nb_kernel010_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,354 行 · 第 1/3 页
S
1,354 行
movl (%esi),%eax_nb_kernel010nf_ia32_sse.nb010nf_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_kernel010nf_ia32_sse.nb010nf_spinlock ## if(nn1>nri) nn1=nri movl nb010nf_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,nb010nf_n(%esp) movl %ebx,nb010nf_nn1(%esp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel010nf_ia32_sse.nb010nf_outerstart jmp _nb_kernel010nf_ia32_sse.nb010nf_end_nb_kernel010nf_ia32_sse.nb010nf_outerstart: ## ebx contains number of outer iterations addl nb010nf_nouter(%esp),%ebx movl %ebx,nb010nf_nouter(%esp)_nb_kernel010nf_ia32_sse.nb010nf_outer: movl nb010nf_shift(%ebp),%eax ## eax = base of shift[] movl (%eax,%esi,4),%ebx ## ebx=shift[n] leal (%ebx,%ebx,2),%ebx ## ebx=3*is movl %ebx,nb010nf_is3(%esp) ## store is3 movl nb010nf_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 nb010nf_iinr(%ebp),%ecx ## ecx = base of iinr[] movl (%ecx,%esi,4),%ebx ## ebx =ii movl nb010nf_type(%ebp),%edx movl (%edx,%ebx,4),%edx imull nb010nf_ntype(%esp),%edx shll %edx movl %edx,nb010nf_ntia(%esp) leal (%ebx,%ebx,2),%ebx ## ebx = 3*ii=ii3 movl nb010nf_pos(%ebp),%eax ## eax = base of pos[] addss (%eax,%ebx,4),%xmm0 addss 4(%eax,%ebx,4),%xmm1 addss 8(%eax,%ebx,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb010nf_ix(%esp) movaps %xmm1,nb010nf_iy(%esp) movaps %xmm2,nb010nf_iz(%esp) movl %ebx,nb010nf_ii3(%esp) ## clear Vvdwtot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb010nf_Vvdwtot(%esp) movl nb010nf_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 nb010nf_pos(%ebp),%esi movl nb010nf_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb010nf_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb010nf_ninner(%esp),%ecx movl %ecx,nb010nf_ninner(%esp) addl $0,%edx movl %edx,nb010nf_innerk(%esp) ## number of innerloop atoms jge _nb_kernel010nf_ia32_sse.nb010nf_unroll_loop jmp _nb_kernel010nf_ia32_sse.nb010nf_finish_inner_nb_kernel010nf_ia32_sse.nb010nf_unroll_loop: ## quad-unroll innerloop here movl nb010nf_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 ## advance pointer (unrolled 4) addl $16,nb010nf_innerjjnr(%esp) movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb010nf_type(%ebp),%esi movl (%esi,%eax,4),%eax movl (%esi,%ebx,4),%ebx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb010nf_vdwparam(%ebp),%esi shll %eax shll %ebx shll %ecx shll %edx movl nb010nf_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,nb010nf_c6(%esp) movaps %xmm6,nb010nf_c12(%esp) movl nb010nf_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx mulps %xmm2,%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 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 ix-iz to xmm4-xmm6 movaps nb010nf_ix(%esp),%xmm4 movaps nb010nf_iy(%esp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## constant 1/x lookup seed in xmm5 movaps nb010nf_two(%esp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%esp),%xmm1 mulps nb010nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb010nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb010nf_Vvdwtot(%esp) ## should we do one more iteration? subl $4,nb010nf_innerk(%esp) jl _nb_kernel010nf_ia32_sse.nb010nf_finish_inner jmp _nb_kernel010nf_ia32_sse.nb010nf_unroll_loop_nb_kernel010nf_ia32_sse.nb010nf_finish_inner: ## check if at least two particles remain addl $4,nb010nf_innerk(%esp) movl nb010nf_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel010nf_ia32_sse.nb010nf_dopair jmp _nb_kernel010nf_ia32_sse.nb010nf_checksingle_nb_kernel010nf_ia32_sse.nb010nf_dopair: movl nb010nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb010nf_innerjjnr(%esp) movl nb010nf_type(%ebp),%esi movl %eax,%ecx movl %ebx,%edx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb010nf_vdwparam(%ebp),%esi shll %ecx shll %edx movl nb010nf_ntia(%esp),%edi addl %edi,%ecx addl %edi,%edx movlps (%esi,%ecx,4),%xmm6 movhps (%esi,%edx,4),%xmm6 movl nb010nf_pos(%ebp),%edi xorps %xmm7,%xmm7 movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## constant 00001000 shufps $13,%xmm6,%xmm6 ## constant 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb010nf_c6(%esp) movaps %xmm6,nb010nf_c12(%esp) 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 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 nb010nf_ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb010nf_ix(%esp),%xmm4 movaps nb010nf_iy(%esp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## constant 1/x lookup seed in xmm5 movaps nb010nf_two(%esp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%esp),%xmm1 mulps nb010nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb010nf_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb010nf_Vvdwtot(%esp)_nb_kernel010nf_ia32_sse.nb010nf_checksingle: movl nb010nf_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel010nf_ia32_sse.nb010nf_dosingle jmp _nb_kernel010nf_ia32_sse.nb010nf_updateouterdata_nb_kernel010nf_ia32_sse.nb010nf_dosingle: movl nb010nf_pos(%ebp),%edi movl nb010nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl nb010nf_type(%ebp),%esi movl %eax,%ecx movl (%esi,%ecx,4),%ecx movl nb010nf_vdwparam(%ebp),%esi shll %ecx addl nb010nf_ntia(%esp),%ecx xorps %xmm6,%xmm6 movlps (%esi,%ecx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## constant 11111100 shufps $253,%xmm6,%xmm6 ## constant 11111101 movaps %xmm4,nb010nf_c6(%esp) movaps %xmm6,nb010nf_c12(%esp) 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 xorps %xmm7,%xmm7 movaps nb010nf_ix(%esp),%xmm4 movaps nb010nf_iy(%esp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## constant 1/x lookup seed in xmm5 movaps nb010nf_two(%esp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%esp),%xmm1 mulps nb010nf_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addss nb010nf_Vvdwtot(%esp),%xmm5 movss %xmm5,nb010nf_Vvdwtot(%esp)_nb_kernel010nf_ia32_sse.nb010nf_updateouterdata: ## get n from stack movl nb010nf_n(%esp),%esi ## get group index for i particle movl nb010nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total lj energy and update it movaps nb010nf_Vvdwtot(%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 nb010nf_Vvdw(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb010nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel010nf_ia32_sse.nb010nf_outerend ## not last, iterate outer loop once more! movl %esi,nb010nf_n(%esp) jmp _nb_kernel010nf_ia32_sse.nb010nf_outer_nb_kernel010nf_ia32_sse.nb010nf_outerend: ## check if more outer neighborlists remain movl nb010nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel010nf_ia32_sse.nb010nf_end ## non-zero, do one more workunit jmp _nb_kernel010nf_ia32_sse.nb010nf_threadloop_nb_kernel010nf_ia32_sse.nb010nf_end: emms movl nb010nf_nouter(%esp),%eax movl nb010nf_ninner(%esp),%ebx movl nb010nf_outeriter(%ebp),%ecx movl nb010nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb010nf_salign(%esp),%eax addl %eax,%esp addl $192,%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?