nb_kernel102_ia32_3dnow.intel_syntax.s

最著名最快的分子模拟软件

.equiv		nb102nf_qqHH,		60     	
.equiv		nb102nf_vctot,		68
.equiv		nb102nf_innerjjnr,	76
.equiv		nb102nf_innerk,		80       
.equiv          nb102nf_n,              84 ;# idx for outer loop
.equiv          nb102nf_nn1,            88 ;# number of outer iterations
.equiv          nb102nf_nri,            92
.equiv          nb102nf_nouter,         96
.equiv          nb102nf_ninner,         100
	push ebp
	mov ebp,esp	
    	push eax
    	push ebx
    	push ecx
    	push edx
	push esi
	push edi
	sub esp, 104		;# local stack space 
	femms
	mov ecx, [ebp + nb102nf_p_nri]
	mov esi, [ebp + nb102nf_p_facel]
	mov ecx, [ecx]
	mov [esp + nb102nf_nri], ecx

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb102nf_nouter], eax
	mov [esp + nb102nf_ninner], eax
	;# assume we have at least one i particle - start directly 	

	mov   ecx, [ebp + nb102nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx]	    ;# ebx=ii 

	mov   edx, [ebp + nb102nf_charge]
	movd  mm1, [esi]	;# mm1=facel 
	movd  mm2, [edx + ebx*4]    ;# mm2=charge[ii0] (O) 
	movd  mm3, [edx + ebx*4 + 4]    ;# mm2=charge[ii0+1] (H)  
	movq  mm4, mm2	
	pfmul mm4, mm1
	movq  mm6, mm3
	pfmul mm6, mm1
	movq  mm5, mm4
	pfmul mm4, mm2			;# mm4=qqOO*facel 
	pfmul mm5, mm3			;# mm5=qqOH*facel 
	pfmul mm6, mm3			;# mm6=qqHH*facel 
	punpckldq mm5,mm5	    ;# spread to both halves 
	punpckldq mm6,mm6	    ;# spread to both halves 
	movq  [esp + nb102nf_qqOO], mm4
	movq  [esp + nb102nf_qqOH], mm5
	movq  [esp + nb102nf_qqHH], mm6
.nb102nf_threadloop:
        mov   esi, [ebp + nb102nf_count]          ;# pointer to sync counter
        mov   eax, [esi]
.nb102nf_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 .nb102nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb102nf_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 + nb102nf_n], eax
        mov [esp + nb102nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb102nf_outerstart
        jmp .nb102nf_end

.nb102nf_outerstart:	
	;# ebx contains number of outer iterations
	add ebx, [esp + nb102nf_nouter]
        mov [esp + nb102nf_nouter], ebx
	
.nb102nf_outer:
	mov   eax, [ebp + nb102nf_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [eax + esi*4]		;# ebx=shift[n] 
	
	lea   ebx, [ebx + ebx*2]    ;# ebx=3*is 
	mov   [esp + nb102nf_is3],ebx    	;# store is3 

	mov   eax, [ebp + nb102nf_shiftvec]   ;# eax = base of shiftvec[] 
	
	movq  mm5, [eax + ebx*4]	;# move shX/shY to mm5 and shZ to mm6 
	movd  mm6, [eax + ebx*4 + 8]
	movq  mm0, mm5
	movq  mm1, mm5
	movq  mm2, mm6
	punpckldq mm0,mm0	    ;# also expand shX,Y,Z in mm0--mm2 
	punpckhdq mm1,mm1
	punpckldq mm2,mm2		
	
	mov   ecx, [ebp + nb102nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx + esi*4]	    ;# ebx=ii 

	lea   ebx, [ebx + ebx*2]	;# ebx = 3*ii=ii3 
	mov   eax, [ebp + nb102nf_pos]    ;# eax = base of pos[] 

	pfadd mm5, [eax + ebx*4]    ;# ix = shX + posX (and iy too) 
	movd  mm7, [eax + ebx*4 + 8]    ;# cant use direct memory add for 4 bytes (iz) 
	mov   [esp + nb102nf_ii3], ebx	    ;# (use mm7 as temp storage for iz) 
	pfadd mm6, mm7
	movq  [esp + nb102nf_ixO], mm5	
	movq  [esp + nb102nf_izO], mm6

	movd  mm3, [eax + ebx*4 + 12]
	movd  mm4, [eax + ebx*4 + 16]
	movd  mm5, [eax + ebx*4 + 20]
	punpckldq  mm3, [eax + ebx*4 + 24]
	punpckldq  mm4, [eax + ebx*4 + 28]
	punpckldq  mm5, [eax + ebx*4 + 32] ;# coords of H1 in low mm3-mm5, H2 in high 
	
	pfadd mm0, mm3
	pfadd mm1, mm4
	pfadd mm2, mm5		
	movq [esp + nb102nf_ixH], mm0	
	movq [esp + nb102nf_iyH], mm1	
	movq [esp + nb102nf_izH], mm2	

	;# clear vctot
	pxor  mm7,mm7
	movq  [esp + nb102nf_vctot], mm7

	mov   eax, [ebp + nb102nf_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   [esp + nb102nf_innerk], edx    ;# number of innerloop atoms 
	add   edx, [esp + nb102nf_ninner]
	mov   [esp + nb102nf_ninner], edx

	mov   esi, [ebp + nb102nf_pos]
	mov   edi, [ebp + nb102nf_faction]	
	mov   eax, [ebp + nb102nf_jjnr]
	shl   ecx, 2
	add   eax, ecx
	mov   [esp + nb102nf_innerjjnr], eax     ;# pointer to jjnr[nj0] 
.nb102nf_inner_loop:
	;# a single j particle iteration here - compare with the unrolled code for comments 
	mov   eax, [esp + nb102nf_innerjjnr]
	mov   eax, [eax]	;# eax=jnr offset 
    	add dword ptr [esp + nb102nf_innerjjnr],  4 ;# advance pointer 

	movd  mm6, [esp + nb102nf_qqOO]
	movq  mm7, [esp + nb102nf_qqOH]

	lea   eax, [eax + eax*2]
	movq  mm0, [esi + eax*4]
	movd  mm1, [esi + eax*4 + 8]
	;# copy & expand to mm2-mm4 for the H interactions 
	movq  mm2, mm0
	movq  mm3, mm0
	movq  mm4, mm1
	punpckldq mm2,mm2
	punpckhdq mm3,mm3
	punpckldq mm4,mm4
	
	pfsubr mm0, [esp + nb102nf_ixO]
	pfsubr mm1, [esp + nb102nf_izO]
		
	pfmul mm0,mm0
	pfmul mm1,mm1
	pfacc mm0, mm0
	pfadd mm0, mm1		;# mm0=rsqO 
	
	punpckldq mm2, mm2
	punpckldq mm3, mm3
	punpckldq mm4, mm4  ;# mm2-mm4 is jx-jz 
	pfsubr mm2, [esp + nb102nf_ixH]
	pfsubr mm3, [esp + nb102nf_iyH]
	pfsubr mm4, [esp + nb102nf_izH] ;# mm2-mm4 is dxH-dzH 
	
	pfmul mm2,mm2
	pfmul mm3,mm3
	pfmul mm4,mm4

	pfadd mm3,mm2
	pfadd mm3,mm4		;# mm3=rsqH 

    	pfrsqrt mm1,mm0

    	movq mm2,mm1
    	pfmul mm1,mm1
    	pfrsqit1 mm1,mm0				
    	pfrcpit2 mm1,mm2	;# mm1=invsqrt 
	;# calculate potential 
	pfmul mm6, mm1		;# mm6=vcoul 

	pfrsqrt mm5, mm3
	pswapd mm3,mm3
	pfrsqrt mm2, mm3
	pswapd mm3,mm3
	punpckldq mm5,mm2	;# seeds are in mm5 now, and rsq in mm3 

	movq mm2, mm5
	pfmul mm5,mm5
    	pfrsqit1 mm5,mm3				
    	pfrcpit2 mm5,mm2	;# mm5=invsqrt 
	pfmul mm7, mm5		;# mm7=vcoul 

	;# update vctot 
	pfadd mm7, mm6
	pfadd mm7, [esp + nb102nf_vctot]
	movq [esp + nb102nf_vctot], mm7
	
	;# interactions with j H1 
	movq  mm0, [esi + eax*4 + 12]
	movd  mm1, [esi + eax*4 + 20]
	;# copy & expand to mm2-mm4 for the H interactions 
	movq  mm2, mm0
	movq  mm3, mm0
	movq  mm4, mm1
	punpckldq mm2,mm2
	punpckhdq mm3,mm3
	punpckldq mm4,mm4
	
	movd mm6, [esp + nb102nf_qqOH]
	movq mm7, [esp + nb102nf_qqHH]
	
	pfsubr mm0, [esp + nb102nf_ixO]
	pfsubr mm1, [esp + nb102nf_izO]
		
	pfmul mm0,mm0
	pfmul mm1,mm1
	pfacc mm0, mm1
	pfadd mm0, mm1		;# mm0=rsqO 
	
	punpckldq mm2, mm2
	punpckldq mm3, mm3
	punpckldq mm4, mm4  ;# mm2-mm4 is jx-jz 
	pfsubr mm2, [esp + nb102nf_ixH]
	pfsubr mm3, [esp + nb102nf_iyH]
	pfsubr mm4, [esp + nb102nf_izH] ;# mm2-mm4 is dxH-dzH 
	
	pfmul mm2,mm2
	pfmul mm3,mm3
	pfmul mm4,mm4

	pfadd mm3,mm2
	pfadd mm3,mm4		;# mm3=rsqH 

    	pfrsqrt mm1,mm0

    	movq mm2,mm1
    	pfmul mm1,mm1
    	pfrsqit1 mm1,mm0				
    	pfrcpit2 mm1,mm2	;# mm1=invsqrt 
	;# calculate potential 
	pfmul mm6, mm1		;# mm6=vcoul 

	pfrsqrt mm5, mm3
	pswapd mm3,mm3
	pfrsqrt mm2, mm3
	pswapd mm3,mm3
	punpckldq mm5,mm2	;# seeds are in mm5 now, and rsq in mm3 

	movq mm2, mm5
	pfmul mm5,mm5
    	pfrsqit1 mm5,mm3				
    	pfrcpit2 mm5,mm2	;# mm5=invsqrt 
	pfmul mm7, mm5		;# mm7=vcoul 

	;# update vctot 
	pfadd mm7, mm6
	pfadd mm7, [esp + nb102nf_vctot]
	movq [esp + nb102nf_vctot], mm7
	
	;# interactions with j H2 
	movq  mm0, [esi + eax*4 + 24]
	movd  mm1, [esi + eax*4 + 32]
	;# copy & expand to mm2-mm4 for the H interactions 
	movq  mm2, mm0
	movq  mm3, mm0
	movq  mm4, mm1
	punpckldq mm2,mm2
	punpckhdq mm3,mm3
	punpckldq mm4,mm4

	movd mm6, [esp + nb102nf_qqOH]
	movq mm7, [esp + nb102nf_qqHH]

	pfsubr mm0, [esp + nb102nf_ixO]
	pfsubr mm1, [esp + nb102nf_izO]
		
	pfmul mm0,mm0
	pfmul mm1,mm1
	pfacc mm0, mm1
	pfadd mm0, mm1		;# mm0=rsqO 
	
	punpckldq mm2, mm2
	punpckldq mm3, mm3
	punpckldq mm4, mm4  ;# mm2-mm4 is jx-jz 
	pfsubr mm2, [esp + nb102nf_ixH]
	pfsubr mm3, [esp + nb102nf_iyH]
	pfsubr mm4, [esp + nb102nf_izH] ;# mm2-mm4 is dxH-dzH 
	
	pfmul mm2,mm2
	pfmul mm3,mm3
	pfmul mm4,mm4

	pfadd mm3,mm2
	pfadd mm3,mm4		;# mm3=rsqH 

    	pfrsqrt mm1,mm0

    	movq mm2,mm1
    	pfmul mm1,mm1
    	pfrsqit1 mm1,mm0				
    	pfrcpit2 mm1,mm2	;# mm1=invsqrt 
	;# calculate potential 
	pfmul mm6, mm1		;# mm6=vcoul 

	pfrsqrt mm5, mm3
	pswapd mm3,mm3
	pfrsqrt mm2, mm3
	pswapd mm3,mm3
	punpckldq mm5,mm2	;# seeds are in mm5 now, and rsq in mm3 

	movq mm2, mm5
	pfmul mm5,mm5
    	pfrsqit1 mm5,mm3				
    	pfrcpit2 mm5,mm2	;# mm5=invsqrt 
	pfmul mm7, mm5		;# mm7=vcoul 

	;# update vctot 
	pfadd mm7, mm6
	pfadd mm7, [esp + nb102nf_vctot]
	movq [esp + nb102nf_vctot], mm7
	
	;#  done  - one more? 
	dec dword ptr [esp + nb102nf_innerk]
	jz  .nb102nf_updateouterdata
	jmp .nb102nf_inner_loop	
.nb102nf_updateouterdata:	
	;# get n from stack
	mov esi, [esp + nb102nf_n]
        ;# get group index for i particle 
        mov   edx, [ebp + nb102nf_gid]      	;# base of gid[]
        mov   edx, [edx + esi*4]		;# ggid=gid[n]

	movq  mm7, [esp + nb102nf_vctot]     
	pfacc mm7,mm7	          ;# get and sum the two parts of total potential 

	mov   eax, [ebp + nb102nf_Vc]
	movd  mm6, [eax + edx*4] 
	pfadd mm6, mm7
	movd  [eax + edx*4], mm6          ;# increment vc[gid] 
       ;# finish if last 
        mov ecx, [esp + nb102nf_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jecxz .nb102nf_outerend

        ;# not last, iterate outer loop once more!  
        mov [esp + nb102nf_n], esi
        jmp .nb102nf_outer
.nb102nf_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [esp + nb102nf_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jecxz .nb102nf_end
        ;# non-zero, do one more workunit
        jmp   .nb102nf_threadloop
.nb102nf_end:
	femms

	mov eax, [esp + nb102nf_nouter] 	
	mov ebx, [esp + nb102nf_ninner]
	mov ecx, [ebp + nb102nf_outeriter]
	mov edx, [ebp + nb102nf_inneriter]
	mov [ecx], eax
	mov [edx], ebx

	add esp, 104
	pop edi
	pop esi
    	pop edx
    	pop ecx
    	pop ebx
    	pop eax
	leave
	ret