nb_kernel112_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,220 行 · 第 1/5 页
S
2,220 行
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 + nb112nf_ixH1], xmm0 movaps [rsp + nb112nf_iyH1], xmm1 movaps [rsp + nb112nf_izH1], xmm2 movaps [rsp + nb112nf_ixH2], xmm3 movaps [rsp + nb112nf_iyH2], xmm4 movaps [rsp + nb112nf_izH2], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [rsp + nb112nf_vctot], xmm4 movaps [rsp + nb112nf_Vvdwtot], xmm4 mov rax, [rsp + nb112nf_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 + nb112nf_pos] mov rax, [rsp + nb112nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb112nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [rsp + nb112nf_ninner] mov [rsp + nb112nf_ninner], ecx add edx, 0 mov [rsp + nb112nf_innerk], edx ;# number of innerloop atoms jge .nb112nf_unroll_loop jmp .nb112nf_single_check.nb112nf_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb112nf_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 + nb112nf_innerjjnr], 16 ;# advance pointer (unrolled 4) mov rsi, [rbp + nb112nf_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 j coordinates to local temp variables movlps xmm2, [rsi + rax*4] movlps xmm3, [rsi + rax*4 + 12] movlps xmm4, [rsi + rax*4 + 24] movlps xmm5, [rsi + rbx*4] movlps xmm6, [rsi + rbx*4 + 12] movlps xmm7, [rsi + rbx*4 + 24] movhps xmm2, [rsi + rcx*4] movhps xmm3, [rsi + rcx*4 + 12] movhps xmm4, [rsi + rcx*4 + 24] movhps xmm5, [rsi + rdx*4] movhps xmm6, [rsi + rdx*4 + 12] movhps xmm7, [rsi + rdx*4 + 24] ;# 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 xmm0, xmm2 movaps xmm1, xmm3 unpcklps xmm0, xmm5 ;# xmm0= jxOa jxOb jyOa jyOb unpcklps xmm1, xmm6 ;# xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps xmm2, xmm5 ;# xmm2= jxOc jxOd jyOc jyOd unpckhps xmm3, xmm6 ;# xmm3= jxH1c jxH1d jyH1c jyH1d movaps xmm5, xmm4 movaps xmm6, xmm0 unpcklps xmm4, xmm7 ;# xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps xmm5, xmm7 ;# xmm5= jxH2c jxH2d jyH2c jyH2d movaps xmm7, xmm1 movlhps xmm0, xmm2 ;# xmm0= jxOa jxOb jxOc jxOd movaps [rsp + nb112nf_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [rsp + nb112nf_jyO], xmm2 movlhps xmm1, xmm3 movaps [rsp + nb112nf_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [rsp + nb112nf_jyH1], xmm3 movlhps xmm4, xmm5 movaps [rsp + nb112nf_jxH2], xmm4 movhlps xmm5, xmm6 movaps [rsp + nb112nf_jyH2], xmm5 movss xmm0, [rsi + rax*4 + 8] movss xmm1, [rsi + rax*4 + 20] movss xmm2, [rsi + rax*4 + 32] movss xmm3, [rsi + rcx*4 + 8] movss xmm4, [rsi + rcx*4 + 20] movss xmm5, [rsi + rcx*4 + 32] movhps xmm0, [rsi + rbx*4 + 4] movhps xmm1, [rsi + rbx*4 + 16] movhps xmm2, [rsi + rbx*4 + 28] movhps xmm3, [rsi + rdx*4 + 4] movhps xmm4, [rsi + rdx*4 + 16] movhps xmm5, [rsi + rdx*4 + 28] shufps xmm0, xmm3, 204 ;# 11001100 shufps xmm1, xmm4, 204 ;# 11001100 shufps xmm2, xmm5, 204 ;# 11001100 movaps [rsp + nb112nf_jzO], xmm0 movaps [rsp + nb112nf_jzH1], xmm1 movaps [rsp + nb112nf_jzH2], xmm2 movaps xmm0, [rsp + nb112nf_ixO] movaps xmm1, [rsp + nb112nf_iyO] movaps xmm2, [rsp + nb112nf_izO] movaps xmm3, [rsp + nb112nf_ixO] movaps xmm4, [rsp + nb112nf_iyO] movaps xmm5, [rsp + nb112nf_izO] subps xmm0, [rsp + nb112nf_jxO] subps xmm1, [rsp + nb112nf_jyO] subps xmm2, [rsp + nb112nf_jzO] subps xmm3, [rsp + nb112nf_jxH1] subps xmm4, [rsp + nb112nf_jyH1] subps xmm5, [rsp + nb112nf_jzH1] mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [rsp + nb112nf_rsqOO], xmm0 movaps [rsp + nb112nf_rsqOH1], xmm3 movaps xmm0, [rsp + nb112nf_ixO] movaps xmm1, [rsp + nb112nf_iyO] movaps xmm2, [rsp + nb112nf_izO] movaps xmm3, [rsp + nb112nf_ixH1] movaps xmm4, [rsp + nb112nf_iyH1] movaps xmm5, [rsp + nb112nf_izH1] subps xmm0, [rsp + nb112nf_jxH2] subps xmm1, [rsp + nb112nf_jyH2] subps xmm2, [rsp + nb112nf_jzH2] subps xmm3, [rsp + nb112nf_jxO] subps xmm4, [rsp + nb112nf_jyO] subps xmm5, [rsp + nb112nf_jzO] mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [rsp + nb112nf_rsqOH2], xmm0 movaps [rsp + nb112nf_rsqH1O], xmm3 movaps xmm0, [rsp + nb112nf_ixH1] movaps xmm1, [rsp + nb112nf_iyH1] movaps xmm2, [rsp + nb112nf_izH1] movaps xmm3, [rsp + nb112nf_ixH1] movaps xmm4, [rsp + nb112nf_iyH1] movaps xmm5, [rsp + nb112nf_izH1] subps xmm0, [rsp + nb112nf_jxH1] subps xmm1, [rsp + nb112nf_jyH1] subps xmm2, [rsp + nb112nf_jzH1] subps xmm3, [rsp + nb112nf_jxH2] subps xmm4, [rsp + nb112nf_jyH2] subps xmm5, [rsp + nb112nf_jzH2] mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [rsp + nb112nf_rsqH1H1], xmm0 movaps [rsp + nb112nf_rsqH1H2], xmm3 movaps xmm0, [rsp + nb112nf_ixH2] movaps xmm1, [rsp + nb112nf_iyH2] movaps xmm2, [rsp + nb112nf_izH2] movaps xmm3, [rsp + nb112nf_ixH2] movaps xmm4, [rsp + nb112nf_iyH2] movaps xmm5, [rsp + nb112nf_izH2] subps xmm0, [rsp + nb112nf_jxO] subps xmm1, [rsp + nb112nf_jyO] subps xmm2, [rsp + nb112nf_jzO] subps xmm3, [rsp + nb112nf_jxH1] subps xmm4, [rsp + nb112nf_jyH1] subps xmm5, [rsp + nb112nf_jzH1] mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm4, xmm3 addps xmm4, xmm5 movaps [rsp + nb112nf_rsqH2O], xmm0 movaps [rsp + nb112nf_rsqH2H1], xmm4 movaps xmm0, [rsp + nb112nf_ixH2] movaps xmm1, [rsp + nb112nf_iyH2] movaps xmm2, [rsp + nb112nf_izH2] subps xmm0, [rsp + nb112nf_jxH2] subps xmm1, [rsp + nb112nf_jyH2] subps xmm2, [rsp + nb112nf_jzH2] mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [rsp + nb112nf_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb112nf_three] movaps xmm7, xmm3 mulps xmm1, xmm0 mulps xmm5, xmm4 subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [rsp + nb112nf_half] ;# rinvH2H2 mulps xmm7, [rsp + nb112nf_half] ;# rinvH2H1 movaps [rsp + nb112nf_rinvH2H2], xmm3 movaps [rsp + nb112nf_rinvH2H1], xmm7 rsqrtps xmm1, [rsp + nb112nf_rsqOO] rsqrtps xmm5, [rsp + nb112nf_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb112nf_three] movaps xmm7, xmm3 mulps xmm1, [rsp + nb112nf_rsqOO] mulps xmm5, [rsp + nb112nf_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [rsp + nb112nf_half] mulps xmm7, [rsp + nb112nf_half] movaps [rsp + nb112nf_rinvOO], xmm3 movaps [rsp + nb112nf_rinvOH1], xmm7 rsqrtps xmm1, [rsp + nb112nf_rsqOH2] rsqrtps xmm5, [rsp + nb112nf_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb112nf_three] movaps xmm7, xmm3 mulps xmm1, [rsp + nb112nf_rsqOH2] mulps xmm5, [rsp + nb112nf_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [rsp + nb112nf_half] mulps xmm7, [rsp + nb112nf_half] movaps [rsp + nb112nf_rinvOH2], xmm3 movaps [rsp + nb112nf_rinvH1O], xmm7 rsqrtps xmm1, [rsp + nb112nf_rsqH1H1] rsqrtps xmm5, [rsp + nb112nf_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [rsp + nb112nf_three] movaps xmm7, xmm3 mulps xmm1, [rsp + nb112nf_rsqH1H1] mulps xmm5, [rsp + nb112nf_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [rsp + nb112nf_half] mulps xmm7, [rsp + nb112nf_half] movaps [rsp + nb112nf_rinvH1H1], xmm3 movaps [rsp + nb112nf_rinvH1H2], xmm7 rsqrtps xmm1, [rsp + nb112nf_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [rsp + nb112nf_three] mulps xmm1, [rsp + nb112nf_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [rsp + nb112nf_half] movaps [rsp + nb112nf_rinvH2O], xmm3 ;# start with OO interaction movaps xmm0, [rsp + nb112nf_rinvOO] movaps xmm7, xmm0 mulps xmm0, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix mulps xmm7, [rsp + nb112nf_qqOO] movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [rsp + nb112nf_c6] mulps xmm2, [rsp + nb112nf_c12] subps xmm2, xmm1 ;# xmm3=Vvdw12-Vvdw6 addps xmm2, [rsp + nb112nf_Vvdwtot] movaps [rsp + nb112nf_Vvdwtot], xmm2 addps xmm7, [rsp + nb112nf_vctot] ;# all other interaction movaps xmm0, [rsp + nb112nf_rinvOH1] movaps xmm1, [rsp + nb112nf_rinvH1H1] addps xmm0, [rsp + nb112nf_rinvOH2] addps xmm1, [rsp + nb112nf_rinvH1H2] addps xmm0, [rsp + nb112nf_rinvH1O] addps xmm1, [rsp + nb112nf_rinvH2H1] addps xmm0, [rsp + nb112nf_rinvH2O] addps xmm1, [rsp + nb112nf_rinvH2H2] mulps xmm0, [rsp + nb112nf_qqOH] mulps xmm1, [rsp + nb112nf_qqHH] addps xmm7, xmm0 addps xmm7, xmm1 movaps [rsp + nb112nf_vctot], xmm7 ;# should we do one more iteration? sub dword ptr [rsp + nb112nf_innerk], 4 jl .nb112nf_single_check jmp .nb112nf_unroll_loop.nb112nf_single_check: add dword ptr [rsp + nb112nf_innerk], 4 jnz .nb112nf_single_loop jmp .nb112nf_updateouterdata.nb112nf_single_loop: mov rdx, [rsp + nb112nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] add qword ptr [rsp + nb112nf_innerjjnr], 4 mov rsi, [rbp + nb112nf_pos] lea rax, [rax + rax*2] ;# fetch j coordinates xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 movss xmm3, [rsi + rax*4] ;# jxO - - - movss xmm4, [rsi + rax*4 + 4] ;# jyO - - - movss xmm5, [rsi + rax*4 + 8] ;# jzO - - - movlps xmm6, [rsi + rax*4 + 12] ;# xmm6 = jxH1 jyH1 - - movss xmm7, [rsi + rax*4 + 20] ;# xmm7 = jzH1 - - - movhps xmm6, [rsi + rax*4 + 24] ;# xmm6 = jxH1 jyH1 jxH2 jyH2 movss xmm2, [rsi + rax*4 + 32] ;# xmm2 = jzH2 - - - ;# have all coords, time for some shuffling. shufps xmm6, xmm6, 216 ;# 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps xmm7, xmm2 ;# xmm7 = jzH1 jzH2 - - movaps xmm0, [rsp + nb112nf_ixO] movaps xmm1, [rsp + nb112nf_iyO] movaps xmm2, [rsp + nb112nf_izO] movlhps xmm3, xmm6 ;# xmm3 = jxO 0 jxH1 jxH2 shufps xmm4, xmm6, 228 ;# 11100100 ;# xmm4 = jyO 0 jyH1 jyH2 shufps xmm5, xmm7, 68 ;# 01000100 ;# xmm5 = jzO 0 jzH1 jzH2 ;# store all j coordinates in jO movaps [rsp + nb112nf_jxO], xmm3 movaps [rsp + nb112nf_jyO], xmm4 movaps [rsp + nb112nf_jzO], xmm5 subps xmm0, xmm3 subps xmm1, xmm4 subps xmm2, xmm5 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 ;# have rsq in xmm0 ;# do invsqrt rsqrtps xmm1, xmm0 movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [rsp + nb112nf_three] mulps xmm1, xmm0 subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [rsp + nb112nf_half] ;# rinv iO - j water xorps xmm1, xmm1 movaps xmm0, xmm3 xorps xmm4, xmm4 mulps xmm0, xmm0 ;# xmm0=rinvsq ;# fetch charges to xmm4 (temporary) movss xmm4, [rsp + nb112nf_qqOO] movss xmm1, xmm0 movhps xmm4, [rsp + nb112nf_qqOH] mulss xmm1, xmm0 mulps xmm3, xmm4 ;# xmm3=vcoul mulss xmm1, xmm0 ;# xmm1(0)=rinvsix movaps xmm2, xmm1 ;# zero everything else in xmm2 mulss xmm2, xmm2 ;# xmm2=rinvtwelve mulss xmm1, [rsp + nb112nf_c6] mulss xmm2, [rsp + nb112nf_c12] movaps xmm4, xmm2 subss xmm4, x⌨️ 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?