nb_kernel113_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,821 行 · 第 1/4 页
S
1,821 行
#### $Id: nb_kernel113_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:48 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_kernel113_x86_64_sse.globl _nb_kernel113_x86_64_ssenb_kernel113_x86_64_sse: _nb_kernel113_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb113_fshift, 16.set nb113_gid, 24.set nb113_pos, 32.set nb113_faction, 40.set nb113_charge, 48.set nb113_p_facel, 56.set nb113_argkrf, 64.set nb113_argcrf, 72.set nb113_Vc, 80.set nb113_type, 88.set nb113_p_ntype, 96.set nb113_vdwparam, 104.set nb113_Vvdw, 112.set nb113_p_tabscale, 120.set nb113_VFtab, 128.set nb113_invsqrta, 136.set nb113_dvda, 144.set nb113_p_gbtabscale, 152.set nb113_GBtab, 160.set nb113_p_nthreads, 168.set nb113_count, 176.set nb113_mtx, 184.set nb113_outeriter, 192.set nb113_inneriter, 200.set nb113_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb113_ixO, 0.set nb113_iyO, 16.set nb113_izO, 32.set nb113_ixH1, 48.set nb113_iyH1, 64.set nb113_izH1, 80.set nb113_ixH2, 96.set nb113_iyH2, 112.set nb113_izH2, 128.set nb113_ixM, 144.set nb113_iyM, 160.set nb113_izM, 176.set nb113_iqM, 192.set nb113_iqH, 208.set nb113_dxO, 224.set nb113_dyO, 240.set nb113_dzO, 256.set nb113_dxH1, 272.set nb113_dyH1, 288.set nb113_dzH1, 304.set nb113_dxH2, 320.set nb113_dyH2, 336.set nb113_dzH2, 352.set nb113_dxM, 368.set nb113_dyM, 384.set nb113_dzM, 400.set nb113_qqM, 416.set nb113_qqH, 432.set nb113_rinvH1, 448.set nb113_rinvH2, 464.set nb113_rinvM, 480.set nb113_two, 496.set nb113_c6, 512.set nb113_c12, 528.set nb113_six, 544.set nb113_twelve, 560.set nb113_vctot, 576.set nb113_Vvdwtot, 592.set nb113_fixO, 608.set nb113_fiyO, 624.set nb113_fizO, 640.set nb113_fixH1, 656.set nb113_fiyH1, 672.set nb113_fizH1, 688.set nb113_fixH2, 704.set nb113_fiyH2, 720.set nb113_fizH2, 736.set nb113_fixM, 752.set nb113_fiyM, 768.set nb113_fizM, 784.set nb113_fjx, 800.set nb113_fjy, 816.set nb113_fjz, 832.set nb113_half, 848.set nb113_three, 864.set nb113_nri, 880.set nb113_iinr, 888.set nb113_jindex, 896.set nb113_jjnr, 904.set nb113_shift, 912.set nb113_shiftvec, 920.set nb113_facel, 928.set nb113_innerjjnr, 936.set nb113_ntia, 944.set nb113_is3, 948.set nb113_ii3, 952.set nb113_innerk, 956.set nb113_n, 960.set nb113_nn1, 964.set nb113_nouter, 968.set nb113_ninner, 972 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $984,%rsp ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb113_nouter(%rsp) movl %eax,nb113_ninner(%rsp) movl (%rdi),%edi movl %edi,nb113_nri(%rsp) movq %rsi,nb113_iinr(%rsp) movq %rdx,nb113_jindex(%rsp) movq %rcx,nb113_jjnr(%rsp) movq %r8,nb113_shift(%rsp) movq %r9,nb113_shiftvec(%rsp) movq nb113_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb113_facel(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb113_half(%rsp) movss nb113_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,nb113_half(%rsp) movaps %xmm2,nb113_two(%rsp) movaps %xmm3,nb113_three(%rsp) movaps %xmm4,nb113_six(%rsp) movaps %xmm5,nb113_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb113_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb113_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm4 movss 12(%rdx,%rbx,4),%xmm3 movq nb113_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb113_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb113_iqM(%rsp) movaps %xmm4,nb113_iqH(%rsp) movq nb113_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb113_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb113_ntia(%rsp)_nb_kernel113_x86_64_sse.nb113_threadloop: movq nb113_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel113_x86_64_sse.nb113_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_kernel113_x86_64_sse.nb113_spinlock ## if(nn1>nri) nn1=nri movl nb113_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,nb113_n(%rsp) movl %ebx,nb113_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel113_x86_64_sse.nb113_outerstart jmp _nb_kernel113_x86_64_sse.nb113_end_nb_kernel113_x86_64_sse.nb113_outerstart: ## ebx contains number of outer iterations addl nb113_nouter(%rsp),%ebx movl %ebx,nb113_nouter(%rsp)_nb_kernel113_x86_64_sse.nb113_outer: movq nb113_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,nb113_is3(%rsp) ## store is3 movq nb113_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 nb113_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 nb113_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb113_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,nb113_ixO(%rsp) movaps %xmm4,nb113_iyO(%rsp) movaps %xmm5,nb113_izO(%rsp) movaps %xmm6,nb113_ixH1(%rsp) movaps %xmm7,nb113_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,nb113_izH1(%rsp) movaps %xmm0,nb113_ixH2(%rsp) movaps %xmm1,nb113_iyH2(%rsp) movaps %xmm2,nb113_izH2(%rsp) movaps %xmm3,nb113_ixM(%rsp) movaps %xmm4,nb113_iyM(%rsp) movaps %xmm5,nb113_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb113_vctot(%rsp) movaps %xmm4,nb113_Vvdwtot(%rsp) movaps %xmm4,nb113_fixO(%rsp) movaps %xmm4,nb113_fiyO(%rsp) movaps %xmm4,nb113_fizO(%rsp) movaps %xmm4,nb113_fixH1(%rsp) movaps %xmm4,nb113_fiyH1(%rsp) movaps %xmm4,nb113_fizH1(%rsp) movaps %xmm4,nb113_fixH2(%rsp) movaps %xmm4,nb113_fiyH2(%rsp) movaps %xmm4,nb113_fizH2(%rsp) movaps %xmm4,nb113_fixM(%rsp) movaps %xmm4,nb113_fiyM(%rsp) movaps %xmm4,nb113_fizM(%rsp) movq nb113_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 nb113_pos(%rbp),%rsi movq nb113_faction(%rbp),%rdi movq nb113_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb113_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb113_ninner(%rsp),%ecx movl %ecx,nb113_ninner(%rsp) addl $0,%edx movl %edx,nb113_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel113_x86_64_sse.nb113_unroll_loop jmp _nb_kernel113_x86_64_sse.nb113_odd_inner_nb_kernel113_x86_64_sse.nb113_unroll_loop: ## quad-unroll innerloop here movq nb113_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,nb113_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb113_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 nb113_iqM(%rsp),%xmm3 mulps nb113_iqH(%rsp),%xmm4 movaps %xmm3,nb113_qqM(%rsp) movaps %xmm4,nb113_qqH(%rsp) movq nb113_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movl (%rsi,%rcx,4),%r10d movl (%rsi,%rdx,4),%r11d movq nb113_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movl nb113_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,nb113_c6(%rsp) movaps %xmm6,nb113_c12(%rsp) movq nb113_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 j 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 ## xmm0 = jx ## xmm1 = jy ## xmm2 = jz ## interaction ## copy to xmm3-xmm5 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb113_ixO(%rsp),%xmm3 subps nb113_iyO(%rsp),%xmm4 subps nb113_izO(%rsp),%xmm5 movaps %xmm3,%xmm13 movaps %xmm4,%xmm14 movaps %xmm5,%xmm15 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm3 addps %xmm5,%xmm3 ## calc 1/rsq rcpps %xmm3,%xmm5 movaps nb113_two(%rsp),%xmm4 mulps %xmm5,%xmm3 subps %xmm3,%xmm4 mulps %xmm5,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm3 ## rinvsq mulps %xmm4,%xmm4 ## rinv4⌨️ 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?