nb_kernel313_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,144 行 · 第 1/5 页
S
2,144 行
movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm6 ## 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb313nf_c6(%rsp) movaps %xmm6,nb313nf_c12(%rsp) movq nb313nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx ## replace jnr with j3 lea (%rdx,%rdx,2),%rdx ## move four coordinates to xmm0-xmm2 movlps (%rsi,%rax,4),%xmm4 movlps (%rsi,%rcx,4),%xmm5 movss 8(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm6 movhps (%rsi,%rbx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movss 8(%rsi,%rbx,4),%xmm0 movss 8(%rsi,%rdx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb313nf_ixO(%rsp),%xmm4 movaps nb313nf_iyO(%rsp),%xmm5 movaps nb313nf_izO(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb313nf_ixH1(%rsp),%xmm4 movaps nb313nf_iyH1(%rsp),%xmm5 movaps nb313nf_izH1(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb313nf_ixH2(%rsp),%xmm3 movaps nb313nf_iyH2(%rsp),%xmm4 movaps nb313nf_izH2(%rsp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## move ixM-izM to xmm2-xmm4 movaps nb313nf_iyM(%rsp),%xmm3 movaps nb313nf_izM(%rsp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb313nf_ixM(%rsp),%xmm2 subps %xmm0,%xmm2 ## square it mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 addps %xmm3,%xmm4 addps %xmm2,%xmm4 ## rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb313nf_three(%rsp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb313nf_half(%rsp),%xmm0 movaps %xmm0,nb313nf_rinvH1(%rsp) ## rinvH1 in xmm4 mulps %xmm0,%xmm6 movaps %xmm6,nb313nf_rH1(%rsp) ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb313nf_three(%rsp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb313nf_half(%rsp),%xmm0 movaps %xmm0,nb313nf_rinvH2(%rsp) ## rinvH2 in xmm4 mulps %xmm0,%xmm5 movaps %xmm5,nb313nf_rH2(%rsp) ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb313nf_three(%rsp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb313nf_half(%rsp),%xmm0 movaps %xmm0,nb313nf_rinvM(%rsp) ## rinvM in xmm5 mulps %xmm0,%xmm4 movaps %xmm4,nb313nf_rM(%rsp) ## Do the O LJ-only interaction directly. rcpps %xmm7,%xmm2 movaps nb313nf_two(%rsp),%xmm1 mulps %xmm2,%xmm7 subps %xmm7,%xmm1 mulps %xmm1,%xmm2 ## rinvsq movaps %xmm2,%xmm0 mulps %xmm2,%xmm0 ## r4 mulps %xmm2,%xmm0 ## r6 movaps %xmm0,%xmm1 mulps %xmm1,%xmm1 ## r12 mulps nb313nf_c6(%rsp),%xmm0 mulps nb313nf_c12(%rsp),%xmm1 movaps %xmm1,%xmm3 subps %xmm0,%xmm3 ## Vvdw12-Vvdw6 addps nb313nf_Vvdwtot(%rsp),%xmm3 movaps %xmm3,nb313nf_Vvdwtot(%rsp) ## Do H1 interaction movq nb313nf_VFtab(%rbp),%rsi movaps nb313nf_rH1(%rsp),%xmm7 mulps nb313nf_tsc(%rsp),%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 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 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,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 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 movaps nb313nf_qqH(%rsp),%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 nb313nf_vctot(%rsp),%xmm5 movaps %xmm5,nb313nf_vctot(%rsp) ## Done with H1, do H2 interactions movaps nb313nf_rH2(%rsp),%xmm7 mulps nb313nf_tsc(%rsp),%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 (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## shuffle 10001000 shufps $221,%xmm7,%xmm5 ## shuffle 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## shuf 10001000 shufps $221,%xmm3,%xmm7 ## shuf 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 nb313nf_qqH(%rsp),%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 nb313nf_vctot(%rsp),%xmm5 movaps %xmm5,nb313nf_vctot(%rsp) ## Done with H2, do M interactions movaps nb313nf_rM(%rsp),%xmm7 mulps nb313nf_tsc(%rsp),%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 (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 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,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 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 movaps nb313nf_qqM(%rsp),%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 nb313nf_vctot(%rsp),%xmm5 movaps %xmm5,nb313nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb313nf_innerk(%rsp) jl _nb_kernel313nf_x86_64_sse.nb313nf_odd_inner jmp _nb_kernel313nf_x86_64_sse.nb313nf_unroll_loop_nb_kernel313nf_x86_64_sse.nb313nf_odd_inner: addl $4,nb313nf_innerk(%rsp) jnz _nb_kernel313nf_x86_64_sse.nb313nf_odd_loop jmp _nb_kernel313nf_x86_64_sse.nb313nf_updateouterdata_nb_kernel313nf_x86_64_sse.nb313nf_odd_loop: movq nb313nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb313nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 ## clear reg. movss nb313nf_iqM(%rsp),%xmm4 movq nb313nf_charge(%rbp),%rsi movhps nb313nf_iqH(%rsp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb313nf_qqM(%rsp) ## use dummy qq for storage xorps %xmm6,%xmm6 movq nb313nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb313nf_vdwparam(%rbp),%rsi shll %ebx addl nb313nf_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb313nf_c6(%rsp) movaps %xmm7,nb313nf_c12(%rsp) movq nb313nf_pos(%rbp),%rsi lea (%rax,%rax,2),%rax movss nb313nf_ixO(%rsp),%xmm3 movss nb313nf_iyO(%rsp),%xmm4 movss nb313nf_izO(%rsp),%xmm5 movss nb313nf_ixH1(%rsp),%xmm0 movss nb313nf_iyH1(%rsp),%xmm1 movss nb313nf_izH1(%rsp),%xmm2 unpcklps nb313nf_ixH2(%rsp),%xmm3 ## ixO ixH2 - - unpcklps nb313nf_iyH2(%rsp),%xmm4 ## iyO iyH2 - - unpcklps nb313nf_izH2(%rsp),%xmm5 ## izO izH2 - - unpcklps nb313nf_ixM(%rsp),%xmm0 ## ixH1 ixM - - unpcklps nb313nf_iyM(%rsp),%xmm1 ## iyH1 iyM - - unpcklps nb313nf_izM(%rsp),%xmm2 ## izH1 izM - - unpcklps %xmm0,%xmm3 ## ixO ixH1 ixH2 ixM unpcklps %xmm1,%xmm4 ## same for y unpcklps %xmm2,%xmm5 ## same for z ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm0 movss 4(%rsi,%rax,4),%xmm1 movss 8(%rsi,%rax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 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 nb313nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb313nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ⌨️ 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?