nb_kernel111_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,657 行 · 第 1/3 页
S
1,657 行
mov [rsp + nb111nf_iinr], rsi mov [rsp + nb111nf_jindex], rdx mov [rsp + nb111nf_jjnr], rcx mov [rsp + nb111nf_shift], r8 mov [rsp + nb111nf_shiftvec], r9 mov rsi, [rbp + nb111nf_p_facel] movss xmm0, [rsi] movss [rsp + nb111nf_facel], xmm0 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb111nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb111nf_charge] movss xmm3, [rdx + rbx*4] movss xmm4, [rdx + rbx*4 + 4] mov rsi, [rbp + nb111nf_p_facel] movss xmm0, [rsi] movss xmm5, [rsp + nb111nf_facel] mulss xmm3, xmm5 mulss xmm4, xmm5 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 movaps [rsp + nb111nf_iqO], xmm3 movaps [rsp + nb111nf_iqH], xmm4 mov rdx, [rbp + nb111nf_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov rdi, [rbp + nb111nf_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] mov [rsp + nb111nf_ntia], ecx ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# half in IEEE (hex) mov [rsp + nb111nf_half], eax movss xmm1, [rsp + nb111nf_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# one movaps xmm3, xmm2 addps xmm2, xmm2 ;# two addps xmm3, xmm2 ;# three movaps [rsp + nb111nf_half], xmm1 movaps [rsp + nb111nf_three], xmm3.nb111nf_threadloop: mov rsi, [rbp + nb111nf_count] ;# pointer to sync counter mov eax, [rsi].nb111nf_spinlock: mov ebx, eax ;# ebx=*count=nn0 add ebx, 1 ;# ebx=nn1=nn0+10 lock cmpxchg [esi], ebx ;# write nn1 to *counter, ;# if it hasnt changed. ;# or reread *counter to eax. pause ;# -> better p4 performance jnz .nb111nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb111nf_nri] mov edx, ecx sub ecx, ebx cmovle ebx, edx ;# if(nn1>nri) nn1=nri ;# Cleared the spinlock if we got here. ;# eax contains nn0, ebx contains nn1. mov [rsp + nb111nf_n], eax mov [rsp + nb111nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb111nf_outerstart jmp .nb111nf_end.nb111nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb111nf_nouter] mov [rsp + nb111nf_nouter], ebx.nb111nf_outer: mov rax, [rsp + nb111nf_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb111nf_is3],ebx ;# store is3 mov rax, [rsp + nb111nf_shiftvec] ;# rax = base of shiftvec[] movss xmm0, [rax + rbx*4] movss xmm1, [rax + rbx*4 + 4] movss xmm2, [rax + rbx*4 + 8] mov rcx, [rsp + nb111nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb111nf_pos] ;# rax = base of pos[] mov [rsp + nb111nf_ii3], ebx addss xmm3, [rax + rbx*4] addss xmm4, [rax + rbx*4 + 4] addss xmm5, [rax + rbx*4 + 8] shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb111nf_ixO], xmm3 movaps [rsp + nb111nf_iyO], xmm4 movaps [rsp + nb111nf_izO], xmm5 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm0, [rax + rbx*4 + 12] addss xmm1, [rax + rbx*4 + 16] addss xmm2, [rax + rbx*4 + 20] addss xmm3, [rax + rbx*4 + 24] addss xmm4, [rax + rbx*4 + 28] addss xmm5, [rax + rbx*4 + 32] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb111nf_ixH1], xmm0 movaps [rsp + nb111nf_iyH1], xmm1 movaps [rsp + nb111nf_izH1], xmm2 movaps [rsp + nb111nf_ixH2], xmm3 movaps [rsp + nb111nf_iyH2], xmm4 movaps [rsp + nb111nf_izH2], xmm5 ;# clear vctot and Vvdwtot xorps xmm4, xmm4 movaps [rsp + nb111nf_vctot], xmm4 movaps [rsp + nb111nf_Vvdwtot], xmm4 mov rax, [rsp + nb111nf_jindex] mov ecx, [rax + rsi*4] ;# jindex[n] mov edx, [rax + rsi*4 + 4] ;# jindex[n+1] sub edx, ecx ;# number of innerloop atoms mov rsi, [rbp + nb111nf_pos] mov rax, [rsp + nb111nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb111nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [rsp + nb111nf_ninner] mov [rsp + nb111nf_ninner], ecx add edx, 0 mov [rsp + nb111nf_innerk], edx ;# number of innerloop atoms jge .nb111nf_unroll_loop jmp .nb111nf_odd_inner.nb111nf_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb111nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] mov ecx, [rdx + 8] mov edx, [rdx + 12] ;# eax-edx=jnr1-4 add qword ptr [rsp + nb111nf_innerjjnr], 16 ;# advance pointer (unrolled 4) mov rsi, [rbp + nb111nf_charge] ;# base of charge[] movss xmm3, [rsi + rax*4] movss xmm4, [rsi + rcx*4] movss xmm6, [rsi + rbx*4] movss xmm7, [rsi + rdx*4] shufps xmm3, xmm6, 0 shufps xmm4, xmm7, 0 shufps xmm3, xmm4, 136 ;# 10001000 ;# all charges in xmm3 movaps xmm4, xmm3 ;# and in xmm4 mulps xmm3, [rsp + nb111nf_iqO] mulps xmm4, [rsp + nb111nf_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movd mm2, ecx movd mm3, edx movaps [rsp + nb111nf_qqO], xmm3 movaps [rsp + nb111nf_qqH], xmm4 mov rsi, [rbp + nb111nf_type] mov eax, [rsi + rax*4] mov ebx, [rsi + rbx*4] mov ecx, [rsi + rcx*4] mov edx, [rsi + rdx*4] mov rsi, [rbp + nb111nf_vdwparam] shl eax, 1 shl ebx, 1 shl ecx, 1 shl edx, 1 mov edi, [rsp + nb111nf_ntia] add eax, edi add ebx, edi add ecx, edi add edx, edi movlps xmm6, [rsi + rax*4] movlps xmm7, [rsi + rcx*4] movhps xmm6, [rsi + rbx*4] movhps xmm7, [rsi + rdx*4] movaps xmm4, xmm6 shufps xmm4, xmm7, 136 ;# 10001000 shufps xmm6, xmm7, 221 ;# 11011101 movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 movaps [rsp + nb111nf_c6], xmm4 movaps [rsp + nb111nf_c12], xmm6 mov rsi, [rbp + nb111nf_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] lea rcx, [rcx + rcx*2] ;# replace jnr with j3 lea rdx, [rdx + rdx*2] ;# move four coordinates to xmm0-xmm2 movlps xmm4, [rsi + rax*4] movlps xmm5, [rsi + rcx*4] movss xmm2, [rsi + rax*4 + 8] movss xmm6, [rsi + rcx*4 + 8] movhps xmm4, [rsi + rbx*4] movhps xmm5, [rsi + rdx*4] movss xmm0, [rsi + rbx*4 + 8] movss xmm1, [rsi + rdx*4 + 8] shufps xmm2, xmm0, 0 shufps xmm6, xmm1, 0 movaps xmm0, xmm4 movaps xmm1, xmm4 shufps xmm2, xmm6, 136 ;# 10001000 shufps xmm0, xmm5, 136 ;# 10001000 shufps xmm1, xmm5, 221 ;# 11011101 ;# move ixO-izO to xmm4-xmm6 movaps xmm4, [rsp + nb111nf_ixO] movaps xmm5, [rsp + nb111nf_iyO] movaps xmm6, [rsp + nb111nf_izO] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 movaps xmm7, xmm4 ;# rsqO in xmm7 ;# move ixH1-izH1 to xmm4-xmm6 movaps xmm4, [rsp + nb111nf_ixH1] movaps xmm5, [rsp + nb111nf_iyH1] movaps xmm6, [rsp + nb111nf_izH1] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm6, xmm5 addps xmm6, xmm4 ;# rsqH1 in xmm6 ;# move ixH2-izH2 to xmm3-xmm5 movaps xmm3, [rsp + nb111nf_ixH2] movaps xmm4, [rsp + nb111nf_iyH2] movaps xmm5, [rsp + nb111nf_izH2] ;# calc dr subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# square it mulps xmm3,xmm3 mulps xmm4,xmm4 mulps xmm5,xmm5 addps xmm5, xmm4 addps xmm5, xmm3 ;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ;# start with rsqO - seed in xmm2 rsqrtps xmm2, xmm7 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [rsp + nb111nf_three] mulps xmm2, xmm7 ;# rsq*lu*lu subps xmm4, xmm2 ;# 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [rsp + nb111nf_half] movaps xmm7, xmm4 ;# rinvO in xmm7 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [rsp + nb111nf_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm4, xmm2 ;# 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [rsp + nb111nf_half] movaps xmm6, xmm4 ;# rinvH1 in xmm6 ;# rsqH2 - seed in xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [rsp + nb111nf_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm4, xmm2 ;# 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [rsp + nb111nf_half] movaps xmm5, xmm4 ;# rinvH2 in xmm5 ;# do O interactions movaps xmm4, xmm7 mulps xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq movaps xmm1, xmm4 mulps xmm1, xmm4 mulps xmm1, xmm4 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm7, [rsp + nb111nf_qqO] ;# xmm7=vcoul mulps xmm1, [rsp + nb111nf_c6] mulps xmm2, [rsp + nb111nf_c12] movaps xmm3, xmm2 subps xmm3, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addps xmm3, [rsp + nb111nf_Vvdwtot] addps xmm7, [rsp + nb111nf_vctot] movaps [rsp + nb111nf_Vvdwtot], xmm3 movaps [rsp + nb111nf_vctot], xmm7 ;# H1 & H2 interactions addps xmm6, xmm5 ;# add H2 rinv mulps xmm6, [rsp + nb111nf_qqH] ;# xmm6=vcoul addps xmm6, [rsp + nb111nf_vctot] movaps [rsp + nb111nf_vctot], xmm6 ;# should we do one more iteration? sub dword ptr [rsp + nb111nf_innerk], 4 jl .nb111nf_odd_inner jmp .nb111nf_unroll_loop.nb111nf_odd_inner: add dword ptr [rsp + nb111nf_innerk], 4 jnz .nb111nf_odd_loop jmp .nb111nf_updateouterdata.nb111nf_odd_loop: mov rdx, [rsp + nb111nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] add qword ptr [rsp + nb111nf_innerjjnr], 4 xorps xmm4, xmm4 movss xmm4, [rsp + nb111nf_iqO] mov rsi, [rbp + nb111nf_charge] movhps xmm4, [rsp + nb111nf_iqH] movss xmm3, [rsi + rax*4] ;# charge in xmm3 shufps xmm3, xmm3, 0 mulps xmm3, xmm4 movaps [rsp + nb111nf_qqO], xmm3 ;# use oxygen qq for storage xorps xmm6, xmm6 mov rsi, [rbp + nb111nf_type] mov ebx, [rsi + rax*4] mov rsi, [rbp + nb111nf_vdwparam] shl ebx, 1 add ebx, [rsp + nb111nf_ntia] movlps xmm6, [rsi + rbx*4] movaps xmm7, xmm6 shufps xmm6, xmm6, 252 ;# 11111100 shufps xmm7, xmm7, 253 ;# 11111101 movaps [rsp + nb111nf_c6], xmm6 movaps [rsp + nb111nf_c12], xmm7 mov rsi, [rbp + nb111nf_pos] lea rax, [rax + rax*2] ;# move j coords to xmm0-xmm2 movss xmm0, [rsi + rax*4] movss xmm1, [rsi + rax*4 + 4] movss xmm2, [rsi + rax*4 + 8] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movss xmm3, [rsp + nb111nf_ixO] movss xmm4, [rsp + nb111nf_iyO] movss xmm5, [rsp + nb111nf_izO] movlps xmm6, [rsp + nb111nf_ixH1] movlps xmm7, [rsp + nb111nf_ixH2] unpcklps xmm6, xmm7 movlhps xmm3, xmm6 movlps xmm6, [rsp + nb111nf_iyH1] movlps xmm7, [rsp + nb111nf_iyH2] unpcklps xmm6, xmm7 movlhps xmm4, xmm6 movlps xmm6, [rsp + nb111nf_izH1] movlps xmm7, [rsp + nb111nf_izH2] unpcklps xmm6, xmm7 movlhps xmm5, xmm6 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm4, xmm3 addps xmm4, xmm5 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [rsp + nb111nf_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb111nf_half] subps xmm1, xmm5 ;# 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv ;# a little trick to avoid NaNs: ;# positions 0,2,and 3 are valid, but not 1. ;# If it contains NaN it doesnt help to mult by 0, ;# So we shuffle it and copy pos 0 to pos1! shufps xmm0, xmm0, 224 ;# 11100000 movaps xmm4, xmm0 mulps xmm4, xmm4 ;# xmm4=rinvsq movaps xmm1, xmm4 mulss xmm1, xmm4 movaps xmm3, [rsp + nb111nf_qqO] mulss xmm1, xmm4 ;# xmm1=rinvsix movaps xmm2, xmm1 mulss xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm3, xmm0 ;# xmm3=vcoul mulps xmm1, [rsp + nb111nf_c6] mulps xmm2, [rsp + nb111nf_c12] movaps xmm5, xmm2 subss xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addps xmm5, [rsp + nb111nf_Vvdwtot] addps xmm3, [rsp + nb111nf_vctot] movaps [rsp + nb111nf_vctot], xmm3 movaps [rsp + nb111nf_Vvdwtot], xmm5 dec dword ptr [rsp + nb111nf_innerk] jz .nb111nf_updateouterdata jmp .nb111nf_odd_loop.nb111nf_updateouterdata: ;# get n from stack mov esi, [rsp + nb111nf_n] ;# get group index for i particle mov rdx, [rbp + nb111nf_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] movaps xmm7, [rsp + nb111nf_vctot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb111nf_Vc] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# accumulate total lj energy and update it movaps xmm7, [rsp + nb111nf_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb111nf_Vvdw] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# finish if last mov ecx, [rsp + nb111nf_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb111nf_outerend ;# not last, iterate outer loop once more! mov [rsp + nb111nf_n], esi jmp .nb111nf_outer.nb111nf_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb111nf_nri] ;# esi already loaded with n above sub ecx, esi jz .nb111nf_end ;# non-zero, do one more workunit jmp .nb111nf_threadloop.nb111nf_end: mov eax, [rsp + nb111nf_nouter] mov ebx, [rsp + nb111nf_ninner] mov rcx, [rbp + nb111nf_outeriter] mov rdx, [rbp + nb111nf_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 424 emms pop rbx pop rbp ret⌨️ 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?