nb_kernel333_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,082 行 · 第 1/5 页
S
2,082 行
.set nb333nf_c12, 416.set nb333nf_vctot, 432.set nb333nf_Vvdwtot, 448.set nb333nf_half, 464.set nb333nf_three, 480.set nb333nf_is3, 496.set nb333nf_ii3, 500.set nb333nf_ntia, 504.set nb333nf_innerjjnr, 508.set nb333nf_innerk, 512.set nb333nf_n, 516.set nb333nf_nn1, 520.set nb333nf_nri, 524.set nb333nf_nouter, 528.set nb333nf_ninner, 532.set nb333nf_salign, 536 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $540,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb333nf_salign(%esp) emms ## Move args passed by reference to stack movl nb333nf_p_nri(%ebp),%ecx movl (%ecx),%ecx movl %ecx,nb333nf_nri(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb333nf_nouter(%esp) movl %eax,nb333nf_ninner(%esp) movl nb333nf_p_tabscale(%ebp),%eax movss (%eax),%xmm5 shufps $0,%xmm5,%xmm5 movaps %xmm5,nb333nf_tsc(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## constant 0.5 in IEEE (hex) movl %eax,nb333nf_half(%esp) movss nb333nf_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,nb333nf_half(%esp) movaps %xmm3,nb333nf_three(%esp) ## assume we have at least one i particle - start directly movl nb333nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx),%ebx ## ebx =ii movl nb333nf_charge(%ebp),%edx movss 4(%edx,%ebx,4),%xmm4 movss 12(%edx,%ebx,4),%xmm3 movl nb333nf_p_facel(%ebp),%esi movss (%esi),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb333nf_iqM(%esp) movaps %xmm4,nb333nf_iqH(%esp) movl nb333nf_type(%ebp),%edx movl (%edx,%ebx,4),%ecx shll %ecx movl nb333nf_p_ntype(%ebp),%edi imull (%edi),%ecx ## ecx = ntia = 2*ntype*type[ii0] movl %ecx,nb333nf_ntia(%esp)_nb_kernel333nf_ia32_sse.nb333nf_threadloop: movl nb333nf_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel333nf_ia32_sse.nb333nf_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_kernel333nf_ia32_sse.nb333nf_spinlock ## if(nn1>nri) nn1=nri movl nb333nf_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,nb333nf_n(%esp) movl %ebx,nb333nf_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel333nf_ia32_sse.nb333nf_outerstart jmp _nb_kernel333nf_ia32_sse.nb333nf_end_nb_kernel333nf_ia32_sse.nb333nf_outerstart: ## ebx contains number of outer iterations addl nb333nf_nouter(%esp),%ebx movl %ebx,nb333nf_nouter(%esp)_nb_kernel333nf_ia32_sse.nb333nf_outer: movl nb333nf_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,nb333nf_is3(%esp) ## store is3 movl nb333nf_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 nb333nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb333nf_pos(%ebp),%eax ## eax = base of pos[] movl %ebx,nb333nf_ii3(%esp) addss (%eax,%ebx,4),%xmm3 ## ox addss 4(%eax,%ebx,4),%xmm4 ## oy addss 8(%eax,%ebx,4),%xmm5 ## oz addss 12(%eax,%ebx,4),%xmm6 ## h1x addss 16(%eax,%ebx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb333nf_ixO(%esp) movaps %xmm4,nb333nf_iyO(%esp) movaps %xmm5,nb333nf_izO(%esp) movaps %xmm6,nb333nf_ixH1(%esp) movaps %xmm7,nb333nf_iyH1(%esp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%eax,%ebx,4),%xmm6 ## h1z addss 24(%eax,%ebx,4),%xmm0 ## h2x addss 28(%eax,%ebx,4),%xmm1 ## h2y addss 32(%eax,%ebx,4),%xmm2 ## h2z addss 36(%eax,%ebx,4),%xmm3 ## mx addss 40(%eax,%ebx,4),%xmm4 ## my addss 44(%eax,%ebx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 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 %xmm6,nb333nf_izH1(%esp) movaps %xmm0,nb333nf_ixH2(%esp) movaps %xmm1,nb333nf_iyH2(%esp) movaps %xmm2,nb333nf_izH2(%esp) movaps %xmm3,nb333nf_ixM(%esp) movaps %xmm4,nb333nf_iyM(%esp) movaps %xmm5,nb333nf_izM(%esp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb333nf_vctot(%esp) movaps %xmm4,nb333nf_Vvdwtot(%esp) movl nb333nf_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 nb333nf_pos(%ebp),%esi movl nb333nf_faction(%ebp),%edi movl nb333nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb333nf_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb333nf_ninner(%esp),%ecx movl %ecx,nb333nf_ninner(%esp) addl $0,%edx movl %edx,nb333nf_innerk(%esp) ## number of innerloop atoms jge _nb_kernel333nf_ia32_sse.nb333nf_unroll_loop jmp _nb_kernel333nf_ia32_sse.nb333nf_odd_inner_nb_kernel333nf_ia32_sse.nb333nf_unroll_loop: ## quad-unroll innerloop here movl nb333nf_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,nb333nf_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb333nf_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 nb333nf_iqM(%esp),%xmm3 mulps nb333nf_iqH(%esp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb333nf_qqM(%esp) movaps %xmm4,nb333nf_qqH(%esp) movl nb333nf_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb333nf_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb333nf_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,nb333nf_c6(%esp) movaps %xmm6,nb333nf_c12(%esp) movl nb333nf_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 nb333nf_ixO(%esp),%xmm4 movaps nb333nf_iyO(%esp),%xmm5 movaps nb333nf_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 nb333nf_ixH1(%esp),%xmm4 movaps nb333nf_iyH1(%esp),%xmm5 movaps nb333nf_izH1(%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,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb333nf_ixH2(%esp),%xmm3 movaps nb333nf_iyH2(%esp),%xmm4 movaps nb333nf_izH2(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## move ixM-izM to xmm2-xmm4 movaps nb333nf_iyM(%esp),%xmm3 movaps nb333nf_izM(%esp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb333nf_ixM(%esp),%xmm2 subps %xmm0,%xmm2 ## square it mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 addps %xmm3,%xmm4 addps %xmm2,%xmm4 ## rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%esp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%esp),%xmm0 movaps %xmm0,nb333nf_rinvH1(%esp) ## rinvH1 mulps %xmm0,%xmm6 movaps %xmm6,nb333nf_rH1(%esp) ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%esp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%esp),%xmm0 movaps %xmm0,nb333nf_rinvH2(%esp) ## rinvH2 mulps %xmm0,%xmm5 movaps %xmm5,nb333nf_rH2(%esp) ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333nf_three(%esp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333nf_half(%esp),%xmm0 movaps %xmm0,nb333nf_rinvM(%esp) ## rinvM mulps %xmm0,%xmm4 movaps %xmm4,nb333nf_rM(%esp) ## Do the O LJ table interaction dir⌨️ 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?