nb_kernel430_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,138 行 · 第 1/5 页
S
2,138 行
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 nb430_GBtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx ## load coulomb table movaps (%esi,%eax,4),%xmm4 movaps (%esi,%ebx,4),%xmm5 movaps (%esi,%ecx,4),%xmm6 movaps (%esi,%edx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm6,%xmm3 shufps $0xEE,%xmm7,%xmm3 shufps $0x44,%xmm7,%xmm6 movaps %xmm4,%xmm7 shufps $0xEE,%xmm5,%xmm7 shufps $0x44,%xmm5,%xmm4 movaps %xmm4,%xmm5 shufps $0xDD,%xmm6,%xmm5 shufps $0x88,%xmm6,%xmm4 movaps %xmm7,%xmm6 shufps $0x88,%xmm3,%xmm6 shufps $0xDD,%xmm3,%xmm7 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb430_two(%esp),%xmm7 ## two*Heps2 movaps nb430_qq(%esp),%xmm3 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm3 ## fijC=FF*qq ## get jnr from stack movl nb430_jnra(%esp),%eax movl nb430_jnrb(%esp),%ebx movl nb430_jnrc(%esp),%ecx movl nb430_jnrd(%esp),%edx movl nb430_dvda(%ebp),%esi ## Calculate dVda xorps %xmm7,%xmm7 mulps nb430_gbscale(%esp),%xmm3 movaps %xmm3,%xmm6 mulps nb430_r(%esp),%xmm6 addps %xmm5,%xmm6 addps nb430_vctot(%esp),%xmm5 movaps %xmm5,nb430_vctot(%esp) ## xmm6=(vcoul+fijC*r) subps %xmm6,%xmm7 movaps %xmm7,%xmm6 ## update dvdasum addps nb430_dvdasum(%esp),%xmm7 movaps %xmm7,nb430_dvdasum(%esp) ## update j atoms dvdaj movhlps %xmm6,%xmm7 movaps %xmm6,%xmm5 movaps %xmm7,%xmm4 shufps $0x1,%xmm5,%xmm5 shufps $0x1,%xmm4,%xmm4 ## xmm6=dvdaj1 xmm5=dvdaj2 xmm7=dvdaj3 xmm4=dvdaj4 addss (%esi,%eax,4),%xmm6 addss (%esi,%ebx,4),%xmm5 addss (%esi,%ecx,4),%xmm7 addss (%esi,%edx,4),%xmm4 movss %xmm6,(%esi,%eax,4) movss %xmm5,(%esi,%ebx,4) movss %xmm7,(%esi,%ecx,4) movss %xmm4,(%esi,%edx,4) ## put scalar force on stack temporarily movaps %xmm3,nb430_fscal(%esp) movaps nb430_r(%esp),%xmm4 mulps nb430_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 $3,%mm6 pslld $3,%mm7 movl nb430_VFtab(%ebp),%esi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx ## dispersion movaps (%esi,%eax,4),%xmm4 movaps (%esi,%ebx,4),%xmm5 movaps (%esi,%ecx,4),%xmm6 movaps (%esi,%edx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm6,%xmm3 shufps $0xEE,%xmm7,%xmm3 shufps $0x44,%xmm7,%xmm6 movaps %xmm4,%xmm7 shufps $0xEE,%xmm5,%xmm7 shufps $0x44,%xmm5,%xmm4 movaps %xmm4,%xmm5 shufps $0xDD,%xmm6,%xmm5 shufps $0x88,%xmm6,%xmm4 movaps %xmm7,%xmm6 shufps $0x88,%xmm3,%xmm6 shufps $0xDD,%xmm3,%xmm7 ## dispersion table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb430_two(%esp),%xmm7 ## two*Heps2 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb430_c6(%esp),%xmm4 mulps %xmm4,%xmm7 ## fijD mulps %xmm4,%xmm5 ## Vvdw6 mulps nb430_tsc(%esp),%xmm7 addps nb430_fscal(%esp),%xmm7 ## add to fscal ## put scalar force on stack Update Vvdwtot directly addps nb430_Vvdwtot(%esp),%xmm5 movaps %xmm7,nb430_fscal(%esp) movaps %xmm5,nb430_Vvdwtot(%esp) ## repulsion movaps 16(%esi,%eax,4),%xmm4 movaps 16(%esi,%ebx,4),%xmm5 movaps 16(%esi,%ecx,4),%xmm6 movaps 16(%esi,%edx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm6,%xmm3 shufps $0xEE,%xmm7,%xmm3 shufps $0x44,%xmm7,%xmm6 movaps %xmm4,%xmm7 shufps $0xEE,%xmm5,%xmm7 shufps $0x44,%xmm5,%xmm4 movaps %xmm4,%xmm5 shufps $0xDD,%xmm6,%xmm5 shufps $0x88,%xmm6,%xmm4 movaps %xmm7,%xmm6 shufps $0x88,%xmm3,%xmm6 shufps $0xDD,%xmm3,%xmm7 ## table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb430_two(%esp),%xmm7 ## two*Heps2 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb430_c12(%esp),%xmm4 mulps %xmm4,%xmm7 ## fijR mulps %xmm4,%xmm5 ## Vvdw12 mulps nb430_tsc(%esp),%xmm7 addps nb430_fscal(%esp),%xmm7 addps nb430_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb430_Vvdwtot(%esp) xorps %xmm4,%xmm4 mulps %xmm0,%xmm7 subps %xmm7,%xmm4 movaps nb430_dx(%esp),%xmm0 movaps nb430_dy(%esp),%xmm1 movaps nb430_dz(%esp),%xmm2 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movl nb430_faction(%ebp),%edi mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movaps nb430_fix(%esp),%xmm3 movaps nb430_fiy(%esp),%xmm4 movaps nb430_fiz(%esp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb430_fix(%esp) movaps %xmm4,nb430_fiy(%esp) movaps %xmm5,nb430_fiz(%esp) ## the fj's - start by accumulating x & y forces from memory movlps (%edi,%eax,4),%xmm4 movlps (%edi,%ecx,4),%xmm6 movhps (%edi,%ebx,4),%xmm4 movhps (%edi,%edx,4),%xmm6 movaps %xmm4,%xmm3 shufps $136,%xmm6,%xmm3 ## constant 10001000 shufps $221,%xmm6,%xmm4 ## constant 11011101 ## now xmm3-xmm5 contains fjx, fjy, fjz subps %xmm0,%xmm3 subps %xmm1,%xmm4 ## unpack them back so we can store them - first x & y in xmm3/xmm4 movaps %xmm3,%xmm6 unpcklps %xmm4,%xmm6 unpckhps %xmm4,%xmm3 ## xmm6(l)=x & y for j1, (h) for j2 ## xmm3(l)=x & y for j3, (h) for j4 movlps %xmm6,(%edi,%eax,4) movlps %xmm3,(%edi,%ecx,4) movhps %xmm6,(%edi,%ebx,4) movhps %xmm3,(%edi,%edx,4) ## and the z forces movss 8(%edi,%eax,4),%xmm4 movss 8(%edi,%ebx,4),%xmm5 movss 8(%edi,%ecx,4),%xmm6 movss 8(%edi,%edx,4),%xmm7 subss %xmm2,%xmm4 shufps $229,%xmm2,%xmm2 ## constant 11100101 subss %xmm2,%xmm5 shufps $234,%xmm2,%xmm2 ## constant 11101010 subss %xmm2,%xmm6 shufps $255,%xmm2,%xmm2 ## constant 11111111 subss %xmm2,%xmm7 movss %xmm4,8(%edi,%eax,4) movss %xmm5,8(%edi,%ebx,4) movss %xmm6,8(%edi,%ecx,4) movss %xmm7,8(%edi,%edx,4) ## should we do one more iteration? subl $4,nb430_innerk(%esp) jl _nb_kernel430_ia32_sse.nb430_finish_inner jmp _nb_kernel430_ia32_sse.nb430_unroll_loop_nb_kernel430_ia32_sse.nb430_finish_inner: ## check if at least two particles remain addl $4,nb430_innerk(%esp) movl nb430_innerk(%esp),%edx andl $2,%edx jnz _nb_kernel430_ia32_sse.nb430_dopair jmp _nb_kernel430_ia32_sse.nb430_checksingle_nb_kernel430_ia32_sse.nb430_dopair: movl nb430_innerjjnr(%esp),%ecx movl (%ecx),%eax movl 4(%ecx),%ebx addl $8,nb430_innerjjnr(%esp) xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 ## load isaj movl nb430_invsqrta(%ebp),%esi movss (%esi,%eax,4),%xmm2 movss (%esi,%ebx,4),%xmm3 unpcklps %xmm3,%xmm2 ## isaj in xmm3(0,1) mulps nb430_isai(%esp),%xmm2 movaps %xmm2,nb430_isaprod(%esp) movaps %xmm2,%xmm1 mulps nb430_gbtsc(%esp),%xmm1 movaps %xmm1,nb430_gbscale(%esp) movl nb430_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 nb430_iq(%esp),%xmm2 mulps %xmm2,%xmm3 movaps %xmm3,nb430_qq(%esp) movl nb430_type(%ebp),%esi movl %eax,%ecx movl %ebx,%edx movl (%esi,%ecx,4),%ecx movl (%esi,%edx,4),%edx movl nb430_vdwparam(%ebp),%esi shll %ecx shll %edx movl nb430_ntia(%esp),%edi addl %edi,%ecx addl %edi,%edx movlps (%esi,%ecx,4),%xmm6 movhps (%esi,%edx,4),%xmm6 movl nb430_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,nb430_c6(%esp) movaps %xmm6,nb430_c12(%esp) movd %eax,%mm0 ## copy jnr to mm0/mm1 movd %ebx,%mm1 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 nb430_faction(%ebp),%edi ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb430_ix(%esp),%xmm4 movaps nb430_iy(%esp),%xmm5 movaps nb430_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb430_dx(%esp) movaps %xmm5,nb430_dy(%esp) movaps %xmm6,nb430_dz(%esp) ## 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 nb430_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb430_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm4,nb430_r(%esp) mulps nb430_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 nb430_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 mulps nb430_two(%esp),%xmm7 ## two*Heps2 movaps nb430_qq(%esp),%xmm3⌨️ 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?