nb_kernel102_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,225 行 · 第 1/5 页
S
2,225 行
#### $Id: nb_kernel102_x86_64_sse2.s,v 1.4.2.3 2006/09/22 08:40:32 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_kernel102_x86_64_sse2.globl _nb_kernel102_x86_64_sse2nb_kernel102_x86_64_sse2: _nb_kernel102_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb102_fshift, 16.set nb102_gid, 24.set nb102_pos, 32.set nb102_faction, 40.set nb102_charge, 48.set nb102_p_facel, 56.set nb102_argkrf, 64.set nb102_argcrf, 72.set nb102_Vc, 80.set nb102_type, 88.set nb102_p_ntype, 96.set nb102_vdwparam, 104.set nb102_Vvdw, 112.set nb102_p_tabscale, 120.set nb102_VFtab, 128.set nb102_invsqrta, 136.set nb102_dvda, 144.set nb102_p_gbtabscale, 152.set nb102_GBtab, 160.set nb102_p_nthreads, 168.set nb102_count, 176.set nb102_mtx, 184.set nb102_outeriter, 192.set nb102_inneriter, 200.set nb102_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb102_ixO, 0.set nb102_iyO, 16.set nb102_izO, 32.set nb102_ixH1, 48.set nb102_iyH1, 64.set nb102_izH1, 80.set nb102_ixH2, 96.set nb102_iyH2, 112.set nb102_izH2, 128.set nb102_jxO, 144.set nb102_jyO, 160.set nb102_jzO, 176.set nb102_jxH1, 192.set nb102_jyH1, 208.set nb102_jzH1, 224.set nb102_jxH2, 240.set nb102_jyH2, 256.set nb102_jzH2, 272.set nb102_dxOO, 288.set nb102_dyOO, 304.set nb102_dzOO, 320.set nb102_dxOH1, 336.set nb102_dyOH1, 352.set nb102_dzOH1, 368.set nb102_dxOH2, 384.set nb102_dyOH2, 400.set nb102_dzOH2, 416.set nb102_dxH1O, 432.set nb102_dyH1O, 448.set nb102_dzH1O, 464.set nb102_dxH1H1, 480.set nb102_dyH1H1, 496.set nb102_dzH1H1, 512.set nb102_dxH1H2, 528.set nb102_dyH1H2, 544.set nb102_dzH1H2, 560.set nb102_dxH2O, 576.set nb102_dyH2O, 592.set nb102_dzH2O, 608.set nb102_dxH2H1, 624.set nb102_dyH2H1, 640.set nb102_dzH2H1, 656.set nb102_dxH2H2, 672.set nb102_dyH2H2, 688.set nb102_dzH2H2, 704.set nb102_qqOO, 720.set nb102_qqOH, 736.set nb102_qqHH, 752.set nb102_vctot, 768.set nb102_fixO, 784.set nb102_fiyO, 800.set nb102_fizO, 816.set nb102_fixH1, 832.set nb102_fiyH1, 848.set nb102_fizH1, 864.set nb102_fixH2, 880.set nb102_fiyH2, 896.set nb102_fizH2, 912.set nb102_fjxO, 928.set nb102_fjyO, 944.set nb102_fjzO, 960.set nb102_fjxH1, 976.set nb102_fjyH1, 992.set nb102_fjzH1, 1008.set nb102_fjxH2, 1024.set nb102_fjyH2, 1040.set nb102_fjzH2, 1056.set nb102_half, 1072.set nb102_three, 1088.set nb102_rsqOO, 1104.set nb102_rsqOH1, 1120.set nb102_rsqOH2, 1136.set nb102_rsqH1O, 1152.set nb102_rsqH1H1, 1168.set nb102_rsqH1H2, 1184.set nb102_rsqH2O, 1200.set nb102_rsqH2H1, 1216.set nb102_rsqH2H2, 1232.set nb102_rinvOO, 1248.set nb102_rinvOH1, 1264.set nb102_rinvOH2, 1280.set nb102_rinvH1O, 1296.set nb102_rinvH1H1, 1312.set nb102_rinvH1H2, 1328.set nb102_rinvH2O, 1344.set nb102_rinvH2H1, 1360.set nb102_rinvH2H2, 1376.set nb102_is3, 1392.set nb102_ii3, 1396.set nb102_innerjjnr, 1400.set nb102_nri, 1408.set nb102_iinr, 1416.set nb102_jindex, 1424.set nb102_jjnr, 1432.set nb102_shift, 1440.set nb102_shiftvec, 1448.set nb102_facel, 1456.set nb102_innerk, 1464.set nb102_n, 1468.set nb102_nn1, 1472.set nb102_nouter, 1476.set nb102_ninner, 1480 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1496,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb102_nouter(%rsp) movl %eax,nb102_ninner(%rsp) movl (%rdi),%edi movl %edi,nb102_nri(%rsp) movq %rsi,nb102_iinr(%rsp) movq %rdx,nb102_jindex(%rsp) movq %rcx,nb102_jjnr(%rsp) movq %r8,nb102_shift(%rsp) movq %r9,nb102_shiftvec(%rsp) movq nb102_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb102_facel(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb102_half(%rsp) movl %ebx,nb102_half+4(%rsp) movsd nb102_half(%rsp),%xmm1 shufpd $0,%xmm1,%xmm1 ## splat to all elements movapd %xmm1,%xmm3 addpd %xmm3,%xmm3 ## one movapd %xmm3,%xmm2 addpd %xmm2,%xmm2 ## two addpd %xmm2,%xmm3 ## three movapd %xmm1,nb102_half(%rsp) movapd %xmm3,nb102_three(%rsp) ## assume we have at least one i particle - start directly movq nb102_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb102_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 ## qO movsd %xmm3,%xmm4 ## qO movsd 8(%rdx,%rbx,8),%xmm5 ## qH movq nb102_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd nb102_facel(%rsp),%xmm6 ## facel mulsd %xmm3,%xmm3 ## qO*qO mulsd %xmm5,%xmm4 ## qO*qH mulsd %xmm5,%xmm5 ## qH*qH mulsd %xmm6,%xmm3 mulsd %xmm6,%xmm4 mulsd %xmm6,%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb102_qqOO(%rsp) movapd %xmm4,nb102_qqOH(%rsp) movapd %xmm5,nb102_qqHH(%rsp)_nb_kernel102_x86_64_sse2.nb102_threadloop: movq nb102_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel102_x86_64_sse2.nb102_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_kernel102_x86_64_sse2.nb102_spinlock ## if(nn1>nri) nn1=nri movl nb102_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,nb102_n(%rsp) movl %ebx,nb102_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel102_x86_64_sse2.nb102_outerstart jmp _nb_kernel102_x86_64_sse2.nb102_end_nb_kernel102_x86_64_sse2.nb102_outerstart: ## ebx contains number of outer iterations addl nb102_nouter(%rsp),%ebx movl %ebx,nb102_nouter(%rsp)_nb_kernel102_x86_64_sse2.nb102_outer: movq nb102_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,nb102_is3(%rsp) ## store is3 movq nb102_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%xmm2 movq nb102_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb102_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb102_ii3(%rsp) movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 addsd (%rax,%rbx,8),%xmm3 addsd 8(%rax,%rbx,8),%xmm4 addsd 16(%rax,%rbx,8),%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb102_ixO(%rsp) movapd %xmm4,nb102_iyO(%rsp) movapd %xmm5,nb102_izO(%rsp) movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 24(%rax,%rbx,8),%xmm0 addsd 32(%rax,%rbx,8),%xmm1 addsd 40(%rax,%rbx,8),%xmm2 addsd 48(%rax,%rbx,8),%xmm3 addsd 56(%rax,%rbx,8),%xmm4 addsd 64(%rax,%rbx,8),%xmm5 shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm0,nb102_ixH1(%rsp) movapd %xmm1,nb102_iyH1(%rsp) movapd %xmm2,nb102_izH1(%rsp) movapd %xmm3,nb102_ixH2(%rsp) movapd %xmm4,nb102_iyH2(%rsp) movapd %xmm5,nb102_izH2(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb102_vctot(%rsp) movapd %xmm4,nb102_fixO(%rsp) movapd %xmm4,nb102_fiyO(%rsp) movapd %xmm4,nb102_fizO(%rsp) movapd %xmm4,nb102_fixH1(%rsp) movapd %xmm4,nb102_fiyH1(%rsp) movapd %xmm4,nb102_fizH1(%rsp) movapd %xmm4,nb102_fixH2(%rsp) movapd %xmm4,nb102_fiyH2(%rsp) movapd %xmm4,nb102_fizH2(%rsp) movq nb102_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 nb102_pos(%rbp),%rsi movq nb102_faction(%rbp),%rdi movq nb102_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb102_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb102_ninner(%rsp),%ecx movl %ecx,nb102_ninner(%rsp) addl $0,%edx movl %edx,nb102_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel102_x86_64_sse2.nb102_unroll_loop jmp _nb_kernel102_x86_64_sse2.nb102_checksingle_nb_kernel102_x86_64_sse2.nb102_unroll_loop: ## twice unrolled innerloop here movq nb102_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb102_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb102_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move j O coordinates to local temp variables movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm0 movhpd 8(%rsi,%rbx,8),%xmm1 movhpd 16(%rsi,%rbx,8),%xmm2 ## xmm0 = Ox ## xmm1 = Oy ## xmm2 = Oz movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb102_ixO(%rsp),%xmm0 subpd nb102_iyO(%rsp),%xmm1 subpd nb102_izO(%rsp),%xmm2 subpd nb102_ixH1(%rsp),%xmm3 subpd nb102_iyH1(%rsp),%xmm4 subpd nb102_izH1(%rsp),%xmm5 subpd nb102_ixH2(%rsp),%xmm6 subpd nb102_iyH2(%rsp),%xmm7 subpd nb102_izH2(%rsp),%xmm8 movapd %xmm0,nb102_dxOO(%rsp) movapd %xmm1,nb102_dyOO(%rsp) movapd %xmm2,nb102_dzOO(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb102_dxH1O(%rsp) movapd %xmm4,nb102_dyH1O(%rsp) movapd %xmm5,nb102_dzH1O(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb102_dxH2O(%rsp) movapd %xmm7,nb102_dyH2O(%rsp) movapd %xmm8,nb102_dzH2O(%rsp) mulpd %xmm6,%xmm6 mulpd %xmm7,%xmm7 mulpd %xmm8,%xmm8 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 addpd %xmm7,%xmm6 addpd %xmm8,%xmm6 ## start doing invsqrt for jO atoms cvtpd2ps %xmm0,%xmm1 cvtpd2ps %xmm3,%xmm4 cvtpd2ps %xmm6,%xmm7 rsqrtps %xmm1,%xmm1 rsqrtps %xmm4,%xmm4 rsqrtps %xmm7,%xmm7 cvtps2pd %xmm1,%xmm1 cvtps2pd %xmm4,%xmm4 cvtps2pd %xmm7,%xmm7 movapd %xmm1,%xmm2 movapd %xmm4,%xmm5 movapd %xmm7,%xmm8 mulpd %xmm1,%xmm1 ## lu*lu mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm7,%xmm7 ## lu*lu movapd nb102_three(%rsp),%xmm9 movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulpd %xmm0,%xmm1 ## rsq*lu*lu mulpd %xmm3,%xmm4 ## rsq*lu*lu mulpd %xmm6,%xmm7 ## rsq*lu*lu subpd %xmm1,%xmm9 subpd %xmm4,%xmm10 subpd %xmm7,%xmm11 ## 3-rsq*lu*lu mulpd %xmm2,%xmm9 mulpd %xmm5,%xmm10⌨️ 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?