nb_kernel132_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,328 行 · 第 1/5 页
S
2,328 行
;#;# $Id: nb_kernel132_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_kernel132_x86_64_sse2.globl _nb_kernel132_x86_64_sse2nb_kernel132_x86_64_sse2: _nb_kernel132_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb132_fshift, 16.equiv nb132_gid, 24.equiv nb132_pos, 32.equiv nb132_faction, 40.equiv nb132_charge, 48.equiv nb132_p_facel, 56.equiv nb132_argkrf, 64.equiv nb132_argcrf, 72.equiv nb132_Vc, 80.equiv nb132_type, 88.equiv nb132_p_ntype, 96.equiv nb132_vdwparam, 104.equiv nb132_Vvdw, 112.equiv nb132_p_tabscale, 120.equiv nb132_VFtab, 128.equiv nb132_invsqrta, 136.equiv nb132_dvda, 144.equiv nb132_p_gbtabscale, 152.equiv nb132_GBtab, 160.equiv nb132_p_nthreads, 168.equiv nb132_count, 176.equiv nb132_mtx, 184.equiv nb132_outeriter, 192.equiv nb132_inneriter, 200.equiv nb132_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb132_ixO, 0.equiv nb132_iyO, 16.equiv nb132_izO, 32.equiv nb132_ixH1, 48.equiv nb132_iyH1, 64.equiv nb132_izH1, 80.equiv nb132_ixH2, 96.equiv nb132_iyH2, 112.equiv nb132_izH2, 128.equiv nb132_jxO, 144.equiv nb132_jyO, 160.equiv nb132_jzO, 176.equiv nb132_jxH1, 192.equiv nb132_jyH1, 208.equiv nb132_jzH1, 224.equiv nb132_jxH2, 240.equiv nb132_jyH2, 256.equiv nb132_jzH2, 272.equiv nb132_dxOO, 288.equiv nb132_dyOO, 304.equiv nb132_dzOO, 320.equiv nb132_dxOH1, 336.equiv nb132_dyOH1, 352.equiv nb132_dzOH1, 368.equiv nb132_dxOH2, 384.equiv nb132_dyOH2, 400.equiv nb132_dzOH2, 416.equiv nb132_dxH1O, 432.equiv nb132_dyH1O, 448.equiv nb132_dzH1O, 464.equiv nb132_dxH1H1, 480.equiv nb132_dyH1H1, 496.equiv nb132_dzH1H1, 512.equiv nb132_dxH1H2, 528.equiv nb132_dyH1H2, 544.equiv nb132_dzH1H2, 560.equiv nb132_dxH2O, 576.equiv nb132_dyH2O, 592.equiv nb132_dzH2O, 608.equiv nb132_dxH2H1, 624.equiv nb132_dyH2H1, 640.equiv nb132_dzH2H1, 656.equiv nb132_dxH2H2, 672.equiv nb132_dyH2H2, 688.equiv nb132_dzH2H2, 704.equiv nb132_qqOO, 720.equiv nb132_qqOH, 736.equiv nb132_qqHH, 752.equiv nb132_c6, 768.equiv nb132_c12, 784.equiv nb132_tsc, 800.equiv nb132_fstmp, 816.equiv nb132_vctot, 832.equiv nb132_Vvdwtot, 848.equiv nb132_fixO, 864.equiv nb132_fiyO, 880.equiv nb132_fizO, 896.equiv nb132_fixH1, 912.equiv nb132_fiyH1, 928.equiv nb132_fizH1, 944.equiv nb132_fixH2, 960.equiv nb132_fiyH2, 976.equiv nb132_fizH2, 992.equiv nb132_fjxO, 1008.equiv nb132_fjyO, 1024.equiv nb132_fjzO, 1040.equiv nb132_fjxH1, 1056.equiv nb132_fjyH1, 1072.equiv nb132_fjzH1, 1088.equiv nb132_fjxH2, 1104.equiv nb132_fjyH2, 1120.equiv nb132_fjzH2, 1136.equiv nb132_half, 1152.equiv nb132_three, 1168.equiv nb132_rsqOO, 1184.equiv nb132_rsqOH1, 1200.equiv nb132_rsqOH2, 1216.equiv nb132_rsqH1O, 1232.equiv nb132_rsqH1H1, 1248.equiv nb132_rsqH1H2, 1264.equiv nb132_rsqH2O, 1280.equiv nb132_rsqH2H1, 1296.equiv nb132_rsqH2H2, 1312.equiv nb132_rinvOO, 1328.equiv nb132_rinvOH1, 1344.equiv nb132_rinvOH2, 1360.equiv nb132_rinvH1O, 1376.equiv nb132_rinvH1H1, 1392.equiv nb132_rinvH1H2, 1408.equiv nb132_rinvH2O, 1424.equiv nb132_rinvH2H1, 1440.equiv nb132_rinvH2H2, 1456.equiv nb132_two, 1472.equiv nb132_krf, 1488.equiv nb132_crf, 1504.equiv nb132_nri, 1520.equiv nb132_iinr, 1528.equiv nb132_jindex, 1536.equiv nb132_jjnr, 1544.equiv nb132_shift, 1552.equiv nb132_shiftvec, 1560.equiv nb132_facel, 1568.equiv nb132_innerjjnr, 1576.equiv nb132_is3, 1584.equiv nb132_ii3, 1588.equiv nb132_innerk, 1592.equiv nb132_n, 1596.equiv nb132_nn1, 1600.equiv nb132_nouter, 1604.equiv nb132_ninner, 1608 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1624 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb132_nouter], eax mov [rsp + nb132_ninner], eax mov edi, [rdi] mov [rsp + nb132_nri], edi mov [rsp + nb132_iinr], rsi mov [rsp + nb132_jindex], rdx mov [rsp + nb132_jjnr], rcx mov [rsp + nb132_shift], r8 mov [rsp + nb132_shiftvec], r9 mov rsi, [rbp + nb132_p_facel] movsd xmm0, [rsi] movsd [rsp + nb132_facel], xmm0 mov rax, [rbp + nb132_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb132_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb132_half], eax mov [rsp + nb132_half+4], ebx movsd xmm1, [rsp + nb132_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 + nb132_half], xmm1 movapd [rsp + nb132_two], xmm2 movapd [rsp + nb132_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb132_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb132_charge] movsd xmm3, [rdx + rbx*8] movsd xmm4, xmm3 movsd xmm5, [rdx + rbx*8 + 8] movsd xmm6, [rsp + nb132_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 + nb132_qqOO], xmm3 movapd [rsp + nb132_qqOH], xmm4 movapd [rsp + nb132_qqHH], xmm5 xorpd xmm0, xmm0 mov rdx, [rbp + nb132_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov edx, ecx mov rdi, [rbp + nb132_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] add edx, ecx mov rax, [rbp + nb132_vdwparam] movlpd xmm0, [rax + rdx*8] movlpd xmm1, [rax + rdx*8 + 8] shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 movapd [rsp + nb132_c6], xmm0 movapd [rsp + nb132_c12], xmm1.nb132_threadloop: mov rsi, [rbp + nb132_count] ;# pointer to sync counter mov eax, [rsi].nb132_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 .nb132_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb132_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 + nb132_n], eax mov [rsp + nb132_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb132_outerstart jmp .nb132_end.nb132_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb132_nouter] mov [rsp + nb132_nouter], ebx.nb132_outer: mov rax, [rsp + nb132_shift] ;# eax = pointer into shift[] mov ebx, [rax + rsi*4] ;# ebx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb132_is3],ebx ;# store is3 mov rax, [rsp + nb132_shiftvec] ;# eax = base of shiftvec[] movlpd xmm0, [rax + rbx*8] movlpd xmm1, [rax + rbx*8 + 8] movlpd xmm2, [rax + rbx*8 + 16] mov rcx, [rsp + nb132_iinr] ;# ecx = pointer into iinr[] mov ebx, [rcx + rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb132_pos] ;# eax = base of pos[] mov [rsp + nb132_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 + nb132_ixO], xmm3 movapd [rsp + nb132_iyO], xmm4 movapd [rsp + nb132_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 + nb132_ixH1], xmm0 movapd [rsp + nb132_iyH1], xmm1 movapd [rsp + nb132_izH1], xmm2 movapd [rsp + nb132_ixH2], xmm3 movapd [rsp + nb132_iyH2], xmm4 movapd [rsp + nb132_izH2], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb132_vctot], xmm4 movapd [rsp + nb132_Vvdwtot], xmm4 movapd [rsp + nb132_fixO], xmm4 movapd [rsp + nb132_fiyO], xmm4 movapd [rsp + nb132_fizO], xmm4 movapd [rsp + nb132_fixH1], xmm4 movapd [rsp + nb132_fiyH1], xmm4 movapd [rsp + nb132_fizH1], xmm4 movapd [rsp + nb132_fixH2], xmm4 movapd [rsp + nb132_fiyH2], xmm4 movapd [rsp + nb132_fizH2], xmm4 mov rax, [rsp + nb132_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 + nb132_pos] mov rdi, [rbp + nb132_faction] mov rax, [rsp + nb132_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb132_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb132_ninner] mov [rsp + nb132_ninner], ecx add edx, 0 mov [rsp + nb132_innerk], edx ;# number of innerloop atoms jge .nb132_unroll_loop jmp .nb132_checksingle.nb132_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb132_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb132_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb132_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 + nb132_ixO] subpd xmm1, [rsp + nb132_iyO] subpd xmm2, [rsp + nb132_izO] subpd xmm3, [rsp + nb132_ixH1] subpd xmm4, [rsp + nb132_iyH1] subpd xmm5, [rsp + nb132_izH1] subpd xmm6, [rsp + nb132_ixH2] subpd xmm7, [rsp + nb132_iyH2] subpd xmm8, [rsp + nb132_izH2] movapd [rsp + nb132_dxOO], xmm0 movapd [rsp + nb132_dyOO], xmm1 movapd [rsp + nb132_dzOO], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [rsp + nb132_dxH1O], xmm3 movapd [rsp + nb132_dyH1O], xmm4 movapd [rsp + nb132_dzH1O], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 movapd [rsp + nb132_dxH2O], xmm6 movapd [rsp + nb132_dyH2O], xmm7 movapd [rsp + nb132_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⌨️ 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?