nb_kernel201_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,605 行 · 第 1/3 页
S
1,605 行
;#;# $Id: nb_kernel201_ia32_sse.intel_syntax.s,v 1.1.2.1 2006/03/01 15:18:29 lindahl Exp $;#;# Gromacs 4.0 Copyright (c) 1991-2003 ;# David van der Spoel, Erik Lindahl;#;# This program is free software; you can redistribute it and/or;# modify it under the terms of the GNU General Public License;# as published by the Free Software Foundation; either version 2;# of the License, or (at your option) any later version.;#;# To help us fund GROMACS development, we humbly ask that you cite;# the research papers on the package. Check out http://www.gromacs.org;# ;# And Hey:;# Gnomes, ROck Monsters And Chili Sauce;#;# These files require GNU binutils 2.10 or later, since we;# use intel syntax for portability, or a recent version ;# of NASM that understands Extended 3DNow and SSE2 instructions.;# (NASM is normally only used with MS Visual C++).;# Since NASM and gnu as disagree on some definitions and use ;# completely different preprocessing options I have to introduce a;# trick: NASM uses ';' for comments, while gnu as uses '#' on x86.;# Gnu as treats ';' as a line break, i.e. ignores it. This is the;# reason why all comments need both symbols...;# The source is written for GNU as, with intel syntax. When you use;# NASM we redefine a couple of things. The false if-statement around ;# the following code is seen by GNU as, but NASM doesn't see it, so ;# the code inside is read by NASM but not gcc.; .if 0 # block below only read by NASM%define .section section%define .long dd%define .align align%define .globl global;# NASM only wants 'dword', not 'dword ptr'.%define ptr.equiv .equiv 2 %1 equ %2%endmacro; .endif # End of NASM-specific block; .intel_syntax noprefix # Line only read by gnu as .globl nb_kernel201_ia32_sse.globl _nb_kernel201_ia32_ssenb_kernel201_ia32_sse: _nb_kernel201_ia32_sse: .equiv nb201_p_nri, 8.equiv nb201_iinr, 12.equiv nb201_jindex, 16.equiv nb201_jjnr, 20.equiv nb201_shift, 24.equiv nb201_shiftvec, 28.equiv nb201_fshift, 32.equiv nb201_gid, 36.equiv nb201_pos, 40.equiv nb201_faction, 44.equiv nb201_charge, 48.equiv nb201_p_facel, 52.equiv nb201_argkrf, 56.equiv nb201_argcrf, 60.equiv nb201_Vc, 64.equiv nb201_type, 68.equiv nb201_p_ntype, 72.equiv nb201_vdwparam, 76.equiv nb201_Vvdw, 80.equiv nb201_p_tabscale, 84.equiv nb201_VFtab, 88.equiv nb201_invsqrta, 92.equiv nb201_dvda, 96.equiv nb201_p_gbtabscale, 100.equiv nb201_GBtab, 104.equiv nb201_p_nthreads, 108.equiv nb201_count, 112.equiv nb201_mtx, 116.equiv nb201_outeriter, 120.equiv nb201_inneriter, 124.equiv nb201_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb201_ixO, 0.equiv nb201_iyO, 16.equiv nb201_izO, 32.equiv nb201_ixH1, 48.equiv nb201_iyH1, 64.equiv nb201_izH1, 80.equiv nb201_ixH2, 96.equiv nb201_iyH2, 112.equiv nb201_izH2, 128.equiv nb201_iqO, 144.equiv nb201_iqH, 160.equiv nb201_dxO, 176.equiv nb201_dyO, 192.equiv nb201_dzO, 208.equiv nb201_dxH1, 224.equiv nb201_dyH1, 240.equiv nb201_dzH1, 256.equiv nb201_dxH2, 272.equiv nb201_dyH2, 288.equiv nb201_dzH2, 304.equiv nb201_qqO, 320.equiv nb201_qqH, 336.equiv nb201_vctot, 352.equiv nb201_fixO, 384.equiv nb201_fiyO, 400.equiv nb201_fizO, 416.equiv nb201_fixH1, 432.equiv nb201_fiyH1, 448.equiv nb201_fizH1, 464.equiv nb201_fixH2, 480.equiv nb201_fiyH2, 496.equiv nb201_fizH2, 512.equiv nb201_fjx, 528.equiv nb201_fjy, 544.equiv nb201_fjz, 560.equiv nb201_half, 576.equiv nb201_three, 592.equiv nb201_two, 608.equiv nb201_krf, 624.equiv nb201_crf, 640.equiv nb201_krsqO, 656.equiv nb201_krsqH1, 672.equiv nb201_krsqH2, 688.equiv nb201_is3, 704.equiv nb201_ii3, 708.equiv nb201_innerjjnr, 712.equiv nb201_innerk, 716.equiv nb201_n, 720.equiv nb201_nn1, 724.equiv nb201_nri, 728.equiv nb201_nouter, 732.equiv nb201_ninner, 736.equiv nb201_salign, 740 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 744 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb201_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb201_p_nri] mov ecx, [ecx] mov [esp + nb201_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb201_nouter], eax mov [esp + nb201_ninner], eax mov esi, [ebp + nb201_argkrf] mov edi, [ebp + nb201_argcrf] movss xmm5, [esi] movss xmm6, [edi] shufps xmm5, xmm5, 0 shufps xmm6, xmm6, 0 movaps [esp + nb201_krf], xmm5 movaps [esp + nb201_crf], xmm6 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb201_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb201_charge] movss xmm3, [edx + ebx*4] movss xmm4, [edx + ebx*4 + 4] mov esi, [ebp + nb201_p_facel] movss xmm5, [esi] mulss xmm3, xmm5 mulss xmm4, xmm5 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 movaps [esp + nb201_iqO], xmm3 movaps [esp + nb201_iqH], xmm4 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb201_half], eax movss xmm1, [esp + nb201_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 + nb201_half], xmm1 movaps [esp + nb201_two], xmm2 movaps [esp + nb201_three], xmm3.nb201_threadloop: mov esi, [ebp + nb201_count] ;# pointer to sync counter mov eax, [esi].nb201_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 .nb201_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb201_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 + nb201_n], eax mov [esp + nb201_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb201_outerstart jmp .nb201_end .nb201_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb201_nouter] mov [esp + nb201_nouter], ebx.nb201_outer: mov eax, [ebp + nb201_shift] ;# eax = pointer into shift[] mov ebx, [eax+esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb201_is3],ebx ;# store is3 mov eax, [ebp + nb201_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 + nb201_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx+esi*4] ;# ebx =ii movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb201_pos] ;# eax = base of pos[] mov [esp + nb201_ii3], ebx addss xmm3, [eax + ebx*4] addss xmm4, [eax + ebx*4 + 4] addss xmm5, [eax + ebx*4 + 8] shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb201_ixO], xmm3 movaps [esp + nb201_iyO], xmm4 movaps [esp + nb201_izO], xmm5 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm0, [eax + ebx*4 + 12] addss xmm1, [eax + ebx*4 + 16] addss xmm2, [eax + ebx*4 + 20] addss xmm3, [eax + ebx*4 + 24] addss xmm4, [eax + ebx*4 + 28] addss xmm5, [eax + ebx*4 + 32] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb201_ixH1], xmm0 movaps [esp + nb201_iyH1], xmm1 movaps [esp + nb201_izH1], xmm2 movaps [esp + nb201_ixH2], xmm3 movaps [esp + nb201_iyH2], xmm4 movaps [esp + nb201_izH2], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [esp + nb201_vctot], xmm4 movaps [esp + nb201_fixO], xmm4 movaps [esp + nb201_fiyO], xmm4 movaps [esp + nb201_fizO], xmm4 movaps [esp + nb201_fixH1], xmm4 movaps [esp + nb201_fiyH1], xmm4 movaps [esp + nb201_fizH1], xmm4 movaps [esp + nb201_fixH2], xmm4 movaps [esp + nb201_fiyH2], xmm4 movaps [esp + nb201_fizH2], xmm4 mov eax, [ebp + nb201_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 + nb201_pos] mov edi, [ebp + nb201_faction] mov eax, [ebp + nb201_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb201_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb201_ninner] mov [esp + nb201_ninner], ecx add edx, 0 mov [esp + nb201_innerk], edx ;# number of innerloop atoms jge .nb201_unroll_loop jmp .nb201_odd_inner.nb201_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb201_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov ebx, [edx + 4] mov ecx, [edx + 8] mov edx, [edx + 12] ;# eax-edx=jnr1-4 add dword ptr [esp + nb201_innerjjnr], 16 ;# advance pointer (unrolled 4) mov esi, [ebp + nb201_charge] ;# base of charge[] movss xmm3, [esi + eax*4] movss xmm4, [esi + ecx*4] movss xmm6, [esi + ebx*4] movss xmm7, [esi + edx*4] shufps xmm3, xmm6, 0 shufps xmm4, xmm7, 0 shufps xmm3, xmm4, 136 ;# constant 10001000 ;# all charges in xmm3 movaps xmm4, xmm3 ;# and in xmm4 mulps xmm3, [esp + nb201_iqO] mulps xmm4, [esp + nb201_iqH] movaps [esp + nb201_qqO], xmm3 movaps [esp + nb201_qqH], xmm4 mov esi, [ebp + nb201_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 lea ebx, [ebx + ebx*2] 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 ixO-izO to xmm4-xmm6 movaps xmm4, [esp + nb201_ixO] movaps xmm5, [esp + nb201_iyO] movaps xmm6, [esp + nb201_izO] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb201_dxO], xmm4 movaps [esp + nb201_dyO], xmm5 movaps [esp + nb201_dzO], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 movaps xmm7, xmm4 ;# rsqO in xmm7 ;# move ixH1-izH1 to xmm4-xmm6 movaps xmm4, [esp + nb201_ixH1] movaps xmm5, [esp + nb201_iyH1] movaps xmm6, [esp + nb201_izH1] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb201_dxH1], xmm4 movaps [esp + nb201_dyH1], xmm5 movaps [esp + nb201_dzH1], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm6, xmm5 addps xmm6, xmm4 ;# rsqH1 in xmm6 ;# move ixH2-izH2 to xmm3-xmm5 movaps xmm3, [esp + nb201_ixH2] movaps xmm4, [esp + nb201_iyH2] movaps xmm5, [esp + nb201_izH2] ;# calc dr subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# store dr movaps [esp + nb201_dxH2], xmm3 movaps [esp + nb201_dyH2], xmm4 movaps [esp + nb201_dzH2], xmm5 ;# square it mulps xmm3,xmm3 mulps xmm4,xmm4 mulps xmm5,xmm5 addps xmm5, xmm4 addps xmm5, xmm3 ;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movaps xmm0, xmm5 movaps xmm1, xmm6 movaps xmm2, xmm7 mulps xmm0, [esp + nb201_krf] mulps xmm1, [esp + nb201_krf] mulps xmm2, [esp + nb201_krf] movaps [esp + nb201_krsqH2], xmm0 movaps [esp + nb201_krsqH1], xmm1 movaps [esp + nb201_krsqO], xmm2 ;# start with rsqO - seed in xmm2 rsqrtps xmm2, xmm7 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb201_three] mulps xmm2, xmm7 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb201_half] movaps xmm7, xmm4 ;# rinvO in xmm7 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb201_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb201_half] movaps xmm6, xmm4 ;# rinvH1 in xmm6 ;# rsqH2 - seed in xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb201_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb201_half] movaps xmm5, xmm4 ;# rinvH2 in xmm5 ;# do O interactions movaps xmm4, xmm7 mulps xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq movaps xmm0, xmm7 movaps xmm1, [esp + nb201_krsqO] addps xmm0, xmm1 subps xmm0, [esp + nb201_crf] ;# xmm0=rinv+ krsq-crf mulps xmm1, [esp + nb201_two] subps xmm7, xmm1 mulps xmm0, [esp + nb201_qqO] mulps xmm7, [esp + nb201_qqO] mulps xmm4, xmm7 ;# total fsO in xmm4 addps xmm0, [esp + nb201_vctot] movaps [esp + nb201_vctot], xmm0 movaps xmm0, [esp + nb201_dxO] movaps xmm1, [esp + nb201_dyO] movaps xmm2, [esp + nb201_dzO] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update O forces movaps xmm3, [esp + nb201_fixO] movaps xmm4, [esp + nb201_fiyO] movaps xmm7, [esp + nb201_fizO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb201_fixO], xmm3 movaps [esp + nb201_fiyO], xmm4 movaps [esp + nb201_fizO], xmm7 ;# update j forces with water O ⌨️ 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?