nb_kernel301_x86_64_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,950 行 · 第 1/4 页

	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 + nb301nf_ixO]
	movaps xmm5, [rsp + nb301nf_iyO]
	movaps xmm6, [rsp + nb301nf_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 + nb301nf_ixH1]
	movaps xmm5, [rsp + nb301nf_iyH1]
	movaps xmm6, [rsp + nb301nf_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 + nb301nf_ixH2]
	movaps xmm4, [rsp + nb301nf_iyH2]
	movaps xmm5, [rsp + nb301nf_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 + nb301nf_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 + nb301nf_half]
	movaps  [rsp + nb301nf_rinvO], xmm4	;# rinvO in xmm4 
	mulps   xmm7, xmm4
	movaps  [rsp + nb301nf_rO], xmm7	

	;# rsqH1 - seed in xmm2 
	rsqrtps xmm2, xmm6
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm4, [rsp + nb301nf_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 + nb301nf_half]
	movaps  [rsp + nb301nf_rinvH1], xmm4	;# rinvH1 in xmm4 
	mulps   xmm6, xmm4
	movaps  [rsp + nb301nf_rH1], xmm6

	;# rsqH2 - seed to xmm2 
	rsqrtps xmm2, xmm5
	movaps  xmm3, xmm2
	mulps   xmm2, xmm2
	movaps  xmm4, [rsp + nb301nf_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 + nb301nf_half]
	movaps  [rsp + nb301nf_rinvH2], xmm4	;# rinvH2 in xmm4 
	mulps   xmm5, xmm4
	movaps  [rsp + nb301nf_rH2], xmm5

	;# do O interactions 
	;# rO is still in xmm7 
	mulps   xmm7, [rsp + nb301nf_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 
    	pslld mm6, 2
    	pslld mm7, 2
	
    	movd mm0, eax   
    	movd mm1, ebx
    	movd mm2, ecx
    	movd mm3, edx

    	mov  rsi, [rbp + nb301nf_VFtab]
    	movd eax, mm6
    	psrlq mm6, 32
    	movd ecx, mm7
    	psrlq mm7, 32
    	movd ebx, mm6
    	movd edx, mm7

    	movlps xmm5, [rsi + rax*4]
    	movlps xmm7, [rsi + rcx*4]
    	movhps xmm5, [rsi + rbx*4]
    	movhps xmm7, [rsi + rdx*4] ;# got half coulomb table 

    	movaps xmm4, xmm5
    	shufps xmm4, xmm7, 136  ;# 10001000
    	shufps xmm5, xmm7, 221  ;# 11011101

    	movlps xmm7, [rsi + rax*4 + 8]
    	movlps xmm3, [rsi + rcx*4 + 8]
    	movhps xmm7, [rsi + rbx*4 + 8]
    	movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table  
    	movaps xmm6, xmm7
    	shufps xmm6, xmm3, 136  ;# 10001000
    	shufps xmm7, xmm3, 221  ;# 11011101
    	;# coulomb table ready, in xmm4-xmm7      
        
    	mulps  xmm6, xmm1       ;# xmm6=Geps 
    	mulps  xmm7, xmm2       ;# xmm7=Heps2 
    	addps  xmm5, xmm6
    	addps  xmm5, xmm7       ;# xmm5=Fp    
    	movaps xmm0, [rsp + nb301nf_qqO]
    	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
    	addps  xmm5, xmm4 ;# xmm5=VV 
    	mulps  xmm5, xmm0 ;# vcoul=qq*VV  

    	;# at this point mm5 contains vcoul 
    	;# increment vcoul - then we can get rid of mm5 
    	addps  xmm5, [rsp + nb301nf_vctot]
    	movaps [rsp + nb301nf_vctot], xmm5
	
	;# Done with O interactions - now H1! 
	movaps xmm7, [rsp + nb301nf_rH1]
	mulps   xmm7, [rsp + nb301nf_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 
    	pslld mm6, 2
    	pslld mm7, 2
	
    	movd eax, mm6
    	psrlq mm6, 32
    	movd ecx, mm7
    	psrlq mm7, 32
    	movd ebx, mm6
    	movd edx, mm7

    	movlps xmm5, [rsi + rax*4]
    	movlps xmm7, [rsi + rcx*4]
    	movhps xmm5, [rsi + rbx*4]
    	movhps xmm7, [rsi + rdx*4] ;# got half coulomb table 

    	movaps xmm4, xmm5
    	shufps xmm4, xmm7, 136  ;# 10001000
    	shufps xmm5, xmm7, 221  ;# 11011101

    	movlps xmm7, [rsi + rax*4 + 8]
    	movlps xmm3, [rsi + rcx*4 + 8]
    	movhps xmm7, [rsi + rbx*4 + 8]
    	movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table  
    	movaps xmm6, xmm7
    	shufps xmm6, xmm3, 136  ;# 10001000
    	shufps xmm7, xmm3, 221  ;# 11011101
    	;# coulomb table ready, in xmm4-xmm7      
        
    	mulps  xmm6, xmm1       ;# xmm6=Geps 
    	mulps  xmm7, xmm2       ;# xmm7=Heps2 
    	addps  xmm5, xmm6
    	addps  xmm5, xmm7       ;# xmm5=Fp        
    	movaps xmm0, [rsp + nb301nf_qqH]
    	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
    	addps  xmm5, xmm4 ;# xmm5=VV 
    	mulps  xmm5, xmm0 ;# vcoul=qq*VV  
    	;# at this point mm5 contains vcoul 
    	;# increment vcoul 
    	addps  xmm5, [rsp + nb301nf_vctot]
    	movaps [rsp + nb301nf_vctot], xmm5 

	;# Done with H1, finally we do H2 interactions 
	movaps xmm7, [rsp + nb301nf_rH2]
	mulps   xmm7, [rsp + nb301nf_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 
    	pslld mm6, 2
    	pslld mm7, 2
	
    	movd eax, mm6
    	psrlq mm6, 32
    	movd ecx, mm7
    	psrlq mm7, 32
    	movd ebx, mm6
    	movd edx, mm7

    	movlps xmm5, [rsi + rax*4]
    	movlps xmm7, [rsi + rcx*4]
    	movhps xmm5, [rsi + rbx*4]
    	movhps xmm7, [rsi + rdx*4] ;# got half coulomb table 

    	movaps xmm4, xmm5
    	shufps xmm4, xmm7, 136  ;# 10001000
    	shufps xmm5, xmm7, 221  ;# 11011101

    	movlps xmm7, [rsi + rax*4 + 8]
    	movlps xmm3, [rsi + rcx*4 + 8]
    	movhps xmm7, [rsi + rbx*4 + 8]
    	movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table  
    	movaps xmm6, xmm7
    	shufps xmm6, xmm3, 136  ;# 10001000
    	shufps xmm7, xmm3, 221  ;# 11011101
    	;# coulomb table ready, in xmm4-xmm7      
        
    	mulps  xmm6, xmm1       ;# xmm6=Geps 
    	mulps  xmm7, xmm2       ;# xmm7=Heps2 
    	addps  xmm5, xmm6
    	addps  xmm5, xmm7       ;# xmm5=Fp        
    	movaps xmm0, [rsp + nb301nf_qqH]
    	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
    	addps  xmm5, xmm4 ;# xmm5=VV 
    	mulps  xmm5, xmm0 ;# vcoul=qq*VV  
    	;# at this point mm5 contains vcoul 
    	;# increment vcoul 
    	addps  xmm5, [rsp + nb301nf_vctot]
    	movaps [rsp + nb301nf_vctot], xmm5
	
	;# should we do one more iteration? 
	sub dword ptr [rsp + nb301nf_innerk],  4
	jl    .nb301nf_odd_inner
	jmp   .nb301nf_unroll_loop
.nb301nf_odd_inner:	
	add dword ptr [rsp + nb301nf_innerk],  4
	jnz   .nb301nf_odd_loop
	jmp   .nb301nf_updateouterdata
.nb301nf_odd_loop:
	mov   rdx, [rsp + nb301nf_innerjjnr]     ;# pointer to jjnr[k] 
	mov   eax, [rdx]	
	add qword ptr [rsp + nb301nf_innerjjnr],  4	

 	xorps xmm4, xmm4
	movss xmm4, [rsp + nb301nf_iqO]
	mov rsi, [rbp + nb301nf_charge] 
	movhps xmm4, [rsp + nb301nf_iqH]     
	movss xmm3, [rsi + rax*4]	;# charge in xmm3 
	shufps xmm3, xmm3, 0
	mulps xmm3, xmm4
	movaps [rsp + nb301nf_qqO], xmm3	;# use oxygen qq for storage 

	mov rsi, [rbp + nb301nf_pos]
	lea   rax, [rax + rax*2]  
	
	;# move j coords to xmm0-xmm2 
	movss xmm0, [rsi + rax*4]
	movss xmm1, [rsi + rax*4 + 4]
	movss xmm2, [rsi + rax*4 + 8]
	shufps xmm0, xmm0, 0
	shufps xmm1, xmm1, 0
	shufps xmm2, xmm2, 0
	
	movss xmm3, [rsp + nb301nf_ixO]
	movss xmm4, [rsp + nb301nf_iyO]
	movss xmm5, [rsp + nb301nf_izO]
	
	movlps xmm6, [rsp + nb301nf_ixH1]
	movlps xmm7, [rsp + nb301nf_ixH2]
	unpcklps xmm6, xmm7
	movlhps xmm3, xmm6
	movlps xmm6, [rsp + nb301nf_iyH1]
	movlps xmm7, [rsp + nb301nf_iyH2]
	unpcklps xmm6, xmm7
	movlhps xmm4, xmm6
	movlps xmm6, [rsp + nb301nf_izH1]
	movlps xmm7, [rsp + nb301nf_izH2]
	unpcklps xmm6, xmm7
	movlhps xmm5, xmm6

	subps xmm3, xmm0
	subps xmm4, xmm1
	subps xmm5, xmm2

	mulps  xmm3, xmm3
	mulps  xmm4, xmm4
	mulps  xmm5, xmm5

	addps  xmm4, xmm3
	addps  xmm4, xmm5
	;# rsq in xmm4 

	rsqrtps xmm5, xmm4
	;# lookup seed in xmm5 
	movaps xmm2, xmm5
	mulps xmm5, xmm5
	movaps xmm1, [rsp + nb301nf_three]
	mulps xmm5, xmm4	;# rsq*lu*lu 			
	movaps xmm0, [rsp + nb301nf_half]
	subps xmm1, xmm5	;# 30-rsq*lu*lu 
	mulps xmm1, xmm2	
	mulps xmm0, xmm1	;# xmm0=rinv 

	;# a little trick to avoid NaNs: 
	;# positions 0,2,and 3 are valid, but not 1. 
	;# If it contains NaN it doesnt help to mult by 0, 
	;# So we shuffle it and copy pos 0 to pos1! 
	shufps xmm0, xmm0, 224 ;# 11100000	
	
	mulps xmm4, xmm0	;# xmm4=r 
	movaps [rsp + nb301nf_rinvO], xmm0
	
	mulps xmm4, [rsp + nb301nf_tsc]
	movhlps xmm7, xmm4
	cvttps2pi mm6, xmm4
	cvttps2pi mm7, xmm7    ;# mm6/mm7 contain lu indices 
    	cvtpi2ps xmm3, mm6
    	cvtpi2ps xmm7, mm7
    	movlhps xmm3, xmm7

	subps   xmm4, xmm3	
	movaps xmm1, xmm4	;# xmm1=eps 
	movaps xmm2, xmm1
	mulps  xmm2, xmm2	;# xmm2=eps2 
    	pslld mm6, 2
    	pslld mm7, 2
	
    	movd mm0, eax   
    	movd mm1, ecx
    	movd mm2, edx

    	mov  rsi, [rbp + nb301nf_VFtab]
    	movd eax, mm6
    	movd ecx, mm7
    	psrlq mm7, 32
    	movd edx, mm7

    	movlps xmm5, [rsi + rax*4]
    	movlps xmm7, [rsi + rcx*4]
    	movhps xmm7, [rsi + rdx*4] ;# got half coulomb table 

    	movaps xmm4, xmm5
    	shufps xmm4, xmm7, 136  ;# 10001000
    	shufps xmm5, xmm7, 221  ;# 11011101

    	movlps xmm7, [rsi + rax*4 + 8]
    	movlps xmm3, [rsi + rcx*4 + 8]
    	movhps xmm3, [rsi + rdx*4 + 8] ;# other half of coulomb table  
    	movaps xmm6, xmm7
    	shufps xmm6, xmm3, 136  ;# 10001000
    	shufps xmm7, xmm3, 221  ;# 11011101
    	;# coulomb table ready, in xmm4-xmm7      
    	mulps  xmm6, xmm1       ;# xmm6=Geps 
    	mulps  xmm7, xmm2       ;# xmm7=Heps2 
    	addps  xmm5, xmm6
    	addps  xmm5, xmm7       ;# xmm5=Fp        
    	movaps xmm0, [rsp + nb301nf_qqO]
    	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
    	addps  xmm5, xmm4 ;# xmm5=VV 
    	mulps  xmm5, xmm0 ;# vcoul=qq*VV  
    	;# at this point mm5 contains vcoul 
    	;# increment vcoul - then we can get rid of mm5 
    	addps  xmm5, [rsp + nb301nf_vctot]
    	movaps [rsp + nb301nf_vctot], xmm5

	dec dword ptr [rsp + nb301nf_innerk]
	jz    .nb301nf_updateouterdata
	jmp   .nb301nf_odd_loop
.nb301nf_updateouterdata:
	;# get n from stack
	mov esi, [rsp + nb301nf_n]
        ;# get group index for i particle 
        mov   rdx, [rbp + nb301nf_gid]      	;# base of gid[]
        mov   edx, [rdx + rsi*4]		;# ggid=gid[n]

	;# accumulate total potential energy and update it 
	movaps xmm7, [rsp + nb301nf_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 + nb301nf_Vc]
	addss xmm7, [rax + rdx*4] 
	;# move back to mem 
	movss [rax + rdx*4], xmm7 

        ;# finish if last 
        mov ecx, [rsp + nb301nf_nn1]
	;# esi already loaded with n
	inc esi
        sub ecx, esi
        jz .nb301nf_outerend

        ;# not last, iterate outer loop once more!  
        mov [rsp + nb301nf_n], esi
        jmp .nb301nf_outer
.nb301nf_outerend:
        ;# check if more outer neighborlists remain
        mov   ecx, [rsp + nb301nf_nri]
	;# esi already loaded with n above
        sub   ecx, esi
        jz .nb301nf_end
        ;# non-zero, do one more workunit
        jmp   .nb301nf_threadloop
.nb301nf_end:

	mov eax, [rsp + nb301nf_nouter]
	mov ebx, [rsp + nb301nf_ninner]
	mov rcx, [rbp + nb301nf_outeriter]
	mov rdx, [rbp + nb301nf_inneriter]
	mov [rcx], eax
	mov [rdx], ebx

	add rsp, 472
	emms


        pop r15
        pop r14
        pop r13
        pop r12

	pop rbx
	pop	rbp
	ret