nb_kernel132_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,146 行 · 第 1/5 页
S
2,146 行
.set nb132nf_jzO, 176.set nb132nf_jxH1, 192.set nb132nf_jyH1, 208.set nb132nf_jzH1, 224.set nb132nf_jxH2, 240.set nb132nf_jyH2, 256.set nb132nf_jzH2, 272.set nb132nf_qqOO, 288.set nb132nf_qqOH, 304.set nb132nf_qqHH, 320.set nb132nf_c6, 336.set nb132nf_c12, 352.set nb132nf_vctot, 368.set nb132nf_Vvdwtot, 384.set nb132nf_half, 400.set nb132nf_three, 416.set nb132nf_rsqOO, 432.set nb132nf_rsqOH1, 448.set nb132nf_rsqOH2, 464.set nb132nf_rsqH1O, 480.set nb132nf_rsqH1H1, 496.set nb132nf_rsqH1H2, 512.set nb132nf_rsqH2O, 528.set nb132nf_rsqH2H1, 544.set nb132nf_rsqH2H2, 560.set nb132nf_rinvOO, 576.set nb132nf_rinvOH1, 592.set nb132nf_rinvOH2, 608.set nb132nf_rinvH1O, 624.set nb132nf_rinvH1H1, 640.set nb132nf_rinvH1H2, 656.set nb132nf_rinvH2O, 672.set nb132nf_rinvH2H1, 688.set nb132nf_rinvH2H2, 704.set nb132nf_krf, 720.set nb132nf_crf, 736.set nb132nf_tsc, 752.set nb132nf_nri, 768.set nb132nf_iinr, 776.set nb132nf_jindex, 784.set nb132nf_jjnr, 792.set nb132nf_shift, 800.set nb132nf_shiftvec, 808.set nb132nf_facel, 816.set nb132nf_innerjjnr, 824.set nb132nf_is3, 832.set nb132nf_ii3, 836.set nb132nf_innerk, 840.set nb132nf_n, 844.set nb132nf_nn1, 848.set nb132nf_nouter, 852.set nb132nf_ninner, 856 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $872,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb132nf_nouter(%rsp) movl %eax,nb132nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb132nf_nri(%rsp) movq %rsi,nb132nf_iinr(%rsp) movq %rdx,nb132nf_jindex(%rsp) movq %rcx,nb132nf_jjnr(%rsp) movq %r8,nb132nf_shift(%rsp) movq %r9,nb132nf_shiftvec(%rsp) movq nb132nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb132nf_facel(%rsp) movq nb132nf_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb132nf_tsc(%rsp) ## assume we have at least one i particle - start directly movq nb132nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb132nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss %xmm3,%xmm4 movss 4(%rdx,%rbx,4),%xmm5 movq nb132nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb132nf_facel(%rsp),%xmm6 mulss %xmm3,%xmm3 mulss %xmm5,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm3 mulss %xmm6,%xmm4 mulss %xmm6,%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb132nf_qqOO(%rsp) movaps %xmm4,nb132nf_qqOH(%rsp) movaps %xmm5,nb132nf_qqHH(%rsp) xorps %xmm0,%xmm0 movq nb132nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movl %ecx,%edx movq nb132nf_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] addq %rcx,%rdx movq nb132nf_vdwparam(%rbp),%rax movlps (%rax,%rdx,4),%xmm0 movaps %xmm0,%xmm1 shufps $0,%xmm0,%xmm0 shufps $85,%xmm1,%xmm1 ## 01010101 movaps %xmm0,nb132nf_c6(%rsp) movaps %xmm1,nb132nf_c12(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb132nf_half(%rsp) movss nb132nf_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,nb132nf_half(%rsp) movaps %xmm3,nb132nf_three(%rsp)_nb_kernel132nf_x86_64_sse.nb132nf_threadloop: movq nb132nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel132nf_x86_64_sse.nb132nf_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_kernel132nf_x86_64_sse.nb132nf_spinlock ## if(nn1>nri) nn1=nri movl nb132nf_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,nb132nf_n(%rsp) movl %ebx,nb132nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel132nf_x86_64_sse.nb132nf_outerstart jmp _nb_kernel132nf_x86_64_sse.nb132nf_end_nb_kernel132nf_x86_64_sse.nb132nf_outerstart: ## ebx contains number of outer iterations addl nb132nf_nouter(%rsp),%ebx movl %ebx,nb132nf_nouter(%rsp)_nb_kernel132nf_x86_64_sse.nb132nf_outer: movq nb132nf_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,nb132nf_is3(%rsp) ## store is3 movq nb132nf_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 nb132nf_iinr(%rsp),%rcx ## ecx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb132nf_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb132nf_ii3(%rsp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 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,nb132nf_ixO(%rsp) movaps %xmm4,nb132nf_iyO(%rsp) movaps %xmm5,nb132nf_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,nb132nf_ixH1(%rsp) movaps %xmm1,nb132nf_iyH1(%rsp) movaps %xmm2,nb132nf_izH1(%rsp) movaps %xmm3,nb132nf_ixH2(%rsp) movaps %xmm4,nb132nf_iyH2(%rsp) movaps %xmm5,nb132nf_izH2(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb132nf_vctot(%rsp) movaps %xmm4,nb132nf_Vvdwtot(%rsp) movq nb132nf_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 nb132nf_pos(%rbp),%rsi movq nb132nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb132nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb132nf_ninner(%rsp),%ecx movl %ecx,nb132nf_ninner(%rsp) addl $0,%edx movl %edx,nb132nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel132nf_x86_64_sse.nb132nf_unroll_loop jmp _nb_kernel132nf_x86_64_sse.nb132nf_single_check_nb_kernel132nf_x86_64_sse.nb132nf_unroll_loop: ## quad-unroll innerloop here movq nb132nf_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,nb132nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb132nf_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 j coordinates to local temp variables movlps (%rsi,%rax,4),%xmm2 movlps 12(%rsi,%rax,4),%xmm3 movlps 24(%rsi,%rax,4),%xmm4 movlps (%rsi,%rbx,4),%xmm5 movlps 12(%rsi,%rbx,4),%xmm6 movlps 24(%rsi,%rbx,4),%xmm7 movhps (%rsi,%rcx,4),%xmm2 movhps 12(%rsi,%rcx,4),%xmm3 movhps 24(%rsi,%rcx,4),%xmm4 movhps (%rsi,%rdx,4),%xmm5 movhps 12(%rsi,%rdx,4),%xmm6 movhps 24(%rsi,%rdx,4),%xmm7 ## current state: ## xmm2= jxOa jyOa jxOc jyOc ## xmm3= jxH1a jyH1a jxH1c jyH1c ## xmm4= jxH2a jyH2a jxH2c jyH2c ## xmm5= jxOb jyOb jxOd jyOd ## xmm6= jxH1b jyH1b jxH1d jyH1d ## xmm7= jxH2b jyH2b jxH2d jyH2d movaps %xmm2,%xmm0 movaps %xmm3,%xmm1 unpcklps %xmm5,%xmm0 ## xmm0= jxOa jxOb jyOa jyOb unpcklps %xmm6,%xmm1 ## xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps %xmm5,%xmm2 ## xmm2= jxOc jxOd jyOc jyOd unpckhps %xmm6,%xmm3 ## xmm3= jxH1c jxH1d jyH1c jyH1d movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 ## xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps %xmm7,%xmm5 ## xmm5= jxH2c jxH2d jyH2c jyH2d movaps %xmm1,%xmm7 movlhps %xmm2,%xmm0 ## xmm0= jxOa jxOb jxOc jxOd movaps %xmm0,nb132nf_jxO(%rsp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb132nf_jyO(%rsp) movlhps %xmm3,%xmm1 movaps %xmm1,nb132nf_jxH1(%rsp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb132nf_jyH1(%rsp) movlhps %xmm5,%xmm4 movaps %xmm4,nb132nf_jxH2(%rsp) movhlps %xmm6,%xmm5 movaps %xmm5,nb132nf_jyH2(%rsp) movss 8(%rsi,%rax,4),%xmm0 movss 20(%rsi,%rax,4),%xmm1 movss 32(%rsi,%rax,4),%xmm2 movss 8(%rsi,%rcx,4),%xmm3 movss 20(%rsi,%rcx,4),%xmm4 movss 32(%rsi,%rcx,4),%xmm5 movhps 4(%rsi,%rbx,4),%xmm0 movhps 16(%rsi,%rbx,4),%xmm1 movhps 28(%rsi,%rbx,4),%xmm2 movhps 4(%rsi,%rdx,4),%xmm3 movhps 16(%rsi,%rdx,4),%xmm4 movhps 28(%rsi,%rdx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb132nf_jzO(%rsp) movaps %xmm1,nb132nf_jzH1(%rsp) movaps %xmm2,nb132nf_jzH2(%rsp) movaps nb132nf_ixO(%rsp),%xmm0 movaps nb132nf_iyO(%rsp),%xmm1 movaps nb132nf_izO(%rsp),%xmm2 movaps nb132nf_ixO(%rsp),%xmm3 movaps nb132nf_iyO(%rsp),%xmm4 movaps nb132nf_izO(%rsp),%xmm5 subps nb132nf_jxO(%rsp),%xmm0 subps nb132nf_jyO(%rsp),%xmm1 subps nb132nf_jzO(%rsp),%xmm2 subps nb132nf_jxH1(%rsp),%xmm3 subps nb132nf_jyH1(%rsp),%xmm4 subps nb132nf_jzH1(%rsp),%xmm5 mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb132nf_rsqOO(%rsp) movaps %xmm3,nb132nf_rsqOH1(%rsp) movaps nb132nf_ixO(%rsp),%xmm0 movaps nb132nf_iyO(%rsp),%xmm1 movaps nb132nf_izO(%rsp),%xmm2 movaps nb132nf_ixH1(%rsp),%xmm3 movaps nb132nf_iyH1(%rsp),%xmm4 movaps nb132nf_izH1(%rsp),%xmm5 subps nb132nf_jxH2(%rsp),%xmm0 subps nb132nf_jyH2(%rsp),%xmm1 subps nb132nf_jzH2(%rsp),%xmm2 subps nb132nf_jxO(%rsp),%xmm3 subps nb132nf_jyO(%rsp),%xmm4 subps nb132nf_jzO(%rsp),%xmm5 mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb132nf_rsqOH2(%rsp) movaps %xmm3,nb132nf_rsqH1O(%rsp) movaps nb132nf_ixH1(%rsp),%xmm0 movaps nb132nf_iyH1(%rsp),%xmm1 movaps nb132nf_izH1(%rsp),%xmm2 movaps nb132nf_ixH1(%rsp),%xmm3 movaps nb132nf_iyH1(%rsp),%xmm4 movaps nb132nf_izH1(%rsp),%xmm5 subps nb132nf_jxH1(%rsp),%xmm0 subps nb132nf_jyH1(%rsp),%xmm1 subps nb132nf_jzH1(%rsp),%xmm2 subps nb132nf_jxH2(%rsp),%xmm3 subps nb132nf_jyH2(%rsp),%xmm4 subps nb132nf_jzH2(%rsp),%xmm5 mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb132nf_rsqH1H1(%rsp) movaps %xmm3,nb132nf_rsqH1H2(%rsp) movaps nb132nf_ixH2(%rsp),%xmm0 movaps nb132nf_iyH2(%rsp),%xmm1 movaps nb132nf_izH2(%rsp),%xmm2 movaps nb132nf_ixH2(%rsp),%xmm3 movaps nb132nf_iyH2(%rsp),%xmm4 movaps nb132nf_izH2(%rsp),%xmm5 subps nb132nf_jxO(%rsp),%xmm0 subps nb132nf_jyO(%rsp),%xmm1 subps nb132⌨️ 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?