nb_kernel200_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,212 行 · 第 1/3 页
S
1,212 行
movss (%rsi,%rbx,4),%xmm1 unpcklps %xmm1,%xmm0 ## jqa jqb - - mulps nb200_iq(%rsp),%xmm0 ##qq lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## load coordinates movq nb200_pos(%rbp),%rdi movlps (%rdi,%rax,4),%xmm4 ## x1 y1 - - movlps (%rdi,%rbx,4),%xmm5 ## x2 y2 - - movss 8(%rdi,%rax,4),%xmm6 ## z1 - - - movss 8(%rdi,%rbx,4),%xmm7 ## z2 - - - unpcklps %xmm5,%xmm4 ## x1 x2 y1 y2 movhlps %xmm4,%xmm5 ## y1 y2 - - unpcklps %xmm7,%xmm6 ## z1 z2 - - ## calc dr subps nb200_ix(%rsp),%xmm4 subps nb200_iy(%rsp),%xmm5 subps nb200_iz(%rsp),%xmm6 ## store dr in xmm9-xmm11 movaps %xmm4,%xmm9 movaps %xmm5,%xmm10 movaps %xmm6,%xmm11 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 movaps nb200_krf(%rsp),%xmm7 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb200_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb200_half(%rsp),%xmm3 mulps %xmm4,%xmm7 ## xmm7=krsq subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm3 ## xmm3=rinv movaps %xmm3,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm3,%xmm6 addps %xmm7,%xmm6 ## xmm6=rinv+ krsq subps nb200_crf(%rsp),%xmm6 ## xmm6=rinv+ krsq-crf mulps %xmm0,%xmm6 ## xmm6=vcoul=qq*(rinv+ krsq) addps %xmm7,%xmm7 subps %xmm7,%xmm3 mulps %xmm3,%xmm0 mulps %xmm0,%xmm4 ## xmm4=total fscal xorps %xmm5,%xmm5 movlhps %xmm5,%xmm6 ## add potential to vctot (sum in xmm12) addps %xmm6,%xmm12 ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm4,%xmm9 mulps %xmm4,%xmm10 mulps %xmm4,%xmm11 movlhps %xmm5,%xmm9 movlhps %xmm5,%xmm10 movlhps %xmm5,%xmm11 ## xmm0-xmm2 contains tx-tz (partial force) ## accumulate i forces addps %xmm9,%xmm13 addps %xmm10,%xmm14 addps %xmm11,%xmm15 movq nb200_faction(%rbp),%rsi ## the fj's - start by accumulating x & y forces from memory movlps (%rsi,%rax,4),%xmm0 ## x1 y1 - - movhps (%rsi,%rbx,4),%xmm0 ## x1 y1 x2 y2 unpcklps %xmm10,%xmm9 ## x1 y1 x2 y2 addps %xmm9,%xmm0 movlps %xmm0,(%rsi,%rax,4) movhps %xmm0,(%rsi,%rbx,4) ## z forces pshufd $1,%xmm11,%xmm8 addss 8(%rsi,%rax,4),%xmm11 addss 8(%rsi,%rbx,4),%xmm8 movss %xmm11,8(%rsi,%rax,4) movss %xmm8,8(%rsi,%rbx,4)_nb_kernel200_x86_64_sse.nb200_checksingle: movl nb200_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel200_x86_64_sse.nb200_dosingle jmp _nb_kernel200_x86_64_sse.nb200_updateouterdata_nb_kernel200_x86_64_sse.nb200_dosingle: movq nb200_innerjjnr(%rsp),%rcx movl (%rcx),%eax movq nb200_charge(%rbp),%rsi movss (%rsi,%rax,4),%xmm0 ## jq mulss nb200_iq(%rsp),%xmm0 ## qq lea (%rax,%rax,2),%rax ## replace jnr with j3 movq nb200_pos(%rbp),%rdi movss (%rdi,%rax,4),%xmm4 ## x1 - - - movss 4(%rdi,%rax,4),%xmm5 ## y2 - - - movss 8(%rdi,%rax,4),%xmm6 ## 13 - - - ## calc dr subss nb200_ix(%rsp),%xmm4 subss nb200_iy(%rsp),%xmm5 subss nb200_iz(%rsp),%xmm6 ## store dr in xmm9-xmm11 movaps %xmm4,%xmm9 movaps %xmm5,%xmm10 movaps %xmm6,%xmm11 ## square it mulss %xmm4,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm6 addss %xmm5,%xmm4 addss %xmm6,%xmm4 ## rsq in xmm4 movaps nb200_krf(%rsp),%xmm7 rsqrtss %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulss %xmm5,%xmm5 movaps nb200_three(%rsp),%xmm1 mulss %xmm4,%xmm5 ## rsq*lu*lu movaps nb200_half(%rsp),%xmm3 mulss %xmm4,%xmm7 ## xmm7=krsq subss %xmm5,%xmm1 ## 30-rsq*lu*lu mulss %xmm2,%xmm1 mulss %xmm1,%xmm3 ## xmm3=rinv movaps %xmm3,%xmm4 mulss %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm3,%xmm6 addss %xmm7,%xmm6 ## xmm6=rinv+ krsq subss nb200_crf(%rsp),%xmm6 ## xmm6=rinv+ krsq-crf mulss %xmm0,%xmm6 ## xmm6=vcoul=qq*(rinv+krsq-crf) addss %xmm7,%xmm7 subss %xmm7,%xmm3 mulss %xmm3,%xmm0 mulss %xmm0,%xmm4 ## xmm4=total fscal ## add potential to vctot (sum in xmm12) addss %xmm6,%xmm12 ## calculate scalar force by multiplying dx/dy/dz with fscal mulss %xmm4,%xmm9 mulss %xmm4,%xmm10 mulss %xmm4,%xmm11 ## xmm0-xmm2 contains tx-tz (partial force) ## accumulate i forces addss %xmm9,%xmm13 addss %xmm10,%xmm14 addss %xmm11,%xmm15 movq nb200_faction(%rbp),%rsi ## add to j forces addss (%rsi,%rax,4),%xmm9 addss 4(%rsi,%rax,4),%xmm10 addss 8(%rsi,%rax,4),%xmm11 movss %xmm9,(%rsi,%rax,4) movss %xmm10,4(%rsi,%rax,4) movss %xmm11,8(%rsi,%rax,4)_nb_kernel200_x86_64_sse.nb200_updateouterdata: movl nb200_ii3(%rsp),%ecx movq nb200_faction(%rbp),%rdi movq nb200_fshift(%rbp),%rsi movl nb200_is3(%rsp),%edx ## accumulate i forces in xmm13, xmm14, xmm15 movhlps %xmm13,%xmm0 movhlps %xmm14,%xmm1 movhlps %xmm15,%xmm2 addps %xmm13,%xmm0 addps %xmm14,%xmm1 addps %xmm15,%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 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%rdi,%rcx,4) movss %xmm4,4(%rdi,%rcx,4) movss %xmm5,8(%rdi,%rcx,4) ## increment fshift force movss (%rsi,%rdx,4),%xmm3 movss 4(%rsi,%rdx,4),%xmm4 movss 8(%rsi,%rdx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%rsi,%rdx,4) movss %xmm4,4(%rsi,%rdx,4) movss %xmm5,8(%rsi,%rdx,4) ## get n from stack movl nb200_n(%rsp),%esi ## get group index for i particle movq nb200_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it ## accumulate movhlps %xmm12,%xmm6 addps %xmm6,%xmm12 ## pos 0-1 in xmm12 have the sum now movaps %xmm12,%xmm6 shufps $1,%xmm6,%xmm6 addss %xmm6,%xmm12 ## add earlier value from mem movq nb200_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm12 ## move back to mem movss %xmm12,(%rax,%rdx,4) ## finish if last movl nb200_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel200_x86_64_sse.nb200_outerend ## not last, iterate outer loop once more! movl %esi,nb200_n(%rsp) jmp _nb_kernel200_x86_64_sse.nb200_outer_nb_kernel200_x86_64_sse.nb200_outerend: ## check if more outer neighborlists remain movl nb200_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel200_x86_64_sse.nb200_end ## non-zero, do one more workunit jmp _nb_kernel200_x86_64_sse.nb200_threadloop_nb_kernel200_x86_64_sse.nb200_end: movl nb200_nouter(%rsp),%eax movl nb200_ninner(%rsp),%ebx movq nb200_outeriter(%rbp),%rcx movq nb200_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $360,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel200nf_x86_64_sse.globl _nb_kernel200nf_x86_64_ssenb_kernel200nf_x86_64_sse: _nb_kernel200nf_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb200nf_fshift, 16.set nb200nf_gid, 24.set nb200nf_pos, 32.set nb200nf_faction, 40.set nb200nf_charge, 48.set nb200nf_p_facel, 56.set nb200nf_argkrf, 64.set nb200nf_argcrf, 72.set nb200nf_Vc, 80.set nb200nf_type, 88.set nb200nf_p_ntype, 96.set nb200nf_vdwparam, 104.set nb200nf_Vvdw, 112.set nb200nf_p_tabscale, 120.set nb200nf_VFtab, 128.set nb200nf_invsqrta, 136.set nb200nf_dvda, 144.set nb200nf_p_gbtabscale, 152.set nb200nf_GBtab, 160.set nb200nf_p_nthreads, 168.set nb200nf_count, 176.set nb200nf_mtx, 184.set nb200nf_outeriter, 192.set nb200nf_inneriter, 200.set nb208nf_work, 200 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb200nf_ix, 0.set nb200nf_iy, 16.set nb200nf_iz, 32.set nb200nf_iq, 48.set nb200nf_vctot, 64.set nb200nf_half, 80.set nb200nf_three, 96.set nb200nf_krf, 112.set nb200nf_crf, 128.set nb200nf_is3, 144.set nb200nf_ii3, 148.set nb200nf_innerjjnr, 152.set nb200nf_nri, 160.set nb200nf_iinr, 168.set nb200nf_jindex, 176.set nb200nf_jjnr, 184.set nb200nf_shift, 192.set nb200nf_shiftvec, 200.set nb200nf_facel, 208.set nb200nf_innerk, 216.set nb200nf_n, 220.set nb200nf_nn1, 224.set nb200nf_nouter, 228.set nb200nf_ninner, 232 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $248,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb200nf_nouter(%rsp) movl %eax,nb200nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb200nf_nri(%rsp) movq %rsi,nb200nf_iinr(%rsp) movq %rdx,nb200nf_jindex(%rsp) movq %rcx,nb200nf_jjnr(%rsp) movq %r8,nb200nf_shift(%rsp) movq %r9,nb200nf_shiftvec(%rsp) movq nb200nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb200nf_facel(%rsp) movq nb200nf_argkrf(%rbp),%rsi movq nb200nf_argcrf(%rbp),%rdi movss (%rsi),%xmm1 movss (%rdi),%xmm2 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm1,nb200nf_krf(%rsp) movaps %xmm2,nb200nf_crf(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb200nf_half(%rsp) movss nb200nf_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2⌨️ 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?