nb_kernel210_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,550 行 · 第 1/3 页
S
1,550 行
movlps xmm3, [rsi + r12*4] movhps xmm3, [rsi + r13*4] xorps xmm7, xmm7 movaps xmm0, xmm3 shufps xmm0, xmm7, 136 ;# 10001000 shufps xmm3, xmm7, 221 ;# 11011101 movaps [rsp + nb210_c6], xmm0 movaps [rsp + nb210_c12], xmm3 lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# load coordinates mov rdi, [rbp + nb210_pos] movlps xmm1, [rdi + rax*4] ;# x1 y1 - - movlps xmm2, [rdi + rbx*4] ;# x2 y2 - - movss xmm5, [rdi + rax*4 + 8] ;# z1 - - - movss xmm6, [rdi + rbx*4 + 8] ;# z2 - - - unpcklps xmm1, xmm2 ;# x1 x2 y1 y2 movhlps xmm2, xmm1 ;# y1 y2 - - unpcklps xmm5, xmm6 ;# z1 z2 - - ;# calc dr subps xmm1, [rsp + nb210_ix] subps xmm2, [rsp + nb210_iy] subps xmm5, [rsp + nb210_iz] ;# store dr in xmm9-xmm11 movaps xmm9, xmm1 movaps xmm10, xmm2 movaps xmm11, xmm5 ;# square it mulps xmm1,xmm1 mulps xmm2,xmm2 mulps xmm5,xmm5 addps xmm1, xmm2 addps xmm1, xmm5 ;# rsq in xmm1 movaps xmm7, [rsp + nb210_krf] mulps xmm7, xmm1 ;# krsq ;# calculate rinv=1/sqrt(rsq) rsqrtps xmm5, xmm1 movaps xmm6, xmm5 mulps xmm5, xmm5 movaps xmm4, [rsp + nb210_three] mulps xmm5, xmm1 ;# rsq*lu*lu subps xmm4, xmm5 ;# 30-rsq*lu*lu mulps xmm4, xmm6 mulps xmm4, [rsp + nb210_half] movaps xmm1, xmm4 mulps xmm4, xmm4 ;# xmm1=rinv ;# xmm4=rinvsq movaps xmm3, xmm1 subps xmm1, xmm7 subps xmm1, xmm7 ;# rinv-2*krsq addps xmm3, xmm7 ;# rinv+krsq subps xmm3, [rsp + nb210_crf] ;# rinv+krsq-crf mulps xmm3, xmm12 ;# vcoul=qq*(rinv+krsq-crf) mulps xmm1, xmm12 ;# fijC movaps xmm5, xmm4 mulps xmm5, xmm4 ;# rinv4 mulps xmm5, xmm4 ;# rinv6 movaps xmm6, xmm5 mulps xmm5, xmm5 ;# xmm5=rinv12 ;# add to vctot addps xmm3, [rsp + nb210_vctot] movlps [rsp + nb210_vctot], xmm3 mulps xmm6, [rsp + nb210_c6] ;# vvdw6=c6*rinv6 mulps xmm5, [rsp + nb210_c12] ;# vvdw12=c12*rinv12 movaps xmm7, xmm5 subps xmm5, xmm6 ;# Vvdw=Vvdw12-Vvdw6 mulps xmm6, [rsp + nb210_six] mulps xmm7, [rsp + nb210_twelve] subps xmm7, xmm6 addps xmm1, xmm7 mulps xmm4, xmm1 ;# xmm4=total fscal xorps xmm7, xmm7 movlhps xmm5, xmm7 ;# add potential to Vvdwtot addps xmm5, [rsp + nb210_Vvdwtot] movaps [rsp + nb210_Vvdwtot], xmm5 ;# calculate scalar force by multiplying dx/dy/dz with fscal mulps xmm9, xmm4 mulps xmm10, xmm4 mulps xmm11, xmm4 movlhps xmm9, xmm7 movlhps xmm10, xmm7 movlhps xmm11, xmm7 ;# accumulate i forces addps xmm13, xmm9 addps xmm14, xmm10 addps xmm15, xmm11 mov rsi, [rbp + nb210_faction] ;# the fj's - start by accumulating x & y forces from memory movlps xmm0, [rsi + rax*4] ;# x1 y1 - - movhps xmm0, [rsi + rbx*4] ;# x1 y1 x2 y2 unpcklps xmm9, xmm10 ;# x1 y1 x2 y2 addps xmm0, xmm9 movlps [rsi + rax*4], xmm0 movhps [rsi + rbx*4], xmm0 ;# z forces pshufd xmm8, xmm11, 1 addss xmm11, [rsi + rax*4 + 8] addss xmm8, [rsi + rbx*4 + 8] movss [rsi + rax*4 + 8], xmm11 movss [rsi + rbx*4 + 8], xmm8.nb210_checksingle: mov edx, [rsp + nb210_innerk] and edx, 1 jnz .nb210_dosingle jmp .nb210_updateouterdata.nb210_dosingle: mov rcx, [rsp + nb210_innerjjnr] mov eax, [rcx] mov rsi, [rbp + nb210_charge] movss xmm12, [rsi + rax*4] mulss xmm12, [rsp + nb210_iq] ;# qq mov rsi, [rbp + nb210_type] mov r12d, [rsi + rax*4] shl r12d, 1 mov edi, [rsp + nb210_ntia] add r12d, edi mov rsi, [rbp + nb210_vdwparam] movss xmm0, [rsi + r12*4] movss xmm3, [rsi + r12*4 + 4] movaps [rsp + nb210_c6], xmm0 movaps [rsp + nb210_c12], xmm3 lea rax, [rax + rax*2] ;# replace jnr with j3 mov rdi, [rbp + nb210_pos] ;# load coordinates movss xmm1, [rdi + rax*4] ;# x1 - - - movss xmm2, [rdi + rax*4 + 4] ;# y2 - - - movss xmm5, [rdi + rax*4 + 8] ;# 13 - - - ;# calc dr subss xmm1, [rsp + nb210_ix] subss xmm2, [rsp + nb210_iy] subss xmm5, [rsp + nb210_iz] ;# store dr in xmm9-xmm11 movaps xmm9, xmm1 movaps xmm10, xmm2 movaps xmm11, xmm5 ;# square it mulss xmm1,xmm1 mulss xmm2,xmm2 mulss xmm5,xmm5 addss xmm1, xmm2 addss xmm1, xmm5 ;# rsq in xmm1 movaps xmm7, [rsp + nb210_krf] mulss xmm7, xmm1 ;# krsq ;# calculate rinv=1/sqrt(rsq) rsqrtss xmm5, xmm1 movaps xmm6, xmm5 mulss xmm5, xmm5 movaps xmm4, [rsp + nb210_three] mulss xmm5, xmm1 ;# rsq*lu*lu subss xmm4, xmm5 ;# 30-rsq*lu*lu mulss xmm4, xmm6 mulss xmm4, [rsp + nb210_half] movaps xmm1, xmm4 mulss xmm4, xmm4 ;# xmm1=rinv ;# xmm4=rinvsq movaps xmm3, xmm1 subss xmm1, xmm7 subss xmm1, xmm7 ;# rinv-2*krsq addss xmm3, xmm7 ;# rinv+krsq subss xmm3, [rsp + nb210_crf] ;# rinv+krsq-crf mulss xmm3, xmm12 ;# vcoul=qq*(rinv+krsq-crf) mulss xmm1, xmm12 ;# fijC movaps xmm5, xmm4 mulss xmm5, xmm4 ;# rinv4 mulss xmm5, xmm4 ;# rinv6 movaps xmm6, xmm5 mulss xmm5, xmm5 ;# xmm5=rinv12 ;# add to vctot addss xmm3, [rsp + nb210_vctot] movss [rsp + nb210_vctot], xmm3 mulss xmm6, [rsp + nb210_c6] ;# vvdw6=c6*rinv6 mulss xmm5, [rsp + nb210_c12] ;# vvdw12=c12*rinv12 movaps xmm7, xmm5 subss xmm5, xmm6 ;# Vvdw=Vvdw12-Vvdw6 mulss xmm6, [rsp + nb210_six] mulss xmm7, [rsp + nb210_twelve] subss xmm7, xmm6 addss xmm1, xmm7 mulss xmm4, xmm1 ;# xmm4=total fscal ;# add potential to Vvdwtot addss xmm5, [rsp + nb210_Vvdwtot] movss [rsp + nb210_Vvdwtot], xmm5 ;# calculate scalar force by multiplying dx/dy/dz with fscal mulss xmm9, xmm4 mulss xmm10, xmm4 mulss xmm11, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# accumulate i forces addss xmm13, xmm9 addss xmm14, xmm10 addss xmm15, xmm11 mov rsi, [rbp + nb210_faction] ;# add to j forces addss xmm9, [rsi + rax*4] addss xmm10, [rsi + rax*4 + 4] addss xmm11, [rsi + rax*4 + 8] movss [rsi + rax*4], xmm9 movss [rsi + rax*4 + 4], xmm10 movss [rsi + rax*4 + 8], xmm11 .nb210_updateouterdata: mov ecx, [rsp + nb210_ii3] mov rdi, [rbp + nb210_faction] mov rsi, [rbp + nb210_fshift] mov edx, [rsp + nb210_is3] ;# accumulate i forces in xmm13, xmm14, xmm15 movhlps xmm0, xmm13 movhlps xmm1, xmm14 movhlps xmm2, xmm15 addps xmm0, xmm13 addps xmm1, xmm14 addps xmm2, xmm15 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [rdi + rcx*4] movss xmm4, [rdi + rcx*4 + 4] movss xmm5, [rdi + rcx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rdi + rcx*4], xmm3 movss [rdi + rcx*4 + 4], xmm4 movss [rdi + rcx*4 + 8], xmm5 ;# increment fshift force movss xmm3, [rsi + rdx*4] movss xmm4, [rsi + rdx*4 + 4] movss xmm5, [rsi + rdx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [rsi + rdx*4], xmm3 movss [rsi + rdx*4 + 4], xmm4 movss [rsi + rdx*4 + 8], xmm5 ;# get n from stack mov esi, [rsp + nb210_n] ;# get group index for i particle mov rdx, [rbp + nb210_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [rsp + nb210_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 + nb210_Vc] addss xmm7, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm7 ;# accumulate total potential energy and update it movaps xmm12, [rsp + nb210_Vvdwtot] ;# accumulate movhlps xmm6, xmm12 addps xmm12, xmm6 ;# pos 0-1 in xmm12 have the sum now movaps xmm6, xmm12 shufps xmm6, xmm6, 1 addss xmm12, xmm6 ;# add earlier value from mem mov rax, [rbp + nb210_Vvdw] addss xmm12, [rax + rdx*4] ;# move back to mem movss [rax + rdx*4], xmm12 ;# finish if last mov ecx, [rsp + nb210_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb210_outerend ;# not last, iterate outer loop once more! mov [rsp + nb210_n], esi jmp .nb210_outer.nb210_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb210_nri] ;# esi already loaded with n above sub ecx, esi jz .nb210_end ;# non-zero, do one more workunit jmp .nb210_threadloop.nb210_end: mov eax, [rsp + nb210_nouter] mov ebx, [rsp + nb210_ninner] mov rcx, [rbp + nb210_outeriter] mov rdx, [rbp + nb210_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 440 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel210nf_x86_64_sse.globl _nb_kernel210nf_x86_64_ssenb_kernel210nf_x86_64_sse: _nb_kernel210nf_x86_64_sse: ;# Room for return address and rbp (16 bytes).equiv nb210nf_fshift, 16.equiv nb210nf_gid, 24.equiv nb210nf_pos, 32.equiv nb210nf_faction, 40.equiv nb210nf_charge, 48.equiv nb210nf_p_facel, 56.equiv nb210nf_argkrf, 64.equiv nb210nf_argcrf, 72.equiv nb210nf_Vc, 80.equiv nb210nf_type, 88.equiv nb210nf_p_ntype, 96.equiv nb210nf_vdwparam, 104.equiv nb210nf_Vvdw, 112.equiv nb210nf_p_tabscale, 120.equiv nb210nf_VFtab, 128.equiv nb210nf_invsqrta, 136.equiv nb210nf_dvda, 144.equiv nb210nf_p_gbtabscale, 152.equiv nb210nf_GBtab, 160.equiv nb210nf_p_nthreads, 168.equiv nb210nf_count, 176.equiv nb210nf_mtx, 184.equiv nb210nf_outeriter, 192.equiv nb210nf_inneriter, 200.equiv nb210nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb210nf_ix, 0.equiv nb210nf_iy, 16.equiv nb210nf_iz, 32.equiv nb210nf_iq, 48.equiv nb210nf_c6, 64.equiv nb210nf_c12, 80.equiv nb210nf_vctot, 96.equiv nb210nf_Vvdwtot, 112.equiv nb210nf_half, 128.equiv nb210nf_three, 144.equiv nb210nf_krf, 160.equiv nb210nf_crf, 176.equiv nb210nf_nri, 192.equiv nb210nf_iinr, 200.equiv nb210nf_jindex, 208.equiv nb210nf_jjnr, 216.equiv nb210nf_shift, 224.equiv nb210nf_shiftvec, 232.equiv nb210nf_facel, 240.equiv nb210nf_innerjjnr, 248.equiv nb210nf_is3, 256.equiv nb210nf_ii3, 260.equiv nb210nf_ntia, 264.equiv nb210nf_innerk, 268.equiv nb210nf_n, 272.equiv nb210nf_nn1, 276.equiv nb210nf_ntype, 280.equiv nb210nf_nouter, 284.equiv nb210nf_ninner, 288 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 312 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb210nf_nouter], eax mov [rsp + nb210nf_ninner], eax mov edi, [rdi] mov [rsp + nb210nf_nri], edi mov [rsp + nb210nf_iinr], rsi mov [rsp + nb210nf_jindex], rdx mov [rsp + nb210nf_jjnr], rcx mov [rsp + nb210nf_shift], r8 mov [rsp + nb210nf_shiftvec], r9 mov rdi, [rbp + nb210nf_p_ntype] mov edi, [rdi] mov [rsp + nb210nf_ntype], edi mov rsi, [rbp + nb210nf_p_facel] movss xmm0, [rsi] movss [rsp + nb210nf_facel], xmm0 mov rsi, [rbp + nb210nf_argkrf] mov rdi, [rbp + nb210nf_argcrf] movss xmm1, [rsi] movss xmm2, [rdi] shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movaps [rsp + nb210nf_krf], xmm1 movaps [rsp + nb210nf_crf], xmm2 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# half in IEEE (hex) mov [rsp + nb210nf_half], eax movss xmm1, [rsp + nb210nf_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# one movaps xmm3, xmm2 addps xmm2, xmm2 ;# two addps xmm3, xmm2 ;# three movaps [rsp + nb210nf_half], xmm1 movaps [rsp + nb210nf_three], xmm3.nb210nf_threadloop: mov rsi, [rbp + nb210nf_count] ;# pointer to sync counter mov eax, [rsi].nb210nf_spinlock: mov ebx, eax ;# ebx=*count=nn0 add ebx, 1 ;# ebx=nn1=nn0+10 lock cmpxchg [esi], ebx ;# write nn1 to *counter, ;# if it hasnt changed. ;# or reread *counter to eax. pause ;# -> better p4 performance jnz .nb210nf_spinlock ;# if(nn1>nri) nn1=nri⌨️ 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?