nb_kernel214_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,201 行 · 第 1/5 页

	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 Mi forces in xmm0, xmm1, xmm2 
	movaps xmm0, [rsp + nb214_fixM]
	movaps xmm1, [rsp + nb214_fiyM]
	movaps xmm2, [rsp + nb214_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 + nb214_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb214_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movaps xmm7, [rsp + nb214_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 + nb214_Vc]
	addss xmm7, [rax + rdx*4] 
	;# move back to mem 
	movss [rax + rdx*4], xmm7 
	
	;# accumulate total lj energy and update it 
	movaps xmm7, [rsp + nb214_Vvdwtot]
	;# 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 + nb214_Vvdw]
	addss xmm7, [rax + rdx*4] 
	;# move back to mem 
	movss [rax + rdx*4], xmm7 
	
        ;# finish if last 
        mov ecx, [rsp + nb214_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb214_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb214_n], esi
        jmp .nb214_outer
.nb214_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb214_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb214_end
        ;# non-zero, do one more workunit
        jmp   .nb214_threadloop
.nb214_end:
	mov eax, [rsp + nb214_nouter]
	mov ebx, [rsp + nb214_ninner]
	mov rcx, [rbp + nb214_outeriter]
	mov rdx, [rbp + nb214_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 1912
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret


	
.globl nb_kernel214nf_x86_64_sse
.globl _nb_kernel214nf_x86_64_sse
nb_kernel214nf_x86_64_sse:	
_nb_kernel214nf_x86_64_sse:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb214nf_fshift,         16
.equiv          nb214nf_gid,            24
.equiv          nb214nf_pos,            32
.equiv          nb214nf_faction,        40
.equiv          nb214nf_charge,         48
.equiv          nb214nf_p_facel,        56
.equiv          nb214nf_argkrf,         64
.equiv          nb214nf_argcrf,         72
.equiv          nb214nf_Vc,             80
.equiv          nb214nf_type,           88
.equiv          nb214nf_p_ntype,        96
.equiv          nb214nf_vdwparam,       104
.equiv          nb214nf_Vvdw,           112
.equiv          nb214nf_p_tabscale,     120
.equiv          nb214nf_VFtab,          128
.equiv          nb214nf_invsqrta,       136
.equiv          nb214nf_dvda,           144
.equiv          nb214nf_p_gbtabscale,   152
.equiv          nb214nf_GBtab,          160
.equiv          nb214nf_p_nthreads,     168
.equiv          nb214nf_count,          176
.equiv          nb214nf_mtx,            184
.equiv          nb214nf_outeriter,      192
.equiv          nb214nf_inneriter,      200
.equiv          nb214nf_work,           208
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb214nf_ixO,            0
.equiv          nb214nf_iyO,            16
.equiv          nb214nf_izO,            32
.equiv          nb214nf_ixH1,           48
.equiv          nb214nf_iyH1,           64
.equiv          nb214nf_izH1,           80
.equiv          nb214nf_ixH2,           96
.equiv          nb214nf_iyH2,           112
.equiv          nb214nf_izH2,           128
.equiv          nb214nf_ixM,            144
.equiv          nb214nf_iyM,            160
.equiv          nb214nf_izM,            176
.equiv          nb214nf_jxO,            192
.equiv          nb214nf_jyO,            208
.equiv          nb214nf_jzO,            224
.equiv          nb214nf_jxH1,           240
.equiv          nb214nf_jyH1,           256
.equiv          nb214nf_jzH1,           272
.equiv          nb214nf_jxH2,           288
.equiv          nb214nf_jyH2,           304
.equiv          nb214nf_jzH2,           320
.equiv          nb214nf_jxM,            336
.equiv          nb214nf_jyM,            352
.equiv          nb214nf_jzM,            368
.equiv          nb214nf_qqMM,           384
.equiv          nb214nf_qqMH,           400
.equiv          nb214nf_qqHH,           416
.equiv          nb214nf_two,            432
.equiv          nb214nf_c6,             448
.equiv          nb214nf_c12,            464
.equiv          nb214nf_vctot,          480
.equiv          nb214nf_Vvdwtot,        496
.equiv          nb214nf_half,           512
.equiv          nb214nf_three,          528
.equiv          nb214nf_rsqOO,          544
.equiv          nb214nf_rsqH1H1,        560
.equiv          nb214nf_rsqH1H2,        576
.equiv          nb214nf_rsqH1M,         592
.equiv          nb214nf_rsqH2H1,        608
.equiv          nb214nf_rsqH2H2,        624
.equiv          nb214nf_rsqH2M,         640
.equiv          nb214nf_rsqMH1,         656
.equiv          nb214nf_rsqMH2,         672
.equiv          nb214nf_rsqMM,          688
.equiv          nb214nf_rinvsqOO,       704
.equiv          nb214nf_rinvH1H1,       720
.equiv          nb214nf_rinvH1H2,       736
.equiv          nb214nf_rinvH1M,        752
.equiv          nb214nf_rinvH2H1,       768
.equiv          nb214nf_rinvH2H2,       784
.equiv          nb214nf_rinvH2M,        800
.equiv          nb214nf_rinvMH1,        816
.equiv          nb214nf_rinvMH2,        832
.equiv          nb214nf_rinvMM,         848
.equiv          nb214nf_krf,            864
.equiv          nb214nf_crf,            880
.equiv          nb214nf_is3,            896
.equiv          nb214nf_ii3,            900
.equiv          nb214nf_innerjjnr,      904
.equiv          nb214nf_nri,            912
.equiv          nb214nf_iinr,           920
.equiv          nb214nf_jindex,         928
.equiv          nb214nf_jjnr,           936
.equiv          nb214nf_shift,          944
.equiv          nb214nf_shiftvec,       952
.equiv          nb214nf_facel,          960
.equiv          nb214nf_innerk,         968
.equiv          nb214nf_n,              972
.equiv          nb214nf_nn1,            976
.equiv          nb214nf_nouter,         980
.equiv          nb214nf_ninner,         984


	push rbp
	mov  rbp, rsp
	push rbx
	
	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 1000		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb214nf_nouter], eax
	mov [rsp + nb214nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb214nf_nri], edi
	mov [rsp + nb214nf_iinr], rsi
	mov [rsp + nb214nf_jindex], rdx
	mov [rsp + nb214nf_jjnr], rcx
	mov [rsp + nb214nf_shift], r8
	mov [rsp + nb214nf_shiftvec], r9
	mov rsi, [rbp + nb214nf_p_facel]
	movss xmm0, [rsi]
	movss [rsp + nb214nf_facel], xmm0


	mov rsi, [rbp + nb214nf_argkrf]
	mov rdi, [rbp + nb214nf_argcrf]
	movss xmm1, [rsi]
	movss xmm2, [rdi]
	shufps xmm1, xmm1, 0
	shufps xmm2, xmm2, 0
	movaps [rsp + nb214nf_krf], xmm1
	movaps [rsp + nb214nf_crf], xmm2

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# half in IEEE (hex)
	mov [rsp + nb214nf_half], eax
	movss xmm1, [rsp + nb214nf_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 + nb214nf_half],  xmm1
	movaps [rsp + nb214nf_two],  xmm2
	movaps [rsp + nb214nf_three],  xmm3

	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb214nf_iinr]   ;# rcx = pointer into iinr[]
	mov   ebx, [rcx]		;# ebx =ii 

	mov   rdx, [rbp + nb214nf_charge]
	movss xmm5, [rdx + rbx*4 + 4]	
	movss xmm3, [rdx + rbx*4 + 12]	
	movss xmm4, xmm3	
	mov rsi, [rbp + nb214nf_p_facel]
	movss xmm0, [rsi]
	movss xmm6, [rsp + nb214nf_facel]
	mulss  xmm3, xmm3
	mulss  xmm4, xmm5
	mulss  xmm5, xmm5
	mulss  xmm3, xmm6
	mulss  xmm4, xmm6
	mulss  xmm5, xmm6
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	movaps [rsp + nb214nf_qqMM], xmm3
	movaps [rsp + nb214nf_qqMH], xmm4
	movaps [rsp + nb214nf_qqHH], xmm5
	
	xorps xmm0, xmm0
	mov   rdx, [rbp + nb214nf_type]
	mov   ecx, [rdx + rbx*4]
	shl   ecx, 1
	mov   edx, ecx
	mov rdi, [rbp + nb214nf_p_ntype]
	imul  ecx, [rdi]  ;# rcx = ntia = 2*ntype*type[ii0] 
	add   rdx, rcx
	mov   rax, [rbp + nb214nf_vdwparam]
	movlps xmm0, [rax + rdx*4] 
	movaps xmm1, xmm0
	shufps xmm0, xmm0, 0
	shufps xmm1, xmm1, 0x55
	movaps [rsp + nb214nf_c6], xmm0
	movaps [rsp + nb214nf_c12], xmm1

.nb214nf_threadloop:
        mov   rsi, [rbp + nb214nf_count]          ;# pointer to sync counter
        mov   eax, [rsi]
.nb214nf_spinlock:
        mov   ebx, eax                          ;# ebx=*count=nn0
        add   rbx, 1                           ;# rbx=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 .nb214nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb214nf_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 + nb214nf_n], eax
        mov [rsp + nb214nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb214nf_outerstart
        jmp .nb214nf_end
	
.nb214nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb214nf_nouter]
	mov [rsp + nb214nf_nouter], ebx

.nb214nf_outer:
	mov   rax, [rsp + nb214nf_shift]  	;# rax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]	;# rbx=3*is 
	mov   [rsp + nb214nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb214nf_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 + nb214nf_iinr]   	;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx + rsi*4]		;# ebx =ii 

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb214nf_pos]	;# rax = base of pos[]  
	mov   [rsp + nb214nf_ii3], ebx	
	
	movaps xmm3, xmm0
	movaps xmm4, xmm1
	movaps xmm5, xmm2
	movaps xmm6, xmm0
	movaps xmm7, xmm1

	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 + nb214nf_ixO], xmm3
	movaps [rsp + nb214nf_iyO], xmm4
	movaps [rsp + nb214nf_izO], xmm5
	movaps [rsp + nb214nf_ixH1], xmm6
	movaps [rsp + nb214nf_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