nb_kernel010_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,354 行 · 第 1/3 页
S
1,354 行
movhps (%esi,%edx,4),%xmm6 movl nb010_pos(%ebp),%edi xorps %xmm7,%xmm7 movaps %xmm6,%xmm4 shufps $8,%xmm4,%xmm4 ## constant 00001000 shufps $13,%xmm6,%xmm6 ## constant 00001101 movlhps %xmm7,%xmm4 movlhps %xmm7,%xmm6 movaps %xmm4,nb010_c6(%esp) movaps %xmm6,nb010_c12(%esp) leal (%eax,%eax,2),%eax leal (%ebx,%ebx,2),%ebx ## move coordinates to xmm0-xmm2 movlps (%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 movhps (%edi,%ebx,4),%xmm1 movss 8(%edi,%ebx,4),%xmm0 movlhps %xmm7,%xmm3 shufps $0,%xmm0,%xmm2 movaps %xmm1,%xmm0 shufps $136,%xmm2,%xmm2 ## constant 10001000 shufps $136,%xmm0,%xmm0 ## constant 10001000 shufps $221,%xmm1,%xmm1 ## constant 11011101 movl nb010_faction(%ebp),%edi ## move nb010_ix-iz to xmm4-xmm6 xorps %xmm7,%xmm7 movaps nb010_ix(%esp),%xmm4 movaps nb010_iy(%esp),%xmm5 movaps nb010_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb010_dx(%esp) movaps %xmm5,nb010_dy(%esp) movaps %xmm6,nb010_dz(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rcpps %xmm4,%xmm5 ## constant 1/x lookup seed in xmm5 movaps nb010_two(%esp),%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 nb010_c6(%esp),%xmm1 mulps nb010_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb010_Vvdwtot(%esp),%xmm5 mulps nb010_six(%esp),%xmm1 mulps nb010_twelve(%esp),%xmm2 subps %xmm1,%xmm2 mulps %xmm2,%xmm4 ## xmm4=total fscal movaps nb010_dx(%esp),%xmm0 movaps nb010_dy(%esp),%xmm1 movaps nb010_dz(%esp),%xmm2 movaps %xmm5,nb010_Vvdwtot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movaps nb010_fix(%esp),%xmm3 movaps nb010_fiy(%esp),%xmm4 movaps nb010_fiz(%esp),%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,nb010_fix(%esp) movaps %xmm4,nb010_fiy(%esp) movaps %xmm5,nb010_fiz(%esp) ## update the fj's movss (%edi,%eax,4),%xmm3 movss 4(%edi,%eax,4),%xmm4 movss 8(%edi,%eax,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%edi,%eax,4) movss %xmm4,4(%edi,%eax,4) movss %xmm5,8(%edi,%eax,4) shufps $225,%xmm0,%xmm0 ## constant 11100001 shufps $225,%xmm1,%xmm1 ## constant 11100001 shufps $225,%xmm2,%xmm2 ## constant 11100001 movss (%edi,%ebx,4),%xmm3 movss 4(%edi,%ebx,4),%xmm4 movss 8(%edi,%ebx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%edi,%ebx,4) movss %xmm4,4(%edi,%ebx,4) movss %xmm5,8(%edi,%ebx,4)_nb_kernel010_ia32_sse.nb010_checksingle: movl nb010_innerk(%esp),%edx andl $1,%edx jnz _nb_kernel010_ia32_sse.nb010_dosingle jmp _nb_kernel010_ia32_sse.nb010_updateouterdata_nb_kernel010_ia32_sse.nb010_dosingle: movl nb010_pos(%ebp),%edi movl nb010_innerjjnr(%esp),%ecx movl (%ecx),%eax movl nb010_type(%ebp),%esi movl %eax,%ecx movl (%esi,%ecx,4),%ecx movl nb010_vdwparam(%ebp),%esi shll %ecx addl nb010_ntia(%esp),%ecx xorps %xmm6,%xmm6 movlps (%esi,%ecx,4),%xmm6 movaps %xmm6,%xmm4 shufps $252,%xmm4,%xmm4 ## constant 11111100 shufps $253,%xmm6,%xmm6 ## constant 11111101 movaps %xmm4,nb010_c6(%esp) movaps %xmm6,nb010_c12(%esp) leal (%eax,%eax,2),%eax ## move coordinates to xmm0-xmm2 movss (%edi,%eax,4),%xmm0 movss 4(%edi,%eax,4),%xmm1 movss 8(%edi,%eax,4),%xmm2 xorps %xmm7,%xmm7 movaps nb010_ix(%esp),%xmm4 movaps nb010_iy(%esp),%xmm5 movaps nb010_iz(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb010_dx(%esp) movaps %xmm5,nb010_dy(%esp) movaps %xmm6,nb010_dz(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 rcpps %xmm4,%xmm5 ## constant 1/x lookup seed in xmm5 movaps nb010_two(%esp),%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 nb010_c6(%esp),%xmm1 mulps nb010_c12(%esp),%xmm2 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addss nb010_Vvdwtot(%esp),%xmm5 mulps nb010_six(%esp),%xmm1 mulps nb010_twelve(%esp),%xmm2 subps %xmm1,%xmm2 mulps %xmm2,%xmm4 ## xmm4=total fscal movl nb010_faction(%ebp),%edi movaps nb010_dx(%esp),%xmm0 movaps nb010_dy(%esp),%xmm1 movaps nb010_dz(%esp),%xmm2 movss %xmm5,nb010_Vvdwtot(%esp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) ## now update f_i movaps nb010_fix(%esp),%xmm3 movaps nb010_fiy(%esp),%xmm4 movaps nb010_fiz(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movaps %xmm3,nb010_fix(%esp) movaps %xmm4,nb010_fiy(%esp) movaps %xmm5,nb010_fiz(%esp) ## update fj movss (%edi,%eax,4),%xmm3 movss 4(%edi,%eax,4),%xmm4 movss 8(%edi,%eax,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%edi,%eax,4) movss %xmm4,4(%edi,%eax,4) movss %xmm5,8(%edi,%eax,4)_nb_kernel010_ia32_sse.nb010_updateouterdata: movl nb010_ii3(%esp),%ecx movl nb010_faction(%ebp),%edi movl nb010_fshift(%ebp),%esi movl nb010_is3(%esp),%edx ## accumulate i forces in xmm0, xmm1, xmm2 movaps nb010_fix(%esp),%xmm0 movaps nb010_fiy(%esp),%xmm1 movaps nb010_fiz(%esp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss (%edi,%ecx,4),%xmm3 movss 4(%edi,%ecx,4),%xmm4 movss 8(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,(%edi,%ecx,4) movss %xmm4,4(%edi,%ecx,4) movss %xmm5,8(%edi,%ecx,4) ## increment fshift force movss (%esi,%edx,4),%xmm3 movss 4(%esi,%edx,4),%xmm4 movss 8(%esi,%edx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,(%esi,%edx,4) movss %xmm4,4(%esi,%edx,4) movss %xmm5,8(%esi,%edx,4) ## get n from stack movl nb010_n(%esp),%esi ## get group index for i particle movl nb010_gid(%ebp),%edx ## base of gid[] movl (%edx,%esi,4),%edx ## ggid=gid[n] ## accumulate total lj energy and update it movaps nb010_Vvdwtot(%esp),%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 movl nb010_Vvdw(%ebp),%eax addss (%eax,%edx,4),%xmm7 ## move back to mem movss %xmm7,(%eax,%edx,4) ## finish if last movl nb010_nn1(%esp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel010_ia32_sse.nb010_outerend ## not last, iterate outer loop once more! movl %esi,nb010_n(%esp) jmp _nb_kernel010_ia32_sse.nb010_outer_nb_kernel010_ia32_sse.nb010_outerend: ## check if more outer neighborlists remain movl nb010_nri(%esp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel010_ia32_sse.nb010_end ## non-zero, do one more workunit jmp _nb_kernel010_ia32_sse.nb010_threadloop_nb_kernel010_ia32_sse.nb010_end: emms movl nb010_nouter(%esp),%eax movl nb010_ninner(%esp),%ebx movl nb010_outeriter(%ebp),%ecx movl nb010_inneriter(%ebp),%edx movl %eax,(%ecx) movl %ebx,(%edx) movl nb010_salign(%esp),%eax addl %eax,%esp ## account for stack alignment addl $320,%esp popl %edi popl %esi popl %edx popl %ecx popl %ebx popl %eax leave ret.globl nb_kernel010nf_ia32_sse.globl _nb_kernel010nf_ia32_ssenb_kernel010nf_ia32_sse: _nb_kernel010nf_ia32_sse: .set nb010nf_p_nri, 8.set nb010nf_iinr, 12.set nb010nf_jindex, 16.set nb010nf_jjnr, 20.set nb010nf_shift, 24.set nb010nf_shiftvec, 28.set nb010nf_fshift, 32.set nb010nf_gid, 36.set nb010nf_pos, 40.set nb010nf_faction, 44.set nb010nf_charge, 48.set nb010nf_p_facel, 52.set nb010nf_p_krf, 56.set nb010nf_p_crf, 60.set nb010nf_Vc, 64.set nb010nf_type, 68.set nb010nf_p_ntype, 72.set nb010nf_vdwparam, 76.set nb010nf_Vvdw, 80.set nb010nf_p_tabscale, 84.set nb010nf_VFtab, 88.set nb010nf_invsqrta, 92.set nb010nf_dvda, 96.set nb010nf_p_gbtabscale, 100.set nb010nf_GBtab, 104.set nb010nf_p_nthreads, 108.set nb010nf_count, 112.set nb010nf_mtx, 116.set nb010nf_outeriter, 120.set nb010nf_inneriter, 124.set nb010nf_work, 128 ## The mutex (last arg) is not used in assembly. ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb010nf_ix, 0.set nb010nf_iy, 16.set nb010nf_iz, 32.set nb010nf_two, 48.set nb010nf_c6, 64.set nb010nf_c12, 80.set nb010nf_Vvdwtot, 96.set nb010nf_half, 112.set nb010nf_three, 128.set nb010nf_is3, 144.set nb010nf_ii3, 148.set nb010nf_ntia, 152.set nb010nf_innerjjnr, 156.set nb010nf_innerk, 160.set nb010nf_n, 164.set nb010nf_nn1, 168.set nb010nf_nri, 172.set nb010nf_ntype, 176.set nb010nf_nouter, 180.set nb010nf_ninner, 184.set nb010nf_salign, 188 pushl %ebp movl %esp,%ebp pushl %eax pushl %ebx pushl %ecx pushl %edx pushl %esi pushl %edi subl $192,%esp ## local stack space movl %esp,%eax andl $0xf,%eax subl %eax,%esp movl %eax,nb010nf_salign(%esp) emms ## Move args passed by reference to stack movl nb010nf_p_nri(%ebp),%ecx movl nb010nf_p_ntype(%ebp),%edi movl (%ecx),%ecx movl (%edi),%edi movl %ecx,nb010nf_nri(%esp) movl %edi,nb010nf_ntype(%esp) ## zero iteration counters movl $0,%eax movl %eax,nb010nf_nouter(%esp) movl %eax,nb010nf_ninner(%esp) ## create constant floating-point factors on stack movl $0x40000000,%eax ## constant 2.0 in IEEE (hex) movl %eax,nb010nf_two(%esp) movss nb010nf_two(%esp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,nb010nf_two(%esp)_nb_kernel010nf_ia32_sse.nb010nf_threadloop: movl nb010nf_count(%ebp),%esi ## pointer to sync counter⌨️ 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?