nb_kernel030_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,724 行 · 第 1/4 页
S
1,724 行
#### $Id: nb_kernel030_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_kernel030_x86_64_sse.globl _nb_kernel030_x86_64_ssenb_kernel030_x86_64_sse: _nb_kernel030_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb030_fshift, 16.set nb030_gid, 24.set nb030_pos, 32.set nb030_faction, 40.set nb030_charge, 48.set nb030_p_facel, 56.set nb030_argkrf, 64.set nb030_argcrf, 72.set nb030_Vc, 80.set nb030_type, 88.set nb030_p_ntype, 96.set nb030_vdwparam, 104.set nb030_Vvdw, 112.set nb030_p_tabscale, 120.set nb030_VFtab, 128.set nb030_invsqrta, 136.set nb030_dvda, 144.set nb030_p_gbtabscale, 152.set nb030_GBtab, 160.set nb030_p_nthreads, 168.set nb030_count, 176.set nb030_mtx, 184.set nb030_outeriter, 192.set nb030_inneriter, 200.set nb030_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb030_ix, 0.set nb030_iy, 16.set nb030_iz, 32.set nb030_dx, 48.set nb030_dy, 64.set nb030_dz, 80.set nb030_two, 96.set nb030_tsc, 112.set nb030_c6, 128.set nb030_c12, 144.set nb030_fscal, 160.set nb030_Vvdwtot, 176.set nb030_fix, 192.set nb030_fiy, 208.set nb030_fiz, 224.set nb030_half, 240.set nb030_three, 256.set nb030_is3, 272.set nb030_ii3, 276.set nb030_nri, 280.set nb030_iinr, 288.set nb030_jindex, 296.set nb030_jjnr, 304.set nb030_shift, 312.set nb030_shiftvec, 320.set nb030_innerjjnr, 328.set nb030_ntia, 336.set nb030_innerk, 340.set nb030_n, 344.set nb030_nn1, 348.set nb030_ntype, 352.set nb030_nouter, 356.set nb030_ninner, 360 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 subq $376,%rsp # # local variable stack space (n*16+8) emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb030_nouter(%rsp) movl %eax,nb030_ninner(%rsp) movl (%rdi),%edi movl %edi,nb030_nri(%rsp) movq %rsi,nb030_iinr(%rsp) movq %rdx,nb030_jindex(%rsp) movq %rcx,nb030_jjnr(%rsp) movq %r8,nb030_shift(%rsp) movq %r9,nb030_shiftvec(%rsp) movq nb030_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb030_ntype(%rsp) movq nb030_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb030_tsc(%rsp) movq nb030_pos(%rbp),%r8 movq nb030_faction(%rbp),%r9 movq nb030_type(%rbp),%r10 movq nb030_vdwparam(%rbp),%r11 ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb030_half(%rsp) movss nb030_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 %xmm1,nb030_half(%rsp) movaps %xmm2,nb030_two(%rsp) movaps %xmm3,nb030_three(%rsp)_nb_kernel030_x86_64_sse.nb030_threadloop: movq nb030_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel030_x86_64_sse.nb030_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_kernel030_x86_64_sse.nb030_spinlock ## if(nn1>nri) nn1=nri movl nb030_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,nb030_n(%rsp) movl %ebx,nb030_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel030_x86_64_sse.nb030_outerstart jmp _nb_kernel030_x86_64_sse.nb030_end_nb_kernel030_x86_64_sse.nb030_outerstart: ## ebx contains number of outer iterations addl nb030_nouter(%rsp),%ebx movl %ebx,nb030_nouter(%rsp)_nb_kernel030_x86_64_sse.nb030_outer: movq nb030_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,nb030_is3(%rsp) ## store is3 movq nb030_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 nb030_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb030_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb030_ntype(%rsp),%edx shll %edx movl %edx,nb030_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb030_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb030_ix(%rsp) movaps %xmm1,nb030_iy(%rsp) movaps %xmm2,nb030_iz(%rsp) movl %ebx,nb030_ii3(%rsp) ## clear tot potential and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb030_Vvdwtot(%rsp) movaps %xmm4,nb030_fix(%rsp) movaps %xmm4,nb030_fiy(%rsp) movaps %xmm4,nb030_fiz(%rsp) movq nb030_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 nb030_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb030_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb030_ninner(%rsp),%ecx movl %ecx,nb030_ninner(%rsp) addl $0,%edx movl %edx,nb030_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel030_x86_64_sse.nb030_unroll_loop jmp _nb_kernel030_x86_64_sse.nb030_finish_inner_nb_kernel030_x86_64_sse.nb030_unroll_loop: ## quad-unroll innerloop here movq nb030_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,nb030_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%rax,%rax,2),%r12 ## j3 lea (%rbx,%rbx,2),%r13 lea (%rcx,%rcx,2),%r14 lea (%rdx,%rdx,2),%r15 movq nb030_pos(%rbp),%rdi ## load coordinates movlps (%rdi,%r12,4),%xmm1 ## x1 y1 - - movlps (%rdi,%r14,4),%xmm2 ## x3 y3 - - movhps (%rdi,%r13,4),%xmm1 ## x2 y2 - - movhps (%rdi,%r15,4),%xmm2 ## x4 y4 - - movss 8(%rdi,%r12,4),%xmm5 ## z1 - - - movss 8(%rdi,%r14,4),%xmm6 ## z2 - - - movss 8(%rdi,%r13,4),%xmm7 ## z3 - - - movss 8(%rdi,%r15,4),%xmm8 ## z4 - - - movlhps %xmm7,%xmm5 ## jzOa - jzOb - movlhps %xmm8,%xmm6 ## jzOc - jzOd - movaps %xmm1,%xmm4 unpcklps %xmm2,%xmm1 ## jxa jxc jya jyc unpckhps %xmm2,%xmm4 ## jxb jxd jyb jyd movaps %xmm1,%xmm2 unpcklps %xmm4,%xmm1 ## x unpckhps %xmm4,%xmm2 ## y shufps $136,%xmm6,%xmm5 ## 10001000 => jzH2a jzH2b jzH2c jzH2d movq nb030_type(%rbp),%rsi ## calc dr subps nb030_ix(%rsp),%xmm1 subps nb030_iy(%rsp),%xmm2 subps nb030_iz(%rsp),%xmm5 ## store dr movaps %xmm1,nb030_dx(%rsp) movaps %xmm2,nb030_dy(%rsp) movaps %xmm5,nb030_dz(%rsp) ## square it mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm5,%xmm5 addps %xmm2,%xmm1 addps %xmm5,%xmm1 ## rsq in xmm1 ## load vdw types movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm1,%xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb030_three(%rsp),%xmm4 mulps %xmm1,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm4 ## 30-rsq*lu*lu mulps %xmm2,%xmm4 mulps nb030_half(%rsp),%xmm4 movaps %xmm4,%xmm2 mulps %xmm4,%xmm1 ## xmm2=rinv ## xmm1=r movl nb030_ntia(%rsp),%edi ## type*=2 shll %eax shll %ebx shll %ecx shll %edx mulps nb030_tsc(%rsp),%xmm1 ## rtab ## truncate and convert to integers cvttps2dq %xmm1,%xmm5 addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx ## convert back to float cvtdq2ps %xmm5,%xmm4 ## multiply by 8 pslld $3,%xmm5 ## calculate eps subps %xmm4,%xmm1 ## move to integer registers movhlps %xmm5,%xmm6 movd %xmm5,%r8d movd %xmm6,%r10d pshufd $1,%xmm5,%xmm5 pshufd $1,%xmm6,%xmm6 movd %xmm5,%r9d movd %xmm6,%r11d ## xmm1=eps ## xmm2=rinv movq nb030_VFtab(%rbp),%rsi ## calculate LJ table movlps (%rsi,%r8,4),%xmm5 movlps 16(%rsi,%r8,4),%xmm9 movlps (%rsi,%r10,4),%xmm7 movlps 16(%rsi,%r10,4),%xmm11 movhps (%rsi,%r9,4),%xmm5 movhps 16(%rsi,%r9,4),%xmm9 movhps (%rsi,%r11,4),%xmm7 movhps 16(%rsi,%r11,4),%xmm11 movaps %xmm5,%xmm4 movaps %xmm9,%xmm8 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $136,%xmm11,%xmm8 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 shufps $221,%xmm11,%xmm9 ## 11011101 movlps 8(%rsi,%r8,4),%xmm7 movlps 24(%rsi,%r8,4),%xmm11 movlps 8(%rsi,%r10,4),%xmm13 movlps 24(%rsi,%r10,4),%xmm14 movhps 8(%rsi,%r9,4),%xmm7 movhps 24(%rsi,%r9,4),%xmm11 movhps 8(%rsi,%r11,4),%xmm13 movhps 24(%rsi,%r11,4),%xmm14 movaps %xmm7,%xmm6 movaps %xmm11,%xmm10 shufps $136,%xmm13,%xmm6 ## 10001000 shufps $136,%xmm14,%xmm10 ## 10001000 shufps $221,%xmm13,%xmm7 ## 11011101 shufps $221,%xmm14,%xmm11 ## 11011101 ## dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11 movq nb030_vdwparam(%rbp),%rsi mulps %xmm1,%xmm7 ## Heps mulps %xmm1,%xmm11 mulps %xmm1,%xmm6 ## Geps mulps %xmm1,%xmm10 mulps %xmm1,%xmm7 ## Heps2 mulps %xmm1,%xmm11 ## load c6/c12 movlps (%rsi,%rax,4),%xmm13 movlps (%rsi,%rcx,4),%xmm14 addps %xmm6,%xmm5 ## F+Geps addps %xmm10,%xmm9 addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm11,%xmm9 movhps (%rsi,%rbx,4),%xmm13 movhps (%rsi,%rdx,4),%xmm14 addps %xmm7,%xmm7 ## 2*Heps2 addps %xmm11,%xmm11 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm10,%xmm11 ## shuffle c6/c12 movaps %xmm13,%xmm12 shufps $136,%xmm14,%xmm12 ## 10001000 => c6 shufps $221,%xmm14,%xmm13 ## 11011101 => c12 addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps addps %xmm9,%xmm11 mulps %xmm1,%xmm5 ## eps*Fp mulps %xmm1,%xmm9 addps %xmm4,%xmm5 ## VV addps %xmm8,%xmm9⌨️ 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?