nb_kernel302_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,217 行 · 第 1/5 页

	mulpd  xmm3, xmm7 ;# fijC=FF*qq 
    ;# at this point mm5 contains vcoul and xmm3 fijC 

    addpd  xmm5, [esp + nb302_vctot]
    movapd [esp + nb302_vctot], xmm5
	xorpd  xmm1, xmm1
	mulpd  xmm3,  [esp + nb302_tsc]
	mulpd  xmm3, xmm0
	subpd  xmm1, xmm3

	movapd xmm0, xmm1
	movapd xmm2, xmm1
	
	movapd xmm3, [esp + nb302_fjxH1]
	movapd xmm4, [esp + nb302_fjyH1]
	movapd xmm5, [esp + nb302_fjzH1]
	mulpd xmm0, [esp + nb302_dxH2H1]
	mulpd xmm1, [esp + nb302_dyH2H1]
	mulpd xmm2, [esp + nb302_dzH2H1]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb302_fixH2]
	addpd xmm1, [esp + nb302_fiyH2]
	addpd xmm2, [esp + nb302_fizH2]
	movapd [esp + nb302_fjxH1], xmm3
	movapd [esp + nb302_fjyH1], xmm4
	movapd [esp + nb302_fjzH1], xmm5
	movapd [esp + nb302_fixH2], xmm0
	movapd [esp + nb302_fiyH2], xmm1
	movapd [esp + nb302_fizH2], xmm2

	;# H2-H2 interaction 
	movapd xmm0, [esp + nb302_rinvH2H2]
	movapd xmm1, xmm0
	mulpd  xmm1, [esp + nb302_rsqH2H2] ;# xmm1=r 
	mulpd  xmm1, [esp + nb302_tsc]	
	cvttpd2pi mm6, xmm1	;# mm6 = lu idx 
	cvtpi2pd xmm6, mm6
	subpd xmm1, xmm6	;# xmm1=eps 
	movapd xmm2, xmm1	
	mulpd  xmm2, xmm2	;# xmm2=eps2 
	
	pslld mm6, 2		;# idx *= 4 
	mov  esi, [ebp + nb302_VFtab]
	movd eax, mm6
	psrlq mm6, 32
	movd ebx, mm6		;# indices in eax/ebx 

	movlpd xmm4, [esi + eax*8]	;# Y1	
	movlpd xmm3, [esi + ebx*8]	;# Y2
	movhpd xmm4, [esi + eax*8 + 8]	;# Y1 F1 	
	movhpd xmm3, [esi + ebx*8 + 8]	;# Y2 F2 
	movapd xmm5, xmm4
	unpcklpd xmm4, xmm3	;# Y1 Y2 
	unpckhpd xmm5, xmm3	;# F1 F2 

	movlpd xmm6, [esi + eax*8 + 16]	;# G1
	movlpd xmm3, [esi + ebx*8 + 16]	;# G2
	movhpd xmm6, [esi + eax*8 + 24]	;# G1 H1 	
	movhpd xmm3, [esi + ebx*8 + 24]	;# G2 H2 

	movapd xmm7, xmm6
	unpcklpd xmm6, xmm3	;# G1 G2 
	unpckhpd xmm7, xmm3	;# H1 H2 
	;# coulomb table ready, in xmm4-xmm7  		
	mulpd  xmm6, xmm1	;# xmm6=Geps 
	mulpd  xmm7, xmm2	;# xmm7=Heps2 
	addpd  xmm5, xmm6
	addpd  xmm5, xmm7	;# xmm5=Fp 	
	mulpd  xmm7, [esp + nb302_two]	;# two*Heps2 
	movapd xmm3, [esp + nb302_qqHH]
	addpd  xmm7, xmm6
	addpd  xmm7, xmm5 ;# xmm7=FF 
	mulpd  xmm5, xmm1 ;# xmm5=eps*Fp 
	addpd  xmm5, xmm4 ;# xmm5=VV 
	mulpd  xmm5, xmm3 ;# vcoul=qq*VV  
	mulpd  xmm3, xmm7 ;# fijC=FF*qq 
    ;# at this point mm5 contains vcoul and xmm3 fijC 
	
    addpd  xmm5, [esp + nb302_vctot]
    movapd [esp + nb302_vctot], xmm5
	xorpd  xmm1, xmm1
	mulpd  xmm3,  [esp + nb302_tsc]
	mulpd  xmm3, xmm0
	subpd  xmm1, xmm3

	movapd xmm0, xmm1
	movapd xmm2, xmm1
	
	movapd xmm3, [esp + nb302_fjxH2]
	movapd xmm4, [esp + nb302_fjyH2]
	movapd xmm5, [esp + nb302_fjzH2]
	mulpd xmm0, [esp + nb302_dxH2H2]
	mulpd xmm1, [esp + nb302_dyH2H2]
	mulpd xmm2, [esp + nb302_dzH2H2]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb302_fixH2]
	addpd xmm1, [esp + nb302_fiyH2]
	addpd xmm2, [esp + nb302_fizH2]
	movapd [esp + nb302_fjxH2], xmm3
	movapd [esp + nb302_fjyH2], xmm4
	movapd [esp + nb302_fjzH2], xmm5
	movapd [esp + nb302_fixH2], xmm0
	movapd [esp + nb302_fiyH2], xmm1
	movapd [esp + nb302_fizH2], xmm2

	mov edi, [ebp + nb302_faction]

	movd eax, mm0
	movd ebx, mm1
	
	;# Did all interactions - now update j forces 
	movlpd xmm0, [edi + eax*8]
	movlpd xmm1, [edi + eax*8 + 8]
	movlpd xmm2, [edi + eax*8 + 16]
	movlpd xmm3, [edi + eax*8 + 24]
	movlpd xmm4, [edi + eax*8 + 32]
	movlpd xmm5, [edi + eax*8 + 40]
	movlpd xmm6, [edi + eax*8 + 48]
	movlpd xmm7, [edi + eax*8 + 56]
	movhpd xmm0, [edi + ebx*8]
	movhpd xmm1, [edi + ebx*8 + 8]
	movhpd xmm2, [edi + ebx*8 + 16]
	movhpd xmm3, [edi + ebx*8 + 24]
	movhpd xmm4, [edi + ebx*8 + 32]
	movhpd xmm5, [edi + ebx*8 + 40]
	movhpd xmm6, [edi + ebx*8 + 48]
	movhpd xmm7, [edi + ebx*8 + 56]
	addpd xmm0, [esp + nb302_fjxO]
	addpd xmm1, [esp + nb302_fjyO]
	addpd xmm2, [esp + nb302_fjzO]
	addpd xmm3, [esp + nb302_fjxH1]
	addpd xmm4, [esp + nb302_fjyH1]
	addpd xmm5, [esp + nb302_fjzH1]
	addpd xmm6, [esp + nb302_fjxH2]
	addpd xmm7, [esp + nb302_fjyH2]
	movlpd [edi + eax*8], xmm0
	movlpd [edi + eax*8 + 8], xmm1
	movlpd [edi + eax*8 + 16], xmm2
	movlpd [edi + eax*8 + 24], xmm3
	movlpd [edi + eax*8 + 32], xmm4
	movlpd [edi + eax*8 + 40], xmm5
	movlpd [edi + eax*8 + 48], xmm6
	movlpd [edi + eax*8 + 56], xmm7
	movhpd [edi + ebx*8], xmm0
	movhpd [edi + ebx*8 + 8], xmm1
	movhpd [edi + ebx*8 + 16], xmm2
	movhpd [edi + ebx*8 + 24], xmm3
	movhpd [edi + ebx*8 + 32], xmm4
	movhpd [edi + ebx*8 + 40], xmm5
	movhpd [edi + ebx*8 + 48], xmm6
	movhpd [edi + ebx*8 + 56], xmm7

	movlpd xmm0, [edi + eax*8 + 64]
	movhpd xmm0, [edi + ebx*8 + 64]
	addpd xmm0, [esp + nb302_fjzH2]
	movlpd [edi + eax*8 + 64], xmm0
	movhpd [edi + ebx*8 + 64], xmm0
	
	;# should we do one more iteration? 
	sub dword ptr [esp + nb302_innerk],  2
	jl    .nb302_checksingle
	jmp   .nb302_unroll_loop
.nb302_checksingle:
	mov   edx, [esp + nb302_innerk]
	and   edx, 1
	jnz   .nb302_dosingle
	jmp   .nb302_updateouterdata
.nb302_dosingle:
	mov   edx, [esp + nb302_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [edx]

	mov esi, [ebp + nb302_pos]
	lea   eax, [eax + eax*2]  

	;# fetch j coordinates 
	movlpd xmm2, [esi + eax*8]
	movlpd xmm3, [esi + eax*8 + 8]
	movlpd xmm4, [esi + eax*8 + 16]
	movlpd xmm5, [esi + eax*8 + 24]
	movlpd xmm6, [esi + eax*8 + 32]
	movlpd xmm7, [esi + eax*8 + 40]
	movapd 	[esp + nb302_jxO], xmm2
	movapd 	[esp + nb302_jyO], xmm3
	movapd 	[esp + nb302_jzO], xmm4
	movapd 	[esp + nb302_jxH1], xmm5
	movapd 	[esp + nb302_jyH1], xmm6
	movapd 	[esp + nb302_jzH1], xmm7
	movlpd xmm2, [esi + eax*8 + 48]
	movlpd xmm3, [esi + eax*8 + 56]
	movlpd xmm4, [esi + eax*8 + 64]
	movapd 	[esp + nb302_jxH2], xmm2
	movapd 	[esp + nb302_jyH2], xmm3
	movapd 	[esp + nb302_jzH2], xmm4
	
	movapd xmm0, [esp + nb302_ixO]
	movapd xmm1, [esp + nb302_iyO]
	movapd xmm2, [esp + nb302_izO]
	movapd xmm3, [esp + nb302_ixO]
	movapd xmm4, [esp + nb302_iyO]
	movapd xmm5, [esp + nb302_izO]
	subsd  xmm0, [esp + nb302_jxO]
	subsd  xmm1, [esp + nb302_jyO]
	subsd  xmm2, [esp + nb302_jzO]
	subsd  xmm3, [esp + nb302_jxH1]
	subsd  xmm4, [esp + nb302_jyH1]
	subsd  xmm5, [esp + nb302_jzH1]
	movapd [esp + nb302_dxOO], xmm0
	movapd [esp + nb302_dyOO], xmm1
	movapd [esp + nb302_dzOO], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb302_dxOH1], xmm3
	movapd [esp + nb302_dyOH1], xmm4
	movapd [esp + nb302_dzOH1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb302_rsqOO], xmm0
	movapd [esp + nb302_rsqOH1], xmm3

	movapd xmm0, [esp + nb302_ixO]
	movapd xmm1, [esp + nb302_iyO]
	movapd xmm2, [esp + nb302_izO]
	movapd xmm3, [esp + nb302_ixH1]
	movapd xmm4, [esp + nb302_iyH1]
	movapd xmm5, [esp + nb302_izH1]
	subsd  xmm0, [esp + nb302_jxH2]
	subsd  xmm1, [esp + nb302_jyH2]
	subsd  xmm2, [esp + nb302_jzH2]
	subsd  xmm3, [esp + nb302_jxO]
	subsd  xmm4, [esp + nb302_jyO]
	subsd  xmm5, [esp + nb302_jzO]
	movapd [esp + nb302_dxOH2], xmm0
	movapd [esp + nb302_dyOH2], xmm1
	movapd [esp + nb302_dzOH2], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb302_dxH1O], xmm3
	movapd [esp + nb302_dyH1O], xmm4
	movapd [esp + nb302_dzH1O], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb302_rsqOH2], xmm0
	movapd [esp + nb302_rsqH1O], xmm3

	movapd xmm0, [esp + nb302_ixH1]
	movapd xmm1, [esp + nb302_iyH1]
	movapd xmm2, [esp + nb302_izH1]
	movapd xmm3, [esp + nb302_ixH1]
	movapd xmm4, [esp + nb302_iyH1]
	movapd xmm5, [esp + nb302_izH1]
	subsd  xmm0, [esp + nb302_jxH1]
	subsd  xmm1, [esp + nb302_jyH1]
	subsd  xmm2, [esp + nb302_jzH1]
	subsd  xmm3, [esp + nb302_jxH2]
	subsd  xmm4, [esp + nb302_jyH2]
	subsd  xmm5, [esp + nb302_jzH2]
	movapd [esp + nb302_dxH1H1], xmm0
	movapd [esp + nb302_dyH1H1], xmm1
	movapd [esp + nb302_dzH1H1], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb302_dxH1H2], xmm3
	movapd [esp + nb302_dyH1H2], xmm4
	movapd [esp + nb302_dzH1H2], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb302_rsqH1H1], xmm0
	movapd [esp + nb302_rsqH1H2], xmm3

	movapd xmm0, [esp + nb302_ixH2]
	movapd xmm1, [esp + nb302_iyH2]
	movapd xmm2, [esp + nb302_izH2]
	movapd xmm3, [esp + nb302_ixH2]
	movapd xmm4, [esp + nb302_iyH2]
	movapd xmm5, [esp + nb302_izH2]
	subsd  xmm0, [esp + nb302_jxO]
	subsd  xmm1, [esp + nb302_jyO]
	subsd  xmm2, [esp + nb302_jzO]
	subsd  xmm3, [esp + nb302_jxH1]
	subsd  xmm4, [esp + nb302_jyH1]
	subsd  xmm5, [esp + nb302_jzH1]
	movapd [esp + nb302_dxH2O], xmm0
	movapd [esp + nb302_dyH2O], xmm1
	movapd [esp + nb302_dzH2O], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb302_dxH2H1], xmm3
	movapd [esp + nb302_dyH2H1], xmm4
	movapd [esp + nb302_dzH2H1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm4, xmm3
	addsd  xmm4, xmm5
	movapd [esp + nb302_rsqH2O], xmm0
	movapd [esp + nb302_rsqH2H1], xmm4

	movapd xmm0, [esp + nb302_ixH2]
	movapd xmm1, [esp + nb302_iyH2]
	movapd xmm2, [esp + nb302_izH2]
	subsd  xmm0, [esp + nb302_jxH2]
	subsd  xmm1, [esp + nb302_jyH2]
	subsd  xmm2, [esp + nb302_jzH2]
	movapd [esp + nb302_dxH2H2], xmm0
	movapd [esp + nb302_dyH2H2], xmm1
	movapd [esp + nb302_dzH2H2], xmm2
	mulsd xmm0, xmm0
	mulsd xmm1, xmm1
	mulsd xmm2, xmm2
	addsd xmm0, xmm1
	addsd xmm0, xmm2
	movapd [esp + nb302_rsqH2H2], xmm0
		
	;# start doing invsqrt use rsq values in xmm0, xmm4 
	cvtsd2ss xmm1, xmm0	
	cvtsd2ss xmm5, xmm4	
	rsqrtss xmm1, xmm1
	rsqrtss xmm5, xmm5
	cvtss2sd xmm1, xmm1
	cvtss2sd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulsd   xmm1, xmm1	;# luA*luA 
	mulsd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [esp + nb302_three]
	mulsd   xmm1, xmm0	;# rsqA*luA*luA 
	mulsd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subsd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subsd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulsd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulsd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulsd   xmm3, [esp + nb302_half] ;# iter1 
	mulsd   xmm7, [esp + nb302_half] ;# iter1 

	movapd  xmm2, xmm3	;# copy of luA 
	movapd  xmm6, xmm7	;# copy of luB 
	mulsd   xmm3, xmm3	;# luA*luA 
	mulsd   xmm7, xmm7	;# luB*luB 
	movapd  xmm1, [esp + nb302_three]
	mulsd   xmm3, xmm0	;# rsqA*luA*luA 
	mulsd   xmm7, xmm4	;# rsqB*luB*luB 	
	movapd  xmm5, xmm1
	subsd   xmm1, xmm3	;# 3-rsqA*luA*luA 
	subsd   xmm5, xmm7	;# 3-rsqB*luB*luB 
	mulsd   xmm1, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulsd   xmm5, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulsd   xmm1, [esp + nb302_half] ;# rinv 
	mulsd   xmm5, [esp + nb302_half] ;# rinv 
	movapd [esp + nb302_rinvH2H2], xmm1
	movapd [esp + nb302_rinvH2H1], xmm5

	movapd xmm0, [esp + nb302_rsqOO]
	movapd xmm4, [esp + nb302_rsqOH1]	
	cvtsd2ss xmm1, xmm0	
	cvtsd2ss xmm5, xmm4	
	rsqrtss xmm1, xmm1
	rsqrtss xmm5, xmm5
	cvtss2sd xmm1, xmm1
	cvtss2sd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulsd   xmm1, xmm1	;# luA*luA 
	mulsd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [esp + nb302_three]
	mulsd   xmm1, xmm0	;# rsqA*luA*luA 
	mulsd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subsd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subsd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulsd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulsd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulsd   xmm3, [esp + nb302_half] ;# iter1 of  
	mulsd   xmm7, [esp + nb302_half] ;# iter1 of  

	movapd  xmm2, xmm3	;# copy of luA 
	movapd  xmm6, xmm7	;# copy of luB 
	mulsd   xmm3, xmm3	;# luA*luA 
	mulsd   xmm7, xmm7	;# luB*luB 
	movapd  xmm1, [esp + nb302_three]
	mulsd   xmm3, xmm0	;# rsqA*luA*luA 
	mulsd   xmm7, xmm4	;# rsqB*luB*luB 	
	movapd  xmm5, xmm1
	subsd   xmm1, xmm3	;# 3-rsqA*luA*luA 
	subsd   xmm5, xmm7	;# 3-rsqB*luB*luB 
	mulsd   xmm1, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulsd   xmm5, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulsd   xmm1, [esp + nb302_half] ;# rinv 
	mulsd   xmm5, [esp + nb302_half] ;# rinv
	movapd [esp + nb302_rinvOO], xmm1
	movapd [esp + nb302_rinvOH1], xmm5

	movapd xmm0, [esp + nb302_rsqOH2]
	movapd xmm4, [esp + nb302_rsqH1O]	
	cvtsd2ss xmm1, xmm0	
	cvtsd2ss xmm5, xmm4	
	rsqrtss xmm1, xmm1
	rsqrtss xmm5, xmm5
	cvtss2sd xmm1, xmm1
	cvtss2sd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulsd   xmm1, xmm1	;# luA*luA 
	mulsd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [esp + nb302_three]
	mulsd   xmm1, xmm0	;# rsqA*luA*luA 
	mulsd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subsd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subsd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulsd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulsd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulsd   xmm3, [esp + nb302_half] ;# iter1 
	mulsd   xmm7, [esp + nb302_half] ;# iter1 

	movapd  xmm2, xmm3	;# copy of luA