nb_kernel213_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,930 行 · 第 1/4 页
S
1,930 行
#### $Id: nb_kernel213_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:50 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_kernel213_x86_64_sse.globl _nb_kernel213_x86_64_ssenb_kernel213_x86_64_sse: _nb_kernel213_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb213_fshift, 16.set nb213_gid, 24.set nb213_pos, 32.set nb213_faction, 40.set nb213_charge, 48.set nb213_p_facel, 56.set nb213_argkrf, 64.set nb213_argcrf, 72.set nb213_Vc, 80.set nb213_type, 88.set nb213_p_ntype, 96.set nb213_vdwparam, 104.set nb213_Vvdw, 112.set nb213_p_tabscale, 120.set nb213_VFtab, 128.set nb213_invsqrta, 136.set nb213_dvda, 144.set nb213_p_gbtabscale, 152.set nb213_GBtab, 160.set nb213_p_nthreads, 168.set nb213_count, 176.set nb213_mtx, 184.set nb213_outeriter, 192.set nb213_inneriter, 200.set nb213_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb213_ixO, 0.set nb213_iyO, 16.set nb213_izO, 32.set nb213_ixH1, 48.set nb213_iyH1, 64.set nb213_izH1, 80.set nb213_ixH2, 96.set nb213_iyH2, 112.set nb213_izH2, 128.set nb213_ixM, 144.set nb213_iyM, 160.set nb213_izM, 176.set nb213_iqM, 192.set nb213_iqH, 208.set nb213_dxO, 224.set nb213_dyO, 240.set nb213_dzO, 256.set nb213_dxH1, 272.set nb213_dyH1, 288.set nb213_dzH1, 304.set nb213_dxH2, 320.set nb213_dyH2, 336.set nb213_dzH2, 352.set nb213_dxM, 368.set nb213_dyM, 384.set nb213_dzM, 400.set nb213_qqM, 416.set nb213_qqH, 432.set nb213_rinvH1, 448.set nb213_rinvH2, 464.set nb213_rinvM, 480.set nb213_two, 496.set nb213_c6, 512.set nb213_c12, 528.set nb213_six, 544.set nb213_twelve, 560.set nb213_krf, 576.set nb213_crf, 592.set nb213_krsqH1, 608.set nb213_krsqH2, 624.set nb213_krsqM, 640.set nb213_vctot, 656.set nb213_Vvdwtot, 672.set nb213_fixO, 688.set nb213_fiyO, 704.set nb213_fizO, 720.set nb213_fixH1, 736.set nb213_fiyH1, 752.set nb213_fizH1, 768.set nb213_fixH2, 784.set nb213_fiyH2, 800.set nb213_fizH2, 816.set nb213_fixM, 832.set nb213_fiyM, 848.set nb213_fizM, 864.set nb213_fjx, 880.set nb213_fjy, 896.set nb213_fjz, 912.set nb213_half, 928.set nb213_three, 944.set nb213_is3, 960.set nb213_ii3, 964.set nb213_nri, 968.set nb213_iinr, 976.set nb213_jindex, 984.set nb213_jjnr, 992.set nb213_shift, 1000.set nb213_shiftvec, 1008.set nb213_facel, 1016.set nb213_innerjjnr, 1024.set nb213_ntia, 1032.set nb213_innerk, 1036.set nb213_n, 1040.set nb213_nn1, 1044.set nb213_nouter, 1048.set nb213_ninner, 1052 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1064,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb213_nouter(%rsp) movl %eax,nb213_ninner(%rsp) movl (%rdi),%edi movl %edi,nb213_nri(%rsp) movq %rsi,nb213_iinr(%rsp) movq %rdx,nb213_jindex(%rsp) movq %rcx,nb213_jjnr(%rsp) movq %r8,nb213_shift(%rsp) movq %r9,nb213_shiftvec(%rsp) movq nb213_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb213_facel(%rsp) movq nb213_argkrf(%rbp),%rsi movq nb213_argcrf(%rbp),%rdi movss (%rsi),%xmm1 movss (%rdi),%xmm2 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm1,nb213_krf(%rsp) movaps %xmm2,nb213_crf(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb213_half(%rsp) movss nb213_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,nb213_half(%rsp) movaps %xmm2,nb213_two(%rsp) movaps %xmm3,nb213_three(%rsp) movaps %xmm4,nb213_six(%rsp) movaps %xmm5,nb213_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb213_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb213_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm4 movss 12(%rdx,%rbx,4),%xmm3 movq nb213_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb213_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb213_iqM(%rsp) movaps %xmm4,nb213_iqH(%rsp) movq nb213_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb213_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb213_ntia(%rsp)_nb_kernel213_x86_64_sse.nb213_threadloop: movq nb213_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel213_x86_64_sse.nb213_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addq $1,%rbx ## rbx=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_kernel213_x86_64_sse.nb213_spinlock ## if(nn1>nri) nn1=nri movl nb213_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,nb213_n(%rsp) movl %ebx,nb213_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel213_x86_64_sse.nb213_outerstart jmp _nb_kernel213_x86_64_sse.nb213_end_nb_kernel213_x86_64_sse.nb213_outerstart: ## ebx contains number of outer iterations addl nb213_nouter(%rsp),%ebx movl %ebx,nb213_nouter(%rsp)_nb_kernel213_x86_64_sse.nb213_outer: movq nb213_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,nb213_is3(%rsp) ## store is3 movq nb213_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 nb213_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 nb213_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb213_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,nb213_ixO(%rsp) movaps %xmm4,nb213_iyO(%rsp) movaps %xmm5,nb213_izO(%rsp) movaps %xmm6,nb213_ixH1(%rsp) movaps %xmm7,nb213_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,nb213_izH1(%rsp) movaps %xmm0,nb213_ixH2(%rsp) movaps %xmm1,nb213_iyH2(%rsp) movaps %xmm2,nb213_izH2(%rsp) movaps %xmm3,nb213_ixM(%rsp) movaps %xmm4,nb213_iyM(%rsp) movaps %xmm5,nb213_izM(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb213_vctot(%rsp) movaps %xmm4,nb213_Vvdwtot(%rsp) movaps %xmm4,nb213_fixO(%rsp) movaps %xmm4,nb213_fiyO(%rsp) movaps %xmm4,nb213_fizO(%rsp) movaps %xmm4,nb213_fixH1(%rsp) movaps %xmm4,nb213_fiyH1(%rsp) movaps %xmm4,nb213_fizH1(%rsp) movaps %xmm4,nb213_fixH2(%rsp) movaps %xmm4,nb213_fiyH2(%rsp) movaps %xmm4,nb213_fizH2(%rsp) movaps %xmm4,nb213_fixM(%rsp) movaps %xmm4,nb213_fiyM(%rsp) movaps %xmm4,nb213_fizM(%rsp) movq nb213_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 nb213_pos(%rbp),%rsi movq nb213_faction(%rbp),%rdi movq nb213_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb213_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb213_ninner(%rsp),%ecx movl %ecx,nb213_ninner(%rsp) addl $0,%edx movl %edx,nb213_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel213_x86_64_sse.nb213_unroll_loop jmp _nb_kernel213_x86_64_sse.nb213_odd_inner_nb_kernel213_x86_64_sse.nb213_unroll_loop: ## quad-unroll innerloop here movq nb213_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,nb213_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb213_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 nb213_iqM(%rsp),%xmm3 mulps nb213_iqH(%rsp),%xmm4 movaps %xmm3,nb213_qqM(%rsp) movaps %xmm4,nb213_qqH(%rsp) movq nb213_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movl (%rsi,%rcx,4),%r10d movl (%rsi,%rdx,4),%r11d movq nb213_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movl nb213_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,nb213_c6(%rsp) movaps %xmm6,nb213_c12(%rsp) movq nb213_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 ## xmm0 = jx ## xmm1 = jy ## xmm2 = jz ## O interaction ## copy to xmm3-xmm5 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb213_ixO(%rsp),%xmm3 subps nb213_iyO(%rsp),%xmm4 subps nb213_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 nb213_two(%rsp),%xmm4 mulps %xmm5,%xmm3 subps %xmm3,%xmm4 mulps %xmm5,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm3 ## rinvsq mulps %xmm4,%xmm4 ## rinv4 mulps %xmm3,%xmm4 ## rinv6 movaps %xmm4,%xmm5 mulps %xmm5,%xmm5 ## rinv12 mulps nb213_c6(%rsp),%xmm4 mulps nb213_c12(%rsp),%xmm5 movaps %xmm5,%xmm6 subps %xmm4,%xmm6 ## Vvdw=vvdw12-vvdw6 mulps nb213_six(%rsp),%xmm4 mulps nb213_twelve(%rsp),%xmm5 subps %xmm4,%xmm5 mulps %xmm5,%xmm3 ## fscal⌨️ 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?