nb_kernel313_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,333 行 · 第 1/5 页

        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret

	
	



.globl nb_kernel313nf_x86_64_sse
.globl _nb_kernel313nf_x86_64_sse
nb_kernel313nf_x86_64_sse:	
_nb_kernel313nf_x86_64_sse:	
;#	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 sse 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_rinvH1,         256
.equiv          nb313nf_rinvH2,         272
.equiv          nb313nf_rinvM,          288
.equiv          nb313nf_rH1,            304
.equiv          nb313nf_rH2,            320
.equiv          nb313nf_rM,             336
.equiv          nb313nf_tsc,            352
.equiv          nb313nf_two,            368
.equiv          nb313nf_c6,             384
.equiv          nb313nf_c12,            400
.equiv          nb313nf_vctot,          416
.equiv          nb313nf_Vvdwtot,        432
.equiv          nb313nf_half,           448
.equiv          nb313nf_three,          464
.equiv          nb313nf_is3,            480
.equiv          nb313nf_ii3,            484
.equiv          nb313nf_nri,            488
.equiv          nb313nf_iinr,           496
.equiv          nb313nf_jindex,         504
.equiv          nb313nf_jjnr,           512
.equiv          nb313nf_shift,          520
.equiv          nb313nf_shiftvec,       528
.equiv          nb313nf_facel,          536
.equiv          nb313nf_innerjjnr,      544
.equiv          nb313nf_ntia,           552
.equiv          nb313nf_innerk,         556
.equiv          nb313nf_n,              560
.equiv          nb313nf_nn1,            564
.equiv          nb313nf_nouter,         568
.equiv          nb313nf_ninner,         572
	push rbp
	mov  rbp, rsp
	push rbx

	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 584		;# 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]
	movss xmm0, [rsi]
	movss [rsp + nb313nf_facel], xmm0

	mov rax, [rbp + nb313nf_p_tabscale]
	movss xmm3, [rax]
	shufps xmm3, xmm3, 0
	movaps [rsp + nb313nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# half in IEEE (hex)
	mov [rsp + nb313nf_half], eax
	movss xmm1, [rsp + nb313nf_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 + nb313nf_half],  xmm1
	movaps [rsp + nb313nf_two],  xmm2
	movaps [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]
	movss xmm4, [rdx + rbx*4 + 4]	
	movss xmm3, [rdx + rbx*4 + 12]	
	mov rsi, [rbp + nb313nf_p_facel]
	movss xmm0, [rsi]
	movss xmm5, [rsp + nb313nf_facel]
	mulss  xmm3, xmm5
	mulss  xmm4, xmm5

	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	movaps [rsp + nb313nf_iqM], xmm3
	movaps [rsp + nb313nf_iqH], 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[] 

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

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

	movaps xmm3, xmm0
	movaps xmm4, xmm1
	movaps xmm5, xmm2	
	movaps xmm6, xmm0
	movaps 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

	addss xmm3, [rax + rbx*4]  	;# ox
	addss xmm4, [rax + rbx*4 + 4]  ;# oy
	addss xmm5, [rax + rbx*4 + 8]  ;# oz
	addss xmm6, [rax + rbx*4 + 12] ;# h1x
	addss xmm7, [rax + rbx*4 + 16] ;# h1y
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	shufps xmm6, xmm6, 0
	shufps xmm7, xmm7, 0
	movaps [rsp + nb313nf_ixO], xmm3
	movaps [rsp + nb313nf_iyO], xmm4
	movaps [rsp + nb313nf_izO], xmm5
	movaps [rsp + nb313nf_ixH1], xmm6
	movaps [rsp + nb313nf_iyH1], xmm7

	movss xmm6, xmm2
	movss xmm3, xmm0
	movss xmm4, xmm1
	movss xmm5, xmm2
	addss xmm6, [rax + rbx*4 + 20] ;# h1z
	addss xmm0, [rax + rbx*4 + 24] ;# h2x
	addss xmm1, [rax + rbx*4 + 28] ;# h2y
	addss xmm2, [rax + rbx*4 + 32] ;# h2z
	addss xmm3, [rax + rbx*4 + 36] ;# mx
	addss xmm4, [rax + rbx*4 + 40] ;# my
	addss xmm5, [rax + rbx*4 + 44] ;# mz

	shufps xmm6, xmm6, 0
	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 + nb313nf_izH1], xmm6
	movaps [rsp + nb313nf_ixH2], xmm0
	movaps [rsp + nb313nf_iyH2], xmm1
	movaps [rsp + nb313nf_izH2], xmm2
	movaps [rsp + nb313nf_ixM], xmm3
	movaps [rsp + nb313nf_iyM], xmm4
	movaps [rsp + nb313nf_izM], xmm5
	
	;# clear vctot
	xorps xmm4, xmm4
	movaps [rsp + nb313nf_vctot], xmm4
	movaps [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,  4
	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_odd_inner
.nb313nf_unroll_loop:
	;# quad-unroll innerloop here 
	mov   rdx, [rsp + nb313nf_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 + nb313nf_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	mov rsi, [rbp + nb313nf_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 + nb313nf_iqM]
	mulps  xmm4, [rsp + nb313nf_iqH]

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

	movaps  [rsp + nb313nf_qqM], xmm3
	movaps  [rsp + nb313nf_qqH], xmm4
	
	mov rsi, [rbp + nb313nf_type]
	mov eax, [rsi + rax*4]
	mov ebx, [rsi + rbx*4]
	mov ecx, [rsi + rcx*4]
	mov edx, [rsi + rdx*4]
	mov rsi, [rbp + nb313nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	shl ecx, 1	
	shl edx, 1	
	mov edi, [rsp + nb313nf_ntia]
	add eax, edi
	add ebx, edi
	add ecx, edi
	add edx, edi

	movlps xmm6, [rsi + rax*4]
	movlps xmm7, [rsi + rcx*4]
	movhps xmm6, [rsi + rbx*4]
	movhps xmm7, [rsi + rdx*4]

	movaps xmm4, xmm6
	shufps xmm4, xmm7, 136  ;# 10001000
	shufps xmm6, xmm7, 221  ;# 11011101
	
	movd  eax, mm0		
	movd  ebx, mm1
	movd  ecx, mm2
	movd  edx, mm3

	movaps [rsp + nb313nf_c6], xmm4
	movaps [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]	
	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]

	movhps xmm4, [rsi + rbx*4]
	movhps xmm5, [rsi + rdx*4]

	movss xmm0, [rsi + rbx*4 + 8]
	movss xmm1, [rsi + rdx*4 + 8]

	shufps xmm2, xmm0, 0
	shufps xmm6, xmm1, 0
	
	movaps xmm0, xmm4
	movaps xmm1, xmm4

	shufps xmm2, xmm6, 136  ;# 10001000
	
	shufps xmm0, xmm5, 136  ;# 10001000
	shufps xmm1, xmm5, 221  ;# 11011101		

	;# move ixO-izO to xmm4-xmm6 
	movaps xmm4, [rsp + nb313nf_ixO]
	movaps xmm5, [rsp + nb313nf_iyO]
	movaps xmm6, [rsp + nb313nf_izO]

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

	;# square it 
	mulps xmm4,xmm4
	mulps xmm5,xmm5
	mulps xmm6,xmm6
	addps xmm4, xmm5
	addps xmm4, xmm6
	movaps xmm7, xmm4
	;# rsqO in xmm7 

	;# move ixH1-izH1 to xmm4-xmm6 
	movaps xmm4, [rsp + nb313nf_ixH1]
	movaps xmm5, [rsp + nb313nf_iyH1]
	movaps xmm6, [rsp + nb313nf_izH1]

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

	;# square it 
	mulps xmm4,xmm4
	mulps xmm5,xmm5
	mulps xmm6,xmm6
	addps xmm6, xmm5
	addps xmm6, xmm4
	;# rsqH1 in xmm6 

	;# move ixH2-izH2 to xmm3-xmm5  
	movaps xmm3, [rsp + nb313nf_ixH2]
	movaps xmm4, [rsp + nb313nf_iyH2]
	movaps xmm5, [rsp + nb313nf_izH2]

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

	;# square it 
	mulps xmm3,xmm3
	mulps xmm4,xmm4
	mulps xmm5,xmm5
	addps xmm5, xmm4
	addps xmm5, xmm3
	
	;# move ixM-izM to xmm2-xmm4  
	movaps xmm3, [rsp + nb313nf_iyM]
	movaps xmm4, [rsp + nb313nf_izM]
	subps  xmm3, xmm1
	subps  xmm4, xmm2
	movaps xmm2, [rsp + nb313nf_ixM]
	subps  xmm2, xmm0	

	;# square it 
	mulps xmm2,xmm2
	mulps xmm3,xmm3
	mulps xmm4,xmm4
	addps xmm4, xmm3
	addps xmm4, xmm2	
	;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

	;# rsqH1 - seed in xmm2 
	rsqrtps xmm2, xmm6
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm0, [rsp + nb313nf_three]
	mulps   xmm2, xmm6	;# rsq*lu*lu 
	subps   xmm0, xmm2	;# 30-rsq*lu*lu 
	mulps   xmm0, xmm3	;# lu*(3-rsq*lu*lu) 
	mulps   xmm0, [rsp + nb313nf_half]
	movaps  [rsp + nb313nf_rinvH1], xmm0	;# rinvH1 in xmm4 
	mulps   xmm6, xmm0
	movaps  [rsp + nb313nf_rH1], xmm6

	;# rsqH2 - seed to xmm2 
	rsqrtps xmm2, xmm5