nb_kernel100_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,206 行 · 第 1/3 页
S
1,206 行
movl %eax,nb100nf_nouter(%esp) movl %eax,nb100nf_ninner(%esp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb100nf_half(%esp) movss nb100nf_half(%esp),%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,nb100nf_half(%esp) movaps %xmm3,nb100nf_three(%esp)_nb_kernel100nf_ia32_sse.nb100nf_threadloop: movl nb100nf_count(%ebp),%esi ## pointer to sync counter movl (%esi),%eax_nb_kernel100nf_ia32_sse.nb100nf_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_kernel100nf_ia32_sse.nb100nf_spinlock ## if(nn1>nri) nn1=nri movl nb100nf_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,nb100nf_n(%esp) movl %ebx,nb100nf_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel100nf_ia32_sse.nb100nf_outerstart jmp _nb_kernel100nf_ia32_sse.nb100nf_end_nb_kernel100nf_ia32_sse.nb100nf_outerstart: ## ebx contains number of outer iterations addl nb100nf_nouter(%esp),%ebx movl %ebx,nb100nf_nouter(%esp)_nb_kernel100nf_ia32_sse.nb100nf_outer: movl nb100nf_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,nb100nf_is3(%esp) ## store is3 movl nb100nf_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 nb100nf_iinr(%ebp),%ecx ## ecx = pointer into iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movl nb100nf_charge(%ebp),%edx movss (%edx,%ebx,4),%xmm3 mulss nb100nf_facel(%esp),%xmm3 shufps $0,%xmm3,%xmm3 leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb100nf_pos(%ebp),%eax ## eax = base of pos[] addss (%eax,%ebx,4),%xmm0 addss 4(%eax,%ebx,4),%xmm1 addss 8(%eax,%ebx,4),%xmm2 movaps %xmm3,nb100nf_iq(%esp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb100nf_ix(%esp) movaps %xmm1,nb100nf_iy(%esp) movaps %xmm2,nb100nf_iz(%esp) movl %ebx,nb100nf_ii3(%esp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb100nf_vctot(%esp) movl nb100nf_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 nb100nf_pos(%ebp),%esi movl nb100nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb100nf_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb100nf_ninner(%esp),%ecx movl %ecx,nb100nf_ninner(%esp) addl $0,%edx movl %edx,nb100nf_innerk(%esp) ## number of innerloop atoms jge _nb_kernel100nf_ia32_sse.nb100nf_unroll_loop jmp _nb_kernel100nf_ia32_sse.nb100nf_finish_inner_nb_kernel100nf_ia32_sse.nb100nf_unroll_loop: ## quad-unrolled innerloop here movl nb100nf_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,nb100nf_innerjjnr(%esp) ## advance pointer (unrolled 4) movl nb100nf_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 movaps nb100nf_iq(%esp),%xmm5 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 movl nb100nf_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx mulps %xmm5,%xmm3 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 ## x1 y1 - - movlps (%esi,%ecx,4),%xmm5 ## x3 y3 - - movss 8(%esi,%eax,4),%xmm2 ## z1 - - - movss 8(%esi,%ecx,4),%xmm6 ## z3 - - - movhps (%esi,%ebx,4),%xmm4 ## x1 y1 x2 y2 movhps (%esi,%edx,4),%xmm5 ## x3 y3 x4 y4 movss 8(%esi,%ebx,4),%xmm0 ## z2 - - - movss 8(%esi,%edx,4),%xmm1 ## z4 - - - shufps $0,%xmm0,%xmm2 ## z1 z1 z2 z2 shufps $0,%xmm1,%xmm6 ## z3 z3 z4 z4 movaps %xmm4,%xmm0 ## x1 y1 x2 y2 movaps %xmm4,%xmm1 ## x1 y1 x2 y2 shufps $136,%xmm6,%xmm2 ## constant 10001000 ;# z1 z2 z3 z4 shufps $136,%xmm5,%xmm0 ## constant 10001000 ;# x1 x2 x3 x4 shufps $221,%xmm5,%xmm1 ## constant 11011101 ;# y1 y2 y3 y4 ## move nb100nf_ix-iz to xmm4-xmm6 movaps nb100nf_ix(%esp),%xmm4 movaps nb100nf_iy(%esp),%xmm5 movaps nb100nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps nb100nf_vctot(%esp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addps %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%esp) ## should we do one more iteration? subl $4,nb100nf_innerk(%esp) jl _nb_kernel100nf_ia32_sse.nb100nf_finish_inner jmp _nb_kernel100nf_ia32_sse.nb100nf_unroll_loop_nb_kernel100nf_ia32_sse.nb100nf_finish_inner: ## check if at least two particles remain addl $4,nb100nf_innerk(%esp) movl nb100nf_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel100nf_ia32_sse.nb100nf_dopair jmp _nb_kernel100nf_ia32_sse.nb100nf_checksingle_nb_kernel100nf_ia32_sse.nb100nf_dopair: movl nb100nf_charge(%ebp),%esi movl nb100nf_pos(%ebp),%edi movl nb100nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb100nf_innerjjnr(%esp) movss (%esi,%eax,4),%xmm3 movss (%esi,%ebx,4),%xmm6 shufps $0,%xmm6,%xmm3 shufps $8,%xmm3,%xmm3 ## constant 00001000 ;# xmm3(0,1) has the charges leal (%eax,%eax,2),%eax leal (%ebx,%ebx,2),%ebx ## move coordinates to xmm0-xmm2 movlps (%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 movhps (%edi,%ebx,4),%xmm1 movss 8(%edi,%ebx,4),%xmm0 mulps nb100nf_iq(%esp),%xmm3 xorps %xmm7,%xmm7 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## constant 10001000 shufps $136,%xmm0,%xmm0 ## constant 10001000 shufps $221,%xmm1,%xmm1 ## constant 11011101 ## move nb100nf_ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb100nf_ix(%esp),%xmm4 movaps nb100nf_iy(%esp),%xmm5 movaps nb100nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps nb100nf_vctot(%esp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addps %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%esp)_nb_kernel100nf_ia32_sse.nb100nf_checksingle: movl nb100nf_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel100nf_ia32_sse.nb100nf_dosingle jmp _nb_kernel100nf_ia32_sse.nb100nf_updateouterdata_nb_kernel100nf_ia32_sse.nb100nf_dosingle: movl nb100nf_charge(%ebp),%esi movl nb100nf_pos(%ebp),%edi movl nb100nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movss (%esi,%eax,4),%xmm3 ## xmm3(0) has the charge leal (%eax,%eax,2),%eax ## move coordinates to xmm0-xmm2 movss (%edi,%eax,4),%xmm0 movss 4(%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 mulps nb100nf_iq(%esp),%xmm3 xorps %xmm7,%xmm7 movaps nb100nf_ix(%esp),%xmm4 movaps nb100nf_iy(%esp),%xmm5 movaps nb100nf_iz(%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 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps nb100nf_vctot(%esp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addss %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%esp)_nb_kernel100nf_ia32_sse.nb100nf_updateouterdata: ## get n from stack movl nb100nf_n(%esp),%esi ## get group index for i particle movl nb100nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb100nf_vctot(%esp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addps %xmm6,%xmm7 ## pos 0-1 in xmm7 have the sum now movaps %xmm7,%xmm6 shufps $1,%xmm6,%xmm6 addss %xmm6,%xmm7 ## add earlier value from mem movl nb100nf_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb100nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel100nf_ia32_sse.nb100nf_outerend ## not last, iterate outer loop once more! movl %esi,nb100nf_n(%esp) jmp _nb_kernel100nf_ia32_sse.nb100nf_outer_nb_kernel100nf_ia32_sse.nb100nf_outerend: ## check if more outer neighborlists remain movl nb100nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel100nf_ia32_sse.nb100nf_end ## non-zero, do one more workunit jmp _nb_kernel100nf_ia32_sse.nb100nf_threadloop_nb_kernel100nf_ia32_sse.nb100nf_end: emms movl nb100nf_nouter(%esp),%eax movl nb100nf_ninner(%esp),%ebx movl nb100nf_outeriter(%ebp),%ecx movl nb100nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb100nf_salign(%esp),%eax addl %eax,%esp addl $156,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret⌨️ 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?