nb_kernel311_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,136 行 · 第 1/5 页
S
2,136 行
#### $Id: nb_kernel311_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:51 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_kernel311_x86_64_sse.globl _nb_kernel311_x86_64_ssenb_kernel311_x86_64_sse: _nb_kernel311_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb311_fshift, 16.set nb311_gid, 24.set nb311_pos, 32.set nb311_faction, 40.set nb311_charge, 48.set nb311_p_facel, 56.set nb311_argkrf, 64.set nb311_argcrf, 72.set nb311_Vc, 80.set nb311_type, 88.set nb311_p_ntype, 96.set nb311_vdwparam, 104.set nb311_Vvdw, 112.set nb311_p_tabscale, 120.set nb311_VFtab, 128.set nb311_invsqrta, 136.set nb311_dvda, 144.set nb311_p_gbtabscale, 152.set nb311_GBtab, 160.set nb311_p_nthreads, 168.set nb311_count, 176.set nb311_mtx, 184.set nb311_outeriter, 192.set nb311_inneriter, 200.set nb311_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb311_ixO, 0.set nb311_iyO, 16.set nb311_izO, 32.set nb311_ixH1, 48.set nb311_iyH1, 64.set nb311_izH1, 80.set nb311_ixH2, 96.set nb311_iyH2, 112.set nb311_izH2, 128.set nb311_iqO, 144.set nb311_iqH, 160.set nb311_dxO, 176.set nb311_dyO, 192.set nb311_dzO, 208.set nb311_dxH1, 224.set nb311_dyH1, 240.set nb311_dzH1, 256.set nb311_dxH2, 272.set nb311_dyH2, 288.set nb311_dzH2, 304.set nb311_qqO, 320.set nb311_qqH, 336.set nb311_rinvO, 352.set nb311_rinvH1, 368.set nb311_rinvH2, 384.set nb311_rO, 400.set nb311_rH1, 416.set nb311_rH2, 432.set nb311_tsc, 448.set nb311_fstmp, 464.set nb311_c6, 480.set nb311_c12, 496.set nb311_six, 512.set nb311_twelve, 528.set nb311_vctot, 544.set nb311_Vvdwtot, 560.set nb311_fixO, 576.set nb311_fiyO, 592.set nb311_fizO, 608.set nb311_fixH1, 624.set nb311_fiyH1, 640.set nb311_fizH1, 656.set nb311_fixH2, 672.set nb311_fiyH2, 688.set nb311_fizH2, 704.set nb311_epsO, 720.set nb311_epsH1, 736.set nb311_epsH2, 752.set nb311_half, 768.set nb311_three, 784.set nb311_is3, 800.set nb311_ii3, 804.set nb311_nri, 808.set nb311_iinr, 816.set nb311_jindex, 824.set nb311_jjnr, 832.set nb311_shift, 840.set nb311_shiftvec, 848.set nb311_facel, 856.set nb311_innerjjnr, 864.set nb311_ntia, 872.set nb311_innerk, 876.set nb311_n, 880.set nb311_nn1, 884.set nb311_nouter, 888.set nb311_ninner, 892 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $904,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb311_nouter(%rsp) movl %eax,nb311_ninner(%rsp) movl (%rdi),%edi movl %edi,nb311_nri(%rsp) movq %rsi,nb311_iinr(%rsp) movq %rdx,nb311_jindex(%rsp) movq %rcx,nb311_jjnr(%rsp) movq %r8,nb311_shift(%rsp) movq %r9,nb311_shiftvec(%rsp) movq nb311_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb311_facel(%rsp) movq nb311_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb311_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb311_half(%rsp) movss nb311_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 %xmm3,%xmm4 addps %xmm4,%xmm4 ## six movaps %xmm4,%xmm5 addps %xmm5,%xmm5 ## twelve movaps %xmm1,nb311_half(%rsp) movaps %xmm3,nb311_three(%rsp) movaps %xmm4,nb311_six(%rsp) movaps %xmm5,nb311_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb311_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb311_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb311_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb311_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb311_iqO(%rsp) movaps %xmm4,nb311_iqH(%rsp) movq nb311_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb311_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb311_ntia(%rsp)_nb_kernel311_x86_64_sse.nb311_threadloop: movq nb311_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel311_x86_64_sse.nb311_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_kernel311_x86_64_sse.nb311_spinlock ## if(nn1>nri) nn1=nri movl nb311_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,nb311_n(%rsp) movl %ebx,nb311_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel311_x86_64_sse.nb311_outerstart jmp _nb_kernel311_x86_64_sse.nb311_end_nb_kernel311_x86_64_sse.nb311_outerstart: ## ebx contains number of outer iterations addl nb311_nouter(%rsp),%ebx movl %ebx,nb311_nouter(%rsp)_nb_kernel311_x86_64_sse.nb311_outer: movq nb311_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,nb311_is3(%rsp) ## store is3 movq nb311_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 nb311_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 nb311_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb311_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 addss 4(%rax,%rbx,4),%xmm4 addss 8(%rax,%rbx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb311_ixO(%rsp) movaps %xmm4,nb311_iyO(%rsp) movaps %xmm5,nb311_izO(%rsp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%rax,%rbx,4),%xmm0 addss 16(%rax,%rbx,4),%xmm1 addss 20(%rax,%rbx,4),%xmm2 addss 24(%rax,%rbx,4),%xmm3 addss 28(%rax,%rbx,4),%xmm4 addss 32(%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,nb311_ixH1(%rsp) movaps %xmm1,nb311_iyH1(%rsp) movaps %xmm2,nb311_izH1(%rsp) movaps %xmm3,nb311_ixH2(%rsp) movaps %xmm4,nb311_iyH2(%rsp) movaps %xmm5,nb311_izH2(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb311_vctot(%rsp) movaps %xmm4,nb311_Vvdwtot(%rsp) movaps %xmm4,nb311_fixO(%rsp) movaps %xmm4,nb311_fiyO(%rsp) movaps %xmm4,nb311_fizO(%rsp) movaps %xmm4,nb311_fixH1(%rsp) movaps %xmm4,nb311_fiyH1(%rsp) movaps %xmm4,nb311_fizH1(%rsp) movaps %xmm4,nb311_fixH2(%rsp) movaps %xmm4,nb311_fiyH2(%rsp) movaps %xmm4,nb311_fizH2(%rsp) movq nb311_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 nb311_pos(%rbp),%rsi movq nb311_faction(%rbp),%rdi movq nb311_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb311_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb311_ninner(%rsp),%ecx movl %ecx,nb311_ninner(%rsp) addl $0,%edx movl %edx,nb311_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel311_x86_64_sse.nb311_unroll_loop jmp _nb_kernel311_x86_64_sse.nb311_odd_inner_nb_kernel311_x86_64_sse.nb311_unroll_loop: ## quad-unroll innerloop here movq nb311_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,nb311_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb311_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 nb311_iqO(%rsp),%xmm3 mulps nb311_iqH(%rsp),%xmm4 movaps %xmm3,nb311_qqO(%rsp) movaps %xmm4,nb311_qqH(%rsp) movq nb311_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movl (%rsi,%rcx,4),%r10d movl (%rsi,%rdx,4),%r11d movq nb311_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movl nb311_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d addl %edi,%r10d addl %edi,%r11d movlps (%rsi,%r8,4),%xmm6 movlps (%rsi,%r10,4),%xmm7 movhps (%rsi,%r9,4),%xmm6 movhps (%rsi,%r11,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm6 ## 11011101 movaps %xmm4,nb311_c6(%rsp) movaps %xmm6,nb311_c12(%rsp) movq nb311_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 ## replace jnr with j3 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 movd %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 subps nb311_ixO(%rsp),%xmm0 subps nb311_iyO(%rsp),%xmm1 subps nb311_izO(%rsp),%xmm2 subps nb311_ixH1(%rsp),%xmm3 subps nb311_iyH1(%rsp),%xmm4 subps nb311_izH1(%rsp),%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?