nb_kernel104_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

	mulpd xmm1, [esp + nb104_dyH1H1]
	mulpd xmm2, [esp + nb104_dzH1H1]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb104_fixH1]
	addpd xmm1, [esp + nb104_fiyH1]
	addpd xmm2, [esp + nb104_fizH1]
	movapd [esp + nb104_fjxH1], xmm3
	movapd [esp + nb104_fjyH1], xmm4
	movapd [esp + nb104_fjzH1], xmm5
	movapd [esp + nb104_fixH1], xmm0
	movapd [esp + nb104_fiyH1], xmm1
	movapd [esp + nb104_fizH1], xmm2

	;# H1-H2 interaction 
	movapd xmm0, [esp + nb104_rinvH1H2]
	movapd xmm1, xmm0
	mulpd xmm0, xmm0
	mulpd xmm1, [esp + nb104_qqHH]
	mulpd xmm0, xmm1	;# fsMH2  
	addpd xmm7, xmm1	;# add to local vctot 
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	movapd xmm3, [esp + nb104_fjxH2]
	movapd xmm4, [esp + nb104_fjyH2]
	movapd xmm5, [esp + nb104_fjzH2]
	mulpd xmm0, [esp + nb104_dxH1H2]
	mulpd xmm1, [esp + nb104_dyH1H2]
	mulpd xmm2, [esp + nb104_dzH1H2]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb104_fixH1]
	addpd xmm1, [esp + nb104_fiyH1]
	addpd xmm2, [esp + nb104_fizH1]
	movapd [esp + nb104_fjxH2], xmm3
	movapd [esp + nb104_fjyH2], xmm4
	movapd [esp + nb104_fjzH2], xmm5
	movapd [esp + nb104_fixH1], xmm0
	movapd [esp + nb104_fiyH1], xmm1
	movapd [esp + nb104_fizH1], xmm2

	;# H2-M interaction 
	movapd xmm0, [esp + nb104_rinvH2M]
	movapd xmm1, xmm0
	mulpd xmm0, xmm0
	mulpd xmm1, [esp + nb104_qqMH]
	mulpd xmm0, xmm1	;# fsH2M 
	addpd xmm7, xmm1	;# add to local vctot 
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	movapd xmm3, [esp + nb104_fjxM]
	movapd xmm4, [esp + nb104_fjyM]
	movapd xmm5, [esp + nb104_fjzM]
	mulpd xmm0, [esp + nb104_dxH2M]
	mulpd xmm1, [esp + nb104_dyH2M]
	mulpd xmm2, [esp + nb104_dzH2M]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb104_fixH2]
	addpd xmm1, [esp + nb104_fiyH2]
	addpd xmm2, [esp + nb104_fizH2]
	movapd [esp + nb104_fjxM], xmm3
	movapd [esp + nb104_fjyM], xmm4
	movapd [esp + nb104_fjzM], xmm5
	movapd [esp + nb104_fixH2], xmm0
	movapd [esp + nb104_fiyH2], xmm1
	movapd [esp + nb104_fizH2], xmm2

	;# H2-H1 interaction 
	movapd xmm0, [esp + nb104_rinvH2H1]
	movapd xmm1, xmm0
	mulpd xmm0, xmm0
	mulpd xmm1, [esp + nb104_qqHH]
	mulpd xmm0, xmm1	;# fsH2H1 
	addpd xmm7, xmm1	;# add to local vctot 
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	movapd xmm3, [esp + nb104_fjxH1]
	movapd xmm4, [esp + nb104_fjyH1]
	movapd xmm5, [esp + nb104_fjzH1]
	mulpd xmm0, [esp + nb104_dxH2H1]
	mulpd xmm1, [esp + nb104_dyH2H1]
	mulpd xmm2, [esp + nb104_dzH2H1]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb104_fixH2]
	addpd xmm1, [esp + nb104_fiyH2]
	addpd xmm2, [esp + nb104_fizH2]
	movapd [esp + nb104_fjxH1], xmm3
	movapd [esp + nb104_fjyH1], xmm4
	movapd [esp + nb104_fjzH1], xmm5
	movapd [esp + nb104_fixH2], xmm0
	movapd [esp + nb104_fiyH2], xmm1
	movapd [esp + nb104_fizH2], xmm2

	;# H2-H2 interaction 
	movapd xmm0, [esp + nb104_rinvH2H2]
	movapd xmm1, xmm0
	mulpd xmm0, xmm0
	mulpd xmm1, [esp + nb104_qqHH]
	mulpd xmm0, xmm1	;# fsH2H2 
	addpd xmm7, xmm1	;# add to local vctot 
	movapd xmm1, xmm0
	movapd [esp + nb104_vctot], xmm7
	movapd xmm2, xmm0
	movapd xmm3, [esp + nb104_fjxH2]
	movapd xmm4, [esp + nb104_fjyH2]
	movapd xmm5, [esp + nb104_fjzH2]
	mulpd xmm0, [esp + nb104_dxH2H2]
	mulpd xmm1, [esp + nb104_dyH2H2]
	mulpd xmm2, [esp + nb104_dzH2H2]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb104_fixH2]
	addpd xmm1, [esp + nb104_fiyH2]
	addpd xmm2, [esp + nb104_fizH2]
	movapd [esp + nb104_fjxH2], xmm3
	movapd [esp + nb104_fjyH2], xmm4
	movapd [esp + nb104_fjzH2], xmm5
	movapd [esp + nb104_fixH2], xmm0
	movapd [esp + nb104_fiyH2], xmm1
	movapd [esp + nb104_fizH2], xmm2

	mov edi, [ebp + nb104_faction]
		
	;# Did all interactions - now update j forces 
	movlpd xmm0, [edi + eax*8 + 24]
	movlpd xmm1, [edi + eax*8 + 32]
	movlpd xmm2, [edi + eax*8 + 40]
	movlpd xmm3, [edi + eax*8 + 48]
	movlpd xmm4, [edi + eax*8 + 56]
	movlpd xmm5, [edi + eax*8 + 64]
	movlpd xmm6, [edi + eax*8 + 72]
	movlpd xmm7, [edi + eax*8 + 80]
	movhpd xmm0, [edi + ebx*8 + 24]
	movhpd xmm1, [edi + ebx*8 + 32]
	movhpd xmm2, [edi + ebx*8 + 40]
	movhpd xmm3, [edi + ebx*8 + 48]
	movhpd xmm4, [edi + ebx*8 + 56]
	movhpd xmm5, [edi + ebx*8 + 64]
	movhpd xmm6, [edi + ebx*8 + 72]
	movhpd xmm7, [edi + ebx*8 + 80]
	addpd xmm0, [esp + nb104_fjxH1]
	addpd xmm1, [esp + nb104_fjyH1]
	addpd xmm2, [esp + nb104_fjzH1]
	addpd xmm3, [esp + nb104_fjxH2]
	addpd xmm4, [esp + nb104_fjyH2]
	addpd xmm5, [esp + nb104_fjzH2]
	addpd xmm6, [esp + nb104_fjxM]
	addpd xmm7, [esp + nb104_fjyM]
	movlpd [edi + eax*8 + 24], xmm0
	movlpd [edi + eax*8 + 32], xmm1
	movlpd [edi + eax*8 + 40], xmm2
	movlpd [edi + eax*8 + 48], xmm3
	movlpd [edi + eax*8 + 56], xmm4
	movlpd [edi + eax*8 + 64], xmm5
	movlpd [edi + eax*8 + 72], xmm6
	movlpd [edi + eax*8 + 80], xmm7
	movhpd [edi + ebx*8 + 24], xmm0
	movhpd [edi + ebx*8 + 32], xmm1
	movhpd [edi + ebx*8 + 40], xmm2
	movhpd [edi + ebx*8 + 48], xmm3
	movhpd [edi + ebx*8 + 56], xmm4
	movhpd [edi + ebx*8 + 64], xmm5
	movhpd [edi + ebx*8 + 72], xmm6
	movhpd [edi + ebx*8 + 80], xmm7

	movlpd xmm0, [edi + eax*8 + 88]
	movhpd xmm0, [edi + ebx*8 + 88]
	addpd xmm0, [esp + nb104_fjzM]
	movlpd [edi + eax*8 + 88], xmm0
	movhpd [edi + ebx*8 + 88], xmm0
	
	;# should we do one more iteration? 
	sub dword ptr [esp + nb104_innerk],  2
	jl    .nb104_checksingle
	jmp   .nb104_unroll_loop
.nb104_checksingle:
	mov   edx, [esp + nb104_innerk]
	and   edx, 1
	jnz   .nb104_dosingle
	jmp   .nb104_updateouterdata
.nb104_dosingle:
	mov   edx, [esp + nb104_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [edx]	

	mov esi, [ebp + nb104_pos]
	lea   eax, [eax + eax*2]  

	;# move j coordinates to local temp variables 
	movlpd xmm2, [esi + eax*8 + 24]
	movlpd xmm3, [esi + eax*8 + 32]
	movlpd xmm4, [esi + eax*8 + 40]
	movlpd xmm5, [esi + eax*8 + 48]
	movlpd xmm6, [esi + eax*8 + 56]
	movlpd xmm7, [esi + eax*8 + 64]
	movapd 	[esp + nb104_jxH1], xmm2
	movapd 	[esp + nb104_jyH1], xmm3
	movapd 	[esp + nb104_jzH1], xmm4
	movapd 	[esp + nb104_jxH2], xmm5
	movapd 	[esp + nb104_jyH2], xmm6
	movapd 	[esp + nb104_jzH2], xmm7
	movlpd xmm2, [esi + eax*8 + 72]
	movlpd xmm3, [esi + eax*8 + 80]
	movlpd xmm4, [esi + eax*8 + 88]
	movapd 	[esp + nb104_jxM], xmm2
	movapd 	[esp + nb104_jyM], xmm3
	movapd 	[esp + nb104_jzM], xmm4
	
	movapd xmm0, [esp + nb104_ixM]
	movapd xmm1, [esp + nb104_iyM]
	movapd xmm2, [esp + nb104_izM]
	movapd xmm3, [esp + nb104_ixM]
	movapd xmm4, [esp + nb104_iyM]
	movapd xmm5, [esp + nb104_izM]
	subsd  xmm0, [esp + nb104_jxM]
	subsd  xmm1, [esp + nb104_jyM]
	subsd  xmm2, [esp + nb104_jzM]
	subsd  xmm3, [esp + nb104_jxH1]
	subsd  xmm4, [esp + nb104_jyH1]
	subsd  xmm5, [esp + nb104_jzH1]
	movapd [esp + nb104_dxMM], xmm0
	movapd [esp + nb104_dyMM], xmm1
	movapd [esp + nb104_dzMM], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb104_dxMH1], xmm3
	movapd [esp + nb104_dyMH1], xmm4
	movapd [esp + nb104_dzMH1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb104_rsqMM], xmm0
	movapd [esp + nb104_rsqMH1], xmm3

	movapd xmm0, [esp + nb104_ixM]
	movapd xmm1, [esp + nb104_iyM]
	movapd xmm2, [esp + nb104_izM]
	movapd xmm3, [esp + nb104_ixH1]
	movapd xmm4, [esp + nb104_iyH1]
	movapd xmm5, [esp + nb104_izH1]
	subsd  xmm0, [esp + nb104_jxH2]
	subsd  xmm1, [esp + nb104_jyH2]
	subsd  xmm2, [esp + nb104_jzH2]
	subsd  xmm3, [esp + nb104_jxM]
	subsd  xmm4, [esp + nb104_jyM]
	subsd  xmm5, [esp + nb104_jzM]
	movapd [esp + nb104_dxMH2], xmm0
	movapd [esp + nb104_dyMH2], xmm1
	movapd [esp + nb104_dzMH2], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb104_dxH1M], xmm3
	movapd [esp + nb104_dyH1M], xmm4
	movapd [esp + nb104_dzH1M], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb104_rsqMH2], xmm0
	movapd [esp + nb104_rsqH1M], xmm3

	movapd xmm0, [esp + nb104_ixH1]
	movapd xmm1, [esp + nb104_iyH1]
	movapd xmm2, [esp + nb104_izH1]
	movapd xmm3, [esp + nb104_ixH1]
	movapd xmm4, [esp + nb104_iyH1]
	movapd xmm5, [esp + nb104_izH1]
	subsd  xmm0, [esp + nb104_jxH1]
	subsd  xmm1, [esp + nb104_jyH1]
	subsd  xmm2, [esp + nb104_jzH1]
	subsd  xmm3, [esp + nb104_jxH2]
	subsd  xmm4, [esp + nb104_jyH2]
	subsd  xmm5, [esp + nb104_jzH2]
	movapd [esp + nb104_dxH1H1], xmm0
	movapd [esp + nb104_dyH1H1], xmm1
	movapd [esp + nb104_dzH1H1], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb104_dxH1H2], xmm3
	movapd [esp + nb104_dyH1H2], xmm4
	movapd [esp + nb104_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 + nb104_rsqH1H1], xmm0
	movapd [esp + nb104_rsqH1H2], xmm3

	movapd xmm0, [esp + nb104_ixH2]
	movapd xmm1, [esp + nb104_iyH2]
	movapd xmm2, [esp + nb104_izH2]
	movapd xmm3, [esp + nb104_ixH2]
	movapd xmm4, [esp + nb104_iyH2]
	movapd xmm5, [esp + nb104_izH2]
	subsd  xmm0, [esp + nb104_jxM]
	subsd  xmm1, [esp + nb104_jyM]
	subsd  xmm2, [esp + nb104_jzM]
	subsd  xmm3, [esp + nb104_jxH1]
	subsd  xmm4, [esp + nb104_jyH1]
	subsd  xmm5, [esp + nb104_jzH1]
	movapd [esp + nb104_dxH2M], xmm0
	movapd [esp + nb104_dyH2M], xmm1
	movapd [esp + nb104_dzH2M], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb104_dxH2H1], xmm3
	movapd [esp + nb104_dyH2H1], xmm4
	movapd [esp + nb104_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 + nb104_rsqH2M], xmm0
	movapd [esp + nb104_rsqH2H1], xmm4

	movapd xmm0, [esp + nb104_ixH2]
	movapd xmm1, [esp + nb104_iyH2]
	movapd xmm2, [esp + nb104_izH2]
	subsd  xmm0, [esp + nb104_jxH2]
	subsd  xmm1, [esp + nb104_jyH2]
	subsd  xmm2, [esp + nb104_jzH2]
	movapd [esp + nb104_dxH2H2], xmm0
	movapd [esp + nb104_dyH2H2], xmm1
	movapd [esp + nb104_dzH2H2], xmm2
	mulsd xmm0, xmm0
	mulsd xmm1, xmm1
	mulsd xmm2, xmm2
	addsd xmm0, xmm1
	addsd xmm0, xmm2
	movapd [esp + nb104_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 + nb104_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 + nb104_half] ;# iter1 
	mulsd   xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv 
	mulsd   xmm5, [esp + nb104_half] ;# rinv 
	movapd [esp + nb104_rinvH2H2], xmm1
	movapd [esp + nb104_rinvH2H1], xmm5

	movapd xmm0, [esp + nb104_rsqMM]
	movapd xmm4, [esp + nb104_rsqMH1]	
	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 + nb104_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 + nb104_half] ;# iter1 of  
	mulsd   xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv