nb_kernel232_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,152 行 · 第 1/5 页

	movapd xmm7, xmm6
	unpcklpd xmm6, xmm3	;# G1 G2 
	unpckhpd xmm7, xmm3	;# H1 H2 
	
	;# table ready, in xmm4-xmm7 	
	mulsd  xmm6, xmm1	;# xmm6=Geps 
	mulsd  xmm7, xmm2	;# xmm7=Heps2 
	addsd  xmm5, xmm6
	addsd  xmm5, xmm7	;# xmm5=Fp 	
	mulsd  xmm7, [esp + nb232_two]	;# two*Heps2 
	addsd  xmm7, xmm6
	addsd  xmm7, xmm5 ;# xmm7=FF 
	mulsd  xmm5, xmm1 ;# xmm5=eps*Fp 
	addsd  xmm5, xmm4 ;# xmm5=VV 
	
	movsd xmm4, [esp + nb232_c12]
	mulsd  xmm7, xmm4 
	mulsd  xmm5, xmm4  
	
	addsd  xmm5, [esp + nb232_Vvdwtot]
	movsd xmm3, [esp + nb232_fstmp]
	mulsd  xmm7, [esp + nb232_tsc]
	subsd  xmm3, xmm7
	movsd [esp + nb232_Vvdwtot], xmm5

	mulsd  xmm0, xmm3
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	
	mov    edi, [ebp + nb232_faction]

	;# update forces
	xorpd xmm3, xmm3
	movapd xmm4, xmm3
	movapd xmm5, xmm3
	mulsd xmm0, [esp + nb232_dxOO]
	mulsd xmm1, [esp + nb232_dyOO]
	mulsd xmm2, [esp + nb232_dzOO]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixO]
	addsd xmm1, [esp + nb232_fiyO]
	addsd xmm2, [esp + nb232_fizO]
	movlpd [esp + nb232_fjxO], xmm3
	movlpd [esp + nb232_fjyO], xmm4
	movlpd [esp + nb232_fjzO], xmm5
	movlpd [esp + nb232_fixO], xmm0
	movlpd [esp + nb232_fiyO], xmm1
	movlpd [esp + nb232_fizO], xmm2

	;# O-H1 interaction 
	movapd xmm0, [esp + nb232_rinvOH1]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqOH1] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=rinv+ krsq 
	mulsd  xmm0, xmm0
	subsd  xmm4, [esp + nb232_crf]
	mulsd  xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal 
	movsd  xmm6, [esp + nb232_vctot]
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH1  
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	
	xorpd xmm3, xmm3
	movapd xmm4, xmm3
	movapd xmm5, xmm3
	mulsd xmm0, [esp + nb232_dxOH1]
	mulsd xmm1, [esp + nb232_dyOH1]
	mulsd xmm2, [esp + nb232_dzOH1]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixO]
	addsd xmm1, [esp + nb232_fiyO]
	addsd xmm2, [esp + nb232_fizO]
	movlpd [esp + nb232_fjxH1], xmm3
	movlpd [esp + nb232_fjyH1], xmm4
	movlpd [esp + nb232_fjzH1], xmm5
	movlpd [esp + nb232_fixO], xmm0
	movlpd [esp + nb232_fiyO], xmm1
	movlpd [esp + nb232_fizO], xmm2

	;# O-H2 interaction  
	movapd xmm0, [esp + nb232_rinvOH2]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqOH2] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	mulsd  xmm0, xmm0
	subsd  xmm4, [esp + nb232_crf]
	mulsd  xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	xorpd xmm3, xmm3
	movapd xmm4, xmm3
	movapd xmm5, xmm3
	mulsd xmm0, [esp + nb232_dxOH2]
	mulsd xmm1, [esp + nb232_dyOH2]
	mulsd xmm2, [esp + nb232_dzOH2]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixO]
	addsd xmm1, [esp + nb232_fiyO]
	addsd xmm2, [esp + nb232_fizO]
	movlpd [esp + nb232_fjxH2], xmm3
	movlpd [esp + nb232_fjyH2], xmm4
	movlpd [esp + nb232_fjzH2], xmm5
	movlpd [esp + nb232_fixO], xmm0
	movlpd [esp + nb232_fiyO], xmm1
	movlpd [esp + nb232_fizO], xmm2

	;# H1-O interaction 
	movapd xmm0, [esp + nb232_rinvH1O]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH1O] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=rinv+ krsq 
	mulsd xmm0, xmm0
	subsd  xmm4, [esp + nb232_crf]
	mulsd  xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm3, [esp + nb232_fjxO]
	movapd xmm4, [esp + nb232_fjyO]
	movapd xmm5, [esp + nb232_fjzO]
	mulsd xmm0, [esp + nb232_dxH1O]
	mulsd xmm1, [esp + nb232_dyH1O]
	mulsd xmm2, [esp + nb232_dzH1O]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH1]
	addsd xmm1, [esp + nb232_fiyH1]
	addsd xmm2, [esp + nb232_fizH1]
	movlpd [esp + nb232_fjxO], xmm3
	movlpd [esp + nb232_fjyO], xmm4
	movlpd [esp + nb232_fjzO], xmm5
	movlpd [esp + nb232_fixH1], xmm0
	movlpd [esp + nb232_fiyH1], xmm1
	movlpd [esp + nb232_fizH1], xmm2

	;# H1-H1 interaction 
	movapd xmm0, [esp + nb232_rinvH1H1]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH1H1] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	subsd  xmm4, [esp + nb232_crf]
	mulsd xmm0, xmm0
	mulsd  xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm3, [esp + nb232_fjxH1]
	movapd xmm4, [esp + nb232_fjyH1]
	movapd xmm5, [esp + nb232_fjzH1]
	mulsd xmm0, [esp + nb232_dxH1H1]
	mulsd xmm1, [esp + nb232_dyH1H1]
	mulsd xmm2, [esp + nb232_dzH1H1]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH1]
	addsd xmm1, [esp + nb232_fiyH1]
	addsd xmm2, [esp + nb232_fizH1]
	movlpd [esp + nb232_fjxH1], xmm3
	movlpd [esp + nb232_fjyH1], xmm4
	movlpd [esp + nb232_fjzH1], xmm5
	movlpd [esp + nb232_fixH1], xmm0
	movlpd [esp + nb232_fiyH1], xmm1
	movlpd [esp + nb232_fizH1], xmm2

	;# H1-H2 interaction 
	movapd xmm0, [esp + nb232_rinvH1H2]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH1H2] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	mulsd xmm0, xmm0
	subsd  xmm4, [esp + nb232_crf]
	mulsd  xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	
	movapd xmm3, [esp + nb232_fjxH2]
	movapd xmm4, [esp + nb232_fjyH2]
	movapd xmm5, [esp + nb232_fjzH2]
	mulsd xmm0, [esp + nb232_dxH1H2]
	mulsd xmm1, [esp + nb232_dyH1H2]
	mulsd xmm2, [esp + nb232_dzH1H2]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH1]
	addsd xmm1, [esp + nb232_fiyH1]
	addsd xmm2, [esp + nb232_fizH1]
	movlpd [esp + nb232_fjxH2], xmm3
	movlpd [esp + nb232_fjyH2], xmm4
	movlpd [esp + nb232_fjzH2], xmm5
	movlpd [esp + nb232_fixH1], xmm0
	movlpd [esp + nb232_fiyH1], xmm1
	movlpd [esp + nb232_fizH1], xmm2

	;# H2-O interaction 
	movapd xmm0, [esp + nb232_rinvH2O]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH2O] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	subsd  xmm4, [esp + nb232_crf]
	mulsd xmm0, xmm0
	mulsd  xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm3, [esp + nb232_fjxO]
	movapd xmm4, [esp + nb232_fjyO]
	movapd xmm5, [esp + nb232_fjzO]
	mulsd xmm0, [esp + nb232_dxH2O]
	mulsd xmm1, [esp + nb232_dyH2O]
	mulsd xmm2, [esp + nb232_dzH2O]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH2]
	addsd xmm1, [esp + nb232_fiyH2]
	addsd xmm2, [esp + nb232_fizH2]
	movlpd [esp + nb232_fjxO], xmm3
	movlpd [esp + nb232_fjyO], xmm4
	movlpd [esp + nb232_fjzO], xmm5
	movlpd [esp + nb232_fixH2], xmm0
	movlpd [esp + nb232_fiyH2], xmm1
	movlpd [esp + nb232_fizH2], xmm2

	;# H2-H1 interaction 
	movapd xmm0, [esp + nb232_rinvH2H1]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH2H1] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	subsd  xmm4, [esp + nb232_crf]
	mulsd xmm0, xmm0
	mulsd  xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm3, [esp + nb232_fjxH1]
	movapd xmm4, [esp + nb232_fjyH1]
	movapd xmm5, [esp + nb232_fjzH1]
	mulsd xmm0, [esp + nb232_dxH2H1]
	mulsd xmm1, [esp + nb232_dyH2H1]
	mulsd xmm2, [esp + nb232_dzH2H1]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH2]
	addsd xmm1, [esp + nb232_fiyH2]
	addsd xmm2, [esp + nb232_fizH2]
	movlpd [esp + nb232_fjxH1], xmm3
	movlpd [esp + nb232_fjyH1], xmm4
	movlpd [esp + nb232_fjzH1], xmm5
	movlpd [esp + nb232_fixH2], xmm0
	movlpd [esp + nb232_fiyH2], xmm1
	movlpd [esp + nb232_fizH2], xmm2

	;# H2-H2 interaction 
	movapd xmm0, [esp + nb232_rinvH2H2]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb232_krf]	
	movapd xmm1, xmm0
	mulsd  xmm5, [esp + nb232_rsqH2H2] ;# xmm5=krsq 
	movapd xmm4, xmm5
	addsd  xmm4, xmm7	;# xmm4=r inv+ krsq 
	subsd  xmm4, [esp + nb232_crf]
	mulsd xmm0, xmm0
	mulsd  xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) 
	mulsd  xmm5, [esp + nb232_two]
	subsd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulsd  xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal 
	addsd  xmm6, xmm4	;# add to local vctot 
	mulsd xmm0, xmm7	;# fsOH2 
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm1, xmm0
	movsd [esp + nb232_vctot], xmm6
	movapd xmm2, xmm0
	
	movapd xmm3, [esp + nb232_fjxH2]
	movapd xmm4, [esp + nb232_fjyH2]
	movapd xmm5, [esp + nb232_fjzH2]
	mulsd xmm0, [esp + nb232_dxH2H2]
	mulsd xmm1, [esp + nb232_dyH2H2]
	mulsd xmm2, [esp + nb232_dzH2H2]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	addsd xmm0, [esp + nb232_fixH2]
	addsd xmm1, [esp + nb232_fiyH2]
	addsd xmm2, [esp + nb232_fizH2]
	movlpd [esp + nb232_fjxH2], xmm3
	movlpd [esp + nb232_fjyH2], xmm4
	movlpd [esp + nb232_fjzH2], xmm5
	movlpd [esp + nb232_fixH2], xmm0
	movlpd [esp + nb232_fiyH2], xmm1
	movlpd [esp + nb232_fizH2], xmm2

	mov edi, [ebp + nb232_faction]
	;# 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]
	addsd xmm0, [esp + nb232_fjxO]
	addsd xmm1, [esp + nb232_fjyO]
	addsd xmm2, [esp + nb232_fjzO]
	addsd xmm3, [esp + nb232_fjxH1]
	addsd xmm4, [esp + nb232_fjyH1]
	addsd xmm5, [esp + nb232_fjzH1]
	addsd xmm6, [esp + nb232_fjxH2]
	addsd xmm7, [esp + nb232_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

	movlpd xmm0, [edi + eax*8 + 64]
	addsd xmm0, [esp + nb232_fjzH2]
	movlpd [edi + eax*8 + 64], xmm0
	
.nb232_updateouterdata:
	mov   ecx, [esp + nb232_ii3]
	mov   edi, [ebp + nb232_faction]
	mov   esi, [ebp + nb232_fshift]
	mov   edx, [esp + nb232_is3]

	;# accumulate  Oi forces in xmm0, xmm1, xmm2 
	movapd xmm0, [esp + nb232_fixO]
	movapd xmm1, [esp + nb232_fiyO]
	movapd xmm2, [esp + nb232_fizO]

	movhlps xmm3, xmm0
	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]
	movsd  xmm4, [edi + ecx*8 + 8]
	movsd  xmm5, [edi + ecx*8 + 16]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm5, xmm2
	movsd  [edi + ecx*8],     xmm3
	movsd  [edi + ecx*8 + 8], xmm4
	movsd  [edi + ecx*8 + 16], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	movapd xmm6, xmm0
	movsd xmm7,