nb_kernel313_x86_64_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,374 行 · 第 1/5 页

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addsd  xmm0, xmm3
	addsd  xmm1, xmm4
	addsd  xmm2, xmm5 ;# sum is in low xmm0-xmm2 

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

	;# increment i force 
	movsd  xmm3, [rdi + rcx*8 + 72]
	movsd  xmm4, [rdi + rcx*8 + 80]
	movsd  xmm5, [rdi + rcx*8 + 88]
	subsd  xmm3, xmm0
	subsd  xmm4, xmm1
	subsd  xmm5, xmm2
	movsd  [rdi + rcx*8 + 72], xmm3
	movsd  [rdi + rcx*8 + 80], xmm4
	movsd  [rdi + rcx*8 + 88], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	addsd xmm7, xmm2
	unpcklpd xmm0, xmm1
	addpd xmm6, xmm0

	;# increment fshift force 
	movlpd xmm3, [rsi + rdx*8]
	movhpd xmm3, [rsi + rdx*8 + 8]
	movsd  xmm4, [rsi + rdx*8 + 16]
	subpd  xmm3, xmm6
	subsd  xmm4, xmm7
	movlpd [rsi + rdx*8],      xmm3
	movhpd [rsi + rdx*8 + 8],  xmm3
	movsd  [rsi + rdx*8 + 16], xmm4

	;# get n from stack
	mov esi, [rsp + nb313_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb313_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movapd xmm7, [rsp + nb313_vctot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 
        
	;# add earlier value from mem 
	mov   rax, [rbp + nb313_Vc]
	addsd xmm7, [rax + rdx*8] 
	;# move back to mem 
	movsd [rax + rdx*8], xmm7 
	
	;# accumulate total lj energy and update it 
	movapd xmm7, [rsp + nb313_Vvdwtot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 

	;# add earlier value from mem 
	mov   rax, [rbp + nb313_Vvdw]
	addsd xmm7, [rax + rdx*8] 
	;# move back to mem 
	movsd [rax + rdx*8], xmm7 
	
        ;# finish if last 
        mov ecx, [rsp + nb313_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb313_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb313_n], esi
        jmp .nb313_outer
.nb313_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb313_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb313_end
        ;# non-zero, do one more workunit
        jmp   .nb313_threadloop
.nb313_end:
	mov eax, [rsp + nb313_nouter]
	mov ebx, [rsp + nb313_ninner]
	mov rcx, [rbp + nb313_outeriter]
	mov rdx, [rbp + nb313_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 1080
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret


	



.globl nb_kernel313nf_x86_64_sse2
.globl _nb_kernel313nf_x86_64_sse2
nb_kernel313nf_x86_64_sse2:	
_nb_kernel313nf_x86_64_sse2:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb313nf_fshift,         16
.equiv          nb313nf_gid,            24
.equiv          nb313nf_pos,            32
.equiv          nb313nf_faction,        40
.equiv          nb313nf_charge,         48
.equiv          nb313nf_p_facel,        56
.equiv          nb313nf_argkrf,         64
.equiv          nb313nf_argcrf,         72
.equiv          nb313nf_Vc,             80
.equiv          nb313nf_type,           88
.equiv          nb313nf_p_ntype,        96
.equiv          nb313nf_vdwparam,       104
.equiv          nb313nf_Vvdw,           112
.equiv          nb313nf_p_tabscale,     120
.equiv          nb313nf_VFtab,          128
.equiv          nb313nf_invsqrta,       136
.equiv          nb313nf_dvda,           144
.equiv          nb313nf_p_gbtabscale,   152
.equiv          nb313nf_GBtab,          160
.equiv          nb313nf_p_nthreads,     168
.equiv          nb313nf_count,          176
.equiv          nb313nf_mtx,            184
.equiv          nb313nf_outeriter,      192
.equiv          nb313nf_inneriter,      200
.equiv          nb313nf_work,           208
	;# stack offsets for local variables 
	;# bottom of stack is cache-aligned for sse2 use 
.equiv          nb313nf_ixO,            0
.equiv          nb313nf_iyO,            16
.equiv          nb313nf_izO,            32
.equiv          nb313nf_ixH1,           48
.equiv          nb313nf_iyH1,           64
.equiv          nb313nf_izH1,           80
.equiv          nb313nf_ixH2,           96
.equiv          nb313nf_iyH2,           112
.equiv          nb313nf_izH2,           128
.equiv          nb313nf_ixM,            144
.equiv          nb313nf_iyM,            160
.equiv          nb313nf_izM,            176
.equiv          nb313nf_iqM,            192
.equiv          nb313nf_iqH,            208
.equiv          nb313nf_qqM,            224
.equiv          nb313nf_qqH,            240
.equiv          nb313nf_rinvsqO,        256
.equiv          nb313nf_rinvH1,         272
.equiv          nb313nf_rinvH2,         288
.equiv          nb313nf_rinvM,          304
.equiv          nb313nf_rO,             320
.equiv          nb313nf_rH1,            336
.equiv          nb313nf_rH2,            352
.equiv          nb313nf_rM,             368
.equiv          nb313nf_tsc,            384
.equiv          nb313nf_two,            400
.equiv          nb313nf_c6,             416
.equiv          nb313nf_c12,            432
.equiv          nb313nf_vctot,          448
.equiv          nb313nf_Vvdwtot,        464
.equiv          nb313nf_half,           480
.equiv          nb313nf_three,          496
.equiv          nb313nf_is3,            512
.equiv          nb313nf_ii3,            516
.equiv          nb313nf_nri,            520
.equiv          nb313nf_iinr,           528
.equiv          nb313nf_jindex,         536
.equiv          nb313nf_jjnr,           544
.equiv          nb313nf_shift,          552
.equiv          nb313nf_shiftvec,       560
.equiv          nb313nf_facel,          568
.equiv          nb313nf_innerjjnr,      576
.equiv          nb313nf_ntia,           584
.equiv          nb313nf_innerk,         588
.equiv          nb313nf_n,              592
.equiv          nb313nf_nn1,            596
.equiv          nb313nf_nouter,         600
.equiv          nb313nf_ninner,         604
	push rbp
	mov  rbp, rsp
	push rbx
	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 616		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb313nf_nouter], eax
	mov [rsp + nb313nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb313nf_nri], edi
	mov [rsp + nb313nf_iinr], rsi
	mov [rsp + nb313nf_jindex], rdx
	mov [rsp + nb313nf_jjnr], rcx
	mov [rsp + nb313nf_shift], r8
	mov [rsp + nb313nf_shiftvec], r9
	mov rsi, [rbp + nb313nf_p_facel]
	movsd xmm0, [rsi]
	movsd [rsp + nb313nf_facel], xmm0

	mov rax, [rbp + nb313nf_p_tabscale]
	movsd xmm3, [rax]
	shufpd xmm3, xmm3, 0
	movapd [rsp + nb313nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double half IEEE (hex)
	mov ebx, 0x3fe00000
	mov [rsp + nb313nf_half], eax
	mov [rsp + nb313nf_half+4], ebx
	movsd xmm1, [rsp + nb313nf_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 + nb313nf_half], xmm1
	movapd [rsp + nb313nf_two], xmm2
	movapd [rsp + nb313nf_three], xmm3
	
	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb313nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb313nf_charge]
	movsd xmm3, [rdx + rbx*8 + 8]	
	movsd xmm4, [rdx + rbx*8 + 24]	
	mov rsi, [rbp + nb313nf_p_facel]
	movsd xmm0, [rsi]
	movsd xmm5, [rsp + nb313nf_facel]
	mulsd  xmm3, xmm5
	mulsd  xmm4, xmm5

	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	movapd [rsp + nb313nf_iqH], xmm3
	movapd [rsp + nb313nf_iqM], xmm4
	
	mov   rdx, [rbp + nb313nf_type]
	mov   ecx, [rdx + rbx*4]
	shl   ecx, 1
	mov rdi, [rbp + nb313nf_p_ntype]
	imul  ecx, [rdi]      ;# rcx = ntia = 2*ntype*type[ii0] 
	mov   [rsp + nb313nf_ntia], ecx		
.nb313nf_threadloop:
        mov   rsi, [rbp + nb313nf_count]          ;# pointer to sync counter
        mov   eax, [rsi]
.nb313nf_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 .nb313nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb313nf_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 + nb313nf_n], eax
        mov [rsp + nb313nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb313nf_outerstart
        jmp .nb313nf_end

.nb313nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb313nf_nouter]
	mov [rsp + nb313nf_nouter], ebx

.nb313nf_outer:
	mov   rax, [rsp + nb313nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax+rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 
	mov   [rsp + nb313nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb313nf_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 + nb313nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx+rsi*4]	    ;# ebx =ii 

	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	movapd xmm6, xmm0
	movapd xmm7, xmm1

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb313nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb313nf_ii3], ebx

	addsd xmm3, [rax + rbx*8] 	;# ox
	addsd xmm4, [rax + rbx*8 + 8] 	;# oy
	addsd xmm5, [rax + rbx*8 + 16]	;# oz	
	addsd xmm6, [rax + rbx*8 + 24] 	;# h1x
	addsd xmm7, [rax + rbx*8 + 32] 	;# h1y
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	shufpd xmm6, xmm6, 0
	shufpd xmm7, xmm7, 0
	movapd [rsp + nb313nf_ixO], xmm3
	movapd [rsp + nb313nf_iyO], xmm4
	movapd [rsp + nb313nf_izO], xmm5
	movapd [rsp + nb313nf_ixH1], xmm6
	movapd [rsp + nb313nf_iyH1], xmm7

	movsd xmm6, xmm2
	movsd xmm3, xmm0
	movsd xmm4, xmm1
	movsd xmm5, xmm2
	addsd xmm6, [rax + rbx*8 + 40] ;# h1z
	addsd xmm0, [rax + rbx*8 + 48] ;# h2x
	addsd xmm1, [rax + rbx*8 + 56] ;# h2y
	addsd xmm2, [rax + rbx*8 + 64] ;# h2z
	addsd xmm3, [rax + rbx*8 + 72] ;# mx
	addsd xmm4, [rax + rbx*8 + 80] ;# my
	addsd xmm5, [rax + rbx*8 + 88] ;# mz

	shufpd xmm6, xmm6, 0
	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 + nb313nf_izH1], xmm6
	movapd [rsp + nb313nf_ixH2], xmm0
	movapd [rsp + nb313nf_iyH2], xmm1
	movapd [rsp + nb313nf_izH2], xmm2
	movapd [rsp + nb313nf_ixM], xmm3
	movapd [rsp + nb313nf_iyM], xmm4
	movapd [rsp + nb313nf_izM], xmm5
	
	;# clear vctot
	xorpd xmm4, xmm4
	movapd [rsp + nb313nf_vctot], xmm4
	movapd [rsp + nb313nf_Vvdwtot], xmm4
	
	mov   rax, [rsp + nb313nf_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 + nb313nf_pos]
	mov   rdi, [rbp + nb313nf_faction]	
	mov   rax, [rsp + nb313nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb313nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [rsp + nb313nf_ninner]
	mov   [rsp + nb313nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb313nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb313nf_unroll_loop
	jmp   .nb313nf_checksingle
.nb313nf_unroll_loop:
	;# twice unrolled innerloop here 
	mov   rdx, [rsp + nb313nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	mov   ebx, [rdx + 4]
	
	add qword ptr [rsp + nb313nf_innerjjnr],  8 ;# advance pointer (unrolled 2) 

	mov rsi, [rbp + nb313nf_charge]    ;# base of charge[] 
	
	movlpd xmm3, [rsi + rax*8]
	movhpd xmm3, [rsi + rbx*8]
	movapd xmm4, xmm3	     
	mulpd  xmm3, [rsp + nb313nf_iqM]
	mulpd  xmm4, [rsp + nb313nf_iqH]
	movapd  [rsp + nb313nf_qqM], xmm3
	movapd  [rsp + nb313nf_qqH], xmm4	

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx
	mov rsi, [rbp + nb313nf_type]
	mov eax, [rsi + rax*4]
	mov ebx, [rsi + rbx*4]
	mov rsi, [rbp + nb313nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	mov edi, [rsp + nb313nf_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 + nb313nf_c6], xmm4
	movapd [rsp + nb313nf_c12], xmm6

	mov rsi, [rbp + nb313nf_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 + nb313nf_ixO]
	movapd xmm5, [rsp + nb313nf_iyO]
	movapd xmm6, [rsp + nb313nf_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 + nb313nf_ixH1]
	movapd xmm5, [rsp + nb313nf_iyH1]
	movapd xmm6, [rsp + nb313nf_izH1]

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

	;# square it