nb_kernel313_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,374 行 · 第 1/5 页
S
2,374 行
;#;# $Id: nb_kernel313_x86_64_sse2.intel_syntax.s,v 1.1.2.2 2006/09/22 08:40:36 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_kernel313_x86_64_sse2.globl _nb_kernel313_x86_64_sse2nb_kernel313_x86_64_sse2: _nb_kernel313_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb313_fshift, 16.equiv nb313_gid, 24.equiv nb313_pos, 32.equiv nb313_faction, 40.equiv nb313_charge, 48.equiv nb313_p_facel, 56.equiv nb313_argkrf, 64.equiv nb313_argcrf, 72.equiv nb313_Vc, 80.equiv nb313_type, 88.equiv nb313_p_ntype, 96.equiv nb313_vdwparam, 104.equiv nb313_Vvdw, 112.equiv nb313_p_tabscale, 120.equiv nb313_VFtab, 128.equiv nb313_invsqrta, 136.equiv nb313_dvda, 144.equiv nb313_p_gbtabscale, 152.equiv nb313_GBtab, 160.equiv nb313_p_nthreads, 168.equiv nb313_count, 176.equiv nb313_mtx, 184.equiv nb313_outeriter, 192.equiv nb313_inneriter, 200.equiv nb313_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb313_ixO, 0.equiv nb313_iyO, 16.equiv nb313_izO, 32.equiv nb313_ixH1, 48.equiv nb313_iyH1, 64.equiv nb313_izH1, 80.equiv nb313_ixH2, 96.equiv nb313_iyH2, 112.equiv nb313_izH2, 128.equiv nb313_ixM, 144.equiv nb313_iyM, 160.equiv nb313_izM, 176.equiv nb313_iqM, 192.equiv nb313_iqH, 208.equiv nb313_dxO, 224.equiv nb313_dyO, 240.equiv nb313_dzO, 256.equiv nb313_dxH1, 272.equiv nb313_dyH1, 288.equiv nb313_dzH1, 304.equiv nb313_dxH2, 320.equiv nb313_dyH2, 336.equiv nb313_dzH2, 352.equiv nb313_dxM, 368.equiv nb313_dyM, 384.equiv nb313_dzM, 400.equiv nb313_qqM, 416.equiv nb313_qqH, 432.equiv nb313_rinvsqO, 448.equiv nb313_rinvH1, 464.equiv nb313_rinvH2, 480.equiv nb313_rinvM, 496.equiv nb313_rO, 512.equiv nb313_rH1, 528.equiv nb313_rH2, 544.equiv nb313_rM, 560.equiv nb313_tsc, 576.equiv nb313_two, 592.equiv nb313_c6, 608.equiv nb313_c12, 624.equiv nb313_six, 640.equiv nb313_twelve, 656.equiv nb313_vctot, 672.equiv nb313_Vvdwtot, 688.equiv nb313_fixO, 704.equiv nb313_fiyO, 720.equiv nb313_fizO, 736.equiv nb313_fixH1, 752.equiv nb313_fiyH1, 768.equiv nb313_fizH1, 784.equiv nb313_fixH2, 800.equiv nb313_fiyH2, 816.equiv nb313_fizH2, 832.equiv nb313_fixM, 848.equiv nb313_fiyM, 864.equiv nb313_fizM, 880.equiv nb313_fjx, 896.equiv nb313_fjy, 912.equiv nb313_fjz, 928.equiv nb313_half, 944.equiv nb313_three, 960.equiv nb313_is3, 976.equiv nb313_ii3, 980.equiv nb313_nri, 984.equiv nb313_iinr, 992.equiv nb313_jindex, 1000.equiv nb313_jjnr, 1008.equiv nb313_shift, 1016.equiv nb313_shiftvec, 1024.equiv nb313_facel, 1032.equiv nb313_innerjjnr, 1040.equiv nb313_ntia, 1048.equiv nb313_innerk, 1052.equiv nb313_n, 1056.equiv nb313_nn1, 1060.equiv nb313_nouter, 1064.equiv nb313_ninner, 1068 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1080 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb313_nouter], eax mov [rsp + nb313_ninner], eax mov edi, [rdi] mov [rsp + nb313_nri], edi mov [rsp + nb313_iinr], rsi mov [rsp + nb313_jindex], rdx mov [rsp + nb313_jjnr], rcx mov [rsp + nb313_shift], r8 mov [rsp + nb313_shiftvec], r9 mov rsi, [rbp + nb313_p_facel] movsd xmm0, [rsi] movsd [rsp + nb313_facel], xmm0 mov rax, [rbp + nb313_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb313_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb313_half], eax mov [rsp + nb313_half+4], ebx movsd xmm1, [rsp + nb313_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 xmm4, xmm3 addpd xmm4, xmm4 ;# six movapd xmm5, xmm4 addpd xmm5, xmm5 ;# twelve movapd [rsp + nb313_half], xmm1 movapd [rsp + nb313_two], xmm2 movapd [rsp + nb313_three], xmm3 movapd [rsp + nb313_six], xmm4 movapd [rsp + nb313_twelve], xmm5 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb313_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb313_charge] movsd xmm3, [rdx + rbx*8 + 8] movsd xmm4, [rdx + rbx*8 + 24] mov rsi, [rbp + nb313_p_facel] movsd xmm0, [rsi] movsd xmm5, [rsp + nb313_facel] mulsd xmm3, xmm5 mulsd xmm4, xmm5 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 movapd [rsp + nb313_iqH], xmm3 movapd [rsp + nb313_iqM], xmm4 mov rdx, [rbp + nb313_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov rdi, [rbp + nb313_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] mov [rsp + nb313_ntia], ecx .nb313_threadloop: mov rsi, [rbp + nb313_count] ;# pointer to sync counter mov eax, [rsi].nb313_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 .nb313_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb313_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 + nb313_n], eax mov [rsp + nb313_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb313_outerstart jmp .nb313_end.nb313_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb313_nouter] mov [rsp + nb313_nouter], ebx.nb313_outer: mov rax, [rsp + nb313_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb313_is3],ebx ;# store is3 mov rax, [rsp + nb313_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 + nb313_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb313_pos] ;# rax = base of pos[] mov [rsp + nb313_ii3], ebx addsd xmm3, [rax + rbx*8] ;# ox addsd xmm4, [rax + rbx*8 + 8] ;# oy addsd xmm5, [rax + rbx*8 + 16] ;# oz addsd xmm6, [rax + rbx*8 + 24] ;# h1x addsd xmm7, [rax + rbx*8 + 32] ;# h1y shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 shufpd xmm6, xmm6, 0 shufpd xmm7, xmm7, 0 movapd [rsp + nb313_ixO], xmm3 movapd [rsp + nb313_iyO], xmm4 movapd [rsp + nb313_izO], xmm5 movapd [rsp + nb313_ixH1], xmm6 movapd [rsp + nb313_iyH1], xmm7 movsd xmm6, xmm2 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm6, [rax + rbx*8 + 40] ;# h1z addsd xmm0, [rax + rbx*8 + 48] ;# h2x addsd xmm1, [rax + rbx*8 + 56] ;# h2y addsd xmm2, [rax + rbx*8 + 64] ;# h2z addsd xmm3, [rax + rbx*8 + 72] ;# mx addsd xmm4, [rax + rbx*8 + 80] ;# my addsd xmm5, [rax + rbx*8 + 88] ;# mz shufpd xmm6, xmm6, 0 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 + nb313_izH1], xmm6 movapd [rsp + nb313_ixH2], xmm0 movapd [rsp + nb313_iyH2], xmm1 movapd [rsp + nb313_izH2], xmm2 movapd [rsp + nb313_ixM], xmm3 movapd [rsp + nb313_iyM], xmm4 movapd [rsp + nb313_izM], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb313_vctot], xmm4 movapd [rsp + nb313_Vvdwtot], xmm4 movapd [rsp + nb313_fixO], xmm4 movapd [rsp + nb313_fiyO], xmm4 movapd [rsp + nb313_fizO], xmm4 movapd [rsp + nb313_fixH1], xmm4 movapd [rsp + nb313_fiyH1], xmm4 movapd [rsp + nb313_fizH1], xmm4 movapd [rsp + nb313_fixH2], xmm4 movapd [rsp + nb313_fiyH2], xmm4 movapd [rsp + nb313_fizH2], xmm4 movapd [rsp + nb313_fixM], xmm4 movapd [rsp + nb313_fiyM], xmm4 movapd [rsp + nb313_fizM], xmm4 mov rax, [rsp + nb313_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 + nb313_pos] mov rdi, [rbp + nb313_faction] mov rax, [rsp + nb313_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb313_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb313_ninner] mov [rsp + nb313_ninner], ecx add edx, 0 mov [rsp + nb313_innerk], edx ;# number of innerloop atoms jge .nb313_unroll_loop jmp .nb313_checksingle.nb313_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb313_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb313_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb313_charge] ;# base of charge[] movlpd xmm3, [rsi + rax*8] movhpd xmm3, [rsi + rbx*8] movapd xmm4, xmm3 mulpd xmm3, [rsp + nb313_iqM] mulpd xmm4, [rsp + nb313_iqH] movapd [rsp + nb313_qqM], xmm3 movapd [rsp + nb313_qqH], xmm4 mov rsi, [rbp + nb313_type] mov r8d, [rsi + rax*4] mov r9d, [rsi + rbx*4] mov rsi, [rbp + nb313_vdwparam] shl r8d, 1 shl r9d, 1 mov edi, [rsp + nb313_ntia] add r8d, edi add r9d, edi movlpd xmm6, [rsi + r8*8] ;# c6a movlpd xmm7, [rsi + r9*8] ;# c6b movhpd xmm6, [rsi + r8*8 + 8] ;# c6a c12a movhpd xmm7, [rsi + r9*8 + 8] ;# c6b c12b movapd xmm4, xmm6 unpcklpd xmm4, xmm7 unpckhpd xmm6, xmm7 movapd [rsp + nb313_c6], xmm4 movapd [rsp + nb313_c12], xmm6 mov rsi, [rbp + nb313_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# move j 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 = jx ;# xmm1 = jy ;# xmm2 = jz ;# O interaction ;# copy to xmm3-xmm5 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subpd xmm3, [rsp + nb313_ixO] subpd xmm4, [rsp + nb313_iyO] subpd xmm5, [rsp + nb313_izO] movapd xmm13, xmm3 movapd xmm14, xmm4 movapd xmm15, xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm3, xmm4 addpd xmm3, xmm5 ;# calc 1/rsq cvtpd2ps xmm6, xmm3 rcpps xmm6, xmm6 cvtps2pd xmm6, xmm6 ;# lu in low xmm6 ;# 1/x lookup seed in xmm6 movapd xmm4, [rsp + nb313_two] movapd xmm5, xmm3 mulpd xmm3, xmm6 ;# lu*rsq subpd xmm4, xmm3 ;# 2-lu*rsq mulpd xmm6, xmm4 ;# (new lu) movapd xmm4, [rsp + nb313_two] mulpd xmm5, xmm6 ;# lu*rsq subpd xmm4, xmm5 ;# 2-lu*rsq mulpd xmm4, xmm6 ;# xmm4=rinvsq movapd xmm3, xmm4 ;# rinvsq mulpd xmm4, xmm4 ;# rinv4 mulpd xmm4, xmm3 ;# rinv6 movapd xmm5, xmm4 mulpd xmm5, xmm5 ;# rinv12 mulpd xmm4, [rsp + nb313_c6]⌨️ 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?