nb_kernel313_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,129 行 · 第 1/5 页
S
2,129 行
## calc 1/rsq cvtpd2ps %xmm3,%xmm6 rcpps %xmm6,%xmm6 cvtps2pd %xmm6,%xmm6 ## lu in low xmm6 ## 1/x lookup seed in xmm6 movapd nb313_two(%rsp),%xmm4 movapd %xmm3,%xmm5 mulpd %xmm6,%xmm3 ## lu*rsq subpd %xmm3,%xmm4 ## 2-lu*rsq mulpd %xmm4,%xmm6 ## (new lu) movapd nb313_two(%rsp),%xmm4 mulpd %xmm6,%xmm5 ## lu*rsq subpd %xmm5,%xmm4 ## 2-lu*rsq mulpd %xmm6,%xmm4 ## xmm4=rinvsq movapd %xmm4,%xmm3 ## rinvsq mulpd %xmm4,%xmm4 ## rinv4 mulpd %xmm3,%xmm4 ## rinv6 movapd %xmm4,%xmm5 mulpd %xmm5,%xmm5 ## rinv12 mulpd nb313_c6(%rsp),%xmm4 mulpd nb313_c12(%rsp),%xmm5 movapd %xmm5,%xmm6 subpd %xmm4,%xmm6 ## Vvdw=vvdw12-vvdw6 mulpd nb313_six(%rsp),%xmm4 mulpd nb313_twelve(%rsp),%xmm5 subpd %xmm4,%xmm5 mulpd %xmm5,%xmm3 ## fscal addpd nb313_Vvdwtot(%rsp),%xmm6 movapd %xmm6,nb313_Vvdwtot(%rsp) mulpd %xmm3,%xmm13 ## fx mulpd %xmm3,%xmm14 ## fy mulpd %xmm3,%xmm15 ## fz ## save j force temporarily movapd %xmm13,nb313_fjx(%rsp) movapd %xmm14,nb313_fjy(%rsp) movapd %xmm15,nb313_fjz(%rsp) ## increment i O force addpd nb313_fixO(%rsp),%xmm13 addpd nb313_fiyO(%rsp),%xmm14 addpd nb313_fizO(%rsp),%xmm15 movapd %xmm13,nb313_fixO(%rsp) movapd %xmm14,nb313_fiyO(%rsp) movapd %xmm15,nb313_fizO(%rsp) ## finished O LJ interaction. ## do H1, H2, and M interactions in parallel. ## xmm0-xmm2 still contain j coordinates. movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subpd nb313_ixH1(%rsp),%xmm0 subpd nb313_iyH1(%rsp),%xmm1 subpd nb313_izH1(%rsp),%xmm2 subpd nb313_ixH2(%rsp),%xmm3 subpd nb313_iyH2(%rsp),%xmm4 subpd nb313_izH2(%rsp),%xmm5 subpd nb313_ixM(%rsp),%xmm6 subpd nb313_iyM(%rsp),%xmm7 subpd nb313_izM(%rsp),%xmm8 movapd %xmm0,nb313_dxH1(%rsp) movapd %xmm1,nb313_dyH1(%rsp) movapd %xmm2,nb313_dzH1(%rsp) mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 movapd %xmm3,nb313_dxH2(%rsp) movapd %xmm4,nb313_dyH2(%rsp) movapd %xmm5,nb313_dzH2(%rsp) mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 movapd %xmm6,nb313_dxM(%rsp) movapd %xmm7,nb313_dyM(%rsp) movapd %xmm8,nb313_dzM(%rsp) mulpd %xmm6,%xmm6 mulpd %xmm7,%xmm7 mulpd %xmm8,%xmm8 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 addpd %xmm7,%xmm6 addpd %xmm8,%xmm6 ## start doing invsqrt for j atoms cvtpd2ps %xmm0,%xmm1 cvtpd2ps %xmm3,%xmm4 cvtpd2ps %xmm6,%xmm7 rsqrtps %xmm1,%xmm1 rsqrtps %xmm4,%xmm4 rsqrtps %xmm7,%xmm7 cvtps2pd %xmm1,%xmm1 cvtps2pd %xmm4,%xmm4 cvtps2pd %xmm7,%xmm7 movapd %xmm1,%xmm2 movapd %xmm4,%xmm5 movapd %xmm7,%xmm8 mulpd %xmm1,%xmm1 ## lu*lu mulpd %xmm4,%xmm4 ## lu*lu mulpd %xmm7,%xmm7 ## lu*lu movapd nb313_three(%rsp),%xmm9 movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulpd %xmm0,%xmm1 ## rsq*lu*lu mulpd %xmm3,%xmm4 ## rsq*lu*lu mulpd %xmm6,%xmm7 ## rsq*lu*lu subpd %xmm1,%xmm9 subpd %xmm4,%xmm10 subpd %xmm7,%xmm11 ## 3-rsq*lu*lu mulpd %xmm2,%xmm9 mulpd %xmm5,%xmm10 mulpd %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb313_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## first iteration for rinvH1 mulpd %xmm15,%xmm10 ## first iteration for rinvH2 mulpd %xmm15,%xmm11 ## first iteration for rinvM ## second iteration step movapd %xmm9,%xmm2 movapd %xmm10,%xmm5 movapd %xmm11,%xmm8 mulpd %xmm2,%xmm2 ## lu*lu mulpd %xmm5,%xmm5 ## lu*lu mulpd %xmm8,%xmm8 ## lu*lu movapd nb313_three(%rsp),%xmm1 movapd %xmm1,%xmm4 movapd %xmm1,%xmm7 mulpd %xmm0,%xmm2 ## rsq*lu*lu mulpd %xmm3,%xmm5 ## rsq*lu*lu mulpd %xmm6,%xmm8 ## rsq*lu*lu subpd %xmm2,%xmm1 subpd %xmm5,%xmm4 subpd %xmm8,%xmm7 ## 3-rsq*lu*lu mulpd %xmm1,%xmm9 mulpd %xmm4,%xmm10 mulpd %xmm7,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb313_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## rinvH1 mulpd %xmm15,%xmm10 ## rinvH2 mulpd %xmm15,%xmm11 ## rinvM movapd %xmm9,nb313_rinvH1(%rsp) movapd %xmm10,nb313_rinvH2(%rsp) movapd %xmm11,nb313_rinvM(%rsp) ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movapd nb313_tsc(%rsp),%xmm1 mulpd %xmm9,%xmm0 ## r mulpd %xmm10,%xmm3 mulpd %xmm11,%xmm6 mulpd %xmm1,%xmm0 ## rtab mulpd %xmm1,%xmm3 mulpd %xmm1,%xmm6 ## truncate and convert to integers cvttpd2dq %xmm0,%xmm1 cvttpd2dq %xmm3,%xmm4 cvttpd2dq %xmm6,%xmm7 ## convert back to float cvtdq2pd %xmm1,%xmm2 cvtdq2pd %xmm4,%xmm5 cvtdq2pd %xmm7,%xmm8 ## multiply by 4 pslld $2,%xmm1 pslld $2,%xmm4 pslld $2,%xmm7 ## move to integer registers pshufd $1,%xmm1,%xmm13 pshufd $1,%xmm4,%xmm14 pshufd $1,%xmm7,%xmm15 movd %xmm1,%r8d movd %xmm4,%r10d movd %xmm7,%r12d movd %xmm13,%r9d movd %xmm14,%r11d movd %xmm15,%r13d movq nb313_VFtab(%rbp),%rsi ## calculate eps subpd %xmm2,%xmm0 subpd %xmm5,%xmm3 subpd %xmm8,%xmm6 movapd %xmm0,%xmm12 ## epsH1 movapd %xmm3,%xmm13 ## epsH2 movapd %xmm6,%xmm14 ## epsM ## Load LOTS of table data movlpd (%rsi,%r8,8),%xmm0 movlpd 8(%rsi,%r8,8),%xmm1 movlpd 16(%rsi,%r8,8),%xmm2 movlpd 24(%rsi,%r8,8),%xmm3 movlpd (%rsi,%r10,8),%xmm4 movlpd 8(%rsi,%r10,8),%xmm5 movlpd 16(%rsi,%r10,8),%xmm6 movlpd 24(%rsi,%r10,8),%xmm7 movlpd (%rsi,%r12,8),%xmm8 movlpd 8(%rsi,%r12,8),%xmm9 movlpd 16(%rsi,%r12,8),%xmm10 movlpd 24(%rsi,%r12,8),%xmm11 movhpd (%rsi,%r9,8),%xmm0 movhpd 8(%rsi,%r9,8),%xmm1 movhpd 16(%rsi,%r9,8),%xmm2 movhpd 24(%rsi,%r9,8),%xmm3 movhpd (%rsi,%r11,8),%xmm4 movhpd 8(%rsi,%r11,8),%xmm5 movhpd 16(%rsi,%r11,8),%xmm6 movhpd 24(%rsi,%r11,8),%xmm7 movhpd (%rsi,%r13,8),%xmm8 movhpd 8(%rsi,%r13,8),%xmm9 movhpd 16(%rsi,%r13,8),%xmm10 movhpd 24(%rsi,%r13,8),%xmm11 ## table data ready in xmm0-xmm3 , xmm4-xmm7 , and xmm8-xmm11 mulpd %xmm12,%xmm3 ## Heps mulpd %xmm13,%xmm7 mulpd %xmm14,%xmm11 mulpd %xmm12,%xmm2 ## Geps mulpd %xmm13,%xmm6 mulpd %xmm14,%xmm10 mulpd %xmm12,%xmm3 ## Heps2 mulpd %xmm13,%xmm7 mulpd %xmm14,%xmm11 addpd %xmm2,%xmm1 ## F+Geps addpd %xmm6,%xmm5 addpd %xmm10,%xmm9 addpd %xmm3,%xmm1 ## F+Geps+Heps2 = Fp addpd %xmm7,%xmm5 addpd %xmm11,%xmm9 addpd %xmm3,%xmm3 ## 2*Heps2 addpd %xmm7,%xmm7 addpd %xmm11,%xmm11 addpd %xmm2,%xmm3 ## 2*Heps2+Geps addpd %xmm6,%xmm7 addpd %xmm10,%xmm11 addpd %xmm1,%xmm3 ## FF = Fp + 2*Heps2 + Geps addpd %xmm5,%xmm7 addpd %xmm9,%xmm11 mulpd %xmm12,%xmm1 ## eps*Fp mulpd %xmm13,%xmm5 mulpd %xmm14,%xmm9 movapd nb313_qqH(%rsp),%xmm12 movapd nb313_qqM(%rsp),%xmm13 addpd %xmm0,%xmm1 ## VV addpd %xmm4,%xmm5 addpd %xmm8,%xmm9 mulpd %xmm12,%xmm1 ## VV*qq = vcoul mulpd %xmm12,%xmm5 mulpd %xmm13,%xmm9 mulpd %xmm12,%xmm3 ## FF*qq = fij mulpd %xmm12,%xmm7 mulpd %xmm13,%xmm11 ## accumulate vctot addpd nb313_vctot(%rsp),%xmm1 addpd %xmm9,%xmm5 addpd %xmm5,%xmm1 movapd %xmm1,nb313_vctot(%rsp) movapd nb313_tsc(%rsp),%xmm10 mulpd %xmm10,%xmm3 ## fscal mulpd %xmm10,%xmm7 mulpd %xmm11,%xmm10 xorpd %xmm4,%xmm4 xorpd %xmm8,%xmm8 xorpd %xmm11,%xmm11 subpd %xmm3,%xmm4 subpd %xmm7,%xmm8 subpd %xmm10,%xmm11 mulpd nb313_rinvH1(%rsp),%xmm4 mulpd nb313_rinvH2(%rsp),%xmm8 mulpd nb313_rinvM(%rsp),%xmm11 ## move j forces to xmm0-xmm2 movq nb313_faction(%rbp),%rdi movlpd (%rdi,%rax,8),%xmm0 movlpd 8(%rdi,%rax,8),%xmm1 movlpd 16(%rdi,%rax,8),%xmm2 movhpd (%rdi,%rbx,8),%xmm0 movhpd 8(%rdi,%rbx,8),%xmm1 movhpd 16(%rdi,%rbx,8),%xmm2 movapd %xmm4,%xmm3 movapd %xmm4,%xmm5 movapd %xmm8,%xmm7 movapd %xmm8,%xmm9 movapd %xmm11,%xmm10 movapd %xmm11,%xmm12 ## add forces from O interaction addpd nb313_fjx(%rsp),%xmm0 addpd nb313_fjy(%rsp),%xmm1 addpd nb313_fjz(%rsp),%xmm2 mulpd nb313_dxH1(%rsp),%xmm3 mulpd nb313_dyH1(%rsp),%xmm4 mulpd nb313_dzH1(%rsp),%xmm5 mulpd nb313_dxH2(%rsp),%xmm7 mulpd nb313_dyH2(%rsp),%xmm8 mulpd nb313_dzH2(%rsp),%xmm9 mulpd nb313_dxM(%rsp),%xmm10 mulpd nb313_dyM(%rsp),%xmm11 mulpd nb313_dzM(%rsp),%xmm12 addpd %xmm3,%xmm0 addpd %xmm4,%xmm1 addpd %xmm5,%xmm2 addpd nb313_fixH1(%rsp),%xmm3 addpd nb313_fiyH1(%rsp),%xmm4 addpd nb313_fizH1(%rsp),%xmm5 addpd %xmm7,%xmm0 addpd %xmm8,%xmm1 addpd %xmm9,%xmm2 addpd nb313_fixH2(%rsp),%xmm7 addpd nb313_fiyH2(%rsp),%xmm8 addpd nb313_fizH2(%rsp),%xmm9 addpd %xmm10,%xmm0 addpd %xmm11,%xmm1 addpd %xmm12,%xmm2 addpd nb313_fixM(%rsp),%xmm10 addpd nb313_fiyM(%rsp),%xmm11 addpd nb313_fizM(%rsp),%xmm12 movapd %xmm3,nb313_fixH1(%rsp) movapd %xmm4,nb313_fiyH1(%rsp) movapd %xmm5,nb313_fizH1(%rsp) movapd %xmm7,nb313_fixH2(%rsp) movapd %xmm8,nb313_fiyH2(%rsp) movapd %xmm9,nb313_fizH2(%rsp) movapd %xmm10,nb313_fixM(%rsp) movapd %xmm11,nb313_fiyM(%rsp) movapd %xmm12,nb313_fizM(%rsp) ## store back j forces from xmm0-xmm2 movlpd %xmm0,(%rdi,%rax,8) movlpd %xmm1,8(%rdi,%rax,8) movlpd %xmm2,16(%rdi,%rax,8) movhpd %xmm0,(%rdi,%rbx,8) movhpd %xmm1,8(%rdi,%rbx,8) movhpd %xmm2,16(%rdi,%rbx,8) ## should we do one more iteration? subl $2,nb313_innerk(%rsp) jl _nb_kernel313_x86_64_sse2.nb313_checksingle jmp _nb_kernel313_x86_64_sse2.nb313_unroll_loop_nb_kernel313_x86_64_sse2.nb313_checksingle: movl nb313_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel313_x86_64_sse2.nb313_dosingle jmp _nb_kernel313_x86_64_sse2.nb313_updateouterdata_nb_kernel313_x86_64_sse2.nb313_dosingle: movq nb313_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movq nb313_charge(%rbp),%rsi ## base of charge[] xorpd %xmm3,%xmm3 movlpd (%rsi,%rax,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb313_iqM(%rsp),%xmm3 mulpd nb313_iqH(%rsp),%xmm4 movapd %xmm3,nb313_qqM(%rsp) movapd %xmm4,nb313_qqH(%rsp) movq nb313_type(%rbp),%rsi movl (%rsi,%rax,4),%r8d movq nb313_vdwparam(%rbp),%rsi shll %r8d movl nb313_ntia(%rsp),%edi addl %edi,%r8d movlpd (%rsi,%r8,8),%xmm6 ## c6a movhpd 8(%rsi,%r8,8),%xmm6 ## c6a c12a xorpd %xmm7,%xmm7 movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movapd %xmm4,nb313_c6(%rsp) movapd %xmm6,nb313_c12(%rsp) movq nb313_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move coordinates to xmm0-xmm2 and xmm4-xmm6 movlpd (%rsi,%rax,8),%xmm4⌨️ 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?