nb_kernel010_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,380 行 · 第 1/3 页
S
1,380 行
;# create constant floating-point factors on stack mov eax, 0x40000000 ;# constant 2.0 in IEEE (hex) mov [esp + nb010nf_two], eax movss xmm1, [esp + nb010nf_two] shufps xmm1, xmm1, 0 ;# splat to all elements movaps [esp + nb010nf_two], xmm1.nb010nf_threadloop: mov esi, [ebp + nb010nf_count] ;# pointer to sync counter mov eax, [esi].nb010nf_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 .nb010nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb010nf_nri] mov edx, ecx sub ecx, ebx cmovle ebx, edx ;# if(nn1>nri) nn1=nri ;# Cleared the spinlock if we got here. ;# eax contains nn0, ebx contains nn1. mov [esp + nb010nf_n], eax mov [esp + nb010nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb010nf_outerstart jmp .nb010nf_end.nb010nf_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb010nf_nouter] mov [esp + nb010nf_nouter], ebx.nb010nf_outer: mov eax, [ebp + nb010nf_shift] ;# eax = base of shift[] mov ebx, [eax +esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb010nf_is3],ebx ;# store is3 mov eax, [ebp + nb010nf_shiftvec] ;# eax = base of shiftvec[] movss xmm0, [eax + ebx*4] movss xmm1, [eax + ebx*4 + 4] movss xmm2, [eax + ebx*4 + 8] mov ecx, [ebp + nb010nf_iinr] ;# ecx = base of iinr[] mov ebx, [ecx + esi*4] ;# ebx =ii mov edx, [ebp + nb010nf_type] mov edx, [edx + ebx*4] imul edx, [esp + nb010nf_ntype] shl edx, 1 mov [esp + nb010nf_ntia], edx lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb010nf_pos] ;# eax = base of pos[] addss xmm0, [eax + ebx*4] addss xmm1, [eax + ebx*4 + 4] addss xmm2, [eax + ebx*4 + 8] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movaps [esp + nb010nf_ix], xmm0 movaps [esp + nb010nf_iy], xmm1 movaps [esp + nb010nf_iz], xmm2 mov [esp + nb010nf_ii3], ebx ;# clear Vvdwtot and i forces xorps xmm4, xmm4 movaps [esp + nb010nf_Vvdwtot], xmm4 mov eax, [ebp + nb010nf_jindex] mov ecx, [eax + esi*4] ;# jindex[n] mov edx, [eax + esi*4 + 4] ;# jindex[n+1] sub edx, ecx ;# number of innerloop atoms mov esi, [ebp + nb010nf_pos] mov eax, [ebp + nb010nf_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb010nf_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb010nf_ninner] mov [esp + nb010nf_ninner], ecx add edx, 0 mov [esp + nb010nf_innerk], edx ;# number of innerloop atoms jge .nb010nf_unroll_loop jmp .nb010nf_finish_inner.nb010nf_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb010nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov ebx, [edx + 4] mov ecx, [edx + 8] mov edx, [edx + 12] ;# eax-edx=jnr1-4 ;# advance pointer (unrolled 4) add dword ptr [esp + nb010nf_innerjjnr], 16 movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movd mm2, ecx movd mm3, edx mov esi, [ebp + nb010nf_type] mov eax, [esi + eax*4] mov ebx, [esi + ebx*4] mov ecx, [esi + ecx*4] mov edx, [esi + edx*4] mov esi, [ebp + nb010nf_vdwparam] shl eax, 1 shl ebx, 1 shl ecx, 1 shl edx, 1 mov edi, [esp + nb010nf_ntia] add eax, edi add ebx, edi add ecx, edi add edx, edi movlps xmm6, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm6, [esi + ebx*4] movhps xmm7, [esi + edx*4] movaps xmm4, xmm6 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm6, xmm7, 221 ;# constant 11011101 movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 movaps [esp + nb010nf_c6], xmm4 movaps [esp + nb010nf_c12], xmm6 mov esi, [ebp + nb010nf_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 lea ebx, [ebx + ebx*2] mulps xmm3, xmm2 lea ecx, [ecx + ecx*2] ;# replace jnr with j3 lea edx, [edx + edx*2] ;# move four coordinates to xmm0-xmm2 movlps xmm4, [esi + eax*4] movlps xmm5, [esi + ecx*4] movss xmm2, [esi + eax*4 + 8] movss xmm6, [esi + ecx*4 + 8] movhps xmm4, [esi + ebx*4] movhps xmm5, [esi + edx*4] movss xmm0, [esi + ebx*4 + 8] movss xmm1, [esi + edx*4 + 8] shufps xmm2, xmm0, 0 shufps xmm6, xmm1, 0 movaps xmm0, xmm4 movaps xmm1, xmm4 shufps xmm2, xmm6, 136 ;# constant 10001000 shufps xmm0, xmm5, 136 ;# constant 10001000 shufps xmm1, xmm5, 221 ;# constant 11011101 ;# move ix-iz to xmm4-xmm6 movaps xmm4, [esp + nb010nf_ix] movaps xmm5, [esp + nb010nf_iy] movaps xmm6, [esp + nb010nf_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 rcpps xmm5, xmm4 ;# constant 1/x lookup seed in xmm5 movaps xmm0, [esp + nb010nf_two] mulps xmm4, xmm5 subps xmm0, xmm4 mulps xmm0, xmm5 ;# xmm0=rinvsq movaps xmm4, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb010nf_c6] mulps xmm2, [esp + nb010nf_c12] movaps xmm5, xmm2 subps xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addps xmm5, [esp + nb010nf_Vvdwtot] movaps [esp + nb010nf_Vvdwtot], xmm5 ;# should we do one more iteration? sub dword ptr [esp + nb010nf_innerk], 4 jl .nb010nf_finish_inner jmp .nb010nf_unroll_loop.nb010nf_finish_inner: ;# check if at least two particles remain add dword ptr [esp + nb010nf_innerk], 4 mov edx, [esp + nb010nf_innerk] and edx, 2 jnz .nb010nf_dopair jmp .nb010nf_checksingle.nb010nf_dopair: mov ecx, [esp + nb010nf_innerjjnr] mov eax, [ecx] mov ebx, [ecx + 4] add dword ptr [esp + nb010nf_innerjjnr], 8 mov esi, [ebp + nb010nf_type] mov ecx, eax mov edx, ebx mov ecx, [esi + ecx*4] mov edx, [esi + edx*4] mov esi, [ebp + nb010nf_vdwparam] shl ecx, 1 shl edx, 1 mov edi, [esp + nb010nf_ntia] add ecx, edi add edx, edi movlps xmm6, [esi + ecx*4] movhps xmm6, [esi + edx*4] mov edi, [ebp + nb010nf_pos] xorps xmm7,xmm7 movaps xmm4, xmm6 shufps xmm4, xmm4, 8 ;# constant 00001000 shufps xmm6, xmm6, 13 ;# constant 00001101 movlhps xmm4, xmm7 movlhps xmm6, xmm7 movaps [esp + nb010nf_c6], xmm4 movaps [esp + nb010nf_c12], xmm6 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 ;# move nb010nf_ix-iz to xmm4-xmm6 xorps xmm7, xmm7 movaps xmm4, [esp + nb010nf_ix] movaps xmm5, [esp + nb010nf_iy] movaps xmm6, [esp + nb010nf_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 rcpps xmm5, xmm4 ;# constant 1/x lookup seed in xmm5 movaps xmm0, [esp + nb010nf_two] mulps xmm4, xmm5 subps xmm0, xmm4 mulps xmm0, xmm5 ;# xmm0=rinvsq movaps xmm4, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb010nf_c6] mulps xmm2, [esp + nb010nf_c12] movaps xmm5, xmm2 subps xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addps xmm5, [esp + nb010nf_Vvdwtot] movaps [esp + nb010nf_Vvdwtot], xmm5.nb010nf_checksingle: mov edx, [esp + nb010nf_innerk] and edx, 1 jnz .nb010nf_dosingle jmp .nb010nf_updateouterdata.nb010nf_dosingle: mov edi, [ebp + nb010nf_pos] mov ecx, [esp + nb010nf_innerjjnr] mov eax, [ecx] mov esi, [ebp + nb010nf_type] mov ecx, eax mov ecx, [esi + ecx*4] mov esi, [ebp + nb010nf_vdwparam] shl ecx, 1 add ecx, [esp + nb010nf_ntia] xorps xmm6, xmm6 movlps xmm6, [esi + ecx*4] movaps xmm4, xmm6 shufps xmm4, xmm4, 252 ;# constant 11111100 shufps xmm6, xmm6, 253 ;# constant 11111101 movaps [esp + nb010nf_c6], xmm4 movaps [esp + nb010nf_c12], 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] xorps xmm7, xmm7 movaps xmm4, [esp + nb010nf_ix] movaps xmm5, [esp + nb010nf_iy] movaps xmm6, [esp + nb010nf_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 rcpps xmm5, xmm4 ;# constant 1/x lookup seed in xmm5 movaps xmm0, [esp + nb010nf_two] mulps xmm4, xmm5 subps xmm0, xmm4 mulps xmm0, xmm5 ;# xmm0=rinvsq movaps xmm4, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb010nf_c6] mulps xmm2, [esp + nb010nf_c12] movaps xmm5, xmm2 subps xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addss xmm5, [esp + nb010nf_Vvdwtot] movss [esp + nb010nf_Vvdwtot], xmm5 .nb010nf_updateouterdata: ;# get n from stack mov esi, [esp + nb010nf_n] ;# get group index for i particle mov edx, [ebp + nb010nf_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total lj energy and update it movaps xmm7, [esp + nb010nf_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 eax, [ebp + nb010nf_Vvdw] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb010nf_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb010nf_outerend ;# not last, iterate outer loop once more! mov [esp + nb010nf_n], esi jmp .nb010nf_outer.nb010nf_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb010nf_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb010nf_end ;# non-zero, do one more workunit jmp .nb010nf_threadloop.nb010nf_end: emms mov eax, [esp + nb010nf_nouter] mov ebx, [esp + nb010nf_ninner] mov ecx, [ebp + nb010nf_outeriter] mov edx, [ebp + nb010nf_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb010nf_salign] add esp, eax add esp, 192 pop edi pop esi pop edx pop ecx pop ebx pop eax leave 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?