nb_kernel330_ia32_3dnow.intel_syntax.s

最著名最快的分子模拟软件

	pfmul mm7, mm6	;# fijR 
	pfmul mm5, mm6	;# Vvdw12 
	pfadd mm3, mm7	;# total fscal fijC+ fijD+ fijR 

	;# change sign of mm3 
    	pxor mm1,mm1
	pfsub mm1, mm3	
	pfmul mm0, [esp + nb330_tsc]
 	pfmul mm0, mm1    ;# mm0 is total fscal now 	

	prefetchw [esp + nb330_dx1]	;# prefetch i forces to cache 

	;# spread fscalar to both positions 
	movq mm1,mm0
	punpckldq mm0,mm0
	punpckhdq mm1,mm1

	;# calc vector force 
	prefetchw [edi + eax*4]	;# prefetch the 1st faction to cache 
	movq mm2,  [esp + nb330_dx1]	;# fetch dr 
	movd mm3,  [esp + nb330_dz1]

	;# update Vvdwtot 
	pfadd mm5, [esp + nb330_Vvdwtot]      ;# add the earlier value 
	movq [esp + nb330_Vvdwtot], mm5       ;# store the sum       

	prefetchw [edi + ebx*4]	;# prefetch the 2nd faction to cache 
	pfmul mm2, mm0		;# mult by fs  
	pfmul mm3, mm0

	movq mm4,  [esp + nb330_dx2] 	;# fetch dr 
	movd mm5,  [esp + nb330_dz2]
	pfmul mm4, mm1   	;# mult by fs  
	pfmul mm5, mm1
	;# update i forces 

	movq mm0,  [esp + nb330_fix]
	movd mm1,  [esp + nb330_fiz]
	pfadd mm0, mm2
	pfadd mm1, mm3

	pfadd mm0, mm4
	pfadd mm1, mm5
	movq [esp + nb330_fix], mm0
	movd [esp + nb330_fiz], mm1
	;# update j forces 

	movq mm0,  [edi + eax*4]
	movd mm1,  [edi + eax*4 + 8]
	movq mm6,  [edi + ebx*4]
	movd mm7,  [edi + ebx*4 + 8]
	
	pfsub mm0, mm2
	pfsub mm1, mm3
	pfsub mm6, mm4
	pfsub mm7, mm5
	
	movq [edi + eax*4], mm0
	movd [edi + eax*4 +8], mm1
	movq [edi + ebx*4], mm6
	movd [edi + ebx*4 + 8], mm7
	
	;# should we do one more iteration? 
	sub dword ptr [esp + nb330_innerk],  2
	jl    .nb330_finish_inner
	jmp   .nb330_unroll_loop
.nb330_finish_inner:	
	and dword ptr [esp + nb330_innerk],  1
	jnz  .nb330_single_inner
	jmp  .nb330_updateouterdata		
.nb330_single_inner:
	;# a single j particle iteration here - compare with the unrolled code for comments. 
	mov   eax, [esp + nb330_innerjjnr]
	mov   eax, [eax]	;# eax=jnr offset 

	mov ecx, [ebp + nb330_charge]
	movd mm5, [esp + nb330_iq]
	movd mm3, [ecx + eax*4]
	pfmul mm3, mm5	  	;# mm3=qq 

	mov esi, [ebp + nb330_vdwparam]
	mov ecx, [ebp + nb330_type]
	mov edx, [ecx + eax*4]        	 ;# type [jnr1] 
	shl edx, 1
	add edx, [esp + nb330_ntia]	     ;# tja = ntia + 2*type 
	movd mm5, [esi + edx*4]		;# mm5 = 1st c6  		
	movq [esp + nb330_c6], mm5
	movd mm5, [esi + edx*4 + 4]	;# mm5 = 1st c12  		
	movq [esp + nb330_c12], mm5

	mov   esi, [ebp + nb330_pos]
	lea   eax, [eax + eax*2]

	movq  mm0, [esp + nb330_ix]
	movd  mm1, [esp + nb330_iz]
	movq  mm4, [esi + eax*4]
	movd  mm5, [esi + eax*4 + 8]
	pfsubr mm4, mm0
	pfsubr mm5, mm1
	movq  [esp + nb330_dx1], mm4
	pfmul mm4,mm4
	movd  [esp + nb330_dz1], mm5	
	pfmul mm5,mm5
	pfacc mm4, mm5
	pfacc mm4, mm5		;# mm0=rsq 
	
    	pfrsqrt mm0,mm4
    	movq mm2,mm0
    	pfmul mm0,mm0
    	pfrsqit1 mm0,mm4				
    	pfrcpit2 mm0,mm2	;# mm0=invsqrt 
	pfmul mm4, mm0
	movq mm1, mm4
	;# mm0 is invsqrt, and mm1 r. 

	;# calculate potentials and scalar force 
	pfmul mm1, [esp + nb330_tsc]	;# mm1=rt 
	pf2iw mm4,mm1
	movd [esp + nb330_n1], mm4
	pi2fd mm4,mm4
	pfsub mm1, mm4               ;# now mm1 is eps and mm4 is n0 

	movq mm2,mm1
	pfmul mm2,mm2	;# mm1 is eps, mm2 is eps2 
	
	;# coulomb table 
	mov edx, [ebp + nb330_VFtab]
	mov ecx, [esp + nb330_n1]
	lea ecx, [ecx + ecx*2]
	shl ecx, 2
	;# load all the table values we need 
	movd mm4, [edx + ecx*4]
	movd mm5, [edx + ecx*4 + 4]
	movd mm6, [edx + ecx*4 + 8]
	movd mm7, [edx + ecx*4 + 12]

	pfmul mm6, mm1  ;# mm6 = Geps 		
	pfmul mm7, mm2	;# mm7 = Heps2 

	pfadd mm5, mm6
	pfadd mm5, mm7	;# mm5 = Fp 

	pfmul mm7, [esp + nb330_two]	;# two*Heps2 
	pfadd mm7, mm6
	pfadd mm7, mm5	;# mm7=FF 

	pfmul mm5, mm1  ;# mm5=eps*Fp 
	pfadd mm5, mm4	;#  mm5= VV 

	pfmul mm5, mm3	;# vcoul=qq*VV 
	pfmul mm3, mm7	;# fijC=FF*qq  

	;# at this point mm5 contains vcoul and mm3 fijC 
	;# increment vcoul - then we can get rid of mm5 
	;# update vctot 
	pfadd mm5, [esp + nb330_vctot]      ;# add the earlier value 
	movq [esp + nb330_vctot], mm5       ;# store the sum       
	
	;# dispersion table 
	;# load all the table values we need 
	movd mm4, [edx + ecx*4 + 16]
	movd mm5, [edx + ecx*4 + 20]
	movd mm6, [edx + ecx*4 + 24]
	movd mm7, [edx + ecx*4 + 28]
	pfmul mm6, mm1  ;# mm6 = Geps 		
	pfmul mm7, mm2	;# mm7 = Heps2 
	pfadd mm5, mm6
	pfadd mm5, mm7	;# mm5 = Fp 
	pfmul mm7, [esp + nb330_two]	;# two*Heps2 
	pfadd mm7, mm6
	pfadd mm7, mm5	;# mm7=FF 
	pfmul mm5, mm1  ;# mm5=eps*Fp 
	pfadd mm5, mm4	;#  mm5= VV 	

	movq mm4, [esp + nb330_c6]
	pfmul mm7, mm4	;# fijD 
	pfmul mm5, mm4	;# Vvdw6            
	pfadd mm3, mm7	;# add to fscal 

	;# update Vvdwtot to release mm5! 
	pfadd mm5, [esp + nb330_Vvdwtot]      ;# add the earlier value 
	movq [esp + nb330_Vvdwtot], mm5       ;# store the sum       

	;# repulsion table 
	;# load all the table values we need 
	movd mm4, [edx + ecx*4 + 32]
	movd mm5, [edx + ecx*4 + 36]
	movd mm6, [edx + ecx*4 + 40]
	movd mm7, [edx + ecx*4 + 44]

	pfmul mm6, mm1  ;# mm6 = Geps 		
	pfmul mm7, mm2	;# mm7 = Heps2 
	pfadd mm5, mm6
	pfadd mm5, mm7	;# mm5 = Fp 
	pfmul mm7, [esp + nb330_two]	;# two*Heps2 
	pfadd mm7, mm6
	pfadd mm7, mm5	;# mm7=FF 
	pfmul mm5, mm1  ;# mm5=eps*Fp 
	pfadd mm5, mm4	;#  mm5= VV 

	movq mm6, [esp + nb330_c12]
	pfmul mm7, mm6	;# fijR 
	pfmul mm5, mm6	;# Vvdw12 
	pfadd mm3, mm7	;# total fscal fijC+ fijD+ fijR 

	;# change sign of mm3 
    	pxor mm1,mm1
	pfsub mm1, mm3	
	pfmul mm0, [esp + nb330_tsc]
 	pfmul mm0, mm1    ;# mm0 is total fscal now 	

	;# update Vvdwtot 
	pfadd mm5, [esp + nb330_Vvdwtot]      ;# add the earlier value 
	movq [esp + nb330_Vvdwtot], mm5       ;# store the sum       

	;# spread fscalar to both positions 
	punpckldq mm0,mm0
	;# calc vectorial force 
	prefetchw [edi + eax*4]	;# prefetch faction to cache  
	movq mm2,  [esp + nb330_dx1]
	movd mm3,  [esp + nb330_dz1]


	pfmul mm2, mm0
	pfmul mm3, mm0

	;# update i particle force 
	movq mm0,  [esp + nb330_fix]
	movd mm1,  [esp + nb330_fiz]
	pfadd mm0, mm2
	pfadd mm1, mm3
	movq [esp + nb330_fix], mm0
	movd [esp + nb330_fiz], mm1
	;# update j particle force 
	movq mm0,  [edi + eax*4]
	movd mm1,  [edi + eax*4 + 8]
	pfsub mm0, mm2
	pfsub mm1, mm3
	movq [edi + eax*4], mm0
	movd [edi + eax*4 +8], mm1
	;# done! 
.nb330_updateouterdata:	
	mov   ecx, [esp + nb330_ii3]

	movq  mm6, [edi + ecx*4]       ;# increment i force 
	movd  mm7, [edi + ecx*4 + 8]	
	pfadd mm6, [esp + nb330_fix]
	pfadd mm7, [esp + nb330_fiz]
	movq  [edi + ecx*4],    mm6
	movd  [edi + ecx*4 +8], mm7

	mov   ebx, [ebp + nb330_fshift]    ;# increment fshift force 
	mov   edx, [esp + nb330_is3]

	movq  mm6, [ebx + edx*4]	
	movd  mm7, [ebx + edx*4 + 8]	
	pfadd mm6, [esp + nb330_fix]
	pfadd mm7, [esp + nb330_fiz]
	movq  [ebx + edx*4],     mm6
	movd  [ebx + edx*4 + 8], mm7

	;# get n from stack
	mov esi, [esp + nb330_n]
        ;# get group index for i particle 
        mov   edx, [ebp + nb330_gid]      	;# base of gid[]
        mov   edx, [edx + esi*4]		;# ggid=gid[n]

	movq  mm7, [esp + nb330_vctot]     
	pfacc mm7,mm7	          ;# get and sum the two parts of total potential 
	
	mov   eax, [ebp + nb330_Vc]
	movd  mm6, [eax + edx*4] 
	pfadd mm6, mm7
	movd  [eax + edx*4], mm6          ;# increment vc[gid] 

	movq  mm7, [esp + nb330_Vvdwtot]     
	pfacc mm7,mm7	          ;# get and sum the two parts of total potential 
	
	mov   eax, [ebp + nb330_Vvdw]
	movd  mm6, [eax + edx*4] 
	pfadd mm6, mm7
	movd  [eax + edx*4], mm6          ;# increment Vvdw[gid] 

       	;# finish if last 
        mov ecx, [esp + nb330_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jecxz .nb330_outerend

        ;# not last, iterate outer loop once more!  
        mov [esp + nb330_n], esi
        jmp .nb330_outer
.nb330_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [esp + nb330_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jecxz .nb330_end
        ;# non-zero, do one more workunit
        jmp   .nb330_threadloop
.nb330_end:
	femms

	mov eax, [esp + nb330_nouter] 	
	mov ebx, [esp + nb330_ninner]
	mov ecx, [ebp + nb330_outeriter]
	mov edx, [ebp + nb330_inneriter]
	mov [ecx], eax
	mov [edx], ebx

	add esp, 160
	pop edi
	pop esi
    	pop edx
    	pop ecx
    	pop ebx
    	pop eax
	leave
	ret



.globl nb_kernel330nf_ia32_3dnow
.globl _nb_kernel330nf_ia32_3dnow
nb_kernel330nf_ia32_3dnow:	
_nb_kernel330nf_ia32_3dnow:	
.equiv		nb330nf_p_nri,		8
.equiv		nb330nf_iinr,		12
.equiv		nb330nf_jindex,		16
.equiv		nb330nf_jjnr,		20
.equiv		nb330nf_shift,		24
.equiv		nb330nf_shiftvec,	28
.equiv		nb330nf_fshift,		32
.equiv		nb330nf_gid,		36
.equiv		nb330nf_pos,		40		
.equiv		nb330nf_faction,	44
.equiv		nb330nf_charge,		48
.equiv		nb330nf_p_facel,	52
.equiv		nb330nf_p_krf,		56	
.equiv		nb330nf_p_crf,		60	
.equiv		nb330nf_Vc,		64	
.equiv		nb330nf_type,		68
.equiv		nb330nf_p_ntype,	72
.equiv		nb330nf_vdwparam,	76	
.equiv		nb330nf_Vvdw,		80	
.equiv		nb330nf_p_tabscale,	84	
.equiv		nb330nf_VFtab,		88
.equiv		nb330nf_invsqrta,	92	
.equiv		nb330nf_dvda,		96
.equiv          nb330nf_p_gbtabscale,   100
.equiv          nb330nf_GBtab,          104
.equiv          nb330nf_p_nthreads,     108
.equiv          nb330nf_count,          112
.equiv          nb330nf_mtx,            116
.equiv          nb330nf_outeriter,      120
.equiv          nb330nf_inneriter,      124
.equiv          nb330nf_work,           128
	;# stack offsets for local variables 
.equiv		nb330nf_is3,		0
.equiv		nb330nf_ii3,		4
.equiv		nb330nf_ix,		8
.equiv		nb330nf_iy,		12
.equiv		nb330nf_iz,		16
.equiv		nb330nf_iq,		20  
.equiv		nb330nf_vctot,		28  
.equiv		nb330nf_Vvdwtot,	36  
.equiv		nb330nf_c6,		44  
.equiv		nb330nf_c12,		52  
.equiv		nb330nf_n1,		60
.equiv		nb330nf_tsc,		68  
.equiv		nb330nf_ntia,		76
.equiv		nb330nf_innerjjnr,	80
.equiv		nb330nf_innerk,		84					
.equiv          nb330nf_n,              88 ;# idx for outer loop
.equiv          nb330nf_nn1,            92 ;# number of outer iterations
.equiv          nb330nf_nri,            96
.equiv          nb330nf_facel,          100
.equiv          nb330nf_ntype,          104
.equiv          nb330nf_nouter,         108
.equiv          nb330nf_ninner,         112
	push ebp
	mov ebp,esp	
    	push eax
    	push ebx
    	push ecx
    	push edx
	push esi
	push edi
	sub esp, 116		;# local stack space 
	femms
	;# move data to local stack  
	mov ecx, [ebp + nb330nf_p_nri]
	mov edx, [ebp + nb330nf_p_ntype]
	mov esi, [ebp + nb330nf_p_facel]
	mov edi, [ebp + nb330nf_p_tabscale]
	mov ecx, [ecx]
	mov edx, [edx]
	mov esi, [esi]
	mov [esp + nb330nf_nri], ecx
	mov [esp + nb330nf_ntype], edx
	mov [esp + nb330nf_facel], esi

	movd  mm3, [edi]
	punpckldq mm3,mm3
	movq  [esp + nb330nf_tsc], mm3	

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb330nf_nouter], eax
	mov [esp + nb330nf_ninner], eax

	;# assume we have at least one i particle - start directly 	
.nb330nf_threadloop:
        mov   esi, [ebp + nb330nf_count]          ;# pointer to sync counter
        mov   eax, [esi]
.nb330nf_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 .nb330nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb330nf_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 + nb330nf_n], eax
        mov [esp + nb330nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi