nb_kernel313_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,144 行 · 第 1/5 页
S
2,144 行
#### $Id: nb_kernel313_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:51 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_kernel313_x86_64_sse.globl _nb_kernel313_x86_64_ssenb_kernel313_x86_64_sse: _nb_kernel313_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb313_fshift, 16.set nb313_gid, 24.set nb313_pos, 32.set nb313_faction, 40.set nb313_charge, 48.set nb313_p_facel, 56.set nb313_argkrf, 64.set nb313_argcrf, 72.set nb313_Vc, 80.set nb313_type, 88.set nb313_p_ntype, 96.set nb313_vdwparam, 104.set nb313_Vvdw, 112.set nb313_p_tabscale, 120.set nb313_VFtab, 128.set nb313_invsqrta, 136.set nb313_dvda, 144.set nb313_p_gbtabscale, 152.set nb313_GBtab, 160.set nb313_p_nthreads, 168.set nb313_count, 176.set nb313_mtx, 184.set nb313_outeriter, 192.set nb313_inneriter, 200.set nb313_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb313_ixO, 0.set nb313_iyO, 16.set nb313_izO, 32.set nb313_ixH1, 48.set nb313_iyH1, 64.set nb313_izH1, 80.set nb313_ixH2, 96.set nb313_iyH2, 112.set nb313_izH2, 128.set nb313_ixM, 144.set nb313_iyM, 160.set nb313_izM, 176.set nb313_iqM, 192.set nb313_iqH, 208.set nb313_epsH1, 224.set nb313_epsH2, 240.set nb313_epsM, 256.set nb313_dxH1, 272.set nb313_dyH1, 288.set nb313_dzH1, 304.set nb313_dxH2, 320.set nb313_dyH2, 336.set nb313_dzH2, 352.set nb313_dxM, 368.set nb313_dyM, 384.set nb313_dzM, 400.set nb313_qqM, 416.set nb313_qqH, 432.set nb313_rinvH1, 448.set nb313_rinvH2, 464.set nb313_rinvM, 480.set nb313_rH1, 496.set nb313_rH2, 512.set nb313_rM, 528.set nb313_tsc, 544.set nb313_two, 560.set nb313_c6, 576.set nb313_c12, 592.set nb313_six, 608.set nb313_twelve, 624.set nb313_vctot, 640.set nb313_Vvdwtot, 656.set nb313_fixO, 672.set nb313_fiyO, 688.set nb313_fizO, 704.set nb313_fixH1, 720.set nb313_fiyH1, 736.set nb313_fizH1, 752.set nb313_fixH2, 768.set nb313_fiyH2, 784.set nb313_fizH2, 800.set nb313_fixM, 816.set nb313_fiyM, 832.set nb313_fizM, 848.set nb313_fjx, 864.set nb313_fjy, 880.set nb313_fjz, 896.set nb313_half, 912.set nb313_three, 928.set nb313_is3, 944.set nb313_ii3, 948.set nb313_nri, 952.set nb313_iinr, 960.set nb313_jindex, 968.set nb313_jjnr, 976.set nb313_shift, 984.set nb313_shiftvec, 992.set nb313_facel, 1000.set nb313_innerjjnr, 1008.set nb313_ntia, 1016.set nb313_innerk, 1020.set nb313_n, 1024.set nb313_nn1, 1028.set nb313_nouter, 1032.set nb313_ninner, 1036 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1048,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb313_nouter(%rsp) movl %eax,nb313_ninner(%rsp) movl (%rdi),%edi movl %edi,nb313_nri(%rsp) movq %rsi,nb313_iinr(%rsp) movq %rdx,nb313_jindex(%rsp) movq %rcx,nb313_jjnr(%rsp) movq %r8,nb313_shift(%rsp) movq %r9,nb313_shiftvec(%rsp) movq nb313_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb313_facel(%rsp) movq nb313_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb313_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb313_half(%rsp) movss nb313_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm3,%xmm4 addps %xmm4,%xmm4 ## six movaps %xmm4,%xmm5 addps %xmm5,%xmm5 ## twelve movaps %xmm1,nb313_half(%rsp) movaps %xmm2,nb313_two(%rsp) movaps %xmm3,nb313_three(%rsp) movaps %xmm4,nb313_six(%rsp) movaps %xmm5,nb313_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb313_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb313_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm4 movss 12(%rdx,%rbx,4),%xmm3 movq nb313_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb313_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb313_iqM(%rsp) movaps %xmm4,nb313_iqH(%rsp) movq nb313_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb313_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb313_ntia(%rsp)_nb_kernel313_x86_64_sse.nb313_threadloop: movq nb313_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel313_x86_64_sse.nb313_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_kernel313_x86_64_sse.nb313_spinlock ## if(nn1>nri) nn1=nri movl nb313_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,nb313_n(%rsp) movl %ebx,nb313_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel313_x86_64_sse.nb313_outerstart jmp _nb_kernel313_x86_64_sse.nb313_end_nb_kernel313_x86_64_sse.nb313_outerstart: ## ebx contains number of outer iterations addl nb313_nouter(%rsp),%ebx movl %ebx,nb313_nouter(%rsp)_nb_kernel313_x86_64_sse.nb313_outer: movq nb313_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,nb313_is3(%rsp) ## store is3 movq nb313_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb313_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb313_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb313_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 ## ox addss 4(%rax,%rbx,4),%xmm4 ## oy addss 8(%rax,%rbx,4),%xmm5 ## oz addss 12(%rax,%rbx,4),%xmm6 ## h1x addss 16(%rax,%rbx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb313_ixO(%rsp) movaps %xmm4,nb313_iyO(%rsp) movaps %xmm5,nb313_izO(%rsp) movaps %xmm6,nb313_ixH1(%rsp) movaps %xmm7,nb313_iyH1(%rsp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%rax,%rbx,4),%xmm6 ## h1z addss 24(%rax,%rbx,4),%xmm0 ## h2x addss 28(%rax,%rbx,4),%xmm1 ## h2y addss 32(%rax,%rbx,4),%xmm2 ## h2z addss 36(%rax,%rbx,4),%xmm3 ## mx addss 40(%rax,%rbx,4),%xmm4 ## my addss 44(%rax,%rbx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm6,nb313_izH1(%rsp) movaps %xmm0,nb313_ixH2(%rsp) movaps %xmm1,nb313_iyH2(%rsp) movaps %xmm2,nb313_izH2(%rsp) movaps %xmm3,nb313_ixM(%rsp) movaps %xmm4,nb313_iyM(%rsp) movaps %xmm5,nb313_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb313_vctot(%rsp) movaps %xmm4,nb313_Vvdwtot(%rsp) movaps %xmm4,nb313_fixO(%rsp) movaps %xmm4,nb313_fiyO(%rsp) movaps %xmm4,nb313_fizO(%rsp) movaps %xmm4,nb313_fixH1(%rsp) movaps %xmm4,nb313_fiyH1(%rsp) movaps %xmm4,nb313_fizH1(%rsp) movaps %xmm4,nb313_fixH2(%rsp) movaps %xmm4,nb313_fiyH2(%rsp) movaps %xmm4,nb313_fizH2(%rsp) movaps %xmm4,nb313_fixM(%rsp) movaps %xmm4,nb313_fiyM(%rsp) movaps %xmm4,nb313_fizM(%rsp) movq nb313_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 nb313_pos(%rbp),%rsi movq nb313_faction(%rbp),%rdi movq nb313_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb313_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb313_ninner(%rsp),%ecx movl %ecx,nb313_ninner(%rsp) addl $0,%edx movl %edx,nb313_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel313_x86_64_sse.nb313_unroll_loop jmp _nb_kernel313_x86_64_sse.nb313_odd_inner_nb_kernel313_x86_64_sse.nb313_unroll_loop: ## quad-unroll innerloop here movq nb313_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb313_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb313_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb313_iqM(%rsp),%xmm3 mulps nb313_iqH(%rsp),%xmm4 movaps %xmm3,nb313_qqM(%rsp) movaps %xmm4,nb313_qqH(%rsp) movq nb313_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movl (%rsi,%rcx,4),%r10d movl (%rsi,%rdx,4),%r11d movq nb313_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movl nb313_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d addl %edi,%r10d addl %edi,%r11d movlps (%rsi,%r8,4),%xmm6 movlps (%rsi,%r10,4),%xmm7 movhps (%rsi,%r9,4),%xmm6 movhps (%rsi,%r11,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm6 ## 11011101 movaps %xmm4,nb313_c6(%rsp) movaps %xmm6,nb313_c12(%rsp) movq nb313_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx ## replace jnr with j3 lea (%rdx,%rdx,2),%rdx ## move four coordinates to xmm0-xmm2 movlps (%rsi,%rax,4),%xmm4 movlps (%rsi,%rcx,4),%xmm5 movss 8(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm6 movhps (%rsi,%rbx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movss 8(%rsi,%rbx,4),%xmm0 movss 8(%rsi,%rdx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ⌨️ 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?