nb_kernel410_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,051 行 · 第 1/4 页

	mov edx, [esp + nb410_jnrd]
	
	mov esi, [ebp + nb410_dvda]
	
	;# Calculate dVda
	xorps xmm7, xmm7
	mulps xmm3, [esp + nb410_gbscale]
	movaps xmm6, xmm3
	mulps  xmm6, [esp + nb410_r]
	addps  xmm6, xmm5
	addps  xmm5, [esp + nb410_vctot]
	movaps [esp + nb410_vctot], xmm5 

	;# xmm6=(vcoul+fijC*r)
	subps  xmm7, xmm6
	movaps xmm6, xmm7
	
	;# update dvdasum
	addps  xmm7, [esp + nb410_dvdasum]
	movaps [esp + nb410_dvdasum], xmm7 

	;# update j atoms dvdaj
	movhlps xmm7, xmm6
	movaps  xmm5, xmm6
	movaps  xmm4, xmm7
	shufps  xmm5, xmm5, 0x1
	shufps  xmm4, xmm4, 0x1
	;# xmm6=dvdaj1 xmm5=dvdaj2 xmm7=dvdaj3 xmm4=dvdaj4
	addss  xmm6, [esi + eax*4]
	addss  xmm5, [esi + ebx*4]
	addss  xmm7, [esi + ecx*4]
	addss  xmm4, [esi + edx*4]
	movss  [esi + eax*4], xmm6
	movss  [esi + ebx*4], xmm5
	movss  [esi + ecx*4], xmm7
	movss  [esi + edx*4], xmm4
	
	;# L-J 
	movaps xmm4, xmm0
	mulps  xmm4, xmm0	;# xmm4=rinvsq 

	movaps xmm6, xmm4
	mulps  xmm6, xmm4

	mulps  xmm6, xmm4	;# xmm6=rinvsix 
	movaps xmm4, xmm6
	mulps  xmm4, xmm4	;# xmm4=rinvtwelve 
	mulps  xmm6, [esp + nb410_c6]
	mulps  xmm4, [esp + nb410_c12]
	movaps xmm7, [esp + nb410_Vvdwtot]
	addps  xmm7, xmm4
	mulps  xmm4, [esp + nb410_twelve]
	subps  xmm7, xmm6
	mulps  xmm6, [esp + nb410_six]
	movaps [esp + nb410_Vvdwtot], xmm7
	subps  xmm4, xmm6
	mulps  xmm4, xmm0
	subps  xmm4, xmm3
	mulps  xmm4, xmm0

	movaps xmm0, [esp + nb410_dx]
	movaps xmm1, [esp + nb410_dy]
	movaps xmm2, [esp + nb410_dz]

	movd eax, mm0	
	movd ebx, mm1
	movd ecx, mm2
	movd edx, mm3

	mov    edi, [ebp + nb410_faction]
	mulps  xmm0, xmm4
	mulps  xmm1, xmm4
	mulps  xmm2, xmm4
	;# xmm0-xmm2 contains tx-tz (partial force) 
	;# now update f_i 
	movaps xmm3, [esp + nb410_fix]
	movaps xmm4, [esp + nb410_fiy]
	movaps xmm5, [esp + nb410_fiz]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	addps  xmm5, xmm2
	movaps [esp + nb410_fix], xmm3
	movaps [esp + nb410_fiy], xmm4
	movaps [esp + nb410_fiz], xmm5
	;# the fj's - start by accumulating x & y forces from memory 
	movlps xmm4, [edi + eax*4]
	movlps xmm6, [edi + ecx*4]
	movhps xmm4, [edi + ebx*4]
	movhps xmm6, [edi + edx*4]

	movaps xmm3, xmm4
	shufps xmm3, xmm6, 136  ;# constant 10001000
	shufps xmm4, xmm6, 221  ;# constant 11011101			      

	;# now xmm3-xmm5 contains fjx, fjy, fjz 
	subps  xmm3, xmm0
	subps  xmm4, xmm1
	
	;# unpack them back so we can store them - first x & y in xmm3/xmm4 

	movaps xmm6, xmm3
	unpcklps xmm6, xmm4
	unpckhps xmm3, xmm4	
	;# xmm6(l)=x & y for j1, (h) for j2 
	;# xmm3(l)=x & y for j3, (h) for j4 
	movlps [edi + eax*4], xmm6
	movlps [edi + ecx*4], xmm3
	
	movhps [edi + ebx*4], xmm6
	movhps [edi + edx*4], xmm3

	;# and the z forces 
	movss  xmm4, [edi + eax*4 + 8]
	movss  xmm5, [edi + ebx*4 + 8]
	movss  xmm6, [edi + ecx*4 + 8]
	movss  xmm7, [edi + edx*4 + 8]
	subss  xmm4, xmm2
	shufps xmm2, xmm2, 229  ;# constant 11100101
	subss  xmm5, xmm2
	shufps xmm2, xmm2, 234  ;# constant 11101010
	subss  xmm6, xmm2
	shufps xmm2, xmm2, 255  ;# constant 11111111
	subss  xmm7, xmm2
	movss  [edi + eax*4 + 8], xmm4
	movss  [edi + ebx*4 + 8], xmm5
	movss  [edi + ecx*4 + 8], xmm6
	movss  [edi + edx*4 + 8], xmm7
	
	;# should we do one more iteration? 
	sub dword ptr [esp + nb410_innerk],  4
	jl    .nb410_finish_inner
	jmp   .nb410_unroll_loop
.nb410_finish_inner:
	;# check if at least two particles remain 
	add dword ptr [esp + nb410_innerk],  4
	mov   edx, [esp + nb410_innerk]
	and   edx, 2
	jnz   .nb410_dopair
	jmp   .nb410_checksingle
.nb410_dopair:	
	mov   ecx, [esp + nb410_innerjjnr]
	mov   eax, [ecx]	
	mov   ebx, [ecx + 4]              
	add dword ptr [esp + nb410_innerjjnr],  8

	xorps xmm2, xmm2
	movaps xmm6, xmm2
	
	;# load isaj
	mov esi, [ebp + nb410_invsqrta]
	movss xmm2, [esi + eax*4]
	movss xmm3, [esi + ebx*4]
	unpcklps xmm2, xmm3	;# isaj in xmm2(0,1)
	mulps  xmm2, [esp + nb410_isai]
	movaps [esp + nb410_isaprod], xmm2	
	movaps xmm1, xmm2
	mulps xmm1, [esp + nb410_gbtsc]
	movaps [esp + nb410_gbscale], xmm1	
	
	mov esi, [ebp + nb410_charge]    ;# base of charge[] 	
	movss xmm3, [esi + eax*4]		
	movss xmm6, [esi + ebx*4]
	unpcklps xmm3, xmm6 ;# constant 00001000 ;# xmm3(0,1) has the charges 

	mulps  xmm2, [esp + nb410_iq]
	mulps  xmm3, xmm2
	movaps [esp + nb410_qq], xmm3

	mov esi, [ebp + nb410_type]
	mov   ecx, eax
	mov   edx, ebx
	mov ecx, [esi + ecx*4]
	mov edx, [esi + edx*4]	
	mov esi, [ebp + nb410_vdwparam]
	shl ecx, 1	
	shl edx, 1	
	mov edi, [esp + nb410_ntia]
	add ecx, edi
	add edx, edi
	movlps xmm6, [esi + ecx*4]
	movhps xmm6, [esi + edx*4]
	mov edi, [ebp + nb410_pos]	
	
	movaps xmm4, xmm6
	shufps xmm4, xmm4, 8 ;# constant 00001000 	
	shufps xmm6, xmm6, 13 ;# constant 00001101
	movlhps xmm4, xmm7
	movlhps xmm6, xmm7
	
	movaps [esp + nb410_c6], xmm4
	movaps [esp + nb410_c12], xmm6	
		
	movd  mm0, eax
	movd  mm1, ebx
		
	lea   eax, [eax + eax*2]
	lea   ebx, [ebx + ebx*2]
	;# move coordinates to xmm0-xmm2 
	movlps xmm1, [edi + eax*4]
	movss xmm2, [edi + eax*4 + 8]	
	movhps xmm1, [edi + ebx*4]
	movss xmm0, [edi + ebx*4 + 8]	

	movlhps xmm3, xmm7
	
	shufps xmm2, xmm0, 0
	
	movaps xmm0, xmm1

	shufps xmm2, xmm2, 136  ;# constant 10001000
	
	shufps xmm0, xmm0, 136  ;# constant 10001000
	shufps xmm1, xmm1, 221  ;# constant 11011101
			
	mov    edi, [ebp + nb410_faction]
	;# move ix-iz to xmm4-xmm6 
	xorps   xmm7, xmm7
	
	movaps xmm4, [esp + nb410_ix]
	movaps xmm5, [esp + nb410_iy]
	movaps xmm6, [esp + nb410_iz]

	;# calc dr 
	subps xmm4, xmm0
	subps xmm5, xmm1
	subps xmm6, xmm2

	;# store dr 
	movaps [esp + nb410_dx], xmm4
	movaps [esp + nb410_dy], xmm5
	movaps [esp + nb410_dz], xmm6
	;# square it 
	mulps xmm4,xmm4
	mulps xmm5,xmm5
	mulps xmm6,xmm6
	addps xmm4, xmm5
	addps xmm4, xmm6
	;# rsq in xmm4 

	rsqrtps xmm5, xmm4
	;# lookup seed in xmm5 
	movaps xmm2, xmm5
	mulps xmm5, xmm5
	movaps xmm1, [esp + nb410_three]
	mulps xmm5, xmm4	;# rsq*lu*lu 			
	movaps xmm0, [esp + nb410_half]
	subps xmm1, xmm5	;# constant 30-rsq*lu*lu 
	mulps xmm1, xmm2	
	mulps xmm0, xmm1	;# xmm0=rinv 
	mulps xmm4, xmm0	;# xmm4=r 
	movaps [esp + nb410_r], xmm4
	mulps xmm4, [esp + nb410_gbscale]

	cvttps2pi mm6, xmm4     ;# mm6 contain lu indices 
	cvtpi2ps xmm6, mm6
	subps xmm4, xmm6	
	movaps xmm1, xmm4	;# xmm1=eps 
	movaps xmm2, xmm1	
	mulps  xmm2, xmm2	;# xmm2=eps2 

	pslld mm6, 2

	mov  esi, [ebp + nb410_GBtab]
	movd ecx, mm6
	psrlq mm6, 32
	movd edx, mm6

	;# load coulomb table
	movaps xmm4, [esi + ecx*4]
	movaps xmm7, [esi + edx*4]
	;# transpose, using xmm3 for scratch
	movaps xmm6, xmm4
	unpcklps xmm4, xmm7  	;# Y1 Y2 F1 F2 
	unpckhps xmm6, xmm7     ;# G1 G2 H1 H2
	movhlps  xmm5, xmm4    	;# F1 F2 
	movhlps  xmm7, xmm6     ;# H1 H2
	;# coulomb table ready, in xmm4-xmm7  	

	mulps  xmm6, xmm1	;# xmm6=Geps 
	mulps  xmm7, xmm2	;# xmm7=Heps2 
	addps  xmm5, xmm6
	addps  xmm5, xmm7	;# xmm5=Fp 	
	mulps  xmm7, [esp + nb410_two]	;# two*Heps2 
	movaps xmm3, [esp + nb410_qq]
	addps  xmm7, xmm6
	addps  xmm7, xmm5 ;# xmm7=FF 
	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 
	addps  xmm5, xmm4 ;# xmm5=VV 
	mulps  xmm5, xmm3 ;# vcoul=qq*VV  
	mulps  xmm3, xmm7 ;# fijC=FF*qq 
	;# get jnr from regs
	movd ecx, mm0
	movd edx, mm1
	
	mov esi, [ebp + nb410_dvda]
	;# Calculate dVda
	xorps xmm7, xmm7
	mulps xmm3, [esp + nb410_gbscale]
	movaps xmm6, xmm3
	mulps  xmm6, [esp + nb410_r]
	addps  xmm6, xmm5
	addps  xmm5, [esp + nb410_vctot]
	movaps [esp + nb410_vctot], xmm5 

	;# xmm6=(vcoul+fijC*r)
	subps  xmm7, xmm6
	movaps xmm6, xmm7
	
	;# update dvdasum
	addps  xmm7, [esp + nb410_dvdasum]
	movaps [esp + nb410_dvdasum], xmm7 
	
	;# update j atoms dvdaj
	movaps xmm7, xmm6
	shufps xmm7, xmm7, 0x1
	addss  xmm6, [esi + ecx*4]
	addss  xmm7, [esi + edx*4]
	movss  [esi + ecx*4], xmm6
	movss  [esi + edx*4], xmm7
		
	;# L-J 
	movaps xmm4, xmm0
	mulps  xmm4, xmm0	;# xmm4=rinvsq 

	;# at this point mm5 contains vcoul and mm3 fijC 
	;# increment vcoul - then we can get rid of mm5 
	;# update vctot 

	movaps xmm6, xmm4
	mulps  xmm6, xmm4

	mulps  xmm6, xmm4	;# xmm6=rinvsix 
	movaps xmm4, xmm6
	mulps  xmm4, xmm4	;# xmm4=rinvtwelve 
	mulps  xmm6, [esp + nb410_c6]
	mulps  xmm4, [esp + nb410_c12]
	movaps xmm7, [esp + nb410_Vvdwtot]
	addps  xmm7, xmm4
	mulps  xmm4, [esp + nb410_twelve]
	subps  xmm7, xmm6
	mulps  xmm6, [esp + nb410_six]
	movaps [esp + nb410_Vvdwtot], xmm7
	subps  xmm4, xmm6
	mulps  xmm4, xmm0
	subps  xmm4, xmm3
	mulps  xmm4, xmm0

	movaps xmm0, [esp + nb410_dx]
	movaps xmm1, [esp + nb410_dy]
	movaps xmm2, [esp + nb410_dz]

	mulps  xmm0, xmm4
	mulps  xmm1, xmm4
	mulps  xmm2, xmm4
	;# xmm0-xmm2 contains tx-tz (partial force) 
	;# now update f_i 
	movaps xmm3, [esp + nb410_fix]
	movaps xmm4, [esp + nb410_fiy]
	movaps xmm5, [esp + nb410_fiz]
	addps  xmm3, xmm0
	addps  xmm4, xmm1
	addps  xmm5, xmm2
	movaps [esp + nb410_fix], xmm3
	movaps [esp + nb410_fiy], xmm4
	movaps [esp + nb410_fiz], xmm5
	;# update the fj's 
	movss   xmm3, [edi + eax*4]
	movss   xmm4, [edi + eax*4 + 4]
	movss   xmm5, [edi + eax*4 + 8]
	subss   xmm3, xmm0
	subss   xmm4, xmm1
	subss   xmm5, xmm2	
	movss   [edi + eax*4], xmm3
	movss   [edi + eax*4 + 4], xmm4
	movss   [edi + eax*4 + 8], xmm5	

	shufps  xmm0, xmm0, 225  ;# constant 11100001
	shufps  xmm1, xmm1, 225  ;# constant 11100001
	shufps  xmm2, xmm2, 225  ;# constant 11100001

	movss   xmm3, [edi + ebx*4]
	movss   xmm4, [edi + ebx*4 + 4]
	movss   xmm5, [edi + ebx*4 + 8]
	subss   xmm3, xmm0
	subss   xmm4, xmm1
	subss   xmm5, xmm2	
	movss   [edi + ebx*4], xmm3
	movss   [edi + ebx*4 + 4], xmm4
	movss   [edi + ebx*4 + 8], xmm5	

.nb410_checksingle:				
	mov   edx, [esp + nb410_innerk]
	and   edx, 1
	jnz    .nb410_dosingle
	jmp    .nb410_updateouterdata
.nb410_dosingle:
	mov esi, [ebp + nb410_charge]
	mov edx, [ebp + nb410_invsqrta]
	mov edi, [ebp + nb410_pos]
	mov   ecx, [esp + nb410_innerjjnr]
	mov   eax, [ecx]	
	xorps  xmm2, xmm2
	movaps xmm6, xmm2
	movss xmm2, [edx + eax*4]	;# isaj
	mulss xmm2, [esp + nb410_isai]
	movss [esp + nb410_isaprod], xmm2	
	movss xmm1, xmm2
	mulss xmm1, [esp + nb410_gbtsc]
	movss [esp + nb410_gbscale], xmm1	
	
	mulss  xmm2, [esp + nb410_iq]
	movss xmm6, [esi + eax*4]	;# xmm6(0) has the charge 	
	mulss  xmm6, xmm2
	movss [esp + nb410_qq], xmm6
		
	mov esi, [ebp + nb410_type]
	mov ecx, eax
	mov ecx, [esi + ecx*4]	
	mov esi, [ebp + nb410_vdwparam]
	shl ecx, 1
	add ecx, [esp + nb410_ntia]
	movlps xmm6, [esi + ecx*4]
	movaps xmm4, xmm6
	shufps xmm4, xmm4, 252  ;# constant 11111100	
	shufps xmm6, xmm6, 253  ;# constant 11111101	
			
	movaps [esp + nb410_c6], xmm4
	movaps [esp + nb410_c12], xmm6	

	movd  mm0, eax
	lea   eax, [eax + eax*2]
	
	;# move coordinates to xmm0-xmm2 
	movss xmm0, [edi + eax*4]	
	movss xmm1, [edi + eax*4 + 4]	
	movss xmm2, [edi + eax*4 + 8]	 
	
	movaps xmm4, [esp + nb410_ix]
	movaps xmm5, [esp + nb410_iy]
	movaps xmm6, [esp + nb410_iz]

	;# calc dr 
	subss xmm4, xmm0
	subss xmm5, xmm1
	subss xmm6, xmm2

	;# store dr 
	movss [esp + nb410_dx], xmm4
	movss [esp + nb410_dy], xmm5
	movss [esp + nb410_dz], xmm6
	;# square it 
	mulss xmm4,xmm4
	mulss xmm5,xmm5
	mulss xmm6,xmm6
	addss xmm4, xmm5
	addss xmm4, xmm6
	;# rsq in xmm4 

	rsqrtss xmm5, xmm4
	;# lookup seed in xmm5 
	movaps xmm2, xmm5
	mulss xmm5, xmm5
	movss xmm1, [esp + nb410_three]
	mulss xmm5, xmm4	;# rsq*lu*lu 			
	movss xmm0, [esp + nb410_half]
	subss xmm1, xmm5	;# constant 30-rsq*lu*lu 
	mulss xmm1, xmm2	
	mulss xmm0, xmm1	;# xmm0=rinv 

	mulss xmm4, xmm0	;# xmm4=r 
	movss [esp + nb410_r], xmm4
	mulss xmm4, [esp + nb410_gbscale]

	cvttss2si ebx, xmm4     ;# mm6 contain lu indices 
	cvtsi2ss xmm6, ebx
	subss xmm4, xmm6	
	movaps xmm1, xmm4	;# xmm1=eps 
	movaps xmm2, xmm1	
	mulss  xmm2, xmm2	;# xmm2=eps2 

	shl ebx, 2
	mov  esi, [ebp + nb410_GBtab]
	
	movaps xmm4, [esi + ebx*4]	
	movhlps xmm6, xmm4
	movaps xmm5, xmm4
	movaps xmm7, xmm6
	shufps xmm5, xmm5, 1
	shufps xmm7, xmm7, 1
	;# table ready in xmm4-xmm7 

	mulss  xmm6, xmm1	;# xmm6=Geps 
	mulss  xmm7, xmm2	;# xmm7=Heps2 
	addss  xmm5, xmm6
	addss  xmm5, xmm7	;# xmm5=Fp 	
	mulss  xmm7, [esp + nb410_two]	;# two*Heps2 
	movss xmm3, [esp + nb410_qq]
	addss  xmm7, xmm6
	addss  xmm7, xmm5 ;# xmm7=FF 
	mulss  xmm5, xmm1 ;# xmm5=eps*Fp 
	addss  xmm5, xmm4 ;# xmm5=VV 
	mulss  xmm5, xmm3 ;# vcoul=qq*VV  
	mulss  xmm3, xmm7 ;# fijC=FF*qq 

	movd ebx, mm0
	mov esi, [ebp + nb410_dvda]
	
	;# Calculate dVda
	xorps xmm7, xmm7
	mulss xmm3, [esp + nb410_gbscale]
	movaps xmm6, xmm3
	mulss  xmm6, [esp + nb410_r]
	addss  xmm6, xmm5