nb_kernel430_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,642 行 · 第 1/4 页
S
1,642 行
#### $Id: nb_kernel430_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_kernel430_x86_64_sse2.globl _nb_kernel430_x86_64_sse2nb_kernel430_x86_64_sse2: _nb_kernel430_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb430_fshift, 16.set nb430_gid, 24.set nb430_pos, 32.set nb430_faction, 40.set nb430_charge, 48.set nb430_p_facel, 56.set nb430_argkrf, 64.set nb430_argcrf, 72.set nb430_Vc, 80.set nb430_type, 88.set nb430_p_ntype, 96.set nb430_vdwparam, 104.set nb430_Vvdw, 112.set nb430_p_tabscale, 120.set nb430_VFtab, 128.set nb430_invsqrta, 136.set nb430_dvda, 144.set nb430_p_gbtabscale, 152.set nb430_GBtab, 160.set nb430_p_nthreads, 168.set nb430_count, 176.set nb430_mtx, 184.set nb430_outeriter, 192.set nb430_inneriter, 200.set nb430_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb430_ix, 0.set nb430_iy, 16.set nb430_iz, 32.set nb430_iq, 48.set nb430_dx, 64.set nb430_dy, 80.set nb430_dz, 96.set nb430_eps, 112.set nb430_gbtsc, 128.set nb430_tsc, 144.set nb430_qq, 160.set nb430_c6, 176.set nb430_c12, 192.set nb430_epsgb, 208.set nb430_vctot, 224.set nb430_Vvdwtot, 240.set nb430_fix, 256.set nb430_fiy, 272.set nb430_fiz, 288.set nb430_half, 304.set nb430_three, 320.set nb430_r, 336.set nb430_isai, 352.set nb430_isaprod, 368.set nb430_dvdasum, 384.set nb430_gbscale, 400.set nb430_rinv, 416.set nb430_nri, 432.set nb430_iinr, 440.set nb430_jindex, 448.set nb430_jjnr, 456.set nb430_shift, 464.set nb430_shiftvec, 472.set nb430_facel, 480.set nb430_innerjjnr, 488.set nb430_ii, 496.set nb430_is3, 500.set nb430_ii3, 504.set nb430_ntia, 508.set nb430_innerk, 512.set nb430_n, 516.set nb430_nn1, 520.set nb430_ntype, 524.set nb430_nouter, 528.set nb430_ninner, 532 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $536,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb430_nouter(%rsp) movl %eax,nb430_ninner(%rsp) movl (%rdi),%edi movl %edi,nb430_nri(%rsp) movq %rsi,nb430_iinr(%rsp) movq %rdx,nb430_jindex(%rsp) movq %rcx,nb430_jjnr(%rsp) movq %r8,nb430_shift(%rsp) movq %r9,nb430_shiftvec(%rsp) movq nb430_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb430_ntype(%rsp) movq nb430_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb430_facel(%rsp) movq nb430_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb430_tsc(%rsp) movq nb430_p_gbtabscale(%rbp),%rbx movsd (%rbx),%xmm4 shufpd $0,%xmm4,%xmm4 movapd %xmm4,nb430_gbtsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb430_half(%rsp) movl %ebx,nb430_half+4(%rsp) movsd nb430_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,nb430_half(%rsp) movapd %xmm3,nb430_three(%rsp)_nb_kernel430_x86_64_sse2.nb430_threadloop: movq nb430_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel430_x86_64_sse2.nb430_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_kernel430_x86_64_sse2.nb430_spinlock ## if(nn1>nri) nn1=nri movl nb430_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,nb430_n(%rsp) movl %ebx,nb430_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel430_x86_64_sse2.nb430_outerstart jmp _nb_kernel430_x86_64_sse2.nb430_end_nb_kernel430_x86_64_sse2.nb430_outerstart: ## ebx contains number of outer iterations addl nb430_nouter(%rsp),%ebx movl %ebx,nb430_nouter(%rsp)_nb_kernel430_x86_64_sse2.nb430_outer: movq nb430_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,nb430_is3(%rsp) ## store is3 movq nb430_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 nb430_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movl %ebx,nb430_ii(%rsp) movq nb430_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 mulsd nb430_facel(%rsp),%xmm3 shufpd $0,%xmm3,%xmm3 movq nb430_invsqrta(%rbp),%rdx ## load invsqrta[ii] movsd (%rdx,%rbx,8),%xmm4 shufpd $0,%xmm4,%xmm4 movq nb430_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb430_ntype(%rsp),%edx shll %edx movl %edx,nb430_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb430_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,nb430_iq(%rsp) movapd %xmm4,nb430_isai(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb430_ix(%rsp) movapd %xmm1,nb430_iy(%rsp) movapd %xmm2,nb430_iz(%rsp) movl %ebx,nb430_ii3(%rsp) ## clear vctot and i forces xorpd %xmm4,%xmm4 movapd %xmm4,nb430_vctot(%rsp) movapd %xmm4,nb430_Vvdwtot(%rsp) movapd %xmm4,nb430_dvdasum(%rsp) movapd %xmm4,nb430_fix(%rsp) movapd %xmm4,nb430_fiy(%rsp) movapd %xmm4,nb430_fiz(%rsp) movq nb430_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 nb430_pos(%rbp),%rsi movq nb430_faction(%rbp),%rdi movq nb430_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb430_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb430_ninner(%rsp),%ecx movl %ecx,nb430_ninner(%rsp) addl $0,%edx movl %edx,nb430_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel430_x86_64_sse2.nb430_unroll_loop jmp _nb_kernel430_x86_64_sse2.nb430_checksingle_nb_kernel430_x86_64_sse2.nb430_unroll_loop: ## twice unrolled innerloop here movq nb430_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb430_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb430_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%r10 ## j3 lea (%rbx,%rbx,2),%r11 ## move two coordinates to xmm4-xmm6 movlpd (%rsi,%r10,8),%xmm4 movlpd 8(%rsi,%r10,8),%xmm5 movlpd 16(%rsi,%r10,8),%xmm6 movhpd (%rsi,%r11,8),%xmm4 movhpd 8(%rsi,%r11,8),%xmm5 movhpd 16(%rsi,%r11,8),%xmm6 ## calc dr subpd nb430_ix(%rsp),%xmm4 subpd nb430_iy(%rsp),%xmm5 subpd nb430_iz(%rsp),%xmm6 ## store dr movapd %xmm4,nb430_dx(%rsp) movapd %xmm5,nb430_dy(%rsp) movapd %xmm6,nb430_dz(%rsp) ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 ## rsq in xmm4 ## load isaj movq nb430_invsqrta(%rbp),%rsi movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 mulpd nb430_isai(%rsp),%xmm3 movapd %xmm3,nb430_isaprod(%rsp) movapd %xmm3,%xmm6 mulpd nb430_gbtsc(%rsp),%xmm3 movapd %xmm3,nb430_gbscale(%rsp) ##invsqrt cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb430_three(%rsp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb430_half(%rsp),%xmm0 subpd %xmm2,%xmm1 ## 30-rsq*lu*lu mulpd %xmm5,%xmm1 mulpd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) mulpd nb430_iq(%rsp),%xmm6 movq nb430_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 mulpd %xmm6,%xmm3 movapd %xmm3,nb430_qq(%rsp) movapd %xmm1,%xmm5 ## copy of lu mulpd %xmm1,%xmm1 ## lu*lu movapd nb430_three(%rsp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb430_half(%rsp),%xmm0 subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=iter2 of rinv mulpd %xmm0,%xmm4 ## xmm4=r movapd %xmm4,nb430_r(%rsp) movapd %xmm0,nb430_rinv(%rsp) movq nb430_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movl (%rsi,%rbx,4),%r9d shll %r8d shll %r9d movl nb430_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d movapd %xmm4,%xmm8 ## r mulpd nb430_gbscale(%rsp),%xmm4 mulpd nb430_tsc(%rsp),%xmm8 ## truncate and convert to integers cvttpd2pi %xmm4,%mm0 ## gb cvttpd2pi %xmm8,%mm1 ## lj ## convert back to float cvtpi2pd %mm0,%xmm6 ## gb cvtpi2pd %mm1,%xmm10 ## lj ## multiply by 4 and 8, respectively pslld $2,%mm0 ## gb pslld $3,%mm1 ## lj ## move to integer registers movd %mm0,%r12d ## gb movd %mm1,%r14d ## lj psrlq $32,%mm0 psrlq $32,%mm1 movd %mm0,%r13d ## gb movd %mm1,%r15d ## lj ## GB indices: r10-11 LJ indices: r12-r13 ## calculate eps subpd %xmm6,%xmm4 ## gb subpd %xmm10,%xmm8 ## lj movapd %xmm4,nb430_epsgb(%rsp) ## gb eps movapd %xmm8,nb430_eps(%rsp) ## lj eps movq nb430_GBtab(%rbp),%rsi movq nb430_VFtab(%rbp),%rdi ## load GB table data to xmm0-xmm3, disp to xmm4-xmm7, rep. to xmm8-xmm11 movapd (%rsi,%r12,8),%xmm0 ## Y1c F1c movapd (%rsi,%r13,8),%xmm12 ## Y2c F2c movapd (%rdi,%r14,8),%xmm4 ## Y1d F1d movapd (%rdi,%r15,8),%xmm13 ## Y2d F2d movapd 32(%rdi,%r14,8),%xmm8 ## Y1r F1r movapd 32(%rdi,%r15,8),%xmm14 ## Y2r F2r movapd %xmm0,%xmm1 movapd %xmm4,%xmm5 movapd %xmm8,%xmm9 unpcklpd %xmm12,%xmm0 ## Y1c Y2c unpckhpd %xmm12,%xmm1 ## F1c F2c unpcklpd %xmm13,%xmm4 ## Y1d Y2d unpckhpd %xmm13,%xmm5 ## F1d F2d unpcklpd %xmm14,%xmm8 ## Y1r Y2r unpckhpd %xmm14,%xmm9 ## F1r F2r ⌨️ 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?