nb_kernel213_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,954 行 · 第 1/4 页
S
1,954 行
addss xmm4, [rax + rbx*4 + 40] ;# my addss xmm5, [rax + rbx*4 + 44] ;# mz shufps xmm6, xmm6, 0 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 + nb213nf_izH1], xmm6 movaps [rsp + nb213nf_ixH2], xmm0 movaps [rsp + nb213nf_iyH2], xmm1 movaps [rsp + nb213nf_izH2], xmm2 movaps [rsp + nb213nf_ixM], xmm3 movaps [rsp + nb213nf_iyM], xmm4 movaps [rsp + nb213nf_izM], xmm5 ;# clear vctot xorps xmm4, xmm4 movaps [rsp + nb213nf_vctot], xmm4 movaps [rsp + nb213nf_Vvdwtot], xmm4 mov rax, [rsp + nb213nf_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 + nb213nf_pos] mov rdi, [rbp + nb213nf_faction] mov rax, [rsp + nb213nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb213nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [rsp + nb213nf_ninner] mov [rsp + nb213nf_ninner], ecx add edx, 0 mov [rsp + nb213nf_innerk], edx ;# number of innerloop atoms jge .nb213nf_unroll_loop jmp .nb213nf_odd_inner.nb213nf_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb213nf_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 + nb213nf_innerjjnr], 16 ;# advance pointer (unrolled 4) mov rsi, [rbp + nb213nf_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 + nb213nf_iqM] mulps xmm4, [rsp + nb213nf_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movd mm2, ecx movd mm3, edx movaps [rsp + nb213nf_qqM], xmm3 movaps [rsp + nb213nf_qqH], xmm4 mov rsi, [rbp + nb213nf_type] mov eax, [rsi + rax*4] mov ebx, [rsi + rbx*4] mov ecx, [rsi + rcx*4] mov edx, [rsi + rdx*4] mov rsi, [rbp + nb213nf_vdwparam] shl eax, 1 shl ebx, 1 shl ecx, 1 shl edx, 1 mov edi, [rsp + nb213nf_ntia] add rax, rdi add rbx, rdi add rcx, rdi add rdx, rdi 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 + nb213nf_c6], xmm4 movaps [rsp + nb213nf_c12], xmm6 mov rsi, [rbp + nb213nf_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 + nb213nf_ixO] movaps xmm5, [rsp + nb213nf_iyO] movaps xmm6, [rsp + nb213nf_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 + nb213nf_ixH1] movaps xmm5, [rsp + nb213nf_iyH1] movaps xmm6, [rsp + nb213nf_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 + nb213nf_ixH2] movaps xmm4, [rsp + nb213nf_iyH2] movaps xmm5, [rsp + nb213nf_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 ;# move ixM-izM to xmm2-xmm4 movaps xmm3, [rsp + nb213nf_iyM] movaps xmm4, [rsp + nb213nf_izM] subps xmm3, xmm1 subps xmm4, xmm2 movaps xmm2, [rsp + nb213nf_ixM] subps xmm2, xmm0 ;# square it mulps xmm2,xmm2 mulps xmm3,xmm3 mulps xmm4,xmm4 addps xmm4, xmm3 addps xmm4, xmm2 ;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movaps xmm0, xmm4 movaps xmm1, xmm5 movaps xmm2, xmm6 mulps xmm0, [rsp + nb213nf_krf] mulps xmm1, [rsp + nb213nf_krf] mulps xmm2, [rsp + nb213nf_krf] movaps [rsp + nb213nf_krsqM], xmm0 movaps [rsp + nb213nf_krsqH2], xmm1 movaps [rsp + nb213nf_krsqH1], xmm2 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [rsp + nb213nf_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm0, xmm2 ;# 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [rsp + nb213nf_half] movaps [rsp + nb213nf_rinvH1], xmm0 ;# rinvH1 ;# rsqH2 - seed to xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [rsp + nb213nf_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm0, xmm2 ;# 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [rsp + nb213nf_half] movaps [rsp + nb213nf_rinvH2], xmm0 ;# rinvH2 ;# rsqM - seed to xmm2 rsqrtps xmm2, xmm4 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [rsp + nb213nf_three] mulps xmm2, xmm4 ;# rsq*lu*lu subps xmm0, xmm2 ;# 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [rsp + nb213nf_half] movaps [rsp + nb213nf_rinvM], xmm0 ;# Do the O LJ-only interaction directly. rcpps xmm2, xmm7 movaps xmm1, [rsp + nb213nf_two] mulps xmm7, xmm2 subps xmm1, xmm7 mulps xmm2, xmm1 ;# rinvsq movaps xmm0, xmm2 mulps xmm0, xmm2 ;# r4 mulps xmm0, xmm2 ;# r6 movaps xmm1, xmm0 mulps xmm1, xmm1 ;# r12 mulps xmm0, [rsp + nb213nf_c6] mulps xmm1, [rsp + nb213nf_c12] movaps xmm3, xmm1 subps xmm3, xmm0 ;# Vvdw12-Vvdw6 addps xmm3, [rsp + nb213nf_Vvdwtot] movaps [rsp + nb213nf_Vvdwtot], xmm3 ;# do H1 interactions movaps xmm7, [rsp + nb213nf_rinvH1] addps xmm7, [rsp + nb213nf_krsqH1] subps xmm7, [rsp + nb213nf_crf] ;# xmm7=rinv+ krsq-crf mulps xmm7, [rsp + nb213nf_qqH] addps xmm7, [rsp + nb213nf_vctot] ;# H2 interactions movaps xmm6, [rsp + nb213nf_rinvH2] addps xmm6, [rsp + nb213nf_krsqH2] subps xmm6, [rsp + nb213nf_crf] ;# xmm6=rinv+ krsq-crf mulps xmm6, [rsp + nb213nf_qqH] addps xmm6, xmm7 ;# M interactions movaps xmm5, [rsp + nb213nf_rinvM] addps xmm5, [rsp + nb213nf_krsqM] subps xmm5, [rsp + nb213nf_crf] ;# xmm5=rinv+ krsq-crf mulps xmm5, [rsp + nb213nf_qqM] addps xmm5, xmm6 movaps [rsp + nb213nf_vctot], xmm5 ;# should we do one more iteration? sub dword ptr [rsp + nb213nf_innerk], 4 jl .nb213nf_odd_inner jmp .nb213nf_unroll_loop.nb213nf_odd_inner: add dword ptr [rsp + nb213nf_innerk], 4 jnz .nb213nf_odd_loop jmp .nb213nf_updateouterdata.nb213nf_odd_loop: mov rdx, [rsp + nb213nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] add qword ptr [rsp + nb213nf_innerjjnr], 4 xorps xmm4, xmm4 ;# clear reg. movss xmm4, [rsp + nb213nf_iqM] mov rsi, [rbp + nb213nf_charge] movhps xmm4, [rsp + nb213nf_iqH] ;# [qM 0 qH qH] shufps xmm4, xmm4, 41 ;# [0 qH qH qM] movss xmm3, [rsi + rax*4] ;# charge in xmm3 shufps xmm3, xmm3, 0 mulps xmm3, xmm4 movaps [rsp + nb213nf_qqM], xmm3 ;# use dummy qq for storage xorps xmm6, xmm6 mov rsi, [rbp + nb213nf_type] mov ebx, [rsi + rax*4] mov rsi, [rbp + nb213nf_vdwparam] shl ebx, 1 add ebx, [rsp + nb213nf_ntia] movlps xmm6, [rsi + rbx*4] movaps xmm7, xmm6 shufps xmm6, xmm6, 252 ;# 11111100 shufps xmm7, xmm7, 253 ;# 11111101 movaps [rsp + nb213nf_c6], xmm6 movaps [rsp + nb213nf_c12], xmm7 mov rsi, [rbp + nb213nf_pos] lea rax, [rax + rax*2] movss xmm3, [rsp + nb213nf_ixO] movss xmm4, [rsp + nb213nf_iyO] movss xmm5, [rsp + nb213nf_izO] movss xmm0, [rsp + nb213nf_ixH1] movss xmm1, [rsp + nb213nf_iyH1] movss xmm2, [rsp + nb213nf_izH1] unpcklps xmm3, [rsp + nb213nf_ixH2] ;# ixO ixH2 - - unpcklps xmm4, [rsp + nb213nf_iyH2] ;# iyO iyH2 - - unpcklps xmm5, [rsp + nb213nf_izH2] ;# izO izH2 - - unpcklps xmm0, [rsp + nb213nf_ixM] ;# ixH1 ixM - - unpcklps xmm1, [rsp + nb213nf_iyM] ;# iyH1 iyM - - unpcklps xmm2, [rsp + nb213nf_izM] ;# izH1 izM - - unpcklps xmm3, xmm0 ;# ixO ixH1 ixH2 ixM unpcklps xmm4, xmm1 ;# same for y unpcklps xmm5, xmm2 ;# same for z ;# 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 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 movaps xmm0, xmm4 mulps xmm0, [rsp + nb213nf_krf] movaps [rsp + nb213nf_krsqM], xmm0 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [rsp + nb213nf_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb213nf_half] subps xmm1, xmm5 ;# 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv movaps xmm1, xmm0 ;# xmm1=rinv addps xmm1, [rsp + nb213nf_krsqM] subps xmm1, [rsp + nb213nf_crf] ;# xmm0=rinv+ krsq-crf mulps xmm1, [rsp + nb213nf_qqM] addps xmm1, [rsp + nb213nf_vctot] movaps [rsp + nb213nf_vctot], xmm1 movaps xmm1, xmm0 mulps xmm1, xmm1 movaps xmm2, xmm1 mulss xmm1, xmm1 mulss xmm1, xmm2 ;# xmm1=rinvsix xorps xmm4, xmm4 movss xmm4, xmm1 mulss xmm4, xmm4 ;# xmm4=rinvtwelve mulss xmm1, [rsp + nb213nf_c6] mulss xmm4, [rsp + nb213nf_c12] movaps xmm3, xmm4 subss xmm3, xmm1 ;# xmm3=Vvdw12-Vvdw6 addss xmm3, [rsp + nb213nf_Vvdwtot] movss [rsp + nb213nf_Vvdwtot], xmm3 dec dword ptr [rsp + nb213nf_innerk] jz .nb213nf_updateouterdata jmp .nb213nf_odd_loop.nb213nf_updateouterdata: ;# get n from stack mov esi, [rsp + nb213nf_n] ;# get group index for i particle mov rdx, [rbp + nb213nf_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [rsp + nb213nf_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 + nb213nf_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 + nb213nf_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 + nb213nf_Vvdw] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# finish if last mov ecx, [rsp + nb213nf_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb213nf_outerend ;# not last, iterate outer loop once more! mov [rsp + nb213nf_n], esi jmp .nb213nf_outer.nb213nf_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb213nf_nri] ;# esi already loaded with n above sub ecx, esi jz .nb213nf_end ;# non-zero, do one more workunit jmp .nb213nf_threadloop.nb213nf_end: mov eax, [rsp + nb213nf_nouter] mov ebx, [rsp + nb213nf_ninner] mov rcx, [rbp + nb213nf_outeriter] mov rdx, [rbp + nb213nf_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 600 emms pop r15 pop r14 pop r13 pop r12 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?