nb_kernel331_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,337 行 · 第 1/5 页

	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addsd  xmm0, xmm3
	addsd  xmm1, xmm4
	addsd  xmm2, xmm5 ;# sum is in low xmm0-xmm2 

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

	;# increment i force 
	movsd  xmm3, [edi + ecx*8 + 48]
	movsd  xmm4, [edi + ecx*8 + 56]
	movsd  xmm5, [edi + ecx*8 + 64]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm5, xmm2
	movsd  [edi + ecx*8 + 48], xmm3
	movsd  [edi + ecx*8 + 56], xmm4
	movsd  [edi + ecx*8 + 64], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	addsd xmm7, xmm2
	unpcklpd xmm0, xmm1
	addpd xmm6, xmm0

	;# increment fshift force 
	movlpd xmm3, [esi + edx*8]
	movhpd xmm3, [esi + edx*8 + 8]
	movsd  xmm4, [esi + edx*8 + 16]
	addpd  xmm3, xmm6
	addsd  xmm4, xmm7
	movlpd [esi + edx*8],      xmm3
	movhpd [esi + edx*8 + 8],  xmm3
	movsd  [esi + edx*8 + 16], xmm4

	;# get n from stack
	mov esi, [esp + nb331_n]
        ;# get group index for i particle 
        mov   edx, [ebp + nb331_gid]      	;# base of gid[]
        mov   edx, [edx + esi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movapd xmm7, [esp + nb331_vctot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 
        
	;# add earlier value from mem 
	mov   eax, [ebp + nb331_Vc]
	addsd xmm7, [eax + edx*8] 
	;# move back to mem 
	movsd [eax + edx*8], xmm7 
	
	;# accumulate total lj energy and update it 
	movapd xmm7, [esp + nb331_Vvdwtot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 

	;# add earlier value from mem 
	mov   eax, [ebp + nb331_Vvdw]
	addsd xmm7, [eax + edx*8] 
	;# move back to mem 
	movsd [eax + edx*8], xmm7 
	
        ;# finish if last 
        mov ecx, [esp + nb331_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb331_outerend

        ;# not last, iterate outer loop once more!  
        mov [esp + nb331_n], esi
        jmp .nb331_outer
.nb331_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [esp + nb331_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb331_end
        ;# non-zero, do one more workunit
        jmp   .nb331_threadloop
.nb331_end:
	emms

	mov eax, [esp + nb331_nouter]
	mov ebx, [esp + nb331_ninner]
	mov ecx, [ebp + nb331_outeriter]
	mov edx, [ebp + nb331_inneriter]
	mov [ecx], eax
	mov [edx], ebx

	mov eax, [esp + nb331_salign]
	add esp, eax
	add esp, 812
	pop edi
	pop esi
    	pop edx
    	pop ecx
    	pop ebx
    	pop eax
	leave
	ret
	



.globl nb_kernel331nf_ia32_sse2
.globl _nb_kernel331nf_ia32_sse2
nb_kernel331nf_ia32_sse2:	
_nb_kernel331nf_ia32_sse2:	
.equiv          nb331nf_p_nri,          8
.equiv          nb331nf_iinr,           12
.equiv          nb331nf_jindex,         16
.equiv          nb331nf_jjnr,           20
.equiv          nb331nf_shift,          24
.equiv          nb331nf_shiftvec,       28
.equiv          nb331nf_fshift,         32
.equiv          nb331nf_gid,            36
.equiv          nb331nf_pos,            40
.equiv          nb331nf_faction,        44
.equiv          nb331nf_charge,         48
.equiv          nb331nf_p_facel,        52
.equiv          nb331nf_argkrf,         56
.equiv          nb331nf_argcrf,         60
.equiv          nb331nf_Vc,             64
.equiv          nb331nf_type,           68
.equiv          nb331nf_p_ntype,        72
.equiv          nb331nf_vdwparam,       76
.equiv          nb331nf_Vvdw,           80
.equiv          nb331nf_p_tabscale,     84
.equiv          nb331nf_VFtab,          88
.equiv          nb331nf_invsqrta,       92
.equiv          nb331nf_dvda,           96
.equiv          nb331nf_p_gbtabscale,   100
.equiv          nb331nf_GBtab,          104
.equiv          nb331nf_p_nthreads,     108
.equiv          nb331nf_count,          112
.equiv          nb331nf_mtx,            116
.equiv          nb331nf_outeriter,      120
.equiv          nb331nf_inneriter,      124
.equiv          nb331nf_work,           128
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb331nf_ixO,            0
.equiv          nb331nf_iyO,            16
.equiv          nb331nf_izO,            32
.equiv          nb331nf_ixH1,           48
.equiv          nb331nf_iyH1,           64
.equiv          nb331nf_izH1,           80
.equiv          nb331nf_ixH2,           96
.equiv          nb331nf_iyH2,           112
.equiv          nb331nf_izH2,           128
.equiv          nb331nf_iqO,            144
.equiv          nb331nf_iqH,            160
.equiv          nb331nf_qqO,            176
.equiv          nb331nf_qqH,            192
.equiv          nb331nf_rinvO,          208
.equiv          nb331nf_rinvH1,         224
.equiv          nb331nf_rinvH2,         240
.equiv          nb331nf_rO,             256
.equiv          nb331nf_rH1,            272
.equiv          nb331nf_rH2,            288
.equiv          nb331nf_tsc,            304
.equiv          nb331nf_c6,             320
.equiv          nb331nf_c12,            336
.equiv          nb331nf_vctot,          352
.equiv          nb331nf_Vvdwtot,        368
.equiv          nb331nf_half,           384
.equiv          nb331nf_three,          400
.equiv          nb331nf_is3,            416
.equiv          nb331nf_ii3,            420
.equiv          nb331nf_ntia,           424
.equiv          nb331nf_innerjjnr,      428
.equiv          nb331nf_innerk,         432
.equiv          nb331nf_n,              436
.equiv          nb331nf_nn1,            440
.equiv          nb331nf_nri,            444
.equiv          nb331nf_nouter,         448
.equiv          nb331nf_ninner,         452
.equiv          nb331nf_salign,         456
	push ebp
	mov ebp,esp	
    	push eax
    	push ebx
    	push ecx
    	push edx
	push esi
	push edi
	sub esp, 460		;# local stack space 
	mov  eax, esp
	and  eax, 0xf
	sub esp, eax
	mov [esp + nb331nf_salign], eax

	emms

	;# Move args passed by reference to stack
	mov ecx, [ebp + nb331nf_p_nri]
	mov ecx, [ecx]
	mov [esp + nb331nf_nri], ecx

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb331nf_nouter], eax
	mov [esp + nb331nf_ninner], eax


	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double 0.5 IEEE (hex)
	mov ebx, 0x3fe00000
	mov [esp + nb331nf_half], eax
	mov [esp + nb331nf_half+4], ebx
	movsd xmm1, [esp + nb331nf_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 + nb331nf_half], xmm1
	movapd [esp + nb331nf_three], xmm3
	mov eax, [ebp + nb331nf_p_tabscale]
	movsd xmm5, [eax]
	shufpd xmm5, xmm5, 0
	movapd [esp + nb331nf_tsc], xmm5
	
	;# assume we have at least one i particle - start directly 
	mov   ecx, [ebp + nb331nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx]	    ;# ebx =ii 

	mov   edx, [ebp + nb331nf_charge]
	movsd xmm3, [edx + ebx*8]	
	movsd xmm4, [edx + ebx*8 + 8]	
	mov esi, [ebp + nb331nf_p_facel]
	movsd xmm5, [esi]
	mulsd  xmm3, xmm5
	mulsd  xmm4, xmm5

	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	movapd [esp + nb331nf_iqO], xmm3
	movapd [esp + nb331nf_iqH], xmm4
	
	mov   edx, [ebp + nb331nf_type]
	mov   ecx, [edx + ebx*4]
	shl   ecx, 1
	mov edi, [ebp + nb331nf_p_ntype]
	imul  ecx, [edi]      ;# ecx = ntia = 2*ntype*type[ii0] 
	mov   [esp + nb331nf_ntia], ecx		
.nb331nf_threadloop:
        mov   esi, [ebp + nb331nf_count]        ;# pointer to sync counter
        mov   eax, [esi]
.nb331nf_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 .nb331nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb331nf_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 + nb331nf_n], eax
        mov [esp + nb331nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb331nf_outerstart
        jmp .nb331nf_end

.nb331nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [esp + nb331nf_nouter]
	mov [esp + nb331nf_nouter], ebx

.nb331nf_outer:
	mov   eax, [ebp + nb331nf_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [eax + esi*4]		;# ebx=shift[n] 
	
	lea   ebx, [ebx + ebx*2]    ;# ebx=3*is 

	mov   eax, [ebp + nb331nf_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 + nb331nf_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 + nb331nf_pos]    ;# eax = base of pos[]  
	mov   [esp + nb331nf_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 + nb331nf_ixO], xmm3
	movapd [esp + nb331nf_iyO], xmm4
	movapd [esp + nb331nf_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 + nb331nf_ixH1], xmm0
	movapd [esp + nb331nf_iyH1], xmm1
	movapd [esp + nb331nf_izH1], xmm2
	movapd [esp + nb331nf_ixH2], xmm3
	movapd [esp + nb331nf_iyH2], xmm4
	movapd [esp + nb331nf_izH2], xmm5
	
	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [esp + nb331nf_vctot], xmm4
	movapd [esp + nb331nf_Vvdwtot], xmm4
	
	mov   eax, [ebp + nb331nf_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 + nb331nf_pos]
	mov   eax, [ebp + nb331nf_jjnr]
	shl   ecx, 2
	add   eax, ecx
	mov   [esp + nb331nf_innerjjnr], eax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [esp + nb331nf_ninner]
	mov   [esp + nb331nf_ninner], ecx
	add   edx, 0
	mov   [esp + nb331nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb331nf_unroll_loop
	jmp   .nb331nf_checksingle
.nb331nf_unroll_loop:
	;# twice unrolled innerloop here 
	mov   edx, [esp + nb331nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [edx]	
	mov   ebx, [edx + 4]
	
	add dword ptr [esp + nb331nf_innerjjnr],  8 ;# advance pointer (unrolled 2) 

	mov esi, [ebp + nb331nf_charge]    ;# base of charge[] 
	
	movlpd xmm3, [esi + eax*8]
	movhpd xmm3, [esi + ebx*8]
	movapd xmm4, xmm3	     
	mulpd  xmm3, [esp + nb331nf_iqO]
	mulpd  xmm4, [esp + nb331nf_iqH]
	movapd  [esp + nb331nf_qqO], xmm3
	movapd  [esp + nb331nf_qqH], xmm4	

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx
	mov esi, [ebp + nb331nf_type]
	mov eax, [esi + eax*4]
	mov ebx, [esi + ebx*4]
	mov esi, [ebp + nb331nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	mov edi, [esp + nb331nf_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 + nb331nf_c6], xmm4
	movapd [esp + nb331nf_c12], xmm6

	mov esi, [ebp + nb331nf_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 + nb331nf_ixO]
	movapd xmm5, [esp + nb331nf_iyO]
	movapd xmm6, [esp + nb331nf_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 

	;# move ixH1-izH1 to xmm4-xmm6 
	movapd xmm4, [esp + nb331nf_ixH1]
	movapd xmm5, [esp + nb331nf_iyH1]
	movapd xmm6, [esp + nb331nf_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 + nb331nf_ixH2]
	movapd xmm4, [esp + nb331nf_iyH2]
	movapd xmm5, [esp + nb331nf_izH2]

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