nb_kernel303_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,073 行 · 第 1/5 页
S
2,073 行
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 nb303_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 nb303_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## rinvH1 mulpd %xmm15,%xmm10 ## rinvH2 mulpd %xmm15,%xmm11 ## rinvM movapd %xmm9,nb303_rinvH1(%rsp) movapd %xmm10,nb303_rinvH2(%rsp) movapd %xmm11,nb303_rinvM(%rsp) ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movapd nb303_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 nb303_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 nb303_qqH(%rsp),%xmm12 movapd nb303_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 nb303_vctot(%rsp),%xmm1 addpd %xmm9,%xmm5 addpd %xmm5,%xmm1 movapd %xmm1,nb303_vctot(%rsp) movapd nb303_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 nb303_rinvH1(%rsp),%xmm4 mulpd nb303_rinvH2(%rsp),%xmm8 mulpd nb303_rinvM(%rsp),%xmm11 ## move j forces to xmm0-xmm2 movq nb303_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 mulpd nb303_dxH1(%rsp),%xmm3 mulpd nb303_dyH1(%rsp),%xmm4 mulpd nb303_dzH1(%rsp),%xmm5 mulpd nb303_dxH2(%rsp),%xmm7 mulpd nb303_dyH2(%rsp),%xmm8 mulpd nb303_dzH2(%rsp),%xmm9 mulpd nb303_dxM(%rsp),%xmm10 mulpd nb303_dyM(%rsp),%xmm11 mulpd nb303_dzM(%rsp),%xmm12 addpd %xmm3,%xmm0 addpd %xmm4,%xmm1 addpd %xmm5,%xmm2 addpd nb303_fixH1(%rsp),%xmm3 addpd nb303_fiyH1(%rsp),%xmm4 addpd nb303_fizH1(%rsp),%xmm5 addpd %xmm7,%xmm0 addpd %xmm8,%xmm1 addpd %xmm9,%xmm2 addpd nb303_fixH2(%rsp),%xmm7 addpd nb303_fiyH2(%rsp),%xmm8 addpd nb303_fizH2(%rsp),%xmm9 addpd %xmm10,%xmm0 addpd %xmm11,%xmm1 addpd %xmm12,%xmm2 addpd nb303_fixM(%rsp),%xmm10 addpd nb303_fiyM(%rsp),%xmm11 addpd nb303_fizM(%rsp),%xmm12 movapd %xmm3,nb303_fixH1(%rsp) movapd %xmm4,nb303_fiyH1(%rsp) movapd %xmm5,nb303_fizH1(%rsp) movapd %xmm7,nb303_fixH2(%rsp) movapd %xmm8,nb303_fiyH2(%rsp) movapd %xmm9,nb303_fizH2(%rsp) movapd %xmm10,nb303_fixM(%rsp) movapd %xmm11,nb303_fiyM(%rsp) movapd %xmm12,nb303_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,nb303_innerk(%rsp) jl _nb_kernel303_x86_64_sse2.nb303_checksingle jmp _nb_kernel303_x86_64_sse2.nb303_unroll_loop_nb_kernel303_x86_64_sse2.nb303_checksingle: movl nb303_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel303_x86_64_sse2.nb303_dosingle jmp _nb_kernel303_x86_64_sse2.nb303_updateouterdata_nb_kernel303_x86_64_sse2.nb303_dosingle: movq nb303_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movq nb303_charge(%rbp),%rsi ## base of charge[] xorpd %xmm3,%xmm3 movlpd (%rsi,%rax,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb303_iqM(%rsp),%xmm3 mulpd nb303_iqH(%rsp),%xmm4 movapd %xmm3,nb303_qqM(%rsp) movapd %xmm4,nb303_qqH(%rsp) movq nb303_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## move coordinates to xmm4-xmm6 & xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm4 movlpd 8(%rsi,%rax,8),%xmm5 movlpd 16(%rsi,%rax,8),%xmm6 movapd %xmm4,%xmm0 movapd %xmm5,%xmm1 movapd %xmm6,%xmm2 ## calc dr subsd nb303_ixM(%rsp),%xmm4 subsd nb303_iyM(%rsp),%xmm5 subsd nb303_izM(%rsp),%xmm6 ## store dr movapd %xmm4,nb303_dxM(%rsp) movapd %xmm5,nb303_dyM(%rsp) movapd %xmm6,nb303_dzM(%rsp) ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqM in xmm7 ## move j coords to xmm4-xmm6 movapd %xmm0,%xmm4 movapd %xmm1,%xmm5 movapd %xmm2,%xmm6 ## calc dr subsd nb303_ixH1(%rsp),%xmm4 subsd nb303_iyH1(%rsp),%xmm5 subsd nb303_izH1(%rsp),%xmm6 ## store dr movapd %xmm4,nb303_dxH1(%rsp) movapd %xmm5,nb303_dyH1(%rsp) movapd %xmm6,nb303_dzH1(%rsp) ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm6 addsd %xmm4,%xmm6 ## rsqH1 in xmm6 ## move j coords to xmm3-xmm5 movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 ## calc dr subsd nb303_ixH2(%rsp),%xmm3 subsd nb303_iyH2(%rsp),%xmm4 subsd nb303_izH2(%rsp),%xmm5 ## store dr movapd %xmm3,nb303_dxH2(%rsp) movapd %xmm4,nb303_dyH2(%rsp) movapd %xmm5,nb303_dzH2(%rsp) ## square it mulsd %xmm3,%xmm3 mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 addsd %xmm4,%xmm5 addsd %xmm3,%xmm5 ## rsqH2 in xmm5, rsqH1 in xmm6, rsqM in xmm7 ## start with rsqM - put seed in xmm2 cvtsd2ss %xmm7,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb303_three(%rsp),%xmm4 mulsd %xmm7,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm7,%xmm2 movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb303_three(%rsp),%xmm4 subsd %xmm2,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## rinv movapd %xmm4,nb303_rinvM(%rsp) ## rinvM in xmm4 mulsd %xmm4,%xmm7 movapd %xmm7,nb303_rM(%rsp) ## r in xmm7 ## rsqH1 - seed in xmm2 cvtsd2ss %xmm6,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb303_three(%rsp),%xmm4 mulsd %xmm6,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm6,%xmm2 movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb303_three(%rsp),%xmm4 subsd %xmm2,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## rinv movapd %xmm4,nb303_rinvH1(%rsp) ## rinvH1 mulsd %xmm4,%xmm6 movapd %xmm6,nb303_rH1(%rsp) ## rH1 ## rsqH2 - seed in xmm2 cvtsd2ss %xmm5,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb303_three(%rsp),%xmm4 mulsd %xmm5,%xmm2 ## rsq*lu*lu subsd %xmm2,%xmm4 ## 30-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*(3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm5,%xmm2 movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb303_three(%rsp),%xmm4 subsd %xmm2,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb303_half(%rsp),%xmm4 ## rinv movapd %xmm4,nb303_rinvH2(%rsp) ## rinv mulsd %xmm4,%xmm5 movapd %xmm5,nb303_rH2(%rsp) ## r ## do M interactions ## rM is still in xmm7 ⌨️ 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?