nb_kernel233_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,128 行 · 第 1/5 页
S
2,128 行
movq nb233_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb233_vdwparam(%rbp),%rsi shll %ebx addl nb233_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb233_c6(%rsp) movaps %xmm7,nb233_c12(%rsp) movq nb233_pos(%rbp),%rsi lea (%rax,%rax,2),%rax movss nb233_ixO(%rsp),%xmm0 movss nb233_iyO(%rsp),%xmm1 movss nb233_izO(%rsp),%xmm2 movss nb233_ixH1(%rsp),%xmm3 movss nb233_iyH1(%rsp),%xmm4 movss nb233_izH1(%rsp),%xmm5 unpcklps nb233_ixH2(%rsp),%xmm0 ## ixO ixH2 - - unpcklps nb233_iyH2(%rsp),%xmm1 ## iyO iyH2 - - unpcklps nb233_izH2(%rsp),%xmm2 ## izO izH2 - - unpcklps nb233_ixM(%rsp),%xmm3 ## ixH1 ixM - - unpcklps nb233_iyM(%rsp),%xmm4 ## iyH1 iyM - - unpcklps nb233_izM(%rsp),%xmm5 ## izH1 izM - - unpcklps %xmm3,%xmm0 ## ixO ixH1 ixH2 ixM unpcklps %xmm4,%xmm1 ## same for y unpcklps %xmm5,%xmm2 ## same for z ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm3 movss 4(%rsi,%rax,4),%xmm4 movss 8(%rsi,%rax,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use O distances for storage movaps %xmm3,nb233_dxO(%rsp) movaps %xmm4,nb233_dyO(%rsp) movaps %xmm5,nb233_dzO(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 movaps %xmm4,%xmm0 mulps nb233_krf(%rsp),%xmm0 movaps %xmm0,nb233_krsqM(%rsp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb233_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb233_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv, xmm4=rsq mulps %xmm0,%xmm4 mulps nb233_tsc(%rsp),%xmm4 ## rtab cvttps2pi %xmm4,%mm6 cvtpi2ps %mm6,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 movd %eax,%mm0 movq nb233_VFtab(%rbp),%rsi movd %mm6,%eax ## dispersion movlps (%rsi,%rax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%rsi,%rax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## 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 nb233_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb233_c6(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijD mulss %xmm4,%xmm5 ## Vvdw6 mulss nb233_tsc(%rsp),%xmm7 ## put scalar force on stack Update Vvdwtot directly addss nb233_Vvdwtot(%rsp),%xmm5 movss %xmm7,nb233_fstmp(%rsp) movss %xmm5,nb233_Vvdwtot(%rsp) ## repulsion movlps 16(%rsi,%rax,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%rsi,%rax,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb233_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb233_c12(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijR mulss %xmm4,%xmm5 ## Vvdw12 mulss nb233_tsc(%rsp),%xmm7 addss nb233_fstmp(%rsp),%xmm7 movss %xmm7,nb233_fstmp(%rsp) addss nb233_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb233_Vvdwtot(%rsp) movd %mm0,%eax movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 movaps %xmm0,%xmm5 mulps %xmm4,%xmm4 ## xmm1=rinv, xmm4=rinvsq movaps nb233_krsqM(%rsp),%xmm3 addps %xmm3,%xmm5 ## xmm0=rinv+ krsq subps nb233_crf(%rsp),%xmm5 ## xmm0=rinv+ krsq-crf mulps nb233_two(%rsp),%xmm3 subps %xmm3,%xmm1 ## xmm1=rinv-2*krsq movaps %xmm5,%xmm2 mulps nb233_qqM(%rsp),%xmm2 ## xmm2=vcoul mulps nb233_qqM(%rsp),%xmm1 ## xmm1=coul part of fs mulps %xmm0,%xmm1 subss nb233_fstmp(%rsp),%xmm1 mulps %xmm0,%xmm1 movaps %xmm1,%xmm4 addps nb233_vctot(%rsp),%xmm2 movaps %xmm2,nb233_vctot(%rsp) movaps nb233_dxO(%rsp),%xmm0 movaps nb233_dyO(%rsp),%xmm1 movaps nb233_dzO(%rsp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb233_fixO(%rsp),%xmm3 movss nb233_fiyO(%rsp),%xmm4 movss nb233_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb233_fixO(%rsp) movss %xmm4,nb233_fiyO(%rsp) movss %xmm5,nb233_fizO(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb233_fixH1(%rsp),%xmm3 addss nb233_fiyH1(%rsp),%xmm4 addss nb233_fizH1(%rsp),%xmm5 movss %xmm3,nb233_fixH1(%rsp) movss %xmm4,nb233_fiyH1(%rsp) movss %xmm5,nb233_fizH1(%rsp) ## updated the H1 force shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb233_fixH2(%rsp),%xmm3 addss nb233_fiyH2(%rsp),%xmm4 addss nb233_fizH2(%rsp),%xmm5 movss %xmm3,nb233_fixH2(%rsp) movss %xmm4,nb233_fiyH2(%rsp) movss %xmm5,nb233_fizH2(%rsp) ## updated the H2 force movq nb233_faction(%rbp),%rdi shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb233_fixM(%rsp),%xmm3 addss nb233_fiyM(%rsp),%xmm4 addss nb233_fizM(%rsp),%xmm5 movss %xmm3,nb233_fixM(%rsp) movss %xmm4,nb233_fiyM(%rsp) movss %xmm5,nb233_fizM(%rsp) ## updated the M force ## the fj's - move in from mem start by acc. tx/ty/tz in xmm0, xmm1 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,4),%xmm7 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,%xmm0 movaps %xmm4,%xmm1 movaps %xmm5,%xmm2 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 unpcklps %xmm1,%xmm0 ## x,y sum in xmm0, z sum in xmm2 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb233_innerk(%rsp) jz _nb_kernel233_x86_64_sse.nb233_updateouterdata jmp _nb_kernel233_x86_64_sse.nb233_odd_loop_nb_kernel233_x86_64_sse.nb233_updateouterdata: movl nb233_ii3(%rsp),%ecx movq nb233_faction(%rbp),%rdi movq nb233_fshift(%rbp),%rsi movl nb233_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb233_fixO(%rsp),%xmm0 movaps nb233_fiyO(%rsp),%xmm1 movaps nb233_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 ## constant 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb233_fixH1(%rsp),%xmm0 movaps nb233_fiyH1(%rsp),%xmm1 movaps nb233_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 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb233_fixH2(%rsp),%xmm0 movaps nb233_fiyH2(%rsp),%xmm1 movaps nb233_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 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(%rdi,%rcx,4),%xmm3 movss 28(%rdi,%rcx,4),%xmm4 movss 32(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,24(%rdi,%rcx,4) movss %xmm4,28(%rdi,%rcx,4) movss %xmm5,32(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## constant 00001000 addps %xmm0,%xmm6 ## accumulate Mi forces in xmm0, xmm1, xmm2 movaps nb233_fixM(%rsp),%xmm0 movaps nb233_fiyM(%rsp),%xmm1 movaps nb233_fizM(%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 36(%rdi,%rcx,4),%xmm3 movss 40(%rdi,%rcx,4),%xmm4 movss 44(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,36(%rdi,%rcx,4) movss %xmm4,40(%rdi,%rcx,4) movss %xmm5,44(%rdi,%rcx,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 (%rsi,%rdx,4),%xmm3 movss 8(%rsi,%rdx,4),%xmm4 subps %xmm6,%xmm3 subss %xmm7,%xmm4 movlps %xmm3,(%rsi,%rdx,4) movss %xmm4,8(%rsi,%rdx,4)⌨️ 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?