nb_kernel202_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,206 行 · 第 1/5 页
S
2,206 行
;#;# $Id: nb_kernel202_x86_64_sse.intel_syntax.s,v 1.1.2.2 2006/09/22 08:32:49 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_kernel202_x86_64_sse.globl _nb_kernel202_x86_64_ssenb_kernel202_x86_64_sse: _nb_kernel202_x86_64_sse: .equiv nb202_fshift, 16.equiv nb202_gid, 24.equiv nb202_pos, 32.equiv nb202_faction, 40.equiv nb202_charge, 48.equiv nb202_p_facel, 56.equiv nb202_argkrf, 64.equiv nb202_argcrf, 72.equiv nb202_Vc, 80.equiv nb202_type, 88.equiv nb202_p_ntype, 96.equiv nb202_vdwparam, 104.equiv nb202_Vvdw, 112.equiv nb202_p_tabscale, 120.equiv nb202_VFtab, 128.equiv nb202_invsqrta, 136.equiv nb202_dvda, 144.equiv nb202_p_gbtabscale, 152.equiv nb202_GBtab, 160.equiv nb202_p_nthreads, 168.equiv nb202_count, 176.equiv nb202_mtx, 184.equiv nb202_outeriter, 192.equiv nb202_inneriter, 200.equiv nb202_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb202_ixO, 0.equiv nb202_iyO, 16.equiv nb202_izO, 32.equiv nb202_ixH1, 48.equiv nb202_iyH1, 64.equiv nb202_izH1, 80.equiv nb202_ixH2, 96.equiv nb202_iyH2, 112.equiv nb202_izH2, 128.equiv nb202_jxO, 144.equiv nb202_jyO, 160.equiv nb202_jzO, 176.equiv nb202_jxH1, 192.equiv nb202_jyH1, 208.equiv nb202_jzH1, 224.equiv nb202_jxH2, 240.equiv nb202_jyH2, 256.equiv nb202_jzH2, 272.equiv nb202_dxOO, 288.equiv nb202_dyOO, 304.equiv nb202_dzOO, 320.equiv nb202_dxOH1, 336.equiv nb202_dyOH1, 352.equiv nb202_dzOH1, 368.equiv nb202_dxOH2, 384.equiv nb202_dyOH2, 400.equiv nb202_dzOH2, 416.equiv nb202_dxH1O, 432.equiv nb202_dyH1O, 448.equiv nb202_dzH1O, 464.equiv nb202_dxH1H1, 480.equiv nb202_dyH1H1, 496.equiv nb202_dzH1H1, 512.equiv nb202_dxH1H2, 528.equiv nb202_dyH1H2, 544.equiv nb202_dzH1H2, 560.equiv nb202_dxH2O, 576.equiv nb202_dyH2O, 592.equiv nb202_dzH2O, 608.equiv nb202_dxH2H1, 624.equiv nb202_dyH2H1, 640.equiv nb202_dzH2H1, 656.equiv nb202_dxH2H2, 672.equiv nb202_dyH2H2, 688.equiv nb202_dzH2H2, 704.equiv nb202_qqOO, 720.equiv nb202_qqOH, 736.equiv nb202_qqHH, 752.equiv nb202_vctot, 768.equiv nb202_fixO, 784.equiv nb202_fiyO, 800.equiv nb202_fizO, 816.equiv nb202_fixH1, 832.equiv nb202_fiyH1, 848.equiv nb202_fizH1, 864.equiv nb202_fixH2, 880.equiv nb202_fiyH2, 896.equiv nb202_fizH2, 912.equiv nb202_fjxO, 928.equiv nb202_fjyO, 944.equiv nb202_fjzO, 960.equiv nb202_fjxH1, 976.equiv nb202_fjyH1, 992.equiv nb202_fjzH1, 1008.equiv nb202_fjxH2, 1024.equiv nb202_fjyH2, 1040.equiv nb202_fjzH2, 1056.equiv nb202_half, 1072.equiv nb202_three, 1088.equiv nb202_rsqOO, 1104.equiv nb202_rsqOH1, 1120.equiv nb202_rsqOH2, 1136.equiv nb202_rsqH1O, 1152.equiv nb202_rsqH1H1, 1168.equiv nb202_rsqH1H2, 1184.equiv nb202_rsqH2O, 1200.equiv nb202_rsqH2H1, 1216.equiv nb202_rsqH2H2, 1232.equiv nb202_rinvOO, 1248.equiv nb202_rinvOH1, 1264.equiv nb202_rinvOH2, 1280.equiv nb202_rinvH1O, 1296.equiv nb202_rinvH1H1, 1312.equiv nb202_rinvH1H2, 1328.equiv nb202_rinvH2O, 1344.equiv nb202_rinvH2H1, 1360.equiv nb202_rinvH2H2, 1376.equiv nb202_two, 1392.equiv nb202_krf, 1408.equiv nb202_crf, 1424.equiv nb202_innerjjnr, 1440.equiv nb202_nri, 1448.equiv nb202_iinr, 1456.equiv nb202_jindex, 1464.equiv nb202_jjnr, 1472.equiv nb202_shift, 1480.equiv nb202_shiftvec, 1488.equiv nb202_facel, 1496.equiv nb202_is3, 1504.equiv nb202_ii3, 1508.equiv nb202_innerk, 1512.equiv nb202_n, 1516.equiv nb202_nn1, 1520.equiv nb202_nouter, 1524.equiv nb202_ninner, 1528 push rbp mov rbp, rsp push rbx push r12 push r13 push r14 push r15 sub rsp, 1544 ;# local variable stack space (n*16+8) emms ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb202_nouter], eax mov [rsp + nb202_ninner], eax mov edi, [rdi] mov [rsp + nb202_nri], edi mov [rsp + nb202_iinr], rsi mov [rsp + nb202_jindex], rdx mov [rsp + nb202_jjnr], rcx mov [rsp + nb202_shift], r8 mov [rsp + nb202_shiftvec], r9 mov rsi, [rbp + nb202_p_facel] movss xmm0, [rsi] movss [rsp + nb202_facel], xmm0 mov rsi, [rbp + nb202_argkrf] mov rdi, [rbp + nb202_argcrf] movss xmm1, [rsi] movss xmm2, [rdi] shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movaps [rsp + nb202_krf], xmm1 movaps [rsp + nb202_crf], xmm2 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb202_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb202_charge] movss xmm3, [rdx + rbx*4] movss xmm4, xmm3 movss xmm5, [rdx + rbx*4 + 4] mov rsi, [rbp + nb202_p_facel] movss xmm0, [rsi] movss xmm6, [rsp + nb202_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 + nb202_qqOO], xmm3 movaps [rsp + nb202_qqOH], xmm4 movaps [rsp + nb202_qqHH], xmm5 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# half in IEEE (hex) mov [rsp + nb202_half], eax movss xmm1, [rsp + nb202_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 + nb202_half], xmm1 movaps [rsp + nb202_two], xmm2 movaps [rsp + nb202_three], xmm3.nb202_threadloop: mov rsi, [rbp + nb202_count] ;# pointer to sync counter mov eax, [rsi].nb202_spinlock: mov ebx, eax ;# ebx=*count=nn0 add rbx, 1 ;# rbx=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 .nb202_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb202_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 + nb202_n], eax mov [rsp + nb202_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb202_outerstart jmp .nb202_end .nb202_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb202_nouter] mov [rsp + nb202_nouter], ebx.nb202_outer: mov rax, [rsp + nb202_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb202_is3],ebx ;# store is3 mov rax, [rsp + nb202_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 + nb202_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx + rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb202_pos] ;# rax = base of pos[] mov [rsp + nb202_ii3], ebx movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 addss xmm3, [rax + rbx*4] addss xmm4, [rax + rbx*4 + 4] addss xmm5, [rax + rbx*4 + 8] shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb202_ixO], xmm3 movaps [rsp + nb202_iyO], xmm4 movaps [rsp + nb202_izO], xmm5 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm0, [rax + rbx*4 + 12] addss xmm1, [rax + rbx*4 + 16] addss xmm2, [rax + rbx*4 + 20] addss xmm3, [rax + rbx*4 + 24] addss xmm4, [rax + rbx*4 + 28] addss xmm5, [rax + rbx*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 [rsp + nb202_ixH1], xmm0 movaps [rsp + nb202_iyH1], xmm1 movaps [rsp + nb202_izH1], xmm2 movaps [rsp + nb202_ixH2], xmm3 movaps [rsp + nb202_iyH2], xmm4 movaps [rsp + nb202_izH2], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [rsp + nb202_vctot], xmm4 movaps [rsp + nb202_fixO], xmm4 movaps [rsp + nb202_fiyO], xmm4 movaps [rsp + nb202_fizO], xmm4 movaps [rsp + nb202_fixH1], xmm4 movaps [rsp + nb202_fiyH1], xmm4 movaps [rsp + nb202_fizH1], xmm4 movaps [rsp + nb202_fixH2], xmm4 movaps [rsp + nb202_fiyH2], xmm4 movaps [rsp + nb202_fizH2], xmm4 mov rax, [rsp + nb202_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 + nb202_pos] mov rdi, [rbp + nb202_faction] mov rax, [rsp + nb202_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb202_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [rsp + nb202_ninner] mov [rsp + nb202_ninner], ecx add edx, 0 mov [rsp + nb202_innerk], edx ;# number of innerloop atoms jge .nb202_unroll_loop jmp .nb202_single_check.nb202_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb202_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 + nb202_innerjjnr], 16 ;# advance pointer (unrolled 4) mov rsi, [rbp + nb202_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] ;# move j O coordinates to local temp variables 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 - movaps xmm4, xmm0 unpcklps xmm0, xmm1 ;# jxOa jxOc jyOa jyOc unpckhps xmm4, xmm1 ;# jxOb jxOd jyOb jyOd movaps xmm1, xmm0 unpcklps xmm0, xmm4 ;# x unpckhps xmm1, xmm4 ;# y shufps xmm2, xmm3, 136 ;# 10001000 => jzOa jzOb jzOc jzOd ;# xmm0 = Ox ;# xmm1 = Oy ;# xmm2 = Oz movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 movaps xmm6, xmm0 movaps xmm7, xmm1 movaps xmm8, xmm2 subps xmm0, [rsp + nb202_ixO] subps xmm1, [rsp + nb202_iyO] subps xmm2, [rsp + nb202_izO] subps xmm3, [rsp + nb202_ixH1] subps xmm4, [rsp + nb202_iyH1] subps xmm5, [rsp + nb202_izH1] subps xmm6, [rsp + nb202_ixH2] subps xmm7, [rsp + nb202_iyH2] subps xmm8, [rsp + nb202_izH2] ⌨️ 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?