nb_kernel330_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,027 行 · 第 1/4 页
S
2,027 行
#### $Id: nb_kernel330_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_kernel330_x86_64_sse.globl _nb_kernel330_x86_64_ssenb_kernel330_x86_64_sse: _nb_kernel330_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb330_fshift, 16.set nb330_gid, 24.set nb330_pos, 32.set nb330_faction, 40.set nb330_charge, 48.set nb330_p_facel, 56.set nb330_argkrf, 64.set nb330_argcrf, 72.set nb330_Vc, 80.set nb330_type, 88.set nb330_p_ntype, 96.set nb330_vdwparam, 104.set nb330_Vvdw, 112.set nb330_p_tabscale, 120.set nb330_VFtab, 128.set nb330_invsqrta, 136.set nb330_dvda, 144.set nb330_p_gbtabscale, 152.set nb330_GBtab, 160.set nb330_p_nthreads, 168.set nb330_count, 176.set nb330_mtx, 184.set nb330_outeriter, 192.set nb330_inneriter, 200.set nb330_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb330_ix, 0.set nb330_iy, 16.set nb330_iz, 32.set nb330_iq, 48.set nb330_dx, 64.set nb330_dy, 80.set nb330_dz, 96.set nb330_rinv, 112.set nb330_tsc, 128.set nb330_qq, 144.set nb330_c6, 160.set nb330_c12, 176.set nb330_eps, 192.set nb330_vctot, 208.set nb330_Vvdwtot, 224.set nb330_fix, 240.set nb330_fiy, 256.set nb330_fiz, 272.set nb330_half, 288.set nb330_three, 304.set nb330_nri, 320.set nb330_iinr, 328.set nb330_jindex, 336.set nb330_jjnr, 344.set nb330_shift, 352.set nb330_shiftvec, 360.set nb330_facel, 368.set nb330_innerjjnr, 376.set nb330_is3, 384.set nb330_ii3, 388.set nb330_ntia, 392.set nb330_innerk, 396.set nb330_n, 400.set nb330_nn1, 404.set nb330_ntype, 408.set nb330_nouter, 412.set nb330_ninner, 416 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $440,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb330_nouter(%rsp) movl %eax,nb330_ninner(%rsp) movl (%rdi),%edi movl %edi,nb330_nri(%rsp) movq %rsi,nb330_iinr(%rsp) movq %rdx,nb330_jindex(%rsp) movq %rcx,nb330_jjnr(%rsp) movq %r8,nb330_shift(%rsp) movq %r9,nb330_shiftvec(%rsp) movq nb330_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb330_ntype(%rsp) movq nb330_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb330_facel(%rsp) movq nb330_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb330_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb330_half(%rsp) movss nb330_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,nb330_half(%rsp) movaps %xmm3,nb330_three(%rsp)_nb_kernel330_x86_64_sse.nb330_threadloop: movq nb330_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel330_x86_64_sse.nb330_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_kernel330_x86_64_sse.nb330_spinlock ## if(nn1>nri) nn1=nri movl nb330_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,nb330_n(%rsp) movl %ebx,nb330_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel330_x86_64_sse.nb330_outerstart jmp _nb_kernel330_x86_64_sse.nb330_end_nb_kernel330_x86_64_sse.nb330_outerstart: ## ebx contains number of outer iterations addl nb330_nouter(%rsp),%ebx movl %ebx,nb330_nouter(%rsp)_nb_kernel330_x86_64_sse.nb330_outer: movq nb330_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,nb330_is3(%rsp) ## store is3 movq nb330_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 nb330_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb330_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb330_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 movq nb330_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb330_ntype(%rsp),%edx shll %edx movl %edx,nb330_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb330_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,nb330_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb330_ix(%rsp) movaps %xmm1,nb330_iy(%rsp) movaps %xmm2,nb330_iz(%rsp) movl %ebx,nb330_ii3(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb330_vctot(%rsp) movaps %xmm4,nb330_Vvdwtot(%rsp) movaps %xmm4,nb330_fix(%rsp) movaps %xmm4,nb330_fiy(%rsp) movaps %xmm4,nb330_fiz(%rsp) movq nb330_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 movl nb330_charge(%rsp),%esi movl nb330_ntia(%rsp),%edi movq nb330_pos(%rbp),%rsi movq nb330_faction(%rbp),%rdi movq nb330_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb330_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb330_ninner(%rsp),%ecx movl %ecx,nb330_ninner(%rsp) addl $0,%edx movl %edx,nb330_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel330_x86_64_sse.nb330_unroll_loop jmp _nb_kernel330_x86_64_sse.nb330_finish_inner_nb_kernel330_x86_64_sse.nb330_unroll_loop: ## quad-unroll innerloop here movq nb330_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r12d movl 4(%rdx),%r13d movl 8(%rdx),%r14d movl 12(%rdx),%r15d ## eax-edx=jnr1-4 addq $16,nb330_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%r12,%r12,2),%rax ## replace jnr with j3 lea (%r13,%r13,2),%rbx lea (%r14,%r14,2),%rcx lea (%r15,%r15,2),%rdx movq nb330_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 - 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 nb330_charge(%rbp),%rsi ## calc dr subps nb330_ix(%rsp),%xmm1 subps nb330_iy(%rsp),%xmm2 subps nb330_iz(%rsp),%xmm5 ## store dr movaps %xmm1,nb330_dx(%rsp) movaps %xmm2,nb330_dy(%rsp) movaps %xmm5,nb330_dz(%rsp) movss (%rsi,%r12,4),%xmm0 movss (%rsi,%r14,4),%xmm3 movss (%rsi,%r13,4),%xmm4 movss (%rsi,%r15,4),%xmm6 ## square it mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm5,%xmm5 addps %xmm2,%xmm1 addps %xmm5,%xmm1 ## rsq in xmm1 unpcklps %xmm3,%xmm0 ## jqa jqc - - unpcklps %xmm6,%xmm4 ## jqb jqd - - unpcklps %xmm4,%xmm0 ## jqa jqb jqc jqd mulps nb330_iq(%rsp),%xmm0 movaps %xmm0,nb330_qq(%rsp) movq nb330_type(%rbp),%rsi ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm1,%xmm5 movaps %xmm5,%xmm12 mulps %xmm5,%xmm5 movaps nb330_three(%rsp),%xmm4 mulps %xmm1,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm4 ## 30-rsq*lu*lu mulps %xmm12,%xmm4 mulps nb330_half(%rsp),%xmm4 movaps %xmm4,%xmm15 mulps %xmm4,%xmm1 ## xmm15=rinv ## xmm1=r ## vdw parameters movl (%rsi,%r12,4),%r12d movl (%rsi,%r13,4),%r13d movl (%rsi,%r14,4),%r14d movl (%rsi,%r15,4),%r15d mulps nb330_tsc(%rsp),%xmm1 ## rtab ## truncate and convert to integers cvttps2dq %xmm1,%xmm5 shll %r12d shll %r13d shll %r14d shll %r15d movl nb330_ntia(%rsp),%edi ## convert back to float cvtdq2ps %xmm5,%xmm4 ## multiply by 4 pslld $2,%xmm5 ## multiply by three (copy, mult. by two, add back) movaps %xmm5,%xmm6 pslld $1,%xmm5 paddd %xmm6,%xmm5 ## calculate eps subps %xmm4,%xmm1 addl %edi,%r12d addl %edi,%r13d addl %edi,%r14d addl %edi,%r15d ## 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 movq nb330_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,nb330_c6(%rsp) movaps %xmm3,nb330_c12(%rsp) movaps %xmm1,nb330_eps(%rsp) ## xmm15=rinv movq nb330_VFtab(%rbp),%rsi ## load Coulomb and LJ table data in parallel movlps (%rsi,%r8,4),%xmm1 ## Y1c F1c movlps 16(%rsi,%r8,4),%xmm5 ## Y1d F1d movlps 32(%rsi,%r8,4),%xmm9 ## Y1r F1r movlps (%rsi,%r10,4),%xmm3 ## Y3c F3c movlps 16(%rsi,%r10,4),%xmm7 ## Y3d F3d movlps 32(%rsi,%r10,4),%xmm11 ## Y3r F3r movhps (%rsi,%r9,4),%xmm1 ## Y1c F1c Y2c F2c movhps 16(%rsi,%r9,4),%xmm5 ## Y1d F1d Y2d F2d movhps 32(%rsi,%r9,4),%xmm9 ## Y1r F1r Y2r F2r movhps (%rsi,%r11,4),%xmm3 ## Y3c F3c Y4c F4c movhps 16(%rsi,%r11,4),%xmm7 ## Y3d F3d Y4d F4d movhps 32(%rsi,%r11,4),%xmm11 ## Y3r F3r Y4r F4r movaps %xmm1,%xmm0 movaps %xmm5,%xmm4 movaps %xmm9,%xmm8 shufps $136,%xmm3,%xmm0 ## 10001000 => Y1c Y2c Y3c Y4c shufps $136,%xmm7,%xmm4 ## 10001000 => Y1d Y2d Y3d Y4d shufps $136,%xmm11,%xmm8 ## 10001000 => Y1r Y2r Y3r Y4r shufps $221,%xmm3,%xmm1 ## 11011101 => F1c F2c F3c F4c shufps $221,%xmm7,%xmm5 ## 11011101 => F1d F2d F3d F4d shufps $221,%xmm11,%xmm9 ## 11011101 => F1r F2r F3r F4r movlps 8(%rsi,%r8,4),%xmm3 ## G1c H1c movlps 24(%rsi,%r8,4),%xmm7 ## G1d H1d movlps 40(%rsi,%r8,4),%xmm11 ## G1r H1r movlps 8(%rsi,%r10,4),%xmm12 ## G3c H3c movlps 24(%rsi,%r10,4),%xmm13 ## G3d H3d movlps 40(%rsi,%r10,4),%xmm14 ## G3r H3r movhps 8(%rsi,%r9,4),%xmm3 ## G1c H1c G2c H2c movhps 24(%rsi,%r9,4),%xmm7 ## G1d H1d G2d H2d movhps 40(%rsi,%r9,4),%xmm11 ## G1r H1r G2r H2r movhps 8(%rsi,%r11,4),%xmm12 ## G3c H3c G4c H4c movhps 24(%rsi,%r11,4),%xmm13 ## G3d H3d G4d H4d movhps 40(%rsi,%r11,4),%xmm14 ## G3r H3r G4r H4r movaps %xmm3,%xmm2 movaps %xmm7,%xmm6 movaps %xmm11,%xmm10 shufps $136,%xmm12,%xmm2 ## 10001000 => G1c G2c G3c G4c shufps $136,%xmm13,%xmm6 ## 10001000 => G1d G2d G3d G4d shufps $136,%xmm14,%xmm10 ## 10001000 => G1r G2r G3r G4r shufps $221,%xmm12,%xmm3 ## 11011101 => H1c H2c H3c H4c shufps $221,%xmm13,%xmm7 ## 11011101 => H1d H2d H3d H4d shufps $221,%xmm14,%xmm11 ## 11011101 => H1r H2r H3r H4r ## table data ready. Coul in xmm0-xmm3 , disp in xmm4-xmm7 , rep. in xmm8-xmm11 movaps nb330_eps(%rsp),%xmm12 mulps %xmm12,%xmm3 ## Heps mulps %xmm12,%xmm7 mulps %xmm12,%xmm11 mulps %xmm12,%xmm2 ## Geps mulps %xmm12,%xmm6 mulps %xmm12,%xmm10 mulps %xmm12,%xmm3 ## Heps2 mulps %xmm12,%xmm7 mulps %xmm12,%xmm11 addps %xmm2,%xmm1 ## F+Geps addps %xmm6,%xmm5 addps %xmm10,%xmm9 addps %xmm3,%xmm1 ## F+Geps+Heps2 = Fp addps %xmm7,%xmm5 addps %xmm11,%xmm9 addps %xmm3,%xmm3 ## 2*Heps2 addps %xmm7,%xmm7 addps %xmm11,%xmm11 addps %xmm2,%xmm3 ## 2*Heps2+Geps addps %xmm6,%xmm7 addps %xmm10,%xmm11 addps %xmm1,%xmm3 ## FF = Fp + 2*Heps2 + Geps addps %xmm5,%xmm7 addps %xmm9,%xmm11 mulps %xmm12,%xmm1 ## eps*Fp mulps %xmm12,%xmm5 mulps %xmm12,%xmm9 addps %xmm0,%xmm1 ## VV addps %xmm4,%xmm5 addps %xmm8,%xmm9 mulps nb330_qq(%rsp),%xmm1 ## VV*qq = vcoul mulps nb330_c6(%rsp),%xmm5 ## vnb6 mulps nb330_c12(%rsp),%xmm9 ## vnb12 mulps nb330_qq(%rsp),%xmm3 ## FF*qq = fij mulps nb330_c6(%rsp),%xmm7 ## fijD mulps nb330_c12(%rsp),%xmm11 ##fijR ## accumulate vctot addps nb330_vctot(%rsp),%xmm1 ## accumulate Vvdwtot addps nb330_Vvdwtot(%rsp),%xmm5 addps %xmm9,%xmm5 xorps %xmm9,%xmm9 movaps %xmm1,nb330_vctot(%rsp) movaps %xmm5,nb330_Vvdwtot(%rsp) movq nb330_faction(%rbp),%rsi ## the fj's - start by accumulating x & y forces from memory movlps (%rsi,%rax,4),%xmm0 ## x1 y1 - -⌨️ 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?