nb_kernel104_x86_64_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,223 行 · 第 1/5 页
S
2,223 行
;#;# $Id: nb_kernel104_x86_64_sse.intel_syntax.s,v 1.1.2.2 2006/09/22 08:32:48 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_sse.globl _nb_kernel104_x86_64_ssenb_kernel104_x86_64_sse: _nb_kernel104_x86_64_sse: .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 sse use .equiv nb104_ixH1, 0.equiv nb104_iyH1, 16.equiv nb104_izH1, 32.equiv nb104_ixH2, 48.equiv nb104_iyH2, 64.equiv nb104_izH2, 80.equiv nb104_ixM, 96.equiv nb104_iyM, 112.equiv nb104_izM, 128.equiv nb104_jxH1, 144.equiv nb104_jyH1, 160.equiv nb104_jzH1, 176.equiv nb104_jxH2, 192.equiv nb104_jyH2, 208.equiv nb104_jzH2, 224.equiv nb104_jxM, 240.equiv nb104_jyM, 256.equiv nb104_jzM, 272.equiv nb104_dxH1H1, 288.equiv nb104_dyH1H1, 304.equiv nb104_dzH1H1, 320.equiv nb104_dxH1H2, 336.equiv nb104_dyH1H2, 352.equiv nb104_dzH1H2, 368.equiv nb104_dxH1M, 384.equiv nb104_dyH1M, 400.equiv nb104_dzH1M, 416.equiv nb104_dxH2H1, 432.equiv nb104_dyH2H1, 448.equiv nb104_dzH2H1, 464.equiv nb104_dxH2H2, 480.equiv nb104_dyH2H2, 496.equiv nb104_dzH2H2, 512.equiv nb104_dxH2M, 528.equiv nb104_dyH2M, 544.equiv nb104_dzH2M, 560.equiv nb104_dxMH1, 576.equiv nb104_dyMH1, 592.equiv nb104_dzMH1, 608.equiv nb104_dxMH2, 624.equiv nb104_dyMH2, 640.equiv nb104_dzMH2, 656.equiv nb104_dxMM, 672.equiv nb104_dyMM, 688.equiv nb104_dzMM, 704.equiv nb104_qqHH, 720.equiv nb104_qqMH, 736.equiv nb104_qqMM, 752.equiv nb104_vctot, 768.equiv nb104_fixH1, 784.equiv nb104_fiyH1, 800.equiv nb104_fizH1, 816.equiv nb104_fixH2, 832.equiv nb104_fiyH2, 848.equiv nb104_fizH2, 864.equiv nb104_fixM, 880.equiv nb104_fiyM, 896.equiv nb104_fizM, 912.equiv nb104_fjxH1, 928.equiv nb104_fjyH1, 944.equiv nb104_fjzH1, 960.equiv nb104_fjxH2, 976.equiv nb104_fjyH2, 992.equiv nb104_fjzH2, 1008.equiv nb104_fjxM, 1024.equiv nb104_fjyM, 1040.equiv nb104_fjzM, 1056.equiv nb104_half, 1072.equiv nb104_three, 1088.equiv nb104_rsqH1H1, 1104.equiv nb104_rsqH1H2, 1120.equiv nb104_rsqH1M, 1136.equiv nb104_rsqH2H1, 1152.equiv nb104_rsqH2H2, 1168.equiv nb104_rsqH2M, 1184.equiv nb104_rsqMH1, 1200.equiv nb104_rsqMH2, 1216.equiv nb104_rsqMM, 1232.equiv nb104_rinvH1H1, 1248.equiv nb104_rinvH1H2, 1264.equiv nb104_rinvH1M, 1280.equiv nb104_rinvH2H1, 1296.equiv nb104_rinvH2H2, 1312.equiv nb104_rinvH2M, 1328.equiv nb104_rinvMH1, 1344.equiv nb104_rinvMH2, 1360.equiv nb104_rinvMM, 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, 1464.equiv nb104_innerk, 1472.equiv nb104_n, 1480.equiv nb104_nn1, 1484.equiv nb104_nouter, 1488.equiv nb104_ninner, 1492 push rbp mov rbp, rsp push rbx push r12 push r13 push r14 push r15 sub rsp, 1512 emms ;# 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] movss xmm0, [rsi] movss [rsp + nb104_facel], xmm0 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# half in IEEE (hex) mov [rsp + nb104_half], eax movss xmm1, [rsp + nb104_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# one movaps xmm3, xmm2 addps xmm2, xmm2 ;# two addps xmm3, xmm2 ;# three movaps [rsp + nb104_half], xmm1 movaps [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] movss xmm3, [rdx + rbx*4 + 4] movss xmm4, xmm3 movss xmm5, [rdx + rbx*4 + 12] mov rsi, [rbp + nb104_p_facel] movss xmm0, [rsi] movss xmm6, [rsp + nb104_facel] mulss xmm3, xmm3 mulss xmm4, xmm5 mulss xmm5, xmm5 mulss xmm3, xmm6 mulss xmm4, xmm6 mulss xmm5, xmm6 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb104_qqHH], xmm3 movaps [rsp + nb104_qqMH], xmm4 movaps [rsp + nb104_qqMM], 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[] movss xmm0, [rax + rbx*4] movss xmm1, [rax + rbx*4 + 4] movss xmm2, [rax + rbx*4 + 8] 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 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 addss xmm3, [rax + rbx*4 + 12] addss xmm4, [rax + rbx*4 + 16] addss xmm5, [rax + rbx*4 + 20] shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb104_ixH1], xmm3 movaps [rsp + nb104_iyH1], xmm4 movaps [rsp + nb104_izH1], xmm5 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm0, [rax + rbx*4 + 24] addss xmm1, [rax + rbx*4 + 28] addss xmm2, [rax + rbx*4 + 32] addss xmm3, [rax + rbx*4 + 36] addss xmm4, [rax + rbx*4 + 40] addss xmm5, [rax + rbx*4 + 44] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [rsp + nb104_ixH2], xmm0 movaps [rsp + nb104_iyH2], xmm1 movaps [rsp + nb104_izH2], xmm2 movaps [rsp + nb104_ixM], xmm3 movaps [rsp + nb104_iyM], xmm4 movaps [rsp + nb104_izM], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [rsp + nb104_vctot], xmm4 movaps [rsp + nb104_fixH1], xmm4 movaps [rsp + nb104_fiyH1], xmm4 movaps [rsp + nb104_fizH1], xmm4 movaps [rsp + nb104_fixH2], xmm4 movaps [rsp + nb104_fiyH2], xmm4 movaps [rsp + nb104_fizH2], xmm4 movaps [rsp + nb104_fixM], xmm4 movaps [rsp + nb104_fiyM], xmm4 movaps [rsp + nb104_fizM], 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, 4 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_single_check.nb104_unroll_loop: ;# quad-unroll innerloop here mov rdx, [rsp + nb104_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] mov ecx, [rdx + 8] mov edx, [rdx + 12] ;# eax-edx=jnr1-4 add qword ptr [rsp + nb104_innerjjnr], 16 ;# advance pointer (unrolled 4) mov rsi, [rbp + nb104_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] lea rcx, [rcx + rcx*2] ;# replace jnr with j3 lea rdx, [rdx + rdx*2] ;# move j H1 coordinates to local temp variables movlps xmm0, [rsi + rax*4 + 12] ;# jxH1a jyH1a - - movlps xmm1, [rsi + rcx*4 + 12] ;# jxH1c jyH1c - - movhps xmm0, [rsi + rbx*4 + 12] ;# jxH1a jyH1a jxH1b jyH1b movhps xmm1, [rsi + rdx*4 + 12] ;# jxH1c jyH1c jxH1d jyH1d movss xmm2, [rsi + rax*4 + 20] ;# jzH1a - - - movss xmm3, [rsi + rcx*4 + 20] ;# jzH1c - - - movhps xmm2, [rsi + rbx*4 + 20] ;# jzH1a - jzH1b - movhps xmm3, [rsi + rdx*4 + 20] ;# jzH1c - jzH1d - movaps xmm4, xmm0 unpcklps xmm0, xmm1 ;# jxH1a jxH1c jyH1a jyH1c unpckhps xmm4, xmm1 ;# jxH1b jxH1d jyH1b jyH1d movaps xmm1, xmm0 unpcklps xmm0, xmm4 ;# x unpckhps xmm1, xmm4 ;# y shufps xmm2, xmm3, 136 ;# 10001000 => jzH1a jzH1b jzH1c jzH1d ;# xmm0 = H1x ;# xmm1 = H1y ;# xmm2 = H1z movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 movaps xmm6, xmm0 movaps xmm7, xmm1 movaps xmm8, xmm2 subps xmm0, [rsp + nb104_ixH1] subps xmm1, [rsp + nb104_iyH1] subps xmm2, [rsp + nb104_izH1] subps xmm3, [rsp + nb104_ixH2] subps xmm4, [rsp + nb104_iyH2] subps xmm5, [rsp + nb104_izH2] subps xmm6, [rsp + nb104_ixM] subps xmm7, [rsp + nb104_iyM] subps xmm8, [rsp + nb104_izM] movaps [rsp + nb104_dxH1H1], xmm0 movaps [rsp + nb104_dyH1H1], xmm1 movaps [rsp + nb104_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [rsp + nb104_dxH2H1], xmm3 movaps [rsp + nb104_dyH2H1], xmm4 movaps [rsp + nb104_dzH2H1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 movaps [rsp + nb104_dxMH1], xmm6 movaps [rsp + nb104_dyMH1], xmm7 movaps [rsp + nb104_dzMH1], xmm8 mulps xmm6, xmm6 mulps xmm7, xmm7 mulps xmm8, xmm8 addps xmm0, xmm1⌨️ 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?