nb_kernel303_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,919 行 · 第 1/4 页
S
1,919 行
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 ixH1-izH1 to xmm4-xmm6 movaps nb303nf_ixH1(%rsp),%xmm4 movaps nb303nf_iyH1(%rsp),%xmm5 movaps nb303nf_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,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqH1 in xmm7 ## move ixH2-izH2 to xmm4-xmm6 movaps nb303nf_ixH2(%rsp),%xmm4 movaps nb303nf_iyH2(%rsp),%xmm5 movaps nb303nf_izH2(%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 ## rsqH2 in xmm6 ## move ixM-izM to xmm3-xmm5 movaps nb303nf_ixM(%rsp),%xmm3 movaps nb303nf_iyM(%rsp),%xmm4 movaps nb303nf_izM(%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 ## rsqM in xmm5, rsqH2 in xmm6, rsqH1 in xmm7 ## start with rsqH1 - seed to xmm2 rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb303nf_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 nb303nf_half(%rsp),%xmm4 movaps %xmm4,nb303nf_rinvH1(%rsp) ## rinvH1 in xmm4 mulps %xmm4,%xmm7 movaps %xmm7,nb303nf_rH1(%rsp) ## rsqH2 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb303nf_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 nb303nf_half(%rsp),%xmm4 movaps %xmm4,nb303nf_rinvH2(%rsp) ## rinvH2 in xmm4 mulps %xmm4,%xmm6 movaps %xmm6,nb303nf_rH2(%rsp) ## rsqM - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb303nf_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 nb303nf_half(%rsp),%xmm4 movaps %xmm4,nb303nf_rinvM(%rsp) ## rinvM in xmm4 mulps %xmm4,%xmm5 movaps %xmm5,nb303nf_rM(%rsp) ## do H1 interactions ## rH1 is still in xmm7 mulps nb303nf_tsc(%rsp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movq nb303nf_VFtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb303nf_qqH(%rsp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%rsp),%xmm5 movaps %xmm5,nb303nf_vctot(%rsp) ## Done with H1 interactions - now H2! movaps nb303nf_rH2(%rsp),%xmm7 mulps nb303nf_tsc(%rsp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb303nf_qqH(%rsp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%rsp),%xmm5 movaps %xmm5,nb303nf_vctot(%rsp) ## Done with H2, finally we do M interactions movaps nb303nf_rM(%rsp),%xmm7 mulps nb303nf_tsc(%rsp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm5 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rbx,4),%xmm7 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb303nf_qqM(%rsp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%rsp),%xmm5 movaps %xmm5,nb303nf_vctot(%rsp) ## should we do one more iteration? subl $4,nb303nf_innerk(%rsp) jl _nb_kernel303nf_x86_64_sse.nb303nf_odd_inner jmp _nb_kernel303nf_x86_64_sse.nb303nf_unroll_loop_nb_kernel303nf_x86_64_sse.nb303nf_odd_inner: addl $4,nb303nf_innerk(%rsp) jnz _nb_kernel303nf_x86_64_sse.nb303nf_odd_loop jmp _nb_kernel303nf_x86_64_sse.nb303nf_updateouterdata_nb_kernel303nf_x86_64_sse.nb303nf_odd_loop: movq nb303nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb303nf_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb303nf_iqM(%rsp),%xmm4 movq nb303nf_charge(%rbp),%rsi movhps nb303nf_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb303nf_qqM(%rsp) ## use dummy qq for storage movq nb303nf_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 nb303nf_ixM(%rsp),%xmm3 movss nb303nf_iyM(%rsp),%xmm4 movss nb303nf_izM(%rsp),%xmm5 movlps nb303nf_ixH1(%rsp),%xmm6 movlps nb303nf_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm3 movlps nb303nf_iyH1(%rsp),%xmm6 movlps nb303nf_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm4 movlps nb303nf_izH1(%rsp),%xmm6 movlps nb303nf_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 nb303nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb303nf_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 mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm0,nb303nf_rinvM(%rsp) mulps nb303nf_tsc(%rsp),%xmm4 movhlps %xmm4,%xmm7 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm7,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm7 movlhps %xmm7,%xmm3 subps %xmm3,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %eax,%mm0 movd %ecx,%mm1 movd %edx,%mm2 movq nb303nf_VFtab(%rbp),%rsi movd %mm6,%eax movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp movaps nb303nf_qqM(%rsp),%xmm0 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV ## at this point mm5 contains vcoul ## increment vcoul addps nb303nf_vctot(%rsp),%xmm5 movaps %xmm5,nb303nf_vctot(%rsp) decl nb303nf_innerk(%rsp) jz _nb_kernel303nf_x86_64_sse.nb303nf_updateouterdata jmp _nb_kernel303nf_x86_64_sse.nb303nf_odd_loop_nb_kernel303nf_x86_64_sse.nb303nf_updateouterdata: ## get n from stack movl nb303nf_n(%rsp),%esi ## get group index for i particle movq nb303nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb303nf_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 nb303nf_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb303nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel303nf_x86_64_sse.nb303nf_outerend ## not last, iterate outer loop once more! movl %esi,nb303nf_n(%rsp) jmp _nb_kernel303nf_x86_64_sse.nb303nf_outer_nb_kernel303nf_x86_64_sse.nb303nf_outerend: ## check if more outer neighborlists remain movl nb303nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel303nf_x86_64_sse.nb303nf_end ## non-zero, do one more workunit jmp _nb_kernel303nf_x86_64_sse.nb303nf_threadloop_nb_kernel303nf_x86_64_sse.nb303nf_end: movl nb303nf_nouter(%rsp),%eax movl nb303nf_ninner(%rsp),%ebx movq nb303nf_outeriter(%rbp),%rcx movq nb303nf_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $472,%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?