nb_kernel111_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,633 行 · 第 1/4 页
S
1,633 行
#### $Id: nb_kernel111_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_kernel111_x86_64_sse.globl _nb_kernel111_x86_64_ssenb_kernel111_x86_64_sse: _nb_kernel111_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb111_fshift, 16.set nb111_gid, 24.set nb111_pos, 32.set nb111_faction, 40.set nb111_charge, 48.set nb111_p_facel, 56.set nb111_argkrf, 64.set nb111_argcrf, 72.set nb111_Vc, 80.set nb111_type, 88.set nb111_p_ntype, 96.set nb111_vdwparam, 104.set nb111_Vvdw, 112.set nb111_p_tabscale, 120.set nb111_VFtab, 128.set nb111_invsqrta, 136.set nb111_dvda, 144.set nb111_p_gbtabscale, 152.set nb111_GBtab, 160.set nb111_p_nthreads, 168.set nb111_count, 176.set nb111_mtx, 184.set nb111_outeriter, 192.set nb111_inneriter, 200.set nb111_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb111_ixO, 0.set nb111_iyO, 16.set nb111_izO, 32.set nb111_ixH1, 48.set nb111_iyH1, 64.set nb111_izH1, 80.set nb111_ixH2, 96.set nb111_iyH2, 112.set nb111_izH2, 128.set nb111_iqO, 144.set nb111_iqH, 160.set nb111_dxO, 176.set nb111_dyO, 192.set nb111_dzO, 208.set nb111_dxH1, 224.set nb111_dyH1, 240.set nb111_dzH1, 256.set nb111_dxH2, 272.set nb111_dyH2, 288.set nb111_dzH2, 304.set nb111_qqO, 320.set nb111_qqH, 336.set nb111_c6, 352.set nb111_c12, 368.set nb111_six, 384.set nb111_twelve, 400.set nb111_vctot, 416.set nb111_Vvdwtot, 432.set nb111_fixO, 448.set nb111_fiyO, 464.set nb111_fizO, 480.set nb111_fixH1, 496.set nb111_fiyH1, 512.set nb111_fizH1, 528.set nb111_fixH2, 544.set nb111_fiyH2, 560.set nb111_fizH2, 576.set nb111_fjx, 592.set nb111_fjy, 608.set nb111_fjz, 624.set nb111_half, 640.set nb111_three, 656.set nb111_is3, 672.set nb111_ii3, 676.set nb111_nri, 692.set nb111_iinr, 700.set nb111_jindex, 708.set nb111_jjnr, 716.set nb111_shift, 724.set nb111_shiftvec, 732.set nb111_facel, 740.set nb111_innerjjnr, 748.set nb111_ntia, 756.set nb111_innerk, 760.set nb111_n, 764.set nb111_nn1, 768.set nb111_nouter, 772.set nb111_ninner, 776 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 subq $792,%rsp emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb111_nouter(%rsp) movl %eax,nb111_ninner(%rsp) movl (%rdi),%edi movl %edi,nb111_nri(%rsp) movq %rsi,nb111_iinr(%rsp) movq %rdx,nb111_jindex(%rsp) movq %rcx,nb111_jjnr(%rsp) movq %r8,nb111_shift(%rsp) movq %r9,nb111_shiftvec(%rsp) movq nb111_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb111_facel(%rsp) ## assume we have at least one i particle - start directly movq nb111_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb111_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb111_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb111_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb111_iqO(%rsp) movaps %xmm4,nb111_iqH(%rsp) movq nb111_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb111_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb111_ntia(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb111_half(%rsp) movss nb111_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,nb111_half(%rsp) movaps %xmm3,nb111_three(%rsp) movaps %xmm4,nb111_six(%rsp) movaps %xmm5,nb111_twelve(%rsp)_nb_kernel111_x86_64_sse.nb111_threadloop: movq nb111_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel111_x86_64_sse.nb111_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_kernel111_x86_64_sse.nb111_spinlock ## if(nn1>nri) nn1=nri movl nb111_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,nb111_n(%rsp) movl %ebx,nb111_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel111_x86_64_sse.nb111_outerstart jmp _nb_kernel111_x86_64_sse.nb111_end_nb_kernel111_x86_64_sse.nb111_outerstart: ## ebx contains number of outer iterations addl nb111_nouter(%rsp),%ebx movl %ebx,nb111_nouter(%rsp)_nb_kernel111_x86_64_sse.nb111_outer: movq nb111_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,nb111_is3(%rsp) ## store is3 movq nb111_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 nb111_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 nb111_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb111_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,nb111_ixO(%rsp) movaps %xmm4,nb111_iyO(%rsp) movaps %xmm5,nb111_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,nb111_ixH1(%rsp) movaps %xmm1,nb111_iyH1(%rsp) movaps %xmm2,nb111_izH1(%rsp) movaps %xmm3,nb111_ixH2(%rsp) movaps %xmm4,nb111_iyH2(%rsp) movaps %xmm5,nb111_izH2(%rsp) ## clear potentials and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb111_vctot(%rsp) movaps %xmm4,nb111_Vvdwtot(%rsp) movaps %xmm4,nb111_fixO(%rsp) movaps %xmm4,nb111_fiyO(%rsp) movaps %xmm4,nb111_fizO(%rsp) movaps %xmm4,nb111_fixH1(%rsp) movaps %xmm4,nb111_fiyH1(%rsp) movaps %xmm4,nb111_fizH1(%rsp) movaps %xmm4,nb111_fixH2(%rsp) movaps %xmm4,nb111_fiyH2(%rsp) movaps %xmm4,nb111_fizH2(%rsp) movq nb111_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 nb111_pos(%rbp),%rsi movq nb111_faction(%rbp),%rdi movq nb111_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb111_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb111_ninner(%rsp),%ecx movl %ecx,nb111_ninner(%rsp) addl $0,%edx movl %edx,nb111_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel111_x86_64_sse.nb111_unroll_loop jmp _nb_kernel111_x86_64_sse.nb111_odd_inner_nb_kernel111_x86_64_sse.nb111_unroll_loop: ## quad-unroll innerloop here movq nb111_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,nb111_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb111_pos(%rbp),%rsi ## base of pos[] lea (%r8,%r8,2),%rax ## replace jnr with j3 lea (%r9,%r9,2),%rbx lea (%r10,%r10,2),%rcx ## replace jnr with j3 lea (%r11,%r11,2),%rdx ## move four j 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 movq nb111_charge(%rbp),%rsi ## base of charge[] movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb111_ixO(%rsp),%xmm0 subps nb111_iyO(%rsp),%xmm1 subps nb111_izO(%rsp),%xmm2 subps nb111_ixH1(%rsp),%xmm3 subps nb111_iyH1(%rsp),%xmm4 subps nb111_izH1(%rsp),%xmm5 subps nb111_ixH2(%rsp),%xmm6 subps nb111_iyH2(%rsp),%xmm7 subps nb111_izH2(%rsp),%xmm8 movaps %xmm0,nb111_dxO(%rsp) movaps %xmm1,nb111_dyO(%rsp) movaps %xmm2,nb111_dzO(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb111_dxH1(%rsp) movaps %xmm4,nb111_dyH1(%rsp) movaps %xmm5,nb111_dzH1(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb111_dxH2(%rsp) movaps %xmm7,nb111_dyH2(%rsp) movaps %xmm8,nb111_dzH2(%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 movss (%rsi,%r8,4),%xmm9 movss (%rsi,%r10,4),%xmm10 movss (%rsi,%r9,4),%xmm11 movss (%rsi,%r11,4),%xmm12 unpcklps %xmm10,%xmm9 unpcklps %xmm12,%xmm11 unpcklps %xmm11,%xmm9 movaps %xmm9,%xmm7 mulps nb111_iqO(%rsp),%xmm9 mulps nb111_iqH(%rsp),%xmm7 movaps %xmm9,nb111_qqO(%rsp)⌨️ 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?