nb_kernel211_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,749 行 · 第 1/4 页
S
1,749 行
mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb211_dxH1(%rsp) movaps %xmm4,nb211_dyH1(%rsp) movaps %xmm5,nb211_dzH1(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 movaps %xmm6,nb211_dxH2(%rsp) movaps %xmm7,nb211_dyH2(%rsp) movaps %xmm8,nb211_dzH2(%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 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 nb211_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 nb211_half(%rsp),%xmm4 mulps %xmm4,%xmm9 ## rinvO mulps %xmm4,%xmm10 ## rinvH1 mulps %xmm4,%xmm11 ## rinvH2 ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movaps %xmm9,%xmm1 ## copy of rinv movaps %xmm10,%xmm4 movaps %xmm11,%xmm7 movaps nb211_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 %xmm9,%xmm12 mulps %xmm12,%xmm12 ## rinv4 mulps %xmm9,%xmm12 ## rinv6 subps nb211_crf(%rsp),%xmm2 ## rinv+krsq-crf subps nb211_crf(%rsp),%xmm5 subps nb211_crf(%rsp),%xmm8 mulps nb211_qqO(%rsp),%xmm2 ## voul=qq*(rinv+ krsq-crf) mulps nb211_qqH(%rsp),%xmm5 ## voul=qq*(rinv+ krsq-crf) mulps nb211_qqH(%rsp),%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 movaps %xmm12,%xmm13 ## rinv6 mulps %xmm12,%xmm12 ## rinv12 mulps nb211_c6(%rsp),%xmm13 mulps nb211_c12(%rsp),%xmm12 movaps %xmm12,%xmm14 subps %xmm13,%xmm14 mulps nb211_qqO(%rsp),%xmm1 ## (rinv-2*krsq)*qq mulps nb211_qqH(%rsp),%xmm4 mulps nb211_qqH(%rsp),%xmm7 addps nb211_vctot(%rsp),%xmm2 addps %xmm8,%xmm5 addps %xmm5,%xmm2 movaps %xmm2,nb211_vctot(%rsp) addps nb211_Vvdwtot(%rsp),%xmm14 mulps nb211_six(%rsp),%xmm13 mulps nb211_twelve(%rsp),%xmm12 movaps %xmm14,nb211_Vvdwtot(%rsp) subps %xmm13,%xmm12 ## LJ fscal addps %xmm12,%xmm1 mulps %xmm9,%xmm1 ## fscal mulps %xmm10,%xmm4 mulps %xmm11,%xmm7 movq nb211_faction(%rbp),%rdi ## move j forces to local temp variables 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 nb211_dxO(%rsp),%xmm0 mulps nb211_dyO(%rsp),%xmm1 mulps nb211_dzO(%rsp),%xmm2 mulps nb211_dxH1(%rsp),%xmm3 mulps nb211_dyH1(%rsp),%xmm4 mulps nb211_dzH1(%rsp),%xmm5 mulps nb211_dxH2(%rsp),%xmm6 mulps nb211_dyH2(%rsp),%xmm7 mulps nb211_dzH2(%rsp),%xmm8 movaps %xmm0,%xmm13 movaps %xmm1,%xmm14 addps %xmm2,%xmm11 addps nb211_fixO(%rsp),%xmm0 addps nb211_fiyO(%rsp),%xmm1 addps nb211_fizO(%rsp),%xmm2 addps %xmm3,%xmm13 addps %xmm4,%xmm14 addps %xmm5,%xmm11 addps nb211_fixH1(%rsp),%xmm3 addps nb211_fiyH1(%rsp),%xmm4 addps nb211_fizH1(%rsp),%xmm5 addps %xmm6,%xmm13 addps %xmm7,%xmm14 addps %xmm8,%xmm11 addps nb211_fixH2(%rsp),%xmm6 addps nb211_fiyH2(%rsp),%xmm7 addps nb211_fizH2(%rsp),%xmm8 movaps %xmm0,nb211_fixO(%rsp) movaps %xmm1,nb211_fiyO(%rsp) movaps %xmm2,nb211_fizO(%rsp) movaps %xmm3,nb211_fixH1(%rsp) movaps %xmm4,nb211_fiyH1(%rsp) movaps %xmm5,nb211_fizH1(%rsp) movaps %xmm6,nb211_fixH2(%rsp) movaps %xmm7,nb211_fiyH2(%rsp) movaps %xmm8,nb211_fizH2(%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 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,nb211_innerk(%rsp) jl _nb_kernel211_x86_64_sse.nb211_odd_inner jmp _nb_kernel211_x86_64_sse.nb211_unroll_loop_nb_kernel211_x86_64_sse.nb211_odd_inner: addl $4,nb211_innerk(%rsp) jnz _nb_kernel211_x86_64_sse.nb211_odd_loop jmp _nb_kernel211_x86_64_sse.nb211_updateouterdata_nb_kernel211_x86_64_sse.nb211_odd_loop: movq nb211_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb211_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb211_iqO(%rsp),%xmm4 movq nb211_charge(%rbp),%rsi movhps nb211_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb211_qqO(%rsp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movq nb211_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb211_vdwparam(%rbp),%rsi shll %ebx addl nb211_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb211_c6(%rsp) movaps %xmm7,nb211_c12(%rsp) movq nb211_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## 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 movss nb211_ixO(%rsp),%xmm0 movss nb211_iyO(%rsp),%xmm1 movss nb211_izO(%rsp),%xmm2 movlps nb211_ixH1(%rsp),%xmm6 movlps nb211_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm0 movlps nb211_iyH1(%rsp),%xmm6 movlps nb211_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm1 movlps nb211_izH1(%rsp),%xmm6 movlps nb211_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb211_dxO(%rsp) movaps %xmm4,nb211_dyO(%rsp) movaps %xmm5,nb211_dzO(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 movaps %xmm4,%xmm0 mulps nb211_krf(%rsp),%xmm0 movaps %xmm0,nb211_krsqO(%rsp) rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb211_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb211_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv ## a little trick to avoid NaNs: ## positions 0,2,and 3 are valid, but not 1. ## If it contains NaN it doesnt help to mult by 0, ## So we shuffle it and copy pos 0 to pos1! shufps $224,%xmm0,%xmm0 ## 11100000 movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulss %xmm4,%xmm1 mulss %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb211_c6(%rsp),%xmm1 mulps nb211_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subss %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb211_Vvdwtot(%rsp),%xmm5 mulss nb211_six(%rsp),%xmm1 mulss nb211_twelve(%rsp),%xmm2 subss %xmm1,%xmm2 movaps %xmm0,%xmm1 ## xmm1=rinv movaps nb211_krsqO(%rsp),%xmm3 addps %xmm3,%xmm0 ## xmm0=rinv+ krsq mulps nb211_two(%rsp),%xmm3 subps nb211_crf(%rsp),%xmm0 ## xmm0=rinv+ krsq-crf subps %xmm3,%xmm1 ## xmm1=rinv-2*krsq mulps nb211_qqO(%rsp),%xmm0 ## xmm0=vcoul mulps nb211_qqO(%rsp),%xmm1 ## xmm1=coul part of fs addps %xmm1,%xmm2 ## total fs mulps %xmm2,%xmm4 ## xmm4=total fscal addps nb211_vctot(%rsp),%xmm0 movaps %xmm0,nb211_vctot(%rsp) movaps nb211_dxO(%rsp),%xmm0 movaps nb211_dyO(%rsp),%xmm1 movaps nb211_dzO(%rsp),%xmm2 movaps %xmm5,nb211_Vvdwtot(%rsp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb211_fixO(%rsp),%xmm3 movss nb211_fiyO(%rsp),%xmm4 movss nb211_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb211_fixO(%rsp) movss %xmm4,nb211_fiyO(%rsp) movss %xmm5,nb211_fizO(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## 11100110 shufps $230,%xmm5,%xmm5 ## 11100110 addss nb211_fixH1(%rsp),%xmm3 addss nb211_fiyH1(%rsp),%xmm4 addss nb211_fizH1(%rsp),%xmm5 movss %xmm3,nb211_fixH1(%rsp) movss %xmm4,nb211_fiyH1(%rsp) movss %xmm5,nb211_fizH1(%rsp) ## updated the H1 force movq nb211_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## 11100111 shufps $231,%xmm5,%xmm5 ## 11100111 addss nb211_fixH2(%rsp),%xmm3 addss nb211_fiyH2(%rsp),%xmm4 addss nb211_fizH2(%rsp),%xmm5 movss %xmm3,nb211_fixH2(%rsp) movss %xmm4,nb211_fiyH2(%rsp) movss %xmm5,nb211_fizH2(%rsp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb211_innerk(%rsp) jz _nb_kernel211_x86_64_sse.nb211_updateouterdata jmp _nb_kernel211_x86_64_sse.nb211_odd_loop_nb_kernel211_x86_64_sse.nb211_updateouterdata: movl nb211_ii3(%rsp),%ecx movq nb211_faction(%rbp),%rdi movq nb211_fshift(%rbp),%rsi movl nb211_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb211_fixO(%rsp),%xmm0 movaps nb211_fiyO(%rsp),%xmm1 movaps nb211_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⌨️ 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?