nb_kernel302_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,240 行 · 第 1/5 页
S
2,240 行
#### $Id: nb_kernel302_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_kernel302_x86_64_sse.globl _nb_kernel302_x86_64_ssenb_kernel302_x86_64_sse: _nb_kernel302_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb302_fshift, 16.set nb302_gid, 24.set nb302_pos, 32.set nb302_faction, 40.set nb302_charge, 48.set nb302_p_facel, 56.set nb302_argkrf, 64.set nb302_argcrf, 72.set nb302_Vc, 80.set nb302_type, 88.set nb302_p_ntype, 96.set nb302_vdwparam, 104.set nb302_Vvdw, 112.set nb302_p_tabscale, 120.set nb302_VFtab, 128.set nb302_invsqrta, 136.set nb302_dvda, 144.set nb302_p_gbtabscale, 152.set nb302_GBtab, 160.set nb302_p_nthreads, 168.set nb302_count, 176.set nb302_mtx, 184.set nb302_outeriter, 192.set nb302_inneriter, 200.set nb302_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb302_ixO, 0.set nb302_iyO, 16.set nb302_izO, 32.set nb302_ixH1, 48.set nb302_iyH1, 64.set nb302_izH1, 80.set nb302_ixH2, 96.set nb302_iyH2, 112.set nb302_izH2, 128.set nb302_jxO, 144.set nb302_jyO, 160.set nb302_jzO, 176.set nb302_jxH1, 192.set nb302_jyH1, 208.set nb302_jzH1, 224.set nb302_jxH2, 240.set nb302_jyH2, 256.set nb302_jzH2, 272.set nb302_dxOO, 288.set nb302_dyOO, 304.set nb302_dzOO, 320.set nb302_dxOH1, 336.set nb302_dyOH1, 352.set nb302_dzOH1, 368.set nb302_dxOH2, 384.set nb302_dyOH2, 400.set nb302_dzOH2, 416.set nb302_dxH1O, 432.set nb302_dyH1O, 448.set nb302_dzH1O, 464.set nb302_dxH1H1, 480.set nb302_dyH1H1, 496.set nb302_dzH1H1, 512.set nb302_dxH1H2, 528.set nb302_dyH1H2, 544.set nb302_dzH1H2, 560.set nb302_dxH2O, 576.set nb302_dyH2O, 592.set nb302_dzH2O, 608.set nb302_dxH2H1, 624.set nb302_dyH2H1, 640.set nb302_dzH2H1, 656.set nb302_dxH2H2, 672.set nb302_dyH2H2, 688.set nb302_dzH2H2, 704.set nb302_qqOO, 720.set nb302_qqOH, 736.set nb302_qqHH, 752.set nb302_two, 768.set nb302_tsc, 784.set nb302_vctot, 800.set nb302_fixO, 816.set nb302_fiyO, 832.set nb302_fizO, 848.set nb302_fixH1, 864.set nb302_fiyH1, 880.set nb302_fizH1, 896.set nb302_fixH2, 912.set nb302_fiyH2, 928.set nb302_fizH2, 944.set nb302_fjxO, 960.set nb302_fjyO, 976.set nb302_fjzO, 992.set nb302_fjxH1, 1008.set nb302_fjyH1, 1024.set nb302_fjzH1, 1040.set nb302_fjxH2, 1056.set nb302_fjyH2, 1072.set nb302_fjzH2, 1088.set nb302_half, 1104.set nb302_three, 1120.set nb302_epsO, 1136.set nb302_epsH1, 1152.set nb302_epsH2, 1168.set nb302_rsqH1O, 1184.set nb302_rsqH1H1, 1200.set nb302_rsqH1H2, 1216.set nb302_rsqH2O, 1232.set nb302_rsqH2H1, 1248.set nb302_rsqH2H2, 1264.set nb302_rinvOO, 1280.set nb302_rinvOH1, 1296.set nb302_rinvOH2, 1312.set nb302_rinvH1O, 1328.set nb302_rinvH1H1, 1344.set nb302_rinvH1H2, 1360.set nb302_rinvH2O, 1376.set nb302_rinvH2H1, 1392.set nb302_rinvH2H2, 1408.set nb302_is3, 1424.set nb302_ii3, 1428.set nb302_nri, 1432.set nb302_iinr, 1440.set nb302_jindex, 1448.set nb302_jjnr, 1456.set nb302_shift, 1464.set nb302_shiftvec, 1472.set nb302_facel, 1480.set nb302_innerjjnr, 1488.set nb302_innerk, 1496.set nb302_n, 1500.set nb302_nn1, 1504.set nb302_nouter, 1508.set nb302_ninner, 1512 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 emms subq $1528,%rsp ## local variable stack space (n*4+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb302_nouter(%rsp) movl %eax,nb302_ninner(%rsp) movl (%rdi),%edi movl %edi,nb302_nri(%rsp) movq %rsi,nb302_iinr(%rsp) movq %rdx,nb302_jindex(%rsp) movq %rcx,nb302_jjnr(%rsp) movq %r8,nb302_shift(%rsp) movq %r9,nb302_shiftvec(%rsp) movq nb302_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb302_facel(%rsp) movq nb302_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb302_tsc(%rsp) movq $0,%r8 movq $0,%r9 movq $0,%r10 movq $0,%r11 ## assume we have at least one i particle - start directly movq nb302_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb302_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss %xmm3,%xmm4 movss 4(%rdx,%rbx,4),%xmm5 movq nb302_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb302_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,nb302_qqOO(%rsp) movaps %xmm4,nb302_qqOH(%rsp) movaps %xmm5,nb302_qqHH(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb302_half(%rsp) movss nb302_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,nb302_half(%rsp) movaps %xmm2,nb302_two(%rsp) movaps %xmm3,nb302_three(%rsp)_nb_kernel302_x86_64_sse.nb302_threadloop: movq nb302_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel302_x86_64_sse.nb302_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_kernel302_x86_64_sse.nb302_spinlock ## if(nn1>nri) nn1=nri movl nb302_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,nb302_n(%rsp) movl %ebx,nb302_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel302_x86_64_sse.nb302_outerstart jmp _nb_kernel302_x86_64_sse.nb302_end_nb_kernel302_x86_64_sse.nb302_outerstart: ## ebx contains number of outer iterations addl nb302_nouter(%rsp),%ebx movl %ebx,nb302_nouter(%rsp)_nb_kernel302_x86_64_sse.nb302_outer: movq nb302_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,nb302_is3(%rsp) ## store is3 movq nb302_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 nb302_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb302_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb302_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,nb302_ixO(%rsp) movaps %xmm4,nb302_iyO(%rsp) movaps %xmm5,nb302_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,nb302_ixH1(%rsp) movaps %xmm1,nb302_iyH1(%rsp) movaps %xmm2,nb302_izH1(%rsp) movaps %xmm3,nb302_ixH2(%rsp) movaps %xmm4,nb302_iyH2(%rsp) movaps %xmm5,nb302_izH2(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb302_vctot(%rsp) movaps %xmm4,nb302_fixO(%rsp) movaps %xmm4,nb302_fiyO(%rsp) movaps %xmm4,nb302_fizO(%rsp) movaps %xmm4,nb302_fixH1(%rsp) movaps %xmm4,nb302_fiyH1(%rsp) movaps %xmm4,nb302_fizH1(%rsp) movaps %xmm4,nb302_fixH2(%rsp) movaps %xmm4,nb302_fiyH2(%rsp) movaps %xmm4,nb302_fizH2(%rsp) movq nb302_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 nb302_pos(%rbp),%rsi movq nb302_faction(%rbp),%rdi movq nb302_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb302_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb302_ninner(%rsp),%ecx movl %ecx,nb302_ninner(%rsp) addl $0,%edx movl %edx,nb302_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel302_x86_64_sse.nb302_unroll_loop jmp _nb_kernel302_x86_64_sse.nb302_single_check_nb_kernel302_x86_64_sse.nb302_unroll_loop: ## quad-unroll innerloop here movq nb302_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,nb302_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb302_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 j O coordinates to local temp variables 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 - movd %eax,%mm0 ## save j3 in mm0-mm3 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 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 nb302_ixO(%rsp),%xmm0 subps nb302_iyO(%rsp),%xmm1 subps nb302_izO(%rsp),%xmm2 subps nb302_ixH1(%rsp),%xmm3 subps nb302_iyH1(%rsp),%xmm4 subps nb302_izH1(%rsp),%xmm5 subps nb302_ixH2(%rsp),%xmm6 subps nb302_iyH2(%rsp),%xmm7 subps nb302_izH2(%rsp),%xmm8 movaps %xmm0,nb302_dxOO(%rsp) movaps %xmm1,nb302_dyOO(%rsp) movaps %xmm2,nb302_dzOO(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb302_dxH1O(%rsp) movaps %xmm4,nb302_dyH1O(%rsp) movaps %xmm5,nb302_dzH1O(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb302_dxH2O(%rsp) movaps %xmm7,nb302_dyH2O(%rsp) movaps %xmm8,nb302_dzH2O(%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 jO atoms rsqrtps %xmm0,%xmm1 rsqrtps %xmm3,%xmm4 rsqrtps %xmm6,%xmm7⌨️ 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?