nb_kernel303_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,919 行 · 第 1/4 页
S
1,919 行
movd %xmm1,%eax movd %xmm4,%r8d movd %xmm7,%r12d movd %xmm13,%ecx movd %xmm14,%r10d movd %xmm15,%r14d pshufd $1,%xmm1,%xmm1 pshufd $1,%xmm4,%xmm4 pshufd $1,%xmm7,%xmm7 pshufd $1,%xmm13,%xmm13 pshufd $1,%xmm14,%xmm14 pshufd $1,%xmm15,%xmm15 movd %xmm1,%ebx movd %xmm4,%r9d movd %xmm7,%r13d movd %xmm13,%edx movd %xmm14,%r11d movd %xmm15,%r15d movq nb303_VFtab(%rbp),%rsi ## calculate eps subps %xmm2,%xmm0 subps %xmm5,%xmm3 subps %xmm8,%xmm6 movaps %xmm0,nb303_epsH1(%rsp) movaps %xmm3,nb303_epsH2(%rsp) movaps %xmm6,nb303_epsM(%rsp) ## Load LOTS of table data movlps (%rsi,%rax,4),%xmm1 movlps (%rsi,%r8,4),%xmm5 movlps (%rsi,%r12,4),%xmm9 movlps (%rsi,%rcx,4),%xmm3 movlps (%rsi,%r10,4),%xmm7 movlps (%rsi,%r14,4),%xmm11 movhps (%rsi,%rbx,4),%xmm1 movhps (%rsi,%r9,4),%xmm5 movhps (%rsi,%r13,4),%xmm9 movhps (%rsi,%rdx,4),%xmm3 movhps (%rsi,%r11,4),%xmm7 movhps (%rsi,%r15,4),%xmm11 movaps %xmm1,%xmm0 movaps %xmm5,%xmm4 movaps %xmm9,%xmm8 shufps $136,%xmm3,%xmm0 ## 10001000 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $136,%xmm11,%xmm8 ## 10001000 shufps $221,%xmm3,%xmm1 ## 11011101 shufps $221,%xmm7,%xmm5 ## 11011101 shufps $221,%xmm11,%xmm9 ## 11011101 movlps 8(%rsi,%rax,4),%xmm3 movlps 8(%rsi,%r8,4),%xmm7 movlps 8(%rsi,%r12,4),%xmm11 movlps 8(%rsi,%rcx,4),%xmm12 movlps 8(%rsi,%r10,4),%xmm13 movlps 8(%rsi,%r14,4),%xmm14 movhps 8(%rsi,%rbx,4),%xmm3 movhps 8(%rsi,%r9,4),%xmm7 movhps 8(%rsi,%r13,4),%xmm11 movhps 8(%rsi,%rdx,4),%xmm12 movhps 8(%rsi,%r11,4),%xmm13 movhps 8(%rsi,%r15,4),%xmm14 movaps %xmm3,%xmm2 movaps %xmm7,%xmm6 movaps %xmm11,%xmm10 shufps $136,%xmm12,%xmm2 ## 10001000 shufps $136,%xmm13,%xmm6 ## 10001000 shufps $136,%xmm14,%xmm10 ## 10001000 shufps $221,%xmm12,%xmm3 ## 11011101 shufps $221,%xmm13,%xmm7 ## 11011101 shufps $221,%xmm14,%xmm11 ## 11011101 ## table data ready in xmm0-xmm3 , xmm4-xmm7 , and xmm8-xmm11 movaps nb303_epsH1(%rsp),%xmm12 movaps nb303_epsH2(%rsp),%xmm13 movaps nb303_epsM(%rsp),%xmm14 mulps %xmm12,%xmm3 ## Heps mulps %xmm13,%xmm7 mulps %xmm14,%xmm11 mulps %xmm12,%xmm2 ## Geps mulps %xmm13,%xmm6 mulps %xmm14,%xmm10 mulps %xmm12,%xmm3 ## Heps2 mulps %xmm13,%xmm7 mulps %xmm14,%xmm11 addps %xmm2,%xmm1 ## F+Geps addps %xmm6,%xmm5 addps %xmm10,%xmm9 addps %xmm3,%xmm1 ## F+Geps+Heps2 = Fp addps %xmm7,%xmm5 addps %xmm11,%xmm9 addps %xmm3,%xmm3 ## 2*Heps2 addps %xmm7,%xmm7 addps %xmm11,%xmm11 addps %xmm2,%xmm3 ## 2*Heps2+Geps addps %xmm6,%xmm7 addps %xmm10,%xmm11 addps %xmm1,%xmm3 ## FF = Fp + 2*Heps2 + Geps addps %xmm5,%xmm7 addps %xmm9,%xmm11 mulps %xmm12,%xmm1 ## eps*Fp mulps %xmm13,%xmm5 mulps %xmm14,%xmm9 movaps nb303_qqH(%rsp),%xmm12 movaps nb303_qqM(%rsp),%xmm13 addps %xmm0,%xmm1 ## VV addps %xmm4,%xmm5 addps %xmm8,%xmm9 mulps %xmm12,%xmm1 ## VV*qq = vcoul mulps %xmm12,%xmm5 mulps %xmm13,%xmm9 mulps %xmm12,%xmm3 ## FF*qq = fij mulps %xmm12,%xmm7 mulps %xmm13,%xmm11 ## accumulate vctot addps nb303_vctot(%rsp),%xmm1 addps %xmm9,%xmm5 addps %xmm5,%xmm1 movaps %xmm1,nb303_vctot(%rsp) movaps nb303_tsc(%rsp),%xmm10 mulps %xmm10,%xmm3 ## fscal mulps %xmm10,%xmm7 mulps %xmm11,%xmm10 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx ## move j forces to local temp variables 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 nb303_rinvH1(%rsp),%xmm3 mulps nb303_rinvH2(%rsp),%xmm7 mulps nb303_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 nb303_dxH1(%rsp),%xmm0 mulps nb303_dyH1(%rsp),%xmm1 mulps nb303_dzH1(%rsp),%xmm2 mulps nb303_dxH2(%rsp),%xmm3 mulps nb303_dyH2(%rsp),%xmm4 mulps nb303_dzH2(%rsp),%xmm5 mulps nb303_dxM(%rsp),%xmm6 mulps nb303_dyM(%rsp),%xmm7 mulps nb303_dzM(%rsp),%xmm8 movaps %xmm0,%xmm14 movaps %xmm1,%xmm15 addps %xmm2,%xmm13 addps nb303_fixH1(%rsp),%xmm0 addps nb303_fiyH1(%rsp),%xmm1 addps nb303_fizH1(%rsp),%xmm2 addps %xmm3,%xmm14 addps %xmm4,%xmm15 addps %xmm5,%xmm13 addps nb303_fixH2(%rsp),%xmm3 addps nb303_fiyH2(%rsp),%xmm4 addps nb303_fizH2(%rsp),%xmm5 addps %xmm6,%xmm14 addps %xmm7,%xmm15 addps %xmm8,%xmm13 addps nb303_fixM(%rsp),%xmm6 addps nb303_fiyM(%rsp),%xmm7 addps nb303_fizM(%rsp),%xmm8 movaps %xmm0,nb303_fixH1(%rsp) movaps %xmm1,nb303_fiyH1(%rsp) movaps %xmm2,nb303_fizH1(%rsp) movaps %xmm3,nb303_fixH2(%rsp) movaps %xmm4,nb303_fiyH2(%rsp) movaps %xmm5,nb303_fizH2(%rsp) movaps %xmm6,nb303_fixM(%rsp) movaps %xmm7,nb303_fiyM(%rsp) movaps %xmm8,nb303_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,nb303_innerk(%rsp) jl _nb_kernel303_x86_64_sse.nb303_odd_inner jmp _nb_kernel303_x86_64_sse.nb303_unroll_loop_nb_kernel303_x86_64_sse.nb303_odd_inner: addl $4,nb303_innerk(%rsp) jnz _nb_kernel303_x86_64_sse.nb303_odd_loop jmp _nb_kernel303_x86_64_sse.nb303_updateouterdata_nb_kernel303_x86_64_sse.nb303_odd_loop: movq nb303_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb303_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb303_iqM(%rsp),%xmm4 movq nb303_charge(%rbp),%rsi movhps nb303_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb303_qqM(%rsp) ## use dummy qq for storage movq nb303_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 nb303_ixM(%rsp),%xmm0 movss nb303_iyM(%rsp),%xmm1 movss nb303_izM(%rsp),%xmm2 movlps nb303_ixH1(%rsp),%xmm6 movlps nb303_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm0 movlps nb303_iyH1(%rsp),%xmm6 movlps nb303_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm1 movlps nb303_izH1(%rsp),%xmm6 movlps nb303_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## use dummy dx for storage movaps %xmm3,nb303_dxM(%rsp) movaps %xmm4,nb303_dyM(%rsp) movaps %xmm5,nb303_dzM(%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 nb303_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb303_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 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 ## 11100000 mulps %xmm0,%xmm4 ## xmm4=r movaps %xmm0,nb303_rinvM(%rsp) mulps nb303_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 movd %ecx,%mm1 movd %edx,%mm2 movq nb303_VFtab(%rbp),%rsi movd %mm6,%eax movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx movlps (%rsi,%rax,4),%xmm5 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%rax,4),%xmm7 movlps 8(%rsi,%rcx,4),%xmm3 movhps 8(%rsi,%rdx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## 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 nb303_two(%rsp),%xmm7 ## two*Heps2 movaps nb303_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 nb303_vctot(%rsp),%xmm5 movaps %xmm5,nb303_vctot(%rsp) xorps %xmm4,%xmm4 mulps nb303_tsc(%rsp),%xmm0 mulps nb303_rinvM(%rsp),%xmm0 subps %xmm0,%xmm4 movd %mm0,%eax movd %mm1,%ecx movd %mm2,%edx movaps nb303_dxM(%rsp),%xmm0 movaps nb303_dyM(%rsp),%xmm1 movaps nb303_dzM(%rsp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 now contains tx-tz (partial force) movss nb303_fixM(%rsp),%xmm3 movss nb303_fiyM(%rsp),%xmm4 movss nb303_fizM(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb303_fixM(%rsp) movss %xmm4,nb303_fiyM(%rsp) movss %xmm5,nb303_fizM(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## 11100110 shufps $230,%xmm5,%xmm5 ## 11100110 addss nb303_fixH1(%rsp),%xmm3 addss nb303_fiyH1(%rsp),%xmm4 addss nb303_fizH1(%rsp),%xmm5 movss %xmm3,nb303_fixH1(%rsp) movss %xmm4,nb303_fiyH1(%rsp) movss %xmm5,nb303_fizH1(%rsp) ## updated the H1 force movq nb303_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## 11100111 shufps $231,%xmm5,%xmm5 ## 11100111 addss nb303_fixH2(%rsp),%xmm3 addss nb303_fiyH2(%rsp),%xmm4 addss nb303_fizH2(%rsp),%xmm5 movss %xmm3,nb303_fixH2(%rsp) movss %xmm4,nb303_fiyH2(%rsp) movss %xmm5,nb303_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 nb303_innerk(%rsp) jz _nb_kernel303_x86_64_sse.nb303_updateouterdata jmp _nb_kernel303_x86_64_sse.nb303_odd_loop_nb_kernel303_x86_64_sse.nb303_updateouterdata: movl nb303_ii3(%rsp),%ecx movq nb303_faction(%rbp),%rdi movq nb303_fshift(%rbp),%rsi movl nb303_is3(%rsp),%edx ## accumulate H1 i forces in xmm0, xmm1, xmm2 movaps nb303_fixH1(%rsp),%xmm0 movaps nb303_fiyH1(%rsp),%xmm1 movaps nb303_fizH1(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1⌨️ 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?