nb_kernel100_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 946 行 · 第 1/2 页
S
946 行
#### $Id: nb_kernel100_x86_64_sse2.s,v 1.4.2.3 2006/09/22 08:40:31 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_kernel100_x86_64_sse2.globl _nb_kernel100_x86_64_sse2nb_kernel100_x86_64_sse2: _nb_kernel100_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb100_fshift, 16.set nb100_gid, 24.set nb100_pos, 32.set nb100_faction, 40.set nb100_charge, 48.set nb100_p_facel, 56.set nb100_argkrf, 64.set nb100_argcrf, 72.set nb100_Vc, 80.set nb100_type, 88.set nb100_p_ntype, 96.set nb100_vdwparam, 104.set nb100_Vvdw, 112.set nb100_p_tabscale, 120.set nb100_VFtab, 128.set nb100_invsqrta, 136.set nb100_dvda, 144.set nb100_p_gbtabscale, 152.set nb100_GBtab, 160.set nb100_p_nthreads, 168.set nb100_count, 176.set nb100_mtx, 184.set nb100_outeriter, 192.set nb100_inneriter, 200.set nb100_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb100_ix, 0.set nb100_iy, 16.set nb100_iz, 32.set nb100_iq, 48.set nb100_dx, 64.set nb100_dy, 80.set nb100_dz, 96.set nb100_vctot, 112.set nb100_fix, 128.set nb100_fiy, 144.set nb100_fiz, 160.set nb100_half, 176.set nb100_three, 192.set nb100_is3, 208.set nb100_ii3, 212.set nb100_nri, 216.set nb100_iinr, 224.set nb100_jindex, 232.set nb100_jjnr, 240.set nb100_shift, 248.set nb100_shiftvec, 256.set nb100_facel, 264.set nb100_innerjjnr, 272.set nb100_innerk, 280.set nb100_n, 284.set nb100_nn1, 288.set nb100_nouter, 292.set nb100_ninner, 296 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $312,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb100_nouter(%rsp) movl %eax,nb100_ninner(%rsp) movl (%rdi),%edi movl %edi,nb100_nri(%rsp) movq %rsi,nb100_iinr(%rsp) movq %rdx,nb100_jindex(%rsp) movq %rcx,nb100_jjnr(%rsp) movq %r8,nb100_shift(%rsp) movq %r9,nb100_shiftvec(%rsp) movq nb100_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb100_facel(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb100_half(%rsp) movsd nb100_half(%rsp),%xmm1 shufpd $0,%xmm1,%xmm1 ## splat to all elements movapd %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movapd %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movapd %xmm1,nb100_half(%rsp) movapd %xmm3,nb100_three(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb100_half(%rsp) movl %ebx,nb100_half+4(%rsp) movsd nb100_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,nb100_half(%rsp) movapd %xmm3,nb100_three(%rsp)_nb_kernel100_x86_64_sse2.nb100_threadloop: movq nb100_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel100_x86_64_sse2.nb100_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_kernel100_x86_64_sse2.nb100_spinlock ## if(nn1>nri) nn1=nri movl nb100_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,nb100_n(%rsp) movl %ebx,nb100_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel100_x86_64_sse2.nb100_outerstart jmp _nb_kernel100_x86_64_sse2.nb100_end_nb_kernel100_x86_64_sse2.nb100_outerstart: ## ebx contains number of outer iterations addl nb100_nouter(%rsp),%ebx movl %ebx,nb100_nouter(%rsp)_nb_kernel100_x86_64_sse2.nb100_outer: movq nb100_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,nb100_is3(%rsp) ## store is3 movq nb100_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 nb100_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb100_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 mulsd nb100_facel(%rsp),%xmm3 shufpd $0,%xmm3,%xmm3 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb100_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,nb100_iq(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb100_ix(%rsp) movapd %xmm1,nb100_iy(%rsp) movapd %xmm2,nb100_iz(%rsp) movl %ebx,nb100_ii3(%rsp) ## clear vctot (xmm12) and i forces (xmm13-xmm15) xorpd %xmm12,%xmm12 movapd %xmm12,%xmm13 movapd %xmm12,%xmm14 movapd %xmm12,%xmm15 movq nb100_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 nb100_pos(%rbp),%rsi movq nb100_faction(%rbp),%rdi movq nb100_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb100_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb100_ninner(%rsp),%ecx movl %ecx,nb100_ninner(%rsp) addl $0,%edx movl %edx,nb100_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel100_x86_64_sse2.nb100_unroll_loop jmp _nb_kernel100_x86_64_sse2.nb100_checksingle_nb_kernel100_x86_64_sse2.nb100_unroll_loop: ## twice unrolled innerloop here movq nb100_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r8d movl 4(%rdx),%r9d addq $8,nb100_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb100_pos(%rbp),%rsi ## base of pos[] lea (%r8,%r8,2),%rax ## replace jnr with j3 lea (%r9,%r9,2),%rbx ## move two coordinates to xmm4-xmm6 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 movq nb100_faction(%rbp),%rdi ## calc dr subpd nb100_ix(%rsp),%xmm4 subpd nb100_iy(%rsp),%xmm5 subpd nb100_iz(%rsp),%xmm6 ## store dr movapd %xmm4,%xmm9 movapd %xmm5,%xmm10 movapd %xmm6,%xmm11 movq nb100_charge(%rbp),%rsi ## base of charge[] ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 ## rsq in xmm4 cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm5,%xmm2 ## lu in low xmm2 movlpd (%rsi,%r8,8),%xmm3 ## jq A ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulpd %xmm2,%xmm2 ## lu*lu movapd nb100_three(%rsp),%xmm1 mulpd %xmm4,%xmm2 ## rsq*lu*lu movapd nb100_half(%rsp),%xmm0 subpd %xmm2,%xmm1 ## 30-rsq*lu*lu mulpd %xmm5,%xmm1 mulpd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movhpd (%rsi,%r9,8),%xmm3 ## jq B movapd %xmm1,%xmm5 ## copy of lu mulpd %xmm1,%xmm1 ## lu*lu movapd nb100_three(%rsp),%xmm2 mulpd %xmm4,%xmm1 ## rsq*lu*lu movapd nb100_half(%rsp),%xmm0 mulpd nb100_iq(%rsp),%xmm3 ## qq subpd %xmm1,%xmm2 ## 30-rsq*lu*lu mulpd %xmm5,%xmm2 mulpd %xmm2,%xmm0 ## xmm0=iter2 of rinv (new lu) movapd %xmm0,%xmm4 mulpd %xmm4,%xmm4 ## xmm4=rinvsq mulpd %xmm0,%xmm3 ## xmm3=vcoul mulpd %xmm3,%xmm4 ## xmm4=fscal ## the fj's - start by combining forces from memory movlpd (%rdi,%rax,8),%xmm0 movlpd 8(%rdi,%rax,8),%xmm1 movlpd 16(%rdi,%rax,8),%xmm2 ## increment vctot addpd %xmm3,%xmm12 mulpd %xmm4,%xmm9 mulpd %xmm4,%xmm10 mulpd %xmm4,%xmm11 movhpd (%rdi,%rbx,8),%xmm0 movhpd 8(%rdi,%rbx,8),%xmm1 movhpd 16(%rdi,%rbx,8),%xmm2 addpd %xmm9,%xmm0 addpd %xmm10,%xmm1 addpd %xmm11,%xmm2 ## xmm9-xmm11 contains tx-tz (partial force) ## now update f_i addpd %xmm9,%xmm13 addpd %xmm10,%xmm14 addpd %xmm11,%xmm15 movlpd %xmm0,(%rdi,%rax,8) movlpd %xmm1,8(%rdi,%rax,8) movlpd %xmm2,16(%rdi,%rax,8) movhpd %xmm0,(%rdi,%rbx,8) movhpd %xmm1,8(%rdi,%rbx,8) movhpd %xmm2,16(%rdi,%rbx,8) ## should we do one more iteration? subl $2,nb100_innerk(%rsp) jl _nb_kernel100_x86_64_sse2.nb100_checksingle jmp _nb_kernel100_x86_64_sse2.nb100_unroll_loop_nb_kernel100_x86_64_sse2.nb100_checksingle: movl nb100_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel100_x86_64_sse2.nb100_dosingle jmp _nb_kernel100_x86_64_sse2.nb100_updateouterdata_nb_kernel100_x86_64_sse2.nb100_dosingle: movq nb100_charge(%rbp),%rsi movq nb100_pos(%rbp),%rdi movq nb100_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movq nb100_charge(%rbp),%rsi ## base of charge[] movsd (%rsi,%rax,8),%xmm3 ## jq A mulsd nb100_iq(%rsp),%xmm3 ## qq movq nb100_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move two coordinates to xmm4-xmm6 movsd (%rsi,%rax,8),%xmm4 movsd 8(%rsi,%rax,8),%xmm5 movsd 16(%rsi,%rax,8),%xmm6 movq nb100_faction(%rbp),%rdi ## calc dr subsd nb100_ix(%rsp),%xmm4 subsd nb100_iy(%rsp),%xmm5 subsd nb100_iz(%rsp),%xmm6 ## store dr movapd %xmm4,%xmm9 movapd %xmm5,%xmm10 movapd %xmm6,%xmm11 ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 ## rsq in xmm4 cvtsd2ss %xmm4,%xmm5 rsqrtss %xmm5,%xmm5 cvtss2sd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulsd %xmm2,%xmm2 ## lu*lu movapd nb100_three(%rsp),%xmm1 mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb100_half(%rsp),%xmm0 subsd %xmm2,%xmm1 ## 30-rsq*lu*lu mulsd %xmm5,%xmm1 mulsd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movapd %xmm1,%xmm5 ## copy of lu mulsd %xmm1,%xmm1 ## lu*lu movapd nb100_three(%rsp),%xmm2 mulsd %xmm4,%xmm1 ## rsq*lu*lu movapd nb100_half(%rsp),%xmm0 subsd %xmm1,%xmm2 ## 30-rsq*lu*lu mulsd %xmm5,%xmm2 mulsd %xmm2,%xmm0 ## xmm0=iter2 of rinv (new lu) movapd %xmm0,%xmm4 mulsd %xmm4,%xmm4 ## xmm4=rinvsq mulsd %xmm0,%xmm3 ## xmm3=vcoul mulsd %xmm3,%xmm4 ## xmm4=fscal ## increment vctot addsd %xmm3,%xmm12 mulsd %xmm4,%xmm9 mulsd %xmm4,%xmm10 mulsd %xmm4,%xmm11 ## xmm9-xmm11 contains tx-tz (partial force) ## now update f_i addsd %xmm9,%xmm13 addsd %xmm10,%xmm14 addsd %xmm11,%xmm15 ## the fj's - start by accumulating forces from memory addsd (%rdi,%rax,8),%xmm9 addsd 8(%rdi,%rax,8),%xmm10 addsd 16(%rdi,%rax,8),%xmm11 movsd %xmm9,(%rdi,%rax,8) movsd %xmm10,8(%rdi,%rax,8) movsd %xmm11,16(%rdi,%rax,8)_nb_kernel100_x86_64_sse2.nb100_updateouterdata: movl nb100_ii3(%rsp),%ecx movq nb100_faction(%rbp),%rdi movq nb100_fshift(%rbp),%rsi movl nb100_is3(%rsp),%edx ## accumulate i forces in xmm13, xmm14, xmm15 movhlps %xmm13,%xmm3 movhlps %xmm14,%xmm4 movhlps %xmm15,%xmm5 addsd %xmm3,%xmm13 addsd %xmm4,%xmm14 addsd %xmm5,%xmm15 ## sum is in low xmm13-xmm15 ## increment i force movsd (%rdi,%rcx,8),%xmm3 movsd 8(%rdi,%rcx,8),%xmm4 movsd 16(%rdi,%rcx,8),%xmm5 subsd %xmm13,%xmm3 subsd %xmm14,%xmm4 subsd %xmm15,%xmm5 movsd %xmm3,(%rdi,%rcx,8) movsd %xmm4,8(%rdi,%rcx,8) movsd %xmm5,16(%rdi,%rcx,8) ## increment fshift force movsd (%rsi,%rdx,8),%xmm3 movsd 8(%rsi,%rdx,8),%xmm4 movsd 16(%rsi,%rdx,8),%xmm5⌨️ 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?