nb_kernel231_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,053 行 · 第 1/5 页
S
2,053 行
.set nb231nf_Vc, 80.set nb231nf_type, 88.set nb231nf_p_ntype, 96.set nb231nf_vdwparam, 104.set nb231nf_Vvdw, 112.set nb231nf_p_tabscale, 120.set nb231nf_VFtab, 128.set nb231nf_invsqrta, 136.set nb231nf_dvda, 144.set nb231nf_p_gbtabscale, 152.set nb231nf_GBtab, 160.set nb231nf_p_nthreads, 168.set nb231nf_count, 176.set nb231nf_mtx, 184.set nb231nf_outeriter, 192.set nb231nf_inneriter, 200.set nb231nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb231nf_ixO, 0.set nb231nf_iyO, 16.set nb231nf_izO, 32.set nb231nf_ixH1, 48.set nb231nf_iyH1, 64.set nb231nf_izH1, 80.set nb231nf_ixH2, 96.set nb231nf_iyH2, 112.set nb231nf_izH2, 128.set nb231nf_iqO, 144.set nb231nf_iqH, 160.set nb231nf_qqO, 176.set nb231nf_qqH, 192.set nb231nf_c6, 208.set nb231nf_c12, 224.set nb231nf_vctot, 240.set nb231nf_Vvdwtot, 256.set nb231nf_half, 272.set nb231nf_three, 288.set nb231nf_krf, 304.set nb231nf_crf, 320.set nb231nf_rinvO, 336.set nb231nf_rinvH1, 352.set nb231nf_rinvH2, 368.set nb231nf_krsqO, 384.set nb231nf_krsqH1, 400.set nb231nf_krsqH2, 416.set nb231nf_tsc, 432.set nb231nf_facel, 448.set nb231nf_iinr, 456.set nb231nf_jindex, 464.set nb231nf_jjnr, 472.set nb231nf_shift, 480.set nb231nf_shiftvec, 488.set nb231nf_innerjjnr, 496.set nb231nf_nri, 504.set nb231nf_ntia, 508.set nb231nf_is3, 512.set nb231nf_ii3, 516.set nb231nf_innerk, 520.set nb231nf_n, 524.set nb231nf_nn1, 528.set nb231nf_nouter, 532.set nb231nf_ninner, 536 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $552,%rsp # # local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb231nf_nouter(%rsp) movl %eax,nb231nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb231nf_nri(%rsp) movq %rsi,nb231nf_iinr(%rsp) movq %rdx,nb231nf_jindex(%rsp) movq %rcx,nb231nf_jjnr(%rsp) movq %r8,nb231nf_shift(%rsp) movq %r9,nb231nf_shiftvec(%rsp) movq nb231nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb231nf_facel(%rsp) movq nb231nf_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb231nf_tsc(%rsp) movq nb231nf_argkrf(%rbp),%rsi movq nb231nf_argcrf(%rbp),%rdi movss (%rsi),%xmm1 movss (%rdi),%xmm2 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm1,nb231nf_krf(%rsp) movaps %xmm2,nb231nf_crf(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb231nf_half(%rsp) movss nb231nf_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,nb231nf_half(%rsp) movaps %xmm3,nb231nf_three(%rsp) ## assume we have at least one i particle - start directly movq nb231nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb231nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss 4(%rdx,%rbx,4),%xmm4 movq nb231nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb231nf_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb231nf_iqO(%rsp) movaps %xmm4,nb231nf_iqH(%rsp) movq nb231nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb231nf_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb231nf_ntia(%rsp)_nb_kernel231nf_x86_64_sse.nb231nf_threadloop: movq nb231nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel231nf_x86_64_sse.nb231nf_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_kernel231nf_x86_64_sse.nb231nf_spinlock ## if(nn1>nri) nn1=nri movl nb231nf_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,nb231nf_n(%rsp) movl %ebx,nb231nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel231nf_x86_64_sse.nb231nf_outerstart jmp _nb_kernel231nf_x86_64_sse.nb231nf_end_nb_kernel231nf_x86_64_sse.nb231nf_outerstart: ## ebx contains number of outer iterations addl nb231nf_nouter(%rsp),%ebx movl %ebx,nb231nf_nouter(%rsp)_nb_kernel231nf_x86_64_sse.nb231nf_outer: movq nb231nf_shift(%rsp),%rax ## eax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb231nf_is3(%rsp) ## store is3 movq nb231nf_shiftvec(%rsp),%rax ## eax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb231nf_iinr(%rsp),%rcx ## ecx = 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 nb231nf_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb231nf_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,nb231nf_ixO(%rsp) movaps %xmm4,nb231nf_iyO(%rsp) movaps %xmm5,nb231nf_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,nb231nf_ixH1(%rsp) movaps %xmm1,nb231nf_iyH1(%rsp) movaps %xmm2,nb231nf_izH1(%rsp) movaps %xmm3,nb231nf_ixH2(%rsp) movaps %xmm4,nb231nf_iyH2(%rsp) movaps %xmm5,nb231nf_izH2(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb231nf_vctot(%rsp) movaps %xmm4,nb231nf_Vvdwtot(%rsp) movq nb231nf_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 nb231nf_pos(%rbp),%rsi movq nb231nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb231nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb231nf_ninner(%rsp),%ecx movl %ecx,nb231nf_ninner(%rsp) addl $0,%edx movl %edx,nb231nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel231nf_x86_64_sse.nb231nf_unroll_loop jmp _nb_kernel231nf_x86_64_sse.nb231nf_odd_inner_nb_kernel231nf_x86_64_sse.nb231nf_unroll_loop: ## quad-unroll innerloop here movq nb231nf_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 addq $16,nb231nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb231nf_charge(%rbp),%rsi ## base of charge[] movss (%rsi,%rax,4),%xmm3 movss (%rsi,%rcx,4),%xmm4 movss (%rsi,%rbx,4),%xmm6 movss (%rsi,%rdx,4),%xmm7 shufps $0,%xmm6,%xmm3 shufps $0,%xmm7,%xmm4 shufps $136,%xmm4,%xmm3 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb231nf_iqO(%rsp),%xmm3 mulps nb231nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb231nf_qqO(%rsp) movaps %xmm4,nb231nf_qqH(%rsp) movq nb231nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb231nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb231nf_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 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb231nf_c6(%rsp) movaps %xmm6,nb231nf_c12(%rsp) movq nb231nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx 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 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb231nf_ixO(%rsp),%xmm4 movaps nb231nf_iyO(%rsp),%xmm5 movaps nb231nf_izO(%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 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb231nf_ixH1(%rsp),%xmm4 movaps nb231nf_iyH1(%rsp),%xmm5 movaps nb231nf_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,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb231nf_ixH2(%rsp),%xmm3 movaps nb231nf_iyH2(%rsp),%xmm4 movaps nb231nf_izH2(%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 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movaps %xmm5,%xmm0 movaps %xmm6,%xmm1 movaps %xmm7,%xmm2⌨️ 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?