nb_kernel103_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,598 行 · 第 1/4 页
S
1,598 行
## twice unrolled innerloop here movq nb103nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb103nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb103nf_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm6 ## jq A movhpd (%rsi,%rbx,8),%xmm6 ## jq B movapd nb103nf_iqM(%rsp),%xmm3 movapd nb103nf_iqH(%rsp),%xmm4 mulpd %xmm6,%xmm3 ## qqM mulpd %xmm6,%xmm4 ## qqH movapd %xmm3,nb103nf_qqM(%rsp) movapd %xmm4,nb103nf_qqH(%rsp) movq nb103nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move two coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm0 movhpd 8(%rsi,%rbx,8),%xmm1 movhpd 16(%rsi,%rbx,8),%xmm2 ## move ixM-izM to xmm4-xmm6 movapd nb103nf_ixM(%rsp),%xmm4 movapd nb103nf_iyM(%rsp),%xmm5 movapd nb103nf_izM(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqM in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movapd nb103nf_ixH1(%rsp),%xmm4 movapd nb103nf_iyH1(%rsp),%xmm5 movapd nb103nf_izH1(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm6 addpd %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movapd nb103nf_ixH2(%rsp),%xmm3 movapd nb103nf_iyH2(%rsp),%xmm4 movapd nb103nf_izH2(%rsp),%xmm5 ## calc dr subpd %xmm0,%xmm3 subpd %xmm1,%xmm4 subpd %xmm2,%xmm5 ## square it mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 addpd %xmm4,%xmm5 addpd %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqM in xmm7 ## start with rsqM - put seed in xmm2 cvtpd2ps %xmm7,%xmm2 rsqrtps %xmm2,%xmm2 cvtps2pd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulpd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulpd %xmm7,%xmm2 ## rsq*lu*lu subpd %xmm2,%xmm4 ## 30-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm4,%xmm7 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subpd %xmm7,%xmm4 ## 3-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm7 ## rinvM in xmm7 ## rsqH1 - seed in xmm2 cvtpd2ps %xmm6,%xmm2 rsqrtps %xmm2,%xmm2 cvtps2pd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulpd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulpd %xmm6,%xmm2 ## rsq*lu*lu subpd %xmm2,%xmm4 ## 30-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm4,%xmm6 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subpd %xmm6,%xmm4 ## 3-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 cvtpd2ps %xmm5,%xmm2 rsqrtps %xmm2,%xmm2 cvtps2pd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulpd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulpd %xmm5,%xmm2 ## rsq*lu*lu subpd %xmm2,%xmm4 ## 30-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm4,%xmm5 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subpd %xmm5,%xmm4 ## 3-rsq*lu*lu mulpd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulpd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm5 ## rinvH2 in xmm5 ## do M interactions mulpd nb103nf_qqM(%rsp),%xmm7 ## xmm7=vcoul addpd nb103nf_vctot(%rsp),%xmm7 movapd %xmm7,nb103nf_vctot(%rsp) ## H1 interactions mulpd nb103nf_qqH(%rsp),%xmm6 ## xmm6=vcoul addpd nb103nf_vctot(%rsp),%xmm6 movapd %xmm6,nb103nf_vctot(%rsp) ## H2 interactions mulpd nb103nf_qqH(%rsp),%xmm5 ## xmm5=vcoul addpd nb103nf_vctot(%rsp),%xmm5 movapd %xmm5,nb103nf_vctot(%rsp) ## should we do one more iteration? subl $2,nb103nf_innerk(%rsp) jl _nb_kernel103nf_x86_64_sse2.nb103nf_checksingle jmp _nb_kernel103nf_x86_64_sse2.nb103nf_unroll_loop_nb_kernel103nf_x86_64_sse2.nb103nf_checksingle: movl nb103nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel103nf_x86_64_sse2.nb103nf_dosingle jmp _nb_kernel103nf_x86_64_sse2.nb103nf_updateouterdata_nb_kernel103nf_x86_64_sse2.nb103nf_dosingle: movq nb103nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movq nb103nf_charge(%rbp),%rsi ## base of charge[] xorpd %xmm6,%xmm6 movlpd (%rsi,%rax,8),%xmm6 ## jq A movapd nb103nf_iqM(%rsp),%xmm3 movapd nb103nf_iqH(%rsp),%xmm4 mulsd %xmm6,%xmm3 ## qqM mulsd %xmm6,%xmm4 ## qqH movapd %xmm3,nb103nf_qqM(%rsp) movapd %xmm4,nb103nf_qqH(%rsp) movq nb103nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 ## move ixM-izM to xmm4-xmm6 movapd nb103nf_ixM(%rsp),%xmm4 movapd nb103nf_iyM(%rsp),%xmm5 movapd nb103nf_izM(%rsp),%xmm6 ## calc dr subsd %xmm0,%xmm4 subsd %xmm1,%xmm5 subsd %xmm2,%xmm6 ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqM in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movapd nb103nf_ixH1(%rsp),%xmm4 movapd nb103nf_iyH1(%rsp),%xmm5 movapd nb103nf_izH1(%rsp),%xmm6 ## calc dr subsd %xmm0,%xmm4 subsd %xmm1,%xmm5 subsd %xmm2,%xmm6 ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm6 addsd %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movapd nb103nf_ixH2(%rsp),%xmm3 movapd nb103nf_iyH2(%rsp),%xmm4 movapd nb103nf_izH2(%rsp),%xmm5 ## calc dr subsd %xmm0,%xmm3 subsd %xmm1,%xmm4 subsd %xmm2,%xmm5 ## square it mulsd %xmm3,%xmm3 mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 addsd %xmm4,%xmm5 addsd %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqM in xmm7 ## start with rsqM - put seed in xmm2 cvtsd2ss %xmm7,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulsd %xmm7,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm7 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subsd %xmm7,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm7 ## rinvM in xmm7 ## rsqH1 - seed in xmm2 cvtsd2ss %xmm6,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulsd %xmm6,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm6 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subsd %xmm6,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 cvtsd2ss %xmm5,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb103nf_three(%rsp),%xmm4 mulsd %xmm5,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm5 ## rsq*lu*lu movapd nb103nf_three(%rsp),%xmm4 subsd %xmm5,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb103nf_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm5 ## rinvH2 in xmm5 ## do M interactions mulsd nb103nf_qqM(%rsp),%xmm7 ## xmm7=vcoul addsd nb103nf_vctot(%rsp),%xmm7 movlpd %xmm7,nb103nf_vctot(%rsp) ## H1 interactions mulsd nb103nf_qqH(%rsp),%xmm6 ## xmm6=vcoul addsd nb103nf_vctot(%rsp),%xmm6 movlpd %xmm6,nb103nf_vctot(%rsp) ## H2 interactions mulsd nb103nf_qqH(%rsp),%xmm5 ## xmm5=vcoul addsd nb103nf_vctot(%rsp),%xmm5 movlpd %xmm5,nb103nf_vctot(%rsp)_nb_kernel103nf_x86_64_sse2.nb103nf_updateouterdata: ## get n from stack movl nb103nf_n(%rsp),%esi ## get group index for i particle movq nb103nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] movapd nb103nf_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## pos 0-1 in xmm7 have the sum now ## add earlier value from mem movq nb103nf_Vc(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## finish if last movl nb103nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel103nf_x86_64_sse2.nb103nf_outerend ## not last, iterate outer loop once more! movl %esi,nb103nf_n(%rsp) jmp _nb_kernel103nf_x86_64_sse2.nb103nf_outer_nb_kernel103nf_x86_64_sse2.nb103nf_outerend: ## check if more outer neighborlists remain movl nb103nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel103nf_x86_64_sse2.nb103nf_end ## non-zero, do one more workunit jmp _nb_kernel103nf_x86_64_sse2.nb103nf_threadloop_nb_kernel103nf_x86_64_sse2.nb103nf_end: movl nb103nf_nouter(%rsp),%eax movl nb103nf_ninner(%rsp),%ebx movq nb103nf_outeriter(%rbp),%rcx movq nb103nf_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⌨️ 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?