nb_kernel203_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,708 行 · 第 1/4 页
S
1,708 行
mulpd %xmm2,%xmm2 ## lu*lu mulpd %xmm5,%xmm5 ## lu*lu mulpd %xmm8,%xmm8 ## lu*lu movapd nb203_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 nb203_half(%rsp),%xmm15 mulpd %xmm15,%xmm9 ## rinvH1 mulpd %xmm15,%xmm10 ## rinvH2 mulpd %xmm15,%xmm11 ## rinvM ## interactions ## rsq in xmm0,xmm3,xmm6 ## rinv in xmm9, xmm10, xmm11 movapd %xmm9,%xmm1 ## copy of rinv movapd %xmm10,%xmm4 movapd %xmm11,%xmm7 movapd nb203_krf(%rsp),%xmm2 mulpd %xmm9,%xmm9 ## rinvsq mulpd %xmm10,%xmm10 mulpd %xmm11,%xmm11 mulpd %xmm2,%xmm0 ## k*rsq mulpd %xmm2,%xmm3 mulpd %xmm2,%xmm6 movapd %xmm0,%xmm2 ## copy of k*rsq movapd %xmm3,%xmm5 movapd %xmm6,%xmm8 addpd %xmm1,%xmm2 ## rinv+krsq addpd %xmm4,%xmm5 addpd %xmm7,%xmm8 movapd nb203_crf(%rsp),%xmm14 subpd %xmm14,%xmm2 ## rinv+krsq-crf subpd %xmm14,%xmm5 subpd %xmm14,%xmm8 movapd nb203_qqH(%rsp),%xmm12 movapd nb203_qqM(%rsp),%xmm13 mulpd %xmm12,%xmm2 ## voul=qq*(rinv+ krsq-crf) mulpd %xmm12,%xmm5 ## voul=qq*(rinv+ krsq-crf) mulpd %xmm13,%xmm8 ## voul=qq*(rinv+ krsq-crf) addpd %xmm0,%xmm0 ## 2*krsq addpd %xmm3,%xmm3 addpd %xmm6,%xmm6 subpd %xmm0,%xmm1 ## rinv-2*krsq subpd %xmm3,%xmm4 subpd %xmm6,%xmm7 mulpd %xmm12,%xmm1 ## (rinv-2*krsq)*qq mulpd %xmm12,%xmm4 mulpd %xmm13,%xmm7 addpd nb203_vctot(%rsp),%xmm2 addpd %xmm8,%xmm5 addpd %xmm5,%xmm2 movapd %xmm2,nb203_vctot(%rsp) mulpd %xmm1,%xmm9 ## fscal mulpd %xmm4,%xmm10 mulpd %xmm7,%xmm11 ## move j forces to xmm0-xmm2 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 %xmm9,%xmm7 movapd %xmm9,%xmm8 movapd %xmm11,%xmm13 movapd %xmm11,%xmm14 movapd %xmm11,%xmm15 movapd %xmm10,%xmm11 movapd %xmm10,%xmm12 mulpd nb203_dxH1(%rsp),%xmm7 mulpd nb203_dyH1(%rsp),%xmm8 mulpd nb203_dzH1(%rsp),%xmm9 mulpd nb203_dxH2(%rsp),%xmm10 mulpd nb203_dyH2(%rsp),%xmm11 mulpd nb203_dzH2(%rsp),%xmm12 mulpd nb203_dxM(%rsp),%xmm13 mulpd nb203_dyM(%rsp),%xmm14 mulpd nb203_dzM(%rsp),%xmm15 addpd %xmm7,%xmm0 addpd %xmm8,%xmm1 addpd %xmm9,%xmm2 addpd nb203_fixH1(%rsp),%xmm7 addpd nb203_fiyH1(%rsp),%xmm8 addpd nb203_fizH1(%rsp),%xmm9 addpd %xmm10,%xmm0 addpd %xmm11,%xmm1 addpd %xmm12,%xmm2 addpd nb203_fixH2(%rsp),%xmm10 addpd nb203_fiyH2(%rsp),%xmm11 addpd nb203_fizH2(%rsp),%xmm12 addpd %xmm13,%xmm0 addpd %xmm14,%xmm1 addpd %xmm15,%xmm2 addpd nb203_fixM(%rsp),%xmm13 addpd nb203_fiyM(%rsp),%xmm14 addpd nb203_fizM(%rsp),%xmm15 movapd %xmm7,nb203_fixH1(%rsp) movapd %xmm8,nb203_fiyH1(%rsp) movapd %xmm9,nb203_fizH1(%rsp) movapd %xmm10,nb203_fixH2(%rsp) movapd %xmm11,nb203_fiyH2(%rsp) movapd %xmm12,nb203_fizH2(%rsp) movapd %xmm13,nb203_fixM(%rsp) movapd %xmm14,nb203_fiyM(%rsp) movapd %xmm15,nb203_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,nb203_innerk(%rsp) jl _nb_kernel203_x86_64_sse2.nb203_checksingle jmp _nb_kernel203_x86_64_sse2.nb203_unroll_loop_nb_kernel203_x86_64_sse2.nb203_checksingle: movl nb203_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel203_x86_64_sse2.nb203_dosingle jmp _nb_kernel203_x86_64_sse2.nb203_updateouterdata_nb_kernel203_x86_64_sse2.nb203_dosingle: movq nb203_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb203_innerjjnr(%rsp) movq nb203_charge(%rbp),%rsi ## base of charge[] xorpd %xmm3,%xmm3 movlpd (%rsi,%rax,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb203_iqM(%rsp),%xmm3 mulpd nb203_iqH(%rsp),%xmm4 movapd %xmm3,nb203_qqM(%rsp) movapd %xmm4,nb203_qqH(%rsp) movq nb203_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 nb203_ixM(%rsp),%xmm4 subsd nb203_iyM(%rsp),%xmm5 subsd nb203_izM(%rsp),%xmm6 ## store dr movapd %xmm4,nb203_dxM(%rsp) movapd %xmm5,nb203_dyM(%rsp) movapd %xmm6,nb203_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 nb203_ixH1(%rsp),%xmm4 subsd nb203_iyH1(%rsp),%xmm5 subsd nb203_izH1(%rsp),%xmm6 ## store dr movapd %xmm4,nb203_dxH1(%rsp) movapd %xmm5,nb203_dyH1(%rsp) movapd %xmm6,nb203_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 nb203_ixH2(%rsp),%xmm3 subsd nb203_iyH2(%rsp),%xmm4 subsd nb203_izH2(%rsp),%xmm5 ## store dr movapd %xmm3,nb203_dxH2(%rsp) movapd %xmm4,nb203_dyH2(%rsp) movapd %xmm5,nb203_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 movapd %xmm5,%xmm0 movapd %xmm6,%xmm1 movapd %xmm7,%xmm2 mulsd nb203_krf(%rsp),%xmm0 mulsd nb203_krf(%rsp),%xmm1 mulsd nb203_krf(%rsp),%xmm2 movapd %xmm0,nb203_krsqH2(%rsp) movapd %xmm1,nb203_krsqH1(%rsp) movapd %xmm2,nb203_krsqM(%rsp) ## start with rsqM - put seed in xmm2 cvtsd2ss %xmm7,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb203_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 nb203_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm7 ## rsq*lu*lu movapd nb203_three(%rsp),%xmm4 subsd %xmm7,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb203_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm7 ## rinvM in xmm7 ## rsqH1 - seed in xmm2 cvtsd2ss %xmm6,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb203_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 nb203_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm6 ## rsq*lu*lu movapd nb203_three(%rsp),%xmm4 subsd %xmm6,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb203_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm6 ## rinvH1 in xmm6 ## rsqH2 - seed in xmm2 cvtsd2ss %xmm5,%xmm2 rsqrtss %xmm2,%xmm2 cvtss2sd %xmm2,%xmm2 movapd %xmm2,%xmm3 mulsd %xmm2,%xmm2 movapd nb203_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 nb203_half(%rsp),%xmm4 ## iter1 ( new lu) movapd %xmm4,%xmm3 mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm4,%xmm5 ## rsq*lu*lu movapd nb203_three(%rsp),%xmm4 subsd %xmm5,%xmm4 ## 3-rsq*lu*lu mulsd %xmm3,%xmm4 ## lu*( 3-rsq*lu*lu) mulsd nb203_half(%rsp),%xmm4 ## rinv movapd %xmm4,%xmm5 ## rinvH2 in xmm5 ## do M interactions movapd %xmm7,%xmm4 mulsd %xmm4,%xmm4 ## xmm6=rinv, xmm4=rinvsq movapd %xmm7,%xmm3 movapd nb203_krsqM(%rsp),%xmm0 addsd %xmm0,%xmm7 ## xmm6=rinv+ krsq mulsd nb203_two(%rsp),%xmm0 subsd nb203_crf(%rsp),%xmm7 subsd %xmm0,%xmm3 ## xmm7=rinv-2*krsq mulsd nb203_qqM(%rsp),%xmm7 ## vcoul mulsd nb203_qqM(%rsp),%xmm3 mulsd %xmm3,%xmm4 ## total fsH1 in xmm4 addsd nb203_vctot(%rsp),%xmm7 movapd nb203_dxM(%rsp),%xmm0 movapd nb203_dyM(%rsp),%xmm1 movapd nb203_dzM(%rsp),%xmm2 movlpd %xmm7,nb203_vctot(%rsp) mulsd %xmm4,%xmm0 mulsd %xmm4,%xmm1 mulsd %xmm4,%xmm2 ## update M forces movapd nb203_fixM(%rsp),%xmm3 movapd nb203_fiyM(%rsp),%xmm4 movapd nb203_fizM(%rsp),%xmm7 addsd %xmm0,%xmm3 addsd %xmm1,%xmm4 addsd %xmm2,%xmm7 movlpd %xmm3,nb203_fixM(%rsp) movlpd %xmm4,nb203_fiyM(%rsp) movlpd %xmm7,nb203_fizM(%rsp) ## update j forces with water M movlpd %xmm0,nb203_fjx(%rsp) movlpd %xmm1,nb203_fjy(%rsp) movlpd %xmm2,nb203_fjz(%rsp) ## H1 interactions movapd %xmm6,%xmm4 mulsd %xmm4,%xmm4 ## xmm6=rinv, xmm4=rinvsq movapd %xmm6,%xmm7 movapd nb203_krsqH1(%rsp),%xmm0 addsd %xmm0,%xmm6 ## xmm6=rinv+ krsq mulsd nb203_two(%rsp),%xmm0 subsd nb203_crf(%rsp),%xmm6 subsd %xmm0,%xmm7 ## xmm7=rinv-2*krsq mulsd nb203_qqH(%rsp),%xmm6 ## vcoul mulsd nb203_qqH(%rsp),%xmm7 mulsd %xmm7,%xmm4 ## total fsH1 in xmm4 addsd nb203_vctot(%rsp),%xmm6 movapd nb203_dxH1(%rsp),%xmm0 movapd nb203_dyH1(%rsp),%xmm1 movapd nb203_dzH1(%rsp),%xmm2 movlpd %xmm6,nb203_vctot(%rsp) mulsd %xmm4,%xmm0 mulsd %xmm4,%xmm1 mulsd %xmm4,%xmm2 ## update H1 forces movapd nb203_fixH1(%rsp),%xmm3 movapd nb203_fiyH1(%rsp),%xmm4 movapd nb203_fizH1(%rsp),%xmm7 addsd %xmm0,%xmm3 addsd %xmm1,%xmm4 addsd %xmm2,%xmm7 movlpd %xmm3,nb203_fixH1(%rsp) movlpd %xmm4,nb203_fiyH1(%rsp) movlpd %xmm7,nb203_fizH1(%rsp) ## update j forces with water H1 addsd nb203_fjx(%rsp),%xmm0 addsd nb203_fjy(%rsp),%xmm1 addsd nb203_fjz(%rsp),%xmm2 movlpd %xmm0,nb203_fjx(%rsp) movlpd %xmm1,nb203_fjy(%rsp) movlpd %xmm2,nb203_fjz(%rsp) ## H2 interactions movapd %xmm5,%xmm4 mulsd %xmm4,%xmm4 ## xmm5=rinv, xmm4=rinvsq movapd %xmm5,%xmm7 movapd nb203_krsqH2(%rsp),%xmm0 addsd %xmm0,%xmm5 ## xmm5=rinv+ krsq mulsd nb203_two(%rsp),%xmm0 subsd nb203_crf(%rsp),%xmm5 subsd %xmm0,%xmm7 ## xmm7=rinv-2*krsq mulsd nb203_qqH(%rsp),%xmm5 ## vcoul mulsd nb203_qqH(%rsp),%xmm7 mulsd %xmm7,%xmm4 ## total fsH2 in xmm4 addsd nb203_vctot(%rsp),%xmm5 movapd nb203_dxH2(%rsp),%xmm0 movapd nb203_dyH2(%rsp),%xmm1 movapd nb203_dzH2(%rsp),%xmm2 movlpd %xmm5,nb203_vctot(%rsp) mulsd %xmm4,%xmm0 mulsd %xmm4,%xmm1 mulsd %xmm4,%xmm2 ## update H2 forces movapd nb203_fixH2(%rsp),%xmm3 movapd nb203_fiyH2(%rsp),%xmm4 movapd nb203_fizH2(%rsp),%xmm7 addsd %xmm0,%xmm3 addsd %xmm1,%xmm4 addsd %xmm2,%xmm7 movlpd %xmm3,nb203_fixH2(%rsp) movlpd %xmm4,nb203_fiyH2(%rsp) movlpd %xmm7,nb203_fizH2(%rsp) movq nb203_faction(%rbp),%rdi ## update j forces addsd nb203_fjx(%rsp),%xmm0⌨️ 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?