nb_kernel010_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,275 行 · 第 1/3 页
S
1,275 行
movl %ebx,nb010nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel010nf_x86_64_sse.nb010nf_outerstart jmp _nb_kernel010nf_x86_64_sse.nb010nf_end_nb_kernel010nf_x86_64_sse.nb010nf_outerstart: ## ebx contains number of outer iterations addl nb010nf_nouter(%rsp),%ebx movl %ebx,nb010nf_nouter(%rsp)_nb_kernel010nf_x86_64_sse.nb010nf_outer: movq nb010nf_shift(%rsp),%rax ## rax = base of shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb010nf_is3(%rsp) ## store is3 movq nb010nf_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 nb010nf_iinr(%rsp),%rcx ## rcx = base of iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb010nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb010nf_ntype(%rsp),%edx shll %edx movl %edx,nb010nf_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb010nf_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb010nf_ix(%rsp) movaps %xmm1,nb010nf_iy(%rsp) movaps %xmm2,nb010nf_iz(%rsp) movl %ebx,nb010nf_ii3(%rsp) ## clear Vvdwtot and i forces xorps %xmm4,%xmm4 movaps %xmm4,nb010nf_Vvdwtot(%rsp) movq nb010nf_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 nb010nf_pos(%rbp),%rsi movq nb010nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb010nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb010nf_ninner(%rsp),%ecx movl %ecx,nb010nf_ninner(%rsp) addl $0,%edx movl %edx,nb010nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel010nf_x86_64_sse.nb010nf_unroll_loop jmp _nb_kernel010nf_x86_64_sse.nb010nf_finish_inner_nb_kernel010nf_x86_64_sse.nb010nf_unroll_loop: ## quad-unroll innerloop here movq nb010nf_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 ## advance pointer (unrolled 4) addq $16,nb010nf_innerjjnr(%rsp) movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movq nb010nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb010nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb010nf_ntia(%rsp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rax,4),%xmm6 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm6 ## 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb010nf_c6(%rsp) movaps %xmm6,nb010nf_c12(%rsp) movq nb010nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx mulps %xmm2,%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 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 ix-iz to xmm4-xmm6 movaps nb010nf_ix(%rsp),%xmm4 movaps nb010nf_iy(%rsp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## 1/x lookup seed in xmm5 movaps nb010nf_two(%rsp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%rsp),%xmm1 mulps nb010nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb010nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb010nf_Vvdwtot(%rsp) ## should we do one more iteration? subl $4,nb010nf_innerk(%rsp) jl _nb_kernel010nf_x86_64_sse.nb010nf_finish_inner jmp _nb_kernel010nf_x86_64_sse.nb010nf_unroll_loop_nb_kernel010nf_x86_64_sse.nb010nf_finish_inner: ## check if at least two particles remain addl $4,nb010nf_innerk(%rsp) movl nb010nf_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel010nf_x86_64_sse.nb010nf_dopair jmp _nb_kernel010nf_x86_64_sse.nb010nf_checksingle_nb_kernel010nf_x86_64_sse.nb010nf_dopair: movq nb010nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movl 4(%rcx),%ebx addq $8,nb010nf_innerjjnr(%rsp) movq nb010nf_type(%rbp),%rsi movl %eax,%ecx movl %ebx,%edx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb010nf_vdwparam(%rbp),%rsi shll %ecx shll %edx movl nb010nf_ntia(%rsp),%edi addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rcx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm6 movq nb010nf_pos(%rbp),%rdi xorps %xmm7,%xmm7 movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## 00001000 shufps $13,%xmm6,%xmm6 ## 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb010nf_c6(%rsp) movaps %xmm6,nb010nf_c12(%rsp) 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 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 nb010nf_ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb010nf_ix(%rsp),%xmm4 movaps nb010nf_iy(%rsp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## 1/x lookup seed in xmm5 movaps nb010nf_two(%rsp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%rsp),%xmm1 mulps nb010nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb010nf_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb010nf_Vvdwtot(%rsp)_nb_kernel010nf_x86_64_sse.nb010nf_checksingle: movl nb010nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel010nf_x86_64_sse.nb010nf_dosingle jmp _nb_kernel010nf_x86_64_sse.nb010nf_updateouterdata_nb_kernel010nf_x86_64_sse.nb010nf_dosingle: movq nb010nf_pos(%rbp),%rdi movq nb010nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movq nb010nf_type(%rbp),%rsi movl %eax,%ecx movl (%rsi,%rcx,4),%ecx movq nb010nf_vdwparam(%rbp),%rsi shll %ecx addl nb010nf_ntia(%rsp),%ecx xorps %xmm6,%xmm6 movlps (%rsi,%rcx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## 11111100 shufps $253,%xmm6,%xmm6 ## 11111101 movaps %xmm4,nb010nf_c6(%rsp) movaps %xmm6,nb010nf_c12(%rsp) 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 xorps %xmm7,%xmm7 movaps nb010nf_ix(%rsp),%xmm4 movaps nb010nf_iy(%rsp),%xmm5 movaps nb010nf_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 rcpps %xmm4,%xmm5 ## 1/x lookup seed in xmm5 movaps nb010nf_two(%rsp),%xmm0 mulps %xmm5,%xmm4 subps %xmm4,%xmm0 mulps %xmm5,%xmm0 ## xmm0=rinvsq movaps %xmm0,%xmm4 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb010nf_c6(%rsp),%xmm1 mulps nb010nf_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addss nb010nf_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb010nf_Vvdwtot(%rsp)_nb_kernel010nf_x86_64_sse.nb010nf_updateouterdata: ## get n from stack movl nb010nf_n(%rsp),%esi ## get group index for i particle movq nb010nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total lj energy and update it movaps nb010nf_Vvdwtot(%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 nb010nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb010nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel010nf_x86_64_sse.nb010nf_outerend ## not last, iterate outer loop once more! movl %esi,nb010nf_n(%rsp) jmp _nb_kernel010nf_x86_64_sse.nb010nf_outer_nb_kernel010nf_x86_64_sse.nb010nf_outerend: ## check if more outer neighborlists remain movl nb010nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel010nf_x86_64_sse.nb010nf_end ## non-zero, do one more workunit jmp _nb_kernel010nf_x86_64_sse.nb010nf_threadloop_nb_kernel010nf_x86_64_sse.nb010nf_end: movl nb010nf_nouter(%rsp),%eax movl nb010nf_ninner(%rsp),%ebx movq nb010nf_outeriter(%rbp),%rcx movq nb010nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $264,%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?