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

.equiv          nb331nf_rinvO,          208
.equiv          nb331nf_rinvH1,         224
.equiv          nb331nf_rinvH2,         240
.equiv          nb331nf_rO,             256
.equiv          nb331nf_rH1,            272
.equiv          nb331nf_rH2,            288
.equiv          nb331nf_tsc,            304
.equiv          nb331nf_c6,             320
.equiv          nb331nf_c12,            336
.equiv          nb331nf_vctot,          352
.equiv          nb331nf_Vvdwtot,        368
.equiv          nb331nf_half,           384
.equiv          nb331nf_three,          400
.equiv          nb331nf_is3,            416
.equiv          nb331nf_ii3,            420
.equiv          nb331nf_nri,            424
.equiv          nb331nf_iinr,           432
.equiv          nb331nf_jindex,         440
.equiv          nb331nf_jjnr,           448
.equiv          nb331nf_shift,          456
.equiv          nb331nf_shiftvec,       464
.equiv          nb331nf_facel,          472
.equiv          nb331nf_innerjjnr,      480
.equiv          nb331nf_ntia,           488
.equiv          nb331nf_innerk,         492
.equiv          nb331nf_n,              496
.equiv          nb331nf_nn1,            500
.equiv          nb331nf_nouter,         504
.equiv          nb331nf_ninner,         508
	push rbp
	mov  rbp, rsp
	push rbx
	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 520		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb331nf_nouter], eax
	mov [rsp + nb331nf_ninner], eax



	mov edi, [rdi]
	mov [rsp + nb331nf_nri], edi
	mov [rsp + nb331nf_iinr], rsi
	mov [rsp + nb331nf_jindex], rdx
	mov [rsp + nb331nf_jjnr], rcx
	mov [rsp + nb331nf_shift], r8
	mov [rsp + nb331nf_shiftvec], r9
	mov rsi, [rbp + nb331nf_p_facel]
	movsd xmm0, [rsi]
	movsd [rsp + nb331nf_facel], xmm0

	mov rax, [rbp + nb331nf_p_tabscale]
	movsd xmm3, [rax]
	shufpd xmm3, xmm3, 0
	movapd [rsp + nb331nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double half IEEE (hex)
	mov ebx, 0x3fe00000
	mov [rsp + nb331nf_half], eax
	mov [rsp + nb331nf_half+4], ebx
	movsd xmm1, [rsp + nb331nf_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 + nb331nf_half], xmm1
	movapd [rsp + nb331nf_three], xmm3

	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb331nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb331nf_charge]
	movsd xmm3, [rdx + rbx*8]	
	movsd xmm4, [rdx + rbx*8 + 8]	
	mov rsi, [rbp + nb331nf_p_facel]
	movsd xmm0, [rsi]
	movsd xmm5, [rsp + nb331nf_facel]
	mulsd  xmm3, xmm5
	mulsd  xmm4, xmm5

	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	movapd [rsp + nb331nf_iqO], xmm3
	movapd [rsp + nb331nf_iqH], xmm4
	
	mov   rdx, [rbp + nb331nf_type]
	mov   ecx, [rdx + rbx*4]
	shl   ecx, 1
	mov rdi, [rbp + nb331nf_p_ntype]
	imul  ecx, [rdi]      ;# rcx = ntia = 2*ntype*type[ii0] 
	mov   [rsp + nb331nf_ntia], ecx		
.nb331nf_threadloop:
        mov   rsi, [rbp + nb331nf_count]        ;# pointer to sync counter
        mov   eax, [rsi]
.nb331nf_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 .nb331nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb331nf_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 + nb331nf_n], eax
        mov [rsp + nb331nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb331nf_outerstart
        jmp .nb331nf_end

.nb331nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb331nf_nouter]
	mov [rsp + nb331nf_nouter], ebx

.nb331nf_outer:
	mov   rax, [rsp + nb331nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 

	mov   rax, [rsp + nb331nf_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 + nb331nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx + rsi*4]	    ;# ebx =ii 

	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb331nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb331nf_ii3], ebx

	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 + nb331nf_ixO], xmm3
	movapd [rsp + nb331nf_iyO], xmm4
	movapd [rsp + nb331nf_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 + nb331nf_ixH1], xmm0
	movapd [rsp + nb331nf_iyH1], xmm1
	movapd [rsp + nb331nf_izH1], xmm2
	movapd [rsp + nb331nf_ixH2], xmm3
	movapd [rsp + nb331nf_iyH2], xmm4
	movapd [rsp + nb331nf_izH2], xmm5
	
	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [rsp + nb331nf_vctot], xmm4
	movapd [rsp + nb331nf_Vvdwtot], xmm4
	
	mov   rax, [rsp + nb331nf_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 + nb331nf_pos]
	mov   rax, [rsp + nb331nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb331nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [rsp + nb331nf_ninner]
	mov   [rsp + nb331nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb331nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb331nf_unroll_loop
	jmp   .nb331nf_checksingle
.nb331nf_unroll_loop:
	;# twice unrolled innerloop here 
	mov   rdx, [rsp + nb331nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	mov   ebx, [rdx + 4]
	
	add qword ptr [rsp + nb331nf_innerjjnr],  8 ;# advance pointer (unrolled 2) 

	mov rsi, [rbp + nb331nf_charge]    ;# base of charge[] 
	
	movlpd xmm3, [rsi + rax*8]
	movhpd xmm3, [rsi + rbx*8]
	movapd xmm4, xmm3	     
	mulpd  xmm3, [rsp + nb331nf_iqO]
	mulpd  xmm4, [rsp + nb331nf_iqH]
	movapd  [rsp + nb331nf_qqO], xmm3
	movapd  [rsp + nb331nf_qqH], xmm4	

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx
	mov rsi, [rbp + nb331nf_type]
	mov eax, [rsi + rax*4]
	mov ebx, [rsi + rbx*4]
	mov rsi, [rbp + nb331nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	mov edi, [rsp + nb331nf_ntia]
	add eax, edi
	add ebx, edi

	movlpd xmm6, [rsi + rax*8]	;# c6a
	movlpd xmm7, [rsi + rbx*8]	;# c6b
	movhpd xmm6, [rsi + rax*8 + 8]	;# c6a c12a 
	movhpd xmm7, [rsi + rbx*8 + 8]	;# c6b c12b 
	movapd xmm4, xmm6
	unpcklpd xmm4, xmm7
	unpckhpd xmm6, xmm7
	
	movd  eax, mm0
	movd  ebx, mm1
	movapd [rsp + nb331nf_c6], xmm4
	movapd [rsp + nb331nf_c12], xmm6

	mov rsi, [rbp + nb331nf_pos]       ;# base of pos[] 

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

	;# move two coordinates to xmm0-xmm2 	
	movlpd xmm0, [rsi + rax*8]
	movlpd xmm1, [rsi + rax*8 + 8]
	movlpd xmm2, [rsi + rax*8 + 16]
	movhpd xmm0, [rsi + rbx*8]
	movhpd xmm1, [rsi + rbx*8 + 8]
	movhpd xmm2, [rsi + rbx*8 + 16]		

	;# move ixO-izO to xmm4-xmm6 
	movapd xmm4, [rsp + nb331nf_ixO]
	movapd xmm5, [rsp + nb331nf_iyO]
	movapd xmm6, [rsp + nb331nf_izO]

	;# calc dr 
	subpd xmm4, xmm0
	subpd xmm5, xmm1
	subpd xmm6, xmm2

	;# square it 
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	mulpd xmm6,xmm6
	addpd xmm4, xmm5
	addpd xmm4, xmm6
	movapd xmm7, xmm4
	;# rsqO in xmm7 

	;# move ixH1-izH1 to xmm4-xmm6 
	movapd xmm4, [rsp + nb331nf_ixH1]
	movapd xmm5, [rsp + nb331nf_iyH1]
	movapd xmm6, [rsp + nb331nf_izH1]

	;# calc dr 
	subpd xmm4, xmm0
	subpd xmm5, xmm1
	subpd xmm6, xmm2

	;# square it 
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	mulpd xmm6,xmm6
	addpd xmm6, xmm5
	addpd xmm6, xmm4
	;# rsqH1 in xmm6 

	;# move ixH2-izH2 to xmm3-xmm5  
	movapd xmm3, [rsp + nb331nf_ixH2]
	movapd xmm4, [rsp + nb331nf_iyH2]
	movapd xmm5, [rsp + nb331nf_izH2]

	;# calc dr 
	subpd xmm3, xmm0
	subpd xmm4, xmm1
	subpd xmm5, xmm2

	;# square it 
	mulpd xmm3,xmm3
	mulpd xmm4,xmm4
	mulpd xmm5,xmm5
	addpd xmm5, xmm4
	addpd xmm5, xmm3
	;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

	;# start with rsqO - put seed in xmm2 
	cvtpd2ps xmm2, xmm7	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [rsp + nb331nf_three]
	mulpd   xmm2, xmm7	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [rsp + nb331nf_half] ;# iter1 ( new lu) 

	movapd xmm2, xmm7
	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm2, xmm4	;# rsq*lu*lu 
	movapd xmm4, [rsp + nb331nf_three]
	subpd xmm4, xmm2	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [rsp + nb331nf_half] ;# rinv 
	movapd  [rsp + nb331nf_rinvO], xmm4	;# rinvO in xmm4 
	mulpd   xmm7, xmm4
	movapd  [rsp + nb331nf_rO], xmm7	;# r in xmm7 
	
	;# rsqH1 - seed in xmm2 
	cvtpd2ps xmm2, xmm6	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [rsp + nb331nf_three]
	mulpd   xmm2, xmm6	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [rsp + nb331nf_half] ;# iter1 ( new lu) 

	movapd xmm2, xmm6
	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm2, xmm4	;# rsq*lu*lu 
	movapd xmm4, [rsp + nb331nf_three]
	subpd xmm4, xmm2	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [rsp + nb331nf_half] ;# rinv 
	movapd [rsp + nb331nf_rinvH1], xmm4	;# rinvH1 
	mulpd  xmm6, xmm4
	movapd [rsp + nb331nf_rH1], xmm6	;# rH1 
	
	;# rsqH2 - seed in xmm2 
	cvtpd2ps xmm2, xmm5	
	rsqrtps xmm2, xmm2
	cvtps2pd xmm2, xmm2

	movapd  xmm3, xmm2
	mulpd   xmm2, xmm2
	movapd  xmm4, [rsp + nb331nf_three]
	mulpd   xmm2, xmm5	;# rsq*lu*lu 
	subpd   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulpd   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulpd   xmm4, [rsp + nb331nf_half] ;# iter1 ( new lu) 

	movapd xmm2, xmm5
	movapd xmm3, xmm4
	mulpd xmm4, xmm4	;# lu*lu 
	mulpd xmm2, xmm4	;# rsq*lu*lu 
	movapd xmm4, [rsp + nb331nf_three]
	subpd xmm4, xmm2	;# 3-rsq*lu*lu 
	mulpd xmm4, xmm3	;# lu*(	3-rsq*lu*lu) 
	mulpd xmm4, [rsp + nb331nf_half] ;# rinv 
	movapd [rsp + nb331nf_rinvH2], xmm4 ;# rinv 
	mulpd xmm5, xmm4
	movapd [rsp + nb331nf_rH2], xmm5 ;# r 

	;# do O interactions 
	;# rO is still in xmm7 
	movapd xmm0, [rsp + nb331nf_rinvO]
	mulpd   xmm7, [rsp + nb331nf_tsc]
	cvttpd2pi mm6, xmm7	;# mm6 = lu idx 
	cvtpi2pd xmm6, mm6
	subpd xmm7, xmm6
	movapd xmm1, xmm7	;# xmm1=eps 
	movapd xmm2, xmm1	
	mulpd  xmm2, xmm2	;# xmm2=eps2 
	
	pslld mm6, 2		;# idx *= 4 
	movd mm0, eax	
	movd mm1, ebx
	mov  rsi, [rbp + nb331nf_VFtab]
	movd eax, mm6
	psrlq mm6, 32
	movd ebx, mm6		;# indices in eax/ebx 
	lea   rax, [rax + rax*2] ;# idx *= 3 (total *=12 now) 
	lea   rbx, [rbx + rbx*2]
	
	movapd xmm4, [rsi + rax*8]	;# Y1 F1 	
	movapd xmm3, [rsi + rbx*8]	;# Y2 F2 
	movapd xmm5, xmm4
	unpcklpd xmm4, xmm3	;# Y1 Y2 
	unpckhpd xmm5, xmm3	;# F1 F2 

	movapd xmm6, [rsi + rax*8 + 16]	;# G1 H1 	
	movapd xmm3, [rsi + rbx*8 + 16]	;# G2 H2 
	movapd xmm7, xmm6
	unpcklpd xmm6, xmm3	;# G1 G2 
	unpckhpd xmm7, xmm3	;# H1 H2 
	;# coulomb table ready, in xmm4-xmm7  
    mulpd  xmm6, xmm1       ;# xmm6=Geps 
    mulpd  xmm7, xmm2       ;# xmm7=Heps2 
    addpd  xmm5, xmm6
    addpd  xmm5, xmm7       ;# xmm5=Fp        
    movapd xmm0, [rsp + nb331nf_qqO]
    mulpd  xmm5, xmm1 ;# xmm5=eps*Fp 
    addpd  xmm5, xmm4 ;# xmm5=VV 
    mulpd  xmm5, xmm0 ;# vcoul=qq*VV  
	;# at this point xmm5 contains vcoul 
	;# increment vcoul - then we can get rid of mm5 
	;# update vctot 
	addpd  xmm5, [rsp + nb331nf_vctot]
	movapd [rsp + nb331nf_vctot], xmm5 
	
	;# Dispersion 
	movapd xmm4, [rsi + rax*8 + 32]	;# Y1 F1 	
	movapd xmm3, [rsi + rbx*8 + 32]	;# Y2 F2 
	movapd xmm5, xmm4
	unpcklpd xmm4, xmm3	;# Y1 Y2 
	unpckhpd xmm5, xmm3	;# F1 F2 

	movapd xmm6, [rsi + rax*8 + 48]	;# G1 H1 	
	movapd xmm3, [rsi + rbx*8 + 48]	;# G2 H2 
	movapd xmm7, xmm6
	unpcklpd xmm6, xmm3	;# G1 G2 
	unpckhpd xmm7, xmm3	;# H1 H2 
	;# Dispersion table ready, in xmm4-xmm7  		
	mulpd  xmm6, xmm1	;# xmm6=Geps 
	mulpd  xmm7, xmm2	;# xmm7=Heps2 
	addpd  xmm5, xmm6
	addpd  xmm5, xmm7	;# xmm5=Fp 	
	mulpd  xmm5, xmm1 ;# xmm5=eps*Fp 
	addpd  xmm5, xmm4 ;# xmm5=VV 

	mulpd  xmm5, [rsp + nb331nf_c6]	 ;# Vvdw6 
	
	addpd  xmm5, [rsp + nb331nf_Vvdwtot]
	movapd [rsp + nb331nf_Vvdwtot], xmm5