nb_kernel211_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,925 行 · 第 1/4 页
S
1,925 行
#### $Id: nb_kernel211_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_kernel211_x86_64_sse2.globl _nb_kernel211_x86_64_sse2nb_kernel211_x86_64_sse2: _nb_kernel211_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb211_fshift, 16.set nb211_gid, 24.set nb211_pos, 32.set nb211_faction, 40.set nb211_charge, 48.set nb211_p_facel, 56.set nb211_argkrf, 64.set nb211_argcrf, 72.set nb211_Vc, 80.set nb211_type, 88.set nb211_p_ntype, 96.set nb211_vdwparam, 104.set nb211_Vvdw, 112.set nb211_p_tabscale, 120.set nb211_VFtab, 128.set nb211_invsqrta, 136.set nb211_dvda, 144.set nb211_p_gbtabscale, 152.set nb211_GBtab, 160.set nb211_p_nthreads, 168.set nb211_count, 176.set nb211_mtx, 184.set nb211_outeriter, 192.set nb211_inneriter, 200.set nb211_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb211_ixO, 0.set nb211_iyO, 16.set nb211_izO, 32.set nb211_ixH1, 48.set nb211_iyH1, 64.set nb211_izH1, 80.set nb211_ixH2, 96.set nb211_iyH2, 112.set nb211_izH2, 128.set nb211_iqO, 144.set nb211_iqH, 160.set nb211_dxO, 176.set nb211_dyO, 192.set nb211_dzO, 208.set nb211_dxH1, 224.set nb211_dyH1, 240.set nb211_dzH1, 256.set nb211_dxH2, 272.set nb211_dyH2, 288.set nb211_dzH2, 304.set nb211_qqO, 320.set nb211_qqH, 336.set nb211_c6, 352.set nb211_c12, 368.set nb211_six, 384.set nb211_twelve, 400.set nb211_vctot, 416.set nb211_Vvdwtot, 432.set nb211_fixO, 448.set nb211_fiyO, 464.set nb211_fizO, 480.set nb211_fixH1, 496.set nb211_fiyH1, 512.set nb211_fizH1, 528.set nb211_fixH2, 544.set nb211_fiyH2, 560.set nb211_fizH2, 576.set nb211_fjx, 592.set nb211_fjy, 608.set nb211_fjz, 624.set nb211_half, 640.set nb211_three, 656.set nb211_two, 672.set nb211_krf, 688.set nb211_crf, 704.set nb211_krsqO, 720.set nb211_krsqH1, 736.set nb211_krsqH2, 752.set nb211_nri, 768.set nb211_iinr, 776.set nb211_jindex, 784.set nb211_jjnr, 792.set nb211_shift, 800.set nb211_shiftvec, 808.set nb211_facel, 816.set nb211_innerjjnr, 824.set nb211_is3, 832.set nb211_ii3, 836.set nb211_ntia, 840.set nb211_innerk, 844.set nb211_n, 848.set nb211_nn1, 852.set nb211_nouter, 856.set nb211_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,nb211_nouter(%rsp) movl %eax,nb211_ninner(%rsp) movl (%rdi),%edi movl %edi,nb211_nri(%rsp) movq %rsi,nb211_iinr(%rsp) movq %rdx,nb211_jindex(%rsp) movq %rcx,nb211_jjnr(%rsp) movq %r8,nb211_shift(%rsp) movq %r9,nb211_shiftvec(%rsp) movq nb211_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb211_facel(%rsp) movq nb211_argkrf(%rbp),%rsi movq nb211_argcrf(%rbp),%rdi movsd (%rsi),%xmm1 movsd (%rdi),%xmm2 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm1,nb211_krf(%rsp) movapd %xmm2,nb211_crf(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb211_half(%rsp) movl %ebx,nb211_half+4(%rsp) movsd nb211_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 %xmm3,%xmm4 addpd %xmm4,%xmm4 ## six movapd %xmm4,%xmm5 addpd %xmm5,%xmm5 ## twelve movapd %xmm1,nb211_half(%rsp) movapd %xmm2,nb211_two(%rsp) movapd %xmm3,nb211_three(%rsp) movapd %xmm4,nb211_six(%rsp) movapd %xmm5,nb211_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb211_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb211_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 movsd 8(%rdx,%rbx,8),%xmm4 movsd nb211_facel(%rsp),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb211_iqO(%rsp) movapd %xmm4,nb211_iqH(%rsp) movq nb211_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb211_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb211_ntia(%rsp)_nb_kernel211_x86_64_sse2.nb211_threadloop: movq nb211_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel211_x86_64_sse2.nb211_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_kernel211_x86_64_sse2.nb211_spinlock ## if(nn1>nri) nn1=nri movl nb211_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,nb211_n(%rsp) movl %ebx,nb211_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel211_x86_64_sse2.nb211_outerstart jmp _nb_kernel211_x86_64_sse2.nb211_end_nb_kernel211_x86_64_sse2.nb211_outerstart: ## ebx contains number of outer iterations addl nb211_nouter(%rsp),%ebx movl %ebx,nb211_nouter(%rsp)_nb_kernel211_x86_64_sse2.nb211_outer: movq nb211_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,nb211_is3(%rsp) ## store is3 movq nb211_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 nb211_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 nb211_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb211_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,nb211_ixO(%rsp) movapd %xmm4,nb211_iyO(%rsp) movapd %xmm5,nb211_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,nb211_ixH1(%rsp) movapd %xmm1,nb211_iyH1(%rsp) movapd %xmm2,nb211_izH1(%rsp) movapd %xmm3,nb211_ixH2(%rsp) movapd %xmm4,nb211_iyH2(%rsp) movapd %xmm5,nb211_izH2(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb211_vctot(%rsp) movapd %xmm4,nb211_Vvdwtot(%rsp) movapd %xmm4,nb211_fixO(%rsp) movapd %xmm4,nb211_fiyO(%rsp) movapd %xmm4,nb211_fizO(%rsp) movapd %xmm4,nb211_fixH1(%rsp) movapd %xmm4,nb211_fiyH1(%rsp) movapd %xmm4,nb211_fizH1(%rsp) movapd %xmm4,nb211_fixH2(%rsp) movapd %xmm4,nb211_fiyH2(%rsp) movapd %xmm4,nb211_fizH2(%rsp) movq nb211_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 nb211_pos(%rbp),%rsi movq nb211_faction(%rbp),%rdi movq nb211_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb211_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb211_ninner(%rsp),%ecx movl %ecx,nb211_ninner(%rsp) addl $0,%edx movl %edx,nb211_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel211_x86_64_sse2.nb211_unroll_loop jmp _nb_kernel211_x86_64_sse2.nb211_checksingle_nb_kernel211_x86_64_sse2.nb211_unroll_loop: ## twice unrolled innerloop here movq nb211_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb211_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb211_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb211_iqO(%rsp),%xmm3 mulpd nb211_iqH(%rsp),%xmm4 movapd %xmm3,nb211_qqO(%rsp) movapd %xmm4,nb211_qqH(%rsp) movq nb211_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movq nb211_vdwparam(%rbp),%rsi shll %r8d shll %r9d movl nb211_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,nb211_c6(%rsp) movapd %xmm6,nb211_c12(%rsp) movq nb211_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 nb211_ixO(%rsp),%xmm0 subpd nb211_iyO(%rsp),%xmm1 subpd nb211_izO(%rsp),%xmm2 subpd nb211_ixH1(%rsp),%xmm3 subpd nb211_iyH1(%rsp),%xmm4 subpd nb211_izH1(%rsp),%xmm5 subpd nb211_ixH2(%rsp),%xmm6 subpd nb211_iyH2(%rsp),%xmm7 subpd nb211_izH2(%rsp),%xmm8 movapd %xmm0,nb211_dxO(%rsp) movapd %xmm1,nb211_dyO(%rsp) movapd %xmm2,nb211_dzO(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb211_dxH1(%rsp) movapd %xmm4,nb211_dyH1(%rsp) movapd %xmm5,nb211_dzH1(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb211_dxH2(%rsp) movapd %xmm7,nb211_dyH2(%rsp) movapd %xmm8,nb211_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 nb211_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) movapd nb211_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## first iteration for rinvO mulpd %xmm15,%xmm10 ## first iteration for rinvH1 mulpd %xmm15,%xmm11 ## first iteration for rinvH2 ## second iteration step movapd %xmm9,%xmm2 movapd %xmm10,%xmm5 movapd %xmm11,%xmm8 mulpd %xmm2,%xmm2 ## lu*lu mulpd %xmm5,%xmm5 ## lu*lu mulpd %xmm8,%xmm8 ## lu*lu movapd nb211_three(%rsp),%xmm1 movapd %xmm1,%xmm4 movapd %xmm1,%xmm7 mulpd %xmm0,%xmm2 ## rsq*lu*lu mulpd %xmm3,%xmm5 ## rsq*lu*lu mulpd %xmm6,%xmm8 ## rsq*lu*lu subpd %xmm2,%xmm1 subpd %xmm5,%xmm4 subpd %xmm8,%xmm7 ## 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?