nb_kernel310_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,805 行 · 第 1/4 页
S
1,805 行
addps xmm4, xmm6 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [rsp + nb310nf_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb310nf_half] subps xmm1, xmm5 ;# 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv mulps xmm4, xmm0 ;# xmm4=r mulps xmm4, [rsp + nb310nf_tsc] movhlps xmm5, xmm4 cvttps2pi mm6, xmm4 cvttps2pi mm7, xmm5 ;# mm6/mm7 contain lu indices cvtpi2ps xmm6, mm6 cvtpi2ps xmm5, mm7 movlhps xmm6, xmm5 subps xmm4, xmm6 movaps xmm1, xmm4 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd mm0, eax movd mm1, ebx movd mm2, ecx movd mm3, edx mov rsi, [rbp + nb310nf_VFtab] movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 movlps xmm5, [rsi + rax*4] movlps xmm7, [rsi + rcx*4] movhps xmm5, [rsi + rbx*4] movhps xmm7, [rsi + rdx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# 10001000 shufps xmm5, xmm7, 221 ;# 11011101 movlps xmm7, [rsi + rax*4 + 8] movlps xmm3, [rsi + rcx*4 + 8] movhps xmm7, [rsi + rbx*4 + 8] movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# 10001000 shufps xmm7, xmm3, 221 ;# 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp movaps xmm3, [rsp + nb310nf_qq] mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV ;# L-J movaps xmm4, xmm0 mulps xmm4, xmm0 ;# xmm4=rinvsq ;# at this point mm5 contains vcoul ;# increment vcoul - then we can get rid of mm5 ;# update vctot addps xmm5, [rsp + nb310nf_vctot] movaps xmm6, xmm4 mulps xmm6, xmm4 movaps [rsp + nb310nf_vctot], xmm5 mulps xmm6, xmm4 ;# xmm6=rinvsix movaps xmm4, xmm6 mulps xmm4, xmm4 ;# xmm4=rinvtwelve mulps xmm6, [rsp + nb310nf_c6] mulps xmm4, [rsp + nb310nf_c12] movaps xmm7, [rsp + nb310nf_Vvdwtot] addps xmm7, xmm4 subps xmm7, xmm6 movaps [rsp + nb310nf_Vvdwtot], xmm7 ;# should we do one more iteration? sub dword ptr [rsp + nb310nf_innerk], 4 jl .nb310nf_finish_inner jmp .nb310nf_unroll_loop.nb310nf_finish_inner: ;# check if at least two particles remain add dword ptr [rsp + nb310nf_innerk], 4 mov edx, [rsp + nb310nf_innerk] and edx, 2 jnz .nb310nf_dopair jmp .nb310nf_checksingle.nb310nf_dopair: mov rsi, [rbp + nb310nf_charge] mov rcx, [rsp + nb310nf_innerjjnr] mov eax, [rcx] mov ebx, [rcx + 4] add qword ptr [rsp + nb310nf_innerjjnr], 8 xorps xmm7, xmm7 movss xmm3, [rsi + rax*4] movss xmm6, [rsi + rbx*4] shufps xmm3, xmm6, 0 shufps xmm3, xmm3, 8 ;# 00001000 ;# xmm3(0,1) has the charges mulps xmm3, [rsp + nb310nf_iq] movlhps xmm3, xmm7 movaps [rsp + nb310nf_qq], xmm3 mov rsi, [rbp + nb310nf_type] mov ecx, eax mov edx, ebx mov ecx, [rsi + rcx*4] mov edx, [rsi + rdx*4] mov rsi, [rbp + nb310nf_vdwparam] shl ecx, 1 shl edx, 1 mov edi, [rsp + nb310nf_ntia] add ecx, edi add edx, edi movlps xmm6, [rsi + rcx*4] movhps xmm6, [rsi + rdx*4] mov rdi, [rbp + nb310nf_pos] movaps xmm4, xmm6 shufps xmm4, xmm4, 8 ;# 00001000 shufps xmm6, xmm6, 13 ;# 00001101 movlhps xmm4, xmm7 movlhps xmm6, xmm7 movaps [rsp + nb310nf_c6], xmm4 movaps [rsp + nb310nf_c12], xmm6 lea rax, [rax + rax*2] lea rbx, [rbx + rbx*2] ;# move coordinates to xmm0-xmm2 movlps xmm1, [rdi + rax*4] movss xmm2, [rdi + rax*4 + 8] movhps xmm1, [rdi + rbx*4] movss xmm0, [rdi + rbx*4 + 8] movlhps xmm3, xmm7 shufps xmm2, xmm0, 0 movaps xmm0, xmm1 shufps xmm2, xmm2, 136 ;# 10001000 shufps xmm0, xmm0, 136 ;# 10001000 shufps xmm1, xmm1, 221 ;# 11011101 ;# move ix-iz to xmm4-xmm6 xorps xmm7, xmm7 movaps xmm4, [rsp + nb310nf_ix] movaps xmm5, [rsp + nb310nf_iy] movaps xmm6, [rsp + nb310nf_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [rsp + nb310nf_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb310nf_half] subps xmm1, xmm5 ;# 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv mulps xmm4, xmm0 ;# xmm4=r mulps xmm4, [rsp + nb310nf_tsc] cvttps2pi mm6, xmm4 ;# mm6 contain lu indices cvtpi2ps xmm6, mm6 subps xmm4, xmm6 movaps xmm1, xmm4 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 mov rsi, [rbp + nb310nf_VFtab] movd ecx, mm6 psrlq mm6, 32 movd edx, mm6 movlps xmm5, [rsi + rcx*4] movhps xmm5, [rsi + rdx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm4, 136 ;# 10001000 shufps xmm5, xmm7, 221 ;# 11011101 movlps xmm7, [rsi + rcx*4 + 8] movhps xmm7, [rsi + rdx*4 + 8] movaps xmm6, xmm7 shufps xmm6, xmm6, 136 ;# 10001000 shufps xmm7, xmm7, 221 ;# 11011101 ;# table ready in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp movaps xmm3, [rsp + nb310nf_qq] mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV ;# L-J movaps xmm4, xmm0 mulps xmm4, xmm0 ;# xmm4=rinvsq ;# at this point mm5 contains vcoul ;# increment vcoul - then we can get rid of mm5 ;# update vctot addps xmm5, [rsp + nb310nf_vctot] movaps xmm6, xmm4 mulps xmm6, xmm4 movaps [rsp + nb310nf_vctot], xmm5 mulps xmm6, xmm4 ;# xmm6=rinvsix movaps xmm4, xmm6 mulps xmm4, xmm4 ;# xmm4=rinvtwelve mulps xmm6, [rsp + nb310nf_c6] mulps xmm4, [rsp + nb310nf_c12] movaps xmm7, [rsp + nb310nf_Vvdwtot] addps xmm7, xmm4 subps xmm7, xmm6 movaps [rsp + nb310nf_Vvdwtot], xmm7.nb310nf_checksingle: mov edx, [rsp + nb310nf_innerk] and edx, 1 jnz .nb310nf_dosingle jmp .nb310nf_updateouterdata.nb310nf_dosingle: mov rsi, [rbp + nb310nf_charge] mov rdi, [rbp + nb310nf_pos] mov rcx, [rsp + nb310nf_innerjjnr] mov eax, [rcx] xorps xmm6, xmm6 movss xmm6, [rsi + rax*4] ;# xmm6(0) has the charge mulps xmm6, [rsp + nb310nf_iq] movaps [rsp + nb310nf_qq], xmm6 mov rsi, [rbp + nb310nf_type] mov ecx, eax mov ecx, [rsi + rcx*4] mov rsi, [rbp + nb310nf_vdwparam] shl ecx, 1 add ecx, [rsp + nb310nf_ntia] movlps xmm6, [rsi + rcx*4] movaps xmm4, xmm6 shufps xmm4, xmm4, 252 ;# 11111100 shufps xmm6, xmm6, 253 ;# 11111101 movaps [rsp + nb310nf_c6], xmm4 movaps [rsp + nb310nf_c12], xmm6 lea rax, [rax + rax*2] ;# move coordinates to xmm0-xmm2 movss xmm0, [rdi + rax*4] movss xmm1, [rdi + rax*4 + 4] movss xmm2, [rdi + rax*4 + 8] movaps xmm4, [rsp + nb310nf_ix] movaps xmm5, [rsp + nb310nf_iy] movaps xmm6, [rsp + nb310nf_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [rsp + nb310nf_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [rsp + nb310nf_half] subps xmm1, xmm5 ;# 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv mulps xmm4, xmm0 ;# xmm4=r mulps xmm4, [rsp + nb310nf_tsc] cvttps2pi mm6, xmm4 ;# mm6 contain lu indices cvtpi2ps xmm6, mm6 subps xmm4, xmm6 movaps xmm1, xmm4 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 mov rsi, [rbp + nb310nf_VFtab] movd ebx, mm6 movlps xmm4, [rsi + rbx*4] movlps xmm6, [rsi + rbx*4 + 8] movaps xmm5, xmm4 movaps xmm7, xmm6 shufps xmm5, xmm5, 1 shufps xmm7, xmm7, 1 ;# table ready in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp movaps xmm3, [rsp + nb310nf_qq] mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV ;# L-J movaps xmm4, xmm0 mulps xmm4, xmm0 ;# xmm4=rinvsq ;# at this point mm5 contains vcoul ;# increment vcoul - then we can get rid of mm5 ;# update vctot addss xmm5, [rsp + nb310nf_vctot] movaps xmm6, xmm4 mulps xmm6, xmm4 movss [rsp + nb310nf_vctot], xmm5 mulps xmm6, xmm4 ;# xmm6=rinvsix movaps xmm4, xmm6 mulps xmm4, xmm4 ;# xmm4=rinvtwelve mulps xmm6, [rsp + nb310nf_c6] mulps xmm4, [rsp + nb310nf_c12] movss xmm7, [rsp + nb310nf_Vvdwtot] addps xmm7, xmm4 subps xmm7, xmm6 movss [rsp + nb310nf_Vvdwtot], xmm7.nb310nf_updateouterdata: ;# get n from stack mov esi, [rsp + nb310nf_n] ;# get group index for i particle mov rdx, [rbp + nb310nf_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [rsp + nb310nf_vctot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb310nf_Vc] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# accumulate total lj energy and update it movaps xmm7, [rsp + nb310nf_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov rax, [rbp + nb310nf_Vvdw] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# finish if last mov ecx, [rsp + nb310nf_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb310nf_outerend ;# not last, iterate outer loop once more! mov [rsp + nb310nf_n], esi jmp .nb310nf_outer.nb310nf_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb310nf_nri] ;# esi already loaded with n above sub ecx, esi jz .nb310nf_end ;# non-zero, do one more workunit jmp .nb310nf_threadloop.nb310nf_end: mov eax, [rsp + nb310nf_nouter] mov ebx, [rsp + nb310nf_ninner] mov rcx, [rbp + nb310nf_outeriter] mov rdx, [rbp + nb310nf_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 312 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret⌨️ 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?