nb_kernel134_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,348 行 · 第 1/5 页
S
2,348 行
;#;# $Id: nb_kernel134_x86_64_sse2.intel_syntax.s,v 1.1.2.2 2006/09/22 08:40:33 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_kernel134_x86_64_sse2.globl _nb_kernel134_x86_64_sse2nb_kernel134_x86_64_sse2: _nb_kernel134_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb134_fshift, 16.equiv nb134_gid, 24.equiv nb134_pos, 32.equiv nb134_faction, 40.equiv nb134_charge, 48.equiv nb134_p_facel, 56.equiv nb134_argkrf, 64.equiv nb134_argcrf, 72.equiv nb134_Vc, 80.equiv nb134_type, 88.equiv nb134_p_ntype, 96.equiv nb134_vdwparam, 104.equiv nb134_Vvdw, 112.equiv nb134_p_tabscale, 120.equiv nb134_VFtab, 128.equiv nb134_invsqrta, 136.equiv nb134_dvda, 144.equiv nb134_p_gbtabscale, 152.equiv nb134_GBtab, 160.equiv nb134_p_nthreads, 168.equiv nb134_count, 176.equiv nb134_mtx, 184.equiv nb134_outeriter, 192.equiv nb134_inneriter, 200.equiv nb134_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb134_ixO, 0.equiv nb134_iyO, 16.equiv nb134_izO, 32.equiv nb134_ixH1, 48.equiv nb134_iyH1, 64.equiv nb134_izH1, 80.equiv nb134_ixH2, 96.equiv nb134_iyH2, 112.equiv nb134_izH2, 128.equiv nb134_ixM, 144.equiv nb134_iyM, 160.equiv nb134_izM, 176.equiv nb134_jxO, 192.equiv nb134_jyO, 208.equiv nb134_jzO, 224.equiv nb134_jxH1, 240.equiv nb134_jyH1, 256.equiv nb134_jzH1, 272.equiv nb134_jxH2, 288.equiv nb134_jyH2, 304.equiv nb134_jzH2, 320.equiv nb134_jxM, 336.equiv nb134_jyM, 352.equiv nb134_jzM, 368.equiv nb134_dxOO, 384.equiv nb134_dyOO, 400.equiv nb134_dzOO, 416.equiv nb134_dxH1H1, 432.equiv nb134_dyH1H1, 448.equiv nb134_dzH1H1, 464.equiv nb134_dxH1H2, 480.equiv nb134_dyH1H2, 496.equiv nb134_dzH1H2, 512.equiv nb134_dxH1M, 528.equiv nb134_dyH1M, 544.equiv nb134_dzH1M, 560.equiv nb134_dxH2H1, 576.equiv nb134_dyH2H1, 592.equiv nb134_dzH2H1, 608.equiv nb134_dxH2H2, 624.equiv nb134_dyH2H2, 640.equiv nb134_dzH2H2, 656.equiv nb134_dxH2M, 672.equiv nb134_dyH2M, 688.equiv nb134_dzH2M, 704.equiv nb134_dxMH1, 720.equiv nb134_dyMH1, 736.equiv nb134_dzMH1, 752.equiv nb134_dxMH2, 768.equiv nb134_dyMH2, 784.equiv nb134_dzMH2, 800.equiv nb134_dxMM, 816.equiv nb134_dyMM, 832.equiv nb134_dzMM, 848.equiv nb134_qqMM, 864.equiv nb134_qqMH, 880.equiv nb134_qqHH, 896.equiv nb134_two, 912.equiv nb134_c6, 944.equiv nb134_c12, 960.equiv nb134_vctot, 976.equiv nb134_Vvdwtot, 992.equiv nb134_fixO, 1008.equiv nb134_fiyO, 1024.equiv nb134_fizO, 1040.equiv nb134_fixH1, 1056.equiv nb134_fiyH1, 1072.equiv nb134_fizH1, 1088.equiv nb134_fixH2, 1104.equiv nb134_fiyH2, 1120.equiv nb134_fizH2, 1136.equiv nb134_fixM, 1152.equiv nb134_fiyM, 1168.equiv nb134_fizM, 1184.equiv nb134_fjxO, 1200.equiv nb134_fjyO, 1216.equiv nb134_fjzO, 1232.equiv nb134_fjxH1, 1248.equiv nb134_fjyH1, 1264.equiv nb134_fjzH1, 1280.equiv nb134_fjxH2, 1296.equiv nb134_fjyH2, 1312.equiv nb134_fjzH2, 1328.equiv nb134_fjxM, 1344.equiv nb134_fjyM, 1360.equiv nb134_fjzM, 1376.equiv nb134_half, 1392.equiv nb134_three, 1408.equiv nb134_tsc, 1424.equiv nb134_fstmp, 1440.equiv nb134_rsqOO, 1456.equiv nb134_rsqH1H1, 1472.equiv nb134_rsqH1H2, 1488.equiv nb134_rsqH1M, 1504.equiv nb134_rsqH2H1, 1520.equiv nb134_rsqH2H2, 1536.equiv nb134_rsqH2M, 1552.equiv nb134_rsqMH1, 1568.equiv nb134_rsqMH2, 1584.equiv nb134_rsqMM, 1600.equiv nb134_rinvOO, 1616.equiv nb134_rinvH1H1, 1632.equiv nb134_rinvH1H2, 1648.equiv nb134_rinvH1M, 1664.equiv nb134_rinvH2H1, 1680.equiv nb134_rinvH2H2, 1696.equiv nb134_rinvH2M, 1712.equiv nb134_rinvMH1, 1728.equiv nb134_rinvMH2, 1744.equiv nb134_rinvMM, 1760.equiv nb134_krf, 1776.equiv nb134_crf, 1792.equiv nb134_is3, 1808.equiv nb134_ii3, 1812.equiv nb134_nri, 1816.equiv nb134_iinr, 1824.equiv nb134_jindex, 1832.equiv nb134_jjnr, 1840.equiv nb134_shift, 1848.equiv nb134_shiftvec, 1856.equiv nb134_facel, 1864.equiv nb134_innerjjnr, 1872.equiv nb134_innerk, 1880.equiv nb134_n, 1884.equiv nb134_nn1, 1888.equiv nb134_nouter, 1892.equiv nb134_ninner, 1896 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1912 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb134_nouter], eax mov [rsp + nb134_ninner], eax mov edi, [rdi] mov [rsp + nb134_nri], edi mov [rsp + nb134_iinr], rsi mov [rsp + nb134_jindex], rdx mov [rsp + nb134_jjnr], rcx mov [rsp + nb134_shift], r8 mov [rsp + nb134_shiftvec], r9 mov rsi, [rbp + nb134_p_facel] movsd xmm0, [rsi] movsd [rsp + nb134_facel], xmm0 mov rax, [rbp + nb134_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb134_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb134_half], eax mov [rsp + nb134_half+4], ebx movsd xmm1, [rsp + nb134_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 + nb134_half], xmm1 movapd [rsp + nb134_two], xmm2 movapd [rsp + nb134_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb134_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb134_charge] movsd xmm3, [rdx + rbx*8 + 24] movsd xmm4, xmm3 movsd xmm5, [rdx + rbx*8 + 8] movsd xmm6, [rsp + nb134_facel] mulsd xmm3, xmm3 mulsd xmm4, xmm5 mulsd xmm5, xmm5 mulsd xmm3, xmm6 mulsd xmm4, xmm6 mulsd xmm5, xmm6 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb134_qqMM], xmm3 movapd [rsp + nb134_qqMH], xmm4 movapd [rsp + nb134_qqHH], xmm5 xorpd xmm0, xmm0 mov rdx, [rbp + nb134_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov edx, ecx mov rdi, [rbp + nb134_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] add edx, ecx mov rax, [rbp + nb134_vdwparam] movlpd xmm0, [rax + rdx*8] movlpd xmm1, [rax + rdx*8 + 8] shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 movapd [rsp + nb134_c6], xmm0 movapd [rsp + nb134_c12], xmm1.nb134_threadloop: mov rsi, [rbp + nb134_count] ;# pointer to sync counter mov eax, [rsi].nb134_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 .nb134_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb134_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 + nb134_n], eax mov [rsp + nb134_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb134_outerstart jmp .nb134_end.nb134_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb134_nouter] mov [rsp + nb134_nouter], ebx.nb134_outer: mov rax, [rsp + nb134_shift] ;# eax = pointer into shift[] mov ebx, [rax+rsi*4] ;# ebx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb134_is3],ebx ;# store is3 mov rax, [rsp + nb134_shiftvec] ;# eax = base of shiftvec[] movsd xmm0, [rax + rbx*8] movsd xmm1, [rax + rbx*8 + 8] movsd xmm2, [rax + rbx*8 + 16] mov rcx, [rsp + nb134_iinr] ;# ecx = 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 + nb134_pos] ;# eax = base of pos[] mov [rsp + nb134_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 + nb134_ixO], xmm3 movapd [rsp + nb134_iyO], xmm4 movapd [rsp + nb134_izO], xmm5 movapd [rsp + nb134_ixH1], xmm6 movapd [rsp + nb134_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 + nb134_izH1], xmm6 movapd [rsp + nb134_ixH2], xmm0 movapd [rsp + nb134_iyH2], xmm1 movapd [rsp + nb134_izH2], xmm2 movapd [rsp + nb134_ixM], xmm3 movapd [rsp + nb134_iyM], xmm4 movapd [rsp + nb134_izM], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb134_vctot], xmm4 movapd [rsp + nb134_Vvdwtot], xmm4 movapd [rsp + nb134_fixO], xmm4 movapd [rsp + nb134_fiyO], xmm4 movapd [rsp + nb134_fizO], xmm4 movapd [rsp + nb134_fixH1], xmm4 movapd [rsp + nb134_fiyH1], xmm4 movapd [rsp + nb134_fizH1], xmm4 movapd [rsp + nb134_fixH2], xmm4 movapd [rsp + nb134_fiyH2], xmm4 movapd [rsp + nb134_fizH2], xmm4 movapd [rsp + nb134_fixM], xmm4 movapd [rsp + nb134_fiyM], xmm4 movapd [rsp + nb134_fizM], xmm4 mov rax, [rsp + nb134_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 + nb134_pos] mov rdi, [rbp + nb134_faction] mov rax, [rsp + nb134_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb134_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb134_ninner] mov [rsp + nb134_ninner], ecx add edx, 0 mov [rsp + nb134_innerk], edx ;# number of innerloop atoms jge .nb134_unroll_loop jmp .nb134_checksingle.nb134_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb134_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb134_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb134_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# load j O coordinates movlpd xmm4, [rsi + rax*8] movlpd xmm5, [rsi + rax*8 + 8] movlpd xmm6, [rsi + rax*8 + 16] movhpd xmm4, [rsi + rbx*8] movhpd xmm5, [rsi + rbx*8 + 8] movhpd xmm6, [rsi + rbx*8 + 16] ;# xmm4 = Ox ;# xmm5 = Oy ;# xmm6 = Oz subpd xmm4, [rsp + nb134_ixO] subpd xmm5, [rsp + nb134_iyO] subpd xmm6, [rsp + nb134_izO] ;# store dx/dy/dz movapd xmm13, xmm4 movapd xmm14, xmm5 movapd xmm15, xmm6 ;# square it mulpd xmm4, xmm4 mulpd xmm5, xmm5 mulpd xmm6, xmm6 ⌨️ 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?