nb_kernel104_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,174 行 · 第 1/5 页

	mulps xmm2, [esp + nb104_dzMH1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb104_fixM]
	addps xmm1, [esp + nb104_fiyM]
	addps xmm2, [esp + nb104_fizM]
	movaps [esp + nb104_fjxH1], xmm3
	movaps [esp + nb104_fjyH1], xmm4
	movaps [esp + nb104_fjzH1], xmm5
	movaps [esp + nb104_fixM], xmm0
	movaps [esp + nb104_fiyM], xmm1
	movaps [esp + nb104_fizM], xmm2
	
	;# M-H2 interaction 
	movaps xmm0, [esp + nb104_rinvMH2]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb104_qqMH]
	mulps xmm0, xmm1	;# fs M-H2 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb104_fjxH2]
	movaps xmm4, [esp + nb104_fjyH2]
	movaps xmm5, [esp + nb104_fjzH2]
	mulps xmm0, [esp + nb104_dxMH2]
	mulps xmm1, [esp + nb104_dyMH2]
	mulps xmm2, [esp + nb104_dzMH2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb104_fixM]
	addps xmm1, [esp + nb104_fiyM]
	addps xmm2, [esp + nb104_fizM]
	movaps [esp + nb104_fjxH2], xmm3
	movaps [esp + nb104_fjyH2], xmm4
	movaps [esp + nb104_fjzH2], xmm5
	movaps [esp + nb104_fixM], xmm0
	movaps [esp + nb104_fiyM], xmm1
	movaps [esp + nb104_fizM], xmm2

	;# M-M interaction 
	movaps xmm0, [esp + nb104_rinvMM]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb104_qqMM]
	mulps xmm0, xmm1	;# fs M-M 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps [esp + nb104_vctot], xmm7
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb104_fjxM]
	movaps xmm4, [esp + nb104_fjyM]
	movaps xmm5, [esp + nb104_fjzM]
	mulps xmm0, [esp + nb104_dxMM]
	mulps xmm1, [esp + nb104_dyMM]
	mulps xmm2, [esp + nb104_dzMM]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb104_fixM]
	addps xmm1, [esp + nb104_fiyM]
	addps xmm2, [esp + nb104_fizM]
	movaps [esp + nb104_fjxM], xmm3
	movaps [esp + nb104_fjyM], xmm4
	movaps [esp + nb104_fjzM], xmm5
	movaps [esp + nb104_fixM], xmm0
	movaps [esp + nb104_fiyM], xmm1
	movaps [esp + nb104_fizM], xmm2

	mov edi, [ebp + nb104_faction]
		
	;# Did all interactions - now update j forces 
	;# At this stage forces are still on the stack, in positions:
	;# fjxH1, fjyH1, fjzH1, ... , fjzM.
	;# Each position is a quadruplet of forces for the four 
	;# corresponding j waters, so we need to transpose them before
	;# adding to the memory positions.
	;# 
	;# This _used_ to be a simple transpose, but the resulting high number
	;# of unaligned 128-bit load/stores might trigger a possible hardware 
	;# bug on Athlon and Opteron chips, so I have worked around it
	;# to use 64-bit load/stores instead. The performance hit should be
	;# very modest, since the 128-bit unaligned memory instructions were
	;# slow anyway. 

	;# 4 j waters with three atoms each - first do 1st Hydrogen X & Y forces for 4 j particles 
	movaps xmm0, [esp + nb104_fjxH1] ;# xmm0= fjxH1a  fjxH1b  fjxH1c  fjxH1d 
	movaps xmm2, [esp + nb104_fjyH1] ;# xmm1= fjyH1a  fjyH1b  fjyH1c  fjyH1d
	movlps xmm3, [edi + eax*4 + 12]
	movlps xmm4, [edi + ecx*4 + 12]
	movaps xmm1, xmm0
	unpcklps xmm0, xmm2    	   ;# xmm0= fjxH1a  fjyH1a  fjxH1b  fjyH1b
	unpckhps xmm1, xmm2        ;# xmm1= fjxH1c  fjyH1c  fjxH1d  fjyH1d
	movhps xmm3, [edi + ebx*4 + 12]
	movhps xmm4, [edi + edx*4 + 12]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	movlps [edi + eax*4 + 12], xmm3
	movlps [edi + ecx*4 + 12], xmm4
	movhps [edi + ebx*4 + 12], xmm3
	movhps [edi + edx*4 + 12], xmm4

	;# constant 1st Hydrogen Z & 2nd hydrogen X forces for 4 j particles 
	movaps xmm0, [esp + nb104_fjzH1]  ;# xmm0= fjzH1a   fjzH1b   fjzH1c   fjzH1d 
	movaps xmm2, [esp + nb104_fjxH2] ;# xmm1= fjxH2a  fjxH2b  fjxH2c  fjxH2d
	movlps xmm3, [edi + eax*4 + 20]
	movlps xmm4, [edi + ecx*4 + 20]
	movaps xmm1, xmm0
	unpcklps xmm0, xmm2    	   ;# xmm0= fjzH1a  fjxH2a  fjzH1b  fjxH2b
	unpckhps xmm1, xmm2        ;# xmm1= fjzH1c  fjxH2c  fjzH1d  fjxH2d
	movhps xmm3, [edi + ebx*4 + 20]
	movhps xmm4, [edi + edx*4 + 20]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	movlps [edi + eax*4 + 20], xmm3
	movlps [edi + ecx*4 + 20], xmm4
	movhps [edi + ebx*4 + 20], xmm3
	movhps [edi + edx*4 + 20], xmm4
	
	;# constant 2nd hydrogen Y & Z forces for 4 j particles 
	movaps xmm0, [esp + nb104_fjyH2] ;# xmm0= fjyH2a  fjyH2b  fjyH2c  fjyH2d 
	movaps xmm2, [esp + nb104_fjzH2] ;# xmm1= fjzH2a  fjzH2b  fjzH2c  fjzH2d
	movlps xmm3, [edi + eax*4 + 28]
	movlps xmm4, [edi + ecx*4 + 28]
	movaps xmm1, xmm0
	unpcklps xmm0, xmm2		;# xmm0= fjyH2a  fjzH2a  fjyH2b  fjzH2b
	unpckhps xmm1, xmm2		;# xmm1= fjyH2c  fjzH2c  fjyH2d  fjzH2d
	movhps xmm3, [edi + ebx*4 + 28]
	movhps xmm4, [edi + edx*4 + 28]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	movlps [edi + eax*4 + 28], xmm3
	movlps [edi + ecx*4 + 28], xmm4
	movhps [edi + ebx*4 + 28], xmm3
	movhps [edi + edx*4 + 28], xmm4

	;# Dummy (M) X & Y forces for 4 j particles 
	movaps xmm0, [esp + nb104_fjxM] ;# xmm0= fjxMa  fjxMb  fjxMc  fjxMd 
	movaps xmm2, [esp + nb104_fjyM] ;# xmm1= fjyMa  fjyMb  fjyMc  fjyMd
	movlps xmm3, [edi + eax*4 + 36]
	movlps xmm4, [edi + ecx*4 + 36]
	movaps xmm1, xmm0
	unpcklps xmm0, xmm2		;# xmm0= fjxMa  fjyMa  fjxMb  fjyMb
	unpckhps xmm1, xmm2		;# xmm1= fjxMc  fjyMc  fjxMd  fjyMd
	movhps xmm3, [edi + ebx*4 + 36]
	movhps xmm4, [edi + edx*4 + 36]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	movlps [edi + eax*4 + 36], xmm3
	movlps [edi + ecx*4 + 36], xmm4
	movhps [edi + ebx*4 + 36], xmm3
	movhps [edi + edx*4 + 36], xmm4

	
	;# Dummy (M) Z forces for 4 j particles 
	;# Just load the four Z coords into one reg. each
	movss xmm4, [edi + eax*4 + 44]
	movss xmm5, [edi + ebx*4 + 44]
	movss xmm6, [edi + ecx*4 + 44]
	movss xmm7, [edi + edx*4 + 44]
	;# add what we have on the stack
	addss xmm4, [esp + nb104_fjzM] 
	addss xmm5, [esp + nb104_fjzMb] 
	addss xmm6, [esp + nb104_fjzMc] 
	addss xmm7, [esp + nb104_fjzMd]
	;# store back
	movss [edi + eax*4 + 44], xmm4
	movss [edi + ebx*4 + 44], xmm5
	movss [edi + ecx*4 + 44], xmm6
	movss [edi + edx*4 + 44], xmm7
		
	;# should we do one more iteration? 
	sub dword ptr [esp + nb104_innerk],  4
	jl    .nb104_single_check
	jmp   .nb104_unroll_loop
.nb104_single_check:
	add dword ptr [esp + nb104_innerk],  4
	jnz   .nb104_single_loop
	jmp   .nb104_updateouterdata
.nb104_single_loop:
	mov   edx, [esp + nb104_innerjjnr] 	;# pointer to jjnr[k] 
	mov   eax, [edx]	
	add dword ptr [esp + nb104_innerjjnr],  4	

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

	;# fetch j coordinates 
	xorps xmm3, xmm3
	xorps xmm4, xmm4
	xorps xmm5, xmm5
	movss xmm3, [esi + eax*4 + 36]		;# jxM  -  -  -
	movss xmm4, [esi + eax*4 + 40]		;# jyM  -  -  -
	movss xmm5, [esi + eax*4 + 44]		;# jzM  -  -  -  

	movlps xmm6, [esi + eax*4 + 12]		;# xmm6 = jxH1 jyH1   -    -
	movss  xmm7, [esi + eax*4 + 20]		;# xmm7 = jzH1   -    -    - 
	movhps xmm6, [esi + eax*4 + 24]		;# xmm6 = jxH1 jyH1 jxH2 jyH2
	movss  xmm2, [esi + eax*4 + 32]		;# xmm2 = jzH2   -    -    -
	
	;# have all coords, time for some shuffling.

	shufps xmm6, xmm6, 216 ;# constant 11011000	;# xmm6 = jxH1 jxH2 jyH1 jyH2 
	unpcklps xmm7, xmm2			;# xmm7 = jzH1 jzH2   -    -
	movaps  xmm0, [esp + nb104_ixM]     
	movaps  xmm1, [esp + nb104_iyM]
	movaps  xmm2, [esp + nb104_izM]	
	movlhps xmm3, xmm6			;# xmm3 = jxM   0   jxH1 jxH2 
	shufps  xmm4, xmm6, 228 ;# constant 11100100	;# xmm4 = jyM   0   jyH1 jyH2 
	shufps  xmm5, xmm7, 68  ;# constant 01000100	;# xmm5 = jzM   0   jzH1 jzH2
	
	;# store all j coordinates in jM 
	movaps [esp + nb104_jxM], xmm3
	movaps [esp + nb104_jyM], xmm4
	movaps [esp + nb104_jzM], xmm5
	subps  xmm0, xmm3
	subps  xmm1, xmm4
	subps  xmm2, xmm5
	movaps [esp + nb104_dxMM], xmm0
	movaps [esp + nb104_dyMM], xmm1
	movaps [esp + nb104_dzMM], xmm2
	mulps xmm0, xmm0
	mulps xmm1, xmm1
	mulps xmm2, xmm2
	addps xmm0, xmm1
	addps xmm0, xmm2	;# have rsq in xmm0 
	
	;# do invsqrt 
	rsqrtps xmm1, xmm0
	movaps  xmm2, xmm1	
	mulps   xmm1, xmm1
	movaps  xmm3, [esp + nb104_three]
	mulps   xmm1, xmm0
	subps   xmm3, xmm1
	mulps   xmm3, xmm2
	mulps   xmm3, [esp + nb104_half] ;# rinv iM- j water 

	xorps   xmm1, xmm1
	movaps  xmm0, xmm3
	xorps   xmm4, xmm4
	mulps   xmm0, xmm0	;# xmm0=rinvsq
	
	;# fetch charges to xmm4
	movss   xmm4, [esp + nb104_qqMM] 
	movhps  xmm4, [esp + nb104_qqMH]
	
	mulps   xmm3, xmm4	;# xmm3=vcoul 
	mulps   xmm0, xmm3	;# total fscal 
	addps   xmm3, [esp + nb104_vctot]
	movaps  [esp + nb104_vctot], xmm3	

	movaps  xmm1, xmm0
	movaps  xmm2, xmm0
	mulps   xmm0, [esp + nb104_dxMM]
	mulps   xmm1, [esp + nb104_dyMM]
	mulps   xmm2, [esp + nb104_dzMM]
	;# initial update for j forces 
	xorps   xmm3, xmm3
	xorps   xmm4, xmm4
	xorps   xmm5, xmm5
	subps   xmm3, xmm0
	subps   xmm4, xmm1
	subps   xmm5, xmm2
	movaps  [esp + nb104_fjxM], xmm3
	movaps  [esp + nb104_fjyM], xmm4
	movaps  [esp + nb104_fjzM], xmm5
	addps   xmm0, [esp + nb104_fixM]
	addps   xmm1, [esp + nb104_fiyM]
	addps   xmm2, [esp + nb104_fizM]
	movaps  [esp + nb104_fixM], xmm0
	movaps  [esp + nb104_fiyM], xmm1
	movaps  [esp + nb104_fizM], xmm2

	;# done with i M Now do i H1 & H2 simultaneously first get i particle coords: 
	movaps  xmm0, [esp + nb104_ixH1]
	movaps  xmm1, [esp + nb104_iyH1]
	movaps  xmm2, [esp + nb104_izH1]	
	movaps  xmm3, [esp + nb104_ixH2] 
	movaps  xmm4, [esp + nb104_iyH2] 
	movaps  xmm5, [esp + nb104_izH2] 
	subps   xmm0, [esp + nb104_jxM]
	subps   xmm1, [esp + nb104_jyM]
	subps   xmm2, [esp + nb104_jzM]
	subps   xmm3, [esp + nb104_jxM]
	subps   xmm4, [esp + nb104_jyM]
	subps   xmm5, [esp + nb104_jzM]
	movaps [esp + nb104_dxH1M], xmm0
	movaps [esp + nb104_dyH1M], xmm1
	movaps [esp + nb104_dzH1M], xmm2
	movaps [esp + nb104_dxH2M], xmm3
	movaps [esp + nb104_dyH2M], xmm4
	movaps [esp + nb104_dzH2M], xmm5
	mulps xmm0, xmm0
	mulps xmm1, xmm1
	mulps xmm2, xmm2
	mulps xmm3, xmm3
	mulps xmm4, xmm4
	mulps xmm5, xmm5
	addps xmm0, xmm1
	addps xmm4, xmm3
	addps xmm0, xmm2	;# have rsqH1 in xmm0 
	addps xmm4, xmm5	;# have rsqH2 in xmm4 

	;# do invsqrt 
	rsqrtps xmm1, xmm0
	rsqrtps xmm5, xmm4
	movaps  xmm2, xmm1
	movaps  xmm6, xmm5
	mulps   xmm1, xmm1
	mulps   xmm5, xmm5
	movaps  xmm3, [esp + nb104_three]
	movaps  xmm7, xmm3
	mulps   xmm1, xmm0
	mulps   xmm5, xmm4
	subps   xmm3, xmm1
	subps   xmm7, xmm5
	mulps   xmm3, xmm2
	mulps   xmm7, xmm6
	mulps   xmm3, [esp + nb104_half] ;# rinv H1 - j water 
	mulps   xmm7, [esp + nb104_half] ;# rinv H2 - j water  

	;# assemble charges in xmm6 
	xorps   xmm6, xmm6
	movss   xmm6, [esp + nb104_qqMH]
	movhps  xmm6, [esp + nb104_qqHH]
	
	;# do coulomb interaction 
	movaps  xmm0, xmm3
	movaps  xmm4, xmm7
	mulps   xmm0, xmm0	;# rinvsq 
	mulps   xmm4, xmm4	;# rinvsq 
	mulps   xmm3, xmm6	;# vcoul 
	mulps   xmm7, xmm6	;# vcoul 
	movaps  xmm2, xmm3
	addps   xmm2, xmm7	;# total vcoul 
	mulps   xmm0, xmm3	;# fscal 
	
	addps   xmm2, [esp + nb104_vctot]
	mulps   xmm7, xmm4	;# fscal 
	movaps  [esp + nb104_vctot], xmm2
	movaps  xmm1, xmm0
	movaps  xmm2, xmm0
	mulps   xmm0, [esp + nb104_dxH1M]
	mulps   xmm1, [esp + nb104_dyH1M]
	mulps   xmm2, [esp + nb104_dzH1M]
	;# update forces H1 - j water 
	movaps  xmm3, [esp + nb104_fjxM]
	movaps  xmm4, [esp + nb104_fjyM]
	movaps  xmm5, [esp + nb104_fjzM]
	subps   xmm3, xmm0
	subps   xmm4, xmm1
	subps   xmm5, xmm2
	movaps  [esp + nb104_fjxM], xmm3
	movaps  [esp + nb104_fjyM], xmm4
	movaps  [esp + nb104_fjzM], xmm5
	addps   xmm0, [esp + nb104_fixH1]
	addps   xmm1, [esp + nb104_fiyH1]
	addps   xmm2, [esp + nb104_fizH1]
	movaps  [esp + nb104_fixH1], xmm0
	movaps  [esp + nb104_fiyH1], xmm1
	movaps  [esp + nb104_fizH1], xmm2
	;# do forces H2 - j water 
	movaps xmm0, xmm7
	movaps xmm1, xmm7
	movaps xmm2, xmm7
	mulps   xmm0, [esp + nb104_dxH2M]
	mulps   xmm1, [esp + nb104_dyH2M]
	mulps   xmm2, [esp + nb104_dzH2M]
	movaps  xmm3, [esp + nb104_fjxM]
	movaps  xmm4, [esp + nb104_fjyM]
	movaps  xmm5, [esp + nb104_fjzM]
	subps   xmm3, xmm0
	subps   xmm4, xmm1
	subps   xmm5, xmm2
	mov     esi, [ebp + nb104_faction]
	movaps  [esp + nb104_fjxM], xmm3
	movaps  [esp + nb104_fjyM], xmm4
	movaps  [esp + nb104_fjzM], xmm5
	addps   xmm0, [esp + nb104_fixH2]
	addps   xmm1, [esp + nb104_fiyH2]
	addps   xmm2, [esp + nb104_fizH2]
	movaps  [esp + nb104_fixH2], xmm0
	movaps  [esp + nb104_fiyH2], xmm1
	movaps  [esp + nb104_fizH2], xmm2

	;# update j water forces from local variables 
	movlps  xmm0, [esi + eax*4 + 36]
	movlps  xmm1, [esi + eax*4 + 12]
	movhps  xmm1, [esi + eax*4 + 24]
	movaps  xmm3, [esp + nb104_fjxM]
	movaps  xmm4, [esp + nb104_fjyM]
	movaps  xmm5, [esp + nb104_fjzM]
	movaps  xmm6, xmm5
	movaps  xmm7, xmm5
	shufps  xmm6, xmm6, 2 ;# constant 00000010
	shufps  xmm7, xmm7, 3 ;# constant 00000011
	addss   xmm5, [esi + eax*4 + 44]
	addss   xmm6, [esi + eax*4 + 20]
	addss   xmm7, [esi + eax*4 + 32]
	movss   [esi + eax*4 + 44], xmm5
	movss   [esi + eax*4 + 20], xmm6
	movss   [esi + eax*4 + 32], xmm7
	movaps   xmm5, xmm3
	unpcklps xmm3, xmm4
	unpckhps xmm5, xmm4
	addps    xmm0, xmm3
	addps    xmm1, xmm5
	movlps  [esi + eax*4 + 36], xmm0 
	movlps  [esi + eax*4 + 12], xmm1 
	movhps  [esi + eax*4 + 24], xmm1 
	
	dec   dword ptr [esp + nb104_innerk]
	jz    .nb104_updateouterdata
	jmp   .nb104_single_loop
.nb104_updateouterdata:
	mov   ecx, [esp + nb104_ii3]
	mov   edi, [ebp + nb104_faction]
	mov   esi, [ebp + nb104_fshift]
	mov   edx, [esp + nb104_is3]

	;# accumulate H1i forces in xmm0, xmm1, xmm2 
	movaps xmm0, [esp + nb104_fixH1]
	movaps xmm1, [esp + nb104_fiyH1] 
	movaps xmm2, [esp + nb104_fizH1]

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addps  xmm0, xmm3
	addps  xmm1, xmm4
	addps  xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 

	movaps xmm3, xmm0