nb_kernel334_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,355 行 · 第 1/5 页
S
2,355 行
movapd nb334_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 ## multiply by three (copy, mult. by two, add back) movdqa %xmm1,%xmm10 movdqa %xmm4,%xmm11 movdqa %xmm7,%xmm12 pslld $1,%xmm1 pslld $1,%xmm4 pslld $1,%xmm7 paddd %xmm10,%xmm1 paddd %xmm11,%xmm4 paddd %xmm12,%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 nb334_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 nb334_qqMH(%rsp),%xmm12 movapd nb334_qqMM(%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 nb334_vctot(%rsp),%xmm1 addpd %xmm9,%xmm5 addpd %xmm5,%xmm1 movapd %xmm1,nb334_vctot(%rsp) movapd nb334_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 nb334_rinvH1M(%rsp),%xmm4 mulpd nb334_rinvH2M(%rsp),%xmm8 mulpd nb334_rinvMM(%rsp),%xmm11 ## move j M forces to xmm0-xmm2 movq nb334_faction(%rbp),%rdi movlpd 72(%rdi,%rax,8),%xmm0 movlpd 80(%rdi,%rax,8),%xmm1 movlpd 88(%rdi,%rax,8),%xmm2 movhpd 72(%rdi,%rbx,8),%xmm0 movhpd 80(%rdi,%rbx,8),%xmm1 movhpd 88(%rdi,%rbx,8),%xmm2 movapd %xmm4,%xmm3 movapd %xmm4,%xmm5 movapd %xmm8,%xmm7 movapd %xmm8,%xmm9 movapd %xmm11,%xmm10 movapd %xmm11,%xmm12 mulpd nb334_dxH1M(%rsp),%xmm3 mulpd nb334_dyH1M(%rsp),%xmm4 mulpd nb334_dzH1M(%rsp),%xmm5 mulpd nb334_dxH2M(%rsp),%xmm7 mulpd nb334_dyH2M(%rsp),%xmm8 mulpd nb334_dzH2M(%rsp),%xmm9 mulpd nb334_dxMM(%rsp),%xmm10 mulpd nb334_dyMM(%rsp),%xmm11 mulpd nb334_dzMM(%rsp),%xmm12 addpd %xmm3,%xmm0 addpd %xmm4,%xmm1 addpd %xmm5,%xmm2 addpd nb334_fixH1(%rsp),%xmm3 addpd nb334_fiyH1(%rsp),%xmm4 addpd nb334_fizH1(%rsp),%xmm5 addpd %xmm7,%xmm0 addpd %xmm8,%xmm1 addpd %xmm9,%xmm2 addpd nb334_fixH2(%rsp),%xmm7 addpd nb334_fiyH2(%rsp),%xmm8 addpd nb334_fizH2(%rsp),%xmm9 addpd %xmm10,%xmm0 addpd %xmm11,%xmm1 addpd %xmm12,%xmm2 addpd nb334_fixM(%rsp),%xmm10 addpd nb334_fiyM(%rsp),%xmm11 addpd nb334_fizM(%rsp),%xmm12 movapd %xmm3,nb334_fixH1(%rsp) movapd %xmm4,nb334_fiyH1(%rsp) movapd %xmm5,nb334_fizH1(%rsp) movapd %xmm7,nb334_fixH2(%rsp) movapd %xmm8,nb334_fiyH2(%rsp) movapd %xmm9,nb334_fizH2(%rsp) movapd %xmm10,nb334_fixM(%rsp) movapd %xmm11,nb334_fiyM(%rsp) movapd %xmm12,nb334_fizM(%rsp) ## store back j M forces from xmm0-xmm2 movlpd %xmm0,72(%rdi,%rax,8) movlpd %xmm1,80(%rdi,%rax,8) movlpd %xmm2,88(%rdi,%rax,8) movhpd %xmm0,72(%rdi,%rbx,8) movhpd %xmm1,80(%rdi,%rbx,8) movhpd %xmm2,88(%rdi,%rbx,8) ## should we do one more iteration? subl $2,nb334_innerk(%rsp) jl _nb_kernel334_x86_64_sse2.nb334_checksingle jmp _nb_kernel334_x86_64_sse2.nb334_unroll_loop_nb_kernel334_x86_64_sse2.nb334_checksingle: movl nb334_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel334_x86_64_sse2.nb334_dosingle jmp _nb_kernel334_x86_64_sse2.nb334_updateouterdata_nb_kernel334_x86_64_sse2.nb334_dosingle: movq nb334_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movq nb334_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 ## load j O coordinates movsd (%rsi,%rax,8),%xmm4 movsd 8(%rsi,%rax,8),%xmm5 movsd 16(%rsi,%rax,8),%xmm6 ## xmm4 = Ox ## xmm5 = Oy ## xmm6 = Oz subsd nb334_ixO(%rsp),%xmm4 subsd nb334_iyO(%rsp),%xmm5 subsd nb334_izO(%rsp),%xmm6 ## store dx/dy/dz movapd %xmm4,%xmm13 movapd %xmm5,%xmm14 movapd %xmm6,%xmm15 ## square it mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 mulsd %xmm6,%xmm6 addsd %xmm5,%xmm4 addsd %xmm6,%xmm4 ## rsq in xmm4 cvtsd2ss %xmm4,%xmm5 rsqrtss %xmm5,%xmm5 cvtss2sd %xmm5,%xmm2 ## lu in low xmm2 ## lookup seed in xmm2 movapd %xmm2,%xmm5 ## copy of lu mulsd %xmm2,%xmm2 ## lu*lu movapd nb334_three(%rsp),%xmm1 mulsd %xmm4,%xmm2 ## rsq*lu*lu movapd nb334_half(%rsp),%xmm0 subsd %xmm2,%xmm1 ## 30-rsq*lu*lu mulsd %xmm5,%xmm1 mulsd %xmm0,%xmm1 ## xmm0=iter1 of rinv (new lu) movapd %xmm1,%xmm5 ## copy of lu mulsd %xmm1,%xmm1 ## lu*lu movapd nb334_three(%rsp),%xmm2 mulsd %xmm4,%xmm1 ## rsq*lu*lu movapd nb334_half(%rsp),%xmm0 subsd %xmm1,%xmm2 ## 30-rsq*lu*lu mulsd %xmm5,%xmm2 mulsd %xmm0,%xmm2 ## xmm0=iter2 of rinv (new lu) mulsd %xmm2,%xmm4 ## xmm4=r mulsd nb334_tsc(%rsp),%xmm4 cvttsd2si %xmm4,%r10d ## mm6 = lu idx cvtsi2sd %r10d,%xmm5 subsd %xmm5,%xmm4 movapd %xmm4,%xmm1 ## xmm1=eps ## xmm2=rinv movapd %xmm4,%xmm3 ## eps shll $2,%r10d ## idx *= 4 movq nb334_VFtab(%rbp),%rsi ## multiply by 3 lea (%r10,%r10,2),%r10 ## indices in r10, r11. Load dispersion and repulsion tables in parallel. ## NB: We are using a combined (coul+lj) table, so LJ data is offest ## 4*8=32 bytes. movsd 32(%rsi,%r10,8),%xmm4 movsd 40(%rsi,%r10,8),%xmm5 movsd 48(%rsi,%r10,8),%xmm6 movsd 56(%rsi,%r10,8),%xmm7 movsd 64(%rsi,%r10,8),%xmm8 movsd 72(%rsi,%r10,8),%xmm9 movsd 80(%rsi,%r10,8),%xmm10 movsd 88(%rsi,%r10,8),%xmm11 ## tables ready, in xmm4-xmm7 and xmm8-xmm11 mulsd %xmm1,%xmm7 ## Heps mulsd %xmm1,%xmm11 mulsd %xmm1,%xmm6 ## Geps mulsd %xmm1,%xmm10 mulsd %xmm1,%xmm7 ## Heps2 mulsd %xmm1,%xmm11 addsd %xmm6,%xmm5 ## F+Geps addsd %xmm10,%xmm9 addsd %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addsd %xmm11,%xmm9 addsd %xmm7,%xmm7 ## 2*Heps2 addsd %xmm11,%xmm11 addsd %xmm6,%xmm7 ## 2*Heps2+Geps addsd %xmm10,%xmm11 addsd %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps addsd %xmm9,%xmm11 mulsd %xmm1,%xmm5 ## eps*Fp mulsd %xmm1,%xmm9 addsd %xmm4,%xmm5 ## VV addsd %xmm8,%xmm9 mulsd nb334_c6(%rsp),%xmm5 ## VV*c6 = vnb6 mulsd nb334_c12(%rsp),%xmm9 ## VV*c12 = vnb12 addsd %xmm9,%xmm5 addsd nb334_Vvdwtot(%rsp),%xmm5 movsd %xmm5,nb334_Vvdwtot(%rsp) mulsd nb334_c6(%rsp),%xmm7 ## FF*c6 = fnb6 mulsd nb334_c12(%rsp),%xmm11 ## FF*c12 = fnb12 addsd %xmm11,%xmm7 mulsd nb334_tsc(%rsp),%xmm7 mulsd %xmm2,%xmm7 xorpd %xmm9,%xmm9 subsd %xmm7,%xmm9 mulsd %xmm9,%xmm13 mulsd %xmm9,%xmm14 mulsd %xmm9,%xmm15 movapd nb334_fixO(%rsp),%xmm0 movapd nb334_fiyO(%rsp),%xmm1 movapd nb334_fizO(%rsp),%xmm2 ## accumulate i forces addsd %xmm13,%xmm0 addsd %xmm14,%xmm1 addsd %xmm15,%xmm2 movsd %xmm0,nb334_fixO(%rsp) movsd %xmm1,nb334_fiyO(%rsp) movsd %xmm2,nb334_fizO(%rsp) ## the fj's - start by accumulating forces from memory movq nb334_faction(%rbp),%rdi movsd (%rdi,%rax,8),%xmm3 movsd 8(%rdi,%rax,8),%xmm4 movsd 16(%rdi,%rax,8),%xmm5 addsd %xmm13,%xmm3 addsd %xmm14,%xmm4 addsd %xmm15,%xmm5 movsd %xmm3,(%rdi,%rax,8) movsd %xmm4,8(%rdi,%rax,8) movsd %xmm5,16(%rdi,%rax,8) ## done with OO interaction ## move j H1 coordinates to local temp variables movq nb334_pos(%rbp),%rsi movsd 24(%rsi,%rax,8),%xmm0 movsd 32(%rsi,%rax,8),%xmm1 movsd 40(%rsi,%rax,8),%xmm2 ## xmm0 = H1x ## xmm1 = H1y ## xmm2 = H1z movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 movapd %xmm0,%xmm6 movapd %xmm1,%xmm7 movapd %xmm2,%xmm8 subsd nb334_ixH1(%rsp),%xmm0 subsd nb334_iyH1(%rsp),%xmm1 subsd nb334_izH1(%rsp),%xmm2 subsd nb334_ixH2(%rsp),%xmm3 subsd nb334_iyH2(%rsp),%xmm4 subsd nb334_izH2(%rsp),%xmm5 subsd nb334_ixM(%rsp),%xmm6 subsd nb334_iyM(%rsp),%xmm7 subsd nb334_izM(%rsp),%xmm8 movapd %xmm0,nb334_dxH1H1(%rsp) movapd %xmm1,nb334_dyH1H1(%rsp) movapd %xmm2,nb334_dzH1H1(%rsp) mulsd %xmm0,%xmm0 mulsd %xmm1,%xmm1 mulsd %xmm2,%xmm2 movapd %xmm3,nb334_dxH2H1(%rsp) movapd %xmm4,nb334_dyH2H1(%rsp) movapd %xmm5,nb334_dzH2H1(%rsp) mulsd %xmm3,%xmm3 mulsd %xmm4,%xmm4 mulsd %xmm5,%xmm5 movapd %xmm6,nb334_dxMH1(%rsp) movapd %xmm7,nb334_dyMH1(%rsp) movapd %xmm8,nb334_dzMH1(%rsp) mulsd %xmm6,%xmm6 mulsd %xmm7,%xmm7 mulsd %xmm8,%xmm8 addsd %xmm1,%xmm0 addsd %xmm2,%xmm0 addsd %xmm4,%xmm3 addsd %xmm5,%xmm3 addsd %xmm7,%xmm6 addsd %xmm8,%xmm6 ## start doing invsqrt for jH1 atoms cvtsd2ss %xmm0,%xmm1 cvtsd2ss %xmm3,%xmm4 cvtsd2ss %xmm6,%xmm7 rsqrtss %xmm1,%xmm1 rsqrtss %xmm4,%xmm4 rsqrtss %xmm7,%xmm7 cvtss2sd %xmm1,%xmm1 cvtss2sd %xmm4,%xmm4 cvtss2sd %xmm7,%xmm7 movapd %xmm1,%xmm2 movapd %xmm4,%xmm5 movapd %xmm7,%xmm8 mulsd %xmm1,%xmm1 ## lu*lu mulsd %xmm4,%xmm4 ## lu*lu mulsd %xmm7,%xmm7 ## lu*lu movapd nb334_three(%rsp),%xmm9 movapd %xmm9,%xmm10 movapd %xmm9,%xmm11 mulsd %xmm0,%xmm1 ## rsq*lu*lu mulsd %xmm3,%xmm4 ## rsq*lu*lu mulsd %xmm6,%xmm7 ## rsq*lu*lu subsd %xmm1,%xmm9 subsd %xmm4,%xmm10 subsd %xmm7,%xmm11 ## 3-rsq*lu*lu mulsd %xmm2,%xmm9 mulsd %xmm5,%xmm10 mulsd %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movapd nb334_half(%rsp),%xmm15 mulsd %xmm15,%xmm9 ## first iteration for rinvH1H1 mulsd %xmm15,%xmm10 ## first iteration for rinvH2H1 mulsd %xmm15,%xmm11 ## first iteration for rinvMH1 ## second iteration step movapd %xmm9,%xmm2⌨️ 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?