nb_kernel110_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,440 行 · 第 1/3 页
S
1,440 行
#### $Id: nb_kernel110_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:48 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_kernel110_x86_64_sse.globl _nb_kernel110_x86_64_ssenb_kernel110_x86_64_sse: _nb_kernel110_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb110_fshift, 16.set nb110_gid, 24.set nb110_pos, 32.set nb110_faction, 40.set nb110_charge, 48.set nb110_p_facel, 56.set nb110_argkrf, 64.set nb110_argcrf, 72.set nb110_Vc, 80.set nb110_type, 88.set nb110_p_ntype, 96.set nb110_vdwparam, 104.set nb110_Vvdw, 112.set nb110_p_tabscale, 120.set nb110_VFtab, 128.set nb110_invsqrta, 136.set nb110_dvda, 144.set nb110_p_gbtabscale, 152.set nb110_GBtab, 160.set nb110_p_nthreads, 168.set nb110_count, 176.set nb110_mtx, 184.set nb110_outeriter, 192.set nb110_inneriter, 200.set nb110_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb110_ix, 0.set nb110_iy, 16.set nb110_iz, 32.set nb110_iq, 48.set nb110_qq, 64.set nb110_c6, 112.set nb110_c12, 128.set nb110_six, 144.set nb110_twelve, 160.set nb110_vctot, 176.set nb110_Vvdwtot, 192.set nb110_fix, 208.set nb110_fiy, 224.set nb110_fiz, 240.set nb110_half, 256.set nb110_three, 272.set nb110_nri, 288.set nb110_iinr, 296.set nb110_jindex, 304.set nb110_jjnr, 312.set nb110_shift, 320.set nb110_shiftvec, 328.set nb110_facel, 336.set nb110_innerjjnr, 344.set nb110_is3, 352.set nb110_ii3, 356.set nb110_ntia, 360.set nb110_innerk, 364.set nb110_n, 368.set nb110_nn1, 372.set nb110_ntype, 376.set nb110_nouter, 380.set nb110_ninner, 384 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 emms subq $408,%rsp ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb110_nouter(%rsp) movl %eax,nb110_ninner(%rsp) movl (%rdi),%edi movl %edi,nb110_nri(%rsp) movq %rsi,nb110_iinr(%rsp) movq %rdx,nb110_jindex(%rsp) movq %rcx,nb110_jjnr(%rsp) movq %r8,nb110_shift(%rsp) movq %r9,nb110_shiftvec(%rsp) movq nb110_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb110_ntype(%rsp) movq nb110_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb110_facel(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb110_half(%rsp) movss nb110_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm3,%xmm4 addps %xmm4,%xmm4 ## six movaps %xmm4,%xmm5 addps %xmm5,%xmm5 ## twelve movaps %xmm1,nb110_half(%rsp) movaps %xmm3,nb110_three(%rsp) movaps %xmm4,nb110_six(%rsp) movaps %xmm5,nb110_twelve(%rsp)_nb_kernel110_x86_64_sse.nb110_threadloop: movq nb110_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel110_x86_64_sse.nb110_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_kernel110_x86_64_sse.nb110_spinlock ## if(nn1>nri) nn1=nri movl nb110_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,nb110_n(%rsp) movl %ebx,nb110_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel110_x86_64_sse.nb110_outerstart jmp _nb_kernel110_x86_64_sse.nb110_end_nb_kernel110_x86_64_sse.nb110_outerstart: ## ebx contains number of outer iterations addl nb110_nouter(%rsp),%ebx movl %ebx,nb110_nouter(%rsp)_nb_kernel110_x86_64_sse.nb110_outer: movq nb110_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb110_is3(%rsp) ## store is3 movq nb110_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb110_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb110_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb110_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 movq nb110_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb110_ntype(%rsp),%edx shll %edx movl %edx,nb110_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb110_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 movaps %xmm3,nb110_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb110_ix(%rsp) movaps %xmm1,nb110_iy(%rsp) movaps %xmm2,nb110_iz(%rsp) movl %ebx,nb110_ii3(%rsp) ## clear vctot and i forces xorps %xmm12,%xmm12 movaps %xmm12,nb110_vctot(%rsp) movaps %xmm12,%xmm13 movaps %xmm12,%xmm14 movaps %xmm12,%xmm15 movq nb110_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 nb110_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb110_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb110_ninner(%rsp),%ecx movl %ecx,nb110_ninner(%rsp) addl $0,%edx movl %edx,nb110_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel110_x86_64_sse.nb110_unroll_loop jmp _nb_kernel110_x86_64_sse.nb110_finish_inner_nb_kernel110_x86_64_sse.nb110_unroll_loop: ## quad-unrolled innerloop here movq nb110_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r8d movl 4(%rdx),%r9d movl 8(%rdx),%r10d movl 12(%rdx),%r11d ## eax-edx=jnr1-4 addq $16,nb110_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%r8,%r8,2),%rax ## replace jnr with j3 lea (%r9,%r9,2),%rbx lea (%r10,%r10,2),%rcx lea (%r11,%r11,2),%rdx movq nb110_pos(%rbp),%rdi ## load coordinates movlps (%rdi,%rax,4),%xmm1 ## x1 y1 - - movlps (%rdi,%rcx,4),%xmm2 ## x3 y3 - - movhps (%rdi,%rbx,4),%xmm1 ## x2 y2 - - movhps (%rdi,%rdx,4),%xmm2 ## x4 y4 - - movss 8(%rdi,%rax,4),%xmm5 ## z1 - - - movss 8(%rdi,%rcx,4),%xmm6 ## z2 - - - movss 8(%rdi,%rbx,4),%xmm7 ## z3 - - - movss 8(%rdi,%rdx,4),%xmm8 ## z4 - - - movlhps %xmm7,%xmm5 ## jzOa - jzOb - movlhps %xmm8,%xmm6 ## jzOc - jzOd - movq nb110_charge(%rbp),%rsi movq nb110_type(%rbp),%rdi movaps %xmm1,%xmm4 unpcklps %xmm2,%xmm1 ## jxa jxc jya jyc unpckhps %xmm2,%xmm4 ## jxb jxd jyb jyd movss (%rsi,%r8,4),%xmm7 movss (%rsi,%r9,4),%xmm8 movss (%rsi,%r10,4),%xmm10 movss (%rsi,%r11,4),%xmm11 movaps %xmm1,%xmm2 unpcklps %xmm4,%xmm1 ## x unpckhps %xmm4,%xmm2 ## y shufps $136,%xmm6,%xmm5 ## 10001000 => jzH2a jzH2b jzH2c jzH2d movl (%rdi,%r8,4),%r8d movl (%rdi,%r9,4),%r9d movl (%rdi,%r10,4),%r10d movl (%rdi,%r11,4),%r11d unpcklps %xmm10,%xmm7 unpcklps %xmm11,%xmm8 ## calc dr subps nb110_ix(%rsp),%xmm1 subps nb110_iy(%rsp),%xmm2 subps nb110_iz(%rsp),%xmm5 unpcklps %xmm8,%xmm7 ## jq ## store dr in xmm9-xmm11 movaps %xmm1,%xmm9 movaps %xmm2,%xmm10 movaps %xmm5,%xmm11 mulps nb110_iq(%rsp),%xmm7 shll %r8d shll %r9d shll %r10d shll %r11d movl nb110_ntia(%rsp),%edi ## square it mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm5,%xmm5 addps %xmm2,%xmm1 addps %xmm5,%xmm1 ## rsq in xmm1 movq nb110_vdwparam(%rbp),%rsi addl %edi,%r8d addl %edi,%r9d addl %edi,%r10d addl %edi,%r11d ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm1,%xmm5 movaps %xmm5,%xmm6 mulps %xmm5,%xmm5 movaps nb110_three(%rsp),%xmm4 mulps %xmm1,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm4 ## 30-rsq*lu*lu mulps %xmm6,%xmm4 mulps nb110_half(%rsp),%xmm4 movaps %xmm4,%xmm1 mulps %xmm4,%xmm4 ## xmm1=rinv ## xmm4=rinvsq movlps (%rsi,%r8,4),%xmm3 movlps (%rsi,%r10,4),%xmm2 movhps (%rsi,%r9,4),%xmm3 movhps (%rsi,%r11,4),%xmm2 movaps %xmm4,%xmm5 mulps %xmm4,%xmm5 ## rinv4 mulps %xmm4,%xmm5 ## rinv6 movaps %xmm5,%xmm6 mulps %xmm5,%xmm5 ## xmm5=rinv12 ## coulomb stuff mulps %xmm7,%xmm1 ## vcoul=rinv*qq movaps %xmm1,%xmm8 ## fijC movaps %xmm3,%xmm0 shufps $136,%xmm2,%xmm0 ## 10001000 shufps $221,%xmm2,%xmm3 ## 11011101 ## add to vctot addps nb110_vctot(%rsp),%xmm1 movaps %xmm1,nb110_vctot(%rsp) mulps %xmm0,%xmm6 ## vvdw6=c6*rinv6 mulps %xmm3,%xmm5 ## vvdw12=c12*rinv12 movaps %xmm5,%xmm7 subps %xmm6,%xmm5 ## Vvdw=Vvdw12-Vvdw6 mulps nb110_six(%rsp),%xmm6 mulps nb110_twelve(%rsp),%xmm7 subps %xmm6,%xmm7 addps %xmm7,%xmm8 mulps %xmm8,%xmm4 ## xmm4=total fscal movq nb110_faction(%rbp),%rsi ## the fj's - start by combining x & y forces from memory movlps (%rsi,%rax,4),%xmm0 ## x1 y1 - - movlps (%rsi,%rcx,4),%xmm1 ## x3 y3 - - movhps (%rsi,%rbx,4),%xmm0 ## x1 y1 x2 y2 movhps (%rsi,%rdx,4),%xmm1 ## x3 y3 x4 y4 ## add potential to Vvdwtot (sum in xmm12) addps %xmm5,%xmm12 ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm4,%xmm9 mulps %xmm4,%xmm10 mulps %xmm4,%xmm11 ## accumulate i forces addps %xmm9,%xmm13 addps %xmm10,%xmm14 addps %xmm11,%xmm15 movaps %xmm9,%xmm8 unpcklps %xmm10,%xmm9 ## x1 y1 x2 y2 unpckhps %xmm10,%xmm8 ## x3 y3 x4 y4 ## update fjx and fjy addps %xmm9,%xmm0 addps %xmm8,%xmm1 movlps %xmm0,(%rsi,%rax,4) movlps %xmm1,(%rsi,%rcx,4) movhps %xmm0,(%rsi,%rbx,4) movhps %xmm1,(%rsi,%rdx,4) ## xmm11: fjz1 fjz2 fjz3 fjz4 pshufd $1,%xmm11,%xmm10 ## fjz2 - - - movhlps %xmm11,%xmm9 ## fjz3 - - - pshufd $3,%xmm11,%xmm8 ## fjz4 - - - addss 8(%rsi,%rax,4),%xmm11 addss 8(%rsi,%rbx,4),%xmm10 addss 8(%rsi,%rcx,4),%xmm9 addss 8(%rsi,%rdx,4),%xmm8 movss %xmm11,8(%rsi,%rax,4) movss %xmm10,8(%rsi,%rbx,4) movss %xmm9,8(%rsi,%rcx,4) movss %xmm8,8(%rsi,%rdx,4) ## should we do one more iteration? subl $4,nb110_innerk(%rsp) jl _nb_kernel110_x86_64_sse.nb110_finish_inner jmp _nb_kernel110_x86_64_sse.nb110_unroll_loop_nb_kernel110_x86_64_sse.nb110_finish_inner: ## check if at least two particles remain addl $4,nb110_innerk(%rsp) movl nb110_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel110_x86_64_sse.nb110_dopair jmp _nb_kernel110_x86_64_sse.nb110_checksingle_nb_kernel110_x86_64_sse.nb110_dopair: ## twice-unrolled innerloop here movq nb110_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r8d movl 4(%rdx),%r9d addq $8,nb110_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb110_charge(%rbp),%rsi movss (%rsi,%r8,4),%xmm0 movss (%rsi,%r9,4),%xmm2 unpcklps %xmm2,%xmm0 ## jqa jqb - - mulps nb110_iq(%rsp),%xmm0 movaps %xmm0,nb110_qq(%rsp) movq nb110_type(%rbp),%rsi movl (%rsi,%r8,4),%r12d movl (%rsi,%r9,4),%r13d shll %r12d shll %r13d movl nb110_ntia(%rsp),%edi addl %edi,%r12d addl %edi,%r13d movq nb110_vdwparam(%rbp),%rsi movlps (%rsi,%r12,4),%xmm3 movhps (%rsi,%r13,4),%xmm3 xorps %xmm7,%xmm7 movaps %xmm3,%xmm0 shufps $136,%xmm7,%xmm0 ## 10001000 shufps $221,%xmm7,%xmm3 ## 11011101 ## xmm0=c6 ## xmm3=c12 lea (%r8,%r8,2),%rax ## j3 lea (%r9,%r9,2),%rbx⌨️ 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?