nb_kernel410_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,987 行 · 第 1/4 页
S
1,987 行
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 nb410nf_ix(%rsp),%xmm4 movaps nb410nf_iy(%rsp),%xmm5 movaps nb410nf_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 nb410nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb410nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r mulps nb410nf_gbscale(%rsp),%xmm4 movhlps %xmm4,%xmm5 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm5,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm6 cvtpi2ps %mm7,%xmm5 movlhps %xmm5,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%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 nb410nf_GBtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm7,%ecx psrlq $32,%mm7 movd %mm6,%ebx movd %mm7,%edx ## load coulomb table movaps (%rsi,%rax,4),%xmm4 movaps (%rsi,%rbx,4),%xmm5 movaps (%rsi,%rcx,4),%xmm6 movaps (%rsi,%rdx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm6,%xmm3 shufps $0xEE,%xmm7,%xmm3 shufps $0x44,%xmm7,%xmm6 movaps %xmm4,%xmm7 shufps $0xEE,%xmm5,%xmm7 shufps $0x44,%xmm5,%xmm4 movaps %xmm4,%xmm5 shufps $0xDD,%xmm6,%xmm5 shufps $0x88,%xmm6,%xmm4 movaps %xmm7,%xmm6 shufps $0x88,%xmm3,%xmm6 shufps $0xDD,%xmm3,%xmm7 ## 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 nb410nf_qq(%rsp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV ## update vctot addps nb410nf_vctot(%rsp),%xmm5 movaps %xmm5,nb410nf_vctot(%rsp) ## L-J movaps %xmm0,%xmm4 mulps %xmm0,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm6 mulps %xmm4,%xmm6 mulps %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvtwelve mulps nb410nf_c6(%rsp),%xmm6 mulps nb410nf_c12(%rsp),%xmm4 movaps nb410nf_Vvdwtot(%rsp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movaps %xmm7,nb410nf_Vvdwtot(%rsp) ## should we do one more iteration? subl $4,nb410nf_innerk(%rsp) jl _nb_kernel410nf_x86_64_sse.nb410nf_finish_inner jmp _nb_kernel410nf_x86_64_sse.nb410nf_unroll_loop_nb_kernel410nf_x86_64_sse.nb410nf_finish_inner: ## check if at least two particles remain addl $4,nb410nf_innerk(%rsp) movl nb410nf_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel410nf_x86_64_sse.nb410nf_dopair jmp _nb_kernel410nf_x86_64_sse.nb410nf_checksingle_nb_kernel410nf_x86_64_sse.nb410nf_dopair: movq nb410nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax movl 4(%rcx),%ebx addq $8,nb410nf_innerjjnr(%rsp) xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 ## load isa2 movq nb410nf_invsqrta(%rbp),%rsi movss (%rsi,%rax,4),%xmm2 movss (%rsi,%rbx,4),%xmm3 unpcklps %xmm3,%xmm2 ## isa2 in xmm3(0,1) mulps nb410nf_isai(%rsp),%xmm2 movaps %xmm2,nb410nf_isaprod(%rsp) movaps %xmm2,%xmm1 mulps nb410nf_gbtsc(%rsp),%xmm1 movaps %xmm1,nb410nf_gbscale(%rsp) movq nb410nf_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rbx,4),%xmm6 unpcklps %xmm6,%xmm3 ## 00001000 ;# xmm3(0,1) has the charges mulps nb410nf_iq(%rsp),%xmm2 mulps %xmm2,%xmm3 movaps %xmm3,nb410nf_qq(%rsp) movq nb410nf_type(%rbp),%rsi movl %eax,%ecx movl %ebx,%edx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb410nf_vdwparam(%rbp),%rsi shll %ecx shll %edx movl nb410nf_ntia(%rsp),%edi addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rcx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm6 movq nb410nf_pos(%rbp),%rdi movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## 00001000 shufps $13,%xmm6,%xmm6 ## 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb410nf_c6(%rsp) movaps %xmm6,nb410nf_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 movq nb410nf_faction(%rbp),%rdi ## move ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb410nf_ix(%rsp),%xmm4 movaps nb410nf_iy(%rsp),%xmm5 movaps nb410nf_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 nb410nf_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb410nf_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv mulps %xmm0,%xmm4 ## xmm4=r mulps nb410nf_gbscale(%rsp),%xmm4 cvttps2pi %xmm4,%mm6 ## mm6 contain lu indices cvtpi2ps %mm6,%xmm6 subps %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 movq nb410nf_GBtab(%rbp),%rsi movd %mm6,%ecx psrlq $32,%mm6 movd %mm6,%edx ## load coulomb table movaps (%rsi,%rcx,4),%xmm4 movaps (%rsi,%rdx,4),%xmm7 ## transpose, using xmm3 for scratch movaps %xmm4,%xmm6 unpcklps %xmm7,%xmm4 ## Y1 Y2 F1 F2 unpckhps %xmm7,%xmm6 ## G1 G2 H1 H2 movhlps %xmm4,%xmm5 ## F1 F2 movhlps %xmm6,%xmm7 ## H1 H2 ## 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 nb410nf_qq(%rsp),%xmm3 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV addps nb410nf_vctot(%rsp),%xmm5 movaps %xmm5,nb410nf_vctot(%rsp) ## L-J movaps %xmm0,%xmm4 mulps %xmm0,%xmm4 ## xmm4=rinvsq ## at this point mm5 contains vcoul and mm3 fijC ## increment vcoul - then we can get rid of mm5 ## update vctot movaps %xmm4,%xmm6 mulps %xmm4,%xmm6 mulps %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvtwelve mulps nb410nf_c6(%rsp),%xmm6 mulps nb410nf_c12(%rsp),%xmm4 movaps nb410nf_Vvdwtot(%rsp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movaps %xmm7,nb410nf_Vvdwtot(%rsp)_nb_kernel410nf_x86_64_sse.nb410nf_checksingle: movl nb410nf_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel410nf_x86_64_sse.nb410nf_dosingle jmp _nb_kernel410nf_x86_64_sse.nb410nf_updateouterdata_nb_kernel410nf_x86_64_sse.nb410nf_dosingle: movq nb410nf_charge(%rbp),%rsi movq nb410nf_invsqrta(%rbp),%rdx movq nb410nf_pos(%rbp),%rdi movq nb410nf_innerjjnr(%rsp),%rcx movl (%rcx),%eax xorps %xmm2,%xmm2 movaps %xmm2,%xmm6 movss (%rdx,%rax,4),%xmm2 ## isa2 mulss nb410nf_isai(%rsp),%xmm2 movss %xmm2,nb410nf_isaprod(%rsp) movss %xmm2,%xmm1 mulss nb410nf_gbtsc(%rsp),%xmm1 movss %xmm1,nb410nf_gbscale(%rsp) mulss nb410nf_iq(%rsp),%xmm2 movss (%rsi,%rax,4),%xmm6 ## xmm6(0) has the charge mulss %xmm2,%xmm6 movss %xmm6,nb410nf_qq(%rsp) movq nb410nf_type(%rbp),%rsi movl %eax,%ecx movl (%rsi,%rcx,4),%ecx movq nb410nf_vdwparam(%rbp),%rsi shll %ecx addl nb410nf_ntia(%rsp),%ecx movlps (%rsi,%rcx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## 11111100 shufps $253,%xmm6,%xmm6 ## 11111101 movaps %xmm4,nb410nf_c6(%rsp) movaps %xmm6,nb410nf_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 movaps nb410nf_ix(%rsp),%xmm4 movaps nb410nf_iy(%rsp),%xmm5 movaps nb410nf_iz(%rsp),%xmm6 ## calc dr subss %xmm0,%xmm4 subss %xmm1,%xmm5 subss %xmm2,%xmm6 ## square it mulss %xmm4,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm6 addss %xmm5,%xmm4 addss %xmm6,%xmm4 ## rsq in xmm4 rsqrtss %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulss %xmm5,%xmm5 movss nb410nf_three(%rsp),%xmm1 mulss %xmm4,%xmm5 ## rsq*lu*lu movss nb410nf_half(%rsp),%xmm0 subss %xmm5,%xmm1 ## 30-rsq*lu*lu mulss %xmm2,%xmm1 mulss %xmm1,%xmm0 ## xmm0=rinv mulss %xmm0,%xmm4 ## xmm4=r mulss nb410nf_gbscale(%rsp),%xmm4 cvttss2si %xmm4,%ebx ## mm6 contain lu indices cvtsi2ss %ebx,%xmm6 subss %xmm6,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 shll $2,%ebx movq nb410nf_GBtab(%rbp),%rsi movaps (%rsi,%rbx,4),%xmm4 movhlps %xmm4,%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $1,%xmm5,%xmm5 shufps $1,%xmm7,%xmm7 ## table ready in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp movss nb410nf_qq(%rsp),%xmm3 mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV mulss %xmm3,%xmm5 ## vcoul=qq*VV addss nb410nf_vctot(%rsp),%xmm5 movss %xmm5,nb410nf_vctot(%rsp) ## L-J movaps %xmm0,%xmm4 mulss %xmm0,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm6 mulss %xmm4,%xmm6 mulss %xmm4,%xmm6 ## xmm6=rinvsix movaps %xmm6,%xmm4 mulss %xmm4,%xmm4 ## xmm4=rinvtwelve mulss nb410nf_c6(%rsp),%xmm6 mulss nb410nf_c12(%rsp),%xmm4 movss nb410nf_Vvdwtot(%rsp),%xmm7 addps %xmm4,%xmm7 subps %xmm6,%xmm7 movss %xmm7,nb410nf_Vvdwtot(%rsp)_nb_kernel410nf_x86_64_sse.nb410nf_updateouterdata: ## get n from stack movl nb410nf_n(%rsp),%esi ## get group index for i particle movq nb410nf_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb410nf_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 nb410nf_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## accumulate total lj energy and update it movaps nb410nf_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 nb410nf_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb410nf_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel410nf_x86_64_sse.nb410nf_outerend ## not last, iterate outer loop once more! movl %esi,nb410nf_n(%rsp) jmp _nb_kernel410nf_x86_64_sse.nb410nf_outer_nb_kernel410nf_x86_64_sse.nb410nf_outerend: ## check if more outer neighborlists remain movl nb410nf_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel410nf_x86_64_sse.nb410nf_end ## non-zero, do one more workunit jmp _nb_kernel410nf_x86_64_sse.nb410nf_threadloop_nb_kernel410nf_x86_64_sse.nb410nf_end: movl nb410nf_nouter(%rsp),%eax movl nb410nf_ninner(%rsp),%ebx movq nb410nf_outeriter(%rbp),%rcx movq nb410nf_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?