nb_kernel111_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,633 行 · 第 1/4 页
S
1,633 行
subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,(%rdi,%rcx,4) movss %xmm4,4(%rdi,%rcx,4) movss %xmm5,8(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb111_fixH1(%rsp),%xmm0 movaps nb111_fiyH1(%rsp),%xmm1 movaps nb111_fizH1(%rsp),%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 12(%rdi,%rcx,4),%xmm3 movss 16(%rdi,%rcx,4),%xmm4 movss 20(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,12(%rdi,%rcx,4) movss %xmm4,16(%rdi,%rcx,4) movss %xmm5,20(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## 00001000 addps %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movaps nb111_fixH2(%rsp),%xmm0 movaps nb111_fiyH2(%rsp),%xmm1 movaps nb111_fizH2(%rsp),%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 24(%rdi,%rcx,4),%xmm3 movss 28(%rdi,%rcx,4),%xmm4 movss 32(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3 subss %xmm1,%xmm4 subss %xmm2,%xmm5 movss %xmm3,24(%rdi,%rcx,4) movss %xmm4,28(%rdi,%rcx,4) movss %xmm5,32(%rdi,%rcx,4) ## accumulate force in xmm6/xmm7 for fshift addss %xmm2,%xmm7 movlhps %xmm1,%xmm0 shufps $8,%xmm0,%xmm0 ## 00001000 addps %xmm0,%xmm6 ## increment fshift force movlps (%rsi,%rdx,4),%xmm3 movss 8(%rsi,%rdx,4),%xmm4 subps %xmm6,%xmm3 subss %xmm7,%xmm4 movlps %xmm3,(%rsi,%rdx,4) movss %xmm4,8(%rsi,%rdx,4) ## get n from stack movl nb111_n(%rsp),%esi ## get group index for i particle movq nb111_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb111_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 nb111_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 nb111_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 nb111_Vvdw(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb111_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel111_x86_64_sse.nb111_outerend ## not last, iterate outer loop once more! movl %esi,nb111_n(%rsp) jmp _nb_kernel111_x86_64_sse.nb111_outer_nb_kernel111_x86_64_sse.nb111_outerend: ## check if more outer neighborlists remain movl nb111_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel111_x86_64_sse.nb111_end ## non-zero, do one more workunit jmp _nb_kernel111_x86_64_sse.nb111_threadloop_nb_kernel111_x86_64_sse.nb111_end: movl nb111_nouter(%rsp),%eax movl nb111_ninner(%rsp),%ebx movq nb111_outeriter(%rbp),%rcx movq nb111_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $792,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel111nf_x86_64_sse.globl _nb_kernel111nf_x86_64_ssenb_kernel111nf_x86_64_sse: _nb_kernel111nf_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb111nf_fshift, 16.set nb111nf_gid, 24.set nb111nf_pos, 32.set nb111nf_faction, 40.set nb111nf_charge, 48.set nb111nf_p_facel, 56.set nb111nf_argkrf, 64.set nb111nf_argcrf, 72.set nb111nf_Vc, 80.set nb111nf_type, 88.set nb111nf_p_ntype, 96.set nb111nf_vdwparam, 104.set nb111nf_Vvdw, 112.set nb111nf_p_tabscale, 120.set nb111nf_VFtab, 128.set nb111nf_invsqrta, 136.set nb111nf_dvda, 144.set nb111nf_p_gbtabscale, 152.set nb111nf_GBtab, 160.set nb111nf_p_nthreads, 168.set nb111nf_count, 176.set nb111nf_mtx, 184.set nb111nf_outeriter, 192.set nb111nf_inneriter, 200.set nb111nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb111nf_ixO, 0.set nb111nf_iyO, 16.set nb111nf_izO, 32.set nb111nf_ixH1, 48.set nb111nf_iyH1, 64.set nb111nf_izH1, 80.set nb111nf_ixH2, 96.set nb111nf_iyH2, 112.set nb111nf_izH2, 128.set nb111nf_iqO, 144.set nb111nf_iqH, 160.set nb111nf_qqO, 176.set nb111nf_qqH, 192.set nb111nf_c6, 208.set nb111nf_c12, 224.set nb111nf_vctot, 240.set nb111nf_Vvdwtot, 256.set nb111nf_half, 272.set nb111nf_three, 288.set nb111nf_is3, 304.set nb111nf_ii3, 308.set nb111nf_nri, 324.set nb111nf_iinr, 332.set nb111nf_jindex, 340.set nb111nf_jjnr, 348.set nb111nf_shift, 356.set nb111nf_shiftvec, 364.set nb111nf_facel, 372.set nb111nf_innerjjnr, 380.set nb111nf_ntia, 388.set nb111nf_innerk, 392.set nb111nf_n, 396.set nb111nf_nn1, 400.set nb111nf_nouter, 404.set nb111nf_ninner, 408 push %rbp movq %rsp,%rbp push %rbx subq $424,%rsp emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb111nf_nouter(%rsp) movl %eax,nb111nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb111nf_nri(%rsp) movq %rsi,nb111nf_iinr(%rsp) movq %rdx,nb111nf_jindex(%rsp) movq %rcx,nb111nf_jjnr(%rsp) movq %r8,nb111nf_shift(%rsp) movq %r9,nb111nf_shiftvec(%rsp) movq nb111nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb111nf_facel(%rsp) ## assume we have at least one i particle - start directly movq nb111nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb111nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb111nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb111nf_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb111nf_iqO(%rsp) movaps %xmm4,nb111nf_iqH(%rsp) movq nb111nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb111nf_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb111nf_ntia(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb111nf_half(%rsp) movss nb111nf_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm1,nb111nf_half(%rsp) movaps %xmm3,nb111nf_three(%rsp)_nb_kernel111nf_x86_64_sse.nb111nf_threadloop: movq nb111nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel111nf_x86_64_sse.nb111nf_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%rsi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel111nf_x86_64_sse.nb111nf_spinlock ## if(nn1>nri) nn1=nri movl nb111nf_nri(%rsp),%ecx movl %ecx,%edx subl %ebx,%ecx cmovlel %edx,%ebx ## if(nn1>nri) nn1=nri ## Cleared the spinlock if we got here. ## eax contains nn0, ebx contains nn1. movl %eax,nb111nf_n(%rsp) movl %ebx,nb111nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel111nf_x86_64_sse.nb111nf_outerstart jmp _nb_kernel111nf_x86_64_sse.nb111nf_end_nb_kernel111nf_x86_64_sse.nb111nf_outerstart: ## ebx contains number of outer iterations addl nb111nf_nouter(%rsp),%ebx movl %ebx,nb111nf_nouter(%rsp)_nb_kernel111nf_x86_64_sse.nb111nf_outer: movq nb111nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## rbx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb111nf_is3(%rsp) ## store is3 movq nb111nf_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 nb111nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb111nf_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb111nf_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 addss 4(%rax,%rbx,4),%xmm4 addss 8(%rax,%rbx,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb111nf_ixO(%rsp) movaps %xmm4,nb111nf_iyO(%rsp) movaps %xmm5,nb111nf_izO(%rsp) movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 12(%rax,%rbx,4),%xmm0 addss 16(%rax,%rbx,4),%xmm1 addss 20(%rax,%rbx,4),%xmm2 addss 24(%rax,%rbx,4),%xmm3 addss 28(%rax,%rbx,4),%xmm4 addss 32(%rax,%rbx,4),%xmm5 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm0,nb111nf_ixH1(%rsp) movaps %xmm1,nb111nf_iyH1(%rsp) movaps %xmm2,nb111nf_izH1(%rsp) movaps %xmm3,nb111nf_ixH2(%rsp) movaps %xmm4,nb111nf_iyH2(%rsp) movaps %xmm5,nb111nf_izH2(%rsp) ## clear vctot and Vvdwtot xorps %xmm4,%xmm4 movaps %xmm4,nb111nf_vctot(%rsp) movaps %xmm4,nb111nf_Vvdwtot(%rsp) movq nb111nf_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 ⌨️ 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?