nb_kernel430_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,642 行 · 第 1/4 页
S
1,642 行
addsd %xmm4,%xmm1 addsd %xmm5,%xmm2 ## sum is in low xmm0-xmm2 ## increment i force movsd (%rdi,%rcx,8),%xmm3 movsd 8(%rdi,%rcx,8),%xmm4 movsd 16(%rdi,%rcx,8),%xmm5 subsd %xmm0,%xmm3 subsd %xmm1,%xmm4 subsd %xmm2,%xmm5 movsd %xmm3,(%rdi,%rcx,8) movsd %xmm4,8(%rdi,%rcx,8) movsd %xmm5,16(%rdi,%rcx,8) ## increment fshift force movsd (%rsi,%rdx,8),%xmm3 movsd 8(%rsi,%rdx,8),%xmm4 movsd 16(%rsi,%rdx,8),%xmm5 subsd %xmm0,%xmm3 subsd %xmm1,%xmm4 subsd %xmm2,%xmm5 movsd %xmm3,(%rsi,%rdx,8) movsd %xmm4,8(%rsi,%rdx,8) movsd %xmm5,16(%rsi,%rdx,8) ## get n from stack movl nb430_n(%rsp),%esi ## get group index for i particle movq nb430_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movapd nb430_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb430_Vc(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## accumulate total lj energy and update it movapd nb430_Vvdwtot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb430_Vvdw(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## accumulate dVda and update it movapd nb430_dvdasum(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now movl nb430_ii(%rsp),%edx movq nb430_dvda(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 movsd %xmm7,(%rax,%rdx,8) ## finish if last movl nb430_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel430_x86_64_sse2.nb430_outerend ## not last, iterate outer loop once more! movl %esi,nb430_n(%rsp) jmp _nb_kernel430_x86_64_sse2.nb430_outer_nb_kernel430_x86_64_sse2.nb430_outerend: ## check if more outer neighborlists remain movl nb430_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel430_x86_64_sse2.nb430_end ## non-zero, do one more workunit jmp _nb_kernel430_x86_64_sse2.nb430_threadloop_nb_kernel430_x86_64_sse2.nb430_end: movl nb430_nouter(%rsp),%eax movl nb430_ninner(%rsp),%ebx movq nb430_outeriter(%rbp),%rcx movq nb430_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $536,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel430nf_x86_64_sse2.globl _nb_kernel430nf_x86_64_sse2nb_kernel430nf_x86_64_sse2: _nb_kernel430nf_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb430nf_fshift, 16.set nb430nf_gid, 24.set nb430nf_pos, 32.set nb430nf_faction, 40.set nb430nf_charge, 48.set nb430nf_p_facel, 56.set nb430nf_argkrf, 64.set nb430nf_argcrf, 72.set nb430nf_Vc, 80.set nb430nf_type, 88.set nb430nf_p_ntype, 96.set nb430nf_vdwparam, 104.set nb430nf_Vvdw, 112.set nb430nf_p_tabscale, 120.set nb430nf_VFtab, 128.set nb430nf_invsqrta, 136.set nb430nf_dvda, 144.set nb430nf_p_gbtabscale, 152.set nb430nf_GBtab, 160.set nb430nf_p_nthreads, 168.set nb430nf_count, 176.set nb430nf_mtx, 184.set nb430nf_outeriter, 192.set nb430nf_inneriter, 200.set nb430nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse2 use .set nb430nf_ix, 0.set nb430nf_iy, 16.set nb430nf_iz, 32.set nb430nf_iq, 48.set nb430nf_gbtsc, 64.set nb430nf_tsc, 80.set nb430nf_qq, 96.set nb430nf_c6, 112.set nb430nf_c12, 128.set nb430nf_vctot, 144.set nb430nf_Vvdwtot, 160.set nb430nf_half, 176.set nb430nf_three, 192.set nb430nf_r, 208.set nb430nf_isai, 224.set nb430nf_isaprod, 240.set nb430nf_gbscale, 256.set nb430nf_nri, 272.set nb430nf_iinr, 280.set nb430nf_jindex, 288.set nb430nf_jjnr, 296.set nb430nf_shift, 304.set nb430nf_shiftvec, 312.set nb430nf_facel, 320.set nb430nf_innerjjnr, 328.set nb430nf_is3, 336.set nb430nf_ii3, 340.set nb430nf_ntia, 344.set nb430nf_innerk, 348.set nb430nf_n, 352.set nb430nf_nn1, 356.set nb430nf_ntype, 360.set nb430nf_nouter, 364.set nb430nf_ninner, 368 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $392,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb430nf_nouter(%rsp) movl %eax,nb430nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb430nf_nri(%rsp) movq %rsi,nb430nf_iinr(%rsp) movq %rdx,nb430nf_jindex(%rsp) movq %rcx,nb430nf_jjnr(%rsp) movq %r8,nb430nf_shift(%rsp) movq %r9,nb430nf_shiftvec(%rsp) movq nb430nf_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb430nf_ntype(%rsp) movq nb430nf_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb430nf_facel(%rsp) movq nb430nf_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb430nf_tsc(%rsp) movq nb430nf_p_gbtabscale(%rbp),%rbx movsd (%rbx),%xmm4 shufpd $0,%xmm4,%xmm4 movapd %xmm4,nb430nf_gbtsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb430nf_half(%rsp) movl %ebx,nb430nf_half+4(%rsp) movsd nb430nf_half(%rsp),%xmm1 shufpd $0,%xmm1,%xmm1 ## splat to all elements movapd %xmm1,%xmm3 addpd %xmm3,%xmm3 ## one movapd %xmm3,%xmm2 addpd %xmm2,%xmm2 ## two addpd %xmm2,%xmm3 ## three movapd %xmm1,nb430nf_half(%rsp) movapd %xmm3,nb430nf_three(%rsp)_nb_kernel430nf_x86_64_sse2.nb430nf_threadloop: movq nb430nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel430nf_x86_64_sse2.nb430nf_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_kernel430nf_x86_64_sse2.nb430nf_spinlock ## if(nn1>nri) nn1=nri movl nb430nf_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,nb430nf_n(%rsp) movl %ebx,nb430nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel430nf_x86_64_sse2.nb430nf_outerstart jmp _nb_kernel430nf_x86_64_sse2.nb430nf_end_nb_kernel430nf_x86_64_sse2.nb430nf_outerstart: ## ebx contains number of outer iterations addl nb430nf_nouter(%rsp),%ebx movl %ebx,nb430nf_nouter(%rsp)_nb_kernel430nf_x86_64_sse2.nb430nf_outer: movq nb430nf_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,nb430nf_is3(%rsp) ## store is3 movq nb430nf_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%xmm2 movq nb430nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb430nf_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 mulsd nb430nf_facel(%rsp),%xmm3 shufpd $0,%xmm3,%xmm3 movq nb430nf_invsqrta(%rbp),%rdx ## load invsqrta[ii] movsd (%rdx,%rbx,8),%xmm4 shufpd $0,%xmm4,%xmm4 movq nb430nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb430nf_ntype(%rsp),%edx shll %edx movl %edx,nb430nf_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb430nf_pos(%rbp),%rax ## rax = base of pos[] addsd (%rax,%rbx,8),%xmm0 addsd 8(%rax,%rbx,8),%xmm1 addsd 16(%rax,%rbx,8),%xmm2 movapd %xmm3,nb430nf_iq(%rsp) movapd %xmm4,nb430nf_isai(%rsp) shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 movapd %xmm0,nb430nf_ix(%rsp) movapd %xmm1,nb430nf_iy(%rsp) movapd %xmm2,nb430nf_iz(%rsp) movl %ebx,nb430nf_ii3(%rsp) ## clear vctot xorpd %xmm4,%xmm4 movapd %xmm4,nb430nf_vctot(%rsp) movapd %xmm4,nb430nf_Vvdwtot(%rsp) movq nb430nf_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 nb430nf_pos(%rbp),%rsi movq nb430nf_faction(%rbp),%rdi movq nb430nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb430nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb430nf_ninner(%rsp),%ecx movl %ecx,nb430nf_ninner(%rsp) addl $0,%edx movl %edx,nb430nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel430nf_x86_64_sse2.nb430nf_unroll_loop jmp _nb_kernel430nf_x86_64_sse2.nb430nf_checksingle_nb_kernel430nf_x86_64_sse2.nb430nf_unroll_loop: ## twice unrolled innerloop here movq nb430nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb430nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) ## load isaj movq nb430nf_invsqrta(%rbp),%rsi movlpd (%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm2 mulpd nb430nf_isai(%rsp),%xmm2 movapd %xmm2,nb430nf_isaprod(%rsp) movapd %xmm2,%xmm1 mulpd nb430nf_gbtsc(%rsp),%xmm1 movapd %xmm1,nb430nf_gbscale(%rsp) movq nb430nf_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 mulpd nb430nf_iq(%rsp),%xmm2 mulpd %xmm2,%xmm3 movapd %xmm3,nb430nf_qq(%rsp) movq nb430nf_type(%rbp),%rsi movl (%rsi,%rax,4),%ecx movl (%rsi,%rbx,4),%edx movq nb430nf_vdwparam(%rbp),%rsi shll %ecx shll %edx movl nb430nf_ntia(%rsp),%edi addl %edi,%ecx addl %edi,%edx movlpd (%rsi,%rcx,8),%xmm6 ## c6a movlpd (%rsi,%rdx,8),%xmm7 ## c6b movhpd 8(%rsi,%rcx,8),%xmm6 ## c6a c12a movhpd 8(%rsi,%rdx,8),%xmm7 ## c6b c12b movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movapd %xmm4,nb430nf_c6(%rsp) movapd %xmm6,nb430nf_c12(%rsp) movq nb430nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move two coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm0 movhpd 8(%rsi,%rbx,8),%xmm1 movhpd 16(%rsi,%rbx,8),%xmm2 movq nb430nf_faction(%rbp),%rdi ## move nb430nf_ix-iz to xmm4-xmm6 movapd nb430nf_ix(%rsp),%xmm4 movapd nb430nf_iy(%rsp),%xmm5 movapd nb430nf_iz(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6⌨️ 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?