nb_kernel303_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,919 行 · 第 1/4 页
S
1,919 行
#### $Id: nb_kernel303_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:50 lindahl 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_kernel303_x86_64_sse.globl _nb_kernel303_x86_64_ssenb_kernel303_x86_64_sse: _nb_kernel303_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb303_fshift, 16.set nb303_gid, 24.set nb303_pos, 32.set nb303_faction, 40.set nb303_charge, 48.set nb303_p_facel, 56.set nb303_argkrf, 64.set nb303_argcrf, 72.set nb303_Vc, 80.set nb303_type, 88.set nb303_p_ntype, 96.set nb303_vdwparam, 104.set nb303_Vvdw, 112.set nb303_p_tabscale, 120.set nb303_VFtab, 128.set nb303_invsqrta, 136.set nb303_dvda, 144.set nb303_p_gbtabscale, 152.set nb303_GBtab, 160.set nb303_p_nthreads, 168.set nb303_count, 176.set nb303_mtx, 184.set nb303_outeriter, 192.set nb303_inneriter, 200.set nb303_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb303_ixH1, 0.set nb303_iyH1, 16.set nb303_izH1, 32.set nb303_ixH2, 48.set nb303_iyH2, 64.set nb303_izH2, 80.set nb303_ixM, 96.set nb303_iyM, 112.set nb303_izM, 128.set nb303_iqH, 144.set nb303_iqM, 160.set nb303_dxH1, 176.set nb303_dyH1, 192.set nb303_dzH1, 208.set nb303_dxH2, 224.set nb303_dyH2, 240.set nb303_dzH2, 256.set nb303_dxM, 272.set nb303_dyM, 288.set nb303_dzM, 304.set nb303_qqH, 320.set nb303_qqM, 336.set nb303_rinvH1, 352.set nb303_rinvH2, 368.set nb303_rinvM, 384.set nb303_rH1, 400.set nb303_rH2, 416.set nb303_rM, 432.set nb303_tsc, 448.set nb303_two, 464.set nb303_vctot, 480.set nb303_fixH1, 496.set nb303_fiyH1, 512.set nb303_fizH1, 528.set nb303_fixH2, 544.set nb303_fiyH2, 560.set nb303_fizH2, 576.set nb303_fixM, 592.set nb303_fiyM, 608.set nb303_fizM, 624.set nb303_epsH1, 640.set nb303_epsH2, 656.set nb303_epsM, 672.set nb303_half, 688.set nb303_three, 704.set nb303_is3, 720.set nb303_ii3, 724.set nb303_nri, 728.set nb303_iinr, 736.set nb303_jindex, 744.set nb303_jjnr, 752.set nb303_shift, 760.set nb303_shiftvec, 768.set nb303_facel, 776.set nb303_innerjjnr, 784.set nb303_innerk, 792.set nb303_n, 796.set nb303_nn1, 800.set nb303_nouter, 804.set nb303_ninner, 808 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 subq $824,%rsp ## local variable stack space (n*16+8) emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb303_nouter(%rsp) movl %eax,nb303_ninner(%rsp) movl (%rdi),%edi movl %edi,nb303_nri(%rsp) movq %rsi,nb303_iinr(%rsp) movq %rdx,nb303_jindex(%rsp) movq %rcx,nb303_jjnr(%rsp) movq %r8,nb303_shift(%rsp) movq %r9,nb303_shiftvec(%rsp) movq nb303_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb303_facel(%rsp) movq nb303_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb303_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb303_half(%rsp) movss nb303_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm1,nb303_half(%rsp) movaps %xmm2,nb303_two(%rsp) movaps %xmm3,nb303_three(%rsp) ## assume we have at least one i particle - start directly movq nb303_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb303_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm3 movss 12(%rdx,%rbx,4),%xmm4 movq nb303_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb303_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb303_iqH(%rsp) movaps %xmm4,nb303_iqM(%rsp)_nb_kernel303_x86_64_sse.nb303_threadloop: movq nb303_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel303_x86_64_sse.nb303_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%rsi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel303_x86_64_sse.nb303_spinlock ## if(nn1>nri) nn1=nri movl nb303_nri(%rsp),%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,nb303_n(%rsp) movl %ebx,nb303_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel303_x86_64_sse.nb303_outerstart jmp _nb_kernel303_x86_64_sse.nb303_end_nb_kernel303_x86_64_sse.nb303_outerstart: ## ebx contains number of outer iterations addl nb303_nouter(%rsp),%ebx movl %ebx,nb303_nouter(%rsp)_nb_kernel303_x86_64_sse.nb303_outer: movq nb303_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## rbx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb303_is3(%rsp) ## store is3 movq nb303_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb303_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb303_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb303_ii3(%rsp) addss 12(%rax,%rbx,4),%xmm3 addss 16(%rax,%rbx,4),%xmm4 addss 20(%rax,%rbx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb303_ixH1(%rsp) movaps %xmm4,nb303_iyH1(%rsp) movaps %xmm5,nb303_izH1(%rsp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 24(%rax,%rbx,4),%xmm0 addss 28(%rax,%rbx,4),%xmm1 addss 32(%rax,%rbx,4),%xmm2 addss 36(%rax,%rbx,4),%xmm3 addss 40(%rax,%rbx,4),%xmm4 addss 44(%rax,%rbx,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,nb303_ixH2(%rsp) movaps %xmm1,nb303_iyH2(%rsp) movaps %xmm2,nb303_izH2(%rsp) movaps %xmm3,nb303_ixM(%rsp) movaps %xmm4,nb303_iyM(%rsp) movaps %xmm5,nb303_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb303_vctot(%rsp) movaps %xmm4,nb303_fixH1(%rsp) movaps %xmm4,nb303_fiyH1(%rsp) movaps %xmm4,nb303_fizH1(%rsp) movaps %xmm4,nb303_fixH2(%rsp) movaps %xmm4,nb303_fiyH2(%rsp) movaps %xmm4,nb303_fizH2(%rsp) movaps %xmm4,nb303_fixM(%rsp) movaps %xmm4,nb303_fiyM(%rsp) movaps %xmm4,nb303_fizM(%rsp) movq nb303_jindex(%rsp),%rax movl (%rax,%rsi,4),%ecx ## jindex[n] movl 4(%rax,%rsi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movq nb303_pos(%rbp),%rsi movq nb303_faction(%rbp),%rdi movq nb303_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb303_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb303_ninner(%rsp),%ecx movl %ecx,nb303_ninner(%rsp) addl $0,%edx movl %edx,nb303_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel303_x86_64_sse.nb303_unroll_loop jmp _nb_kernel303_x86_64_sse.nb303_odd_inner_nb_kernel303_x86_64_sse.nb303_unroll_loop: ## quad-unroll innerloop here movq nb303_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb303_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb303_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb303_iqH(%rsp),%xmm3 mulps nb303_iqM(%rsp),%xmm4 movaps %xmm3,nb303_qqH(%rsp) movaps %xmm4,nb303_qqM(%rsp) movq nb303_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx ## move four coordinates to xmm0-xmm2 movlps (%rsi,%rax,4),%xmm4 movlps (%rsi,%rcx,4),%xmm5 movss 8(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm6 movhps (%rsi,%rbx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movss 8(%rsi,%rbx,4),%xmm0 movss 8(%rsi,%rdx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ## xmm0 = jx ## xmm1 = jy ## xmm2 = jz movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb303_ixH1(%rsp),%xmm0 subps nb303_iyH1(%rsp),%xmm1 subps nb303_izH1(%rsp),%xmm2 subps nb303_ixH2(%rsp),%xmm3 subps nb303_iyH2(%rsp),%xmm4 subps nb303_izH2(%rsp),%xmm5 subps nb303_ixM(%rsp),%xmm6 subps nb303_iyM(%rsp),%xmm7 subps nb303_izM(%rsp),%xmm8 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm0,nb303_dxH1(%rsp) movaps %xmm1,nb303_dyH1(%rsp) movaps %xmm2,nb303_dzH1(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb303_dxH2(%rsp) movaps %xmm4,nb303_dyH2(%rsp) movaps %xmm5,nb303_dzH2(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb303_dxM(%rsp) movaps %xmm7,nb303_dyM(%rsp) movaps %xmm8,nb303_dzM(%rsp) mulps %xmm6,%xmm6 mulps %xmm7,%xmm7 mulps %xmm8,%xmm8 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 addps %xmm7,%xmm6 addps %xmm8,%xmm6 ## start doing invsqrt for j atoms rsqrtps %xmm0,%xmm1 rsqrtps %xmm3,%xmm4 rsqrtps %xmm6,%xmm7 movaps %xmm1,%xmm2 movaps %xmm4,%xmm5 movaps %xmm7,%xmm8 mulps %xmm1,%xmm1 ## lu*lu mulps %xmm4,%xmm4 ## lu*lu mulps %xmm7,%xmm7 ## lu*lu movaps nb303_three(%rsp),%xmm9 movaps %xmm9,%xmm10 movaps %xmm9,%xmm11 mulps %xmm0,%xmm1 ## rsq*lu*lu mulps %xmm3,%xmm4 ## rsq*lu*lu mulps %xmm6,%xmm7 ## rsq*lu*lu subps %xmm1,%xmm9 subps %xmm4,%xmm10 subps %xmm7,%xmm11 ## 3-rsq*lu*lu mulps %xmm2,%xmm9 mulps %xmm5,%xmm10 mulps %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movaps nb303_half(%rsp),%xmm4 mulps %xmm4,%xmm9 ## rinvH1 mulps %xmm4,%xmm10 ## rinvH2 mulps %xmm4,%xmm11 ## rinvM movaps %xmm9,nb303_rinvH1(%rsp) movaps %xmm10,nb303_rinvH2(%rsp) movaps %xmm11,nb303_rinvM(%rsp) ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movaps nb303_tsc(%rsp),%xmm1 mulps %xmm9,%xmm0 ## r mulps %xmm10,%xmm3 mulps %xmm11,%xmm6 mulps %xmm1,%xmm0 ## rtab mulps %xmm1,%xmm3 mulps %xmm1,%xmm6 ## truncate and convert to integers cvttps2dq %xmm0,%xmm1 cvttps2dq %xmm3,%xmm4 cvttps2dq %xmm6,%xmm7 ## convert back to float cvtdq2ps %xmm1,%xmm2 cvtdq2ps %xmm4,%xmm5 cvtdq2ps %xmm7,%xmm8 ## multiply by 4 pslld $2,%xmm1 pslld $2,%xmm4 pslld $2,%xmm7 ## move to integer registers movhlps %xmm1,%xmm13 movhlps %xmm4,%xmm14 movhlps %xmm7,%xmm15⌨️ 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?