nb_kernel332_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,234 行 · 第 1/5 页
S
2,234 行
movaps %xmm9,%xmm11 mulps %xmm0,%xmm1 ## rsq*lu*lu mulps %xmm3,%xmm4 ## rsq*lu*lu mulps %xmm6,%xmm7 ## rsq*lu*lu subps %xmm1,%xmm9 subps %xmm4,%xmm10 subps %xmm7,%xmm11 ## 3-rsq*lu*lu mulps %xmm2,%xmm9 mulps %xmm5,%xmm10 mulps %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movaps nb332_half(%rsp),%xmm4 mulps %xmm4,%xmm9 ## rinvOH2 mulps %xmm4,%xmm10 ## rinvH1H2 mulps %xmm4,%xmm11 ## rinvH2H2 movaps %xmm9,nb332_rinvOH2(%rsp) movaps %xmm10,nb332_rinvH1H2(%rsp) movaps %xmm11,nb332_rinvH2H2(%rsp) ## H2 interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movaps nb332_tsc(%rsp),%xmm1 mulps %xmm9,%xmm0 ## r mulps %xmm10,%xmm3 mulps %xmm11,%xmm6 mulps %xmm1,%xmm0 ## rtab mulps %xmm1,%xmm3 mulps %xmm1,%xmm6 ## truncate and convert to integers cvttps2dq %xmm0,%xmm1 cvttps2dq %xmm3,%xmm4 cvttps2dq %xmm6,%xmm7 ## convert back to float cvtdq2ps %xmm1,%xmm2 cvtdq2ps %xmm4,%xmm5 cvtdq2ps %xmm7,%xmm8 ## multiply by 4 pslld $2,%xmm1 pslld $2,%xmm4 pslld $2,%xmm7 ## multiply by three (copy, mult. by two, add back) movaps %xmm1,%xmm10 movaps %xmm4,%xmm11 movaps %xmm7,%xmm12 pslld $1,%xmm1 pslld $1,%xmm4 pslld $1,%xmm7 paddd %xmm10,%xmm1 paddd %xmm11,%xmm4 paddd %xmm12,%xmm7 ## move to integer registers movhlps %xmm1,%xmm13 movhlps %xmm4,%xmm14 movhlps %xmm7,%xmm15 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 nb332_VFtab(%rbp),%rsi ## calculate eps subps %xmm2,%xmm0 subps %xmm5,%xmm3 subps %xmm8,%xmm6 movaps %xmm0,nb332_epsO(%rsp) movaps %xmm3,nb332_epsH1(%rsp) movaps %xmm6,nb332_epsH2(%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 nb332_epsO(%rsp),%xmm12 movaps nb332_epsH1(%rsp),%xmm13 movaps nb332_epsH2(%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 nb332_qqOH(%rsp),%xmm12 movaps nb332_qqHH(%rsp),%xmm13 addps %xmm0,%xmm1 ## VV addps %xmm4,%xmm5 addps %xmm8,%xmm9 mulps %xmm12,%xmm1 ## VV*qq = vcoul mulps %xmm13,%xmm5 mulps %xmm13,%xmm9 mulps %xmm12,%xmm3 ## FF*qq = fij mulps %xmm13,%xmm7 mulps %xmm13,%xmm11 ## accumulate vctot addps nb332_vctot(%rsp),%xmm1 addps %xmm9,%xmm5 addps %xmm5,%xmm1 movaps %xmm1,nb332_vctot(%rsp) movaps nb332_tsc(%rsp),%xmm10 mulps %xmm10,%xmm3 ## fscal mulps %xmm10,%xmm7 mulps %xmm11,%xmm10 movd %mm0,%eax ## restore j3 from mm0-mm3 movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx ## move j H2 forces to local temp variables movq nb332_faction(%rbp),%rdi movlps 24(%rdi,%rax,4),%xmm11 ## jxH2a jyH2a - - movlps 24(%rdi,%rcx,4),%xmm12 ## jxH2c jyH2c - - movhps 24(%rdi,%rbx,4),%xmm11 ## jxH2a jyH2a jxH2b jyH2b movhps 24(%rdi,%rdx,4),%xmm12 ## jxH2c jyH2c jxH2d jyH2d movss 32(%rdi,%rax,4),%xmm13 ## jzH2a - - - movss 32(%rdi,%rcx,4),%xmm14 ## jzH2c - - - movss 32(%rdi,%rbx,4),%xmm1 ## jzH2b - - - movss 32(%rdi,%rdx,4),%xmm2 ## jzH2d - - - movlhps %xmm1,%xmm13 ## jzH2a - jzH2b - movlhps %xmm2,%xmm14 ## jzH2c - jzH2d - shufps $136,%xmm14,%xmm13 ## 10001000 => jzH2a jzH2b jzH2c jzH2d ## xmm11: jxH2a jyH2a jxH2b jyH2b ## xmm12: jxH2c jyH2c jxH2d jyH2d ## xmm13: jzH2a jzH2b jzH2c jzH2d xorps %xmm0,%xmm0 xorps %xmm4,%xmm4 xorps %xmm8,%xmm8 mulps nb332_rinvOH2(%rsp),%xmm3 mulps nb332_rinvH1H2(%rsp),%xmm7 mulps nb332_rinvH2H2(%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 nb332_dxOH2(%rsp),%xmm0 mulps nb332_dyOH2(%rsp),%xmm1 mulps nb332_dzOH2(%rsp),%xmm2 mulps nb332_dxH1H2(%rsp),%xmm3 mulps nb332_dyH1H2(%rsp),%xmm4 mulps nb332_dzH1H2(%rsp),%xmm5 mulps nb332_dxH2H2(%rsp),%xmm6 mulps nb332_dyH2H2(%rsp),%xmm7 mulps nb332_dzH2H2(%rsp),%xmm8 movaps %xmm0,%xmm14 movaps %xmm1,%xmm15 addps %xmm2,%xmm13 addps nb332_fixO(%rsp),%xmm0 addps nb332_fiyO(%rsp),%xmm1 addps nb332_fizO(%rsp),%xmm2 addps %xmm3,%xmm14 addps %xmm4,%xmm15 addps %xmm5,%xmm13 addps nb332_fixH1(%rsp),%xmm3 addps nb332_fiyH1(%rsp),%xmm4 addps nb332_fizH1(%rsp),%xmm5 addps %xmm6,%xmm14 addps %xmm7,%xmm15 addps %xmm8,%xmm13 addps nb332_fixH2(%rsp),%xmm6 addps nb332_fiyH2(%rsp),%xmm7 addps nb332_fizH2(%rsp),%xmm8 movaps %xmm0,nb332_fixO(%rsp) movaps %xmm1,nb332_fiyO(%rsp) movaps %xmm2,nb332_fizO(%rsp) movaps %xmm3,nb332_fixH1(%rsp) movaps %xmm4,nb332_fiyH1(%rsp) movaps %xmm5,nb332_fizH1(%rsp) movaps %xmm6,nb332_fixH2(%rsp) movaps %xmm7,nb332_fiyH2(%rsp) movaps %xmm8,nb332_fizH2(%rsp) ## xmm11 = fH2x ## xmm12 = fH2y ## xmm13 = fH2z movaps %xmm14,%xmm0 unpcklps %xmm15,%xmm14 unpckhps %xmm15,%xmm0 addps %xmm14,%xmm11 addps %xmm0,%xmm12 movhlps %xmm13,%xmm14 ## fH2zc fH2zd movlps %xmm11,24(%rdi,%rax,4) movhps %xmm11,24(%rdi,%rbx,4) movlps %xmm12,24(%rdi,%rcx,4) movhps %xmm12,24(%rdi,%rdx,4) movss %xmm13,32(%rdi,%rax,4) movss %xmm14,32(%rdi,%rcx,4) shufps $1,%xmm13,%xmm13 shufps $1,%xmm14,%xmm14 movss %xmm13,32(%rdi,%rbx,4) movss %xmm14,32(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb332_innerk(%rsp) jl _nb_kernel332_x86_64_sse.nb332_single_check jmp _nb_kernel332_x86_64_sse.nb332_unroll_loop_nb_kernel332_x86_64_sse.nb332_single_check: addl $4,nb332_innerk(%rsp) jnz _nb_kernel332_x86_64_sse.nb332_single_loop jmp _nb_kernel332_x86_64_sse.nb332_updateouterdata_nb_kernel332_x86_64_sse.nb332_single_loop: movq nb332_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb332_innerjjnr(%rsp) movq nb332_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## fetch j coordinates xorps %xmm0,%xmm0 xorps %xmm1,%xmm1 xorps %xmm2,%xmm2 movss (%rsi,%rax,4),%xmm0 ## jxO - - - movss 4(%rsi,%rax,4),%xmm1 ## jyO - - - movss 8(%rsi,%rax,4),%xmm2 ## jzO - - - movlps 12(%rsi,%rax,4),%xmm6 ## xmm6 = jxH1 jyH1 - - movss 20(%rsi,%rax,4),%xmm7 ## xmm7 = jzH1 - - - movhps 24(%rsi,%rax,4),%xmm6 ## xmm6 = jxH1 jyH1 jxH2 jyH2 movss 32(%rsi,%rax,4),%xmm5 ## xmm5 = jzH2 - - - ## have all coords, time for some shuffling. shufps $216,%xmm6,%xmm6 ## 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps %xmm5,%xmm7 ## xmm7 = jzH1 jzH2 - - movlhps %xmm6,%xmm0 ## xmm0 = jxO 0 jxH1 jxH2 shufps $228,%xmm6,%xmm1 ## 11100100 ;# xmm1 = jyO 0 jyH1 jyH2 shufps $68,%xmm7,%xmm2 ## 01000100 ;# xmm2 = jzO 0 jzH1 jzH2 ## store all j coordinates in jO movaps %xmm0,nb332_jxO(%rsp) movaps %xmm1,nb332_jyO(%rsp) movaps %xmm2,nb332_jzO(%rsp) subps nb332_ixO(%rsp),%xmm0 subps nb332_iyO(%rsp),%xmm1 subps nb332_izO(%rsp),%xmm2 movaps %xmm0,nb332_dxOO(%rsp) movaps %xmm1,nb332_dyOO(%rsp) movaps %xmm2,nb332_dzOO(%rsp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 ## have rsq in xmm0 ## do invsqrt rsqrtps %xmm0,%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb332_three(%rsp),%xmm3 mulps %xmm0,%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb332_half(%rsp),%xmm3 ## rinv iO - j water movaps %xmm3,%xmm1 mulps %xmm0,%xmm1 ## xmm1=r movaps %xmm3,%xmm0 ## xmm0=rinv mulps nb332_tsc(%rsp),%xmm1 movhlps %xmm1,%xmm2 cvttps2pi %xmm1,%mm6 cvttps2pi %xmm2,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm2 movlhps %xmm2,%xmm3 subps %xmm3,%xmm1 ## xmm1=eps movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=eps2 pslld $2,%mm6 pslld $2,%mm7 movd %mm6,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx movq nb332_VFtab(%rbp),%rsi lea (%rbx,%rbx,2),%rbx lea (%rcx,%rcx,2),%rcx lea (%rdx,%rdx,2),%rdx movlps (%rsi,%rbx,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,%rbx,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 nb332_two(%rsp),%xmm7 ## two*Heps2 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb332_qqOO(%rsp),%xmm3 movhps nb332_qqOH(%rsp),%xmm3 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm3,%xmm5 ## vcoul=qq*VV mulps %xmm7,%xmm3 ## fijC=FF*qq ## at this point xmm5 contains vcoul and xmm3 fijC addps nb332_vctot(%rsp),%xmm5 movaps %xmm5,nb332_vctot(%rsp) ## put scalar force on stack temporarily movaps %xmm3,nb332_fstmp(%rsp) ## dispersion movss 16(%rsi,%rbx,4),%xmm4 movss 20(%rsi,%rbx,4),%xmm5 movss 24(%rsi,%rbx,4),%xmm6⌨️ 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?