nb_kernel331_x86_64_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,353 行 · 第 1/5 页

	mulpd   xmm10, xmm5
    mulpd   xmm11, xmm8 ;# lu*(3-rsq*lu*lu)

	movapd  xmm15, [rsp + nb331_half]
	mulpd   xmm9, xmm15  ;# first iteration for rinvO
	mulpd   xmm10, xmm15 ;# first iteration for rinvH1
    mulpd   xmm11, xmm15 ;# first iteration for rinvH2	

    ;# second iteration step    
	movapd  xmm2, xmm9
	movapd  xmm5, xmm10
    movapd  xmm8, xmm11
    
	mulpd   xmm2, xmm2 ;# lu*lu
	mulpd   xmm5, xmm5 ;# lu*lu
    mulpd   xmm8, xmm8 ;# lu*lu
		
	movapd  xmm1, [rsp + nb331_three]
	movapd  xmm4, xmm1
    movapd  xmm7, xmm1

	mulpd   xmm2, xmm0 ;# rsq*lu*lu
	mulpd   xmm5, xmm3 ;# rsq*lu*lu 
    mulpd   xmm8, xmm6 ;# rsq*lu*lu
	
	subpd   xmm1, xmm2
	subpd   xmm4, xmm5
    subpd   xmm7, xmm8 ;# 3-rsq*lu*lu

	mulpd   xmm9, xmm1
	mulpd   xmm10, xmm4
    mulpd   xmm11, xmm7 ;# lu*(3-rsq*lu*lu)

	movapd  xmm15, [rsp + nb331_half]
	mulpd   xmm9, xmm15  ;#  rinvO
	mulpd   xmm10, xmm15 ;#   rinvH1
    mulpd   xmm11, xmm15 ;#   rinvH2
	
	movapd  [rsp + nb331_rinvO], xmm9
	movapd  [rsp + nb331_rinvH1], xmm10
	movapd  [rsp + nb331_rinvH2], xmm11
	
	;# interactions 
    ;# rsq in xmm0,xmm3,xmm6  
    ;# rinv in xmm9, xmm10, xmm11
    movapd [rsp + nb331_rinvO], xmm9
    movapd xmm1, [rsp + nb331_tsc]
    
    mulpd  xmm0, xmm9  ;# r
    mulpd  xmm3, xmm10
    mulpd  xmm6, xmm11
    mulpd  xmm0, xmm1 ;# rtab
    mulpd  xmm3, xmm1
    mulpd  xmm6, xmm1
    
    ;# truncate and convert to integers
    cvttpd2dq xmm1, xmm0
    cvttpd2dq xmm4, xmm3
    cvttpd2dq xmm7, xmm6        

    ;# convert back to float
    cvtdq2pd  xmm2, xmm1
    cvtdq2pd  xmm5, xmm4
    cvtdq2pd  xmm8, xmm7
    
    ;# multiply by 4
    pslld   xmm1, 2
    pslld   xmm4, 2
    pslld   xmm7, 2
        
    ;# multiply by three (copy, mult. by two, add back)
    movapd  xmm10, xmm1
    movapd  xmm11, xmm4
    movapd  xmm12, xmm7
    pslld   xmm1, 1
    pslld   xmm4, 1
    pslld   xmm7, 1    
    paddd   xmm1, xmm10
    paddd   xmm4, xmm11
    paddd   xmm7, xmm12    
    
    ;# move to integer registers
    pshufd xmm13, xmm1, 1
    pshufd xmm14, xmm4, 1
    pshufd xmm15, xmm7, 1
    movd    r8d, xmm1
    movd    r10d, xmm4
    movd    r12d, xmm7
    movd    r9d, xmm13
    movd    r11d, xmm14
    movd    r13d, xmm15
        
    mov  rsi, [rbp + nb331_VFtab]

    ;# calculate eps
    subpd     xmm0, xmm2
    subpd     xmm3, xmm5
    subpd     xmm6, xmm8

    movapd    xmm12, xmm0
    movapd    xmm13, xmm3
    movapd    xmm14, xmm6

    ;# Load LOTS of table data
    movlpd xmm0,  [rsi + r8*8]
    movlpd xmm1,  [rsi + r8*8 + 8]
    movlpd xmm2,  [rsi + r8*8 + 16]
    movlpd xmm3,  [rsi + r8*8 + 24]
    movlpd xmm4,  [rsi + r10*8]
    movlpd xmm5,  [rsi + r10*8 + 8]
    movlpd xmm6,  [rsi + r10*8 + 16]
    movlpd xmm7,  [rsi + r10*8 + 24]
    movlpd xmm8,  [rsi + r12*8]
    movlpd xmm9,  [rsi + r12*8 + 8]
    movlpd xmm10, [rsi + r12*8 + 16]
    movlpd xmm11, [rsi + r12*8 + 24]
    movhpd xmm0,  [rsi + r9*8]
    movhpd xmm1,  [rsi + r9*8 + 8]
    movhpd xmm2,  [rsi + r9*8 + 16]
    movhpd xmm3,  [rsi + r9*8 + 24]
    movhpd xmm4,  [rsi + r11*8]
    movhpd xmm5,  [rsi + r11*8 + 8]
    movhpd xmm6,  [rsi + r11*8 + 16]
    movhpd xmm7,  [rsi + r11*8 + 24]
    movhpd xmm8,  [rsi + r13*8]
    movhpd xmm9,  [rsi + r13*8 + 8]
    movhpd xmm10, [rsi + r13*8 + 16]
    movhpd xmm11, [rsi + r13*8 + 24]
    ;# table data ready in xmm0-xmm3 , xmm4-xmm7 , and xmm8-xmm11
    
    mulpd  xmm3, xmm12   ;# Heps
    mulpd  xmm7, xmm13
    mulpd  xmm11, xmm14
    mulpd  xmm2, xmm12  ;# Geps
    mulpd  xmm6, xmm13
    mulpd  xmm10, xmm14
    mulpd  xmm3, xmm12   ;# Heps2
    mulpd  xmm7, xmm13
    mulpd  xmm11, xmm14

    addpd  xmm1, xmm2   ;# F+Geps
    addpd  xmm5, xmm6
    addpd  xmm9, xmm10 
    addpd  xmm1, xmm3   ;# F+Geps+Heps2 = Fp
    addpd  xmm5, xmm7
    addpd  xmm9, xmm11 
    addpd  xmm3, xmm3    ;# 2*Heps2
    addpd  xmm7, xmm7
    addpd  xmm11, xmm11
    addpd  xmm3, xmm2    ;# 2*Heps2+Geps
    addpd  xmm7, xmm6  
    addpd  xmm11, xmm10
    addpd  xmm3, xmm1   ;# FF = Fp + 2*Heps2 + Geps
    addpd  xmm7, xmm5
    addpd  xmm11, xmm9
    mulpd  xmm1, xmm12   ;# eps*Fp
    mulpd  xmm5, xmm13
    mulpd  xmm9, xmm14
    addpd  xmm1, xmm0     ;# VV
    addpd  xmm5, xmm4
    addpd  xmm9, xmm8
    mulpd  xmm1, [rsp + nb331_qqO]   ;# VV*qq = vcoul
    mulpd  xmm5, [rsp + nb331_qqH]
    mulpd  xmm9, [rsp + nb331_qqH]
    mulpd  xmm3, [rsp + nb331_qqO]    ;# FF*qq = fij
    mulpd  xmm7, [rsp + nb331_qqH]
    mulpd  xmm11, [rsp + nb331_qqH] 

    ;# accumulate vctot
    addpd  xmm1, [rsp + nb331_vctot]
    addpd  xmm5, xmm9
    addpd  xmm1, xmm5
    movapd [rsp + nb331_vctot], xmm1

    movapd xmm2, xmm7
    movapd xmm1, xmm11

    ;# fij coul in xmm3, xmm2, xmm1    
    
    ;# calculate LJ table
    movlpd xmm4,  [rsi + r8*8 + 32]
    movlpd xmm5,  [rsi + r8*8 + 40]
    movlpd xmm6,  [rsi + r8*8 + 48]
    movlpd xmm7,  [rsi + r8*8 + 56]
    movlpd xmm8,  [rsi + r8*8 + 64]
    movlpd xmm9,  [rsi + r8*8 + 72]
    movlpd xmm10, [rsi + r8*8 + 80]
    movlpd xmm11, [rsi + r8*8 + 88]
    movhpd xmm4,  [rsi + r9*8 + 32]
    movhpd xmm5,  [rsi + r9*8 + 40]
    movhpd xmm6,  [rsi + r9*8 + 48]
    movhpd xmm7,  [rsi + r9*8 + 56]
    movhpd xmm8,  [rsi + r9*8 + 64]
    movhpd xmm9,  [rsi + r9*8 + 72]
    movhpd xmm10, [rsi + r9*8 + 80]
    movhpd xmm11, [rsi + r9*8 + 88]
    ;# dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11
    
    ;# xmm12 = epsO
    
    mulpd  xmm7, xmm12    ;# Heps
    mulpd  xmm11, xmm12 
    mulpd  xmm6, xmm12   ;# Geps
    mulpd  xmm10, xmm12 
    mulpd  xmm7, xmm12   ;# Heps2
    mulpd  xmm11, xmm12 
    addpd  xmm5, xmm6  ;# F+Geps
    addpd  xmm9, xmm10 
    addpd  xmm5, xmm7   ;# F+Geps+Heps2 = Fp
    addpd  xmm9, xmm11 
    addpd  xmm7, xmm7    ;# 2*Heps2
    addpd  xmm11, xmm11
    addpd  xmm7, xmm6   ;# 2*Heps2+Geps
    addpd  xmm11, xmm10
    
    addpd  xmm7, xmm5  ;# FF = Fp + 2*Heps2 + Geps
    addpd  xmm11, xmm9
    mulpd  xmm5, xmm12  ;# eps*Fp
    mulpd  xmm9, xmm12
    movapd xmm12, [rsp + nb331_c6]
    movapd xmm13, [rsp + nb331_c12]
    addpd  xmm5, xmm4 ;# VV
    addpd  xmm9, xmm8

    mulpd  xmm5, xmm12  ;# VV*c6 = vnb6
    mulpd  xmm9, xmm13  ;# VV*c12 = vnb12
    addpd  xmm5, xmm9
    addpd  xmm5, [rsp + nb331_Vvdwtot]
    movapd [rsp + nb331_Vvdwtot], xmm5
        
    mulpd  xmm7, xmm12   ;# FF*c6 = fnb6
    mulpd  xmm11, xmm13   ;# FF*c12  = fnb12
    addpd  xmm7, xmm11
    
    addpd  xmm3, xmm7
    movapd xmm10, [rsp + nb331_tsc]

    mulpd  xmm3, xmm10  ;# fscal
    mulpd  xmm2, xmm10
    mulpd  xmm1, xmm10
        
    ;# move j forces to xmm11-xmm13
    mov rdi, [rbp + nb331_faction]
	movlpd xmm11, [rdi + rax*8]
	movlpd xmm12, [rdi + rax*8 + 8]
	movlpd xmm13, [rdi + rax*8 + 16]
	movhpd xmm11, [rdi + rbx*8]
	movhpd xmm12, [rdi + rbx*8 + 8]
	movhpd xmm13, [rdi + rbx*8 + 16]

    xorpd  xmm0, xmm0
    xorpd  xmm4, xmm4
    xorpd  xmm8, xmm8
    
    subpd  xmm0, xmm3
    subpd  xmm4, xmm2
    subpd  xmm8, xmm1

    mulpd  xmm0, [rsp + nb331_rinvO]
    mulpd  xmm4, [rsp + nb331_rinvH1]
    mulpd  xmm8, [rsp + nb331_rinvH2]
    
    movapd xmm1, xmm0
    movapd xmm2, xmm0
    movapd xmm3, xmm4
    movapd xmm5, xmm4
    movapd xmm6, xmm8
    movapd xmm7, xmm8

	mulpd xmm0, [rsp + nb331_dxO]
	mulpd xmm1, [rsp + nb331_dyO]
	mulpd xmm2, [rsp + nb331_dzO]
	mulpd xmm3, [rsp + nb331_dxH1]
	mulpd xmm4, [rsp + nb331_dyH1]
	mulpd xmm5, [rsp + nb331_dzH1]
	mulpd xmm6, [rsp + nb331_dxH2]
	mulpd xmm7, [rsp + nb331_dyH2]
	mulpd xmm8, [rsp + nb331_dzH2]

    addpd xmm11,  xmm0
    addpd xmm12, xmm1
    addpd xmm13, xmm2
    addpd xmm0, [rsp + nb331_fixO]
    addpd xmm1, [rsp + nb331_fiyO]
    addpd xmm2, [rsp + nb331_fizO]

    addpd xmm11,  xmm3
    addpd xmm12, xmm4
    addpd xmm13, xmm5
    addpd xmm3, [rsp + nb331_fixH1]
    addpd xmm4, [rsp + nb331_fiyH1]
    addpd xmm5, [rsp + nb331_fizH1]

    addpd xmm11,  xmm6
    addpd xmm12, xmm7
    addpd xmm13, xmm8
    addpd xmm6, [rsp + nb331_fixH2]
    addpd xmm7, [rsp + nb331_fiyH2]
    addpd xmm8, [rsp + nb331_fizH2]

    movapd [rsp + nb331_fixO], xmm0
    movapd [rsp + nb331_fiyO], xmm1
    movapd [rsp + nb331_fizO], xmm2
    movapd [rsp + nb331_fixH1], xmm3
    movapd [rsp + nb331_fiyH1], xmm4
    movapd [rsp + nb331_fizH1], xmm5
    movapd [rsp + nb331_fixH2], xmm6
    movapd [rsp + nb331_fiyH2], xmm7
    movapd [rsp + nb331_fizH2], xmm8
       
    ;# store back j forces from xmm11-xmm13
	movlpd [rdi + rax*8],      xmm11
	movlpd [rdi + rax*8 + 8],  xmm12
	movlpd [rdi + rax*8 + 16], xmm13
	movhpd [rdi + rbx*8],      xmm11
	movhpd [rdi + rbx*8 + 8],  xmm12
	movhpd [rdi + rbx*8 + 16], xmm13

	;# should we do one more iteration? 
	sub dword ptr [rsp + nb331_innerk],  2
	jl    .nb331_checksingle
	jmp   .nb331_unroll_loop
.nb331_checksingle:	
	mov   edx, [rsp + nb331_innerk]
	and   edx, 1
	jnz   .nb331_dosingle
	jmp   .nb331_updateouterdata
.nb331_dosingle:
	mov   rdx, [rsp + nb331_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	
	mov rsi, [rbp + nb331_charge]    ;# base of charge[] 
	
	movsd xmm3, [rsi + rax*8]
	movapd xmm4, xmm3
    mulsd xmm3, [rsp + nb331_iqO]
    mulsd xmm4, [rsp + nb331_iqH]    
	movapd  [rsp + nb331_qqO], xmm3
	movapd  [rsp + nb331_qqH], xmm4	

	mov rsi, [rbp + nb331_type]
	mov r8d, [rsi + rax*4]
	mov rsi, [rbp + nb331_vdwparam]
	shl r8d, 1	
	mov edi, [rsp + nb331_ntia]
	add r8d, edi

	movsd xmm6, [rsi + r8*8]	     ;# c6a
	movsd xmm7, [rsi + r8*8 + 8]	 ;# c12a
	
	movapd [rsp + nb331_c6], xmm6
	movapd [rsp + nb331_c12], xmm7

	mov rsi, [rbp + nb331_pos]       ;# base of pos[] 

	lea   rax, [rax + rax*2]     ;# replace jnr with j3 

	;# move j coordinates to local temp variables 
    movsd xmm0, [rsi + rax*8] 
    movsd xmm1, [rsi + rax*8 + 8] 
    movsd xmm2, [rsi + rax*8 + 16] 

    ;# xmm0 = jx
    ;# xmm1 = jy
    ;# xmm2 = jz
        
    movapd xmm3, xmm0
    movapd xmm4, xmm1
    movapd xmm5, xmm2
    movapd xmm6, xmm0
    movapd xmm7, xmm1
    movapd xmm8, xmm2
    
    subsd xmm0, [rsp + nb331_ixO]
    subsd xmm1, [rsp + nb331_iyO]
    subsd xmm2, [rsp + nb331_izO]
    subsd xmm3, [rsp + nb331_ixH1]
    subsd xmm4, [rsp + nb331_iyH1]
    subsd xmm5, [rsp + nb331_izH1]
    subsd xmm6, [rsp + nb331_ixH2]
    subsd xmm7, [rsp + nb331_iyH2]
    subsd xmm8, [rsp + nb331_izH2]
    
	movapd [rsp + nb331_dxO], xmm0
	movapd [rsp + nb331_dyO], xmm1
	movapd [rsp + nb331_dzO], xmm2
	mulsd  xmm0, xmm0
	mulsd  xmm1, xmm1
	mulsd  xmm2, xmm2
	movapd [rsp + nb331_dxH1], xmm3
	movapd [rsp + nb331_dyH1], xmm4
	movapd [rsp + nb331_dzH1], xmm5
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm4
	mulsd  xmm5, xmm5
	movapd [rsp + nb331_dxH2], xmm6
	movapd [rsp + nb331_dyH2], xmm7
	movapd [rsp + nb331_dzH2], xmm8
	mulsd  xmm6, xmm6
	mulsd  xmm7, xmm7
	mulsd  xmm8, xmm8
	addsd  xmm0, xmm1
	addsd  xmm0, xmm2
	addsd  xmm3, xmm4
	addsd  xmm3, xmm5
    addsd  xmm6, xmm7
    addsd  xmm6, xmm8

	;# start doing invsqrt for j atoms
    cvtsd2ss xmm1, xmm0
    cvtsd2ss xmm4, xmm3
    cvtsd2ss xmm7, xmm6
	rsqrtss xmm1, xmm1
	rsqrtss xmm4, xmm4
    rsqrtss xmm7, xmm7
    cvtss2sd xmm1, xmm1
    cvtss2sd xmm4, xmm4
    cvtss2sd xmm7, xmm7
	
	movapd  xmm2, xmm1
	movapd  xmm5, xmm4
    movapd  xmm8, xmm7
    
	mulsd   xmm1, xmm1 ;# lu*lu
	mulsd   xmm4, xmm4 ;# lu*lu
    mulsd   xmm7, xmm7 ;# lu*lu
		
	movapd  xmm9, [rsp + nb331_three]
	movapd  xmm10, xmm9
    movapd  xmm11, xmm9

	mulsd   xmm1, xmm0 ;# rsq*lu*lu
	mulsd   xmm4, xmm3 ;# rsq*lu*lu 
    mulsd   xmm7, xmm6 ;# rsq*lu*lu
	
	subsd   xmm9, xmm1
	subsd   xmm10, xmm4
    subsd   xmm11, xmm7 ;# 3-rsq*lu*lu

	mulsd   xmm9, xmm2
	mulsd   xmm10, xmm5
    mulsd   xmm11, xmm8 ;# lu*(3-rsq*lu*lu)

	movapd  xmm15, [rsp + nb331_half]
	mulsd   xmm9, xmm15  ;# first iteration for rinvO
	mulsd   xmm10, xmm15 ;# first iteration for rinvH1
    mulsd   xmm11, xmm15 ;# first iteration for rinvH2	

    ;# second iteration step    
	movapd  xmm2, xmm9
	movapd  xmm5, xmm10
    movapd  xmm8, xmm11
    
	mulsd   xmm2, xmm2 ;# lu*lu
	mulsd   xmm5, xmm5 ;# lu*lu
    mulsd   xmm8, xmm8 ;# lu*lu
		
	movapd  xmm1, [rsp + nb331_three]
	movapd  xmm4, xmm1
    movapd  xmm7, xmm1

	mulsd   xmm2, xmm0 ;# rsq*lu*lu
	mulsd   xmm5, xmm3 ;# rsq*lu*lu 
    mulsd   xmm8, xmm6 ;# rsq*lu*lu
	
	subsd   xmm1, xmm2
	subsd   xmm4, xmm5
    subsd   xmm7, xmm8 ;# 3-rsq*lu*lu

	mulsd   xmm9, xmm1
	mulsd   xmm10, xmm4