nb_kernel231_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,366 行 · 第 1/5 页

.equiv          nb231nf_charge,           48
.equiv          nb231nf_p_facel,          52
.equiv          nb231nf_argkrf,           56
.equiv          nb231nf_argcrf,           60
.equiv          nb231nf_Vc,               64
.equiv          nb231nf_type,             68
.equiv          nb231nf_p_ntype,          72
.equiv          nb231nf_vdwparam,         76
.equiv          nb231nf_Vvdw,             80
.equiv          nb231nf_p_tabscale,       84
.equiv          nb231nf_VFtab,            88
.equiv          nb231nf_invsqrta,         92
.equiv          nb231nf_dvda,             96
.equiv          nb231nf_p_gbtabscale,     100
.equiv          nb231nf_GBtab,            104
.equiv          nb231nf_p_nthreads,       108
.equiv          nb231nf_count,            112
.equiv          nb231nf_mtx,              116
.equiv          nb231nf_outeriter,        120
.equiv          nb231nf_inneriter,        124
.equiv          nb231nf_work,             128
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse2 use 
.equiv          nb231nf_ixO,              0
.equiv          nb231nf_iyO,              16
.equiv          nb231nf_izO,              32
.equiv          nb231nf_ixH1,             48
.equiv          nb231nf_iyH1,             64
.equiv          nb231nf_izH1,             80
.equiv          nb231nf_ixH2,             96
.equiv          nb231nf_iyH2,             112
.equiv          nb231nf_izH2,             128
.equiv          nb231nf_iqO,              144
.equiv          nb231nf_iqH,              160
.equiv          nb231nf_qqO,              176
.equiv          nb231nf_qqH,              192
.equiv          nb231nf_c6,               208
.equiv          nb231nf_c12,              224
.equiv          nb231nf_tsc,              240
.equiv          nb231nf_vctot,            256
.equiv          nb231nf_Vvdwtot,          272
.equiv          nb231nf_half,             288
.equiv          nb231nf_three,            304
.equiv          nb231nf_two,              320
.equiv          nb231nf_krf,              336
.equiv          nb231nf_crf,              352
.equiv          nb231nf_krsqO,            368
.equiv          nb231nf_krsqH1,           384
.equiv          nb231nf_krsqH2,           400
.equiv          nb231nf_rsqO,             416
.equiv          nb231nf_rinvH1,           432
.equiv          nb231nf_rinvH2,           448
.equiv          nb231nf_is3,              464
.equiv          nb231nf_ii3,              468
.equiv          nb231nf_ntia,             472
.equiv          nb231nf_innerjjnr,        476
.equiv          nb231nf_innerk,           480
.equiv          nb231nf_n,                484
.equiv          nb231nf_nn1,              488
.equiv          nb231nf_nri,              492
.equiv          nb231nf_nouter,           496
.equiv          nb231nf_ninner,           500
.equiv          nb231nf_salign,           504
	push ebp
	mov ebp,esp	
    push eax
    push ebx
    push ecx
    push edx
	push esi
	push edi
	sub esp, 508		;# local stack space 
	mov  eax, esp
	and  eax, 0xf
	sub esp, eax
	mov [esp + nb231nf_salign], eax

	emms

	;# Move args passed by reference to stack
	mov ecx, [ebp + nb231nf_p_nri]
	mov ecx, [ecx]
	mov [esp + nb231nf_nri], ecx

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb231nf_nouter], eax
	mov [esp + nb231nf_ninner], eax

	mov eax, [ebp + nb231nf_p_tabscale]
	movsd xmm3, [eax]
	shufpd xmm3, xmm3, 0
	movapd [esp + nb231nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double 0.5 IEEE (hex)
	mov ebx, 0x3fe00000
	mov [esp + nb231nf_half], eax
	mov [esp + nb231nf_half+4], ebx
	movsd xmm1, [esp + nb231nf_half]
	shufpd xmm1, xmm1, 0    ;# splat to all elements
	movapd xmm3, xmm1
	addpd  xmm3, xmm3       ;# 1.0
	movapd xmm2, xmm3
	addpd  xmm2, xmm2       ;# 2.0
	addpd  xmm3, xmm2	;# 3.0
	movapd [esp + nb231nf_half], xmm1
	movapd [esp + nb231nf_two], xmm2
	movapd [esp + nb231nf_three], xmm3

	mov esi, [ebp + nb231nf_argkrf]
	mov edi, [ebp + nb231nf_argcrf]
	movsd xmm5, [esi]
	movsd xmm6, [edi]
	shufpd xmm5, xmm5, 0
	shufpd xmm6, xmm6, 0
	movapd [esp + nb231nf_krf], xmm5
	movapd [esp + nb231nf_crf], xmm6
	
	;# assume we have at least one i particle - start directly 
	mov   ecx, [ebp + nb231nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx]	    ;# ebx =ii 

	mov   edx, [ebp + nb231nf_charge]
	movsd xmm3, [edx + ebx*8]	
	movsd xmm4, [edx + ebx*8 + 8]	
	mov esi, [ebp + nb231nf_p_facel]
	movsd xmm5, [esi]
	mulsd  xmm3, xmm5
	mulsd  xmm4, xmm5

	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	movapd [esp + nb231nf_iqO], xmm3
	movapd [esp + nb231nf_iqH], xmm4
	
	mov   edx, [ebp + nb231nf_type]
	mov   ecx, [edx + ebx*4]
	shl   ecx, 1
	mov edi, [ebp + nb231nf_p_ntype]
	imul  ecx, [edi]      ;# ecx = ntia = 2*ntype*type[ii0] 
	mov   [esp + nb231nf_ntia], ecx		
.nb231nf_threadloop:
        mov   esi, [ebp + nb231nf_count]          ;# pointer to sync counter
        mov   eax, [esi]
.nb231nf_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 .nb231nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb231nf_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 + nb231nf_n], eax
        mov [esp + nb231nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb231nf_outerstart
        jmp .nb231nf_end

.nb231nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [esp + nb231nf_nouter]
	mov [esp + nb231nf_nouter], ebx

.nb231nf_outer:
	mov   eax, [ebp + nb231nf_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [eax+esi*4]		;# ebx=shift[n] 
	
	lea   ebx, [ebx + ebx*2]    ;# ebx=3*is 
	mov   [esp + nb231nf_is3],ebx    	;# store is3 

	mov   eax, [ebp + nb231nf_shiftvec]   ;# eax = base of shiftvec[] 

	movsd xmm0, [eax + ebx*8]
	movsd xmm1, [eax + ebx*8 + 8]
	movsd xmm2, [eax + ebx*8 + 16] 

	mov   ecx, [ebp + nb231nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx+esi*4]	    ;# ebx =ii 

	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2

	lea   ebx, [ebx + ebx*2]	;# ebx = 3*ii=ii3 
	mov   eax, [ebp + nb231nf_pos]    ;# eax = base of pos[]  
	mov   [esp + nb231nf_ii3], ebx

	addsd xmm3, [eax + ebx*8]
	addsd xmm4, [eax + ebx*8 + 8]
	addsd xmm5, [eax + ebx*8 + 16]		
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	movapd [esp + nb231nf_ixO], xmm3
	movapd [esp + nb231nf_iyO], xmm4
	movapd [esp + nb231nf_izO], xmm5

	movsd xmm3, xmm0
	movsd xmm4, xmm1
	movsd xmm5, xmm2
	addsd xmm0, [eax + ebx*8 + 24]
	addsd xmm1, [eax + ebx*8 + 32]
	addsd xmm2, [eax + ebx*8 + 40]		
	addsd xmm3, [eax + ebx*8 + 48]
	addsd xmm4, [eax + ebx*8 + 56]
	addsd xmm5, [eax + ebx*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 [esp + nb231nf_ixH1], xmm0
	movapd [esp + nb231nf_iyH1], xmm1
	movapd [esp + nb231nf_izH1], xmm2
	movapd [esp + nb231nf_ixH2], xmm3
	movapd [esp + nb231nf_iyH2], xmm4
	movapd [esp + nb231nf_izH2], xmm5
	
	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [esp + nb231nf_vctot], xmm4
	movapd [esp + nb231nf_Vvdwtot], xmm4
	
	mov   eax, [ebp + nb231nf_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 + nb231nf_pos]
	mov   edi, [ebp + nb231nf_faction]	
	mov   eax, [ebp + nb231nf_jjnr]
	shl   ecx, 2
	add   eax, ecx
	mov   [esp + nb231nf_innerjjnr], eax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [esp + nb231nf_ninner]
	mov   [esp + nb231nf_ninner], ecx
	add   edx, 0
	mov   [esp + nb231nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb231nf_unroll_loop
	jmp   .nb231nf_checksingle
.nb231nf_unroll_loop:
	;# twice unrolled innerloop here 
	mov   edx, [esp + nb231nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [edx]	
	mov   ebx, [edx + 4]

	add dword ptr [esp + nb231nf_innerjjnr],  8	;# advance pointer (unrolled 2) 

	mov esi, [ebp + nb231nf_charge]    ;# base of charge[] 
	
	movlpd xmm3, [esi + eax*8]
	movhpd xmm3, [esi + ebx*8]
	movapd xmm4, xmm3
	mulpd  xmm3, [esp + nb231nf_iqO]
	mulpd  xmm4, [esp + nb231nf_iqH]

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx

	movapd  [esp + nb231nf_qqO], xmm3
	movapd  [esp + nb231nf_qqH], xmm4
	
	mov esi, [ebp + nb231nf_type]
	mov eax, [esi + eax*4]
	mov ebx, [esi + ebx*4]
	mov esi, [ebp + nb231nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	mov edi, [esp + nb231nf_ntia]
	add eax, edi
	add ebx, edi

	movlpd xmm6, [esi + eax*8]	;# c6a
	movlpd xmm7, [esi + ebx*8]	;# c6b
	movhpd xmm6, [esi + eax*8 + 8]	;# c6a c12a 
	movhpd xmm7, [esi + ebx*8 + 8]	;# c6b c12b 
	movapd xmm4, xmm6
	unpcklpd xmm4, xmm7
	unpckhpd xmm6, xmm7
	
	movd  eax, mm0
	movd  ebx, mm1
	movapd [esp + nb231nf_c6], xmm4
	movapd [esp + nb231nf_c12], xmm6
	
	mov esi, [ebp + nb231nf_pos]       ;# base of pos[] 

	lea   eax, [eax + eax*2]     ;# replace jnr with j3 
	lea   ebx, [ebx + ebx*2]	

	;# move two coordinates to xmm0-xmm2 
	movlpd xmm0, [esi + eax*8]
	movlpd xmm1, [esi + eax*8 + 8]
	movlpd xmm2, [esi + eax*8 + 16]
	movhpd xmm0, [esi + ebx*8]
	movhpd xmm1, [esi + ebx*8 + 8]
	movhpd xmm2, [esi + ebx*8 + 16]		

	;# move ixO-izO to xmm4-xmm6 
	movapd xmm4, [esp + nb231nf_ixO]
	movapd xmm5, [esp + nb231nf_iyO]
	movapd xmm6, [esp + nb231nf_izO]

	;# calc dr 
	subpd xmm4, xmm0
	subpd xmm5, xmm1
	subpd xmm6, xmm2

	;# square it 
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	mulpd xmm6,xmm6
	addpd xmm4, xmm5
	addpd xmm4, xmm6
	movapd xmm7, xmm4
	;# rsqO in xmm7 
	movapd [esp + nb231nf_rsqO], xmm7
	
	;# move ixH1-izH1 to xmm4-xmm6 
	movapd xmm4, [esp + nb231nf_ixH1]
	movapd xmm5, [esp + nb231nf_iyH1]
	movapd xmm6, [esp + nb231nf_izH1]

	;# calc dr 
	subpd xmm4, xmm0
	subpd xmm5, xmm1
	subpd xmm6, xmm2

	;# square it 
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	mulpd xmm6,xmm6
	addpd xmm6, xmm5
	addpd xmm6, xmm4
	;# rsqH1 in xmm6 

	;# move ixH2-izH2 to xmm3-xmm5  
	movapd xmm3, [esp + nb231nf_ixH2]
	movapd xmm4, [esp + nb231nf_iyH2]
	movapd xmm5, [esp + nb231nf_izH2]

	;# calc dr 
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2

	;# square it 
	mulpd xmm3,xmm3
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	addpd xmm5, xmm4
	addpd xmm5, xmm3
	;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

	movapd xmm0, xmm5
	movapd xmm1, xmm6
	movapd xmm2, xmm7

	mulpd  xmm0, [esp + nb231nf_krf]	
	mulpd  xmm1, [esp + nb231nf_krf]	
	mulpd  xmm2, [esp + nb231nf_krf]	

	movapd [esp + nb231nf_krsqH2], xmm0
	movapd [esp + nb231nf_krsqH1], xmm1
	movapd [esp + nb231nf_krsqO], xmm2
		
	;# start with rsqO - put seed in xmm2 
	cvtpd2ps xmm2, xmm7	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [esp + nb231nf_three]
	mulpd   xmm2, xmm7	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [esp + nb231nf_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm7, xmm4	;# rsq*lu*lu 
	movapd xmm4, [esp + nb231nf_three]
	subpd xmm4, xmm7	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [esp + nb231nf_half] ;# rinv 
	movapd  xmm7, xmm4	;# rinvO in xmm7 
	
	;# rsqH1 - seed in xmm2 
	cvtpd2ps xmm2, xmm6	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [esp + nb231nf_three]
	mulpd   xmm2, xmm6	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [esp + nb231nf_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm6, xmm4	;# rsq*lu*lu 
	movapd xmm4, [esp + nb231nf_three]
	subpd xmm4, xmm6	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [esp + nb231nf_half] ;# rinv 
	movapd  xmm6, xmm4	;# rinvH1 in xmm6 
	movapd  [esp + nb231nf_rinvH1], xmm6
	
	;# rsqH2 - seed in xmm2 
	cvtpd2ps xmm2, xmm5	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [esp + nb231nf_three]
	mulpd   xmm2, xmm5	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [esp + nb231nf_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm5, xmm4	;# rsq*lu*lu 
	movapd xmm4, [esp + nb231nf_three]
	subpd xmm4, xmm5	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [esp + nb231nf_half] ;# rinv 
	movapd  xmm5, xmm4	;# rinvH2 in xmm5 
	movapd  [esp + nb231nf_rinvH2], xmm5

	;# do O interactions 
	movapd xmm0, xmm7
	movapd xmm1, [esp + nb231nf_krsqO]
	addpd  xmm0, xmm1
	subpd  xmm0, [esp + nb231nf_crf] ;# xmm0=rinv+ krsq-crf 
	mulpd  xmm0, [esp + nb231nf_qqO]

	mulpd  xmm6, xmm7

	addpd  xmm0, [esp + nb231nf_vctot]
	movapd [esp + nb231nf_vctot], xmm0

	movapd xmm0, xmm7
	movapd xmm4, [esp + nb231nf_rsqO]
	;# LJ table interaction. xmm0=rinv, xmm4=rsq
	
	mulpd xmm4, xmm0	;# xmm4=r 
	mulpd xmm4, [esp + nb231nf_tsc]
	
	cvttpd2pi mm6, xmm4	;# mm6 = lu idx 
	cvtpi2pd xmm5, mm6
	subpd xmm4, xmm5
	movapd xmm1, xmm4	;# xmm1=eps 
	movapd xmm2, xmm1	
	mulpd  xmm2, xmm2	;# xmm2=eps2