nb_kernel303_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,963 行 · 第 1/5 页
S
1,963 行
## rsqM - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb303nf_three(%esp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb303nf_half(%esp),%xmm4 movaps %xmm4,nb303nf_rinvM(%esp) ## rinvM in xmm4 mulps %xmm4,%xmm5 movaps %xmm5,nb303nf_rM(%esp) ## do H1 interactions ## rH1 is still in xmm7 mulps nb303nf_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%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 nb303nf_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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb303nf_qqH(%esp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%esp),%xmm5 movaps %xmm5,nb303nf_vctot(%esp) ## Done with H1 interactions - now H2! movaps nb303nf_rH2(%esp),%xmm7 mulps nb303nf_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb303nf_qqH(%esp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%esp),%xmm5 movaps %xmm5,nb303nf_vctot(%esp) ## Done with H2, finally we do M interactions movaps nb303nf_rM(%esp),%xmm7 mulps nb303nf_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb303nf_qqM(%esp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%esp),%xmm5 movaps %xmm5,nb303nf_vctot(%esp) ## should we do one more iteration? subl $4,nb303nf_innerk(%esp) jl _nb_kernel303nf_ia32_sse.nb303nf_odd_inner jmp _nb_kernel303nf_ia32_sse.nb303nf_unroll_loop_nb_kernel303nf_ia32_sse.nb303nf_odd_inner: addl $4,nb303nf_innerk(%esp) jnz _nb_kernel303nf_ia32_sse.nb303nf_odd_loop jmp _nb_kernel303nf_ia32_sse.nb303nf_updateouterdata_nb_kernel303nf_ia32_sse.nb303nf_odd_loop: movl nb303nf_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb303nf_innerjjnr(%esp) xorps %xmm4,%xmm4 movss nb303nf_iqM(%esp),%xmm4 movl nb303nf_charge(%ebp),%esi movhps nb303nf_iqH(%esp),%xmm4 movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb303nf_qqM(%esp) ## use dummy qq for storage movl nb303nf_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## move j coords to xmm0-xmm2 movss (%esi,%eax,4),%xmm0 movss 4(%esi,%eax,4),%xmm1 movss 8(%esi,%eax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movss nb303nf_ixM(%esp),%xmm3 movss nb303nf_iyM(%esp),%xmm4 movss nb303nf_izM(%esp),%xmm5 movlps nb303nf_ixH1(%esp),%xmm6 movlps nb303nf_ixH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb303nf_iyH1(%esp),%xmm6 movlps nb303nf_iyH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb303nf_izH1(%esp),%xmm6 movlps nb303nf_izH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb303nf_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb303nf_half(%esp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv ## a little trick to avoid NaNs: ## positions 0,2,and 3 are valid, but not 1. ## If it contains NaN it doesnt help to mult by 0, ## So we shuffle it and copy pos 0 to pos1! shufps $224,%xmm0,%xmm0 ## 11100000 mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm0,nb303nf_rinvM(%esp) mulps nb303nf_tsc(%esp),%xmm4 movhlps %xmm4,%xmm7 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm7,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm7 movlhps %xmm7,%xmm3 subps %xmm3,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ecx,%mm1 movd %edx,%mm2 movl nb303nf_VFtab(%ebp),%esi movd %mm6,%eax movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx movlps (%esi,%eax,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 movhps (%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%esi,%eax,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 movhps 8(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb303nf_qqM(%esp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%esp),%xmm5 movaps %xmm5,nb303nf_vctot(%esp) decl nb303nf_innerk(%esp) jz _nb_kernel303nf_ia32_sse.nb303nf_updateouterdata jmp _nb_kernel303nf_ia32_sse.nb303nf_odd_loop_nb_kernel303nf_ia32_sse.nb303nf_updateouterdata: ## get n from stack movl nb303nf_n(%esp),%esi ## get group index for i particle movl nb303nf_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb303nf_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 nb303nf_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb303nf_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel303nf_ia32_sse.nb303nf_outerend ## not last, iterate outer loop once more! movl %esi,nb303nf_n(%esp) jmp _nb_kernel303nf_ia32_sse.nb303nf_outer_nb_kernel303nf_ia32_sse.nb303nf_outerend: ## check if more outer neighborlists remain movl nb303nf_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel303nf_ia32_sse.nb303nf_end ## non-zero, do one more workunit jmp _nb_kernel303nf_ia32_sse.nb303nf_threadloop_nb_kernel303nf_ia32_sse.nb303nf_end: emms movl nb303nf_nouter(%esp),%eax movl nb303nf_ninner(%esp),%ebx movl nb303nf_outeriter(%ebp),%ecx movl nb303nf_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb303nf_salign(%esp),%eax addl %eax,%esp addl $408,%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?