nb_kernel234_ia32_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,167 行 · 第 1/5 页

	subpd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulpd  xmm7, [esp + nb234_qqMH] ;# xmm7 = coul part of fscal 
	
	addpd  xmm6, [esp + nb234_vctot] ;# local vctot summation variable 
	mulpd  xmm0, xmm7
	movapd [esp + nb234_vctot], xmm6
	
	movapd xmm1, xmm0
	movapd xmm2, xmm0
	
	movapd xmm3, [esp + nb234_fjxH2]
	movapd xmm4, [esp + nb234_fjyH2]
	movapd xmm5, [esp + nb234_fjzH2]
	mulpd xmm0, [esp + nb234_dxMH2]
	mulpd xmm1, [esp + nb234_dyMH2]
	mulpd xmm2, [esp + nb234_dzMH2]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb234_fixM]
	addpd xmm1, [esp + nb234_fiyM]
	addpd xmm2, [esp + nb234_fizM]
	movapd [esp + nb234_fjxH2], xmm3
	movapd [esp + nb234_fjyH2], xmm4
	movapd [esp + nb234_fjzH2], xmm5
	movapd [esp + nb234_fixM], xmm0
	movapd [esp + nb234_fiyM], xmm1
	movapd [esp + nb234_fizM], xmm2

	;# M-M interaction 
	movapd xmm0, [esp + nb234_rinvMM]
	movapd xmm7, xmm0	;# xmm7=rinv 
	movapd xmm5, [esp + nb234_krf]
	mulpd  xmm0, xmm0	;# xmm0=rinvsq 

	mulpd  xmm5, [esp + nb234_rsqMM] ;# xmm5=krsq 
	movapd xmm6, xmm5
	addpd  xmm6, xmm7	;# xmm6=rinv+ krsq 
	subpd  xmm6, [esp + nb234_crf]
	mulpd  xmm6, [esp + nb234_qqMM] ;# xmm6=voul=qq*(rinv+ krsq-crf) 
	mulpd  xmm5, [esp + nb234_two]
	subpd  xmm7, xmm5	;# xmm7=rinv-2*krsq 
	mulpd  xmm7, [esp + nb234_qqMM] ;# xmm7 = coul part of fscal 
	
	addpd  xmm6, [esp + nb234_vctot] ;# local vctot summation variable 
	mulpd  xmm0, xmm7
	movapd [esp + nb234_vctot], xmm6
	
	movapd xmm1, xmm0
	movapd xmm2, xmm0

	movapd xmm3, [esp + nb234_fjxM]
	movapd xmm4, [esp + nb234_fjyM] 
	movapd xmm5, [esp + nb234_fjzM]
	mulpd xmm0, [esp + nb234_dxMM]
	mulpd xmm1, [esp + nb234_dyMM]
	mulpd xmm2, [esp + nb234_dzMM]
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2
	addpd xmm0, [esp + nb234_fixM]
	addpd xmm1, [esp + nb234_fiyM]
	addpd xmm2, [esp + nb234_fizM]
	movapd [esp + nb234_fjxM], xmm3
	movapd [esp + nb234_fjyM], xmm4
	movapd [esp + nb234_fjzM], xmm5
	movapd [esp + nb234_fixM], xmm0
	movapd [esp + nb234_fiyM], xmm1
	movapd [esp + nb234_fizM], xmm2

	mov edi, [ebp + nb234_faction]
	
	;# Did all interactions - now update j forces 
	;# Step1 - transpose fjxO, fjyO and fjzO, fjxH1
	movapd xmm0, [esp + nb234_fjxO]
	movapd xmm1, [esp + nb234_fjyO]
	movapd xmm2, [esp + nb234_fjzO]
	movapd xmm3, [esp + nb234_fjxH1]
	movapd xmm4, xmm0
	movapd xmm5, xmm2
	unpcklpd xmm0, xmm1   ;# fjOxA fjOyA
	unpckhpd xmm4, xmm1   ;# fjOxB fjOyB
	unpcklpd xmm2, xmm3   ;# fjOzA fjH1xA
	unpckhpd xmm5, xmm3   ;# fjOzB fjH1xB
	movlpd xmm1, [edi + eax*8]
	movlpd xmm3, [edi + ebx*8]
	movhpd xmm1, [edi + eax*8 + 8]
	movhpd xmm3, [edi + ebx*8 + 8]
	movlpd xmm6, [edi + eax*8 + 16]
	movlpd xmm7, [edi + ebx*8 + 16]
	movhpd xmm6, [edi + eax*8 + 24]
	movhpd xmm7, [edi + ebx*8 + 24]
	addpd  xmm1, xmm0
	addpd  xmm3, xmm4
	addpd  xmm6, xmm2
	addpd  xmm7, xmm5
	movlpd [edi + eax*8],      xmm1
	movlpd [edi + ebx*8],      xmm3
	movhpd [edi + eax*8 + 8],  xmm1
	movhpd [edi + ebx*8 + 8],  xmm3
	movlpd [edi + eax*8 + 16], xmm6
	movlpd [edi + ebx*8 + 16], xmm7
	movhpd [edi + eax*8 + 24], xmm6
	movhpd [edi + ebx*8 + 24], xmm7

	;# Step2 - transpose fjyH1, fjzH1 and fjxH2, fjyH2
	movapd xmm0, [esp + nb234_fjyH1]
	movapd xmm1, [esp + nb234_fjzH1]
	movapd xmm2, [esp + nb234_fjxH2]
	movapd xmm3, [esp + nb234_fjyH2]
	movapd xmm4, xmm0
	movapd xmm5, xmm2
	unpcklpd xmm0, xmm1   ;# fjOxA fjOyA
	unpckhpd xmm4, xmm1   ;# fjOxB fjOyB
	unpcklpd xmm2, xmm3   ;# fjOzA fjH1xA
	unpckhpd xmm5, xmm3   ;# fjOzB fjH1xB
	movlpd xmm1, [edi + eax*8 + 32]
	movlpd xmm3, [edi + ebx*8 + 32]
	movhpd xmm1, [edi + eax*8 + 40]
	movhpd xmm3, [edi + ebx*8 + 40]
	movlpd xmm6, [edi + eax*8 + 48]
	movlpd xmm7, [edi + ebx*8 + 48]
	movhpd xmm6, [edi + eax*8 + 56]
	movhpd xmm7, [edi + ebx*8 + 56]
	addpd  xmm1, xmm0
	addpd  xmm3, xmm4
	addpd  xmm6, xmm2
	addpd  xmm7, xmm5
	movlpd [edi + eax*8 + 32], xmm1
	movlpd [edi + ebx*8 + 32], xmm3
	movhpd [edi + eax*8 + 40], xmm1
	movhpd [edi + ebx*8 + 40], xmm3
	movlpd [edi + eax*8 + 48], xmm6
	movlpd [edi + ebx*8 + 48], xmm7
	movhpd [edi + eax*8 + 56], xmm6
	movhpd [edi + ebx*8 + 56], xmm7

	;# Step3 - transpose fjzH2, fjxM and fjyM, fjzM
	movapd xmm0, [esp + nb234_fjzH2]
	movapd xmm1, [esp + nb234_fjxM]
	movapd xmm2, [esp + nb234_fjyM]
	movapd xmm3, [esp + nb234_fjzM]
	movapd xmm4, xmm0
	movapd xmm5, xmm2
	unpcklpd xmm0, xmm1   ;# fjOxA fjOyA
	unpckhpd xmm4, xmm1   ;# fjOxB fjOyB
	unpcklpd xmm2, xmm3   ;# fjOzA fjH1xA
	unpckhpd xmm5, xmm3   ;# fjOzB fjH1xB
	movlpd xmm1, [edi + eax*8 + 64]
	movlpd xmm3, [edi + ebx*8 + 64]
	movhpd xmm1, [edi + eax*8 + 72]
	movhpd xmm3, [edi + ebx*8 + 72]
	movlpd xmm6, [edi + eax*8 + 80]
	movlpd xmm7, [edi + ebx*8 + 80]
	movhpd xmm6, [edi + eax*8 + 88]
	movhpd xmm7, [edi + ebx*8 + 88]
	addpd  xmm1, xmm0
	addpd  xmm3, xmm4
	addpd  xmm6, xmm2
	addpd  xmm7, xmm5
	movlpd [edi + eax*8 + 64], xmm1
	movlpd [edi + ebx*8 + 64], xmm3
	movhpd [edi + eax*8 + 72], xmm1
	movhpd [edi + ebx*8 + 72], xmm3
	movlpd [edi + eax*8 + 80], xmm6
	movlpd [edi + ebx*8 + 80], xmm7
	movhpd [edi + eax*8 + 88], xmm6
	movhpd [edi + ebx*8 + 88], xmm7

	;# should we do one more iteration? 
	sub dword ptr [esp + nb234_innerk],  2
	jl    .nb234_checksingle
	jmp   .nb234_unroll_loop
.nb234_checksingle:
	mov   edx, [esp + nb234_innerk]
	and   edx, 1
	jnz   .nb234_dosingle
	jmp   .nb234_updateouterdata
.nb234_dosingle:
	mov   edx, [esp + nb234_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [edx]

	mov esi, [ebp + nb234_pos]       ;# base of pos[] 

	lea   eax, [eax + eax*2]     ;# replace jnr with j3 
	
	;# move j coordinates to local temp variables 
	;# load ox, oy, oz, h1x
	movlpd xmm0, [esi + eax*8]
	movhpd xmm0, [esi + eax*8 + 8]
	movlpd xmm1, [esi + eax*8 + 16]
	movhpd xmm1, [esi + eax*8 + 24]
	movlpd xmm2, [esi + eax*8 + 32]
	movhpd xmm2, [esi + eax*8 + 40]
	movlpd xmm3, [esi + eax*8 + 48]
	movhpd xmm3, [esi + eax*8 + 56]
	movlpd xmm4, [esi + eax*8 + 64]
	movhpd xmm4, [esi + eax*8 + 72]
	movlpd xmm5, [esi + eax*8 + 80]
	movhpd xmm5, [esi + eax*8 + 88]
	movsd  [esp + nb234_jxO], xmm0
	movsd  [esp + nb234_jzO], xmm1
	movsd  [esp + nb234_jyH1], xmm2
	movsd  [esp + nb234_jxH2], xmm3
	movsd  [esp + nb234_jzH2], xmm4
	movsd  [esp + nb234_jyM], xmm5
	unpckhpd xmm0, xmm0
	unpckhpd xmm1, xmm1
	unpckhpd xmm2, xmm2
	unpckhpd xmm3, xmm3
	unpckhpd xmm4, xmm4
	unpckhpd xmm5, xmm5
	movsd  [esp + nb234_jyO], xmm0
	movsd  [esp + nb234_jxH1], xmm1
	movsd  [esp + nb234_jzH1], xmm2
	movsd  [esp + nb234_jyH2], xmm3
	movsd  [esp + nb234_jxM], xmm4
	movsd  [esp + nb234_jzM], xmm5

	;# start calculating pairwise distances
	movapd xmm0, [esp + nb234_ixO]
	movapd xmm1, [esp + nb234_iyO]
	movapd xmm2, [esp + nb234_izO]
	movapd xmm3, [esp + nb234_ixH1]
	movapd xmm4, [esp + nb234_iyH1]
	movapd xmm5, [esp + nb234_izH1]
	subsd  xmm0, [esp + nb234_jxO]
	subsd  xmm1, [esp + nb234_jyO]
	subsd  xmm2, [esp + nb234_jzO]
	subsd  xmm3, [esp + nb234_jxH1]
	subsd  xmm4, [esp + nb234_jyH1]
	subsd  xmm5, [esp + nb234_jzH1]
	movapd [esp + nb234_dxOO], xmm0
	movapd [esp + nb234_dyOO], xmm1
	movapd [esp + nb234_dzOO], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb234_dxH1H1], xmm3
	movapd [esp + nb234_dyH1H1], xmm4
	movapd [esp + nb234_dzH1H1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb234_rsqOO], xmm0
	movapd [esp + nb234_rsqH1H1], xmm3

	movapd xmm0, [esp + nb234_ixH1]
	movapd xmm1, [esp + nb234_iyH1]
	movapd xmm2, [esp + nb234_izH1]
	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	subsd  xmm0, [esp + nb234_jxH2]
	subsd  xmm1, [esp + nb234_jyH2]
	subsd  xmm2, [esp + nb234_jzH2]
	subsd  xmm3, [esp + nb234_jxM]
	subsd  xmm4, [esp + nb234_jyM]
	subsd  xmm5, [esp + nb234_jzM]
	movapd [esp + nb234_dxH1H2], xmm0
	movapd [esp + nb234_dyH1H2], xmm1
	movapd [esp + nb234_dzH1H2], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb234_dxH1M], xmm3
	movapd [esp + nb234_dyH1M], xmm4
	movapd [esp + nb234_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 + nb234_rsqH1H2], xmm0
	movapd [esp + nb234_rsqH1M], xmm3

	movapd xmm0, [esp + nb234_ixH2]
	movapd xmm1, [esp + nb234_iyH2]
	movapd xmm2, [esp + nb234_izH2]
	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	subsd  xmm0, [esp + nb234_jxH1]
	subsd  xmm1, [esp + nb234_jyH1]
	subsd  xmm2, [esp + nb234_jzH1]
	subsd  xmm3, [esp + nb234_jxH2]
	subsd  xmm4, [esp + nb234_jyH2]
	subsd  xmm5, [esp + nb234_jzH2]
	movapd [esp + nb234_dxH2H1], xmm0
	movapd [esp + nb234_dyH2H1], xmm1
	movapd [esp + nb234_dzH2H1], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb234_dxH2H2], xmm3
	movapd [esp + nb234_dyH2H2], xmm4
	movapd [esp + nb234_dzH2H2], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
	movapd [esp + nb234_rsqH2H1], xmm0
	movapd [esp + nb234_rsqH2H2], xmm3

	movapd xmm0, [esp + nb234_ixH2]
	movapd xmm1, [esp + nb234_iyH2]
	movapd xmm2, [esp + nb234_izH2]
	movapd xmm3, [esp + nb234_ixM]
	movapd xmm4, [esp + nb234_iyM]
	movapd xmm5, [esp + nb234_izM]
	subsd  xmm0, [esp + nb234_jxM]
	subsd  xmm1, [esp + nb234_jyM]
	subsd  xmm2, [esp + nb234_jzM]
	subsd  xmm3, [esp + nb234_jxH1]
	subsd  xmm4, [esp + nb234_jyH1]
	subsd  xmm5, [esp + nb234_jzH1]
	movapd [esp + nb234_dxH2M], xmm0
	movapd [esp + nb234_dyH2M], xmm1
	movapd [esp + nb234_dzH2M], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb234_dxMH1], xmm3
	movapd [esp + nb234_dyMH1], xmm4
	movapd [esp + nb234_dzMH1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm4, xmm3
	addsd  xmm4, xmm5
	movapd [esp + nb234_rsqH2M], xmm0
	movapd [esp + nb234_rsqMH1], xmm4

	movapd xmm0, [esp + nb234_ixM]
	movapd xmm1, [esp + nb234_iyM]
	movapd xmm2, [esp + nb234_izM]
	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	subsd  xmm0, [esp + nb234_jxH2]
	subsd  xmm1, [esp + nb234_jyH2]
	subsd  xmm2, [esp + nb234_jzH2]
	subsd  xmm3, [esp + nb234_jxM]
	subsd  xmm4, [esp + nb234_jyM]
	subsd  xmm5, [esp + nb234_jzM]
	movapd [esp + nb234_dxMH2], xmm0
	movapd [esp + nb234_dyMH2], xmm1
	movapd [esp + nb234_dzMH2], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [esp + nb234_dxMM], xmm3
	movapd [esp + nb234_dyMM], xmm4
	movapd [esp + nb234_dzMM], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm4, xmm3
	addsd  xmm4, xmm5
	movapd [esp + nb234_rsqMH2], xmm0
	movapd [esp + nb234_rsqMM], xmm4

	;# Invsqrt form rsq M-H2 (rsq in xmm0) and MM (rsq in xmm4) 
	cvtsd2ss xmm1, xmm0
	cvtsd2ss xmm5, xmm4
	rsqrtss xmm1, xmm1
	rsqrtss xmm5, xmm5
	cvtss2sd xmm1, xmm1   ;# luA
	cvtss2sd xmm5, xmm5   ;# luB
	
	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 + nb234_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 + nb234_half] ;# iter1 
	mulsd   xmm7, [esp + nb234_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 + nb234_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 + nb234_half] ;# rinv 
	mulsd   xmm5, [esp + nb234_half] ;# rinv 
	movapd [esp + nb234_rinvMH2], xmm1
	movapd [esp + nb234_rinvMM], xmm5

	movapd xmm0, [esp + nb234_rsqOO]
	movapd xmm4, [esp + nb234_rsqH1H1]	
	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 + nb234_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