nb_kernel410_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,987 行 · 第 1/4 页
S
1,987 行
#### $Id: nb_kernel410_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:51 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_kernel410_x86_64_sse.globl _nb_kernel410_x86_64_ssenb_kernel410_x86_64_sse: _nb_kernel410_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb410_fshift, 16.set nb410_gid, 24.set nb410_pos, 32.set nb410_faction, 40.set nb410_charge, 48.set nb410_p_facel, 56.set nb410_argkrf, 64.set nb410_argcrf, 72.set nb410_Vc, 80.set nb410_type, 88.set nb410_p_ntype, 96.set nb410_vdwparam, 104.set nb410_Vvdw, 112.set nb410_p_tabscale, 120.set nb410_VFtab, 128.set nb410_invsqrta, 136.set nb410_dvda, 144.set nb410_p_gbtabscale, 152.set nb410_GBtab, 160.set nb410_p_nthreads, 168.set nb410_count, 176.set nb410_mtx, 184.set nb410_outeriter, 192.set nb410_inneriter, 200.set nb410_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb410_ix, 0.set nb410_iy, 16.set nb410_iz, 32.set nb410_iq, 48.set nb410_dx, 64.set nb410_dy, 80.set nb410_dz, 96.set nb410_two, 112.set nb410_six, 128.set nb410_twelve, 144.set nb410_gbtsc, 160.set nb410_qq, 176.set nb410_c6, 192.set nb410_c12, 208.set nb410_fscal, 224.set nb410_vctot, 240.set nb410_Vvdwtot, 256.set nb410_fix, 272.set nb410_fiy, 288.set nb410_fiz, 304.set nb410_half, 320.set nb410_three, 336.set nb410_r, 352.set nb410_isai, 368.set nb410_isaprod, 384.set nb410_dvdasum, 400.set nb410_gbscale, 416.set nb410_nri, 432.set nb410_iinr, 440.set nb410_jindex, 448.set nb410_jjnr, 456.set nb410_shift, 464.set nb410_shiftvec, 472.set nb410_facel, 480.set nb410_innerjjnr, 488.set nb410_is3, 496.set nb410_ii3, 500.set nb410_ii, 504.set nb410_ntia, 508.set nb410_innerk, 512.set nb410_n, 516.set nb410_nn1, 520.set nb410_ntype, 524.set nb410_nouter, 528.set nb410_ninner, 532.set nb410_jnra, 536.set nb410_jnrb, 540.set nb410_jnrc, 544.set nb410_jnrd, 548 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $568,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb410_nouter(%rsp) movl %eax,nb410_ninner(%rsp) movl (%rdi),%edi movl %edi,nb410_nri(%rsp) movq %rsi,nb410_iinr(%rsp) movq %rdx,nb410_jindex(%rsp) movq %rcx,nb410_jjnr(%rsp) movq %r8,nb410_shift(%rsp) movq %r9,nb410_shiftvec(%rsp) movq nb410_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb410_ntype(%rsp) movq nb410_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb410_facel(%rsp) movq nb410_p_gbtabscale(%rbp),%rbx movss (%rbx),%xmm4 shufps $0,%xmm4,%xmm4 movaps %xmm4,nb410_gbtsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb410_half(%rsp) movss nb410_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,nb410_half(%rsp) movaps %xmm2,nb410_two(%rsp) movaps %xmm3,nb410_three(%rsp) movaps %xmm4,nb410_six(%rsp) movaps %xmm5,nb410_twelve(%rsp)_nb_kernel410_x86_64_sse.nb410_threadloop: movq nb410_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel410_x86_64_sse.nb410_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_kernel410_x86_64_sse.nb410_spinlock ## if(nn1>nri) nn1=nri movl nb410_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,nb410_n(%rsp) movl %ebx,nb410_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel410_x86_64_sse.nb410_outerstart jmp _nb_kernel410_x86_64_sse.nb410_end_nb_kernel410_x86_64_sse.nb410_outerstart: ## ebx contains number of outer iterations addl nb410_nouter(%rsp),%ebx movl %ebx,nb410_nouter(%rsp)_nb_kernel410_x86_64_sse.nb410_outer: movq nb410_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,nb410_is3(%rsp) ## store is3 movq nb410_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 nb410_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movl %ebx,nb410_ii(%rsp) movq nb410_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb410_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 movq nb410_invsqrta(%rbp),%rdx ## load invsqrta[ii] movss (%rdx,%rbx,4),%xmm4 shufps $0,%xmm4,%xmm4 movq nb410_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb410_ntype(%rsp),%edx shll %edx movl %edx,nb410_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb410_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,nb410_iq(%rsp) movaps %xmm4,nb410_isai(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb410_ix(%rsp) movaps %xmm1,nb410_iy(%rsp) movaps %xmm2,nb410_iz(%rsp) movl %ebx,nb410_ii3(%rsp) ## clear vctot and i forces xorps %xmm13,%xmm13 movaps %xmm13,%xmm12 movaps %xmm13,nb410_Vvdwtot(%rsp) movaps %xmm13,nb410_dvdasum(%rsp) movaps %xmm13,%xmm14 movaps %xmm13,%xmm15 movq nb410_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 nb410_pos(%rbp),%rsi movq nb410_faction(%rbp),%rdi movq nb410_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb410_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb410_ninner(%rsp),%ecx movl %ecx,nb410_ninner(%rsp) addl $0,%edx movl %edx,nb410_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel410_x86_64_sse.nb410_unroll_loop jmp _nb_kernel410_x86_64_sse.nb410_finish_inner_nb_kernel410_x86_64_sse.nb410_unroll_loop: ## quad-unroll innerloop here movq nb410_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb410_innerjjnr(%rsp) ## advance pointer (unrolled 4) ## load isaj movq nb410_invsqrta(%rbp),%rsi movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 movaps nb410_isai(%rsp),%xmm2 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all isaj in xmm3 mulps %xmm3,%xmm2 movaps %xmm2,nb410_isaprod(%rsp) movaps %xmm2,%xmm1 mulps nb410_gbtsc(%rsp),%xmm1 movaps %xmm1,nb410_gbscale(%rsp) movq nb410_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 mulps nb410_iq(%rsp),%xmm2 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 mulps %xmm2,%xmm3 movaps %xmm3,nb410_qq(%rsp) ## vdw parameters movq nb410_type(%rbp),%rsi movl (%rsi,%rax,4),%r12d movl (%rsi,%rbx,4),%r13d movl (%rsi,%rcx,4),%r14d movl (%rsi,%rdx,4),%r15d shll %r12d shll %r13d shll %r14d shll %r15d movl nb410_ntia(%rsp),%edi addl %edi,%r12d addl %edi,%r13d addl %edi,%r14d addl %edi,%r15d movq nb410_vdwparam(%rbp),%rsi movlps (%rsi,%r12,4),%xmm3 movlps (%rsi,%r14,4),%xmm7 movhps (%rsi,%r13,4),%xmm3 movhps (%rsi,%r15,4),%xmm7 movaps %xmm3,%xmm0 shufps $136,%xmm7,%xmm0 ## 10001000 shufps $221,%xmm7,%xmm3 ## 11011101 movaps %xmm0,nb410_c6(%rsp) movaps %xmm3,nb410_c12(%rsp) movq nb410_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%r8 ## jnr lea (%rbx,%rbx,2),%r9 lea (%rcx,%rcx,2),%r10 lea (%rdx,%rdx,2),%r11 ## move four coordinates to xmm0-xmm2 movlps (%rsi,%r8,4),%xmm4 movlps (%rsi,%r10,4),%xmm5 movss 8(%rsi,%r8,4),%xmm2 movss 8(%rsi,%r10,4),%xmm6 movhps (%rsi,%r9,4),%xmm4 movhps (%rsi,%r11,4),%xmm5 movss 8(%rsi,%r9,4),%xmm0 movss 8(%rsi,%r11,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ## calc dr subps nb410_ix(%rsp),%xmm0 subps nb410_iy(%rsp),%xmm1 subps nb410_iz(%rsp),%xmm2 ## store dr movaps %xmm0,nb410_dx(%rsp) movaps %xmm1,nb410_dy(%rsp) movaps %xmm2,nb410_dz(%rsp) ## square it mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb410_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb410_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm4,nb410_r(%rsp) mulps nb410_gbscale(%rsp),%xmm4 ## truncate and convert to integers cvttps2dq %xmm4,%xmm5 ## convert back to float cvtdq2ps %xmm5,%xmm6 ## multiply by 4 pslld $2,%xmm5 ## move to integer registers movhlps %xmm5,%xmm7 movd %xmm5,%r12d movd %xmm7,%r14d pshufd $1,%xmm5,%xmm5 pshufd $1,%xmm7,%xmm7 movd %xmm5,%r13d movd %xmm7,%r15d ## calculate eps subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ##eps movq nb410_GBtab(%rbp),%rsi movaps %xmm0,%xmm9 ## rinv mulps %xmm9,%xmm9 ## rinvsq movaps %xmm9,%xmm10 ## rinvsq mulps %xmm10,%xmm10 ## rinv4 mulps %xmm9,%xmm10 ## rinv6 movaps %xmm10,%xmm11 mulps %xmm11,%xmm11 ## rinv12 ## load table data movlps (%rsi,%r12,4),%xmm5 movlps (%rsi,%r14,4),%xmm7 movhps (%rsi,%r13,4),%xmm5 movhps (%rsi,%r15,4),%xmm7 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 mulps nb410_c6(%rsp),%xmm10 ## vvdw6=c6*rinv6 mulps nb410_c12(%rsp),%xmm11 ## vvdw12=c12*rinv12 movaps %xmm11,%xmm9 subps %xmm10,%xmm11 ## Vvdw=Vvdw12-Vvdw6 ## add potential to vvdwtot addps nb410_Vvdwtot(%rsp),%xmm11 movaps %xmm11,nb410_Vvdwtot(%rsp) movlps 8(%rsi,%r12,4),%xmm7 movlps 8(%rsi,%r14,4),%xmm8 movhps 8(%rsi,%r13,4),%xmm7 movhps 8(%rsi,%r15,4),%xmm8 movaps %xmm7,%xmm6 shufps $136,%xmm8,%xmm6 ## 10001000 shufps $221,%xmm8,%xmm7 ## 11011101 ## table data ready in xmm4-xmm7 mulps %xmm1,%xmm7 ## Heps mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm1,%xmm7 ## Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp addps %xmm7,%xmm7 ## two*Heps2 movaps nb410_qq(%rsp),%xmm3 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm3 ## fijC=FF*qq ## at this point xmm5 contains vcoul and xmm3 fijC ## LJ forces mulps nb410_six(%rsp),%xmm10 mulps nb410_twelve(%rsp),%xmm9 subps %xmm10,%xmm9 mulps %xmm0,%xmm9 ## (12*vnb12-6*vnb6)*rinv movq nb410_dvda(%rbp),%rsi ## Calculate dVda xorps %xmm7,%xmm7 mulps nb410_gbscale(%rsp),%xmm3⌨️ 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?