nb_kernel333_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,182 行 · 第 1/5 页
S
2,182 行
## move j forces to local temp variables movq nb333_faction(%rbp),%rdi movlps (%rdi,%rax,4),%xmm11 ## jxa jya - - movlps (%rdi,%rcx,4),%xmm12 ## jxc jyc - - movhps (%rdi,%rbx,4),%xmm11 ## jxa jya jxb jyb movhps (%rdi,%rdx,4),%xmm12 ## jxc jyc jxd jyd movss 8(%rdi,%rax,4),%xmm13 ## jza - - - movss 8(%rdi,%rcx,4),%xmm14 ## jzc - - - movss 8(%rdi,%rbx,4),%xmm2 ## jzb movss 8(%rdi,%rdx,4),%xmm5 ## jzd movlhps %xmm2,%xmm13 ## jza - jzb - movlhps %xmm5,%xmm14 ## jzc - jzd - shufps $136,%xmm14,%xmm13 ## 10001000 => jza jzb jzc jzd ## xmm11: jxa jya jxb jyb ## xmm12: jxc jyc jxd jyd ## xmm13: jza jzb jzc jzd xorps %xmm0,%xmm0 xorps %xmm4,%xmm4 xorps %xmm8,%xmm8 mulps nb333_rinvH1(%rsp),%xmm3 mulps nb333_rinvH2(%rsp),%xmm7 mulps nb333_rinvM(%rsp),%xmm10 subps %xmm3,%xmm0 subps %xmm7,%xmm4 subps %xmm10,%xmm8 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps %xmm4,%xmm3 movaps %xmm4,%xmm5 movaps %xmm8,%xmm6 movaps %xmm8,%xmm7 mulps nb333_dxH1(%rsp),%xmm0 mulps nb333_dyH1(%rsp),%xmm1 mulps nb333_dzH1(%rsp),%xmm2 mulps nb333_dxH2(%rsp),%xmm3 mulps nb333_dyH2(%rsp),%xmm4 mulps nb333_dzH2(%rsp),%xmm5 mulps nb333_dxM(%rsp),%xmm6 mulps nb333_dyM(%rsp),%xmm7 mulps nb333_dzM(%rsp),%xmm8 ## fetch forces from O interaction movaps nb333_fjx(%rsp),%xmm14 movaps nb333_fjy(%rsp),%xmm15 addps nb333_fjz(%rsp),%xmm13 addps %xmm0,%xmm14 addps %xmm1,%xmm15 addps %xmm2,%xmm13 addps nb333_fixH1(%rsp),%xmm0 addps nb333_fiyH1(%rsp),%xmm1 addps nb333_fizH1(%rsp),%xmm2 addps %xmm3,%xmm14 addps %xmm4,%xmm15 addps %xmm5,%xmm13 addps nb333_fixH2(%rsp),%xmm3 addps nb333_fiyH2(%rsp),%xmm4 addps nb333_fizH2(%rsp),%xmm5 addps %xmm6,%xmm14 addps %xmm7,%xmm15 addps %xmm8,%xmm13 addps nb333_fixM(%rsp),%xmm6 addps nb333_fiyM(%rsp),%xmm7 addps nb333_fizM(%rsp),%xmm8 movaps %xmm0,nb333_fixH1(%rsp) movaps %xmm1,nb333_fiyH1(%rsp) movaps %xmm2,nb333_fizH1(%rsp) movaps %xmm3,nb333_fixH2(%rsp) movaps %xmm4,nb333_fiyH2(%rsp) movaps %xmm5,nb333_fizH2(%rsp) movaps %xmm6,nb333_fixM(%rsp) movaps %xmm7,nb333_fiyM(%rsp) movaps %xmm8,nb333_fizM(%rsp) ## xmm14 = fjx ## xmm15 = fjy ## xmm13 = fjz movaps %xmm14,%xmm0 unpcklps %xmm15,%xmm14 unpckhps %xmm15,%xmm0 addps %xmm14,%xmm11 addps %xmm0,%xmm12 movhlps %xmm13,%xmm14 ## fjzc fjzd movlps %xmm11,(%rdi,%rax,4) movhps %xmm11,(%rdi,%rbx,4) movlps %xmm12,(%rdi,%rcx,4) movhps %xmm12,(%rdi,%rdx,4) movss %xmm13,8(%rdi,%rax,4) movss %xmm14,8(%rdi,%rcx,4) shufps $1,%xmm13,%xmm13 shufps $1,%xmm14,%xmm14 movss %xmm13,8(%rdi,%rbx,4) movss %xmm14,8(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb333_innerk(%rsp) jl _nb_kernel333_x86_64_sse.nb333_odd_inner jmp _nb_kernel333_x86_64_sse.nb333_unroll_loop_nb_kernel333_x86_64_sse.nb333_odd_inner: addl $4,nb333_innerk(%rsp) jnz _nb_kernel333_x86_64_sse.nb333_odd_loop jmp _nb_kernel333_x86_64_sse.nb333_updateouterdata_nb_kernel333_x86_64_sse.nb333_odd_loop: movq nb333_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb333_innerjjnr(%rsp) xorps %xmm4,%xmm4 ## clear reg. movss nb333_iqM(%rsp),%xmm4 movq nb333_charge(%rbp),%rsi movhps nb333_iqH(%rsp),%xmm4 ## [qM 0 qH qH] shufps $41,%xmm4,%xmm4 ## [0 qH qH qM] movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb333_qqM(%rsp) ## use dummy qq for storage xorps %xmm6,%xmm6 movq nb333_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb333_vdwparam(%rbp),%rsi shll %ebx addl nb333_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb333_c6(%rsp) movaps %xmm7,nb333_c12(%rsp) movq nb333_pos(%rbp),%rsi lea (%rax,%rax,2),%rax movss nb333_ixO(%rsp),%xmm0 movss nb333_iyO(%rsp),%xmm1 movss nb333_izO(%rsp),%xmm2 movss nb333_ixH1(%rsp),%xmm3 movss nb333_iyH1(%rsp),%xmm4 movss nb333_izH1(%rsp),%xmm5 unpcklps nb333_ixH2(%rsp),%xmm0 ## ixO ixH2 - - unpcklps nb333_iyH2(%rsp),%xmm1 ## iyO iyH2 - - unpcklps nb333_izH2(%rsp),%xmm2 ## izO izH2 - - unpcklps nb333_ixM(%rsp),%xmm3 ## ixH1 ixM - - unpcklps nb333_iyM(%rsp),%xmm4 ## iyH1 iyM - - unpcklps nb333_izM(%rsp),%xmm5 ## 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 (%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 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use O distances for storage movaps %xmm3,nb333_dxO(%rsp) movaps %xmm4,nb333_dyO(%rsp) movaps %xmm5,nb333_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 nb333_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb333_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv movaps %xmm0,nb333_rinvM(%rsp) mulps %xmm0,%xmm4 ## r mulps nb333_tsc(%rsp),%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 movd %eax,%mm0 movq nb333_VFtab(%rbp),%rsi movd %mm6,%eax psrlq $32,%mm6 movd %mm6,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx lea (%rax,%rax,2),%rax lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx ## first do LJ table for O ## load dispersion table data into xmm4 movlps 16(%rsi,%rax,4),%xmm4 movlps 24(%rsi,%rax,4),%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $0x1,%xmm5,%xmm5 shufps $0x1,%xmm7,%xmm7 ## dispersion table YFGH ready in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb333_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movaps nb333_c6(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijD mulss %xmm4,%xmm5 ## Vvdw6 ## save scalar force in xmm3. Update Vvdwtot directly addss nb333_Vvdwtot(%rsp),%xmm5 xorps %xmm3,%xmm3 movss %xmm7,%xmm3 ## fscal movss %xmm5,nb333_Vvdwtot(%rsp) ## load repulsion table data into xmm4 movlps 32(%rsi,%rax,4),%xmm4 movlps 40(%rsi,%rax,4),%xmm6 movaps %xmm4,%xmm5 movaps %xmm6,%xmm7 shufps $0x1,%xmm5,%xmm5 shufps $0x1,%xmm7,%xmm7 ## repulsion table YFGH ready in xmm4-xmm7 mulss %xmm1,%xmm6 ## xmm6=Geps mulss %xmm2,%xmm7 ## xmm7=Heps2 addss %xmm6,%xmm5 addss %xmm7,%xmm5 ## xmm5=Fp mulss nb333_two(%rsp),%xmm7 ## two*Heps2 addss %xmm6,%xmm7 addss %xmm5,%xmm7 ## xmm7=FF mulss %xmm1,%xmm5 ## xmm5=eps*Fp addss %xmm4,%xmm5 ## xmm5=VV movaps nb333_c12(%rsp),%xmm4 mulss %xmm4,%xmm7 ## fijR mulss %xmm4,%xmm5 ## Vvdw12 addss %xmm7,%xmm3 addss nb333_Vvdwtot(%rsp),%xmm5 movss %xmm5,nb333_Vvdwtot(%rsp) movaps %xmm3,nb333_rinvO(%rsp) ## save fscal temp. in rinvO ## do the Coulomb interaction for H1,H2,M xorps %xmm5,%xmm5 movlps (%rsi,%rcx,4),%xmm3 ## data: Y3 F3 - - movhps (%rsi,%rbx,4),%xmm5 ## data: 0 0 Y2 F2 movhps (%rsi,%rdx,4),%xmm3 ## data: Y3 F3 Y4 F4 movaps %xmm5,%xmm4 ## data: 0 0 Y2 F2 shufps $0x88,%xmm3,%xmm4 ## data: 0 Y2 Y3 Y3 shufps $0xDD,%xmm3,%xmm5 ## data: 0 F2 F3 F4 xorps %xmm7,%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 ## data: G3 H3 - - movhps 8(%rsi,%rbx,4),%xmm7 ## data: 0 0 G2 H2 movhps 8(%rsi,%rdx,4),%xmm3 ## data: G3 H3 G4 H4 movaps %xmm7,%xmm6 ## data: 0 0 G2 H2 shufps $0x88,%xmm3,%xmm6 ## data: 0 G2 G3 G3 shufps $0xDD,%xmm3,%xmm7 ## data: 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 nb333_two(%rsp),%xmm7 ## two*Heps2 movaps nb333_qqM(%rsp),%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 nb333_vctot(%rsp),%xmm5 movaps %xmm5,nb333_vctot(%rsp) addps nb333_rinvO(%rsp),%xmm0 ## total fscal (temp. storage in rinvO) xorps %xmm4,%xmm4 mulps nb333_rinvM(%rsp),%xmm0 mulps nb333_tsc(%rsp),%xmm0 subps %xmm0,%xmm4 movaps nb333_dxO(%rsp),%xmm0 movaps nb333_dyO(%rsp),%xmm1 movaps nb333_dzO(%rsp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb333_fixO(%rsp),%xmm3 movss nb333_fiyO(%rsp),%xmm4 movss nb333_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb333_fixO(%rsp) movss %xmm4,nb333_fiyO(%rsp) movss %xmm5,nb333_fizO(%rsp) ## 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 nb333_fixH1(%rsp),%xmm3 addss nb333_fiyH1(%rsp),%xmm4 addss nb333_fizH1(%rsp),%xmm5 movss %xmm3,nb333_fixH1(%rsp) movss %xmm4,nb333_fiyH1(%rsp) movss %xmm5,nb333_fizH1(%rsp) ## updated the H1 force shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb333_fixH2(%rsp),%xmm3 addss nb333_fiyH2(%rsp),%xmm4 addss nb333_fizH2(%rsp),%xmm5 movss %xmm3,nb333_fixH2(%rsp) movss %xmm4,nb333_fiyH2(%rsp) movss %xmm5,nb333_fizH2(%rsp) ## updated the H2 force movq nb333_faction(%rbp),%rdi shufps $0x39,%xmm3,%xmm3 shufps $0x39,%xmm4,%xmm4 shufps $0x39,%xmm5,%xmm5 addss nb333_fixM(%rsp),%xmm3 addss nb333_fiyM(%rsp),%xmm4 addss nb333_fizM(%rsp),%xmm5 movss %xmm3,nb333_fixM(%rsp) movss %xmm4,nb333_fiyM(%rsp) movss %xmm5,nb333_fizM(%rsp) ## updated the M force movd %mm0,%eax ## the fj's - move in from mem start by acc. tx/ty/tz in xmm0, xmm1 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,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 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb333_innerk(%rsp) jz _nb_kernel333_x86_64_sse.nb333_updateouterdata jmp _nb_kernel333_x86_64_sse.nb333_odd_loop_nb_kernel333_x86_64_sse.nb333_updateouterdata: movl nb333_ii3(%rsp),%ecx movq nb333_faction(%rbp),%rdi movq nb333_fshift(%rbp),%rsi movl nb333_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb333_fixO(%rsp),%xmm0 movaps nb333_fiyO(%rsp),%xmm1 movaps nb333_fizO(%rsp),%xmm2 movhlps %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?