nb_kernel312_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,232 行 · 第 1/5 页
S
2,232 行
#### $Id: nb_kernel312_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_kernel312_x86_64_sse.globl _nb_kernel312_x86_64_ssenb_kernel312_x86_64_sse: _nb_kernel312_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb312_fshift, 16.set nb312_gid, 24.set nb312_pos, 32.set nb312_faction, 40.set nb312_charge, 48.set nb312_p_facel, 56.set nb312_argkrf, 64.set nb312_argcrf, 72.set nb312_Vc, 80.set nb312_type, 88.set nb312_p_ntype, 96.set nb312_vdwparam, 104.set nb312_Vvdw, 112.set nb312_p_tabscale, 120.set nb312_VFtab, 128.set nb312_invsqrta, 136.set nb312_dvda, 144.set nb312_p_gbtabscale, 152.set nb312_GBtab, 160.set nb312_p_nthreads, 168.set nb312_count, 176.set nb312_mtx, 184.set nb312_outeriter, 192.set nb312_inneriter, 200.set nb312_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb312_ixO, 0.set nb312_iyO, 16.set nb312_izO, 32.set nb312_ixH1, 48.set nb312_iyH1, 64.set nb312_izH1, 80.set nb312_ixH2, 96.set nb312_iyH2, 112.set nb312_izH2, 128.set nb312_jxO, 144.set nb312_jyO, 160.set nb312_jzO, 176.set nb312_jxH1, 192.set nb312_jyH1, 208.set nb312_jzH1, 224.set nb312_jxH2, 240.set nb312_jyH2, 256.set nb312_jzH2, 272.set nb312_dxOO, 288.set nb312_dyOO, 304.set nb312_dzOO, 320.set nb312_dxOH1, 336.set nb312_dyOH1, 352.set nb312_dzOH1, 368.set nb312_dxOH2, 384.set nb312_dyOH2, 400.set nb312_dzOH2, 416.set nb312_dxH1O, 432.set nb312_dyH1O, 448.set nb312_dzH1O, 464.set nb312_dxH1H1, 480.set nb312_dyH1H1, 496.set nb312_dzH1H1, 512.set nb312_dxH1H2, 528.set nb312_dyH1H2, 544.set nb312_dzH1H2, 560.set nb312_dxH2O, 576.set nb312_dyH2O, 592.set nb312_dzH2O, 608.set nb312_dxH2H1, 624.set nb312_dyH2H1, 640.set nb312_dzH2H1, 656.set nb312_dxH2H2, 672.set nb312_dyH2H2, 688.set nb312_dzH2H2, 704.set nb312_qqOO, 720.set nb312_qqOH, 736.set nb312_qqHH, 752.set nb312_two, 768.set nb312_tsc, 784.set nb312_c6, 800.set nb312_c12, 816.set nb312_six, 832.set nb312_twelve, 848.set nb312_vctot, 864.set nb312_Vvdwtot, 880.set nb312_fixO, 896.set nb312_fiyO, 912.set nb312_fizO, 928.set nb312_fixH1, 944.set nb312_fiyH1, 960.set nb312_fizH1, 976.set nb312_fixH2, 992.set nb312_fiyH2, 1008.set nb312_fizH2, 1024.set nb312_fjxO, 1040.set nb312_fjyO, 1056.set nb312_fjzO, 1072.set nb312_fjxH1, 1088.set nb312_fjyH1, 1104.set nb312_fjzH1, 1120.set nb312_fjxH2, 1136.set nb312_fjyH2, 1152.set nb312_fjzH2, 1168.set nb312_half, 1184.set nb312_three, 1200.set nb312_epsO, 1216.set nb312_epsH1, 1232.set nb312_epsH2, 1248.set nb312_rsqH1O, 1264.set nb312_rsqH1H1, 1280.set nb312_rsqH1H2, 1296.set nb312_rsqH2O, 1312.set nb312_rsqH2H1, 1328.set nb312_rsqH2H2, 1344.set nb312_rinvOO, 1360.set nb312_rinvOH1, 1376.set nb312_rinvOH2, 1392.set nb312_rinvH1O, 1408.set nb312_rinvH1H1, 1424.set nb312_rinvH1H2, 1440.set nb312_rinvH2O, 1456.set nb312_rinvH2H1, 1472.set nb312_rinvH2H2, 1488.set nb312_fstmp, 1504.set nb312_is3, 1520.set nb312_ii3, 1524.set nb312_nri, 1528.set nb312_iinr, 1536.set nb312_jindex, 1544.set nb312_jjnr, 1552.set nb312_shift, 1560.set nb312_shiftvec, 1568.set nb312_facel, 1576.set nb312_innerjjnr, 1584.set nb312_innerk, 1592.set nb312_n, 1596.set nb312_nn1, 1600.set nb312_nouter, 1604.set nb312_ninner, 1608 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 emms subq $1624,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb312_nouter(%rsp) movl %eax,nb312_ninner(%rsp) movl (%rdi),%edi movl %edi,nb312_nri(%rsp) movq %rsi,nb312_iinr(%rsp) movq %rdx,nb312_jindex(%rsp) movq %rcx,nb312_jjnr(%rsp) movq %r8,nb312_shift(%rsp) movq %r9,nb312_shiftvec(%rsp) movq nb312_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb312_facel(%rsp) movq nb312_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb312_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb312_half(%rsp) movss nb312_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,nb312_half(%rsp) movaps %xmm2,nb312_two(%rsp) movaps %xmm3,nb312_three(%rsp) movaps %xmm4,nb312_six(%rsp) movaps %xmm5,nb312_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb312_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb312_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss %xmm3,%xmm4 movss 4(%rdx,%rbx,4),%xmm5 movq nb312_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb312_facel(%rsp),%xmm6 mulss %xmm3,%xmm3 mulss %xmm5,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm3 mulss %xmm6,%xmm4 mulss %xmm6,%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb312_qqOO(%rsp) movaps %xmm4,nb312_qqOH(%rsp) movaps %xmm5,nb312_qqHH(%rsp) xorps %xmm0,%xmm0 movq nb312_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movl %ecx,%edx movq nb312_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] addl %ecx,%edx movq nb312_vdwparam(%rbp),%rax movlps (%rax,%rdx,4),%xmm0 movaps %xmm0,%xmm1 shufps $0,%xmm0,%xmm0 shufps $85,%xmm1,%xmm1 ## 01010101 movaps %xmm0,nb312_c6(%rsp) movaps %xmm1,nb312_c12(%rsp)_nb_kernel312_x86_64_sse.nb312_threadloop: movq nb312_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel312_x86_64_sse.nb312_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_kernel312_x86_64_sse.nb312_spinlock ## if(nn1>nri) nn1=nri movl nb312_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,nb312_n(%rsp) movl %ebx,nb312_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel312_x86_64_sse.nb312_outerstart jmp _nb_kernel312_x86_64_sse.nb312_end_nb_kernel312_x86_64_sse.nb312_outerstart: ## ebx contains number of outer iterations addl nb312_nouter(%rsp),%ebx movl %ebx,nb312_nouter(%rsp)_nb_kernel312_x86_64_sse.nb312_outer: movq nb312_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,nb312_is3(%rsp) ## store is3 movq nb312_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 nb312_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb312_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb312_ii3(%rsp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 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,nb312_ixO(%rsp) movaps %xmm4,nb312_iyO(%rsp) movaps %xmm5,nb312_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,nb312_ixH1(%rsp) movaps %xmm1,nb312_iyH1(%rsp) movaps %xmm2,nb312_izH1(%rsp) movaps %xmm3,nb312_ixH2(%rsp) movaps %xmm4,nb312_iyH2(%rsp) movaps %xmm5,nb312_izH2(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb312_vctot(%rsp) movaps %xmm4,nb312_Vvdwtot(%rsp) movaps %xmm4,nb312_fixO(%rsp) movaps %xmm4,nb312_fiyO(%rsp) movaps %xmm4,nb312_fizO(%rsp) movaps %xmm4,nb312_fixH1(%rsp) movaps %xmm4,nb312_fiyH1(%rsp) movaps %xmm4,nb312_fizH1(%rsp) movaps %xmm4,nb312_fixH2(%rsp) movaps %xmm4,nb312_fiyH2(%rsp) movaps %xmm4,nb312_fizH2(%rsp) movq nb312_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 nb312_pos(%rbp),%rsi movq nb312_faction(%rbp),%rdi movq nb312_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb312_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb312_ninner(%rsp),%ecx movl %ecx,nb312_ninner(%rsp) addl $0,%edx movl %edx,nb312_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel312_x86_64_sse.nb312_unroll_loop jmp _nb_kernel312_x86_64_sse.nb312_single_check_nb_kernel312_x86_64_sse.nb312_unroll_loop: ## quad-unroll innerloop here movq nb312_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,nb312_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb312_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 ## load j O coordinates movlps (%rsi,%rax,4),%xmm0 ## jxOa jyOa - - movlps (%rsi,%rcx,4),%xmm1 ## jxOc jyOc - - movhps (%rsi,%rbx,4),%xmm0 ## jxOa jyOa jxOb jyOb movhps (%rsi,%rdx,4),%xmm1 ## jxOc jyOc jxOd jyOd movss 8(%rsi,%rax,4),%xmm2 ## jzOa - - - movss 8(%rsi,%rcx,4),%xmm3 ## jzOc - - - movhps 8(%rsi,%rbx,4),%xmm2 ## jzOa - jzOb - movhps 8(%rsi,%rdx,4),%xmm3 ## jzOc - jzOd - movaps %xmm0,%xmm4 unpcklps %xmm1,%xmm0 ## jxOa jxOc jyOa jyOc unpckhps %xmm1,%xmm4 ## jxOb jxOd jyOb jyOd movaps %xmm0,%xmm1 unpcklps %xmm4,%xmm0 ## x unpckhps %xmm4,%xmm1 ## y shufps $136,%xmm3,%xmm2 ## 10001000 => jzOa jzOb jzOc jzOd ## xmm0 = Ox ## xmm1 = Oy ## xmm2 = Oz movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb312_ixO(%rsp),%xmm0 subps nb312_iyO(%rsp),%xmm1 subps nb312_izO(%rsp),%xmm2 subps nb312_ixH1(%rsp),%xmm3 subps nb312_iyH1(%rsp),%xmm4 subps nb312_izH1(%rsp),%xmm5 subps nb312_ixH2(%rsp),%xmm6 subps nb312_iyH2(%rsp),%xmm7 subps nb312_izH2(%rsp),%xmm8 movaps %xmm0,nb312_dxOO(%rsp) movaps %xmm1,nb312_dyOO(%rsp) movaps %xmm2,nb312_dzOO(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb312_dxH1O(%rsp) movaps %xmm4,nb312_dyH1O(%rsp) movaps %xmm5,nb312_dzH1O(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%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?