nb_kernel410_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,024 行 · 第 1/5 页
S
2,024 行
addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb410nf_qq(%esp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV ## update vctot addps nb410nf_vctot(%esp),%xmm5 movaps %xmm5,nb410nf_vctot(%esp) ## L-J movaps %xmm0,%xmm4 mulps %xmm0,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm6 mulps %xmm4,%xmm6 mulps %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvtwelve mulps nb410nf_c6(%esp),%xmm6 mulps nb410nf_c12(%esp),%xmm4 movaps nb410nf_Vvdwtot(%esp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movaps %xmm7,nb410nf_Vvdwtot(%esp) ## should we do one more iteration? subl $4,nb410nf_innerk(%esp) jl _nb_kernel410nf_ia32_sse.nb410nf_finish_inner jmp _nb_kernel410nf_ia32_sse.nb410nf_unroll_loop_nb_kernel410nf_ia32_sse.nb410nf_finish_inner: ## check if at least two particles remain addl $4,nb410nf_innerk(%esp) movl nb410nf_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel410nf_ia32_sse.nb410nf_dopair jmp _nb_kernel410nf_ia32_sse.nb410nf_checksingle_nb_kernel410nf_ia32_sse.nb410nf_dopair: movl nb410nf_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb410nf_innerjjnr(%esp) xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 ## load isa2 movl nb410nf_invsqrta(%ebp),%esi movss (%esi,%eax,4),%xmm2 movss (%esi,%ebx,4),%xmm3 unpcklps %xmm3,%xmm2 ## isa2 in xmm3(0,1) mulps nb410nf_isai(%esp),%xmm2 movaps %xmm2,nb410nf_isaprod(%esp) movaps %xmm2,%xmm1 mulps nb410nf_gbtsc(%esp),%xmm1 movaps %xmm1,nb410nf_gbscale(%esp) movl nb410nf_charge(%ebp),%esi ## base of charge[] movss (%esi,%eax,4),%xmm3 movss (%esi,%ebx,4),%xmm6 unpcklps %xmm6,%xmm3 ## constant 00001000 ;# xmm3(0,1) has the charges mulps nb410nf_iq(%esp),%xmm2 mulps %xmm2,%xmm3 movaps %xmm3,nb410nf_qq(%esp) movl nb410nf_type(%ebp),%esi movl %eax,%ecx movl %ebx,%edx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb410nf_vdwparam(%ebp),%esi shll %ecx shll %edx movl nb410nf_ntia(%esp),%edi addl %edi,%ecx addl %edi,%edx movlps (%esi,%ecx,4),%xmm6 movhps (%esi,%edx,4),%xmm6 movl nb410nf_pos(%ebp),%edi movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## constant 00001000 shufps $13,%xmm6,%xmm6 ## constant 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb410nf_c6(%esp) movaps %xmm6,nb410nf_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 movl nb410nf_faction(%ebp),%edi ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb410nf_ix(%esp),%xmm4 movaps nb410nf_iy(%esp),%xmm5 movaps nb410nf_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 nb410nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb410nf_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 nb410nf_gbscale(%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 nb410nf_GBtab(%ebp),%esi movd %mm6,%ecx psrlq $32,%mm6 movd %mm6,%edx ## load coulomb table movaps (%esi,%ecx,4),%xmm4 movaps (%esi,%edx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm4,%xmm6 unpcklps %xmm7,%xmm4 ## Y1 Y2 F1 F2 unpckhps %xmm7,%xmm6 ## G1 G2 H1 H2 movhlps %xmm4,%xmm5 ## F1 F2 movhlps %xmm6,%xmm7 ## H1 H2 ## 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 nb410nf_qq(%esp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV addps nb410nf_vctot(%esp),%xmm5 movaps %xmm5,nb410nf_vctot(%esp) ## L-J movaps %xmm0,%xmm4 mulps %xmm0,%xmm4 ## xmm4=rinvsq ## at this point mm5 contains vcoul and mm3 fijC ## increment vcoul - then we can get rid of mm5 ## update vctot movaps %xmm4,%xmm6 mulps %xmm4,%xmm6 mulps %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvtwelve mulps nb410nf_c6(%esp),%xmm6 mulps nb410nf_c12(%esp),%xmm4 movaps nb410nf_Vvdwtot(%esp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movaps %xmm7,nb410nf_Vvdwtot(%esp)_nb_kernel410nf_ia32_sse.nb410nf_checksingle: movl nb410nf_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel410nf_ia32_sse.nb410nf_dosingle jmp _nb_kernel410nf_ia32_sse.nb410nf_updateouterdata_nb_kernel410nf_ia32_sse.nb410nf_dosingle: movl nb410nf_charge(%ebp),%esi movl nb410nf_invsqrta(%ebp),%edx movl nb410nf_pos(%ebp),%edi movl nb410nf_innerjjnr(%esp),%ecx movl (%ecx),%eax xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 movss (%edx,%eax,4),%xmm2 ## isa2 mulss nb410nf_isai(%esp),%xmm2 movss %xmm2,nb410nf_isaprod(%esp) movss %xmm2,%xmm1 mulss nb410nf_gbtsc(%esp),%xmm1 movss %xmm1,nb410nf_gbscale(%esp) mulss nb410nf_iq(%esp),%xmm2 movss (%esi,%eax,4),%xmm6 ## xmm6(0) has the charge mulss %xmm2,%xmm6 movss %xmm6,nb410nf_qq(%esp) movl nb410nf_type(%ebp),%esi movl %eax,%ecx movl (%esi,%ecx,4),%ecx movl nb410nf_vdwparam(%ebp),%esi shll %ecx addl nb410nf_ntia(%esp),%ecx movlps (%esi,%ecx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## constant 11111100 shufps $253,%xmm6,%xmm6 ## constant 11111101 movaps %xmm4,nb410nf_c6(%esp) movaps %xmm6,nb410nf_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 movaps nb410nf_ix(%esp),%xmm4 movaps nb410nf_iy(%esp),%xmm5 movaps nb410nf_iz(%esp),%xmm6 ## calc dr subss %xmm0,%xmm4 subss %xmm1,%xmm5 subss %xmm2,%xmm6 ## square it mulss %xmm4,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm6 addss %xmm5,%xmm4 addss %xmm6,%xmm4 ## rsq in xmm4 rsqrtss %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulss %xmm5,%xmm5 movss nb410nf_three(%esp),%xmm1 mulss %xmm4,%xmm5 ## rsq*lu*lu movss nb410nf_half(%esp),%xmm0 subss %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulss %xmm2,%xmm1 mulss %xmm1,%xmm0 ## xmm0=rinv mulss %xmm0,%xmm4 ## xmm4=r mulss nb410nf_gbscale(%esp),%xmm4 cvttss2si %xmm4,%ebx ## mm6 contain lu indices cvtsi2ss %ebx,%xmm6 subss %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 shll $2,%ebx movl nb410nf_GBtab(%ebp),%esi movaps (%esi,%ebx,4),%xmm4 movhlps %xmm4,%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $1,%xmm5,%xmm5 shufps $1,%xmm7,%xmm7 ## table ready in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp movss nb410nf_qq(%esp),%xmm3 mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV mulss %xmm3,%xmm5 ## vcoul=qq*VV addss nb410nf_vctot(%esp),%xmm5 movss %xmm5,nb410nf_vctot(%esp) ## L-J movaps %xmm0,%xmm4 mulss %xmm0,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm6 mulss %xmm4,%xmm6 mulss %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulss %xmm4,%xmm4 ## xmm4=rinvtwelve mulss nb410nf_c6(%esp),%xmm6 mulss nb410nf_c12(%esp),%xmm4 movss nb410nf_Vvdwtot(%esp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movss %xmm7,nb410nf_Vvdwtot(%esp)_nb_kernel410nf_ia32_sse.nb410nf_updateouterdata: ## get n from stack movl nb410nf_n(%esp),%esi ## get group index for i particle movl nb410nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb410nf_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 nb410nf_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## accumulate total lj energy and update it movaps nb410nf_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 nb410nf_Vvdw(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb410nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel410nf_ia32_sse.nb410nf_outerend ## not last, iterate outer loop once more! movl %esi,nb410nf_n(%esp) jmp _nb_kernel410nf_ia32_sse.nb410nf_outer_nb_kernel410nf_ia32_sse.nb410nf_outerend: ## check if more outer neighborlists remain movl nb410nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel410nf_ia32_sse.nb410nf_end ## non-zero, do one more workunit jmp _nb_kernel410nf_ia32_sse.nb410nf_threadloop_nb_kernel410nf_ia32_sse.nb410nf_end: emms movl nb410nf_nouter(%esp),%eax movl nb410nf_ninner(%esp),%ebx movl nb410nf_outeriter(%ebp),%ecx movl nb410nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb410nf_salign(%esp),%eax addl %eax,%esp addl $292,%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?