nb_kernel103_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,426 行 · 第 1/3 页
S
1,426 行
mulps nb103_dxH2(%rsp),%xmm10 mulps nb103_dyH2(%rsp),%xmm11 mulps nb103_dzH2(%rsp),%xmm12 mulps nb103_dxM(%rsp),%xmm13 mulps nb103_dyM(%rsp),%xmm14 mulps nb103_dzM(%rsp),%xmm15 movaps %xmm7,%xmm3 movaps %xmm8,%xmm4 addps %xmm9,%xmm2 addps nb103_fixH1(%rsp),%xmm7 addps nb103_fiyH1(%rsp),%xmm8 addps nb103_fizH1(%rsp),%xmm9 addps %xmm10,%xmm3 addps %xmm11,%xmm4 addps %xmm12,%xmm2 addps nb103_fixH2(%rsp),%xmm10 addps nb103_fiyH2(%rsp),%xmm11 addps nb103_fizH2(%rsp),%xmm12 addps %xmm13,%xmm3 addps %xmm14,%xmm4 addps %xmm15,%xmm2 addps nb103_fixM(%rsp),%xmm13 addps nb103_fiyM(%rsp),%xmm14 addps nb103_fizM(%rsp),%xmm15 movaps %xmm7,nb103_fixH1(%rsp) movaps %xmm8,nb103_fiyH1(%rsp) movaps %xmm9,nb103_fizH1(%rsp) movaps %xmm10,nb103_fixH2(%rsp) movaps %xmm11,nb103_fiyH2(%rsp) movaps %xmm12,nb103_fizH2(%rsp) movaps %xmm13,nb103_fixM(%rsp) movaps %xmm14,nb103_fiyM(%rsp) movaps %xmm15,nb103_fizM(%rsp) ## xmm3 = fjx , xmm4 = fjy movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movhlps %xmm2,%xmm3 ## fzc fzd movlps %xmm0,(%rdi,%rax,4) movhps %xmm0,(%rdi,%rbx,4) movlps %xmm1,(%rdi,%rcx,4) movhps %xmm1,(%rdi,%rdx,4) movss %xmm2,8(%rdi,%rax,4) movss %xmm3,8(%rdi,%rcx,4) shufps $1,%xmm2,%xmm2 shufps $1,%xmm3,%xmm3 movss %xmm2,8(%rdi,%rbx,4) movss %xmm3,8(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb103_innerk(%rsp) jl _nb_kernel103_x86_64_sse.nb103_odd_inner jmp _nb_kernel103_x86_64_sse.nb103_unroll_loop_nb_kernel103_x86_64_sse.nb103_odd_inner: addl $4,nb103_innerk(%rsp) jnz _nb_kernel103_x86_64_sse.nb103_odd_loop jmp _nb_kernel103_x86_64_sse.nb103_updateouterdata_nb_kernel103_x86_64_sse.nb103_odd_loop: movq nb103_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb103_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb103_iqM(%rsp),%xmm4 movq nb103_charge(%rbp),%rsi movhps nb103_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb103_qqM(%rsp) ## use dummy qq for storage movq nb103_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 nb103_ixM(%rsp),%xmm0 movss nb103_iyM(%rsp),%xmm1 movss nb103_izM(%rsp),%xmm2 movlps nb103_ixH1(%rsp),%xmm6 movlps nb103_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm0 movlps nb103_iyH1(%rsp),%xmm6 movlps nb103_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm1 movlps nb103_izH1(%rsp),%xmm6 movlps nb103_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use dummy dx for storage movaps %xmm3,nb103_dxM(%rsp) movaps %xmm4,nb103_dyM(%rsp) movaps %xmm5,nb103_dzM(%rsp) 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 nb103_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb103_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 nb103_qqM(%rsp),%xmm3 mulps %xmm0,%xmm3 ## xmm3=vcoul mulps %xmm3,%xmm4 ## xmm4=total fscal addps nb103_vctot(%rsp),%xmm3 movaps nb103_dxM(%rsp),%xmm0 movaps nb103_dyM(%rsp),%xmm1 movaps nb103_dzM(%rsp),%xmm2 movaps %xmm3,nb103_vctot(%rsp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb103_fixM(%rsp),%xmm3 movss nb103_fiyM(%rsp),%xmm4 movss nb103_fizM(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb103_fixM(%rsp) movss %xmm4,nb103_fiyM(%rsp) movss %xmm5,nb103_fizM(%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 nb103_fixH1(%rsp),%xmm3 addss nb103_fiyH1(%rsp),%xmm4 addss nb103_fizH1(%rsp),%xmm5 movss %xmm3,nb103_fixH1(%rsp) movss %xmm4,nb103_fiyH1(%rsp) movss %xmm5,nb103_fizH1(%rsp) ## updated the H1 force movq nb103_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## 11100111 shufps $231,%xmm5,%xmm5 ## 11100111 addss nb103_fixH2(%rsp),%xmm3 addss nb103_fiyH2(%rsp),%xmm4 addss nb103_fizH2(%rsp),%xmm5 movss %xmm3,nb103_fixH2(%rsp) movss %xmm4,nb103_fiyH2(%rsp) movss %xmm5,nb103_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 nb103_innerk(%rsp) jz _nb_kernel103_x86_64_sse.nb103_updateouterdata jmp _nb_kernel103_x86_64_sse.nb103_odd_loop_nb_kernel103_x86_64_sse.nb103_updateouterdata: movl nb103_ii3(%rsp),%ecx movq nb103_faction(%rbp),%rdi movq nb103_fshift(%rbp),%rsi movl nb103_is3(%rsp),%edx ## accumulate H1 forces in xmm0, xmm1, xmm2 movaps nb103_fixH1(%rsp),%xmm0 movaps nb103_fiyH1(%rsp),%xmm1 movaps nb103_fizH1(%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 12(%rdi,%rcx,4),%xmm3 movss 16(%rdi,%rcx,4),%xmm4 movss 20(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,12(%rdi,%rcx,4) movss %xmm4,16(%rdi,%rcx,4) movss %xmm5,20(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## 00001000 ## accumulate H2 i forces in xmm0, xmm1, xmm2 movaps nb103_fixH2(%rsp),%xmm0 movaps nb103_fiyH2(%rsp),%xmm1 movaps nb103_fizH2(%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 24(%rdi,%rcx,4),%xmm3 movss 28(%rdi,%rcx,4),%xmm4 movss 32(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,24(%rdi,%rcx,4) movss %xmm4,28(%rdi,%rcx,4) movss %xmm5,32(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## 00001000 addps %xmm0,%xmm6 ## accumulate M i forces in xmm0, xmm1, xmm2 movaps nb103_fixM(%rsp),%xmm0 movaps nb103_fiyM(%rsp),%xmm1 movaps nb103_fizM(%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 36(%rdi,%rcx,4),%xmm3 movss 40(%rdi,%rcx,4),%xmm4 movss 44(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,36(%rdi,%rcx,4) movss %xmm4,40(%rdi,%rcx,4) movss %xmm5,44(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## 00001000 addps %xmm0,%xmm6 ## increment fshift force movlps (%rsi,%rdx,4),%xmm3 movss 8(%rsi,%rdx,4),%xmm4 subps %xmm6,%xmm3 subss %xmm7,%xmm4 movlps %xmm3,(%rsi,%rdx,4) movss %xmm4,8(%rsi,%rdx,4) ## get n from stack movl nb103_n(%rsp),%esi ## get group index for i particle movq nb103_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb103_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addps %xmm6,%xmm7 ## pos 0-1 in xmm7 have the sum now movaps %xmm7,%xmm6 shufps $1,%xmm6,%xmm6 addss %xmm6,%xmm7 ## add earlier value from mem movq nb103_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb103_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel103_x86_64_sse.nb103_outerend ## not last, iterate outer loop once more! movl %esi,nb103_n(%rsp) jmp _nb_kernel103_x86_64_sse.nb103_outer_nb_kernel103_x86_64_sse.nb103_outerend: ## check if more outer neighborlists remain movl nb103_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel103_x86_64_sse.nb103_end ## non-zero, do one more workunit jmp _nb_kernel103_x86_64_sse.nb103_threadloop_nb_kernel103_x86_64_sse.nb103_end: movl nb103_nouter(%rsp),%eax movl nb103_ninner(%rsp),%ebx movq nb103_outeriter(%rbp),%rcx movq nb103_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $696,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel103nf_x86_64_sse.globl _nb_kernel103nf_x86_64_ssenb_kernel103nf_x86_64_sse: _nb_kernel103nf_x86_64_sse: .set nb103nf_fshift, 16.set nb103nf_gid, 24.set nb103nf_pos, 32.set nb103nf_faction, 40.set nb103nf_charge, 48.set nb103nf_p_facel, 56.set nb103nf_argkrf, 64.set nb103nf_argcrf, 72.set nb103nf_Vc, 80.set nb103nf_type, 88.set nb103nf_p_ntype, 96.set nb103nf_vdwparam, 104.set nb103nf_Vvdw, 112.set nb103nf_p_tabscale, 120.set nb103nf_VFtab, 128.set nb103nf_invsqrta, 136.set nb103nf_dvda, 144.set nb103nf_p_gbtabscale, 152.set nb103nf_GBtab, 160.set nb103nf_p_nthreads, 168.set nb103nf_count, 176.set nb103nf_mtx, 184.set nb103nf_outeriter, 192.set nb103nf_inneriter, 200.set nb103nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb103nf_ixH1, 0.set nb103nf_iyH1, 16.set nb103nf_izH1, 32.set nb103nf_ixH2, 48.set nb103nf_iyH2, 64.set nb103nf_izH2, 80.set nb103nf_ixM, 96.set nb103nf_iyM, 112.set nb103nf_izM, 128.set nb103nf_iqH, 144.set nb103nf_iqM, 160.set nb103nf_vctot, 176.set nb103nf_half, 192.set nb103nf_three, 208.set nb103nf_qqH, 224.set nb103nf_qqM, 240.set nb103nf_is3, 256.set nb103nf_ii3, 260.set nb103nf_nri, 264.set nb103nf_iinr, 272.set nb103nf_jindex, 280.set nb103nf_jjnr, 288.set nb103nf_shift, 296.set nb103nf_shiftvec, 304.set nb103nf_facel, 312.set nb103nf_innerjjnr, 320.set nb103nf_innerk, 328⌨️ 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?