nb_kernel131_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,955 行 · 第 1/4 页
S
1,955 行
#### $Id: nb_kernel131_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:49 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_kernel131_x86_64_sse.globl _nb_kernel131_x86_64_ssenb_kernel131_x86_64_sse: _nb_kernel131_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb131_fshift, 16.set nb131_gid, 24.set nb131_pos, 32.set nb131_faction, 40.set nb131_charge, 48.set nb131_p_facel, 56.set nb131_argkrf, 64.set nb131_argcrf, 72.set nb131_Vc, 80.set nb131_type, 88.set nb131_p_ntype, 96.set nb131_vdwparam, 104.set nb131_Vvdw, 112.set nb131_p_tabscale, 120.set nb131_VFtab, 128.set nb131_invsqrta, 136.set nb131_dvda, 144.set nb131_p_gbtabscale, 152.set nb131_GBtab, 160.set nb131_p_nthreads, 168.set nb131_count, 176.set nb131_mtx, 184.set nb131_outeriter, 192.set nb131_inneriter, 200.set nb131_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb131_ixO, 0.set nb131_iyO, 16.set nb131_izO, 32.set nb131_ixH1, 48.set nb131_iyH1, 64.set nb131_izH1, 80.set nb131_ixH2, 96.set nb131_iyH2, 112.set nb131_izH2, 128.set nb131_iqO, 144.set nb131_iqH, 160.set nb131_dxO, 176.set nb131_dyO, 192.set nb131_dzO, 208.set nb131_dxH1, 224.set nb131_dyH1, 240.set nb131_dzH1, 256.set nb131_dxH2, 272.set nb131_dyH2, 288.set nb131_dzH2, 304.set nb131_qqO, 320.set nb131_qqH, 336.set nb131_c6, 352.set nb131_c12, 368.set nb131_tsc, 384.set nb131_fstmp, 400.set nb131_vctot, 416.set nb131_Vvdwtot, 432.set nb131_fixO, 448.set nb131_fiyO, 464.set nb131_fizO, 480.set nb131_fixH1, 496.set nb131_fiyH1, 512.set nb131_fizH1, 528.set nb131_fixH2, 544.set nb131_fiyH2, 560.set nb131_fizH2, 576.set nb131_fjx, 592.set nb131_fjy, 608.set nb131_fjz, 624.set nb131_half, 640.set nb131_three, 656.set nb131_two, 672.set nb131_krf, 688.set nb131_crf, 704.set nb131_rsqO, 720.set nb131_rsqH1, 736.set nb131_rsqH2, 752.set nb131_rinvO, 768.set nb131_rinvH1, 784.set nb131_rinvH2, 800.set nb131_facel, 816.set nb131_iinr, 824.set nb131_jindex, 832.set nb131_jjnr, 840.set nb131_shift, 848.set nb131_shiftvec, 856.set nb131_innerjjnr, 864.set nb131_nri, 872.set nb131_is3, 876.set nb131_ii3, 880.set nb131_ntia, 884.set nb131_innerk, 888.set nb131_n, 892.set nb131_nn1, 896.set nb131_nouter, 900.set nb131_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,nb131_nouter(%rsp) movl %eax,nb131_ninner(%rsp) movl (%rdi),%edi movl %edi,nb131_nri(%rsp) movq %rsi,nb131_iinr(%rsp) movq %rdx,nb131_jindex(%rsp) movq %rcx,nb131_jjnr(%rsp) movq %r8,nb131_shift(%rsp) movq %r9,nb131_shiftvec(%rsp) movq nb131_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb131_facel(%rsp) movq nb131_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb131_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb131_half(%rsp) movss nb131_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 %xmm1,nb131_half(%rsp) movaps %xmm2,nb131_two(%rsp) movaps %xmm3,nb131_three(%rsp) ## assume we have at least one i particle - start directly movq nb131_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb131_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb131_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb131_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb131_iqO(%rsp) movaps %xmm4,nb131_iqH(%rsp) movq nb131_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb131_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb131_ntia(%rsp)_nb_kernel131_x86_64_sse.nb131_threadloop: movq nb131_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel131_x86_64_sse.nb131_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_kernel131_x86_64_sse.nb131_spinlock ## if(nn1>nri) nn1=nri movl nb131_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,nb131_n(%rsp) movl %ebx,nb131_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel131_x86_64_sse.nb131_outerstart jmp _nb_kernel131_x86_64_sse.nb131_end_nb_kernel131_x86_64_sse.nb131_outerstart: ## ebx contains number of outer iterations addl nb131_nouter(%rsp),%ebx movl %ebx,nb131_nouter(%rsp)_nb_kernel131_x86_64_sse.nb131_outer: movq nb131_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,nb131_is3(%rsp) ## store is3 movq nb131_shiftvec(%rsp),%rax ## eax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb131_iinr(%rsp),%rcx ## ecx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb131_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb131_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 addss 4(%rax,%rbx,4),%xmm4 addss 8(%rax,%rbx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb131_ixO(%rsp) movaps %xmm4,nb131_iyO(%rsp) movaps %xmm5,nb131_izO(%rsp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%rax,%rbx,4),%xmm0 addss 16(%rax,%rbx,4),%xmm1 addss 20(%rax,%rbx,4),%xmm2 addss 24(%rax,%rbx,4),%xmm3 addss 28(%rax,%rbx,4),%xmm4 addss 32(%rax,%rbx,4),%xmm5 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 %xmm0,nb131_ixH1(%rsp) movaps %xmm1,nb131_iyH1(%rsp) movaps %xmm2,nb131_izH1(%rsp) movaps %xmm3,nb131_ixH2(%rsp) movaps %xmm4,nb131_iyH2(%rsp) movaps %xmm5,nb131_izH2(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb131_vctot(%rsp) movaps %xmm4,nb131_Vvdwtot(%rsp) movaps %xmm4,nb131_fixO(%rsp) movaps %xmm4,nb131_fiyO(%rsp) movaps %xmm4,nb131_fizO(%rsp) movaps %xmm4,nb131_fixH1(%rsp) movaps %xmm4,nb131_fiyH1(%rsp) movaps %xmm4,nb131_fizH1(%rsp) movaps %xmm4,nb131_fixH2(%rsp) movaps %xmm4,nb131_fiyH2(%rsp) movaps %xmm4,nb131_fizH2(%rsp) movq nb131_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 nb131_pos(%rbp),%rsi movq nb131_faction(%rbp),%rdi movq nb131_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb131_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb131_ninner(%rsp),%ecx movl %ecx,nb131_ninner(%rsp) addl $0,%edx movl %edx,nb131_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel131_x86_64_sse.nb131_unroll_loop jmp _nb_kernel131_x86_64_sse.nb131_odd_inner_nb_kernel131_x86_64_sse.nb131_unroll_loop: ## quad-unroll innerloop here movq nb131_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,nb131_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb131_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 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb131_iqO(%rsp),%xmm3 mulps nb131_iqH(%rsp),%xmm4 movaps %xmm3,nb131_qqO(%rsp) movaps %xmm4,nb131_qqH(%rsp) movq nb131_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d movl (%rsi,%rcx,4),%r10d movl (%rsi,%rdx,4),%r11d movq nb131_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movl nb131_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 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movaps %xmm4,nb131_c6(%rsp) movaps %xmm6,nb131_c12(%rsp) movq nb131_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 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb131_ixO(%rsp),%xmm0 subps nb131_iyO(%rsp),%xmm1 subps nb131_izO(%rsp),%xmm2 subps nb131_ixH1(%rsp),%xmm3 subps nb131_iyH1(%rsp),%xmm4 subps nb131_izH1(%rsp),%xmm5 subps nb131_ixH2(%rsp),%xmm6 subps nb131_iyH2(%rsp),%xmm7 subps nb131_izH2(%rsp),%xmm8 movaps %xmm0,nb131_dxO(%rsp) movaps %xmm1,nb131_dyO(%rsp) movaps %xmm2,nb131_dzO(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb131_dxH1(%rsp) movaps %xmm4,nb131_dyH1(%rsp) movaps %xmm5,nb131_dzH1(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb131_dxH2(%rsp) movaps %xmm7,nb131_dyH2(%rsp) movaps %xmm8,nb131_dzH2(%rsp) mulps %xmm6,%xmm6 mulps %xmm7,%xmm7 mulps %xmm8,%xmm8 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 addps %xmm7,%xmm6 addps %xmm8,%xmm6 ## start doing invsqrt rsqrtps %xmm0,%xmm1 rsqrtps %xmm3,%xmm4 rsqrtps %xmm6,%xmm7 movaps %xmm1,%xmm2 movaps %xmm4,%xmm5 movaps %xmm7,%xmm8 mulps %xmm1,%xmm1 ## lu*lu mulps %xmm4,%xmm4 ## lu*lu mulps %xmm7,%xmm7 ## lu*lu movaps nb131_three(%rsp),%xmm9 movaps %xmm9,%xmm10 movaps %xmm9,%xmm11 mulps %xmm0,%xmm1 ## rsq*lu*lu mulps %xmm3,%xmm4 ## rsq*lu*lu mulps %xmm6,%xmm7 ## rsq*lu*lu subps %xmm1,%xmm9 subps %xmm4,%xmm10 subps %xmm7,%xmm11 ## 3-rsq*lu*lu mulps %xmm2,%xmm9 mulps %xmm5,%xmm10 mulps %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movaps nb131_half(%rsp),%xmm2 mulps %xmm2,%xmm9 ## rinvO mulps %xmm2,%xmm10 ## rinvH1 mulps %xmm2,%xmm11 ## rinvH2 ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movaps %xmm0,nb131_rsqO(%rsp) movaps %xmm3,nb131_rsqH1(%rsp) movaps %xmm6,nb131_rsqH2(%rsp) movaps %xmm9,nb131_rinvO(%rsp)⌨️ 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?