nb_kernel314_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,319 行 · 第 1/5 页
S
2,319 行
;#;# $Id: nb_kernel314_x86_64_sse.intel_syntax.s,v 1.1.2.2 2006/09/22 08:32:51 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_kernel314_x86_64_sse.globl _nb_kernel314_x86_64_ssenb_kernel314_x86_64_sse: _nb_kernel314_x86_64_sse: ;# Room for return address and rbp (16 bytes).equiv nb314_fshift, 16.equiv nb314_gid, 24.equiv nb314_pos, 32.equiv nb314_faction, 40.equiv nb314_charge, 48.equiv nb314_p_facel, 56.equiv nb314_argkrf, 64.equiv nb314_argcrf, 72.equiv nb314_Vc, 80.equiv nb314_type, 88.equiv nb314_p_ntype, 96.equiv nb314_vdwparam, 104.equiv nb314_Vvdw, 112.equiv nb314_p_tabscale, 120.equiv nb314_VFtab, 128.equiv nb314_invsqrta, 136.equiv nb314_dvda, 144.equiv nb314_p_gbtabscale, 152.equiv nb314_GBtab, 160.equiv nb314_p_nthreads, 168.equiv nb314_count, 176.equiv nb314_mtx, 184.equiv nb314_outeriter, 192.equiv nb314_inneriter, 200.equiv nb314_work, 208 ;# bottom of stack is cache-aligned for sse use .equiv nb314_ixO, 0.equiv nb314_iyO, 16.equiv nb314_izO, 32.equiv nb314_ixH1, 48.equiv nb314_iyH1, 64.equiv nb314_izH1, 80.equiv nb314_ixH2, 96.equiv nb314_iyH2, 112.equiv nb314_izH2, 128.equiv nb314_ixM, 144.equiv nb314_iyM, 160.equiv nb314_izM, 176.equiv nb314_jxO, 192.equiv nb314_jyO, 208.equiv nb314_jzO, 224.equiv nb314_jxH1, 240.equiv nb314_jyH1, 256.equiv nb314_jzH1, 272.equiv nb314_jxH2, 288.equiv nb314_jyH2, 304.equiv nb314_jzH2, 320.equiv nb314_jxM, 336.equiv nb314_jyM, 352.equiv nb314_jzM, 368.equiv nb314_dxOO, 384.equiv nb314_dyOO, 400.equiv nb314_dzOO, 416.equiv nb314_dxH1H1, 432.equiv nb314_dyH1H1, 448.equiv nb314_dzH1H1, 464.equiv nb314_dxH1H2, 480.equiv nb314_dyH1H2, 496.equiv nb314_dzH1H2, 512.equiv nb314_dxH1M, 528.equiv nb314_dyH1M, 544.equiv nb314_dzH1M, 560.equiv nb314_dxH2H1, 576.equiv nb314_dyH2H1, 592.equiv nb314_dzH2H1, 608.equiv nb314_dxH2H2, 624.equiv nb314_dyH2H2, 640.equiv nb314_dzH2H2, 656.equiv nb314_dxH2M, 672.equiv nb314_dyH2M, 688.equiv nb314_dzH2M, 704.equiv nb314_dxMH1, 720.equiv nb314_dyMH1, 736.equiv nb314_dzMH1, 752.equiv nb314_dxMH2, 768.equiv nb314_dyMH2, 784.equiv nb314_dzMH2, 800.equiv nb314_dxMM, 816.equiv nb314_dyMM, 832.equiv nb314_dzMM, 848.equiv nb314_qqMM, 864.equiv nb314_qqMH, 880.equiv nb314_qqHH, 896.equiv nb314_two, 912.equiv nb314_tsc, 928.equiv nb314_c6, 944.equiv nb314_c12, 960.equiv nb314_six, 976.equiv nb314_twelve, 992.equiv nb314_vctot, 1008.equiv nb314_Vvdwtot, 1024.equiv nb314_fixO, 1040.equiv nb314_fiyO, 1056.equiv nb314_fizO, 1072.equiv nb314_fixH1, 1088.equiv nb314_fiyH1, 1104.equiv nb314_fizH1, 1120.equiv nb314_fixH2, 1136.equiv nb314_fiyH2, 1152.equiv nb314_fizH2, 1168.equiv nb314_fixM, 1184.equiv nb314_fiyM, 1200.equiv nb314_fizM, 1216.equiv nb314_fjxO, 1232.equiv nb314_fjyO, 1248.equiv nb314_fjzO, 1264.equiv nb314_fjxH1, 1280.equiv nb314_fjyH1, 1296.equiv nb314_fjzH1, 1312.equiv nb314_fjxH2, 1328.equiv nb314_fjyH2, 1344.equiv nb314_fjzH2, 1360.equiv nb314_epsH1, 1376.equiv nb314_epsH2, 1392.equiv nb314_epsM, 1408.equiv nb314_half, 1424.equiv nb314_three, 1440.equiv nb314_rsqOO, 1456.equiv nb314_rsqH1H1, 1472.equiv nb314_rsqH1H2, 1488.equiv nb314_rsqH1M, 1504.equiv nb314_rsqH2H1, 1520.equiv nb314_rsqH2H2, 1536.equiv nb314_rsqH2M, 1552.equiv nb314_rsqMH1, 1568.equiv nb314_rsqMH2, 1584.equiv nb314_rsqMM, 1600.equiv nb314_rinvsqOO, 1616.equiv nb314_rinvH1H1, 1632.equiv nb314_rinvH1H2, 1648.equiv nb314_rinvH1M, 1664.equiv nb314_rinvH2H1, 1680.equiv nb314_rinvH2H2, 1696.equiv nb314_rinvH2M, 1712.equiv nb314_rinvMH1, 1728.equiv nb314_rinvMH2, 1744.equiv nb314_rinvMM, 1760.equiv nb314_fstmp, 1776.equiv nb314_is3, 1792.equiv nb314_ii3, 1796.equiv nb314_nri, 1800.equiv nb314_iinr, 1808.equiv nb314_jindex, 1816.equiv nb314_jjnr, 1824.equiv nb314_shift, 1832.equiv nb314_shiftvec, 1840.equiv nb314_facel, 1848.equiv nb314_innerjjnr, 1856.equiv nb314_innerk, 1864.equiv nb314_n, 1868.equiv nb314_nn1, 1872.equiv nb314_nouter, 1876.equiv nb314_ninner, 1880 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1896 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb314_nouter], eax mov [rsp + nb314_ninner], eax mov edi, [rdi] mov [rsp + nb314_nri], edi mov [rsp + nb314_iinr], rsi mov [rsp + nb314_jindex], rdx mov [rsp + nb314_jjnr], rcx mov [rsp + nb314_shift], r8 mov [rsp + nb314_shiftvec], r9 mov rsi, [rbp + nb314_p_facel] movss xmm0, [rsi] movss [rsp + nb314_facel], xmm0 mov rax, [rbp + nb314_p_tabscale] movss xmm3, [rax] shufps xmm3, xmm3, 0 movaps [rsp + nb314_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# half in IEEE (hex) mov [rsp + nb314_half], eax movss xmm1, [rsp + nb314_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 xmm4, xmm3 addps xmm4, xmm4 ;# six movaps xmm5, xmm4 addps xmm5, xmm5 ;# twelve movaps [rsp + nb314_half], xmm1 movaps [rsp + nb314_two], xmm2 movaps [rsp + nb314_three], xmm3 movaps [rsp + nb314_six], xmm4 movaps [rsp + nb314_twelve], xmm5 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb314_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb314_charge] movss xmm5, [rdx + rbx*4 + 4] movss xmm3, [rdx + rbx*4 + 12] movss xmm4, xmm3 mov rsi, [rbp + nb314_p_facel] movss xmm0, [rsi] movss xmm6, [rsp + nb314_facel] mulss xmm3, xmm3 mulss xmm4, xmm5 mulss xmm5, xmm5 mulss xmm3, xmm6 mulss xmm4, xmm6 mulss xmm5, xmm6 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb314_qqMM], xmm3 movaps [rsp + nb314_qqMH], xmm4 movaps [rsp + nb314_qqHH], xmm5 xorps xmm0, xmm0 mov rdx, [rbp + nb314_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov edx, ecx mov rdi, [rbp + nb314_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] add edx, ecx mov rax, [rbp + nb314_vdwparam] movlps xmm0, [rax + rdx*4] movaps xmm1, xmm0 shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0x55 movaps [rsp + nb314_c6], xmm0 movaps [rsp + nb314_c12], xmm1.nb314_threadloop: mov rsi, [rbp + nb314_count] ;# pointer to sync counter mov eax, [rsi].nb314_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 .nb314_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb314_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 [rsp + nb314_n], eax mov [rsp + nb314_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb314_outerstart jmp .nb314_end .nb314_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb314_nouter] mov [rsp + nb314_nouter], ebx.nb314_outer: mov rax, [rsp + nb314_shift] ;# rax = pointer into shift[] mov ebx, [rax + rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb314_is3],ebx ;# store is3 mov rax, [rsp + nb314_shiftvec] ;# rax = base of shiftvec[] movss xmm0, [rax + rbx*4] movss xmm1, [rax + rbx*4 + 4] movss xmm2, [rax + rbx*4 + 8] mov rcx, [rsp + nb314_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx + rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb314_pos] ;# rax = base of pos[] mov [rsp + nb314_ii3], ebx movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 movaps xmm6, xmm0 movaps xmm7, xmm1 addss xmm3, [rax + rbx*4] ;# ox addss xmm4, [rax + rbx*4 + 4] ;# oy addss xmm5, [rax + rbx*4 + 8] ;# oz addss xmm6, [rax + rbx*4 + 12] ;# h1x addss xmm7, [rax + rbx*4 + 16] ;# h1y shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 shufps xmm6, xmm6, 0 shufps xmm7, xmm7, 0 movaps [rsp + nb314_ixO], xmm3 movaps [rsp + nb314_iyO], xmm4 movaps [rsp + nb314_izO], xmm5 movaps [rsp + nb314_ixH1], xmm6 movaps [rsp + nb314_iyH1], xmm7 movss xmm6, xmm2 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm6, [rax + rbx*4 + 20] ;# h1z addss xmm0, [rax + rbx*4 + 24] ;# h2x addss xmm1, [rax + rbx*4 + 28] ;# h2y addss xmm2, [rax + rbx*4 + 32] ;# h2z addss xmm3, [rax + rbx*4 + 36] ;# mx addss xmm4, [rax + rbx*4 + 40] ;# my addss xmm5, [rax + rbx*4 + 44] ;# mz shufps xmm6, xmm6, 0 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 [rsp + nb314_izH1], xmm6 movaps [rsp + nb314_ixH2], xmm0 movaps [rsp + nb314_iyH2], xmm1 movaps [rsp + nb314_izH2], xmm2 movaps [rsp + nb314_ixM], xmm3 movaps [rsp + nb314_iyM], xmm4 movaps [rsp + nb314_izM], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [rsp + nb314_vctot], xmm4 movaps [rsp + nb314_Vvdwtot], xmm4 movaps [rsp + nb314_fixO], xmm4 movaps [rsp + nb314_fiyO], xmm4 movaps [rsp + nb314_fizO], xmm4 movaps [rsp + nb314_fixH1], xmm4 movaps [rsp + nb314_fiyH1], xmm4 movaps [rsp + nb314_fizH1], xmm4 movaps [rsp + nb314_fixH2], xmm4 movaps [rsp + nb314_fiyH2], xmm4 movaps [rsp + nb314_fizH2], xmm4 movaps [rsp + nb314_fixM], xmm4 movaps [rsp + nb314_fiyM], xmm4 movaps [rsp + nb314_fizM], xmm4 mov rax, [rsp + nb314_jindex] mov ecx, [rax + rsi*4] ;# jindex[n] mov edx, [rax + rsi*4 + 4] ;# jindex[n+1] sub edx, ecx ;# number of innerloop atoms mov rsi, [rbp + nb314_pos] mov rdi, [rbp + nb314_faction] mov rax, [rsp + nb314_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb314_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [rsp + nb314_ninner] mov [rsp + nb314_ninner], ecx add edx, 0 mov [rsp + nb314_innerk], edx ;# number of innerloop atoms jge .nb314_unroll_loop jmp .nb314_single_check.nb314_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb314_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] mov ecx, [rdx + 8] mov edx, [rdx + 12] ;# eax-edx=jnr1-4 add qword ptr [rsp + nb314_innerjjnr], 16 ;# advance pointer (unroll 4) mov rsi, [rbp + nb314_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] lea rcx, [rcx + rcx*2] ;# replace jnr with j3 lea rdx, [rdx + rdx*2] ;# load j O coordinates movlps xmm0, [rsi + rax*4] ;# jxOa jyOa - - movlps xmm1, [rsi + rcx*4] ;# jxOc jyOc - - movhps xmm0, [rsi + rbx*4] ;# jxOa jyOa jxOb jyOb movhps xmm1, [rsi + rdx*4] ;# jxOc jyOc jxOd jyOd movss xmm2, [rsi + rax*4 + 8] ;# jzOa - - - movss xmm3, [rsi + rcx*4 + 8] ;# jzOc - - - movhps xmm2, [rsi + rbx*4 + 8] ;# jzOa - jzOb - movhps xmm3, [rsi + rdx*4 + 8] ;# jzOc - jzOd -⌨️ 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?