nb_kernel132_x86_64_sse2.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,328 行 · 第 1/5 页

	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 + nb132_fixH1]
	movapd xmm1, [rsp + nb132_fiyH1]
	movapd xmm2, [rsp + nb132_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 + nb132_fixH2]
	movapd xmm1, [rsp + nb132_fiyH2]
	movapd xmm2, [rsp + nb132_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 + nb132_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb132_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movapd xmm7, [rsp + nb132_vctot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 

	;# add earlier value from mem 
	mov   rax, [rbp + nb132_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 + nb132_Vvdwtot]
	;# accumulate 
	movhlps xmm6, xmm7
	addsd  xmm7, xmm6	;# low xmm7 has the sum now 
	
	;# add earlier value from mem 
	mov   rax, [rbp + nb132_Vvdw]
	addsd xmm7, [rax + rdx*8] 
	;# move back to mem 
	movsd [rax + rdx*8], xmm7 
	
        ;# finish if last 
        mov ecx, [rsp + nb132_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb132_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb132_n], esi
        jmp .nb132_outer
.nb132_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb132_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb132_end
        ;# non-zero, do one more workunit
        jmp   .nb132_threadloop
.nb132_end:
	mov eax, [rsp + nb132_nouter]
	mov ebx, [rsp + nb132_ninner]
	mov rcx, [rbp + nb132_outeriter]
	mov rdx, [rbp + nb132_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 1624
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret



	
.globl nb_kernel132nf_x86_64_sse2
.globl _nb_kernel132nf_x86_64_sse2
nb_kernel132nf_x86_64_sse2:	
_nb_kernel132nf_x86_64_sse2:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb132nf_fshift,         16
.equiv          nb132nf_gid,            24
.equiv          nb132nf_pos,            32
.equiv          nb132nf_faction,        40
.equiv          nb132nf_charge,         48
.equiv          nb132nf_p_facel,        56
.equiv          nb132nf_argkrf,         64
.equiv          nb132nf_argcrf,         72
.equiv          nb132nf_Vc,             80
.equiv          nb132nf_type,           88
.equiv          nb132nf_p_ntype,        96
.equiv          nb132nf_vdwparam,       104
.equiv          nb132nf_Vvdw,           112
.equiv          nb132nf_p_tabscale,     120
.equiv          nb132nf_VFtab,          128
.equiv          nb132nf_invsqrta,       136
.equiv          nb132nf_dvda,           144
.equiv          nb132nf_p_gbtabscale,   152
.equiv          nb132nf_GBtab,          160
.equiv          nb132nf_p_nthreads,     168
.equiv          nb132nf_count,          176
.equiv          nb132nf_mtx,            184
.equiv          nb132nf_outeriter,      192
.equiv          nb132nf_inneriter,      200
.equiv          nb132nf_work,           208
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb132nf_ixO,            0
.equiv          nb132nf_iyO,            16
.equiv          nb132nf_izO,            32
.equiv          nb132nf_ixH1,           48
.equiv          nb132nf_iyH1,           64
.equiv          nb132nf_izH1,           80
.equiv          nb132nf_ixH2,           96
.equiv          nb132nf_iyH2,           112
.equiv          nb132nf_izH2,           128
.equiv          nb132nf_jxO,            144
.equiv          nb132nf_jyO,            160
.equiv          nb132nf_jzO,            176
.equiv          nb132nf_jxH1,           192
.equiv          nb132nf_jyH1,           208
.equiv          nb132nf_jzH1,           224
.equiv          nb132nf_jxH2,           240
.equiv          nb132nf_jyH2,           256
.equiv          nb132nf_jzH2,           272
.equiv          nb132nf_qqOO,           288
.equiv          nb132nf_qqOH,           304
.equiv          nb132nf_qqHH,           320
.equiv          nb132nf_c6,             336
.equiv          nb132nf_c12,            352
.equiv          nb132nf_vctot,          368
.equiv          nb132nf_Vvdwtot,        384
.equiv          nb132nf_half,           400
.equiv          nb132nf_three,          416
.equiv          nb132nf_rsqOO,          432
.equiv          nb132nf_rsqOH1,         448
.equiv          nb132nf_rsqOH2,         464
.equiv          nb132nf_rsqH1O,         480
.equiv          nb132nf_rsqH1H1,        496
.equiv          nb132nf_rsqH1H2,        512
.equiv          nb132nf_rsqH2O,         528
.equiv          nb132nf_rsqH2H1,        544
.equiv          nb132nf_rsqH2H2,        560
.equiv          nb132nf_rinvOO,         576
.equiv          nb132nf_rinvOH1,        592
.equiv          nb132nf_rinvOH2,        608
.equiv          nb132nf_rinvH1O,        624
.equiv          nb132nf_rinvH1H1,       640
.equiv          nb132nf_rinvH1H2,       656
.equiv          nb132nf_rinvH2O,        672
.equiv          nb132nf_rinvH2H1,       688
.equiv          nb132nf_rinvH2H2,       704
.equiv          nb132nf_krf,            720
.equiv          nb132nf_crf,            736
.equiv          nb132nf_tsc,            752
.equiv          nb132nf_nri,            768
.equiv          nb132nf_iinr,           776
.equiv          nb132nf_jindex,         784
.equiv          nb132nf_jjnr,           792
.equiv          nb132nf_shift,          800
.equiv          nb132nf_shiftvec,       808
.equiv          nb132nf_facel,          816
.equiv          nb132nf_innerjjnr,      824
.equiv          nb132nf_is3,            832
.equiv          nb132nf_ii3,            836
.equiv          nb132nf_innerk,         840
.equiv          nb132nf_n,              844
.equiv          nb132nf_nn1,            848
.equiv          nb132nf_nouter,         852
.equiv          nb132nf_ninner,         856
	push rbp
	mov  rbp, rsp
	push rbx

	
	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 872		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb132nf_nouter], eax
	mov [rsp + nb132nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb132nf_nri], edi
	mov [rsp + nb132nf_iinr], rsi
	mov [rsp + nb132nf_jindex], rdx
	mov [rsp + nb132nf_jjnr], rcx
	mov [rsp + nb132nf_shift], r8
	mov [rsp + nb132nf_shiftvec], r9
	mov rsi, [rbp + nb132nf_p_facel]
	movsd xmm0, [rsi]
	movsd [rsp + nb132nf_facel], xmm0

	mov rax, [rbp + nb132nf_p_tabscale]
	movsd xmm3, [rax]
	shufpd xmm3, xmm3, 0
	movapd [rsp + nb132nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x00000000     ;# lower half of double half IEEE (hex)
	mov ebx, 0x3fe00000
	mov [rsp + nb132nf_half], eax
	mov [rsp + nb132nf_half+4], ebx
	movsd xmm1, [rsp + nb132nf_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 + nb132nf_half], xmm1
	movapd [rsp + nb132nf_three], xmm3

	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb132nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb132nf_charge]
	movsd xmm3, [rdx + rbx*8]	
	movsd xmm4, xmm3	
	movsd xmm5, [rdx + rbx*8 + 8]	

	movsd xmm6, [rsp + nb132nf_facel]
	mulsd  xmm3, xmm3
	mulsd  xmm4, xmm5
	mulsd  xmm5, xmm5
	mulsd  xmm3, xmm6
	mulsd  xmm4, xmm6
	mulsd  xmm5, xmm6
	shufpd xmm3, xmm3, 0
	shufpd xmm4, xmm4, 0
	shufpd xmm5, xmm5, 0
	movapd [rsp + nb132nf_qqOO], xmm3
	movapd [rsp + nb132nf_qqOH], xmm4
	movapd [rsp + nb132nf_qqHH], xmm5
	
	xorpd xmm0, xmm0
	mov   rdx, [rbp + nb132nf_type]
	mov   ecx, [rdx + rbx*4]
	shl   ecx, 1
	mov   edx, ecx
	mov rdi, [rbp + nb132nf_p_ntype]
	imul  ecx, [rdi]      ;# rcx = ntia = 2*ntype*type[ii0] 
	add   edx, ecx
	mov   rax, [rbp + nb132nf_vdwparam]
	movlpd xmm0, [rax + rdx*8] 
	movlpd xmm1, [rax + rdx*8 + 8] 
	shufpd xmm0, xmm0, 0
	shufpd xmm1, xmm1, 0
	movapd [rsp + nb132nf_c6], xmm0
	movapd [rsp + nb132nf_c12], xmm1

.nb132nf_threadloop:
        mov   rsi, [rbp + nb132nf_count]          ;# pointer to sync counter
        mov   eax, [rsi]
.nb132nf_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 .nb132nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb132nf_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 + nb132nf_n], eax
        mov [rsp + nb132nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb132nf_outerstart
        jmp .nb132nf_end

.nb132nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb132nf_nouter]
	mov [rsp + nb132nf_nouter], ebx

.nb132nf_outer:
	mov   rax, [rsp + nb132nf_shift]      ;# eax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# ebx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 
	mov   [rsp + nb132nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb132nf_shiftvec]   ;# eax = base of shiftvec[] 

	movlpd xmm0, [rax + rbx*8]
	movlpd xmm1, [rax + rbx*8 + 8]
	movlpd xmm2, [rax + rbx*8 + 16] 

	mov   rcx, [rsp + nb132nf_iinr]       ;# ecx = pointer into iinr[] 	
	mov   ebx, [rcx + rsi*4]	    ;# ebx =ii 

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb132nf_pos]    ;# eax = base of pos[]  
	mov   [rsp + nb132nf_ii3], ebx	
	
	movapd xmm3, xmm0
	movapd xmm4, xmm1
	movapd xmm5, xmm2
	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 + nb132nf_ixO], xmm3
	movapd [rsp + nb132nf_iyO], xmm4
	movapd [rsp + nb132nf_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 + nb132nf_ixH1], xmm0
	movapd [rsp + nb132nf_iyH1], xmm1
	movapd [rsp + nb132nf_izH1], xmm2
	movapd [rsp + nb132nf_ixH2], xmm3
	movapd [rsp + nb132nf_iyH2], xmm4
	movapd [rsp + nb132nf_izH2], xmm5

	;# clear vctot 
	xorpd xmm4, xmm4
	movapd [rsp + nb132nf_vctot], xmm4
	movapd [rsp + nb132nf_Vvdwtot], xmm4
	
	mov   rax, [rsp + nb132nf_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 + nb132nf_pos]
	mov   rdi, [rbp + nb132nf_faction]	
	mov   rax, [rsp + nb132nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb132nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  2
	add   ecx, [rsp + nb132nf_ninner]
	mov   [rsp + nb132nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb132nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb132nf_unroll_loop
	jmp   .nb132nf_checksingle
.nb132nf_unroll_loop:
	;# twice unrolled innerloop here 
	mov   rdx, [rsp + nb132nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	mov   ebx, [rdx + 4] 
	
	add qword ptr [rsp + nb132nf_innerjjnr],  8	;# advance pointer (unrolled 2) 

	mov rsi, [rbp + nb132nf_pos]       ;# base of pos[] 

	lea   rax, [rax + rax*2]     ;# replace jnr with j3 
	lea   rbx, [rbx + rbx*2]	
	
	;# move j coordinates to local temp variables 
	m