nb_kernel101_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,416 行 · 第 1/3 页
S
1,416 行
.set nb101nf_nn1, 336.set nb101nf_nouter, 340.set nb101nf_ninner, 344 push %rbp movq %rsp,%rbp push %rbx subq $360,%rsp emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb101nf_nouter(%rsp) movl %eax,nb101nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb101nf_nri(%rsp) movq %rsi,nb101nf_iinr(%rsp) movq %rdx,nb101nf_jindex(%rsp) movq %rcx,nb101nf_jjnr(%rsp) movq %r8,nb101nf_shift(%rsp) movq %r9,nb101nf_shiftvec(%rsp) movq nb101nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb101nf_facel(%rsp) ## assume we have at least one i particle - start directly movq nb101nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb101nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb101nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb101nf_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb101nf_iqO(%rsp) movaps %xmm4,nb101nf_iqH(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb101nf_half(%rsp) movss nb101nf_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm1,nb101nf_half(%rsp) movaps %xmm3,nb101nf_three(%rsp)_nb_kernel101nf_x86_64_sse.nb101nf_threadloop: movq nb101nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel101nf_x86_64_sse.nb101nf_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_kernel101nf_x86_64_sse.nb101nf_spinlock ## if(nn1>nri) nn1=nri movl nb101nf_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,nb101nf_n(%rsp) movl %ebx,nb101nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel101nf_x86_64_sse.nb101nf_outerstart jmp _nb_kernel101nf_x86_64_sse.nb101nf_end_nb_kernel101nf_x86_64_sse.nb101nf_outerstart: ## ebx contains number of outer iterations addl nb101nf_nouter(%rsp),%ebx movl %ebx,nb101nf_nouter(%rsp)_nb_kernel101nf_x86_64_sse.nb101nf_outer: movq nb101nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## rbx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb101nf_is3(%rsp) ## store is3 movq nb101nf_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 nb101nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb101nf_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb101nf_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 addss 4(%rax,%rbx,4),%xmm4 addss 8(%rax,%rbx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb101nf_ixO(%rsp) movaps %xmm4,nb101nf_iyO(%rsp) movaps %xmm5,nb101nf_izO(%rsp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%rax,%rbx,4),%xmm0 addss 16(%rax,%rbx,4),%xmm1 addss 20(%rax,%rbx,4),%xmm2 addss 24(%rax,%rbx,4),%xmm3 addss 28(%rax,%rbx,4),%xmm4 addss 32(%rax,%rbx,4),%xmm5 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm0,nb101nf_ixH1(%rsp) movaps %xmm1,nb101nf_iyH1(%rsp) movaps %xmm2,nb101nf_izH1(%rsp) movaps %xmm3,nb101nf_ixH2(%rsp) movaps %xmm4,nb101nf_iyH2(%rsp) movaps %xmm5,nb101nf_izH2(%rsp) ## clear vctot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb101nf_vctot(%rsp) movq nb101nf_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 nb101nf_pos(%rbp),%rsi movq nb101nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb101nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb101nf_ninner(%rsp),%ecx movl %ecx,nb101nf_ninner(%rsp) addl $0,%edx movl %edx,nb101nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel101nf_x86_64_sse.nb101nf_unroll_loop jmp _nb_kernel101nf_x86_64_sse.nb101nf_odd_inner_nb_kernel101nf_x86_64_sse.nb101nf_unroll_loop: ## quad-unroll innerloop here movq nb101nf_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,nb101nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb101nf_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 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb101nf_iqO(%rsp),%xmm3 mulps nb101nf_iqH(%rsp),%xmm4 movaps %xmm3,nb101nf_qqO(%rsp) movaps %xmm4,nb101nf_qqH(%rsp) movq nb101nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx 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 movlps (%rsi,%rcx,4),%xmm5 movss 8(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm6 movhps (%rsi,%rbx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movss 8(%rsi,%rbx,4),%xmm0 movss 8(%rsi,%rdx,4),%xmm1 shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## 10001000 shufps $136,%xmm5,%xmm0 ## 10001000 shufps $221,%xmm5,%xmm1 ## 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb101nf_ixO(%rsp),%xmm4 movaps nb101nf_iyO(%rsp),%xmm5 movaps nb101nf_izO(%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 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb101nf_ixH1(%rsp),%xmm4 movaps nb101nf_iyH1(%rsp),%xmm5 movaps nb101nf_izH1(%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,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb101nf_ixH2(%rsp),%xmm3 movaps nb101nf_iyH2(%rsp),%xmm4 movaps nb101nf_izH2(%rsp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## start with rsqO - seed in xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb101nf_three(%rsp),%xmm4 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb101nf_half(%rsp),%xmm4 movaps %xmm4,%xmm7 ## rinvO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb101nf_three(%rsp),%xmm4 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb101nf_half(%rsp),%xmm4 movaps %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb101nf_three(%rsp),%xmm4 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm4 ## 30-rsq*lu*lu mulps %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulps nb101nf_half(%rsp),%xmm4 movaps %xmm4,%xmm5 ## rinvH2 in xmm5 ## do O interactions mulps nb101nf_qqO(%rsp),%xmm7 ## xmm7=vcoul addps nb101nf_vctot(%rsp),%xmm7 movaps %xmm7,nb101nf_vctot(%rsp) ## H1 interactions mulps nb101nf_qqH(%rsp),%xmm6 ## xmm6=vcoul addps nb101nf_vctot(%rsp),%xmm6 movaps %xmm6,nb101nf_vctot(%rsp) ## H2 interactions mulps nb101nf_qqH(%rsp),%xmm5 ## xmm5=vcoul addps nb101nf_vctot(%rsp),%xmm5 movaps %xmm5,nb101nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb101nf_innerk(%rsp) jl _nb_kernel101nf_x86_64_sse.nb101nf_odd_inner jmp _nb_kernel101nf_x86_64_sse.nb101nf_unroll_loop_nb_kernel101nf_x86_64_sse.nb101nf_odd_inner: addl $4,nb101nf_innerk(%rsp) jnz _nb_kernel101nf_x86_64_sse.nb101nf_odd_loop jmp _nb_kernel101nf_x86_64_sse.nb101nf_updateouterdata_nb_kernel101nf_x86_64_sse.nb101nf_odd_loop: movq nb101nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb101nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb101nf_iqO(%rsp),%xmm4 movq nb101nf_charge(%rbp),%rsi movhps nb101nf_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb101nf_qqO(%rsp) ## use oxygen qq for storage movq nb101nf_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm0 movss 4(%rsi,%rax,4),%xmm1 movss 8(%rsi,%rax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movss nb101nf_ixO(%rsp),%xmm3 movss nb101nf_iyO(%rsp),%xmm4 movss nb101nf_izO(%rsp),%xmm5 movlps nb101nf_ixH1(%rsp),%xmm6 movlps nb101nf_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb101nf_iyH1(%rsp),%xmm6 movlps nb101nf_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb101nf_izH1(%rsp),%xmm6 movlps nb101nf_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 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 nb101nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb101nf_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 nb101nf_qqO(%rsp),%xmm3 mulps %xmm0,%xmm3 ## xmm3=vcoul addps nb101nf_vctot(%rsp),%xmm3 movaps %xmm3,nb101nf_vctot(%rsp) decl nb101nf_innerk(%rsp) jz _nb_kernel101nf_x86_64_sse.nb101nf_updateouterdata jmp _nb_kernel101nf_x86_64_sse.nb101nf_odd_loop_nb_kernel101nf_x86_64_sse.nb101nf_updateouterdata: ## accumulate total potential energy and update it ## get n from stack movl nb101nf_n(%rsp),%esi ## get group index for i particle movq nb101nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] movaps nb101nf_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 nb101nf_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb101nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel101nf_x86_64_sse.nb101nf_outerend ## not last, iterate outer loop once more! movl %esi,nb101nf_n(%rsp) jmp _nb_kernel101nf_x86_64_sse.nb101nf_outer_nb_kernel101nf_x86_64_sse.nb101nf_outerend: ## check if more outer neighborlists remain movl nb101nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel101nf_x86_64_sse.nb101nf_end ## non-zero, do one more workunit jmp _nb_kernel101nf_x86_64_sse.nb101nf_threadloop_nb_kernel101nf_x86_64_sse.nb101nf_end: movl nb101nf_nouter(%rsp),%eax movl nb101nf_ninner(%rsp),%ebx movq nb101nf_outeriter(%rbp),%rcx movq nb101nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $360,%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?