nb_kernel313_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,088 行 · 第 1/5 页
S
2,088 行
addps %xmm7,%xmm5 ## xmm5=Fp mulps nb313_two(%esp),%xmm7 ## two*Heps2 movaps nb313_qqM(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm0,%xmm7 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm0 fijC ## increment vcoul xorps %xmm4,%xmm4 addps nb313_vctot(%esp),%xmm5 mulps nb313_rinvM(%esp),%xmm7 movaps %xmm5,nb313_vctot(%esp) mulps nb313_tsc(%esp),%xmm7 subps %xmm7,%xmm4 movaps nb313_dxM(%esp),%xmm0 movaps nb313_dyM(%esp),%xmm1 movaps nb313_dzM(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx ## update M forces movaps nb313_fixM(%esp),%xmm3 movaps nb313_fiyM(%esp),%xmm4 movaps nb313_fizM(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb313_fixM(%esp) movaps %xmm4,nb313_fiyM(%esp) movaps %xmm7,nb313_fizM(%esp) movl nb313_faction(%ebp),%edi ## update j forces from stored values addps nb313_fjx(%esp),%xmm0 addps nb313_fjy(%esp),%xmm1 addps nb313_fjz(%esp),%xmm2 movlps (%edi,%eax,4),%xmm4 movlps (%edi,%ecx,4),%xmm7 movhps (%edi,%ebx,4),%xmm4 movhps (%edi,%edx,4),%xmm7 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,%xmm3 shufps $136,%xmm7,%xmm3 ## constant 10001000 shufps $221,%xmm7,%xmm4 ## constant 11011101 ## xmm3 has fjx, xmm4 has fjy subps %xmm0,%xmm3 subps %xmm1,%xmm4 ## unpack them back for storing movaps %xmm3,%xmm7 unpcklps %xmm4,%xmm7 unpckhps %xmm4,%xmm3 movlps %xmm7,(%edi,%eax,4) movlps %xmm3,(%edi,%ecx,4) movhps %xmm7,(%edi,%ebx,4) movhps %xmm3,(%edi,%edx,4) ## finally z forces movss 8(%edi,%eax,4),%xmm0 movss 8(%edi,%ebx,4),%xmm1 movss 8(%edi,%ecx,4),%xmm3 movss 8(%edi,%edx,4),%xmm4 subss %xmm2,%xmm0 shufps $229,%xmm2,%xmm2 ## constant 11100101 subss %xmm2,%xmm1 shufps $234,%xmm2,%xmm2 ## constant 11101010 subss %xmm2,%xmm3 shufps $255,%xmm2,%xmm2 ## constant 11111111 subss %xmm2,%xmm4 movss %xmm0,8(%edi,%eax,4) movss %xmm1,8(%edi,%ebx,4) movss %xmm3,8(%edi,%ecx,4) movss %xmm4,8(%edi,%edx,4) ## should we do one more iteration? subl $4,nb313_innerk(%esp) jl _nb_kernel313_ia32_sse.nb313_odd_inner jmp _nb_kernel313_ia32_sse.nb313_unroll_loop_nb_kernel313_ia32_sse.nb313_odd_inner: addl $4,nb313_innerk(%esp) jnz _nb_kernel313_ia32_sse.nb313_odd_loop jmp _nb_kernel313_ia32_sse.nb313_updateouterdata_nb_kernel313_ia32_sse.nb313_odd_loop: movl nb313_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb313_innerjjnr(%esp) xorps %xmm4,%xmm4 ## clear reg. movss nb313_iqM(%esp),%xmm4 movl nb313_charge(%ebp),%esi movhps nb313_iqH(%esp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%esi,%eax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb313_qqM(%esp) ## use dummy qq for storage xorps %xmm6,%xmm6 movl nb313_type(%ebp),%esi movl (%esi,%eax,4),%ebx movl nb313_vdwparam(%ebp),%esi shll %ebx addl nb313_ntia(%esp),%ebx movlps (%esi,%ebx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## constant 11111100 shufps $253,%xmm7,%xmm7 ## constant 11111101 movaps %xmm6,nb313_c6(%esp) movaps %xmm7,nb313_c12(%esp) movl nb313_pos(%ebp),%esi leal (%eax,%eax,2),%eax movss nb313_ixO(%esp),%xmm3 movss nb313_iyO(%esp),%xmm4 movss nb313_izO(%esp),%xmm5 movss nb313_ixH1(%esp),%xmm0 movss nb313_iyH1(%esp),%xmm1 movss nb313_izH1(%esp),%xmm2 unpcklps nb313_ixH2(%esp),%xmm3 ## ixO ixH2 - - unpcklps nb313_iyH2(%esp),%xmm4 ## iyO iyH2 - - unpcklps nb313_izH2(%esp),%xmm5 ## izO izH2 - - unpcklps nb313_ixM(%esp),%xmm0 ## ixH1 ixM - - unpcklps nb313_iyM(%esp),%xmm1 ## iyH1 iyM - - unpcklps nb313_izM(%esp),%xmm2 ## izH1 izM - - unpcklps %xmm0,%xmm3 ## ixO ixH1 ixH2 ixM unpcklps %xmm1,%xmm4 ## same for y unpcklps %xmm2,%xmm5 ## same for z ## move j coords to xmm0-xmm2 movss (%esi,%eax,4),%xmm0 movss 4(%esi,%eax,4),%xmm1 movss 8(%esi,%eax,4),%xmm2 shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use O distances for storage movaps %xmm3,nb313_dxO(%esp) movaps %xmm4,nb313_dyO(%esp) movaps %xmm5,nb313_dzO(%esp) 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 nb313_three(%esp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb313_half(%esp),%xmm0 subps %xmm5,%xmm1 ## constant 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,nb313_rinvM(%esp) mulps %xmm0,%xmm4 ## r mulps nb313_tsc(%esp),%xmm4 movhlps %xmm4,%xmm7 cvttps2pi %xmm4,%mm6 cvttps2pi %xmm7,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm7 movlhps %xmm7,%xmm3 subps %xmm3,%xmm4 movaps %xmm4,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movl nb313_VFtab(%ebp),%esi psrlq $32,%mm6 movd %mm6,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx xorps %xmm5,%xmm5 movlps (%esi,%ecx,4),%xmm3 ## values= Y3 F3 - - movhps (%esi,%ebx,4),%xmm5 ## values= 0 0 Y2 F2 movhps (%esi,%edx,4),%xmm3 ## values= Y3 F3 Y4 F4 movaps %xmm5,%xmm4 ## values= 0 0 Y2 F2 shufps $0x88,%xmm3,%xmm4 ## values= 0 Y2 Y3 Y3 shufps $0xDD,%xmm3,%xmm5 ## values= 0 F2 F3 F4 xorps %xmm7,%xmm7 movlps 8(%esi,%ecx,4),%xmm3 ## values= G3 H3 - - movhps 8(%esi,%ebx,4),%xmm7 ## values= 0 0 G2 H2 movhps 8(%esi,%edx,4),%xmm3 ## values= G3 H3 G4 H4 movaps %xmm7,%xmm6 ## values= 0 0 G2 H2 shufps $0x88,%xmm3,%xmm6 ## values= 0 G2 G3 G3 shufps $0xDD,%xmm3,%xmm7 ## values= 0 H2 H3 H4 ## xmm4 = 0 Y2 Y3 Y4 ## xmm5 = 0 F2 F3 F4 ## xmm6 = 0 G2 G3 G4 ## xmm7 = 0 H2 H3 H4 ## coulomb table ready, in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb313_two(%esp),%xmm7 ## two*Heps2 movaps nb313_qqM(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm0 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm0 fijC ## increment vcoul - then we can get rid of mm5 addps nb313_vctot(%esp),%xmm5 movaps %xmm5,nb313_vctot(%esp) ## do nontable L-J in first element only. movaps nb313_rinvM(%esp),%xmm2 mulss %xmm2,%xmm2 movaps %xmm2,%xmm1 mulss %xmm1,%xmm1 mulss %xmm2,%xmm1 ## xmm1=rinvsix xorps %xmm4,%xmm4 movss %xmm1,%xmm4 mulss %xmm4,%xmm4 ## xmm4=rinvtwelve mulss nb313_c6(%esp),%xmm1 mulss nb313_c12(%esp),%xmm4 movaps %xmm4,%xmm3 subss %xmm1,%xmm3 ## xmm3=Vvdw12-Vvdw6 mulss nb313_six(%esp),%xmm1 mulss nb313_twelve(%esp),%xmm4 subss %xmm1,%xmm4 addss nb313_Vvdwtot(%esp),%xmm3 mulss nb313_rinvM(%esp),%xmm4 ## add back coul stuff from memory, and work on all elements again mulps nb313_tsc(%esp),%xmm0 subps %xmm0,%xmm4 movss %xmm3,nb313_Vvdwtot(%esp) mulps nb313_rinvM(%esp),%xmm4 movaps nb313_dxO(%esp),%xmm0 movaps nb313_dyO(%esp),%xmm1 movaps nb313_dzO(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb313_fixO(%esp),%xmm3 movss nb313_fiyO(%esp),%xmm4 movss nb313_fizO(%esp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb313_fixO(%esp) movss %xmm4,nb313_fiyO(%esp) movss %xmm5,nb313_fizO(%esp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb313_fixH1(%esp),%xmm3 addss nb313_fiyH1(%esp),%xmm4 addss nb313_fizH1(%esp),%xmm5 movss %xmm3,nb313_fixH1(%esp) movss %xmm4,nb313_fiyH1(%esp) movss %xmm5,nb313_fizH1(%esp) ## updated the H1 force shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb313_fixH2(%esp),%xmm3 addss nb313_fiyH2(%esp),%xmm4 addss nb313_fizH2(%esp),%xmm5 movss %xmm3,nb313_fixH2(%esp) movss %xmm4,nb313_fiyH2(%esp) movss %xmm5,nb313_fizH2(%esp) ## updated the H2 force movl nb313_faction(%ebp),%edi shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb313_fixM(%esp),%xmm3 addss nb313_fiyM(%esp),%xmm4 addss nb313_fizM(%esp),%xmm5 movss %xmm3,nb313_fixM(%esp) movss %xmm4,nb313_fiyM(%esp) movss %xmm5,nb313_fizM(%esp) ## updated the M force ## the fj's - move in from mem start by acc. tx/ty/tz in xmm0, xmm1 movlps (%edi,%eax,4),%xmm6 movss 8(%edi,%eax,4),%xmm7 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm5 movaps %xmm3,%xmm0 movaps %xmm4,%xmm1 movaps %xmm5,%xmm2 shufps $0x39,%xmm3,%xmm3 ## shift right shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 unpcklps %xmm1,%xmm0 ## x,y sum in xmm0, z sum in xmm2 subps %xmm0,%xmm6 subss %xmm2,%xmm7 movlps %xmm6,(%edi,%eax,4) movss %xmm7,8(%edi,%eax,4) decl nb313_innerk(%esp) jz _nb_kernel313_ia32_sse.nb313_updateouterdata jmp _nb_kernel313_ia32_sse.nb313_odd_loop_nb_kernel313_ia32_sse.nb313_updateouterdata: movl nb313_ii3(%esp),%ecx movl nb313_faction(%ebp),%edi movl nb313_fshift(%ebp),%esi movl nb313_is3(%esp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb313_fixO(%esp),%xmm0 movaps nb313_fiyO(%esp),%xmm1 movaps nb313_fizO(%esp),%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 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss (%edi,%ecx,4),%xmm3 movss 4(%edi,%ecx,4),%xmm4 movss 8(%edi,%ecx,4),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,(%edi,%ecx,4) movss %xmm4,4(%edi,%ecx,4) movss %xmm5,8(%edi,%ecx,4) ## accumulate force in xmm6/xmm7 for fshift movaps %xmm0,%xmm6 movss %xmm2,%xmm7 movlhps %xmm1,%xmm6 shufps $8,%xmm6,%xmm6 ## constant 00001000 ## accumulate H1i forces in xmm0, xmm1, xmm2 movaps nb313_fixH1(%esp),%xmm0 movaps nb313_fiyH1(%esp),%xmm1 movaps nb313_fizH1(%esp),%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 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5⌨️ 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?