _nb_kernel430_ia32_sse.nb430_end_nb_kernel430_ia32_sse.nb430_outerstart: ## ebx contains number of outer iterations addl nb430_nouter(%esp),%ebx movl %ebx,nb430_nouter(%esp)_nb_kernel430_ia32_sse.nb430_outer: movl nb430_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,nb430_is3(%esp) ## store is3 movl nb430_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 nb430_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movl %ebx,nb430_ii(%esp) movl nb430_charge(%ebp),%edx movss (%edx,%ebx,4),%xmm3 mulss nb430_facel(%esp),%xmm3 shufps $0,%xmm3,%xmm3 movl nb430_invsqrta(%ebp),%edx ## load invsqrta[ii] movss (%edx,%ebx,4),%xmm4 shufps $0,%xmm4,%xmm4 movl nb430_type(%ebp),%edx movl (%edx,%ebx,4),%edx imull nb430_ntype(%esp),%edx shll %edx movl %edx,nb430_ntia(%esp) leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb430_pos(%ebp),%eax ## eax = base of pos[] addss (%eax,%ebx,4),%xmm0 addss 4(%eax,%ebx,4),%xmm1 addss 8(%eax,%ebx,4),%xmm2 movaps %xmm3,nb430_iq(%esp) movaps %xmm4,nb430_isai(%esp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb430_ix(%esp) movaps %xmm1,nb430_iy(%esp) movaps %xmm2,nb430_iz(%esp) movl %ebx,nb430_ii3(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb430_vctot(%esp) movaps %xmm4,nb430_Vvdwtot(%esp) movaps %xmm4,nb430_dvdasum(%esp) movaps %xmm4,nb430_fix(%esp) movaps %xmm4,nb430_fiy(%esp) movaps %xmm4,nb430_fiz(%esp) movl nb430_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 nb430_pos(%ebp),%esi movl nb430_faction(%ebp),%edi movl nb430_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb430_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb430_ninner(%esp),%ecx movl %ecx,nb430_ninner(%esp) addl $0,%edx movl %edx,nb430_innerk(%esp) ## number of innerloop atoms jge _nb_kernel430_ia32_sse.nb430_unroll_loop jmp _nb_kernel430_ia32_sse.nb430_finish_inner_nb_kernel430_ia32_sse.nb430_unroll_loop: ## quad-unroll innerloop here movl nb430_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,nb430_innerjjnr(%esp) ## advance pointer (unrolled 4) ## load isaj movl nb430_invsqrta(%ebp),%esi movss (%esi,%eax,4),%xmm3 movss (%esi,%ecx,4),%xmm4 movss (%esi,%ebx,4),%xmm6 movss (%esi,%edx,4),%xmm7 movaps nb430_isai(%esp),%xmm2 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all isaj in xmm3 mulps %xmm3,%xmm2 movaps %xmm2,nb430_isaprod(%esp) movaps %xmm2,%xmm1 mulps nb430_gbtsc(%esp),%xmm1 movaps %xmm1,nb430_gbscale(%esp) movl nb430_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 mulps nb430_iq(%esp),%xmm2 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all charges in xmm3 mulps %xmm2,%xmm3 movaps %xmm3,nb430_qq(%esp) movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb430_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb430_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb430_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,nb430_c6(%esp) movaps %xmm6,nb430_c12(%esp) movl nb430_pos(%ebp),%esi ## base of pos[] movl %eax,nb430_jnra(%esp) movl %ebx,nb430_jnrb(%esp) movl %ecx,nb430_jnrc(%esp) movl %edx,nb430_jnrd(%esp) leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx leal (%ecx,%ecx,2),%ecx 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 ix-iz to xmm4-xmm6 movaps nb430_ix(%esp),%xmm4 movaps nb430_iy(%esp),%xmm5 movaps nb430_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb430_dx(%esp) movaps %xmm5,nb430_dy(%esp) movaps %xmm6,nb430_dz(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb430_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb430_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm4,nb430_r(%esp) mulps nb430_gbscale(%esp),%xmm4 movhlps %xmm4,%xmm5 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm5,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm6 cvtpi2ps %mm7,%xmm5 movlhps %xmm5,%xmm6 subps %xmm6,%xmm4⌨️ 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?