nb_kernel233_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,397 行 · 第 1/5 页

	mulsd xmm4,xmm4
	mulsd xmm5,xmm5
	mulsd xmm6,xmm6
	addsd xmm4, xmm5
	addsd xmm4, xmm6
	movapd xmm7, xmm4
	;# rsqO in xmm7 
	movapd [esp + nb233_rsqO], xmm7
	
	;# move ixH1-izH1 to xmm4-xmm6 
	movapd xmm4, [esp + nb233_ixH1]
	movapd xmm5, [esp + nb233_iyH1]
	movapd xmm6, [esp + nb233_izH1]

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

	;# store dr 
	movapd [esp + nb233_dxH1], xmm4
	movapd [esp + nb233_dyH1], xmm5
	movapd [esp + nb233_dzH1], xmm6
	;# square it 
	mulsd xmm4,xmm4
	mulsd xmm5,xmm5
	mulsd xmm6,xmm6
	addsd xmm6, xmm5
	addsd xmm6, xmm4
	;# rsqH1 in xmm6 

	;# move ixH2-izH2 to xmm3-xmm5  
	movapd xmm3, [esp + nb233_ixH2]
	movapd xmm4, [esp + nb233_iyH2]
	movapd xmm5, [esp + nb233_izH2]

	;# calc dr 
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2

	;# store dr 
	movapd [esp + nb233_dxH2], xmm3
	movapd [esp + nb233_dyH2], xmm4
	movapd [esp + nb233_dzH2], xmm5
	;# square it 
	mulsd xmm3,xmm3
	mulsd xmm4,xmm4
	mulsd xmm5,xmm5
	addsd xmm5, xmm4
	addsd xmm5, xmm3
	;# move ixM-izM to xmm2-xmm4  
	movapd xmm3, [esp + nb233_iyM]
	movapd xmm4, [esp + nb233_izM]
	subpd  xmm3, xmm1
	subpd  xmm4, xmm2
	movapd xmm2, [esp + nb233_ixM]
	subpd  xmm2, xmm0	

	;# store dr 
	movapd [esp + nb233_dxM], xmm2
	movapd [esp + nb233_dyM], xmm3
	movapd [esp + nb233_dzM], xmm4
	;# square it 
	mulpd xmm2,xmm2
	mulpd xmm3,xmm3
	mulpd xmm4,xmm4
	addpd xmm4, xmm3
	addpd xmm4, xmm2	
	;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

	;# calculate krsq
	movsd xmm0, [esp + nb233_krf]
	movsd xmm1, xmm0
	movsd xmm2, xmm0
	mulsd xmm0, xmm4  
	mulsd xmm1, xmm5
	mulsd xmm2, xmm6
	movsd [esp + nb233_krsqM], xmm0
	movsd [esp + nb233_krsqH2], xmm1
	movsd [esp + nb233_krsqH1], xmm2

	;# start with rsqH1 - put seed in xmm2 
	cvtsd2ss xmm2, xmm6	
	rsqrtss xmm2, xmm2
	cvtss2sd xmm2, xmm2

	movapd  xmm3, xmm2
	mulsd   xmm2, xmm2
	movapd  xmm1, [esp + nb233_three]
	mulsd   xmm2, xmm6	;# rsq*lu*lu 
	subsd   xmm1, xmm2	;# 30-rsq*lu*lu 
	mulsd   xmm1, xmm3	;# lu*(3-rsq*lu*lu) 
	mulsd   xmm1, [esp + nb233_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm1
	mulsd xmm1, xmm1	;# lu*lu 
	mulsd xmm6, xmm1	;# rsq*lu*lu 
	movapd xmm1, [esp + nb233_three]
	subsd xmm1, xmm6	;# 3-rsq*lu*lu 
	mulsd xmm1, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulsd xmm1, [esp + nb233_half] ;# rinv 
	movapd [esp + nb233_rinvH1], xmm1
	
	;# rsqH2 - seed in xmm2 
	cvtsd2ss xmm2, xmm5	
	rsqrtss xmm2, xmm2
	cvtss2sd xmm2, xmm2

	movapd  xmm3, xmm2
	mulsd   xmm2, xmm2
	movapd  xmm1, [esp + nb233_three]
	mulsd   xmm2, xmm5	;# rsq*lu*lu 
	subsd   xmm1, xmm2	;# 30-rsq*lu*lu 
	mulsd   xmm1, xmm3	;# lu*(3-rsq*lu*lu) 
	mulsd   xmm1, [esp + nb233_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm1
	mulsd xmm1, xmm1	;# lu*lu 
	mulsd xmm5, xmm1	;# rsq*lu*lu 
	movapd xmm1, [esp + nb233_three]
	subsd xmm1, xmm5	;# 3-rsq*lu*lu 
	mulsd xmm1, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulsd xmm1, [esp + nb233_half] ;# rinv 
	movapd [esp + nb233_rinvH2], xmm1
	
	;# rsqM - seed in xmm2 
	cvtsd2ss xmm2, xmm4
	rsqrtss xmm2, xmm2
	cvtss2sd xmm2, xmm2

	movapd  xmm3, xmm2
	mulsd   xmm2, xmm2
	movapd  xmm1, [esp + nb233_three]
	mulsd   xmm2, xmm4	;# rsq*lu*lu 
	subsd   xmm1, xmm2	;# 30-rsq*lu*lu 
	mulsd   xmm1, xmm3	;# lu*(3-rsq*lu*lu) 
	mulsd   xmm1, [esp + nb233_half] ;# iter1 ( new lu) 

	movapd xmm3, xmm1
	mulsd xmm1, xmm1	;# lu*lu 
	mulsd xmm4, xmm1	;# rsq*lu*lu 
	movapd xmm1, [esp + nb233_three]
	subsd xmm1, xmm4	;# 3-rsq*lu*lu 
	mulsd xmm1, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulsd xmm1, [esp + nb233_half] ;# rinv 
	movapd [esp + nb233_rinvM], xmm1

	;# rsqO - put seed in xmm2 
	cvtsd2ss xmm2, xmm7	
	rsqrtss xmm2, xmm2
	cvtss2sd xmm2, xmm2

	movsd  xmm3, xmm2
	mulsd   xmm2, xmm2
	movsd  xmm4, [esp + nb233_three]
	mulsd   xmm2, xmm7	;# rsq*lu*lu 
	subsd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulsd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulsd   xmm4, [esp + nb233_half] ;# iter1 ( new lu) 

	movsd xmm3, xmm4
	mulsd xmm4, xmm4	;# lu*lu 
	mulsd xmm7, xmm4	;# rsq*lu*lu 
	movsd xmm4, [esp + nb233_three]
	subsd xmm4, xmm7	;# 3-rsq*lu*lu 
	mulsd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulsd xmm4, [esp + nb233_half] ;# rinv 
	movsd  xmm7, xmm4	;# rinvO in xmm7 
	
	movsd xmm4, [esp + nb233_rsqO]
	movapd xmm0, xmm7
	;# LJ table interaction.
	mulsd xmm4, xmm7	;# xmm4=r 
	mulsd xmm4, [esp + nb233_tsc]
	
	cvttsd2si ebx, xmm4	;# mm6 = lu idx 
	cvtsi2sd xmm5, ebx
	subpd xmm4, xmm5
	movapd xmm1, xmm4	;# xmm1=eps 
	movapd xmm2, xmm1	
	mulpd  xmm2, xmm2	;# xmm2=eps2 

	shl ebx, 3

	mov  esi, [ebp + nb233_VFtab]

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

	movlpd xmm6, [esi + ebx*8 + 16]	;# G1
	movhpd xmm6, [esi + ebx*8 + 24]	;# G1 H1 	
	movapd xmm7, xmm6
	unpcklpd xmm6, xmm3	;# G1 G2 
	unpckhpd xmm7, xmm3	;# H1 H2 
	;# dispersion 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 + nb233_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 + nb233_c6]
	mulsd  xmm7, xmm4	 ;# fijD 
	mulsd  xmm5, xmm4	 ;# Vvdw6 

	;# put scalar force on stack Update Vvdwtot directly 
	addsd  xmm5, [esp + nb233_Vvdwtot]
	xorpd  xmm3, xmm3
	mulsd  xmm7, [esp + nb233_tsc]
	subsd  xmm3, xmm7
	movsd [esp + nb233_fstmp], xmm3
	movsd [esp + nb233_Vvdwtot], xmm5

	;# repulsion 
	movlpd xmm4, [esi + ebx*8 + 32]	;# Y1 	
	movhpd xmm4, [esi + ebx*8 + 40]	;# Y1 F1 	

	movapd xmm5, xmm4
	unpcklpd xmm4, xmm3	;# Y1 Y2 
	unpckhpd xmm5, xmm3	;# F1 F2 

	movlpd xmm6, [esi + ebx*8 + 48]	;# G1
	movhpd xmm6, [esi + ebx*8 + 56]	;# G1 H1 	

	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 + nb233_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 + nb233_c12]
	mulsd  xmm7, xmm4 
	mulsd  xmm5, xmm4  
	
	addsd  xmm5, [esp + nb233_Vvdwtot]
	movsd xmm3, [esp + nb233_fstmp]
	mulsd  xmm7, [esp + nb233_tsc]
	subsd  xmm3, xmm7
	movsd [esp + nb233_Vvdwtot], xmm5

	mulsd  xmm3, xmm0
		
		
	movsd xmm0, [esp + nb233_dxO]
	movsd xmm1, [esp + nb233_dyO]
	movsd xmm2, [esp + nb233_dzO]

	mov    edi, [ebp + nb233_faction]
	mulsd  xmm0, xmm3
	mulsd  xmm1, xmm3
	mulsd  xmm2, xmm3

	;# update O forces 
	movapd xmm3, [esp + nb233_fixO]
	movapd xmm4, [esp + nb233_fiyO]
	movapd xmm7, [esp + nb233_fizO]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm7, xmm2
	movsd [esp + nb233_fixO], xmm3
	movsd [esp + nb233_fiyO], xmm4
	movsd [esp + nb233_fizO], xmm7
	;# update j forces with water O 
	movsd [esp + nb233_fjx], xmm0
	movsd [esp + nb233_fjy], xmm1
	movsd [esp + nb233_fjz], xmm2

	;# H1 interactions
	movsd  xmm6, [esp + nb233_rinvH1] 
	movsd  xmm4, xmm6
	mulsd   xmm4, xmm4	;# xmm6=rinv, xmm4=rinvsq 
	movsd  xmm7, xmm6
	movsd  xmm0, [esp + nb233_krsqH1]
	addsd   xmm6, xmm0	;# xmm6=rinv+ krsq 
	mulsd   xmm0, [esp + nb233_two]
	subsd   xmm6, [esp + nb233_crf]
	subsd   xmm7, xmm0	;# xmm7=rinv-2*krsq 
	mulsd   xmm6, [esp + nb233_qqH] ;# vcoul 
	mulsd   xmm7, [esp + nb233_qqH]
	mulsd  xmm4, xmm7		;# total fsH1 in xmm4 
	
	addsd  xmm6, [esp + nb233_vctot]

	movapd xmm0, [esp + nb233_dxH1]
	movapd xmm1, [esp + nb233_dyH1]
	movapd xmm2, [esp + nb233_dzH1]
	movsd [esp + nb233_vctot], xmm6
	mulsd  xmm0, xmm4
	mulsd  xmm1, xmm4
	mulsd  xmm2, xmm4

	;# update H1 forces 
	movapd xmm3, [esp + nb233_fixH1]
	movapd xmm4, [esp + nb233_fiyH1]
	movapd xmm7, [esp + nb233_fizH1]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm7, xmm2
	movsd [esp + nb233_fixH1], xmm3
	movsd [esp + nb233_fiyH1], xmm4
	movsd [esp + nb233_fizH1], xmm7
	;# update j forces with water H1 
	addsd  xmm0, [esp + nb233_fjx]
	addsd  xmm1, [esp + nb233_fjy]
	addsd  xmm2, [esp + nb233_fjz]
	movsd [esp + nb233_fjx], xmm0
	movsd [esp + nb233_fjy], xmm1
	movsd [esp + nb233_fjz], xmm2

	;# H2 interactions 
	movsd  xmm5, [esp + nb233_rinvH2] 
	movsd  xmm4, xmm5	
	mulsd   xmm4, xmm4	;# xmm5=rinv, xmm4=rinvsq 
	movsd  xmm7, xmm5
	movsd  xmm0, [esp + nb233_krsqH2]
	addsd   xmm5, xmm0	;# xmm5=rinv+ krsq 
	mulsd   xmm0, [esp + nb233_two]
	subsd   xmm5, [esp + nb233_crf]
	subsd   xmm7, xmm0	;# xmm7=rinv-2*krsq 
	mulsd   xmm5, [esp + nb233_qqH] ;# vcoul 
	mulsd   xmm7, [esp + nb233_qqH]
	mulsd  xmm4, xmm7		;# total fsH2 in xmm4 
	
	addsd  xmm5, [esp + nb233_vctot]

	movapd xmm0, [esp + nb233_dxH2]
	movapd xmm1, [esp + nb233_dyH2]
	movapd xmm2, [esp + nb233_dzH2]
	movsd [esp + nb233_vctot], xmm5
	mulsd  xmm0, xmm4
	mulsd  xmm1, xmm4
	mulsd  xmm2, xmm4

	;# update H2 forces 
	movapd xmm3, [esp + nb233_fixH2]
	movapd xmm4, [esp + nb233_fiyH2]
	movapd xmm7, [esp + nb233_fizH2]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm7, xmm2
	movsd [esp + nb233_fixH2], xmm3
	movsd [esp + nb233_fiyH2], xmm4
	movsd [esp + nb233_fizH2], xmm7
	;# update j forces with water H2 
	addsd  xmm0, [esp + nb233_fjx]
	addsd  xmm1, [esp + nb233_fjy]
	addsd  xmm2, [esp + nb233_fjz]
	movsd [esp + nb233_fjx], xmm0
	movsd [esp + nb233_fjy], xmm1
	movsd [esp + nb233_fjz], xmm2

	;# M interactions 
	movsd  xmm5, [esp + nb233_rinvM] 
	movsd  xmm4, xmm5	
	mulsd   xmm4, xmm4	;# xmm5=rinv, xmm4=rinvsq 
	movsd  xmm7, xmm5
	movsd  xmm0, [esp + nb233_krsqM]
	addsd   xmm5, xmm0	;# xmm5=rinv+ krsq 
	mulsd   xmm0, [esp + nb233_two]
	subsd   xmm5, [esp + nb233_crf]
	subsd   xmm7, xmm0	;# xmm7=rinv-2*krsq 
	mulsd   xmm5, [esp + nb233_qqM] ;# vcoul 
	mulsd   xmm7, [esp + nb233_qqM]
	mulsd  xmm4, xmm7		;# total fsH2 in xmm4 
	
	addsd  xmm5, [esp + nb233_vctot]

	movapd xmm0, [esp + nb233_dxM]
	movapd xmm1, [esp + nb233_dyM]
	movapd xmm2, [esp + nb233_dzM]
	movsd [esp + nb233_vctot], xmm5
	mulsd  xmm0, xmm4
	mulsd  xmm1, xmm4
	mulsd  xmm2, xmm4

	;# update M forces 
	movapd xmm3, [esp + nb233_fixM]
	movapd xmm4, [esp + nb233_fiyM]
	movapd xmm7, [esp + nb233_fizM]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm7, xmm2
	movsd [esp + nb233_fixM], xmm3
	movsd [esp + nb233_fiyM], xmm4
	movsd [esp + nb233_fizM], xmm7

	mov edi, [ebp + nb233_faction]
	;# update j forces 
	addsd  xmm0, [esp + nb233_fjx]
	addsd  xmm1, [esp + nb233_fjy]
	addsd  xmm2, [esp + nb233_fjz]
	movlpd xmm3, [edi + eax*8]
	movlpd xmm4, [edi + eax*8 + 8]
	movlpd xmm5, [edi + eax*8 + 16]
	subsd xmm3, xmm0
	subsd xmm4, xmm1
	subsd xmm5, xmm2
	movlpd [edi + eax*8], xmm3
	movlpd [edi + eax*8 + 8], xmm4
	movlpd [edi + eax*8 + 16], xmm5

.nb233_updateouterdata:
	mov   ecx, [esp + nb233_ii3]
	mov   edi, [ebp + nb233_faction]
	mov   esi, [ebp + nb233_fshift]
	mov   edx, [esp + nb233_is3]

	;# accumulate  Oi forces in xmm0, xmm1, xmm2 
	movapd xmm0, [esp + nb233_fixO]
	movapd xmm1, [esp + nb233_fiyO]
	movapd xmm2, [esp + nb233_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 

	;# 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, xmm2
	unpcklpd xmm6,xmm1 

	;# accumulate H1i forces in xmm0, xmm1, xmm2 
	movapd xmm0, [esp + nb233_fixH1]
	movapd xmm1, [esp + nb233_fiyH1]
	movapd xmm2, [esp + nb233_fizH1]

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addsd  xmm0, xmm3
	addsd  xmm1, xmm4
	addsd  xmm2, xmm5 ;# sum is in low xmm0-xmm2 

	;# increment i force 
	movsd  xmm3, [edi + ecx*8 + 24]
	movsd  xmm4, [edi + ecx*8 + 32]
	movsd  xmm5, [edi + ecx*8 + 40]
	addsd  xmm3, xmm0
	addsd  xmm4, xmm1
	addsd  xmm5, xmm2
	movsd  [edi + ecx*8 + 24], xmm3
	movsd  [edi + ecx*8 + 32], xmm4
	movsd  [edi + ecx*8 + 40], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	addsd xmm7, xmm2
	unpcklpd xmm0, xmm1
	addpd xmm6, xmm0