nb_kernel201_x86_64_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,738 行 · 第 1/4 页

	addsd  xmm4, xmm1
	addsd  xmm7, xmm2
	movlpd [rsp + nb201_fixH2], xmm3
	movlpd [rsp + nb201_fiyH2], xmm4
	movlpd [rsp + nb201_fizH2], xmm7

	mov rdi, [rbp + nb201_faction]
	;# update j forces 
	addsd  xmm0, [rsp + nb201_fjx]
	addsd  xmm1, [rsp + nb201_fjy]
	addsd  xmm2, [rsp + nb201_fjz]
	movlpd xmm3, [rdi + rax*8]
	movlpd xmm4, [rdi + rax*8 + 8]
	movlpd xmm5, [rdi + rax*8 + 16]
	addsd xmm3, xmm0
	addsd xmm4, xmm1
	addsd xmm5, xmm2
	movlpd [rdi + rax*8], xmm3
	movlpd [rdi + rax*8 + 8], xmm4
	movlpd [rdi + rax*8 + 16], xmm5	

.nb201_updateouterdata:
	mov   ecx, [rsp + nb201_ii3]
	mov   rdi, [rbp + nb201_faction]
	mov   rsi, [rbp + nb201_fshift]
	mov   edx, [rsp + nb201_is3]

	;# accumulate  Oi forces in xmm0, xmm1, xmm2 
	movapd xmm0, [rsp + nb201_fixO]
	movapd xmm1, [rsp + nb201_fiyO]
	movapd xmm2, [rsp + nb201_fizO]

	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]
	movsd  xmm4, [rdi + rcx*8 + 8]
	movsd  xmm5, [rdi + rcx*8 + 16]
	subsd  xmm3, xmm0
	subsd  xmm4, xmm1
	subsd  xmm5, xmm2
	movsd  [rdi + rcx*8],     xmm3
	movsd  [rdi + rcx*8 + 8], xmm4
	movsd  [rdi + rcx*8 + 16], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	movapd xmm6, xmm0
	movsd xmm7, xmm2
	unpcklpd xmm6, xmm1

	;# accumulate H1i forces in xmm0, xmm1, xmm2 
	movapd xmm0, [rsp + nb201_fixH1]
	movapd xmm1, [rsp + nb201_fiyH1]
	movapd xmm2, [rsp + nb201_fizH1]

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

	;# increment i force 
	movsd  xmm3, [rdi + rcx*8 + 24]
	movsd  xmm4, [rdi + rcx*8 + 32]
	movsd  xmm5, [rdi + rcx*8 + 40]
	subsd  xmm3, xmm0
	subsd  xmm4, xmm1
	subsd  xmm5, xmm2
	movsd  [rdi + rcx*8 + 24], xmm3
	movsd  [rdi + rcx*8 + 32], xmm4
	movsd  [rdi + rcx*8 + 40], xmm5

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

	;# accumulate H2i forces in xmm0, xmm1, xmm2 
	movapd xmm0, [rsp + nb201_fixH2]
	movapd xmm1, [rsp + nb201_fiyH2]
	movapd xmm2, [rsp + nb201_fizH2]

	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 + 48]
	movsd  xmm4, [rdi + rcx*8 + 56]
	movsd  xmm5, [rdi + rcx*8 + 64]
	subsd  xmm3, xmm0
	subsd  xmm4, xmm1
	subsd  xmm5, xmm2
	movsd  [rdi + rcx*8 + 48], xmm3
	movsd  [rdi + rcx*8 + 56], xmm4
	movsd  [rdi + rcx*8 + 64], 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 + nb201_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb201_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movapd xmm7, [rsp + nb201_vctot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 
        
	;# add earlier value from mem 
	mov   rax, [rbp + nb201_Vc]
	addsd xmm7, [rax + rdx*8] 
	;# move back to mem 
	movsd [rax + rdx*8], xmm7 
	
        ;# finish if last 
        mov ecx, [rsp + nb201_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb201_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb201_n], esi
        jmp .nb201_outer
.nb201_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb201_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb201_end
        ;# non-zero, do one more workunit
        jmp   .nb201_threadloop
.nb201_end:
	mov eax, [rsp + nb201_nouter]
	mov ebx, [rsp + nb201_ninner]
	mov rcx, [rbp + nb201_outeriter]
	mov rdx, [rbp + nb201_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 808
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret





.globl nb_kernel201nf_x86_64_sse2
.globl _nb_kernel201nf_x86_64_sse2
nb_kernel201nf_x86_64_sse2:	
_nb_kernel201nf_x86_64_sse2:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb201nf_fshift,         16
.equiv          nb201nf_gid,            24
.equiv          nb201nf_pos,            32
.equiv          nb201nf_faction,        40
.equiv          nb201nf_charge,         48
.equiv          nb201nf_p_facel,        56
.equiv          nb201nf_argkrf,         64
.equiv          nb201nf_argcrf,         72
.equiv          nb201nf_Vc,             80
.equiv          nb201nf_type,           88
.equiv          nb201nf_p_ntype,        96
.equiv          nb201nf_vdwparam,       104
.equiv          nb201nf_Vvdw,           112
.equiv          nb201nf_p_tabscale,     120
.equiv          nb201nf_VFtab,          128
.equiv          nb201nf_invsqrta,       136
.equiv          nb201nf_dvda,           144
.equiv          nb201nf_p_gbtabscale,   152
.equiv          nb201nf_GBtab,          160
.equiv          nb201nf_p_nthreads,     168
.equiv          nb201nf_count,          176
.equiv          nb201nf_mtx,            184
.equiv          nb201nf_outeriter,      192
.equiv          nb201nf_inneriter,      200
.equiv          nb201nf_work,           208
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb201nf_ixO,            0
.equiv          nb201nf_iyO,            16
.equiv          nb201nf_izO,            32
.equiv          nb201nf_ixH1,           48
.equiv          nb201nf_iyH1,           64
.equiv          nb201nf_izH1,           80
.equiv          nb201nf_ixH2,           96
.equiv          nb201nf_iyH2,           112
.equiv          nb201nf_izH2,           128
.equiv          nb201nf_iqO,            144
.equiv          nb201nf_iqH,            160
.equiv          nb201nf_qqO,            176
.equiv          nb201nf_qqH,            192
.equiv          nb201nf_vctot,          208
.equiv          nb201nf_half,           224
.equiv          nb201nf_three,          240
.equiv          nb201nf_krf,            256
.equiv          nb201nf_crf,            272
.equiv          nb201nf_krsqO,          288
.equiv          nb201nf_krsqH1,         304
.equiv          nb201nf_krsqH2,         320
.equiv          nb201nf_nri,            336
.equiv          nb201nf_iinr,           344
.equiv          nb201nf_jindex,         352
.equiv          nb201nf_jjnr,           360
.equiv          nb201nf_shift,          368
.equiv          nb201nf_shiftvec,       376
.equiv          nb201nf_facel,          384
.equiv          nb201nf_innerjjnr,      392
.equiv          nb201nf_is3,            400
.equiv          nb201nf_ii3,            404
.equiv          nb201nf_innerk,         408
.equiv          nb201nf_n,              412
.equiv          nb201nf_nn1,            416
.equiv          nb201nf_nouter,         420
.equiv          nb201nf_ninner,         424
	push rbp
	mov  rbp, rsp
	push rbx

	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 440		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb201nf_nouter], eax
	mov [rsp + nb201nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb201nf_nri], edi
	mov [rsp + nb201nf_iinr], rsi
	mov [rsp + nb201nf_jindex], rdx
	mov [rsp + nb201nf_jjnr], rcx
	mov [rsp + nb201nf_shift], r8
	mov [rsp + nb201nf_shiftvec], r9
	mov rsi, [rbp + nb201nf_p_facel]
	movsd xmm0, [rsi]
	movsd [rsp + nb201nf_facel], xmm0

	mov rsi, [rbp + nb201nf_argkrf]
	mov rdi, [rbp + nb201nf_argcrf]
	movsd xmm1, [rsi]
	movsd xmm2, [rdi]
	shufpd xmm1, xmm1, 0
	shufpd xmm2, xmm2, 0
	movapd [rsp + nb201nf_krf], xmm1
	movapd [rsp + nb201nf_crf], xmm2

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double half IEEE (hex)
	mov ebx, 0x3fe00000
	mov [rsp + nb201nf_half], eax
	mov [rsp + nb201nf_half+4], ebx
	movsd xmm1, [rsp + nb201nf_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 + nb201nf_half], xmm1
	movapd [rsp + nb201nf_three], xmm3

	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb201nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb201nf_charge]
	movsd xmm3, [rdx + rbx*8]	
	movsd xmm4, [rdx + rbx*8 + 8]	

	movsd xmm5, [rsp + nb201nf_facel]
	mulsd  xmm3, xmm5
	mulsd  xmm4, xmm5

	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	movapd [rsp + nb201nf_iqO], xmm3
	movapd [rsp + nb201nf_iqH], xmm4
	
.nb201nf_threadloop:
        mov   rsi, [rbp + nb201nf_count]        ;# pointer to sync counter
        mov   eax, [rsi]
.nb201nf_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 .nb201nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb201nf_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 + nb201nf_n], eax
        mov [rsp + nb201nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb201nf_outerstart
        jmp .nb201nf_end

.nb201nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb201nf_nouter]
	mov [rsp + nb201nf_nouter], ebx

.nb201nf_outer:
	mov   rax, [rsp + nb201nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax+rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 

	mov   rax, [rsp + nb201nf_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 + nb201nf_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 + nb201nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb201nf_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 + nb201nf_ixO], xmm3
	movapd [rsp + nb201nf_iyO], xmm4
	movapd [rsp + nb201nf_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 + nb201nf_ixH1], xmm0
	movapd [rsp + nb201nf_iyH1], xmm1
	movapd [rsp + nb201nf_izH1], xmm2
	movapd [rsp + nb201nf_ixH2], xmm3
	movapd [rsp + nb201nf_iyH2], xmm4
	movapd [rsp + nb201nf_izH2], xmm5
	
	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [rsp + nb201nf_vctot], xmm4
	
	mov   rax, [rsp + nb201nf_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 + nb201nf_pos]
	mov   rax, [rsp + nb201nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb201nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [rsp + nb201nf_ninner]
	mov   [rsp + nb201nf_ninner], ecx
	add   edx, 0