nb_kernel104_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,312 行 · 第 1/5 页
S
2,312 行
;#;# $Id: nb_kernel104_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_kernel104_x86_64_sse2.globl _nb_kernel104_x86_64_sse2nb_kernel104_x86_64_sse2: _nb_kernel104_x86_64_sse2: .equiv nb104_fshift, 16.equiv nb104_gid, 24.equiv nb104_pos, 32.equiv nb104_faction, 40.equiv nb104_charge, 48.equiv nb104_p_facel, 56.equiv nb104_argkrf, 64.equiv nb104_argcrf, 72.equiv nb104_Vc, 80.equiv nb104_type, 88.equiv nb104_p_ntype, 96.equiv nb104_vdwparam, 104.equiv nb104_Vvdw, 112.equiv nb104_p_tabscale, 120.equiv nb104_VFtab, 128.equiv nb104_invsqrta, 136.equiv nb104_dvda, 144.equiv nb104_p_gbtabscale, 152.equiv nb104_GBtab, 160.equiv nb104_p_nthreads, 168.equiv nb104_count, 176.equiv nb104_mtx, 184.equiv nb104_outeriter, 192.equiv nb104_inneriter, 200.equiv nb104_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb104_ixM, 0.equiv nb104_iyM, 16.equiv nb104_izM, 32.equiv nb104_ixH1, 48.equiv nb104_iyH1, 64.equiv nb104_izH1, 80.equiv nb104_ixH2, 96.equiv nb104_iyH2, 112.equiv nb104_izH2, 128.equiv nb104_jxM, 144.equiv nb104_jyM, 160.equiv nb104_jzM, 176.equiv nb104_jxH1, 192.equiv nb104_jyH1, 208.equiv nb104_jzH1, 224.equiv nb104_jxH2, 240.equiv nb104_jyH2, 256.equiv nb104_jzH2, 272.equiv nb104_dxMM, 288.equiv nb104_dyMM, 304.equiv nb104_dzMM, 320.equiv nb104_dxMH1, 336.equiv nb104_dyMH1, 352.equiv nb104_dzMH1, 368.equiv nb104_dxMH2, 384.equiv nb104_dyMH2, 400.equiv nb104_dzMH2, 416.equiv nb104_dxH1M, 432.equiv nb104_dyH1M, 448.equiv nb104_dzH1M, 464.equiv nb104_dxH1H1, 480.equiv nb104_dyH1H1, 496.equiv nb104_dzH1H1, 512.equiv nb104_dxH1H2, 528.equiv nb104_dyH1H2, 544.equiv nb104_dzH1H2, 560.equiv nb104_dxH2M, 576.equiv nb104_dyH2M, 592.equiv nb104_dzH2M, 608.equiv nb104_dxH2H1, 624.equiv nb104_dyH2H1, 640.equiv nb104_dzH2H1, 656.equiv nb104_dxH2H2, 672.equiv nb104_dyH2H2, 688.equiv nb104_dzH2H2, 704.equiv nb104_qqMM, 720.equiv nb104_qqMH, 736.equiv nb104_qqHH, 752.equiv nb104_vctot, 768.equiv nb104_fixM, 784.equiv nb104_fiyM, 800.equiv nb104_fizM, 816.equiv nb104_fixH1, 832.equiv nb104_fiyH1, 848.equiv nb104_fizH1, 864.equiv nb104_fixH2, 880.equiv nb104_fiyH2, 896.equiv nb104_fizH2, 912.equiv nb104_fjxM, 928.equiv nb104_fjyM, 944.equiv nb104_fjzM, 960.equiv nb104_fjxH1, 976.equiv nb104_fjyH1, 992.equiv nb104_fjzH1, 1008.equiv nb104_fjxH2, 1024.equiv nb104_fjyH2, 1040.equiv nb104_fjzH2, 1056.equiv nb104_half, 1072.equiv nb104_three, 1088.equiv nb104_rsqMM, 1104.equiv nb104_rsqMH1, 1120.equiv nb104_rsqMH2, 1136.equiv nb104_rsqH1M, 1152.equiv nb104_rsqH1H1, 1168.equiv nb104_rsqH1H2, 1184.equiv nb104_rsqH2M, 1200.equiv nb104_rsqH2H1, 1216.equiv nb104_rsqH2H2, 1232.equiv nb104_rinvMM, 1248.equiv nb104_rinvMH1, 1264.equiv nb104_rinvMH2, 1280.equiv nb104_rinvH1M, 1296.equiv nb104_rinvH1H1, 1312.equiv nb104_rinvH1H2, 1328.equiv nb104_rinvH2M, 1344.equiv nb104_rinvH2H1, 1360.equiv nb104_rinvH2H2, 1376.equiv nb104_nri, 1392.equiv nb104_innerjjnr, 1400.equiv nb104_iinr, 1408.equiv nb104_jindex, 1416.equiv nb104_jjnr, 1424.equiv nb104_shift, 1432.equiv nb104_shiftvec, 1440.equiv nb104_facel, 1448.equiv nb104_is3, 1456.equiv nb104_ii3, 1460.equiv nb104_innerk, 1464.equiv nb104_n, 1468.equiv nb104_nn1, 1472.equiv nb104_nouter, 1476.equiv nb104_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 + nb104_nouter], eax mov [rsp + nb104_ninner], eax mov edi, [rdi] mov [rsp + nb104_nri], edi mov [rsp + nb104_iinr], rsi mov [rsp + nb104_jindex], rdx mov [rsp + nb104_jjnr], rcx mov [rsp + nb104_shift], r8 mov [rsp + nb104_shiftvec], r9 mov rsi, [rbp + nb104_p_facel] movsd xmm0, [rsi] movsd [rsp + nb104_facel], xmm0 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb104_half], eax mov [rsp + nb104_half+4], ebx movsd xmm1, [rsp + nb104_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 + nb104_half], xmm1 movapd [rsp + nb104_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb104_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb104_charge] movsd xmm3, [rdx + rbx*8 + 24] ;# qM movsd xmm4, xmm3 ;# qM movsd xmm5, [rdx + rbx*8 + 8] ;# qH mov rsi, [rbp + nb104_p_facel] movsd xmm0, [rsi] movsd xmm6, [rsp + nb104_facel] ;# facel mulsd xmm3, xmm3 ;# qM*qM mulsd xmm4, xmm5 ;# qM*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 + nb104_qqMM], xmm3 movapd [rsp + nb104_qqMH], xmm4 movapd [rsp + nb104_qqHH], xmm5.nb104_threadloop: mov rsi, [rbp + nb104_count] ;# pointer to sync counter mov eax, [rsi].nb104_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 .nb104_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb104_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 + nb104_n], eax mov [rsp + nb104_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb104_outerstart jmp .nb104_end.nb104_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb104_nouter] mov [rsp + nb104_nouter], ebx.nb104_outer: mov rax, [rsp + nb104_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb104_is3],ebx ;# store is3 mov rax, [rsp + nb104_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 + nb104_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb104_pos] ;# rax = base of pos[] mov [rsp + nb104_ii3], ebx movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 addsd xmm3, [rax + rbx*8 + 24] addsd xmm4, [rax + rbx*8 + 32] addsd xmm5, [rax + rbx*8 + 40] shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb104_ixH1], xmm3 movapd [rsp + nb104_iyH1], xmm4 movapd [rsp + nb104_izH1], xmm5 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm0, [rax + rbx*8 + 48] addsd xmm1, [rax + rbx*8 + 56] addsd xmm2, [rax + rbx*8 + 64] addsd xmm3, [rax + rbx*8 + 72] addsd xmm4, [rax + rbx*8 + 80] addsd xmm5, [rax + rbx*8 + 88] 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 + nb104_ixH2], xmm0 movapd [rsp + nb104_iyH2], xmm1 movapd [rsp + nb104_izH2], xmm2 movapd [rsp + nb104_ixM], xmm3 movapd [rsp + nb104_iyM], xmm4 movapd [rsp + nb104_izM], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb104_vctot], xmm4 movapd [rsp + nb104_fixM], xmm4 movapd [rsp + nb104_fiyM], xmm4 movapd [rsp + nb104_fizM], xmm4 movapd [rsp + nb104_fixH1], xmm4 movapd [rsp + nb104_fiyH1], xmm4 movapd [rsp + nb104_fizH1], xmm4 movapd [rsp + nb104_fixH2], xmm4 movapd [rsp + nb104_fiyH2], xmm4 movapd [rsp + nb104_fizH2], xmm4 mov rax, [rsp + nb104_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 + nb104_pos] mov rdi, [rbp + nb104_faction] mov rax, [rsp + nb104_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb104_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb104_ninner] mov [rsp + nb104_ninner], ecx add edx, 0 mov [rsp + nb104_innerk], edx ;# number of innerloop atoms jge .nb104_unroll_loop jmp .nb104_checksingle.nb104_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb104_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb104_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb104_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# move j H1 coordinates to local temp variables movlpd xmm0, [rsi + rax*8 + 24] movlpd xmm1, [rsi + rax*8 + 32] movlpd xmm2, [rsi + rax*8 + 40] movhpd xmm0, [rsi + rbx*8 + 24] movhpd xmm1, [rsi + rbx*8 + 32] movhpd xmm2, [rsi + rbx*8 + 40] ;# xmm0 = H1x ;# xmm1 = H1y ;# xmm2 = H1z movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 movapd xmm8, xmm2 subpd xmm0, [rsp + nb104_ixH1] subpd xmm1, [rsp + nb104_iyH1] subpd xmm2, [rsp + nb104_izH1] subpd xmm3, [rsp + nb104_ixH2] subpd xmm4, [rsp + nb104_iyH2] subpd xmm5, [rsp + nb104_izH2] subpd xmm6, [rsp + nb104_ixM] subpd xmm7, [rsp + nb104_iyM] subpd xmm8, [rsp + nb104_izM] movapd [rsp + nb104_dxH1H1], xmm0 movapd [rsp + nb104_dyH1H1], xmm1 movapd [rsp + nb104_dzH1H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [rsp + nb104_dxH2H1], xmm3 movapd [rsp + nb104_dyH2H1], xmm4 movapd [rsp + nb104_dzH2H1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 movapd [rsp + nb104_dxMH1], xmm6 movapd [rsp + nb104_dyMH1], xmm7 movapd [rsp + nb104_dzMH1], 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 jH1 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 + nb104_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?