nb_kernel303_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,943 行 · 第 1/4 页

	mov   rsi, [rbp + nb303_fshift]
	mov   edx, [rsp + nb303_is3]

	;# accumulate  H1 i forces in xmm0, xmm1, xmm2 
	movaps xmm0, [rsp + nb303_fixH1]
	movaps xmm1, [rsp + nb303_fiyH1]
	movaps xmm2, [rsp + nb303_fizH1]

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addps  xmm0, xmm3
	addps  xmm1, xmm4
	addps  xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 

	movaps xmm3, xmm0	
	movaps xmm4, xmm1	
	movaps xmm5, xmm2	

	shufps xmm3, xmm3, 1
	shufps xmm4, xmm4, 1
	shufps xmm5, xmm5, 1
	addss  xmm0, xmm3
	addss  xmm1, xmm4
	addss  xmm2, xmm5	;# xmm0-xmm2 has single force in pos0 

	;# increment i force 
	movss  xmm3, [rdi + rcx*4 + 12]
	movss  xmm4, [rdi + rcx*4 + 16]
	movss  xmm5, [rdi + rcx*4 + 20]
	subss  xmm3, xmm0
	subss  xmm4, xmm1
	subss  xmm5, xmm2
	movss  [rdi + rcx*4 + 12], xmm3
	movss  [rdi + rcx*4 + 16], xmm4
	movss  [rdi + rcx*4 + 20], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	movaps xmm6, xmm0
	movss xmm7, xmm2
	movlhps xmm6, xmm1
	shufps  xmm6, xmm6, 8 ;# 00001000	

	;# accumulate H2i forces in xmm0, xmm1, xmm2 
	movaps xmm0, [rsp + nb303_fixH2]
	movaps xmm1, [rsp + nb303_fiyH2]
	movaps xmm2, [rsp + nb303_fizH2]

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addps  xmm0, xmm3
	addps  xmm1, xmm4
	addps  xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 

	movaps xmm3, xmm0	
	movaps xmm4, xmm1	
	movaps xmm5, xmm2	

	shufps xmm3, xmm3, 1
	shufps xmm4, xmm4, 1
	shufps xmm5, xmm5, 1
	addss  xmm0, xmm3
	addss  xmm1, xmm4
	addss  xmm2, xmm5	;# xmm0-xmm2 has single force in pos0 

	;# increment i force 
	movss  xmm3, [rdi + rcx*4 + 24]
	movss  xmm4, [rdi + rcx*4 + 28]
	movss  xmm5, [rdi + rcx*4 + 32]
	subss  xmm3, xmm0
	subss  xmm4, xmm1
	subss  xmm5, xmm2
	movss  [rdi + rcx*4 + 24], xmm3
	movss  [rdi + rcx*4 + 28], xmm4
	movss  [rdi + rcx*4 + 32], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	addss xmm7, xmm2
	movlhps xmm0, xmm1
	shufps  xmm0, xmm0, 8 ;# 00001000	
	addps   xmm6, xmm0

	;# accumulate M i forces in xmm0, xmm1, xmm2 
	movaps xmm0, [rsp + nb303_fixM]
	movaps xmm1, [rsp + nb303_fiyM]
	movaps xmm2, [rsp + nb303_fizM]

	movhlps xmm3, xmm0
	movhlps xmm4, xmm1
	movhlps xmm5, xmm2
	addps  xmm0, xmm3
	addps  xmm1, xmm4
	addps  xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 

	movaps xmm3, xmm0	
	movaps xmm4, xmm1	
	movaps xmm5, xmm2	

	shufps xmm3, xmm3, 1
	shufps xmm4, xmm4, 1
	shufps xmm5, xmm5, 1
	addss  xmm0, xmm3
	addss  xmm1, xmm4
	addss  xmm2, xmm5	;# xmm0-xmm2 has single force in pos0 

	;# increment i force 
	movss  xmm3, [rdi + rcx*4 + 36]
	movss  xmm4, [rdi + rcx*4 + 40]
	movss  xmm5, [rdi + rcx*4 + 44]
	subss  xmm3, xmm0
	subss  xmm4, xmm1
	subss  xmm5, xmm2
	movss  [rdi + rcx*4 + 36], xmm3
	movss  [rdi + rcx*4 + 40], xmm4
	movss  [rdi + rcx*4 + 44], xmm5

	;# accumulate force in xmm6/xmm7 for fshift 
	addss xmm7, xmm2
	movlhps xmm0, xmm1
	shufps  xmm0, xmm0, 8 ;# 00001000	
	addps   xmm6, xmm0

	;# increment fshift force  
	movlps  xmm3, [rsi + rdx*4]
	movss  xmm4, [rsi + rdx*4 + 8]
	subps  xmm3, xmm6
	subss  xmm4, xmm7
	movlps  [rsi + rdx*4],    xmm3
	movss  [rsi + rdx*4 + 8], xmm4

	;# get n from stack
	mov esi, [rsp + nb303_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb303_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movaps xmm7, [rsp + nb303_vctot]
	;# accumulate 
	movhlps xmm6, xmm7
	addps  xmm7, xmm6	;# pos 0-1 in xmm7 have the sum now 
	movaps xmm6, xmm7
	shufps xmm6, xmm6, 1
	addss  xmm7, xmm6		
        
	;# add earlier value from mem 
	mov   rax, [rbp + nb303_Vc]
	addss xmm7, [rax + rdx*4] 
	;# move back to mem 
	movss [rax + rdx*4], xmm7 

        ;# finish if last 
        mov ecx, [rsp + nb303_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb303_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb303_n], esi
        jmp .nb303_outer
.nb303_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb303_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb303_end
        ;# non-zero, do one more workunit
        jmp   .nb303_threadloop
.nb303_end:
	mov eax, [rsp + nb303_nouter]
	mov ebx, [rsp + nb303_ninner]
	mov rcx, [rbp + nb303_outeriter]
	mov rdx, [rbp + nb303_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 824
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret

	


.globl nb_kernel303nf_x86_64_sse
.globl _nb_kernel303nf_x86_64_sse
nb_kernel303nf_x86_64_sse:	
_nb_kernel303nf_x86_64_sse:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb303nf_fshift,         16
.equiv          nb303nf_gid,            24
.equiv          nb303nf_pos,            32
.equiv          nb303nf_faction,        40
.equiv          nb303nf_charge,         48
.equiv          nb303nf_p_facel,        56
.equiv          nb303nf_argkrf,         64
.equiv          nb303nf_argcrf,         72
.equiv          nb303nf_Vc,             80
.equiv          nb303nf_type,           88
.equiv          nb303nf_p_ntype,        96
.equiv          nb303nf_vdwparam,       104
.equiv          nb303nf_Vvdw,           112
.equiv          nb303nf_p_tabscale,     120
.equiv          nb303nf_VFtab,          128
.equiv          nb303nf_invsqrta,       136
.equiv          nb303nf_dvda,           144
.equiv          nb303nf_p_gbtabscale,   152
.equiv          nb303nf_GBtab,          160
.equiv          nb303nf_p_nthreads,     168
.equiv          nb303nf_count,          176
.equiv          nb303nf_mtx,            184
.equiv          nb303nf_outeriter,      192
.equiv          nb303nf_inneriter,      200
.equiv          nb303nf_work,           208
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb303nf_ixH1,           0
.equiv          nb303nf_iyH1,           16
.equiv          nb303nf_izH1,           32
.equiv          nb303nf_ixH2,           48
.equiv          nb303nf_iyH2,           64
.equiv          nb303nf_izH2,           80
.equiv          nb303nf_ixM,            96
.equiv          nb303nf_iyM,            112
.equiv          nb303nf_izM,            128
.equiv          nb303nf_iqH,            144
.equiv          nb303nf_iqM,            160
.equiv          nb303nf_qqH,            176
.equiv          nb303nf_qqM,            192
.equiv          nb303nf_rinvH1,         208
.equiv          nb303nf_rinvH2,         224
.equiv          nb303nf_rinvM,          240
.equiv          nb303nf_rH1,            256
.equiv          nb303nf_rH2,            272
.equiv          nb303nf_rM,             288
.equiv          nb303nf_tsc,            304
.equiv          nb303nf_vctot,          320
.equiv          nb303nf_half,           336
.equiv          nb303nf_three,          352
.equiv          nb303nf_is3,            368
.equiv          nb303nf_ii3,            372
.equiv          nb303nf_nri,            376
.equiv          nb303nf_iinr,           384
.equiv          nb303nf_jindex,         392
.equiv          nb303nf_jjnr,           400
.equiv          nb303nf_shift,          408
.equiv          nb303nf_shiftvec,       416
.equiv          nb303nf_facel,          424
.equiv          nb303nf_innerjjnr,      432
.equiv          nb303nf_innerk,         440
.equiv          nb303nf_n,              444
.equiv          nb303nf_nn1,            448
.equiv          nb303nf_nouter,         452
.equiv          nb303nf_ninner,         456
	push rbp
	mov  rbp, rsp
	push rbx

	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 472		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb303nf_nouter], eax
	mov [rsp + nb303nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb303nf_nri], edi
	mov [rsp + nb303nf_iinr], rsi
	mov [rsp + nb303nf_jindex], rdx
	mov [rsp + nb303nf_jjnr], rcx
	mov [rsp + nb303nf_shift], r8
	mov [rsp + nb303nf_shiftvec], r9
	mov rsi, [rbp + nb303nf_p_facel]
	movss xmm0, [rsi]
	movss [rsp + nb303nf_facel], xmm0

	mov rax, [rbp + nb303nf_p_tabscale]
	movss xmm3, [rax]
	shufps xmm3, xmm3, 0
	movaps [rsp + nb303nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# half in IEEE (hex)
	mov [rsp + nb303nf_half], eax
	movss xmm1, [rsp + nb303nf_half]
	shufps xmm1, xmm1, 0    ;# splat to all elements
	movaps xmm2, xmm1       
	addps  xmm2, xmm2	;# one
	movaps xmm3, xmm2
	addps  xmm2, xmm2	;# two
	addps  xmm3, xmm2	;# three
	movaps [rsp + nb303nf_half],  xmm1
	movaps [rsp + nb303nf_three],  xmm3
	
	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb303nf_iinr]   	;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]		;# ebx =ii 

	mov   rdx, [rbp + nb303nf_charge]
	movss xmm3, [rdx + rbx*4 + 4]	
	movss xmm4, [rdx + rbx*4 + 12]	
	mov rsi, [rbp + nb303nf_p_facel]
	movss xmm0, [rsi]
	movss xmm5, [rsp + nb303nf_facel]
	mulss  xmm3, xmm5
	mulss  xmm4, xmm5

	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	movaps [rsp + nb303nf_iqH], xmm3
	movaps [rsp + nb303nf_iqM], xmm4
	
.nb303nf_threadloop:
        mov   rsi, [rbp + nb303nf_count]          ;# pointer to sync counter
        mov   eax, [rsi]
.nb303nf_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 .nb303nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb303nf_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 + nb303nf_n], eax
        mov [rsp + nb303nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb303nf_outerstart
        jmp .nb303nf_end
	
.nb303nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb303nf_nouter]
	mov [rsp + nb303nf_nouter], ebx

.nb303nf_outer:
	mov   rax, [rsp + nb303nf_shift]  	;# rax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]	;# rbx=3*is 
	mov   [rsp + nb303nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb303nf_shiftvec]   ;# rax = base of shiftvec[] 

	movss xmm0, [rax + rbx*4]
	movss xmm1, [rax + rbx*4 + 4]
	movss xmm2, [rax + rbx*4 + 8] 

	mov   rcx, [rsp + nb303nf_iinr]   	;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx + rsi*4]		;# ebx =ii 

	movaps xmm3, xmm0
	movaps xmm4, xmm1
	movaps xmm5, xmm2

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb303nf_pos]	;# rax = base of pos[]  
	mov   [rsp + nb303nf_ii3], ebx

	addss xmm3, [rax + rbx*4 + 12]
	addss xmm4, [rax + rbx*4 + 16]
	addss xmm5, [rax + rbx*4 + 20]		
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	movaps [rsp + nb303nf_ixH1], xmm3
	movaps [rsp + nb303nf_iyH1], xmm4
	movaps [rsp + nb303nf_izH1], xmm5

	movss xmm3, xmm0
	movss xmm4, xmm1
	movss xmm5, xmm2
	addss xmm0, [rax + rbx*4 + 24]
	addss xmm1, [rax + rbx*4 + 28]
	addss xmm2, [rax + rbx*4 + 32]		
	addss xmm3, [rax + rbx*4 + 36]
	addss xmm4, [rax + rbx*4 + 40]
	addss xmm5, [rax + rbx*4 + 44]		

	shufps xmm0, xmm0, 0
	shufps xmm1, xmm1, 0
	shufps xmm2, xmm2, 0
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	movaps [rsp + nb303nf_ixH2], xmm0
	movaps [rsp + nb303nf_iyH2], xmm1
	movaps [rsp + nb303nf_izH2], xmm2
	movaps [rsp + nb303nf_ixM], xmm3
	movaps [rsp + nb303nf_iyM], xmm4
	movaps [rsp + nb303nf_izM], xmm5
	
	;# clear vctot 
	xorps xmm4, xmm4
	movaps [rsp + nb303nf_vctot], xmm4
	
	mov   rax, [rsp + nb303nf_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 + nb303nf_pos]
	mov   rdi, [rbp + nb303nf_faction]	
	mov   rax, [rsp + nb303nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb303nf_innerjjnr], rax 	;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  4
	add   ecx, [rsp + nb303nf_ninner]
	mov   [rsp + nb303nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb303nf_innerk], edx	;# number of innerloop atoms 
	jge   .nb303nf_unroll_loop
	jmp   .nb303nf_odd_inner
.nb303nf_unroll_loop:
	;# quad-unroll innerloop here 
	mov   rdx, [rsp + nb303nf_innerjjnr] 	;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	mov   ebx, [rdx + 4]              
	mov   ecx, [rdx + 8]            
	mov   edx, [rdx + 12]     	;# eax-edx=jnr1-4 

	add qword ptr [rsp + nb303nf_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	mov rsi, [rbp + nb303nf_charge]	;# base of charge[] 
	
	movss xmm3, [rsi + rax*4]
	movss xmm4, [rsi + rcx*4]
	movss xmm6, [rsi + rbx*4]
	movss xmm7, [rsi + rdx*4]

	shufps xmm3, xmm6, 0 
	shufps xmm4, xmm7, 0 
	shufps xmm3, xmm4, 136  ;# 10001000 ;# all charges in xmm3  
	movaps xmm4, xmm3	 	;# and in xmm4 
	mulps  xmm3, [rsp + nb303nf_iqH]
	mulps  xmm4, [rsp + nb303nf_iqM]

	movd  mm0, eax		;# use mmx registers as temp storage 
	movd  mm1, ebx
	movd  mm2, ecx
	movd  mm3, edx

	movaps  [rsp + nb303nf_qqH], xmm3
	movaps  [rsp + nb303nf_qqM], xmm4	

	mov rsi, [rbp + nb303nf_pos]   	;# base of pos[] 

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

	;# move four coordinates to xmm0-xmm2 	
	movlps xmm4, [rsi + rax*4]
	movlps xmm5, [rsi + rcx*4]
	movss xmm2, [rsi + rax*4 + 8]
	movss xmm6, [rsi + rcx*4 + 8]