nb_kernel114_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,191 行 · 第 1/5 页

	mulps xmm1, [esp + nb114_dyH1M]
	mulps xmm2, [esp + nb114_dzH1M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixH1]
	addps xmm1, [esp + nb114_fiyH1]
	addps xmm2, [esp + nb114_fizH1]
	movaps [esp + nb114_fjxM], xmm3
	movaps [esp + nb114_fjyM], xmm4
	movaps [esp + nb114_fjzM], xmm5
	movaps [esp + nb114_fixH1], xmm0
	movaps [esp + nb114_fiyH1], xmm1
	movaps [esp + nb114_fizH1], xmm2

	;# H2-H1 interaction 
	movaps xmm0, [esp + nb114_rinvH2H1]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqHH]
	mulps xmm0, xmm1	;# fs H2-H1 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxH1]
	movaps xmm4, [esp + nb114_fjyH1]
	movaps xmm5, [esp + nb114_fjzH1]
	mulps xmm0, [esp + nb114_dxH2H1]
	mulps xmm1, [esp + nb114_dyH2H1]
	mulps xmm2, [esp + nb114_dzH2H1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixH2]
	addps xmm1, [esp + nb114_fiyH2]
	addps xmm2, [esp + nb114_fizH2]
	movaps [esp + nb114_fjxH1], xmm3
	movaps [esp + nb114_fjyH1], xmm4
	movaps [esp + nb114_fjzH1], xmm5
	movaps [esp + nb114_fixH2], xmm0
	movaps [esp + nb114_fiyH2], xmm1
	movaps [esp + nb114_fizH2], xmm2

	;# H2-H2 interaction 
	movaps xmm0, [esp + nb114_rinvH2H2]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqHH]
	mulps xmm0, xmm1	;# fsH2H2
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxH2]
	movaps xmm4, [esp + nb114_fjyH2]
	movaps xmm5, [esp + nb114_fjzH2]
	mulps xmm0, [esp + nb114_dxH2H2]
	mulps xmm1, [esp + nb114_dyH2H2]
	mulps xmm2, [esp + nb114_dzH2H2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixH2]
	addps xmm1, [esp + nb114_fiyH2]
	addps xmm2, [esp + nb114_fizH2]
	movaps [esp + nb114_fjxH2], xmm3
	movaps [esp + nb114_fjyH2], xmm4
	movaps [esp + nb114_fjzH2], xmm5
	movaps [esp + nb114_fixH2], xmm0
	movaps [esp + nb114_fiyH2], xmm1
	movaps [esp + nb114_fizH2], xmm2

	;# H2-M interaction 
	movaps xmm0, [esp + nb114_rinvH2M]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqMH]
	mulps xmm0, xmm1	;# fs H2-M  
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxM]
	movaps xmm4, [esp + nb114_fjyM]
	movaps xmm5, [esp + nb114_fjzM]
	mulps xmm0, [esp + nb114_dxH2M]
	mulps xmm1, [esp + nb114_dyH2M]
	mulps xmm2, [esp + nb114_dzH2M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixH2]
	addps xmm1, [esp + nb114_fiyH2]
	addps xmm2, [esp + nb114_fizH2]
	movaps [esp + nb114_fjxM], xmm3
	movaps [esp + nb114_fjyM], xmm4
	movaps [esp + nb114_fjzM], xmm5
	movaps [esp + nb114_fixH2], xmm0
	movaps [esp + nb114_fiyH2], xmm1
	movaps [esp + nb114_fizH2], xmm2

	;# M-H1 interaction 
	movaps xmm0, [esp + nb114_rinvMH1]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqMH]
	mulps xmm0, xmm1	;# fs M-H1 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxH1]
	movaps xmm4, [esp + nb114_fjyH1]
	movaps xmm5, [esp + nb114_fjzH1]
	mulps xmm0, [esp + nb114_dxMH1]
	mulps xmm1, [esp + nb114_dyMH1]
	mulps xmm2, [esp + nb114_dzMH1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixM]
	addps xmm1, [esp + nb114_fiyM]
	addps xmm2, [esp + nb114_fizM]
	movaps [esp + nb114_fjxH1], xmm3
	movaps [esp + nb114_fjyH1], xmm4
	movaps [esp + nb114_fjzH1], xmm5
	movaps [esp + nb114_fixM], xmm0
	movaps [esp + nb114_fiyM], xmm1
	movaps [esp + nb114_fizM], xmm2
	
	;# M-H2 interaction 
	movaps xmm0, [esp + nb114_rinvMH2]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqMH]
	mulps xmm0, xmm1	;# fs M-H2 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxH2]
	movaps xmm4, [esp + nb114_fjyH2]
	movaps xmm5, [esp + nb114_fjzH2]
	mulps xmm0, [esp + nb114_dxMH2]
	mulps xmm1, [esp + nb114_dyMH2]
	mulps xmm2, [esp + nb114_dzMH2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixM]
	addps xmm1, [esp + nb114_fiyM]
	addps xmm2, [esp + nb114_fizM]
	movaps [esp + nb114_fjxH2], xmm3
	movaps [esp + nb114_fjyH2], xmm4
	movaps [esp + nb114_fjzH2], xmm5
	movaps [esp + nb114_fixM], xmm0
	movaps [esp + nb114_fiyM], xmm1
	movaps [esp + nb114_fizM], xmm2

	;# M-M interaction 
	movaps xmm0, [esp + nb114_rinvMM]
	movaps xmm1, xmm0
	mulps xmm0, xmm0
	mulps xmm1, [esp + nb114_qqMM]
	mulps xmm0, xmm1	;# fs M-M 
	addps xmm7, xmm1	;# add to local vctot 
	movaps xmm1, xmm0
	movaps [esp + nb114_vctot], xmm7
	movaps xmm2, xmm0
	movaps xmm3, [esp + nb114_fjxM]
	movaps xmm4, [esp + nb114_fjyM]
	movaps xmm5, [esp + nb114_fjzM]
	mulps xmm0, [esp + nb114_dxMM]
	mulps xmm1, [esp + nb114_dyMM]
	mulps xmm2, [esp + nb114_dzMM]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb114_fixM]
	addps xmm1, [esp + nb114_fiyM]
	addps xmm2, [esp + nb114_fizM]
	movaps [esp + nb114_fjxM], xmm3
	movaps [esp + nb114_fjyM], xmm4
	movaps [esp + nb114_fjzM], xmm5
	movaps [esp + nb114_fixM], xmm0
	movaps [esp + nb114_fiyM], xmm1
	movaps [esp + nb114_fizM], xmm2

	;# Did all interactions - now update j forces 
	mov edi, [ebp + nb114_faction]
	;# 4 j waters with four atoms each.
	;# step 1 : transpose fjxO, fjyO, fjzO, fjxH1
	movaps xmm0, [esp + nb114_fjxO]
	movaps xmm1, [esp + nb114_fjyO]
	movaps xmm2, [esp + nb114_fjzO]
	movaps xmm3, [esp + nb114_fjxH1]
	movaps xmm4, xmm0
	movaps xmm5, xmm1
	unpcklps xmm4, xmm2 	
	unpcklps xmm5, xmm3	
	unpckhps xmm0, xmm2 
	unpckhps xmm1, xmm3 
	movaps xmm2, xmm4
	movaps xmm3, xmm0

	unpcklps xmm4, xmm5 
	unpckhps xmm2, xmm5 
	unpcklps xmm0, xmm1 
	unpckhps xmm3, xmm1
	
	;# results are now in xmm4, xmm2, xmm0, xmm3
	;# load the corresponding j forces from memory
	movlps   xmm1, [edi + eax*4]
	movlps   xmm5, [edi + ebx*4]
	movlps   xmm6, [edi + ecx*4]
	movlps   xmm7, [edi + edx*4]
	movhps   xmm1, [edi + eax*4 + 8]
	movhps   xmm5, [edi + ebx*4 + 8]
	movhps   xmm6, [edi + ecx*4 + 8]
	movhps   xmm7, [edi + edx*4 + 8]
	;# add
	addps    xmm1, xmm4
	addps    xmm5, xmm2
	addps    xmm6, xmm0
	addps    xmm7, xmm3
	;# store back
	movlps   [edi + eax*4], xmm1
	movlps   [edi + ebx*4], xmm5
	movlps   [edi + ecx*4], xmm6
	movlps   [edi + edx*4], xmm7
	movhps   [edi + eax*4 + 8], xmm1
	movhps   [edi + ebx*4 + 8], xmm5
	movhps   [edi + ecx*4 + 8], xmm6
	movhps   [edi + edx*4 + 8], xmm7

	;# step 2 : transpose fjyH1, fjzH1, fjxH2, fjyH2
	movaps xmm0, [esp + nb114_fjyH1]
	movaps xmm1, [esp + nb114_fjzH1]
	movaps xmm2, [esp + nb114_fjxH2]
	movaps xmm3, [esp + nb114_fjyH2]
	movaps xmm4, xmm0
	movaps xmm5, xmm1
	unpcklps xmm4, xmm2 	
	unpcklps xmm5, xmm3	
	unpckhps xmm0, xmm2 
	unpckhps xmm1, xmm3 
	movaps xmm2, xmm4
	movaps xmm3, xmm0
	unpcklps xmm4, xmm5 
	unpckhps xmm2, xmm5 

	unpcklps xmm0, xmm1 
	unpckhps xmm3, xmm1 
	;# results are now in xmm4, xmm2, xmm0, xmm3
	;# load the corresponding j forces from memory
	movlps   xmm1, [edi + eax*4 + 16]
	movlps   xmm5, [edi + ebx*4 + 16]
	movlps   xmm6, [edi + ecx*4 + 16]
	movlps   xmm7, [edi + edx*4 + 16]
	movhps   xmm1, [edi + eax*4 + 24]
	movhps   xmm5, [edi + ebx*4 + 24]
	movhps   xmm6, [edi + ecx*4 + 24]
	movhps   xmm7, [edi + edx*4 + 24]
	;# add
	addps    xmm1, xmm4
	addps    xmm5, xmm2
	addps    xmm6, xmm0
	addps    xmm7, xmm3
	;# store back
	movlps   [edi + eax*4 + 16], xmm1
	movlps   [edi + ebx*4 + 16], xmm5
	movlps   [edi + ecx*4 + 16], xmm6
	movlps   [edi + edx*4 + 16], xmm7
	movhps   [edi + eax*4 + 24], xmm1
	movhps   [edi + ebx*4 + 24], xmm5
	movhps   [edi + ecx*4 + 24], xmm6
	movhps   [edi + edx*4 + 24], xmm7

	;# step 3 : transpose fjzH2, fjxM, fjyM, fjzM. xmm4 is scratch
	movaps xmm0, [esp + nb114_fjzH2]
	movaps xmm1, [esp + nb114_fjxM]
	movaps xmm2, [esp + nb114_fjyM]
	movaps xmm3, [esp + nb114_fjzM]

        movaps xmm4, xmm0
        movaps xmm5, xmm1
        unpcklps xmm4, xmm2
        unpcklps xmm5, xmm3
        unpckhps xmm0, xmm2
        unpckhps xmm1, xmm3
        movaps xmm2, xmm4
        movaps xmm3, xmm0
        unpcklps xmm4, xmm5
        unpckhps xmm2, xmm5
	
        unpcklps xmm0, xmm1
        unpckhps xmm3, xmm1

	;# results are now in xmm0, xmm1, xmm2, xmm3
	;# load the corresponding j forces from memory
	movlps   xmm1, [edi + eax*4 + 32]
	movlps   xmm5, [edi + ebx*4 + 32]
	movlps   xmm6, [edi + ecx*4 + 32]
	movlps   xmm7, [edi + edx*4 + 32]
	movhps   xmm1, [edi + eax*4 + 40]
	movhps   xmm5, [edi + ebx*4 + 40]
	movhps   xmm6, [edi + ecx*4 + 40]
	movhps   xmm7, [edi + edx*4 + 40]
	;# add
	addps    xmm1, xmm4
	addps    xmm5, xmm2
	addps    xmm6, xmm0
	addps    xmm7, xmm3
	;# store back
	movlps   [edi + eax*4 + 32], xmm1
	movlps   [edi + ebx*4 + 32], xmm5
	movlps   [edi + ecx*4 + 32], xmm6
	movlps   [edi + edx*4 + 32], xmm7
	movhps   [edi + eax*4 + 40], xmm1
	movhps   [edi + ebx*4 + 40], xmm5
	movhps   [edi + ecx*4 + 40], xmm6
	movhps   [edi + edx*4 + 40], xmm7

	;# should we do one more iteration? 
	sub dword ptr [esp + nb114_innerk],  4
	jl    .nb114_single_check
	jmp   .nb114_unroll_loop
.nb114_single_check:
	add dword ptr [esp + nb114_innerk],  4
	jnz   .nb114_single_loop
	jmp   .nb114_updateouterdata
.nb114_single_loop:
	mov   edx, [esp + nb114_innerjjnr] 	;# pointer to jjnr[k] 
	mov   eax, [edx]	
	add dword ptr [esp + nb114_innerjjnr],  4	

	mov esi, [ebp + nb114_pos]
	lea   eax, [eax + eax*2]  

	;# fetch j coordinates	
	movlps xmm3,  [esi + eax*4]		;#  Ox  Oy  
	movlps xmm4,  [esi + eax*4 + 16]	;# H1y H1z 
	movlps xmm5,  [esi + eax*4 + 32]	;# H2z  Mx 
	movhps xmm3,  [esi + eax*4 + 8]   	;#  Ox  Oy  Oz H1x
	movhps xmm4,  [esi + eax*4 + 24]	;# H1y H1z H2x H2y
	movhps xmm5,  [esi + eax*4 + 40]	;# H2z  Mx  My  Mz
	;# transpose
	movaps xmm0, xmm4
	movaps xmm1, xmm3
	movaps xmm2, xmm4
	movaps xmm6, xmm3
	shufps xmm4, xmm5, 18  ;# (00010010)  h2x - Mx  - 
	shufps xmm3, xmm0, 193 ;# (11000001)  Oy  - H1y - 
	shufps xmm2, xmm5, 35  ;# (00100011) H2y - My  - 
 	shufps xmm1, xmm0, 18  ;# (00010010)  Oz  - H1z - 
	;#  xmm6: Ox - - H1x   xmm5: H2z - - Mz 
	shufps xmm6, xmm4, 140 ;# (10001100) Ox H1x H2x Mx 
	shufps xmm3, xmm2, 136 ;# (10001000) Oy H1y H2y My 
	shufps xmm1, xmm5, 200 ;# (11001000) Oz H1z H2z Mz

	;# store all j coordinates in jO  
	movaps [esp + nb114_jxO], xmm6
	movaps [esp + nb114_jyO], xmm3
	movaps [esp + nb114_jzO], xmm1

	;# do O and M in parallel
	movaps xmm0, [esp + nb114_ixO]
	movaps xmm1, [esp + nb114_iyO]
	movaps xmm2, [esp + nb114_izO]
	movaps xmm3, [esp + nb114_ixM]
	movaps xmm4, [esp + nb114_iyM]
	movaps xmm5, [esp + nb114_izM]
	subps  xmm0, [esp + nb114_jxO]
	subps  xmm1, [esp + nb114_jyO]
	subps  xmm2, [esp + nb114_jzO]
	subps  xmm3, [esp + nb114_jxO]
	subps  xmm4, [esp + nb114_jyO]
	subps  xmm5, [esp + nb114_jzO]
	
	movaps [esp + nb114_dxOO], xmm0
	movaps [esp + nb114_dyOO], xmm1
	movaps [esp + nb114_dzOO], xmm2
	movaps [esp + nb114_dxMM], xmm3
	movaps [esp + nb114_dyMM], xmm4
	movaps [esp + nb114_dzMM], xmm5
	
	mulps xmm0, xmm0
	mulps xmm1, xmm1
	mulps xmm2, xmm2
	addps xmm0, xmm1
	addps xmm0, xmm2	;# have rsq in xmm0 
	mulps xmm3, xmm3
	mulps xmm4, xmm4
	mulps xmm5, xmm5
	addps xmm4, xmm3
	addps xmm4, xmm5	;# have rsq in xmm4
	;# Save M data 
	movaps [esp + nb114_rsqMM], xmm4
	
	;# do 1/x for O and 1/sqrt(x) for M
	rcpss  xmm1, xmm0
	rsqrtps xmm5, xmm4
	movss  xmm2, [esp + nb114_two]
	movaps  xmm6, xmm5	
	mulss  xmm0, xmm1
	mulps   xmm5, xmm5
	subss  xmm2, xmm0
	movaps  xmm7, [esp + nb114_three]
	mulss  xmm2, xmm1 	;# constant 1/r2
	
	mulps   xmm5, xmm4
	movss  xmm0, xmm2
	subps   xmm7, xmm5
	mulss  xmm2, xmm2
	mulps   xmm7, xmm6
	mulss  xmm2, xmm0 	;# constant 1/r6
	mulps   xmm7, [esp + nb114_half] ;# rinv iH1 - j water 
	movss  xmm1, xmm2
	movaps [esp + nb114_rinvMM], xmm7

	mulss  xmm2, xmm2 	;# constant 1/r12
	mulss  xmm1, [esp + nb114_c6]
	mulss  xmm2, [esp + nb114_c12]
	movss  xmm3, xmm2
	subss  xmm3, xmm1
	addss  xmm3, [esp + nb114_Vvdwtot]
	movss  [esp + nb114_Vvdwtot], xmm3
	mulss  xmm1, [esp + nb114_six]
	mulss  xmm2, [esp + nb114_twelve]
	subss  xmm2, xmm1
	mulss  xmm0, xmm2 	;# fscal
	movss  xmm1, xmm0
	movss  xmm2, xmm0
	mulss  xmm0, [esp + nb114_dxOO]
	mulss  xmm1, [esp + nb114_dyOO]
	mulss  xmm2, [esp + nb114_dzOO]
	xorps   xmm3, xmm3
	xorps   xmm4, xmm4
	xorps   xmm5, xmm5
	subss   xmm3, xmm0
	subss   xmm4, xmm1
	subss   xmm5, xmm2
	movaps  [esp + nb114_fjxO], xmm3