nb_kernel104_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,085 行 · 第 1/5 页
S
2,085 行
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 jH1 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 nb104_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 nb104_half(%rsp),%xmm0 mulps %xmm0,%xmm9 ## rinvH1H1 mulps %xmm0,%xmm10 ## rinvH2H1 mulps %xmm0,%xmm11 ## rinvMH1 ## H1 interactions movaps %xmm9,%xmm0 movaps %xmm10,%xmm1 movaps %xmm11,%xmm2 mulps %xmm9,%xmm9 mulps %xmm10,%xmm10 mulps %xmm11,%xmm11 mulps nb104_qqHH(%rsp),%xmm0 mulps nb104_qqHH(%rsp),%xmm1 mulps nb104_qqMH(%rsp),%xmm2 mulps %xmm0,%xmm9 mulps %xmm1,%xmm10 mulps %xmm2,%xmm11 addps nb104_vctot(%rsp),%xmm0 addps %xmm2,%xmm1 addps %xmm1,%xmm0 movaps %xmm0,nb104_vctot(%rsp) ## move j H1 forces to local temp variables movlps 12(%rdi,%rax,4),%xmm0 ## jxH1a jyH1a - - movlps 12(%rdi,%rcx,4),%xmm1 ## jxH1c jyH1c - - movhps 12(%rdi,%rbx,4),%xmm0 ## jxH1a jyH1a jxH1b jyH1b movhps 12(%rdi,%rdx,4),%xmm1 ## jxH1c jyH1c jxH1d jyH1d movss 20(%rdi,%rax,4),%xmm2 ## jzH1a - - - movss 20(%rdi,%rcx,4),%xmm3 ## jzH1c - - - movhps 20(%rdi,%rbx,4),%xmm2 ## jzH1a - jzH1b - movhps 20(%rdi,%rdx,4),%xmm3 ## jzH1c - jzH1d - shufps $136,%xmm3,%xmm2 ## 10001000 => jzH1a jzH1b jzH1c jzH1d ## xmm0: jxH1a jyH1a jxH1b jyH1b ## xmm1: jxH1c jyH1c jxH1d jyH1d ## xmm2: jzH1a jzH1b jzH1c jzH1d movaps %xmm9,%xmm7 movaps %xmm9,%xmm8 movaps %xmm11,%xmm13 movaps %xmm11,%xmm14 movaps %xmm11,%xmm15 movaps %xmm10,%xmm11 movaps %xmm10,%xmm12 mulps nb104_dxH1H1(%rsp),%xmm7 mulps nb104_dyH1H1(%rsp),%xmm8 mulps nb104_dzH1H1(%rsp),%xmm9 mulps nb104_dxH2H1(%rsp),%xmm10 mulps nb104_dyH2H1(%rsp),%xmm11 mulps nb104_dzH2H1(%rsp),%xmm12 mulps nb104_dxMH1(%rsp),%xmm13 mulps nb104_dyMH1(%rsp),%xmm14 mulps nb104_dzMH1(%rsp),%xmm15 movaps %xmm7,%xmm3 movaps %xmm8,%xmm4 addps %xmm9,%xmm2 addps nb104_fixH1(%rsp),%xmm7 addps nb104_fiyH1(%rsp),%xmm8 addps nb104_fizH1(%rsp),%xmm9 addps %xmm10,%xmm3 addps %xmm11,%xmm4 addps %xmm12,%xmm2 addps nb104_fixH2(%rsp),%xmm10 addps nb104_fiyH2(%rsp),%xmm11 addps nb104_fizH2(%rsp),%xmm12 addps %xmm13,%xmm3 addps %xmm14,%xmm4 addps %xmm15,%xmm2 addps nb104_fixM(%rsp),%xmm13 addps nb104_fiyM(%rsp),%xmm14 addps nb104_fizM(%rsp),%xmm15 movaps %xmm7,nb104_fixH1(%rsp) movaps %xmm8,nb104_fiyH1(%rsp) movaps %xmm9,nb104_fizH1(%rsp) movaps %xmm10,nb104_fixH2(%rsp) movaps %xmm11,nb104_fiyH2(%rsp) movaps %xmm12,nb104_fizH2(%rsp) movaps %xmm13,nb104_fixM(%rsp) movaps %xmm14,nb104_fiyM(%rsp) movaps %xmm15,nb104_fizM(%rsp) ## xmm3 = fH1x , xmm4 = fH1y movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movhlps %xmm2,%xmm3 ## fH1zc fH1zd movlps %xmm0,12(%rdi,%rax,4) movhps %xmm0,12(%rdi,%rbx,4) movlps %xmm1,12(%rdi,%rcx,4) movhps %xmm1,12(%rdi,%rdx,4) movss %xmm2,20(%rdi,%rax,4) movss %xmm3,20(%rdi,%rcx,4) shufps $1,%xmm2,%xmm2 shufps $1,%xmm3,%xmm3 movss %xmm2,20(%rdi,%rbx,4) movss %xmm3,20(%rdi,%rdx,4) ## move j H2 coordinates to local temp variables movlps 24(%rsi,%rax,4),%xmm0 ## jxH2a jyH2a - - movlps 24(%rsi,%rcx,4),%xmm1 ## jxH2c jyH2c - - movhps 24(%rsi,%rbx,4),%xmm0 ## jxH2a jyH2a jxH2b jyH2b movhps 24(%rsi,%rdx,4),%xmm1 ## jxH2c jyH2c jxH2d jyH2d movss 32(%rsi,%rax,4),%xmm2 ## jzH2a - - - movss 32(%rsi,%rcx,4),%xmm3 ## jzH2c - - - movhps 32(%rsi,%rbx,4),%xmm2 ## jzH2a - jzH2b - movhps 32(%rsi,%rdx,4),%xmm3 ## jzH2c - jzH2d - movaps %xmm0,%xmm4 unpcklps %xmm1,%xmm0 ## jxH2a jxH2c jyH2a jyH2c unpckhps %xmm1,%xmm4 ## jxH2b jxH2d jyH2b jyH2d movaps %xmm0,%xmm1 unpcklps %xmm4,%xmm0 ## x unpckhps %xmm4,%xmm1 ## y shufps $136,%xmm3,%xmm2 ## 10001000 => jzH2a jzH2b jzH2c jzH2d ## xmm0 = H2x ## xmm1 = H2y ## xmm2 = H2z movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb104_ixH1(%rsp),%xmm0 subps nb104_iyH1(%rsp),%xmm1 subps nb104_izH1(%rsp),%xmm2 subps nb104_ixH2(%rsp),%xmm3 subps nb104_iyH2(%rsp),%xmm4 subps nb104_izH2(%rsp),%xmm5 subps nb104_ixM(%rsp),%xmm6 subps nb104_iyM(%rsp),%xmm7 subps nb104_izM(%rsp),%xmm8 movaps %xmm0,nb104_dxH1H2(%rsp) movaps %xmm1,nb104_dyH1H2(%rsp) movaps %xmm2,nb104_dzH1H2(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxH2H2(%rsp) movaps %xmm4,nb104_dyH2H2(%rsp) movaps %xmm5,nb104_dzH2H2(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb104_dxMH2(%rsp) movaps %xmm7,nb104_dyMH2(%rsp) movaps %xmm8,nb104_dzMH2(%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 jH2 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 nb104_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 nb104_half(%rsp),%xmm0 mulps %xmm0,%xmm9 ## rinvH1H2 mulps %xmm0,%xmm10 ## rinvH2H2 mulps %xmm0,%xmm11 ## rinvMH2 ## H2 interactions movaps %xmm9,%xmm0 movaps %xmm10,%xmm1 movaps %xmm11,%xmm2 mulps %xmm9,%xmm9 mulps %xmm10,%xmm10 mulps %xmm11,%xmm11 mulps nb104_qqHH(%rsp),%xmm0 mulps nb104_qqHH(%rsp),%xmm1 mulps nb104_qqMH(%rsp),%xmm2 mulps %xmm0,%xmm9 mulps %xmm1,%xmm10 mulps %xmm2,%xmm11 addps nb104_vctot(%rsp),%xmm0 addps %xmm2,%xmm1 addps %xmm1,%xmm0 movaps %xmm0,nb104_vctot(%rsp) ## move j H2 forces to local temp variables movlps 24(%rdi,%rax,4),%xmm0 ## jxH2a jyH2a - - movlps 24(%rdi,%rcx,4),%xmm1 ## jxH2c jyH2c - - movhps 24(%rdi,%rbx,4),%xmm0 ## jxH2a jyH2a jxH2b jyH2b movhps 24(%rdi,%rdx,4),%xmm1 ## jxH2c jyH2c jxH2d jyH2d movss 32(%rdi,%rax,4),%xmm2 ## jzH2a - - - movss 32(%rdi,%rcx,4),%xmm3 ## jzH2c - - - movhps 32(%rdi,%rbx,4),%xmm2 ## jzH2a - jzH2b - movhps 32(%rdi,%rdx,4),%xmm3 ## jzH2c - jzH2d - shufps $136,%xmm3,%xmm2 ## 10001000 => jzH2a jzH2b jzH2c jzH2d ## xmm0: jxH2a jyH2a jxH2b jyH2b ## xmm1: jxH2c jyH2c jxH2d jyH2d ## xmm2: jzH2a jzH2b jzH2c jzH2d movaps %xmm9,%xmm7 movaps %xmm9,%xmm8 movaps %xmm11,%xmm13 movaps %xmm11,%xmm14 movaps %xmm11,%xmm15 movaps %xmm10,%xmm11 movaps %xmm10,%xmm12 mulps nb104_dxH1H2(%rsp),%xmm7 mulps nb104_dyH1H2(%rsp),%xmm8 mulps nb104_dzH1H2(%rsp),%xmm9 mulps nb104_dxH2H2(%rsp),%xmm10 mulps nb104_dyH2H2(%rsp),%xmm11 mulps nb104_dzH2H2(%rsp),%xmm12 mulps nb104_dxMH2(%rsp),%xmm13 mulps nb104_dyMH2(%rsp),%xmm14 mulps nb104_dzMH2(%rsp),%xmm15 movaps %xmm7,%xmm3 movaps %xmm8,%xmm4 addps %xmm9,%xmm2 addps nb104_fixH1(%rsp),%xmm7 addps nb104_fiyH1(%rsp),%xmm8 addps nb104_fizH1(%rsp),%xmm9 addps %xmm10,%xmm3 addps %xmm11,%xmm4 addps %xmm12,%xmm2 addps nb104_fixH2(%rsp),%xmm10 addps nb104_fiyH2(%rsp),%xmm11 addps nb104_fizH2(%rsp),%xmm12 addps %xmm13,%xmm3 addps %xmm14,%xmm4 addps %xmm15,%xmm2 addps nb104_fixM(%rsp),%xmm13 addps nb104_fiyM(%rsp),%xmm14 addps nb104_fizM(%rsp),%xmm15 movaps %xmm7,nb104_fixH1(%rsp) movaps %xmm8,nb104_fiyH1(%rsp) movaps %xmm9,nb104_fizH1(%rsp) movaps %xmm10,nb104_fixH2(%rsp) movaps %xmm11,nb104_fiyH2(%rsp) movaps %xmm12,nb104_fizH2(%rsp) movaps %xmm13,nb104_fixM(%rsp) movaps %xmm14,nb104_fiyM(%rsp) movaps %xmm15,nb104_fizM(%rsp) ## xmm3 = fH2x , xmm4 = fH2y movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movhlps %xmm2,%xmm3 ## fH2zc fH2zd movlps %xmm0,24(%rdi,%rax,4) movhps %xmm0,24(%rdi,%rbx,4) movlps %xmm1,24(%rdi,%rcx,4) movhps %xmm1,24(%rdi,%rdx,4) movss %xmm2,32(%rdi,%rax,4) movss %xmm3,32(%rdi,%rcx,4) shufps $1,%xmm2,%xmm2 shufps $1,%xmm3,%xmm3 movss %xmm2,32(%rdi,%rbx,4) movss %xmm3,32(%rdi,%rdx,4) ## move j M coordinates to local temp variables movlps 36(%rsi,%rax,4),%xmm0 ## jxMa jyMa - - movlps 36(%rsi,%rcx,4),%xmm1 ## jxMc jyMc - - movhps 36(%rsi,%rbx,4),%xmm0 ## jxMa jyMa jxMb jyMb movhps 36(%rsi,%rdx,4),%xmm1 ## jxMc jyMc jxMd jyMd movss 44(%rsi,%rax,4),%xmm2 ## jzMa - - - movss 44(%rsi,%rcx,4),%xmm3 ## jzMc - - - movss 44(%rsi,%rbx,4),%xmm5 ## jzMb - - - movss 44(%rsi,%rdx,4),%xmm6 ## jzMd - - - movlhps %xmm5,%xmm2 ## jzMa - jzMb - movlhps %xmm6,%xmm3 ## jzMc - jzMd - movaps %xmm0,%xmm4 unpcklps %xmm1,%xmm0 ## jxMa jxMc jyMa jyMc unpckhps %xmm1,%xmm4 ## jxMb jxMd jyMb jyMd movaps %xmm0,%xmm1 unpcklps %xmm4,%xmm0 ## x unpckhps %xmm4,%xmm1 ## y shufps $136,%xmm3,%xmm2 ## 10001000 => jzMa jzMb jzMc jzMd ## xmm0 = Mx ## xmm1 = My ## xmm2 = Mz movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 movaps %xmm2,%xmm8 subps nb104_ixH1(%rsp),%xmm0 subps nb104_iyH1(%rsp),%xmm1 subps nb104_izH1(%rsp),%xmm2 subps nb104_ixH2(%rsp),%xmm3 subps nb104_iyH2(%rsp),%xmm4 subps nb104_izH2(%rsp),%xmm5 subps nb104_ixM(%rsp),%xmm6 subps nb104_iyM(%rsp),%xmm7 subps nb104_izM(%rsp),%xmm8 movaps %xmm0,nb104_dxH1M(%rsp) movaps %xmm1,nb104_dyH1M(%rsp) movaps %xmm2,nb104_dzH1M(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxH2M(%rsp) movaps %xmm4,nb104_dyH2M(%rsp) movaps %xmm5,nb104_dzH2M(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb104_dxMM(%rsp) movaps %xmm7,nb104_dyMM(%rsp) movaps %xmm8,nb104_dzMM(%rsp) mulps %xmm6,%xmm6 mulps %xmm7,%xmm7 mulps %xmm8,%xmm8⌨️ 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?