nb_kernel331_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,310 行 · 第 1/5 页

.globl _nb_kernel331nf_x86_64_sse
nb_kernel331nf_x86_64_sse:	
_nb_kernel331nf_x86_64_sse:	
;#	Room for return address and rbp (16 bytes)
.equiv          nb331nf_fshift,         16
.equiv          nb331nf_gid,            24
.equiv          nb331nf_pos,            32
.equiv          nb331nf_faction,        40
.equiv          nb331nf_charge,         48
.equiv          nb331nf_p_facel,        56
.equiv          nb331nf_argkrf,         64
.equiv          nb331nf_argcrf,         72
.equiv          nb331nf_Vc,             80
.equiv          nb331nf_type,           88
.equiv          nb331nf_p_ntype,        96
.equiv          nb331nf_vdwparam,       104
.equiv          nb331nf_Vvdw,           112
.equiv          nb331nf_p_tabscale,     120
.equiv          nb331nf_VFtab,          128
.equiv          nb331nf_invsqrta,       136
.equiv          nb331nf_dvda,           144
.equiv          nb331nf_p_gbtabscale,   152
.equiv          nb331nf_GBtab,          160
.equiv          nb331nf_p_nthreads,     168
.equiv          nb331nf_count,          176
.equiv          nb331nf_mtx,            184
.equiv          nb331nf_outeriter,      192
.equiv          nb331nf_inneriter,      200
.equiv          nb331nf_work,           208
	;# stack offsets for local variables  
	;# bottom of stack is cache-aligned for sse use 
.equiv          nb331nf_ixO,            0
.equiv          nb331nf_iyO,            16
.equiv          nb331nf_izO,            32
.equiv          nb331nf_ixH1,           48
.equiv          nb331nf_iyH1,           64
.equiv          nb331nf_izH1,           80
.equiv          nb331nf_ixH2,           96
.equiv          nb331nf_iyH2,           112
.equiv          nb331nf_izH2,           128
.equiv          nb331nf_iqO,            144
.equiv          nb331nf_iqH,            160
.equiv          nb331nf_qqO,            176
.equiv          nb331nf_qqH,            192
.equiv          nb331nf_rinvO,          208
.equiv          nb331nf_rinvH1,         224
.equiv          nb331nf_rinvH2,         240
.equiv          nb331nf_rO,             256
.equiv          nb331nf_rH1,            272
.equiv          nb331nf_rH2,            288
.equiv          nb331nf_tsc,            304
.equiv          nb331nf_c6,             320
.equiv          nb331nf_c12,            336
.equiv          nb331nf_vctot,          352
.equiv          nb331nf_Vvdwtot,        368
.equiv          nb331nf_half,           384
.equiv          nb331nf_three,          400
.equiv          nb331nf_is3,            416
.equiv          nb331nf_ii3,            420
.equiv          nb331nf_nri,            424
.equiv          nb331nf_iinr,           432
.equiv          nb331nf_jindex,         440
.equiv          nb331nf_jjnr,           448
.equiv          nb331nf_shift,          456
.equiv          nb331nf_shiftvec,       464
.equiv          nb331nf_facel,          472
.equiv          nb331nf_innerjjnr,      480
.equiv          nb331nf_ntia,           488
.equiv          nb331nf_innerk,         492
.equiv          nb331nf_n,              496
.equiv          nb331nf_nn1,            500
.equiv          nb331nf_nouter,         504
.equiv          nb331nf_ninner,         508
	push rbp
	mov  rbp, rsp
	push rbx

	emms

        push r12
        push r13
        push r14
        push r15

	sub rsp, 520		;# local variable stack space (n*16+8)

	;# zero 32-bit iteration counters
	mov eax, 0
	mov [rsp + nb331nf_nouter], eax
	mov [rsp + nb331nf_ninner], eax

	mov edi, [rdi]
	mov [rsp + nb331nf_nri], edi
	mov [rsp + nb331nf_iinr], rsi
	mov [rsp + nb331nf_jindex], rdx
	mov [rsp + nb331nf_jjnr], rcx
	mov [rsp + nb331nf_shift], r8
	mov [rsp + nb331nf_shiftvec], r9
	mov rsi, [rbp + nb331nf_p_facel]
	movss xmm0, [rsi]
	movss [rsp + nb331nf_facel], xmm0

	mov rax, [rbp + nb331nf_p_tabscale]
	movss xmm3, [rax]
	shufps xmm3, xmm3, 0
	movaps [rsp + nb331nf_tsc], xmm3

	;# create constant floating-point factors on stack
	mov eax, 0x3f000000     ;# half in IEEE (hex)
	mov [rsp + nb331nf_half], eax
	movss xmm1, [rsp + nb331nf_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 + nb331nf_half],  xmm1
	movaps [rsp + nb331nf_three],  xmm3
	
	;# assume we have at least one i particle - start directly 
	mov   rcx, [rsp + nb331nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx]	    ;# ebx =ii 

	mov   rdx, [rbp + nb331nf_charge]
	movss xmm3, [rdx + rbx*4]	
	movss xmm4, [rdx + rbx*4 + 4]	
	mov rsi, [rbp + nb331nf_p_facel]
	movss xmm0, [rsi]
	movss xmm5, [rsp + nb331nf_facel]
	mulss  xmm3, xmm5
	mulss  xmm4, xmm5

	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	movaps [rsp + nb331nf_iqO], xmm3
	movaps [rsp + nb331nf_iqH], xmm4
	
	mov   rdx, [rbp + nb331nf_type]
	mov   ecx, [rdx + rbx*4]
	shl   ecx, 1
	mov rdi, [rbp + nb331nf_p_ntype]
	imul  ecx, [rdi]      ;# rcx = ntia = 2*ntype*type[ii0] 
	mov   [rsp + nb331nf_ntia], ecx		

.nb331nf_threadloop:
        mov   rsi, [rbp + nb331nf_count]          ;# pointer to sync counter
        mov   eax, [rsi]
.nb331nf_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 .nb331nf_spinlock

        ;# if(nn1>nri) nn1=nri
        mov ecx, [rsp + nb331nf_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 + nb331nf_n], eax
        mov [rsp + nb331nf_nn1], ebx
        sub ebx, eax                            ;# calc number of outer lists
	mov esi, eax				;# copy n to esi
        jg  .nb331nf_outerstart
        jmp .nb331nf_end
	
.nb331nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb331nf_nouter]
	mov [rsp + nb331nf_nouter], ebx

.nb331nf_outer:
	mov   rax, [rsp + nb331nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 
	mov   [rsp + nb331nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb331nf_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 + nb331nf_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 + nb331nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb331nf_ii3], ebx

	addss xmm3, [rax + rbx*4]
	addss xmm4, [rax + rbx*4 + 4]
	addss xmm5, [rax + rbx*4 + 8]		
	shufps xmm3, xmm3, 0
	shufps xmm4, xmm4, 0
	shufps xmm5, xmm5, 0
	movaps [rsp + nb331nf_ixO], xmm3
	movaps [rsp + nb331nf_iyO], xmm4
	movaps [rsp + nb331nf_izO], xmm5

	movss xmm3, xmm0
	movss xmm4, xmm1
	movss xmm5, xmm2
	addss xmm0, [rax + rbx*4 + 12]
	addss xmm1, [rax + rbx*4 + 16]
	addss xmm2, [rax + rbx*4 + 20]		
	addss xmm3, [rax + rbx*4 + 24]
	addss xmm4, [rax + rbx*4 + 28]
	addss xmm5, [rax + rbx*4 + 32]		

	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 + nb331nf_ixH1], xmm0
	movaps [rsp + nb331nf_iyH1], xmm1
	movaps [rsp + nb331nf_izH1], xmm2
	movaps [rsp + nb331nf_ixH2], xmm3
	movaps [rsp + nb331nf_iyH2], xmm4
	movaps [rsp + nb331nf_izH2], xmm5
	
	;# clear vctot and i forces 
	xorps xmm4, xmm4
	movaps [rsp + nb331nf_vctot], xmm4
	movaps [rsp + nb331nf_Vvdwtot], xmm4
	
	mov   rax, [rsp + nb331nf_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 + nb331nf_pos]
	mov   rax, [rsp + nb331nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb331nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  4
	add   ecx, [rsp + nb331nf_ninner]
	mov   [rsp + nb331nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb331nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb331nf_unroll_loop
	jmp   .nb331nf_odd_inner
.nb331nf_unroll_loop:
	;# quad-unroll innerloop here 
	mov   rdx, [rsp + nb331nf_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 + nb331nf_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	mov rsi, [rbp + nb331nf_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 + nb331nf_iqO]
	mulps  xmm4, [rsp + nb331nf_iqH]

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

	movaps  [rsp + nb331nf_qqO], xmm3
	movaps  [rsp + nb331nf_qqH], xmm4
	
	mov rsi, [rbp + nb331nf_type]
	mov eax, [rsi + rax*4]
	mov ebx, [rsi + rbx*4]
	mov ecx, [rsi + rcx*4]
	mov edx, [rsi + rdx*4]
	mov rsi, [rbp + nb331nf_vdwparam]
	shl eax, 1	
	shl ebx, 1	
	shl ecx, 1	
	shl edx, 1	
	mov edi, [rsp + nb331nf_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 + nb331nf_c6], xmm4
	movaps [rsp + nb331nf_c12], xmm6

	mov rsi, [rbp + nb331nf_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 + nb331nf_ixO]
	movaps xmm5, [rsp + nb331nf_iyO]
	movaps xmm6, [rsp + nb331nf_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 + nb331nf_ixH1]
	movaps xmm5, [rsp + nb331nf_iyH1]
	movaps xmm6, [rsp + nb331nf_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 + nb331nf_ixH2]
	movaps xmm4, [rsp + nb331nf_iyH2]
	movaps xmm5, [rsp + nb331nf_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
	;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

	;# start with rsqO - seed to xmm2 	
	rsqrtps xmm2, xmm7
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm4, [rsp + nb331nf_three]
	mulps   xmm2, xmm7	;# rsq*lu*lu 
	subps   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulps   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulps   xmm4, [rsp + nb331nf_half]
	movaps  [rsp + nb331nf_rinvO], xmm4	;# rinvO in xmm4 
	mulps   xmm7, xmm4
	movaps  [rsp + nb331nf_rO], xmm7	

	;# rsqH1 - seed in xmm2 
	rsqrtps xmm2, xmm6
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm4, [rsp + nb331nf_three]
	mulps   xmm2, xmm6	;# rsq*lu*lu 
	subps   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulps   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulps   xmm4, [rsp + nb331nf_half]
	movaps  [rsp + nb331nf_rinvH1], xmm4	;# rinvH1 in xmm4 
	mulps   xmm6, xmm4
	movaps  [rsp + nb331nf_rH1], xmm6

	;# rsqH2 - seed to xmm2 
	rsqrtps xmm2, xmm5
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm4, [rsp + nb331nf_three]
	mulps   xmm2, xmm5	;# rsq*lu*lu 
	subps   xmm4, xmm2	;# 30-rsq*lu*lu 
	mulps   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 
	mulps   xmm4, [rsp + nb331nf_half]
	movaps  [rsp + nb331nf_rinvH2], xmm4	;# rinvH2 in xmm4 
	mulps   xmm5, xmm4
	movaps  [rsp + nb331nf_rH2], xmm5

	;# do O interactions 
	;# rO is still in xmm7 
	mulps   xmm7, [rsp + nb331nf_tsc]
	movhlps xmm4, xmm7
	cvttps2pi mm6, xmm7
	cvttps2pi mm7, xmm4    ;# mm6/mm7 contain lu indices 
	
    cvtpi2ps xmm3, mm6
    cvtpi2ps xmm4, mm7
    movlhps xmm3, xmm4
	
    subps xmm7, xmm3

	movaps xmm1, xmm7	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2	;# xmm2=eps2