nb_kernel102_x86_64_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,311 行 · 第 1/5 页

.equiv          nb102nf_rinvOO,         528
.equiv          nb102nf_rinvOH1,        544
.equiv          nb102nf_rinvOH2,        560
.equiv          nb102nf_rinvH1O,        576
.equiv          nb102nf_rinvH1H1,       592
.equiv          nb102nf_rinvH1H2,       608
.equiv          nb102nf_rinvH2O,        624
.equiv          nb102nf_rinvH2H1,       640
.equiv          nb102nf_rinvH2H2,       656
.equiv          nb102nf_is3,            672
.equiv          nb102nf_ii3,            676
.equiv          nb102nf_nri,            680
.equiv          nb102nf_iinr,           688
.equiv          nb102nf_jindex,         696
.equiv          nb102nf_jjnr,           704
.equiv          nb102nf_shift,          712
.equiv          nb102nf_shiftvec,       720
.equiv          nb102nf_facel,          728
.equiv          nb102nf_innerjjnr,      736
.equiv          nb102nf_innerk,         744
.equiv          nb102nf_n,              748
.equiv          nb102nf_nn1,            752
.equiv          nb102nf_nouter,         756
.equiv          nb102nf_ninner,         760
	push rbp
	mov  rbp, rsp
	push rbx
	
	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 792		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb102nf_nouter], eax
	mov [rsp + nb102nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb102nf_nri], edi
	mov [rsp + nb102nf_iinr], rsi
	mov [rsp + nb102nf_jindex], rdx
	mov [rsp + nb102nf_jjnr], rcx
	mov [rsp + nb102nf_shift], r8
	mov [rsp + nb102nf_shiftvec], r9
	mov rsi, [rbp + nb102nf_p_facel]
	movsd xmm0, [rsi]
	movsd [rsp + nb102nf_facel], xmm0

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double half IEEE (hex)
	mov ebx, 0x3fe00000
	mov [rsp + nb102nf_half], eax
	mov [rsp + nb102nf_half+4], ebx
	movsd xmm1, [rsp + nb102nf_half]
	shufpd xmm1, xmm1, 0    ;# splat to all elements
	movapd xmm3, xmm1
	addpd  xmm3, xmm3       ;# one
	movapd xmm2, xmm3
	addpd  xmm2, xmm2       ;# two
	addpd  xmm3, xmm2	;# three
	movapd [rsp + nb102nf_half], xmm1
	movapd [rsp + nb102nf_three], xmm3
	
	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb102nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb102nf_charge]
	movsd xmm3, [rdx + rbx*8]	;# qO 
	movsd xmm4, xmm3		;# qO 
	movsd xmm5, [rdx + rbx*8 + 8]	;# qH 
	mov rsi, [rbp + nb102nf_p_facel]
	movsd xmm0, [rsi]
	movsd xmm6, [rsp + nb102nf_facel]	;# facel 
	mulsd  xmm3, xmm3		;# qO*qO 
	mulsd  xmm4, xmm5		;# qO*qH 
	mulsd  xmm5, xmm5		;# qH*qH 
	mulsd  xmm3, xmm6
	mulsd  xmm4, xmm6
	mulsd  xmm5, xmm6
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	movapd [rsp + nb102nf_qqOO], xmm3
	movapd [rsp + nb102nf_qqOH], xmm4
	movapd [rsp + nb102nf_qqHH], xmm5

.nb102nf_threadloop:
        mov   rsi, [rbp + nb102nf_count]        ;# pointer to sync counter
        mov   eax, [rsi]
.nb102nf_spinlock:
        mov   ebx, eax                          ;# ebx=*count=nn0
        add   ebx, 1                           	;# ebx=nn1=nn0+10
        lock
        cmpxchg [esi], ebx                      ;# write nn1 to *counter,
                                                ;# if it hasnt changed.
                                                ;# or reread *counter to eax.
        pause                                   ;# -> better p4 performance
        jnz .nb102nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb102nf_nri]
        mov edx, ecx
        sub ecx, ebx
        cmovle ebx, edx                         ;# if(nn1>nri) nn1=nri
        ;# Cleared the spinlock if we got here.
        ;# eax contains nn0, ebx contains nn1.
        mov [rsp + nb102nf_n], eax
        mov [rsp + nb102nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb102nf_outerstart
        jmp .nb102nf_end

.nb102nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb102nf_nouter]
	mov [rsp + nb102nf_nouter], ebx

.nb102nf_outer:
	mov   rax, [rsp + nb102nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax+rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 

	mov   rax, [rsp + nb102nf_shiftvec]   ;# rax = base of shiftvec[] 

	movsd xmm0, [rax + rbx*8]
	movsd xmm1, [rax + rbx*8 + 8]
	movsd xmm2, [rax + rbx*8 + 16] 

	mov   rcx, [rsp + nb102nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx+rsi*4]	    ;# ebx =ii 

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb102nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb102nf_ii3], ebx	
	
	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	addsd xmm3, [rax + rbx*8]
	addsd xmm4, [rax + rbx*8 + 8]
	addsd xmm5, [rax + rbx*8 + 16]		
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	movapd [rsp + nb102nf_ixO], xmm3
	movapd [rsp + nb102nf_iyO], xmm4
	movapd [rsp + nb102nf_izO], xmm5

	movsd xmm3, xmm0
	movsd xmm4, xmm1
	movsd xmm5, xmm2
	addsd xmm0, [rax + rbx*8 + 24]
	addsd xmm1, [rax + rbx*8 + 32]
	addsd xmm2, [rax + rbx*8 + 40]		
	addsd xmm3, [rax + rbx*8 + 48]
	addsd xmm4, [rax + rbx*8 + 56]
	addsd xmm5, [rax + rbx*8 + 64]		

	shufpd xmm0, xmm0, 0
	shufpd xmm1, xmm1, 0
	shufpd xmm2, xmm2, 0
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	movapd [rsp + nb102nf_ixH1], xmm0
	movapd [rsp + nb102nf_iyH1], xmm1
	movapd [rsp + nb102nf_izH1], xmm2
	movapd [rsp + nb102nf_ixH2], xmm3
	movapd [rsp + nb102nf_iyH2], xmm4
	movapd [rsp + nb102nf_izH2], xmm5

	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [rsp + nb102nf_vctot], xmm4
	
	mov   rax, [rsp + nb102nf_jindex]
	mov   ecx, [rax + rsi*4]	     ;# jindex[n] 
	mov   edx, [rax + rsi*4 + 4]	     ;# jindex[n+1] 
	sub   edx, ecx               ;# number of innerloop atoms 

	mov   rsi, [rbp + nb102nf_pos]
	mov   rax, [rsp + nb102nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb102nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [rsp + nb102nf_ninner]
	mov   [rsp + nb102nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb102nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb102nf_unroll_loop
	jmp   .nb102nf_checksingle
.nb102nf_unroll_loop:	
	;# twice unrolled innerloop here 
	mov   rdx, [rsp + nb102nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	mov   ebx, [rdx + 4] 
	
	add qword ptr [rsp + nb102nf_innerjjnr], 8 ;# advance pointer (unrolled 2) 

	mov rsi, [rbp + nb102nf_pos]       ;# base of pos[] 

	lea   rax, [rax + rax*2]     ;# replace jnr with j3 
	lea   rbx, [rbx + rbx*2]	
	
	;# move j coordinates to local temp variables 
	movlpd xmm2, [rsi + rax*8]
	movlpd xmm3, [rsi + rax*8 + 8]
	movlpd xmm4, [rsi + rax*8 + 16]
	movlpd xmm5, [rsi + rax*8 + 24]
	movlpd xmm6, [rsi + rax*8 + 32]
	movlpd xmm7, [rsi + rax*8 + 40]
	movhpd xmm2, [rsi + rbx*8]
	movhpd xmm3, [rsi + rbx*8 + 8]
	movhpd xmm4, [rsi + rbx*8 + 16]
	movhpd xmm5, [rsi + rbx*8 + 24]
	movhpd xmm6, [rsi + rbx*8 + 32]
	movhpd xmm7, [rsi + rbx*8 + 40]
	movapd 	[rsp + nb102nf_jxO], xmm2
	movapd 	[rsp + nb102nf_jyO], xmm3
	movapd 	[rsp + nb102nf_jzO], xmm4
	movapd 	[rsp + nb102nf_jxH1], xmm5
	movapd 	[rsp + nb102nf_jyH1], xmm6
	movapd 	[rsp + nb102nf_jzH1], xmm7
	movlpd xmm2, [rsi + rax*8 + 48]
	movlpd xmm3, [rsi + rax*8 + 56]
	movlpd xmm4, [rsi + rax*8 + 64]
	movhpd xmm2, [rsi + rbx*8 + 48]
	movhpd xmm3, [rsi + rbx*8 + 56]
	movhpd xmm4, [rsi + rbx*8 + 64]
	movapd 	[rsp + nb102nf_jxH2], xmm2
	movapd 	[rsp + nb102nf_jyH2], xmm3
	movapd 	[rsp + nb102nf_jzH2], xmm4
	
	movapd xmm0, [rsp + nb102nf_ixO]
	movapd xmm1, [rsp + nb102nf_iyO]
	movapd xmm2, [rsp + nb102nf_izO]
	movapd xmm3, [rsp + nb102nf_ixO]
	movapd xmm4, [rsp + nb102nf_iyO]
	movapd xmm5, [rsp + nb102nf_izO]
	subpd  xmm0, [rsp + nb102nf_jxO]
	subpd  xmm1, [rsp + nb102nf_jyO]
	subpd  xmm2, [rsp + nb102nf_jzO]
	subpd  xmm3, [rsp + nb102nf_jxH1]
	subpd  xmm4, [rsp + nb102nf_jyH1]
	subpd  xmm5, [rsp + nb102nf_jzH1]
	mulpd  xmm0, xmm0
	mulpd  xmm1, xmm1
	mulpd  xmm2, xmm2
	mulpd  xmm3, xmm3
	mulpd  xmm4, xmm4
	mulpd  xmm5, xmm5
	addpd  xmm0, xmm1
	addpd  xmm0, xmm2
	addpd  xmm3, xmm4
	addpd  xmm3, xmm5
	movapd [rsp + nb102nf_rsqOO], xmm0
	movapd [rsp + nb102nf_rsqOH1], xmm3

	movapd xmm0, [rsp + nb102nf_ixO]
	movapd xmm1, [rsp + nb102nf_iyO]
	movapd xmm2, [rsp + nb102nf_izO]
	movapd xmm3, [rsp + nb102nf_ixH1]
	movapd xmm4, [rsp + nb102nf_iyH1]
	movapd xmm5, [rsp + nb102nf_izH1]
	subpd  xmm0, [rsp + nb102nf_jxH2]
	subpd  xmm1, [rsp + nb102nf_jyH2]
	subpd  xmm2, [rsp + nb102nf_jzH2]
	subpd  xmm3, [rsp + nb102nf_jxO]
	subpd  xmm4, [rsp + nb102nf_jyO]
	subpd  xmm5, [rsp + nb102nf_jzO]
	mulpd  xmm0, xmm0
	mulpd  xmm1, xmm1
	mulpd  xmm2, xmm2
	mulpd  xmm3, xmm3
	mulpd  xmm4, xmm4
	mulpd  xmm5, xmm5
	addpd  xmm0, xmm1
	addpd  xmm0, xmm2
	addpd  xmm3, xmm4
	addpd  xmm3, xmm5
	movapd [rsp + nb102nf_rsqOH2], xmm0
	movapd [rsp + nb102nf_rsqH1O], xmm3

	movapd xmm0, [rsp + nb102nf_ixH1]
	movapd xmm1, [rsp + nb102nf_iyH1]
	movapd xmm2, [rsp + nb102nf_izH1]
	movapd xmm3, [rsp + nb102nf_ixH1]
	movapd xmm4, [rsp + nb102nf_iyH1]
	movapd xmm5, [rsp + nb102nf_izH1]
	subpd  xmm0, [rsp + nb102nf_jxH1]
	subpd  xmm1, [rsp + nb102nf_jyH1]
	subpd  xmm2, [rsp + nb102nf_jzH1]
	subpd  xmm3, [rsp + nb102nf_jxH2]
	subpd  xmm4, [rsp + nb102nf_jyH2]
	subpd  xmm5, [rsp + nb102nf_jzH2]
	mulpd  xmm0, xmm0
	mulpd  xmm1, xmm1
	mulpd  xmm2, xmm2
	mulpd  xmm3, xmm3
	mulpd  xmm4, xmm4
	mulpd  xmm5, xmm5
	addpd  xmm0, xmm1
	addpd  xmm0, xmm2
	addpd  xmm3, xmm4
	addpd  xmm3, xmm5
	movapd [rsp + nb102nf_rsqH1H1], xmm0
	movapd [rsp + nb102nf_rsqH1H2], xmm3

	movapd xmm0, [rsp + nb102nf_ixH2]
	movapd xmm1, [rsp + nb102nf_iyH2]
	movapd xmm2, [rsp + nb102nf_izH2]
	movapd xmm3, [rsp + nb102nf_ixH2]
	movapd xmm4, [rsp + nb102nf_iyH2]
	movapd xmm5, [rsp + nb102nf_izH2]
	subpd  xmm0, [rsp + nb102nf_jxO]
	subpd  xmm1, [rsp + nb102nf_jyO]
	subpd  xmm2, [rsp + nb102nf_jzO]
	subpd  xmm3, [rsp + nb102nf_jxH1]
	subpd  xmm4, [rsp + nb102nf_jyH1]
	subpd  xmm5, [rsp + nb102nf_jzH1]
	mulpd  xmm0, xmm0
	mulpd  xmm1, xmm1
	mulpd  xmm2, xmm2
	mulpd  xmm3, xmm3
	mulpd  xmm4, xmm4
	mulpd  xmm5, xmm5
	addpd  xmm0, xmm1
	addpd  xmm0, xmm2
	addpd  xmm4, xmm3
	addpd  xmm4, xmm5
	movapd [rsp + nb102nf_rsqH2O], xmm0
	movapd [rsp + nb102nf_rsqH2H1], xmm4

	movapd xmm0, [rsp + nb102nf_ixH2]
	movapd xmm1, [rsp + nb102nf_iyH2]
	movapd xmm2, [rsp + nb102nf_izH2]
	subpd  xmm0, [rsp + nb102nf_jxH2]
	subpd  xmm1, [rsp + nb102nf_jyH2]
	subpd  xmm2, [rsp + nb102nf_jzH2]
	mulpd xmm0, xmm0
	mulpd xmm1, xmm1
	mulpd xmm2, xmm2
	addpd xmm0, xmm1
	addpd xmm0, xmm2
	movapd [rsp + nb102nf_rsqH2H2], xmm0
	
	;# start doing invsqrt use rsq values in xmm0 (h2h2) , xmm4 (h2h1) 
	cvtpd2ps xmm1, xmm0	
	cvtpd2ps xmm5, xmm4	
	rsqrtps xmm1, xmm1
	rsqrtps xmm5, xmm5
	cvtps2pd xmm1, xmm1
	cvtps2pd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulpd   xmm1, xmm1	;# luA*luA 
	mulpd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [rsp + nb102nf_three]
	mulpd   xmm1, xmm0	;# rsqA*luA*luA 
	mulpd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subpd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subpd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulpd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulpd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulpd   xmm3, [rsp + nb102nf_half] ;# iter1 
	mulpd   xmm7, [rsp + nb102nf_half] ;# iter1 

	movapd  xmm2, xmm3	;# copy of luA 
	movapd  xmm6, xmm7	;# copy of luB 
	mulpd   xmm3, xmm3	;# luA*luA 
	mulpd   xmm7, xmm7	;# luB*luB 
	movapd  xmm1, [rsp + nb102nf_three]
	mulpd   xmm3, xmm0	;# rsqA*luA*luA 
	mulpd   xmm7, xmm4	;# rsqB*luB*luB 	
	movapd  xmm5, xmm1
	subpd   xmm1, xmm3	;# 3-rsqA*luA*luA 
	subpd   xmm5, xmm7	;# 3-rsqB*luB*luB 
	mulpd   xmm1, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulpd   xmm5, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulpd   xmm1, [rsp + nb102nf_half] ;# rinv 
	mulpd   xmm5, [rsp + nb102nf_half] ;# rinv 
	movapd [rsp + nb102nf_rinvH2H2], xmm1
	movapd [rsp + nb102nf_rinvH2H1], xmm5

	movapd xmm0, [rsp + nb102nf_rsqOO]
	movapd xmm4, [rsp + nb102nf_rsqOH1]	
	cvtpd2ps xmm1, xmm0	
	cvtpd2ps xmm5, xmm4	
	rsqrtps xmm1, xmm1
	rsqrtps xmm5, xmm5
	cvtps2pd xmm1, xmm1
	cvtps2pd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulpd   xmm1, xmm1	;# luA*luA 
	mulpd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [rsp + nb102nf_three]
	mulpd   xmm1, xmm0	;# rsqA*luA*luA 
	mulpd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subpd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subpd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulpd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulpd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulpd   xmm3, [rsp + nb102nf_half] ;# iter1 of  
	mulpd   xmm7, [rsp + nb102nf_half] ;# iter1 of  

	movapd  xmm2, xmm3	;# copy of luA 
	movapd  xmm6, xmm7	;# copy of luB 
	mulpd   xmm3, xmm3	;# luA*luA 
	mulpd   xmm7, xmm7	;# luB*luB 
	movapd  xmm1, [rsp + nb102nf_three]
	mulpd   xmm3, xmm0	;# rsqA*luA*luA 
	mulpd   xmm7, xmm4	;# rsqB*luB*luB 	
	movapd  xmm5, xmm1
	subpd   xmm1, xmm3	;# 3-rsqA*luA*luA 
	subpd   xmm5, xmm7	;# 3-rsqB*luB*luB 
	mulpd   xmm1, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulpd   xmm5, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulpd   xmm1, [rsp + nb102nf_half] ;# rinv 
	mulpd   xmm5, [rsp + nb102nf_half] ;# rinv
	movapd [rsp + nb102nf_rinvOO], xmm1
	movapd [rsp + nb102nf_rinvOH1], xmm5

	movapd xmm0, [rsp + nb102nf_rsqOH2]
	movapd xmm4, [rsp + nb102nf_rsqH1O]	
	cvtpd2ps xmm1, xmm0	
	cvtpd2ps xmm5, xmm4	
	rsqrtps xmm1, xmm1
	rsqrtps xmm5, xmm5
	cvtps2pd xmm1, xmm1
	cvtps2pd xmm5, xmm5
	
	movapd  xmm2, xmm1	;# copy of luA 
	movapd  xmm6, xmm5	;# copy of luB 
	mulpd   xmm1, xmm1	;# luA*luA 
	mulpd   xmm5, xmm5	;# luB*luB 
	movapd  xmm3, [rsp + nb102nf_three]
	mulpd   xmm1, xmm0	;# rsqA*luA*luA 
	mulpd   xmm5, xmm4	;# rsqB*luB*luB 	
	movapd  xmm7, xmm3
	subpd   xmm3, xmm1	;# 3-rsqA*luA*luA 
	subpd   xmm7, xmm5	;# 3-rsqB*luB*luB 
	mulpd   xmm3, xmm2	;# luA*(3-rsqA*luA*luA) 
	mulpd   xmm7, xmm6	;# luB*(3-rsqB*luB*luB) 
	mulpd   xm