nb_kernel100_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,140 行 · 第 1/3 页
S
1,140 行
_nb_kernel100nf_x86_64_sse.nb100nf_threadloop: movq nb100nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel100nf_x86_64_sse.nb100nf_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%rsi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel100nf_x86_64_sse.nb100nf_spinlock ## if(nn1>nri) nn1=nri movl nb100nf_nri(%rsp),%ecx movl %ecx,%edx subl %ebx,%ecx cmovlel %edx,%ebx ## if(nn1>nri) nn1=nri ## Cleared the spinlock if we got here. ## eax contains nn0, ebx contains nn1. movl %eax,nb100nf_n(%rsp) movl %ebx,nb100nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel100nf_x86_64_sse.nb100nf_outerstart jmp _nb_kernel100nf_x86_64_sse.nb100nf_end_nb_kernel100nf_x86_64_sse.nb100nf_outerstart: ## ebx contains number of outer iterations addl nb100nf_nouter(%rsp),%ebx movl %ebx,nb100nf_nouter(%rsp)_nb_kernel100nf_x86_64_sse.nb100nf_outer: movq nb100nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb100nf_is3(%rsp) ## store is3 movq nb100nf_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb100nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb100nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb100nf_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb100nf_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 movaps %xmm3,nb100nf_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb100nf_ix(%rsp) movaps %xmm1,nb100nf_iy(%rsp) movaps %xmm2,nb100nf_iz(%rsp) movl %ebx,nb100nf_ii3(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb100nf_vctot(%rsp) movq nb100nf_jindex(%rsp),%rax movl (%rax,%rsi,4),%ecx ## jindex[n] movl 4(%rax,%rsi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movq nb100nf_pos(%rbp),%rsi movq nb100nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb100nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb100nf_ninner(%rsp),%ecx movl %ecx,nb100nf_ninner(%rsp) addl $0,%edx movl %edx,nb100nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel100nf_x86_64_sse.nb100nf_unroll_loop jmp _nb_kernel100nf_x86_64_sse.nb100nf_finish_inner_nb_kernel100nf_x86_64_sse.nb100nf_unroll_loop: ## quad-unrolled innerloop here movq nb100nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb100nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb100nf_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 movaps nb100nf_iq(%rsp),%xmm5 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 movq nb100nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx mulps %xmm5,%xmm3 lea (%rcx,%rcx,2),%rcx ## replace jnr with j3 lea (%rdx,%rdx,2),%rdx ## move four coordinates to xmm0-xmm2 movlps (%rsi,%rax,4),%xmm4 ## x1 y1 - - movlps (%rsi,%rcx,4),%xmm5 ## x3 y3 - - movss 8(%rsi,%rax,4),%xmm2 ## z1 - - - movss 8(%rsi,%rcx,4),%xmm6 ## z3 - - - movhps (%rsi,%rbx,4),%xmm4 ## x1 y1 x2 y2 movhps (%rsi,%rdx,4),%xmm5 ## x3 y3 x4 y4 movss 8(%rsi,%rbx,4),%xmm0 ## z2 - - - movss 8(%rsi,%rdx,4),%xmm1 ## z4 - - - shufps $0,%xmm0,%xmm2 ## z1 z1 z2 z2 shufps $0,%xmm1,%xmm6 ## z3 z3 z4 z4 movaps %xmm4,%xmm0 ## x1 y1 x2 y2 movaps %xmm4,%xmm1 ## x1 y1 x2 y2 shufps $136,%xmm6,%xmm2 ## 10001000 ;# z1 z2 z3 z4 shufps $136,%xmm5,%xmm0 ## 10001000 ;# x1 x2 x3 x4 shufps $221,%xmm5,%xmm1 ## 11011101 ;# y1 y2 y3 y4 ## move nb100nf_ix-iz to xmm4-xmm6 movaps nb100nf_ix(%rsp),%xmm4 movaps nb100nf_iy(%rsp),%xmm5 movaps nb100nf_iz(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps nb100nf_vctot(%rsp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addps %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb100nf_innerk(%rsp) jl _nb_kernel100nf_x86_64_sse.nb100nf_finish_inner jmp _nb_kernel100nf_x86_64_sse.nb100nf_unroll_loop_nb_kernel100nf_x86_64_sse.nb100nf_finish_inner: ## check if at least two particles remain addl $4,nb100nf_innerk(%rsp) movl nb100nf_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel100nf_x86_64_sse.nb100nf_dopair jmp _nb_kernel100nf_x86_64_sse.nb100nf_checksingle_nb_kernel100nf_x86_64_sse.nb100nf_dopair: movq nb100nf_charge(%rbp),%rsi movq nb100nf_pos(%rbp),%rdi movq nb100nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movl 4(%rcx),%ebx addq $8,nb100nf_innerjjnr(%rsp) movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rbx,4),%xmm6 shufps $0,%xmm6,%xmm3 shufps $8,%xmm3,%xmm3 ## 00001000 ;# xmm3(0,1) has the charges lea (%rax,%rax,2),%rax lea (%rbx,%rbx,2),%rbx ## move coordinates to xmm0-xmm2 movlps (%rdi,%rax,4),%xmm1 movss 8(%rdi,%rax,4),%xmm2 movhps (%rdi,%rbx,4),%xmm1 movss 8(%rdi,%rbx,4),%xmm0 mulps nb100nf_iq(%rsp),%xmm3 xorps %xmm7,%xmm7 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## 10001000 shufps $136,%xmm0,%xmm0 ## 10001000 shufps $221,%xmm1,%xmm1 ## 11011101 ## move nb100nf_ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb100nf_ix(%rsp),%xmm4 movaps nb100nf_iy(%rsp),%xmm5 movaps nb100nf_iz(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps nb100nf_vctot(%rsp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addps %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%rsp)_nb_kernel100nf_x86_64_sse.nb100nf_checksingle: movl nb100nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel100nf_x86_64_sse.nb100nf_dosingle jmp _nb_kernel100nf_x86_64_sse.nb100nf_updateouterdata_nb_kernel100nf_x86_64_sse.nb100nf_dosingle: movq nb100nf_charge(%rbp),%rsi movq nb100nf_pos(%rbp),%rdi movq nb100nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movss (%rsi,%rax,4),%xmm3 ## xmm3(0) has the charge lea (%rax,%rax,2),%rax ## move coordinates to xmm0-xmm2 movss (%rdi,%rax,4),%xmm0 movss 4(%rdi,%rax,4),%xmm1 movss 8(%rdi,%rax,4),%xmm2 mulps nb100nf_iq(%rsp),%xmm3 xorps %xmm7,%xmm7 movaps nb100nf_ix(%rsp),%xmm4 movaps nb100nf_iy(%rsp),%xmm5 movaps nb100nf_iz(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb100nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb100nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps nb100nf_vctot(%rsp),%xmm5 mulps %xmm0,%xmm3 ## xmm3=vcoul addss %xmm3,%xmm5 movaps %xmm5,nb100nf_vctot(%rsp)_nb_kernel100nf_x86_64_sse.nb100nf_updateouterdata: ## get n from stack movl nb100nf_n(%rsp),%esi ## get group index for i particle movq nb100nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb100nf_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 nb100nf_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb100nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel100nf_x86_64_sse.nb100nf_outerend ## not last, iterate outer loop once more! movl %esi,nb100nf_n(%rsp) jmp _nb_kernel100nf_x86_64_sse.nb100nf_outer_nb_kernel100nf_x86_64_sse.nb100nf_outerend: ## check if more outer neighborlists remain movl nb100nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel100nf_x86_64_sse.nb100nf_end ## non-zero, do one more workunit jmp _nb_kernel100nf_x86_64_sse.nb100nf_threadloop_nb_kernel100nf_x86_64_sse.nb100nf_end: movl nb100nf_nouter(%rsp),%eax movl nb100nf_ninner(%rsp),%ebx movq nb100nf_outeriter(%rbp),%rcx movq nb100nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $216,%rsp emms pop %rbx pop %rbp ret⌨️ 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?