nb_kernel330_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,474 行 · 第 1/3 页
S
1,474 行
;#;# $Id: nb_kernel330_x86_64_sse2.intel_syntax.s,v 1.1.2.2 2006/09/22 08:40:37 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_kernel330_x86_64_sse2.globl _nb_kernel330_x86_64_sse2nb_kernel330_x86_64_sse2: _nb_kernel330_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb330_fshift, 16.equiv nb330_gid, 24.equiv nb330_pos, 32.equiv nb330_faction, 40.equiv nb330_charge, 48.equiv nb330_p_facel, 56.equiv nb330_argkrf, 64.equiv nb330_argcrf, 72.equiv nb330_Vc, 80.equiv nb330_type, 88.equiv nb330_p_ntype, 96.equiv nb330_vdwparam, 104.equiv nb330_Vvdw, 112.equiv nb330_p_tabscale, 120.equiv nb330_VFtab, 128.equiv nb330_invsqrta, 136.equiv nb330_dvda, 144.equiv nb330_p_gbtabscale, 152.equiv nb330_GBtab, 160.equiv nb330_p_nthreads, 168.equiv nb330_count, 176.equiv nb330_mtx, 184.equiv nb330_outeriter, 192.equiv nb330_inneriter, 200.equiv nb330_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb330_ix, 0.equiv nb330_iy, 16.equiv nb330_iz, 32.equiv nb330_iq, 48.equiv nb330_dx, 64.equiv nb330_dy, 80.equiv nb330_dz, 96.equiv nb330_rinv, 112.equiv nb330_tsc, 128.equiv nb330_qq, 144.equiv nb330_c6, 160.equiv nb330_c12, 176.equiv nb330_eps, 192.equiv nb330_vctot, 208.equiv nb330_Vvdwtot, 224.equiv nb330_fix, 240.equiv nb330_fiy, 256.equiv nb330_fiz, 272.equiv nb330_half, 288.equiv nb330_three, 304.equiv nb330_nri, 320.equiv nb330_iinr, 328.equiv nb330_jindex, 336.equiv nb330_jjnr, 344.equiv nb330_shift, 352.equiv nb330_shiftvec, 360.equiv nb330_facel, 368.equiv nb330_innerjjnr, 376.equiv nb330_is3, 384.equiv nb330_ii3, 388.equiv nb330_ntia, 392.equiv nb330_innerk, 396.equiv nb330_n, 400.equiv nb330_nn1, 404.equiv nb330_ntype, 408.equiv nb330_nouter, 412.equiv nb330_ninner, 416 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 440 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb330_nouter], eax mov [rsp + nb330_ninner], eax mov edi, [rdi] mov [rsp + nb330_nri], edi mov [rsp + nb330_iinr], rsi mov [rsp + nb330_jindex], rdx mov [rsp + nb330_jjnr], rcx mov [rsp + nb330_shift], r8 mov [rsp + nb330_shiftvec], r9 mov rdi, [rbp + nb330_p_ntype] mov edi, [rdi] mov [rsp + nb330_ntype], edi mov rsi, [rbp + nb330_p_facel] movsd xmm0, [rsi] movsd [rsp + nb330_facel], xmm0 mov rax, [rbp + nb330_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb330_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb330_half], eax mov [rsp + nb330_half+4], ebx movsd xmm1, [rsp + nb330_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 + nb330_half], xmm1 movapd [rsp + nb330_three], xmm3.nb330_threadloop: mov rsi, [rbp + nb330_count] ;# pointer to sync counter mov eax, [rsi].nb330_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 .nb330_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb330_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 + nb330_n], eax mov [rsp + nb330_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb330_outerstart jmp .nb330_end.nb330_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb330_nouter] mov [rsp + nb330_nouter], ebx.nb330_outer: mov rax, [rsp + nb330_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb330_is3],ebx ;# store is3 mov rax, [rsp + nb330_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 + nb330_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii mov rdx, [rbp + nb330_charge] movsd xmm3, [rdx + rbx*8] mulsd xmm3, [rsp + nb330_facel] shufpd xmm3, xmm3, 0 mov rdx, [rbp + nb330_type] mov edx, [rdx + rbx*4] imul edx, [rsp + nb330_ntype] shl edx, 1 mov [rsp + nb330_ntia], edx lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb330_pos] ;# rax = base of pos[] addsd xmm0, [rax + rbx*8] addsd xmm1, [rax + rbx*8 + 8] addsd xmm2, [rax + rbx*8 + 16] movapd [rsp + nb330_iq], xmm3 shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 movapd [rsp + nb330_ix], xmm0 movapd [rsp + nb330_iy], xmm1 movapd [rsp + nb330_iz], xmm2 mov [rsp + nb330_ii3], ebx ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb330_vctot], xmm4 movapd [rsp + nb330_Vvdwtot], xmm4 movapd [rsp + nb330_fix], xmm4 movapd [rsp + nb330_fiy], xmm4 movapd [rsp + nb330_fiz], xmm4 mov rax, [rsp + nb330_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 + nb330_pos] mov rdi, [rbp + nb330_faction] mov rax, [rsp + nb330_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb330_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb330_ninner] mov [rsp + nb330_ninner], ecx add edx, 0 mov [rsp + nb330_innerk], edx ;# number of innerloop atoms jge .nb330_unroll_loop jmp .nb330_checksingle.nb330_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb330_innerjjnr] ;# pointer to jjnr[k] mov r12d, [rdx] mov r13d, [rdx + 4] add qword ptr [rsp + nb330_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb330_pos] ;# base of pos[] lea rax, [r12 + r12*2] ;# replace jnr with j3 lea rbx, [r13 + r13*2] ;# move two coordinates to xmm4-xmm6 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] ;# calc dr subpd xmm4, [rsp + nb330_ix] subpd xmm5, [rsp + nb330_iy] subpd xmm6, [rsp + nb330_iz] ;# store dr movapd [rsp + nb330_dx], xmm4 movapd [rsp + nb330_dy], xmm5 movapd [rsp + nb330_dz], xmm6 mov rsi, [rbp + nb330_charge] ;# base of charge[] ;# square it mulpd xmm4,xmm4 mulpd xmm5,xmm5 mulpd xmm6,xmm6 addpd xmm4, xmm5 addpd xmm4, xmm6 ;# rsq in xmm4 movlpd xmm3, [rsi + r12*8] cvtpd2ps xmm5, xmm4 rsqrtps xmm5, xmm5 cvtps2pd xmm2, xmm5 ;# lu in low xmm2 movhpd xmm3, [rsi + r13*8] mov rsi, [rbp + nb330_type] ;# lookup seed in xmm2 movapd xmm5, xmm2 ;# copy of lu mulpd xmm2, xmm2 ;# lu*lu movapd xmm1, [rsp + nb330_three] mulpd xmm3, [rsp + nb330_iq] mulpd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb330_half] subpd xmm1, xmm2 ;# 30-rsq*lu*lu mulpd xmm1, xmm5 mulpd xmm1, xmm0 ;# xmm0=iter1 of rinv (new lu) movapd [rsp + nb330_qq], xmm3 mov r8d, [rsi + r12*4] mov r9d, [rsi + r13*4] movapd xmm5, xmm1 ;# copy of lu mulpd xmm1, xmm1 ;# lu*lu movapd xmm2, [rsp + nb330_three] mulpd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb330_half] subpd xmm2, xmm1 ;# 30-rsq*lu*lu mulpd xmm2, xmm5 mulpd xmm0, xmm2 ;# xmm0=iter2 of rinv (new lu) mulpd xmm4, xmm0 ;# xmm4=r mulpd xmm4, [rsp + nb330_tsc] movapd xmm15, xmm0 ;# copy of rinv shl r8d, 1 shl r9d, 1 mov edi, [rsp + nb330_ntia] cvttpd2pi mm6, xmm4 ;# mm6 = lu idx add r8d, edi add r9d, edi cvtpi2pd xmm5, mm6 subpd xmm4, xmm5 movapd [rsp + nb330_eps], xmm4 pslld mm6, 2 ;# idx *= 4 mov rsi, [rbp + nb330_VFtab] movd r10d, mm6 psrlq mm6, 32 movd r11d, mm6 ;# indices in r10,r11 lea r10, [r10 + r10*2] ;# idx*=3 (12 total now) lea r11, [r11 + r11*2] ;# idx*=3 (12 total now) ;# load Coulomb and LJ table data in parallel movapd xmm0, [rsi + r10*8] ;# Y1c F1c movapd xmm12, [rsi + r11*8] ;# Y2c F2c movapd xmm4, [rsi + r10*8 + 32] ;# Y1d F1d movapd xmm13, [rsi + r11*8 + 32] ;# Y2d F2d movapd xmm8, [rsi + r10*8 + 64] ;# Y1r F1r movapd xmm14, [rsi + r11*8 + 64] ;# Y2r F2r movapd xmm1, xmm0 movapd xmm5, xmm4 movapd xmm9, xmm8 unpcklpd xmm0, xmm12 ;# Y1c Y2c unpckhpd xmm1, xmm12 ;# F1c F2c unpcklpd xmm4, xmm13 ;# Y1d Y2d unpckhpd xmm5, xmm13 ;# F1d F2d unpcklpd xmm8, xmm14 ;# Y1r Y2r unpckhpd xmm9, xmm14 ;# F1r F2r movapd xmm2, [rsi + r10*8 + 16] ;# G1c H1c movapd xmm12, [rsi + r11*8 + 16] ;# G2c H2c movapd xmm6, [rsi + r10*8 + 48] ;# G1d H1d movapd xmm13, [rsi + r11*8 + 48] ;# G2d H2d movapd xmm10, [rsi + r10*8 + 80] ;# G1r H1r movapd xmm14, [rsi + r11*8 + 80] ;# G2r H2r movapd xmm3, xmm2 movapd xmm7, xmm6 movapd xmm11, xmm10 unpcklpd xmm2, xmm12 ;# G1c G2c unpckhpd xmm3, xmm12 ;# H1c H2c unpcklpd xmm6, xmm13 ;# G1d G2d unpckhpd xmm7, xmm13 ;# H1d H2d unpcklpd xmm10, xmm14 ;# G1r G2r unpckhpd xmm11, xmm14 ;# H1r H2r ;# table data ready. Coul in xmm0-xmm3 , disp in xmm4-xmm7 , rep. in xmm8-xmm11 mov rsi, [rbp + nb330_vdwparam] movapd xmm12, [rsp + nb330_eps] mulpd xmm3, xmm12 ;# Heps mulpd xmm7, xmm12 mulpd xmm11, xmm12 mulpd xmm2, xmm12 ;# Geps mulpd xmm6, xmm12 mulpd xmm10, xmm12 mulpd xmm3, xmm12 ;# Heps2 mulpd xmm7, xmm12 mulpd xmm11, xmm12 movlpd xmm13, [rsi + r8*8] ;# c6 movlpd xmm14, [rsi + r8*8 + 8] ;# c12 addpd xmm1, xmm2 ;# F+Geps addpd xmm5, xmm6 addpd xmm9, xmm10 addpd xmm1, xmm3 ;# F+Geps+Heps2 = Fp addpd xmm5, xmm7 addpd xmm9, xmm11 addpd xmm3, xmm3 ;# 2*Heps2 addpd xmm7, xmm7 addpd xmm11, xmm11 addpd xmm3, xmm2 ;# 2*Heps2+Geps addpd xmm7, xmm6 addpd xmm11, xmm10 movhpd xmm13, [rsi + r9*8] ;# c6 movhpd xmm14, [rsi + r9*8 + 8] ;# c12 addpd xmm3, xmm1 ;# FF = Fp + 2*Heps2 + Geps addpd xmm7, xmm5 addpd xmm11, xmm9 mulpd xmm1, xmm12 ;# eps*Fp mulpd xmm5, xmm12 mulpd xmm9, xmm12 addpd xmm1, xmm0 ;# VV addpd xmm5, xmm4 addpd xmm9, xmm8 mulpd xmm1, [rsp + nb330_qq] ;# VV*qq = vcoul mulpd xmm5, xmm13 ;# vnb6 mulpd xmm9, xmm14 ;# vnb12 mulpd xmm3, [rsp + nb330_qq] ;# FF*qq = fij mulpd xmm7, xmm13 ;# fijD mulpd xmm11, xmm14 ;#fijR ;# accumulate vctot addpd xmm1, [rsp + nb330_vctot] movapd [rsp + nb330_vctot], xmm1 ;# accumulate Vvdwtot addpd xmm5, [rsp + nb330_Vvdwtot] addpd xmm5, xmm9 movapd [rsp + nb330_Vvdwtot], xmm5 xorpd xmm9, xmm9 ;# the fj's - start by accumulating forces from memory mov rdi, [rbp + nb330_faction] movlpd xmm4, [rdi + rax*8] movlpd xmm5, [rdi + rax*8 + 8] movlpd xmm6, [rdi + rax*8 + 16] movhpd xmm4, [rdi + rbx*8] movhpd xmm5, [rdi + rbx*8 + 8] movhpd xmm6, [rdi + rbx*8 + 16] addpd xmm3, xmm7 addpd xmm3, xmm11 mulpd xmm3, xmm15 mulpd xmm3, [rsp + nb330_tsc] ;# fscal subpd xmm9, xmm3 movapd xmm10, xmm9 movapd xmm11, xmm9 movapd xmm12, [rsp + nb330_fix] movapd xmm13, [rsp + nb330_fiy] movapd xmm14, [rsp + nb330_fiz] mulpd xmm9, [rsp + nb330_dx] mulpd xmm10, [rsp + nb330_dy] mulpd xmm11, [rsp + nb330_dz] addpd xmm4, xmm9 addpd xmm5, xmm10 addpd xmm6, xmm11 ;# accumulate i forces addpd xmm12, xmm9 addpd xmm13, xmm10 addpd xmm14, xmm11⌨️ 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?