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
shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## xmm0 = jx ## xmm1 = jy ## xmm2 = jz ## O interaction ## copy to xmm3-xmm5 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb233_ixO(%rsp),%xmm3 subps nb233_iyO(%rsp),%xmm4 subps nb233_izO(%rsp),%xmm5 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 %xmm4,%xmm3 addps %xmm5,%xmm3 ## xmm3=rsq ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm3,%xmm5 movaps %xmm5,%xmm15 mulps %xmm5,%xmm5 movaps nb233_three(%rsp),%xmm4 mulps %xmm3,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm4 ## 30-rsq*lu*lu mulps %xmm15,%xmm4 mulps nb233_half(%rsp),%xmm4 movaps %xmm4,%xmm15 mulps %xmm4,%xmm3 ## xmm15=rinv ## xmm3=r mulps nb233_tsc(%rsp),%xmm3 ## rtab ## truncate and convert to integers cvttps2dq %xmm3,%xmm5 ## convert back to float cvtdq2ps %xmm5,%xmm4 ## multiply by 8 pslld $3,%xmm5 ## calculate eps subps %xmm4,%xmm3 ## xmm3=eps ## move to integer registers movhlps %xmm5,%xmm6 movd %xmm5,%r8d movd %xmm6,%r10d pshufd $1,%xmm5,%xmm5 pshufd $1,%xmm6,%xmm6 movd %xmm5,%r9d movd %xmm6,%r11d ## xmm3=eps ## xmm15=rinv movq nb233_VFtab(%rbp),%rsi ## calculate LJ table movlps (%rsi,%r8,4),%xmm5 movlps 16(%rsi,%r8,4),%xmm9 movlps (%rsi,%r10,4),%xmm7 movlps 16(%rsi,%r10,4),%xmm11 movhps (%rsi,%r9,4),%xmm5 movhps 16(%rsi,%r9,4),%xmm9 movhps (%rsi,%r11,4),%xmm7 movhps 16(%rsi,%r11,4),%xmm11 movaps %xmm5,%xmm4 movaps %xmm9,%xmm8 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $136,%xmm11,%xmm8 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 shufps $221,%xmm11,%xmm9 ## 11011101 movlps 8(%rsi,%r8,4),%xmm7 movlps 24(%rsi,%r8,4),%xmm11 movlps 8(%rsi,%r10,4),%xmm13 movlps 24(%rsi,%r10,4),%xmm14 movhps 8(%rsi,%r9,4),%xmm7 movhps 24(%rsi,%r9,4),%xmm11 movhps 8(%rsi,%r11,4),%xmm13 movhps 24(%rsi,%r11,4),%xmm14 movaps %xmm7,%xmm6 movaps %xmm11,%xmm10 shufps $136,%xmm13,%xmm6 ## 10001000 shufps $136,%xmm14,%xmm10 ## 10001000 shufps $221,%xmm13,%xmm7 ## 11011101 shufps $221,%xmm14,%xmm11 ## 11011101 ## dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11 mulps %xmm3,%xmm7 ## Heps mulps %xmm3,%xmm11 mulps %xmm3,%xmm6 ## Geps mulps %xmm3,%xmm10 mulps %xmm3,%xmm7 ## Heps2 mulps %xmm3,%xmm11 addps %xmm6,%xmm5 ## F+Geps addps %xmm10,%xmm9 addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm11,%xmm9 addps %xmm7,%xmm7 ## 2*Heps2 addps %xmm11,%xmm11 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm10,%xmm11 addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps addps %xmm9,%xmm11 mulps %xmm3,%xmm5 ## eps*Fp mulps %xmm3,%xmm9 movaps nb233_c6(%rsp),%xmm12 movaps nb233_c12(%rsp),%xmm13 addps %xmm4,%xmm5 ## VV addps %xmm8,%xmm9 mulps %xmm12,%xmm5 ## VV*c6 = vnb6 mulps %xmm13,%xmm9 ## VV*c12 = vnb12 addps %xmm9,%xmm5 addps nb233_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb233_Vvdwtot(%rsp) mulps %xmm12,%xmm7 ## FF*c6 = fnb6 mulps %xmm13,%xmm11 ## FF*c12 = fnb12 addps %xmm11,%xmm7 mulps nb233_tsc(%rsp),%xmm7 mulps %xmm15,%xmm7 ## -fscal xorps %xmm9,%xmm9 subps %xmm7,%xmm9 ## fscal movaps %xmm9,%xmm10 movaps %xmm9,%xmm11 mulps nb233_dxO(%rsp),%xmm9 ## fx/fy/fz mulps nb233_dyO(%rsp),%xmm10 mulps nb233_dzO(%rsp),%xmm11 ## save j force temporarily movaps %xmm9,nb233_fjx(%rsp) movaps %xmm10,nb233_fjy(%rsp) movaps %xmm11,nb233_fjz(%rsp) ## increment i O force addps nb233_fixO(%rsp),%xmm9 addps nb233_fiyO(%rsp),%xmm10 addps nb233_fizO(%rsp),%xmm11 movaps %xmm9,nb233_fixO(%rsp) movaps %xmm10,nb233_fiyO(%rsp) movaps %xmm11,nb233_fizO(%rsp) ## finished O LJ interaction. ## do H1, H2, and M interactions in parallel. ## xmm0-xmm2 still contain j coordinates. movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb233_ixH1(%rsp),%xmm0 subps nb233_iyH1(%rsp),%xmm1 subps nb233_izH1(%rsp),%xmm2 subps nb233_ixH2(%rsp),%xmm3 subps nb233_iyH2(%rsp),%xmm4 subps nb233_izH2(%rsp),%xmm5 subps nb233_ixM(%rsp),%xmm6 subps nb233_iyM(%rsp),%xmm7 subps nb233_izM(%rsp),%xmm8 movaps %xmm0,nb233_dxH1(%rsp) movaps %xmm1,nb233_dyH1(%rsp) movaps %xmm2,nb233_dzH1(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb233_dxH2(%rsp) movaps %xmm4,nb233_dyH2(%rsp) movaps %xmm5,nb233_dzH2(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb233_dxM(%rsp) movaps %xmm7,nb233_dyM(%rsp) movaps %xmm8,nb233_dzM(%rsp) mulps %xmm6,%xmm6 mulps %xmm7,%xmm7 mulps %xmm8,%xmm8 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 addps %xmm7,%xmm6 addps %xmm8,%xmm6 ## start doing invsqrt for j atoms rsqrtps %xmm0,%xmm1 rsqrtps %xmm3,%xmm4 rsqrtps %xmm6,%xmm7 movaps %xmm1,%xmm2 movaps %xmm4,%xmm5 movaps %xmm7,%xmm8 mulps %xmm1,%xmm1 ## lu*lu mulps %xmm4,%xmm4 ## lu*lu mulps %xmm7,%xmm7 ## lu*lu movaps nb233_three(%rsp),%xmm9 movaps %xmm9,%xmm10 movaps %xmm9,%xmm11 mulps %xmm0,%xmm1 ## rsq*lu*lu mulps %xmm3,%xmm4 ## rsq*lu*lu mulps %xmm6,%xmm7 ## rsq*lu*lu subps %xmm1,%xmm9 subps %xmm4,%xmm10 subps %xmm7,%xmm11 ## 3-rsq*lu*lu mulps %xmm2,%xmm9 mulps %xmm5,%xmm10 mulps %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movaps nb233_half(%rsp),%xmm4 mulps %xmm4,%xmm9 ## rinvH1 mulps %xmm4,%xmm10 ## rinvH2 mulps %xmm4,%xmm11 ## rinvM ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movaps %xmm9,%xmm1 ## copy of rinv movaps %xmm10,%xmm4 movaps %xmm11,%xmm7 movaps nb233_krf(%rsp),%xmm2 mulps %xmm9,%xmm9 ## rinvsq mulps %xmm10,%xmm10 mulps %xmm11,%xmm11 mulps %xmm2,%xmm0 ## k*rsq mulps %xmm2,%xmm3 mulps %xmm2,%xmm6 movaps %xmm0,%xmm2 ## copy of k*rsq movaps %xmm3,%xmm5 movaps %xmm6,%xmm8 addps %xmm1,%xmm2 ## rinv+krsq addps %xmm4,%xmm5 addps %xmm7,%xmm8 movaps nb233_crf(%rsp),%xmm14 subps %xmm14,%xmm2 ## rinv+krsq-crf subps %xmm14,%xmm5 subps %xmm14,%xmm8 movaps nb233_qqH(%rsp),%xmm12 movaps nb233_qqM(%rsp),%xmm13 mulps %xmm12,%xmm2 ## voul=qq*(rinv+ krsq-crf) mulps %xmm12,%xmm5 ## voul=qq*(rinv+ krsq-crf) mulps %xmm13,%xmm8 ## voul=qq*(rinv+ krsq-crf) addps %xmm0,%xmm0 ## 2*krsq addps %xmm3,%xmm3 addps %xmm6,%xmm6 subps %xmm0,%xmm1 ## rinv-2*krsq subps %xmm3,%xmm4 subps %xmm6,%xmm7 mulps %xmm12,%xmm1 ## (rinv-2*krsq)*qq mulps %xmm12,%xmm4 mulps %xmm13,%xmm7 addps nb233_vctot(%rsp),%xmm2 addps %xmm8,%xmm5 addps %xmm5,%xmm2 movaps %xmm2,%xmm15 movaps %xmm2,nb233_vctot(%rsp) mulps %xmm9,%xmm1 ## fscal mulps %xmm10,%xmm4 mulps %xmm11,%xmm7 ## move j forces to local temp variables movq nb233_faction(%rbp),%rdi movlps (%rdi,%rax,4),%xmm9 ## jxa jya - - movlps (%rdi,%rcx,4),%xmm10 ## jxc jyc - - movhps (%rdi,%rbx,4),%xmm9 ## jxa jya jxb jyb movhps (%rdi,%rdx,4),%xmm10 ## jxc jyc jxd jyd movss 8(%rdi,%rax,4),%xmm11 ## jza - - - movss 8(%rdi,%rcx,4),%xmm12 ## jzc - - - movss 8(%rdi,%rbx,4),%xmm6 ## jzb movss 8(%rdi,%rdx,4),%xmm8 ## jzd movlhps %xmm6,%xmm11 ## jza - jzb - movlhps %xmm8,%xmm12 ## jzc - jzd - shufps $136,%xmm12,%xmm11 ## 10001000 => jza jzb jzc jzd ## xmm9: jxa jya jxb jyb ## xmm10: jxc jyc jxd jyd ## xmm11: jza jzb jzc jzd movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps %xmm4,%xmm3 movaps %xmm4,%xmm5 movaps %xmm7,%xmm6 movaps %xmm7,%xmm8 mulps nb233_dxH1(%rsp),%xmm0 mulps nb233_dyH1(%rsp),%xmm1 mulps nb233_dzH1(%rsp),%xmm2 mulps nb233_dxH2(%rsp),%xmm3 mulps nb233_dyH2(%rsp),%xmm4 mulps nb233_dzH2(%rsp),%xmm5 mulps nb233_dxM(%rsp),%xmm6 mulps nb233_dyM(%rsp),%xmm7 mulps nb233_dzM(%rsp),%xmm8 ## fetch forces from O interaction movaps nb233_fjx(%rsp),%xmm13 movaps nb233_fjy(%rsp),%xmm14 addps nb233_fjz(%rsp),%xmm11 addps %xmm0,%xmm13 addps %xmm1,%xmm14 addps %xmm2,%xmm11 addps nb233_fixH1(%rsp),%xmm0 addps nb233_fiyH1(%rsp),%xmm1 addps nb233_fizH1(%rsp),%xmm2 addps %xmm3,%xmm13 addps %xmm4,%xmm14 addps %xmm5,%xmm11 addps nb233_fixH2(%rsp),%xmm3 addps nb233_fiyH2(%rsp),%xmm4 addps nb233_fizH2(%rsp),%xmm5 addps %xmm6,%xmm13 addps %xmm7,%xmm14 addps %xmm8,%xmm11 addps nb233_fixM(%rsp),%xmm6 addps nb233_fiyM(%rsp),%xmm7 addps nb233_fizM(%rsp),%xmm8 movaps %xmm0,nb233_fixH1(%rsp) movaps %xmm1,nb233_fiyH1(%rsp) movaps %xmm2,nb233_fizH1(%rsp) movaps %xmm3,nb233_fixH2(%rsp) movaps %xmm4,nb233_fiyH2(%rsp) movaps %xmm5,nb233_fizH2(%rsp) movaps %xmm6,nb233_fixM(%rsp) movaps %xmm7,nb233_fiyM(%rsp) movaps %xmm8,nb233_fizM(%rsp) ## xmm9 = fjx ## xmm10 = fjy ## xmm11 = fjz movaps %xmm13,%xmm15 unpcklps %xmm14,%xmm13 unpckhps %xmm14,%xmm15 addps %xmm13,%xmm9 addps %xmm15,%xmm10 movhlps %xmm11,%xmm12 ## fjzc fjzd movlps %xmm9,(%rdi,%rax,4) movhps %xmm9,(%rdi,%rbx,4) movlps %xmm10,(%rdi,%rcx,4) movhps %xmm10,(%rdi,%rdx,4) movss %xmm11,8(%rdi,%rax,4) movss %xmm12,8(%rdi,%rcx,4) shufps $1,%xmm11,%xmm11 shufps $1,%xmm12,%xmm12 movss %xmm11,8(%rdi,%rbx,4) movss %xmm12,8(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb233_innerk(%rsp) jl _nb_kernel233_x86_64_sse.nb233_odd_inner jmp _nb_kernel233_x86_64_sse.nb233_unroll_loop_nb_kernel233_x86_64_sse.nb233_odd_inner: addl $4,nb233_innerk(%rsp) jnz _nb_kernel233_x86_64_sse.nb233_odd_loop jmp _nb_kernel233_x86_64_sse.nb233_updateouterdata_nb_kernel233_x86_64_sse.nb233_odd_loop: movq nb233_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb233_innerjjnr(%rsp) xorps %xmm4,%xmm4 ## clear reg. movss nb233_iqM(%rsp),%xmm4 movq nb233_charge(%rbp),%rsi movhps nb233_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,nb233_qqM(%rsp) ## use dummy qq for storage xorps %xmm6,%xmm6⌨️ 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?