nb_kernel310_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,781 行 · 第 1/4 页
S
1,781 行
#### $Id: nb_kernel310_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:50 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_kernel310_x86_64_sse.globl _nb_kernel310_x86_64_ssenb_kernel310_x86_64_sse: _nb_kernel310_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb310_fshift, 16.set nb310_gid, 24.set nb310_pos, 32.set nb310_faction, 40.set nb310_charge, 48.set nb310_p_facel, 56.set nb310_argkrf, 64.set nb310_argcrf, 72.set nb310_Vc, 80.set nb310_type, 88.set nb310_p_ntype, 96.set nb310_vdwparam, 104.set nb310_Vvdw, 112.set nb310_p_tabscale, 120.set nb310_VFtab, 128.set nb310_invsqrta, 136.set nb310_dvda, 144.set nb310_p_gbtabscale, 152.set nb310_GBtab, 160.set nb310_p_nthreads, 168.set nb310_count, 176.set nb310_mtx, 184.set nb310_outeriter, 192.set nb310_inneriter, 200.set nb310_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb310_ix, 0.set nb310_iy, 16.set nb310_iz, 32.set nb310_iq, 48.set nb310_dx, 64.set nb310_dy, 80.set nb310_dz, 96.set nb310_two, 112.set nb310_six, 128.set nb310_twelve, 144.set nb310_tsc, 160.set nb310_qq, 176.set nb310_c6, 192.set nb310_c12, 208.set nb310_fscal, 224.set nb310_vctot, 240.set nb310_Vvdwtot, 256.set nb310_fix, 272.set nb310_fiy, 288.set nb310_fiz, 304.set nb310_half, 320.set nb310_three, 336.set nb310_nri, 352.set nb310_iinr, 360.set nb310_jindex, 368.set nb310_jjnr, 376.set nb310_shift, 384.set nb310_shiftvec, 392.set nb310_facel, 400.set nb310_innerjjnr, 408.set nb310_is3, 416.set nb310_ii3, 420.set nb310_ntia, 424.set nb310_innerk, 428.set nb310_n, 432.set nb310_nn1, 436.set nb310_ntype, 440.set nb310_nouter, 444.set nb310_ninner, 448 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $472,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb310_nouter(%rsp) movl %eax,nb310_ninner(%rsp) movl (%rdi),%edi movl %edi,nb310_nri(%rsp) movq %rsi,nb310_iinr(%rsp) movq %rdx,nb310_jindex(%rsp) movq %rcx,nb310_jjnr(%rsp) movq %r8,nb310_shift(%rsp) movq %r9,nb310_shiftvec(%rsp) movq nb310_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb310_ntype(%rsp) movq nb310_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb310_facel(%rsp) movq nb310_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb310_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb310_half(%rsp) movss nb310_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,nb310_half(%rsp) movaps %xmm2,nb310_two(%rsp) movaps %xmm3,nb310_three(%rsp) movaps %xmm4,nb310_six(%rsp) movaps %xmm5,nb310_twelve(%rsp)_nb_kernel310_x86_64_sse.nb310_threadloop: movq nb310_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel310_x86_64_sse.nb310_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_kernel310_x86_64_sse.nb310_spinlock ## if(nn1>nri) nn1=nri movl nb310_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,nb310_n(%rsp) movl %ebx,nb310_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel310_x86_64_sse.nb310_outerstart jmp _nb_kernel310_x86_64_sse.nb310_end_nb_kernel310_x86_64_sse.nb310_outerstart: ## ebx contains number of outer iterations addl nb310_nouter(%rsp),%ebx movl %ebx,nb310_nouter(%rsp)_nb_kernel310_x86_64_sse.nb310_outer: movq nb310_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,nb310_is3(%rsp) ## store is3 movq nb310_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 nb310_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb310_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb310_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 movq nb310_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb310_ntype(%rsp),%edx shll %edx movl %edx,nb310_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb310_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,nb310_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb310_ix(%rsp) movaps %xmm1,nb310_iy(%rsp) movaps %xmm2,nb310_iz(%rsp) movl %ebx,nb310_ii3(%rsp) ## clear vctot and i forces xorps %xmm15,%xmm15 movaps %xmm15,nb310_vctot(%rsp) movaps %xmm15,nb310_Vvdwtot(%rsp) movaps %xmm15,%xmm14 movaps %xmm15,%xmm13 movq nb310_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 nb310_pos(%rbp),%rsi movq nb310_faction(%rbp),%rdi movq nb310_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb310_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb310_ninner(%rsp),%ecx movl %ecx,nb310_ninner(%rsp) addl $0,%edx movl %edx,nb310_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel310_x86_64_sse.nb310_unroll_loop jmp _nb_kernel310_x86_64_sse.nb310_finish_inner_nb_kernel310_x86_64_sse.nb310_unroll_loop: ## quad-unrolled innerloop here movq nb310_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,nb310_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 ## replace jnr with j3 lea (%r11,%r11,2),%rdx ## load coordinates movq nb310_pos(%rbp),%rdi movlps (%rdi,%rax,4),%xmm1 ## x1 y1 - - movlps (%rdi,%rbx,4),%xmm2 ## x2 y2 - - movlps (%rdi,%rcx,4),%xmm3 ## x3 y3 - - movlps (%rdi,%rdx,4),%xmm4 ## x4 y4 - - movss 8(%rdi,%rax,4),%xmm5 ## z1 - - - movss 8(%rdi,%rbx,4),%xmm6 ## z2 - - - movss 8(%rdi,%rcx,4),%xmm7 ## z3 - - - movss 8(%rdi,%rdx,4),%xmm8 ## z4 - - - unpcklps %xmm3,%xmm1 ## x1 x3 y1 y3 unpcklps %xmm4,%xmm2 ## x2 x4 y2 y4 unpcklps %xmm7,%xmm5 ## z1 z3 - - unpcklps %xmm8,%xmm6 ## z2 z4 - - movq nb310_charge(%rbp),%rsi movaps %xmm1,%xmm3 unpcklps %xmm2,%xmm1 ## x1 x2 x3 x4 unpckhps %xmm2,%xmm3 ## y1 y2 y3 y4 unpcklps %xmm6,%xmm5 ## z1 z2 z3 z4 ## calc dr subps nb310_ix(%rsp),%xmm1 subps nb310_iy(%rsp),%xmm3 subps nb310_iz(%rsp),%xmm5 ## store dr in xmm9-xmm11 movaps %xmm1,%xmm9 movaps %xmm3,%xmm10 movaps %xmm5,%xmm11 movss (%rsi,%r8,4),%xmm0 movss (%rsi,%r10,4),%xmm2 movss (%rsi,%r9,4),%xmm6 movss (%rsi,%r11,4),%xmm8 ## square it mulps %xmm1,%xmm1 mulps %xmm3,%xmm3 mulps %xmm5,%xmm5 addps %xmm1,%xmm3 addps %xmm5,%xmm3 ## rsq in xmm3 movq nb310_type(%rbp),%rsi unpcklps %xmm2,%xmm0 unpcklps %xmm8,%xmm6 unpcklps %xmm6,%xmm0 ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm3,%xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb310_three(%rsp),%xmm1 mulps %xmm3,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps nb310_half(%rsp),%xmm1 ## xmm1=rinv ## xmm3=rsq mulps nb310_iq(%rsp),%xmm0 ## vdw types movl (%rsi,%r8,4),%r8d movl (%rsi,%r9,4),%r9d movl (%rsi,%r10,4),%r10d movl (%rsi,%r11,4),%r11d mulps %xmm1,%xmm3 ## r mulps nb310_tsc(%rsp),%xmm3 ## rtab movaps %xmm0,nb310_qq(%rsp) ## truncate and convert to integers cvttps2dq %xmm3,%xmm2 shll %r8d shll %r9d shll %r10d shll %r11d ## convert back to float cvtdq2ps %xmm2,%xmm0 movl nb310_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d addl %edi,%r10d addl %edi,%r11d ## multiply by 4 pslld $2,%xmm2 ## move to integer registers movhlps %xmm2,%xmm7 movd %xmm2,%r12d movd %xmm7,%r14d pshufd $1,%xmm2,%xmm2 pshufd $1,%xmm7,%xmm7 movd %xmm2,%r13d movd %xmm7,%r15d ## calculate eps subps %xmm0,%xmm3 movq nb310_vdwparam(%rbp),%rsi movlps (%rsi,%r8,4),%xmm7 movlps (%rsi,%r10,4),%xmm8 movhps (%rsi,%r9,4),%xmm7 movhps (%rsi,%r11,4),%xmm8 movaps %xmm7,%xmm12 shufps $136,%xmm8,%xmm12 ## 10001000 shufps $221,%xmm8,%xmm7 ## 11011101 movaps %xmm12,nb310_c6(%rsp) movaps %xmm7,nb310_c12(%rsp) movq nb310_VFtab(%rbp),%rsi ## 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 movaps %xmm1,%xmm0 ## rinv mulps %xmm0,%xmm0 ## rinvsq movaps %xmm0,%xmm2 ## rinvsq mulps %xmm2,%xmm2 ## rinv4 mulps %xmm0,%xmm2 ## rinv6 movaps %xmm2,%xmm12 mulps %xmm12,%xmm12 ## rinv12 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 mulps nb310_c6(%rsp),%xmm2 ## vvdw6=c6*rinv6 mulps nb310_c12(%rsp),%xmm12 ## vvdw12=c12*rinv12 movaps %xmm12,%xmm0 subps %xmm2,%xmm12 ## Vvdw=Vvdw12-Vvdw6 ## add potential to vvdwtot addps nb310_Vvdwtot(%rsp),%xmm12 movaps %xmm12,nb310_Vvdwtot(%rsp) shufps $136,%xmm8,%xmm6 ## 10001000 shufps $221,%xmm8,%xmm7 ## 11011101 ## table data ready in xmm4-xmm7 mulps %xmm3,%xmm7 ## Heps mulps %xmm3,%xmm6 ## Geps mulps %xmm3,%xmm7 ## Heps2 addps %xmm6,%xmm5 ## F+Geps⌨️ 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?