nb_kernel131_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,955 行 · 第 1/4 页
S
1,955 行
movaps %xmm10,nb131_rinvH1(%rsp) movaps %xmm11,nb131_rinvH2(%rsp) ## table LJ interaction mulps %xmm9,%xmm0 mulps nb131_tsc(%rsp),%xmm0 ## rtab ## truncate and convert to integers cvttps2dq %xmm0,%xmm1 ## convert back to float cvtdq2ps %xmm1,%xmm2 ## multiply by 8 pslld $3,%xmm1 ## move to integer registers movhlps %xmm1,%xmm13 movd %xmm1,%r8d movd %xmm13,%r10d shufps $1,%xmm1,%xmm1 shufps $1,%xmm13,%xmm13 movd %xmm1,%r9d movd %xmm13,%r11d ## calculate eps subps %xmm2,%xmm0 movq nb131_VFtab(%rbp),%rsi movlps (%rsi,%r8,4),%xmm5 movlps 16(%rsi,%r8,4),%xmm9 movlps (%rsi,%r10,4),%xmm7 movlps 16(%rsi,%r10,4),%xmm11 movhps (%rsi,%r9,4),%xmm5 movhps 16(%rsi,%r9,4),%xmm9 movhps (%rsi,%r11,4),%xmm7 movhps 16(%rsi,%r11,4),%xmm11 movaps %xmm5,%xmm4 movaps %xmm9,%xmm8 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $136,%xmm11,%xmm8 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 shufps $221,%xmm11,%xmm9 ## 11011101 movlps 8(%rsi,%r8,4),%xmm7 movlps 24(%rsi,%r8,4),%xmm11 movlps 8(%rsi,%r10,4),%xmm13 movlps 24(%rsi,%r10,4),%xmm14 movhps 8(%rsi,%r9,4),%xmm7 movhps 24(%rsi,%r9,4),%xmm11 movhps 8(%rsi,%r11,4),%xmm13 movhps 24(%rsi,%r11,4),%xmm14 movaps %xmm7,%xmm6 movaps %xmm11,%xmm10 shufps $136,%xmm13,%xmm6 ## 10001000 shufps $136,%xmm14,%xmm10 ## 10001000 shufps $221,%xmm13,%xmm7 ## 11011101 shufps $221,%xmm14,%xmm11 ## 11011101 ## dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11 mulps %xmm0,%xmm7 ## Heps mulps %xmm0,%xmm11 mulps %xmm0,%xmm6 ## Geps mulps %xmm0,%xmm10 mulps %xmm0,%xmm7 ## Heps2 mulps %xmm0,%xmm11 addps %xmm6,%xmm5 ## F+Geps addps %xmm10,%xmm9 addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm11,%xmm9 addps %xmm7,%xmm7 ## 2*Heps2 addps %xmm11,%xmm11 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm10,%xmm11 addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps addps %xmm9,%xmm11 mulps %xmm0,%xmm5 ## eps*Fp mulps %xmm0,%xmm9 movaps nb131_c6(%rsp),%xmm12 movaps nb131_c12(%rsp),%xmm13 addps %xmm4,%xmm5 ## VV addps %xmm8,%xmm9 mulps %xmm12,%xmm5 ## VV*c6 = vnb6 mulps %xmm13,%xmm9 ## VV*c12 = vnb12 addps %xmm9,%xmm5 addps nb131_Vvdwtot(%rsp),%xmm5 movaps %xmm5,nb131_Vvdwtot(%rsp) mulps %xmm12,%xmm7 ## FF*c6 = fnb6 mulps %xmm13,%xmm11 ## FF*c12 = fnb12 addps %xmm11,%xmm7 mulps nb131_tsc(%rsp),%xmm7 movaps nb131_rinvO(%rsp),%xmm9 movaps nb131_rinvH1(%rsp),%xmm10 movaps nb131_rinvH2(%rsp),%xmm11 movaps %xmm9,%xmm0 ## rinv movaps %xmm10,%xmm1 movaps %xmm11,%xmm2 mulps %xmm10,%xmm10 ## rinvsq mulps %xmm11,%xmm11 mulps nb131_qqO(%rsp),%xmm0 mulps nb131_qqH(%rsp),%xmm1 mulps nb131_qqH(%rsp),%xmm2 mulps %xmm0,%xmm9 mulps %xmm1,%xmm10 mulps %xmm2,%xmm11 subps %xmm7,%xmm9 mulps nb131_rinvO(%rsp),%xmm9 addps nb131_vctot(%rsp),%xmm0 addps %xmm2,%xmm1 addps %xmm1,%xmm0 movaps %xmm0,nb131_vctot(%rsp) ## move j forces to local temp variables movq nb131_faction(%rbp),%rdi movlps (%rdi,%rax,4),%xmm0 ## jxa jya - - movlps (%rdi,%rcx,4),%xmm1 ## jxc jyc - - movhps (%rdi,%rbx,4),%xmm0 ## jxa jya jxb jyb movhps (%rdi,%rdx,4),%xmm1 ## jxc jyc jxd jyd movss 8(%rdi,%rax,4),%xmm2 ## jza - - - movss 8(%rdi,%rcx,4),%xmm3 ## jzc - - - movss 8(%rdi,%rbx,4),%xmm4 ## jzb movss 8(%rdi,%rdx,4),%xmm5 ## jzd movlhps %xmm4,%xmm2 ## jza - jzb - movlhps %xmm5,%xmm3 ## jzc - jzd - shufps $136,%xmm3,%xmm2 ## 10001000 => jza jzb jzc jzd ## xmm0: jxa jya jxb jyb ## xmm1: jxc jyc jxd jyd ## xmm2: jza jzb jzc jzd movaps %xmm9,%xmm7 movaps %xmm9,%xmm8 movaps %xmm11,%xmm13 movaps %xmm11,%xmm14 movaps %xmm11,%xmm15 movaps %xmm10,%xmm11 movaps %xmm10,%xmm12 mulps nb131_dxO(%rsp),%xmm7 mulps nb131_dyO(%rsp),%xmm8 mulps nb131_dzO(%rsp),%xmm9 mulps nb131_dxH1(%rsp),%xmm10 mulps nb131_dyH1(%rsp),%xmm11 mulps nb131_dzH1(%rsp),%xmm12 mulps nb131_dxH2(%rsp),%xmm13 mulps nb131_dyH2(%rsp),%xmm14 mulps nb131_dzH2(%rsp),%xmm15 movaps %xmm7,%xmm3 movaps %xmm8,%xmm4 addps %xmm9,%xmm2 addps nb131_fixO(%rsp),%xmm7 addps nb131_fiyO(%rsp),%xmm8 addps nb131_fizO(%rsp),%xmm9 addps %xmm10,%xmm3 addps %xmm11,%xmm4 addps %xmm12,%xmm2 addps nb131_fixH1(%rsp),%xmm10 addps nb131_fiyH1(%rsp),%xmm11 addps nb131_fizH1(%rsp),%xmm12 addps %xmm13,%xmm3 addps %xmm14,%xmm4 addps %xmm15,%xmm2 addps nb131_fixH2(%rsp),%xmm13 addps nb131_fiyH2(%rsp),%xmm14 addps nb131_fizH2(%rsp),%xmm15 movaps %xmm7,nb131_fixO(%rsp) movaps %xmm8,nb131_fiyO(%rsp) movaps %xmm9,nb131_fizO(%rsp) movaps %xmm10,nb131_fixH1(%rsp) movaps %xmm11,nb131_fiyH1(%rsp) movaps %xmm12,nb131_fizH1(%rsp) movaps %xmm13,nb131_fixH2(%rsp) movaps %xmm14,nb131_fiyH2(%rsp) movaps %xmm15,nb131_fizH2(%rsp) ## xmm0 = fjx ## xmm1 = fjy ## xmm2 = fjz movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movhlps %xmm2,%xmm3 movlps %xmm0,(%rdi,%rax,4) movhps %xmm0,(%rdi,%rbx,4) movlps %xmm1,(%rdi,%rcx,4) movhps %xmm1,(%rdi,%rdx,4) movss %xmm2,8(%rdi,%rax,4) movss %xmm3,8(%rdi,%rcx,4) shufps $1,%xmm2,%xmm2 shufps $1,%xmm3,%xmm3 movss %xmm2,8(%rdi,%rbx,4) movss %xmm3,8(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb131_innerk(%rsp) jl _nb_kernel131_x86_64_sse.nb131_odd_inner jmp _nb_kernel131_x86_64_sse.nb131_unroll_loop_nb_kernel131_x86_64_sse.nb131_odd_inner: addl $4,nb131_innerk(%rsp) jnz _nb_kernel131_x86_64_sse.nb131_odd_loop jmp _nb_kernel131_x86_64_sse.nb131_updateouterdata_nb_kernel131_x86_64_sse.nb131_odd_loop: movq nb131_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb131_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb131_iqO(%rsp),%xmm4 movq nb131_charge(%rbp),%rsi movhps nb131_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb131_qqO(%rsp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movq nb131_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb131_vdwparam(%rbp),%rsi shll %ebx addl nb131_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb131_c6(%rsp) movaps %xmm7,nb131_c12(%rsp) movq nb131_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm3 movss 4(%rsi,%rax,4),%xmm4 movss 8(%rsi,%rax,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movss nb131_ixO(%rsp),%xmm0 movss nb131_iyO(%rsp),%xmm1 movss nb131_izO(%rsp),%xmm2 movlps nb131_ixH1(%rsp),%xmm6 movlps nb131_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm0 movlps nb131_iyH1(%rsp),%xmm6 movlps nb131_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm1 movlps nb131_izH1(%rsp),%xmm6 movlps nb131_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb131_dxO(%rsp) movaps %xmm4,nb131_dyO(%rsp) movaps %xmm5,nb131_dzO(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb131_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb131_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## 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 $224,%xmm0,%xmm0 ## constant 11100000 ## rsq still in xmm4, rinv in xmm0. mulps %xmm0,%xmm4 mulps nb131_tsc(%rsp),%xmm4 ## rtab cvttps2pi %xmm4,%mm6 cvtpi2ps %mm6,%xmm6 subss %xmm6,%xmm4 movss %xmm4,%xmm1 ## xmm1=eps movss %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=eps2 pslld $3,%mm6 movq nb131_VFtab(%rbp),%rsi movd %mm6,%r12d ## dispersion movlps (%rsi,%r12,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 8(%rsi,%r12,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb131_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb131_c6(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijD mulss %xmm4,%xmm5 ## Vvdw6 mulss nb131_tsc(%rsp),%xmm7 ## put scalar force on stack Update Vvdwtot directly addss nb131_Vvdwtot(%rsp),%xmm5 movss %xmm7,nb131_fstmp(%rsp) movss %xmm5,nb131_Vvdwtot(%rsp) ## repulsion movlps 16(%rsi,%r12,4),%xmm5 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 24(%rsi,%r12,4),%xmm7 movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## table ready, in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb131_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movss nb131_c12(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijR mulss %xmm4,%xmm5 ## Vvdw12 mulss nb131_tsc(%rsp),%xmm7 addss nb131_fstmp(%rsp),%xmm7 movss %xmm7,nb131_fstmp(%rsp) addss nb131_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb131_Vvdwtot(%rsp) ## coulomb in analytical form. xmm0=rinv movaps %xmm0,%xmm2 ## rinv mulps nb131_qqO(%rsp),%xmm2 ## vcoulO movaps %xmm2,%xmm3 ## vcoulO addps nb131_vctot(%rsp),%xmm2 movaps %xmm2,nb131_vctot(%rsp) mulps %xmm0,%xmm3 subss nb131_fstmp(%rsp),%xmm3 mulps %xmm0,%xmm3 movaps nb131_dxO(%rsp),%xmm0 movaps nb131_dyO(%rsp),%xmm1 movaps nb131_dzO(%rsp),%xmm2 mulps %xmm3,%xmm0 mulps %xmm3,%xmm1 mulps %xmm3,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb131_fixO(%rsp),%xmm3 movss nb131_fiyO(%rsp),%xmm4 movss nb131_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb131_fixO(%rsp) movss %xmm4,nb131_fiyO(%rsp) movss %xmm5,nb131_fizO(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## constant 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## constant 11100110 shufps $230,%xmm5,%xmm5 ## constant 11100110 addss nb131_fixH1(%rsp),%xmm3 addss nb131_fiyH1(%rsp),%xmm4 addss nb131_fizH1(%rsp),%xmm5 movss %xmm3,nb131_fixH1(%rsp) movss %xmm4,nb131_fiyH1(%rsp) movss %xmm5,nb131_fizH1(%rsp) ## updated the H1 force movq nb131_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## constant 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## constant 11100111 shufps $231,%xmm5,%xmm5 ## constant 11100111 addss nb131_fixH2(%rsp),%xmm3 addss nb131_fiyH2(%rsp),%xmm4 addss nb131_fizH2(%rsp),%xmm5 movss %xmm3,nb131_fixH2(%rsp) movss %xmm4,nb131_fiyH2(%rsp) movss %xmm5,nb131_fizH2(%rsp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb131_innerk(%rsp) jz _nb_kernel131_x86_64_sse.nb131_updateouterdata jmp _nb_kernel131_x86_64_sse.nb131_odd_loop_nb_kernel131_x86_64_sse.nb131_updateouterdata: movl nb131_ii3(%rsp),%ecx movq nb131_faction(%rbp),%rdi movq nb131_fshift(%rbp),%rsi movl nb131_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb131_fixO(%rsp),%xmm0 movaps nb131_fiyO(%rsp),%xmm1 movaps nb131_fizO(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3⌨️ 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?