nb_kernel314_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,365 行 · 第 1/5 页
S
2,365 行
#### $Id: nb_kernel314_x86_64_sse2.s,v 1.4.2.3 2006/09/22 08:40:37 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_kernel314_x86_64_sse2.globl _nb_kernel314_x86_64_sse2nb_kernel314_x86_64_sse2: _nb_kernel314_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb314_fshift, 16.set nb314_gid, 24.set nb314_pos, 32.set nb314_faction, 40.set nb314_charge, 48.set nb314_p_facel, 56.set nb314_argkrf, 64.set nb314_argcrf, 72.set nb314_Vc, 80.set nb314_type, 88.set nb314_p_ntype, 96.set nb314_vdwparam, 104.set nb314_Vvdw, 112.set nb314_p_tabscale, 120.set nb314_VFtab, 128.set nb314_invsqrta, 136.set nb314_dvda, 144.set nb314_p_gbtabscale, 152.set nb314_GBtab, 160.set nb314_p_nthreads, 168.set nb314_count, 176.set nb314_mtx, 184.set nb314_outeriter, 192.set nb314_inneriter, 200.set nb314_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb314_ixO, 0.set nb314_iyO, 16.set nb314_izO, 32.set nb314_ixH1, 48.set nb314_iyH1, 64.set nb314_izH1, 80.set nb314_ixH2, 96.set nb314_iyH2, 112.set nb314_izH2, 128.set nb314_ixM, 144.set nb314_iyM, 160.set nb314_izM, 176.set nb314_jxO, 192.set nb314_jyO, 208.set nb314_jzO, 224.set nb314_jxH1, 240.set nb314_jyH1, 256.set nb314_jzH1, 272.set nb314_jxH2, 288.set nb314_jyH2, 304.set nb314_jzH2, 320.set nb314_jxM, 336.set nb314_jyM, 352.set nb314_jzM, 368.set nb314_dxOO, 384.set nb314_dyOO, 400.set nb314_dzOO, 416.set nb314_dxH1H1, 432.set nb314_dyH1H1, 448.set nb314_dzH1H1, 464.set nb314_dxH1H2, 480.set nb314_dyH1H2, 496.set nb314_dzH1H2, 512.set nb314_dxH1M, 528.set nb314_dyH1M, 544.set nb314_dzH1M, 560.set nb314_dxH2H1, 576.set nb314_dyH2H1, 592.set nb314_dzH2H1, 608.set nb314_dxH2H2, 624.set nb314_dyH2H2, 640.set nb314_dzH2H2, 656.set nb314_dxH2M, 672.set nb314_dyH2M, 688.set nb314_dzH2M, 704.set nb314_dxMH1, 720.set nb314_dyMH1, 736.set nb314_dzMH1, 752.set nb314_dxMH2, 768.set nb314_dyMH2, 784.set nb314_dzMH2, 800.set nb314_dxMM, 816.set nb314_dyMM, 832.set nb314_dzMM, 848.set nb314_qqMM, 864.set nb314_qqMH, 880.set nb314_qqHH, 896.set nb314_two, 912.set nb314_tsc, 928.set nb314_c6, 944.set nb314_c12, 960.set nb314_vctot, 976.set nb314_Vvdwtot, 992.set nb314_fixO, 1008.set nb314_fiyO, 1024.set nb314_fizO, 1040.set nb314_fixH1, 1056.set nb314_fiyH1, 1072.set nb314_fizH1, 1088.set nb314_fixH2, 1104.set nb314_fiyH2, 1120.set nb314_fizH2, 1136.set nb314_fixM, 1152.set nb314_fiyM, 1168.set nb314_fizM, 1184.set nb314_fjxO, 1200.set nb314_fjyO, 1216.set nb314_fjzO, 1232.set nb314_fjxH1, 1248.set nb314_fjyH1, 1264.set nb314_fjzH1, 1280.set nb314_fjxH2, 1296.set nb314_fjyH2, 1312.set nb314_fjzH2, 1328.set nb314_epsH1, 1344.set nb314_epsH2, 1360.set nb314_epsM, 1376.set nb314_half, 1392.set nb314_three, 1408.set nb314_six, 1424.set nb314_twelve, 1440.set nb314_rsqOO, 1456.set nb314_rsqH1H1, 1472.set nb314_rsqH1H2, 1488.set nb314_rsqH1M, 1504.set nb314_rsqH2H1, 1520.set nb314_rsqH2H2, 1536.set nb314_rsqH2M, 1552.set nb314_rsqMH1, 1568.set nb314_rsqMH2, 1584.set nb314_rsqMM, 1600.set nb314_rinvsqOO, 1616.set nb314_rinvH1H1, 1632.set nb314_rinvH1H2, 1648.set nb314_rinvH1M, 1664.set nb314_rinvH2H1, 1680.set nb314_rinvH2H2, 1696.set nb314_rinvH2M, 1712.set nb314_rinvMH1, 1728.set nb314_rinvMH2, 1744.set nb314_rinvMM, 1760.set nb314_is3, 1776.set nb314_ii3, 1780.set nb314_nri, 1784.set nb314_iinr, 1792.set nb314_jindex, 1800.set nb314_jjnr, 1808.set nb314_shift, 1816.set nb314_shiftvec, 1824.set nb314_facel, 1832.set nb314_innerjjnr, 1840.set nb314_innerk, 1848.set nb314_n, 1852.set nb314_nn1, 1856.set nb314_nouter, 1860.set nb314_ninner, 1864 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1880,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb314_nouter(%rsp) movl %eax,nb314_ninner(%rsp) movl (%rdi),%edi movl %edi,nb314_nri(%rsp) movq %rsi,nb314_iinr(%rsp) movq %rdx,nb314_jindex(%rsp) movq %rcx,nb314_jjnr(%rsp) movq %r8,nb314_shift(%rsp) movq %r9,nb314_shiftvec(%rsp) movq nb314_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb314_facel(%rsp) movq nb314_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb314_tsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb314_half(%rsp) movl %ebx,nb314_half+4(%rsp) movsd nb314_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,nb314_half(%rsp) movapd %xmm2,nb314_two(%rsp) movapd %xmm3,nb314_three(%rsp) movapd %xmm4,nb314_six(%rsp) movapd %xmm5,nb314_twelve(%rsp) ## assume we have at least one i particle - start directly movq nb314_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb314_charge(%rbp),%rdx movsd 24(%rdx,%rbx,8),%xmm3 movsd %xmm3,%xmm4 movsd 8(%rdx,%rbx,8),%xmm5 movq nb314_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd nb314_facel(%rsp),%xmm6 mulsd %xmm3,%xmm3 mulsd %xmm5,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm3 mulsd %xmm6,%xmm4 mulsd %xmm6,%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb314_qqMM(%rsp) movapd %xmm4,nb314_qqMH(%rsp) movapd %xmm5,nb314_qqHH(%rsp) xorpd %xmm0,%xmm0 movq nb314_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movl %ecx,%edx movq nb314_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] addl %ecx,%edx movq nb314_vdwparam(%rbp),%rax movlpd (%rax,%rdx,8),%xmm0 movlpd 8(%rax,%rdx,8),%xmm1 shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 movapd %xmm0,nb314_c6(%rsp) movapd %xmm1,nb314_c12(%rsp)_nb_kernel314_x86_64_sse2.nb314_threadloop: movq nb314_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel314_x86_64_sse2.nb314_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_kernel314_x86_64_sse2.nb314_spinlock ## if(nn1>nri) nn1=nri movl nb314_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,nb314_n(%rsp) movl %ebx,nb314_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel314_x86_64_sse2.nb314_outerstart jmp _nb_kernel314_x86_64_sse2.nb314_end_nb_kernel314_x86_64_sse2.nb314_outerstart: ## ebx contains number of outer iterations addl nb314_nouter(%rsp),%ebx movl %ebx,nb314_nouter(%rsp)_nb_kernel314_x86_64_sse2.nb314_outer: movq nb314_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,nb314_is3(%rsp) ## store is3 movq nb314_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 nb314_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb314_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb314_ii3(%rsp) addsd (%rax,%rbx,8),%xmm3 ## ox addsd 8(%rax,%rbx,8),%xmm4 ## oy addsd 16(%rax,%rbx,8),%xmm5 ## oz addsd 24(%rax,%rbx,8),%xmm6 ## h1x addsd 32(%rax,%rbx,8),%xmm7 ## h1y shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 shufpd $0,%xmm6,%xmm6 shufpd $0,%xmm7,%xmm7 movapd %xmm3,nb314_ixO(%rsp) movapd %xmm4,nb314_iyO(%rsp) movapd %xmm5,nb314_izO(%rsp) movapd %xmm6,nb314_ixH1(%rsp) movapd %xmm7,nb314_iyH1(%rsp) movsd %xmm2,%xmm6 movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 40(%rax,%rbx,8),%xmm6 ## h1z addsd 48(%rax,%rbx,8),%xmm0 ## h2x addsd 56(%rax,%rbx,8),%xmm1 ## h2y addsd 64(%rax,%rbx,8),%xmm2 ## h2z addsd 72(%rax,%rbx,8),%xmm3 ## mx addsd 80(%rax,%rbx,8),%xmm4 ## my addsd 88(%rax,%rbx,8),%xmm5 ## mz shufpd $0,%xmm6,%xmm6 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 %xmm6,nb314_izH1(%rsp) movapd %xmm0,nb314_ixH2(%rsp) movapd %xmm1,nb314_iyH2(%rsp) movapd %xmm2,nb314_izH2(%rsp) movapd %xmm3,nb314_ixM(%rsp) movapd %xmm4,nb314_iyM(%rsp) movapd %xmm5,nb314_izM(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb314_vctot(%rsp) movapd %xmm4,nb314_Vvdwtot(%rsp) movapd %xmm4,nb314_fixO(%rsp) movapd %xmm4,nb314_fiyO(%rsp) movapd %xmm4,nb314_fizO(%rsp) movapd %xmm4,nb314_fixH1(%rsp) movapd %xmm4,nb314_fiyH1(%rsp) movapd %xmm4,nb314_fizH1(%rsp) movapd %xmm4,nb314_fixH2(%rsp) movapd %xmm4,nb314_fiyH2(%rsp) movapd %xmm4,nb314_fizH2(%rsp) movapd %xmm4,nb314_fixM(%rsp) movapd %xmm4,nb314_fiyM(%rsp) movapd %xmm4,nb314_fizM(%rsp) movq nb314_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 nb314_pos(%rbp),%rsi movq nb314_faction(%rbp),%rdi movq nb314_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb314_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb314_ninner(%rsp),%ecx movl %ecx,nb314_ninner(%rsp) addl $0,%edx movl %edx,nb314_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel314_x86_64_sse2.nb314_unroll_loop jmp _nb_kernel314_x86_64_sse2.nb314_checksingle_nb_kernel314_x86_64_sse2.nb314_unroll_loop: ## twice unrolled innerloop here movq nb314_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb314_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb314_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## load j O coordinates movlpd (%rsi,%rax,8),%xmm4 movlpd 8(%rsi,%rax,8),%xmm5 movlpd 16(%rsi,%rax,8),%xmm6 movhpd (%rsi,%rbx,8),%xmm4 movhpd 8(%rsi,%rbx,8),%xmm5 movhpd 16(%rsi,%rbx,8),%xmm6 ## xmm4 = Ox ## xmm5 = Oy ## xmm6 = Oz subpd nb314_ixO(%rsp),%xmm4 subpd nb314_iyO(%rsp),%xmm5 subpd nb314_izO(%rsp),%xmm6 ## store dx/dy/dz movapd %xmm4,%xmm9 movapd %xmm5,%xmm10 movapd %xmm6,%xmm11 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 ## rsq in xmm4 cvtpd2ps %xmm4,%xmm6 rcpps %xmm6,%xmm6 cvtps2pd %xmm6,%xmm6 ## lu in low xmm6 ## 1/x lookup seed in xmm6 movapd nb314_two(%rsp),%xmm0 movapd %xmm4,%xmm5 mulpd %xmm6,%xmm4 ## lu*rsq subpd %xmm4,%xmm0 ## 2-lu*rsq mulpd %xmm0,%xmm6 ## (new lu) movapd nb314_two(%rsp),%xmm0 mulpd %xmm6,%xmm5 ## lu*rsq subpd %xmm5,%xmm0 ## 2-lu*rsq mulpd %xmm6,%xmm0 ## xmm0=rinvsq movapd %xmm0,%xmm1 mulpd %xmm0,%xmm1 mulpd %xmm0,%xmm1 ## xmm1=rinvsix movapd %xmm1,%xmm2⌨️ 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?