nb_kernel102_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,311 行 · 第 1/5 页
S
2,311 行
;#;# $Id: nb_kernel102_x86_64_sse2.intel_syntax.s,v 1.1.2.2 2006/09/22 08:40:32 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_kernel102_x86_64_sse2.globl _nb_kernel102_x86_64_sse2nb_kernel102_x86_64_sse2: _nb_kernel102_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb102_fshift, 16.equiv nb102_gid, 24.equiv nb102_pos, 32.equiv nb102_faction, 40.equiv nb102_charge, 48.equiv nb102_p_facel, 56.equiv nb102_argkrf, 64.equiv nb102_argcrf, 72.equiv nb102_Vc, 80.equiv nb102_type, 88.equiv nb102_p_ntype, 96.equiv nb102_vdwparam, 104.equiv nb102_Vvdw, 112.equiv nb102_p_tabscale, 120.equiv nb102_VFtab, 128.equiv nb102_invsqrta, 136.equiv nb102_dvda, 144.equiv nb102_p_gbtabscale, 152.equiv nb102_GBtab, 160.equiv nb102_p_nthreads, 168.equiv nb102_count, 176.equiv nb102_mtx, 184.equiv nb102_outeriter, 192.equiv nb102_inneriter, 200.equiv nb102_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb102_ixO, 0.equiv nb102_iyO, 16.equiv nb102_izO, 32.equiv nb102_ixH1, 48.equiv nb102_iyH1, 64.equiv nb102_izH1, 80.equiv nb102_ixH2, 96.equiv nb102_iyH2, 112.equiv nb102_izH2, 128.equiv nb102_jxO, 144.equiv nb102_jyO, 160.equiv nb102_jzO, 176.equiv nb102_jxH1, 192.equiv nb102_jyH1, 208.equiv nb102_jzH1, 224.equiv nb102_jxH2, 240.equiv nb102_jyH2, 256.equiv nb102_jzH2, 272.equiv nb102_dxOO, 288.equiv nb102_dyOO, 304.equiv nb102_dzOO, 320.equiv nb102_dxOH1, 336.equiv nb102_dyOH1, 352.equiv nb102_dzOH1, 368.equiv nb102_dxOH2, 384.equiv nb102_dyOH2, 400.equiv nb102_dzOH2, 416.equiv nb102_dxH1O, 432.equiv nb102_dyH1O, 448.equiv nb102_dzH1O, 464.equiv nb102_dxH1H1, 480.equiv nb102_dyH1H1, 496.equiv nb102_dzH1H1, 512.equiv nb102_dxH1H2, 528.equiv nb102_dyH1H2, 544.equiv nb102_dzH1H2, 560.equiv nb102_dxH2O, 576.equiv nb102_dyH2O, 592.equiv nb102_dzH2O, 608.equiv nb102_dxH2H1, 624.equiv nb102_dyH2H1, 640.equiv nb102_dzH2H1, 656.equiv nb102_dxH2H2, 672.equiv nb102_dyH2H2, 688.equiv nb102_dzH2H2, 704.equiv nb102_qqOO, 720.equiv nb102_qqOH, 736.equiv nb102_qqHH, 752.equiv nb102_vctot, 768.equiv nb102_fixO, 784.equiv nb102_fiyO, 800.equiv nb102_fizO, 816.equiv nb102_fixH1, 832.equiv nb102_fiyH1, 848.equiv nb102_fizH1, 864.equiv nb102_fixH2, 880.equiv nb102_fiyH2, 896.equiv nb102_fizH2, 912.equiv nb102_fjxO, 928.equiv nb102_fjyO, 944.equiv nb102_fjzO, 960.equiv nb102_fjxH1, 976.equiv nb102_fjyH1, 992.equiv nb102_fjzH1, 1008.equiv nb102_fjxH2, 1024.equiv nb102_fjyH2, 1040.equiv nb102_fjzH2, 1056.equiv nb102_half, 1072.equiv nb102_three, 1088.equiv nb102_rsqOO, 1104.equiv nb102_rsqOH1, 1120.equiv nb102_rsqOH2, 1136.equiv nb102_rsqH1O, 1152.equiv nb102_rsqH1H1, 1168.equiv nb102_rsqH1H2, 1184.equiv nb102_rsqH2O, 1200.equiv nb102_rsqH2H1, 1216.equiv nb102_rsqH2H2, 1232.equiv nb102_rinvOO, 1248.equiv nb102_rinvOH1, 1264.equiv nb102_rinvOH2, 1280.equiv nb102_rinvH1O, 1296.equiv nb102_rinvH1H1, 1312.equiv nb102_rinvH1H2, 1328.equiv nb102_rinvH2O, 1344.equiv nb102_rinvH2H1, 1360.equiv nb102_rinvH2H2, 1376.equiv nb102_is3, 1392.equiv nb102_ii3, 1396.equiv nb102_innerjjnr, 1400.equiv nb102_nri, 1408.equiv nb102_iinr, 1416.equiv nb102_jindex, 1424.equiv nb102_jjnr, 1432.equiv nb102_shift, 1440.equiv nb102_shiftvec, 1448.equiv nb102_facel, 1456.equiv nb102_innerk, 1464.equiv nb102_n, 1468.equiv nb102_nn1, 1472.equiv nb102_nouter, 1476.equiv nb102_ninner, 1480 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1496 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb102_nouter], eax mov [rsp + nb102_ninner], eax mov edi, [rdi] mov [rsp + nb102_nri], edi mov [rsp + nb102_iinr], rsi mov [rsp + nb102_jindex], rdx mov [rsp + nb102_jjnr], rcx mov [rsp + nb102_shift], r8 mov [rsp + nb102_shiftvec], r9 mov rsi, [rbp + nb102_p_facel] movsd xmm0, [rsi] movsd [rsp + nb102_facel], xmm0 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb102_half], eax mov [rsp + nb102_half+4], ebx movsd xmm1, [rsp + nb102_half] shufpd xmm1, xmm1, 0 ;# splat to all elements movapd xmm3, xmm1 addpd xmm3, xmm3 ;# one movapd xmm2, xmm3 addpd xmm2, xmm2 ;# two addpd xmm3, xmm2 ;# three movapd [rsp + nb102_half], xmm1 movapd [rsp + nb102_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb102_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb102_charge] movsd xmm3, [rdx + rbx*8] ;# qO movsd xmm4, xmm3 ;# qO movsd xmm5, [rdx + rbx*8 + 8] ;# qH mov rsi, [rbp + nb102_p_facel] movsd xmm0, [rsi] movsd xmm6, [rsp + nb102_facel] ;# facel mulsd xmm3, xmm3 ;# qO*qO mulsd xmm4, xmm5 ;# qO*qH mulsd xmm5, xmm5 ;# qH*qH mulsd xmm3, xmm6 mulsd xmm4, xmm6 mulsd xmm5, xmm6 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb102_qqOO], xmm3 movapd [rsp + nb102_qqOH], xmm4 movapd [rsp + nb102_qqHH], xmm5.nb102_threadloop: mov rsi, [rbp + nb102_count] ;# pointer to sync counter mov eax, [rsi].nb102_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 .nb102_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb102_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 + nb102_n], eax mov [rsp + nb102_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb102_outerstart jmp .nb102_end.nb102_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb102_nouter] mov [rsp + nb102_nouter], ebx.nb102_outer: mov rax, [rsp + nb102_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb102_is3],ebx ;# store is3 mov rax, [rsp + nb102_shiftvec] ;# rax = base of shiftvec[] movsd xmm0, [rax + rbx*8] movsd xmm1, [rax + rbx*8 + 8] movsd xmm2, [rax + rbx*8 + 16] mov rcx, [rsp + nb102_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb102_pos] ;# rax = base of pos[] mov [rsp + nb102_ii3], ebx movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 addsd xmm3, [rax + rbx*8] addsd xmm4, [rax + rbx*8 + 8] addsd xmm5, [rax + rbx*8 + 16] shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb102_ixO], xmm3 movapd [rsp + nb102_iyO], xmm4 movapd [rsp + nb102_izO], xmm5 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm0, [rax + rbx*8 + 24] addsd xmm1, [rax + rbx*8 + 32] addsd xmm2, [rax + rbx*8 + 40] addsd xmm3, [rax + rbx*8 + 48] addsd xmm4, [rax + rbx*8 + 56] addsd xmm5, [rax + rbx*8 + 64] shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb102_ixH1], xmm0 movapd [rsp + nb102_iyH1], xmm1 movapd [rsp + nb102_izH1], xmm2 movapd [rsp + nb102_ixH2], xmm3 movapd [rsp + nb102_iyH2], xmm4 movapd [rsp + nb102_izH2], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb102_vctot], xmm4 movapd [rsp + nb102_fixO], xmm4 movapd [rsp + nb102_fiyO], xmm4 movapd [rsp + nb102_fizO], xmm4 movapd [rsp + nb102_fixH1], xmm4 movapd [rsp + nb102_fiyH1], xmm4 movapd [rsp + nb102_fizH1], xmm4 movapd [rsp + nb102_fixH2], xmm4 movapd [rsp + nb102_fiyH2], xmm4 movapd [rsp + nb102_fizH2], xmm4 mov rax, [rsp + nb102_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 + nb102_pos] mov rdi, [rbp + nb102_faction] mov rax, [rsp + nb102_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb102_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb102_ninner] mov [rsp + nb102_ninner], ecx add edx, 0 mov [rsp + nb102_innerk], edx ;# number of innerloop atoms jge .nb102_unroll_loop jmp .nb102_checksingle.nb102_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb102_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb102_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb102_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# move j O coordinates to local temp variables movlpd xmm0, [rsi + rax*8] movlpd xmm1, [rsi + rax*8 + 8] movlpd xmm2, [rsi + rax*8 + 16] movhpd xmm0, [rsi + rbx*8] movhpd xmm1, [rsi + rbx*8 + 8] movhpd xmm2, [rsi + rbx*8 + 16] ;# xmm0 = Ox ;# xmm1 = Oy ;# xmm2 = Oz movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 movapd xmm8, xmm2 subpd xmm0, [rsp + nb102_ixO] subpd xmm1, [rsp + nb102_iyO] subpd xmm2, [rsp + nb102_izO] subpd xmm3, [rsp + nb102_ixH1] subpd xmm4, [rsp + nb102_iyH1] subpd xmm5, [rsp + nb102_izH1] subpd xmm6, [rsp + nb102_ixH2] subpd xmm7, [rsp + nb102_iyH2] subpd xmm8, [rsp + nb102_izH2] movapd [rsp + nb102_dxOO], xmm0 movapd [rsp + nb102_dyOO], xmm1 movapd [rsp + nb102_dzOO], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [rsp + nb102_dxH1O], xmm3 movapd [rsp + nb102_dyH1O], xmm4 movapd [rsp + nb102_dzH1O], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 movapd [rsp + nb102_dxH2O], xmm6 movapd [rsp + nb102_dyH2O], xmm7 movapd [rsp + nb102_dzH2O], xmm8 mulpd xmm6, xmm6 mulpd xmm7, xmm7 mulpd xmm8, xmm8 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 addpd xmm6, xmm7 addpd xmm6, xmm8 ;# start doing invsqrt for jO atoms cvtpd2ps xmm1, xmm0 cvtpd2ps xmm4, xmm3 cvtpd2ps xmm7, xmm6 rsqrtps xmm1, xmm1 rsqrtps xmm4, xmm4 rsqrtps xmm7, xmm7 cvtps2pd xmm1, xmm1 cvtps2pd xmm4, xmm4 cvtps2pd xmm7, xmm7 movapd xmm2, xmm1 movapd xmm5, xmm4 movapd xmm8, xmm7 mulpd xmm1, xmm1 ;# lu*lu mulpd xmm4, xmm4 ;# lu*lu mulpd xmm7, xmm7 ;# lu*lu movapd xmm9, [rsp + nb102_three] movapd xmm10, xmm9 movapd xmm11, xmm9 mulpd xmm1, xmm0 ;# rsq*lu*lu mulpd xmm4, xmm3 ;# rsq*lu*lu mulpd xmm7, xmm6 ;# rsq*lu*lu⌨️ 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?