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
xorps %xmm4,%xmm4 movss nb303_iqM(%esp),%xmm4 movl nb303_charge(%ebp),%esi movhps nb303_iqH(%esp),%xmm4 movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb303_qqM(%esp) ## use dummy qq for storage movl nb303_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 nb303_ixM(%esp),%xmm3 movss nb303_iyM(%esp),%xmm4 movss nb303_izM(%esp),%xmm5 movlps nb303_ixH1(%esp),%xmm6 movlps nb303_ixH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb303_iyH1(%esp),%xmm6 movlps nb303_iyH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb303_izH1(%esp),%xmm6 movlps nb303_izH2(%esp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use dummy dx for storage movaps %xmm3,nb303_dxM(%esp) movaps %xmm4,nb303_dyM(%esp) movaps %xmm5,nb303_dzM(%esp) 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 nb303_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb303_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,nb303_rinvM(%esp) mulps nb303_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 nb303_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 mulps nb303_two(%esp),%xmm7 ## two*Heps2 movaps nb303_qqM(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm0 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm0 fijC ## increment vcoul - then we can get rid of mm5 addps nb303_vctot(%esp),%xmm5 movaps %xmm5,nb303_vctot(%esp) xorps %xmm4,%xmm4 mulps nb303_tsc(%esp),%xmm0 mulps nb303_rinvM(%esp),%xmm0 subps %xmm0,%xmm4 movd %mm0,%eax movd %mm1,%ecx movd %mm2,%edx movaps nb303_dxM(%esp),%xmm0 movaps nb303_dyM(%esp),%xmm1 movaps nb303_dzM(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb303_fixM(%esp),%xmm3 movss nb303_fiyM(%esp),%xmm4 movss nb303_fizM(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb303_fixM(%esp) movss %xmm4,nb303_fiyM(%esp) movss %xmm5,nb303_fizM(%esp) ## updated the M force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## 11100110 shufps $230,%xmm5,%xmm5 ## 11100110 addss nb303_fixH1(%esp),%xmm3 addss nb303_fiyH1(%esp),%xmm4 addss nb303_fizH1(%esp),%xmm5 movss %xmm3,nb303_fixH1(%esp) movss %xmm4,nb303_fiyH1(%esp) movss %xmm5,nb303_fizH1(%esp) ## updated the H1 force movl nb303_faction(%ebp),%edi shufps $231,%xmm3,%xmm3 ## 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## 11100111 shufps $231,%xmm5,%xmm5 ## 11100111 addss nb303_fixH2(%esp),%xmm3 addss nb303_fiyH2(%esp),%xmm4 addss nb303_fizH2(%esp),%xmm5 movss %xmm3,nb303_fixH2(%esp) movss %xmm4,nb303_fiyH2(%esp) movss %xmm5,nb303_fizH2(%esp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%edi,%eax,4),%xmm6 movss 8(%edi,%eax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 subps %xmm0,%xmm6 subss %xmm2,%xmm7 movlps %xmm6,(%edi,%eax,4) movss %xmm7,8(%edi,%eax,4) decl nb303_innerk(%esp) jz _nb_kernel303_ia32_sse.nb303_updateouterdata jmp _nb_kernel303_ia32_sse.nb303_odd_loop_nb_kernel303_ia32_sse.nb303_updateouterdata: movl nb303_ii3(%esp),%ecx movl nb303_faction(%ebp),%edi movl nb303_fshift(%ebp),%esi movl nb303_is3(%esp),%edx ## accumulate H1 i forces in xmm0, xmm1, xmm2 movaps nb303_fixH1(%esp),%xmm0 movaps nb303_fiyH1(%esp),%xmm1 movaps nb303_fizH1(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 12(%edi,%ecx,4),%xmm3 movss 16(%edi,%ecx,4),%xmm4 movss 20(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,12(%edi,%ecx,4) movss %xmm4,16(%edi,%ecx,4) movss %xmm5,20(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## constant 00001000 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb303_fixH2(%esp),%xmm0 movaps nb303_fiyH2(%esp),%xmm1 movaps nb303_fizH2(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 24(%edi,%ecx,4),%xmm3 movss 28(%edi,%ecx,4),%xmm4 movss 32(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,24(%edi,%ecx,4) movss %xmm4,28(%edi,%ecx,4) movss %xmm5,32(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate M i forces in xmm0, xmm1, xmm2 movaps nb303_fixM(%esp),%xmm0 movaps nb303_fiyM(%esp),%xmm1 movaps nb303_fizM(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss 36(%edi,%ecx,4),%xmm3 movss 40(%edi,%ecx,4),%xmm4 movss 44(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,36(%edi,%ecx,4) movss %xmm4,40(%edi,%ecx,4) movss %xmm5,44(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## increment fshift force movlps (%esi,%edx,4),%xmm3 movss 8(%esi,%edx,4),%xmm4 addps %xmm6,%xmm3 addss %xmm7,%xmm4 movlps %xmm3,(%esi,%edx,4) movss %xmm4,8(%esi,%edx,4) ## get n from stack movl nb303_n(%esp),%esi ## get group index for i particle movl nb303_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb303_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 nb303_Vc(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb303_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel303_ia32_sse.nb303_outerend ## not last, iterate outer loop once more! movl %esi,nb303_n(%esp) jmp _nb_kernel303_ia32_sse.nb303_outer_nb_kernel303_ia32_sse.nb303_outerend: ## check if more outer neighborlists remain movl nb303_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel303_ia32_sse.nb303_end ## non-zero, do one more workunit jmp _nb_kernel303_ia32_sse.nb303_threadloop_nb_kernel303_ia32_sse.nb303_end: emms movl nb303_nouter(%esp),%eax movl nb303_ninner(%esp),%ebx movl nb303_outeriter(%ebp),%ecx movl nb303_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb303_salign(%esp),%eax addl %eax,%esp addl $760,%esp popl %edi popl %esi popl %edx⌨️ 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?