nb_kernel400_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,214 行 · 第 1/3 页
S
1,214 行
#### $Id: nb_kernel400_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_kernel400_x86_64_sse2.globl _nb_kernel400_x86_64_sse2nb_kernel400_x86_64_sse2: _nb_kernel400_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb400_fshift, 16.set nb400_gid, 24.set nb400_pos, 32.set nb400_faction, 40.set nb400_charge, 48.set nb400_p_facel, 56.set nb400_argkrf, 64.set nb400_argcrf, 72.set nb400_Vc, 80.set nb400_type, 88.set nb400_p_ntype, 96.set nb400_vdwparam, 104.set nb400_Vvdw, 112.set nb400_p_tabscale, 120.set nb400_VFtab, 128.set nb400_invsqrta, 136.set nb400_dvda, 144.set nb400_p_gbtabscale, 152.set nb400_GBtab, 160.set nb400_p_nthreads, 168.set nb400_count, 176.set nb400_mtx, 184.set nb400_outeriter, 192.set nb400_inneriter, 200.set nb400_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb400_ix, 0.set nb400_iy, 16.set nb400_iz, 32.set nb400_iq, 48.set nb400_dx, 64.set nb400_dy, 80.set nb400_dz, 96.set nb400_two, 112.set nb400_gbtsc, 128.set nb400_qq, 144.set nb400_r, 160.set nb400_vctot, 176.set nb400_fix, 192.set nb400_fiy, 208.set nb400_fiz, 224.set nb400_half, 240.set nb400_three, 256.set nb400_isai, 272.set nb400_isaprod, 288.set nb400_dvdasum, 304.set nb400_gbscale, 320.set nb400_nri, 336.set nb400_iinr, 344.set nb400_jindex, 352.set nb400_jjnr, 360.set nb400_shift, 368.set nb400_shiftvec, 376.set nb400_facel, 384.set nb400_innerjjnr, 392.set nb400_is3, 400.set nb400_ii3, 404.set nb400_ii, 408.set nb400_innerk, 412.set nb400_n, 416.set nb400_nn1, 420.set nb400_nouter, 424.set nb400_ninner, 428 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $440,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb400_nouter(%rsp) movl %eax,nb400_ninner(%rsp) movl (%rdi),%edi movl %edi,nb400_nri(%rsp) movq %rsi,nb400_iinr(%rsp) movq %rdx,nb400_jindex(%rsp) movq %rcx,nb400_jjnr(%rsp) movq %r8,nb400_shift(%rsp) movq %r9,nb400_shiftvec(%rsp) movq nb400_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb400_facel(%rsp) movq nb400_p_gbtabscale(%rbp),%rbx movsd (%rbx),%xmm4 shufpd $0,%xmm4,%xmm4 movapd %xmm4,nb400_gbtsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb400_half(%rsp) movl %ebx,nb400_half+4(%rsp) movsd nb400_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,nb400_half(%rsp) movapd %xmm2,nb400_two(%rsp) movapd %xmm3,nb400_three(%rsp)_nb_kernel400_x86_64_sse2.nb400_threadloop: movq nb400_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel400_x86_64_sse2.nb400_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_kernel400_x86_64_sse2.nb400_spinlock ## if(nn1>nri) nn1=nri movl nb400_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,nb400_n(%rsp) movl %ebx,nb400_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel400_x86_64_sse2.nb400_outerstart jmp _nb_kernel400_x86_64_sse2.nb400_end_nb_kernel400_x86_64_sse2.nb400_outerstart: ## ebx contains number of outer iterations addl nb400_nouter(%rsp),%ebx movl %ebx,nb400_nouter(%rsp)_nb_kernel400_x86_64_sse2.nb400_outer: movq nb400_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,nb400_is3(%rsp) ## store is3 movq nb400_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 nb400_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movl %ebx,nb400_ii(%rsp) movq nb400_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 mulsd nb400_facel(%rsp),%xmm3 shufpd $0,%xmm3,%xmm3 movq nb400_invsqrta(%rbp),%rdx ## load invsqrta[ii] movsd (%rdx,%rbx,8),%xmm4 shufpd $0,%xmm4,%xmm4 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb400_pos(%rbp),%rax ## rax = base of pos[] addsd (%rax,%rbx,8),%xmm0 addsd 8(%rax,%rbx,8),%xmm1 addsd 16(%rax,%rbx,8),%xmm2 movapd %xmm3,nb400_iq(%rsp) movapd %xmm4,nb400_isai(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb400_ix(%rsp) movapd %xmm1,nb400_iy(%rsp) movapd %xmm2,nb400_iz(%rsp) movl %ebx,nb400_ii3(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,%xmm8 movapd %xmm4,%xmm12 movapd %xmm4,%xmm13 movapd %xmm4,%xmm14 movapd %xmm4,%xmm15 movq nb400_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 nb400_pos(%rbp),%rsi movq nb400_faction(%rbp),%rdi movq nb400_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb400_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb400_ninner(%rsp),%ecx movl %ecx,nb400_ninner(%rsp) addl $0,%edx movl %edx,nb400_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel400_x86_64_sse2.nb400_unroll_loop jmp _nb_kernel400_x86_64_sse2.nb400_checksingle_nb_kernel400_x86_64_sse2.nb400_unroll_loop: ## twice unrolled innerloop here movq nb400_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r12d movl 4(%rdx),%r13d addq $8,nb400_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb400_pos(%rbp),%rsi ## base of pos[] lea (%r12,%r12,2),%r8 ## j3 lea (%r13,%r13,2),%r9 ## move two coordinates to xmm4-xmm6 movlpd (%rsi,%r8,8),%xmm4 movlpd 8(%rsi,%r8,8),%xmm5 movlpd 16(%rsi,%r8,8),%xmm6 movhpd (%rsi,%r9,8),%xmm4 movhpd 8(%rsi,%r9,8),%xmm5 movhpd 16(%rsi,%r9,8),%xmm6 ## calc dr subpd nb400_ix(%rsp),%xmm4 subpd nb400_iy(%rsp),%xmm5 subpd nb400_iz(%rsp),%xmm6 ## store dr 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 movq nb400_invsqrta(%rbp),%rsi movlpd (%rsi,%r12,8),%xmm3 cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm5,%xmm2 ## lu in low xmm2 movhpd (%rsi,%r13,8),%xmm3 mulpd nb400_isai(%rsp),%xmm3 movapd %xmm3,nb400_isaprod(%rsp) movapd %xmm3,%xmm6 mulpd nb400_gbtsc(%rsp),%xmm3 movapd %xmm3,nb400_gbscale(%rsp) ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb400_three(%rsp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb400_half(%rsp),%xmm0 subpd %xmm2,%xmm1 ## 30-rsq*lu*lu mulpd %xmm5,%xmm1 mulpd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movq nb400_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%r12,8),%xmm3 movapd %xmm1,%xmm5 ## copy of lu mulpd %xmm1,%xmm1 ## lu*lu movapd nb400_three(%rsp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb400_half(%rsp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=iter2 of rinv (new lu) mulpd %xmm0,%xmm4 ## xmm4=r mulpd nb400_iq(%rsp),%xmm6 movhpd (%rsi,%r13,8),%xmm3 mulpd %xmm6,%xmm3 movapd %xmm3,nb400_qq(%rsp) movapd %xmm4,nb400_r(%rsp) mulpd nb400_gbscale(%rsp),%xmm4 cvttpd2pi %xmm4,%mm6 ## mm6 = lu idx cvtpi2pd %mm6,%xmm5 subpd %xmm5,%xmm4 movapd %xmm4,%xmm1 ## xmm1=eps pslld $2,%mm6 ## idx *= 4 movq nb400_GBtab(%rbp),%rsi movd %mm6,%r10d psrlq $32,%mm6 movd %mm6,%r11d ## indices in r10/r11 movapd (%rsi,%r10,8),%xmm4 ## Y1 F1 movapd (%rsi,%r11,8),%xmm3 ## Y2 F2 movapd %xmm4,%xmm5 unpcklpd %xmm3,%xmm4 ## Y1 Y2 unpckhpd %xmm3,%xmm5 ## F1 F2 movapd 16(%rsi,%r10,8),%xmm6 ## G1 H1 movapd 16(%rsi,%r11,8),%xmm3 ## G2 H2 movapd %xmm6,%xmm7 unpcklpd %xmm3,%xmm6 ## G1 G2 unpckhpd %xmm3,%xmm7 ## H1 H2 ## coulomb table ready, in xmm4-xmm7 mulpd %xmm1,%xmm7 ## xmm7=Heps mulpd %xmm1,%xmm6 ## xmm6=Geps mulpd %xmm1,%xmm7 ## xmm7=Heps2 addpd %xmm6,%xmm5 addpd %xmm7,%xmm5 ## xmm5=Fp addpd %xmm7,%xmm7 ## two*Heps2 movapd nb400_qq(%rsp),%xmm3 addpd %xmm6,%xmm7 addpd %xmm5,%xmm7 ## xmm7=FF mulpd %xmm1,%xmm5 ## xmm5=eps*Fp addpd %xmm4,%xmm5 ## xmm5=VV mulpd %xmm3,%xmm5 ## vcoul=qq*VV mulpd %xmm7,%xmm3 ## fijC=FF*qq movq nb400_dvda(%rbp),%rsi ## Calculate dVda xorpd %xmm7,%xmm7 mulpd nb400_gbscale(%rsp),%xmm3 movapd %xmm3,%xmm6 mulpd nb400_r(%rsp),%xmm6 addpd %xmm5,%xmm6 ## update vctot addpd %xmm5,%xmm12 ## xmm6=(vcoul+fijC*r) subpd %xmm6,%xmm7 movapd %xmm7,%xmm6 ## update dvdasum addpd %xmm7,%xmm8 ## update j atoms dvdaj movhlps %xmm6,%xmm7 addsd (%rsi,%r12,8),%xmm6 addsd (%rsi,%r13,8),%xmm7 movsd %xmm6,(%rsi,%r12,8) movsd %xmm7,(%rsi,%r13,8) ## the fj's - start by accumulating forces from memory movq nb400_faction(%rbp),%rdi movlpd (%rdi,%r8,8),%xmm5 movlpd 8(%rdi,%r8,8),%xmm6 movlpd 16(%rdi,%r8,8),%xmm7 movhpd (%rdi,%r9,8),%xmm5 movhpd 8(%rdi,%r9,8),%xmm6 movhpd 16(%rdi,%r9,8),%xmm7⌨️ 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?