nb_kernel300_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,533 行 · 第 1/3 页
S
1,533 行
mulps xmm3, [esp + nb300_iq] movlhps xmm3, xmm7 movaps [esp + nb300_qq], xmm3 mov edi, [ebp + nb300_pos] lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] ;# move coordinates to xmm0-xmm2 movlps xmm1, [edi + eax*4] movss xmm2, [edi + eax*4 + 8] movhps xmm1, [edi + ebx*4] movss xmm0, [edi + ebx*4 + 8] movlhps xmm3, xmm7 shufps xmm2, xmm0, 0 movaps xmm0, xmm1 shufps xmm2, xmm2, 136 ;# constant 10001000 shufps xmm0, xmm0, 136 ;# constant 10001000 shufps xmm1, xmm1, 221 ;# constant 11011101 mov edi, [ebp + nb300_faction] ;# move ix-iz to xmm4-xmm6 xorps xmm7, xmm7 movaps xmm4, [esp + nb300_ix] movaps xmm5, [esp + nb300_iy] movaps xmm6, [esp + nb300_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb300_dx], xmm4 movaps [esp + nb300_dy], xmm5 movaps [esp + nb300_dz], xmm6 ;# 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, [esp + nb300_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb300_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv mulps xmm4, xmm0 ;# xmm4=r mulps xmm4, [esp + nb300_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 esi, [ebp + nb300_VFtab] movd ecx, mm6 psrlq mm6, 32 movd edx, mm6 movlps xmm5, [esi + ecx*4] movhps xmm5, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm4, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + ecx*4 + 8] movhps xmm7, [esi + edx*4 + 8] movaps xmm6, xmm7 shufps xmm6, xmm6, 136 ;# constant 10001000 shufps xmm7, xmm7, 221 ;# constant 11011101 ;# table ready in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb300_two] ;# two*Heps2 movaps xmm3, [esp + nb300_qq] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC ;# increment vcoul - then we can get rid of mm5 ;# update vctot addps xmm5, [esp + nb300_vctot] movaps [esp + nb300_vctot], xmm5 xorps xmm4, xmm4 mulps xmm3, [esp + nb300_tsc] mulps xmm3, xmm0 subps xmm4, xmm3 movaps xmm0, [esp + nb300_dx] movaps xmm1, [esp + nb300_dy] movaps xmm2, [esp + nb300_dz] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb300_fix] movaps xmm4, [esp + nb300_fiy] movaps xmm5, [esp + nb300_fiz] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [esp + nb300_fix], xmm3 movaps [esp + nb300_fiy], xmm4 movaps [esp + nb300_fiz], xmm5 ;# update the fj's movss xmm3, [edi + eax*4] movss xmm4, [edi + eax*4 + 4] movss xmm5, [edi + eax*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + eax*4], xmm3 movss [edi + eax*4 + 4], xmm4 movss [edi + eax*4 + 8], xmm5 shufps xmm0, xmm0, 225 ;# constant 11100001 shufps xmm1, xmm1, 225 ;# constant 11100001 shufps xmm2, xmm2, 225 ;# constant 11100001 movss xmm3, [edi + ebx*4] movss xmm4, [edi + ebx*4 + 4] movss xmm5, [edi + ebx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + ebx*4], xmm3 movss [edi + ebx*4 + 4], xmm4 movss [edi + ebx*4 + 8], xmm5 .nb300_checksingle: mov edx, [esp + nb300_innerk] and edx, 1 jnz .nb300_dosingle jmp .nb300_updateouterdata.nb300_dosingle: mov esi, [ebp + nb300_charge] mov edi, [ebp + nb300_pos] mov ecx, [esp + nb300_innerjjnr] mov eax, [ecx] xorps xmm6, xmm6 movss xmm6, [esi + eax*4] ;# xmm6(0) has the charge mulps xmm6, [esp + nb300_iq] movaps [esp + nb300_qq], xmm6 lea eax, [eax + eax*2] ;# move coordinates to xmm0-xmm2 movss xmm0, [edi + eax*4] movss xmm1, [edi + eax*4 + 4] movss xmm2, [edi + eax*4 + 8] movaps xmm4, [esp + nb300_ix] movaps xmm5, [esp + nb300_iy] movaps xmm6, [esp + nb300_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb300_dx], xmm4 movaps [esp + nb300_dy], xmm5 movaps [esp + nb300_dz], xmm6 ;# 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, [esp + nb300_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb300_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv mulps xmm4, xmm0 ;# xmm4=r mulps xmm4, [esp + nb300_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 esi, [ebp + nb300_VFtab] movd ebx, mm6 movlps xmm4, [esi + ebx*4] movlps xmm6, [esi + ebx*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 mulps xmm7, [esp + nb300_two] ;# two*Heps2 movaps xmm3, [esp + nb300_qq] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and mm3 fijC ;# increment vcoul - then we can get rid of mm5 ;# update vctot addss xmm5, [esp + nb300_vctot] movss [esp + nb300_vctot], xmm5 xorps xmm4, xmm4 mulps xmm3, [esp + nb300_tsc] mulps xmm3, xmm0 subps xmm4, xmm3 mov edi, [ebp + nb300_faction] movaps xmm0, [esp + nb300_dx] movaps xmm1, [esp + nb300_dy] movaps xmm2, [esp + nb300_dz] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb300_fix] movaps xmm4, [esp + nb300_fiy] movaps xmm5, [esp + nb300_fiz] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movaps [esp + nb300_fix], xmm3 movaps [esp + nb300_fiy], xmm4 movaps [esp + nb300_fiz], xmm5 ;# update fj movss xmm3, [edi + eax*4] movss xmm4, [edi + eax*4 + 4] movss xmm5, [edi + eax*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + eax*4], xmm3 movss [edi + eax*4 + 4], xmm4 movss [edi + eax*4 + 8], xmm5 .nb300_updateouterdata: mov ecx, [esp + nb300_ii3] mov edi, [ebp + nb300_faction] mov esi, [ebp + nb300_fshift] mov edx, [esp + nb300_is3] ;# accumulate i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb300_fix] movaps xmm1, [esp + nb300_fiy] movaps xmm2, [esp + nb300_fiz] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 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, [edi + ecx*4] movss xmm4, [edi + ecx*4 + 4] movss xmm5, [edi + ecx*4 + 8] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4], xmm3 movss [edi + ecx*4 + 4], xmm4 movss [edi + ecx*4 + 8], xmm5 ;# increment fshift force movss xmm3, [esi + edx*4] movss xmm4, [esi + edx*4 + 4] movss xmm5, [esi + edx*4 + 8] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [esi + edx*4], xmm3 movss [esi + edx*4 + 4], xmm4 movss [esi + edx*4 + 8], xmm5 ;# get n from stack mov esi, [esp + nb300_n] ;# get group index for i particle mov edx, [ebp + nb300_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [esp + nb300_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 eax, [ebp + nb300_Vc] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb300_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb300_outerend ;# not last, iterate outer loop once more! mov [esp + nb300_n], esi jmp .nb300_outer.nb300_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb300_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb300_end ;# non-zero, do one more workunit jmp .nb300_threadloop.nb300_end: emms mov eax, [esp + nb300_nouter] mov ebx, [esp + nb300_ninner] mov ecx, [ebp + nb300_outeriter] mov edx, [ebp + nb300_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb300_salign] add esp, eax add esp, 316 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret .globl nb_kernel300nf_ia32_sse.globl _nb_kernel300nf_ia32_ssenb_kernel300nf_ia32_sse: _nb_kernel300nf_ia32_sse: .equiv nb300nf_p_nri, 8.equiv nb300nf_iinr, 12.equiv nb300nf_jindex, 16.equiv nb300nf_jjnr, 20.equiv nb300nf_shift, 24.equiv nb300nf_shiftvec, 28.equiv nb300nf_fshift, 32.equiv nb300nf_gid, 36.equiv nb300nf_pos, 40.equiv nb300nf_faction, 44.equiv nb300nf_charge, 48.equiv nb300nf_p_facel, 52.equiv nb300nf_argkrf, 56.equiv nb300nf_argcrf, 60.equiv nb300nf_Vc, 64.equiv nb300nf_type, 68.equiv nb300nf_p_ntype, 72.equiv nb300nf_vdwparam, 76.equiv nb300nf_Vvdw, 80.equiv nb300nf_p_tabscale, 84.equiv nb300nf_VFtab, 88.equiv nb300nf_invsqrta, 92.equiv nb300nf_dvda, 96.equiv nb300nf_p_gbtabscale, 100.equiv nb300nf_GBtab, 104.equiv nb300nf_p_nthreads, 108.equiv nb300nf_count, 112.equiv nb300nf_mtx, 116.equiv nb300nf_outeriter, 120.equiv nb300nf_inneriter, 124.equiv nb300nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb300nf_ix, 0.equiv nb300nf_iy, 16.equiv nb300nf_iz, 32.equiv nb300nf_iq, 48.equiv nb300nf_tsc, 64.equiv nb300nf_qq, 80.equiv nb300nf_vctot, 96.equiv nb300nf_half, 112.equiv nb300nf_three, 128.equiv nb300nf_is3, 144.equiv nb300nf_ii3, 148.equiv nb300nf_innerjjnr, 152.equiv nb300nf_innerk, 156.equiv nb300nf_n, 160.equiv nb300nf_nn1, 164.equiv nb300nf_nri, 168.equiv nb300nf_facel, 172.equiv nb300nf_nouter, 176.equiv nb300nf_ninner, 180.equiv nb300nf_salign, 184 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 188 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb300nf_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb300nf_p_nri] mov esi, [ebp + nb300nf_p_facel] mov ecx, [ecx] mov esi, [esi] mov [esp + nb300nf_nri], ecx mov [esp + nb300nf_facel], esi ;# zero iteration counters mov eax, 0 mov [esp + nb300nf_nouter], eax mov [esp + nb300nf_ninner], eax mov eax, [ebp + nb300nf_p_tabscale] movss xmm3, [eax] shufps xmm3, xmm3, 0 movaps [esp + nb300nf_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb300nf_half], eax movss xmm1, [esp + nb300nf_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# constant 1.0 movaps xmm3, xmm2 addps xmm2, xmm2 ;# constant 2.0 addps xmm3, xmm2 ;# constant 3.0 movaps [esp + nb300nf_half], xmm1 movaps [esp + nb300nf_three], xmm3.nb300nf_threadloop: mov esi, [ebp + nb300nf_count] ;# pointer to sync counter⌨️ 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?