nb_kernel232_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,146 行 · 第 1/5 页

;#
;# $Id: nb_kernel232_ia32_sse.intel_syntax.s,v 1.1.2.1 2006/03/01 15:18:30 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_kernel232_ia32_sse
.globl _nb_kernel232_ia32_sse
nb_kernel232_ia32_sse:	
_nb_kernel232_ia32_sse:	
.equiv          nb232_p_nri,            8
.equiv          nb232_iinr,             12
.equiv          nb232_jindex,           16
.equiv          nb232_jjnr,             20
.equiv          nb232_shift,            24
.equiv          nb232_shiftvec,         28
.equiv          nb232_fshift,           32
.equiv          nb232_gid,              36
.equiv          nb232_pos,              40
.equiv          nb232_faction,          44
.equiv          nb232_charge,           48
.equiv          nb232_p_facel,          52
.equiv          nb232_argkrf,           56
.equiv          nb232_argcrf,           60
.equiv          nb232_Vc,               64
.equiv          nb232_type,             68
.equiv          nb232_p_ntype,          72
.equiv          nb232_vdwparam,         76
.equiv          nb232_Vvdw,             80
.equiv          nb232_p_tabscale,       84
.equiv          nb232_VFtab,            88
.equiv          nb232_invsqrta,         92
.equiv          nb232_dvda,             96
.equiv          nb232_p_gbtabscale,     100
.equiv          nb232_GBtab,            104
.equiv          nb232_p_nthreads,       108
.equiv          nb232_count,            112
.equiv          nb232_mtx,              116
.equiv          nb232_outeriter,        120
.equiv          nb232_inneriter,        124
.equiv          nb232_work,             128
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb232_ixO,              0
.equiv          nb232_iyO,              16
.equiv          nb232_izO,              32
.equiv          nb232_ixH1,             48
.equiv          nb232_iyH1,             64
.equiv          nb232_izH1,             80
.equiv          nb232_ixH2,             96
.equiv          nb232_iyH2,             112
.equiv          nb232_izH2,             128
.equiv          nb232_jxO,              144
.equiv          nb232_jyO,              160
.equiv          nb232_jzO,              176
.equiv          nb232_jxH1,             192
.equiv          nb232_jyH1,             208
.equiv          nb232_jzH1,             224
.equiv          nb232_jxH2,             240
.equiv          nb232_jyH2,             256
.equiv          nb232_jzH2,             272
.equiv          nb232_dxOO,             288
.equiv          nb232_dyOO,             304
.equiv          nb232_dzOO,             320
.equiv          nb232_dxOH1,            336
.equiv          nb232_dyOH1,            352
.equiv          nb232_dzOH1,            368
.equiv          nb232_dxOH2,            384
.equiv          nb232_dyOH2,            400
.equiv          nb232_dzOH2,            416
.equiv          nb232_dxH1O,            432
.equiv          nb232_dyH1O,            448
.equiv          nb232_dzH1O,            464
.equiv          nb232_dxH1H1,           480
.equiv          nb232_dyH1H1,           496
.equiv          nb232_dzH1H1,           512
.equiv          nb232_dxH1H2,           528
.equiv          nb232_dyH1H2,           544
.equiv          nb232_dzH1H2,           560
.equiv          nb232_dxH2O,            576
.equiv          nb232_dyH2O,            592
.equiv          nb232_dzH2O,            608
.equiv          nb232_dxH2H1,           624
.equiv          nb232_dyH2H1,           640
.equiv          nb232_dzH2H1,           656
.equiv          nb232_dxH2H2,           672
.equiv          nb232_dyH2H2,           688
.equiv          nb232_dzH2H2,           704
.equiv          nb232_qqOO,             720
.equiv          nb232_qqOH,             736
.equiv          nb232_qqHH,             752
.equiv          nb232_c6,               768
.equiv          nb232_c12,              784
.equiv          nb232_tsc,              800
.equiv          nb232_fstmp,            816
.equiv          nb232_vctot,            832
.equiv          nb232_Vvdwtot,          848
.equiv          nb232_fixO,             864
.equiv          nb232_fiyO,             880
.equiv          nb232_fizO,             896
.equiv          nb232_fixH1,            912
.equiv          nb232_fiyH1,            928
.equiv          nb232_fizH1,            944
.equiv          nb232_fixH2,            960
.equiv          nb232_fiyH2,            976
.equiv          nb232_fizH2,            992
.equiv          nb232_fjxO,             1008
.equiv          nb232_fjyO,             1024
.equiv          nb232_fjzO,             1040
.equiv          nb232_fjxH1,            1056
.equiv          nb232_fjyH1,            1072
.equiv          nb232_fjzH1,            1088
.equiv          nb232_fjxH2,            1104
.equiv          nb232_fjyH2,            1120
.equiv          nb232_fjzH2,            1136
.equiv          nb232_fjzH2b,           1140
.equiv          nb232_fjzH2c,           1144
.equiv          nb232_fjzH2d,           1148
.equiv          nb232_half,             1152
.equiv          nb232_three,            1168
.equiv          nb232_rsqOO,            1184
.equiv          nb232_rsqOH1,           1200
.equiv          nb232_rsqOH2,           1216
.equiv          nb232_rsqH1O,           1232
.equiv          nb232_rsqH1H1,          1248
.equiv          nb232_rsqH1H2,          1264
.equiv          nb232_rsqH2O,           1280
.equiv          nb232_rsqH2H1,          1296
.equiv          nb232_rsqH2H2,          1312
.equiv          nb232_rinvOO,           1328
.equiv          nb232_rinvOH1,          1344
.equiv          nb232_rinvOH2,          1360
.equiv          nb232_rinvH1O,          1376
.equiv          nb232_rinvH1H1,         1392
.equiv          nb232_rinvH1H2,         1408
.equiv          nb232_rinvH2O,          1424
.equiv          nb232_rinvH2H1,         1440
.equiv          nb232_rinvH2H2,         1456
.equiv          nb232_two,              1472
.equiv          nb232_krf,              1488
.equiv          nb232_crf,              1504
.equiv          nb232_is3,              1520
.equiv          nb232_ii3,              1524
.equiv          nb232_innerjjnr,        1528
.equiv          nb232_innerk,           1532
.equiv          nb232_n,                1536
.equiv          nb232_nn1,              1540
.equiv          nb232_nri,              1544
.equiv          nb232_nouter,           1548
.equiv          nb232_ninner,           1552
.equiv          nb232_salign,           1556
	push ebp
	mov ebp,esp	
    	push eax
    	push ebx
    	push ecx
    	push edx
	push esi
	push edi
	sub esp, 1560		;# local stack space 
	mov  eax, esp
	and  eax, 0xf
	sub esp, eax
	mov [esp + nb232_salign], eax

	emms

	;# Move args passed by reference to stack
	mov ecx, [ebp + nb232_p_nri]
	mov ecx, [ecx]
	mov [esp + nb232_nri], ecx

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb232_nouter], eax
	mov [esp + nb232_ninner], eax

	mov eax, [ebp + nb232_p_tabscale]
	movss xmm3, [eax]
	shufps xmm3, xmm3, 0
	movaps [esp + nb232_tsc], xmm3

	mov esi, [ebp + nb232_argkrf]
	mov edi, [ebp + nb232_argcrf]
	movss xmm5, [esi]
	movss xmm6, [edi]
	shufps xmm5, xmm5, 0
	shufps xmm6, xmm6, 0
	movaps [esp + nb232_krf], xmm5
	movaps [esp + nb232_crf], xmm6
	
	;# assume we have at least one i particle - start directly 
	mov   ecx, [ebp + nb232_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx]	    ;# ebx =ii 

	mov   edx, [ebp + nb232_charge]
	movss xmm3, [edx + ebx*4]	
	movss xmm4, xmm3	
	movss xmm5, [edx + ebx*4 + 4]	
	mov esi, [ebp + nb232_p_facel]
	movss xmm6, [esi]
	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 [esp + nb232_qqOO], xmm3
	movaps [esp + nb232_qqOH], xmm4
	movaps [esp + nb232_qqHH], xmm5
		
	xorps xmm0, xmm0
	mov   edx, [ebp + nb232_type]
	mov   ecx, [edx + ebx*4]
	shl   ecx, 1
	mov   edx, ecx
	mov edi, [ebp + nb232_p_ntype]
	imul  ecx, [edi]      ;# ecx = ntia = 2*ntype*type[ii0] 
	add   edx, ecx
	mov   eax, [ebp + nb232_vdwparam]
	movlps xmm0, [eax + edx*4] 
	movaps xmm1, xmm0
	shufps xmm0, xmm0, 0
	shufps xmm1, xmm1, 85  ;# constant 01010101
	movaps [esp + nb232_c6], xmm0
	movaps [esp + nb232_c12], xmm1

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# constant 0.5 in IEEE (hex)
	mov [esp + nb232_half], eax
	movss xmm1, [esp + nb232_half]
	shufps xmm1, xmm1, 0    ;# splat to all elements
	movaps xmm2, xmm1       
	addps  xmm2, xmm2	;# constant 1.0
	movaps xmm3, xmm2
	addps  xmm2, xmm2	;# constant 2.0
	addps  xmm3, xmm2	;# constant 3.0
	movaps [esp + nb232_half],  xmm1
	movaps [esp + nb232_two],  xmm2
	movaps [esp + nb232_three],  xmm3

.nb232_threadloop:
        mov   esi, [ebp + nb232_count]          ;# pointer to sync counter
        mov   eax, [esi]
.nb232_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 .nb232_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb232_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 [esp + nb232_n], eax
        mov [esp + nb232_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb232_outerstart
        jmp .nb232_end

.nb232_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [esp + nb232_nouter]
	mov [esp + nb232_nouter], ebx

.nb232_outer:
	mov   eax, [ebp + nb232_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [eax + esi*4]		;# ebx=shift[n] 
	
	lea   ebx, [ebx + ebx*2]    ;# ebx=3*is 
	mov   [esp + nb232_is3],ebx    	;# store is3 

	mov   eax, [ebp + nb232_shiftvec]   ;# eax = base of shiftvec[] 

	movss xmm0, [eax + ebx*4]
	movss xmm1, [eax + ebx*4 + 4]
	movss xmm2, [eax + ebx*4 + 8] 

	mov   ecx, [ebp + nb232_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx + esi*4]	    ;# ebx =ii 

	lea   ebx, [ebx + ebx*2]	;# ebx = 3*ii=ii3 
	mov   eax, [ebp + nb232_pos]    ;# eax = base of pos[]  
	mov   [esp + nb232_ii3], ebx	
	
	movaps xmm3, xmm0
	movaps xmm4, xmm1
	movaps xmm5, xmm2
	addss xmm3, [eax + ebx*4]
	addss xmm4, [eax + ebx*4 + 4]
	addss xmm5, [eax + ebx*4 + 8]		
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	movaps [esp + nb232_ixO], xmm3
	movaps [esp + nb232_iyO], xmm4
	movaps [esp + nb232_izO], xmm5

	movss xmm3, xmm0
	movss xmm4, xmm1
	movss xmm5, xmm2
	addss xmm0, [eax + ebx*4 + 12]
	addss xmm1, [eax + ebx*4 + 16]
	addss xmm2, [eax + ebx*4 + 20]		
	addss xmm3, [eax + ebx*4 + 24]
	addss xmm4, [eax + ebx*4 + 28]
	addss xmm5, [eax + ebx*4 + 32]		

	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 [esp + nb232_ixH1], xmm0
	movaps [esp + nb232_iyH1], xmm1
	movaps [esp + nb232_izH1], xmm2
	movaps [esp + nb232_ixH2], xmm3
	movaps [esp + nb232_iyH2], xmm4
	movaps [esp + nb232_izH2], xmm5

	;# clear vctot and i forces 
	xorps xmm4, xmm4
	movaps [esp + nb232_vctot], xmm4
	movaps [esp + nb232_Vvdwtot], xmm4
	movaps [esp + nb232_fixO], xmm4
	movaps [esp + nb232_fiyO], xmm4
	movaps [esp + nb232_fizO], xmm4
	movaps [esp + nb232_fixH1], xmm4
	movaps [esp + nb232_fiyH1], xmm4
	movaps [esp + nb232_fizH1], xmm4
	movaps [esp + nb232_fixH2], xmm4
	movaps [esp + nb232_fiyH2], xmm4
	movaps [esp + nb232_fizH2], xmm4
	
	mov   eax, [ebp + nb232_jindex]
	mov   ecx, [eax + esi*4]	     ;# jindex[n] 
	mov   edx, [eax + esi*4 + 4]	     ;# jindex[n+1] 
	sub   edx, ecx               ;# number of innerloop atoms 

	mov   esi, [ebp + nb232_pos]
	mov   edi, [ebp + nb232_faction]	
	mov   eax, [ebp + nb232_jjnr]
	shl   ecx, 2
	add   eax, ecx
	mov   [esp + nb232_innerjjnr], eax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  4
	add   ecx, [esp + nb232_ninner]
	mov   [esp + nb232_ninner], ecx
	add   edx, 0
	mov   [esp + nb232_innerk], edx    ;# number of innerloop atoms 
	jge   .nb232_unroll_loop
	jmp   .nb232_single_check
.nb232_unroll_loop:	
	;# quad-unroll innerloop here 
	mov   edx, [esp + nb232_innerjjnr]     ;# pointer to jjnr[k] 

	mov   eax, [edx]	
	mov   ebx, [edx + 4] 
	mov   ecx, [edx + 8]
	mov   edx, [edx + 12]         ;# eax-edx=jnr1-4 
	
	add dword ptr [esp + nb232_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	mov esi, [ebp + nb232_pos]       ;# base of pos[] 

	lea   eax, [eax + eax*2]     ;# replace jnr with j3 
	lea   ebx, [ebx + ebx*2]	
	lea   ecx, [ecx + ecx*2]     ;# replace jnr with j3 
	lea   edx, [edx + edx*2]	
	
	;# move j coordinates to local temp variables 
	movlps xmm2, [esi + eax*4]
	movlps xmm3, [esi + eax*4 + 12]
	movlps xmm4, [esi + eax*4 + 24]

	movlps xmm5, [esi + ebx*4]