nb_kernel104_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,227 行 · 第 1/5 页
S
2,227 行
#### $Id: nb_kernel104_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_kernel104_x86_64_sse2.globl _nb_kernel104_x86_64_sse2nb_kernel104_x86_64_sse2: _nb_kernel104_x86_64_sse2: .set nb104_fshift, 16.set nb104_gid, 24.set nb104_pos, 32.set nb104_faction, 40.set nb104_charge, 48.set nb104_p_facel, 56.set nb104_argkrf, 64.set nb104_argcrf, 72.set nb104_Vc, 80.set nb104_type, 88.set nb104_p_ntype, 96.set nb104_vdwparam, 104.set nb104_Vvdw, 112.set nb104_p_tabscale, 120.set nb104_VFtab, 128.set nb104_invsqrta, 136.set nb104_dvda, 144.set nb104_p_gbtabscale, 152.set nb104_GBtab, 160.set nb104_p_nthreads, 168.set nb104_count, 176.set nb104_mtx, 184.set nb104_outeriter, 192.set nb104_inneriter, 200.set nb104_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb104_ixM, 0.set nb104_iyM, 16.set nb104_izM, 32.set nb104_ixH1, 48.set nb104_iyH1, 64.set nb104_izH1, 80.set nb104_ixH2, 96.set nb104_iyH2, 112.set nb104_izH2, 128.set nb104_jxM, 144.set nb104_jyM, 160.set nb104_jzM, 176.set nb104_jxH1, 192.set nb104_jyH1, 208.set nb104_jzH1, 224.set nb104_jxH2, 240.set nb104_jyH2, 256.set nb104_jzH2, 272.set nb104_dxMM, 288.set nb104_dyMM, 304.set nb104_dzMM, 320.set nb104_dxMH1, 336.set nb104_dyMH1, 352.set nb104_dzMH1, 368.set nb104_dxMH2, 384.set nb104_dyMH2, 400.set nb104_dzMH2, 416.set nb104_dxH1M, 432.set nb104_dyH1M, 448.set nb104_dzH1M, 464.set nb104_dxH1H1, 480.set nb104_dyH1H1, 496.set nb104_dzH1H1, 512.set nb104_dxH1H2, 528.set nb104_dyH1H2, 544.set nb104_dzH1H2, 560.set nb104_dxH2M, 576.set nb104_dyH2M, 592.set nb104_dzH2M, 608.set nb104_dxH2H1, 624.set nb104_dyH2H1, 640.set nb104_dzH2H1, 656.set nb104_dxH2H2, 672.set nb104_dyH2H2, 688.set nb104_dzH2H2, 704.set nb104_qqMM, 720.set nb104_qqMH, 736.set nb104_qqHH, 752.set nb104_vctot, 768.set nb104_fixM, 784.set nb104_fiyM, 800.set nb104_fizM, 816.set nb104_fixH1, 832.set nb104_fiyH1, 848.set nb104_fizH1, 864.set nb104_fixH2, 880.set nb104_fiyH2, 896.set nb104_fizH2, 912.set nb104_fjxM, 928.set nb104_fjyM, 944.set nb104_fjzM, 960.set nb104_fjxH1, 976.set nb104_fjyH1, 992.set nb104_fjzH1, 1008.set nb104_fjxH2, 1024.set nb104_fjyH2, 1040.set nb104_fjzH2, 1056.set nb104_half, 1072.set nb104_three, 1088.set nb104_rsqMM, 1104.set nb104_rsqMH1, 1120.set nb104_rsqMH2, 1136.set nb104_rsqH1M, 1152.set nb104_rsqH1H1, 1168.set nb104_rsqH1H2, 1184.set nb104_rsqH2M, 1200.set nb104_rsqH2H1, 1216.set nb104_rsqH2H2, 1232.set nb104_rinvMM, 1248.set nb104_rinvMH1, 1264.set nb104_rinvMH2, 1280.set nb104_rinvH1M, 1296.set nb104_rinvH1H1, 1312.set nb104_rinvH1H2, 1328.set nb104_rinvH2M, 1344.set nb104_rinvH2H1, 1360.set nb104_rinvH2H2, 1376.set nb104_nri, 1392.set nb104_innerjjnr, 1400.set nb104_iinr, 1408.set nb104_jindex, 1416.set nb104_jjnr, 1424.set nb104_shift, 1432.set nb104_shiftvec, 1440.set nb104_facel, 1448.set nb104_is3, 1456.set nb104_ii3, 1460.set nb104_innerk, 1464.set nb104_n, 1468.set nb104_nn1, 1472.set nb104_nouter, 1476.set nb104_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,nb104_nouter(%rsp) movl %eax,nb104_ninner(%rsp) movl (%rdi),%edi movl %edi,nb104_nri(%rsp) movq %rsi,nb104_iinr(%rsp) movq %rdx,nb104_jindex(%rsp) movq %rcx,nb104_jjnr(%rsp) movq %r8,nb104_shift(%rsp) movq %r9,nb104_shiftvec(%rsp) movq nb104_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb104_facel(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb104_half(%rsp) movl %ebx,nb104_half+4(%rsp) movsd nb104_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,nb104_half(%rsp) movapd %xmm3,nb104_three(%rsp) ## assume we have at least one i particle - start directly movq nb104_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb104_charge(%rbp),%rdx movsd 24(%rdx,%rbx,8),%xmm3 ## qM movsd %xmm3,%xmm4 ## qM movsd 8(%rdx,%rbx,8),%xmm5 ## qH movq nb104_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd nb104_facel(%rsp),%xmm6 ## facel mulsd %xmm3,%xmm3 ## qM*qM mulsd %xmm5,%xmm4 ## qM*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,nb104_qqMM(%rsp) movapd %xmm4,nb104_qqMH(%rsp) movapd %xmm5,nb104_qqHH(%rsp)_nb_kernel104_x86_64_sse2.nb104_threadloop: movq nb104_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel104_x86_64_sse2.nb104_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_kernel104_x86_64_sse2.nb104_spinlock ## if(nn1>nri) nn1=nri movl nb104_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,nb104_n(%rsp) movl %ebx,nb104_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel104_x86_64_sse2.nb104_outerstart jmp _nb_kernel104_x86_64_sse2.nb104_end_nb_kernel104_x86_64_sse2.nb104_outerstart: ## ebx contains number of outer iterations addl nb104_nouter(%rsp),%ebx movl %ebx,nb104_nouter(%rsp)_nb_kernel104_x86_64_sse2.nb104_outer: movq nb104_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,nb104_is3(%rsp) ## store is3 movq nb104_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 nb104_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb104_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb104_ii3(%rsp) movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 addsd 24(%rax,%rbx,8),%xmm3 addsd 32(%rax,%rbx,8),%xmm4 addsd 40(%rax,%rbx,8),%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb104_ixH1(%rsp) movapd %xmm4,nb104_iyH1(%rsp) movapd %xmm5,nb104_izH1(%rsp) movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 48(%rax,%rbx,8),%xmm0 addsd 56(%rax,%rbx,8),%xmm1 addsd 64(%rax,%rbx,8),%xmm2 addsd 72(%rax,%rbx,8),%xmm3 addsd 80(%rax,%rbx,8),%xmm4 addsd 88(%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,nb104_ixH2(%rsp) movapd %xmm1,nb104_iyH2(%rsp) movapd %xmm2,nb104_izH2(%rsp) movapd %xmm3,nb104_ixM(%rsp) movapd %xmm4,nb104_iyM(%rsp) movapd %xmm5,nb104_izM(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb104_vctot(%rsp) movapd %xmm4,nb104_fixM(%rsp) movapd %xmm4,nb104_fiyM(%rsp) movapd %xmm4,nb104_fizM(%rsp) movapd %xmm4,nb104_fixH1(%rsp) movapd %xmm4,nb104_fiyH1(%rsp) movapd %xmm4,nb104_fizH1(%rsp) movapd %xmm4,nb104_fixH2(%rsp) movapd %xmm4,nb104_fiyH2(%rsp) movapd %xmm4,nb104_fizH2(%rsp) movq nb104_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 nb104_pos(%rbp),%rsi movq nb104_faction(%rbp),%rdi movq nb104_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb104_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb104_ninner(%rsp),%ecx movl %ecx,nb104_ninner(%rsp) addl $0,%edx movl %edx,nb104_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel104_x86_64_sse2.nb104_unroll_loop jmp _nb_kernel104_x86_64_sse2.nb104_checksingle_nb_kernel104_x86_64_sse2.nb104_unroll_loop: ## twice unrolled innerloop here movq nb104_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb104_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb104_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move j H1 coordinates to local temp variables movlpd 24(%rsi,%rax,8),%xmm0 movlpd 32(%rsi,%rax,8),%xmm1 movlpd 40(%rsi,%rax,8),%xmm2 movhpd 24(%rsi,%rbx,8),%xmm0 movhpd 32(%rsi,%rbx,8),%xmm1 movhpd 40(%rsi,%rbx,8),%xmm2 ## xmm0 = H1x ## xmm1 = H1y ## xmm2 = H1z movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb104_ixH1(%rsp),%xmm0 subpd nb104_iyH1(%rsp),%xmm1 subpd nb104_izH1(%rsp),%xmm2 subpd nb104_ixH2(%rsp),%xmm3 subpd nb104_iyH2(%rsp),%xmm4 subpd nb104_izH2(%rsp),%xmm5 subpd nb104_ixM(%rsp),%xmm6 subpd nb104_iyM(%rsp),%xmm7 subpd nb104_izM(%rsp),%xmm8 movapd %xmm0,nb104_dxH1H1(%rsp) movapd %xmm1,nb104_dyH1H1(%rsp) movapd %xmm2,nb104_dzH1H1(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb104_dxH2H1(%rsp) movapd %xmm4,nb104_dyH2H1(%rsp) movapd %xmm5,nb104_dzH2H1(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb104_dxMH1(%rsp) movapd %xmm7,nb104_dyMH1(%rsp) movapd %xmm8,nb104_dzMH1(%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 jH1 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 nb104_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 mulpd %xmm8,%xmm11 ## lu*(3-rsq*lu*lu)⌨️ 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?