nb_kernel300_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,505 行 · 第 1/4 页
S
1,505 行
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 nb300nf_ix(%esp),%xmm4 movaps nb300nf_iy(%esp),%xmm5 movaps nb300nf_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 nb300nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb300nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r mulps nb300nf_tsc(%esp),%xmm4 movhlps %xmm4,%xmm5 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm5,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm6 cvtpi2ps %mm7,%xmm5 movlhps %xmm5,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb300nf_VFtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%ebx,4),%xmm5 movhps (%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%ebx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb300nf_qq(%esp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV ## at this point xmm5 contains vcoul ## increment vcoul - then we can get rid of mm5 ## update vctot addps nb300nf_vctot(%esp),%xmm5 movaps %xmm5,nb300nf_vctot(%esp) ## should we do one more iteration? subl $4,nb300nf_innerk(%esp) jl _nb_kernel300nf_ia32_sse.nb300nf_finish_inner jmp _nb_kernel300nf_ia32_sse.nb300nf_unroll_loop_nb_kernel300nf_ia32_sse.nb300nf_finish_inner: ## check if at least two particles remain addl $4,nb300nf_innerk(%esp) movl nb300nf_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel300nf_ia32_sse.nb300nf_dopair jmp _nb_kernel300nf_ia32_sse.nb300nf_checksingle_nb_kernel300nf_ia32_sse.nb300nf_dopair: movl nb300nf_charge(%ebp),%esi movl nb300nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb300nf_innerjjnr(%esp) xorps %xmm7,%xmm7 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 mulps nb300nf_iq(%esp),%xmm3 movlhps %xmm7,%xmm3 movaps %xmm3,nb300nf_qq(%esp) movl nb300nf_pos(%ebp),%edi 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 ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb300nf_ix(%esp),%xmm4 movaps nb300nf_iy(%esp),%xmm5 movaps nb300nf_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 nb300nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb300nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r mulps nb300nf_tsc(%esp),%xmm4 cvttps2pi %xmm4,%mm6 ## mm6 contain lu indices cvtpi2ps %mm6,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 movl nb300nf_VFtab(%ebp),%esi movd %mm6,%ecx psrlq $32,%mm6 movd %mm6,%edx movlps (%esi,%ecx,4),%xmm5 movhps (%esi,%edx,4),%xmm5 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm4,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%esi,%ecx,4),%xmm7 movhps 8(%esi,%edx,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm6,%xmm6 ## constant 10001000 shufps $221,%xmm7,%xmm7 ## constant 11011101 ## table ready in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb300nf_qq(%esp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul - then we can get rid of mm5 ## update vctot addps nb300nf_vctot(%esp),%xmm5 movaps %xmm5,nb300nf_vctot(%esp)_nb_kernel300nf_ia32_sse.nb300nf_checksingle: movl nb300nf_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel300nf_ia32_sse.nb300nf_dosingle jmp _nb_kernel300nf_ia32_sse.nb300nf_updateouterdata_nb_kernel300nf_ia32_sse.nb300nf_dosingle: movl nb300nf_charge(%ebp),%esi movl nb300nf_pos(%ebp),%edi movl nb300nf_innerjjnr(%esp),%ecx movl (%ecx),%eax xorps %xmm6,%xmm6 movss (%esi,%eax,4),%xmm6 ## xmm6(0) has the charge mulps nb300nf_iq(%esp),%xmm6 movaps %xmm6,nb300nf_qq(%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 movaps nb300nf_ix(%esp),%xmm4 movaps nb300nf_iy(%esp),%xmm5 movaps nb300nf_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 nb300nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb300nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r mulps nb300nf_tsc(%esp),%xmm4 cvttps2pi %xmm4,%mm6 ## mm6 contain lu indices cvtpi2ps %mm6,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 movl nb300nf_VFtab(%ebp),%esi movd %mm6,%ebx movlps (%esi,%ebx,4),%xmm4 movlps 8(%esi,%ebx,4),%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $1,%xmm5,%xmm5 shufps $1,%xmm7,%xmm7 ## table ready in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb300nf_qq(%esp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul - then we can get rid of mm5 ## update vctot addss nb300nf_vctot(%esp),%xmm5 movss %xmm5,nb300nf_vctot(%esp)_nb_kernel300nf_ia32_sse.nb300nf_updateouterdata: ## get n from stack movl nb300nf_n(%esp),%esi ## get group index for i particle movl nb300nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb300nf_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 nb300nf_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb300nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel300nf_ia32_sse.nb300nf_outerend ## not last, iterate outer loop once more! movl %esi,nb300nf_n(%esp) jmp _nb_kernel300nf_ia32_sse.nb300nf_outer_nb_kernel300nf_ia32_sse.nb300nf_outerend: ## check if more outer neighborlists remain movl nb300nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel300nf_ia32_sse.nb300nf_end ## non-zero, do one more workunit jmp _nb_kernel300nf_ia32_sse.nb300nf_threadloop_nb_kernel300nf_ia32_sse.nb300nf_end: emms movl nb300nf_nouter(%esp),%eax movl nb300nf_ninner(%esp),%ebx movl nb300nf_outeriter(%ebp),%ecx movl nb300nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb300nf_salign(%esp),%eax addl %eax,%esp addl $188,%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?