nb_kernel202_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,206 行 · 第 1/5 页

        jmp .nb202nf_end
	
.nb202nf_outerstart:
	;# ebx contains number of outer iterations
	add ebx, [rsp + nb202nf_nouter]
	mov [rsp + nb202nf_nouter], ebx

.nb202nf_outer:
	mov   rax, [rsp + nb202nf_shift]      ;# rax = pointer into shift[] 
	mov   ebx, [rax + rsi*4]		;# rbx=shift[n] 
	
	lea   rbx, [rbx + rbx*2]    ;# rbx=3*is 
	mov   [rsp + nb202nf_is3],ebx    	;# store is3 

	mov   rax, [rsp + nb202nf_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 + nb202nf_iinr]       ;# rcx = pointer into iinr[] 	
	mov   ebx, [rcx + rsi*4]	    ;# ebx =ii 

	lea   rbx, [rbx + rbx*2]	;# rbx = 3*ii=ii3 
	mov   rax, [rbp + nb202nf_pos]    ;# rax = base of pos[]  
	mov   [rsp + nb202nf_ii3], ebx	
	
	movaps xmm3, xmm0
	movaps xmm4, xmm1
	movaps xmm5, xmm2
	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 + nb202nf_ixO], xmm3
	movaps [rsp + nb202nf_iyO], xmm4
	movaps [rsp + nb202nf_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 + nb202nf_ixH1], xmm0
	movaps [rsp + nb202nf_iyH1], xmm1
	movaps [rsp + nb202nf_izH1], xmm2
	movaps [rsp + nb202nf_ixH2], xmm3
	movaps [rsp + nb202nf_iyH2], xmm4
	movaps [rsp + nb202nf_izH2], xmm5

	;# clear vctot and i forces 
	xorps xmm4, xmm4
	movaps [rsp + nb202nf_vctot], xmm4
	
	mov   rax, [rsp + nb202nf_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 + nb202nf_pos]	
	mov   rax, [rsp + nb202nf_jjnr]
	shl   ecx, 2
	add   rax, rcx
	mov   [rsp + nb202nf_innerjjnr], rax     ;# pointer to jjnr[nj0] 
	mov   ecx, edx
	sub   edx,  4
	add   ecx, [rsp + nb202nf_ninner]
	mov   [rsp + nb202nf_ninner], ecx
	add   edx, 0
	mov   [rsp + nb202nf_innerk], edx    ;# number of innerloop atoms 
	jge   .nb202nf_unroll_loop
	jmp   .nb202nf_single_check
.nb202nf_unroll_loop:	
	;# quad-unroll innerloop here 
	mov   rdx, [rsp + nb202nf_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 + nb202nf_innerjjnr],  16 ;# advance pointer (unrolled 4) 

	mov rsi, [rbp + nb202nf_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 j coordinates to local temp variables 
	movlps xmm2, [rsi + rax*4]
	movlps xmm3, [rsi + rax*4 + 12]
	movlps xmm4, [rsi + rax*4 + 24]

	movlps xmm5, [rsi + rbx*4]
	movlps xmm6, [rsi + rbx*4 + 12]
	movlps xmm7, [rsi + rbx*4 + 24]

	movhps xmm2, [rsi + rcx*4]
	movhps xmm3, [rsi + rcx*4 + 12]
	movhps xmm4, [rsi + rcx*4 + 24]

	movhps xmm5, [rsi + rdx*4]
	movhps xmm6, [rsi + rdx*4 + 12]
	movhps xmm7, [rsi + rdx*4 + 24]

	;# current state: 	
	;# xmm2= jxOa  jyOa  jxOc  jyOc 
	;# xmm3= jxH1a jyH1a jxH1c jyH1c 
	;# xmm4= jxH2a jyH2a jxH2c jyH2c 
	;# xmm5= jxOb  jyOb  jxOd  jyOd 
	;# xmm6= jxH1b jyH1b jxH1d jyH1d 
	;# xmm7= jxH2b jyH2b jxH2d jyH2d 
	
	movaps xmm0, xmm2
	movaps xmm1, xmm3
	unpcklps xmm0, xmm5	;# xmm0= jxOa  jxOb  jyOa  jyOb 
	unpcklps xmm1, xmm6	;# xmm1= jxH1a jxH1b jyH1a jyH1b 
	unpckhps xmm2, xmm5	;# xmm2= jxOc  jxOd  jyOc  jyOd 
	unpckhps xmm3, xmm6	;# xmm3= jxH1c jxH1d jyH1c jyH1d 
	movaps xmm5, xmm4
	movaps   xmm6, xmm0
	unpcklps xmm4, xmm7	;# xmm4= jxH2a jxH2b jyH2a jyH2b 		
	unpckhps xmm5, xmm7	;# xmm5= jxH2c jxH2d jyH2c jyH2d 
	movaps   xmm7, xmm1
	movlhps  xmm0, xmm2	;# xmm0= jxOa  jxOb  jxOc  jxOd 
	movaps [rsp + nb202nf_jxO], xmm0
	movhlps  xmm2, xmm6	;# xmm2= jyOa  jyOb  jyOc  jyOd 
	movaps [rsp + nb202nf_jyO], xmm2
	movlhps  xmm1, xmm3
	movaps [rsp + nb202nf_jxH1], xmm1
	movhlps  xmm3, xmm7
	movaps   xmm6, xmm4
	movaps [rsp + nb202nf_jyH1], xmm3
	movlhps  xmm4, xmm5
	movaps [rsp + nb202nf_jxH2], xmm4
	movhlps  xmm5, xmm6
	movaps [rsp + nb202nf_jyH2], xmm5

	movss  xmm0, [rsi + rax*4 + 8]
	movss  xmm1, [rsi + rax*4 + 20]
	movss  xmm2, [rsi + rax*4 + 32]

	movss  xmm3, [rsi + rcx*4 + 8]
	movss  xmm4, [rsi + rcx*4 + 20]
	movss  xmm5, [rsi + rcx*4 + 32]

	movhps xmm0, [rsi + rbx*4 + 4]
	movhps xmm1, [rsi + rbx*4 + 16]
	movhps xmm2, [rsi + rbx*4 + 28]
	
	movhps xmm3, [rsi + rdx*4 + 4]
	movhps xmm4, [rsi + rdx*4 + 16]
	movhps xmm5, [rsi + rdx*4 + 28]
	
	shufps xmm0, xmm3, 204  ;# 11001100
	shufps xmm1, xmm4, 204  ;# 11001100
	shufps xmm2, xmm5, 204  ;# 11001100
	movaps [rsp + nb202nf_jzO],  xmm0
	movaps [rsp + nb202nf_jzH1],  xmm1
	movaps [rsp + nb202nf_jzH2],  xmm2

	movaps xmm0, [rsp + nb202nf_ixO]
	movaps xmm1, [rsp + nb202nf_iyO]
	movaps xmm2, [rsp + nb202nf_izO]
	movaps xmm3, [rsp + nb202nf_ixO]
	movaps xmm4, [rsp + nb202nf_iyO]
	movaps xmm5, [rsp + nb202nf_izO]
	subps  xmm0, [rsp + nb202nf_jxO]
	subps  xmm1, [rsp + nb202nf_jyO]
	subps  xmm2, [rsp + nb202nf_jzO]
	subps  xmm3, [rsp + nb202nf_jxH1]
	subps  xmm4, [rsp + nb202nf_jyH1]
	subps  xmm5, [rsp + nb202nf_jzH1]
	mulps  xmm0, xmm0
	mulps  xmm1, xmm1
	mulps  xmm2, xmm2
	mulps  xmm3, xmm3
	mulps  xmm4, xmm4
	mulps  xmm5, xmm5
	addps  xmm0, xmm1
	addps  xmm0, xmm2
	addps  xmm3, xmm4
	addps  xmm3, xmm5
	movaps [rsp + nb202nf_rsqOO], xmm0
	movaps [rsp + nb202nf_rsqOH1], xmm3

	movaps xmm0, [rsp + nb202nf_ixO]
	movaps xmm1, [rsp + nb202nf_iyO]
	movaps xmm2, [rsp + nb202nf_izO]
	movaps xmm3, [rsp + nb202nf_ixH1]
	movaps xmm4, [rsp + nb202nf_iyH1]
	movaps xmm5, [rsp + nb202nf_izH1]
	subps  xmm0, [rsp + nb202nf_jxH2]
	subps  xmm1, [rsp + nb202nf_jyH2]
	subps  xmm2, [rsp + nb202nf_jzH2]
	subps  xmm3, [rsp + nb202nf_jxO]
	subps  xmm4, [rsp + nb202nf_jyO]
	subps  xmm5, [rsp + nb202nf_jzO]
	mulps  xmm0, xmm0
	mulps  xmm1, xmm1
	mulps  xmm2, xmm2
	mulps  xmm3, xmm3
	mulps  xmm4, xmm4
	mulps  xmm5, xmm5
	addps  xmm0, xmm1
	addps  xmm0, xmm2
	addps  xmm3, xmm4
	addps  xmm3, xmm5
	movaps [rsp + nb202nf_rsqOH2], xmm0
	movaps [rsp + nb202nf_rsqH1O], xmm3

	movaps xmm0, [rsp + nb202nf_ixH1]
	movaps xmm1, [rsp + nb202nf_iyH1]
	movaps xmm2, [rsp + nb202nf_izH1]
	movaps xmm3, [rsp + nb202nf_ixH1]
	movaps xmm4, [rsp + nb202nf_iyH1]
	movaps xmm5, [rsp + nb202nf_izH1]
	subps  xmm0, [rsp + nb202nf_jxH1]
	subps  xmm1, [rsp + nb202nf_jyH1]
	subps  xmm2, [rsp + nb202nf_jzH1]
	subps  xmm3, [rsp + nb202nf_jxH2]
	subps  xmm4, [rsp + nb202nf_jyH2]
	subps  xmm5, [rsp + nb202nf_jzH2]
	mulps  xmm0, xmm0
	mulps  xmm1, xmm1
	mulps  xmm2, xmm2
	mulps  xmm3, xmm3
	mulps  xmm4, xmm4
	mulps  xmm5, xmm5
	addps  xmm0, xmm1
	addps  xmm0, xmm2
	addps  xmm3, xmm4
	addps  xmm3, xmm5
	movaps [rsp + nb202nf_rsqH1H1], xmm0
	movaps [rsp + nb202nf_rsqH1H2], xmm3

	movaps xmm0, [rsp + nb202nf_ixH2]
	movaps xmm1, [rsp + nb202nf_iyH2]
	movaps xmm2, [rsp + nb202nf_izH2]
	movaps xmm3, [rsp + nb202nf_ixH2]
	movaps xmm4, [rsp + nb202nf_iyH2]
	movaps xmm5, [rsp + nb202nf_izH2]
	subps  xmm0, [rsp + nb202nf_jxO]
	subps  xmm1, [rsp + nb202nf_jyO]
	subps  xmm2, [rsp + nb202nf_jzO]
	subps  xmm3, [rsp + nb202nf_jxH1]
	subps  xmm4, [rsp + nb202nf_jyH1]
	subps  xmm5, [rsp + nb202nf_jzH1]
	mulps  xmm0, xmm0
	mulps  xmm1, xmm1
	mulps  xmm2, xmm2
	mulps  xmm3, xmm3
	mulps  xmm4, xmm4
	mulps  xmm5, xmm5
	addps  xmm0, xmm1
	addps  xmm0, xmm2
	addps  xmm4, xmm3
	addps  xmm4, xmm5
	movaps [rsp + nb202nf_rsqH2O], xmm0
	movaps [rsp + nb202nf_rsqH2H1], xmm4

	movaps xmm0, [rsp + nb202nf_ixH2]
	movaps xmm1, [rsp + nb202nf_iyH2]
	movaps xmm2, [rsp + nb202nf_izH2]
	subps  xmm0, [rsp + nb202nf_jxH2]
	subps  xmm1, [rsp + nb202nf_jyH2]
	subps  xmm2, [rsp + nb202nf_jzH2]
	mulps xmm0, xmm0
	mulps xmm1, xmm1
	mulps xmm2, xmm2
	addps xmm0, xmm1
	addps xmm0, xmm2
	movaps [rsp + nb202nf_rsqH2H2], xmm0
	
	;# start doing invsqrt use rsq values in xmm0, xmm4 
	rsqrtps xmm1, xmm0
	rsqrtps xmm5, xmm4
	movaps  xmm2, xmm1
	movaps  xmm6, xmm5
	mulps   xmm1, xmm1
	mulps   xmm5, xmm5
	movaps  xmm3, [rsp + nb202nf_three]
	movaps  xmm7, xmm3
	mulps   xmm1, xmm0
	mulps   xmm5, xmm4
	subps   xmm3, xmm1
	subps   xmm7, xmm5
	mulps   xmm3, xmm2
	mulps   xmm7, xmm6
	mulps   xmm3, [rsp + nb202nf_half] ;# rinvH2H2 
	mulps   xmm7, [rsp + nb202nf_half] ;# rinvH2H1 
	movaps  [rsp + nb202nf_rinvH2H2], xmm3
	movaps  [rsp + nb202nf_rinvH2H1], xmm7
	
	rsqrtps xmm1, [rsp + nb202nf_rsqOO]
	rsqrtps xmm5, [rsp + nb202nf_rsqOH1]
	movaps  xmm2, xmm1
	movaps  xmm6, xmm5
	mulps   xmm1, xmm1
	mulps   xmm5, xmm5
	movaps  xmm3, [rsp + nb202nf_three]
	movaps  xmm7, xmm3
	mulps   xmm1, [rsp + nb202nf_rsqOO]
	mulps   xmm5, [rsp + nb202nf_rsqOH1]
	subps   xmm3, xmm1
	subps   xmm7, xmm5
	mulps   xmm3, xmm2
	mulps   xmm7, xmm6
	mulps   xmm3, [rsp + nb202nf_half] 
	mulps   xmm7, [rsp + nb202nf_half]
	movaps  [rsp + nb202nf_rinvOO], xmm3
	movaps  [rsp + nb202nf_rinvOH1], xmm7
	
	rsqrtps xmm1, [rsp + nb202nf_rsqOH2]
	rsqrtps xmm5, [rsp + nb202nf_rsqH1O]
	movaps  xmm2, xmm1
	movaps  xmm6, xmm5
	mulps   xmm1, xmm1
	mulps   xmm5, xmm5
	movaps  xmm3, [rsp + nb202nf_three]
	movaps  xmm7, xmm3
	mulps   xmm1, [rsp + nb202nf_rsqOH2]
	mulps   xmm5, [rsp + nb202nf_rsqH1O]
	subps   xmm3, xmm1
	subps   xmm7, xmm5
	mulps   xmm3, xmm2
	mulps   xmm7, xmm6
	mulps   xmm3, [rsp + nb202nf_half] 
	mulps   xmm7, [rsp + nb202nf_half]
	movaps  [rsp + nb202nf_rinvOH2], xmm3
	movaps  [rsp + nb202nf_rinvH1O], xmm7
	
	rsqrtps xmm1, [rsp + nb202nf_rsqH1H1]
	rsqrtps xmm5, [rsp + nb202nf_rsqH1H2]
	movaps  xmm2, xmm1
	movaps  xmm6, xmm5
	mulps   xmm1, xmm1
	mulps   xmm5, xmm5
	movaps  xmm3, [rsp + nb202nf_three]
	movaps  xmm7, xmm3
	mulps   xmm1, [rsp + nb202nf_rsqH1H1]
	mulps   xmm5, [rsp + nb202nf_rsqH1H2]
	subps   xmm3, xmm1
	subps   xmm7, xmm5
	mulps   xmm3, xmm2
	mulps   xmm7, xmm6
	mulps   xmm3, [rsp + nb202nf_half] 
	mulps   xmm7, [rsp + nb202nf_half]
	movaps  [rsp + nb202nf_rinvH1H1], xmm3
	movaps  [rsp + nb202nf_rinvH1H2], xmm7
	
	rsqrtps xmm1, [rsp + nb202nf_rsqH2O]
	movaps  xmm2, xmm1
	mulps   xmm1, xmm1
	movaps  xmm3, [rsp + nb202nf_three]
	mulps   xmm1, [rsp + nb202nf_rsqH2O]
	subps   xmm3, xmm1
	mulps   xmm3, xmm2
	mulps   xmm3, [rsp + nb202nf_half] 
	movaps  [rsp + nb202nf_rinvH2O], xmm3

	;# start with OO interaction 
	movaps xmm6, [rsp + nb202nf_krf]
	mulps  xmm6, [rsp + nb202nf_rsqOO] ;# xmm5=krsq 
	addps  xmm6, [rsp + nb202nf_rinvOO]	;# xmm6=rinv+ krsq 
	subps  xmm6, [rsp + nb202nf_crf]
	mulps  xmm6, [rsp + nb202nf_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) 
	addps  xmm6, [rsp + nb202nf_vctot] ;# local vctot summation variable 

	;# O-H interactions 
	movaps xmm0, [rsp + nb202nf_krf]
	movaps xmm1, [rsp + nb202nf_krf]
	movaps xmm2, [rsp + nb202nf_krf]
	movaps xmm3, [rsp + nb202nf_krf]
	mulps  xmm0, [rsp + nb202nf_rsqOH1] ;# krsq 
	mulps  xmm1, [rsp + nb202nf_rsqOH2] ;# krsq 
	mulps  xmm2, [rsp + nb202nf_rsqH1O] ;# krsq 
	mulps  xmm3, [rsp + nb202nf_rsqH2O] ;# krsq 
	addps  xmm0, [rsp + nb202nf_rinvOH1]	;# rinv+ krsq 
	addps  xmm1, [rsp + nb202nf_rinvOH2]	;# rinv+ krsq 
	addps  xmm2, [rsp + nb202nf_rinvH1O]	;# rinv+ krsq 
	addps  xmm3, [rsp + nb202nf_rinvH2O]	;# rinv+ krsq 
	subps  xmm0, [rsp + nb202nf_crf]
	subps  xmm1, [rsp + nb202nf_crf]
	subps  xmm2, [rsp + nb202nf_crf]
	subps  xmm3, [rsp + nb202nf_crf]
	mulps  xmm0, [rsp + nb202nf_qqOH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm1, [rsp + nb202nf_qqOH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm2, [rsp + nb202nf_qqOH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm3, [rsp + nb202nf_qqOH] ;# voul=qq*(rinv+ krsq-crf) 
	addps xmm6, xmm0
	addps xmm1, xmm2
	addps xmm6, xmm3
	addps xmm6, xmm1
	
	;# H-H interactions 
	movaps xmm0, [rsp + nb202nf_krf]
	movaps xmm1, [rsp + nb202nf_krf]
	movaps xmm2, [rsp + nb202nf_krf]
	movaps xmm3, [rsp + nb202nf_krf]
	mulps  xmm0, [rsp + nb202nf_rsqH1H1] ;# krsq 
	mulps  xmm1, [rsp + nb202nf_rsqH1H2] ;# krsq 
	mulps  xmm2, [rsp + nb202nf_rsqH2H1] ;# krsq 
	mulps  xmm3, [rsp + nb202nf_rsqH2H2] ;# krsq 
	addps  xmm0, [rsp + nb202nf_rinvH1H1]	;# rinv+ krsq 
	addps  xmm1, [rsp + nb202nf_rinvH1H2]	;# rinv+ krsq 
	addps  xmm2, [rsp + nb202nf_rinvH2H1]	;# rinv+ krsq 
	addps  xmm3, [rsp + nb202nf_rinvH2H2]	;# rinv+ krsq 
	subps  xmm0, [rsp + nb202nf_crf]
	subps  xmm1, [rsp + nb202nf_crf]
	subps  xmm2, [rsp + nb202nf_crf]
	subps  xmm3, [rsp + nb202nf_crf]
	mulps  xmm0, [rsp + nb202nf_qqHH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm1, [rsp + nb202nf_qqHH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm2, [rsp + nb202nf_qqHH] ;# voul=qq*(rinv+ krsq-crf) 
	mulps  xmm3, [rsp + nb202nf_qqHH] ;# voul=qq*(rinv+ krsq-crf) 
	addps xmm6, xmm0
	addps xmm1, xmm2
	addps xmm6, xmm3
	addps xmm6, xmm1
	movaps [rsp + nb202nf_vctot], xmm6
	
	;# should w