nb_kernel430_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,411 行 · 第 1/5 页

.equiv          nb430nf_jindex,         16
.equiv          nb430nf_jjnr,           20
.equiv          nb430nf_shift,          24
.equiv          nb430nf_shiftvec,       28
.equiv          nb430nf_fshift,         32
.equiv          nb430nf_gid,            36
.equiv          nb430nf_pos,            40
.equiv          nb430nf_faction,        44
.equiv          nb430nf_charge,         48
.equiv          nb430nf_p_facel,        52
.equiv          nb430nf_argkrf,         56
.equiv          nb430nf_argcrf,         60
.equiv          nb430nf_Vc,             64
.equiv          nb430nf_type,           68
.equiv          nb430nf_p_ntype,        72
.equiv          nb430nf_vdwparam,       76
.equiv          nb430nf_Vvdw,           80
.equiv          nb430nf_p_tabscale,     84
.equiv          nb430nf_VFtab,          88
.equiv          nb430nf_invsqrta,       92
.equiv          nb430nf_dvda,           96
.equiv          nb430nf_p_gbtabscale,   100
.equiv          nb430nf_GBtab,          104
.equiv          nb430nf_p_nthreads,     108
.equiv          nb430nf_count,          112
.equiv          nb430nf_mtx,            116
.equiv          nb430nf_outeriter,      120
.equiv          nb430nf_inneriter,      124
.equiv          nb430nf_work,           128
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb430nf_ix,             0
.equiv          nb430nf_iy,             16
.equiv          nb430nf_iz,             32
.equiv          nb430nf_iq,             48
.equiv          nb430nf_gbtsc,          64
.equiv          nb430nf_tsc,            80
.equiv          nb430nf_qq,             96
.equiv          nb430nf_c6,             112
.equiv          nb430nf_c12,            128
.equiv          nb430nf_vctot,          144
.equiv          nb430nf_Vvdwtot,        160
.equiv          nb430nf_half,           176
.equiv          nb430nf_three,          192
.equiv          nb430nf_isai,           208
.equiv          nb430nf_isaprod,        224
.equiv          nb430nf_gbscale,        240
.equiv          nb430nf_r,              256
.equiv          nb430nf_is3,            272
.equiv          nb430nf_ii3,            276
.equiv          nb430nf_ntia,           280
.equiv          nb430nf_innerjjnr,      284
.equiv          nb430nf_innerk,         288
.equiv          nb430nf_n,              292
.equiv          nb430nf_nn1,            296
.equiv          nb430nf_nri,            300
.equiv          nb430nf_facel,          304
.equiv          nb430nf_ntype,          308
.equiv          nb430nf_nouter,         312
.equiv          nb430nf_ninner,         316
.equiv          nb430nf_salign,         320
	push ebp
	mov ebp,esp	
    	push eax
    	push ebx
    	push ecx
    	push edx
	push esi
	push edi
	sub esp, 324		;# local stack space 
	mov  eax, esp
	and  eax, 0xf
	sub esp, eax
	mov [esp + nb430nf_salign], eax

	emms

	;# Move args passed by reference to stack
	mov ecx, [ebp + nb430nf_p_nri]
	mov esi, [ebp + nb430nf_p_facel]
	mov edi, [ebp + nb430nf_p_ntype]
	mov ecx, [ecx]
	mov esi, [esi]
	mov edi, [edi]
	mov [esp + nb430nf_nri], ecx
	mov [esp + nb430nf_facel], esi
	mov [esp + nb430nf_ntype], edi

	;# zero iteration counters
	mov eax, 0
	mov [esp + nb430nf_nouter], eax
	mov [esp + nb430nf_ninner], eax


	mov eax, [ebp + nb430nf_p_gbtabscale]
	movss xmm3, [eax]
	mov eax, [ebp + nb430nf_p_tabscale]
	movss xmm4, [eax]
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	movaps [esp + nb430nf_gbtsc], xmm3
	movaps [esp + nb430nf_tsc], xmm4

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# constant 0.5 in IEEE (hex)
	mov [esp + nb430nf_half], eax
	movss xmm1, [esp + nb430nf_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 + nb430nf_half],  xmm1
	movaps [esp + nb430nf_three],  xmm3

.nb430nf_threadloop:
        mov   esi, [ebp + nb430nf_count]          ;# pointer to sync counter
        mov   eax, [esi]
.nb430nf_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 .nb430nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [esp + nb430nf_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 + nb430nf_n], eax
        mov [esp + nb430nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb430nf_outerstart
        jmp .nb430nf_end

.nb430nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [esp + nb430nf_nouter]
	mov [esp + nb430nf_nouter], ebx

.nb430nf_outer:
	mov   eax, [ebp + nb430nf_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [eax + esi*4]		;# ebx=shift[n] 
	
	lea   ebx, [ebx + ebx*2]    ;# ebx=3*is 
	mov   [esp + nb430nf_is3],ebx    	;# store is3 

	mov   eax, [ebp + nb430nf_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 + nb430nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [ecx + esi*4]	    ;# ebx =ii 

	mov   edx, [ebp + nb430nf_charge]
	movss xmm3, [edx + ebx*4]	
	mulss xmm3, [esp + nb430nf_facel]
	shufps xmm3, xmm3, 0

	mov   edx, [ebp + nb430nf_invsqrta]	;# load invsqrta[ii]
	movss xmm4, [edx + ebx*4]
	shufps xmm4, xmm4, 0

    	mov   edx, [ebp + nb430nf_type] 
    	mov   edx, [edx + ebx*4]
    	imul  edx, [esp + nb430nf_ntype]
    	shl   edx, 1
    	mov   [esp + nb430nf_ntia], edx
		
	lea   ebx, [ebx + ebx*2]	;# ebx = 3*ii=ii3 
	mov   eax, [ebp + nb430nf_pos]    ;# eax = base of pos[]  

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

	movaps [esp + nb430nf_iq], xmm3
	movaps [esp + nb430nf_isai], xmm4
	
	shufps xmm0, xmm0, 0
	shufps xmm1, xmm1, 0
	shufps xmm2, xmm2, 0

	movaps [esp + nb430nf_ix], xmm0
	movaps [esp + nb430nf_iy], xmm1
	movaps [esp + nb430nf_iz], xmm2

	mov   [esp + nb430nf_ii3], ebx
	
	;# clear vctot 
	xorps xmm4, xmm4
	movaps [esp + nb430nf_vctot], xmm4
	movaps [esp + nb430nf_Vvdwtot], xmm4
	
	mov   eax, [ebp + nb430nf_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 + nb430nf_pos]
	mov   edi, [ebp + nb430nf_faction]	
	mov   eax, [ebp + nb430nf_jjnr]
	shl   ecx, 2
	add   eax, ecx
	mov   [esp + nb430nf_innerjjnr], eax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  4
	add   ecx, [esp + nb430nf_ninner]
	mov   [esp + nb430nf_ninner], ecx
	add   edx, 0
	mov   [esp + nb430nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb430nf_unroll_loop
	jmp   .nb430nf_finish_inner
.nb430nf_unroll_loop:	
	;# quad-unroll innerloop here 
	mov   edx, [esp + nb430nf_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 + nb430nf_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	;# load isa2
	mov esi, [ebp + nb430nf_invsqrta]
	movss xmm3, [esi + eax*4]
	movss xmm4, [esi + ecx*4]
	movss xmm6, [esi + ebx*4]
	movss xmm7, [esi + edx*4]
	movaps xmm2, [esp + nb430nf_isai]
	shufps xmm3, xmm6, 0 
	shufps xmm4, xmm7, 0 
	shufps xmm3, xmm4, 136  ;# constant 10001000 ;# all charges in xmm3  
	mulps  xmm2, xmm3
		
	movaps [esp + nb430nf_isaprod], xmm2
	movaps xmm1, xmm2
	mulps xmm1, [esp + nb430nf_gbtsc]
	movaps [esp + nb430nf_gbscale], xmm1
	
	mov esi, [ebp + nb430nf_charge]    ;# base of charge[] 
	
	movss xmm3, [esi + eax*4]
	movss xmm4, [esi + ecx*4]
	movss xmm6, [esi + ebx*4]
	movss xmm7, [esi + edx*4]

	mulps xmm2, [esp + nb430nf_iq]
	shufps xmm3, xmm6, 0 
	shufps xmm4, xmm7, 0 
	shufps xmm3, xmm4, 136  ;# constant 10001000 ;# all charges in xmm3  
	mulps  xmm3, xmm2
	movaps [esp + nb430nf_qq], xmm3	

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx
	movd  mm2, ecx
	movd  mm3, edx
	
	mov esi, [ebp + nb430nf_type]
	mov eax, [esi + eax*4]
	mov ebx, [esi + ebx*4]
	mov ecx, [esi + ecx*4]
	mov edx, [esi + edx*4]
	mov esi, [ebp + nb430nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	shl ecx, 1	
	shl edx, 1	
	mov edi, [esp + nb430nf_ntia]
	add eax, edi
	add ebx, edi
	add ecx, edi
	add edx, edi

	movlps xmm6, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm6, [esi + ebx*4]
	movhps xmm7, [esi + edx*4]

	movaps xmm4, xmm6
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm6, xmm7, 221  ;# constant 11011101
	
	movd  eax, mm0		
	movd  ebx, mm1
	movd  ecx, mm2
	movd  edx, mm3

	movaps [esp + nb430nf_c6], xmm4
	movaps [esp + nb430nf_c12], xmm6
	
	mov esi, [ebp + nb430nf_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 four coordinates to xmm0-xmm2 	

	movlps xmm4, [esi + eax*4]
	movlps xmm5, [esi + ecx*4]
	movss xmm2, [esi + eax*4 + 8]
	movss xmm6, [esi + ecx*4 + 8]

	movhps xmm4, [esi + ebx*4]
	movhps xmm5, [esi + edx*4]

	movss xmm0, [esi + ebx*4 + 8]
	movss xmm1, [esi + edx*4 + 8]

	shufps xmm2, xmm0, 0
	shufps xmm6, xmm1, 0
	
	movaps xmm0, xmm4
	movaps xmm1, xmm4

	shufps xmm2, xmm6, 136  ;# constant 10001000
	
	shufps xmm0, xmm5, 136  ;# constant 10001000
	shufps xmm1, xmm5, 221  ;# constant 11011101		

	;# move ix-iz to xmm4-xmm6 
	movaps xmm4, [esp + nb430nf_ix]
	movaps xmm5, [esp + nb430nf_iy]
	movaps xmm6, [esp + nb430nf_iz]

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

	;# square it 
	mulps xmm4,xmm4
	mulps xmm5,xmm5
	mulps xmm6,xmm6
	addps xmm4, xmm5
	addps xmm4, xmm6
	;# rsq in xmm4 

	rsqrtps xmm5, xmm4
	;# lookup seed in xmm5 
	movaps xmm2, xmm5
	mulps xmm5, xmm5
	movaps xmm1, [esp + nb430nf_three]
	mulps xmm5, xmm4	;# rsq*lu*lu 			
	movaps xmm0, [esp + nb430nf_half]
	subps xmm1, xmm5	;# constant 30-rsq*lu*lu 
	mulps xmm1, xmm2	
	mulps xmm0, xmm1	;# xmm0=rinv 
	mulps xmm4, xmm0	;# xmm4=r
	movaps [esp + nb430nf_r], xmm4
	mulps xmm4, [esp + nb430nf_gbscale]

	movhlps xmm5, xmm4
	cvttps2pi mm6, xmm4
	cvttps2pi mm7, xmm5	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm6, mm6
	cvtpi2ps xmm5, mm7
	movlhps xmm6, xmm5
	subps xmm4, xmm6	
	movaps xmm1, xmm4	;# xmm1=eps 
	movaps xmm2, xmm1	
	mulps  xmm2, xmm2	;# xmm2=eps2 
	pslld mm6, 2
	pslld mm7, 2

	movd mm0, eax	
	movd mm1, ebx
	movd mm2, ecx
	movd mm3, edx

	mov  esi, [ebp + nb430nf_GBtab]
	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7
		
	;# load coulomb table
	movaps xmm4, [esi + eax*4]
	movaps xmm5, [esi + ebx*4]
	movaps xmm6, [esi + ecx*4]
	movaps xmm7, [esi + edx*4]
	;# transpose, using xmm3 for scratch
	movaps xmm3, xmm6
	shufps xmm3, xmm7, 0xEE 
	shufps xmm6, xmm7, 0x44
	movaps xmm7, xmm4
	shufps xmm7, xmm5, 0xEE
	shufps xmm4, xmm5, 0x44
	movaps xmm5, xmm4
	shufps xmm5, xmm6, 0xDD
	shufps xmm4, xmm6, 0x88
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 0x88
	shufps xmm7, xmm3, 0xDD
	;# coulomb table ready, in xmm4-xmm7  		
	
	mulps  xmm6, xmm1	;# xmm6=Geps 
	mulps  xmm7, xmm2	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7	;# xmm5=Fp 	
	movaps xmm3, [esp + nb430nf_qq]
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	addps  xmm5, [esp + nb430nf_vctot]
	movaps [esp + nb430nf_vctot], xmm5

	
	movaps xmm4, [esp + nb430nf_r]
	mulps xmm4, [esp + nb430nf_tsc]
	
	movhlps xmm5, xmm4
	cvttps2pi mm6, xmm4
	cvttps2pi mm7, xmm5	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm6, mm6
	cvtpi2ps xmm5, mm7
	movlhps xmm6, xmm5
	subps xmm4, xmm6	
	movaps xmm1, xmm4	;# xmm1=eps 
	movaps xmm2, xmm1	
	mulps  xmm2, xmm2	;# xmm2=eps2 
	pslld mm6, 3
	pslld mm7, 3
	
	mov  esi, [ebp + nb430nf_VFtab]
	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7
		
	;# dispersion 
	movaps xmm4, [esi + eax*4]
	movaps xmm5, [esi + ebx*4]
	movaps xmm6, [esi + ecx*4]
	movaps xmm7, [esi + edx*4]
	;# transpose, using xmm3 for scratch
	movaps xmm3, xmm6
	shufps xmm3, xmm7, 0xEE 
	shufps xmm6, xmm7, 0x44
	movaps xmm7, xmm4
	shufps xmm7, xmm5, 0xEE
	shufps xmm4, xmm5, 0x44
	movaps xmm5, xmm4
	shufps xmm5, xmm6, 0xDD
	shufps xmm4, xmm6, 0x88
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 0x88
	shufps xmm7, xmm3, 0xDD
	;# dispersion table ready, in xmm4-xmm7 	
	mulps  xmm6, xmm1	;# xmm6=Geps 
	mulps  xmm7, xmm2	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7	;# xmm5=Fp 	
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, [esp + nb430nf_c6]	 ;# Vvdw6
	addps  xmm5, [esp + nb430nf_Vvdwtot]
	movaps [esp + nb430nf_Vvdwtot], xmm5

	;# repulsion 
	movaps xmm4, [esi + eax*4 + 16]
	movaps xmm5, [esi + ebx*4 + 16]
	movaps xmm6, [esi + ecx*4 + 16]
	movaps xmm7, [esi + edx*4 + 16]
	;# transpose, using xmm3 for scratch
	movaps xmm3, xmm6