nb_kernel134_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,207 行 · 第 1/5 页

	movaps xmm5, xmm3
	mulps xmm0, [esp + nb134_dxOO]
	mulps xmm1, [esp + nb134_dyOO]
	mulps xmm2, [esp + nb134_dzOO]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixO]
	addps xmm1, [esp + nb134_fiyO]
	addps xmm2, [esp + nb134_fizO]
	movaps [esp + nb134_fjxO], xmm3
	movaps [esp + nb134_fjyO], xmm4
	movaps [esp + nb134_fjzO], xmm5
	movaps [esp + nb134_fixO], xmm0
	movaps [esp + nb134_fiyO], xmm1
	movaps [esp + nb134_fizO], xmm2

	;# Coulomb interactions 
	;# start with H1-H1 interaction 
	movaps  xmm7, [esp + nb134_rinvH1H1]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqHH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	movaps xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0

	xorps xmm3, xmm3
	movaps xmm4, xmm3
	movaps xmm5, xmm3
	mulps xmm0, [esp + nb134_dxH1H1]
	mulps xmm1, [esp + nb134_dyH1H1]
	mulps xmm2, [esp + nb134_dzH1H1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH1]
	addps xmm1, [esp + nb134_fiyH1]
	addps xmm2, [esp + nb134_fizH1]
	movaps [esp + nb134_fjxH1], xmm3
	movaps [esp + nb134_fjyH1], xmm4
	movaps [esp + nb134_fjzH1], xmm5
	movaps [esp + nb134_fixH1], xmm0
	movaps [esp + nb134_fiyH1], xmm1
	movaps [esp + nb134_fizH1], xmm2

	;# H1-H2 interaction 
	movaps  xmm7, [esp + nb134_rinvH1H2]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqHH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	xorps xmm3, xmm3
	movaps xmm4, xmm3
	movaps xmm5, xmm3
	mulps xmm0, [esp + nb134_dxH1H2]
	mulps xmm1, [esp + nb134_dyH1H2]
	mulps xmm2, [esp + nb134_dzH1H2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH1]
	addps xmm1, [esp + nb134_fiyH1]
	addps xmm2, [esp + nb134_fizH1]
	movaps [esp + nb134_fjxH2], xmm3
	movaps [esp + nb134_fjyH2], xmm4
	movaps [esp + nb134_fjzH2], xmm5
	movaps [esp + nb134_fixH1], xmm0
	movaps [esp + nb134_fiyH1], xmm1
	movaps [esp + nb134_fizH1], xmm2

	;# H1-M interaction  
	movaps  xmm7, [esp + nb134_rinvH1M]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqMH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	xorps xmm3, xmm3
	movaps xmm4, xmm3
	movaps xmm5, xmm3
	mulps xmm0, [esp + nb134_dxH1M]
	mulps xmm1, [esp + nb134_dyH1M]
	mulps xmm2, [esp + nb134_dzH1M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH1]
	addps xmm1, [esp + nb134_fiyH1]
	addps xmm2, [esp + nb134_fizH1]
	movaps [esp + nb134_fjxM], xmm3
	movaps [esp + nb134_fjyM], xmm4
	movaps [esp + nb134_fjzM], xmm5
	movaps [esp + nb134_fixH1], xmm0
	movaps [esp + nb134_fiyH1], xmm1
	movaps [esp + nb134_fizH1], xmm2

	;# H2-H1 interaction 
	movaps  xmm7, [esp + nb134_rinvH2H1]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqHH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxH1]
	movaps xmm4, [esp + nb134_fjyH1]
	movaps xmm5, [esp + nb134_fjzH1]
	mulps xmm0, [esp + nb134_dxH2H1]
	mulps xmm1, [esp + nb134_dyH2H1]
	mulps xmm2, [esp + nb134_dzH2H1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH2]
	addps xmm1, [esp + nb134_fiyH2]
	addps xmm2, [esp + nb134_fizH2]
	movaps [esp + nb134_fjxH1], xmm3
	movaps [esp + nb134_fjyH1], xmm4
	movaps [esp + nb134_fjzH1], xmm5
	movaps [esp + nb134_fixH2], xmm0
	movaps [esp + nb134_fiyH2], xmm1
	movaps [esp + nb134_fizH2], xmm2

	;# H2-H2 interaction 
	movaps  xmm7, [esp + nb134_rinvH2H2]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqHH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxH2]
	movaps xmm4, [esp + nb134_fjyH2]
	movaps xmm5, [esp + nb134_fjzH2]
	mulps xmm0, [esp + nb134_dxH2H2]
	mulps xmm1, [esp + nb134_dyH2H2]
	mulps xmm2, [esp + nb134_dzH2H2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH2]
	addps xmm1, [esp + nb134_fiyH2]
	addps xmm2, [esp + nb134_fizH2]
	movaps [esp + nb134_fjxH2], xmm3
	movaps [esp + nb134_fjyH2], xmm4
	movaps [esp + nb134_fjzH2], xmm5
	movaps [esp + nb134_fixH2], xmm0
	movaps [esp + nb134_fiyH2], xmm1
	movaps [esp + nb134_fizH2], xmm2

	;# H2-M interaction 
	movaps  xmm7, [esp + nb134_rinvH2M]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqMH]
	mulps  xmm0, xmm7	;# total fs H1 in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxM]
	movaps xmm4, [esp + nb134_fjyM]
	movaps xmm5, [esp + nb134_fjzM]
	mulps xmm0, [esp + nb134_dxH2M]
	mulps xmm1, [esp + nb134_dyH2M]
	mulps xmm2, [esp + nb134_dzH2M]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixH2]
	addps xmm1, [esp + nb134_fiyH2]
	addps xmm2, [esp + nb134_fizH2]
	movaps [esp + nb134_fjxM], xmm3
	movaps [esp + nb134_fjyM], xmm4
	movaps [esp + nb134_fjzM], xmm5
	movaps [esp + nb134_fixH2], xmm0
	movaps [esp + nb134_fiyH2], xmm1
	movaps [esp + nb134_fizH2], xmm2

	;# M-H1 interaction 
	movaps  xmm7, [esp + nb134_rinvMH1]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqMH]
	mulps  xmm0, xmm7	;# total fs in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxH1]
	movaps xmm4, [esp + nb134_fjyH1]
	movaps xmm5, [esp + nb134_fjzH1]
	mulps xmm0, [esp + nb134_dxMH1]
	mulps xmm1, [esp + nb134_dyMH1]
	mulps xmm2, [esp + nb134_dzMH1]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixM]
	addps xmm1, [esp + nb134_fiyM]
	addps xmm2, [esp + nb134_fizM]
	movaps [esp + nb134_fjxH1], xmm3
	movaps [esp + nb134_fjyH1], xmm4
	movaps [esp + nb134_fjzH1], xmm5
	movaps [esp + nb134_fixM], xmm0
	movaps [esp + nb134_fiyM], xmm1
	movaps [esp + nb134_fizM], xmm2

	;# M-H2 interaction 
	movaps  xmm7, [esp + nb134_rinvMH2]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqMH]
	mulps  xmm0, xmm7	;# total fs in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxH2]
	movaps xmm4, [esp + nb134_fjyH2]
	movaps xmm5, [esp + nb134_fjzH2]
	mulps xmm0, [esp + nb134_dxMH2]
	mulps xmm1, [esp + nb134_dyMH2]
	mulps xmm2, [esp + nb134_dzMH2]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixM]
	addps xmm1, [esp + nb134_fiyM]
	addps xmm2, [esp + nb134_fizM]
	movaps [esp + nb134_fjxH2], xmm3
	movaps [esp + nb134_fjyH2], xmm4
	movaps [esp + nb134_fjzH2], xmm5
	movaps [esp + nb134_fixM], xmm0
	movaps [esp + nb134_fiyM], xmm1
	movaps [esp + nb134_fizM], xmm2

	;# M-M interaction 
	movaps  xmm7, [esp + nb134_rinvMM]
	movaps  xmm0, xmm7	
	mulps   xmm0, xmm0	;# xmm7=rinv, xmm0=rinvsq
	mulps  xmm7, [esp + nb134_qqMM]
	mulps  xmm0, xmm7	;# total fs in xmm0 

	addps  xmm6, xmm7  ;# local vctot summation variable
	
	movaps xmm1, xmm0
	addps  xmm6, [esp + nb134_vctot]
	movaps [esp + nb134_vctot], xmm6
	movaps xmm2, xmm0
	
	movaps xmm3, [esp + nb134_fjxM]
	movaps xmm4, [esp + nb134_fjyM]
	movaps xmm5, [esp + nb134_fjzM]
	mulps xmm0, [esp + nb134_dxMM]
	mulps xmm1, [esp + nb134_dyMM]
	mulps xmm2, [esp + nb134_dzMM]
	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2
	addps xmm0, [esp + nb134_fixM]
	addps xmm1, [esp + nb134_fiyM]
	addps xmm2, [esp + nb134_fizM]
	movaps [esp + nb134_fjxM], xmm3
	movaps [esp + nb134_fjyM], xmm4
	movaps [esp + nb134_fjzM], xmm5
	movaps [esp + nb134_fixM], xmm0
	movaps [esp + nb134_fiyM], xmm1
	movaps [esp + nb134_fizM], xmm2

	mov edi, [ebp + nb134_faction]
	;# update j forces 
	;# 4 j waters with four atoms each.
	;# step 1 : transpose fjxO, fjyO, fjzO, fjxH1
	movaps xmm0, [esp + nb134_fjxO]
	movaps xmm1, [esp + nb134_fjyO]
	movaps xmm2, [esp + nb134_fjzO]
	movaps xmm3, [esp + nb134_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 + nb134_fjyH1]
	movaps xmm1, [esp + nb134_fjzH1]
	movaps xmm2, [esp + nb134_fjxH2]
	movaps xmm3, [esp + nb134_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 + nb134_fjzH2]
	movaps xmm1, [esp + nb134_fjxM]
	movaps xmm2, [esp + nb134_fjyM]
	movaps xmm3, [esp + nb134_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 + nb134_innerk],  4
	jl    .nb134_single_check
	jmp   .nb134_unroll_loop
.nb134_single_check:
	add dword ptr [esp + nb134_innerk],  4
	jnz   .nb134_single_loop
	jmp   .nb134_updateouterdata
.nb134_single_loop: