nb_kernel332_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,234 行 · 第 1/5 页
S
2,234 行
movss 28(%rsi,%rbx,4),%xmm7 ## dispersion table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb332_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV xorps %xmm4,%xmm4 movss nb332_c6(%rsp),%xmm4 mulps %xmm4,%xmm7 ## fijD mulps %xmm4,%xmm5 ## Vvdw6 addps nb332_fstmp(%rsp),%xmm7 ## add to fscal ## put scalar force on stack Update Vvdwtot directly addps nb332_Vvdwtot(%rsp),%xmm5 movaps %xmm7,nb332_fstmp(%rsp) movaps %xmm5,nb332_Vvdwtot(%rsp) ## repulsion movss 32(%rsi,%rbx,4),%xmm4 movss 36(%rsi,%rbx,4),%xmm5 movss 40(%rsi,%rbx,4),%xmm6 movss 44(%rsi,%rbx,4),%xmm7 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb332_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV xorps %xmm4,%xmm4 movss nb332_c12(%rsp),%xmm4 mulps %xmm4,%xmm7 ## fijR mulps %xmm4,%xmm5 ## Vvdw12 addps nb332_fstmp(%rsp),%xmm7 addps nb332_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb332_Vvdwtot(%rsp) xorps %xmm1,%xmm1 mulps nb332_tsc(%rsp),%xmm7 mulps %xmm0,%xmm7 subps %xmm7,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 mulps nb332_dxOO(%rsp),%xmm0 mulps nb332_dyOO(%rsp),%xmm1 mulps nb332_dzOO(%rsp),%xmm2 ## initial update for j forces xorps %xmm3,%xmm3 xorps %xmm4,%xmm4 xorps %xmm5,%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb332_fjxO(%rsp) movaps %xmm4,nb332_fjyO(%rsp) movaps %xmm5,nb332_fjzO(%rsp) addps nb332_fixO(%rsp),%xmm0 addps nb332_fiyO(%rsp),%xmm1 addps nb332_fizO(%rsp),%xmm2 movaps %xmm0,nb332_fixO(%rsp) movaps %xmm1,nb332_fiyO(%rsp) movaps %xmm2,nb332_fizO(%rsp) ## done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps nb332_jxO(%rsp),%xmm0 movaps nb332_jyO(%rsp),%xmm1 movaps nb332_jzO(%rsp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb332_ixH1(%rsp),%xmm0 subps nb332_iyH1(%rsp),%xmm1 subps nb332_izH1(%rsp),%xmm2 subps nb332_ixH2(%rsp),%xmm3 subps nb332_iyH2(%rsp),%xmm4 subps nb332_izH2(%rsp),%xmm5 movaps %xmm0,nb332_dxH1O(%rsp) movaps %xmm1,nb332_dyH1O(%rsp) movaps %xmm2,nb332_dzH1O(%rsp) movaps %xmm3,nb332_dxH2O(%rsp) movaps %xmm4,nb332_dyH2O(%rsp) movaps %xmm5,nb332_dzH2O(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm3,%xmm4 addps %xmm2,%xmm0 ## have rsqH1 in xmm0 addps %xmm5,%xmm4 ## have rsqH2 in xmm4 ## start with H1, save H2 data movaps %xmm4,nb332_rsqH2O(%rsp) ## do invsqrt rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb332_three(%rsp),%xmm3 movaps %xmm3,%xmm7 mulps %xmm0,%xmm1 mulps %xmm4,%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb332_half(%rsp),%xmm3 ## rinv H1 - j water mulps nb332_half(%rsp),%xmm7 ## rinv H2 - j water ## start with H1, save H2 data movaps %xmm7,nb332_rinvH2O(%rsp) movaps %xmm3,%xmm1 mulps %xmm0,%xmm1 ## xmm1=r movaps %xmm3,%xmm0 ## xmm0=rinv mulps nb332_tsc(%rsp),%xmm1 movhlps %xmm1,%xmm2 cvttps2pi %xmm1,%mm6 cvttps2pi %xmm2,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm2 movlhps %xmm2,%xmm3 subps %xmm3,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx movlps (%rsi,%rbx,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rbx,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rdx,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 nb332_two(%rsp),%xmm7 ## two*Heps2 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb332_qqOH(%rsp),%xmm3 movhps nb332_qqHH(%rsp),%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 ## at this point xmm5 contains vcoul and xmm3 fijC addps nb332_vctot(%rsp),%xmm5 movaps %xmm5,nb332_vctot(%rsp) xorps %xmm1,%xmm1 mulps nb332_tsc(%rsp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 mulps nb332_dxH1O(%rsp),%xmm0 mulps nb332_dyH1O(%rsp),%xmm1 mulps nb332_dzH1O(%rsp),%xmm2 ## update forces H1 - j water movaps nb332_fjxO(%rsp),%xmm3 movaps nb332_fjyO(%rsp),%xmm4 movaps nb332_fjzO(%rsp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb332_fjxO(%rsp) movaps %xmm4,nb332_fjyO(%rsp) movaps %xmm5,nb332_fjzO(%rsp) addps nb332_fixH1(%rsp),%xmm0 addps nb332_fiyH1(%rsp),%xmm1 addps nb332_fizH1(%rsp),%xmm2 movaps %xmm0,nb332_fixH1(%rsp) movaps %xmm1,nb332_fiyH1(%rsp) movaps %xmm2,nb332_fizH1(%rsp) ## do table for H2 - j water interaction movaps nb332_rinvH2O(%rsp),%xmm0 movaps nb332_rsqH2O(%rsp),%xmm1 mulps %xmm0,%xmm1 ## xmm0=rinv, xmm1=r mulps nb332_tsc(%rsp),%xmm1 movhlps %xmm1,%xmm2 cvttps2pi %xmm1,%mm6 cvttps2pi %xmm2,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm2 movlhps %xmm2,%xmm3 subps %xmm3,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx movlps (%rsi,%rbx,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rbx,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rdx,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 nb332_two(%rsp),%xmm7 ## two*Heps2 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb332_qqOH(%rsp),%xmm3 movhps nb332_qqHH(%rsp),%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 ## at this point xmm5 contains vcoul and xmm3 fijC addps nb332_vctot(%rsp),%xmm5 movaps %xmm5,nb332_vctot(%rsp) xorps %xmm1,%xmm1 mulps nb332_tsc(%rsp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 mulps nb332_dxH2O(%rsp),%xmm0 mulps nb332_dyH2O(%rsp),%xmm1 mulps nb332_dzH2O(%rsp),%xmm2 movaps nb332_fjxO(%rsp),%xmm3 movaps nb332_fjyO(%rsp),%xmm4 movaps nb332_fjzO(%rsp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movq nb332_faction(%rbp),%rsi movaps %xmm3,nb332_fjxO(%rsp) movaps %xmm4,nb332_fjyO(%rsp) movaps %xmm5,nb332_fjzO(%rsp) addps nb332_fixH2(%rsp),%xmm0 addps nb332_fiyH2(%rsp),%xmm1 addps nb332_fizH2(%rsp),%xmm2 movaps %xmm0,nb332_fixH2(%rsp) movaps %xmm1,nb332_fiyH2(%rsp) movaps %xmm2,nb332_fizH2(%rsp) ## update j water forces from local variables movlps (%rsi,%rax,4),%xmm0 movlps 12(%rsi,%rax,4),%xmm1 movhps 24(%rsi,%rax,4),%xmm1 movaps nb332_fjxO(%rsp),%xmm3 movaps nb332_fjyO(%rsp),%xmm4 movaps nb332_fjzO(%rsp),%xmm5 movaps %xmm5,%xmm6 movaps %xmm5,%xmm7 shufps $2,%xmm6,%xmm6 ## 00000010 shufps $3,%xmm7,%xmm7 ## 00000011 addss 8(%rsi,%rax,4),%xmm5 addss 20(%rsi,%rax,4),%xmm6 addss 32(%rsi,%rax,4),%xmm7 movss %xmm5,8(%rsi,%rax,4) movss %xmm6,20(%rsi,%rax,4) movss %xmm7,32(%rsi,%rax,4) movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movlps %xmm0,(%rsi,%rax,4) movlps %xmm1,12(%rsi,%rax,4) movhps %xmm1,24(%rsi,%rax,4) decl nb332_innerk(%rsp) jz _nb_kernel332_x86_64_sse.nb332_updateouterdata jmp _nb_kernel332_x86_64_sse.nb332_single_loop_nb_kernel332_x86_64_sse.nb332_updateouterdata: movl nb332_ii3(%rsp),%ecx movq nb332_faction(%rbp),%rdi movq nb332_fshift(%rbp),%rsi movl nb332_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb332_fixO(%rsp),%xmm0 movaps nb332_fiyO(%rsp),%xmm1 movaps nb332_fizO(%rsp),%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 (%rdi,%rcx,4),%xmm3 movss 4(%rdi,%rcx,4),%xmm4 movss 8(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%rdi,%rcx,4) movss %xmm4,4(%rdi,%rcx,4) movss %xmm5,8(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb332_fixH1(%rsp),%xmm0 movaps nb332_fiyH1(%rsp),%xmm1 movaps nb332_fizH1(%rsp),%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(%rdi,%rcx,4),%xmm3 movss 16(%rdi,%rcx,4),%xmm4 movss 20(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,12(%rdi,%rcx,4) movss %xmm4,16(%rdi,%rcx,4) movss %xmm5,20(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## 00001000 addps %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb332_fixH2(%rsp),%xmm0 movaps nb332_fiyH2(%rsp),%xmm1 movaps nb332_fizH2(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 i 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 ⌨️ 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?