nb_kernel214_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,140 行 · 第 1/5 页
S
2,140 行
#### $Id: nb_kernel214_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_kernel214_x86_64_sse.globl _nb_kernel214_x86_64_ssenb_kernel214_x86_64_sse: _nb_kernel214_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb214_fshift, 16.set nb214_gid, 24.set nb214_pos, 32.set nb214_faction, 40.set nb214_charge, 48.set nb214_p_facel, 56.set nb214_argkrf, 64.set nb214_argcrf, 72.set nb214_Vc, 80.set nb214_type, 88.set nb214_p_ntype, 96.set nb214_vdwparam, 104.set nb214_Vvdw, 112.set nb214_p_tabscale, 120.set nb214_VFtab, 128.set nb214_invsqrta, 136.set nb214_dvda, 144.set nb214_p_gbtabscale, 152.set nb214_GBtab, 160.set nb214_p_nthreads, 168.set nb214_count, 176.set nb214_mtx, 184.set nb214_outeriter, 192.set nb214_inneriter, 200.set nb214_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb214_ixO, 0.set nb214_iyO, 16.set nb214_izO, 32.set nb214_ixH1, 48.set nb214_iyH1, 64.set nb214_izH1, 80.set nb214_ixH2, 96.set nb214_iyH2, 112.set nb214_izH2, 128.set nb214_ixM, 144.set nb214_iyM, 160.set nb214_izM, 176.set nb214_jxO, 192.set nb214_jyO, 208.set nb214_jzO, 224.set nb214_jxH1, 240.set nb214_jyH1, 256.set nb214_jzH1, 272.set nb214_jxH2, 288.set nb214_jyH2, 304.set nb214_jzH2, 320.set nb214_jxM, 336.set nb214_jyM, 352.set nb214_jzM, 368.set nb214_dxOO, 384.set nb214_dyOO, 400.set nb214_dzOO, 416.set nb214_dxH1H1, 432.set nb214_dyH1H1, 448.set nb214_dzH1H1, 464.set nb214_dxH1H2, 480.set nb214_dyH1H2, 496.set nb214_dzH1H2, 512.set nb214_dxH1M, 528.set nb214_dyH1M, 544.set nb214_dzH1M, 560.set nb214_dxH2H1, 576.set nb214_dyH2H1, 592.set nb214_dzH2H1, 608.set nb214_dxH2H2, 624.set nb214_dyH2H2, 640.set nb214_dzH2H2, 656.set nb214_dxH2M, 672.set nb214_dyH2M, 688.set nb214_dzH2M, 704.set nb214_dxMH1, 720.set nb214_dyMH1, 736.set nb214_dzMH1, 752.set nb214_dxMH2, 768.set nb214_dyMH2, 784.set nb214_dzMH2, 800.set nb214_dxMM, 816.set nb214_dyMM, 832.set nb214_dzMM, 848.set nb214_qqMM, 864.set nb214_qqMH, 880.set nb214_qqHH, 896.set nb214_two, 912.set nb214_c6, 928.set nb214_c12, 944.set nb214_six, 960.set nb214_twelve, 976.set nb214_vctot, 992.set nb214_Vvdwtot, 1008.set nb214_fixO, 1024.set nb214_fiyO, 1040.set nb214_fizO, 1056.set nb214_fixH1, 1072.set nb214_fiyH1, 1088.set nb214_fizH1, 1104.set nb214_fixH2, 1120.set nb214_fiyH2, 1136.set nb214_fizH2, 1152.set nb214_fixM, 1168.set nb214_fiyM, 1184.set nb214_fizM, 1200.set nb214_fjxO, 1216.set nb214_fjyO, 1232.set nb214_fjzO, 1248.set nb214_fjxH1, 1264.set nb214_fjyH1, 1280.set nb214_fjzH1, 1296.set nb214_fjxH2, 1312.set nb214_fjyH2, 1328.set nb214_fjzH2, 1344.set nb214_fjxM, 1360.set nb214_fjyM, 1376.set nb214_fjzM, 1392.set nb214_half, 1408.set nb214_three, 1424.set nb214_rsqOO, 1440.set nb214_rsqH1H1, 1456.set nb214_rsqH1H2, 1472.set nb214_rsqH1M, 1488.set nb214_rsqH2H1, 1504.set nb214_rsqH2H2, 1520.set nb214_rsqH2M, 1536.set nb214_rsqMH1, 1552.set nb214_rsqMH2, 1568.set nb214_rsqMM, 1584.set nb214_rinvsqOO, 1600.set nb214_rinvH1H1, 1616.set nb214_rinvH1H2, 1632.set nb214_rinvH1M, 1648.set nb214_rinvH2H1, 1664.set nb214_rinvH2H2, 1680.set nb214_rinvH2M, 1696.set nb214_rinvMH1, 1712.set nb214_rinvMH2, 1728.set nb214_rinvMM, 1744.set nb214_fstmp, 1760.set nb214_krf, 1776.set nb214_crf, 1792.set nb214_is3, 1808.set nb214_ii3, 1812.set nb214_innerjjnr, 1816.set nb214_nri, 1824.set nb214_iinr, 1832.set nb214_jindex, 1840.set nb214_jjnr, 1848.set nb214_shift, 1856.set nb214_shiftvec, 1864.set nb214_facel, 1872.set nb214_innerk, 1880.set nb214_n, 1884.set nb214_nn1, 1888.set nb214_nouter, 1892.set nb214_ninner, 1896 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1912,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb214_nouter(%rsp) movl %eax,nb214_ninner(%rsp) movl (%rdi),%edi movl %edi,nb214_nri(%rsp) movq %rsi,nb214_iinr(%rsp) movq %rdx,nb214_jindex(%rsp) movq %rcx,nb214_jjnr(%rsp) movq %r8,nb214_shift(%rsp) movq %r9,nb214_shiftvec(%rsp) movq nb214_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb214_facel(%rsp) movq nb214_argkrf(%rbp),%rsi movq nb214_argcrf(%rbp),%rdi movss (%rsi),%xmm1 movss (%rdi),%xmm2 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm1,nb214_krf(%rsp) movaps %xmm2,nb214_crf(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb214_half(%rsp) movss nb214_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,nb214_half(%rsp) movaps %xmm2,nb214_two(%rsp) movaps %xmm3,nb214_three(%rsp) movaps %xmm4,nb214_six(%rsp) movaps %xmm5,nb214_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb214_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb214_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm5 movss 12(%rdx,%rbx,4),%xmm3 movss %xmm3,%xmm4 movq nb214_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb214_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,nb214_qqMM(%rsp) movaps %xmm4,nb214_qqMH(%rsp) movaps %xmm5,nb214_qqHH(%rsp) xorps %xmm0,%xmm0 movq nb214_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movl %ecx,%edx movq nb214_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] addq %rcx,%rdx movq nb214_vdwparam(%rbp),%rax movlps (%rax,%rdx,4),%xmm0 movaps %xmm0,%xmm1 shufps $0,%xmm0,%xmm0 shufps $0x55,%xmm1,%xmm1 movaps %xmm0,nb214_c6(%rsp) movaps %xmm1,nb214_c12(%rsp)_nb_kernel214_x86_64_sse.nb214_threadloop: movq nb214_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel214_x86_64_sse.nb214_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addq $1,%rbx ## rbx=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_kernel214_x86_64_sse.nb214_spinlock ## if(nn1>nri) nn1=nri movl nb214_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,nb214_n(%rsp) movl %ebx,nb214_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel214_x86_64_sse.nb214_outerstart jmp _nb_kernel214_x86_64_sse.nb214_end_nb_kernel214_x86_64_sse.nb214_outerstart: ## ebx contains number of outer iterations addl nb214_nouter(%rsp),%ebx movl %ebx,nb214_nouter(%rsp)_nb_kernel214_x86_64_sse.nb214_outer: movq nb214_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,nb214_is3(%rsp) ## store is3 movq nb214_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 nb214_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb214_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb214_ii3(%rsp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 addss (%rax,%rbx,4),%xmm3 ## ox addss 4(%rax,%rbx,4),%xmm4 ## oy addss 8(%rax,%rbx,4),%xmm5 ## oz addss 12(%rax,%rbx,4),%xmm6 ## h1x addss 16(%rax,%rbx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb214_ixO(%rsp) movaps %xmm4,nb214_iyO(%rsp) movaps %xmm5,nb214_izO(%rsp) movaps %xmm6,nb214_ixH1(%rsp) movaps %xmm7,nb214_iyH1(%rsp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%rax,%rbx,4),%xmm6 ## h1z addss 24(%rax,%rbx,4),%xmm0 ## h2x addss 28(%rax,%rbx,4),%xmm1 ## h2y addss 32(%rax,%rbx,4),%xmm2 ## h2z addss 36(%rax,%rbx,4),%xmm3 ## mx addss 40(%rax,%rbx,4),%xmm4 ## my addss 44(%rax,%rbx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 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 %xmm6,nb214_izH1(%rsp) movaps %xmm0,nb214_ixH2(%rsp) movaps %xmm1,nb214_iyH2(%rsp) movaps %xmm2,nb214_izH2(%rsp) movaps %xmm3,nb214_ixM(%rsp) movaps %xmm4,nb214_iyM(%rsp) movaps %xmm5,nb214_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb214_vctot(%rsp) movaps %xmm4,nb214_Vvdwtot(%rsp) movaps %xmm4,nb214_fixO(%rsp) movaps %xmm4,nb214_fiyO(%rsp) movaps %xmm4,nb214_fizO(%rsp) movaps %xmm4,nb214_fixH1(%rsp) movaps %xmm4,nb214_fiyH1(%rsp) movaps %xmm4,nb214_fizH1(%rsp) movaps %xmm4,nb214_fixH2(%rsp) movaps %xmm4,nb214_fiyH2(%rsp) movaps %xmm4,nb214_fizH2(%rsp) movaps %xmm4,nb214_fixM(%rsp) movaps %xmm4,nb214_fiyM(%rsp) movaps %xmm4,nb214_fizM(%rsp) movq nb214_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 nb214_pos(%rbp),%rsi movq nb214_faction(%rbp),%rdi movq nb214_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb214_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb214_ninner(%rsp),%ecx movl %ecx,nb214_ninner(%rsp) addl $0,%edx movl %edx,nb214_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel214_x86_64_sse.nb214_unroll_loop jmp _nb_kernel214_x86_64_sse.nb214_single_check_nb_kernel214_x86_64_sse.nb214_unroll_loop: ## quad-unroll innerloop here movq nb214_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 ⌨️ 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?