nb_kernel314_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,300 行 · 第 1/5 页

	pslld   mm6, 2
	pslld   mm7, 2
	
	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7   	;# xmm5=Fp 
	mulps  xmm7, [esp + nb314_two]   	;# two*Heps2 
	movaps xmm3, [esp + nb314_qqHH]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# at this point mm5 contains vcoul and mm3 fijC 

	addps  xmm5, [esp + nb314_vctot]
	movaps [esp + nb314_vctot], xmm5
	xorps  xmm1, xmm1
	mulps  xmm3,  [esp + nb314_tsc]
	mulps  xmm3, xmm0
	subps  xmm1, xmm3

	movaps xmm0, xmm1
	movaps xmm2, xmm1
	
	xorps xmm3, xmm3
	movaps xmm4, xmm3
	movaps xmm5, xmm3
	mulps xmm0, [esp + nb314_dxH1H2]
	mulps xmm1, [esp + nb314_dyH1H2]
	mulps xmm2, [esp + nb314_dzH1H2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb314_fixH1]
	addps xmm1, [esp + nb314_fiyH1]
	addps xmm2, [esp + nb314_fizH1]
	movaps [esp + nb314_fjxH2], xmm3
	movaps [esp + nb314_fjyH2], xmm4
	movaps [esp + nb314_fjzH2], xmm5
	movaps [esp + nb314_fixH1], xmm0
	movaps [esp + nb314_fiyH1], xmm1
	movaps [esp + nb314_fizH1], xmm2

	;# H1-M interaction  
	movaps xmm0, [esp + nb314_rinvH1M]
	movaps xmm1, xmm0
	mulps  xmm1, [esp + nb314_rsqH1M] ;# xmm1=r 
	mulps  xmm1, [esp + nb314_tsc]	
	movhlps xmm2, xmm1
	cvttps2pi mm6, xmm1
	cvttps2pi mm7, xmm2 	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm3, mm6
	cvtpi2ps xmm2, mm7
	movlhps  xmm3, xmm2
	subps    xmm1, xmm3	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2   	;# xmm2=eps2 

	pslld   mm6, 2
	pslld   mm7, 2

	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7   	;# xmm5=Fp 
	mulps  xmm7, [esp + nb314_two]   	;# two*Heps2 
	movaps xmm3, [esp + nb314_qqMH]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# at this point mm5 contains vcoul and mm3 fijC 

	addps  xmm5, [esp + nb314_vctot]
	movaps [esp + nb314_vctot], xmm5
	xorps  xmm1, xmm1
	mulps  xmm3,  [esp + nb314_tsc]
	mulps  xmm3, xmm0
	subps  xmm1, xmm3

	movaps xmm0, xmm1
	movaps xmm2, xmm1
	
	xorps xmm3, xmm3
	movaps xmm4, xmm3
	movaps xmm5, xmm3
	mulps xmm0, [esp + nb314_dxH1M]
	mulps xmm1, [esp + nb314_dyH1M]
	mulps xmm2, [esp + nb314_dzH1M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb314_fixH1]
	addps xmm1, [esp + nb314_fiyH1]
	addps xmm2, [esp + nb314_fizH1]
	movaps [esp + nb314_fjxM], xmm3
	movaps [esp + nb314_fjyM], xmm4
	movaps [esp + nb314_fjzM], xmm5
	movaps [esp + nb314_fixH1], xmm0
	movaps [esp + nb314_fiyH1], xmm1
	movaps [esp + nb314_fizH1], xmm2

	;# H2-H1 interaction 
	movaps xmm0, [esp + nb314_rinvH2H1]
	movaps xmm1, xmm0
	mulps  xmm1, [esp + nb314_rsqH2H1] ;# xmm1=r 
	mulps  xmm1, [esp + nb314_tsc]	
	movhlps xmm2, xmm1
	cvttps2pi mm6, xmm1
	cvttps2pi mm7, xmm2 	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm3, mm6
	cvtpi2ps xmm2, mm7
	movlhps  xmm3, xmm2
	subps    xmm1, xmm3	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2   	;# xmm2=eps2 

	pslld   mm6, 2
	pslld   mm7, 2

	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7   	;# xmm5=Fp 
	mulps  xmm7, [esp + nb314_two]   	;# two*Heps2 
	movaps xmm3, [esp + nb314_qqHH]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# at this point mm5 contains vcoul and mm3 fijC 

	addps  xmm5, [esp + nb314_vctot]
	movaps [esp + nb314_vctot], xmm5
	xorps  xmm1, xmm1
	mulps  xmm3,  [esp + nb314_tsc]
	mulps  xmm3, xmm0
	subps  xmm1, xmm3

	movaps xmm0, xmm1
	movaps xmm2, xmm1
	
	movaps xmm3, [esp + nb314_fjxH1]
	movaps xmm4, [esp + nb314_fjyH1]
	movaps xmm5, [esp + nb314_fjzH1]
	mulps xmm0, [esp + nb314_dxH2H1]
	mulps xmm1, [esp + nb314_dyH2H1]
	mulps xmm2, [esp + nb314_dzH2H1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb314_fixH2]
	addps xmm1, [esp + nb314_fiyH2]
	addps xmm2, [esp + nb314_fizH2]
	movaps [esp + nb314_fjxH1], xmm3
	movaps [esp + nb314_fjyH1], xmm4
	movaps [esp + nb314_fjzH1], xmm5
	movaps [esp + nb314_fixH2], xmm0
	movaps [esp + nb314_fiyH2], xmm1
	movaps [esp + nb314_fizH2], xmm2

	;# H2-H2 interaction 
	movaps xmm0, [esp + nb314_rinvH2H2]
	movaps xmm1, xmm0
	mulps  xmm1, [esp + nb314_rsqH2H2] ;# xmm1=r 
	mulps  xmm1, [esp + nb314_tsc]	
	movhlps xmm2, xmm1
	cvttps2pi mm6, xmm1
	cvttps2pi mm7, xmm2 	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm3, mm6
	cvtpi2ps xmm2, mm7
	movlhps  xmm3, xmm2
	subps    xmm1, xmm3	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2   	;# xmm2=eps2 

	pslld   mm6, 2
	pslld   mm7, 2

	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7   	;# xmm5=Fp 
	mulps  xmm7, [esp + nb314_two]   	;# two*Heps2 
	movaps xmm3, [esp + nb314_qqHH]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# at this point mm5 contains vcoul and mm3 fijC 

	addps  xmm5, [esp + nb314_vctot]
	movaps [esp + nb314_vctot], xmm5
	xorps  xmm1, xmm1
	mulps  xmm3,  [esp + nb314_tsc]
	mulps  xmm3, xmm0
	subps  xmm1, xmm3

	movaps xmm0, xmm1
	movaps xmm2, xmm1
	
	movaps xmm3, [esp + nb314_fjxH2]
	movaps xmm4, [esp + nb314_fjyH2]
	movaps xmm5, [esp + nb314_fjzH2]
	mulps xmm0, [esp + nb314_dxH2H2]
	mulps xmm1, [esp + nb314_dyH2H2]
	mulps xmm2, [esp + nb314_dzH2H2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb314_fixH2]
	addps xmm1, [esp + nb314_fiyH2]
	addps xmm2, [esp + nb314_fizH2]
	movaps [esp + nb314_fjxH2], xmm3
	movaps [esp + nb314_fjyH2], xmm4
	movaps [esp + nb314_fjzH2], xmm5
	movaps [esp + nb314_fixH2], xmm0
	movaps [esp + nb314_fiyH2], xmm1
	movaps [esp + nb314_fizH2], xmm2

	;# H2-M interaction 
	movaps xmm0, [esp + nb314_rinvH2M]
	movaps xmm1, xmm0
	mulps  xmm1, [esp + nb314_rsqH2M] ;# xmm1=r 
	mulps  xmm1, [esp + nb314_tsc]
	movhlps xmm2, xmm1
	cvttps2pi mm6, xmm1
	cvttps2pi mm7, xmm2 	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm3, mm6
	cvtpi2ps xmm2, mm7
	movlhps  xmm3, xmm2
	subps    xmm1, xmm3	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2   	;# xmm2=eps2 

	pslld   mm6, 2
	pslld   mm7, 2

	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7   	;# xmm5=Fp 
	mulps  xmm7, [esp + nb314_two]   	;# two*Heps2 
	movaps xmm3, [esp + nb314_qqMH]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# at this point mm5 contains vcoul and mm3 fijC 

	addps  xmm5, [esp + nb314_vctot]
	movaps [esp + nb314_vctot], xmm5
	xorps  xmm1, xmm1
	mulps  xmm3,  [esp + nb314_tsc]
	mulps  xmm3, xmm0
	subps  xmm1, xmm3

	movaps xmm0, xmm1
	movaps xmm2, xmm1
	
	movaps xmm3, [esp + nb314_fjxM]
	movaps xmm4, [esp + nb314_fjyM]
	movaps xmm5, [esp + nb314_fjzM]
	mulps xmm0, [esp + nb314_dxH2M]
	mulps xmm1, [esp + nb314_dyH2M]
	mulps xmm2, [esp + nb314_dzH2M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb314_fixH2]
	addps xmm1, [esp + nb314_fiyH2]
	addps xmm2, [esp + nb314_fizH2]
	movaps [esp + nb314_fjxM], xmm3
	movaps [esp + nb314_fjyM], xmm4
	movaps [esp + nb314_fjzM], xmm5
	movaps [esp + nb314_fixH2], xmm0
	movaps [esp + nb314_fiyH2], xmm1
	movaps [esp + nb314_fizH2], xmm2

	;# M-H1 interaction 
	movaps xmm0, [esp + nb314_rinvMH1]
	movaps xmm1, xmm0
	mulps  xmm1, [esp + nb314_rsqMH1] ;# xmm1=r 
	mulps  xmm1, [esp + nb314_tsc]	
	movhlps xmm2, xmm1
	cvttps2pi mm6, xmm1
	cvttps2pi mm7, xmm2 	;# mm6/mm7 contain lu indices 
	cvtpi2ps xmm3, mm6
	cvtpi2ps xmm2, mm7
	movlhps  xmm3, xmm2
	subps    xmm1, xmm3	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2   	;# xmm2=eps2 

	pslld   mm6, 2
	pslld   mm7, 2

	movd eax, mm6
	psrlq mm6, 32
	movd ecx, mm7
	psrlq mm7, 32
	movd ebx, mm6
	movd edx, mm7

	movlps xmm5, [esi + eax*4]
	movlps xmm7, [esi + ecx*4]
	movhps xmm5, [esi + ebx*4]
	movhps xmm7, [esi + edx*4] ;# got half coulomb table 

	movaps xmm4, xmm5
	shufps xmm4, xmm7, 136  ;# constant 10001000
	shufps xmm5, xmm7, 221  ;# constant 11011101

	movlps xmm7, [esi + eax*4 + 8]
	movlps xmm3, [esi + ecx*4 + 8]
	movhps xmm7, [esi + ebx*4 + 8]
	movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table  
	movaps xmm6, xmm7
	shufps xmm6, xmm3, 136  ;# constant 10001000
	shufps xmm7, xmm3, 221  ;# constant 11011101
	;# coulomb table ready, in xmm4-xmm7  

	mulps  xmm6, xmm1   	;# xmm6=Geps 
	mulps  xmm7, xmm2   	;# xmm7=Heps2