nb_kernel233_ia32_sse.intel_syntax.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,301 行 · 第 1/5 页

S
2,301
字号
	shufps xmm1, xmm5, 221  ;# constant 11011101			;# move ixO-izO to xmm4-xmm6 	movaps xmm4, [esp + nb233_ixO]	movaps xmm5, [esp + nb233_iyO]	movaps xmm6, [esp + nb233_izO]	;# calc dr 	subps xmm4, xmm0	subps xmm5, xmm1	subps xmm6, xmm2	;# store dr 	movaps [esp + nb233_dxO], xmm4	movaps [esp + nb233_dyO], xmm5	movaps [esp + nb233_dzO], xmm6	;# 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, [esp + nb233_ixH1]	movaps xmm5, [esp + nb233_iyH1]	movaps xmm6, [esp + nb233_izH1]	;# calc dr 	subps xmm4, xmm0	subps xmm5, xmm1	subps xmm6, xmm2	;# store dr 	movaps [esp + nb233_dxH1], xmm4	movaps [esp + nb233_dyH1], xmm5	movaps [esp + nb233_dzH1], xmm6	;# 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, [esp + nb233_ixH2]	movaps xmm4, [esp + nb233_iyH2]	movaps xmm5, [esp + nb233_izH2]	;# calc dr 	subps xmm3, xmm0	subps xmm4, xmm1	subps xmm5, xmm2	;# store dr 	movaps [esp + nb233_dxH2], xmm3	movaps [esp + nb233_dyH2], xmm4	movaps [esp + nb233_dzH2], xmm5	;# square it 	mulps xmm3,xmm3	mulps xmm4,xmm4	mulps xmm5,xmm5	addps xmm5, xmm4	addps xmm5, xmm3		;# move ixM-izM to xmm2-xmm4  	movaps xmm3, [esp + nb233_iyM]	movaps xmm4, [esp + nb233_izM]	subps  xmm3, xmm1	subps  xmm4, xmm2	movaps xmm2, [esp + nb233_ixM]	subps  xmm2, xmm0		;# store dr 	movaps [esp + nb233_dxM], xmm2	movaps [esp + nb233_dyM], xmm3	movaps [esp + nb233_dzM], xmm4	;# square it 	mulps xmm2,xmm2	mulps xmm3,xmm3	mulps xmm4,xmm4	addps xmm4, xmm3	addps xmm4, xmm2		;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 	movaps xmm0, xmm4	movaps xmm1, xmm5	movaps xmm2, xmm6	mulps  xmm0, [esp + nb233_krf]		mulps  xmm1, [esp + nb233_krf]		mulps  xmm2, [esp + nb233_krf]		movaps [esp + nb233_krsqM], xmm0	movaps [esp + nb233_krsqH2], xmm1	movaps [esp + nb233_krsqH1], xmm2	;# rsqH1 - seed in xmm2 	rsqrtps xmm2, xmm6	movaps  xmm3, xmm2	mulps   xmm2, xmm2	movaps  xmm0, [esp + nb233_three]	mulps   xmm2, xmm6	;# rsq*lu*lu 	subps   xmm0, xmm2	;# constant 30-rsq*lu*lu 	mulps   xmm0, xmm3	;# lu*(3-rsq*lu*lu) 	mulps   xmm0, [esp + nb233_half]	movaps  [esp + nb233_rinvH1], xmm0	;# rinvH1 	;# rsqH2 - seed to xmm2 	rsqrtps xmm2, xmm5	movaps  xmm3, xmm2	mulps   xmm2, xmm2	movaps  xmm0, [esp + nb233_three]	mulps   xmm2, xmm5	;# rsq*lu*lu 	subps   xmm0, xmm2	;# constant 30-rsq*lu*lu 	mulps   xmm0, xmm3	;# lu*(3-rsq*lu*lu) 	mulps   xmm0, [esp + nb233_half]	movaps  [esp + nb233_rinvH2], xmm0	;# rinvH2 	;# rsqM - seed to xmm2 	rsqrtps xmm2, xmm4	movaps  xmm3, xmm2	mulps   xmm2, xmm2	movaps  xmm0, [esp + nb233_three]	mulps   xmm2, xmm4	;# rsq*lu*lu 	subps   xmm0, xmm2	;# constant 30-rsq*lu*lu 	mulps   xmm0, xmm3	;# lu*(3-rsq*lu*lu) 	mulps   xmm0, [esp + nb233_half]	movaps  [esp + nb233_rinvM], xmm0		;# Do the O LJ-only interaction directly.		;# rsqO is in xmm7	rsqrtps xmm2, xmm7	movaps  xmm3, xmm2	mulps   xmm2, xmm2	movaps  xmm4, [esp + nb233_three]	mulps   xmm2, xmm7	;# rsq*lu*lu 	subps   xmm4, xmm2	;# constant 30-rsq*lu*lu 	mulps   xmm4, xmm3	;# lu*(3-rsq*lu*lu) 	mulps   xmm4, [esp + nb233_half]	movaps  xmm0, xmm4	;# xmm0=rinvO		mulps xmm7, xmm0	mulps xmm7, [esp + nb233_tsc] ;# rtab		movhlps xmm5, xmm7	cvttps2pi mm6, xmm7	cvttps2pi mm7, xmm5	;# mm6/mm7 contain lu indices 	cvtpi2ps xmm6, mm6	cvtpi2ps xmm5, mm7	movlhps xmm6, xmm5	subps  xmm7, xmm6		movaps xmm1, xmm7	;# xmm1=eps 	movaps xmm2, xmm1		mulps  xmm2, xmm2	;# xmm2=eps2 	pslld mm6, 3	pslld mm7, 3	movd mm0, eax		movd mm1, ebx	movd mm2, ecx	movd mm3, edx		mov  esi, [ebp + nb233_VFtab]	movd eax, mm6	psrlq mm6, 32	movd ecx, mm7	psrlq mm7, 32	movd ebx, mm6	movd edx, mm7		;# dispersion 	movlps xmm5, [esi + eax*4]	movlps xmm7, [esi + ecx*4]	movhps xmm5, [esi + ebx*4]	movhps xmm7, [esi + edx*4] ;# got half dispersion table 	movaps xmm4, xmm5	shufps xmm4, xmm7, 136  ;# constant 10001000	shufps xmm5, xmm7, 221  ;# constant 11011101		movlps xmm7, [esi + eax*4 + 8]	movlps xmm3, [esi + ecx*4 + 8]	movhps xmm7, [esi + ebx*4 + 8]	movhps xmm3, [esi + edx*4 + 8] ;# other half of dispersion table 	movaps xmm6, xmm7	shufps xmm6, xmm3, 136  ;# constant 10001000	shufps xmm7, xmm3, 221  ;# constant 11011101	;# dispersion 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 + nb233_two]	;# two*Heps2 	addps  xmm7, xmm6	addps  xmm7, xmm5 ;# xmm7=FF 	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 	addps  xmm5, xmm4 ;# xmm5=VV 	movaps xmm4, [esp + nb233_c6]	mulps  xmm7, xmm4	 ;# fijD 	mulps  xmm5, xmm4	 ;# Vvdw6 	mulps  xmm7, [esp + nb233_tsc]	;# put scalar force on stack Update Vvdwtot directly 	addps  xmm5, [esp + nb233_Vvdwtot]	movaps [esp + nb233_fstmp], xmm7	movaps [esp + nb233_Vvdwtot], xmm5	;# repulsion 	movlps xmm5, [esi + eax*4 + 16]	movlps xmm7, [esi + ecx*4 + 16]	movhps xmm5, [esi + ebx*4 + 16]	movhps xmm7, [esi + edx*4 + 16] ;# got half repulsion table 	movaps xmm4, xmm5	shufps xmm4, xmm7, 136  ;# constant 10001000	shufps xmm5, xmm7, 221  ;# constant 11011101	movlps xmm7, [esi + eax*4 + 24]	movlps xmm3, [esi + ecx*4 + 24]	movhps xmm7, [esi + ebx*4 + 24]	movhps xmm3, [esi + edx*4 + 24] ;# other half of repulsion table 	movaps xmm6, xmm7	shufps xmm6, xmm3, 136  ;# constant 10001000	shufps xmm7, xmm3, 221  ;# constant 11011101	;# 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 + nb233_two]	;# two*Heps2 	addps  xmm7, xmm6	addps  xmm7, xmm5 ;# xmm7=FF 	mulps  xmm5, xmm1 ;# xmm5=eps*Fp 	addps  xmm5, xmm4 ;# xmm5=VV  		movaps xmm4, [esp + nb233_c12]	mulps  xmm7, xmm4 ;# fijR 	mulps  xmm5, xmm4 ;# Vvdw12 	mulps  xmm7, [esp + nb233_tsc]	addps  xmm7, [esp + nb233_fstmp]	addps  xmm5, [esp + nb233_Vvdwtot]	movaps [esp + nb233_Vvdwtot], xmm5	movd  eax, mm0	movd  ebx, mm1	movd  ecx, mm2	movd  edx, mm3	xorps  xmm1, xmm1	mulps  xmm7, xmm0	subps  xmm1, xmm7		movaps xmm3, [esp + nb233_dxO]	movaps xmm4, [esp + nb233_dyO]	movaps xmm5, [esp + nb233_dzO]	mulps  xmm3, xmm1	mulps  xmm4, xmm1	mulps  xmm5, xmm1	;# tx in xmm3-xmm5	;# update O forces 	movaps xmm0, [esp + nb233_fixO]	movaps xmm1, [esp + nb233_fiyO]	movaps xmm2, [esp + nb233_fizO]	addps  xmm0, xmm3	addps  xmm1, xmm4	addps  xmm2, xmm5	movaps [esp + nb233_fixO], xmm0	movaps [esp + nb233_fiyO], xmm1	movaps [esp + nb233_fizO], xmm2	;# update j forces with water O 	movaps [esp + nb233_fjx], xmm3	movaps [esp + nb233_fjy], xmm4	movaps [esp + nb233_fjz], xmm5	;# Do H1 interaction	movaps  xmm7, [esp + nb233_rinvH1]	movaps  xmm4, xmm7		mulps   xmm4, xmm4	;# xmm7=rinv, xmm4=rinvsq	movaps xmm0, xmm7	movaps xmm1, [esp + nb233_krsqH1]	addps  xmm0, xmm1	subps  xmm0, [esp + nb233_crf] ;# xmm0=rinv+ krsq-crf 	mulps  xmm1, [esp + nb233_two]	subps  xmm7, xmm1	mulps  xmm0, [esp + nb233_qqH]	mulps  xmm7, [esp + nb233_qqH]	mulps  xmm4, xmm7	;# total fs H1 in xmm4 	addps  xmm0, [esp + nb233_vctot]		movaps [esp + nb233_vctot], xmm0	movaps xmm0, [esp + nb233_dxH1]	movaps xmm1, [esp + nb233_dyH1]	movaps xmm2, [esp + nb233_dzH1]	mulps  xmm0, xmm4	mulps  xmm1, xmm4	mulps  xmm2, xmm4	;# update H1 forces 	movaps xmm3, [esp + nb233_fixH1]	movaps xmm4, [esp + nb233_fiyH1]	movaps xmm7, [esp + nb233_fizH1]	addps  xmm3, xmm0	addps  xmm4, xmm1	addps  xmm7, xmm2	movaps [esp + nb233_fixH1], xmm3	movaps [esp + nb233_fiyH1], xmm4	movaps [esp + nb233_fizH1], xmm7	;# update j forces with water H1 	addps  xmm0, [esp + nb233_fjx]	addps  xmm1, [esp + nb233_fjy]	addps  xmm2, [esp + nb233_fjz]	movaps [esp + nb233_fjx], xmm0	movaps [esp + nb233_fjy], xmm1	movaps [esp + nb233_fjz], xmm2	;# Done with H1, do H2 interactions	movaps  xmm7, [esp + nb233_rinvH2]	movaps  xmm4, xmm7		mulps   xmm4, xmm4	;# xmm7=rinv, xmm4=rinvsq	movaps xmm0, xmm7	movaps xmm1, [esp + nb233_krsqH2]	addps  xmm0, xmm1	subps  xmm0, [esp + nb233_crf] ;# xmm0=rinv+ krsq-crf 	mulps  xmm1, [esp + nb233_two]	subps  xmm7, xmm1	mulps  xmm0, [esp + nb233_qqH]	mulps  xmm7, [esp + nb233_qqH]	mulps  xmm4, xmm7	;# total fs H2 in xmm4 	addps  xmm0, [esp + nb233_vctot]		movaps [esp + nb233_vctot], xmm0	movaps xmm0, [esp + nb233_dxH2]	movaps xmm1, [esp + nb233_dyH2]	movaps xmm2, [esp + nb233_dzH2]	mulps  xmm0, xmm4	mulps  xmm1, xmm4	mulps  xmm2, xmm4	;# update H2 forces 	movaps xmm3, [esp + nb233_fixH2]	movaps xmm4, [esp + nb233_fiyH2]	movaps xmm7, [esp + nb233_fizH2]	addps  xmm3, xmm0	addps  xmm4, xmm1	addps  xmm7, xmm2	movaps [esp + nb233_fixH2], xmm3	movaps [esp + nb233_fiyH2], xmm4	movaps [esp + nb233_fizH2], xmm7	addps xmm0, [esp + nb233_fjx]        addps xmm1, [esp + nb233_fjy]        addps xmm2, [esp + nb233_fjz]	movaps [esp + nb233_fjx], xmm0	movaps [esp + nb233_fjy], xmm1	movaps [esp + nb233_fjz], xmm2	;# Done with H2, do M interactions	movaps  xmm7, [esp + nb233_rinvM]	movaps  xmm4, xmm7		mulps   xmm4, xmm4	;# xmm7=rinv, xmm4=rinvsq	movaps xmm0, xmm7	movaps xmm1, [esp + nb233_krsqM]	addps  xmm0, xmm1	subps  xmm0, [esp + nb233_crf] ;# xmm0=rinv+ krsq-crf 	mulps  xmm1, [esp + nb233_two]	subps  xmm7, xmm1	mulps  xmm0, [esp + nb233_qqM]	mulps  xmm7, [esp + nb233_qqM]	mulps  xmm4, xmm7	;# total fs M in xmm4 	addps  xmm0, [esp + nb233_vctot]		movaps [esp + nb233_vctot], xmm0	movaps xmm0, [esp + nb233_dxM]	movaps xmm1, [esp + nb233_dyM]	movaps xmm2, [esp + nb233_dzM]	mulps  xmm0, xmm4	mulps  xmm1, xmm4	mulps  xmm2, xmm4		;# update M forces 	movaps xmm3, [esp + nb233_fixM]	movaps xmm4, [esp + nb233_fiyM]	movaps xmm7, [esp + nb233_fizM]	addps  xmm3, xmm0	addps  xmm4, xmm1	addps  xmm7, xmm2	movaps [esp + nb233_fixM], xmm3	movaps [esp + nb233_fiyM], xmm4	movaps [esp + nb233_fizM], xmm7	mov edi, [ebp + nb233_faction]	;# update j forces from stored values	addps xmm0, [esp + nb233_fjx]	addps xmm1, [esp + nb233_fjy]	addps xmm2, [esp + nb233_fjz]	movlps xmm4, [edi + eax*4]	movlps xmm7, [edi + ecx*4]	movhps xmm4, [edi + ebx*4]	movhps xmm7, [edi + edx*4]	movaps xmm3, xmm4	shufps xmm3, xmm7, 136  ;# constant 10001000	shufps xmm4, xmm7, 221  ;# constant 11011101		;# xmm3 has fjx, xmm4 has fjy 	subps xmm3, xmm0	subps xmm4, xmm1	;# unpack them back for storing 	movaps xmm7, xmm3	unpcklps xmm7, xmm4	unpckhps xmm3, xmm4		movlps [edi + eax*4], xmm7	movlps [edi + ecx*4], xmm3	movhps [edi + ebx*4], xmm7	movhps [edi + edx*4], xmm3	;# finally z forces 	movss  xmm0, [edi + eax*4 + 8]	movss  xmm1, [edi + ebx*4 + 8]	movss  xmm3, [edi + ecx*4 + 8]	movss  xmm4, [edi + edx*4 + 8]	subss  xmm0, xmm2	shufps xmm2, xmm2, 229  ;# constant 11100101	subss  xmm1, xmm2	shufps xmm2, xmm2, 234  ;# constant 11101010	subss  xmm3, xmm2	shufps xmm2, xmm2, 255  ;# constant 11111111	subss  xmm4, xmm2	movss  [edi + eax*4 + 8], xmm0	movss  [edi + ebx*4 + 8], xmm1	movss  [edi + ecx*4 + 8], xmm3	movss  [edi + edx*4 + 8], xmm4		;# should we do one more iteration? 	sub dword ptr [esp + nb233_innerk],  4	jl    .nb233_odd_inner	jmp   .nb233_unroll_loop.nb233_odd_inner:		add dword ptr [esp + nb233_innerk],  4	jnz   .nb233_odd_loop	jmp   .nb233_updateouterdata.nb233_odd_loop:	mov   edx, [esp + nb233_innerjjnr] 	;# pointer to jjnr[k] 	mov   eax, [edx]		add dword ptr [esp + nb233_innerjjnr],  4	 	xorps xmm4, xmm4  	;# clear reg.	movss xmm4, [esp + nb233_iqM]	mov esi, [ebp + nb233_charge] 	movhps xmm4, [esp + nb233_iqH]  ;# [qM  0  qH  qH] 	shufps xmm4, xmm4, 41	;# [0 qH qH qM]	movss xmm3, [esi + eax*4]	;# charge in xmm3 	shufps xmm3, xmm3, 0

⌨️ 快捷键说明

复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?