nb_kernel231_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,124 行 · 第 1/5 页
S
2,124 行
#### $Id: nb_kernel231_x86_64_sse2.s,v 1.4.2.3 2006/09/22 08:40:34 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_kernel231_x86_64_sse2.globl _nb_kernel231_x86_64_sse2nb_kernel231_x86_64_sse2: _nb_kernel231_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb231_fshift, 16.set nb231_gid, 24.set nb231_pos, 32.set nb231_faction, 40.set nb231_charge, 48.set nb231_p_facel, 56.set nb231_argkrf, 64.set nb231_argcrf, 72.set nb231_Vc, 80.set nb231_type, 88.set nb231_p_ntype, 96.set nb231_vdwparam, 104.set nb231_Vvdw, 112.set nb231_p_tabscale, 120.set nb231_VFtab, 128.set nb231_invsqrta, 136.set nb231_dvda, 144.set nb231_p_gbtabscale, 152.set nb231_GBtab, 160.set nb231_p_nthreads, 168.set nb231_count, 176.set nb231_mtx, 184.set nb231_outeriter, 192.set nb231_inneriter, 200.set nb231_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb231_ixO, 0.set nb231_iyO, 16.set nb231_izO, 32.set nb231_ixH1, 48.set nb231_iyH1, 64.set nb231_izH1, 80.set nb231_ixH2, 96.set nb231_iyH2, 112.set nb231_izH2, 128.set nb231_iqO, 144.set nb231_iqH, 160.set nb231_dxO, 176.set nb231_dyO, 192.set nb231_dzO, 208.set nb231_dxH1, 224.set nb231_dyH1, 240.set nb231_dzH1, 256.set nb231_dxH2, 272.set nb231_dyH2, 288.set nb231_dzH2, 304.set nb231_qqO, 320.set nb231_qqH, 336.set nb231_c6, 352.set nb231_c12, 368.set nb231_tsc, 384.set nb231_fstmp, 400.set nb231_vctot, 416.set nb231_Vvdwtot, 432.set nb231_fixO, 448.set nb231_fiyO, 464.set nb231_fizO, 480.set nb231_fixH1, 496.set nb231_fiyH1, 512.set nb231_fizH1, 528.set nb231_fixH2, 544.set nb231_fiyH2, 560.set nb231_fizH2, 576.set nb231_fjx, 592.set nb231_fjy, 608.set nb231_fjz, 624.set nb231_half, 640.set nb231_three, 656.set nb231_two, 672.set nb231_krf, 688.set nb231_crf, 704.set nb231_rsqO, 720.set nb231_rsqH1, 736.set nb231_rsqH2, 752.set nb231_rinvO, 768.set nb231_rinvH1, 784.set nb231_rinvH2, 800.set nb231_facel, 816.set nb231_iinr, 824.set nb231_jindex, 832.set nb231_jjnr, 840.set nb231_shift, 848.set nb231_shiftvec, 856.set nb231_innerjjnr, 864.set nb231_nri, 872.set nb231_is3, 876.set nb231_ii3, 880.set nb231_ntia, 884.set nb231_innerk, 888.set nb231_n, 892.set nb231_nn1, 896.set nb231_nouter, 900.set nb231_ninner, 904 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $920,%rsp # # local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb231_nouter(%rsp) movl %eax,nb231_ninner(%rsp) movl (%rdi),%edi movl %edi,nb231_nri(%rsp) movq %rsi,nb231_iinr(%rsp) movq %rdx,nb231_jindex(%rsp) movq %rcx,nb231_jjnr(%rsp) movq %r8,nb231_shift(%rsp) movq %r9,nb231_shiftvec(%rsp) movq nb231_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb231_facel(%rsp) movq nb231_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb231_tsc(%rsp) movq nb231_argkrf(%rbp),%rsi movq nb231_argcrf(%rbp),%rdi movsd (%rsi),%xmm1 movsd (%rdi),%xmm2 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm1,nb231_krf(%rsp) movapd %xmm2,nb231_crf(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb231_half(%rsp) movl %ebx,nb231_half+4(%rsp) movsd nb231_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,nb231_half(%rsp) movapd %xmm2,nb231_two(%rsp) movapd %xmm3,nb231_three(%rsp) ## assume we have at least one i particle - start directly movq nb231_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb231_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 movsd 8(%rdx,%rbx,8),%xmm4 movsd nb231_facel(%rsp),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb231_iqO(%rsp) movapd %xmm4,nb231_iqH(%rsp) movq nb231_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb231_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb231_ntia(%rsp)_nb_kernel231_x86_64_sse2.nb231_threadloop: movq nb231_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel231_x86_64_sse2.nb231_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_kernel231_x86_64_sse2.nb231_spinlock ## if(nn1>nri) nn1=nri movl nb231_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,nb231_n(%rsp) movl %ebx,nb231_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel231_x86_64_sse2.nb231_outerstart jmp _nb_kernel231_x86_64_sse2.nb231_end_nb_kernel231_x86_64_sse2.nb231_outerstart: ## ebx contains number of outer iterations addl nb231_nouter(%rsp),%ebx movl %ebx,nb231_nouter(%rsp)_nb_kernel231_x86_64_sse2.nb231_outer: movq nb231_shift(%rsp),%rax ## eax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb231_is3(%rsp) ## store is3 movq nb231_shiftvec(%rsp),%rax ## eax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%xmm2 movq nb231_iinr(%rsp),%rcx ## ecx = 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 nb231_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb231_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,nb231_ixO(%rsp) movapd %xmm4,nb231_iyO(%rsp) movapd %xmm5,nb231_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,nb231_ixH1(%rsp) movapd %xmm1,nb231_iyH1(%rsp) movapd %xmm2,nb231_izH1(%rsp) movapd %xmm3,nb231_ixH2(%rsp) movapd %xmm4,nb231_iyH2(%rsp) movapd %xmm5,nb231_izH2(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb231_vctot(%rsp) movapd %xmm4,nb231_Vvdwtot(%rsp) movapd %xmm4,nb231_fixO(%rsp) movapd %xmm4,nb231_fiyO(%rsp) movapd %xmm4,nb231_fizO(%rsp) movapd %xmm4,nb231_fixH1(%rsp) movapd %xmm4,nb231_fiyH1(%rsp) movapd %xmm4,nb231_fizH1(%rsp) movapd %xmm4,nb231_fixH2(%rsp) movapd %xmm4,nb231_fiyH2(%rsp) movapd %xmm4,nb231_fizH2(%rsp) movq nb231_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 nb231_pos(%rbp),%rsi movq nb231_faction(%rbp),%rdi movq nb231_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb231_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb231_ninner(%rsp),%ecx movl %ecx,nb231_ninner(%rsp) addl $0,%edx movl %edx,nb231_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel231_x86_64_sse2.nb231_unroll_loop jmp _nb_kernel231_x86_64_sse2.nb231_checksingle_nb_kernel231_x86_64_sse2.nb231_unroll_loop: ## twice unrolled innerloop here movq nb231_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb231_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb231_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb231_iqO(%rsp),%xmm3 mulpd nb231_iqH(%rsp),%xmm4 movapd %xmm3,nb231_qqO(%rsp) movapd %xmm4,nb231_qqH(%rsp) movq nb231_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movq nb231_vdwparam(%rbp),%rsi shll %r8d shll %r9d movl nb231_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,nb231_c6(%rsp) movapd %xmm6,nb231_c12(%rsp) movq nb231_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 nb231_ixO(%rsp),%xmm0 subpd nb231_iyO(%rsp),%xmm1 subpd nb231_izO(%rsp),%xmm2 subpd nb231_ixH1(%rsp),%xmm3 subpd nb231_iyH1(%rsp),%xmm4 subpd nb231_izH1(%rsp),%xmm5 subpd nb231_ixH2(%rsp),%xmm6 subpd nb231_iyH2(%rsp),%xmm7 subpd nb231_izH2(%rsp),%xmm8 movapd %xmm0,nb231_dxO(%rsp) movapd %xmm1,nb231_dyO(%rsp) movapd %xmm2,nb231_dzO(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb231_dxH1(%rsp) movapd %xmm4,nb231_dyH1(%rsp) movapd %xmm5,nb231_dzH1(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb231_dxH2(%rsp) movapd %xmm7,nb231_dyH2(%rsp) movapd %xmm8,nb231_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⌨️ 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?