nb_kernel331_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,175 行 · 第 1/5 页
S
2,175 行
#### $Id: nb_kernel331_x86_64_sse2.s,v 1.5.2.3 2006/09/22 08:40:37 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_kernel331_x86_64_sse2.globl _nb_kernel331_x86_64_sse2nb_kernel331_x86_64_sse2: _nb_kernel331_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb331_fshift, 16.set nb331_gid, 24.set nb331_pos, 32.set nb331_faction, 40.set nb331_charge, 48.set nb331_p_facel, 56.set nb331_argkrf, 64.set nb331_argcrf, 72.set nb331_Vc, 80.set nb331_type, 88.set nb331_p_ntype, 96.set nb331_vdwparam, 104.set nb331_Vvdw, 112.set nb331_p_tabscale, 120.set nb331_VFtab, 128.set nb331_invsqrta, 136.set nb331_dvda, 144.set nb331_p_gbtabscale, 152.set nb331_GBtab, 160.set nb331_p_nthreads, 168.set nb331_count, 176.set nb331_mtx, 184.set nb331_outeriter, 192.set nb331_inneriter, 200.set nb331_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb331_ixO, 0.set nb331_iyO, 16.set nb331_izO, 32.set nb331_ixH1, 48.set nb331_iyH1, 64.set nb331_izH1, 80.set nb331_ixH2, 96.set nb331_iyH2, 112.set nb331_izH2, 128.set nb331_iqO, 144.set nb331_iqH, 160.set nb331_dxO, 176.set nb331_dyO, 192.set nb331_dzO, 208.set nb331_dxH1, 224.set nb331_dyH1, 240.set nb331_dzH1, 256.set nb331_dxH2, 272.set nb331_dyH2, 288.set nb331_dzH2, 304.set nb331_qqO, 320.set nb331_qqH, 336.set nb331_rinvO, 352.set nb331_rinvH1, 368.set nb331_rinvH2, 384.set nb331_rO, 400.set nb331_rH1, 416.set nb331_rH2, 432.set nb331_tsc, 448.set nb331_two, 464.set nb331_c6, 480.set nb331_c12, 496.set nb331_vctot, 512.set nb331_Vvdwtot, 528.set nb331_fixO, 544.set nb331_fiyO, 560.set nb331_fizO, 576.set nb331_fixH1, 592.set nb331_fiyH1, 608.set nb331_fizH1, 624.set nb331_fixH2, 640.set nb331_fiyH2, 656.set nb331_fizH2, 672.set nb331_fjx, 688.set nb331_fjy, 704.set nb331_fjz, 720.set nb331_half, 736.set nb331_three, 752.set nb331_is3, 768.set nb331_ii3, 772.set nb331_nri, 776.set nb331_iinr, 784.set nb331_jindex, 792.set nb331_jjnr, 800.set nb331_shift, 808.set nb331_shiftvec, 816.set nb331_facel, 824.set nb331_innerjjnr, 832.set nb331_ntia, 840.set nb331_innerk, 844.set nb331_n, 848.set nb331_nn1, 852.set nb331_nouter, 856.set nb331_ninner, 860 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $872,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb331_nouter(%rsp) movl %eax,nb331_ninner(%rsp) movl (%rdi),%edi movl %edi,nb331_nri(%rsp) movq %rsi,nb331_iinr(%rsp) movq %rdx,nb331_jindex(%rsp) movq %rcx,nb331_jjnr(%rsp) movq %r8,nb331_shift(%rsp) movq %r9,nb331_shiftvec(%rsp) movq nb331_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb331_facel(%rsp) movq nb331_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb331_tsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb331_half(%rsp) movl %ebx,nb331_half+4(%rsp) movsd nb331_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,nb331_half(%rsp) movapd %xmm2,nb331_two(%rsp) movapd %xmm3,nb331_three(%rsp) ## assume we have at least one i particle - start directly movq nb331_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb331_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 movsd 8(%rdx,%rbx,8),%xmm4 movsd nb331_facel(%rsp),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb331_iqO(%rsp) movapd %xmm4,nb331_iqH(%rsp) movq nb331_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb331_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb331_ntia(%rsp)_nb_kernel331_x86_64_sse2.nb331_threadloop: movq nb331_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel331_x86_64_sse2.nb331_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_kernel331_x86_64_sse2.nb331_spinlock ## if(nn1>nri) nn1=nri movl nb331_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,nb331_n(%rsp) movl %ebx,nb331_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel331_x86_64_sse2.nb331_outerstart jmp _nb_kernel331_x86_64_sse2.nb331_end_nb_kernel331_x86_64_sse2.nb331_outerstart: ## ebx contains number of outer iterations addl nb331_nouter(%rsp),%ebx movl %ebx,nb331_nouter(%rsp)_nb_kernel331_x86_64_sse2.nb331_outer: movq nb331_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,nb331_is3(%rsp) ## store is3 movq nb331_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 nb331_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb331_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb331_ii3(%rsp) 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,nb331_ixO(%rsp) movapd %xmm4,nb331_iyO(%rsp) movapd %xmm5,nb331_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,nb331_ixH1(%rsp) movapd %xmm1,nb331_iyH1(%rsp) movapd %xmm2,nb331_izH1(%rsp) movapd %xmm3,nb331_ixH2(%rsp) movapd %xmm4,nb331_iyH2(%rsp) movapd %xmm5,nb331_izH2(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb331_vctot(%rsp) movapd %xmm4,nb331_Vvdwtot(%rsp) movapd %xmm4,nb331_fixO(%rsp) movapd %xmm4,nb331_fiyO(%rsp) movapd %xmm4,nb331_fizO(%rsp) movapd %xmm4,nb331_fixH1(%rsp) movapd %xmm4,nb331_fiyH1(%rsp) movapd %xmm4,nb331_fizH1(%rsp) movapd %xmm4,nb331_fixH2(%rsp) movapd %xmm4,nb331_fiyH2(%rsp) movapd %xmm4,nb331_fizH2(%rsp) movq nb331_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 nb331_pos(%rbp),%rsi movq nb331_faction(%rbp),%rdi movq nb331_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb331_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb331_ninner(%rsp),%ecx movl %ecx,nb331_ninner(%rsp) addl $0,%edx movl %edx,nb331_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel331_x86_64_sse2.nb331_unroll_loop jmp _nb_kernel331_x86_64_sse2.nb331_checksingle_nb_kernel331_x86_64_sse2.nb331_unroll_loop: ## twice unrolled innerloop here movq nb331_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb331_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb331_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb331_iqO(%rsp),%xmm3 mulpd nb331_iqH(%rsp),%xmm4 movapd %xmm3,nb331_qqO(%rsp) movapd %xmm4,nb331_qqH(%rsp) movq nb331_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movq nb331_vdwparam(%rbp),%rsi shll %r8d shll %r9d movl nb331_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d movlpd (%rsi,%r8,8),%xmm6 ## c6a movlpd (%rsi,%r9,8),%xmm7 ## c6b movhpd 8(%rsi,%r8,8),%xmm6 ## c6a c12a movhpd 8(%rsi,%r9,8),%xmm7 ## c6b c12b movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movapd %xmm4,nb331_c6(%rsp) movapd %xmm6,nb331_c12(%rsp) movq nb331_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move j 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 = jx ## xmm1 = jy ## xmm2 = jz movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb331_ixO(%rsp),%xmm0 subpd nb331_iyO(%rsp),%xmm1 subpd nb331_izO(%rsp),%xmm2 subpd nb331_ixH1(%rsp),%xmm3 subpd nb331_iyH1(%rsp),%xmm4 subpd nb331_izH1(%rsp),%xmm5 subpd nb331_ixH2(%rsp),%xmm6 subpd nb331_iyH2(%rsp),%xmm7 subpd nb331_izH2(%rsp),%xmm8 movapd %xmm0,nb331_dxO(%rsp) movapd %xmm1,nb331_dyO(%rsp) movapd %xmm2,nb331_dzO(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb331_dxH1(%rsp) movapd %xmm4,nb331_dyH1(%rsp) movapd %xmm5,nb331_dzH1(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb331_dxH2(%rsp) movapd %xmm7,nb331_dyH2(%rsp) movapd %xmm8,nb331_dzH2(%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 j 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 nb331_three(%rsp),%xmm9 movapd %xmm9,%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?