nb_kernel331_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,310 行 · 第 1/5 页
S
2,310 行
shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movss xmm0, [rsp + nb331_ixO] movss xmm1, [rsp + nb331_iyO] movss xmm2, [rsp + nb331_izO] movlps xmm6, [rsp + nb331_ixH1] movlps xmm7, [rsp + nb331_ixH2] unpcklps xmm6, xmm7 movlhps xmm0, xmm6 movlps xmm6, [rsp + nb331_iyH1] movlps xmm7, [rsp + nb331_iyH2] unpcklps xmm6, xmm7 movlhps xmm1, xmm6 movlps xmm6, [rsp + nb331_izH1] movlps xmm7, [rsp + nb331_izH2] unpcklps xmm6, xmm7 movlhps xmm2, xmm6 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [rsp + nb331_dxO], xmm3 movaps [rsp + nb331_dyO], xmm4 movaps [rsp + nb331_dzO], xmm5 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 + nb331_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb331_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 mulps xmm4, xmm0 ;# xmm4=r movaps [rsp + nb331_rinvO], xmm0 mulps xmm4, [rsp + nb331_tsc] movhlps xmm7, xmm4 cvttps2pi mm6, xmm4 cvttps2pi mm7, xmm7 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm7, mm7 movlhps xmm3, xmm7 subps xmm4, xmm3 movaps xmm1, xmm4 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd mm0, eax movd mm1, ecx movd mm2, edx mov rsi, [rbp + nb331_VFtab] movd eax, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 lea rax, [rax + rax*2] lea rcx, [rcx + rcx*2] lea rdx, [rdx + rdx*2] movlps xmm5, [rsi + rax*4] movlps xmm7, [rsi + rcx*4] movhps xmm7, [rsi + rdx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# 10001000 shufps xmm5, xmm7, 221 ;# 11011101 movlps xmm7, [rsi + rax*4 + 8] movlps xmm3, [rsi + rcx*4 + 8] movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# 10001000 shufps xmm7, xmm3, 221 ;# 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [rsp + nb331_two] ;# two*Heps2 movaps xmm0, [rsp + nb331_qqO] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm0 ;# vcoul=qq*VV mulps xmm0, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and xmm0 fijC ;# increment vcoul - then we can get rid of mm5 addps xmm5, [rsp + nb331_vctot] movaps [rsp + nb331_vctot], xmm5 ;# dispersion movlps xmm5, [rsi + rax*4 + 16] ;# half table movaps xmm4, xmm5 shufps xmm4, xmm4, 252 ;# 11111100 shufps xmm5, xmm5, 253 ;# 11111101 movlps xmm7, [rsi + rax*4 + 24] ;# other half of dispersion table movaps xmm6, xmm7 shufps xmm6, xmm6, 252 ;# 11111100 shufps xmm7, xmm7, 253 ;# 11111101 ;# dispersion table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 ;# Update Vvdwtot directly addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [rsp + nb331_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movaps xmm4, [rsp + nb331_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 addps xmm0, xmm7 ;# add to fscal ;# Update Vvdwtot directly addps xmm5, [rsp + nb331_Vvdwtot] movaps [rsp + nb331_Vvdwtot], xmm5 ;# repulsion movlps xmm5, [rsi + rax*4 + 32] ;# got half repulsion table movaps xmm4, xmm5 shufps xmm4, xmm4, 136 ;# 10001000 shufps xmm5, xmm5, 221 ;# 11011101 movlps xmm7, [rsi + rax*4 + 40] ;# other half of repulsion table movaps xmm6, xmm7 shufps xmm6, xmm6, 136 ;# 10001000 shufps xmm7, xmm7, 221 ;# 11011101 ;# repulsion table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [rsp + nb331_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movaps xmm4, [rsp + nb331_c12] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, xmm0 ;# add to fscal addps xmm5, [rsp + nb331_Vvdwtot] ;# total nonbonded potential in xmm5 xorps xmm4, xmm4 movd eax, mm0 movd ecx, mm1 movd edx, mm2 mulps xmm7, [rsp + nb331_rinvO] ;# total fscal now in xmm7 movaps [rsp + nb331_Vvdwtot], xmm5 mulps xmm7, [rsp + nb331_tsc] subps xmm4, xmm7 movaps xmm0, [rsp + nb331_dxO] movaps xmm1, [rsp + nb331_dyO] movaps xmm2, [rsp + nb331_dzO] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 now contains tx-tz (partial force) movss xmm3, [rsp + nb331_fixO] movss xmm4, [rsp + nb331_fiyO] movss xmm5, [rsp + nb331_fizO] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [rsp + nb331_fixO], xmm3 movss [rsp + nb331_fiyO], xmm4 movss [rsp + nb331_fizO], xmm5 ;# updated the O force now do the H's movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 230 ;# 11100110 ;# shift right shufps xmm4, xmm4, 230 ;# 11100110 shufps xmm5, xmm5, 230 ;# 11100110 addss xmm3, [rsp + nb331_fixH1] addss xmm4, [rsp + nb331_fiyH1] addss xmm5, [rsp + nb331_fizH1] movss [rsp + nb331_fixH1], xmm3 movss [rsp + nb331_fiyH1], xmm4 movss [rsp + nb331_fizH1], xmm5 ;# updated the H1 force mov rdi, [rbp + nb331_faction] shufps xmm3, xmm3, 231 ;# 11100111 ;# shift right shufps xmm4, xmm4, 231 ;# 11100111 shufps xmm5, xmm5, 231 ;# 11100111 addss xmm3, [rsp + nb331_fixH2] addss xmm4, [rsp + nb331_fiyH2] addss xmm5, [rsp + nb331_fizH2] movss [rsp + nb331_fixH2], xmm3 movss [rsp + nb331_fiyH2], xmm4 movss [rsp + nb331_fizH2], xmm5 ;# updated the H2 force ;# the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps xmm5, xmm5 movaps xmm3, xmm0 movlps xmm6, [rdi + rax*4] movss xmm7, [rdi + rax*4 + 8] unpcklps xmm3, xmm1 movlhps xmm3, xmm5 unpckhps xmm0, xmm1 addps xmm0, xmm3 movhlps xmm3, xmm0 addps xmm0, xmm3 ;# x,y sum in xmm0 movhlps xmm1, xmm2 addss xmm2, xmm1 shufps xmm1, xmm1, 1 addss xmm2, xmm1 ;# z sum in xmm2 addps xmm6, xmm0 addss xmm7, xmm2 movlps [rdi + rax*4], xmm6 movss [rdi + rax*4 + 8], xmm7 dec dword ptr [rsp + nb331_innerk] jz .nb331_updateouterdata jmp .nb331_odd_loop.nb331_updateouterdata: mov ecx, [rsp + nb331_ii3] mov rdi, [rbp + nb331_faction] mov rsi, [rbp + nb331_fshift] mov edx, [rsp + nb331_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb331_fixO] movaps xmm1, [rsp + nb331_fiyO] movaps xmm2, [rsp + nb331_fizO] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4] movss xmm4, [rdi + rcx*4 + 4] movss xmm5, [rdi + rcx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4], xmm3 movss [rdi + rcx*4 + 4], xmm4 movss [rdi + rcx*4 + 8], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movaps xmm6, xmm0 movss xmm7, xmm2 movlhps xmm6, xmm1 shufps xmm6, xmm6, 8 ;# 00001000 ;# accumulate H1i forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb331_fixH1] movaps xmm1, [rsp + nb331_fiyH1] movaps xmm2, [rsp + nb331_fizH1] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4 + 12] movss xmm4, [rdi + rcx*4 + 16] movss xmm5, [rdi + rcx*4 + 20] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4 + 12], xmm3 movss [rdi + rcx*4 + 16], xmm4 movss [rdi + rcx*4 + 20], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# 00001000 addps xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movaps xmm0, [rsp + nb331_fixH2] movaps xmm1, [rsp + nb331_fiyH2] movaps xmm2, [rsp + nb331_fizH2] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4 + 24] movss xmm4, [rdi + rcx*4 + 28] movss xmm5, [rdi + rcx*4 + 32] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4 + 24], xmm3 movss [rdi + rcx*4 + 28], xmm4 movss [rdi + rcx*4 + 32], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# 00001000 addps xmm6, xmm0 ;# increment fshift force movlps xmm3, [rsi + rdx*4] movss xmm4, [rsi + rdx*4 + 8] subps xmm3, xmm6 subss xmm4, xmm7 movlps [rsi + rdx*4], xmm3 movss [rsi + rdx*4 + 8], xmm4 ;# get n from stack mov esi, [rsp + nb331_n] ;# get group index for i particle mov rdx, [rbp + nb331_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [rsp + nb331_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 + nb331_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 + nb331_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 + nb331_Vvdw] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# finish if last mov ecx, [rsp + nb331_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb331_outerend ;# not last, iterate outer loop once more! mov [rsp + nb331_n], esi jmp .nb331_outer.nb331_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb331_nri] ;# esi already loaded with n above sub ecx, esi jz .nb331_end ;# non-zero, do one more workunit jmp .nb331_threadloop.nb331_end: mov eax, [rsp + nb331_nouter] mov ebx, [rsp + nb331_ninner] mov rcx, [rbp + nb331_outeriter] mov rdx, [rbp + nb331_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 872 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret .globl nb_kernel331nf_x86_64_sse⌨️ 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?