nb_kernel304_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,383 行 · 第 1/5 页
S
2,383 行
#### $Id: nb_kernel304_x86_64_sse2.s,v 1.4.2.3 2006/09/22 08:40:36 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_kernel304_x86_64_sse2.globl _nb_kernel304_x86_64_sse2nb_kernel304_x86_64_sse2: _nb_kernel304_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb304_fshift, 16.set nb304_gid, 24.set nb304_pos, 32.set nb304_faction, 40.set nb304_charge, 48.set nb304_p_facel, 56.set nb304_argkrf, 64.set nb304_argcrf, 72.set nb304_Vc, 80.set nb304_type, 88.set nb304_p_ntype, 96.set nb304_vdwparam, 104.set nb304_Vvdw, 112.set nb304_p_tabscale, 120.set nb304_VFtab, 128.set nb304_invsqrta, 136.set nb304_dvda, 144.set nb304_p_gbtabscale, 152.set nb304_GBtab, 160.set nb304_p_nthreads, 168.set nb304_count, 176.set nb304_mtx, 184.set nb304_outeriter, 192.set nb304_inneriter, 200.set nb304_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb304_ixM, 0.set nb304_iyM, 16.set nb304_izM, 32.set nb304_ixH1, 48.set nb304_iyH1, 64.set nb304_izH1, 80.set nb304_ixH2, 96.set nb304_iyH2, 112.set nb304_izH2, 128.set nb304_jxM, 144.set nb304_jyM, 160.set nb304_jzM, 176.set nb304_jxH1, 192.set nb304_jyH1, 208.set nb304_jzH1, 224.set nb304_jxH2, 240.set nb304_jyH2, 256.set nb304_jzH2, 272.set nb304_dxMM, 288.set nb304_dyMM, 304.set nb304_dzMM, 320.set nb304_dxMH1, 336.set nb304_dyMH1, 352.set nb304_dzMH1, 368.set nb304_dxMH2, 384.set nb304_dyMH2, 400.set nb304_dzMH2, 416.set nb304_dxH1M, 432.set nb304_dyH1M, 448.set nb304_dzH1M, 464.set nb304_dxH1H1, 480.set nb304_dyH1H1, 496.set nb304_dzH1H1, 512.set nb304_dxH1H2, 528.set nb304_dyH1H2, 544.set nb304_dzH1H2, 560.set nb304_dxH2M, 576.set nb304_dyH2M, 592.set nb304_dzH2M, 608.set nb304_dxH2H1, 624.set nb304_dyH2H1, 640.set nb304_dzH2H1, 656.set nb304_dxH2H2, 672.set nb304_dyH2H2, 688.set nb304_dzH2H2, 704.set nb304_qqMM, 720.set nb304_qqMH, 736.set nb304_qqHH, 752.set nb304_two, 768.set nb304_tsc, 784.set nb304_vctot, 800.set nb304_fixM, 816.set nb304_fiyM, 832.set nb304_fizM, 848.set nb304_fixH1, 864.set nb304_fiyH1, 880.set nb304_fizH1, 896.set nb304_fixH2, 912.set nb304_fiyH2, 928.set nb304_fizH2, 944.set nb304_epsH1, 960.set nb304_epsH2, 976.set nb304_epsM, 992.set nb304_fjxH1, 1008.set nb304_fjyH1, 1024.set nb304_fjzH1, 1040.set nb304_fjxH2, 1056.set nb304_fjyH2, 1072.set nb304_fjzH2, 1088.set nb304_half, 1104.set nb304_three, 1120.set nb304_rsqMM, 1136.set nb304_rsqMH1, 1152.set nb304_rsqMH2, 1168.set nb304_rsqH1M, 1184.set nb304_rsqH1H1, 1200.set nb304_rsqH1H2, 1216.set nb304_rsqH2M, 1232.set nb304_rsqH2H1, 1248.set nb304_rsqH2H2, 1264.set nb304_rinvMM, 1280.set nb304_rinvMH1, 1296.set nb304_rinvMH2, 1312.set nb304_rinvH1M, 1328.set nb304_rinvH1H1, 1344.set nb304_rinvH1H2, 1360.set nb304_rinvH2M, 1376.set nb304_rinvH2H1, 1392.set nb304_rinvH2H2, 1408.set nb304_is3, 1424.set nb304_ii3, 1428.set nb304_nri, 1432.set nb304_iinr, 1440.set nb304_jindex, 1448.set nb304_jjnr, 1456.set nb304_shift, 1464.set nb304_shiftvec, 1472.set nb304_facel, 1480.set nb304_innerjjnr, 1488.set nb304_innerk, 1496.set nb304_n, 1500.set nb304_nn1, 1504.set nb304_nouter, 1508.set nb304_ninner, 1512 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $1528,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb304_nouter(%rsp) movl %eax,nb304_ninner(%rsp) movl (%rdi),%edi movl %edi,nb304_nri(%rsp) movq %rsi,nb304_iinr(%rsp) movq %rdx,nb304_jindex(%rsp) movq %rcx,nb304_jjnr(%rsp) movq %r8,nb304_shift(%rsp) movq %r9,nb304_shiftvec(%rsp) movq nb304_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb304_facel(%rsp) movq nb304_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb304_tsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb304_half(%rsp) movl %ebx,nb304_half+4(%rsp) movsd nb304_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 %xmm1,nb304_half(%rsp) movapd %xmm2,nb304_two(%rsp) movapd %xmm3,nb304_three(%rsp) ## assume we have at least one i particle - start directly movq nb304_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb304_charge(%rbp),%rdx movsd 24(%rdx,%rbx,8),%xmm3 movsd %xmm3,%xmm4 movsd 8(%rdx,%rbx,8),%xmm5 movq nb304_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd nb304_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,nb304_qqMM(%rsp) movapd %xmm4,nb304_qqMH(%rsp) movapd %xmm5,nb304_qqHH(%rsp)_nb_kernel304_x86_64_sse2.nb304_threadloop: movq nb304_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel304_x86_64_sse2.nb304_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_kernel304_x86_64_sse2.nb304_spinlock ## if(nn1>nri) nn1=nri movl nb304_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,nb304_n(%rsp) movl %ebx,nb304_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel304_x86_64_sse2.nb304_outerstart jmp _nb_kernel304_x86_64_sse2.nb304_end_nb_kernel304_x86_64_sse2.nb304_outerstart: ## ebx contains number of outer iterations addl nb304_nouter(%rsp),%ebx movl %ebx,nb304_nouter(%rsp)_nb_kernel304_x86_64_sse2.nb304_outer: movq nb304_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,nb304_is3(%rsp) ## store is3 movq nb304_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 nb304_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb304_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb304_ii3(%rsp) movapd %xmm0,%xmm3 movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 addsd 24(%rax,%rbx,8),%xmm3 addsd 32(%rax,%rbx,8),%xmm4 addsd 40(%rax,%rbx,8),%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb304_ixH1(%rsp) movapd %xmm4,nb304_iyH1(%rsp) movapd %xmm5,nb304_izH1(%rsp) movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 48(%rax,%rbx,8),%xmm0 addsd 56(%rax,%rbx,8),%xmm1 addsd 64(%rax,%rbx,8),%xmm2 addsd 72(%rax,%rbx,8),%xmm3 addsd 80(%rax,%rbx,8),%xmm4 addsd 88(%rax,%rbx,8),%xmm5 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 %xmm0,nb304_ixH2(%rsp) movapd %xmm1,nb304_iyH2(%rsp) movapd %xmm2,nb304_izH2(%rsp) movapd %xmm3,nb304_ixM(%rsp) movapd %xmm4,nb304_iyM(%rsp) movapd %xmm5,nb304_izM(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb304_vctot(%rsp) movapd %xmm4,nb304_fixM(%rsp) movapd %xmm4,nb304_fiyM(%rsp) movapd %xmm4,nb304_fizM(%rsp) movapd %xmm4,nb304_fixH1(%rsp) movapd %xmm4,nb304_fiyH1(%rsp) movapd %xmm4,nb304_fizH1(%rsp) movapd %xmm4,nb304_fixH2(%rsp) movapd %xmm4,nb304_fiyH2(%rsp) movapd %xmm4,nb304_fizH2(%rsp) movq nb304_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 nb304_pos(%rbp),%rsi movq nb304_faction(%rbp),%rdi movq nb304_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb304_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb304_ninner(%rsp),%ecx movl %ecx,nb304_ninner(%rsp) addl $0,%edx movl %edx,nb304_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel304_x86_64_sse2.nb304_unroll_loop jmp _nb_kernel304_x86_64_sse2.nb304_checksingle_nb_kernel304_x86_64_sse2.nb304_unroll_loop: ## twice unrolled innerloop here movq nb304_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb304_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb304_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move j H1 coordinates to local temp variables movlpd 24(%rsi,%rax,8),%xmm0 movlpd 32(%rsi,%rax,8),%xmm1 movlpd 40(%rsi,%rax,8),%xmm2 movhpd 24(%rsi,%rbx,8),%xmm0 movhpd 32(%rsi,%rbx,8),%xmm1 movhpd 40(%rsi,%rbx,8),%xmm2 ## xmm0 = H1x ## xmm1 = H1y ## xmm2 = H1z movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb304_ixH1(%rsp),%xmm0 subpd nb304_iyH1(%rsp),%xmm1 subpd nb304_izH1(%rsp),%xmm2 subpd nb304_ixH2(%rsp),%xmm3 subpd nb304_iyH2(%rsp),%xmm4 subpd nb304_izH2(%rsp),%xmm5 subpd nb304_ixM(%rsp),%xmm6 subpd nb304_iyM(%rsp),%xmm7 subpd nb304_izM(%rsp),%xmm8 movapd %xmm0,nb304_dxH1H1(%rsp) movapd %xmm1,nb304_dyH1H1(%rsp) movapd %xmm2,nb304_dzH1H1(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb304_dxH2H1(%rsp) movapd %xmm4,nb304_dyH2H1(%rsp) movapd %xmm5,nb304_dzH2H1(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb304_dxMH1(%rsp) movapd %xmm7,nb304_dyMH1(%rsp) movapd %xmm8,nb304_dzMH1(%rsp) mulpd %xmm6,%xmm6 mulpd %xmm7,%xmm7 mulpd %xmm8,%xmm8 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 addpd %xmm7,%xmm6 addpd %xmm8,%xmm6 ## start doing invsqrt for jH1 atoms cvtpd2ps %xmm0,%xmm1 cvtpd2ps %xmm3,%xmm4 cvtpd2ps %xmm6,%xmm7 rsqrtps %xmm1,%xmm1 rsqrtps %xmm4,%xmm4 rsqrtps %xmm7,%xmm7 cvtps2pd %xmm1,%xmm1 cvtps2pd %xmm4,%xmm4 cvtps2pd %xmm7,%xmm7 movapd %xmm1,%xmm2 movapd %xmm4,%xmm5 movapd %xmm7,%xmm8 mulpd %xmm1,%xmm1 ## lu*lu mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm7,%xmm7 ## lu*lu movapd nb304_three(%rsp),%xmm9 movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulpd %xmm0,%xmm1 ## rsq*lu*lu mulpd %xmm3,%xmm4 ## rsq*lu*lu mulpd %xmm6,%xmm7 ## rsq*lu*lu subpd %xmm1,%xmm9 subpd %xmm4,%xmm10 subpd %xmm7,%xmm11 ## 3-rsq*lu*lu mulpd %xmm2,%xmm9 mulpd %xmm5,%xmm10 mulpd %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb304_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## first iteration for rinvH1H1 mulpd %xmm15,%xmm10 ## first iteration for rinvH2H1 mulpd %xmm15,%xmm11 ## first iteration for rinvMH1 ## second iteration step movapd %xmm9,%xmm2 movapd %xmm10,%xmm5 movapd %xmm11,%xmm8 mulpd %xmm2,%xmm2 ## lu*lu mulpd %xmm5,%xmm5 ## lu*lu mulpd %xmm8,%xmm8 ## lu*lu movapd nb304_three(%rsp),%xmm1 movapd %xmm1,%xmm4 movapd %xmm1,%xmm7 mulpd %xmm0,%xmm2 ## rsq*lu*lu mulpd %xmm3,%xmm5 ## rsq*lu*lu mulpd %xmm6,%xmm8 ## rsq*lu*lu⌨️ 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?