nb_kernel103_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,426 行 · 第 1/3 页
S
1,426 行
#### $Id: nb_kernel103_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:48 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_kernel103_x86_64_sse.globl _nb_kernel103_x86_64_ssenb_kernel103_x86_64_sse: _nb_kernel103_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb103_fshift, 16.set nb103_gid, 24.set nb103_pos, 32.set nb103_faction, 40.set nb103_charge, 48.set nb103_p_facel, 56.set nb103_argkrf, 64.set nb103_argcrf, 72.set nb103_Vc, 80.set nb103_type, 88.set nb103_p_ntype, 96.set nb103_vdwparam, 104.set nb103_Vvdw, 112.set nb103_p_tabscale, 120.set nb103_VFtab, 128.set nb103_invsqrta, 136.set nb103_dvda, 144.set nb103_p_gbtabscale, 152.set nb103_GBtab, 160.set nb103_p_nthreads, 168.set nb103_count, 176.set nb103_mtx, 184.set nb103_outeriter, 192.set nb103_inneriter, 200.set nb103_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb103_ixH1, 0.set nb103_iyH1, 16.set nb103_izH1, 32.set nb103_ixH2, 48.set nb103_iyH2, 64.set nb103_izH2, 80.set nb103_ixM, 96.set nb103_iyM, 112.set nb103_izM, 128.set nb103_iqH, 144.set nb103_iqM, 160.set nb103_dxH1, 176.set nb103_dyH1, 192.set nb103_dzH1, 208.set nb103_dxH2, 224.set nb103_dyH2, 240.set nb103_dzH2, 256.set nb103_dxM, 272.set nb103_dyM, 288.set nb103_dzM, 304.set nb103_qqH, 320.set nb103_qqM, 336.set nb103_vctot, 352.set nb103_fixH1, 368.set nb103_fiyH1, 384.set nb103_fizH1, 400.set nb103_fixH2, 416.set nb103_fiyH2, 432.set nb103_fizH2, 448.set nb103_fixM, 464.set nb103_fiyM, 480.set nb103_fizM, 496.set nb103_fjx, 512.set nb103_fjy, 528.set nb103_fjz, 544.set nb103_half, 560.set nb103_three, 576.set nb103_is3, 592.set nb103_ii3, 596.set nb103_nri, 600.set nb103_innerjjnr, 608.set nb103_iinr, 616.set nb103_jindex, 624.set nb103_jjnr, 632.set nb103_shift, 640.set nb103_shiftvec, 648.set nb103_facel, 656.set nb103_innerk, 664.set nb103_n, 668.set nb103_nn1, 672.set nb103_nouter, 676.set nb103_ninner, 680 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 emms subq $696,%rsp # # local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb103_nouter(%rsp) movl %eax,nb103_ninner(%rsp) movl (%rdi),%edi movl %edi,nb103_nri(%rsp) movq %rsi,nb103_iinr(%rsp) movq %rdx,nb103_jindex(%rsp) movq %rcx,nb103_jjnr(%rsp) movq %r8,nb103_shift(%rsp) movq %r9,nb103_shiftvec(%rsp) movq nb103_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb103_facel(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb103_half(%rsp) movss nb103_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,nb103_half(%rsp) movaps %xmm3,nb103_three(%rsp) ## assume we have at least one i particle - start directly movq nb103_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb103_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm3 movss 12(%rdx,%rbx,4),%xmm4 movq nb103_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb103_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb103_iqH(%rsp) movaps %xmm4,nb103_iqM(%rsp)_nb_kernel103_x86_64_sse.nb103_threadloop: movq nb103_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel103_x86_64_sse.nb103_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_kernel103_x86_64_sse.nb103_spinlock ## if(nn1>nri) nn1=nri movl nb103_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,nb103_n(%rsp) movl %ebx,nb103_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel103_x86_64_sse.nb103_outerstart jmp _nb_kernel103_x86_64_sse.nb103_end_nb_kernel103_x86_64_sse.nb103_outerstart: ## ebx contains number of outer iterations addl nb103_nouter(%rsp),%ebx movl %ebx,nb103_nouter(%rsp)_nb_kernel103_x86_64_sse.nb103_outer: movq nb103_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,nb103_is3(%rsp) ## store is3 movq nb103_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 nb103_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 nb103_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb103_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,nb103_ixH1(%rsp) movaps %xmm4,nb103_iyH1(%rsp) movaps %xmm5,nb103_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,nb103_ixH2(%rsp) movaps %xmm1,nb103_iyH2(%rsp) movaps %xmm2,nb103_izH2(%rsp) movaps %xmm3,nb103_ixM(%rsp) movaps %xmm4,nb103_iyM(%rsp) movaps %xmm5,nb103_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb103_vctot(%rsp) movaps %xmm4,nb103_fixH1(%rsp) movaps %xmm4,nb103_fiyH1(%rsp) movaps %xmm4,nb103_fizH1(%rsp) movaps %xmm4,nb103_fixH2(%rsp) movaps %xmm4,nb103_fiyH2(%rsp) movaps %xmm4,nb103_fizH2(%rsp) movaps %xmm4,nb103_fixM(%rsp) movaps %xmm4,nb103_fiyM(%rsp) movaps %xmm4,nb103_fizM(%rsp) movq nb103_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 nb103_pos(%rbp),%rsi movq nb103_faction(%rbp),%rdi movq nb103_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb103_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb103_ninner(%rsp),%ecx movl %ecx,nb103_ninner(%rsp) addl $0,%edx movl %edx,nb103_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel103_x86_64_sse.nb103_unroll_loop jmp _nb_kernel103_x86_64_sse.nb103_odd_inner_nb_kernel103_x86_64_sse.nb103_unroll_loop: ## quad-unroll innerloop here movq nb103_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r8d movl 4(%rdx),%r9d movl 8(%rdx),%r10d movl 12(%rdx),%r11d ## eax-edx=jnr1-4 addq $16,nb103_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb103_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%r8,4),%xmm12 movss (%rsi,%r10,4),%xmm13 movss (%rsi,%r9,4),%xmm14 movss (%rsi,%r11,4),%xmm15 shufps $0,%xmm14,%xmm12 shufps $0,%xmm15,%xmm13 shufps $136,%xmm13,%xmm12 ## 10001000 ;# all charges in xmm3 movaps %xmm12,%xmm13 ## and in xmm4 mulps nb103_iqH(%rsp),%xmm12 mulps nb103_iqM(%rsp),%xmm13 movaps %xmm12,nb103_qqH(%rsp) movaps %xmm13,nb103_qqM(%rsp) movq nb103_pos(%rbp),%rsi ## base of pos[] lea (%r8,%r8,2),%rax ## j3 lea (%r9,%r9,2),%rbx lea (%r10,%r10,2),%rcx lea (%r11,%r11,2),%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 nb103_ixH1(%rsp),%xmm0 subps nb103_iyH1(%rsp),%xmm1 subps nb103_izH1(%rsp),%xmm2 subps nb103_ixH2(%rsp),%xmm3 subps nb103_iyH2(%rsp),%xmm4 subps nb103_izH2(%rsp),%xmm5 subps nb103_ixM(%rsp),%xmm6 subps nb103_iyM(%rsp),%xmm7 subps nb103_izM(%rsp),%xmm8 movaps %xmm0,nb103_dxH1(%rsp) movaps %xmm1,nb103_dyH1(%rsp) movaps %xmm2,nb103_dzH1(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb103_dxH2(%rsp) movaps %xmm4,nb103_dyH2(%rsp) movaps %xmm5,nb103_dzH2(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb103_dxM(%rsp) movaps %xmm7,nb103_dyM(%rsp) movaps %xmm8,nb103_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 nb103_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 nb103_half(%rsp),%xmm0 mulps %xmm0,%xmm9 ## rinvH1 mulps %xmm0,%xmm10 ## rinvH2 mulps %xmm0,%xmm11 ## rinvM ## interactions movaps %xmm9,%xmm0 movaps %xmm10,%xmm1 movaps %xmm11,%xmm2 mulps %xmm9,%xmm9 mulps %xmm10,%xmm10 mulps %xmm11,%xmm11 mulps nb103_qqH(%rsp),%xmm0 mulps nb103_qqH(%rsp),%xmm1 mulps nb103_qqM(%rsp),%xmm2 mulps %xmm0,%xmm9 mulps %xmm1,%xmm10 mulps %xmm2,%xmm11 addps nb103_vctot(%rsp),%xmm0 addps %xmm2,%xmm1 addps %xmm1,%xmm0 movaps %xmm0,nb103_vctot(%rsp) ## move j forces to local temp variables movlps (%rdi,%rax,4),%xmm0 ## jxa jya - - movlps (%rdi,%rcx,4),%xmm1 ## jxc jyc - - movhps (%rdi,%rbx,4),%xmm0 ## jxa jya jxb jyb movhps (%rdi,%rdx,4),%xmm1 ## jxc jyc jxd jyd movss 8(%rdi,%rax,4),%xmm2 ## jza - - - movss 8(%rdi,%rcx,4),%xmm3 ## jzc - - - movss 8(%rdi,%rbx,4),%xmm4 ## jzb movss 8(%rdi,%rdx,4),%xmm5 ## jzd movlhps %xmm4,%xmm2 ## jza - jzb - movlhps %xmm5,%xmm3 ## jzc - jzd - shufps $136,%xmm3,%xmm2 ## 10001000 => jza jzb jzc jzd ## xmm0: jxa jya jxb jyb ## xmm1: jxc jyc jxd jyd ## xmm2: jza jzb jzc jzd movaps %xmm9,%xmm7 movaps %xmm9,%xmm8 movaps %xmm11,%xmm13 movaps %xmm11,%xmm14 movaps %xmm11,%xmm15 movaps %xmm10,%xmm11 movaps %xmm10,%xmm12 mulps nb103_dxH1(%rsp),%xmm7 mulps nb103_dyH1(%rsp),%xmm8 mulps nb103_dzH1(%rsp),%xmm9⌨️ 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?