nb_kernel231_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,071 行 · 第 1/5 页
S
2,071 行
jz _nb_kernel231_ia32_sse.nb231_end ## non-zero, do one more workunit jmp _nb_kernel231_ia32_sse.nb231_threadloop_nb_kernel231_ia32_sse.nb231_end: emms movl nb231_nouter(%esp),%eax movl nb231_ninner(%esp),%ebx movl nb231_outeriter(%ebp),%ecx movl nb231_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb231_salign(%esp),%eax addl %eax,%esp addl $860,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret.globl nb_kernel231nf_ia32_sse.globl _nb_kernel231nf_ia32_ssenb_kernel231nf_ia32_sse: _nb_kernel231nf_ia32_sse: .set nb231nf_p_nri, 8.set nb231nf_iinr, 12.set nb231nf_jindex, 16.set nb231nf_jjnr, 20.set nb231nf_shift, 24.set nb231nf_shiftvec, 28.set nb231nf_fshift, 32.set nb231nf_gid, 36.set nb231nf_pos, 40.set nb231nf_faction, 44.set nb231nf_charge, 48.set nb231nf_p_facel, 52.set nb231nf_argkrf, 56.set nb231nf_argcrf, 60.set nb231nf_Vc, 64.set nb231nf_type, 68.set nb231nf_p_ntype, 72.set nb231nf_vdwparam, 76.set nb231nf_Vvdw, 80.set nb231nf_p_tabscale, 84.set nb231nf_VFtab, 88.set nb231nf_invsqrta, 92.set nb231nf_dvda, 96.set nb231nf_p_gbtabscale, 100.set nb231nf_GBtab, 104.set nb231nf_p_nthreads, 108.set nb231nf_count, 112.set nb231nf_mtx, 116.set nb231nf_outeriter, 120.set nb231nf_inneriter, 124.set nb231nf_work, 128 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb231nf_ixO, 0.set nb231nf_iyO, 16.set nb231nf_izO, 32.set nb231nf_ixH1, 48.set nb231nf_iyH1, 64.set nb231nf_izH1, 80.set nb231nf_ixH2, 96.set nb231nf_iyH2, 112.set nb231nf_izH2, 128.set nb231nf_iqO, 144.set nb231nf_iqH, 160.set nb231nf_qqO, 176.set nb231nf_qqH, 192.set nb231nf_c6, 208.set nb231nf_c12, 224.set nb231nf_tsc, 240.set nb231nf_vctot, 256.set nb231nf_Vvdwtot, 272.set nb231nf_half, 288.set nb231nf_three, 304.set nb231nf_krf, 320.set nb231nf_crf, 336.set nb231nf_krsqO, 352.set nb231nf_krsqH1, 368.set nb231nf_krsqH2, 384.set nb231nf_rinvO, 400.set nb231nf_rinvH1, 416.set nb231nf_rinvH2, 432.set nb231nf_is3, 448.set nb231nf_ii3, 452.set nb231nf_ntia, 456.set nb231nf_innerjjnr, 460.set nb231nf_innerk, 464.set nb231nf_n, 468.set nb231nf_nn1, 472.set nb231nf_nri, 476.set nb231nf_nouter, 480.set nb231nf_ninner, 484.set nb231nf_salign, 488 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $492,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb231nf_salign(%esp) emms ## Move args passed by reference to stack movl nb231nf_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb231nf_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb231nf_nouter(%esp) movl %eax,nb231nf_ninner(%esp) movl nb231nf_p_tabscale(%ebp),%eax movss (%eax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb231nf_tsc(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb231nf_half(%esp) movss nb231nf_half(%esp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## constant 1.0 movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## constant 2.0 addps %xmm2,%xmm3 ## constant 3.0 movaps %xmm1,nb231nf_half(%esp) movaps %xmm3,nb231nf_three(%esp) movl nb231nf_argkrf(%ebp),%esi movl nb231nf_argcrf(%ebp),%edi movss (%esi),%xmm5 movss (%edi),%xmm6 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 movaps %xmm5,nb231nf_krf(%esp) movaps %xmm6,nb231nf_crf(%esp) ## assume we have at least one i particle - start directly movl nb231nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb231nf_charge(%ebp),%edx movss (%edx,%ebx,4),%xmm3 movss 4(%edx,%ebx,4),%xmm4 movl nb231nf_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb231nf_iqO(%esp) movaps %xmm4,nb231nf_iqH(%esp) movl nb231nf_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb231nf_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb231nf_ntia(%esp)_nb_kernel231nf_ia32_sse.nb231nf_threadloop: movl nb231nf_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel231nf_ia32_sse.nb231nf_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%esi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel231nf_ia32_sse.nb231nf_spinlock ## if(nn1>nri) nn1=nri movl nb231nf_nri(%esp),%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,nb231nf_n(%esp) movl %ebx,nb231nf_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel231nf_ia32_sse.nb231nf_outerstart jmp _nb_kernel231nf_ia32_sse.nb231nf_end_nb_kernel231nf_ia32_sse.nb231nf_outerstart: ## ebx contains number of outer iterations addl nb231nf_nouter(%esp),%ebx movl %ebx,nb231nf_nouter(%esp)_nb_kernel231nf_ia32_sse.nb231nf_outer: movl nb231nf_shift(%ebp),%eax ## eax = pointer into shift[] movl (%eax,%esi,4),%ebx ## ebx=shift[n] leal (%ebx,%ebx,2),%ebx ## ebx=3*is movl %ebx,nb231nf_is3(%esp) ## store is3 movl nb231nf_shiftvec(%ebp),%eax ## eax = base of shiftvec[] movss (%eax,%ebx,4),%xmm0 movss 4(%eax,%ebx,4),%xmm1 movss 8(%eax,%ebx,4),%xmm2 movl nb231nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb231nf_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb231nf_ii3(%esp) addss (%eax,%ebx,4),%xmm3 addss 4(%eax,%ebx,4),%xmm4 addss 8(%eax,%ebx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb231nf_ixO(%esp) movaps %xmm4,nb231nf_iyO(%esp) movaps %xmm5,nb231nf_izO(%esp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%eax,%ebx,4),%xmm0 addss 16(%eax,%ebx,4),%xmm1 addss 20(%eax,%ebx,4),%xmm2 addss 24(%eax,%ebx,4),%xmm3 addss 28(%eax,%ebx,4),%xmm4 addss 32(%eax,%ebx,4),%xmm5 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm0,nb231nf_ixH1(%esp) movaps %xmm1,nb231nf_iyH1(%esp) movaps %xmm2,nb231nf_izH1(%esp) movaps %xmm3,nb231nf_ixH2(%esp) movaps %xmm4,nb231nf_iyH2(%esp) movaps %xmm5,nb231nf_izH2(%esp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb231nf_vctot(%esp) movaps %xmm4,nb231nf_Vvdwtot(%esp) movl nb231nf_jindex(%ebp),%eax movl (%eax,%esi,4),%ecx ## jindex[n] movl 4(%eax,%esi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movl nb231nf_pos(%ebp),%esi movl nb231nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb231nf_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb231nf_ninner(%esp),%ecx movl %ecx,nb231nf_ninner(%esp) addl $0,%edx movl %edx,nb231nf_innerk(%esp) ## number of innerloop atoms jge _nb_kernel231nf_ia32_sse.nb231nf_unroll_loop jmp _nb_kernel231nf_ia32_sse.nb231nf_odd_inner_nb_kernel231nf_ia32_sse.nb231nf_unroll_loop: ## quad-unroll innerloop here movl nb231nf_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax movl 4(%edx),%ebx movl 8(%edx),%ecx movl 12(%edx),%edx ## eax-edx=jnr1-4 addl $16,nb231nf_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb231nf_charge(%ebp),%esi ## base of charge[] movss (%esi,%eax,4),%xmm3 movss (%esi,%ecx,4),%xmm4 movss (%esi,%ebx,4),%xmm6 movss (%esi,%edx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb231nf_iqO(%esp),%xmm3 mulps nb231nf_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb231nf_qqO(%esp) movaps %xmm4,nb231nf_qqH(%esp) movl nb231nf_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb231nf_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb231nf_ntia(%esp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%esi,%eax,4),%xmm6 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm6 movhps (%esi,%edx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb231nf_c6(%esp) movaps %xmm6,nb231nf_c12(%esp) movl nb231nf_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx leal (%ecx,%ecx,2),%ecx ## replace jnr with j3 leal (%edx,%edx,2),%edx ## move four coordinates to xmm0-xmm2 movlps (%esi,%eax,4),%xmm4 movlps (%esi,%ecx,4),%xmm5 movss 8(%esi,%eax,4),%xmm2 movss 8(%esi,%ecx,4),%xmm6 movhps (%esi,%ebx,4),%xmm4 movhps (%esi,%edx,4),%xmm5 movss 8(%esi,%ebx,4),%xmm0 movss 8(%esi,%edx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb231nf_ixO(%esp),%xmm4 movaps nb231nf_iyO(%esp),%xmm5 movaps nb231nf_izO(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb231nf_ixH1(%esp),%xmm4 movaps nb231nf_iyH1(%esp),%xmm5 movaps nb231nf_izH1(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5⌨️ 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?