nb_kernel133_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,141 行 · 第 1/5 页
S
2,141 行
mulpd %xmm5,%xmm7 mulpd %xmm6,%xmm7 ## xmm0=iter1 of rinv (new lu) movapd %xmm7,%xmm5 ## copy of lu mulpd %xmm7,%xmm7 ## lu*lu movapd nb133_three(%rsp),%xmm15 mulpd %xmm3,%xmm7 ## rsq*lu*lu movapd nb133_half(%rsp),%xmm6 subpd %xmm7,%xmm15 ## 30-rsq*lu*lu mulpd %xmm5,%xmm15 mulpd %xmm6,%xmm15 ## xmm15=rinv mulpd %xmm15,%xmm3 ## xmm3=r ## xmm15=rinv ## xmm3=r mulpd nb133_tsc(%rsp),%xmm3 ## rtab ## truncate and convert to integers cvttpd2pi %xmm3,%mm6 ## convert back to float cvtpi2pd %mm6,%xmm4 ## multiply by 8 pslld $3,%mm6 ## calculate eps subpd %xmm4,%xmm3 ## xmm3=eps ## move to integer registers movd %mm6,%r10d psrlq $32,%mm6 movd %mm6,%r11d ## xmm3=eps ## xmm15=rinv movq nb133_VFtab(%rbp),%rsi ## indices in r10, r11. Load dispersion and repulsion tables in parallel. movapd (%rsi,%r10,8),%xmm4 ## Y1d F1d movapd (%rsi,%r11,8),%xmm12 ## Y2d F2d movapd 32(%rsi,%r10,8),%xmm8 ## Y1r F1r movapd 32(%rsi,%r11,8),%xmm13 ## Y2r F2r movapd %xmm4,%xmm5 movapd %xmm8,%xmm9 unpcklpd %xmm12,%xmm4 ## Y1d Y2d unpckhpd %xmm12,%xmm5 ## F1d F2d unpcklpd %xmm13,%xmm8 ## Y1r Y2r unpckhpd %xmm13,%xmm9 ## F1r F2r movapd 16(%rsi,%r10,8),%xmm6 ## G1d H1d movapd 16(%rsi,%r11,8),%xmm12 ## G2d H2d movapd 48(%rsi,%r10,8),%xmm10 ## G1r H1r movapd 48(%rsi,%r11,8),%xmm13 ## G2r H2r movapd %xmm6,%xmm7 movapd %xmm10,%xmm11 unpcklpd %xmm12,%xmm6 ## G1d G2d unpckhpd %xmm12,%xmm7 ## H1d H2d unpcklpd %xmm13,%xmm10 ## G1r G2r unpckhpd %xmm13,%xmm11 ## H1r H2r ## dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11 mulpd %xmm3,%xmm7 ## Heps mulpd %xmm3,%xmm11 mulpd %xmm3,%xmm6 ## Geps mulpd %xmm3,%xmm10 mulpd %xmm3,%xmm7 ## Heps2 mulpd %xmm3,%xmm11 addpd %xmm6,%xmm5 ## F+Geps addpd %xmm10,%xmm9 addpd %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addpd %xmm11,%xmm9 addpd %xmm7,%xmm7 ## 2*Heps2 addpd %xmm11,%xmm11 addpd %xmm6,%xmm7 ## 2*Heps2+Geps addpd %xmm10,%xmm11 addpd %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps addpd %xmm9,%xmm11 mulpd %xmm3,%xmm5 ## eps*Fp mulpd %xmm3,%xmm9 movapd nb133_c6(%rsp),%xmm12 movapd nb133_c12(%rsp),%xmm13 addpd %xmm4,%xmm5 ## VV addpd %xmm8,%xmm9 mulpd %xmm12,%xmm5 ## VV*c6 = vnb6 mulpd %xmm13,%xmm9 ## VV*c12 = vnb12 addpd %xmm9,%xmm5 addpd nb133_Vvdwtot(%rsp),%xmm5 movapd %xmm5,nb133_Vvdwtot(%rsp) mulpd %xmm12,%xmm7 ## FF*c6 = fnb6 mulpd %xmm13,%xmm11 ## FF*c12 = fnb12 addpd %xmm11,%xmm7 mulpd nb133_tsc(%rsp),%xmm7 mulpd %xmm15,%xmm7 ## -fscal xorpd %xmm9,%xmm9 subpd %xmm7,%xmm9 ## fscal movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulpd nb133_dxO(%rsp),%xmm9 ## fx/fy/fz mulpd nb133_dyO(%rsp),%xmm10 mulpd nb133_dzO(%rsp),%xmm11 ## save j force temporarily movapd %xmm9,nb133_fjx(%rsp) movapd %xmm10,nb133_fjy(%rsp) movapd %xmm11,nb133_fjz(%rsp) ## increment i O force addpd nb133_fixO(%rsp),%xmm9 addpd nb133_fiyO(%rsp),%xmm10 addpd nb133_fizO(%rsp),%xmm11 movapd %xmm9,nb133_fixO(%rsp) movapd %xmm10,nb133_fiyO(%rsp) movapd %xmm11,nb133_fizO(%rsp) ## finished O LJ interaction. ## do H1, H2, and M interactions in parallel. ## xmm0-xmm2 still contain j coordinates. movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb133_ixH1(%rsp),%xmm0 subpd nb133_iyH1(%rsp),%xmm1 subpd nb133_izH1(%rsp),%xmm2 subpd nb133_ixH2(%rsp),%xmm3 subpd nb133_iyH2(%rsp),%xmm4 subpd nb133_izH2(%rsp),%xmm5 subpd nb133_ixM(%rsp),%xmm6 subpd nb133_iyM(%rsp),%xmm7 subpd nb133_izM(%rsp),%xmm8 movapd %xmm0,nb133_dxH1(%rsp) movapd %xmm1,nb133_dyH1(%rsp) movapd %xmm2,nb133_dzH1(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb133_dxH2(%rsp) movapd %xmm4,nb133_dyH2(%rsp) movapd %xmm5,nb133_dzH2(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb133_dxM(%rsp) movapd %xmm7,nb133_dyM(%rsp) movapd %xmm8,nb133_dzM(%rsp) mulpd %xmm6,%xmm6 mulpd %xmm7,%xmm7 mulpd %xmm8,%xmm8 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 addpd %xmm7,%xmm6 addpd %xmm8,%xmm6 ## start doing invsqrt for j atoms cvtpd2ps %xmm0,%xmm1 cvtpd2ps %xmm3,%xmm4 cvtpd2ps %xmm6,%xmm7 rsqrtps %xmm1,%xmm1 rsqrtps %xmm4,%xmm4 rsqrtps %xmm7,%xmm7 cvtps2pd %xmm1,%xmm1 cvtps2pd %xmm4,%xmm4 cvtps2pd %xmm7,%xmm7 movapd %xmm1,%xmm2 movapd %xmm4,%xmm5 movapd %xmm7,%xmm8 mulpd %xmm1,%xmm1 ## lu*lu mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm7,%xmm7 ## lu*lu movapd nb133_three(%rsp),%xmm9 movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulpd %xmm0,%xmm1 ## rsq*lu*lu mulpd %xmm3,%xmm4 ## rsq*lu*lu mulpd %xmm6,%xmm7 ## rsq*lu*lu subpd %xmm1,%xmm9 subpd %xmm4,%xmm10 subpd %xmm7,%xmm11 ## 3-rsq*lu*lu mulpd %xmm2,%xmm9 mulpd %xmm5,%xmm10 mulpd %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb133_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## first iteration for rinvH1 mulpd %xmm15,%xmm10 ## first iteration for rinvH2 mulpd %xmm15,%xmm11 ## first iteration for rinvM ## second iteration step movapd %xmm9,%xmm2 movapd %xmm10,%xmm5 movapd %xmm11,%xmm8 mulpd %xmm2,%xmm2 ## lu*lu mulpd %xmm5,%xmm5 ## lu*lu mulpd %xmm8,%xmm8 ## lu*lu movapd nb133_three(%rsp),%xmm1 movapd %xmm1,%xmm4 movapd %xmm1,%xmm7 mulpd %xmm0,%xmm2 ## rsq*lu*lu mulpd %xmm3,%xmm5 ## rsq*lu*lu mulpd %xmm6,%xmm8 ## rsq*lu*lu subpd %xmm2,%xmm1 subpd %xmm5,%xmm4 subpd %xmm8,%xmm7 ## 3-rsq*lu*lu mulpd %xmm1,%xmm9 mulpd %xmm4,%xmm10 mulpd %xmm7,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb133_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## rinvH1 mulpd %xmm15,%xmm10 ## rinvH2 mulpd %xmm15,%xmm11 ## rinvM ## interactions movapd %xmm9,%xmm0 movapd %xmm10,%xmm1 movapd %xmm11,%xmm2 mulpd %xmm9,%xmm9 mulpd %xmm10,%xmm10 mulpd %xmm11,%xmm11 mulpd nb133_qqH(%rsp),%xmm0 mulpd nb133_qqH(%rsp),%xmm1 mulpd nb133_qqM(%rsp),%xmm2 mulpd %xmm0,%xmm9 mulpd %xmm1,%xmm10 mulpd %xmm2,%xmm11 addpd nb133_vctot(%rsp),%xmm0 addpd %xmm2,%xmm1 addpd %xmm1,%xmm0 movapd %xmm0,nb133_vctot(%rsp) ## move j forces to xmm0-xmm2 movq nb133_faction(%rbp),%rdi movlpd (%rdi,%rax,8),%xmm0 movlpd 8(%rdi,%rax,8),%xmm1 movlpd 16(%rdi,%rax,8),%xmm2 movhpd (%rdi,%rbx,8),%xmm0 movhpd 8(%rdi,%rbx,8),%xmm1 movhpd 16(%rdi,%rbx,8),%xmm2 movapd %xmm9,%xmm7 movapd %xmm9,%xmm8 movapd %xmm11,%xmm13 movapd %xmm11,%xmm14 movapd %xmm11,%xmm15 movapd %xmm10,%xmm11 movapd %xmm10,%xmm12 ## add forces from O interaction addpd nb133_fjx(%rsp),%xmm0 addpd nb133_fjy(%rsp),%xmm1 addpd nb133_fjz(%rsp),%xmm2 mulpd nb133_dxH1(%rsp),%xmm7 mulpd nb133_dyH1(%rsp),%xmm8 mulpd nb133_dzH1(%rsp),%xmm9 mulpd nb133_dxH2(%rsp),%xmm10 mulpd nb133_dyH2(%rsp),%xmm11 mulpd nb133_dzH2(%rsp),%xmm12 mulpd nb133_dxM(%rsp),%xmm13 mulpd nb133_dyM(%rsp),%xmm14 mulpd nb133_dzM(%rsp),%xmm15 addpd %xmm7,%xmm0 addpd %xmm8,%xmm1 addpd %xmm9,%xmm2 addpd nb133_fixH1(%rsp),%xmm7 addpd nb133_fiyH1(%rsp),%xmm8 addpd nb133_fizH1(%rsp),%xmm9 addpd %xmm10,%xmm0 addpd %xmm11,%xmm1 addpd %xmm12,%xmm2 addpd nb133_fixH2(%rsp),%xmm10 addpd nb133_fiyH2(%rsp),%xmm11 addpd nb133_fizH2(%rsp),%xmm12 addpd %xmm13,%xmm0 addpd %xmm14,%xmm1 addpd %xmm15,%xmm2 addpd nb133_fixM(%rsp),%xmm13 addpd nb133_fiyM(%rsp),%xmm14 addpd nb133_fizM(%rsp),%xmm15 movapd %xmm7,nb133_fixH1(%rsp) movapd %xmm8,nb133_fiyH1(%rsp) movapd %xmm9,nb133_fizH1(%rsp) movapd %xmm10,nb133_fixH2(%rsp) movapd %xmm11,nb133_fiyH2(%rsp) movapd %xmm12,nb133_fizH2(%rsp) movapd %xmm13,nb133_fixM(%rsp) movapd %xmm14,nb133_fiyM(%rsp) movapd %xmm15,nb133_fizM(%rsp) ## store back j forces from xmm0-xmm2 movlpd %xmm0,(%rdi,%rax,8) movlpd %xmm1,8(%rdi,%rax,8) movlpd %xmm2,16(%rdi,%rax,8) movhpd %xmm0,(%rdi,%rbx,8) movhpd %xmm1,8(%rdi,%rbx,8) movhpd %xmm2,16(%rdi,%rbx,8) ## should we do one more iteration? subl $2,nb133_innerk(%rsp) jl _nb_kernel133_x86_64_sse2.nb133_checksingle jmp _nb_kernel133_x86_64_sse2.nb133_unroll_loop_nb_kernel133_x86_64_sse2.nb133_checksingle: movl nb133_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel133_x86_64_sse2.nb133_dosingle jmp _nb_kernel133_x86_64_sse2.nb133_updateouterdata_nb_kernel133_x86_64_sse2.nb133_dosingle: movq nb133_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb133_innerjjnr(%rsp) movq nb133_charge(%rbp),%rsi ## base of charge[] xorpd %xmm3,%xmm3 movlpd (%rsi,%rax,8),%xmm3 movapd %xmm3,%xmm4 mulsd nb133_iqM(%rsp),%xmm3 mulsd nb133_iqH(%rsp),%xmm4 movapd %xmm3,nb133_qqM(%rsp) movapd %xmm4,nb133_qqH(%rsp) movq nb133_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movq nb133_vdwparam(%rbp),%rsi shll %r8d movl nb133_ntia(%rsp),%edi addl %edi,%r8d movlpd (%rsi,%r8,8),%xmm6 ## c6a movhpd 8(%rsi,%r8,8),%xmm6 ## c6a c12a xorpd %xmm7,%xmm7 movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movapd %xmm4,nb133_c6(%rsp) movapd %xmm6,nb133_c12(%rsp) movq nb133_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move coordinates to xmm0-xmm2 and xmm4-xmm6 movlpd (%rsi,%rax,8),%xmm4 movlpd 8(%rsi,%rax,8),%xmm5 movlpd 16(%rsi,%rax,8),%xmm6 movapd %xmm4,%xmm0 movapd %xmm5,%xmm1 movapd %xmm6,%xmm2 ## calc dr subsd nb133_ixO(%rsp),%xmm4 subsd nb133_iyO(%rsp),%xmm5 subsd nb133_izO(%rsp),%xmm6 ## store dr movapd %xmm4,nb133_dxO(%rsp) movapd %xmm5,nb133_dyO(%rsp) movapd %xmm6,nb133_dzO(%rsp) ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqO in xmm7 movapd %xmm7,nb133_rsqO(%rsp) ## move j coords to xmm4-xmm6 movapd %xmm0,%xmm4 movapd %xmm1,%xmm5 movapd %xmm2,%xmm6 ## calc dr subsd nb133_ixH1(%rsp),%xmm4 subsd nb133_iyH1(%rsp),%xmm5 subsd nb133_izH1(%rsp),%xmm6 ## store dr movapd %xmm4,nb133_dxH1(%rsp) movapd %xmm5,nb133_dyH1(%rsp) movapd %xmm6,nb133_dzH1(%rsp) ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm6 addsd %xmm4,%xmm6 ## rsqH1 in xmm6 ## move j coords to xmm3-xmm5 movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5⌨️ 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?