cbf930c6d4bcb460f0174fe5ccedb9ad1ba985da
[unres.git] / source / unres / src_MD / energy_p_new_barrier.F
1       subroutine etotal(energia)
2       implicit real*8 (a-h,o-z)
3       include 'DIMENSIONS'
4 #ifndef ISNAN
5       external proc_proc
6 #ifdef WINPGI
7 cMS$ATTRIBUTES C ::  proc_proc
8 #endif
9 #endif
10 #ifdef MPI
11       include "mpif.h"
12       double precision weights_(n_ene)
13 #endif
14       include 'COMMON.SETUP'
15       include 'COMMON.IOUNITS'
16       double precision energia(0:n_ene)
17       include 'COMMON.LOCAL'
18       include 'COMMON.FFIELD'
19       include 'COMMON.DERIV'
20       include 'COMMON.INTERACT'
21       include 'COMMON.SBRIDGE'
22       include 'COMMON.CHAIN'
23       include 'COMMON.VAR'
24       include 'COMMON.MD'
25       include 'COMMON.CONTROL'
26       include 'COMMON.TIME1'
27 #ifdef MPI      
28 c      print*,"ETOTAL Processor",fg_rank," absolute rank",myrank,
29 c     & " nfgtasks",nfgtasks
30       call flush(iout)
31       if (nfgtasks.gt.1) then
32 #ifdef MPI
33         time00=MPI_Wtime()
34 #else
35         time00=tcpu()
36 #endif
37 C FG slaves call the following matching MPI_Bcast in ERGASTULUM
38         if (fg_rank.eq.0) then
39           call MPI_Bcast(0,1,MPI_INTEGER,king,FG_COMM,IERROR)
40 c          print *,"Processor",myrank," BROADCAST iorder"
41 C FG master sets up the WEIGHTS_ array which will be broadcast to the 
42 C FG slaves as WEIGHTS array.
43           weights_(1)=wsc
44           weights_(2)=wscp
45           weights_(3)=welec
46           weights_(4)=wcorr
47           weights_(5)=wcorr5
48           weights_(6)=wcorr6
49           weights_(7)=wel_loc
50           weights_(8)=wturn3
51           weights_(9)=wturn4
52           weights_(10)=wturn6
53           weights_(11)=wang
54           weights_(12)=wscloc
55           weights_(13)=wtor
56           weights_(14)=wtor_d
57           weights_(15)=wstrain
58           weights_(16)=wvdwpp
59           weights_(17)=wbond
60           weights_(18)=scal14
61           weights_(21)=wsccor
62           weights_(22)=wsct
63 C FG Master broadcasts the WEIGHTS_ array
64           call MPI_Bcast(weights_(1),n_ene,
65      &        MPI_DOUBLE_PRECISION,king,FG_COMM,IERROR)
66         else
67 C FG slaves receive the WEIGHTS array
68           call MPI_Bcast(weights(1),n_ene,
69      &        MPI_DOUBLE_PRECISION,king,FG_COMM,IERROR)
70           wsc=weights(1)
71           wscp=weights(2)
72           welec=weights(3)
73           wcorr=weights(4)
74           wcorr5=weights(5)
75           wcorr6=weights(6)
76           wel_loc=weights(7)
77           wturn3=weights(8)
78           wturn4=weights(9)
79           wturn6=weights(10)
80           wang=weights(11)
81           wscloc=weights(12)
82           wtor=weights(13)
83           wtor_d=weights(14)
84           wstrain=weights(15)
85           wvdwpp=weights(16)
86           wbond=weights(17)
87           scal14=weights(18)
88           wsccor=weights(21)
89           wsct=weights(22)
90         endif
91         time_Bcast=time_Bcast+MPI_Wtime()-time00
92         time_Bcastw=time_Bcastw+MPI_Wtime()-time00
93 c        call chainbuild_cart
94       endif
95 c      write(iout,*) 'Processor',myrank,' calling etotal ipot=',ipot
96 c      print *,'Processor',myrank,' nnt=',nnt,' nct=',nct
97 #else
98 c      if (modecalc.eq.12.or.modecalc.eq.14) then
99 c        call int_from_cart1(.false.)
100 c      endif
101 #endif     
102 #ifndef DFA
103       edfadis=0.0d0
104       edfator=0.0d0
105       edfanei=0.0d0
106       edfabet=0.0d0
107 #endif
108 #ifdef TIMING
109 #ifdef MPI
110       time00=MPI_Wtime()
111 #else
112       time00=tcpu()
113 #endif
114 #endif
115
116 C Compute the side-chain and electrostatic interaction energy
117 C
118       goto (101,102,103,104,105,106) ipot
119 C Lennard-Jones potential.
120   101 call elj(evdw,evdw_p,evdw_m)
121 cd    print '(a)','Exit ELJ'
122       goto 107
123 C Lennard-Jones-Kihara potential (shifted).
124   102 call eljk(evdw,evdw_p,evdw_m)
125       goto 107
126 C Berne-Pechukas potential (dilated LJ, angular dependence).
127   103 call ebp(evdw,evdw_p,evdw_m)
128       goto 107
129 C Gay-Berne potential (shifted LJ, angular dependence).
130   104 call egb(evdw,evdw_p,evdw_m)
131       goto 107
132 C Gay-Berne-Vorobjev potential (shifted LJ, angular dependence).
133   105 call egbv(evdw,evdw_p,evdw_m)
134       goto 107
135 C Soft-sphere potential
136   106 call e_softsphere(evdw)
137 C
138 C Calculate electrostatic (H-bonding) energy of the main chain.
139 C
140   107 continue
141 #ifdef DFA
142 C     BARTEK for dfa test!
143       if (wdfa_dist.gt.0) then 
144         call edfad(edfadis)
145       else
146         edfadis=0
147       endif
148 c      print*, 'edfad is finished!', edfadis
149       if (wdfa_tor.gt.0) then
150         call edfat(edfator)
151       else
152         edfator=0
153       endif
154 c      print*, 'edfat is finished!', edfator
155       if (wdfa_nei.gt.0) then
156         call edfan(edfanei)
157       else
158         edfanei=0
159       endif    
160 c      print*, 'edfan is finished!', edfanei
161       if (wdfa_beta.gt.0) then 
162         call edfab(edfabet)
163       else
164         edfabet=0
165       endif
166 #endif
167 c      print*, 'edfab is finished!', edfabet
168 cmc
169 cmc Sep-06: egb takes care of dynamic ss bonds too
170 cmc
171 c      if (dyn_ss) call dyn_set_nss
172
173 c      print *,"Processor",myrank," computed USCSC"
174 #ifdef TIMING
175 #ifdef MPI
176       time01=MPI_Wtime() 
177 #else
178       time00=tcpu()
179 #endif
180 #endif
181       call vec_and_deriv
182 #ifdef TIMING
183 #ifdef MPI
184       time_vec=time_vec+MPI_Wtime()-time01
185 #else
186       time_vec=time_vec+tcpu()-time01
187 #endif
188 #endif
189 c      print *,"Processor",myrank," left VEC_AND_DERIV"
190       if (ipot.lt.6) then
191 #ifdef SPLITELE
192          if (welec.gt.0d0.or.wvdwpp.gt.0d0.or.wel_loc.gt.0d0.or.
193      &       wturn3.gt.0d0.or.wturn4.gt.0d0 .or. wcorr.gt.0.0d0
194      &       .or. wcorr4.gt.0.0d0 .or. wcorr5.gt.0.d0
195      &       .or. wcorr6.gt.0.0d0 .or. wturn6.gt.0.0d0 ) then
196 #else
197          if (welec.gt.0d0.or.wel_loc.gt.0d0.or.
198      &       wturn3.gt.0d0.or.wturn4.gt.0d0 .or. wcorr.gt.0.0d0
199      &       .or. wcorr4.gt.0.0d0 .or. wcorr5.gt.0.d0 
200      &       .or. wcorr6.gt.0.0d0 .or. wturn6.gt.0.0d0 ) then
201 #endif
202             call eelec(ees,evdw1,eel_loc,eello_turn3,eello_turn4)
203          else
204             ees=0.0d0
205             evdw1=0.0d0
206             eel_loc=0.0d0
207             eello_turn3=0.0d0
208             eello_turn4=0.0d0
209          endif
210       else
211 c        write (iout,*) "Soft-spheer ELEC potential"
212         call eelec_soft_sphere(ees,evdw1,eel_loc,eello_turn3,
213      &   eello_turn4)
214       endif
215 c      print *,"Processor",myrank," computed UELEC"
216 C
217 C Calculate excluded-volume interaction energy between peptide groups
218 C and side chains.
219 C
220       if (ipot.lt.6) then
221        if(wscp.gt.0d0) then
222         call escp(evdw2,evdw2_14)
223        else
224         evdw2=0
225         evdw2_14=0
226        endif
227       else
228 c        write (iout,*) "Soft-sphere SCP potential"
229         call escp_soft_sphere(evdw2,evdw2_14)
230       endif
231 c
232 c Calculate the bond-stretching energy
233 c
234       call ebond(estr)
235
236 C Calculate the disulfide-bridge and other energy and the contributions
237 C from other distance constraints.
238 cd    print *,'Calling EHPB'
239       call edis(ehpb)
240 cd    print *,'EHPB exitted succesfully.'
241 C
242 C Calculate the virtual-bond-angle energy.
243 C
244       if (wang.gt.0d0) then
245         call ebend(ebe)
246       else
247         ebe=0
248       endif
249 c      print *,"Processor",myrank," computed UB"
250 C
251 C Calculate the SC local energy.
252 C
253       call esc(escloc)
254 c      print *,"Processor",myrank," computed USC"
255 C
256 C Calculate the virtual-bond torsional energy.
257 C
258 cd    print *,'nterm=',nterm
259       if (wtor.gt.0) then
260        call etor(etors,edihcnstr)
261       else
262        etors=0
263        edihcnstr=0
264       endif
265
266       if (constr_homology.ge.1.and.waga_homology(iset).ne.0d0) then
267         call e_modeller(ehomology_constr)
268 c        print *,'iset=',iset,'me=',me,ehomology_constr,
269 c     &  'Processor',fg_rank,' CG group',kolor,
270 c     &  ' absolute rank',MyRank
271       else
272         ehomology_constr=0.0d0
273       endif
274
275
276 c      write(iout,*) ehomology_constr
277 c      print *,"Processor",myrank," computed Utor"
278 C
279 C 6/23/01 Calculate double-torsional energy
280 C
281       if (wtor_d.gt.0) then
282        call etor_d(etors_d)
283       else
284        etors_d=0
285       endif
286 c      print *,"Processor",myrank," computed Utord"
287 C
288 C 21/5/07 Calculate local sicdechain correlation energy
289 C
290       if (wsccor.gt.0.0d0) then
291         call eback_sc_corr(esccor)
292       else
293         esccor=0.0d0
294       endif
295 c      print *,"Processor",myrank," computed Usccorr"
296
297 C 12/1/95 Multi-body terms
298 C
299       n_corr=0
300       n_corr1=0
301       if ((wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0 
302      &    .or. wturn6.gt.0.0d0) .and. ipot.lt.6) then
303          call multibody_eello(ecorr,ecorr5,ecorr6,eturn6,n_corr,n_corr1)
304 cd         write(2,*)'multibody_eello n_corr=',n_corr,' n_corr1=',n_corr1,
305 cd     &" ecorr",ecorr," ecorr5",ecorr5," ecorr6",ecorr6," eturn6",eturn6
306       else
307          ecorr=0.0d0
308          ecorr5=0.0d0
309          ecorr6=0.0d0
310          eturn6=0.0d0
311       endif
312       if ((wcorr4.eq.0.0d0 .and. wcorr.gt.0.0d0) .and. ipot.lt.6) then
313          call multibody_hb(ecorr,ecorr5,ecorr6,n_corr,n_corr1)
314 cd         write (iout,*) "multibody_hb ecorr",ecorr
315       endif
316 c      print *,"Processor",myrank," computed Ucorr"
317
318 C If performing constraint dynamics, call the constraint energy
319 C  after the equilibration time
320       if(usampl.and.totT.gt.eq_time) then
321 c         write (iout,*) "CALL TO ECONSTR_BACK"
322          call EconstrQ   
323          call Econstr_back
324       else
325          Uconst=0.0d0
326          Uconst_back=0.0d0
327       endif
328 #ifdef TIMING
329 #ifdef MPI
330       time_enecalc=time_enecalc+MPI_Wtime()-time00
331 #else
332       time_enecalc=time_enecalc+tcpu()-time00
333 #endif
334 #endif
335 c      print *,"Processor",myrank," computed Uconstr"
336 #ifdef TIMING
337 #ifdef MPI
338       time00=MPI_Wtime()
339 #else
340       time00=tcpu()
341 #endif
342 #endif
343 c
344 C Sum the energies
345 C
346       energia(1)=evdw
347 #ifdef SCP14
348       energia(2)=evdw2-evdw2_14
349       energia(18)=evdw2_14
350 #else
351       energia(2)=evdw2
352       energia(18)=0.0d0
353 #endif
354 #ifdef SPLITELE
355       energia(3)=ees
356       energia(16)=evdw1
357 #else
358       energia(3)=ees+evdw1
359       energia(16)=0.0d0
360 #endif
361       energia(4)=ecorr
362       energia(5)=ecorr5
363       energia(6)=ecorr6
364       energia(7)=eel_loc
365       energia(8)=eello_turn3
366       energia(9)=eello_turn4
367       energia(10)=eturn6
368       energia(11)=ebe
369       energia(12)=escloc
370       energia(13)=etors
371       energia(14)=etors_d
372       energia(15)=ehpb
373       energia(19)=edihcnstr
374       energia(17)=estr
375       energia(20)=Uconst+Uconst_back
376       energia(21)=esccor
377       energia(22)=evdw_p
378       energia(23)=evdw_m
379       energia(24)=ehomology_constr
380       energia(25)=edfadis
381       energia(26)=edfator
382       energia(27)=edfanei
383       energia(28)=edfabet
384 c      print *," Processor",myrank," calls SUM_ENERGY"
385       call sum_energy(energia,.true.)
386       if (dyn_ss) call dyn_set_nss
387 c      print *," Processor",myrank," left SUM_ENERGY"
388 #ifdef TIMING
389 #ifdef MPI
390       time_sumene=time_sumene+MPI_Wtime()-time00
391 #else
392       time_sumene=time_sumene+tcpu()-time00
393 #endif
394 #endif
395       return
396       end
397 c-------------------------------------------------------------------------------
398       subroutine sum_energy(energia,reduce)
399       implicit real*8 (a-h,o-z)
400       include 'DIMENSIONS'
401 #ifndef ISNAN
402       external proc_proc
403 #ifdef WINPGI
404 cMS$ATTRIBUTES C ::  proc_proc
405 #endif
406 #endif
407 #ifdef MPI
408       include "mpif.h"
409 #endif
410       include 'COMMON.SETUP'
411       include 'COMMON.IOUNITS'
412       double precision energia(0:n_ene),enebuff(0:n_ene+1)
413       include 'COMMON.FFIELD'
414       include 'COMMON.DERIV'
415       include 'COMMON.INTERACT'
416       include 'COMMON.SBRIDGE'
417       include 'COMMON.CHAIN'
418       include 'COMMON.VAR'
419       include 'COMMON.CONTROL'
420       include 'COMMON.TIME1'
421       logical reduce
422 #ifdef MPI
423       if (nfgtasks.gt.1 .and. reduce) then
424 #ifdef DEBUG
425         write (iout,*) "energies before REDUCE"
426         call enerprint(energia)
427         call flush(iout)
428 #endif
429         do i=0,n_ene
430           enebuff(i)=energia(i)
431         enddo
432         time00=MPI_Wtime()
433         call MPI_Barrier(FG_COMM,IERR)
434         time_barrier_e=time_barrier_e+MPI_Wtime()-time00
435         time00=MPI_Wtime()
436         call MPI_Reduce(enebuff(0),energia(0),n_ene+1,
437      &    MPI_DOUBLE_PRECISION,MPI_SUM,king,FG_COMM,IERR)
438 #ifdef DEBUG
439         write (iout,*) "energies after REDUCE"
440         call enerprint(energia)
441         call flush(iout)
442 #endif
443         time_Reduce=time_Reduce+MPI_Wtime()-time00
444       endif
445       if (fg_rank.eq.0) then
446 #endif
447 #ifdef TSCSC
448       evdw=energia(22)+wsct*energia(23)
449 #else
450       evdw=energia(1)
451 #endif
452 #ifdef SCP14
453       evdw2=energia(2)+energia(18)
454       evdw2_14=energia(18)
455 #else
456       evdw2=energia(2)
457 #endif
458 #ifdef SPLITELE
459       ees=energia(3)
460       evdw1=energia(16)
461 #else
462       ees=energia(3)
463       evdw1=0.0d0
464 #endif
465       ecorr=energia(4)
466       ecorr5=energia(5)
467       ecorr6=energia(6)
468       eel_loc=energia(7)
469       eello_turn3=energia(8)
470       eello_turn4=energia(9)
471       eturn6=energia(10)
472       ebe=energia(11)
473       escloc=energia(12)
474       etors=energia(13)
475       etors_d=energia(14)
476       ehpb=energia(15)
477       edihcnstr=energia(19)
478       estr=energia(17)
479       Uconst=energia(20)
480       esccor=energia(21)
481       ehomology_constr=energia(24)
482       edfadis=energia(25)
483       edfator=energia(26)
484       edfanei=energia(27)
485       edfabet=energia(28)
486 #ifdef SPLITELE
487       etot=wsc*evdw+wscp*evdw2+welec*ees+wvdwpp*evdw1
488      & +wang*ebe+wtor*etors+wscloc*escloc
489      & +wstrain*ehpb+wcorr*ecorr+wcorr5*ecorr5
490      & +wcorr6*ecorr6+wturn4*eello_turn4+wturn3*eello_turn3
491      & +wturn6*eturn6+wel_loc*eel_loc+edihcnstr+wtor_d*etors_d
492      & +wbond*estr+Uconst+wsccor*esccor+ehomology_constr
493      & +wdfa_dist*edfadis+wdfa_tor*edfator+wdfa_nei*edfanei
494      & +wdfa_beta*edfabet    
495 #else
496       etot=wsc*evdw+wscp*evdw2+welec*(ees+evdw1)
497      & +wang*ebe+wtor*etors+wscloc*escloc
498      & +wstrain*ehpb+wcorr*ecorr+wcorr5*ecorr5
499      & +wcorr6*ecorr6+wturn4*eello_turn4+wturn3*eello_turn3
500      & +wturn6*eturn6+wel_loc*eel_loc+edihcnstr+wtor_d*etors_d
501      & +wbond*estr+Uconst+wsccor*esccor+ehomology_constr
502      & +wdfa_dist*edfadis+wdfa_tor*edfator+wdfa_nei*edfanei
503      & +wdfa_beta*edfabet    
504 #endif
505       energia(0)=etot
506 c detecting NaNQ
507 #ifdef ISNAN
508 #ifdef AIX
509       if (isnan(etot).ne.0) energia(0)=1.0d+99
510 #else
511       if (isnan(etot)) energia(0)=1.0d+99
512 #endif
513 #else
514       i=0
515 #ifdef WINPGI
516       idumm=proc_proc(etot,i)
517 #else
518       call proc_proc(etot,i)
519 #endif
520       if(i.eq.1)energia(0)=1.0d+99
521 #endif
522 #ifdef MPI
523       endif
524 #endif
525       return
526       end
527 c-------------------------------------------------------------------------------
528       subroutine sum_gradient
529       implicit real*8 (a-h,o-z)
530       include 'DIMENSIONS'
531 #ifndef ISNAN
532       external proc_proc
533 #ifdef WINPGI
534 cMS$ATTRIBUTES C ::  proc_proc
535 #endif
536 #endif
537 #ifdef MPI
538       include 'mpif.h'
539 #endif
540       double precision gradbufc(3,maxres),gradbufx(3,maxres),
541      &  glocbuf(4*maxres),gradbufc_sum(3,maxres),gloc_scbuf(3,maxres)
542       include 'COMMON.SETUP'
543       include 'COMMON.IOUNITS'
544       include 'COMMON.FFIELD'
545       include 'COMMON.DERIV'
546       include 'COMMON.INTERACT'
547       include 'COMMON.SBRIDGE'
548       include 'COMMON.CHAIN'
549       include 'COMMON.VAR'
550       include 'COMMON.CONTROL'
551       include 'COMMON.TIME1'
552       include 'COMMON.MAXGRAD'
553       include 'COMMON.SCCOR'
554       include 'COMMON.MD'
555 #ifdef TIMING
556 #ifdef MPI
557       time01=MPI_Wtime()
558 #else
559       time01=tcpu()
560 #endif
561 #endif
562 #ifdef DEBUG
563       write (iout,*) "sum_gradient gvdwc, gvdwx"
564       do i=1,nres
565         write (iout,'(i3,3f10.5,5x,3f10.5,5x,3f10.5,5x,3f10.5)') 
566      &   i,(gvdwx(j,i),j=1,3),(gvdwcT(j,i),j=1,3),(gvdwc(j,i),j=1,3),
567      &   (gvdwcT(j,i),j=1,3)
568       enddo
569       call flush(iout)
570 #endif
571 #ifdef MPI
572 C FG slaves call the following matching MPI_Bcast in ERGASTULUM
573         if (nfgtasks.gt.1 .and. fg_rank.eq.0) 
574      &    call MPI_Bcast(1,1,MPI_INTEGER,king,FG_COMM,IERROR)
575 #endif
576 C
577 C 9/29/08 AL Transform parts of gradients in site coordinates to the gradient
578 C            in virtual-bond-vector coordinates
579 C
580 #ifdef DEBUG
581 c      write (iout,*) "gel_loc gel_loc_long and gel_loc_loc"
582 c      do i=1,nres-1
583 c        write (iout,'(i5,3f10.5,2x,3f10.5,2x,f10.5)') 
584 c     &   i,(gel_loc(j,i),j=1,3),(gel_loc_long(j,i),j=1,3),gel_loc_loc(i)
585 c      enddo
586 c      write (iout,*) "gel_loc_tur3 gel_loc_turn4"
587 c      do i=1,nres-1
588 c        write (iout,'(i5,3f10.5,2x,f10.5)') 
589 c     &  i,(gcorr4_turn(j,i),j=1,3),gel_loc_turn4(i)
590 c      enddo
591       write (iout,*) "gradcorr5 gradcorr5_long gradcorr5_loc"
592       do i=1,nres
593         write (iout,'(i3,3f10.5,5x,3f10.5,5x,f10.5)') 
594      &   i,(gradcorr5(j,i),j=1,3),(gradcorr5_long(j,i),j=1,3),
595      &   g_corr5_loc(i)
596       enddo
597       call flush(iout)
598 #endif
599 #ifdef SPLITELE
600 #ifdef TSCSC
601       do i=1,nct
602         do j=1,3
603           gradbufc(j,i)=wsc*gvdwc(j,i)+wsc*wscT*gvdwcT(j,i)+
604      &                wscp*(gvdwc_scp(j,i)+gvdwc_scpp(j,i))+
605      &                welec*gelc_long(j,i)+wvdwpp*gvdwpp(j,i)+
606      &                wel_loc*gel_loc_long(j,i)+
607      &                wcorr*gradcorr_long(j,i)+
608      &                wcorr5*gradcorr5_long(j,i)+
609      &                wcorr6*gradcorr6_long(j,i)+
610      &                wturn6*gcorr6_turn_long(j,i)+
611      &                wstrain*ghpbc(j,i)+
612      &                wdfa_dist*gdfad(j,i)+
613      &                wdfa_tor*gdfat(j,i)+
614      &                wdfa_nei*gdfan(j,i)+
615      &                wdfa_beta*gdfab(j,i)
616         enddo
617       enddo 
618 #else
619       do i=1,nct
620         do j=1,3
621           gradbufc(j,i)=wsc*gvdwc(j,i)+
622      &                wscp*(gvdwc_scp(j,i)+gvdwc_scpp(j,i))+
623      &                welec*gelc_long(j,i)+wvdwpp*gvdwpp(j,i)+
624      &                wel_loc*gel_loc_long(j,i)+
625      &                wcorr*gradcorr_long(j,i)+
626      &                wcorr5*gradcorr5_long(j,i)+
627      &                wcorr6*gradcorr6_long(j,i)+
628      &                wturn6*gcorr6_turn_long(j,i)+
629      &                wstrain*ghpbc(j,i)+
630      &                wdfa_dist*gdfad(j,i)+
631      &                wdfa_tor*gdfat(j,i)+
632      &                wdfa_nei*gdfan(j,i)+
633      &                wdfa_beta*gdfab(j,i)
634         enddo
635       enddo 
636 #endif
637 #else
638       do i=1,nct
639         do j=1,3
640           gradbufc(j,i)=wsc*gvdwc(j,i)+
641      &                wscp*(gvdwc_scp(j,i)+gvdwc_scpp(j,i))+
642      &                welec*gelc_long(j,i)+
643      &                wbond*gradb(j,i)+
644      &                wel_loc*gel_loc_long(j,i)+
645      &                wcorr*gradcorr_long(j,i)+
646      &                wcorr5*gradcorr5_long(j,i)+
647      &                wcorr6*gradcorr6_long(j,i)+
648      &                wturn6*gcorr6_turn_long(j,i)+
649      &                wstrain*ghpbc(j,i)+
650      &                wdfa_dist*gdfad(j,i)+
651      &                wdfa_tor*gdfat(j,i)+
652      &                wdfa_nei*gdfan(j,i)+
653      &                wdfa_beta*gdfab(j,i)
654         enddo
655       enddo 
656 #endif
657 #ifdef MPI
658       if (nfgtasks.gt.1) then
659       time00=MPI_Wtime()
660 #ifdef DEBUG
661       write (iout,*) "gradbufc before allreduce"
662       do i=1,nres
663         write (iout,'(i3,3f10.5)') i,(gradbufc(j,i),j=1,3)
664       enddo
665       call flush(iout)
666 #endif
667       do i=1,nres
668         do j=1,3
669           gradbufc_sum(j,i)=gradbufc(j,i)
670         enddo
671       enddo
672 c      call MPI_AllReduce(gradbufc(1,1),gradbufc_sum(1,1),3*nres,
673 c     &    MPI_DOUBLE_PRECISION,MPI_SUM,FG_COMM,IERR)
674 c      time_reduce=time_reduce+MPI_Wtime()-time00
675 #ifdef DEBUG
676 c      write (iout,*) "gradbufc_sum after allreduce"
677 c      do i=1,nres
678 c        write (iout,'(i3,3f10.5)') i,(gradbufc_sum(j,i),j=1,3)
679 c      enddo
680 c      call flush(iout)
681 #endif
682 #ifdef TIMING
683 c      time_allreduce=time_allreduce+MPI_Wtime()-time00
684 #endif
685       do i=nnt,nres
686         do k=1,3
687           gradbufc(k,i)=0.0d0
688         enddo
689       enddo
690 #ifdef DEBUG
691       write (iout,*) "igrad_start",igrad_start," igrad_end",igrad_end
692       write (iout,*) (i," jgrad_start",jgrad_start(i),
693      &                  " jgrad_end  ",jgrad_end(i),
694      &                  i=igrad_start,igrad_end)
695 #endif
696 c
697 c Obsolete and inefficient code; we can make the effort O(n) and, therefore,
698 c do not parallelize this part.
699 c
700 c      do i=igrad_start,igrad_end
701 c        do j=jgrad_start(i),jgrad_end(i)
702 c          do k=1,3
703 c            gradbufc(k,i)=gradbufc(k,i)+gradbufc_sum(k,j)
704 c          enddo
705 c        enddo
706 c      enddo
707       do j=1,3
708         gradbufc(j,nres-1)=gradbufc_sum(j,nres)
709       enddo
710       do i=nres-2,nnt,-1
711         do j=1,3
712           gradbufc(j,i)=gradbufc(j,i+1)+gradbufc_sum(j,i+1)
713         enddo
714       enddo
715 #ifdef DEBUG
716       write (iout,*) "gradbufc after summing"
717       do i=1,nres
718         write (iout,'(i3,3f10.5)') i,(gradbufc(j,i),j=1,3)
719       enddo
720       call flush(iout)
721 #endif
722       else
723 #endif
724 #ifdef DEBUG
725       write (iout,*) "gradbufc"
726       do i=1,nres
727         write (iout,'(i3,3f10.5)') i,(gradbufc(j,i),j=1,3)
728       enddo
729       call flush(iout)
730 #endif
731       do i=1,nres
732         do j=1,3
733           gradbufc_sum(j,i)=gradbufc(j,i)
734           gradbufc(j,i)=0.0d0
735         enddo
736       enddo
737       do j=1,3
738         gradbufc(j,nres-1)=gradbufc_sum(j,nres)
739       enddo
740       do i=nres-2,nnt,-1
741         do j=1,3
742           gradbufc(j,i)=gradbufc(j,i+1)+gradbufc_sum(j,i+1)
743         enddo
744       enddo
745 c      do i=nnt,nres-1
746 c        do k=1,3
747 c          gradbufc(k,i)=0.0d0
748 c        enddo
749 c        do j=i+1,nres
750 c          do k=1,3
751 c            gradbufc(k,i)=gradbufc(k,i)+gradbufc(k,j)
752 c          enddo
753 c        enddo
754 c      enddo
755 #ifdef DEBUG
756       write (iout,*) "gradbufc after summing"
757       do i=1,nres
758         write (iout,'(i3,3f10.5)') i,(gradbufc(j,i),j=1,3)
759       enddo
760       call flush(iout)
761 #endif
762 #ifdef MPI
763       endif
764 #endif
765       do k=1,3
766         gradbufc(k,nres)=0.0d0
767       enddo
768       do i=1,nct
769         do j=1,3
770 #ifdef SPLITELE
771           gradc(j,i,icg)=gradbufc(j,i)+welec*gelc(j,i)+
772      &                wel_loc*gel_loc(j,i)+
773      &                0.5d0*(wscp*gvdwc_scpp(j,i)+
774      &                welec*gelc_long(j,i)+wvdwpp*gvdwpp(j,i)+
775      &                wel_loc*gel_loc_long(j,i)+
776      &                wcorr*gradcorr_long(j,i)+
777      &                wcorr5*gradcorr5_long(j,i)+
778      &                wcorr6*gradcorr6_long(j,i)+
779      &                wturn6*gcorr6_turn_long(j,i))+
780      &                wbond*gradb(j,i)+
781      &                wcorr*gradcorr(j,i)+
782      &                wturn3*gcorr3_turn(j,i)+
783      &                wturn4*gcorr4_turn(j,i)+
784      &                wcorr5*gradcorr5(j,i)+
785      &                wcorr6*gradcorr6(j,i)+
786      &                wturn6*gcorr6_turn(j,i)+
787      &                wsccor*gsccorc(j,i)
788      &               +wscloc*gscloc(j,i)
789 #else
790           gradc(j,i,icg)=gradbufc(j,i)+welec*gelc(j,i)+
791      &                wel_loc*gel_loc(j,i)+
792      &                0.5d0*(wscp*gvdwc_scpp(j,i)+
793      &                welec*gelc_long(j,i)+
794      &                wel_loc*gel_loc_long(j,i)+
795      &                wcorr*gcorr_long(j,i)+
796      &                wcorr5*gradcorr5_long(j,i)+
797      &                wcorr6*gradcorr6_long(j,i)+
798      &                wturn6*gcorr6_turn_long(j,i))+
799      &                wbond*gradb(j,i)+
800      &                wcorr*gradcorr(j,i)+
801      &                wturn3*gcorr3_turn(j,i)+
802      &                wturn4*gcorr4_turn(j,i)+
803      &                wcorr5*gradcorr5(j,i)+
804      &                wcorr6*gradcorr6(j,i)+
805      &                wturn6*gcorr6_turn(j,i)+
806      &                wsccor*gsccorc(j,i)
807      &               +wscloc*gscloc(j,i)
808 #endif
809 #ifdef TSCSC
810           gradx(j,i,icg)=wsc*gvdwx(j,i)+wsc*wscT*gvdwxT(j,i)+
811      &                  wscp*gradx_scp(j,i)+
812      &                  wbond*gradbx(j,i)+
813      &                  wstrain*ghpbx(j,i)+wcorr*gradxorr(j,i)+
814      &                  wsccor*gsccorx(j,i)
815      &                 +wscloc*gsclocx(j,i)
816 #else
817           gradx(j,i,icg)=wsc*gvdwx(j,i)+wscp*gradx_scp(j,i)+
818      &                  wbond*gradbx(j,i)+
819      &                  wstrain*ghpbx(j,i)+wcorr*gradxorr(j,i)+
820      &                  wsccor*gsccorx(j,i)
821      &                 +wscloc*gsclocx(j,i)
822 #endif
823         enddo
824       enddo 
825       if (constr_homology.gt.0.and.waga_homology(iset).ne.0d0) then
826         do i=1,nct
827           do j=1,3
828             gradc(j,i,icg)=gradc(j,i,icg)+duscdiff(j,i)
829             gradx(j,i,icg)=gradx(j,i,icg)+duscdiffx(j,i)
830           enddo
831         enddo
832       endif
833 #ifdef DEBUG
834       write (iout,*) "gloc before adding corr"
835       do i=1,4*nres
836         write (iout,*) i,gloc(i,icg)
837       enddo
838 #endif
839       do i=1,nres-3
840         gloc(i,icg)=gloc(i,icg)+wcorr*gcorr_loc(i)
841      &   +wcorr5*g_corr5_loc(i)
842      &   +wcorr6*g_corr6_loc(i)
843      &   +wturn4*gel_loc_turn4(i)
844      &   +wturn3*gel_loc_turn3(i)
845      &   +wturn6*gel_loc_turn6(i)
846      &   +wel_loc*gel_loc_loc(i)
847       enddo
848 #ifdef DEBUG
849       write (iout,*) "gloc after adding corr"
850       do i=1,4*nres
851         write (iout,*) i,gloc(i,icg)
852       enddo
853 #endif
854 #ifdef MPI
855       if (nfgtasks.gt.1) then
856         do j=1,3
857           do i=1,nres
858             gradbufc(j,i)=gradc(j,i,icg)
859             gradbufx(j,i)=gradx(j,i,icg)
860           enddo
861         enddo
862         do i=1,4*nres
863           glocbuf(i)=gloc(i,icg)
864         enddo
865 #ifdef DEBUG
866       write (iout,*) "gloc_sc before reduce"
867       do i=1,nres
868        do j=1,3
869         write (iout,*) i,j,gloc_sc(j,i,icg)
870        enddo
871       enddo
872 #endif
873         do i=1,nres
874          do j=1,3
875           gloc_scbuf(j,i)=gloc_sc(j,i,icg)
876          enddo
877         enddo
878         time00=MPI_Wtime()
879         call MPI_Barrier(FG_COMM,IERR)
880         time_barrier_g=time_barrier_g+MPI_Wtime()-time00
881         time00=MPI_Wtime()
882         call MPI_Reduce(gradbufc(1,1),gradc(1,1,icg),3*nres,
883      &    MPI_DOUBLE_PRECISION,MPI_SUM,king,FG_COMM,IERR)
884         call MPI_Reduce(gradbufx(1,1),gradx(1,1,icg),3*nres,
885      &    MPI_DOUBLE_PRECISION,MPI_SUM,king,FG_COMM,IERR)
886         call MPI_Reduce(glocbuf(1),gloc(1,icg),4*nres,
887      &    MPI_DOUBLE_PRECISION,MPI_SUM,king,FG_COMM,IERR)
888         call MPI_Reduce(gloc_scbuf(1,1),gloc_sc(1,1,icg),3*nres,
889      &    MPI_DOUBLE_PRECISION,MPI_SUM,king,FG_COMM,IERR)
890         time_reduce=time_reduce+MPI_Wtime()-time00
891 #ifdef DEBUG
892       write (iout,*) "gloc_sc after reduce"
893       do i=1,nres
894        do j=1,3
895         write (iout,*) i,j,gloc_sc(j,i,icg)
896        enddo
897       enddo
898 #endif
899 #ifdef DEBUG
900       write (iout,*) "gloc after reduce"
901       do i=1,4*nres
902         write (iout,*) i,gloc(i,icg)
903       enddo
904 #endif
905       endif
906 #endif
907       if (gnorm_check) then
908 c
909 c Compute the maximum elements of the gradient
910 c
911       gvdwc_max=0.0d0
912       gvdwc_scp_max=0.0d0
913       gelc_max=0.0d0
914       gvdwpp_max=0.0d0
915       gradb_max=0.0d0
916       ghpbc_max=0.0d0
917       gradcorr_max=0.0d0
918       gel_loc_max=0.0d0
919       gcorr3_turn_max=0.0d0
920       gcorr4_turn_max=0.0d0
921       gradcorr5_max=0.0d0
922       gradcorr6_max=0.0d0
923       gcorr6_turn_max=0.0d0
924       gsccorc_max=0.0d0
925       gscloc_max=0.0d0
926       gvdwx_max=0.0d0
927       gradx_scp_max=0.0d0
928       ghpbx_max=0.0d0
929       gradxorr_max=0.0d0
930       gsccorx_max=0.0d0
931       gsclocx_max=0.0d0
932       do i=1,nct
933         gvdwc_norm=dsqrt(scalar(gvdwc(1,i),gvdwc(1,i)))
934         if (gvdwc_norm.gt.gvdwc_max) gvdwc_max=gvdwc_norm
935 #ifdef TSCSC
936         gvdwc_norm=dsqrt(scalar(gvdwcT(1,i),gvdwcT(1,i)))
937         if (gvdwc_norm.gt.gvdwc_max) gvdwc_max=gvdwc_norm          
938 #endif
939         gvdwc_scp_norm=dsqrt(scalar(gvdwc_scp(1,i),gvdwc_scp(1,i)))
940         if (gvdwc_scp_norm.gt.gvdwc_scp_max) 
941      &   gvdwc_scp_max=gvdwc_scp_norm
942         gelc_norm=dsqrt(scalar(gelc(1,i),gelc(1,i)))
943         if (gelc_norm.gt.gelc_max) gelc_max=gelc_norm
944         gvdwpp_norm=dsqrt(scalar(gvdwpp(1,i),gvdwpp(1,i)))
945         if (gvdwpp_norm.gt.gvdwpp_max) gvdwpp_max=gvdwpp_norm
946         gradb_norm=dsqrt(scalar(gradb(1,i),gradb(1,i)))
947         if (gradb_norm.gt.gradb_max) gradb_max=gradb_norm
948         ghpbc_norm=dsqrt(scalar(ghpbc(1,i),ghpbc(1,i)))
949         if (ghpbc_norm.gt.ghpbc_max) ghpbc_max=ghpbc_norm
950         gradcorr_norm=dsqrt(scalar(gradcorr(1,i),gradcorr(1,i)))
951         if (gradcorr_norm.gt.gradcorr_max) gradcorr_max=gradcorr_norm
952         gel_loc_norm=dsqrt(scalar(gel_loc(1,i),gel_loc(1,i)))
953         if (gel_loc_norm.gt.gel_loc_max) gel_loc_max=gel_loc_norm
954         gcorr3_turn_norm=dsqrt(scalar(gcorr3_turn(1,i),
955      &    gcorr3_turn(1,i)))
956         if (gcorr3_turn_norm.gt.gcorr3_turn_max) 
957      &    gcorr3_turn_max=gcorr3_turn_norm
958         gcorr4_turn_norm=dsqrt(scalar(gcorr4_turn(1,i),
959      &    gcorr4_turn(1,i)))
960         if (gcorr4_turn_norm.gt.gcorr4_turn_max) 
961      &    gcorr4_turn_max=gcorr4_turn_norm
962         gradcorr5_norm=dsqrt(scalar(gradcorr5(1,i),gradcorr5(1,i)))
963         if (gradcorr5_norm.gt.gradcorr5_max) 
964      &    gradcorr5_max=gradcorr5_norm
965         gradcorr6_norm=dsqrt(scalar(gradcorr6(1,i),gradcorr6(1,i)))
966         if (gradcorr6_norm.gt.gradcorr6_max) gcorr6_max=gradcorr6_norm
967         gcorr6_turn_norm=dsqrt(scalar(gcorr6_turn(1,i),
968      &    gcorr6_turn(1,i)))
969         if (gcorr6_turn_norm.gt.gcorr6_turn_max) 
970      &    gcorr6_turn_max=gcorr6_turn_norm
971         gsccorr_norm=dsqrt(scalar(gsccorc(1,i),gsccorc(1,i)))
972         if (gsccorr_norm.gt.gsccorr_max) gsccorr_max=gsccorr_norm
973         gscloc_norm=dsqrt(scalar(gscloc(1,i),gscloc(1,i)))
974         if (gscloc_norm.gt.gscloc_max) gscloc_max=gscloc_norm
975         gvdwx_norm=dsqrt(scalar(gvdwx(1,i),gvdwx(1,i)))
976         if (gvdwx_norm.gt.gvdwx_max) gvdwx_max=gvdwx_norm
977 #ifdef TSCSC
978         gvdwx_norm=dsqrt(scalar(gvdwxT(1,i),gvdwxT(1,i)))
979         if (gvdwx_norm.gt.gvdwx_max) gvdwx_max=gvdwx_norm
980 #endif
981         gradx_scp_norm=dsqrt(scalar(gradx_scp(1,i),gradx_scp(1,i)))
982         if (gradx_scp_norm.gt.gradx_scp_max) 
983      &    gradx_scp_max=gradx_scp_norm
984         ghpbx_norm=dsqrt(scalar(ghpbx(1,i),ghpbx(1,i)))
985         if (ghpbx_norm.gt.ghpbx_max) ghpbx_max=ghpbx_norm
986         gradxorr_norm=dsqrt(scalar(gradxorr(1,i),gradxorr(1,i)))
987         if (gradxorr_norm.gt.gradxorr_max) gradxorr_max=gradxorr_norm
988         gsccorrx_norm=dsqrt(scalar(gsccorx(1,i),gsccorx(1,i)))
989         if (gsccorrx_norm.gt.gsccorrx_max) gsccorrx_max=gsccorrx_norm
990         gsclocx_norm=dsqrt(scalar(gsclocx(1,i),gsclocx(1,i)))
991         if (gsclocx_norm.gt.gsclocx_max) gsclocx_max=gsclocx_norm
992       enddo 
993       if (gradout) then
994 #ifdef AIX
995         open(istat,file=statname,position="append")
996 #else
997         open(istat,file=statname,access="append")
998 #endif
999         write (istat,'(1h#,21f10.2)') gvdwc_max,gvdwc_scp_max,
1000      &     gelc_max,gvdwpp_max,gradb_max,ghpbc_max,
1001      &     gradcorr_max,gel_loc_max,gcorr3_turn_max,gcorr4_turn_max,
1002      &     gradcorr5_max,gradcorr6_max,gcorr6_turn_max,gsccorc_max,
1003      &     gscloc_max,gvdwx_max,gradx_scp_max,ghpbx_max,gradxorr_max,
1004      &     gsccorx_max,gsclocx_max
1005         close(istat)
1006         if (gvdwc_max.gt.1.0d4) then
1007           write (iout,*) "gvdwc gvdwx gradb gradbx"
1008           do i=nnt,nct
1009             write(iout,'(i5,4(3f10.2,5x))') i,(gvdwc(j,i),gvdwx(j,i),
1010      &        gradb(j,i),gradbx(j,i),j=1,3)
1011           enddo
1012           call pdbout(0.0d0,'cipiszcze',iout)
1013           call flush(iout)
1014         endif
1015       endif
1016       endif
1017 #ifdef DEBUG
1018       write (iout,*) "gradc gradx gloc"
1019       do i=1,nres
1020         write (iout,'(i5,3f10.5,5x,3f10.5,5x,f10.5)') 
1021      &   i,(gradc(j,i,icg),j=1,3),(gradx(j,i,icg),j=1,3),gloc(i,icg)
1022       enddo 
1023 #endif
1024 #ifdef TIMING
1025 #ifdef MPI
1026       time_sumgradient=time_sumgradient+MPI_Wtime()-time01
1027 #else
1028       time_sumgradient=time_sumgradient+tcpu()-time01
1029 #endif
1030 #endif
1031       return
1032       end
1033 c-------------------------------------------------------------------------------
1034       subroutine rescale_weights(t_bath)
1035       implicit real*8 (a-h,o-z)
1036       include 'DIMENSIONS'
1037       include 'COMMON.IOUNITS'
1038       include 'COMMON.FFIELD'
1039       include 'COMMON.SBRIDGE'
1040       double precision kfac /2.4d0/
1041       double precision x,x2,x3,x4,x5,licznik /1.12692801104297249644/
1042 c      facT=temp0/t_bath
1043 c      facT=2*temp0/(t_bath+temp0)
1044       if (rescale_mode.eq.0) then
1045         facT=1.0d0
1046         facT2=1.0d0
1047         facT3=1.0d0
1048         facT4=1.0d0
1049         facT5=1.0d0
1050       else if (rescale_mode.eq.1) then
1051         facT=kfac/(kfac-1.0d0+t_bath/temp0)
1052         facT2=kfac**2/(kfac**2-1.0d0+(t_bath/temp0)**2)
1053         facT3=kfac**3/(kfac**3-1.0d0+(t_bath/temp0)**3)
1054         facT4=kfac**4/(kfac**4-1.0d0+(t_bath/temp0)**4)
1055         facT5=kfac**5/(kfac**5-1.0d0+(t_bath/temp0)**5)
1056       else if (rescale_mode.eq.2) then
1057         x=t_bath/temp0
1058         x2=x*x
1059         x3=x2*x
1060         x4=x3*x
1061         x5=x4*x
1062         facT=licznik/dlog(dexp(x)+dexp(-x))
1063         facT2=licznik/dlog(dexp(x2)+dexp(-x2))
1064         facT3=licznik/dlog(dexp(x3)+dexp(-x3))
1065         facT4=licznik/dlog(dexp(x4)+dexp(-x4))
1066         facT5=licznik/dlog(dexp(x5)+dexp(-x5))
1067       else
1068         write (iout,*) "Wrong RESCALE_MODE",rescale_mode
1069         write (*,*) "Wrong RESCALE_MODE",rescale_mode
1070 #ifdef MPI
1071        call MPI_Finalize(MPI_COMM_WORLD,IERROR)
1072 #endif
1073        stop 555
1074       endif
1075       welec=weights(3)*fact
1076       wcorr=weights(4)*fact3
1077       wcorr5=weights(5)*fact4
1078       wcorr6=weights(6)*fact5
1079       wel_loc=weights(7)*fact2
1080       wturn3=weights(8)*fact2
1081       wturn4=weights(9)*fact3
1082       wturn6=weights(10)*fact5
1083       wtor=weights(13)*fact
1084       wtor_d=weights(14)*fact2
1085       wsccor=weights(21)*fact
1086 #ifdef TSCSC
1087 c      wsct=t_bath/temp0
1088       wsct=(320.0+80.0*dtanh((t_bath-320.0)/80.0))/320.0
1089 #endif
1090       return
1091       end
1092 C------------------------------------------------------------------------
1093       subroutine enerprint(energia)
1094       implicit real*8 (a-h,o-z)
1095       include 'DIMENSIONS'
1096       include 'COMMON.IOUNITS'
1097       include 'COMMON.FFIELD'
1098       include 'COMMON.SBRIDGE'
1099       include 'COMMON.MD'
1100       double precision energia(0:n_ene)
1101       etot=energia(0)
1102 #ifdef TSCSC
1103       evdw=energia(22)+wsct*energia(23)
1104 #else
1105       evdw=energia(1)
1106 #endif
1107       evdw2=energia(2)
1108 #ifdef SCP14
1109       evdw2=energia(2)+energia(18)
1110 #else
1111       evdw2=energia(2)
1112 #endif
1113       ees=energia(3)
1114 #ifdef SPLITELE
1115       evdw1=energia(16)
1116 #endif
1117       ecorr=energia(4)
1118       ecorr5=energia(5)
1119       ecorr6=energia(6)
1120       eel_loc=energia(7)
1121       eello_turn3=energia(8)
1122       eello_turn4=energia(9)
1123       eello_turn6=energia(10)
1124       ebe=energia(11)
1125       escloc=energia(12)
1126       etors=energia(13)
1127       etors_d=energia(14)
1128       ehpb=energia(15)
1129       edihcnstr=energia(19)
1130       estr=energia(17)
1131       Uconst=energia(20)
1132       esccor=energia(21)
1133       ehomology_constr=energia(24)
1134 C     Bartek
1135       edfadis = energia(25)
1136       edfator = energia(26)
1137       edfanei = energia(27)
1138       edfabet = energia(28)
1139
1140 #ifdef SPLITELE
1141       write (iout,10) evdw,wsc,evdw2,wscp,ees,welec,evdw1,wvdwpp,
1142      &  estr,wbond,ebe,wang,
1143      &  escloc,wscloc,etors,wtor,etors_d,wtor_d,ehpb,wstrain,
1144      &  ecorr,wcorr,
1145      &  ecorr5,wcorr5,ecorr6,wcorr6,eel_loc,wel_loc,eello_turn3,wturn3,
1146      &  eello_turn4,wturn4,eello_turn6,wturn6,esccor,wsccor,
1147      &  edihcnstr,ehomology_constr, ebr*nss,
1148      &  Uconst,edfadis,wdfa_dist,edfator,wdfa_tor,edfanei,wdfa_nei,
1149      &  edfabet,wdfa_beta,etot
1150    10 format (/'Virtual-chain energies:'//
1151      & 'EVDW=  ',1pE16.6,' WEIGHT=',1pE16.6,' (SC-SC)'/
1152      & 'EVDW2= ',1pE16.6,' WEIGHT=',1pE16.6,' (SC-p)'/
1153      & 'EES=   ',1pE16.6,' WEIGHT=',1pE16.6,' (p-p)'/
1154      & 'EVDWPP=',1pE16.6,' WEIGHT=',1pE16.6,' (p-p VDW)'/
1155      & 'ESTR=  ',1pE16.6,' WEIGHT=',1pE16.6,' (stretching)'/
1156      & 'EBE=   ',1pE16.6,' WEIGHT=',1pE16.6,' (bending)'/
1157      & 'ESC=   ',1pE16.6,' WEIGHT=',1pE16.6,' (SC local)'/
1158      & 'ETORS= ',1pE16.6,' WEIGHT=',1pE16.6,' (torsional)'/
1159      & 'ETORSD=',1pE16.6,' WEIGHT=',1pE16.6,' (double torsional)'/
1160      & 'EHPB=  ',1pE16.6,' WEIGHT=',1pE16.6,
1161      & ' (SS bridges & dist. cnstr.)'/
1162      & 'ECORR4=',1pE16.6,' WEIGHT=',1pE16.6,' (multi-body)'/
1163      & 'ECORR5=',1pE16.6,' WEIGHT=',1pE16.6,' (multi-body)'/
1164      & 'ECORR6=',1pE16.6,' WEIGHT=',1pE16.6,' (multi-body)'/
1165      & 'EELLO= ',1pE16.6,' WEIGHT=',1pE16.6,' (electrostatic-local)'/
1166      & 'ETURN3=',1pE16.6,' WEIGHT=',1pE16.6,' (turns, 3rd order)'/
1167      & 'ETURN4=',1pE16.6,' WEIGHT=',1pE16.6,' (turns, 4th order)'/
1168      & 'ETURN6=',1pE16.6,' WEIGHT=',1pE16.6,' (turns, 6th order)'/
1169      & 'ESCCOR=',1pE16.6,' WEIGHT=',1pE16.6,' (backbone-rotamer corr)'/
1170      & 'EDIHC= ',1pE16.6,' (dihedral angle constraints)'/
1171      & 'H_CONS=',1pE16.6,' (Homology model constraints energy)'/
1172      & 'ESS=   ',1pE16.6,' (disulfide-bridge intrinsic energy)'/
1173      & 'UCONST= ',1pE16.6,' (Constraint energy)'/ 
1174      & 'EDFAD= ',1pE16.6,' WEIGHT=',1pE16.6,' (DFA distance energy)'/ 
1175      & 'EDFAT= ',1pE16.6,' WEIGHT=',1pE16.6,' (DFA torsion energy)'/ 
1176      & 'EDFAN= ',1pE16.6,' WEIGHT=',1pE16.6,' (DFA NCa energy)'/ 
1177      & 'EDFAB= ',1pE16.6,' WEIGHT=',1pE16.6,' (DFA Beta energy)'/ 
1178      & 'ETOT=  ',1pE16.6,' (total)')
1179 #else
1180       write (iout,10) evdw,wsc,evdw2,wscp,ees,welec,
1181      &  estr,wbond,ebe,wang,
1182      &  escloc,wscloc,etors,wtor,etors_d,wtor_d,ehpb,wstrain,
1183      &  ecorr,wcorr,
1184      &  ecorr5,wcorr5,ecorr6,wcorr6,eel_loc,wel_loc,eello_turn3,wturn3,
1185      &  eello_turn4,wturn4,eello_turn6,wturn6,esccor,wsccro,edihcnstr,
1186      &  ehomology_constr,ebr*nss,Uconst,edfadis,wdfa_dist,edfator,
1187      &  wdfa_tor,edfanei,wdfa_nei,edfabet,wdfa_beta,
1188      &  etot
1189    10 format (/'Virtual-chain energies:'//
1190      & 'EVDW=  ',1pE16.6,' WEIGHT=',1pD16.6,' (SC-SC)'/
1191      & 'EVDW2= ',1pE16.6,' WEIGHT=',1pD16.6,' (SC-p)'/
1192      & 'EES=   ',1pE16.6,' WEIGHT=',1pD16.6,' (p-p)'/
1193      & 'ESTR=  ',1pE16.6,' WEIGHT=',1pD16.6,' (stretching)'/
1194      & 'EBE=   ',1pE16.6,' WEIGHT=',1pD16.6,' (bending)'/
1195      & 'ESC=   ',1pE16.6,' WEIGHT=',1pD16.6,' (SC local)'/
1196      & 'ETORS= ',1pE16.6,' WEIGHT=',1pD16.6,' (torsional)'/
1197      & 'ETORSD=',1pE16.6,' WEIGHT=',1pD16.6,' (double torsional)'/
1198      & 'EHBP=  ',1pE16.6,' WEIGHT=',1pD16.6,
1199      & ' (SS bridges & dist. cnstr.)'/
1200      & 'ECORR4=',1pE16.6,' WEIGHT=',1pD16.6,' (multi-body)'/
1201      & 'ECORR5=',1pE16.6,' WEIGHT=',1pD16.6,' (multi-body)'/
1202      & 'ECORR6=',1pE16.6,' WEIGHT=',1pD16.6,' (multi-body)'/
1203      & 'EELLO= ',1pE16.6,' WEIGHT=',1pD16.6,' (electrostatic-local)'/
1204      & 'ETURN3=',1pE16.6,' WEIGHT=',1pD16.6,' (turns, 3rd order)'/
1205      & 'ETURN4=',1pE16.6,' WEIGHT=',1pD16.6,' (turns, 4th order)'/
1206      & 'ETURN6=',1pE16.6,' WEIGHT=',1pD16.6,' (turns, 6th order)'/
1207      & 'ESCCOR=',1pE16.6,' WEIGHT=',1pD16.6,' (backbone-rotamer corr)'/
1208      & 'EDIHC= ',1pE16.6,' (dihedral angle constraints)'/
1209      & 'H_CONS=',1pE16.6,' (Homology model constraints energy)'/
1210      & 'ESS=   ',1pE16.6,' (disulfide-bridge intrinsic energy)'/
1211      & 'UCONST=',1pE16.6,' (Constraint energy)'/ 
1212      & 'EDFAD= ',1pE16.6,' WEIGHT=',1pD16.6,' (DFA distance energy)'/ 
1213      & 'EDFAT= ',1pE16.6,' WEIGHT=',1pD16.6,' (DFA torsion energy)'/ 
1214      & 'EDFAN= ',1pE16.6,' WEIGHT=',1pD16.6,' (DFA NCa energy)'/ 
1215      & 'EDFAB= ',1pE16.6,' WEIGHT=',1pD16.6,' (DFA Beta energy)'/ 
1216      & 'ETOT=  ',1pE16.6,' (total)')
1217 #endif
1218       return
1219       end
1220 C-----------------------------------------------------------------------
1221       subroutine elj(evdw,evdw_p,evdw_m)
1222 C
1223 C This subroutine calculates the interaction energy of nonbonded side chains
1224 C assuming the LJ potential of interaction.
1225 C
1226       implicit real*8 (a-h,o-z)
1227       include 'DIMENSIONS'
1228       parameter (accur=1.0d-10)
1229       include 'COMMON.GEO'
1230       include 'COMMON.VAR'
1231       include 'COMMON.LOCAL'
1232       include 'COMMON.CHAIN'
1233       include 'COMMON.DERIV'
1234       include 'COMMON.INTERACT'
1235       include 'COMMON.TORSION'
1236       include 'COMMON.SBRIDGE'
1237       include 'COMMON.NAMES'
1238       include 'COMMON.IOUNITS'
1239       include 'COMMON.CONTACTS'
1240       dimension gg(3)
1241 c      write(iout,*)'Entering ELJ nnt=',nnt,' nct=',nct,' expon=',expon
1242       evdw=0.0D0
1243       do i=iatsc_s,iatsc_e
1244         itypi=itype(i)
1245         itypi1=itype(i+1)
1246         xi=c(1,nres+i)
1247         yi=c(2,nres+i)
1248         zi=c(3,nres+i)
1249 C Change 12/1/95
1250         num_conti=0
1251 C
1252 C Calculate SC interaction energy.
1253 C
1254         do iint=1,nint_gr(i)
1255 cd        write (iout,*) 'i=',i,' iint=',iint,' istart=',istart(i,iint),
1256 cd   &                  'iend=',iend(i,iint)
1257           do j=istart(i,iint),iend(i,iint)
1258             itypj=itype(j)
1259             xj=c(1,nres+j)-xi
1260             yj=c(2,nres+j)-yi
1261             zj=c(3,nres+j)-zi
1262 C Change 12/1/95 to calculate four-body interactions
1263             rij=xj*xj+yj*yj+zj*zj
1264             rrij=1.0D0/rij
1265 c           write (iout,*)'i=',i,' j=',j,' itypi=',itypi,' itypj=',itypj
1266             eps0ij=eps(itypi,itypj)
1267             fac=rrij**expon2
1268             e1=fac*fac*aa(itypi,itypj)
1269             e2=fac*bb(itypi,itypj)
1270             evdwij=e1+e2
1271 cd          sigm=dabs(aa(itypi,itypj)/bb(itypi,itypj))**(1.0D0/6.0D0)
1272 cd          epsi=bb(itypi,itypj)**2/aa(itypi,itypj)
1273 cd          write (iout,'(2(a3,i3,2x),6(1pd12.4)/2(3(1pd12.4),5x)/)')
1274 cd   &        restyp(itypi),i,restyp(itypj),j,aa(itypi,itypj),
1275 cd   &        bb(itypi,itypj),1.0D0/dsqrt(rrij),evdwij,epsi,sigm,
1276 cd   &        (c(k,i),k=1,3),(c(k,j),k=1,3)
1277 #ifdef TSCSC
1278             if (bb(itypi,itypj).gt.0) then
1279                evdw_p=evdw_p+evdwij
1280             else
1281                evdw_m=evdw_m+evdwij
1282             endif
1283 #else
1284             evdw=evdw+evdwij
1285 #endif
1286
1287 C Calculate the components of the gradient in DC and X
1288 C
1289             fac=-rrij*(e1+evdwij)
1290             gg(1)=xj*fac
1291             gg(2)=yj*fac
1292             gg(3)=zj*fac
1293 #ifdef TSCSC
1294             if (bb(itypi,itypj).gt.0.0d0) then
1295               do k=1,3
1296                 gvdwx(k,i)=gvdwx(k,i)-gg(k)
1297                 gvdwx(k,j)=gvdwx(k,j)+gg(k)
1298                 gvdwc(k,i)=gvdwc(k,i)-gg(k)
1299                 gvdwc(k,j)=gvdwc(k,j)+gg(k)
1300               enddo
1301             else
1302               do k=1,3
1303                 gvdwxT(k,i)=gvdwxT(k,i)-gg(k)
1304                 gvdwxT(k,j)=gvdwxT(k,j)+gg(k)
1305                 gvdwcT(k,i)=gvdwcT(k,i)-gg(k)
1306                 gvdwcT(k,j)=gvdwcT(k,j)+gg(k)
1307               enddo
1308             endif
1309 #else
1310             do k=1,3
1311               gvdwx(k,i)=gvdwx(k,i)-gg(k)
1312               gvdwx(k,j)=gvdwx(k,j)+gg(k)
1313               gvdwc(k,i)=gvdwc(k,i)-gg(k)
1314               gvdwc(k,j)=gvdwc(k,j)+gg(k)
1315             enddo
1316 #endif
1317 cgrad            do k=i,j-1
1318 cgrad              do l=1,3
1319 cgrad                gvdwc(l,k)=gvdwc(l,k)+gg(l)
1320 cgrad              enddo
1321 cgrad            enddo
1322 C
1323 C 12/1/95, revised on 5/20/97
1324 C
1325 C Calculate the contact function. The ith column of the array JCONT will 
1326 C contain the numbers of atoms that make contacts with the atom I (of numbers
1327 C greater than I). The arrays FACONT and GACONT will contain the values of
1328 C the contact function and its derivative.
1329 C
1330 C Uncomment next line, if the correlation interactions include EVDW explicitly.
1331 c           if (j.gt.i+1 .and. evdwij.le.0.0D0) then
1332 C Uncomment next line, if the correlation interactions are contact function only
1333             if (j.gt.i+1.and. eps0ij.gt.0.0D0) then
1334               rij=dsqrt(rij)
1335               sigij=sigma(itypi,itypj)
1336               r0ij=rs0(itypi,itypj)
1337 C
1338 C Check whether the SC's are not too far to make a contact.
1339 C
1340               rcut=1.5d0*r0ij
1341               call gcont(rij,rcut,1.0d0,0.2d0*rcut,fcont,fprimcont)
1342 C Add a new contact, if the SC's are close enough, but not too close (r<sigma).
1343 C
1344               if (fcont.gt.0.0D0) then
1345 C If the SC-SC distance if close to sigma, apply spline.
1346 cAdam           call gcont(-rij,-1.03d0*sigij,2.0d0*sigij,1.0d0,
1347 cAdam &             fcont1,fprimcont1)
1348 cAdam           fcont1=1.0d0-fcont1
1349 cAdam           if (fcont1.gt.0.0d0) then
1350 cAdam             fprimcont=fprimcont*fcont1+fcont*fprimcont1
1351 cAdam             fcont=fcont*fcont1
1352 cAdam           endif
1353 C Uncomment following 4 lines to have the geometric average of the epsilon0's
1354 cga             eps0ij=1.0d0/dsqrt(eps0ij)
1355 cga             do k=1,3
1356 cga               gg(k)=gg(k)*eps0ij
1357 cga             enddo
1358 cga             eps0ij=-evdwij*eps0ij
1359 C Uncomment for AL's type of SC correlation interactions.
1360 cadam           eps0ij=-evdwij
1361                 num_conti=num_conti+1
1362                 jcont(num_conti,i)=j
1363                 facont(num_conti,i)=fcont*eps0ij
1364                 fprimcont=eps0ij*fprimcont/rij
1365                 fcont=expon*fcont
1366 cAdam           gacont(1,num_conti,i)=-fprimcont*xj+fcont*gg(1)
1367 cAdam           gacont(2,num_conti,i)=-fprimcont*yj+fcont*gg(2)
1368 cAdam           gacont(3,num_conti,i)=-fprimcont*zj+fcont*gg(3)
1369 C Uncomment following 3 lines for Skolnick's type of SC correlation.
1370                 gacont(1,num_conti,i)=-fprimcont*xj
1371                 gacont(2,num_conti,i)=-fprimcont*yj
1372                 gacont(3,num_conti,i)=-fprimcont*zj
1373 cd              write (iout,'(2i5,2f10.5)') i,j,rij,facont(num_conti,i)
1374 cd              write (iout,'(2i3,3f10.5)') 
1375 cd   &           i,j,(gacont(kk,num_conti,i),kk=1,3)
1376               endif
1377             endif
1378           enddo      ! j
1379         enddo        ! iint
1380 C Change 12/1/95
1381         num_cont(i)=num_conti
1382       enddo          ! i
1383       do i=1,nct
1384         do j=1,3
1385           gvdwc(j,i)=expon*gvdwc(j,i)
1386           gvdwx(j,i)=expon*gvdwx(j,i)
1387         enddo
1388       enddo
1389 C******************************************************************************
1390 C
1391 C                              N O T E !!!
1392 C
1393 C To save time, the factor of EXPON has been extracted from ALL components
1394 C of GVDWC and GRADX. Remember to multiply them by this factor before further 
1395 C use!
1396 C
1397 C******************************************************************************
1398       return
1399       end
1400 C-----------------------------------------------------------------------------
1401       subroutine eljk(evdw,evdw_p,evdw_m)
1402 C
1403 C This subroutine calculates the interaction energy of nonbonded side chains
1404 C assuming the LJK potential of interaction.
1405 C
1406       implicit real*8 (a-h,o-z)
1407       include 'DIMENSIONS'
1408       include 'COMMON.GEO'
1409       include 'COMMON.VAR'
1410       include 'COMMON.LOCAL'
1411       include 'COMMON.CHAIN'
1412       include 'COMMON.DERIV'
1413       include 'COMMON.INTERACT'
1414       include 'COMMON.IOUNITS'
1415       include 'COMMON.NAMES'
1416       dimension gg(3)
1417       logical scheck
1418 c     print *,'Entering ELJK nnt=',nnt,' nct=',nct,' expon=',expon
1419       evdw=0.0D0
1420       do i=iatsc_s,iatsc_e
1421         itypi=itype(i)
1422         itypi1=itype(i+1)
1423         xi=c(1,nres+i)
1424         yi=c(2,nres+i)
1425         zi=c(3,nres+i)
1426 C
1427 C Calculate SC interaction energy.
1428 C
1429         do iint=1,nint_gr(i)
1430           do j=istart(i,iint),iend(i,iint)
1431             itypj=itype(j)
1432             xj=c(1,nres+j)-xi
1433             yj=c(2,nres+j)-yi
1434             zj=c(3,nres+j)-zi
1435             rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
1436             fac_augm=rrij**expon
1437             e_augm=augm(itypi,itypj)*fac_augm
1438             r_inv_ij=dsqrt(rrij)
1439             rij=1.0D0/r_inv_ij 
1440             r_shift_inv=1.0D0/(rij+r0(itypi,itypj)-sigma(itypi,itypj))
1441             fac=r_shift_inv**expon
1442             e1=fac*fac*aa(itypi,itypj)
1443             e2=fac*bb(itypi,itypj)
1444             evdwij=e_augm+e1+e2
1445 cd          sigm=dabs(aa(itypi,itypj)/bb(itypi,itypj))**(1.0D0/6.0D0)
1446 cd          epsi=bb(itypi,itypj)**2/aa(itypi,itypj)
1447 cd          write (iout,'(2(a3,i3,2x),8(1pd12.4)/2(3(1pd12.4),5x)/)')
1448 cd   &        restyp(itypi),i,restyp(itypj),j,aa(itypi,itypj),
1449 cd   &        bb(itypi,itypj),augm(itypi,itypj),epsi,sigm,
1450 cd   &        sigma(itypi,itypj),1.0D0/dsqrt(rrij),evdwij,
1451 cd   &        (c(k,i),k=1,3),(c(k,j),k=1,3)
1452 #ifdef TSCSC
1453             if (bb(itypi,itypj).gt.0) then
1454                evdw_p=evdw_p+evdwij
1455             else
1456                evdw_m=evdw_m+evdwij
1457             endif
1458 #else
1459             evdw=evdw+evdwij
1460 #endif
1461
1462 C Calculate the components of the gradient in DC and X
1463 C
1464             fac=-2.0D0*rrij*e_augm-r_inv_ij*r_shift_inv*(e1+e1+e2)
1465             gg(1)=xj*fac
1466             gg(2)=yj*fac
1467             gg(3)=zj*fac
1468 #ifdef TSCSC
1469             if (bb(itypi,itypj).gt.0.0d0) then
1470               do k=1,3
1471                 gvdwx(k,i)=gvdwx(k,i)-gg(k)
1472                 gvdwx(k,j)=gvdwx(k,j)+gg(k)
1473                 gvdwc(k,i)=gvdwc(k,i)-gg(k)
1474                 gvdwc(k,j)=gvdwc(k,j)+gg(k)
1475               enddo
1476             else
1477               do k=1,3
1478                 gvdwxT(k,i)=gvdwxT(k,i)-gg(k)
1479                 gvdwxT(k,j)=gvdwxT(k,j)+gg(k)
1480                 gvdwcT(k,i)=gvdwcT(k,i)-gg(k)
1481                 gvdwcT(k,j)=gvdwcT(k,j)+gg(k)
1482               enddo
1483             endif
1484 #else
1485             do k=1,3
1486               gvdwx(k,i)=gvdwx(k,i)-gg(k)
1487               gvdwx(k,j)=gvdwx(k,j)+gg(k)
1488               gvdwc(k,i)=gvdwc(k,i)-gg(k)
1489               gvdwc(k,j)=gvdwc(k,j)+gg(k)
1490             enddo
1491 #endif
1492 cgrad            do k=i,j-1
1493 cgrad              do l=1,3
1494 cgrad                gvdwc(l,k)=gvdwc(l,k)+gg(l)
1495 cgrad              enddo
1496 cgrad            enddo
1497           enddo      ! j
1498         enddo        ! iint
1499       enddo          ! i
1500       do i=1,nct
1501         do j=1,3
1502           gvdwc(j,i)=expon*gvdwc(j,i)
1503           gvdwx(j,i)=expon*gvdwx(j,i)
1504         enddo
1505       enddo
1506       return
1507       end
1508 C-----------------------------------------------------------------------------
1509       subroutine ebp(evdw,evdw_p,evdw_m)
1510 C
1511 C This subroutine calculates the interaction energy of nonbonded side chains
1512 C assuming the Berne-Pechukas potential of interaction.
1513 C
1514       implicit real*8 (a-h,o-z)
1515       include 'DIMENSIONS'
1516       include 'COMMON.GEO'
1517       include 'COMMON.VAR'
1518       include 'COMMON.LOCAL'
1519       include 'COMMON.CHAIN'
1520       include 'COMMON.DERIV'
1521       include 'COMMON.NAMES'
1522       include 'COMMON.INTERACT'
1523       include 'COMMON.IOUNITS'
1524       include 'COMMON.CALC'
1525       common /srutu/ icall
1526 c     double precision rrsave(maxdim)
1527       logical lprn
1528       evdw=0.0D0
1529 c     print *,'Entering EBP nnt=',nnt,' nct=',nct,' expon=',expon
1530       evdw=0.0D0
1531 c     if (icall.eq.0) then
1532 c       lprn=.true.
1533 c     else
1534         lprn=.false.
1535 c     endif
1536       ind=0
1537       do i=iatsc_s,iatsc_e
1538         itypi=itype(i)
1539         itypi1=itype(i+1)
1540         xi=c(1,nres+i)
1541         yi=c(2,nres+i)
1542         zi=c(3,nres+i)
1543         dxi=dc_norm(1,nres+i)
1544         dyi=dc_norm(2,nres+i)
1545         dzi=dc_norm(3,nres+i)
1546 c        dsci_inv=dsc_inv(itypi)
1547         dsci_inv=vbld_inv(i+nres)
1548 C
1549 C Calculate SC interaction energy.
1550 C
1551         do iint=1,nint_gr(i)
1552           do j=istart(i,iint),iend(i,iint)
1553             ind=ind+1
1554             itypj=itype(j)
1555 c            dscj_inv=dsc_inv(itypj)
1556             dscj_inv=vbld_inv(j+nres)
1557             chi1=chi(itypi,itypj)
1558             chi2=chi(itypj,itypi)
1559             chi12=chi1*chi2
1560             chip1=chip(itypi)
1561             chip2=chip(itypj)
1562             chip12=chip1*chip2
1563             alf1=alp(itypi)
1564             alf2=alp(itypj)
1565             alf12=0.5D0*(alf1+alf2)
1566 C For diagnostics only!!!
1567 c           chi1=0.0D0
1568 c           chi2=0.0D0
1569 c           chi12=0.0D0
1570 c           chip1=0.0D0
1571 c           chip2=0.0D0
1572 c           chip12=0.0D0
1573 c           alf1=0.0D0
1574 c           alf2=0.0D0
1575 c           alf12=0.0D0
1576             xj=c(1,nres+j)-xi
1577             yj=c(2,nres+j)-yi
1578             zj=c(3,nres+j)-zi
1579             dxj=dc_norm(1,nres+j)
1580             dyj=dc_norm(2,nres+j)
1581             dzj=dc_norm(3,nres+j)
1582             rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
1583 cd          if (icall.eq.0) then
1584 cd            rrsave(ind)=rrij
1585 cd          else
1586 cd            rrij=rrsave(ind)
1587 cd          endif
1588             rij=dsqrt(rrij)
1589 C Calculate the angle-dependent terms of energy & contributions to derivatives.
1590             call sc_angular
1591 C Calculate whole angle-dependent part of epsilon and contributions
1592 C to its derivatives
1593             fac=(rrij*sigsq)**expon2
1594             e1=fac*fac*aa(itypi,itypj)
1595             e2=fac*bb(itypi,itypj)
1596             evdwij=eps1*eps2rt*eps3rt*(e1+e2)
1597             eps2der=evdwij*eps3rt
1598             eps3der=evdwij*eps2rt
1599             evdwij=evdwij*eps2rt*eps3rt
1600 #ifdef TSCSC
1601             if (bb(itypi,itypj).gt.0) then
1602                evdw_p=evdw_p+evdwij
1603             else
1604                evdw_m=evdw_m+evdwij
1605             endif
1606 #else
1607             evdw=evdw+evdwij
1608 #endif
1609             if (lprn) then
1610             sigm=dabs(aa(itypi,itypj)/bb(itypi,itypj))**(1.0D0/6.0D0)
1611             epsi=bb(itypi,itypj)**2/aa(itypi,itypj)
1612 cd            write (iout,'(2(a3,i3,2x),15(0pf7.3))')
1613 cd     &        restyp(itypi),i,restyp(itypj),j,
1614 cd     &        epsi,sigm,chi1,chi2,chip1,chip2,
1615 cd     &        eps1,eps2rt**2,eps3rt**2,1.0D0/dsqrt(sigsq),
1616 cd     &        om1,om2,om12,1.0D0/dsqrt(rrij),
1617 cd     &        evdwij
1618             endif
1619 C Calculate gradient components.
1620             e1=e1*eps1*eps2rt**2*eps3rt**2
1621             fac=-expon*(e1+evdwij)
1622             sigder=fac/sigsq
1623             fac=rrij*fac
1624 C Calculate radial part of the gradient
1625             gg(1)=xj*fac
1626             gg(2)=yj*fac
1627             gg(3)=zj*fac
1628 C Calculate the angular part of the gradient and sum add the contributions
1629 C to the appropriate components of the Cartesian gradient.
1630 #ifdef TSCSC
1631             if (bb(itypi,itypj).gt.0) then
1632                call sc_grad
1633             else
1634                call sc_grad_T
1635             endif
1636 #else
1637             call sc_grad
1638 #endif
1639           enddo      ! j
1640         enddo        ! iint
1641       enddo          ! i
1642 c     stop
1643       return
1644       end
1645 C-----------------------------------------------------------------------------
1646       subroutine egb(evdw,evdw_p,evdw_m)
1647 C
1648 C This subroutine calculates the interaction energy of nonbonded side chains
1649 C assuming the Gay-Berne potential of interaction.
1650 C
1651       implicit real*8 (a-h,o-z)
1652       include 'DIMENSIONS'
1653       include 'COMMON.GEO'
1654       include 'COMMON.VAR'
1655       include 'COMMON.LOCAL'
1656       include 'COMMON.CHAIN'
1657       include 'COMMON.DERIV'
1658       include 'COMMON.NAMES'
1659       include 'COMMON.INTERACT'
1660       include 'COMMON.IOUNITS'
1661       include 'COMMON.CALC'
1662       include 'COMMON.CONTROL'
1663       include 'COMMON.SBRIDGE'
1664       logical lprn
1665       evdw=0.0D0
1666 ccccc      energy_dec=.false.
1667 c     print *,'Entering EGB nnt=',nnt,' nct=',nct,' expon=',expon
1668       evdw=0.0D0
1669       evdw_p=0.0D0
1670       evdw_m=0.0D0
1671       lprn=.false.
1672 c     if (icall.eq.0) lprn=.false.
1673       ind=0
1674       do i=iatsc_s,iatsc_e
1675         itypi=itype(i)
1676         itypi1=itype(i+1)
1677         xi=c(1,nres+i)
1678         yi=c(2,nres+i)
1679         zi=c(3,nres+i)
1680         dxi=dc_norm(1,nres+i)
1681         dyi=dc_norm(2,nres+i)
1682         dzi=dc_norm(3,nres+i)
1683 c        dsci_inv=dsc_inv(itypi)
1684         dsci_inv=vbld_inv(i+nres)
1685 c        write (iout,*) "i",i,dsc_inv(itypi),dsci_inv,1.0d0/vbld(i+nres)
1686 c        write (iout,*) "dcnori",dxi*dxi+dyi*dyi+dzi*dzi
1687 C
1688 C Calculate SC interaction energy.
1689 C
1690         do iint=1,nint_gr(i)
1691           do j=istart(i,iint),iend(i,iint)
1692             IF (dyn_ss_mask(i).and.dyn_ss_mask(j)) THEN
1693               call dyn_ssbond_ene(i,j,evdwij)
1694               evdw=evdw+evdwij
1695               if (energy_dec) write (iout,'(a6,2i5,0pf7.3,a3)') 
1696      &                        'evdw',i,j,evdwij,' ss'
1697             ELSE
1698             ind=ind+1
1699             itypj=itype(j)
1700 c            dscj_inv=dsc_inv(itypj)
1701             dscj_inv=vbld_inv(j+nres)
1702 c            write (iout,*) "j",j,dsc_inv(itypj),dscj_inv,
1703 c     &       1.0d0/vbld(j+nres)
1704 c            write (iout,*) "i",i," j", j," itype",itype(i),itype(j)
1705             sig0ij=sigma(itypi,itypj)
1706             chi1=chi(itypi,itypj)
1707             chi2=chi(itypj,itypi)
1708             chi12=chi1*chi2
1709             chip1=chip(itypi)
1710             chip2=chip(itypj)
1711             chip12=chip1*chip2
1712             alf1=alp(itypi)
1713             alf2=alp(itypj)
1714             alf12=0.5D0*(alf1+alf2)
1715 C For diagnostics only!!!
1716 c           chi1=0.0D0
1717 c           chi2=0.0D0
1718 c           chi12=0.0D0
1719 c           chip1=0.0D0
1720 c           chip2=0.0D0
1721 c           chip12=0.0D0
1722 c           alf1=0.0D0
1723 c           alf2=0.0D0
1724 c           alf12=0.0D0
1725             xj=c(1,nres+j)-xi
1726             yj=c(2,nres+j)-yi
1727             zj=c(3,nres+j)-zi
1728             dxj=dc_norm(1,nres+j)
1729             dyj=dc_norm(2,nres+j)
1730             dzj=dc_norm(3,nres+j)
1731 c            write (iout,*) "dcnorj",dxi*dxi+dyi*dyi+dzi*dzi
1732 c            write (iout,*) "j",j," dc_norm",
1733 c     &       dc_norm(1,nres+j),dc_norm(2,nres+j),dc_norm(3,nres+j)
1734             rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
1735             rij=dsqrt(rrij)
1736 C Calculate angle-dependent terms of energy and contributions to their
1737 C derivatives.
1738             call sc_angular
1739             sigsq=1.0D0/sigsq
1740             sig=sig0ij*dsqrt(sigsq)
1741             rij_shift=1.0D0/rij-sig+sig0ij
1742 c for diagnostics; uncomment
1743 c            rij_shift=1.2*sig0ij
1744 C I hate to put IF's in the loops, but here don't have another choice!!!!
1745             if (rij_shift.le.0.0D0) then
1746               evdw=1.0D20
1747 cd              write (iout,'(2(a3,i3,2x),17(0pf7.3))')
1748 cd     &        restyp(itypi),i,restyp(itypj),j,
1749 cd     &        rij_shift,1.0D0/rij,sig,sig0ij,sigsq,1-dsqrt(sigsq) 
1750               return
1751             endif
1752             sigder=-sig*sigsq
1753 c---------------------------------------------------------------
1754             rij_shift=1.0D0/rij_shift 
1755             fac=rij_shift**expon
1756             e1=fac*fac*aa(itypi,itypj)
1757             e2=fac*bb(itypi,itypj)
1758             evdwij=eps1*eps2rt*eps3rt*(e1+e2)
1759             eps2der=evdwij*eps3rt
1760             eps3der=evdwij*eps2rt
1761 c            write (iout,*) "sigsq",sigsq," sig",sig," eps2rt",eps2rt,
1762 c     &        " eps3rt",eps3rt," eps1",eps1," e1",e1," e2",e2
1763             evdwij=evdwij*eps2rt*eps3rt
1764 #ifdef TSCSC
1765             if (bb(itypi,itypj).gt.0) then
1766                evdw_p=evdw_p+evdwij
1767             else
1768                evdw_m=evdw_m+evdwij
1769             endif
1770 #else
1771             evdw=evdw+evdwij
1772 #endif
1773             if (lprn) then
1774             sigm=dabs(aa(itypi,itypj)/bb(itypi,itypj))**(1.0D0/6.0D0)
1775             epsi=bb(itypi,itypj)**2/aa(itypi,itypj)
1776             write (iout,'(2(a3,i3,2x),17(0pf7.3))')
1777      &        restyp(itypi),i,restyp(itypj),j,
1778      &        epsi,sigm,chi1,chi2,chip1,chip2,
1779      &        eps1,eps2rt**2,eps3rt**2,sig,sig0ij,
1780      &        om1,om2,om12,1.0D0/rij,1.0D0/rij_shift,
1781      &        evdwij
1782             endif
1783
1784             if (energy_dec) then
1785               write (iout,'(a6,2i5,0pf7.3)') 'evdw',i,j,evdwij
1786               call flush(iout)
1787             endif
1788 C Calculate gradient components.
1789             e1=e1*eps1*eps2rt**2*eps3rt**2
1790             fac=-expon*(e1+evdwij)*rij_shift
1791             sigder=fac*sigder
1792             fac=rij*fac
1793 c            fac=0.0d0
1794 C Calculate the radial part of the gradient
1795             gg(1)=xj*fac
1796             gg(2)=yj*fac
1797             gg(3)=zj*fac
1798 C Calculate angular part of the gradient.
1799 #ifdef TSCSC
1800             if (bb(itypi,itypj).gt.0) then
1801                call sc_grad
1802             else
1803                call sc_grad_T
1804             endif
1805 #else
1806             call sc_grad
1807 #endif
1808             ENDIF    ! dyn_ss            
1809           enddo      ! j
1810         enddo        ! iint
1811       enddo          ! i
1812 c      write (iout,*) "Number of loop steps in EGB:",ind
1813 cccc      energy_dec=.false.
1814       return
1815       end
1816 C-----------------------------------------------------------------------------
1817       subroutine egbv(evdw,evdw_p,evdw_m)
1818 C
1819 C This subroutine calculates the interaction energy of nonbonded side chains
1820 C assuming the Gay-Berne-Vorobjev potential of interaction.
1821 C
1822       implicit real*8 (a-h,o-z)
1823       include 'DIMENSIONS'
1824       include 'COMMON.GEO'
1825       include 'COMMON.VAR'
1826       include 'COMMON.LOCAL'
1827       include 'COMMON.CHAIN'
1828       include 'COMMON.DERIV'
1829       include 'COMMON.NAMES'
1830       include 'COMMON.INTERACT'
1831       include 'COMMON.IOUNITS'
1832       include 'COMMON.CALC'
1833       common /srutu/ icall
1834       logical lprn
1835       evdw=0.0D0
1836 c     print *,'Entering EGB nnt=',nnt,' nct=',nct,' expon=',expon
1837       evdw=0.0D0
1838       lprn=.false.
1839 c     if (icall.eq.0) lprn=.true.
1840       ind=0
1841       do i=iatsc_s,iatsc_e
1842         itypi=itype(i)
1843         itypi1=itype(i+1)
1844         xi=c(1,nres+i)
1845         yi=c(2,nres+i)
1846         zi=c(3,nres+i)
1847         dxi=dc_norm(1,nres+i)
1848         dyi=dc_norm(2,nres+i)
1849         dzi=dc_norm(3,nres+i)
1850 c        dsci_inv=dsc_inv(itypi)
1851         dsci_inv=vbld_inv(i+nres)
1852 C
1853 C Calculate SC interaction energy.
1854 C
1855         do iint=1,nint_gr(i)
1856           do j=istart(i,iint),iend(i,iint)
1857             ind=ind+1
1858             itypj=itype(j)
1859 c            dscj_inv=dsc_inv(itypj)
1860             dscj_inv=vbld_inv(j+nres)
1861             sig0ij=sigma(itypi,itypj)
1862             r0ij=r0(itypi,itypj)
1863             chi1=chi(itypi,itypj)
1864             chi2=chi(itypj,itypi)
1865             chi12=chi1*chi2
1866             chip1=chip(itypi)
1867             chip2=chip(itypj)
1868             chip12=chip1*chip2
1869             alf1=alp(itypi)
1870             alf2=alp(itypj)
1871             alf12=0.5D0*(alf1+alf2)
1872 C For diagnostics only!!!
1873 c           chi1=0.0D0
1874 c           chi2=0.0D0
1875 c           chi12=0.0D0
1876 c           chip1=0.0D0
1877 c           chip2=0.0D0
1878 c           chip12=0.0D0
1879 c           alf1=0.0D0
1880 c           alf2=0.0D0
1881 c           alf12=0.0D0
1882             xj=c(1,nres+j)-xi
1883             yj=c(2,nres+j)-yi
1884             zj=c(3,nres+j)-zi
1885             dxj=dc_norm(1,nres+j)
1886             dyj=dc_norm(2,nres+j)
1887             dzj=dc_norm(3,nres+j)
1888             rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
1889             rij=dsqrt(rrij)
1890 C Calculate angle-dependent terms of energy and contributions to their
1891 C derivatives.
1892             call sc_angular
1893             sigsq=1.0D0/sigsq
1894             sig=sig0ij*dsqrt(sigsq)
1895             rij_shift=1.0D0/rij-sig+r0ij
1896 C I hate to put IF's in the loops, but here don't have another choice!!!!
1897             if (rij_shift.le.0.0D0) then
1898               evdw=1.0D20
1899               return
1900             endif
1901             sigder=-sig*sigsq
1902 c---------------------------------------------------------------
1903             rij_shift=1.0D0/rij_shift 
1904             fac=rij_shift**expon
1905             e1=fac*fac*aa(itypi,itypj)
1906             e2=fac*bb(itypi,itypj)
1907             evdwij=eps1*eps2rt*eps3rt*(e1+e2)
1908             eps2der=evdwij*eps3rt
1909             eps3der=evdwij*eps2rt
1910             fac_augm=rrij**expon
1911             e_augm=augm(itypi,itypj)*fac_augm
1912             evdwij=evdwij*eps2rt*eps3rt
1913 #ifdef TSCSC
1914             if (bb(itypi,itypj).gt.0) then
1915                evdw_p=evdw_p+evdwij+e_augm
1916             else
1917                evdw_m=evdw_m+evdwij+e_augm
1918             endif
1919 #else
1920             evdw=evdw+evdwij+e_augm
1921 #endif
1922             if (lprn) then
1923             sigm=dabs(aa(itypi,itypj)/bb(itypi,itypj))**(1.0D0/6.0D0)
1924             epsi=bb(itypi,itypj)**2/aa(itypi,itypj)
1925             write (iout,'(2(a3,i3,2x),17(0pf7.3))')
1926      &        restyp(itypi),i,restyp(itypj),j,
1927      &        epsi,sigm,sig,(augm(itypi,itypj)/epsi)**(1.0D0/12.0D0),
1928      &        chi1,chi2,chip1,chip2,
1929      &        eps1,eps2rt**2,eps3rt**2,
1930      &        om1,om2,om12,1.0D0/rij,1.0D0/rij_shift,
1931      &        evdwij+e_augm
1932             endif
1933 C Calculate gradient components.
1934             e1=e1*eps1*eps2rt**2*eps3rt**2
1935             fac=-expon*(e1+evdwij)*rij_shift
1936             sigder=fac*sigder
1937             fac=rij*fac-2*expon*rrij*e_augm
1938 C Calculate the radial part of the gradient
1939             gg(1)=xj*fac
1940             gg(2)=yj*fac
1941             gg(3)=zj*fac
1942 C Calculate angular part of the gradient.
1943 #ifdef TSCSC
1944             if (bb(itypi,itypj).gt.0) then
1945                call sc_grad
1946             else
1947                call sc_grad_T
1948             endif
1949 #else
1950             call sc_grad
1951 #endif
1952           enddo      ! j
1953         enddo        ! iint
1954       enddo          ! i
1955       end
1956 C-----------------------------------------------------------------------------
1957       subroutine sc_angular
1958 C Calculate eps1,eps2,eps3,sigma, and parts of their derivatives in om1,om2,
1959 C om12. Called by ebp, egb, and egbv.
1960       implicit none
1961       include 'COMMON.CALC'
1962       include 'COMMON.IOUNITS'
1963       erij(1)=xj*rij
1964       erij(2)=yj*rij
1965       erij(3)=zj*rij
1966       om1=dxi*erij(1)+dyi*erij(2)+dzi*erij(3)
1967       om2=dxj*erij(1)+dyj*erij(2)+dzj*erij(3)
1968       om12=dxi*dxj+dyi*dyj+dzi*dzj
1969       chiom12=chi12*om12
1970 C Calculate eps1(om12) and its derivative in om12
1971       faceps1=1.0D0-om12*chiom12
1972       faceps1_inv=1.0D0/faceps1
1973       eps1=dsqrt(faceps1_inv)
1974 C Following variable is eps1*deps1/dom12
1975       eps1_om12=faceps1_inv*chiom12
1976 c diagnostics only
1977 c      faceps1_inv=om12
1978 c      eps1=om12
1979 c      eps1_om12=1.0d0
1980 c      write (iout,*) "om12",om12," eps1",eps1
1981 C Calculate sigma(om1,om2,om12) and the derivatives of sigma**2 in om1,om2,
1982 C and om12.
1983       om1om2=om1*om2
1984       chiom1=chi1*om1
1985       chiom2=chi2*om2
1986       facsig=om1*chiom1+om2*chiom2-2.0D0*om1om2*chiom12
1987       sigsq=1.0D0-facsig*faceps1_inv
1988       sigsq_om1=(chiom1-chiom12*om2)*faceps1_inv
1989       sigsq_om2=(chiom2-chiom12*om1)*faceps1_inv
1990       sigsq_om12=-chi12*(om1om2*faceps1-om12*facsig)*faceps1_inv**2
1991 c diagnostics only
1992 c      sigsq=1.0d0
1993 c      sigsq_om1=0.0d0
1994 c      sigsq_om2=0.0d0
1995 c      sigsq_om12=0.0d0
1996 c      write (iout,*) "chiom1",chiom1," chiom2",chiom2," chiom12",chiom12
1997 c      write (iout,*) "faceps1",faceps1," faceps1_inv",faceps1_inv,
1998 c     &    " eps1",eps1
1999 C Calculate eps2 and its derivatives in om1, om2, and om12.
2000       chipom1=chip1*om1
2001       chipom2=chip2*om2
2002       chipom12=chip12*om12
2003       facp=1.0D0-om12*chipom12
2004       facp_inv=1.0D0/facp
2005       facp1=om1*chipom1+om2*chipom2-2.0D0*om1om2*chipom12
2006 c      write (iout,*) "chipom1",chipom1," chipom2",chipom2,
2007 c     &  " chipom12",chipom12," facp",facp," facp_inv",facp_inv
2008 C Following variable is the square root of eps2
2009       eps2rt=1.0D0-facp1*facp_inv
2010 C Following three variables are the derivatives of the square root of eps
2011 C in om1, om2, and om12.
2012       eps2rt_om1=-4.0D0*(chipom1-chipom12*om2)*facp_inv
2013       eps2rt_om2=-4.0D0*(chipom2-chipom12*om1)*facp_inv
2014       eps2rt_om12=4.0D0*chip12*(om1om2*facp-om12*facp1)*facp_inv**2 
2015 C Evaluate the "asymmetric" factor in the VDW constant, eps3
2016       eps3rt=1.0D0-alf1*om1+alf2*om2-alf12*om12 
2017 c      write (iout,*) "eps2rt",eps2rt," eps3rt",eps3rt
2018 c      write (iout,*) "eps2rt_om1",eps2rt_om1," eps2rt_om2",eps2rt_om2,
2019 c     &  " eps2rt_om12",eps2rt_om12
2020 C Calculate whole angle-dependent part of epsilon and contributions
2021 C to its derivatives
2022       return
2023       end
2024
2025 C----------------------------------------------------------------------------
2026       subroutine sc_grad_T
2027       implicit real*8 (a-h,o-z)
2028       include 'DIMENSIONS'
2029       include 'COMMON.CHAIN'
2030       include 'COMMON.DERIV'
2031       include 'COMMON.CALC'
2032       include 'COMMON.IOUNITS'
2033       double precision dcosom1(3),dcosom2(3)
2034       eom1=eps2der*eps2rt_om1-2.0D0*alf1*eps3der+sigder*sigsq_om1
2035       eom2=eps2der*eps2rt_om2+2.0D0*alf2*eps3der+sigder*sigsq_om2
2036       eom12=evdwij*eps1_om12+eps2der*eps2rt_om12
2037      &     -2.0D0*alf12*eps3der+sigder*sigsq_om12
2038 c diagnostics only
2039 c      eom1=0.0d0
2040 c      eom2=0.0d0
2041 c      eom12=evdwij*eps1_om12
2042 c end diagnostics
2043 c      write (iout,*) "eps2der",eps2der," eps3der",eps3der,
2044 c     &  " sigder",sigder
2045 c      write (iout,*) "eps1_om12",eps1_om12," eps2rt_om12",eps2rt_om12
2046 c      write (iout,*) "eom1",eom1," eom2",eom2," eom12",eom12
2047       do k=1,3
2048         dcosom1(k)=rij*(dc_norm(k,nres+i)-om1*erij(k))
2049         dcosom2(k)=rij*(dc_norm(k,nres+j)-om2*erij(k))
2050       enddo
2051       do k=1,3
2052         gg(k)=gg(k)+eom1*dcosom1(k)+eom2*dcosom2(k)
2053       enddo 
2054 c      write (iout,*) "gg",(gg(k),k=1,3)
2055       do k=1,3
2056         gvdwxT(k,i)=gvdwxT(k,i)-gg(k)
2057      &            +(eom12*(dc_norm(k,nres+j)-om12*dc_norm(k,nres+i))
2058      &            +eom1*(erij(k)-om1*dc_norm(k,nres+i)))*dsci_inv
2059         gvdwxT(k,j)=gvdwxT(k,j)+gg(k)
2060      &            +(eom12*(dc_norm(k,nres+i)-om12*dc_norm(k,nres+j))
2061      &            +eom2*(erij(k)-om2*dc_norm(k,nres+j)))*dscj_inv
2062 c        write (iout,*)(eom12*(dc_norm(k,nres+j)-om12*dc_norm(k,nres+i))
2063 c     &            +eom1*(erij(k)-om1*dc_norm(k,nres+i)))*dsci_inv
2064 c        write (iout,*)(eom12*(dc_norm(k,nres+i)-om12*dc_norm(k,nres+j))
2065 c     &            +eom2*(erij(k)-om2*dc_norm(k,nres+j)))*dscj_inv
2066       enddo
2067
2068 C Calculate the components of the gradient in DC and X
2069 C
2070 cgrad      do k=i,j-1
2071 cgrad        do l=1,3
2072 cgrad          gvdwc(l,k)=gvdwc(l,k)+gg(l)
2073 cgrad        enddo
2074 cgrad      enddo
2075       do l=1,3
2076         gvdwcT(l,i)=gvdwcT(l,i)-gg(l)
2077         gvdwcT(l,j)=gvdwcT(l,j)+gg(l)
2078       enddo
2079       return
2080       end
2081
2082 C----------------------------------------------------------------------------
2083       subroutine sc_grad
2084       implicit real*8 (a-h,o-z)
2085       include 'DIMENSIONS'
2086       include 'COMMON.CHAIN'
2087       include 'COMMON.DERIV'
2088       include 'COMMON.CALC'
2089       include 'COMMON.IOUNITS'
2090       double precision dcosom1(3),dcosom2(3)
2091       eom1=eps2der*eps2rt_om1-2.0D0*alf1*eps3der+sigder*sigsq_om1
2092       eom2=eps2der*eps2rt_om2+2.0D0*alf2*eps3der+sigder*sigsq_om2
2093       eom12=evdwij*eps1_om12+eps2der*eps2rt_om12
2094      &     -2.0D0*alf12*eps3der+sigder*sigsq_om12
2095 c diagnostics only
2096 c      eom1=0.0d0
2097 c      eom2=0.0d0
2098 c      eom12=evdwij*eps1_om12
2099 c end diagnostics
2100 c      write (iout,*) "eps2der",eps2der," eps3der",eps3der,
2101 c     &  " sigder",sigder
2102 c      write (iout,*) "eps1_om12",eps1_om12," eps2rt_om12",eps2rt_om12
2103 c      write (iout,*) "eom1",eom1," eom2",eom2," eom12",eom12
2104       do k=1,3
2105         dcosom1(k)=rij*(dc_norm(k,nres+i)-om1*erij(k))
2106         dcosom2(k)=rij*(dc_norm(k,nres+j)-om2*erij(k))
2107       enddo
2108       do k=1,3
2109         gg(k)=gg(k)+eom1*dcosom1(k)+eom2*dcosom2(k)
2110       enddo 
2111 c      write (iout,*) "gg",(gg(k),k=1,3)
2112       do k=1,3
2113         gvdwx(k,i)=gvdwx(k,i)-gg(k)
2114      &            +(eom12*(dc_norm(k,nres+j)-om12*dc_norm(k,nres+i))
2115      &            +eom1*(erij(k)-om1*dc_norm(k,nres+i)))*dsci_inv
2116         gvdwx(k,j)=gvdwx(k,j)+gg(k)
2117      &            +(eom12*(dc_norm(k,nres+i)-om12*dc_norm(k,nres+j))
2118      &            +eom2*(erij(k)-om2*dc_norm(k,nres+j)))*dscj_inv
2119 c        write (iout,*)(eom12*(dc_norm(k,nres+j)-om12*dc_norm(k,nres+i))
2120 c     &            +eom1*(erij(k)-om1*dc_norm(k,nres+i)))*dsci_inv
2121 c        write (iout,*)(eom12*(dc_norm(k,nres+i)-om12*dc_norm(k,nres+j))
2122 c     &            +eom2*(erij(k)-om2*dc_norm(k,nres+j)))*dscj_inv
2123       enddo
2124
2125 C Calculate the components of the gradient in DC and X
2126 C
2127 cgrad      do k=i,j-1
2128 cgrad        do l=1,3
2129 cgrad          gvdwc(l,k)=gvdwc(l,k)+gg(l)
2130 cgrad        enddo
2131 cgrad      enddo
2132       do l=1,3
2133         gvdwc(l,i)=gvdwc(l,i)-gg(l)
2134         gvdwc(l,j)=gvdwc(l,j)+gg(l)
2135       enddo
2136       return
2137       end
2138 C-----------------------------------------------------------------------
2139       subroutine e_softsphere(evdw)
2140 C
2141 C This subroutine calculates the interaction energy of nonbonded side chains
2142 C assuming the LJ potential of interaction.
2143 C
2144       implicit real*8 (a-h,o-z)
2145       include 'DIMENSIONS'
2146       parameter (accur=1.0d-10)
2147       include 'COMMON.GEO'
2148       include 'COMMON.VAR'
2149       include 'COMMON.LOCAL'
2150       include 'COMMON.CHAIN'
2151       include 'COMMON.DERIV'
2152       include 'COMMON.INTERACT'
2153       include 'COMMON.TORSION'
2154       include 'COMMON.SBRIDGE'
2155       include 'COMMON.NAMES'
2156       include 'COMMON.IOUNITS'
2157       include 'COMMON.CONTACTS'
2158       dimension gg(3)
2159 cd    print *,'Entering Esoft_sphere nnt=',nnt,' nct=',nct
2160       evdw=0.0D0
2161       do i=iatsc_s,iatsc_e
2162         itypi=itype(i)
2163         itypi1=itype(i+1)
2164         xi=c(1,nres+i)
2165         yi=c(2,nres+i)
2166         zi=c(3,nres+i)
2167 C
2168 C Calculate SC interaction energy.
2169 C
2170         do iint=1,nint_gr(i)
2171 cd        write (iout,*) 'i=',i,' iint=',iint,' istart=',istart(i,iint),
2172 cd   &                  'iend=',iend(i,iint)
2173           do j=istart(i,iint),iend(i,iint)
2174             itypj=itype(j)
2175             xj=c(1,nres+j)-xi
2176             yj=c(2,nres+j)-yi
2177             zj=c(3,nres+j)-zi
2178             rij=xj*xj+yj*yj+zj*zj
2179 c           write (iout,*)'i=',i,' j=',j,' itypi=',itypi,' itypj=',itypj
2180             r0ij=r0(itypi,itypj)
2181             r0ijsq=r0ij*r0ij
2182 c            print *,i,j,r0ij,dsqrt(rij)
2183             if (rij.lt.r0ijsq) then
2184               evdwij=0.25d0*(rij-r0ijsq)**2
2185               fac=rij-r0ijsq
2186             else
2187               evdwij=0.0d0
2188               fac=0.0d0
2189             endif
2190             evdw=evdw+evdwij
2191
2192 C Calculate the components of the gradient in DC and X
2193 C
2194             gg(1)=xj*fac
2195             gg(2)=yj*fac
2196             gg(3)=zj*fac
2197             do k=1,3
2198               gvdwx(k,i)=gvdwx(k,i)-gg(k)
2199               gvdwx(k,j)=gvdwx(k,j)+gg(k)
2200               gvdwc(k,i)=gvdwc(k,i)-gg(k)
2201               gvdwc(k,j)=gvdwc(k,j)+gg(k)
2202             enddo
2203 cgrad            do k=i,j-1
2204 cgrad              do l=1,3
2205 cgrad                gvdwc(l,k)=gvdwc(l,k)+gg(l)
2206 cgrad              enddo
2207 cgrad            enddo
2208           enddo ! j
2209         enddo ! iint
2210       enddo ! i
2211       return
2212       end
2213 C--------------------------------------------------------------------------
2214       subroutine eelec_soft_sphere(ees,evdw1,eel_loc,eello_turn3,
2215      &              eello_turn4)
2216 C
2217 C Soft-sphere potential of p-p interaction
2218
2219       implicit real*8 (a-h,o-z)
2220       include 'DIMENSIONS'
2221       include 'COMMON.CONTROL'
2222       include 'COMMON.IOUNITS'
2223       include 'COMMON.GEO'
2224       include 'COMMON.VAR'
2225       include 'COMMON.LOCAL'
2226       include 'COMMON.CHAIN'
2227       include 'COMMON.DERIV'
2228       include 'COMMON.INTERACT'
2229       include 'COMMON.CONTACTS'
2230       include 'COMMON.TORSION'
2231       include 'COMMON.VECTORS'
2232       include 'COMMON.FFIELD'
2233       dimension ggg(3)
2234 cd      write(iout,*) 'In EELEC_soft_sphere'
2235       ees=0.0D0
2236       evdw1=0.0D0
2237       eel_loc=0.0d0 
2238       eello_turn3=0.0d0
2239       eello_turn4=0.0d0
2240       ind=0
2241       do i=iatel_s,iatel_e
2242         dxi=dc(1,i)
2243         dyi=dc(2,i)
2244         dzi=dc(3,i)
2245         xmedi=c(1,i)+0.5d0*dxi
2246         ymedi=c(2,i)+0.5d0*dyi
2247         zmedi=c(3,i)+0.5d0*dzi
2248         num_conti=0
2249 c        write (iout,*) 'i',i,' ielstart',ielstart(i),' ielend',ielend(i)
2250         do j=ielstart(i),ielend(i)
2251           ind=ind+1
2252           iteli=itel(i)
2253           itelj=itel(j)
2254           if (j.eq.i+2 .and. itelj.eq.2) iteli=2
2255           r0ij=rpp(iteli,itelj)
2256           r0ijsq=r0ij*r0ij 
2257           dxj=dc(1,j)
2258           dyj=dc(2,j)
2259           dzj=dc(3,j)
2260           xj=c(1,j)+0.5D0*dxj-xmedi
2261           yj=c(2,j)+0.5D0*dyj-ymedi
2262           zj=c(3,j)+0.5D0*dzj-zmedi
2263           rij=xj*xj+yj*yj+zj*zj
2264           if (rij.lt.r0ijsq) then
2265             evdw1ij=0.25d0*(rij-r0ijsq)**2
2266             fac=rij-r0ijsq
2267           else
2268             evdw1ij=0.0d0
2269             fac=0.0d0
2270           endif
2271           evdw1=evdw1+evdw1ij
2272 C
2273 C Calculate contributions to the Cartesian gradient.
2274 C
2275           ggg(1)=fac*xj
2276           ggg(2)=fac*yj
2277           ggg(3)=fac*zj
2278           do k=1,3
2279             gvdwpp(k,i)=gvdwpp(k,i)-ggg(k)
2280             gvdwpp(k,j)=gvdwpp(k,j)+ggg(k)
2281           enddo
2282 *
2283 * Loop over residues i+1 thru j-1.
2284 *
2285 cgrad          do k=i+1,j-1
2286 cgrad            do l=1,3
2287 cgrad              gelc(l,k)=gelc(l,k)+ggg(l)
2288 cgrad            enddo
2289 cgrad          enddo
2290         enddo ! j
2291       enddo   ! i
2292 cgrad      do i=nnt,nct-1
2293 cgrad        do k=1,3
2294 cgrad          gelc(k,i)=gelc(k,i)+0.5d0*gelc(k,i)
2295 cgrad        enddo
2296 cgrad        do j=i+1,nct-1
2297 cgrad          do k=1,3
2298 cgrad            gelc(k,i)=gelc(k,i)+gelc(k,j)
2299 cgrad          enddo
2300 cgrad        enddo
2301 cgrad      enddo
2302       return
2303       end
2304 c------------------------------------------------------------------------------
2305       subroutine vec_and_deriv
2306       implicit real*8 (a-h,o-z)
2307       include 'DIMENSIONS'
2308 #ifdef MPI
2309       include 'mpif.h'
2310 #endif
2311       include 'COMMON.IOUNITS'
2312       include 'COMMON.GEO'
2313       include 'COMMON.VAR'
2314       include 'COMMON.LOCAL'
2315       include 'COMMON.CHAIN'
2316       include 'COMMON.VECTORS'
2317       include 'COMMON.SETUP'
2318       include 'COMMON.TIME1'
2319       dimension uyder(3,3,2),uzder(3,3,2),vbld_inv_temp(2)
2320 C Compute the local reference systems. For reference system (i), the
2321 C X-axis points from CA(i) to CA(i+1), the Y axis is in the 
2322 C CA(i)-CA(i+1)-CA(i+2) plane, and the Z axis is perpendicular to this plane.
2323 #ifdef PARVEC
2324       do i=ivec_start,ivec_end
2325 #else
2326       do i=1,nres-1
2327 #endif
2328           if (i.eq.nres-1) then
2329 C Case of the last full residue
2330 C Compute the Z-axis
2331             call vecpr(dc_norm(1,i),dc_norm(1,i-1),uz(1,i))
2332             costh=dcos(pi-theta(nres))
2333             fac=1.0d0/dsqrt(1.0d0-costh*costh)
2334             do k=1,3
2335               uz(k,i)=fac*uz(k,i)
2336             enddo
2337 C Compute the derivatives of uz
2338             uzder(1,1,1)= 0.0d0
2339             uzder(2,1,1)=-dc_norm(3,i-1)
2340             uzder(3,1,1)= dc_norm(2,i-1) 
2341             uzder(1,2,1)= dc_norm(3,i-1)
2342             uzder(2,2,1)= 0.0d0
2343             uzder(3,2,1)=-dc_norm(1,i-1)
2344             uzder(1,3,1)=-dc_norm(2,i-1)
2345             uzder(2,3,1)= dc_norm(1,i-1)
2346             uzder(3,3,1)= 0.0d0
2347             uzder(1,1,2)= 0.0d0
2348             uzder(2,1,2)= dc_norm(3,i)
2349             uzder(3,1,2)=-dc_norm(2,i) 
2350             uzder(1,2,2)=-dc_norm(3,i)
2351             uzder(2,2,2)= 0.0d0
2352             uzder(3,2,2)= dc_norm(1,i)
2353             uzder(1,3,2)= dc_norm(2,i)
2354             uzder(2,3,2)=-dc_norm(1,i)
2355             uzder(3,3,2)= 0.0d0
2356 C Compute the Y-axis
2357             facy=fac
2358             do k=1,3
2359               uy(k,i)=fac*(dc_norm(k,i-1)-costh*dc_norm(k,i))
2360             enddo
2361 C Compute the derivatives of uy
2362             do j=1,3
2363               do k=1,3
2364                 uyder(k,j,1)=2*dc_norm(k,i-1)*dc_norm(j,i)
2365      &                        -dc_norm(k,i)*dc_norm(j,i-1)
2366                 uyder(k,j,2)=-dc_norm(j,i)*dc_norm(k,i)
2367               enddo
2368               uyder(j,j,1)=uyder(j,j,1)-costh
2369               uyder(j,j,2)=1.0d0+uyder(j,j,2)
2370             enddo
2371             do j=1,2
2372               do k=1,3
2373                 do l=1,3
2374                   uygrad(l,k,j,i)=uyder(l,k,j)
2375                   uzgrad(l,k,j,i)=uzder(l,k,j)
2376                 enddo
2377               enddo
2378             enddo 
2379             call unormderiv(uy(1,i),uyder(1,1,1),facy,uygrad(1,1,1,i))
2380             call unormderiv(uy(1,i),uyder(1,1,2),facy,uygrad(1,1,2,i))
2381             call unormderiv(uz(1,i),uzder(1,1,1),fac,uzgrad(1,1,1,i))
2382             call unormderiv(uz(1,i),uzder(1,1,2),fac,uzgrad(1,1,2,i))
2383           else
2384 C Other residues
2385 C Compute the Z-axis
2386             call vecpr(dc_norm(1,i),dc_norm(1,i+1),uz(1,i))
2387             costh=dcos(pi-theta(i+2))
2388             fac=1.0d0/dsqrt(1.0d0-costh*costh)
2389             do k=1,3
2390               uz(k,i)=fac*uz(k,i)
2391             enddo
2392 C Compute the derivatives of uz
2393             uzder(1,1,1)= 0.0d0
2394             uzder(2,1,1)=-dc_norm(3,i+1)
2395             uzder(3,1,1)= dc_norm(2,i+1) 
2396             uzder(1,2,1)= dc_norm(3,i+1)
2397             uzder(2,2,1)= 0.0d0
2398             uzder(3,2,1)=-dc_norm(1,i+1)
2399             uzder(1,3,1)=-dc_norm(2,i+1)
2400             uzder(2,3,1)= dc_norm(1,i+1)
2401             uzder(3,3,1)= 0.0d0
2402             uzder(1,1,2)= 0.0d0
2403             uzder(2,1,2)= dc_norm(3,i)
2404             uzder(3,1,2)=-dc_norm(2,i) 
2405             uzder(1,2,2)=-dc_norm(3,i)
2406             uzder(2,2,2)= 0.0d0
2407             uzder(3,2,2)= dc_norm(1,i)
2408             uzder(1,3,2)= dc_norm(2,i)
2409             uzder(2,3,2)=-dc_norm(1,i)
2410             uzder(3,3,2)= 0.0d0
2411 C Compute the Y-axis
2412             facy=fac
2413             do k=1,3
2414               uy(k,i)=facy*(dc_norm(k,i+1)-costh*dc_norm(k,i))
2415             enddo
2416 C Compute the derivatives of uy
2417             do j=1,3
2418               do k=1,3
2419                 uyder(k,j,1)=2*dc_norm(k,i+1)*dc_norm(j,i)
2420      &                        -dc_norm(k,i)*dc_norm(j,i+1)
2421                 uyder(k,j,2)=-dc_norm(j,i)*dc_norm(k,i)
2422               enddo
2423               uyder(j,j,1)=uyder(j,j,1)-costh
2424               uyder(j,j,2)=1.0d0+uyder(j,j,2)
2425             enddo
2426             do j=1,2
2427               do k=1,3
2428                 do l=1,3
2429                   uygrad(l,k,j,i)=uyder(l,k,j)
2430                   uzgrad(l,k,j,i)=uzder(l,k,j)
2431                 enddo
2432               enddo
2433             enddo 
2434             call unormderiv(uy(1,i),uyder(1,1,1),facy,uygrad(1,1,1,i))
2435             call unormderiv(uy(1,i),uyder(1,1,2),facy,uygrad(1,1,2,i))
2436             call unormderiv(uz(1,i),uzder(1,1,1),fac,uzgrad(1,1,1,i))
2437             call unormderiv(uz(1,i),uzder(1,1,2),fac,uzgrad(1,1,2,i))
2438           endif
2439       enddo
2440       do i=1,nres-1
2441         vbld_inv_temp(1)=vbld_inv(i+1)
2442         if (i.lt.nres-1) then
2443           vbld_inv_temp(2)=vbld_inv(i+2)
2444           else
2445           vbld_inv_temp(2)=vbld_inv(i)
2446           endif
2447         do j=1,2
2448           do k=1,3
2449             do l=1,3
2450               uygrad(l,k,j,i)=vbld_inv_temp(j)*uygrad(l,k,j,i)
2451               uzgrad(l,k,j,i)=vbld_inv_temp(j)*uzgrad(l,k,j,i)
2452             enddo
2453           enddo
2454         enddo
2455       enddo
2456 #if defined(PARVEC) && defined(MPI)
2457       if (nfgtasks1.gt.1) then
2458         time00=MPI_Wtime()
2459 c        print *,"Processor",fg_rank1,kolor1," ivec_start",ivec_start,
2460 c     &   " ivec_displ",(ivec_displ(i),i=0,nfgtasks1-1),
2461 c     &   " ivec_count",(ivec_count(i),i=0,nfgtasks1-1)
2462         call MPI_Allgatherv(uy(1,ivec_start),ivec_count(fg_rank1),
2463      &   MPI_UYZ,uy(1,1),ivec_count(0),ivec_displ(0),MPI_UYZ,
2464      &   FG_COMM1,IERR)
2465         call MPI_Allgatherv(uz(1,ivec_start),ivec_count(fg_rank1),
2466      &   MPI_UYZ,uz(1,1),ivec_count(0),ivec_displ(0),MPI_UYZ,
2467      &   FG_COMM1,IERR)
2468         call MPI_Allgatherv(uygrad(1,1,1,ivec_start),
2469      &   ivec_count(fg_rank1),MPI_UYZGRAD,uygrad(1,1,1,1),ivec_count(0),
2470      &   ivec_displ(0),MPI_UYZGRAD,FG_COMM1,IERR)
2471         call MPI_Allgatherv(uzgrad(1,1,1,ivec_start),
2472      &   ivec_count(fg_rank1),MPI_UYZGRAD,uzgrad(1,1,1,1),ivec_count(0),
2473      &   ivec_displ(0),MPI_UYZGRAD,FG_COMM1,IERR)
2474         time_gather=time_gather+MPI_Wtime()-time00
2475       endif
2476 c      if (fg_rank.eq.0) then
2477 c        write (iout,*) "Arrays UY and UZ"
2478 c        do i=1,nres-1
2479 c          write (iout,'(i5,3f10.5,5x,3f10.5)') i,(uy(k,i),k=1,3),
2480 c     &     (uz(k,i),k=1,3)
2481 c        enddo
2482 c      endif
2483 #endif
2484       return
2485       end
2486 C-----------------------------------------------------------------------------
2487       subroutine check_vecgrad
2488       implicit real*8 (a-h,o-z)
2489       include 'DIMENSIONS'
2490       include 'COMMON.IOUNITS'
2491       include 'COMMON.GEO'
2492       include 'COMMON.VAR'
2493       include 'COMMON.LOCAL'
2494       include 'COMMON.CHAIN'
2495       include 'COMMON.VECTORS'
2496       dimension uygradt(3,3,2,maxres),uzgradt(3,3,2,maxres)
2497       dimension uyt(3,maxres),uzt(3,maxres)
2498       dimension uygradn(3,3,2),uzgradn(3,3,2),erij(3)
2499       double precision delta /1.0d-7/
2500       call vec_and_deriv
2501 cd      do i=1,nres
2502 crc          write(iout,'(2i5,2(3f10.5,5x))') i,1,dc_norm(:,i)
2503 crc          write(iout,'(2i5,2(3f10.5,5x))') i,2,uy(:,i)
2504 crc          write(iout,'(2i5,2(3f10.5,5x)/)')i,3,uz(:,i)
2505 cd          write(iout,'(2i5,2(3f10.5,5x))') i,1,
2506 cd     &     (dc_norm(if90,i),if90=1,3)
2507 cd          write(iout,'(2i5,2(3f10.5,5x))') i,2,(uy(if90,i),if90=1,3)
2508 cd          write(iout,'(2i5,2(3f10.5,5x)/)')i,3,(uz(if90,i),if90=1,3)
2509 cd          write(iout,'(a)')
2510 cd      enddo
2511       do i=1,nres
2512         do j=1,2
2513           do k=1,3
2514             do l=1,3
2515               uygradt(l,k,j,i)=uygrad(l,k,j,i)
2516               uzgradt(l,k,j,i)=uzgrad(l,k,j,i)
2517             enddo
2518           enddo
2519         enddo
2520       enddo
2521       call vec_and_deriv
2522       do i=1,nres
2523         do j=1,3
2524           uyt(j,i)=uy(j,i)
2525           uzt(j,i)=uz(j,i)
2526         enddo
2527       enddo
2528       do i=1,nres
2529 cd        write (iout,*) 'i=',i
2530         do k=1,3
2531           erij(k)=dc_norm(k,i)
2532         enddo
2533         do j=1,3
2534           do k=1,3
2535             dc_norm(k,i)=erij(k)
2536           enddo
2537           dc_norm(j,i)=dc_norm(j,i)+delta
2538 c          fac=dsqrt(scalar(dc_norm(1,i),dc_norm(1,i)))
2539 c          do k=1,3
2540 c            dc_norm(k,i)=dc_norm(k,i)/fac
2541 c          enddo
2542 c          write (iout,*) (dc_norm(k,i),k=1,3)
2543 c          write (iout,*) (erij(k),k=1,3)
2544           call vec_and_deriv
2545           do k=1,3
2546             uygradn(k,j,1)=(uy(k,i)-uyt(k,i))/delta
2547             uygradn(k,j,2)=(uy(k,i-1)-uyt(k,i-1))/delta
2548             uzgradn(k,j,1)=(uz(k,i)-uzt(k,i))/delta
2549             uzgradn(k,j,2)=(uz(k,i-1)-uzt(k,i-1))/delta
2550           enddo 
2551 c          write (iout,'(i5,3f8.5,3x,3f8.5,5x,3f8.5,3x,3f8.5)') 
2552 c     &      j,(uzgradt(k,j,1,i),k=1,3),(uzgradn(k,j,1),k=1,3),
2553 c     &      (uzgradt(k,j,2,i-1),k=1,3),(uzgradn(k,j,2),k=1,3)
2554         enddo
2555         do k=1,3
2556           dc_norm(k,i)=erij(k)
2557         enddo
2558 cd        do k=1,3
2559 cd          write (iout,'(i5,3f8.5,3x,3f8.5,5x,3f8.5,3x,3f8.5)') 
2560 cd     &      k,(uygradt(k,l,1,i),l=1,3),(uygradn(k,l,1),l=1,3),
2561 cd     &      (uygradt(k,l,2,i-1),l=1,3),(uygradn(k,l,2),l=1,3)
2562 cd          write (iout,'(i5,3f8.5,3x,3f8.5,5x,3f8.5,3x,3f8.5)') 
2563 cd     &      k,(uzgradt(k,l,1,i),l=1,3),(uzgradn(k,l,1),l=1,3),
2564 cd     &      (uzgradt(k,l,2,i-1),l=1,3),(uzgradn(k,l,2),l=1,3)
2565 cd          write (iout,'(a)')
2566 cd        enddo
2567       enddo
2568       return
2569       end
2570 C--------------------------------------------------------------------------
2571       subroutine set_matrices
2572       implicit real*8 (a-h,o-z)
2573       include 'DIMENSIONS'
2574 #ifdef MPI
2575       include "mpif.h"
2576       include "COMMON.SETUP"
2577       integer IERR
2578       integer status(MPI_STATUS_SIZE)
2579 #endif
2580       include 'COMMON.IOUNITS'
2581       include 'COMMON.GEO'
2582       include 'COMMON.VAR'
2583       include 'COMMON.LOCAL'
2584       include 'COMMON.CHAIN'
2585       include 'COMMON.DERIV'
2586       include 'COMMON.INTERACT'
2587       include 'COMMON.CONTACTS'
2588       include 'COMMON.TORSION'
2589       include 'COMMON.VECTORS'
2590       include 'COMMON.FFIELD'
2591       double precision auxvec(2),auxmat(2,2)
2592 C
2593 C Compute the virtual-bond-torsional-angle dependent quantities needed
2594 C to calculate the el-loc multibody terms of various order.
2595 C
2596 #ifdef PARMAT
2597       do i=ivec_start+2,ivec_end+2
2598 #else
2599       do i=3,nres+1
2600 #endif
2601         if (i .lt. nres+1) then
2602           sin1=dsin(phi(i))
2603           cos1=dcos(phi(i))
2604           sintab(i-2)=sin1
2605           costab(i-2)=cos1
2606           obrot(1,i-2)=cos1
2607           obrot(2,i-2)=sin1
2608           sin2=dsin(2*phi(i))
2609           cos2=dcos(2*phi(i))
2610           sintab2(i-2)=sin2
2611           costab2(i-2)=cos2
2612           obrot2(1,i-2)=cos2
2613           obrot2(2,i-2)=sin2
2614           Ug(1,1,i-2)=-cos1
2615           Ug(1,2,i-2)=-sin1
2616           Ug(2,1,i-2)=-sin1
2617           Ug(2,2,i-2)= cos1
2618           Ug2(1,1,i-2)=-cos2
2619           Ug2(1,2,i-2)=-sin2
2620           Ug2(2,1,i-2)=-sin2
2621           Ug2(2,2,i-2)= cos2
2622         else
2623           costab(i-2)=1.0d0
2624           sintab(i-2)=0.0d0
2625           obrot(1,i-2)=1.0d0
2626           obrot(2,i-2)=0.0d0
2627           obrot2(1,i-2)=0.0d0
2628           obrot2(2,i-2)=0.0d0
2629           Ug(1,1,i-2)=1.0d0
2630           Ug(1,2,i-2)=0.0d0
2631           Ug(2,1,i-2)=0.0d0
2632           Ug(2,2,i-2)=1.0d0
2633           Ug2(1,1,i-2)=0.0d0
2634           Ug2(1,2,i-2)=0.0d0
2635           Ug2(2,1,i-2)=0.0d0
2636           Ug2(2,2,i-2)=0.0d0
2637         endif
2638         if (i .gt. 3 .and. i .lt. nres+1) then
2639           obrot_der(1,i-2)=-sin1
2640           obrot_der(2,i-2)= cos1
2641           Ugder(1,1,i-2)= sin1
2642           Ugder(1,2,i-2)=-cos1
2643           Ugder(2,1,i-2)=-cos1
2644           Ugder(2,2,i-2)=-sin1
2645           dwacos2=cos2+cos2
2646           dwasin2=sin2+sin2
2647           obrot2_der(1,i-2)=-dwasin2
2648           obrot2_der(2,i-2)= dwacos2
2649           Ug2der(1,1,i-2)= dwasin2
2650           Ug2der(1,2,i-2)=-dwacos2
2651           Ug2der(2,1,i-2)=-dwacos2
2652           Ug2der(2,2,i-2)=-dwasin2
2653         else
2654           obrot_der(1,i-2)=0.0d0
2655           obrot_der(2,i-2)=0.0d0
2656           Ugder(1,1,i-2)=0.0d0
2657           Ugder(1,2,i-2)=0.0d0
2658           Ugder(2,1,i-2)=0.0d0
2659           Ugder(2,2,i-2)=0.0d0
2660           obrot2_der(1,i-2)=0.0d0
2661           obrot2_der(2,i-2)=0.0d0
2662           Ug2der(1,1,i-2)=0.0d0
2663           Ug2der(1,2,i-2)=0.0d0
2664           Ug2der(2,1,i-2)=0.0d0
2665           Ug2der(2,2,i-2)=0.0d0
2666         endif
2667 c        if (i.gt. iatel_s+2 .and. i.lt.iatel_e+5) then
2668         if (i.gt. nnt+2 .and. i.lt.nct+2) then
2669           iti = itortyp(itype(i-2))
2670         else
2671           iti=ntortyp+1
2672         endif
2673 c        if (i.gt. iatel_s+1 .and. i.lt.iatel_e+4) then
2674         if (i.gt. nnt+1 .and. i.lt.nct+1) then
2675           iti1 = itortyp(itype(i-1))
2676         else
2677           iti1=ntortyp+1
2678         endif
2679 cd        write (iout,*) '*******i',i,' iti1',iti
2680 cd        write (iout,*) 'b1',b1(:,iti)
2681 cd        write (iout,*) 'b2',b2(:,iti)
2682 cd        write (iout,*) 'Ug',Ug(:,:,i-2)
2683 c        if (i .gt. iatel_s+2) then
2684         if (i .gt. nnt+2) then
2685           call matvec2(Ug(1,1,i-2),b2(1,iti),Ub2(1,i-2))
2686           call matmat2(EE(1,1,iti),Ug(1,1,i-2),EUg(1,1,i-2))
2687           if (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0) 
2688      &    then
2689           call matmat2(CC(1,1,iti),Ug(1,1,i-2),CUg(1,1,i-2))
2690           call matmat2(DD(1,1,iti),Ug(1,1,i-2),DUg(1,1,i-2))
2691           call matmat2(Dtilde(1,1,iti),Ug2(1,1,i-2),DtUg2(1,1,i-2))
2692           call matvec2(Ctilde(1,1,iti1),obrot(1,i-2),Ctobr(1,i-2))
2693           call matvec2(Dtilde(1,1,iti),obrot2(1,i-2),Dtobr2(1,i-2))
2694           endif
2695         else
2696           do k=1,2
2697             Ub2(k,i-2)=0.0d0
2698             Ctobr(k,i-2)=0.0d0 
2699             Dtobr2(k,i-2)=0.0d0
2700             do l=1,2
2701               EUg(l,k,i-2)=0.0d0
2702               CUg(l,k,i-2)=0.0d0
2703               DUg(l,k,i-2)=0.0d0
2704               DtUg2(l,k,i-2)=0.0d0
2705             enddo
2706           enddo
2707         endif
2708         call matvec2(Ugder(1,1,i-2),b2(1,iti),Ub2der(1,i-2))
2709         call matmat2(EE(1,1,iti),Ugder(1,1,i-2),EUgder(1,1,i-2))
2710         do k=1,2
2711           muder(k,i-2)=Ub2der(k,i-2)
2712         enddo
2713 c        if (i.gt. iatel_s+1 .and. i.lt.iatel_e+4) then
2714         if (i.gt. nnt+1 .and. i.lt.nct+1) then
2715           iti1 = itortyp(itype(i-1))
2716         else
2717           iti1=ntortyp+1
2718         endif
2719         do k=1,2
2720           mu(k,i-2)=Ub2(k,i-2)+b1(k,iti1)
2721         enddo
2722 cd        write (iout,*) 'mu ',mu(:,i-2)
2723 cd        write (iout,*) 'mu1',mu1(:,i-2)
2724 cd        write (iout,*) 'mu2',mu2(:,i-2)
2725         if (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or.wcorr6.gt.0.0d0)
2726      &  then  
2727         call matmat2(CC(1,1,iti1),Ugder(1,1,i-2),CUgder(1,1,i-2))
2728         call matmat2(DD(1,1,iti),Ugder(1,1,i-2),DUgder(1,1,i-2))
2729         call matmat2(Dtilde(1,1,iti),Ug2der(1,1,i-2),DtUg2der(1,1,i-2))
2730         call matvec2(Ctilde(1,1,iti1),obrot_der(1,i-2),Ctobrder(1,i-2))
2731         call matvec2(Dtilde(1,1,iti),obrot2_der(1,i-2),Dtobr2der(1,i-2))
2732 C Vectors and matrices dependent on a single virtual-bond dihedral.
2733         call matvec2(DD(1,1,iti),b1tilde(1,iti1),auxvec(1))
2734         call matvec2(Ug2(1,1,i-2),auxvec(1),Ug2Db1t(1,i-2)) 
2735         call matvec2(Ug2der(1,1,i-2),auxvec(1),Ug2Db1tder(1,i-2)) 
2736         call matvec2(CC(1,1,iti1),Ub2(1,i-2),CUgb2(1,i-2))
2737         call matvec2(CC(1,1,iti1),Ub2der(1,i-2),CUgb2der(1,i-2))
2738         call matmat2(EUg(1,1,i-2),CC(1,1,iti1),EUgC(1,1,i-2))
2739         call matmat2(EUgder(1,1,i-2),CC(1,1,iti1),EUgCder(1,1,i-2))
2740         call matmat2(EUg(1,1,i-2),DD(1,1,iti1),EUgD(1,1,i-2))
2741         call matmat2(EUgder(1,1,i-2),DD(1,1,iti1),EUgDder(1,1,i-2))
2742         endif
2743       enddo
2744 C Matrices dependent on two consecutive virtual-bond dihedrals.
2745 C The order of matrices is from left to right.
2746       if (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or.wcorr6.gt.0.0d0)
2747      &then
2748 c      do i=max0(ivec_start,2),ivec_end
2749       do i=2,nres-1
2750         call matmat2(DtUg2(1,1,i-1),EUg(1,1,i),DtUg2EUg(1,1,i))
2751         call matmat2(DtUg2der(1,1,i-1),EUg(1,1,i),DtUg2EUgder(1,1,1,i))
2752         call matmat2(DtUg2(1,1,i-1),EUgder(1,1,i),DtUg2EUgder(1,1,2,i))
2753         call transpose2(DtUg2(1,1,i-1),auxmat(1,1))
2754         call matmat2(auxmat(1,1),EUg(1,1,i),Ug2DtEUg(1,1,i))
2755         call matmat2(auxmat(1,1),EUgder(1,1,i),Ug2DtEUgder(1,1,2,i))
2756         call transpose2(DtUg2der(1,1,i-1),auxmat(1,1))
2757         call matmat2(auxmat(1,1),EUg(1,1,i),Ug2DtEUgder(1,1,1,i))
2758       enddo
2759       endif
2760 #if defined(MPI) && defined(PARMAT)
2761 #ifdef DEBUG
2762 c      if (fg_rank.eq.0) then
2763         write (iout,*) "Arrays UG and UGDER before GATHER"
2764         do i=1,nres-1
2765           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2766      &     ((ug(l,k,i),l=1,2),k=1,2),
2767      &     ((ugder(l,k,i),l=1,2),k=1,2)
2768         enddo
2769         write (iout,*) "Arrays UG2 and UG2DER"
2770         do i=1,nres-1
2771           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2772      &     ((ug2(l,k,i),l=1,2),k=1,2),
2773      &     ((ug2der(l,k,i),l=1,2),k=1,2)
2774         enddo
2775         write (iout,*) "Arrays OBROT OBROT2 OBROTDER and OBROT2DER"
2776         do i=1,nres-1
2777           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2778      &     (obrot(k,i),k=1,2),(obrot2(k,i),k=1,2),
2779      &     (obrot_der(k,i),k=1,2),(obrot2_der(k,i),k=1,2)
2780         enddo
2781         write (iout,*) "Arrays COSTAB SINTAB COSTAB2 and SINTAB2"
2782         do i=1,nres-1
2783           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2784      &     costab(i),sintab(i),costab2(i),sintab2(i)
2785         enddo
2786         write (iout,*) "Array MUDER"
2787         do i=1,nres-1
2788           write (iout,'(i5,2f10.5)') i,muder(1,i),muder(2,i)
2789         enddo
2790 c      endif
2791 #endif
2792       if (nfgtasks.gt.1) then
2793         time00=MPI_Wtime()
2794 c        write(iout,*)"Processor",fg_rank,kolor," ivec_start",ivec_start,
2795 c     &   " ivec_displ",(ivec_displ(i),i=0,nfgtasks-1),
2796 c     &   " ivec_count",(ivec_count(i),i=0,nfgtasks-1)
2797 #ifdef MATGATHER
2798         call MPI_Allgatherv(Ub2(1,ivec_start),ivec_count(fg_rank1),
2799      &   MPI_MU,Ub2(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2800      &   FG_COMM1,IERR)
2801         call MPI_Allgatherv(Ub2der(1,ivec_start),ivec_count(fg_rank1),
2802      &   MPI_MU,Ub2der(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2803      &   FG_COMM1,IERR)
2804         call MPI_Allgatherv(mu(1,ivec_start),ivec_count(fg_rank1),
2805      &   MPI_MU,mu(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2806      &   FG_COMM1,IERR)
2807         call MPI_Allgatherv(muder(1,ivec_start),ivec_count(fg_rank1),
2808      &   MPI_MU,muder(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2809      &   FG_COMM1,IERR)
2810         call MPI_Allgatherv(Eug(1,1,ivec_start),ivec_count(fg_rank1),
2811      &   MPI_MAT1,Eug(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2812      &   FG_COMM1,IERR)
2813         call MPI_Allgatherv(Eugder(1,1,ivec_start),ivec_count(fg_rank1),
2814      &   MPI_MAT1,Eugder(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2815      &   FG_COMM1,IERR)
2816         call MPI_Allgatherv(costab(ivec_start),ivec_count(fg_rank1),
2817      &   MPI_DOUBLE_PRECISION,costab(1),ivec_count(0),ivec_displ(0),
2818      &   MPI_DOUBLE_PRECISION,FG_COMM1,IERR)
2819         call MPI_Allgatherv(sintab(ivec_start),ivec_count(fg_rank1),
2820      &   MPI_DOUBLE_PRECISION,sintab(1),ivec_count(0),ivec_displ(0),
2821      &   MPI_DOUBLE_PRECISION,FG_COMM1,IERR)
2822         call MPI_Allgatherv(costab2(ivec_start),ivec_count(fg_rank1),
2823      &   MPI_DOUBLE_PRECISION,costab2(1),ivec_count(0),ivec_displ(0),
2824      &   MPI_DOUBLE_PRECISION,FG_COMM1,IERR)
2825         call MPI_Allgatherv(sintab2(ivec_start),ivec_count(fg_rank1),
2826      &   MPI_DOUBLE_PRECISION,sintab2(1),ivec_count(0),ivec_displ(0),
2827      &   MPI_DOUBLE_PRECISION,FG_COMM1,IERR)
2828         if (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0)
2829      &  then
2830         call MPI_Allgatherv(Ctobr(1,ivec_start),ivec_count(fg_rank1),
2831      &   MPI_MU,Ctobr(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2832      &   FG_COMM1,IERR)
2833         call MPI_Allgatherv(Ctobrder(1,ivec_start),ivec_count(fg_rank1),
2834      &   MPI_MU,Ctobrder(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2835      &   FG_COMM1,IERR)
2836         call MPI_Allgatherv(Dtobr2(1,ivec_start),ivec_count(fg_rank1),
2837      &   MPI_MU,Dtobr2(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2838      &   FG_COMM1,IERR)
2839        call MPI_Allgatherv(Dtobr2der(1,ivec_start),ivec_count(fg_rank1),
2840      &   MPI_MU,Dtobr2der(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2841      &   FG_COMM1,IERR)
2842         call MPI_Allgatherv(Ug2Db1t(1,ivec_start),ivec_count(fg_rank1),
2843      &   MPI_MU,Ug2Db1t(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2844      &   FG_COMM1,IERR)
2845         call MPI_Allgatherv(Ug2Db1tder(1,ivec_start),
2846      &   ivec_count(fg_rank1),
2847      &   MPI_MU,Ug2Db1tder(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2848      &   FG_COMM1,IERR)
2849         call MPI_Allgatherv(CUgb2(1,ivec_start),ivec_count(fg_rank1),
2850      &   MPI_MU,CUgb2(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2851      &   FG_COMM1,IERR)
2852         call MPI_Allgatherv(CUgb2der(1,ivec_start),ivec_count(fg_rank1),
2853      &   MPI_MU,CUgb2der(1,1),ivec_count(0),ivec_displ(0),MPI_MU,
2854      &   FG_COMM1,IERR)
2855         call MPI_Allgatherv(Cug(1,1,ivec_start),ivec_count(fg_rank1),
2856      &   MPI_MAT1,Cug(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2857      &   FG_COMM1,IERR)
2858         call MPI_Allgatherv(Cugder(1,1,ivec_start),ivec_count(fg_rank1),
2859      &   MPI_MAT1,Cugder(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2860      &   FG_COMM1,IERR)
2861         call MPI_Allgatherv(Dug(1,1,ivec_start),ivec_count(fg_rank1),
2862      &   MPI_MAT1,Dug(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2863      &   FG_COMM1,IERR)
2864         call MPI_Allgatherv(Dugder(1,1,ivec_start),ivec_count(fg_rank1),
2865      &   MPI_MAT1,Dugder(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2866      &   FG_COMM1,IERR)
2867         call MPI_Allgatherv(Dtug2(1,1,ivec_start),ivec_count(fg_rank1),
2868      &   MPI_MAT1,Dtug2(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2869      &   FG_COMM1,IERR)
2870         call MPI_Allgatherv(Dtug2der(1,1,ivec_start),
2871      &   ivec_count(fg_rank1),
2872      &   MPI_MAT1,Dtug2der(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2873      &   FG_COMM1,IERR)
2874         call MPI_Allgatherv(EugC(1,1,ivec_start),ivec_count(fg_rank1),
2875      &   MPI_MAT1,EugC(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2876      &   FG_COMM1,IERR)
2877        call MPI_Allgatherv(EugCder(1,1,ivec_start),ivec_count(fg_rank1),
2878      &   MPI_MAT1,EugCder(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2879      &   FG_COMM1,IERR)
2880         call MPI_Allgatherv(EugD(1,1,ivec_start),ivec_count(fg_rank1),
2881      &   MPI_MAT1,EugD(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2882      &   FG_COMM1,IERR)
2883        call MPI_Allgatherv(EugDder(1,1,ivec_start),ivec_count(fg_rank1),
2884      &   MPI_MAT1,EugDder(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2885      &   FG_COMM1,IERR)
2886         call MPI_Allgatherv(DtUg2EUg(1,1,ivec_start),
2887      &   ivec_count(fg_rank1),
2888      &   MPI_MAT1,DtUg2EUg(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2889      &   FG_COMM1,IERR)
2890         call MPI_Allgatherv(Ug2DtEUg(1,1,ivec_start),
2891      &   ivec_count(fg_rank1),
2892      &   MPI_MAT1,Ug2DtEUg(1,1,1),ivec_count(0),ivec_displ(0),MPI_MAT1,
2893      &   FG_COMM1,IERR)
2894         call MPI_Allgatherv(DtUg2EUgder(1,1,1,ivec_start),
2895      &   ivec_count(fg_rank1),
2896      &   MPI_MAT2,DtUg2EUgder(1,1,1,1),ivec_count(0),ivec_displ(0),
2897      &   MPI_MAT2,FG_COMM1,IERR)
2898         call MPI_Allgatherv(Ug2DtEUgder(1,1,1,ivec_start),
2899      &   ivec_count(fg_rank1),
2900      &   MPI_MAT2,Ug2DtEUgder(1,1,1,1),ivec_count(0),ivec_displ(0),
2901      &   MPI_MAT2,FG_COMM1,IERR)
2902         endif
2903 #else
2904 c Passes matrix info through the ring
2905       isend=fg_rank1
2906       irecv=fg_rank1-1
2907       if (irecv.lt.0) irecv=nfgtasks1-1 
2908       iprev=irecv
2909       inext=fg_rank1+1
2910       if (inext.ge.nfgtasks1) inext=0
2911       do i=1,nfgtasks1-1
2912 c        write (iout,*) "isend",isend," irecv",irecv
2913 c        call flush(iout)
2914         lensend=lentyp(isend)
2915         lenrecv=lentyp(irecv)
2916 c        write (iout,*) "lensend",lensend," lenrecv",lenrecv
2917 c        call MPI_SENDRECV(ug(1,1,ivec_displ(isend)+1),1,
2918 c     &   MPI_ROTAT1(lensend),inext,2200+isend,
2919 c     &   ug(1,1,ivec_displ(irecv)+1),1,MPI_ROTAT1(lenrecv),
2920 c     &   iprev,2200+irecv,FG_COMM,status,IERR)
2921 c        write (iout,*) "Gather ROTAT1"
2922 c        call flush(iout)
2923 c        call MPI_SENDRECV(obrot(1,ivec_displ(isend)+1),1,
2924 c     &   MPI_ROTAT2(lensend),inext,3300+isend,
2925 c     &   obrot(1,ivec_displ(irecv)+1),1,MPI_ROTAT2(lenrecv),
2926 c     &   iprev,3300+irecv,FG_COMM,status,IERR)
2927 c        write (iout,*) "Gather ROTAT2"
2928 c        call flush(iout)
2929         call MPI_SENDRECV(costab(ivec_displ(isend)+1),1,
2930      &   MPI_ROTAT_OLD(lensend),inext,4400+isend,
2931      &   costab(ivec_displ(irecv)+1),1,MPI_ROTAT_OLD(lenrecv),
2932      &   iprev,4400+irecv,FG_COMM,status,IERR)
2933 c        write (iout,*) "Gather ROTAT_OLD"
2934 c        call flush(iout)
2935         call MPI_SENDRECV(mu(1,ivec_displ(isend)+1),1,
2936      &   MPI_PRECOMP11(lensend),inext,5500+isend,
2937      &   mu(1,ivec_displ(irecv)+1),1,MPI_PRECOMP11(lenrecv),
2938      &   iprev,5500+irecv,FG_COMM,status,IERR)
2939 c        write (iout,*) "Gather PRECOMP11"
2940 c        call flush(iout)
2941         call MPI_SENDRECV(Eug(1,1,ivec_displ(isend)+1),1,
2942      &   MPI_PRECOMP12(lensend),inext,6600+isend,
2943      &   Eug(1,1,ivec_displ(irecv)+1),1,MPI_PRECOMP12(lenrecv),
2944      &   iprev,6600+irecv,FG_COMM,status,IERR)
2945 c        write (iout,*) "Gather PRECOMP12"
2946 c        call flush(iout)
2947         if (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0) 
2948      &  then
2949         call MPI_SENDRECV(ug2db1t(1,ivec_displ(isend)+1),1,
2950      &   MPI_ROTAT2(lensend),inext,7700+isend,
2951      &   ug2db1t(1,ivec_displ(irecv)+1),1,MPI_ROTAT2(lenrecv),
2952      &   iprev,7700+irecv,FG_COMM,status,IERR)
2953 c        write (iout,*) "Gather PRECOMP21"
2954 c        call flush(iout)
2955         call MPI_SENDRECV(EUgC(1,1,ivec_displ(isend)+1),1,
2956      &   MPI_PRECOMP22(lensend),inext,8800+isend,
2957      &   EUgC(1,1,ivec_displ(irecv)+1),1,MPI_PRECOMP22(lenrecv),
2958      &   iprev,8800+irecv,FG_COMM,status,IERR)
2959 c        write (iout,*) "Gather PRECOMP22"
2960 c        call flush(iout)
2961         call MPI_SENDRECV(Ug2DtEUgder(1,1,1,ivec_displ(isend)+1),1,
2962      &   MPI_PRECOMP23(lensend),inext,9900+isend,
2963      &   Ug2DtEUgder(1,1,1,ivec_displ(irecv)+1),1,
2964      &   MPI_PRECOMP23(lenrecv),
2965      &   iprev,9900+irecv,FG_COMM,status,IERR)
2966 c        write (iout,*) "Gather PRECOMP23"
2967 c        call flush(iout)
2968         endif
2969         isend=irecv
2970         irecv=irecv-1
2971         if (irecv.lt.0) irecv=nfgtasks1-1
2972       enddo
2973 #endif
2974         time_gather=time_gather+MPI_Wtime()-time00
2975       endif
2976 #ifdef DEBUG
2977 c      if (fg_rank.eq.0) then
2978         write (iout,*) "Arrays UG and UGDER"
2979         do i=1,nres-1
2980           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2981      &     ((ug(l,k,i),l=1,2),k=1,2),
2982      &     ((ugder(l,k,i),l=1,2),k=1,2)
2983         enddo
2984         write (iout,*) "Arrays UG2 and UG2DER"
2985         do i=1,nres-1
2986           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2987      &     ((ug2(l,k,i),l=1,2),k=1,2),
2988      &     ((ug2der(l,k,i),l=1,2),k=1,2)
2989         enddo
2990         write (iout,*) "Arrays OBROT OBROT2 OBROTDER and OBROT2DER"
2991         do i=1,nres-1
2992           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2993      &     (obrot(k,i),k=1,2),(obrot2(k,i),k=1,2),
2994      &     (obrot_der(k,i),k=1,2),(obrot2_der(k,i),k=1,2)
2995         enddo
2996         write (iout,*) "Arrays COSTAB SINTAB COSTAB2 and SINTAB2"
2997         do i=1,nres-1
2998           write (iout,'(i5,4f10.5,5x,4f10.5)') i,
2999      &     costab(i),sintab(i),costab2(i),sintab2(i)
3000         enddo
3001         write (iout,*) "Array MUDER"
3002         do i=1,nres-1
3003           write (iout,'(i5,2f10.5)') i,muder(1,i),muder(2,i)
3004         enddo
3005 c      endif
3006 #endif
3007 #endif
3008 cd      do i=1,nres
3009 cd        iti = itortyp(itype(i))
3010 cd        write (iout,*) i
3011 cd        do j=1,2
3012 cd        write (iout,'(2f10.5,5x,2f10.5,5x,2f10.5)') 
3013 cd     &  (EE(j,k,iti),k=1,2),(Ug(j,k,i),k=1,2),(EUg(j,k,i),k=1,2)
3014 cd        enddo
3015 cd      enddo
3016       return
3017       end
3018 C--------------------------------------------------------------------------
3019       subroutine eelec(ees,evdw1,eel_loc,eello_turn3,eello_turn4)
3020 C
3021 C This subroutine calculates the average interaction energy and its gradient
3022 C in the virtual-bond vectors between non-adjacent peptide groups, based on 
3023 C the potential described in Liwo et al., Protein Sci., 1993, 2, 1715. 
3024 C The potential depends both on the distance of peptide-group centers and on 
3025 C the orientation of the CA-CA virtual bonds.
3026
3027       implicit real*8 (a-h,o-z)
3028 #ifdef MPI
3029       include 'mpif.h'
3030 #endif
3031       include 'DIMENSIONS'
3032       include 'COMMON.CONTROL'
3033       include 'COMMON.SETUP'
3034       include 'COMMON.IOUNITS'
3035       include 'COMMON.GEO'
3036       include 'COMMON.VAR'
3037       include 'COMMON.LOCAL'
3038       include 'COMMON.CHAIN'
3039       include 'COMMON.DERIV'
3040       include 'COMMON.INTERACT'
3041       include 'COMMON.CONTACTS'
3042       include 'COMMON.TORSION'
3043       include 'COMMON.VECTORS'
3044       include 'COMMON.FFIELD'
3045       include 'COMMON.TIME1'
3046       dimension ggg(3),gggp(3),gggm(3),erij(3),dcosb(3),dcosg(3),
3047      &          erder(3,3),uryg(3,3),urzg(3,3),vryg(3,3),vrzg(3,3)
3048       double precision acipa(2,2),agg(3,4),aggi(3,4),aggi1(3,4),
3049      &    aggj(3,4),aggj1(3,4),a_temp(2,2),muij(4)
3050       common /locel/ a_temp,agg,aggi,aggi1,aggj,aggj1,a22,a23,a32,a33,
3051      &    dxi,dyi,dzi,dx_normi,dy_normi,dz_normi,xmedi,ymedi,zmedi,
3052      &    num_conti,j1,j2
3053 c 4/26/02 - AL scaling factor for 1,4 repulsive VDW interactions
3054 #ifdef MOMENT
3055       double precision scal_el /1.0d0/
3056 #else
3057       double precision scal_el /0.5d0/
3058 #endif
3059 C 12/13/98 
3060 C 13-go grudnia roku pamietnego... 
3061       double precision unmat(3,3) /1.0d0,0.0d0,0.0d0,
3062      &                   0.0d0,1.0d0,0.0d0,
3063      &                   0.0d0,0.0d0,1.0d0/
3064 cd      write(iout,*) 'In EELEC'
3065 cd      do i=1,nloctyp
3066 cd        write(iout,*) 'Type',i
3067 cd        write(iout,*) 'B1',B1(:,i)
3068 cd        write(iout,*) 'B2',B2(:,i)
3069 cd        write(iout,*) 'CC',CC(:,:,i)
3070 cd        write(iout,*) 'DD',DD(:,:,i)
3071 cd        write(iout,*) 'EE',EE(:,:,i)
3072 cd      enddo
3073 cd      call check_vecgrad
3074 cd      stop
3075       if (icheckgrad.eq.1) then
3076         do i=1,nres-1
3077           fac=1.0d0/dsqrt(scalar(dc(1,i),dc(1,i)))
3078           do k=1,3
3079             dc_norm(k,i)=dc(k,i)*fac
3080           enddo
3081 c          write (iout,*) 'i',i,' fac',fac
3082         enddo
3083       endif
3084       if (wel_loc.gt.0.0d0 .or. wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 
3085      &    .or. wcorr6.gt.0.0d0 .or. wturn3.gt.0.0d0 .or. 
3086      &    wturn4.gt.0.0d0 .or. wturn6.gt.0.0d0) then
3087 c        call vec_and_deriv
3088 #ifdef TIMING
3089         time01=MPI_Wtime()
3090 #endif
3091         call set_matrices
3092 #ifdef TIMING
3093         time_mat=time_mat+MPI_Wtime()-time01
3094 #endif
3095       endif
3096 cd      do i=1,nres-1
3097 cd        write (iout,*) 'i=',i
3098 cd        do k=1,3
3099 cd        write (iout,'(i5,2f10.5)') k,uy(k,i),uz(k,i)
3100 cd        enddo
3101 cd        do k=1,3
3102 cd          write (iout,'(f10.5,2x,3f10.5,2x,3f10.5)') 
3103 cd     &     uz(k,i),(uzgrad(k,l,1,i),l=1,3),(uzgrad(k,l,2,i),l=1,3)
3104 cd        enddo
3105 cd      enddo
3106       t_eelecij=0.0d0
3107       ees=0.0D0
3108       evdw1=0.0D0
3109       eel_loc=0.0d0 
3110       eello_turn3=0.0d0
3111       eello_turn4=0.0d0
3112       ind=0
3113       do i=1,nres
3114         num_cont_hb(i)=0
3115       enddo
3116 cd      print '(a)','Enter EELEC'
3117 cd      write (iout,*) 'iatel_s=',iatel_s,' iatel_e=',iatel_e
3118       do i=1,nres
3119         gel_loc_loc(i)=0.0d0
3120         gcorr_loc(i)=0.0d0
3121       enddo
3122 c
3123 c
3124 c 9/27/08 AL Split the interaction loop to ensure load balancing of turn terms
3125 C
3126 C Loop over i,i+2 and i,i+3 pairs of the peptide groups
3127 C
3128       do i=iturn3_start,iturn3_end
3129         dxi=dc(1,i)
3130         dyi=dc(2,i)
3131         dzi=dc(3,i)
3132         dx_normi=dc_norm(1,i)
3133         dy_normi=dc_norm(2,i)
3134         dz_normi=dc_norm(3,i)
3135         xmedi=c(1,i)+0.5d0*dxi
3136         ymedi=c(2,i)+0.5d0*dyi
3137         zmedi=c(3,i)+0.5d0*dzi
3138         num_conti=0
3139         call eelecij(i,i+2,ees,evdw1,eel_loc)
3140         if (wturn3.gt.0.0d0) call eturn3(i,eello_turn3)
3141         num_cont_hb(i)=num_conti
3142       enddo
3143       do i=iturn4_start,iturn4_end
3144         dxi=dc(1,i)
3145         dyi=dc(2,i)
3146         dzi=dc(3,i)
3147         dx_normi=dc_norm(1,i)
3148         dy_normi=dc_norm(2,i)
3149         dz_normi=dc_norm(3,i)
3150         xmedi=c(1,i)+0.5d0*dxi
3151         ymedi=c(2,i)+0.5d0*dyi
3152         zmedi=c(3,i)+0.5d0*dzi
3153         num_conti=num_cont_hb(i)
3154         call eelecij(i,i+3,ees,evdw1,eel_loc)
3155         if (wturn4.gt.0.0d0) call eturn4(i,eello_turn4)
3156         num_cont_hb(i)=num_conti
3157       enddo   ! i
3158 c
3159 c Loop over all pairs of interacting peptide groups except i,i+2 and i,i+3
3160 c
3161       do i=iatel_s,iatel_e
3162         dxi=dc(1,i)
3163         dyi=dc(2,i)
3164         dzi=dc(3,i)
3165         dx_normi=dc_norm(1,i)
3166         dy_normi=dc_norm(2,i)
3167         dz_normi=dc_norm(3,i)
3168         xmedi=c(1,i)+0.5d0*dxi
3169         ymedi=c(2,i)+0.5d0*dyi
3170         zmedi=c(3,i)+0.5d0*dzi
3171 c        write (iout,*) 'i',i,' ielstart',ielstart(i),' ielend',ielend(i)
3172         num_conti=num_cont_hb(i)
3173         do j=ielstart(i),ielend(i)
3174           call eelecij(i,j,ees,evdw1,eel_loc)
3175         enddo ! j
3176         num_cont_hb(i)=num_conti
3177       enddo   ! i
3178 c      write (iout,*) "Number of loop steps in EELEC:",ind
3179 cd      do i=1,nres
3180 cd        write (iout,'(i3,3f10.5,5x,3f10.5)') 
3181 cd     &     i,(gel_loc(k,i),k=1,3),gel_loc_loc(i)
3182 cd      enddo
3183 c 12/7/99 Adam eello_turn3 will be considered as a separate energy term
3184 ccc      eel_loc=eel_loc+eello_turn3
3185 cd      print *,"Processor",fg_rank," t_eelecij",t_eelecij
3186       return
3187       end
3188 C-------------------------------------------------------------------------------
3189       subroutine eelecij(i,j,ees,evdw1,eel_loc)
3190       implicit real*8 (a-h,o-z)
3191       include 'DIMENSIONS'
3192 #ifdef MPI
3193       include "mpif.h"
3194 #endif
3195       include 'COMMON.CONTROL'
3196       include 'COMMON.IOUNITS'
3197       include 'COMMON.GEO'
3198       include 'COMMON.VAR'
3199       include 'COMMON.LOCAL'
3200       include 'COMMON.CHAIN'
3201       include 'COMMON.DERIV'
3202       include 'COMMON.INTERACT'
3203       include 'COMMON.CONTACTS'
3204       include 'COMMON.TORSION'
3205       include 'COMMON.VECTORS'
3206       include 'COMMON.FFIELD'
3207       include 'COMMON.TIME1'
3208       dimension ggg(3),gggp(3),gggm(3),erij(3),dcosb(3),dcosg(3),
3209      &          erder(3,3),uryg(3,3),urzg(3,3),vryg(3,3),vrzg(3,3)
3210       double precision acipa(2,2),agg(3,4),aggi(3,4),aggi1(3,4),
3211      &    aggj(3,4),aggj1(3,4),a_temp(2,2),muij(4)
3212       common /locel/ a_temp,agg,aggi,aggi1,aggj,aggj1,a22,a23,a32,a33,
3213      &    dxi,dyi,dzi,dx_normi,dy_normi,dz_normi,xmedi,ymedi,zmedi,
3214      &    num_conti,j1,j2
3215 c 4/26/02 - AL scaling factor for 1,4 repulsive VDW interactions
3216 #ifdef MOMENT
3217       double precision scal_el /1.0d0/
3218 #else
3219       double precision scal_el /0.5d0/
3220 #endif
3221 C 12/13/98 
3222 C 13-go grudnia roku pamietnego... 
3223       double precision unmat(3,3) /1.0d0,0.0d0,0.0d0,
3224      &                   0.0d0,1.0d0,0.0d0,
3225      &                   0.0d0,0.0d0,1.0d0/
3226 c          time00=MPI_Wtime()
3227 cd      write (iout,*) "eelecij",i,j
3228 c          ind=ind+1
3229           iteli=itel(i)
3230           itelj=itel(j)
3231           if (j.eq.i+2 .and. itelj.eq.2) iteli=2
3232           aaa=app(iteli,itelj)
3233           bbb=bpp(iteli,itelj)
3234           ael6i=ael6(iteli,itelj)
3235           ael3i=ael3(iteli,itelj) 
3236           dxj=dc(1,j)
3237           dyj=dc(2,j)
3238           dzj=dc(3,j)
3239           dx_normj=dc_norm(1,j)
3240           dy_normj=dc_norm(2,j)
3241           dz_normj=dc_norm(3,j)
3242           xj=c(1,j)+0.5D0*dxj-xmedi
3243           yj=c(2,j)+0.5D0*dyj-ymedi
3244           zj=c(3,j)+0.5D0*dzj-zmedi
3245           rij=xj*xj+yj*yj+zj*zj
3246           rrmij=1.0D0/rij
3247           rij=dsqrt(rij)
3248           rmij=1.0D0/rij
3249           r3ij=rrmij*rmij
3250           r6ij=r3ij*r3ij  
3251           cosa=dx_normi*dx_normj+dy_normi*dy_normj+dz_normi*dz_normj
3252           cosb=(xj*dx_normi+yj*dy_normi+zj*dz_normi)*rmij
3253           cosg=(xj*dx_normj+yj*dy_normj+zj*dz_normj)*rmij
3254           fac=cosa-3.0D0*cosb*cosg
3255           ev1=aaa*r6ij*r6ij
3256 c 4/26/02 - AL scaling down 1,4 repulsive VDW interactions
3257           if (j.eq.i+2) ev1=scal_el*ev1
3258           ev2=bbb*r6ij
3259           fac3=ael6i*r6ij
3260           fac4=ael3i*r3ij
3261           evdwij=ev1+ev2
3262           el1=fac3*(4.0D0+fac*fac-3.0D0*(cosb*cosb+cosg*cosg))
3263           el2=fac4*fac       
3264           eesij=el1+el2
3265 C 12/26/95 - for the evaluation of multi-body H-bonding interactions
3266           ees0ij=4.0D0+fac*fac-3.0D0*(cosb*cosb+cosg*cosg)
3267           ees=ees+eesij
3268           evdw1=evdw1+evdwij
3269 cd          write(iout,'(2(2i3,2x),7(1pd12.4)/2(3(1pd12.4),5x)/)')
3270 cd     &      iteli,i,itelj,j,aaa,bbb,ael6i,ael3i,
3271 cd     &      1.0D0/dsqrt(rrmij),evdwij,eesij,
3272 cd     &      xmedi,ymedi,zmedi,xj,yj,zj
3273
3274           if (energy_dec) then 
3275               write (iout,'(a6,2i5,0pf7.3)') 'evdw1',i,j,evdwij
3276               write (iout,'(a6,2i5,0pf7.3)') 'ees',i,j,eesij
3277           endif
3278
3279 C
3280 C Calculate contributions to the Cartesian gradient.
3281 C
3282 #ifdef SPLITELE
3283           facvdw=-6*rrmij*(ev1+evdwij)
3284           facel=-3*rrmij*(el1+eesij)
3285           fac1=fac
3286           erij(1)=xj*rmij
3287           erij(2)=yj*rmij
3288           erij(3)=zj*rmij
3289 *
3290 * Radial derivatives. First process both termini of the fragment (i,j)
3291 *
3292           ggg(1)=facel*xj
3293           ggg(2)=facel*yj
3294           ggg(3)=facel*zj
3295 c          do k=1,3
3296 c            ghalf=0.5D0*ggg(k)
3297 c            gelc(k,i)=gelc(k,i)+ghalf
3298 c            gelc(k,j)=gelc(k,j)+ghalf
3299 c          enddo
3300 c 9/28/08 AL Gradient compotents will be summed only at the end
3301           do k=1,3
3302             gelc_long(k,j)=gelc_long(k,j)+ggg(k)
3303             gelc_long(k,i)=gelc_long(k,i)-ggg(k)
3304           enddo
3305 *
3306 * Loop over residues i+1 thru j-1.
3307 *
3308 cgrad          do k=i+1,j-1
3309 cgrad            do l=1,3
3310 cgrad              gelc(l,k)=gelc(l,k)+ggg(l)
3311 cgrad            enddo
3312 cgrad          enddo
3313           ggg(1)=facvdw*xj
3314           ggg(2)=facvdw*yj
3315           ggg(3)=facvdw*zj
3316 c          do k=1,3
3317 c            ghalf=0.5D0*ggg(k)
3318 c            gvdwpp(k,i)=gvdwpp(k,i)+ghalf
3319 c            gvdwpp(k,j)=gvdwpp(k,j)+ghalf
3320 c          enddo
3321 c 9/28/08 AL Gradient compotents will be summed only at the end
3322           do k=1,3
3323             gvdwpp(k,j)=gvdwpp(k,j)+ggg(k)
3324             gvdwpp(k,i)=gvdwpp(k,i)-ggg(k)
3325           enddo
3326 *
3327 * Loop over residues i+1 thru j-1.
3328 *
3329 cgrad          do k=i+1,j-1
3330 cgrad            do l=1,3
3331 cgrad              gvdwpp(l,k)=gvdwpp(l,k)+ggg(l)
3332 cgrad            enddo
3333 cgrad          enddo
3334 #else
3335           facvdw=ev1+evdwij 
3336           facel=el1+eesij  
3337           fac1=fac
3338           fac=-3*rrmij*(facvdw+facvdw+facel)
3339           erij(1)=xj*rmij
3340           erij(2)=yj*rmij
3341           erij(3)=zj*rmij
3342 *
3343 * Radial derivatives. First process both termini of the fragment (i,j)
3344
3345           ggg(1)=fac*xj
3346           ggg(2)=fac*yj
3347           ggg(3)=fac*zj
3348 c          do k=1,3
3349 c            ghalf=0.5D0*ggg(k)
3350 c            gelc(k,i)=gelc(k,i)+ghalf
3351 c            gelc(k,j)=gelc(k,j)+ghalf
3352 c          enddo
3353 c 9/28/08 AL Gradient compotents will be summed only at the end
3354           do k=1,3
3355             gelc_long(k,j)=gelc(k,j)+ggg(k)
3356             gelc_long(k,i)=gelc(k,i)-ggg(k)
3357           enddo
3358 *
3359 * Loop over residues i+1 thru j-1.
3360 *
3361 cgrad          do k=i+1,j-1
3362 cgrad            do l=1,3
3363 cgrad              gelc(l,k)=gelc(l,k)+ggg(l)
3364 cgrad            enddo
3365 cgrad          enddo
3366 c 9/28/08 AL Gradient compotents will be summed only at the end
3367           ggg(1)=facvdw*xj
3368           ggg(2)=facvdw*yj
3369           ggg(3)=facvdw*zj
3370           do k=1,3
3371             gvdwpp(k,j)=gvdwpp(k,j)+ggg(k)
3372             gvdwpp(k,i)=gvdwpp(k,i)-ggg(k)
3373           enddo
3374 #endif
3375 *
3376 * Angular part
3377 *          
3378           ecosa=2.0D0*fac3*fac1+fac4
3379           fac4=-3.0D0*fac4
3380           fac3=-6.0D0*fac3
3381           ecosb=(fac3*(fac1*cosg+cosb)+cosg*fac4)
3382           ecosg=(fac3*(fac1*cosb+cosg)+cosb*fac4)
3383           do k=1,3
3384             dcosb(k)=rmij*(dc_norm(k,i)-erij(k)*cosb)
3385             dcosg(k)=rmij*(dc_norm(k,j)-erij(k)*cosg)
3386           enddo
3387 cd        print '(2i3,2(3(1pd14.5),3x))',i,j,(dcosb(k),k=1,3),
3388 cd   &          (dcosg(k),k=1,3)
3389           do k=1,3
3390             ggg(k)=ecosb*dcosb(k)+ecosg*dcosg(k) 
3391           enddo
3392 c          do k=1,3
3393 c            ghalf=0.5D0*ggg(k)
3394 c            gelc(k,i)=gelc(k,i)+ghalf
3395 c     &               +(ecosa*(dc_norm(k,j)-cosa*dc_norm(k,i))
3396 c     &               + ecosb*(erij(k)-cosb*dc_norm(k,i)))*vbld_inv(i+1)
3397 c            gelc(k,j)=gelc(k,j)+ghalf
3398 c     &               +(ecosa*(dc_norm(k,i)-cosa*dc_norm(k,j))
3399 c     &               + ecosg*(erij(k)-cosg*dc_norm(k,j)))*vbld_inv(j+1)
3400 c          enddo
3401 cgrad          do k=i+1,j-1
3402 cgrad            do l=1,3
3403 cgrad              gelc(l,k)=gelc(l,k)+ggg(l)
3404 cgrad            enddo
3405 cgrad          enddo
3406           do k=1,3
3407             gelc(k,i)=gelc(k,i)
3408      &               +(ecosa*(dc_norm(k,j)-cosa*dc_norm(k,i))
3409      &               + ecosb*(erij(k)-cosb*dc_norm(k,i)))*vbld_inv(i+1)
3410             gelc(k,j)=gelc(k,j)
3411      &               +(ecosa*(dc_norm(k,i)-cosa*dc_norm(k,j))
3412      &               + ecosg*(erij(k)-cosg*dc_norm(k,j)))*vbld_inv(j+1)
3413             gelc_long(k,j)=gelc_long(k,j)+ggg(k)
3414             gelc_long(k,i)=gelc_long(k,i)-ggg(k)
3415           enddo
3416           IF (wel_loc.gt.0.0d0 .or. wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0
3417      &        .or. wcorr6.gt.0.0d0 .or. wturn3.gt.0.0d0 
3418      &        .or. wturn4.gt.0.0d0 .or. wturn6.gt.0.0d0) THEN
3419 C
3420 C 9/25/99 Mixed third-order local-electrostatic terms. The local-interaction 
3421 C   energy of a peptide unit is assumed in the form of a second-order 
3422 C   Fourier series in the angles lambda1 and lambda2 (see Nishikawa et al.
3423 C   Macromolecules, 1974, 7, 797-806 for definition). This correlation terms
3424 C   are computed for EVERY pair of non-contiguous peptide groups.
3425 C
3426           if (j.lt.nres-1) then
3427             j1=j+1
3428             j2=j-1
3429           else
3430             j1=j-1
3431             j2=j-2
3432           endif
3433           kkk=0
3434           do k=1,2
3435             do l=1,2
3436               kkk=kkk+1
3437               muij(kkk)=mu(k,i)*mu(l,j)
3438             enddo
3439           enddo  
3440 cd         write (iout,*) 'EELEC: i',i,' j',j
3441 cd          write (iout,*) 'j',j,' j1',j1,' j2',j2
3442 cd          write(iout,*) 'muij',muij
3443           ury=scalar(uy(1,i),erij)
3444           urz=scalar(uz(1,i),erij)
3445           vry=scalar(uy(1,j),erij)
3446           vrz=scalar(uz(1,j),erij)
3447           a22=scalar(uy(1,i),uy(1,j))-3*ury*vry
3448           a23=scalar(uy(1,i),uz(1,j))-3*ury*vrz
3449           a32=scalar(uz(1,i),uy(1,j))-3*urz*vry
3450           a33=scalar(uz(1,i),uz(1,j))-3*urz*vrz
3451           fac=dsqrt(-ael6i)*r3ij
3452           a22=a22*fac
3453           a23=a23*fac
3454           a32=a32*fac
3455           a33=a33*fac
3456 cd          write (iout,'(4i5,4f10.5)')
3457 cd     &     i,itortyp(itype(i)),j,itortyp(itype(j)),a22,a23,a32,a33
3458 cd          write (iout,'(6f10.5)') (muij(k),k=1,4),fac,eel_loc_ij
3459 cd          write (iout,'(2(3f10.5,5x)/2(3f10.5,5x))') uy(:,i),uz(:,i),
3460 cd     &      uy(:,j),uz(:,j)
3461 cd          write (iout,'(4f10.5)') 
3462 cd     &      scalar(uy(1,i),uy(1,j)),scalar(uy(1,i),uz(1,j)),
3463 cd     &      scalar(uz(1,i),uy(1,j)),scalar(uz(1,i),uz(1,j))
3464 cd          write (iout,'(4f10.5)') ury,urz,vry,vrz
3465 cd           write (iout,'(9f10.5/)') 
3466 cd     &      fac22,a22,fac23,a23,fac32,a32,fac33,a33,eel_loc_ij
3467 C Derivatives of the elements of A in virtual-bond vectors
3468           call unormderiv(erij(1),unmat(1,1),rmij,erder(1,1))
3469           do k=1,3
3470             uryg(k,1)=scalar(erder(1,k),uy(1,i))
3471             uryg(k,2)=scalar(uygrad(1,k,1,i),erij(1))
3472             uryg(k,3)=scalar(uygrad(1,k,2,i),erij(1))
3473             urzg(k,1)=scalar(erder(1,k),uz(1,i))
3474             urzg(k,2)=scalar(uzgrad(1,k,1,i),erij(1))
3475             urzg(k,3)=scalar(uzgrad(1,k,2,i),erij(1))
3476             vryg(k,1)=scalar(erder(1,k),uy(1,j))
3477             vryg(k,2)=scalar(uygrad(1,k,1,j),erij(1))
3478             vryg(k,3)=scalar(uygrad(1,k,2,j),erij(1))
3479             vrzg(k,1)=scalar(erder(1,k),uz(1,j))
3480             vrzg(k,2)=scalar(uzgrad(1,k,1,j),erij(1))
3481             vrzg(k,3)=scalar(uzgrad(1,k,2,j),erij(1))
3482           enddo
3483 C Compute radial contributions to the gradient
3484           facr=-3.0d0*rrmij
3485           a22der=a22*facr
3486           a23der=a23*facr
3487           a32der=a32*facr
3488           a33der=a33*facr
3489           agg(1,1)=a22der*xj
3490           agg(2,1)=a22der*yj
3491           agg(3,1)=a22der*zj
3492           agg(1,2)=a23der*xj
3493           agg(2,2)=a23der*yj
3494           agg(3,2)=a23der*zj
3495           agg(1,3)=a32der*xj
3496           agg(2,3)=a32der*yj
3497           agg(3,3)=a32der*zj
3498           agg(1,4)=a33der*xj
3499           agg(2,4)=a33der*yj
3500           agg(3,4)=a33der*zj
3501 C Add the contributions coming from er
3502           fac3=-3.0d0*fac
3503           do k=1,3
3504             agg(k,1)=agg(k,1)+fac3*(uryg(k,1)*vry+vryg(k,1)*ury)
3505             agg(k,2)=agg(k,2)+fac3*(uryg(k,1)*vrz+vrzg(k,1)*ury)
3506             agg(k,3)=agg(k,3)+fac3*(urzg(k,1)*vry+vryg(k,1)*urz)
3507             agg(k,4)=agg(k,4)+fac3*(urzg(k,1)*vrz+vrzg(k,1)*urz)
3508           enddo
3509           do k=1,3
3510 C Derivatives in DC(i) 
3511 cgrad            ghalf1=0.5d0*agg(k,1)
3512 cgrad            ghalf2=0.5d0*agg(k,2)
3513 cgrad            ghalf3=0.5d0*agg(k,3)
3514 cgrad            ghalf4=0.5d0*agg(k,4)
3515             aggi(k,1)=fac*(scalar(uygrad(1,k,1,i),uy(1,j))
3516      &      -3.0d0*uryg(k,2)*vry)!+ghalf1
3517             aggi(k,2)=fac*(scalar(uygrad(1,k,1,i),uz(1,j))
3518      &      -3.0d0*uryg(k,2)*vrz)!+ghalf2
3519             aggi(k,3)=fac*(scalar(uzgrad(1,k,1,i),uy(1,j))
3520      &      -3.0d0*urzg(k,2)*vry)!+ghalf3
3521             aggi(k,4)=fac*(scalar(uzgrad(1,k,1,i),uz(1,j))
3522      &      -3.0d0*urzg(k,2)*vrz)!+ghalf4
3523 C Derivatives in DC(i+1)
3524             aggi1(k,1)=fac*(scalar(uygrad(1,k,2,i),uy(1,j))
3525      &      -3.0d0*uryg(k,3)*vry)!+agg(k,1)
3526             aggi1(k,2)=fac*(scalar(uygrad(1,k,2,i),uz(1,j))
3527      &      -3.0d0*uryg(k,3)*vrz)!+agg(k,2)
3528             aggi1(k,3)=fac*(scalar(uzgrad(1,k,2,i),uy(1,j))
3529      &      -3.0d0*urzg(k,3)*vry)!+agg(k,3)
3530             aggi1(k,4)=fac*(scalar(uzgrad(1,k,2,i),uz(1,j))
3531      &      -3.0d0*urzg(k,3)*vrz)!+agg(k,4)
3532 C Derivatives in DC(j)
3533             aggj(k,1)=fac*(scalar(uygrad(1,k,1,j),uy(1,i))
3534      &      -3.0d0*vryg(k,2)*ury)!+ghalf1
3535             aggj(k,2)=fac*(scalar(uzgrad(1,k,1,j),uy(1,i))
3536      &      -3.0d0*vrzg(k,2)*ury)!+ghalf2
3537             aggj(k,3)=fac*(scalar(uygrad(1,k,1,j),uz(1,i))
3538      &      -3.0d0*vryg(k,2)*urz)!+ghalf3
3539             aggj(k,4)=fac*(scalar(uzgrad(1,k,1,j),uz(1,i)) 
3540      &      -3.0d0*vrzg(k,2)*urz)!+ghalf4
3541 C Derivatives in DC(j+1) or DC(nres-1)
3542             aggj1(k,1)=fac*(scalar(uygrad(1,k,2,j),uy(1,i))
3543      &      -3.0d0*vryg(k,3)*ury)
3544             aggj1(k,2)=fac*(scalar(uzgrad(1,k,2,j),uy(1,i))
3545      &      -3.0d0*vrzg(k,3)*ury)
3546             aggj1(k,3)=fac*(scalar(uygrad(1,k,2,j),uz(1,i))
3547      &      -3.0d0*vryg(k,3)*urz)
3548             aggj1(k,4)=fac*(scalar(uzgrad(1,k,2,j),uz(1,i)) 
3549      &      -3.0d0*vrzg(k,3)*urz)
3550 cgrad            if (j.eq.nres-1 .and. i.lt.j-2) then
3551 cgrad              do l=1,4
3552 cgrad                aggj1(k,l)=aggj1(k,l)+agg(k,l)
3553 cgrad              enddo
3554 cgrad            endif
3555           enddo
3556           acipa(1,1)=a22
3557           acipa(1,2)=a23
3558           acipa(2,1)=a32
3559           acipa(2,2)=a33
3560           a22=-a22
3561           a23=-a23
3562           do l=1,2
3563             do k=1,3
3564               agg(k,l)=-agg(k,l)
3565               aggi(k,l)=-aggi(k,l)
3566               aggi1(k,l)=-aggi1(k,l)
3567               aggj(k,l)=-aggj(k,l)
3568               aggj1(k,l)=-aggj1(k,l)
3569             enddo
3570           enddo
3571           if (j.lt.nres-1) then
3572             a22=-a22
3573             a32=-a32
3574             do l=1,3,2
3575               do k=1,3
3576                 agg(k,l)=-agg(k,l)
3577                 aggi(k,l)=-aggi(k,l)
3578                 aggi1(k,l)=-aggi1(k,l)
3579                 aggj(k,l)=-aggj(k,l)
3580                 aggj1(k,l)=-aggj1(k,l)
3581               enddo
3582             enddo
3583           else
3584             a22=-a22
3585             a23=-a23
3586             a32=-a32
3587             a33=-a33
3588             do l=1,4
3589               do k=1,3
3590                 agg(k,l)=-agg(k,l)
3591                 aggi(k,l)=-aggi(k,l)
3592                 aggi1(k,l)=-aggi1(k,l)
3593                 aggj(k,l)=-aggj(k,l)
3594                 aggj1(k,l)=-aggj1(k,l)
3595               enddo
3596             enddo 
3597           endif    
3598           ENDIF ! WCORR
3599           IF (wel_loc.gt.0.0d0) THEN
3600 C Contribution to the local-electrostatic energy coming from the i-j pair
3601           eel_loc_ij=a22*muij(1)+a23*muij(2)+a32*muij(3)
3602      &     +a33*muij(4)
3603 cd          write (iout,*) 'i',i,' j',j,' eel_loc_ij',eel_loc_ij
3604
3605           if (energy_dec) write (iout,'(a6,2i5,0pf7.3)')
3606      &            'eelloc',i,j,eel_loc_ij
3607
3608           eel_loc=eel_loc+eel_loc_ij
3609 C Partial derivatives in virtual-bond dihedral angles gamma
3610           if (i.gt.1)
3611      &    gel_loc_loc(i-1)=gel_loc_loc(i-1)+ 
3612      &            a22*muder(1,i)*mu(1,j)+a23*muder(1,i)*mu(2,j)
3613      &           +a32*muder(2,i)*mu(1,j)+a33*muder(2,i)*mu(2,j)
3614           gel_loc_loc(j-1)=gel_loc_loc(j-1)+ 
3615      &            a22*mu(1,i)*muder(1,j)+a23*mu(1,i)*muder(2,j)
3616      &           +a32*mu(2,i)*muder(1,j)+a33*mu(2,i)*muder(2,j)
3617 C Derivatives of eello in DC(i+1) thru DC(j-1) or DC(nres-2)
3618           do l=1,3
3619             ggg(l)=agg(l,1)*muij(1)+
3620      &          agg(l,2)*muij(2)+agg(l,3)*muij(3)+agg(l,4)*muij(4)
3621             gel_loc_long(l,j)=gel_loc_long(l,j)+ggg(l)
3622             gel_loc_long(l,i)=gel_loc_long(l,i)-ggg(l)
3623 cgrad            ghalf=0.5d0*ggg(l)
3624 cgrad            gel_loc(l,i)=gel_loc(l,i)+ghalf
3625 cgrad            gel_loc(l,j)=gel_loc(l,j)+ghalf
3626           enddo
3627 cgrad          do k=i+1,j2
3628 cgrad            do l=1,3
3629 cgrad              gel_loc(l,k)=gel_loc(l,k)+ggg(l)
3630 cgrad            enddo
3631 cgrad          enddo
3632 C Remaining derivatives of eello
3633           do l=1,3
3634             gel_loc(l,i)=gel_loc(l,i)+aggi(l,1)*muij(1)+
3635      &          aggi(l,2)*muij(2)+aggi(l,3)*muij(3)+aggi(l,4)*muij(4)
3636             gel_loc(l,i+1)=gel_loc(l,i+1)+aggi1(l,1)*muij(1)+
3637      &          aggi1(l,2)*muij(2)+aggi1(l,3)*muij(3)+aggi1(l,4)*muij(4)
3638             gel_loc(l,j)=gel_loc(l,j)+aggj(l,1)*muij(1)+
3639      &          aggj(l,2)*muij(2)+aggj(l,3)*muij(3)+aggj(l,4)*muij(4)
3640             gel_loc(l,j1)=gel_loc(l,j1)+aggj1(l,1)*muij(1)+
3641      &          aggj1(l,2)*muij(2)+aggj1(l,3)*muij(3)+aggj1(l,4)*muij(4)
3642           enddo
3643           ENDIF
3644 C Change 12/26/95 to calculate four-body contributions to H-bonding energy
3645 c          if (j.gt.i+1 .and. num_conti.le.maxconts) then
3646           if (wcorr+wcorr4+wcorr5+wcorr6.gt.0.0d0
3647      &       .and. num_conti.le.maxconts) then
3648 c            write (iout,*) i,j," entered corr"
3649 C
3650 C Calculate the contact function. The ith column of the array JCONT will 
3651 C contain the numbers of atoms that make contacts with the atom I (of numbers
3652 C greater than I). The arrays FACONT and GACONT will contain the values of
3653 C the contact function and its derivative.
3654 c           r0ij=1.02D0*rpp(iteli,itelj)
3655 c           r0ij=1.11D0*rpp(iteli,itelj)
3656             r0ij=2.20D0*rpp(iteli,itelj)
3657 c           r0ij=1.55D0*rpp(iteli,itelj)
3658             call gcont(rij,r0ij,1.0D0,0.2d0*r0ij,fcont,fprimcont)
3659             if (fcont.gt.0.0D0) then
3660               num_conti=num_conti+1
3661               if (num_conti.gt.maxconts) then
3662                 write (iout,*) 'WARNING - max. # of contacts exceeded;',
3663      &                         ' will skip next contacts for this conf.'
3664               else
3665                 jcont_hb(num_conti,i)=j
3666 cd                write (iout,*) "i",i," j",j," num_conti",num_conti,
3667 cd     &           " jcont_hb",jcont_hb(num_conti,i)
3668                 IF (wcorr4.gt.0.0d0 .or. wcorr5.gt.0.0d0 .or. 
3669      &          wcorr6.gt.0.0d0 .or. wturn6.gt.0.0d0) THEN
3670 C 9/30/99 (AL) - store components necessary to evaluate higher-order loc-el
3671 C  terms.
3672                 d_cont(num_conti,i)=rij
3673 cd                write (2,'(3e15.5)') rij,r0ij+0.2d0*r0ij,rij
3674 C     --- Electrostatic-interaction matrix --- 
3675                 a_chuj(1,1,num_conti,i)=a22
3676                 a_chuj(1,2,num_conti,i)=a23
3677                 a_chuj(2,1,num_conti,i)=a32
3678                 a_chuj(2,2,num_conti,i)=a33
3679 C     --- Gradient of rij
3680                 do kkk=1,3
3681                   grij_hb_cont(kkk,num_conti,i)=erij(kkk)
3682                 enddo
3683                 kkll=0
3684                 do k=1,2
3685                   do l=1,2
3686                     kkll=kkll+1
3687                     do m=1,3
3688                       a_chuj_der(k,l,m,1,num_conti,i)=agg(m,kkll)
3689                       a_chuj_der(k,l,m,2,num_conti,i)=aggi(m,kkll)
3690                       a_chuj_der(k,l,m,3,num_conti,i)=aggi1(m,kkll)
3691                       a_chuj_der(k,l,m,4,num_conti,i)=aggj(m,kkll)
3692                       a_chuj_der(k,l,m,5,num_conti,i)=aggj1(m,kkll)
3693                     enddo
3694                   enddo
3695                 enddo
3696                 ENDIF
3697                 IF (wcorr4.eq.0.0d0 .and. wcorr.gt.0.0d0) THEN
3698 C Calculate contact energies
3699                 cosa4=4.0D0*cosa
3700                 wij=cosa-3.0D0*cosb*cosg
3701                 cosbg1=cosb+cosg
3702                 cosbg2=cosb-cosg
3703 c               fac3=dsqrt(-ael6i)/r0ij**3     
3704                 fac3=dsqrt(-ael6i)*r3ij
3705 c                 ees0pij=dsqrt(4.0D0+cosa4+wij*wij-3.0D0*cosbg1*cosbg1)
3706                 ees0tmp=4.0D0+cosa4+wij*wij-3.0D0*cosbg1*cosbg1
3707                 if (ees0tmp.gt.0) then
3708                   ees0pij=dsqrt(ees0tmp)
3709                 else
3710                   ees0pij=0
3711                 endif
3712 c                ees0mij=dsqrt(4.0D0-cosa4+wij*wij-3.0D0*cosbg2*cosbg2)
3713                 ees0tmp=4.0D0-cosa4+wij*wij-3.0D0*cosbg2*cosbg2
3714                 if (ees0tmp.gt.0) then
3715                   ees0mij=dsqrt(ees0tmp)
3716                 else
3717                   ees0mij=0
3718                 endif
3719 c               ees0mij=0.0D0
3720                 ees0p(num_conti,i)=0.5D0*fac3*(ees0pij+ees0mij)
3721                 ees0m(num_conti,i)=0.5D0*fac3*(ees0pij-ees0mij)
3722 C Diagnostics. Comment out or remove after debugging!
3723 c               ees0p(num_conti,i)=0.5D0*fac3*ees0pij
3724 c               ees0m(num_conti,i)=0.5D0*fac3*ees0mij
3725 c               ees0m(num_conti,i)=0.0D0
3726 C End diagnostics.
3727 c               write (iout,*) 'i=',i,' j=',j,' rij=',rij,' r0ij=',r0ij,
3728 c    & ' ees0ij=',ees0p(num_conti,i),ees0m(num_conti,i),' fcont=',fcont
3729 C Angular derivatives of the contact function
3730                 ees0pij1=fac3/ees0pij 
3731                 ees0mij1=fac3/ees0mij
3732                 fac3p=-3.0D0*fac3*rrmij
3733                 ees0pijp=0.5D0*fac3p*(ees0pij+ees0mij)
3734                 ees0mijp=0.5D0*fac3p*(ees0pij-ees0mij)
3735 c               ees0mij1=0.0D0
3736                 ecosa1=       ees0pij1*( 1.0D0+0.5D0*wij)
3737                 ecosb1=-1.5D0*ees0pij1*(wij*cosg+cosbg1)
3738                 ecosg1=-1.5D0*ees0pij1*(wij*cosb+cosbg1)
3739                 ecosa2=       ees0mij1*(-1.0D0+0.5D0*wij)
3740                 ecosb2=-1.5D0*ees0mij1*(wij*cosg+cosbg2) 
3741                 ecosg2=-1.5D0*ees0mij1*(wij*cosb-cosbg2)
3742                 ecosap=ecosa1+ecosa2
3743                 ecosbp=ecosb1+ecosb2
3744                 ecosgp=ecosg1+ecosg2
3745                 ecosam=ecosa1-ecosa2
3746                 ecosbm=ecosb1-ecosb2
3747                 ecosgm=ecosg1-ecosg2
3748 C Diagnostics
3749 c               ecosap=ecosa1
3750 c               ecosbp=ecosb1
3751 c               ecosgp=ecosg1
3752 c               ecosam=0.0D0
3753 c               ecosbm=0.0D0
3754 c               ecosgm=0.0D0
3755 C End diagnostics
3756                 facont_hb(num_conti,i)=fcont
3757                 fprimcont=fprimcont/rij
3758 cd              facont_hb(num_conti,i)=1.0D0
3759 C Following line is for diagnostics.
3760 cd              fprimcont=0.0D0
3761                 do k=1,3
3762                   dcosb(k)=rmij*(dc_norm(k,i)-erij(k)*cosb)
3763                   dcosg(k)=rmij*(dc_norm(k,j)-erij(k)*cosg)
3764                 enddo
3765                 do k=1,3
3766                   gggp(k)=ecosbp*dcosb(k)+ecosgp*dcosg(k)
3767                   gggm(k)=ecosbm*dcosb(k)+ecosgm*dcosg(k)
3768                 enddo
3769                 gggp(1)=gggp(1)+ees0pijp*xj
3770                 gggp(2)=gggp(2)+ees0pijp*yj
3771                 gggp(3)=gggp(3)+ees0pijp*zj
3772                 gggm(1)=gggm(1)+ees0mijp*xj
3773                 gggm(2)=gggm(2)+ees0mijp*yj
3774                 gggm(3)=gggm(3)+ees0mijp*zj
3775 C Derivatives due to the contact function
3776                 gacont_hbr(1,num_conti,i)=fprimcont*xj
3777                 gacont_hbr(2,num_conti,i)=fprimcont*yj
3778                 gacont_hbr(3,num_conti,i)=fprimcont*zj
3779                 do k=1,3
3780 c
3781 c 10/24/08 cgrad and ! comments indicate the parts of the code removed 
3782 c          following the change of gradient-summation algorithm.
3783 c
3784 cgrad                  ghalfp=0.5D0*gggp(k)
3785 cgrad                  ghalfm=0.5D0*gggm(k)
3786                   gacontp_hb1(k,num_conti,i)=!ghalfp
3787      &              +(ecosap*(dc_norm(k,j)-cosa*dc_norm(k,i))
3788      &              + ecosbp*(erij(k)-cosb*dc_norm(k,i)))*vbld_inv(i+1)
3789                   gacontp_hb2(k,num_conti,i)=!ghalfp
3790      &              +(ecosap*(dc_norm(k,i)-cosa*dc_norm(k,j))
3791      &              + ecosgp*(erij(k)-cosg*dc_norm(k,j)))*vbld_inv(j+1)
3792                   gacontp_hb3(k,num_conti,i)=gggp(k)
3793                   gacontm_hb1(k,num_conti,i)=!ghalfm
3794      &              +(ecosam*(dc_norm(k,j)-cosa*dc_norm(k,i))
3795      &              + ecosbm*(erij(k)-cosb*dc_norm(k,i)))*vbld_inv(i+1)
3796                   gacontm_hb2(k,num_conti,i)=!ghalfm
3797      &              +(ecosam*(dc_norm(k,i)-cosa*dc_norm(k,j))
3798      &              + ecosgm*(erij(k)-cosg*dc_norm(k,j)))*vbld_inv(j+1)
3799                   gacontm_hb3(k,num_conti,i)=gggm(k)
3800                 enddo
3801 C Diagnostics. Comment out or remove after debugging!
3802 cdiag           do k=1,3
3803 cdiag             gacontp_hb1(k,num_conti,i)=0.0D0
3804 cdiag             gacontp_hb2(k,num_conti,i)=0.0D0
3805 cdiag             gacontp_hb3(k,num_conti,i)=0.0D0
3806 cdiag             gacontm_hb1(k,num_conti,i)=0.0D0
3807 cdiag             gacontm_hb2(k,num_conti,i)=0.0D0
3808 cdiag             gacontm_hb3(k,num_conti,i)=0.0D0
3809 cdiag           enddo
3810               ENDIF ! wcorr
3811               endif  ! num_conti.le.maxconts
3812             endif  ! fcont.gt.0
3813           endif    ! j.gt.i+1
3814           if (wturn3.gt.0.0d0 .or. wturn4.gt.0.0d0) then
3815             do k=1,4
3816               do l=1,3
3817                 ghalf=0.5d0*agg(l,k)
3818                 aggi(l,k)=aggi(l,k)+ghalf
3819                 aggi1(l,k)=aggi1(l,k)+agg(l,k)
3820                 aggj(l,k)=aggj(l,k)+ghalf
3821               enddo
3822             enddo
3823             if (j.eq.nres-1 .and. i.lt.j-2) then
3824               do k=1,4
3825                 do l=1,3
3826                   aggj1(l,k)=aggj1(l,k)+agg(l,k)
3827                 enddo
3828               enddo
3829             endif
3830           endif
3831 c          t_eelecij=t_eelecij+MPI_Wtime()-time00
3832       return
3833       end
3834 C-----------------------------------------------------------------------------
3835       subroutine eturn3(i,eello_turn3)
3836 C Third- and fourth-order contributions from turns
3837       implicit real*8 (a-h,o-z)
3838       include 'DIMENSIONS'
3839       include 'COMMON.IOUNITS'
3840       include 'COMMON.GEO'
3841       include 'COMMON.VAR'
3842       include 'COMMON.LOCAL'
3843       include 'COMMON.CHAIN'
3844       include 'COMMON.DERIV'
3845       include 'COMMON.INTERACT'
3846       include 'COMMON.CONTACTS'
3847       include 'COMMON.TORSION'
3848       include 'COMMON.VECTORS'
3849       include 'COMMON.FFIELD'
3850       include 'COMMON.CONTROL'
3851       dimension ggg(3)
3852       double precision auxmat(2,2),auxmat1(2,2),auxmat2(2,2),pizda(2,2),
3853      &  e1t(2,2),e2t(2,2),e3t(2,2),e1tder(2,2),e2tder(2,2),e3tder(2,2),
3854      &  e1a(2,2),ae3(2,2),ae3e2(2,2),auxvec(2),auxvec1(2)
3855       double precision agg(3,4),aggi(3,4),aggi1(3,4),
3856      &    aggj(3,4),aggj1(3,4),a_temp(2,2),auxmat3(2,2)
3857       common /locel/ a_temp,agg,aggi,aggi1,aggj,aggj1,a22,a23,a32,a33,
3858      &    dxi,dyi,dzi,dx_normi,dy_normi,dz_normi,xmedi,ymedi,zmedi,
3859      &    num_conti,j1,j2
3860       j=i+2
3861 c      write (iout,*) "eturn3",i,j,j1,j2
3862       a_temp(1,1)=a22
3863       a_temp(1,2)=a23
3864       a_temp(2,1)=a32
3865       a_temp(2,2)=a33
3866 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
3867 C
3868 C               Third-order contributions
3869 C        
3870 C                 (i+2)o----(i+3)
3871 C                      | |
3872 C                      | |
3873 C                 (i+1)o----i
3874 C
3875 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC   
3876 cd        call checkint_turn3(i,a_temp,eello_turn3_num)
3877         call matmat2(EUg(1,1,i+1),EUg(1,1,i+2),auxmat(1,1))
3878         call transpose2(auxmat(1,1),auxmat1(1,1))
3879         call matmat2(a_temp(1,1),auxmat1(1,1),pizda(1,1))
3880         eello_turn3=eello_turn3+0.5d0*(pizda(1,1)+pizda(2,2))
3881         if (energy_dec) write (iout,'(a6,2i5,0pf7.3)')
3882      &          'eturn3',i,j,0.5d0*(pizda(1,1)+pizda(2,2))
3883 cd        write (2,*) 'i,',i,' j',j,'eello_turn3',
3884 cd     &    0.5d0*(pizda(1,1)+pizda(2,2)),
3885 cd     &    ' eello_turn3_num',4*eello_turn3_num
3886 C Derivatives in gamma(i)
3887         call matmat2(EUgder(1,1,i+1),EUg(1,1,i+2),auxmat2(1,1))
3888         call transpose2(auxmat2(1,1),auxmat3(1,1))
3889         call matmat2(a_temp(1,1),auxmat3(1,1),pizda(1,1))
3890         gel_loc_turn3(i)=gel_loc_turn3(i)+0.5d0*(pizda(1,1)+pizda(2,2))
3891 C Derivatives in gamma(i+1)
3892         call matmat2(EUg(1,1,i+1),EUgder(1,1,i+2),auxmat2(1,1))
3893         call transpose2(auxmat2(1,1),auxmat3(1,1))
3894         call matmat2(a_temp(1,1),auxmat3(1,1),pizda(1,1))
3895         gel_loc_turn3(i+1)=gel_loc_turn3(i+1)
3896      &    +0.5d0*(pizda(1,1)+pizda(2,2))
3897 C Cartesian derivatives
3898 !DIR$ UNROLL(0)
3899         do l=1,3
3900 c            ghalf1=0.5d0*agg(l,1)
3901 c            ghalf2=0.5d0*agg(l,2)
3902 c            ghalf3=0.5d0*agg(l,3)
3903 c            ghalf4=0.5d0*agg(l,4)
3904           a_temp(1,1)=aggi(l,1)!+ghalf1
3905           a_temp(1,2)=aggi(l,2)!+ghalf2
3906           a_temp(2,1)=aggi(l,3)!+ghalf3
3907           a_temp(2,2)=aggi(l,4)!+ghalf4
3908           call matmat2(a_temp(1,1),auxmat1(1,1),pizda(1,1))
3909           gcorr3_turn(l,i)=gcorr3_turn(l,i)
3910      &      +0.5d0*(pizda(1,1)+pizda(2,2))
3911           a_temp(1,1)=aggi1(l,1)!+agg(l,1)
3912           a_temp(1,2)=aggi1(l,2)!+agg(l,2)
3913           a_temp(2,1)=aggi1(l,3)!+agg(l,3)
3914           a_temp(2,2)=aggi1(l,4)!+agg(l,4)
3915           call matmat2(a_temp(1,1),auxmat1(1,1),pizda(1,1))
3916           gcorr3_turn(l,i+1)=gcorr3_turn(l,i+1)
3917      &      +0.5d0*(pizda(1,1)+pizda(2,2))
3918           a_temp(1,1)=aggj(l,1)!+ghalf1
3919           a_temp(1,2)=aggj(l,2)!+ghalf2
3920           a_temp(2,1)=aggj(l,3)!+ghalf3
3921           a_temp(2,2)=aggj(l,4)!+ghalf4
3922           call matmat2(a_temp(1,1),auxmat1(1,1),pizda(1,1))
3923           gcorr3_turn(l,j)=gcorr3_turn(l,j)
3924      &      +0.5d0*(pizda(1,1)+pizda(2,2))
3925           a_temp(1,1)=aggj1(l,1)
3926           a_temp(1,2)=aggj1(l,2)
3927           a_temp(2,1)=aggj1(l,3)
3928           a_temp(2,2)=aggj1(l,4)
3929           call matmat2(a_temp(1,1),auxmat1(1,1),pizda(1,1))
3930           gcorr3_turn(l,j1)=gcorr3_turn(l,j1)
3931      &      +0.5d0*(pizda(1,1)+pizda(2,2))
3932         enddo
3933       return
3934       end
3935 C-------------------------------------------------------------------------------
3936       subroutine eturn4(i,eello_turn4)
3937 C Third- and fourth-order contributions from turns
3938       implicit real*8 (a-h,o-z)
3939       include 'DIMENSIONS'
3940       include 'COMMON.IOUNITS'
3941       include 'COMMON.GEO'
3942       include 'COMMON.VAR'
3943       include 'COMMON.LOCAL'
3944       include 'COMMON.CHAIN'
3945       include 'COMMON.DERIV'
3946       include 'COMMON.INTERACT'
3947       include 'COMMON.CONTACTS'
3948       include 'COMMON.TORSION'
3949       include 'COMMON.VECTORS'
3950       include 'COMMON.FFIELD'
3951       include 'COMMON.CONTROL'
3952       dimension ggg(3)
3953       double precision auxmat(2,2),auxmat1(2,2),auxmat2(2,2),pizda(2,2),
3954      &  e1t(2,2),e2t(2,2),e3t(2,2),e1tder(2,2),e2tder(2,2),e3tder(2,2),
3955      &  e1a(2,2),ae3(2,2),ae3e2(2,2),auxvec(2),auxvec1(2)
3956       double precision agg(3,4),aggi(3,4),aggi1(3,4),
3957      &    aggj(3,4),aggj1(3,4),a_temp(2,2),auxmat3(2,2)
3958       common /locel/ a_temp,agg,aggi,aggi1,aggj,aggj1,a22,a23,a32,a33,
3959      &    dxi,dyi,dzi,dx_normi,dy_normi,dz_normi,xmedi,ymedi,zmedi,
3960      &    num_conti,j1,j2
3961       j=i+3
3962 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
3963 C
3964 C               Fourth-order contributions
3965 C        
3966 C                 (i+3)o----(i+4)
3967 C                     /  |
3968 C               (i+2)o   |
3969 C                     \  |
3970 C                 (i+1)o----i
3971 C
3972 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC   
3973 cd        call checkint_turn4(i,a_temp,eello_turn4_num)
3974 c        write (iout,*) "eturn4 i",i," j",j," j1",j1," j2",j2
3975         a_temp(1,1)=a22
3976         a_temp(1,2)=a23
3977         a_temp(2,1)=a32
3978         a_temp(2,2)=a33
3979         iti1=itortyp(itype(i+1))
3980         iti2=itortyp(itype(i+2))
3981         iti3=itortyp(itype(i+3))
3982 c        write(iout,*) "iti1",iti1," iti2",iti2," iti3",iti3
3983         call transpose2(EUg(1,1,i+1),e1t(1,1))
3984         call transpose2(Eug(1,1,i+2),e2t(1,1))
3985         call transpose2(Eug(1,1,i+3),e3t(1,1))
3986         call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
3987         call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
3988         s1=scalar2(b1(1,iti2),auxvec(1))
3989         call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
3990         call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
3991         s2=scalar2(b1(1,iti1),auxvec(1))
3992         call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
3993         call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
3994         s3=0.5d0*(pizda(1,1)+pizda(2,2))
3995         eello_turn4=eello_turn4-(s1+s2+s3)
3996         if (energy_dec) write (iout,'(a6,2i5,0pf7.3)')
3997      &      'eturn4',i,j,-(s1+s2+s3)
3998 cd        write (2,*) 'i,',i,' j',j,'eello_turn4',-(s1+s2+s3),
3999 cd     &    ' eello_turn4_num',8*eello_turn4_num
4000 C Derivatives in gamma(i)
4001         call transpose2(EUgder(1,1,i+1),e1tder(1,1))
4002         call matmat2(e1tder(1,1),a_temp(1,1),auxmat(1,1))
4003         call matvec2(auxmat(1,1),Ub2(1,i+3),auxvec(1))
4004         s1=scalar2(b1(1,iti2),auxvec(1))
4005         call matmat2(ae3e2(1,1),e1tder(1,1),pizda(1,1))
4006         s3=0.5d0*(pizda(1,1)+pizda(2,2))
4007         gel_loc_turn4(i)=gel_loc_turn4(i)-(s1+s3)
4008 C Derivatives in gamma(i+1)
4009         call transpose2(EUgder(1,1,i+2),e2tder(1,1))
4010         call matvec2(ae3(1,1),Ub2der(1,i+2),auxvec(1)) 
4011         s2=scalar2(b1(1,iti1),auxvec(1))
4012         call matmat2(ae3(1,1),e2tder(1,1),auxmat(1,1))
4013         call matmat2(auxmat(1,1),e1t(1,1),pizda(1,1))
4014         s3=0.5d0*(pizda(1,1)+pizda(2,2))
4015         gel_loc_turn4(i+1)=gel_loc_turn4(i+1)-(s2+s3)
4016 C Derivatives in gamma(i+2)
4017         call transpose2(EUgder(1,1,i+3),e3tder(1,1))
4018         call matvec2(e1a(1,1),Ub2der(1,i+3),auxvec(1))
4019         s1=scalar2(b1(1,iti2),auxvec(1))
4020         call matmat2(a_temp(1,1),e3tder(1,1),auxmat(1,1))
4021         call matvec2(auxmat(1,1),Ub2(1,i+2),auxvec(1)) 
4022         s2=scalar2(b1(1,iti1),auxvec(1))
4023         call matmat2(auxmat(1,1),e2t(1,1),auxmat3(1,1))
4024         call matmat2(auxmat3(1,1),e1t(1,1),pizda(1,1))
4025         s3=0.5d0*(pizda(1,1)+pizda(2,2))
4026         gel_loc_turn4(i+2)=gel_loc_turn4(i+2)-(s1+s2+s3)
4027 C Cartesian derivatives
4028 C Derivatives of this turn contributions in DC(i+2)
4029         if (j.lt.nres-1) then
4030           do l=1,3
4031             a_temp(1,1)=agg(l,1)
4032             a_temp(1,2)=agg(l,2)
4033             a_temp(2,1)=agg(l,3)
4034             a_temp(2,2)=agg(l,4)
4035             call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
4036             call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
4037             s1=scalar2(b1(1,iti2),auxvec(1))
4038             call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
4039             call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
4040             s2=scalar2(b1(1,iti1),auxvec(1))
4041             call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
4042             call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
4043             s3=0.5d0*(pizda(1,1)+pizda(2,2))
4044             ggg(l)=-(s1+s2+s3)
4045             gcorr4_turn(l,i+2)=gcorr4_turn(l,i+2)-(s1+s2+s3)
4046           enddo
4047         endif
4048 C Remaining derivatives of this turn contribution
4049         do l=1,3
4050           a_temp(1,1)=aggi(l,1)
4051           a_temp(1,2)=aggi(l,2)
4052           a_temp(2,1)=aggi(l,3)
4053           a_temp(2,2)=aggi(l,4)
4054           call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
4055           call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
4056           s1=scalar2(b1(1,iti2),auxvec(1))
4057           call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
4058           call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
4059           s2=scalar2(b1(1,iti1),auxvec(1))
4060           call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
4061           call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
4062           s3=0.5d0*(pizda(1,1)+pizda(2,2))
4063           gcorr4_turn(l,i)=gcorr4_turn(l,i)-(s1+s2+s3)
4064           a_temp(1,1)=aggi1(l,1)
4065           a_temp(1,2)=aggi1(l,2)
4066           a_temp(2,1)=aggi1(l,3)
4067           a_temp(2,2)=aggi1(l,4)
4068           call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
4069           call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
4070           s1=scalar2(b1(1,iti2),auxvec(1))
4071           call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
4072           call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
4073           s2=scalar2(b1(1,iti1),auxvec(1))
4074           call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
4075           call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
4076           s3=0.5d0*(pizda(1,1)+pizda(2,2))
4077           gcorr4_turn(l,i+1)=gcorr4_turn(l,i+1)-(s1+s2+s3)
4078           a_temp(1,1)=aggj(l,1)
4079           a_temp(1,2)=aggj(l,2)
4080           a_temp(2,1)=aggj(l,3)
4081           a_temp(2,2)=aggj(l,4)
4082           call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
4083           call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
4084           s1=scalar2(b1(1,iti2),auxvec(1))
4085           call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
4086           call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
4087           s2=scalar2(b1(1,iti1),auxvec(1))
4088           call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
4089           call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
4090           s3=0.5d0*(pizda(1,1)+pizda(2,2))
4091           gcorr4_turn(l,j)=gcorr4_turn(l,j)-(s1+s2+s3)
4092           a_temp(1,1)=aggj1(l,1)
4093           a_temp(1,2)=aggj1(l,2)
4094           a_temp(2,1)=aggj1(l,3)
4095           a_temp(2,2)=aggj1(l,4)
4096           call matmat2(e1t(1,1),a_temp(1,1),e1a(1,1))
4097           call matvec2(e1a(1,1),Ub2(1,i+3),auxvec(1))
4098           s1=scalar2(b1(1,iti2),auxvec(1))
4099           call matmat2(a_temp(1,1),e3t(1,1),ae3(1,1))
4100           call matvec2(ae3(1,1),Ub2(1,i+2),auxvec(1)) 
4101           s2=scalar2(b1(1,iti1),auxvec(1))
4102           call matmat2(ae3(1,1),e2t(1,1),ae3e2(1,1))
4103           call matmat2(ae3e2(1,1),e1t(1,1),pizda(1,1))
4104           s3=0.5d0*(pizda(1,1)+pizda(2,2))
4105 c          write (iout,*) "s1",s1," s2",s2," s3",s3," s1+s2+s3",s1+s2+s3
4106           gcorr4_turn(l,j1)=gcorr4_turn(l,j1)-(s1+s2+s3)
4107         enddo
4108       return
4109       end
4110 C-----------------------------------------------------------------------------
4111       subroutine vecpr(u,v,w)
4112       implicit real*8(a-h,o-z)
4113       dimension u(3),v(3),w(3)
4114       w(1)=u(2)*v(3)-u(3)*v(2)
4115       w(2)=-u(1)*v(3)+u(3)*v(1)
4116       w(3)=u(1)*v(2)-u(2)*v(1)
4117       return
4118       end
4119 C-----------------------------------------------------------------------------
4120       subroutine unormderiv(u,ugrad,unorm,ungrad)
4121 C This subroutine computes the derivatives of a normalized vector u, given
4122 C the derivatives computed without normalization conditions, ugrad. Returns
4123 C ungrad.
4124       implicit none
4125       double precision u(3),ugrad(3,3),unorm,ungrad(3,3)
4126       double precision vec(3)
4127       double precision scalar
4128       integer i,j
4129 c      write (2,*) 'ugrad',ugrad
4130 c      write (2,*) 'u',u
4131       do i=1,3
4132         vec(i)=scalar(ugrad(1,i),u(1))
4133       enddo
4134 c      write (2,*) 'vec',vec
4135       do i=1,3
4136         do j=1,3
4137           ungrad(j,i)=(ugrad(j,i)-u(j)*vec(i))*unorm
4138         enddo
4139       enddo
4140 c      write (2,*) 'ungrad',ungrad
4141       return
4142       end
4143 C-----------------------------------------------------------------------------
4144       subroutine escp_soft_sphere(evdw2,evdw2_14)
4145 C
4146 C This subroutine calculates the excluded-volume interaction energy between
4147 C peptide-group centers and side chains and its gradient in virtual-bond and
4148 C side-chain vectors.
4149 C
4150       implicit real*8 (a-h,o-z)
4151       include 'DIMENSIONS'
4152       include 'COMMON.GEO'
4153       include 'COMMON.VAR'
4154       include 'COMMON.LOCAL'
4155       include 'COMMON.CHAIN'
4156       include 'COMMON.DERIV'
4157       include 'COMMON.INTERACT'
4158       include 'COMMON.FFIELD'
4159       include 'COMMON.IOUNITS'
4160       include 'COMMON.CONTROL'
4161       dimension ggg(3)
4162       evdw2=0.0D0
4163       evdw2_14=0.0d0
4164       r0_scp=4.5d0
4165 cd    print '(a)','Enter ESCP'
4166 cd    write (iout,*) 'iatscp_s=',iatscp_s,' iatscp_e=',iatscp_e
4167       do i=iatscp_s,iatscp_e
4168         iteli=itel(i)
4169         xi=0.5D0*(c(1,i)+c(1,i+1))
4170         yi=0.5D0*(c(2,i)+c(2,i+1))
4171         zi=0.5D0*(c(3,i)+c(3,i+1))
4172
4173         do iint=1,nscp_gr(i)
4174
4175         do j=iscpstart(i,iint),iscpend(i,iint)
4176           itypj=itype(j)
4177 C Uncomment following three lines for SC-p interactions
4178 c         xj=c(1,nres+j)-xi
4179 c         yj=c(2,nres+j)-yi
4180 c         zj=c(3,nres+j)-zi
4181 C Uncomment following three lines for Ca-p interactions
4182           xj=c(1,j)-xi
4183           yj=c(2,j)-yi
4184           zj=c(3,j)-zi
4185           rij=xj*xj+yj*yj+zj*zj
4186           r0ij=r0_scp
4187           r0ijsq=r0ij*r0ij
4188           if (rij.lt.r0ijsq) then
4189             evdwij=0.25d0*(rij-r0ijsq)**2
4190             fac=rij-r0ijsq
4191           else
4192             evdwij=0.0d0
4193             fac=0.0d0
4194           endif 
4195           evdw2=evdw2+evdwij
4196 C
4197 C Calculate contributions to the gradient in the virtual-bond and SC vectors.
4198 C
4199           ggg(1)=xj*fac
4200           ggg(2)=yj*fac
4201           ggg(3)=zj*fac
4202 cgrad          if (j.lt.i) then
4203 cd          write (iout,*) 'j<i'
4204 C Uncomment following three lines for SC-p interactions
4205 c           do k=1,3
4206 c             gradx_scp(k,j)=gradx_scp(k,j)+ggg(k)
4207 c           enddo
4208 cgrad          else
4209 cd          write (iout,*) 'j>i'
4210 cgrad            do k=1,3
4211 cgrad              ggg(k)=-ggg(k)
4212 C Uncomment following line for SC-p interactions
4213 c             gradx_scp(k,j)=gradx_scp(k,j)-ggg(k)
4214 cgrad            enddo
4215 cgrad          endif
4216 cgrad          do k=1,3
4217 cgrad            gvdwc_scp(k,i)=gvdwc_scp(k,i)-0.5D0*ggg(k)
4218 cgrad          enddo
4219 cgrad          kstart=min0(i+1,j)
4220 cgrad          kend=max0(i-1,j-1)
4221 cd        write (iout,*) 'i=',i,' j=',j,' kstart=',kstart,' kend=',kend
4222 cd        write (iout,*) ggg(1),ggg(2),ggg(3)
4223 cgrad          do k=kstart,kend
4224 cgrad            do l=1,3
4225 cgrad              gvdwc_scp(l,k)=gvdwc_scp(l,k)-ggg(l)
4226 cgrad            enddo
4227 cgrad          enddo
4228           do k=1,3
4229             gvdwc_scpp(k,i)=gvdwc_scpp(k,i)-ggg(k)
4230             gvdwc_scp(k,j)=gvdwc_scp(k,j)+ggg(k)
4231           enddo
4232         enddo
4233
4234         enddo ! iint
4235       enddo ! i
4236       return
4237       end
4238 C-----------------------------------------------------------------------------
4239       subroutine escp(evdw2,evdw2_14)
4240 C
4241 C This subroutine calculates the excluded-volume interaction energy between
4242 C peptide-group centers and side chains and its gradient in virtual-bond and
4243 C side-chain vectors.
4244 C
4245       implicit real*8 (a-h,o-z)
4246       include 'DIMENSIONS'
4247       include 'COMMON.GEO'
4248       include 'COMMON.VAR'
4249       include 'COMMON.LOCAL'
4250       include 'COMMON.CHAIN'
4251       include 'COMMON.DERIV'
4252       include 'COMMON.INTERACT'
4253       include 'COMMON.FFIELD'
4254       include 'COMMON.IOUNITS'
4255       include 'COMMON.CONTROL'
4256       dimension ggg(3)
4257       evdw2=0.0D0
4258       evdw2_14=0.0d0
4259 cd    print '(a)','Enter ESCP'
4260 cd    write (iout,*) 'iatscp_s=',iatscp_s,' iatscp_e=',iatscp_e
4261       do i=iatscp_s,iatscp_e
4262         iteli=itel(i)
4263         xi=0.5D0*(c(1,i)+c(1,i+1))
4264         yi=0.5D0*(c(2,i)+c(2,i+1))
4265         zi=0.5D0*(c(3,i)+c(3,i+1))
4266
4267         do iint=1,nscp_gr(i)
4268
4269         do j=iscpstart(i,iint),iscpend(i,iint)
4270           itypj=itype(j)
4271 C Uncomment following three lines for SC-p interactions
4272 c         xj=c(1,nres+j)-xi
4273 c         yj=c(2,nres+j)-yi
4274 c         zj=c(3,nres+j)-zi
4275 C Uncomment following three lines for Ca-p interactions
4276           xj=c(1,j)-xi
4277           yj=c(2,j)-yi
4278           zj=c(3,j)-zi
4279           rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
4280           fac=rrij**expon2
4281           e1=fac*fac*aad(itypj,iteli)
4282           e2=fac*bad(itypj,iteli)
4283           if (iabs(j-i) .le. 2) then
4284             e1=scal14*e1
4285             e2=scal14*e2
4286             evdw2_14=evdw2_14+e1+e2
4287           endif
4288           evdwij=e1+e2
4289           evdw2=evdw2+evdwij
4290           if (energy_dec) write (iout,'(a6,2i5,0pf7.3)')
4291      &        'evdw2',i,j,evdwij
4292 C
4293 C Calculate contributions to the gradient in the virtual-bond and SC vectors.
4294 C
4295           fac=-(evdwij+e1)*rrij
4296           ggg(1)=xj*fac
4297           ggg(2)=yj*fac
4298           ggg(3)=zj*fac
4299 cgrad          if (j.lt.i) then
4300 cd          write (iout,*) 'j<i'
4301 C Uncomment following three lines for SC-p interactions
4302 c           do k=1,3
4303 c             gradx_scp(k,j)=gradx_scp(k,j)+ggg(k)
4304 c           enddo
4305 cgrad          else
4306 cd          write (iout,*) 'j>i'
4307 cgrad            do k=1,3
4308 cgrad              ggg(k)=-ggg(k)
4309 C Uncomment following line for SC-p interactions
4310 ccgrad             gradx_scp(k,j)=gradx_scp(k,j)-ggg(k)
4311 c             gradx_scp(k,j)=gradx_scp(k,j)+ggg(k)
4312 cgrad            enddo
4313 cgrad          endif
4314 cgrad          do k=1,3
4315 cgrad            gvdwc_scp(k,i)=gvdwc_scp(k,i)-0.5D0*ggg(k)
4316 cgrad          enddo
4317 cgrad          kstart=min0(i+1,j)
4318 cgrad          kend=max0(i-1,j-1)
4319 cd        write (iout,*) 'i=',i,' j=',j,' kstart=',kstart,' kend=',kend
4320 cd        write (iout,*) ggg(1),ggg(2),ggg(3)
4321 cgrad          do k=kstart,kend
4322 cgrad            do l=1,3
4323 cgrad              gvdwc_scp(l,k)=gvdwc_scp(l,k)-ggg(l)
4324 cgrad            enddo
4325 cgrad          enddo
4326           do k=1,3
4327             gvdwc_scpp(k,i)=gvdwc_scpp(k,i)-ggg(k)
4328             gvdwc_scp(k,j)=gvdwc_scp(k,j)+ggg(k)
4329           enddo
4330         enddo
4331
4332         enddo ! iint
4333       enddo ! i
4334       do i=1,nct
4335         do j=1,3
4336           gvdwc_scp(j,i)=expon*gvdwc_scp(j,i)
4337           gvdwc_scpp(j,i)=expon*gvdwc_scpp(j,i)
4338           gradx_scp(j,i)=expon*gradx_scp(j,i)
4339         enddo
4340       enddo
4341 C******************************************************************************
4342 C
4343 C                              N O T E !!!
4344 C
4345 C To save time the factor EXPON has been extracted from ALL components
4346 C of GVDWC and GRADX. Remember to multiply them by this factor before further 
4347 C use!
4348 C
4349 C******************************************************************************
4350       return
4351       end
4352 C--------------------------------------------------------------------------
4353       subroutine edis(ehpb)
4354
4355 C Evaluate bridge-strain energy and its gradient in virtual-bond and SC vectors.
4356 C
4357       implicit real*8 (a-h,o-z)
4358       include 'DIMENSIONS'
4359       include 'COMMON.SBRIDGE'
4360       include 'COMMON.CHAIN'
4361       include 'COMMON.DERIV'
4362       include 'COMMON.VAR'
4363       include 'COMMON.INTERACT'
4364       include 'COMMON.IOUNITS'
4365       include 'COMMON.CONTROL'
4366       dimension ggg(3)
4367       ehpb=0.0D0
4368 cd      write(iout,*)'edis: nhpb=',nhpb,' fbr=',fbr
4369 cd      write(iout,*)'link_start=',link_start,' link_end=',link_end
4370       if (link_end.eq.0) return
4371       do i=link_start,link_end
4372 C If ihpb(i) and jhpb(i) > NRES, this is a SC-SC distance, otherwise a
4373 C CA-CA distance used in regularization of structure.
4374         ii=ihpb(i)
4375         jj=jhpb(i)
4376 C iii and jjj point to the residues for which the distance is assigned.
4377         if (ii.gt.nres) then
4378           iii=ii-nres
4379           jjj=jj-nres 
4380         else
4381           iii=ii
4382           jjj=jj
4383         endif
4384 c        write (iout,*) "i",i," ii",ii," iii",iii," jj",jj," jjj",jjj,
4385 c     &    dhpb(i),dhpb1(i),forcon(i)
4386 C 24/11/03 AL: SS bridges handled separately because of introducing a specific
4387 C    distance and angle dependent SS bond potential.
4388 cmc        if (ii.gt.nres .and. itype(iii).eq.1 .and. itype(jjj).eq.1) then
4389 C 18/07/06 MC: Use the convention that the first nss pairs are SS bonds
4390         if (.not.dyn_ss .and. i.le.nss) then
4391 C 15/02/13 CC dynamic SSbond - additional check
4392          if (ii.gt.nres 
4393      &       .and. itype(iii).eq.1 .and. itype(jjj).eq.1) then 
4394           call ssbond_ene(iii,jjj,eij)
4395           ehpb=ehpb+2*eij
4396          endif
4397 cd          write (iout,*) "eij",eij
4398         else if (ii.gt.nres .and. jj.gt.nres) then
4399 c Restraints from contact prediction
4400           dd=dist(ii,jj)
4401           if (constr_dist.eq.11) then
4402             ehpb=ehpb+fordepth(i)**4.0d0
4403      &          *rlornmr1(dd,dhpb(i),dhpb1(i),forcon(i))
4404             fac=fordepth(i)**4.0d0
4405      &          *rlornmr1prim(dd,dhpb(i),dhpb1(i),forcon(i))/dd
4406           if (energy_dec) write (iout,'(a6,2i5,f15.6,2f8.3)') 
4407      &     "edisl",ii,jj,
4408      &     fordepth(i)**4.0d0*rlornmr1(dd,dhpb(i),dhpb1(i),forcon(i)),
4409      &     fordepth(i),dd
4410            else
4411           if (dhpb1(i).gt.0.0d0) then
4412             ehpb=ehpb+2*forcon(i)*gnmr1(dd,dhpb(i),dhpb1(i))
4413             fac=forcon(i)*gnmr1prim(dd,dhpb(i),dhpb1(i))/dd
4414 c            write (iout,*) "beta nmr",
4415 c     &        dd,2*forcon(i)*gnmr1(dd,dhpb(i),dhpb1(i))
4416           else
4417             dd=dist(ii,jj)
4418             rdis=dd-dhpb(i)
4419 C Get the force constant corresponding to this distance.
4420             waga=forcon(i)
4421 C Calculate the contribution to energy.
4422             ehpb=ehpb+waga*rdis*rdis
4423 c            write (iout,*) "beta reg",dd,waga*rdis*rdis
4424 C
4425 C Evaluate gradient.
4426 C
4427             fac=waga*rdis/dd
4428           endif
4429           endif
4430           do j=1,3
4431             ggg(j)=fac*(c(j,jj)-c(j,ii))
4432           enddo
4433           do j=1,3
4434             ghpbx(j,iii)=ghpbx(j,iii)-ggg(j)
4435             ghpbx(j,jjj)=ghpbx(j,jjj)+ggg(j)
4436           enddo
4437           do k=1,3
4438             ghpbc(k,jjj)=ghpbc(k,jjj)+ggg(k)
4439             ghpbc(k,iii)=ghpbc(k,iii)-ggg(k)
4440           enddo
4441         else
4442 C Calculate the distance between the two points and its difference from the
4443 C target distance.
4444           dd=dist(ii,jj)
4445          if (constr_dist.eq.11) then
4446             ehpb=ehpb+fordepth(i)**4.0d0
4447      &           *rlornmr1(dd,dhpb(i),dhpb1(i),forcon(i))
4448             fac=fordepth(i)**4.0d0
4449      &           *rlornmr1prim(dd,dhpb(i),dhpb1(i),forcon(i))/dd
4450           if (energy_dec) write (iout,'(a6,2i5,f15.6,2f8.3)') 
4451      7     "edisl",ii,jj,
4452      &     fordepth(i)**4.0d0*rlornmr1(dd,dhpb(i),dhpb1(i),forcon(i)),
4453      &     fordepth(i),dd
4454 c          if (energy_dec)
4455 c     &      write (iout,*) fac
4456          else   
4457           if (dhpb1(i).gt.0.0d0) then
4458             ehpb=ehpb+2*forcon(i)*gnmr1(dd,dhpb(i),dhpb1(i))
4459             fac=forcon(i)*gnmr1prim(dd,dhpb(i),dhpb1(i))/dd
4460 c            write (iout,*) "alph nmr",
4461 c     &        dd,2*forcon(i)*gnmr1(dd,dhpb(i),dhpb1(i))
4462           else
4463             rdis=dd-dhpb(i)
4464 C Get the force constant corresponding to this distance.
4465             waga=forcon(i)
4466 C Calculate the contribution to energy.
4467             ehpb=ehpb+waga*rdis*rdis
4468 c            write (iout,*) "alpha reg",dd,waga*rdis*rdis
4469 C
4470 C Evaluate gradient.
4471 C
4472             fac=waga*rdis/dd
4473           endif
4474          endif
4475 cd      print *,'i=',i,' ii=',ii,' jj=',jj,' dhpb=',dhpb(i),' dd=',dd,
4476 cd   &   ' waga=',waga,' fac=',fac
4477             do j=1,3
4478               ggg(j)=fac*(c(j,jj)-c(j,ii))
4479             enddo
4480 cd      print '(i3,3(1pe14.5))',i,(ggg(j),j=1,3)
4481 C If this is a SC-SC distance, we need to calculate the contributions to the
4482 C Cartesian gradient in the SC vectors (ghpbx).
4483           if (iii.lt.ii) then
4484           do j=1,3
4485             ghpbx(j,iii)=ghpbx(j,iii)-ggg(j)
4486             ghpbx(j,jjj)=ghpbx(j,jjj)+ggg(j)
4487           enddo
4488           endif
4489 cgrad        do j=iii,jjj-1
4490 cgrad          do k=1,3
4491 cgrad            ghpbc(k,j)=ghpbc(k,j)+ggg(k)
4492 cgrad          enddo
4493 cgrad        enddo
4494           do k=1,3
4495             ghpbc(k,jjj)=ghpbc(k,jjj)+ggg(k)
4496             ghpbc(k,iii)=ghpbc(k,iii)-ggg(k)
4497           enddo
4498         endif
4499       enddo
4500       if (constr_dist.ne.11) ehpb=0.5D0*ehpb
4501 c      do i=1,nres
4502 c        write (iout,*) "ghpbc",i,(ghpbc(j,i),j=1,3)
4503 c      enddo
4504       return
4505       end
4506 C--------------------------------------------------------------------------
4507       subroutine ssbond_ene(i,j,eij)
4508
4509 C Calculate the distance and angle dependent SS-bond potential energy
4510 C using a free-energy function derived based on RHF/6-31G** ab initio
4511 C calculations of diethyl disulfide.
4512 C
4513 C A. Liwo and U. Kozlowska, 11/24/03
4514 C
4515       implicit real*8 (a-h,o-z)
4516       include 'DIMENSIONS'
4517       include 'COMMON.SBRIDGE'
4518       include 'COMMON.CHAIN'
4519       include 'COMMON.DERIV'
4520       include 'COMMON.LOCAL'
4521       include 'COMMON.INTERACT'
4522       include 'COMMON.VAR'
4523       include 'COMMON.IOUNITS'
4524       double precision erij(3),dcosom1(3),dcosom2(3),gg(3)
4525       itypi=itype(i)
4526       xi=c(1,nres+i)
4527       yi=c(2,nres+i)
4528       zi=c(3,nres+i)
4529       dxi=dc_norm(1,nres+i)
4530       dyi=dc_norm(2,nres+i)
4531       dzi=dc_norm(3,nres+i)
4532 c      dsci_inv=dsc_inv(itypi)
4533       dsci_inv=vbld_inv(nres+i)
4534       itypj=itype(j)
4535 c      dscj_inv=dsc_inv(itypj)
4536       dscj_inv=vbld_inv(nres+j)
4537       xj=c(1,nres+j)-xi
4538       yj=c(2,nres+j)-yi
4539       zj=c(3,nres+j)-zi
4540       dxj=dc_norm(1,nres+j)
4541       dyj=dc_norm(2,nres+j)
4542       dzj=dc_norm(3,nres+j)
4543       rrij=1.0D0/(xj*xj+yj*yj+zj*zj)
4544       rij=dsqrt(rrij)
4545       erij(1)=xj*rij
4546       erij(2)=yj*rij
4547       erij(3)=zj*rij
4548       om1=dxi*erij(1)+dyi*erij(2)+dzi*erij(3)
4549       om2=dxj*erij(1)+dyj*erij(2)+dzj*erij(3)
4550       om12=dxi*dxj+dyi*dyj+dzi*dzj
4551       do k=1,3
4552         dcosom1(k)=rij*(dc_norm(k,nres+i)-om1*erij(k))
4553         dcosom2(k)=rij*(dc_norm(k,nres+j)-om2*erij(k))
4554       enddo
4555       rij=1.0d0/rij
4556       deltad=rij-d0cm
4557       deltat1=1.0d0-om1
4558       deltat2=1.0d0+om2
4559       deltat12=om2-om1+2.0d0
4560       cosphi=om12-om1*om2
4561       eij=akcm*deltad*deltad+akth*(deltat1*deltat1+deltat2*deltat2)
4562      &  +akct*deltad*deltat12+ebr
4563      &  +v1ss*cosphi+v2ss*cosphi*cosphi+v3ss*cosphi*cosphi*cosphi
4564 c      write(iout,*) i,j,"rij",rij,"d0cm",d0cm," akcm",akcm," akth",akth,
4565 c     &  " akct",akct," deltad",deltad," deltat",deltat1,deltat2,
4566 c     &  " deltat12",deltat12," eij",eij 
4567       ed=2*akcm*deltad+akct*deltat12
4568       pom1=akct*deltad
4569       pom2=v1ss+2*v2ss*cosphi+3*v3ss*cosphi*cosphi
4570       eom1=-2*akth*deltat1-pom1-om2*pom2
4571       eom2= 2*akth*deltat2+pom1-om1*pom2
4572       eom12=pom2
4573       do k=1,3
4574         ggk=ed*erij(k)+eom1*dcosom1(k)+eom2*dcosom2(k)
4575         ghpbx(k,i)=ghpbx(k,i)-ggk
4576      &            +(eom12*(dc_norm(k,nres+j)-om12*dc_norm(k,nres+i))
4577      &            +eom1*(erij(k)-om1*dc_norm(k,nres+i)))*dsci_inv
4578         ghpbx(k,j)=ghpbx(k,j)+ggk
4579      &            +(eom12*(dc_norm(k,nres+i)-om12*dc_norm(k,nres+j))
4580      &            +eom2*(erij(k)-om2*dc_norm(k,nres+j)))*dscj_inv
4581         ghpbc(k,i)=ghpbc(k,i)-ggk
4582         ghpbc(k,j)=ghpbc(k,j)+ggk
4583       enddo
4584 C
4585 C Calculate the components of the gradient in DC and X
4586 C
4587 cgrad      do k=i,j-1
4588 cgrad        do l=1,3
4589 cgrad          ghpbc(l,k)=ghpbc(l,k)+gg(l)
4590 cgrad        enddo
4591 cgrad      enddo
4592       return
4593       end
4594 C--------------------------------------------------------------------------
4595       subroutine ebond(estr)
4596 c
4597 c Evaluate the energy of stretching of the CA-CA and CA-SC virtual bonds
4598 c
4599       implicit real*8 (a-h,o-z)
4600       include 'DIMENSIONS'
4601       include 'COMMON.LOCAL'
4602       include 'COMMON.GEO'
4603       include 'COMMON.INTERACT'
4604       include 'COMMON.DERIV'
4605       include 'COMMON.VAR'
4606       include 'COMMON.CHAIN'
4607       include 'COMMON.IOUNITS'
4608       include 'COMMON.NAMES'
4609       include 'COMMON.FFIELD'
4610       include 'COMMON.CONTROL'
4611       include 'COMMON.SETUP'
4612       double precision u(3),ud(3)
4613       estr=0.0d0
4614       do i=ibondp_start,ibondp_end
4615         diff = vbld(i)-vbldp0
4616 c        write (iout,*) i,vbld(i),vbldp0,diff,AKP*diff*diff
4617         if (energy_dec)    write (iout,'(a7,i5,4f7.3)') 
4618      &     "estr bb",i,vbld(i),vbldp0,diff,AKP*diff*diff
4619         estr=estr+diff*diff
4620         do j=1,3
4621           gradb(j,i-1)=AKP*diff*dc(j,i-1)/vbld(i)
4622         enddo
4623 c        write (iout,'(i5,3f10.5)') i,(gradb(j,i-1),j=1,3)
4624       enddo
4625       estr=0.5d0*AKP*estr
4626 c
4627 c 09/18/07 AL: multimodal bond potential based on AM1 CA-SC PMF's included
4628 c
4629       do i=ibond_start,ibond_end
4630         iti=itype(i)
4631         if (iti.ne.10) then
4632           nbi=nbondterm(iti)
4633           if (nbi.eq.1) then
4634             diff=vbld(i+nres)-vbldsc0(1,iti)
4635 c            write (iout,*) i,iti,vbld(i+nres),vbldsc0(1,iti),diff,
4636 c     &      AKSC(1,iti),AKSC(1,iti)*diff*diff
4637             if (energy_dec)  then
4638               write (iout,*) 
4639      &         "estr sc",i,iti,vbld(i+nres),vbldsc0(1,iti),diff,
4640      &         AKSC(1,iti),AKSC(1,iti)*diff*diff
4641               call flush(iout)
4642             endif
4643             estr=estr+0.5d0*AKSC(1,iti)*diff*diff
4644             do j=1,3
4645               gradbx(j,i)=AKSC(1,iti)*diff*dc(j,i+nres)/vbld(i+nres)
4646             enddo
4647           else
4648             do j=1,nbi
4649               diff=vbld(i+nres)-vbldsc0(j,iti) 
4650               ud(j)=aksc(j,iti)*diff
4651               u(j)=abond0(j,iti)+0.5d0*ud(j)*diff
4652             enddo
4653             uprod=u(1)
4654             do j=2,nbi
4655               uprod=uprod*u(j)
4656             enddo
4657             usum=0.0d0
4658             usumsqder=0.0d0
4659             do j=1,nbi
4660               uprod1=1.0d0
4661               uprod2=1.0d0
4662               do k=1,nbi
4663                 if (k.ne.j) then
4664                   uprod1=uprod1*u(k)
4665                   uprod2=uprod2*u(k)*u(k)
4666                 endif
4667               enddo
4668               usum=usum+uprod1
4669               usumsqder=usumsqder+ud(j)*uprod2   
4670             enddo
4671             estr=estr+uprod/usum
4672             do j=1,3
4673              gradbx(j,i)=usumsqder/(usum*usum)*dc(j,i+nres)/vbld(i+nres)
4674             enddo
4675           endif
4676         endif
4677       enddo
4678       return
4679       end 
4680 #ifdef CRYST_THETA
4681 C--------------------------------------------------------------------------
4682       subroutine ebend(etheta)
4683 C
4684 C Evaluate the virtual-bond-angle energy given the virtual-bond dihedral
4685 C angles gamma and its derivatives in consecutive thetas and gammas.
4686 C
4687       implicit real*8 (a-h,o-z)
4688       include 'DIMENSIONS'
4689       include 'COMMON.LOCAL'
4690       include 'COMMON.GEO'
4691       include 'COMMON.INTERACT'
4692       include 'COMMON.DERIV'
4693       include 'COMMON.VAR'
4694       include 'COMMON.CHAIN'
4695       include 'COMMON.IOUNITS'
4696       include 'COMMON.NAMES'
4697       include 'COMMON.FFIELD'
4698       include 'COMMON.CONTROL'
4699       common /calcthet/ term1,term2,termm,diffak,ratak,
4700      & ak,aktc,termpre,termexp,sigc,sig0i,time11,time12,sigcsq,
4701      & delthe0,sig0inv,sigtc,sigsqtc,delthec,it
4702       double precision y(2),z(2)
4703       delta=0.02d0*pi
4704 c      time11=dexp(-2*time)
4705 c      time12=1.0d0
4706       etheta=0.0D0
4707 c     write (*,'(a,i2)') 'EBEND ICG=',icg
4708       do i=ithet_start,ithet_end
4709 C Zero the energy function and its derivative at 0 or pi.
4710         call splinthet(theta(i),0.5d0*delta,ss,ssd)
4711         it=itype(i-1)
4712         if (i.gt.3) then
4713 #ifdef OSF
4714           phii=phi(i)
4715           if (phii.ne.phii) phii=150.0
4716 #else
4717           phii=phi(i)
4718 #endif
4719           y(1)=dcos(phii)
4720           y(2)=dsin(phii)
4721         else 
4722           y(1)=0.0D0
4723           y(2)=0.0D0
4724         endif
4725         if (i.lt.nres) then
4726 #ifdef OSF
4727           phii1=phi(i+1)
4728           if (phii1.ne.phii1) phii1=150.0
4729           phii1=pinorm(phii1)
4730           z(1)=cos(phii1)
4731 #else
4732           phii1=phi(i+1)
4733           z(1)=dcos(phii1)
4734 #endif
4735           z(2)=dsin(phii1)
4736         else
4737           z(1)=0.0D0
4738           z(2)=0.0D0
4739         endif  
4740 C Calculate the "mean" value of theta from the part of the distribution
4741 C dependent on the adjacent virtual-bond-valence angles (gamma1 & gamma2).
4742 C In following comments this theta will be referred to as t_c.
4743         thet_pred_mean=0.0d0
4744         do k=1,2
4745           athetk=athet(k,it)
4746           bthetk=bthet(k,it)
4747           thet_pred_mean=thet_pred_mean+athetk*y(k)+bthetk*z(k)
4748         enddo
4749         dthett=thet_pred_mean*ssd
4750         thet_pred_mean=thet_pred_mean*ss+a0thet(it)
4751 C Derivatives of the "mean" values in gamma1 and gamma2.
4752         dthetg1=(-athet(1,it)*y(2)+athet(2,it)*y(1))*ss
4753         dthetg2=(-bthet(1,it)*z(2)+bthet(2,it)*z(1))*ss
4754         if (theta(i).gt.pi-delta) then
4755           call theteng(pi-delta,thet_pred_mean,theta0(it),f0,fprim0,
4756      &         E_tc0)
4757           call mixder(pi-delta,thet_pred_mean,theta0(it),fprim_tc0)
4758           call theteng(pi,thet_pred_mean,theta0(it),f1,fprim1,E_tc1)
4759           call spline1(theta(i),pi-delta,delta,f0,f1,fprim0,ethetai,
4760      &        E_theta)
4761           call spline2(theta(i),pi-delta,delta,E_tc0,E_tc1,fprim_tc0,
4762      &        E_tc)
4763         else if (theta(i).lt.delta) then
4764           call theteng(delta,thet_pred_mean,theta0(it),f0,fprim0,E_tc0)
4765           call theteng(0.0d0,thet_pred_mean,theta0(it),f1,fprim1,E_tc1)
4766           call spline1(theta(i),delta,-delta,f0,f1,fprim0,ethetai,
4767      &        E_theta)
4768           call mixder(delta,thet_pred_mean,theta0(it),fprim_tc0)
4769           call spline2(theta(i),delta,-delta,E_tc0,E_tc1,fprim_tc0,
4770      &        E_tc)
4771         else
4772           call theteng(theta(i),thet_pred_mean,theta0(it),ethetai,
4773      &        E_theta,E_tc)
4774         endif
4775         etheta=etheta+ethetai
4776         if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
4777      &      'ebend',i,ethetai
4778         if (i.gt.3) gloc(i-3,icg)=gloc(i-3,icg)+wang*E_tc*dthetg1
4779         if (i.lt.nres) gloc(i-2,icg)=gloc(i-2,icg)+wang*E_tc*dthetg2
4780         gloc(nphi+i-2,icg)=wang*(E_theta+E_tc*dthett)+gloc(nphi+i-2,icg)
4781       enddo
4782 C Ufff.... We've done all this!!! 
4783       return
4784       end
4785 C---------------------------------------------------------------------------
4786       subroutine theteng(thetai,thet_pred_mean,theta0i,ethetai,E_theta,
4787      &     E_tc)
4788       implicit real*8 (a-h,o-z)
4789       include 'DIMENSIONS'
4790       include 'COMMON.LOCAL'
4791       include 'COMMON.IOUNITS'
4792       common /calcthet/ term1,term2,termm,diffak,ratak,
4793      & ak,aktc,termpre,termexp,sigc,sig0i,time11,time12,sigcsq,
4794      & delthe0,sig0inv,sigtc,sigsqtc,delthec,it
4795 C Calculate the contributions to both Gaussian lobes.
4796 C 6/6/97 - Deform the Gaussians using the factor of 1/(1+time)
4797 C The "polynomial part" of the "standard deviation" of this part of 
4798 C the distribution.
4799         sig=polthet(3,it)
4800         do j=2,0,-1
4801           sig=sig*thet_pred_mean+polthet(j,it)
4802         enddo
4803 C Derivative of the "interior part" of the "standard deviation of the" 
4804 C gamma-dependent Gaussian lobe in t_c.
4805         sigtc=3*polthet(3,it)
4806         do j=2,1,-1
4807           sigtc=sigtc*thet_pred_mean+j*polthet(j,it)
4808         enddo
4809         sigtc=sig*sigtc
4810 C Set the parameters of both Gaussian lobes of the distribution.
4811 C "Standard deviation" of the gamma-dependent Gaussian lobe (sigtc)
4812         fac=sig*sig+sigc0(it)
4813         sigcsq=fac+fac
4814         sigc=1.0D0/sigcsq
4815 C Following variable (sigsqtc) is -(1/2)d[sigma(t_c)**(-2))]/dt_c
4816         sigsqtc=-4.0D0*sigcsq*sigtc
4817 c       print *,i,sig,sigtc,sigsqtc
4818 C Following variable (sigtc) is d[sigma(t_c)]/dt_c
4819         sigtc=-sigtc/(fac*fac)
4820 C Following variable is sigma(t_c)**(-2)
4821         sigcsq=sigcsq*sigcsq
4822         sig0i=sig0(it)
4823         sig0inv=1.0D0/sig0i**2
4824         delthec=thetai-thet_pred_mean
4825         delthe0=thetai-theta0i
4826         term1=-0.5D0*sigcsq*delthec*delthec
4827         term2=-0.5D0*sig0inv*delthe0*delthe0
4828 C Following fuzzy logic is to avoid underflows in dexp and subsequent INFs and
4829 C NaNs in taking the logarithm. We extract the largest exponent which is added
4830 C to the energy (this being the log of the distribution) at the end of energy
4831 C term evaluation for this virtual-bond angle.
4832         if (term1.gt.term2) then
4833           termm=term1
4834           term2=dexp(term2-termm)
4835           term1=1.0d0
4836         else
4837           termm=term2
4838           term1=dexp(term1-termm)
4839           term2=1.0d0
4840         endif
4841 C The ratio between the gamma-independent and gamma-dependent lobes of
4842 C the distribution is a Gaussian function of thet_pred_mean too.
4843         diffak=gthet(2,it)-thet_pred_mean
4844         ratak=diffak/gthet(3,it)**2
4845         ak=dexp(gthet(1,it)-0.5D0*diffak*ratak)
4846 C Let's differentiate it in thet_pred_mean NOW.
4847         aktc=ak*ratak
4848 C Now put together the distribution terms to make complete distribution.
4849         termexp=term1+ak*term2
4850         termpre=sigc+ak*sig0i
4851 C Contribution of the bending energy from this theta is just the -log of
4852 C the sum of the contributions from the two lobes and the pre-exponential
4853 C factor. Simple enough, isn't it?
4854         ethetai=(-dlog(termexp)-termm+dlog(termpre))
4855 C NOW the derivatives!!!
4856 C 6/6/97 Take into account the deformation.
4857         E_theta=(delthec*sigcsq*term1
4858      &       +ak*delthe0*sig0inv*term2)/termexp
4859         E_tc=((sigtc+aktc*sig0i)/termpre
4860      &      -((delthec*sigcsq+delthec*delthec*sigsqtc)*term1+
4861      &       aktc*term2)/termexp)
4862       return
4863       end
4864 c-----------------------------------------------------------------------------
4865       subroutine mixder(thetai,thet_pred_mean,theta0i,E_tc_t)
4866       implicit real*8 (a-h,o-z)
4867       include 'DIMENSIONS'
4868       include 'COMMON.LOCAL'
4869       include 'COMMON.IOUNITS'
4870       common /calcthet/ term1,term2,termm,diffak,ratak,
4871      & ak,aktc,termpre,termexp,sigc,sig0i,time11,time12,sigcsq,
4872      & delthe0,sig0inv,sigtc,sigsqtc,delthec,it
4873       delthec=thetai-thet_pred_mean
4874       delthe0=thetai-theta0i
4875 C "Thank you" to MAPLE (probably spared one day of hand-differentiation).
4876       t3 = thetai-thet_pred_mean
4877       t6 = t3**2
4878       t9 = term1
4879       t12 = t3*sigcsq
4880       t14 = t12+t6*sigsqtc
4881       t16 = 1.0d0
4882       t21 = thetai-theta0i
4883       t23 = t21**2
4884       t26 = term2
4885       t27 = t21*t26
4886       t32 = termexp
4887       t40 = t32**2
4888       E_tc_t = -((sigcsq+2.D0*t3*sigsqtc)*t9-t14*sigcsq*t3*t16*t9
4889      & -aktc*sig0inv*t27)/t32+(t14*t9+aktc*t26)/t40
4890      & *(-t12*t9-ak*sig0inv*t27)
4891       return
4892       end
4893 #else
4894 C--------------------------------------------------------------------------
4895       subroutine ebend(etheta)
4896 C
4897 C Evaluate the virtual-bond-angle energy given the virtual-bond dihedral
4898 C angles gamma and its derivatives in consecutive thetas and gammas.
4899 C ab initio-derived potentials from 
4900 c Kozlowska et al., J. Phys.: Condens. Matter 19 (2007) 285203
4901 C
4902       implicit real*8 (a-h,o-z)
4903       include 'DIMENSIONS'
4904       include 'COMMON.LOCAL'
4905       include 'COMMON.GEO'
4906       include 'COMMON.INTERACT'
4907       include 'COMMON.DERIV'
4908       include 'COMMON.VAR'
4909       include 'COMMON.CHAIN'
4910       include 'COMMON.IOUNITS'
4911       include 'COMMON.NAMES'
4912       include 'COMMON.FFIELD'
4913       include 'COMMON.CONTROL'
4914       double precision coskt(mmaxtheterm),sinkt(mmaxtheterm),
4915      & cosph1(maxsingle),sinph1(maxsingle),cosph2(maxsingle),
4916      & sinph2(maxsingle),cosph1ph2(maxdouble,maxdouble),
4917      & sinph1ph2(maxdouble,maxdouble)
4918       logical lprn /.false./, lprn1 /.false./
4919       etheta=0.0D0
4920 c      write (iout,*) "EBEND ithet_start",ithet_start,
4921 c     &     " ithet_end",ithet_end
4922       do i=ithet_start,ithet_end
4923         if ((itype(i-1).eq.ntyp1).or.(itype(i-2).eq.ntyp1).or.
4924      &(itype(i).eq.ntyp1)) cycle
4925         dethetai=0.0d0
4926         dephii=0.0d0
4927         dephii1=0.0d0
4928         theti2=0.5d0*theta(i)
4929         ityp2=ithetyp(itype(i-1))
4930         do k=1,nntheterm
4931           coskt(k)=dcos(k*theti2)
4932           sinkt(k)=dsin(k*theti2)
4933         enddo
4934 C        if (i.gt.3) then
4935         if (i.gt.3 .and. itype(max0(i-3,1)).ne.ntyp1) then
4936 #ifdef OSF
4937           phii=phi(i)
4938           if (phii.ne.phii) phii=150.0
4939 #else
4940           phii=phi(i)
4941 #endif
4942           ityp1=ithetyp(itype(i-2))
4943           do k=1,nsingle
4944             cosph1(k)=dcos(k*phii)
4945             sinph1(k)=dsin(k*phii)
4946           enddo
4947         else
4948           phii=0.0d0
4949           ityp1=ithetyp(itype(i-2))
4950           do k=1,nsingle
4951             cosph1(k)=0.0d0
4952             sinph1(k)=0.0d0
4953           enddo 
4954         endif
4955         if ((i.lt.nres).and. itype(i+1).ne.ntyp1) then
4956 #ifdef OSF
4957           phii1=phi(i+1)
4958           if (phii1.ne.phii1) phii1=150.0
4959           phii1=pinorm(phii1)
4960 #else
4961           phii1=phi(i+1)
4962 #endif
4963           ityp3=ithetyp(itype(i))
4964           do k=1,nsingle
4965             cosph2(k)=dcos(k*phii1)
4966             sinph2(k)=dsin(k*phii1)
4967           enddo
4968         else
4969           phii1=0.0d0
4970           ityp3=ithetyp(itype(i))
4971           do k=1,nsingle
4972             cosph2(k)=0.0d0
4973             sinph2(k)=0.0d0
4974           enddo
4975         endif  
4976         ethetai=aa0thet(ityp1,ityp2,ityp3)
4977         do k=1,ndouble
4978           do l=1,k-1
4979             ccl=cosph1(l)*cosph2(k-l)
4980             ssl=sinph1(l)*sinph2(k-l)
4981             scl=sinph1(l)*cosph2(k-l)
4982             csl=cosph1(l)*sinph2(k-l)
4983             cosph1ph2(l,k)=ccl-ssl
4984             cosph1ph2(k,l)=ccl+ssl
4985             sinph1ph2(l,k)=scl+csl
4986             sinph1ph2(k,l)=scl-csl
4987           enddo
4988         enddo
4989         if (lprn) then
4990         write (iout,*) "i",i," ityp1",ityp1," ityp2",ityp2,
4991      &    " ityp3",ityp3," theti2",theti2," phii",phii," phii1",phii1
4992         write (iout,*) "coskt and sinkt"
4993         do k=1,nntheterm
4994           write (iout,*) k,coskt(k),sinkt(k)
4995         enddo
4996         endif
4997         do k=1,ntheterm
4998           ethetai=ethetai+aathet(k,ityp1,ityp2,ityp3)*sinkt(k)
4999           dethetai=dethetai+0.5d0*k*aathet(k,ityp1,ityp2,ityp3)
5000      &      *coskt(k)
5001           if (lprn)
5002      &    write (iout,*) "k",k," aathet",aathet(k,ityp1,ityp2,ityp3),
5003      &     " ethetai",ethetai
5004         enddo
5005         if (lprn) then
5006         write (iout,*) "cosph and sinph"
5007         do k=1,nsingle
5008           write (iout,*) k,cosph1(k),sinph1(k),cosph2(k),sinph2(k)
5009         enddo
5010         write (iout,*) "cosph1ph2 and sinph2ph2"
5011         do k=2,ndouble
5012           do l=1,k-1
5013             write (iout,*) l,k,cosph1ph2(l,k),cosph1ph2(k,l),
5014      &         sinph1ph2(l,k),sinph1ph2(k,l) 
5015           enddo
5016         enddo
5017         write(iout,*) "ethetai",ethetai
5018         endif
5019         do m=1,ntheterm2
5020           do k=1,nsingle
5021             aux=bbthet(k,m,ityp1,ityp2,ityp3)*cosph1(k)
5022      &         +ccthet(k,m,ityp1,ityp2,ityp3)*sinph1(k)
5023      &         +ddthet(k,m,ityp1,ityp2,ityp3)*cosph2(k)
5024      &         +eethet(k,m,ityp1,ityp2,ityp3)*sinph2(k)
5025             ethetai=ethetai+sinkt(m)*aux
5026             dethetai=dethetai+0.5d0*m*aux*coskt(m)
5027             dephii=dephii+k*sinkt(m)*(
5028      &          ccthet(k,m,ityp1,ityp2,ityp3)*cosph1(k)-
5029      &          bbthet(k,m,ityp1,ityp2,ityp3)*sinph1(k))
5030             dephii1=dephii1+k*sinkt(m)*(
5031      &          eethet(k,m,ityp1,ityp2,ityp3)*cosph2(k)-
5032      &          ddthet(k,m,ityp1,ityp2,ityp3)*sinph2(k))
5033             if (lprn)
5034      &      write (iout,*) "m",m," k",k," bbthet",
5035      &         bbthet(k,m,ityp1,ityp2,ityp3)," ccthet",
5036      &         ccthet(k,m,ityp1,ityp2,ityp3)," ddthet",
5037      &         ddthet(k,m,ityp1,ityp2,ityp3)," eethet",
5038      &         eethet(k,m,ityp1,ityp2,ityp3)," ethetai",ethetai
5039           enddo
5040         enddo
5041         if (lprn)
5042      &  write(iout,*) "ethetai",ethetai
5043         do m=1,ntheterm3
5044           do k=2,ndouble
5045             do l=1,k-1
5046               aux=ffthet(l,k,m,ityp1,ityp2,ityp3)*cosph1ph2(l,k)+
5047      &            ffthet(k,l,m,ityp1,ityp2,ityp3)*cosph1ph2(k,l)+
5048      &            ggthet(l,k,m,ityp1,ityp2,ityp3)*sinph1ph2(l,k)+
5049      &            ggthet(k,l,m,ityp1,ityp2,ityp3)*sinph1ph2(k,l)
5050               ethetai=ethetai+sinkt(m)*aux
5051               dethetai=dethetai+0.5d0*m*coskt(m)*aux
5052               dephii=dephii+l*sinkt(m)*(
5053      &           -ffthet(l,k,m,ityp1,ityp2,ityp3)*sinph1ph2(l,k)-
5054      &            ffthet(k,l,m,ityp1,ityp2,ityp3)*sinph1ph2(k,l)+
5055      &            ggthet(l,k,m,ityp1,ityp2,ityp3)*cosph1ph2(l,k)+
5056      &            ggthet(k,l,m,ityp1,ityp2,ityp3)*cosph1ph2(k,l))
5057               dephii1=dephii1+(k-l)*sinkt(m)*(
5058      &           -ffthet(l,k,m,ityp1,ityp2,ityp3)*sinph1ph2(l,k)+
5059      &            ffthet(k,l,m,ityp1,ityp2,ityp3)*sinph1ph2(k,l)+
5060      &            ggthet(l,k,m,ityp1,ityp2,ityp3)*cosph1ph2(l,k)-
5061      &            ggthet(k,l,m,ityp1,ityp2,ityp3)*cosph1ph2(k,l))
5062               if (lprn) then
5063               write (iout,*) "m",m," k",k," l",l," ffthet",
5064      &            ffthet(l,k,m,ityp1,ityp2,ityp3),
5065      &            ffthet(k,l,m,ityp1,ityp2,ityp3)," ggthet",
5066      &            ggthet(l,k,m,ityp1,ityp2,ityp3),
5067      &            ggthet(k,l,m,ityp1,ityp2,ityp3)," ethetai",ethetai
5068               write (iout,*) cosph1ph2(l,k)*sinkt(m),
5069      &            cosph1ph2(k,l)*sinkt(m),
5070      &            sinph1ph2(l,k)*sinkt(m),sinph1ph2(k,l)*sinkt(m)
5071               endif
5072             enddo
5073           enddo
5074         enddo
5075 10      continue
5076 c        lprn1=.true.
5077         if (lprn1) write (iout,'(a4,i2,3f8.1,9h ethetai ,f10.5)') 
5078      &  'ebe', i,theta(i)*rad2deg,phii*rad2deg,
5079      &   phii1*rad2deg,ethetai
5080 c        lprn1=.false.
5081         etheta=etheta+ethetai
5082         if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
5083      &      'ebend',i,ethetai
5084         if (i.gt.3) gloc(i-3,icg)=gloc(i-3,icg)+wang*dephii
5085         if (i.lt.nres) gloc(i-2,icg)=gloc(i-2,icg)+wang*dephii1
5086         gloc(nphi+i-2,icg)=gloc(nphi+i-2,icg)+wang*dethetai
5087       enddo
5088       return
5089       end
5090 #endif
5091 #ifdef CRYST_SC
5092 c-----------------------------------------------------------------------------
5093       subroutine esc(escloc)
5094 C Calculate the local energy of a side chain and its derivatives in the
5095 C corresponding virtual-bond valence angles THETA and the spherical angles 
5096 C ALPHA and OMEGA.
5097       implicit real*8 (a-h,o-z)
5098       include 'DIMENSIONS'
5099       include 'COMMON.GEO'
5100       include 'COMMON.LOCAL'
5101       include 'COMMON.VAR'
5102       include 'COMMON.INTERACT'
5103       include 'COMMON.DERIV'
5104       include 'COMMON.CHAIN'
5105       include 'COMMON.IOUNITS'
5106       include 'COMMON.NAMES'
5107       include 'COMMON.FFIELD'
5108       include 'COMMON.CONTROL'
5109       double precision x(3),dersc(3),xemp(3),dersc0(3),dersc1(3),
5110      &     ddersc0(3),ddummy(3),xtemp(3),temp(3)
5111       common /sccalc/ time11,time12,time112,theti,it,nlobit
5112       delta=0.02d0*pi
5113       escloc=0.0D0
5114 c     write (iout,'(a)') 'ESC'
5115       do i=loc_start,loc_end
5116         it=itype(i)
5117         if (it.eq.10) goto 1
5118         nlobit=nlob(it)
5119 c       print *,'i=',i,' it=',it,' nlobit=',nlobit
5120 c       write (iout,*) 'i=',i,' ssa=',ssa,' ssad=',ssad
5121         theti=theta(i+1)-pipol
5122         x(1)=dtan(theti)
5123         x(2)=alph(i)
5124         x(3)=omeg(i)
5125
5126         if (x(2).gt.pi-delta) then
5127           xtemp(1)=x(1)
5128           xtemp(2)=pi-delta
5129           xtemp(3)=x(3)
5130           call enesc(xtemp,escloci0,dersc0,ddersc0,.true.)
5131           xtemp(2)=pi
5132           call enesc(xtemp,escloci1,dersc1,ddummy,.false.)
5133           call spline1(x(2),pi-delta,delta,escloci0,escloci1,dersc0(2),
5134      &        escloci,dersc(2))
5135           call spline2(x(2),pi-delta,delta,dersc0(1),dersc1(1),
5136      &        ddersc0(1),dersc(1))
5137           call spline2(x(2),pi-delta,delta,dersc0(3),dersc1(3),
5138      &        ddersc0(3),dersc(3))
5139           xtemp(2)=pi-delta
5140           call enesc_bound(xtemp,esclocbi0,dersc0,dersc12,.true.)
5141           xtemp(2)=pi
5142           call enesc_bound(xtemp,esclocbi1,dersc1,chuju,.false.)
5143           call spline1(x(2),pi-delta,delta,esclocbi0,esclocbi1,
5144      &            dersc0(2),esclocbi,dersc02)
5145           call spline2(x(2),pi-delta,delta,dersc0(1),dersc1(1),
5146      &            dersc12,dersc01)
5147           call splinthet(x(2),0.5d0*delta,ss,ssd)
5148           dersc0(1)=dersc01
5149           dersc0(2)=dersc02
5150           dersc0(3)=0.0d0
5151           do k=1,3
5152             dersc(k)=ss*dersc(k)+(1.0d0-ss)*dersc0(k)
5153           enddo
5154           dersc(2)=dersc(2)+ssd*(escloci-esclocbi)
5155 c         write (iout,*) 'i=',i,x(2)*rad2deg,escloci0,escloci,
5156 c    &             esclocbi,ss,ssd
5157           escloci=ss*escloci+(1.0d0-ss)*esclocbi
5158 c         escloci=esclocbi
5159 c         write (iout,*) escloci
5160         else if (x(2).lt.delta) then
5161           xtemp(1)=x(1)
5162           xtemp(2)=delta
5163           xtemp(3)=x(3)
5164           call enesc(xtemp,escloci0,dersc0,ddersc0,.true.)
5165           xtemp(2)=0.0d0
5166           call enesc(xtemp,escloci1,dersc1,ddummy,.false.)
5167           call spline1(x(2),delta,-delta,escloci0,escloci1,dersc0(2),
5168      &        escloci,dersc(2))
5169           call spline2(x(2),delta,-delta,dersc0(1),dersc1(1),
5170      &        ddersc0(1),dersc(1))
5171           call spline2(x(2),delta,-delta,dersc0(3),dersc1(3),
5172      &        ddersc0(3),dersc(3))
5173           xtemp(2)=delta
5174           call enesc_bound(xtemp,esclocbi0,dersc0,dersc12,.true.)
5175           xtemp(2)=0.0d0
5176           call enesc_bound(xtemp,esclocbi1,dersc1,chuju,.false.)
5177           call spline1(x(2),delta,-delta,esclocbi0,esclocbi1,
5178      &            dersc0(2),esclocbi,dersc02)
5179           call spline2(x(2),delta,-delta,dersc0(1),dersc1(1),
5180      &            dersc12,dersc01)
5181           dersc0(1)=dersc01
5182           dersc0(2)=dersc02
5183           dersc0(3)=0.0d0
5184           call splinthet(x(2),0.5d0*delta,ss,ssd)
5185           do k=1,3
5186             dersc(k)=ss*dersc(k)+(1.0d0-ss)*dersc0(k)
5187           enddo
5188           dersc(2)=dersc(2)+ssd*(escloci-esclocbi)
5189 c         write (iout,*) 'i=',i,x(2)*rad2deg,escloci0,escloci,
5190 c    &             esclocbi,ss,ssd
5191           escloci=ss*escloci+(1.0d0-ss)*esclocbi
5192 c         write (iout,*) escloci
5193         else
5194           call enesc(x,escloci,dersc,ddummy,.false.)
5195         endif
5196
5197         escloc=escloc+escloci
5198         if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
5199      &     'escloc',i,escloci
5200 c       write (iout,*) 'i=',i,' escloci=',escloci,' dersc=',dersc
5201
5202         gloc(nphi+i-1,icg)=gloc(nphi+i-1,icg)+
5203      &   wscloc*dersc(1)
5204         gloc(ialph(i,1),icg)=wscloc*dersc(2)
5205         gloc(ialph(i,1)+nside,icg)=wscloc*dersc(3)
5206     1   continue
5207       enddo
5208       return
5209       end
5210 C---------------------------------------------------------------------------
5211       subroutine enesc(x,escloci,dersc,ddersc,mixed)
5212       implicit real*8 (a-h,o-z)
5213       include 'DIMENSIONS'
5214       include 'COMMON.GEO'
5215       include 'COMMON.LOCAL'
5216       include 'COMMON.IOUNITS'
5217       common /sccalc/ time11,time12,time112,theti,it,nlobit
5218       double precision x(3),z(3),Ax(3,maxlob,-1:1),dersc(3),ddersc(3)
5219       double precision contr(maxlob,-1:1)
5220       logical mixed
5221 c       write (iout,*) 'it=',it,' nlobit=',nlobit
5222         escloc_i=0.0D0
5223         do j=1,3
5224           dersc(j)=0.0D0
5225           if (mixed) ddersc(j)=0.0d0
5226         enddo
5227         x3=x(3)
5228
5229 C Because of periodicity of the dependence of the SC energy in omega we have
5230 C to add up the contributions from x(3)-2*pi, x(3), and x(3+2*pi).
5231 C To avoid underflows, first compute & store the exponents.
5232
5233         do iii=-1,1
5234
5235           x(3)=x3+iii*dwapi
5236  
5237           do j=1,nlobit
5238             do k=1,3
5239               z(k)=x(k)-censc(k,j,it)
5240             enddo
5241             do k=1,3
5242               Axk=0.0D0
5243               do l=1,3
5244                 Axk=Axk+gaussc(l,k,j,it)*z(l)
5245               enddo
5246               Ax(k,j,iii)=Axk
5247             enddo 
5248             expfac=0.0D0 
5249             do k=1,3
5250               expfac=expfac+Ax(k,j,iii)*z(k)
5251             enddo
5252             contr(j,iii)=expfac
5253           enddo ! j
5254
5255         enddo ! iii
5256
5257         x(3)=x3
5258 C As in the case of ebend, we want to avoid underflows in exponentiation and
5259 C subsequent NaNs and INFs in energy calculation.
5260 C Find the largest exponent
5261         emin=contr(1,-1)
5262         do iii=-1,1
5263           do j=1,nlobit
5264             if (emin.gt.contr(j,iii)) emin=contr(j,iii)
5265           enddo 
5266         enddo
5267         emin=0.5D0*emin
5268 cd      print *,'it=',it,' emin=',emin
5269
5270 C Compute the contribution to SC energy and derivatives
5271         do iii=-1,1
5272
5273           do j=1,nlobit
5274 #ifdef OSF
5275             adexp=bsc(j,it)-0.5D0*contr(j,iii)+emin
5276             if(adexp.ne.adexp) adexp=1.0
5277             expfac=dexp(adexp)
5278 #else
5279             expfac=dexp(bsc(j,it)-0.5D0*contr(j,iii)+emin)
5280 #endif
5281 cd          print *,'j=',j,' expfac=',expfac
5282             escloc_i=escloc_i+expfac
5283             do k=1,3
5284               dersc(k)=dersc(k)+Ax(k,j,iii)*expfac
5285             enddo
5286             if (mixed) then
5287               do k=1,3,2
5288                 ddersc(k)=ddersc(k)+(-Ax(2,j,iii)*Ax(k,j,iii)
5289      &            +gaussc(k,2,j,it))*expfac
5290               enddo
5291             endif
5292           enddo
5293
5294         enddo ! iii
5295
5296         dersc(1)=dersc(1)/cos(theti)**2
5297         ddersc(1)=ddersc(1)/cos(theti)**2
5298         ddersc(3)=ddersc(3)
5299
5300         escloci=-(dlog(escloc_i)-emin)
5301         do j=1,3
5302           dersc(j)=dersc(j)/escloc_i
5303         enddo
5304         if (mixed) then
5305           do j=1,3,2
5306             ddersc(j)=(ddersc(j)/escloc_i+dersc(2)*dersc(j))
5307           enddo
5308         endif
5309       return
5310       end
5311 C------------------------------------------------------------------------------
5312       subroutine enesc_bound(x,escloci,dersc,dersc12,mixed)
5313       implicit real*8 (a-h,o-z)
5314       include 'DIMENSIONS'
5315       include 'COMMON.GEO'
5316       include 'COMMON.LOCAL'
5317       include 'COMMON.IOUNITS'
5318       common /sccalc/ time11,time12,time112,theti,it,nlobit
5319       double precision x(3),z(3),Ax(3,maxlob),dersc(3)
5320       double precision contr(maxlob)
5321       logical mixed
5322
5323       escloc_i=0.0D0
5324
5325       do j=1,3
5326         dersc(j)=0.0D0
5327       enddo
5328
5329       do j=1,nlobit
5330         do k=1,2
5331           z(k)=x(k)-censc(k,j,it)
5332         enddo
5333         z(3)=dwapi
5334         do k=1,3
5335           Axk=0.0D0
5336           do l=1,3
5337             Axk=Axk+gaussc(l,k,j,it)*z(l)
5338           enddo
5339           Ax(k,j)=Axk
5340         enddo 
5341         expfac=0.0D0 
5342         do k=1,3
5343           expfac=expfac+Ax(k,j)*z(k)
5344         enddo
5345         contr(j)=expfac
5346       enddo ! j
5347
5348 C As in the case of ebend, we want to avoid underflows in exponentiation and
5349 C subsequent NaNs and INFs in energy calculation.
5350 C Find the largest exponent
5351       emin=contr(1)
5352       do j=1,nlobit
5353         if (emin.gt.contr(j)) emin=contr(j)
5354       enddo 
5355       emin=0.5D0*emin
5356  
5357 C Compute the contribution to SC energy and derivatives
5358
5359       dersc12=0.0d0
5360       do j=1,nlobit
5361         expfac=dexp(bsc(j,it)-0.5D0*contr(j)+emin)
5362         escloc_i=escloc_i+expfac
5363         do k=1,2
5364           dersc(k)=dersc(k)+Ax(k,j)*expfac
5365         enddo
5366         if (mixed) dersc12=dersc12+(-Ax(2,j)*Ax(1,j)
5367      &            +gaussc(1,2,j,it))*expfac
5368         dersc(3)=0.0d0
5369       enddo
5370
5371       dersc(1)=dersc(1)/cos(theti)**2
5372       dersc12=dersc12/cos(theti)**2
5373       escloci=-(dlog(escloc_i)-emin)
5374       do j=1,2
5375         dersc(j)=dersc(j)/escloc_i
5376       enddo
5377       if (mixed) dersc12=(dersc12/escloc_i+dersc(2)*dersc(1))
5378       return
5379       end
5380 #else
5381 c----------------------------------------------------------------------------------
5382       subroutine esc(escloc)
5383 C Calculate the local energy of a side chain and its derivatives in the
5384 C corresponding virtual-bond valence angles THETA and the spherical angles 
5385 C ALPHA and OMEGA derived from AM1 all-atom calculations.
5386 C added by Urszula Kozlowska. 07/11/2007
5387 C
5388       implicit real*8 (a-h,o-z)
5389       include 'DIMENSIONS'
5390       include 'COMMON.GEO'
5391       include 'COMMON.LOCAL'
5392       include 'COMMON.VAR'
5393       include 'COMMON.SCROT'
5394       include 'COMMON.INTERACT'
5395       include 'COMMON.DERIV'
5396       include 'COMMON.CHAIN'
5397       include 'COMMON.IOUNITS'
5398       include 'COMMON.NAMES'
5399       include 'COMMON.FFIELD'
5400       include 'COMMON.CONTROL'
5401       include 'COMMON.VECTORS'
5402       double precision x_prime(3),y_prime(3),z_prime(3)
5403      &    , sumene,dsc_i,dp2_i,x(65),
5404      &     xx,yy,zz,sumene1,sumene2,sumene3,sumene4,s1,s1_6,s2,s2_6,
5405      &    de_dxx,de_dyy,de_dzz,de_dt
5406       double precision s1_t,s1_6_t,s2_t,s2_6_t
5407       double precision 
5408      & dXX_Ci1(3),dYY_Ci1(3),dZZ_Ci1(3),dXX_Ci(3),
5409      & dYY_Ci(3),dZZ_Ci(3),dXX_XYZ(3),dYY_XYZ(3),dZZ_XYZ(3),
5410      & dt_dCi(3),dt_dCi1(3)
5411       common /sccalc/ time11,time12,time112,theti,it,nlobit
5412       delta=0.02d0*pi
5413       escloc=0.0D0
5414 c      write(iout,*) "ESC: loc_start",loc_start," loc_end",loc_end
5415       do i=loc_start,loc_end
5416         costtab(i+1) =dcos(theta(i+1))
5417         sinttab(i+1) =dsqrt(1-costtab(i+1)*costtab(i+1))
5418         cost2tab(i+1)=dsqrt(0.5d0*(1.0d0+costtab(i+1)))
5419         sint2tab(i+1)=dsqrt(0.5d0*(1.0d0-costtab(i+1)))
5420         cosfac2=0.5d0/(1.0d0+costtab(i+1))
5421         cosfac=dsqrt(cosfac2)
5422         sinfac2=0.5d0/(1.0d0-costtab(i+1))
5423         sinfac=dsqrt(sinfac2)
5424         it=itype(i)
5425         if (it.eq.10) goto 1
5426 c
5427 C  Compute the axes of tghe local cartesian coordinates system; store in
5428 c   x_prime, y_prime and z_prime 
5429 c
5430         do j=1,3
5431           x_prime(j) = 0.00
5432           y_prime(j) = 0.00
5433           z_prime(j) = 0.00
5434         enddo
5435 C        write(2,*) "dc_norm", dc_norm(1,i+nres),dc_norm(2,i+nres),
5436 C     &   dc_norm(3,i+nres)
5437         do j = 1,3
5438           x_prime(j) = (dc_norm(j,i) - dc_norm(j,i-1))*cosfac
5439           y_prime(j) = (dc_norm(j,i) + dc_norm(j,i-1))*sinfac
5440         enddo
5441         do j = 1,3
5442           z_prime(j) = -uz(j,i-1)
5443         enddo     
5444 c       write (2,*) "i",i
5445 c       write (2,*) "x_prime",(x_prime(j),j=1,3)
5446 c       write (2,*) "y_prime",(y_prime(j),j=1,3)
5447 c       write (2,*) "z_prime",(z_prime(j),j=1,3)
5448 c       write (2,*) "xx",scalar(x_prime(1),x_prime(1)),
5449 c      & " xy",scalar(x_prime(1),y_prime(1)),
5450 c      & " xz",scalar(x_prime(1),z_prime(1)),
5451 c      & " yy",scalar(y_prime(1),y_prime(1)),
5452 c      & " yz",scalar(y_prime(1),z_prime(1)),
5453 c      & " zz",scalar(z_prime(1),z_prime(1))
5454 c
5455 C Transform the unit vector of the ith side-chain centroid, dC_norm(*,i),
5456 C to local coordinate system. Store in xx, yy, zz.
5457 c
5458         xx=0.0d0
5459         yy=0.0d0
5460         zz=0.0d0
5461         do j = 1,3
5462           xx = xx + x_prime(j)*dc_norm(j,i+nres)
5463           yy = yy + y_prime(j)*dc_norm(j,i+nres)
5464           zz = zz + z_prime(j)*dc_norm(j,i+nres)
5465         enddo
5466
5467         xxtab(i)=xx
5468         yytab(i)=yy
5469         zztab(i)=zz
5470 C
5471 C Compute the energy of the ith side cbain
5472 C
5473 c        write (2,*) "xx",xx," yy",yy," zz",zz
5474         it=itype(i)
5475         do j = 1,65
5476           x(j) = sc_parmin(j,it) 
5477         enddo
5478 #ifdef CHECK_COORD
5479 Cc diagnostics - remove later
5480         xx1 = dcos(alph(2))
5481         yy1 = dsin(alph(2))*dcos(omeg(2))
5482         zz1 = -dsin(alph(2))*dsin(omeg(2))
5483         write(2,'(3f8.1,3f9.3,1x,3f9.3)') 
5484      &    alph(2)*rad2deg,omeg(2)*rad2deg,theta(3)*rad2deg,xx,yy,zz,
5485      &    xx1,yy1,zz1
5486 C,"  --- ", xx_w,yy_w,zz_w
5487 c end diagnostics
5488 #endif
5489         sumene1= x(1)+  x(2)*xx+  x(3)*yy+  x(4)*zz+  x(5)*xx**2
5490      &   + x(6)*yy**2+  x(7)*zz**2+  x(8)*xx*zz+  x(9)*xx*yy
5491      &   + x(10)*yy*zz
5492         sumene2=  x(11) + x(12)*xx + x(13)*yy + x(14)*zz + x(15)*xx**2
5493      & + x(16)*yy**2 + x(17)*zz**2 + x(18)*xx*zz + x(19)*xx*yy
5494      & + x(20)*yy*zz
5495         sumene3=  x(21) +x(22)*xx +x(23)*yy +x(24)*zz +x(25)*xx**2
5496      &  +x(26)*yy**2 +x(27)*zz**2 +x(28)*xx*zz +x(29)*xx*yy
5497      &  +x(30)*yy*zz +x(31)*xx**3 +x(32)*yy**3 +x(33)*zz**3
5498      &  +x(34)*(xx**2)*yy +x(35)*(xx**2)*zz +x(36)*(yy**2)*xx
5499      &  +x(37)*(yy**2)*zz +x(38)*(zz**2)*xx +x(39)*(zz**2)*yy
5500      &  +x(40)*xx*yy*zz
5501         sumene4= x(41) +x(42)*xx +x(43)*yy +x(44)*zz +x(45)*xx**2
5502      &  +x(46)*yy**2 +x(47)*zz**2 +x(48)*xx*zz +x(49)*xx*yy
5503      &  +x(50)*yy*zz +x(51)*xx**3 +x(52)*yy**3 +x(53)*zz**3
5504      &  +x(54)*(xx**2)*yy +x(55)*(xx**2)*zz +x(56)*(yy**2)*xx
5505      &  +x(57)*(yy**2)*zz +x(58)*(zz**2)*xx +x(59)*(zz**2)*yy
5506      &  +x(60)*xx*yy*zz
5507         dsc_i   = 0.743d0+x(61)
5508         dp2_i   = 1.9d0+x(62)
5509         dscp1=dsqrt(dsc_i**2+dp2_i**2-2*dsc_i*dp2_i
5510      &          *(xx*cost2tab(i+1)+yy*sint2tab(i+1)))
5511         dscp2=dsqrt(dsc_i**2+dp2_i**2-2*dsc_i*dp2_i
5512      &          *(xx*cost2tab(i+1)-yy*sint2tab(i+1)))
5513         s1=(1+x(63))/(0.1d0 + dscp1)
5514         s1_6=(1+x(64))/(0.1d0 + dscp1**6)
5515         s2=(1+x(65))/(0.1d0 + dscp2)
5516         s2_6=(1+x(65))/(0.1d0 + dscp2**6)
5517         sumene = ( sumene3*sint2tab(i+1) + sumene1)*(s1+s1_6)
5518      & + (sumene4*cost2tab(i+1) +sumene2)*(s2+s2_6)
5519 c        write(2,'(i2," sumene",7f9.3)') i,sumene1,sumene2,sumene3,
5520 c     &   sumene4,
5521 c     &   dscp1,dscp2,sumene
5522 c        sumene = enesc(x,xx,yy,zz,cost2tab(i+1),sint2tab(i+1))
5523         escloc = escloc + sumene
5524 c        write (2,*) "i",i," escloc",sumene,escloc
5525 #ifdef DEBUG
5526 C
5527 C This section to check the numerical derivatives of the energy of ith side
5528 C chain in xx, yy, zz, and theta. Use the -DDEBUG compiler option or insert
5529 C #define DEBUG in the code to turn it on.
5530 C
5531         write (2,*) "sumene               =",sumene
5532         aincr=1.0d-7
5533         xxsave=xx
5534         xx=xx+aincr
5535         write (2,*) xx,yy,zz
5536         sumenep = enesc(x,xx,yy,zz,cost2tab(i+1),sint2tab(i+1))
5537         de_dxx_num=(sumenep-sumene)/aincr
5538         xx=xxsave
5539         write (2,*) "xx+ sumene from enesc=",sumenep
5540         yysave=yy
5541         yy=yy+aincr
5542         write (2,*) xx,yy,zz
5543         sumenep = enesc(x,xx,yy,zz,cost2tab(i+1),sint2tab(i+1))
5544         de_dyy_num=(sumenep-sumene)/aincr
5545         yy=yysave
5546         write (2,*) "yy+ sumene from enesc=",sumenep
5547         zzsave=zz
5548         zz=zz+aincr
5549         write (2,*) xx,yy,zz
5550         sumenep = enesc(x,xx,yy,zz,cost2tab(i+1),sint2tab(i+1))
5551         de_dzz_num=(sumenep-sumene)/aincr
5552         zz=zzsave
5553         write (2,*) "zz+ sumene from enesc=",sumenep
5554         costsave=cost2tab(i+1)
5555         sintsave=sint2tab(i+1)
5556         cost2tab(i+1)=dcos(0.5d0*(theta(i+1)+aincr))
5557         sint2tab(i+1)=dsin(0.5d0*(theta(i+1)+aincr))
5558         sumenep = enesc(x,xx,yy,zz,cost2tab(i+1),sint2tab(i+1))
5559         de_dt_num=(sumenep-sumene)/aincr
5560         write (2,*) " t+ sumene from enesc=",sumenep
5561         cost2tab(i+1)=costsave
5562         sint2tab(i+1)=sintsave
5563 C End of diagnostics section.
5564 #endif
5565 C        
5566 C Compute the gradient of esc
5567 C
5568         pom_s1=(1.0d0+x(63))/(0.1d0 + dscp1)**2
5569         pom_s16=6*(1.0d0+x(64))/(0.1d0 + dscp1**6)**2
5570         pom_s2=(1.0d0+x(65))/(0.1d0 + dscp2)**2
5571         pom_s26=6*(1.0d0+x(65))/(0.1d0 + dscp2**6)**2
5572         pom_dx=dsc_i*dp2_i*cost2tab(i+1)
5573         pom_dy=dsc_i*dp2_i*sint2tab(i+1)
5574         pom_dt1=-0.5d0*dsc_i*dp2_i*(xx*sint2tab(i+1)-yy*cost2tab(i+1))
5575         pom_dt2=-0.5d0*dsc_i*dp2_i*(xx*sint2tab(i+1)+yy*cost2tab(i+1))
5576         pom1=(sumene3*sint2tab(i+1)+sumene1)
5577      &     *(pom_s1/dscp1+pom_s16*dscp1**4)
5578         pom2=(sumene4*cost2tab(i+1)+sumene2)
5579      &     *(pom_s2/dscp2+pom_s26*dscp2**4)
5580         sumene1x=x(2)+2*x(5)*xx+x(8)*zz+ x(9)*yy
5581         sumene3x=x(22)+2*x(25)*xx+x(28)*zz+x(29)*yy+3*x(31)*xx**2
5582      &  +2*x(34)*xx*yy +2*x(35)*xx*zz +x(36)*(yy**2) +x(38)*(zz**2)
5583      &  +x(40)*yy*zz
5584         sumene2x=x(12)+2*x(15)*xx+x(18)*zz+ x(19)*yy
5585         sumene4x=x(42)+2*x(45)*xx +x(48)*zz +x(49)*yy +3*x(51)*xx**2
5586      &  +2*x(54)*xx*yy+2*x(55)*xx*zz+x(56)*(yy**2)+x(58)*(zz**2)
5587      &  +x(60)*yy*zz
5588         de_dxx =(sumene1x+sumene3x*sint2tab(i+1))*(s1+s1_6)
5589      &        +(sumene2x+sumene4x*cost2tab(i+1))*(s2+s2_6)
5590      &        +(pom1+pom2)*pom_dx
5591 #ifdef DEBUG
5592         write(2,*), "de_dxx = ", de_dxx,de_dxx_num
5593 #endif
5594 C
5595         sumene1y=x(3) + 2*x(6)*yy + x(9)*xx + x(10)*zz
5596         sumene3y=x(23) +2*x(26)*yy +x(29)*xx +x(30)*zz +3*x(32)*yy**2
5597      &  +x(34)*(xx**2) +2*x(36)*yy*xx +2*x(37)*yy*zz +x(39)*(zz**2)
5598      &  +x(40)*xx*zz
5599         sumene2y=x(13) + 2*x(16)*yy + x(19)*xx + x(20)*zz
5600         sumene4y=x(43)+2*x(46)*yy+x(49)*xx +x(50)*zz
5601      &  +3*x(52)*yy**2+x(54)*xx**2+2*x(56)*yy*xx +2*x(57)*yy*zz
5602      &  +x(59)*zz**2 +x(60)*xx*zz
5603         de_dyy =(sumene1y+sumene3y*sint2tab(i+1))*(s1+s1_6)
5604      &        +(sumene2y+sumene4y*cost2tab(i+1))*(s2+s2_6)
5605      &        +(pom1-pom2)*pom_dy
5606 #ifdef DEBUG
5607         write(2,*), "de_dyy = ", de_dyy,de_dyy_num
5608 #endif
5609 C
5610         de_dzz =(x(24) +2*x(27)*zz +x(28)*xx +x(30)*yy
5611      &  +3*x(33)*zz**2 +x(35)*xx**2 +x(37)*yy**2 +2*x(38)*zz*xx 
5612      &  +2*x(39)*zz*yy +x(40)*xx*yy)*sint2tab(i+1)*(s1+s1_6) 
5613      &  +(x(4) + 2*x(7)*zz+  x(8)*xx + x(10)*yy)*(s1+s1_6) 
5614      &  +(x(44)+2*x(47)*zz +x(48)*xx   +x(50)*yy  +3*x(53)*zz**2   
5615      &  +x(55)*xx**2 +x(57)*(yy**2)+2*x(58)*zz*xx +2*x(59)*zz*yy  
5616      &  +x(60)*xx*yy)*cost2tab(i+1)*(s2+s2_6)
5617      &  + ( x(14) + 2*x(17)*zz+  x(18)*xx + x(20)*yy)*(s2+s2_6)
5618 #ifdef DEBUG
5619         write(2,*), "de_dzz = ", de_dzz,de_dzz_num
5620 #endif
5621 C
5622         de_dt =  0.5d0*sumene3*cost2tab(i+1)*(s1+s1_6) 
5623      &  -0.5d0*sumene4*sint2tab(i+1)*(s2+s2_6)
5624      &  +pom1*pom_dt1+pom2*pom_dt2
5625 #ifdef DEBUG
5626         write(2,*), "de_dt = ", de_dt,de_dt_num
5627 #endif
5628
5629 C
5630        cossc=scalar(dc_norm(1,i),dc_norm(1,i+nres))
5631        cossc1=scalar(dc_norm(1,i-1),dc_norm(1,i+nres))
5632        cosfac2xx=cosfac2*xx
5633        sinfac2yy=sinfac2*yy
5634        do k = 1,3
5635          dt_dCi(k) = -(dc_norm(k,i-1)+costtab(i+1)*dc_norm(k,i))*
5636      &      vbld_inv(i+1)
5637          dt_dCi1(k)= -(dc_norm(k,i)+costtab(i+1)*dc_norm(k,i-1))*
5638      &      vbld_inv(i)
5639          pom=(dC_norm(k,i+nres)-cossc*dC_norm(k,i))*vbld_inv(i+1)
5640          pom1=(dC_norm(k,i+nres)-cossc1*dC_norm(k,i-1))*vbld_inv(i)
5641 c         write (iout,*) "i",i," k",k," pom",pom," pom1",pom1,
5642 c     &    " dt_dCi",dt_dCi(k)," dt_dCi1",dt_dCi1(k)
5643 c         write (iout,*) "dC_norm",(dC_norm(j,i),j=1,3),
5644 c     &   (dC_norm(j,i-1),j=1,3)," vbld_inv",vbld_inv(i+1),vbld_inv(i)
5645          dXX_Ci(k)=pom*cosfac-dt_dCi(k)*cosfac2xx
5646          dXX_Ci1(k)=-pom1*cosfac-dt_dCi1(k)*cosfac2xx
5647          dYY_Ci(k)=pom*sinfac+dt_dCi(k)*sinfac2yy
5648          dYY_Ci1(k)=pom1*sinfac+dt_dCi1(k)*sinfac2yy
5649          dZZ_Ci1(k)=0.0d0
5650          dZZ_Ci(k)=0.0d0
5651          do j=1,3
5652            dZZ_Ci(k)=dZZ_Ci(k)-uzgrad(j,k,2,i-1)*dC_norm(j,i+nres)
5653            dZZ_Ci1(k)=dZZ_Ci1(k)-uzgrad(j,k,1,i-1)*dC_norm(j,i+nres)
5654          enddo
5655           
5656          dXX_XYZ(k)=vbld_inv(i+nres)*(x_prime(k)-xx*dC_norm(k,i+nres))
5657          dYY_XYZ(k)=vbld_inv(i+nres)*(y_prime(k)-yy*dC_norm(k,i+nres))
5658          dZZ_XYZ(k)=vbld_inv(i+nres)*(z_prime(k)-zz*dC_norm(k,i+nres))
5659 c
5660          dt_dCi(k) = -dt_dCi(k)/sinttab(i+1)
5661          dt_dCi1(k)= -dt_dCi1(k)/sinttab(i+1)
5662        enddo
5663
5664        do k=1,3
5665          dXX_Ctab(k,i)=dXX_Ci(k)
5666          dXX_C1tab(k,i)=dXX_Ci1(k)
5667          dYY_Ctab(k,i)=dYY_Ci(k)
5668          dYY_C1tab(k,i)=dYY_Ci1(k)
5669          dZZ_Ctab(k,i)=dZZ_Ci(k)
5670          dZZ_C1tab(k,i)=dZZ_Ci1(k)
5671          dXX_XYZtab(k,i)=dXX_XYZ(k)
5672          dYY_XYZtab(k,i)=dYY_XYZ(k)
5673          dZZ_XYZtab(k,i)=dZZ_XYZ(k)
5674        enddo
5675
5676        do k = 1,3
5677 c         write (iout,*) "k",k," dxx_ci1",dxx_ci1(k)," dyy_ci1",
5678 c     &    dyy_ci1(k)," dzz_ci1",dzz_ci1(k)
5679 c         write (iout,*) "k",k," dxx_ci",dxx_ci(k)," dyy_ci",
5680 c     &    dyy_ci(k)," dzz_ci",dzz_ci(k)
5681 c         write (iout,*) "k",k," dt_dci",dt_dci(k)," dt_dci",
5682 c     &    dt_dci(k)
5683 c         write (iout,*) "k",k," dxx_XYZ",dxx_XYZ(k)," dyy_XYZ",
5684 c     &    dyy_XYZ(k)," dzz_XYZ",dzz_XYZ(k) 
5685          gscloc(k,i-1)=gscloc(k,i-1)+de_dxx*dxx_ci1(k)
5686      &    +de_dyy*dyy_ci1(k)+de_dzz*dzz_ci1(k)+de_dt*dt_dCi1(k)
5687          gscloc(k,i)=gscloc(k,i)+de_dxx*dxx_Ci(k)
5688      &    +de_dyy*dyy_Ci(k)+de_dzz*dzz_Ci(k)+de_dt*dt_dCi(k)
5689          gsclocx(k,i)=                 de_dxx*dxx_XYZ(k)
5690      &    +de_dyy*dyy_XYZ(k)+de_dzz*dzz_XYZ(k)
5691        enddo
5692 c       write(iout,*) "ENERGY GRAD = ", (gscloc(k,i-1),k=1,3),
5693 c     &  (gscloc(k,i),k=1,3),(gsclocx(k,i),k=1,3)  
5694
5695 C to check gradient call subroutine check_grad
5696
5697     1 continue
5698       enddo
5699       return
5700       end
5701 c------------------------------------------------------------------------------
5702       double precision function enesc(x,xx,yy,zz,cost2,sint2)
5703       implicit none
5704       double precision x(65),xx,yy,zz,cost2,sint2,sumene1,sumene2,
5705      & sumene3,sumene4,sumene,dsc_i,dp2_i,dscp1,dscp2,s1,s1_6,s2,s2_6
5706       sumene1= x(1)+  x(2)*xx+  x(3)*yy+  x(4)*zz+  x(5)*xx**2
5707      &   + x(6)*yy**2+  x(7)*zz**2+  x(8)*xx*zz+  x(9)*xx*yy
5708      &   + x(10)*yy*zz
5709       sumene2=  x(11) + x(12)*xx + x(13)*yy + x(14)*zz + x(15)*xx**2
5710      & + x(16)*yy**2 + x(17)*zz**2 + x(18)*xx*zz + x(19)*xx*yy
5711      & + x(20)*yy*zz
5712       sumene3=  x(21) +x(22)*xx +x(23)*yy +x(24)*zz +x(25)*xx**2
5713      &  +x(26)*yy**2 +x(27)*zz**2 +x(28)*xx*zz +x(29)*xx*yy
5714      &  +x(30)*yy*zz +x(31)*xx**3 +x(32)*yy**3 +x(33)*zz**3
5715      &  +x(34)*(xx**2)*yy +x(35)*(xx**2)*zz +x(36)*(yy**2)*xx
5716      &  +x(37)*(yy**2)*zz +x(38)*(zz**2)*xx +x(39)*(zz**2)*yy
5717      &  +x(40)*xx*yy*zz
5718       sumene4= x(41) +x(42)*xx +x(43)*yy +x(44)*zz +x(45)*xx**2
5719      &  +x(46)*yy**2 +x(47)*zz**2 +x(48)*xx*zz +x(49)*xx*yy
5720      &  +x(50)*yy*zz +x(51)*xx**3 +x(52)*yy**3 +x(53)*zz**3
5721      &  +x(54)*(xx**2)*yy +x(55)*(xx**2)*zz +x(56)*(yy**2)*xx
5722      &  +x(57)*(yy**2)*zz +x(58)*(zz**2)*xx +x(59)*(zz**2)*yy
5723      &  +x(60)*xx*yy*zz
5724       dsc_i   = 0.743d0+x(61)
5725       dp2_i   = 1.9d0+x(62)
5726       dscp1=dsqrt(dsc_i**2+dp2_i**2-2*dsc_i*dp2_i
5727      &          *(xx*cost2+yy*sint2))
5728       dscp2=dsqrt(dsc_i**2+dp2_i**2-2*dsc_i*dp2_i
5729      &          *(xx*cost2-yy*sint2))
5730       s1=(1+x(63))/(0.1d0 + dscp1)
5731       s1_6=(1+x(64))/(0.1d0 + dscp1**6)
5732       s2=(1+x(65))/(0.1d0 + dscp2)
5733       s2_6=(1+x(65))/(0.1d0 + dscp2**6)
5734       sumene = ( sumene3*sint2 + sumene1)*(s1+s1_6)
5735      & + (sumene4*cost2 +sumene2)*(s2+s2_6)
5736       enesc=sumene
5737       return
5738       end
5739 #endif
5740 c------------------------------------------------------------------------------
5741       subroutine gcont(rij,r0ij,eps0ij,delta,fcont,fprimcont)
5742 C
5743 C This procedure calculates two-body contact function g(rij) and its derivative:
5744 C
5745 C           eps0ij                                     !       x < -1
5746 C g(rij) =  esp0ij*(-0.9375*x+0.625*x**3-0.1875*x**5)  ! -1 =< x =< 1
5747 C            0                                         !       x > 1
5748 C
5749 C where x=(rij-r0ij)/delta
5750 C
5751 C rij - interbody distance, r0ij - contact distance, eps0ij - contact energy
5752 C
5753       implicit none
5754       double precision rij,r0ij,eps0ij,fcont,fprimcont
5755       double precision x,x2,x4,delta
5756 c     delta=0.02D0*r0ij
5757 c      delta=0.2D0*r0ij
5758       x=(rij-r0ij)/delta
5759       if (x.lt.-1.0D0) then
5760         fcont=eps0ij
5761         fprimcont=0.0D0
5762       else if (x.le.1.0D0) then  
5763         x2=x*x
5764         x4=x2*x2
5765         fcont=eps0ij*(x*(-0.9375D0+0.6250D0*x2-0.1875D0*x4)+0.5D0)
5766         fprimcont=eps0ij * (-0.9375D0+1.8750D0*x2-0.9375D0*x4)/delta
5767       else
5768         fcont=0.0D0
5769         fprimcont=0.0D0
5770       endif
5771       return
5772       end
5773 c------------------------------------------------------------------------------
5774       subroutine splinthet(theti,delta,ss,ssder)
5775       implicit real*8 (a-h,o-z)
5776       include 'DIMENSIONS'
5777       include 'COMMON.VAR'
5778       include 'COMMON.GEO'
5779       thetup=pi-delta
5780       thetlow=delta
5781       if (theti.gt.pipol) then
5782         call gcont(theti,thetup,1.0d0,delta,ss,ssder)
5783       else
5784         call gcont(-theti,-thetlow,1.0d0,delta,ss,ssder)
5785         ssder=-ssder
5786       endif
5787       return
5788       end
5789 c------------------------------------------------------------------------------
5790       subroutine spline1(x,x0,delta,f0,f1,fprim0,f,fprim)
5791       implicit none
5792       double precision x,x0,delta,f0,f1,fprim0,f,fprim
5793       double precision ksi,ksi2,ksi3,a1,a2,a3
5794       a1=fprim0*delta/(f1-f0)
5795       a2=3.0d0-2.0d0*a1
5796       a3=a1-2.0d0
5797       ksi=(x-x0)/delta
5798       ksi2=ksi*ksi
5799       ksi3=ksi2*ksi  
5800       f=f0+(f1-f0)*ksi*(a1+ksi*(a2+a3*ksi))
5801       fprim=(f1-f0)/delta*(a1+ksi*(2*a2+3*ksi*a3))
5802       return
5803       end
5804 c------------------------------------------------------------------------------
5805       subroutine spline2(x,x0,delta,f0x,f1x,fprim0x,fx)
5806       implicit none
5807       double precision x,x0,delta,f0x,f1x,fprim0x,fx
5808       double precision ksi,ksi2,ksi3,a1,a2,a3
5809       ksi=(x-x0)/delta  
5810       ksi2=ksi*ksi
5811       ksi3=ksi2*ksi
5812       a1=fprim0x*delta
5813       a2=3*(f1x-f0x)-2*fprim0x*delta
5814       a3=fprim0x*delta-2*(f1x-f0x)
5815       fx=f0x+a1*ksi+a2*ksi2+a3*ksi3
5816       return
5817       end
5818 C-----------------------------------------------------------------------------
5819 #ifdef CRYST_TOR
5820 C-----------------------------------------------------------------------------
5821       subroutine etor(etors,edihcnstr)
5822       implicit real*8 (a-h,o-z)
5823       include 'DIMENSIONS'
5824       include 'COMMON.VAR'
5825       include 'COMMON.GEO'
5826       include 'COMMON.LOCAL'
5827       include 'COMMON.TORSION'
5828       include 'COMMON.INTERACT'
5829       include 'COMMON.DERIV'
5830       include 'COMMON.CHAIN'
5831       include 'COMMON.NAMES'
5832       include 'COMMON.IOUNITS'
5833       include 'COMMON.FFIELD'
5834       include 'COMMON.TORCNSTR'
5835       include 'COMMON.CONTROL'
5836       logical lprn
5837 C Set lprn=.true. for debugging
5838       lprn=.false.
5839 c      lprn=.true.
5840       etors=0.0D0
5841       do i=iphi_start,iphi_end
5842       etors_ii=0.0D0
5843         itori=itortyp(itype(i-2))
5844         itori1=itortyp(itype(i-1))
5845         phii=phi(i)
5846         gloci=0.0D0
5847 C Proline-Proline pair is a special case...
5848         if (itori.eq.3 .and. itori1.eq.3) then
5849           if (phii.gt.-dwapi3) then
5850             cosphi=dcos(3*phii)
5851             fac=1.0D0/(1.0D0-cosphi)
5852             etorsi=v1(1,3,3)*fac
5853             etorsi=etorsi+etorsi
5854             etors=etors+etorsi-v1(1,3,3)
5855             if (energy_dec) etors_ii=etors_ii+etorsi-v1(1,3,3)      
5856             gloci=gloci-3*fac*etorsi*dsin(3*phii)
5857           endif
5858           do j=1,3
5859             v1ij=v1(j+1,itori,itori1)
5860             v2ij=v2(j+1,itori,itori1)
5861             cosphi=dcos(j*phii)
5862             sinphi=dsin(j*phii)
5863             etors=etors+v1ij*cosphi+v2ij*sinphi+dabs(v1ij)+dabs(v2ij)
5864             if (energy_dec) etors_ii=etors_ii+
5865      &                              v2ij*sinphi+dabs(v1ij)+dabs(v2ij)
5866             gloci=gloci+j*(v2ij*cosphi-v1ij*sinphi)
5867           enddo
5868         else 
5869           do j=1,nterm_old
5870             v1ij=v1(j,itori,itori1)
5871             v2ij=v2(j,itori,itori1)
5872             cosphi=dcos(j*phii)
5873             sinphi=dsin(j*phii)
5874             etors=etors+v1ij*cosphi+v2ij*sinphi+dabs(v1ij)+dabs(v2ij)
5875             if (energy_dec) etors_ii=etors_ii+
5876      &                  v1ij*cosphi+v2ij*sinphi+dabs(v1ij)+dabs(v2ij)
5877             gloci=gloci+j*(v2ij*cosphi-v1ij*sinphi)
5878           enddo
5879         endif
5880         if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
5881      &        'etor',i,etors_ii
5882         if (lprn)
5883      &  write (iout,'(2(a3,2x,i3,2x),2i3,6f8.3/26x,6f8.3/)')
5884      &  restyp(itype(i-2)),i-2,restyp(itype(i-1)),i-1,itori,itori1,
5885      &  (v1(j,itori,itori1),j=1,6),(v2(j,itori,itori1),j=1,6)
5886         gloc(i-3,icg)=gloc(i-3,icg)+wtor*gloci
5887         write (iout,*) 'i=',i,' gloc=',gloc(i-3,icg)
5888       enddo
5889 ! 6/20/98 - dihedral angle constraints
5890       edihcnstr=0.0d0
5891       do i=1,ndih_constr
5892         itori=idih_constr(i)
5893         phii=phi(itori)
5894         difi=phii-phi0(i)
5895         if (difi.gt.drange(i)) then
5896           difi=difi-drange(i)
5897           edihcnstr=edihcnstr+0.25d0*ftors*difi**4
5898           gloc(itori-3,icg)=gloc(itori-3,icg)+ftors*difi**3
5899         else if (difi.lt.-drange(i)) then
5900           difi=difi+drange(i)
5901           edihcnstr=edihcnstr+0.25d0*ftors*difi**4
5902           gloc(itori-3,icg)=gloc(itori-3,icg)+ftors*difi**3
5903         endif
5904 !        write (iout,'(2i5,2f8.3,2e14.5)') i,itori,rad2deg*phii,
5905 !     &    rad2deg*difi,0.25d0*ftors*difi**4,gloc(itori-3,icg)
5906       enddo
5907 !      write (iout,*) 'edihcnstr',edihcnstr
5908       return
5909       end
5910 c------------------------------------------------------------------------------
5911 c LICZENIE WIEZOW Z ROWNANIA ENERGII MODELLERA
5912       subroutine e_modeller(ehomology_constr)
5913       ehomology_constr=0.0d0
5914       write (iout,*) "!!!!!UWAGA, JESTEM W DZIWNEJ PETLI, TEST!!!!!"
5915       return
5916       end
5917 C !!!!!!!! NIE CZYTANE !!!!!!!!!!!
5918
5919 c------------------------------------------------------------------------------
5920       subroutine etor_d(etors_d)
5921       etors_d=0.0d0
5922       return
5923       end
5924 c----------------------------------------------------------------------------
5925 #else
5926       subroutine etor(etors,edihcnstr)
5927       implicit real*8 (a-h,o-z)
5928       include 'DIMENSIONS'
5929       include 'COMMON.VAR'
5930       include 'COMMON.GEO'
5931       include 'COMMON.LOCAL'
5932       include 'COMMON.TORSION'
5933       include 'COMMON.INTERACT'
5934       include 'COMMON.DERIV'
5935       include 'COMMON.CHAIN'
5936       include 'COMMON.NAMES'
5937       include 'COMMON.IOUNITS'
5938       include 'COMMON.FFIELD'
5939       include 'COMMON.TORCNSTR'
5940       include 'COMMON.CONTROL'
5941       logical lprn
5942 C Set lprn=.true. for debugging
5943       lprn=.false.
5944 c     lprn=.true.
5945       etors=0.0D0
5946       do i=iphi_start,iphi_end
5947       etors_ii=0.0D0
5948         itori=itortyp(itype(i-2))
5949         itori1=itortyp(itype(i-1))
5950         phii=phi(i)
5951         gloci=0.0D0
5952 C Regular cosine and sine terms
5953         do j=1,nterm(itori,itori1)
5954           v1ij=v1(j,itori,itori1)
5955           v2ij=v2(j,itori,itori1)
5956           cosphi=dcos(j*phii)
5957           sinphi=dsin(j*phii)
5958           etors=etors+v1ij*cosphi+v2ij*sinphi
5959           if (energy_dec) etors_ii=etors_ii+
5960      &                v1ij*cosphi+v2ij*sinphi
5961           gloci=gloci+j*(v2ij*cosphi-v1ij*sinphi)
5962         enddo
5963 C Lorentz terms
5964 C                         v1
5965 C  E = SUM ----------------------------------- - v1
5966 C          [v2 cos(phi/2)+v3 sin(phi/2)]^2 + 1
5967 C
5968         cosphi=dcos(0.5d0*phii)
5969         sinphi=dsin(0.5d0*phii)
5970         do j=1,nlor(itori,itori1)
5971           vl1ij=vlor1(j,itori,itori1)
5972           vl2ij=vlor2(j,itori,itori1)
5973           vl3ij=vlor3(j,itori,itori1)
5974           pom=vl2ij*cosphi+vl3ij*sinphi
5975           pom1=1.0d0/(pom*pom+1.0d0)
5976           etors=etors+vl1ij*pom1
5977           if (energy_dec) etors_ii=etors_ii+
5978      &                vl1ij*pom1
5979           pom=-pom*pom1*pom1
5980           gloci=gloci+vl1ij*(vl3ij*cosphi-vl2ij*sinphi)*pom
5981         enddo
5982 C Subtract the constant term
5983         etors=etors-v0(itori,itori1)
5984           if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
5985      &         'etor',i,etors_ii-v0(itori,itori1)
5986         if (lprn)
5987      &  write (iout,'(2(a3,2x,i3,2x),2i3,6f8.3/26x,6f8.3/)')
5988      &  restyp(itype(i-2)),i-2,restyp(itype(i-1)),i-1,itori,itori1,
5989      &  (v1(j,itori,itori1),j=1,6),(v2(j,itori,itori1),j=1,6)
5990         gloc(i-3,icg)=gloc(i-3,icg)+wtor*gloci
5991 c       write (iout,*) 'i=',i,' gloc=',gloc(i-3,icg)
5992       enddo
5993 ! 6/20/98 - dihedral angle constraints
5994       edihcnstr=0.0d0
5995 c      do i=1,ndih_constr
5996       do i=idihconstr_start,idihconstr_end
5997         itori=idih_constr(i)
5998         phii=phi(itori)
5999         difi=pinorm(phii-phi0(i))
6000         if (difi.gt.drange(i)) then
6001           difi=difi-drange(i)
6002           edihcnstr=edihcnstr+0.25d0*ftors*difi**4
6003           gloc(itori-3,icg)=gloc(itori-3,icg)+ftors*difi**3
6004         else if (difi.lt.-drange(i)) then
6005           difi=difi+drange(i)
6006           edihcnstr=edihcnstr+0.25d0*ftors*difi**4
6007           gloc(itori-3,icg)=gloc(itori-3,icg)+ftors*difi**3
6008         else
6009           difi=0.0
6010         endif
6011 c        write (iout,*) "gloci", gloc(i-3,icg)
6012 cd        write (iout,'(2i5,4f8.3,2e14.5)') i,itori,rad2deg*phii,
6013 cd     &    rad2deg*phi0(i),  rad2deg*drange(i),
6014 cd     &    rad2deg*difi,0.25d0*ftors*difi**4,gloc(itori-3,icg)
6015       enddo
6016 cd       write (iout,*) 'edihcnstr',edihcnstr
6017       return
6018       end
6019 c----------------------------------------------------------------------------
6020 c MODELLER restraint function
6021       subroutine e_modeller(ehomology_constr)
6022       implicit real*8 (a-h,o-z)
6023       include 'DIMENSIONS'
6024
6025       integer nnn, i, j, k, ki, irec, l
6026       integer katy, odleglosci, test7
6027       real*8 odleg, odleg2, odleg3, kat, kat2, kat3, gdih(max_template)
6028       real*8 Eval,Erot
6029       real*8 distance(max_template),distancek(max_template),
6030      &    min_odl,godl(max_template),dih_diff(max_template)
6031
6032 c
6033 c     FP - 30/10/2014 Temporary specifications for homology restraints
6034 c
6035       double precision utheta_i,gutheta_i,sum_gtheta,sum_sgtheta,
6036      &                 sgtheta      
6037       double precision, dimension (maxres) :: guscdiff,usc_diff
6038       double precision, dimension (max_template) ::  
6039      &           gtheta,dscdiff,uscdiffk,guscdiff2,guscdiff3,
6040      &           theta_diff
6041 c
6042
6043       include 'COMMON.SBRIDGE'
6044       include 'COMMON.CHAIN'
6045       include 'COMMON.GEO'
6046       include 'COMMON.DERIV'
6047       include 'COMMON.LOCAL'
6048       include 'COMMON.INTERACT'
6049       include 'COMMON.VAR'
6050       include 'COMMON.IOUNITS'
6051       include 'COMMON.MD'
6052       include 'COMMON.CONTROL'
6053 c
6054 c     From subroutine Econstr_back
6055 c
6056       include 'COMMON.NAMES'
6057       include 'COMMON.TIME1'
6058 c
6059
6060
6061       do i=1,max_template
6062         distancek(i)=9999999.9
6063       enddo
6064
6065
6066       odleg=0.0d0
6067
6068 c Pseudo-energy and gradient from homology restraints (MODELLER-like
6069 c function)
6070 C AL 5/2/14 - Introduce list of restraints
6071 c     write(iout,*) "waga_theta",waga_theta,"waga_d",waga_d
6072 #ifdef DEBUG
6073       write(iout,*) "------- dist restrs start -------"
6074 #endif
6075       do ii = link_start_homo,link_end_homo
6076          i = ires_homo(ii)
6077          j = jres_homo(ii)
6078          dij=dist(i,j)
6079 c        write (iout,*) "dij(",i,j,") =",dij
6080          nexl=0
6081          do k=1,constr_homology
6082 c           write(iout,*) ii,k,i,j,l_homo(k,ii),dij,odl(k,ii)
6083            if(.not.l_homo(k,ii)) then
6084              nexl=nexl+1
6085              cycle
6086            endif
6087            distance(k)=odl(k,ii)-dij
6088 c          write (iout,*) "distance(",k,") =",distance(k)
6089 c
6090 c          For Gaussian-type Urestr
6091 c
6092            distancek(k)=0.5d0*distance(k)**2*sigma_odl(k,ii) ! waga_dist rmvd from Gaussian argument
6093 c          write (iout,*) "sigma_odl(",k,ii,") =",sigma_odl(k,ii)
6094 c          write (iout,*) "distancek(",k,") =",distancek(k)
6095 c          distancek(k)=0.5d0*waga_dist*distance(k)**2*sigma_odl(k,ii)
6096 c
6097 c          For Lorentzian-type Urestr
6098 c
6099            if (waga_dist.lt.0.0d0) then
6100               sigma_odlir(k,ii)=dsqrt(1/sigma_odl(k,ii))
6101               distancek(k)=distance(k)**2/(sigma_odlir(k,ii)*
6102      &                     (distance(k)**2+sigma_odlir(k,ii)**2))
6103            endif
6104          enddo
6105 c         write (iout,*) "distance: ii",ii," nexl",nexl
6106          
6107
6108 c         min_odl=minval(distancek)
6109          do kk=1,constr_homology
6110           if(l_homo(kk,ii)) then 
6111             min_odl=distancek(kk)
6112             exit
6113           endif
6114          enddo
6115          do kk=1,constr_homology
6116           if(l_homo(kk,ii) .and. distancek(kk).lt.min_odl) 
6117      &              min_odl=distancek(kk)
6118          enddo
6119 c        write (iout,* )"min_odl",min_odl
6120 #ifdef DEBUG
6121          write (iout,*) "ij dij",i,j,dij
6122          write (iout,*) "distance",(distance(k),k=1,constr_homology)
6123          write (iout,*) "distancek",(distancek(k),k=1,constr_homology)
6124          write (iout,* )"min_odl",min_odl
6125 #endif
6126 #ifdef OLDRESTR
6127          odleg2=0.0d0
6128 #else
6129          if (waga_dist.ge.0.0d0) then
6130            odleg2=nexl
6131          else 
6132            odleg2=0.0d0
6133          endif 
6134 #endif
6135          do k=1,constr_homology
6136 c Nie wiem po co to liczycie jeszcze raz!
6137 c            odleg3=-waga_dist(iset)*((distance(i,j,k)**2)/ 
6138 c     &              (2*(sigma_odl(i,j,k))**2))
6139            if(.not.l_homo(k,ii)) cycle
6140            if (waga_dist.ge.0.0d0) then
6141 c
6142 c          For Gaussian-type Urestr
6143 c
6144             godl(k)=dexp(-distancek(k)+min_odl)
6145             odleg2=odleg2+godl(k)
6146 c
6147 c          For Lorentzian-type Urestr
6148 c
6149            else
6150             odleg2=odleg2+distancek(k)
6151            endif
6152
6153 ccc       write(iout,779) i,j,k, "odleg2=",odleg2, "odleg3=", odleg3,
6154 ccc     & "dEXP(odleg3)=", dEXP(odleg3),"distance(i,j,k)^2=",
6155 ccc     & distance(i,j,k)**2, "dist(i+1,j+1)=", dist(i+1,j+1),
6156 ccc     & "sigma_odl(i,j,k)=", sigma_odl(i,j,k)
6157
6158          enddo
6159 c        write (iout,*) "godl",(godl(k),k=1,constr_homology) ! exponents
6160 c        write (iout,*) "ii i j",ii,i,j," odleg2",odleg2 ! sum of exps
6161 #ifdef DEBUG
6162          write (iout,*) "godl",(godl(k),k=1,constr_homology) ! exponents
6163          write (iout,*) "ii i j",ii,i,j," odleg2",odleg2 ! sum of exps
6164 #endif
6165            if (waga_dist.ge.0.0d0) then
6166 c
6167 c          For Gaussian-type Urestr
6168 c
6169               odleg=odleg-dLOG(odleg2/constr_homology)+min_odl
6170 c
6171 c          For Lorentzian-type Urestr
6172 c
6173            else
6174               odleg=odleg+odleg2/constr_homology
6175            endif
6176 c
6177 c        write (iout,*) "odleg",odleg ! sum of -ln-s
6178 c Gradient
6179 c
6180 c          For Gaussian-type Urestr
6181 c
6182          if (waga_dist.ge.0.0d0) sum_godl=odleg2
6183          sum_sgodl=0.0d0
6184          do k=1,constr_homology
6185 c            godl=dexp(((-(distance(i,j,k)**2)/(2*(sigma_odl(i,j,k))**2))
6186 c     &           *waga_dist)+min_odl
6187 c          sgodl=-godl(k)*distance(k)*sigma_odl(k,ii)*waga_dist
6188 c
6189          if(.not.l_homo(k,ii)) cycle
6190          if (waga_dist.ge.0.0d0) then
6191 c          For Gaussian-type Urestr
6192 c
6193            sgodl=-godl(k)*distance(k)*sigma_odl(k,ii) ! waga_dist rmvd
6194 c
6195 c          For Lorentzian-type Urestr
6196 c
6197          else
6198            sgodl=-2*sigma_odlir(k,ii)*(distance(k)/(distance(k)**2+
6199      &           sigma_odlir(k,ii)**2)**2)
6200          endif
6201            sum_sgodl=sum_sgodl+sgodl
6202
6203 c            sgodl2=sgodl2+sgodl
6204 c      write(iout,*) i, j, k, distance(i,j,k), "W GRADIENCIE1"
6205 c      write(iout,*) "constr_homology=",constr_homology
6206 c      write(iout,*) i, j, k, "TEST K"
6207          enddo
6208          if (waga_dist.ge.0.0d0) then
6209 c
6210 c          For Gaussian-type Urestr
6211 c
6212             grad_odl3=waga_homology(iset)*waga_dist
6213      &                *sum_sgodl/(sum_godl*dij)
6214 c
6215 c          For Lorentzian-type Urestr
6216 c
6217          else
6218 c Original grad expr modified by analogy w Gaussian-type Urestr grad
6219 c           grad_odl3=-waga_homology(iset)*waga_dist*sum_sgodl
6220             grad_odl3=-waga_homology(iset)*waga_dist*
6221      &                sum_sgodl/(constr_homology*dij)
6222          endif
6223 c
6224 c        grad_odl3=sum_sgodl/(sum_godl*dij)
6225
6226
6227 c      write(iout,*) i, j, k, distance(i,j,k), "W GRADIENCIE2"
6228 c      write(iout,*) (distance(i,j,k)**2), (2*(sigma_odl(i,j,k))**2),
6229 c     &              (-(distance(i,j,k)**2)/(2*(sigma_odl(i,j,k))**2))
6230
6231 ccc      write(iout,*) godl, sgodl, grad_odl3
6232
6233 c          grad_odl=grad_odl+grad_odl3
6234
6235          do jik=1,3
6236             ggodl=grad_odl3*(c(jik,i)-c(jik,j))
6237 ccc      write(iout,*) c(jik,i+1), c(jik,j+1), (c(jik,i+1)-c(jik,j+1))
6238 ccc      write(iout,746) "GRAD_ODL_1", i, j, jik, ggodl, 
6239 ccc     &              ghpbc(jik,i+1), ghpbc(jik,j+1)
6240             ghpbc(jik,i)=ghpbc(jik,i)+ggodl
6241             ghpbc(jik,j)=ghpbc(jik,j)-ggodl
6242 ccc      write(iout,746) "GRAD_ODL_2", i, j, jik, ggodl,
6243 ccc     &              ghpbc(jik,i+1), ghpbc(jik,j+1)
6244 c         if (i.eq.25.and.j.eq.27) then
6245 c         write(iout,*) "jik",jik,"i",i,"j",j
6246 c         write(iout,*) "sum_sgodl",sum_sgodl,"sgodl",sgodl
6247 c         write(iout,*) "grad_odl3",grad_odl3
6248 c         write(iout,*) "c(",jik,i,")",c(jik,i),"c(",jik,j,")",c(jik,j)
6249 c         write(iout,*) "ggodl",ggodl
6250 c         write(iout,*) "ghpbc(",jik,i,")",
6251 c     &                 ghpbc(jik,i),"ghpbc(",jik,j,")",
6252 c     &                 ghpbc(jik,j)   
6253 c         endif
6254          enddo
6255 ccc       write(iout,778)"TEST: odleg2=", odleg2, "DLOG(odleg2)=", 
6256 ccc     & dLOG(odleg2),"-odleg=", -odleg
6257
6258       enddo ! ii-loop for dist
6259 #ifdef DEBUG
6260       write(iout,*) "------- dist restrs end -------"
6261 c     if (waga_angle.eq.1.0d0 .or. waga_theta.eq.1.0d0 .or. 
6262 c    &     waga_d.eq.1.0d0) call sum_gradient
6263 #endif
6264 c Pseudo-energy and gradient from dihedral-angle restraints from
6265 c homology templates
6266 c      write (iout,*) "End of distance loop"
6267 c      call flush(iout)
6268       kat=0.0d0
6269 c      write (iout,*) idihconstr_start_homo,idihconstr_end_homo
6270 #ifdef DEBUG
6271       write(iout,*) "------- dih restrs start -------"
6272       do i=idihconstr_start_homo,idihconstr_end_homo
6273         write (iout,*) "gloc_init(",i,icg,")",gloc(i,icg)
6274       enddo
6275 #endif
6276       do i=idihconstr_start_homo,idihconstr_end_homo
6277 c        betai=beta(i,i+1,i+2,i+3)
6278         betai = phi(i)
6279 c       write (iout,*) "betai =",betai
6280         kat2=0.0d0
6281         do k=1,constr_homology
6282           dih_diff(k)=pinorm(dih(k,i)-betai)
6283 c         write (iout,*) "dih_diff(",k,") =",dih_diff(k)
6284 c          if (dih_diff(i,k).gt.3.14159) dih_diff(i,k)=
6285 c     &                                   -(6.28318-dih_diff(i,k))
6286 c          if (dih_diff(i,k).lt.-3.14159) dih_diff(i,k)=
6287 c     &                                   6.28318+dih_diff(i,k)
6288 #ifdef OLD_DIHED
6289           kat3=-0.5d0*dih_diff(k)**2*sigma_dih(k,i) ! waga_angle rmvd from Gaussian argument
6290 #else
6291           kat3=(dcos(dih_diff(k))-1)*sigma_dih(k,i) ! waga_angle rmvd from Gaussian argument
6292 #endif
6293 c         kat3=-0.5d0*waga_angle*dih_diff(k)**2*sigma_dih(k,i)
6294           gdih(k)=dexp(kat3)
6295           kat2=kat2+gdih(k)
6296 c          write(iout,*) "i",i," k",k," sigma",sigma_dih(k,i),
6297 c     &     " kat2=", kat2, "gdih=",gdih(k)
6298         enddo
6299 c       write (iout,*) "gdih",(gdih(k),k=1,constr_homology) ! exps
6300 c       write (iout,*) "i",i," betai",betai," kat2",kat2 ! sum of exps
6301 #ifdef DEBUG
6302         write (iout,*) "i",i," betai",betai," kat2",kat2
6303         write (iout,*) "gdih",(gdih(k),k=1,constr_homology)
6304 #endif
6305         if (kat2.le.1.0d-14) cycle
6306         kat=kat-dLOG(kat2/constr_homology)
6307 c       write (iout,*) "kat",kat ! sum of -ln-s
6308
6309 ccc       write(iout,778)"TEST: kat2=", kat2, "DLOG(kat2)=",
6310 ccc     & dLOG(kat2), "-kat=", -kat
6311
6312 c ----------------------------------------------------------------------
6313 c Gradient
6314 c ----------------------------------------------------------------------
6315
6316         sum_gdih=kat2
6317         sum_sgdih=0.0d0
6318         do k=1,constr_homology
6319 #ifdef OLD_DIHED
6320           sgdih=-gdih(k)*dih_diff(k)*sigma_dih(k,i)  ! waga_angle rmvd
6321 #else
6322           sgdih=-gdih(k)*dsin(dih_diff(k))*sigma_dih(k,i)  ! waga_angle rmvd
6323 #endif
6324 c         sgdih=-gdih(k)*dih_diff(k)*sigma_dih(k,i)*waga_angle
6325           sum_sgdih=sum_sgdih+sgdih
6326         enddo
6327 c       grad_dih3=sum_sgdih/sum_gdih
6328         grad_dih3=waga_homology(iset)*waga_angle*sum_sgdih/sum_gdih
6329
6330 c      write(iout,*)i,k,gdih,sgdih,beta(i+1,i+2,i+3,i+4),grad_dih3
6331 ccc      write(iout,747) "GRAD_KAT_1", i, nphi, icg, grad_dih3,
6332 ccc     & gloc(nphi+i-3,icg)
6333         gloc(i-3,icg)=gloc(i-3,icg)+grad_dih3
6334 c        if (i.eq.25) then
6335 c        write(iout,*) "i",i,"icg",icg,"gloc(",i,icg,")",gloc(i,icg)
6336 c        endif
6337 ccc      write(iout,747) "GRAD_KAT_2", i, nphi, icg, grad_dih3,
6338 ccc     & gloc(nphi+i-3,icg)
6339
6340       enddo ! i-loop for dih
6341 #ifdef DEBUG
6342       write(iout,*) "------- dih restrs end -------"
6343 #endif
6344
6345 c Pseudo-energy and gradient for theta angle restraints from
6346 c homology templates
6347 c FP 01/15 - inserted from econstr_local_test.F, loop structure
6348 c adapted
6349
6350 c
6351 c     For constr_homology reference structures (FP)
6352 c     
6353 c     Uconst_back_tot=0.0d0
6354       Eval=0.0d0
6355       Erot=0.0d0
6356 c     Econstr_back legacy
6357       do i=1,nres
6358 c     do i=ithet_start,ithet_end
6359        dutheta(i)=0.0d0
6360 c     enddo
6361 c     do i=loc_start,loc_end
6362         do j=1,3
6363           duscdiff(j,i)=0.0d0
6364           duscdiffx(j,i)=0.0d0
6365         enddo
6366       enddo
6367 c
6368 c     do iref=1,nref
6369 c     write (iout,*) "ithet_start =",ithet_start,"ithet_end =",ithet_end
6370 c     write (iout,*) "waga_theta",waga_theta
6371       if (waga_theta.gt.0.0d0) then
6372 #ifdef DEBUG
6373       write (iout,*) "usampl",usampl
6374       write(iout,*) "------- theta restrs start -------"
6375 c     do i=ithet_start,ithet_end
6376 c       write (iout,*) "gloc_init(",nphi+i,icg,")",gloc(nphi+i,icg)
6377 c     enddo
6378 #endif
6379 c     write (iout,*) "maxres",maxres,"nres",nres
6380
6381       do i=ithet_start,ithet_end
6382 c
6383 c     do i=1,nfrag_back
6384 c       ii = ifrag_back(2,i,iset)-ifrag_back(1,i,iset)
6385 c
6386 c Deviation of theta angles wrt constr_homology ref structures
6387 c
6388         utheta_i=0.0d0 ! argument of Gaussian for single k
6389         gutheta_i=0.0d0 ! Sum of Gaussians over constr_homology ref structures
6390 c       do j=ifrag_back(1,i,iset)+2,ifrag_back(2,i,iset) ! original loop
6391 c       over residues in a fragment
6392 c       write (iout,*) "theta(",i,")=",theta(i)
6393         do k=1,constr_homology
6394 c
6395 c         dtheta_i=theta(j)-thetaref(j,iref)
6396 c         dtheta_i=thetaref(k,i)-theta(i) ! original form without indexing
6397           theta_diff(k)=thetatpl(k,i)-theta(i)
6398 c
6399           utheta_i=-0.5d0*theta_diff(k)**2*sigma_theta(k,i) ! waga_theta rmvd from Gaussian argument
6400 c         utheta_i=-0.5d0*waga_theta*theta_diff(k)**2*sigma_theta(k,i) ! waga_theta?
6401           gtheta(k)=dexp(utheta_i) ! + min_utheta_i?
6402           gutheta_i=gutheta_i+gtheta(k)  ! Sum of Gaussians (pk)
6403 c          write (iout,*) "i",i," k",k," sigma_theta",sigma_theta(k,i),
6404 c     &     " gtheta",gtheta(k)
6405 c         Gradient for single Gaussian restraint in subr Econstr_back
6406 c         dutheta(j-2)=dutheta(j-2)+wfrag_back(1,i,iset)*dtheta_i/(ii-1)
6407 c
6408         enddo
6409 c       write (iout,*) "gtheta",(gtheta(k),k=1,constr_homology) ! exps
6410 c       write (iout,*) "i",i," gutheta_i",gutheta_i ! sum of exps
6411
6412 c
6413 c         Gradient for multiple Gaussian restraint
6414         sum_gtheta=gutheta_i
6415         sum_sgtheta=0.0d0
6416         do k=1,constr_homology
6417 c        New generalized expr for multiple Gaussian from Econstr_back
6418          sgtheta=-gtheta(k)*theta_diff(k)*sigma_theta(k,i) ! waga_theta rmvd
6419 c
6420 c        sgtheta=-gtheta(k)*theta_diff(k)*sigma_theta(k,i)*waga_theta ! right functional form?
6421           sum_sgtheta=sum_sgtheta+sgtheta ! cum variable
6422         enddo
6423 c       Final value of gradient using same var as in Econstr_back
6424         gloc(nphi+i-2,icg)=gloc(nphi+i-2,icg)
6425      &      +sum_sgtheta/sum_gtheta*waga_theta
6426      &               *waga_homology(iset)
6427 c        dutheta(i-2)=sum_sgtheta/sum_gtheta*waga_theta
6428 c     &               *waga_homology(iset)
6429 c       dutheta(i)=sum_sgtheta/sum_gtheta
6430 c
6431 c       Uconst_back=Uconst_back+waga_theta*utheta(i) ! waga_theta added as weight
6432         Eval=Eval-dLOG(gutheta_i/constr_homology)
6433 c       write (iout,*) "utheta(",i,")=",utheta(i) ! -ln of sum of exps
6434 c       write (iout,*) "Uconst_back",Uconst_back ! sum of -ln-s
6435 c       Uconst_back=Uconst_back+utheta(i)
6436       enddo ! (i-loop for theta)
6437 #ifdef DEBUG
6438       write(iout,*) "------- theta restrs end -------"
6439 #endif
6440       endif
6441 c
6442 c Deviation of local SC geometry
6443 c
6444 c Separation of two i-loops (instructed by AL - 11/3/2014)
6445 c
6446 c     write (iout,*) "loc_start =",loc_start,"loc_end =",loc_end
6447 c     write (iout,*) "waga_d",waga_d
6448
6449 #ifdef DEBUG
6450       write(iout,*) "------- SC restrs start -------"
6451       write (iout,*) "Initial duscdiff,duscdiffx"
6452       do i=loc_start,loc_end
6453         write (iout,*) i,(duscdiff(jik,i),jik=1,3),
6454      &                 (duscdiffx(jik,i),jik=1,3)
6455       enddo
6456 #endif
6457       do i=loc_start,loc_end
6458         usc_diff_i=0.0d0 ! argument of Gaussian for single k
6459         guscdiff(i)=0.0d0 ! Sum of Gaussians over constr_homology ref structures
6460 c       do j=ifrag_back(1,i,iset)+1,ifrag_back(2,i,iset)-1 ! Econstr_back legacy
6461 c       write(iout,*) "xxtab, yytab, zztab"
6462 c       write(iout,'(i5,3f8.2)') i,xxtab(i),yytab(i),zztab(i)
6463         do k=1,constr_homology
6464 c
6465           dxx=-xxtpl(k,i)+xxtab(i) ! Diff b/w x component of ith SC vector in model and kth ref str?
6466 c                                    Original sign inverted for calc of gradients (s. Econstr_back)
6467           dyy=-yytpl(k,i)+yytab(i) ! ibid y
6468           dzz=-zztpl(k,i)+zztab(i) ! ibid z
6469 c         write(iout,*) "dxx, dyy, dzz"
6470 c         write(iout,'(2i5,3f8.2)') k,i,dxx,dyy,dzz
6471 c
6472           usc_diff_i=-0.5d0*(dxx**2+dyy**2+dzz**2)*sigma_d(k,i)  ! waga_d rmvd from Gaussian argument
6473 c         usc_diff(i)=-0.5d0*waga_d*(dxx**2+dyy**2+dzz**2)*sigma_d(k,i) ! waga_d?
6474 c         uscdiffk(k)=usc_diff(i)
6475           guscdiff2(k)=dexp(usc_diff_i) ! without min_scdiff
6476 c          write(iout,*) "i",i," k",k," sigma_d",sigma_d(k,i),
6477 c     &       " guscdiff2",guscdiff2(k)
6478           guscdiff(i)=guscdiff(i)+guscdiff2(k)  !Sum of Gaussians (pk)
6479 c          write (iout,'(i5,6f10.5)') j,xxtab(j),yytab(j),zztab(j),
6480 c     &      xxref(j),yyref(j),zzref(j)
6481         enddo
6482 c
6483 c       Gradient 
6484 c
6485 c       Generalized expression for multiple Gaussian acc to that for a single 
6486 c       Gaussian in Econstr_back as instructed by AL (FP - 03/11/2014)
6487 c
6488 c       Original implementation
6489 c       sum_guscdiff=guscdiff(i)
6490 c
6491 c       sum_sguscdiff=0.0d0
6492 c       do k=1,constr_homology
6493 c          sguscdiff=-guscdiff2(k)*dscdiff(k)*sigma_d(k,i)*waga_d !waga_d? 
6494 c          sguscdiff=-guscdiff3(k)*dscdiff(k)*sigma_d(k,i)*waga_d ! w min_uscdiff
6495 c          sum_sguscdiff=sum_sguscdiff+sguscdiff
6496 c       enddo
6497 c
6498 c       Implementation of new expressions for gradient (Jan. 2015)
6499 c
6500 c       grad_uscdiff=sum_sguscdiff/(sum_guscdiff*dtab) !?
6501         do k=1,constr_homology 
6502 c
6503 c       New calculation of dxx, dyy, and dzz corrected by AL (07/11), was missing and wrong
6504 c       before. Now the drivatives should be correct
6505 c
6506           dxx=-xxtpl(k,i)+xxtab(i) ! Diff b/w x component of ith SC vector in model and kth ref str?
6507 c                                  Original sign inverted for calc of gradients (s. Econstr_back)
6508           dyy=-yytpl(k,i)+yytab(i) ! ibid y
6509           dzz=-zztpl(k,i)+zztab(i) ! ibid z
6510 c
6511 c         New implementation
6512 c
6513           sum_guscdiff=guscdiff2(k)*!(dsqrt(dxx*dxx+dyy*dyy+dzz*dzz))* -> wrong!
6514      &                 sigma_d(k,i) ! for the grad wrt r' 
6515 c         sum_sguscdiff=sum_sguscdiff+sum_guscdiff
6516 c
6517 c
6518 c        New implementation
6519          sum_guscdiff = waga_homology(iset)*waga_d*sum_guscdiff
6520          do jik=1,3
6521             duscdiff(jik,i-1)=duscdiff(jik,i-1)+
6522      &      sum_guscdiff*(dXX_C1tab(jik,i)*dxx+
6523      &      dYY_C1tab(jik,i)*dyy+dZZ_C1tab(jik,i)*dzz)/guscdiff(i)
6524             duscdiff(jik,i)=duscdiff(jik,i)+
6525      &      sum_guscdiff*(dXX_Ctab(jik,i)*dxx+
6526      &      dYY_Ctab(jik,i)*dyy+dZZ_Ctab(jik,i)*dzz)/guscdiff(i)
6527             duscdiffx(jik,i)=duscdiffx(jik,i)+
6528      &      sum_guscdiff*(dXX_XYZtab(jik,i)*dxx+
6529      &      dYY_XYZtab(jik,i)*dyy+dZZ_XYZtab(jik,i)*dzz)/guscdiff(i)
6530 c
6531 #ifdef DEBUG
6532              write(iout,*) "jik",jik,"i",i
6533              write(iout,*) "dxx, dyy, dzz"
6534              write(iout,'(2i5,3f8.2)') k,i,dxx,dyy,dzz
6535              write(iout,*) "guscdiff2(",k,")",guscdiff2(k)
6536 c            write(iout,*) "sum_sguscdiff",sum_sguscdiff
6537 cc           write(iout,*) "dXX_Ctab(",jik,i,")",dXX_Ctab(jik,i)
6538 c            write(iout,*) "dYY_Ctab(",jik,i,")",dYY_Ctab(jik,i)
6539 c            write(iout,*) "dZZ_Ctab(",jik,i,")",dZZ_Ctab(jik,i)
6540 c            write(iout,*) "dXX_C1tab(",jik,i,")",dXX_C1tab(jik,i)
6541 c            write(iout,*) "dYY_C1tab(",jik,i,")",dYY_C1tab(jik,i)
6542 c            write(iout,*) "dZZ_C1tab(",jik,i,")",dZZ_C1tab(jik,i)
6543 c            write(iout,*) "dXX_XYZtab(",jik,i,")",dXX_XYZtab(jik,i)
6544 c            write(iout,*) "dYY_XYZtab(",jik,i,")",dYY_XYZtab(jik,i)
6545 c            write(iout,*) "dZZ_XYZtab(",jik,i,")",dZZ_XYZtab(jik,i)
6546 c            write(iout,*) "duscdiff(",jik,i-1,")",duscdiff(jik,i-1)
6547 c            write(iout,*) "duscdiff(",jik,i,")",duscdiff(jik,i)
6548 c            write(iout,*) "duscdiffx(",jik,i,")",duscdiffx(jik,i)
6549 c            endif
6550 #endif
6551          enddo
6552         enddo
6553 c
6554 c       uscdiff(i)=-dLOG(guscdiff(i)/(ii-1))      ! Weighting by (ii-1) required?
6555 c        usc_diff(i)=-dLOG(guscdiff(i)/constr_homology) ! + min_uscdiff ?
6556 c
6557 c        write (iout,*) i," uscdiff",uscdiff(i)
6558 c
6559 c Put together deviations from local geometry
6560
6561 c       Uconst_back=Uconst_back+wfrag_back(1,i,iset)*utheta(i)+
6562 c      &            wfrag_back(3,i,iset)*uscdiff(i)
6563         Erot=Erot-dLOG(guscdiff(i)/constr_homology)
6564 c       write (iout,*) "usc_diff(",i,")=",usc_diff(i) ! -ln of sum of exps
6565 c       write (iout,*) "Uconst_back",Uconst_back ! cum sum of -ln-s
6566 c       Uconst_back=Uconst_back+usc_diff(i)
6567 c
6568 c     Gradient of multiple Gaussian restraint (FP - 04/11/2014 - right?)
6569 c
6570 c     New implment: multiplied by sum_sguscdiff
6571 c
6572
6573       enddo ! (i-loop for dscdiff)
6574
6575 c      endif
6576
6577 #ifdef DEBUG
6578       write(iout,*) "------- SC restrs end -------"
6579         write (iout,*) "------ After SC loop in e_modeller ------"
6580         do i=loc_start,loc_end
6581          write (iout,*) "i",i," gradc",(gradc(j,i,icg),j=1,3)
6582          write (iout,*) "i",i," gradx",(gradx(j,i,icg),j=1,3)
6583         enddo
6584       if (waga_theta.eq.1.0d0) then
6585       write (iout,*) "in e_modeller after SC restr end: dutheta"
6586       do i=ithet_start,ithet_end
6587         write (iout,*) i,dutheta(i)
6588       enddo
6589       endif
6590       if (waga_d.eq.1.0d0) then
6591       write (iout,*) "e_modeller after SC loop: duscdiff/x"
6592       do i=1,nres
6593         write (iout,*) i,(duscdiff(j,i),j=1,3)
6594         write (iout,*) i,(duscdiffx(j,i),j=1,3)
6595       enddo
6596       endif
6597 #endif
6598
6599 c Total energy from homology restraints
6600 #ifdef DEBUG
6601       write (iout,*) "odleg",odleg," kat",kat
6602 #endif
6603 c
6604 c Addition of energy of theta angle and SC local geom over constr_homologs ref strs
6605 c
6606 c     ehomology_constr=odleg+kat
6607 c
6608 c     For Lorentzian-type Urestr
6609 c
6610
6611       if (waga_dist.ge.0.0d0) then
6612 c
6613 c          For Gaussian-type Urestr
6614 c
6615         ehomology_constr=(waga_dist*odleg+waga_angle*kat+
6616      &              waga_theta*Eval+waga_d*Erot)*waga_homology(iset)
6617 c     write (iout,*) "ehomology_constr=",ehomology_constr
6618       else
6619 c
6620 c          For Lorentzian-type Urestr
6621 c  
6622         ehomology_constr=(-waga_dist*odleg+waga_angle*kat+
6623      &              waga_theta*Eval+waga_d*Erot)*waga_homology(iset)
6624 c     write (iout,*) "ehomology_constr=",ehomology_constr
6625       endif
6626 #ifdef DEBUG
6627       write (iout,*) "odleg",waga_dist,odleg," kat",waga_angle,kat,
6628      & "Eval",waga_theta,eval,
6629      &   "Erot",waga_d,Erot
6630       write (iout,*) "ehomology_constr",ehomology_constr
6631 #endif
6632       return
6633 c
6634 c FP 01/15 end
6635 c
6636   748 format(a8,f12.3,a6,f12.3,a7,f12.3)
6637   747 format(a12,i4,i4,i4,f8.3,f8.3)
6638   746 format(a12,i4,i4,i4,f8.3,f8.3,f8.3)
6639   778 format(a7,1X,f10.3,1X,a4,1X,f10.3,1X,a5,1X,f10.3)
6640   779 format(i3,1X,i3,1X,i2,1X,a7,1X,f7.3,1X,a7,1X,f7.3,1X,a13,1X,
6641      &       f7.3,1X,a17,1X,f9.3,1X,a10,1X,f8.3,1X,a10,1X,f8.3)
6642       end
6643
6644 c------------------------------------------------------------------------------
6645       subroutine etor_d(etors_d)
6646 C 6/23/01 Compute double torsional energy
6647       implicit real*8 (a-h,o-z)
6648       include 'DIMENSIONS'
6649       include 'COMMON.VAR'
6650       include 'COMMON.GEO'
6651       include 'COMMON.LOCAL'
6652       include 'COMMON.TORSION'
6653       include 'COMMON.INTERACT'
6654       include 'COMMON.DERIV'
6655       include 'COMMON.CHAIN'
6656       include 'COMMON.NAMES'
6657       include 'COMMON.IOUNITS'
6658       include 'COMMON.FFIELD'
6659       include 'COMMON.TORCNSTR'
6660       include 'COMMON.CONTROL'
6661       logical lprn
6662 C Set lprn=.true. for debugging
6663       lprn=.false.
6664 c     lprn=.true.
6665       etors_d=0.0D0
6666       do i=iphid_start,iphid_end
6667         etors_d_ii=0.0D0
6668         itori=itortyp(itype(i-2))
6669         itori1=itortyp(itype(i-1))
6670         itori2=itortyp(itype(i))
6671         phii=phi(i)
6672         phii1=phi(i+1)
6673         gloci1=0.0D0
6674         gloci2=0.0D0
6675         do j=1,ntermd_1(itori,itori1,itori2)
6676           v1cij=v1c(1,j,itori,itori1,itori2)
6677           v1sij=v1s(1,j,itori,itori1,itori2)
6678           v2cij=v1c(2,j,itori,itori1,itori2)
6679           v2sij=v1s(2,j,itori,itori1,itori2)
6680           cosphi1=dcos(j*phii)
6681           sinphi1=dsin(j*phii)
6682           cosphi2=dcos(j*phii1)
6683           sinphi2=dsin(j*phii1)
6684           etors_d=etors_d+v1cij*cosphi1+v1sij*sinphi1+
6685      &     v2cij*cosphi2+v2sij*sinphi2
6686           if (energy_dec) etors_d_ii=etors_d_ii+
6687      &     v1cij*cosphi1+v1sij*sinphi1+v2cij*cosphi2+v2sij*sinphi2
6688           gloci1=gloci1+j*(v1sij*cosphi1-v1cij*sinphi1)
6689           gloci2=gloci2+j*(v2sij*cosphi2-v2cij*sinphi2)
6690         enddo
6691         do k=2,ntermd_2(itori,itori1,itori2)
6692           do l=1,k-1
6693             v1cdij = v2c(k,l,itori,itori1,itori2)
6694             v2cdij = v2c(l,k,itori,itori1,itori2)
6695             v1sdij = v2s(k,l,itori,itori1,itori2)
6696             v2sdij = v2s(l,k,itori,itori1,itori2)
6697             cosphi1p2=dcos(l*phii+(k-l)*phii1)
6698             cosphi1m2=dcos(l*phii-(k-l)*phii1)
6699             sinphi1p2=dsin(l*phii+(k-l)*phii1)
6700             sinphi1m2=dsin(l*phii-(k-l)*phii1)
6701             etors_d=etors_d+v1cdij*cosphi1p2+v2cdij*cosphi1m2+
6702      &        v1sdij*sinphi1p2+v2sdij*sinphi1m2
6703             if (energy_dec) etors_d_ii=etors_d_ii+
6704      &        v1cdij*cosphi1p2+v2cdij*cosphi1m2+
6705      &        v1sdij*sinphi1p2+v2sdij*sinphi1m2
6706             gloci1=gloci1+l*(v1sdij*cosphi1p2+v2sdij*cosphi1m2
6707      &        -v1cdij*sinphi1p2-v2cdij*sinphi1m2)
6708             gloci2=gloci2+(k-l)*(v1sdij*cosphi1p2-v2sdij*cosphi1m2
6709      &        -v1cdij*sinphi1p2+v2cdij*sinphi1m2) 
6710           enddo
6711         enddo
6712         if (energy_dec) write (iout,'(a6,i5,0pf7.3)')
6713      &        'etor_d',i,etors_d_ii
6714         gloc(i-3,icg)=gloc(i-3,icg)+wtor_d*gloci1
6715         gloc(i-2,icg)=gloc(i-2,icg)+wtor_d*gloci2
6716 c        write (iout,*) "gloci", gloc(i-3,icg)
6717       enddo
6718       return
6719       end
6720 #endif
6721 c------------------------------------------------------------------------------
6722       subroutine eback_sc_corr(esccor)
6723 c 7/21/2007 Correlations between the backbone-local and side-chain-local
6724 c        conformational states; temporarily implemented as differences
6725 c        between UNRES torsional potentials (dependent on three types of
6726 c        residues) and the torsional potentials dependent on all 20 types
6727 c        of residues computed from AM1  energy surfaces of terminally-blocked
6728 c        amino-acid residues.
6729       implicit real*8 (a-h,o-z)
6730       include 'DIMENSIONS'
6731       include 'COMMON.VAR'
6732       include 'COMMON.GEO'
6733       include 'COMMON.LOCAL'
6734       include 'COMMON.TORSION'
6735       include 'COMMON.SCCOR'
6736       include 'COMMON.INTERACT'
6737       include 'COMMON.DERIV'
6738       include 'COMMON.CHAIN'
6739       include 'COMMON.NAMES'
6740       include 'COMMON.IOUNITS'
6741       include 'COMMON.FFIELD'
6742       include 'COMMON.CONTROL'
6743       logical lprn
6744 C Set lprn=.true. for debugging
6745       lprn=.false.
6746 c      lprn=.true.
6747 c      write (iout,*) "EBACK_SC_COR",iphi_start,iphi_end,nterm_sccor
6748       esccor=0.0D0
6749       do i=itau_start,itau_end
6750         esccor_ii=0.0D0
6751         if ((itype(i-2).eq.ntyp1).or.(itype(i-1).eq.ntyp1)) cycle
6752         isccori=isccortyp(itype(i-2))
6753         isccori1=isccortyp(itype(i-1))
6754         phii=phi(i)
6755 cccc  Added 9 May 2012
6756 cc Tauangle is torsional engle depending on the value of first digit 
6757 c(see comment below)
6758 cc Omicron is flat angle depending on the value of first digit 
6759 c(see comment below)
6760
6761         
6762         do intertyp=1,3 !intertyp
6763 cc Added 09 May 2012 (Adasko)
6764 cc  Intertyp means interaction type of backbone mainchain correlation: 
6765 c   1 = SC...Ca...Ca...Ca
6766 c   2 = Ca...Ca...Ca...SC
6767 c   3 = SC...Ca...Ca...SCi
6768         gloci=0.0D0
6769         if (((intertyp.eq.3).and.((itype(i-2).eq.10).or.
6770      &      (itype(i-1).eq.10).or.(itype(i-2).eq.21).or.
6771      &      (itype(i-1).eq.21)))
6772      &    .or. ((intertyp.eq.1).and.((itype(i-2).eq.10)
6773      &     .or.(itype(i-2).eq.21)))
6774      &    .or.((intertyp.eq.2).and.((itype(i-1).eq.10).or.
6775      &      (itype(i-1).eq.21)))) cycle  
6776         if ((intertyp.eq.2).and.(i.eq.4).and.(itype(1).eq.21)) cycle
6777         if ((intertyp.eq.1).and.(i.eq.nres).and.(itype(nres).eq.21))
6778      & cycle
6779         do j=1,nterm_sccor(isccori,isccori1)
6780           v1ij=v1sccor(j,intertyp,isccori,isccori1)
6781           v2ij=v2sccor(j,intertyp,isccori,isccori1)
6782           cosphi=dcos(j*tauangle(intertyp,i))
6783           sinphi=dsin(j*tauangle(intertyp,i))
6784           esccor=esccor+v1ij*cosphi+v2ij*sinphi
6785           gloci=gloci+j*(v2ij*cosphi-v1ij*sinphi)
6786         enddo
6787         gloc_sc(intertyp,i-3,icg)=gloc_sc(intertyp,i-3,icg)+wsccor*gloci
6788 c        write (iout,*) "WTF",intertyp,i,itype(i),v1ij*cosphi+v2ij*sinphi
6789 c     &gloc_sc(intertyp,i-3,icg)
6790         if (lprn)
6791      &  write (iout,'(2(a3,2x,i3,2x),2i3,6f8.3/26x,6f8.3/)')
6792      &  restyp(itype(i-2)),i-2,restyp(itype(i-1)),i-1,itori,itori1,
6793      &  (v1sccor(j,intertyp,itori,itori1),j=1,6)
6794      & ,(v2sccor(j,intertyp,itori,itori1),j=1,6)
6795         gsccor_loc(i-3)=gsccor_loc(i-3)+gloci
6796        enddo !intertyp
6797       enddo
6798 c        do i=1,nres
6799 c        write (iout,*) "W@T@F",  gloc_sc(1,i,icg),gloc(i,icg)
6800 c        enddo
6801       return
6802       end
6803 c----------------------------------------------------------------------------
6804       subroutine multibody(ecorr)
6805 C This subroutine calculates multi-body contributions to energy following
6806 C the idea of Skolnick et al. If side chains I and J make a contact and
6807 C at the same time side chains I+1 and J+1 make a contact, an extra 
6808 C contribution equal to sqrt(eps(i,j)*eps(i+1,j+1)) is added.
6809       implicit real*8 (a-h,o-z)
6810       include 'DIMENSIONS'
6811       include 'COMMON.IOUNITS'
6812       include 'COMMON.DERIV'
6813       include 'COMMON.INTERACT'
6814       include 'COMMON.CONTACTS'
6815       double precision gx(3),gx1(3)
6816       logical lprn
6817
6818 C Set lprn=.true. for debugging
6819       lprn=.false.
6820
6821       if (lprn) then
6822         write (iout,'(a)') 'Contact function values:'
6823         do i=nnt,nct-2
6824           write (iout,'(i2,20(1x,i2,f10.5))') 
6825      &        i,(jcont(j,i),facont(j,i),j=1,num_cont(i))
6826         enddo
6827       endif
6828       ecorr=0.0D0
6829       do i=nnt,nct
6830         do j=1,3
6831           gradcorr(j,i)=0.0D0
6832           gradxorr(j,i)=0.0D0
6833         enddo
6834       enddo
6835       do i=nnt,nct-2
6836
6837         DO ISHIFT = 3,4
6838
6839         i1=i+ishift
6840         num_conti=num_cont(i)
6841         num_conti1=num_cont(i1)
6842         do jj=1,num_conti
6843           j=jcont(jj,i)
6844           do kk=1,num_conti1
6845             j1=jcont(kk,i1)
6846             if (j1.eq.j+ishift .or. j1.eq.j-ishift) then
6847 cd          write(iout,*)'i=',i,' j=',j,' i1=',i1,' j1=',j1,
6848 cd   &                   ' ishift=',ishift
6849 C Contacts I--J and I+ISHIFT--J+-ISHIFT1 occur simultaneously. 
6850 C The system gains extra energy.
6851               ecorr=ecorr+esccorr(i,j,i1,j1,jj,kk)
6852             endif   ! j1==j+-ishift
6853           enddo     ! kk  
6854         enddo       ! jj
6855
6856         ENDDO ! ISHIFT
6857
6858       enddo         ! i
6859       return
6860       end
6861 c------------------------------------------------------------------------------
6862       double precision function esccorr(i,j,k,l,jj,kk)
6863       implicit real*8 (a-h,o-z)
6864       include 'DIMENSIONS'
6865       include 'COMMON.IOUNITS'
6866       include 'COMMON.DERIV'
6867       include 'COMMON.INTERACT'
6868       include 'COMMON.CONTACTS'
6869       double precision gx(3),gx1(3)
6870       logical lprn
6871       lprn=.false.
6872       eij=facont(jj,i)
6873       ekl=facont(kk,k)
6874 cd    write (iout,'(4i5,3f10.5)') i,j,k,l,eij,ekl,-eij*ekl
6875 C Calculate the multi-body contribution to energy.
6876 C Calculate multi-body contributions to the gradient.
6877 cd    write (iout,'(2(2i3,3f10.5))')i,j,(gacont(m,jj,i),m=1,3),
6878 cd   & k,l,(gacont(m,kk,k),m=1,3)
6879       do m=1,3
6880         gx(m) =ekl*gacont(m,jj,i)
6881         gx1(m)=eij*gacont(m,kk,k)
6882         gradxorr(m,i)=gradxorr(m,i)-gx(m)
6883         gradxorr(m,j)=gradxorr(m,j)+gx(m)
6884         gradxorr(m,k)=gradxorr(m,k)-gx1(m)
6885         gradxorr(m,l)=gradxorr(m,l)+gx1(m)
6886       enddo
6887       do m=i,j-1
6888         do ll=1,3
6889           gradcorr(ll,m)=gradcorr(ll,m)+gx(ll)
6890         enddo
6891       enddo
6892       do m=k,l-1
6893         do ll=1,3
6894           gradcorr(ll,m)=gradcorr(ll,m)+gx1(ll)
6895         enddo
6896       enddo 
6897       esccorr=-eij*ekl
6898       return
6899       end
6900 c------------------------------------------------------------------------------
6901       subroutine multibody_hb(ecorr,ecorr5,ecorr6,n_corr,n_corr1)
6902 C This subroutine calculates multi-body contributions to hydrogen-bonding 
6903       implicit real*8 (a-h,o-z)
6904       include 'DIMENSIONS'
6905       include 'COMMON.IOUNITS'
6906 #ifdef MPI
6907       include "mpif.h"
6908       parameter (max_cont=maxconts)
6909       parameter (max_dim=26)
6910       integer source,CorrelType,CorrelID,CorrelType1,CorrelID1,Error
6911       double precision zapas(max_dim,maxconts,max_fg_procs),
6912      &  zapas_recv(max_dim,maxconts,max_fg_procs)
6913       common /przechowalnia/ zapas
6914       integer status(MPI_STATUS_SIZE),req(maxconts*2),
6915      &  status_array(MPI_STATUS_SIZE,maxconts*2)
6916 #endif
6917       include 'COMMON.SETUP'
6918       include 'COMMON.FFIELD'
6919       include 'COMMON.DERIV'
6920       include 'COMMON.INTERACT'
6921       include 'COMMON.CONTACTS'
6922       include 'COMMON.CONTROL'
6923       include 'COMMON.LOCAL'
6924       double precision gx(3),gx1(3),time00
6925       logical lprn,ldone
6926
6927 C Set lprn=.true. for debugging
6928       lprn=.false.
6929 #ifdef MPI
6930       n_corr=0
6931       n_corr1=0
6932       if (nfgtasks.le.1) goto 30
6933       if (lprn) then
6934         write (iout,'(a)') 'Contact function values before RECEIVE:'
6935         do i=nnt,nct-2
6936           write (iout,'(2i3,50(1x,i2,f5.2))') 
6937      &    i,num_cont_hb(i),(jcont_hb(j,i),facont_hb(j,i),
6938      &    j=1,num_cont_hb(i))
6939         enddo
6940       endif
6941       call flush(iout)
6942       do i=1,ntask_cont_from
6943         ncont_recv(i)=0
6944       enddo
6945       do i=1,ntask_cont_to
6946         ncont_sent(i)=0
6947       enddo
6948 c      write (iout,*) "ntask_cont_from",ntask_cont_from," ntask_cont_to",
6949 c     & ntask_cont_to
6950 C Make the list of contacts to send to send to other procesors
6951 c      write (iout,*) "limits",max0(iturn4_end-1,iatel_s),iturn3_end
6952 c      call flush(iout)
6953       do i=iturn3_start,iturn3_end
6954 c        write (iout,*) "make contact list turn3",i," num_cont",
6955 c     &    num_cont_hb(i)
6956         call add_hb_contact(i,i+2,iturn3_sent_local(1,i))
6957       enddo
6958       do i=iturn4_start,iturn4_end
6959 c        write (iout,*) "make contact list turn4",i," num_cont",
6960 c     &   num_cont_hb(i)
6961         call add_hb_contact(i,i+3,iturn4_sent_local(1,i))
6962       enddo
6963       do ii=1,nat_sent
6964         i=iat_sent(ii)
6965 c        write (iout,*) "make contact list longrange",i,ii," num_cont",
6966 c     &    num_cont_hb(i)
6967         do j=1,num_cont_hb(i)
6968         do k=1,4
6969           jjc=jcont_hb(j,i)
6970           iproc=iint_sent_local(k,jjc,ii)
6971 c          write (iout,*) "i",i," j",j," k",k," jjc",jjc," iproc",iproc
6972           if (iproc.gt.0) then
6973             ncont_sent(iproc)=ncont_sent(iproc)+1
6974             nn=ncont_sent(iproc)
6975             zapas(1,nn,iproc)=i
6976             zapas(2,nn,iproc)=jjc
6977             zapas(3,nn,iproc)=facont_hb(j,i)
6978             zapas(4,nn,iproc)=ees0p(j,i)
6979             zapas(5,nn,iproc)=ees0m(j,i)
6980             zapas(6,nn,iproc)=gacont_hbr(1,j,i)
6981             zapas(7,nn,iproc)=gacont_hbr(2,j,i)
6982             zapas(8,nn,iproc)=gacont_hbr(3,j,i)
6983             zapas(9,nn,iproc)=gacontm_hb1(1,j,i)
6984             zapas(10,nn,iproc)=gacontm_hb1(2,j,i)
6985             zapas(11,nn,iproc)=gacontm_hb1(3,j,i)
6986             zapas(12,nn,iproc)=gacontp_hb1(1,j,i)
6987             zapas(13,nn,iproc)=gacontp_hb1(2,j,i)
6988             zapas(14,nn,iproc)=gacontp_hb1(3,j,i)
6989             zapas(15,nn,iproc)=gacontm_hb2(1,j,i)
6990             zapas(16,nn,iproc)=gacontm_hb2(2,j,i)
6991             zapas(17,nn,iproc)=gacontm_hb2(3,j,i)
6992             zapas(18,nn,iproc)=gacontp_hb2(1,j,i)
6993             zapas(19,nn,iproc)=gacontp_hb2(2,j,i)
6994             zapas(20,nn,iproc)=gacontp_hb2(3,j,i)
6995             zapas(21,nn,iproc)=gacontm_hb3(1,j,i)
6996             zapas(22,nn,iproc)=gacontm_hb3(2,j,i)
6997             zapas(23,nn,iproc)=gacontm_hb3(3,j,i)
6998             zapas(24,nn,iproc)=gacontp_hb3(1,j,i)
6999             zapas(25,nn,iproc)=gacontp_hb3(2,j,i)
7000             zapas(26,nn,iproc)=gacontp_hb3(3,j,i)
7001           endif
7002         enddo
7003         enddo
7004       enddo
7005       if (lprn) then
7006       write (iout,*) 
7007      &  "Numbers of contacts to be sent to other processors",
7008      &  (ncont_sent(i),i=1,ntask_cont_to)
7009       write (iout,*) "Contacts sent"
7010       do ii=1,ntask_cont_to
7011         nn=ncont_sent(ii)
7012         iproc=itask_cont_to(ii)
7013         write (iout,*) nn," contacts to processor",iproc,
7014      &   " of CONT_TO_COMM group"
7015         do i=1,nn
7016           write(iout,'(2f5.0,4f10.5)')(zapas(j,i,ii),j=1,5)
7017         enddo
7018       enddo
7019       call flush(iout)
7020       endif
7021       CorrelType=477
7022       CorrelID=fg_rank+1
7023       CorrelType1=478
7024       CorrelID1=nfgtasks+fg_rank+1
7025       ireq=0
7026 C Receive the numbers of needed contacts from other processors 
7027       do ii=1,ntask_cont_from
7028         iproc=itask_cont_from(ii)
7029         ireq=ireq+1
7030         call MPI_Irecv(ncont_recv(ii),1,MPI_INTEGER,iproc,CorrelType,
7031      &    FG_COMM,req(ireq),IERR)
7032       enddo
7033 c      write (iout,*) "IRECV ended"
7034 c      call flush(iout)
7035 C Send the number of contacts needed by other processors
7036       do ii=1,ntask_cont_to
7037         iproc=itask_cont_to(ii)
7038         ireq=ireq+1
7039         call MPI_Isend(ncont_sent(ii),1,MPI_INTEGER,iproc,CorrelType,
7040      &    FG_COMM,req(ireq),IERR)
7041       enddo
7042 c      write (iout,*) "ISEND ended"
7043 c      write (iout,*) "number of requests (nn)",ireq
7044       call flush(iout)
7045       if (ireq.gt.0) 
7046      &  call MPI_Waitall(ireq,req,status_array,ierr)
7047 c      write (iout,*) 
7048 c     &  "Numbers of contacts to be received from other processors",
7049 c     &  (ncont_recv(i),i=1,ntask_cont_from)
7050 c      call flush(iout)
7051 C Receive contacts
7052       ireq=0
7053       do ii=1,ntask_cont_from
7054         iproc=itask_cont_from(ii)
7055         nn=ncont_recv(ii)
7056 c        write (iout,*) "Receiving",nn," contacts from processor",iproc,
7057 c     &   " of CONT_TO_COMM group"
7058         call flush(iout)
7059         if (nn.gt.0) then
7060           ireq=ireq+1
7061           call MPI_Irecv(zapas_recv(1,1,ii),nn*max_dim,
7062      &    MPI_DOUBLE_PRECISION,iproc,CorrelType1,FG_COMM,req(ireq),IERR)
7063 c          write (iout,*) "ireq,req",ireq,req(ireq)
7064         endif
7065       enddo
7066 C Send the contacts to processors that need them
7067       do ii=1,ntask_cont_to
7068         iproc=itask_cont_to(ii)
7069         nn=ncont_sent(ii)
7070 c        write (iout,*) nn," contacts to processor",iproc,
7071 c     &   " of CONT_TO_COMM group"
7072         if (nn.gt.0) then
7073           ireq=ireq+1 
7074           call MPI_Isend(zapas(1,1,ii),nn*max_dim,MPI_DOUBLE_PRECISION,
7075      &      iproc,CorrelType1,FG_COMM,req(ireq),IERR)
7076 c          write (iout,*) "ireq,req",ireq,req(ireq)
7077 c          do i=1,nn
7078 c            write(iout,'(2f5.0,4f10.5)')(zapas(j,i,ii),j=1,5)
7079 c          enddo
7080         endif  
7081       enddo
7082 c      write (iout,*) "number of requests (contacts)",ireq
7083 c      write (iout,*) "req",(req(i),i=1,4)
7084 c      call flush(iout)
7085       if (ireq.gt.0) 
7086      & call MPI_Waitall(ireq,req,status_array,ierr)
7087       do iii=1,ntask_cont_from
7088         iproc=itask_cont_from(iii)
7089         nn=ncont_recv(iii)
7090         if (lprn) then
7091         write (iout,*) "Received",nn," contacts from processor",iproc,
7092      &   " of CONT_FROM_COMM group"
7093         call flush(iout)
7094         do i=1,nn
7095           write(iout,'(2f5.0,4f10.5)')(zapas_recv(j,i,iii),j=1,5)
7096         enddo
7097         call flush(iout)
7098         endif
7099         do i=1,nn
7100           ii=zapas_recv(1,i,iii)
7101 c Flag the received contacts to prevent double-counting
7102           jj=-zapas_recv(2,i,iii)
7103 c          write (iout,*) "iii",iii," i",i," ii",ii," jj",jj
7104 c          call flush(iout)
7105           nnn=num_cont_hb(ii)+1
7106           num_cont_hb(ii)=nnn
7107           jcont_hb(nnn,ii)=jj
7108           facont_hb(nnn,ii)=zapas_recv(3,i,iii)
7109           ees0p(nnn,ii)=zapas_recv(4,i,iii)
7110           ees0m(nnn,ii)=zapas_recv(5,i,iii)
7111           gacont_hbr(1,nnn,ii)=zapas_recv(6,i,iii)
7112           gacont_hbr(2,nnn,ii)=zapas_recv(7,i,iii)
7113           gacont_hbr(3,nnn,ii)=zapas_recv(8,i,iii)
7114           gacontm_hb1(1,nnn,ii)=zapas_recv(9,i,iii)
7115           gacontm_hb1(2,nnn,ii)=zapas_recv(10,i,iii)
7116           gacontm_hb1(3,nnn,ii)=zapas_recv(11,i,iii)
7117           gacontp_hb1(1,nnn,ii)=zapas_recv(12,i,iii)
7118           gacontp_hb1(2,nnn,ii)=zapas_recv(13,i,iii)
7119           gacontp_hb1(3,nnn,ii)=zapas_recv(14,i,iii)
7120           gacontm_hb2(1,nnn,ii)=zapas_recv(15,i,iii)
7121           gacontm_hb2(2,nnn,ii)=zapas_recv(16,i,iii)
7122           gacontm_hb2(3,nnn,ii)=zapas_recv(17,i,iii)
7123           gacontp_hb2(1,nnn,ii)=zapas_recv(18,i,iii)
7124           gacontp_hb2(2,nnn,ii)=zapas_recv(19,i,iii)
7125           gacontp_hb2(3,nnn,ii)=zapas_recv(20,i,iii)
7126           gacontm_hb3(1,nnn,ii)=zapas_recv(21,i,iii)
7127           gacontm_hb3(2,nnn,ii)=zapas_recv(22,i,iii)
7128           gacontm_hb3(3,nnn,ii)=zapas_recv(23,i,iii)
7129           gacontp_hb3(1,nnn,ii)=zapas_recv(24,i,iii)
7130           gacontp_hb3(2,nnn,ii)=zapas_recv(25,i,iii)
7131           gacontp_hb3(3,nnn,ii)=zapas_recv(26,i,iii)
7132         enddo
7133       enddo
7134       call flush(iout)
7135       if (lprn) then
7136         write (iout,'(a)') 'Contact function values after receive:'
7137         do i=nnt,nct-2
7138           write (iout,'(2i3,50(1x,i3,f5.2))') 
7139      &    i,num_cont_hb(i),(jcont_hb(j,i),facont_hb(j,i),
7140      &    j=1,num_cont_hb(i))
7141         enddo
7142         call flush(iout)
7143       endif
7144    30 continue
7145 #endif
7146       if (lprn) then
7147         write (iout,'(a)') 'Contact function values:'
7148         do i=nnt,nct-2
7149           write (iout,'(2i3,50(1x,i3,f5.2))') 
7150      &    i,num_cont_hb(i),(jcont_hb(j,i),facont_hb(j,i),
7151      &    j=1,num_cont_hb(i))
7152         enddo
7153       endif
7154       ecorr=0.0D0
7155 C Remove the loop below after debugging !!!
7156       do i=nnt,nct
7157         do j=1,3
7158           gradcorr(j,i)=0.0D0
7159           gradxorr(j,i)=0.0D0
7160         enddo
7161       enddo
7162 C Calculate the local-electrostatic correlation terms
7163       do i=min0(iatel_s,iturn4_start),max0(iatel_e,iturn3_end)
7164         i1=i+1
7165         num_conti=num_cont_hb(i)
7166         num_conti1=num_cont_hb(i+1)
7167         do jj=1,num_conti
7168           j=jcont_hb(jj,i)
7169           jp=iabs(j)
7170           do kk=1,num_conti1
7171             j1=jcont_hb(kk,i1)
7172             jp1=iabs(j1)
7173 c            write (iout,*) 'i=',i,' j=',j,' i1=',i1,' j1=',j1,
7174 c     &         ' jj=',jj,' kk=',kk
7175             if ((j.gt.0 .and. j1.gt.0 .or. j.gt.0 .and. j1.lt.0 
7176      &          .or. j.lt.0 .and. j1.gt.0) .and.
7177      &         (jp1.eq.jp+1 .or. jp1.eq.jp-1)) then
7178 C Contacts I-J and (I+1)-(J+1) or (I+1)-(J-1) occur simultaneously. 
7179 C The system gains extra energy.
7180               ecorr=ecorr+ehbcorr(i,jp,i+1,jp1,jj,kk,0.72D0,0.32D0)
7181               if (energy_dec) write (iout,'(a6,2i5,0pf7.3)')
7182      &            'ecorrh',i,j,ehbcorr(i,j,i+1,j1,jj,kk,0.72D0,0.32D0)
7183               n_corr=n_corr+1
7184             else if (j1.eq.j) then
7185 C Contacts I-J and I-(J+1) occur simultaneously. 
7186 C The system loses extra energy.
7187 c             ecorr=ecorr+ehbcorr(i,j,i+1,j,jj,kk,0.60D0,-0.40D0) 
7188             endif
7189           enddo ! kk
7190           do kk=1,num_conti
7191             j1=jcont_hb(kk,i)
7192 c           write (iout,*) 'i=',i,' j=',j,' i1=',i1,' j1=',j1,
7193 c    &         ' jj=',jj,' kk=',kk
7194             if (j1.eq.j+1) then
7195 C Contacts I-J and (I+1)-J occur simultaneously. 
7196 C The system loses extra energy.
7197 c             ecorr=ecorr+ehbcorr(i,j,i,j+1,jj,kk,0.60D0,-0.40D0)
7198             endif ! j1==j+1
7199           enddo ! kk
7200         enddo ! jj
7201       enddo ! i
7202       return
7203       end
7204 c------------------------------------------------------------------------------
7205       subroutine add_hb_contact(ii,jj,itask)
7206       implicit real*8 (a-h,o-z)
7207       include "DIMENSIONS"
7208       include "COMMON.IOUNITS"
7209       integer max_cont
7210       integer max_dim
7211       parameter (max_cont=maxconts)
7212       parameter (max_dim=26)
7213       include "COMMON.CONTACTS"
7214       double precision zapas(max_dim,maxconts,max_fg_procs),
7215      &  zapas_recv(max_dim,maxconts,max_fg_procs)
7216       common /przechowalnia/ zapas
7217       integer i,j,ii,jj,iproc,itask(4),nn
7218 c      write (iout,*) "itask",itask
7219       do i=1,2
7220         iproc=itask(i)
7221         if (iproc.gt.0) then
7222           do j=1,num_cont_hb(ii)
7223             jjc=jcont_hb(j,ii)
7224 c            write (iout,*) "i",ii," j",jj," jjc",jjc
7225             if (jjc.eq.jj) then
7226               ncont_sent(iproc)=ncont_sent(iproc)+1
7227               nn=ncont_sent(iproc)
7228               zapas(1,nn,iproc)=ii
7229               zapas(2,nn,iproc)=jjc
7230               zapas(3,nn,iproc)=facont_hb(j,ii)
7231               zapas(4,nn,iproc)=ees0p(j,ii)
7232               zapas(5,nn,iproc)=ees0m(j,ii)
7233               zapas(6,nn,iproc)=gacont_hbr(1,j,ii)
7234               zapas(7,nn,iproc)=gacont_hbr(2,j,ii)
7235               zapas(8,nn,iproc)=gacont_hbr(3,j,ii)
7236               zapas(9,nn,iproc)=gacontm_hb1(1,j,ii)
7237               zapas(10,nn,iproc)=gacontm_hb1(2,j,ii)
7238               zapas(11,nn,iproc)=gacontm_hb1(3,j,ii)
7239               zapas(12,nn,iproc)=gacontp_hb1(1,j,ii)
7240               zapas(13,nn,iproc)=gacontp_hb1(2,j,ii)
7241               zapas(14,nn,iproc)=gacontp_hb1(3,j,ii)
7242               zapas(15,nn,iproc)=gacontm_hb2(1,j,ii)
7243               zapas(16,nn,iproc)=gacontm_hb2(2,j,ii)
7244               zapas(17,nn,iproc)=gacontm_hb2(3,j,ii)
7245               zapas(18,nn,iproc)=gacontp_hb2(1,j,ii)
7246               zapas(19,nn,iproc)=gacontp_hb2(2,j,ii)
7247               zapas(20,nn,iproc)=gacontp_hb2(3,j,ii)
7248               zapas(21,nn,iproc)=gacontm_hb3(1,j,ii)
7249               zapas(22,nn,iproc)=gacontm_hb3(2,j,ii)
7250               zapas(23,nn,iproc)=gacontm_hb3(3,j,ii)
7251               zapas(24,nn,iproc)=gacontp_hb3(1,j,ii)
7252               zapas(25,nn,iproc)=gacontp_hb3(2,j,ii)
7253               zapas(26,nn,iproc)=gacontp_hb3(3,j,ii)
7254               exit
7255             endif
7256           enddo
7257         endif
7258       enddo
7259       return
7260       end
7261 c------------------------------------------------------------------------------
7262       subroutine multibody_eello(ecorr,ecorr5,ecorr6,eturn6,n_corr,
7263      &  n_corr1)
7264 C This subroutine calculates multi-body contributions to hydrogen-bonding 
7265       implicit real*8 (a-h,o-z)
7266       include 'DIMENSIONS'
7267       include 'COMMON.IOUNITS'
7268 #ifdef MPI
7269       include "mpif.h"
7270       parameter (max_cont=maxconts)
7271       parameter (max_dim=70)
7272       integer source,CorrelType,CorrelID,CorrelType1,CorrelID1,Error
7273       double precision zapas(max_dim,maxconts,max_fg_procs),
7274      &  zapas_recv(max_dim,maxconts,max_fg_procs)
7275       common /przechowalnia/ zapas
7276       integer status(MPI_STATUS_SIZE),req(maxconts*2),
7277      &  status_array(MPI_STATUS_SIZE,maxconts*2)
7278 #endif
7279       include 'COMMON.SETUP'
7280       include 'COMMON.FFIELD'
7281       include 'COMMON.DERIV'
7282       include 'COMMON.LOCAL'
7283       include 'COMMON.INTERACT'
7284       include 'COMMON.CONTACTS'
7285       include 'COMMON.CHAIN'
7286       include 'COMMON.CONTROL'
7287       double precision gx(3),gx1(3)
7288       integer num_cont_hb_old(maxres)
7289       logical lprn,ldone
7290       double precision eello4,eello5,eelo6,eello_turn6
7291       external eello4,eello5,eello6,eello_turn6
7292 C Set lprn=.true. for debugging
7293       lprn=.false.
7294       eturn6=0.0d0
7295 #ifdef MPI
7296       do i=1,nres
7297         num_cont_hb_old(i)=num_cont_hb(i)
7298       enddo
7299       n_corr=0
7300       n_corr1=0
7301       if (nfgtasks.le.1) goto 30
7302       if (lprn) then
7303         write (iout,'(a)') 'Contact function values before RECEIVE:'
7304         do i=nnt,nct-2
7305           write (iout,'(2i3,50(1x,i2,f5.2))') 
7306      &    i,num_cont_hb(i),(jcont_hb(j,i),facont_hb(j,i),
7307      &    j=1,num_cont_hb(i))
7308         enddo
7309       endif
7310       call flush(iout)
7311       do i=1,ntask_cont_from
7312         ncont_recv(i)=0
7313       enddo
7314       do i=1,ntask_cont_to
7315         ncont_sent(i)=0
7316       enddo
7317 c      write (iout,*) "ntask_cont_from",ntask_cont_from," ntask_cont_to",
7318 c     & ntask_cont_to
7319 C Make the list of contacts to send to send to other procesors
7320       do i=iturn3_start,iturn3_end
7321 c        write (iout,*) "make contact list turn3",i," num_cont",
7322 c     &    num_cont_hb(i)
7323         call add_hb_contact_eello(i,i+2,iturn3_sent_local(1,i))
7324       enddo
7325       do i=iturn4_start,iturn4_end
7326 c        write (iout,*) "make contact list turn4",i," num_cont",
7327 c     &   num_cont_hb(i)
7328         call add_hb_contact_eello(i,i+3,iturn4_sent_local(1,i))
7329       enddo
7330       do ii=1,nat_sent
7331         i=iat_sent(ii)
7332 c        write (iout,*) "make contact list longrange",i,ii," num_cont",
7333 c     &    num_cont_hb(i)
7334         do j=1,num_cont_hb(i)
7335         do k=1,4
7336           jjc=jcont_hb(j,i)
7337           iproc=iint_sent_local(k,jjc,ii)
7338 c          write (iout,*) "i",i," j",j," k",k," jjc",jjc," iproc",iproc
7339           if (iproc.ne.0) then
7340             ncont_sent(iproc)=ncont_sent(iproc)+1
7341             nn=ncont_sent(iproc)
7342             zapas(1,nn,iproc)=i
7343             zapas(2,nn,iproc)=jjc
7344             zapas(3,nn,iproc)=d_cont(j,i)
7345             ind=3
7346             do kk=1,3
7347               ind=ind+1
7348               zapas(ind,nn,iproc)=grij_hb_cont(kk,j,i)
7349             enddo
7350             do kk=1,2
7351               do ll=1,2
7352                 ind=ind+1
7353                 zapas(ind,nn,iproc)=a_chuj(ll,kk,j,i)
7354               enddo
7355             enddo
7356             do jj=1,5
7357               do kk=1,3
7358                 do ll=1,2
7359                   do mm=1,2
7360                     ind=ind+1
7361                     zapas(ind,nn,iproc)=a_chuj_der(mm,ll,kk,jj,j,i)
7362                   enddo
7363                 enddo
7364               enddo
7365             enddo
7366           endif
7367         enddo
7368         enddo
7369       enddo
7370       if (lprn) then
7371       write (iout,*) 
7372      &  "Numbers of contacts to be sent to other processors",
7373      &  (ncont_sent(i),i=1,ntask_cont_to)
7374       write (iout,*) "Contacts sent"
7375       do ii=1,ntask_cont_to
7376         nn=ncont_sent(ii)
7377         iproc=itask_cont_to(ii)
7378         write (iout,*) nn," contacts to processor",iproc,
7379      &   " of CONT_TO_COMM group"
7380         do i=1,nn
7381           write(iout,'(2f5.0,10f10.5)')(zapas(j,i,ii),j=1,10)
7382         enddo
7383       enddo
7384       call flush(iout)
7385       endif
7386       CorrelType=477
7387       CorrelID=fg_rank+1
7388       CorrelType1=478
7389       CorrelID1=nfgtasks+fg_rank+1
7390       ireq=0
7391 C Receive the numbers of needed contacts from other processors 
7392       do ii=1,ntask_cont_from
7393         iproc=itask_cont_from(ii)
7394         ireq=ireq+1
7395         call MPI_Irecv(ncont_recv(ii),1,MPI_INTEGER,iproc,CorrelType,
7396      &    FG_COMM,req(ireq),IERR)
7397       enddo
7398 c      write (iout,*) "IRECV ended"
7399 c      call flush(iout)
7400 C Send the number of contacts needed by other processors
7401       do ii=1,ntask_cont_to
7402         iproc=itask_cont_to(ii)
7403         ireq=ireq+1
7404         call MPI_Isend(ncont_sent(ii),1,MPI_INTEGER,iproc,CorrelType,
7405      &    FG_COMM,req(ireq),IERR)
7406       enddo
7407 c      write (iout,*) "ISEND ended"
7408 c      write (iout,*) "number of requests (nn)",ireq
7409       call flush(iout)
7410       if (ireq.gt.0) 
7411      &  call MPI_Waitall(ireq,req,status_array,ierr)
7412 c      write (iout,*) 
7413 c     &  "Numbers of contacts to be received from other processors",
7414 c     &  (ncont_recv(i),i=1,ntask_cont_from)
7415 c      call flush(iout)
7416 C Receive contacts
7417       ireq=0
7418       do ii=1,ntask_cont_from
7419         iproc=itask_cont_from(ii)
7420         nn=ncont_recv(ii)
7421 c        write (iout,*) "Receiving",nn," contacts from processor",iproc,
7422 c     &   " of CONT_TO_COMM group"
7423         call flush(iout)
7424         if (nn.gt.0) then
7425           ireq=ireq+1
7426           call MPI_Irecv(zapas_recv(1,1,ii),nn*max_dim,
7427      &    MPI_DOUBLE_PRECISION,iproc,CorrelType1,FG_COMM,req(ireq),IERR)
7428 c          write (iout,*) "ireq,req",ireq,req(ireq)
7429         endif
7430       enddo
7431 C Send the contacts to processors that need them
7432       do ii=1,ntask_cont_to
7433         iproc=itask_cont_to(ii)
7434         nn=ncont_sent(ii)
7435 c        write (iout,*) nn," contacts to processor",iproc,
7436 c     &   " of CONT_TO_COMM group"
7437         if (nn.gt.0) then
7438           ireq=ireq+1 
7439           call MPI_Isend(zapas(1,1,ii),nn*max_dim,MPI_DOUBLE_PRECISION,
7440      &      iproc,CorrelType1,FG_COMM,req(ireq),IERR)
7441 c          write (iout,*) "ireq,req",ireq,req(ireq)
7442 c          do i=1,nn
7443 c            write(iout,'(2f5.0,4f10.5)')(zapas(j,i,ii),j=1,5)
7444 c          enddo
7445         endif  
7446       enddo
7447 c      write (iout,*) "number of requests (contacts)",ireq
7448 c      write (iout,*) "req",(req(i),i=1,4)
7449 c      call flush(iout)
7450       if (ireq.gt.0) 
7451      & call MPI_Waitall(ireq,req,status_array,ierr)
7452       do iii=1,ntask_cont_from
7453         iproc=itask_cont_from(iii)
7454         nn=ncont_recv(iii)
7455         if (lprn) then
7456         write (iout,*) "Received",nn," contacts from processor",iproc,
7457      &   " of CONT_FROM_COMM group"
7458         call flush(iout)
7459         do i=1,nn
7460           write(iout,'(2f5.0,10f10.5)')(zapas_recv(j,i,iii),j=1,10)
7461         enddo
7462         call flush(iout)
7463         endif
7464         do i=1,nn
7465           ii=zapas_recv(1,i,iii)
7466 c Flag the received contacts to prevent double-counting
7467           jj=-zapas_recv(2,i,iii)
7468 c          write (iout,*) "iii",iii," i",i," ii",ii," jj",jj
7469 c          call flush(iout)
7470           nnn=num_cont_hb(ii)+1
7471           num_cont_hb(ii)=nnn
7472           jcont_hb(nnn,ii)=jj
7473           d_cont(nnn,ii)=zapas_recv(3,i,iii)
7474           ind=3
7475           do kk=1,3
7476             ind=ind+1
7477             grij_hb_cont(kk,nnn,ii)=zapas_recv(ind,i,iii)
7478           enddo
7479           do kk=1,2
7480             do ll=1,2
7481               ind=ind+1
7482               a_chuj(ll,kk,nnn,ii)=zapas_recv(ind,i,iii)
7483             enddo
7484           enddo
7485           do jj=1,5
7486             do kk=1,3
7487               do ll=1,2
7488                 do mm=1,2
7489                   ind=ind+1
7490                   a_chuj_der(mm,ll,kk,jj,nnn,ii)=zapas_recv(ind,i,iii)
7491                 enddo
7492               enddo
7493             enddo
7494           enddo
7495         enddo
7496       enddo
7497       call flush(iout)
7498       if (lprn) then
7499         write (iout,'(a)') 'Contact function values after receive:'
7500         do i=nnt,nct-2
7501           write (iout,'(2i3,50(1x,i3,5f6.3))') 
7502      &    i,num_cont_hb(i),(jcont_hb(j,i),d_cont(j,i),
7503      &    ((a_chuj(ll,kk,j,i),ll=1,2),kk=1,2),j=1,num_cont_hb(i))
7504         enddo
7505         call flush(iout)
7506       endif
7507    30 continue
7508 #endif
7509       if (lprn) then
7510         write (iout,'(a)') 'Contact function values:'
7511         do i=nnt,nct-2
7512           write (iout,'(2i3,50(1x,i2,5f6.3))') 
7513      &    i,num_cont_hb(i),(jcont_hb(j,i),d_cont(j,i),
7514      &    ((a_chuj(ll,kk,j,i),ll=1,2),kk=1,2),j=1,num_cont_hb(i))
7515         enddo
7516       endif
7517       ecorr=0.0D0
7518       ecorr5=0.0d0
7519       ecorr6=0.0d0
7520 C Remove the loop below after debugging !!!
7521       do i=nnt,nct
7522         do j=1,3
7523           gradcorr(j,i)=0.0D0
7524           gradxorr(j,i)=0.0D0
7525         enddo
7526       enddo
7527 C Calculate the dipole-dipole interaction energies
7528       if (wcorr6.gt.0.0d0 .or. wturn6.gt.0.0d0) then
7529       do i=iatel_s,iatel_e+1
7530         num_conti=num_cont_hb(i)
7531         do jj=1,num_conti
7532           j=jcont_hb(jj,i)
7533 #ifdef MOMENT
7534           call dipole(i,j,jj)
7535 #endif
7536         enddo
7537       enddo
7538       endif
7539 C Calculate the local-electrostatic correlation terms
7540 c                write (iout,*) "gradcorr5 in eello5 before loop"
7541 c                do iii=1,nres
7542 c                  write (iout,'(i5,3f10.5)') 
7543 c     &             iii,(gradcorr5(jjj,iii),jjj=1,3)
7544 c                enddo
7545       do i=min0(iatel_s,iturn4_start),max0(iatel_e+1,iturn3_end+1)
7546 c        write (iout,*) "corr loop i",i
7547         i1=i+1
7548         num_conti=num_cont_hb(i)
7549         num_conti1=num_cont_hb(i+1)
7550         do jj=1,num_conti
7551           j=jcont_hb(jj,i)
7552           jp=iabs(j)
7553           do kk=1,num_conti1
7554             j1=jcont_hb(kk,i1)
7555             jp1=iabs(j1)
7556 c            write (iout,*) 'i=',i,' j=',j,' i1=',i1,' j1=',j1,
7557 c     &         ' jj=',jj,' kk=',kk
7558 c            if (j1.eq.j+1 .or. j1.eq.j-1) then
7559             if ((j.gt.0 .and. j1.gt.0 .or. j.gt.0 .and. j1.lt.0 
7560      &          .or. j.lt.0 .and. j1.gt.0) .and.
7561      &         (jp1.eq.jp+1 .or. jp1.eq.jp-1)) then
7562 C Contacts I-J and (I+1)-(J+1) or (I+1)-(J-1) occur simultaneously. 
7563 C The system gains extra energy.
7564               n_corr=n_corr+1
7565               sqd1=dsqrt(d_cont(jj,i))
7566               sqd2=dsqrt(d_cont(kk,i1))
7567               sred_geom = sqd1*sqd2
7568               IF (sred_geom.lt.cutoff_corr) THEN
7569                 call gcont(sred_geom,r0_corr,1.0D0,delt_corr,
7570      &            ekont,fprimcont)
7571 cd               write (iout,*) 'i=',i,' j=',jp,' i1=',i1,' j1=',jp1,
7572 cd     &         ' jj=',jj,' kk=',kk
7573                 fac_prim1=0.5d0*sqd2/sqd1*fprimcont
7574                 fac_prim2=0.5d0*sqd1/sqd2*fprimcont
7575                 do l=1,3
7576                   g_contij(l,1)=fac_prim1*grij_hb_cont(l,jj,i)
7577                   g_contij(l,2)=fac_prim2*grij_hb_cont(l,kk,i1)
7578                 enddo
7579                 n_corr1=n_corr1+1
7580 cd               write (iout,*) 'sred_geom=',sred_geom,
7581 cd     &          ' ekont=',ekont,' fprim=',fprimcont,
7582 cd     &          ' fac_prim1',fac_prim1,' fac_prim2',fac_prim2
7583 cd               write (iout,*) "g_contij",g_contij
7584 cd               write (iout,*) "grij_hb_cont i",grij_hb_cont(:,jj,i)
7585 cd               write (iout,*) "grij_hb_cont i1",grij_hb_cont(:,jj,i1)
7586                 call calc_eello(i,jp,i+1,jp1,jj,kk)
7587                 if (wcorr4.gt.0.0d0) 
7588      &            ecorr=ecorr+eello4(i,jp,i+1,jp1,jj,kk)
7589                   if (energy_dec.and.wcorr4.gt.0.0d0) 
7590      1                 write (iout,'(a6,4i5,0pf7.3)')
7591      2                'ecorr4',i,j,i+1,j1,eello4(i,jp,i+1,jp1,jj,kk)
7592 c                write (iout,*) "gradcorr5 before eello5"
7593 c                do iii=1,nres
7594 c                  write (iout,'(i5,3f10.5)') 
7595 c     &             iii,(gradcorr5(jjj,iii),jjj=1,3)
7596 c                enddo
7597                 if (wcorr5.gt.0.0d0)
7598      &            ecorr5=ecorr5+eello5(i,jp,i+1,jp1,jj,kk)
7599 c                write (iout,*) "gradcorr5 after eello5"
7600 c                do iii=1,nres
7601 c                  write (iout,'(i5,3f10.5)') 
7602 c     &             iii,(gradcorr5(jjj,iii),jjj=1,3)
7603 c                enddo
7604                   if (energy_dec.and.wcorr5.gt.0.0d0) 
7605      1                 write (iout,'(a6,4i5,0pf7.3)')
7606      2                'ecorr5',i,j,i+1,j1,eello5(i,jp,i+1,jp1,jj,kk)
7607 cd                write(2,*)'wcorr6',wcorr6,' wturn6',wturn6
7608 cd                write(2,*)'ijkl',i,jp,i+1,jp1 
7609                 if (wcorr6.gt.0.0d0 .and. (jp.ne.i+4 .or. jp1.ne.i+3
7610      &               .or. wturn6.eq.0.0d0))then
7611 cd                  write (iout,*) '******ecorr6: i,j,i+1,j1',i,j,i+1,j1
7612                   ecorr6=ecorr6+eello6(i,jp,i+1,jp1,jj,kk)
7613                   if (energy_dec) write (iout,'(a6,4i5,0pf7.3)')
7614      1                'ecorr6',i,j,i+1,j1,eello6(i,jp,i+1,jp1,jj,kk)
7615 cd                write (iout,*) 'ecorr',ecorr,' ecorr5=',ecorr5,
7616 cd     &            'ecorr6=',ecorr6
7617 cd                write (iout,'(4e15.5)') sred_geom,
7618 cd     &          dabs(eello4(i,jp,i+1,jp1,jj,kk)),
7619 cd     &          dabs(eello5(i,jp,i+1,jp1,jj,kk)),
7620 cd     &          dabs(eello6(i,jp,i+1,jp1,jj,kk))
7621                 else if (wturn6.gt.0.0d0
7622      &            .and. (jp.eq.i+4 .and. jp1.eq.i+3)) then
7623 cd                  write (iout,*) '******eturn6: i,j,i+1,j1',i,jip,i+1,jp1
7624                   eturn6=eturn6+eello_turn6(i,jj,kk)
7625                   if (energy_dec) write (iout,'(a6,4i5,0pf7.3)')
7626      1                 'eturn6',i,j,i+1,j1,eello_turn6(i,jj,kk)
7627 cd                  write (2,*) 'multibody_eello:eturn6',eturn6
7628                 endif
7629               ENDIF
7630 1111          continue
7631             endif
7632           enddo ! kk
7633         enddo ! jj
7634       enddo ! i
7635       do i=1,nres
7636         num_cont_hb(i)=num_cont_hb_old(i)
7637       enddo
7638 c                write (iout,*) "gradcorr5 in eello5"
7639 c                do iii=1,nres
7640 c                  write (iout,'(i5,3f10.5)') 
7641 c     &             iii,(gradcorr5(jjj,iii),jjj=1,3)
7642 c                enddo
7643       return
7644       end
7645 c------------------------------------------------------------------------------
7646       subroutine add_hb_contact_eello(ii,jj,itask)
7647       implicit real*8 (a-h,o-z)
7648       include "DIMENSIONS"
7649       include "COMMON.IOUNITS"
7650       integer max_cont
7651       integer max_dim
7652       parameter (max_cont=maxconts)
7653       parameter (max_dim=70)
7654       include "COMMON.CONTACTS"
7655       double precision zapas(max_dim,maxconts,max_fg_procs),
7656      &  zapas_recv(max_dim,maxconts,max_fg_procs)
7657       common /przechowalnia/ zapas
7658       integer i,j,ii,jj,iproc,itask(4),nn
7659 c      write (iout,*) "itask",itask
7660       do i=1,2
7661         iproc=itask(i)
7662         if (iproc.gt.0) then
7663           do j=1,num_cont_hb(ii)
7664             jjc=jcont_hb(j,ii)
7665 c            write (iout,*) "send turns i",ii," j",jj," jjc",jjc
7666             if (jjc.eq.jj) then
7667               ncont_sent(iproc)=ncont_sent(iproc)+1
7668               nn=ncont_sent(iproc)
7669               zapas(1,nn,iproc)=ii
7670               zapas(2,nn,iproc)=jjc
7671               zapas(3,nn,iproc)=d_cont(j,ii)
7672               ind=3
7673               do kk=1,3
7674                 ind=ind+1
7675                 zapas(ind,nn,iproc)=grij_hb_cont(kk,j,ii)
7676               enddo
7677               do kk=1,2
7678                 do ll=1,2
7679                   ind=ind+1
7680                   zapas(ind,nn,iproc)=a_chuj(ll,kk,j,ii)
7681                 enddo
7682               enddo
7683               do jj=1,5
7684                 do kk=1,3
7685                   do ll=1,2
7686                     do mm=1,2
7687                       ind=ind+1
7688                       zapas(ind,nn,iproc)=a_chuj_der(mm,ll,kk,jj,j,ii)
7689                     enddo
7690                   enddo
7691                 enddo
7692               enddo
7693               exit
7694             endif
7695           enddo
7696         endif
7697       enddo
7698       return
7699       end
7700 c------------------------------------------------------------------------------
7701       double precision function ehbcorr(i,j,k,l,jj,kk,coeffp,coeffm)
7702       implicit real*8 (a-h,o-z)
7703       include 'DIMENSIONS'
7704       include 'COMMON.IOUNITS'
7705       include 'COMMON.DERIV'
7706       include 'COMMON.INTERACT'
7707       include 'COMMON.CONTACTS'
7708       double precision gx(3),gx1(3)
7709       logical lprn
7710       lprn=.false.
7711       eij=facont_hb(jj,i)
7712       ekl=facont_hb(kk,k)
7713       ees0pij=ees0p(jj,i)
7714       ees0pkl=ees0p(kk,k)
7715       ees0mij=ees0m(jj,i)
7716       ees0mkl=ees0m(kk,k)
7717       ekont=eij*ekl
7718       ees=-(coeffp*ees0pij*ees0pkl+coeffm*ees0mij*ees0mkl)
7719 cd    ees=-(coeffp*ees0pkl+coeffm*ees0mkl)
7720 C Following 4 lines for diagnostics.
7721 cd    ees0pkl=0.0D0
7722 cd    ees0pij=1.0D0
7723 cd    ees0mkl=0.0D0
7724 cd    ees0mij=1.0D0
7725 c      write (iout,'(2(a,2i3,a,f10.5,a,2f10.5),a,f10.5,a,$)')
7726 c     & 'Contacts ',i,j,
7727 c     & ' eij',eij,' eesij',ees0pij,ees0mij,' and ',k,l
7728 c     & ,' fcont ',ekl,' eeskl',ees0pkl,ees0mkl,' energy=',ekont*ees,
7729 c     & 'gradcorr_long'
7730 C Calculate the multi-body contribution to energy.
7731 c      ecorr=ecorr+ekont*ees
7732 C Calculate multi-body contributions to the gradient.
7733       coeffpees0pij=coeffp*ees0pij
7734       coeffmees0mij=coeffm*ees0mij
7735       coeffpees0pkl=coeffp*ees0pkl
7736       coeffmees0mkl=coeffm*ees0mkl
7737       do ll=1,3
7738 cgrad        ghalfi=ees*ekl*gacont_hbr(ll,jj,i)
7739         gradcorr(ll,i)=gradcorr(ll,i)!+0.5d0*ghalfi
7740      &  -ekont*(coeffpees0pkl*gacontp_hb1(ll,jj,i)+
7741      &  coeffmees0mkl*gacontm_hb1(ll,jj,i))
7742         gradcorr(ll,j)=gradcorr(ll,j)!+0.5d0*ghalfi
7743      &  -ekont*(coeffpees0pkl*gacontp_hb2(ll,jj,i)+
7744      &  coeffmees0mkl*gacontm_hb2(ll,jj,i))
7745 cgrad        ghalfk=ees*eij*gacont_hbr(ll,kk,k)
7746         gradcorr(ll,k)=gradcorr(ll,k)!+0.5d0*ghalfk
7747      &  -ekont*(coeffpees0pij*gacontp_hb1(ll,kk,k)+
7748      &  coeffmees0mij*gacontm_hb1(ll,kk,k))
7749         gradcorr(ll,l)=gradcorr(ll,l)!+0.5d0*ghalfk
7750      &  -ekont*(coeffpees0pij*gacontp_hb2(ll,kk,k)+
7751      &  coeffmees0mij*gacontm_hb2(ll,kk,k))
7752         gradlongij=ees*ekl*gacont_hbr(ll,jj,i)-
7753      &     ekont*(coeffpees0pkl*gacontp_hb3(ll,jj,i)+
7754      &     coeffmees0mkl*gacontm_hb3(ll,jj,i))
7755         gradcorr_long(ll,j)=gradcorr_long(ll,j)+gradlongij
7756         gradcorr_long(ll,i)=gradcorr_long(ll,i)-gradlongij
7757         gradlongkl=ees*eij*gacont_hbr(ll,kk,k)-
7758      &     ekont*(coeffpees0pij*gacontp_hb3(ll,kk,k)+
7759      &     coeffmees0mij*gacontm_hb3(ll,kk,k))
7760         gradcorr_long(ll,l)=gradcorr_long(ll,l)+gradlongkl
7761         gradcorr_long(ll,k)=gradcorr_long(ll,k)-gradlongkl
7762 c        write (iout,'(2f10.5,2x,$)') gradlongij,gradlongkl
7763       enddo
7764 c      write (iout,*)
7765 cgrad      do m=i+1,j-1
7766 cgrad        do ll=1,3
7767 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+
7768 cgrad     &     ees*ekl*gacont_hbr(ll,jj,i)-
7769 cgrad     &     ekont*(coeffp*ees0pkl*gacontp_hb3(ll,jj,i)+
7770 cgrad     &     coeffm*ees0mkl*gacontm_hb3(ll,jj,i))
7771 cgrad        enddo
7772 cgrad      enddo
7773 cgrad      do m=k+1,l-1
7774 cgrad        do ll=1,3
7775 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+
7776 cgrad     &     ees*eij*gacont_hbr(ll,kk,k)-
7777 cgrad     &     ekont*(coeffp*ees0pij*gacontp_hb3(ll,kk,k)+
7778 cgrad     &     coeffm*ees0mij*gacontm_hb3(ll,kk,k))
7779 cgrad        enddo
7780 cgrad      enddo 
7781 c      write (iout,*) "ehbcorr",ekont*ees
7782       ehbcorr=ekont*ees
7783       return
7784       end
7785 #ifdef MOMENT
7786 C---------------------------------------------------------------------------
7787       subroutine dipole(i,j,jj)
7788       implicit real*8 (a-h,o-z)
7789       include 'DIMENSIONS'
7790       include 'COMMON.IOUNITS'
7791       include 'COMMON.CHAIN'
7792       include 'COMMON.FFIELD'
7793       include 'COMMON.DERIV'
7794       include 'COMMON.INTERACT'
7795       include 'COMMON.CONTACTS'
7796       include 'COMMON.TORSION'
7797       include 'COMMON.VAR'
7798       include 'COMMON.GEO'
7799       dimension dipi(2,2),dipj(2,2),dipderi(2),dipderj(2),auxvec(2),
7800      &  auxmat(2,2)
7801       iti1 = itortyp(itype(i+1))
7802       if (j.lt.nres-1) then
7803         itj1 = itortyp(itype(j+1))
7804       else
7805         itj1=ntortyp+1
7806       endif
7807       do iii=1,2
7808         dipi(iii,1)=Ub2(iii,i)
7809         dipderi(iii)=Ub2der(iii,i)
7810         dipi(iii,2)=b1(iii,iti1)
7811         dipj(iii,1)=Ub2(iii,j)
7812         dipderj(iii)=Ub2der(iii,j)
7813         dipj(iii,2)=b1(iii,itj1)
7814       enddo
7815       kkk=0
7816       do iii=1,2
7817         call matvec2(a_chuj(1,1,jj,i),dipj(1,iii),auxvec(1)) 
7818         do jjj=1,2
7819           kkk=kkk+1
7820           dip(kkk,jj,i)=scalar2(dipi(1,jjj),auxvec(1))
7821         enddo
7822       enddo
7823       do kkk=1,5
7824         do lll=1,3
7825           mmm=0
7826           do iii=1,2
7827             call matvec2(a_chuj_der(1,1,lll,kkk,jj,i),dipj(1,iii),
7828      &        auxvec(1))
7829             do jjj=1,2
7830               mmm=mmm+1
7831               dipderx(lll,kkk,mmm,jj,i)=scalar2(dipi(1,jjj),auxvec(1))
7832             enddo
7833           enddo
7834         enddo
7835       enddo
7836       call transpose2(a_chuj(1,1,jj,i),auxmat(1,1))
7837       call matvec2(auxmat(1,1),dipderi(1),auxvec(1))
7838       do iii=1,2
7839         dipderg(iii,jj,i)=scalar2(auxvec(1),dipj(1,iii))
7840       enddo
7841       call matvec2(a_chuj(1,1,jj,i),dipderj(1),auxvec(1))
7842       do iii=1,2
7843         dipderg(iii+2,jj,i)=scalar2(auxvec(1),dipi(1,iii))
7844       enddo
7845       return
7846       end
7847 #endif
7848 C---------------------------------------------------------------------------
7849       subroutine calc_eello(i,j,k,l,jj,kk)
7850
7851 C This subroutine computes matrices and vectors needed to calculate 
7852 C the fourth-, fifth-, and sixth-order local-electrostatic terms.
7853 C
7854       implicit real*8 (a-h,o-z)
7855       include 'DIMENSIONS'
7856       include 'COMMON.IOUNITS'
7857       include 'COMMON.CHAIN'
7858       include 'COMMON.DERIV'
7859       include 'COMMON.INTERACT'
7860       include 'COMMON.CONTACTS'
7861       include 'COMMON.TORSION'
7862       include 'COMMON.VAR'
7863       include 'COMMON.GEO'
7864       include 'COMMON.FFIELD'
7865       double precision aa1(2,2),aa2(2,2),aa1t(2,2),aa2t(2,2),
7866      &  aa1tder(2,2,3,5),aa2tder(2,2,3,5),auxmat(2,2)
7867       logical lprn
7868       common /kutas/ lprn
7869 cd      write (iout,*) 'calc_eello: i=',i,' j=',j,' k=',k,' l=',l,
7870 cd     & ' jj=',jj,' kk=',kk
7871 cd      if (i.ne.2 .or. j.ne.4 .or. k.ne.3 .or. l.ne.5) return
7872 cd      write (iout,*) "a_chujij",((a_chuj(iii,jjj,jj,i),iii=1,2),jjj=1,2)
7873 cd      write (iout,*) "a_chujkl",((a_chuj(iii,jjj,kk,k),iii=1,2),jjj=1,2)
7874       do iii=1,2
7875         do jjj=1,2
7876           aa1(iii,jjj)=a_chuj(iii,jjj,jj,i)
7877           aa2(iii,jjj)=a_chuj(iii,jjj,kk,k)
7878         enddo
7879       enddo
7880       call transpose2(aa1(1,1),aa1t(1,1))
7881       call transpose2(aa2(1,1),aa2t(1,1))
7882       do kkk=1,5
7883         do lll=1,3
7884           call transpose2(a_chuj_der(1,1,lll,kkk,jj,i),
7885      &      aa1tder(1,1,lll,kkk))
7886           call transpose2(a_chuj_der(1,1,lll,kkk,kk,k),
7887      &      aa2tder(1,1,lll,kkk))
7888         enddo
7889       enddo 
7890       if (l.eq.j+1) then
7891 C parallel orientation of the two CA-CA-CA frames.
7892         if (i.gt.1) then
7893           iti=itortyp(itype(i))
7894         else
7895           iti=ntortyp+1
7896         endif
7897         itk1=itortyp(itype(k+1))
7898         itj=itortyp(itype(j))
7899         if (l.lt.nres-1) then
7900           itl1=itortyp(itype(l+1))
7901         else
7902           itl1=ntortyp+1
7903         endif
7904 C A1 kernel(j+1) A2T
7905 cd        do iii=1,2
7906 cd          write (iout,'(3f10.5,5x,3f10.5)') 
7907 cd     &     (EUg(iii,jjj,k),jjj=1,2),(EUg(iii,jjj,l),jjj=1,2)
7908 cd        enddo
7909         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
7910      &   aa2tder(1,1,1,1),1,.false.,EUg(1,1,l),EUgder(1,1,l),
7911      &   AEA(1,1,1),AEAderg(1,1,1),AEAderx(1,1,1,1,1,1))
7912 C Following matrices are needed only for 6-th order cumulants
7913         IF (wcorr6.gt.0.0d0) THEN
7914         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
7915      &   aa2tder(1,1,1,1),1,.false.,EUgC(1,1,l),EUgCder(1,1,l),
7916      &   AECA(1,1,1),AECAderg(1,1,1),AECAderx(1,1,1,1,1,1))
7917         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
7918      &   aa2tder(1,1,1,1),2,.false.,Ug2DtEUg(1,1,l),
7919      &   Ug2DtEUgder(1,1,1,l),ADtEA(1,1,1),ADtEAderg(1,1,1,1),
7920      &   ADtEAderx(1,1,1,1,1,1))
7921         lprn=.false.
7922         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
7923      &   aa2tder(1,1,1,1),2,.false.,DtUg2EUg(1,1,l),
7924      &   DtUg2EUgder(1,1,1,l),ADtEA1(1,1,1),ADtEA1derg(1,1,1,1),
7925      &   ADtEA1derx(1,1,1,1,1,1))
7926         ENDIF
7927 C End 6-th order cumulants
7928 cd        lprn=.false.
7929 cd        if (lprn) then
7930 cd        write (2,*) 'In calc_eello6'
7931 cd        do iii=1,2
7932 cd          write (2,*) 'iii=',iii
7933 cd          do kkk=1,5
7934 cd            write (2,*) 'kkk=',kkk
7935 cd            do jjj=1,2
7936 cd              write (2,'(3(2f10.5),5x)') 
7937 cd     &        ((ADtEA1derx(jjj,mmm,lll,kkk,iii,1),mmm=1,2),lll=1,3)
7938 cd            enddo
7939 cd          enddo
7940 cd        enddo
7941 cd        endif
7942         call transpose2(EUgder(1,1,k),auxmat(1,1))
7943         call matmat2(auxmat(1,1),AEA(1,1,1),EAEAderg(1,1,1,1))
7944         call transpose2(EUg(1,1,k),auxmat(1,1))
7945         call matmat2(auxmat(1,1),AEA(1,1,1),EAEA(1,1,1))
7946         call matmat2(auxmat(1,1),AEAderg(1,1,1),EAEAderg(1,1,2,1))
7947         do iii=1,2
7948           do kkk=1,5
7949             do lll=1,3
7950               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,1),
7951      &          EAEAderx(1,1,lll,kkk,iii,1))
7952             enddo
7953           enddo
7954         enddo
7955 C A1T kernel(i+1) A2
7956         call kernel(aa1t(1,1),aa2(1,1),aa1tder(1,1,1,1),
7957      &   a_chuj_der(1,1,1,1,kk,k),1,.false.,EUg(1,1,k),EUgder(1,1,k),
7958      &   AEA(1,1,2),AEAderg(1,1,2),AEAderx(1,1,1,1,1,2))
7959 C Following matrices are needed only for 6-th order cumulants
7960         IF (wcorr6.gt.0.0d0) THEN
7961         call kernel(aa1t(1,1),aa2(1,1),aa1tder(1,1,1,1),
7962      &   a_chuj_der(1,1,1,1,kk,k),1,.false.,EUgC(1,1,k),EUgCder(1,1,k),
7963      &   AECA(1,1,2),AECAderg(1,1,2),AECAderx(1,1,1,1,1,2))
7964         call kernel(aa1t(1,1),aa2(1,1),aa1tder(1,1,1,1),
7965      &   a_chuj_der(1,1,1,1,kk,k),2,.false.,Ug2DtEUg(1,1,k),
7966      &   Ug2DtEUgder(1,1,1,k),ADtEA(1,1,2),ADtEAderg(1,1,1,2),
7967      &   ADtEAderx(1,1,1,1,1,2))
7968         call kernel(aa1t(1,1),aa2(1,1),aa1tder(1,1,1,1),
7969      &   a_chuj_der(1,1,1,1,kk,k),2,.false.,DtUg2EUg(1,1,k),
7970      &   DtUg2EUgder(1,1,1,k),ADtEA1(1,1,2),ADtEA1derg(1,1,1,2),
7971      &   ADtEA1derx(1,1,1,1,1,2))
7972         ENDIF
7973 C End 6-th order cumulants
7974         call transpose2(EUgder(1,1,l),auxmat(1,1))
7975         call matmat2(auxmat(1,1),AEA(1,1,2),EAEAderg(1,1,1,2))
7976         call transpose2(EUg(1,1,l),auxmat(1,1))
7977         call matmat2(auxmat(1,1),AEA(1,1,2),EAEA(1,1,2))
7978         call matmat2(auxmat(1,1),AEAderg(1,1,2),EAEAderg(1,1,2,2))
7979         do iii=1,2
7980           do kkk=1,5
7981             do lll=1,3
7982               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,2),
7983      &          EAEAderx(1,1,lll,kkk,iii,2))
7984             enddo
7985           enddo
7986         enddo
7987 C AEAb1 and AEAb2
7988 C Calculate the vectors and their derivatives in virtual-bond dihedral angles.
7989 C They are needed only when the fifth- or the sixth-order cumulants are
7990 C indluded.
7991         IF (wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0) THEN
7992         call transpose2(AEA(1,1,1),auxmat(1,1))
7993         call matvec2(auxmat(1,1),b1(1,iti),AEAb1(1,1,1))
7994         call matvec2(auxmat(1,1),Ub2(1,i),AEAb2(1,1,1))
7995         call matvec2(auxmat(1,1),Ub2der(1,i),AEAb2derg(1,2,1,1))
7996         call transpose2(AEAderg(1,1,1),auxmat(1,1))
7997         call matvec2(auxmat(1,1),b1(1,iti),AEAb1derg(1,1,1))
7998         call matvec2(auxmat(1,1),Ub2(1,i),AEAb2derg(1,1,1,1))
7999         call matvec2(AEA(1,1,1),b1(1,itk1),AEAb1(1,2,1))
8000         call matvec2(AEAderg(1,1,1),b1(1,itk1),AEAb1derg(1,2,1))
8001         call matvec2(AEA(1,1,1),Ub2(1,k+1),AEAb2(1,2,1))
8002         call matvec2(AEAderg(1,1,1),Ub2(1,k+1),AEAb2derg(1,1,2,1))
8003         call matvec2(AEA(1,1,1),Ub2der(1,k+1),AEAb2derg(1,2,2,1))
8004         call transpose2(AEA(1,1,2),auxmat(1,1))
8005         call matvec2(auxmat(1,1),b1(1,itj),AEAb1(1,1,2))
8006         call matvec2(auxmat(1,1),Ub2(1,j),AEAb2(1,1,2))
8007         call matvec2(auxmat(1,1),Ub2der(1,j),AEAb2derg(1,2,1,2))
8008         call transpose2(AEAderg(1,1,2),auxmat(1,1))
8009         call matvec2(auxmat(1,1),b1(1,itj),AEAb1derg(1,1,2))
8010         call matvec2(auxmat(1,1),Ub2(1,j),AEAb2derg(1,1,1,2))
8011         call matvec2(AEA(1,1,2),b1(1,itl1),AEAb1(1,2,2))
8012         call matvec2(AEAderg(1,1,2),b1(1,itl1),AEAb1derg(1,2,2))
8013         call matvec2(AEA(1,1,2),Ub2(1,l+1),AEAb2(1,2,2))
8014         call matvec2(AEAderg(1,1,2),Ub2(1,l+1),AEAb2derg(1,1,2,2))
8015         call matvec2(AEA(1,1,2),Ub2der(1,l+1),AEAb2derg(1,2,2,2))
8016 C Calculate the Cartesian derivatives of the vectors.
8017         do iii=1,2
8018           do kkk=1,5
8019             do lll=1,3
8020               call transpose2(AEAderx(1,1,lll,kkk,iii,1),auxmat(1,1))
8021               call matvec2(auxmat(1,1),b1(1,iti),
8022      &          AEAb1derx(1,lll,kkk,iii,1,1))
8023               call matvec2(auxmat(1,1),Ub2(1,i),
8024      &          AEAb2derx(1,lll,kkk,iii,1,1))
8025               call matvec2(AEAderx(1,1,lll,kkk,iii,1),b1(1,itk1),
8026      &          AEAb1derx(1,lll,kkk,iii,2,1))
8027               call matvec2(AEAderx(1,1,lll,kkk,iii,1),Ub2(1,k+1),
8028      &          AEAb2derx(1,lll,kkk,iii,2,1))
8029               call transpose2(AEAderx(1,1,lll,kkk,iii,2),auxmat(1,1))
8030               call matvec2(auxmat(1,1),b1(1,itj),
8031      &          AEAb1derx(1,lll,kkk,iii,1,2))
8032               call matvec2(auxmat(1,1),Ub2(1,j),
8033      &          AEAb2derx(1,lll,kkk,iii,1,2))
8034               call matvec2(AEAderx(1,1,lll,kkk,iii,2),b1(1,itl1),
8035      &          AEAb1derx(1,lll,kkk,iii,2,2))
8036               call matvec2(AEAderx(1,1,lll,kkk,iii,2),Ub2(1,l+1),
8037      &          AEAb2derx(1,lll,kkk,iii,2,2))
8038             enddo
8039           enddo
8040         enddo
8041         ENDIF
8042 C End vectors
8043       else
8044 C Antiparallel orientation of the two CA-CA-CA frames.
8045         if (i.gt.1) then
8046           iti=itortyp(itype(i))
8047         else
8048           iti=ntortyp+1
8049         endif
8050         itk1=itortyp(itype(k+1))
8051         itl=itortyp(itype(l))
8052         itj=itortyp(itype(j))
8053         if (j.lt.nres-1) then
8054           itj1=itortyp(itype(j+1))
8055         else 
8056           itj1=ntortyp+1
8057         endif
8058 C A2 kernel(j-1)T A1T
8059         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
8060      &   aa2tder(1,1,1,1),1,.true.,EUg(1,1,j),EUgder(1,1,j),
8061      &   AEA(1,1,1),AEAderg(1,1,1),AEAderx(1,1,1,1,1,1))
8062 C Following matrices are needed only for 6-th order cumulants
8063         IF (wcorr6.gt.0.0d0 .or. (wturn6.gt.0.0d0 .and.
8064      &     j.eq.i+4 .and. l.eq.i+3)) THEN
8065         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
8066      &   aa2tder(1,1,1,1),1,.true.,EUgC(1,1,j),EUgCder(1,1,j),
8067      &   AECA(1,1,1),AECAderg(1,1,1),AECAderx(1,1,1,1,1,1))
8068         call kernel(aa2(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
8069      &   aa2tder(1,1,1,1),2,.true.,Ug2DtEUg(1,1,j),
8070      &   Ug2DtEUgder(1,1,1,j),ADtEA(1,1,1),ADtEAderg(1,1,1,1),
8071      &   ADtEAderx(1,1,1,1,1,1))
8072         call kernel(aa1(1,1),aa2t(1,1),a_chuj_der(1,1,1,1,jj,i),
8073      &   aa2tder(1,1,1,1),2,.true.,DtUg2EUg(1,1,j),
8074      &   DtUg2EUgder(1,1,1,j),ADtEA1(1,1,1),ADtEA1derg(1,1,1,1),
8075      &   ADtEA1derx(1,1,1,1,1,1))
8076         ENDIF
8077 C End 6-th order cumulants
8078         call transpose2(EUgder(1,1,k),auxmat(1,1))
8079         call matmat2(auxmat(1,1),AEA(1,1,1),EAEAderg(1,1,1,1))
8080         call transpose2(EUg(1,1,k),auxmat(1,1))
8081         call matmat2(auxmat(1,1),AEA(1,1,1),EAEA(1,1,1))
8082         call matmat2(auxmat(1,1),AEAderg(1,1,1),EAEAderg(1,1,2,1))
8083         do iii=1,2
8084           do kkk=1,5
8085             do lll=1,3
8086               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,1),
8087      &          EAEAderx(1,1,lll,kkk,iii,1))
8088             enddo
8089           enddo
8090         enddo
8091 C A2T kernel(i+1)T A1
8092         call kernel(aa2t(1,1),aa1(1,1),aa2tder(1,1,1,1),
8093      &   a_chuj_der(1,1,1,1,jj,i),1,.true.,EUg(1,1,k),EUgder(1,1,k),
8094      &   AEA(1,1,2),AEAderg(1,1,2),AEAderx(1,1,1,1,1,2))
8095 C Following matrices are needed only for 6-th order cumulants
8096         IF (wcorr6.gt.0.0d0 .or. (wturn6.gt.0.0d0 .and.
8097      &     j.eq.i+4 .and. l.eq.i+3)) THEN
8098         call kernel(aa2t(1,1),aa1(1,1),aa2tder(1,1,1,1),
8099      &   a_chuj_der(1,1,1,1,jj,i),1,.true.,EUgC(1,1,k),EUgCder(1,1,k),
8100      &   AECA(1,1,2),AECAderg(1,1,2),AECAderx(1,1,1,1,1,2))
8101         call kernel(aa2t(1,1),aa1(1,1),aa2tder(1,1,1,1),
8102      &   a_chuj_der(1,1,1,1,jj,i),2,.true.,Ug2DtEUg(1,1,k),
8103      &   Ug2DtEUgder(1,1,1,k),ADtEA(1,1,2),ADtEAderg(1,1,1,2),
8104      &   ADtEAderx(1,1,1,1,1,2))
8105         call kernel(aa2t(1,1),aa1(1,1),aa2tder(1,1,1,1),
8106      &   a_chuj_der(1,1,1,1,jj,i),2,.true.,DtUg2EUg(1,1,k),
8107      &   DtUg2EUgder(1,1,1,k),ADtEA1(1,1,2),ADtEA1derg(1,1,1,2),
8108      &   ADtEA1derx(1,1,1,1,1,2))
8109         ENDIF
8110 C End 6-th order cumulants
8111         call transpose2(EUgder(1,1,j),auxmat(1,1))
8112         call matmat2(auxmat(1,1),AEA(1,1,1),EAEAderg(1,1,2,2))
8113         call transpose2(EUg(1,1,j),auxmat(1,1))
8114         call matmat2(auxmat(1,1),AEA(1,1,2),EAEA(1,1,2))
8115         call matmat2(auxmat(1,1),AEAderg(1,1,2),EAEAderg(1,1,2,2))
8116         do iii=1,2
8117           do kkk=1,5
8118             do lll=1,3
8119               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,2),
8120      &          EAEAderx(1,1,lll,kkk,iii,2))
8121             enddo
8122           enddo
8123         enddo
8124 C AEAb1 and AEAb2
8125 C Calculate the vectors and their derivatives in virtual-bond dihedral angles.
8126 C They are needed only when the fifth- or the sixth-order cumulants are
8127 C indluded.
8128         IF (wcorr5.gt.0.0d0 .or. wcorr6.gt.0.0d0 .or.
8129      &    (wturn6.gt.0.0d0 .and. j.eq.i+4 .and. l.eq.i+3)) THEN
8130         call transpose2(AEA(1,1,1),auxmat(1,1))
8131         call matvec2(auxmat(1,1),b1(1,iti),AEAb1(1,1,1))
8132         call matvec2(auxmat(1,1),Ub2(1,i),AEAb2(1,1,1))
8133         call matvec2(auxmat(1,1),Ub2der(1,i),AEAb2derg(1,2,1,1))
8134         call transpose2(AEAderg(1,1,1),auxmat(1,1))
8135         call matvec2(auxmat(1,1),b1(1,iti),AEAb1derg(1,1,1))
8136         call matvec2(auxmat(1,1),Ub2(1,i),AEAb2derg(1,1,1,1))
8137         call matvec2(AEA(1,1,1),b1(1,itk1),AEAb1(1,2,1))
8138         call matvec2(AEAderg(1,1,1),b1(1,itk1),AEAb1derg(1,2,1))
8139         call matvec2(AEA(1,1,1),Ub2(1,k+1),AEAb2(1,2,1))
8140         call matvec2(AEAderg(1,1,1),Ub2(1,k+1),AEAb2derg(1,1,2,1))
8141         call matvec2(AEA(1,1,1),Ub2der(1,k+1),AEAb2derg(1,2,2,1))
8142         call transpose2(AEA(1,1,2),auxmat(1,1))
8143         call matvec2(auxmat(1,1),b1(1,itj1),AEAb1(1,1,2))
8144         call matvec2(auxmat(1,1),Ub2(1,l),AEAb2(1,1,2))
8145         call matvec2(auxmat(1,1),Ub2der(1,l),AEAb2derg(1,2,1,2))
8146         call transpose2(AEAderg(1,1,2),auxmat(1,1))
8147         call matvec2(auxmat(1,1),b1(1,itl),AEAb1(1,1,2))
8148         call matvec2(auxmat(1,1),Ub2(1,l),AEAb2derg(1,1,1,2))
8149         call matvec2(AEA(1,1,2),b1(1,itj1),AEAb1(1,2,2))
8150         call matvec2(AEAderg(1,1,2),b1(1,itj1),AEAb1derg(1,2,2))
8151         call matvec2(AEA(1,1,2),Ub2(1,j),AEAb2(1,2,2))
8152         call matvec2(AEAderg(1,1,2),Ub2(1,j),AEAb2derg(1,1,2,2))
8153         call matvec2(AEA(1,1,2),Ub2der(1,j),AEAb2derg(1,2,2,2))
8154 C Calculate the Cartesian derivatives of the vectors.
8155         do iii=1,2
8156           do kkk=1,5
8157             do lll=1,3
8158               call transpose2(AEAderx(1,1,lll,kkk,iii,1),auxmat(1,1))
8159               call matvec2(auxmat(1,1),b1(1,iti),
8160      &          AEAb1derx(1,lll,kkk,iii,1,1))
8161               call matvec2(auxmat(1,1),Ub2(1,i),
8162      &          AEAb2derx(1,lll,kkk,iii,1,1))
8163               call matvec2(AEAderx(1,1,lll,kkk,iii,1),b1(1,itk1),
8164      &          AEAb1derx(1,lll,kkk,iii,2,1))
8165               call matvec2(AEAderx(1,1,lll,kkk,iii,1),Ub2(1,k+1),
8166      &          AEAb2derx(1,lll,kkk,iii,2,1))
8167               call transpose2(AEAderx(1,1,lll,kkk,iii,2),auxmat(1,1))
8168               call matvec2(auxmat(1,1),b1(1,itl),
8169      &          AEAb1derx(1,lll,kkk,iii,1,2))
8170               call matvec2(auxmat(1,1),Ub2(1,l),
8171      &          AEAb2derx(1,lll,kkk,iii,1,2))
8172               call matvec2(AEAderx(1,1,lll,kkk,iii,2),b1(1,itj1),
8173      &          AEAb1derx(1,lll,kkk,iii,2,2))
8174               call matvec2(AEAderx(1,1,lll,kkk,iii,2),Ub2(1,j),
8175      &          AEAb2derx(1,lll,kkk,iii,2,2))
8176             enddo
8177           enddo
8178         enddo
8179         ENDIF
8180 C End vectors
8181       endif
8182       return
8183       end
8184 C---------------------------------------------------------------------------
8185       subroutine kernel(aa1,aa2t,aa1derx,aa2tderx,nderg,transp,
8186      &  KK,KKderg,AKA,AKAderg,AKAderx)
8187       implicit none
8188       integer nderg
8189       logical transp
8190       double precision aa1(2,2),aa2t(2,2),aa1derx(2,2,3,5),
8191      &  aa2tderx(2,2,3,5),KK(2,2),KKderg(2,2,nderg),AKA(2,2),
8192      &  AKAderg(2,2,nderg),AKAderx(2,2,3,5,2)
8193       integer iii,kkk,lll
8194       integer jjj,mmm
8195       logical lprn
8196       common /kutas/ lprn
8197       call prodmat3(aa1(1,1),aa2t(1,1),KK(1,1),transp,AKA(1,1))
8198       do iii=1,nderg 
8199         call prodmat3(aa1(1,1),aa2t(1,1),KKderg(1,1,iii),transp,
8200      &    AKAderg(1,1,iii))
8201       enddo
8202 cd      if (lprn) write (2,*) 'In kernel'
8203       do kkk=1,5
8204 cd        if (lprn) write (2,*) 'kkk=',kkk
8205         do lll=1,3
8206           call prodmat3(aa1derx(1,1,lll,kkk),aa2t(1,1),
8207      &      KK(1,1),transp,AKAderx(1,1,lll,kkk,1))
8208 cd          if (lprn) then
8209 cd            write (2,*) 'lll=',lll
8210 cd            write (2,*) 'iii=1'
8211 cd            do jjj=1,2
8212 cd              write (2,'(3(2f10.5),5x)') 
8213 cd     &        (AKAderx(jjj,mmm,lll,kkk,1),mmm=1,2)
8214 cd            enddo
8215 cd          endif
8216           call prodmat3(aa1(1,1),aa2tderx(1,1,lll,kkk),
8217      &      KK(1,1),transp,AKAderx(1,1,lll,kkk,2))
8218 cd          if (lprn) then
8219 cd            write (2,*) 'lll=',lll
8220 cd            write (2,*) 'iii=2'
8221 cd            do jjj=1,2
8222 cd              write (2,'(3(2f10.5),5x)') 
8223 cd     &        (AKAderx(jjj,mmm,lll,kkk,2),mmm=1,2)
8224 cd            enddo
8225 cd          endif
8226         enddo
8227       enddo
8228       return
8229       end
8230 C---------------------------------------------------------------------------
8231       double precision function eello4(i,j,k,l,jj,kk)
8232       implicit real*8 (a-h,o-z)
8233       include 'DIMENSIONS'
8234       include 'COMMON.IOUNITS'
8235       include 'COMMON.CHAIN'
8236       include 'COMMON.DERIV'
8237       include 'COMMON.INTERACT'
8238       include 'COMMON.CONTACTS'
8239       include 'COMMON.TORSION'
8240       include 'COMMON.VAR'
8241       include 'COMMON.GEO'
8242       double precision pizda(2,2),ggg1(3),ggg2(3)
8243 cd      if (i.ne.1 .or. j.ne.5 .or. k.ne.2 .or.l.ne.4) then
8244 cd        eello4=0.0d0
8245 cd        return
8246 cd      endif
8247 cd      print *,'eello4:',i,j,k,l,jj,kk
8248 cd      write (2,*) 'i',i,' j',j,' k',k,' l',l
8249 cd      call checkint4(i,j,k,l,jj,kk,eel4_num)
8250 cold      eij=facont_hb(jj,i)
8251 cold      ekl=facont_hb(kk,k)
8252 cold      ekont=eij*ekl
8253       eel4=-EAEA(1,1,1)-EAEA(2,2,1)
8254 cd      eel41=-EAEA(1,1,2)-EAEA(2,2,2)
8255       gcorr_loc(k-1)=gcorr_loc(k-1)
8256      &   -ekont*(EAEAderg(1,1,1,1)+EAEAderg(2,2,1,1))
8257       if (l.eq.j+1) then
8258         gcorr_loc(l-1)=gcorr_loc(l-1)
8259      &     -ekont*(EAEAderg(1,1,2,1)+EAEAderg(2,2,2,1))
8260       else
8261         gcorr_loc(j-1)=gcorr_loc(j-1)
8262      &     -ekont*(EAEAderg(1,1,2,1)+EAEAderg(2,2,2,1))
8263       endif
8264       do iii=1,2
8265         do kkk=1,5
8266           do lll=1,3
8267             derx(lll,kkk,iii)=-EAEAderx(1,1,lll,kkk,iii,1)
8268      &                        -EAEAderx(2,2,lll,kkk,iii,1)
8269 cd            derx(lll,kkk,iii)=0.0d0
8270           enddo
8271         enddo
8272       enddo
8273 cd      gcorr_loc(l-1)=0.0d0
8274 cd      gcorr_loc(j-1)=0.0d0
8275 cd      gcorr_loc(k-1)=0.0d0
8276 cd      eel4=1.0d0
8277 cd      write (iout,*)'Contacts have occurred for peptide groups',
8278 cd     &  i,j,' fcont:',eij,' eij',' and ',k,l,
8279 cd     &  ' fcont ',ekl,' eel4=',eel4,' eel4_num',16*eel4_num
8280       if (j.lt.nres-1) then
8281         j1=j+1
8282         j2=j-1
8283       else
8284         j1=j-1
8285         j2=j-2
8286       endif
8287       if (l.lt.nres-1) then
8288         l1=l+1
8289         l2=l-1
8290       else
8291         l1=l-1
8292         l2=l-2
8293       endif
8294       do ll=1,3
8295 cgrad        ggg1(ll)=eel4*g_contij(ll,1)
8296 cgrad        ggg2(ll)=eel4*g_contij(ll,2)
8297         glongij=eel4*g_contij(ll,1)+ekont*derx(ll,1,1)
8298         glongkl=eel4*g_contij(ll,2)+ekont*derx(ll,1,2)
8299 cgrad        ghalf=0.5d0*ggg1(ll)
8300         gradcorr(ll,i)=gradcorr(ll,i)+ekont*derx(ll,2,1)
8301         gradcorr(ll,i+1)=gradcorr(ll,i+1)+ekont*derx(ll,3,1)
8302         gradcorr(ll,j)=gradcorr(ll,j)+ekont*derx(ll,4,1)
8303         gradcorr(ll,j1)=gradcorr(ll,j1)+ekont*derx(ll,5,1)
8304         gradcorr_long(ll,j)=gradcorr_long(ll,j)+glongij
8305         gradcorr_long(ll,i)=gradcorr_long(ll,i)-glongij
8306 cgrad        ghalf=0.5d0*ggg2(ll)
8307         gradcorr(ll,k)=gradcorr(ll,k)+ekont*derx(ll,2,2)
8308         gradcorr(ll,k+1)=gradcorr(ll,k+1)+ekont*derx(ll,3,2)
8309         gradcorr(ll,l)=gradcorr(ll,l)+ekont*derx(ll,4,2)
8310         gradcorr(ll,l1)=gradcorr(ll,l1)+ekont*derx(ll,5,2)
8311         gradcorr_long(ll,l)=gradcorr_long(ll,l)+glongkl
8312         gradcorr_long(ll,k)=gradcorr_long(ll,k)-glongkl
8313       enddo
8314 cgrad      do m=i+1,j-1
8315 cgrad        do ll=1,3
8316 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+ggg1(ll)
8317 cgrad        enddo
8318 cgrad      enddo
8319 cgrad      do m=k+1,l-1
8320 cgrad        do ll=1,3
8321 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+ggg2(ll)
8322 cgrad        enddo
8323 cgrad      enddo
8324 cgrad      do m=i+2,j2
8325 cgrad        do ll=1,3
8326 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+ekont*derx(ll,1,1)
8327 cgrad        enddo
8328 cgrad      enddo
8329 cgrad      do m=k+2,l2
8330 cgrad        do ll=1,3
8331 cgrad          gradcorr(ll,m)=gradcorr(ll,m)+ekont*derx(ll,1,2)
8332 cgrad        enddo
8333 cgrad      enddo 
8334 cd      do iii=1,nres-3
8335 cd        write (2,*) iii,gcorr_loc(iii)
8336 cd      enddo
8337       eello4=ekont*eel4
8338 cd      write (2,*) 'ekont',ekont
8339 cd      write (iout,*) 'eello4',ekont*eel4
8340       return
8341       end
8342 C---------------------------------------------------------------------------
8343       double precision function eello5(i,j,k,l,jj,kk)
8344       implicit real*8 (a-h,o-z)
8345       include 'DIMENSIONS'
8346       include 'COMMON.IOUNITS'
8347       include 'COMMON.CHAIN'
8348       include 'COMMON.DERIV'
8349       include 'COMMON.INTERACT'
8350       include 'COMMON.CONTACTS'
8351       include 'COMMON.TORSION'
8352       include 'COMMON.VAR'
8353       include 'COMMON.GEO'
8354       double precision pizda(2,2),auxmat(2,2),auxmat1(2,2),vv(2)
8355       double precision ggg1(3),ggg2(3)
8356 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
8357 C                                                                              C
8358 C                            Parallel chains                                   C
8359 C                                                                              C
8360 C          o             o                   o             o                   C
8361 C         /l\           / \             \   / \           / \   /              C
8362 C        /   \         /   \             \ /   \         /   \ /               C
8363 C       j| o |l1       | o |              o| o |         | o |o                C
8364 C     \  |/k\|         |/ \|  /            |/ \|         |/ \|                 C
8365 C      \i/   \         /   \ /             /   \         /   \                 C
8366 C       o    k1             o                                                  C
8367 C         (I)          (II)                (III)          (IV)                 C
8368 C                                                                              C
8369 C      eello5_1        eello5_2            eello5_3       eello5_4             C
8370 C                                                                              C
8371 C                            Antiparallel chains                               C
8372 C                                                                              C
8373 C          o             o                   o             o                   C
8374 C         /j\           / \             \   / \           / \   /              C
8375 C        /   \         /   \             \ /   \         /   \ /               C
8376 C      j1| o |l        | o |              o| o |         | o |o                C
8377 C     \  |/k\|         |/ \|  /            |/ \|         |/ \|                 C
8378 C      \i/   \         /   \ /             /   \         /   \                 C
8379 C       o     k1            o                                                  C
8380 C         (I)          (II)                (III)          (IV)                 C
8381 C                                                                              C
8382 C      eello5_1        eello5_2            eello5_3       eello5_4             C
8383 C                                                                              C
8384 C o denotes a local interaction, vertical lines an electrostatic interaction.  C
8385 C                                                                              C
8386 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
8387 cd      if (i.ne.2 .or. j.ne.6 .or. k.ne.3 .or. l.ne.5) then
8388 cd        eello5=0.0d0
8389 cd        return
8390 cd      endif
8391 cd      write (iout,*)
8392 cd     &   'EELLO5: Contacts have occurred for peptide groups',i,j,
8393 cd     &   ' and',k,l
8394       itk=itortyp(itype(k))
8395       itl=itortyp(itype(l))
8396       itj=itortyp(itype(j))
8397       eello5_1=0.0d0
8398       eello5_2=0.0d0
8399       eello5_3=0.0d0
8400       eello5_4=0.0d0
8401 cd      call checkint5(i,j,k,l,jj,kk,eel5_1_num,eel5_2_num,
8402 cd     &   eel5_3_num,eel5_4_num)
8403       do iii=1,2
8404         do kkk=1,5
8405           do lll=1,3
8406             derx(lll,kkk,iii)=0.0d0
8407           enddo
8408         enddo
8409       enddo
8410 cd      eij=facont_hb(jj,i)
8411 cd      ekl=facont_hb(kk,k)
8412 cd      ekont=eij*ekl
8413 cd      write (iout,*)'Contacts have occurred for peptide groups',
8414 cd     &  i,j,' fcont:',eij,' eij',' and ',k,l
8415 cd      goto 1111
8416 C Contribution from the graph I.
8417 cd      write (2,*) 'AEA  ',AEA(1,1,1),AEA(2,1,1),AEA(1,2,1),AEA(2,2,1)
8418 cd      write (2,*) 'AEAb2',AEAb2(1,1,1),AEAb2(2,1,1)
8419       call transpose2(EUg(1,1,k),auxmat(1,1))
8420       call matmat2(AEA(1,1,1),auxmat(1,1),pizda(1,1))
8421       vv(1)=pizda(1,1)-pizda(2,2)
8422       vv(2)=pizda(1,2)+pizda(2,1)
8423       eello5_1=scalar2(AEAb2(1,1,1),Ub2(1,k))
8424      & +0.5d0*scalar2(vv(1),Dtobr2(1,i))
8425 C Explicit gradient in virtual-dihedral angles.
8426       if (i.gt.1) g_corr5_loc(i-1)=g_corr5_loc(i-1)
8427      & +ekont*(scalar2(AEAb2derg(1,2,1,1),Ub2(1,k))
8428      & +0.5d0*scalar2(vv(1),Dtobr2der(1,i)))
8429       call transpose2(EUgder(1,1,k),auxmat1(1,1))
8430       call matmat2(AEA(1,1,1),auxmat1(1,1),pizda(1,1))
8431       vv(1)=pizda(1,1)-pizda(2,2)
8432       vv(2)=pizda(1,2)+pizda(2,1)
8433       g_corr5_loc(k-1)=g_corr5_loc(k-1)
8434      & +ekont*(scalar2(AEAb2(1,1,1),Ub2der(1,k))
8435      & +0.5d0*scalar2(vv(1),Dtobr2(1,i)))
8436       call matmat2(AEAderg(1,1,1),auxmat(1,1),pizda(1,1))
8437       vv(1)=pizda(1,1)-pizda(2,2)
8438       vv(2)=pizda(1,2)+pizda(2,1)
8439       if (l.eq.j+1) then
8440         if (l.lt.nres-1) g_corr5_loc(l-1)=g_corr5_loc(l-1)
8441      &   +ekont*(scalar2(AEAb2derg(1,1,1,1),Ub2(1,k))
8442      &   +0.5d0*scalar2(vv(1),Dtobr2(1,i)))
8443       else
8444         if (j.lt.nres-1) g_corr5_loc(j-1)=g_corr5_loc(j-1)
8445      &   +ekont*(scalar2(AEAb2derg(1,1,1,1),Ub2(1,k))
8446      &   +0.5d0*scalar2(vv(1),Dtobr2(1,i)))
8447       endif 
8448 C Cartesian gradient
8449       do iii=1,2
8450         do kkk=1,5
8451           do lll=1,3
8452             call matmat2(AEAderx(1,1,lll,kkk,iii,1),auxmat(1,1),
8453      &        pizda(1,1))
8454             vv(1)=pizda(1,1)-pizda(2,2)
8455             vv(2)=pizda(1,2)+pizda(2,1)
8456             derx(lll,kkk,iii)=derx(lll,kkk,iii)
8457      &       +scalar2(AEAb2derx(1,lll,kkk,iii,1,1),Ub2(1,k))
8458      &       +0.5d0*scalar2(vv(1),Dtobr2(1,i))
8459           enddo
8460         enddo
8461       enddo
8462 c      goto 1112
8463 c1111  continue
8464 C Contribution from graph II 
8465       call transpose2(EE(1,1,itk),auxmat(1,1))
8466       call matmat2(auxmat(1,1),AEA(1,1,1),pizda(1,1))
8467       vv(1)=pizda(1,1)+pizda(2,2)
8468       vv(2)=pizda(2,1)-pizda(1,2)
8469       eello5_2=scalar2(AEAb1(1,2,1),b1(1,itk))
8470      & -0.5d0*scalar2(vv(1),Ctobr(1,k))
8471 C Explicit gradient in virtual-dihedral angles.
8472       g_corr5_loc(k-1)=g_corr5_loc(k-1)
8473      & -0.5d0*ekont*scalar2(vv(1),Ctobrder(1,k))
8474       call matmat2(auxmat(1,1),AEAderg(1,1,1),pizda(1,1))
8475       vv(1)=pizda(1,1)+pizda(2,2)
8476       vv(2)=pizda(2,1)-pizda(1,2)
8477       if (l.eq.j+1) then
8478         g_corr5_loc(l-1)=g_corr5_loc(l-1)
8479      &   +ekont*(scalar2(AEAb1derg(1,2,1),b1(1,itk))
8480      &   -0.5d0*scalar2(vv(1),Ctobr(1,k)))
8481       else
8482         g_corr5_loc(j-1)=g_corr5_loc(j-1)
8483      &   +ekont*(scalar2(AEAb1derg(1,2,1),b1(1,itk))
8484      &   -0.5d0*scalar2(vv(1),Ctobr(1,k)))
8485       endif
8486 C Cartesian gradient
8487       do iii=1,2
8488         do kkk=1,5
8489           do lll=1,3
8490             call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,1),
8491      &        pizda(1,1))
8492             vv(1)=pizda(1,1)+pizda(2,2)
8493             vv(2)=pizda(2,1)-pizda(1,2)
8494             derx(lll,kkk,iii)=derx(lll,kkk,iii)
8495      &       +scalar2(AEAb1derx(1,lll,kkk,iii,2,1),b1(1,itk))
8496      &       -0.5d0*scalar2(vv(1),Ctobr(1,k))
8497           enddo
8498         enddo
8499       enddo
8500 cd      goto 1112
8501 cd1111  continue
8502       if (l.eq.j+1) then
8503 cd        goto 1110
8504 C Parallel orientation
8505 C Contribution from graph III
8506         call transpose2(EUg(1,1,l),auxmat(1,1))
8507         call matmat2(AEA(1,1,2),auxmat(1,1),pizda(1,1))
8508         vv(1)=pizda(1,1)-pizda(2,2)
8509         vv(2)=pizda(1,2)+pizda(2,1)
8510         eello5_3=scalar2(AEAb2(1,1,2),Ub2(1,l))
8511      &   +0.5d0*scalar2(vv(1),Dtobr2(1,j))
8512 C Explicit gradient in virtual-dihedral angles.
8513         g_corr5_loc(j-1)=g_corr5_loc(j-1)
8514      &   +ekont*(scalar2(AEAb2derg(1,2,1,2),Ub2(1,l))
8515      &   +0.5d0*scalar2(vv(1),Dtobr2der(1,j)))
8516         call matmat2(AEAderg(1,1,2),auxmat(1,1),pizda(1,1))
8517         vv(1)=pizda(1,1)-pizda(2,2)
8518         vv(2)=pizda(1,2)+pizda(2,1)
8519         g_corr5_loc(k-1)=g_corr5_loc(k-1)
8520      &   +ekont*(scalar2(AEAb2derg(1,1,1,2),Ub2(1,l))
8521      &   +0.5d0*scalar2(vv(1),Dtobr2(1,j)))
8522         call transpose2(EUgder(1,1,l),auxmat1(1,1))
8523         call matmat2(AEA(1,1,2),auxmat1(1,1),pizda(1,1))
8524         vv(1)=pizda(1,1)-pizda(2,2)
8525         vv(2)=pizda(1,2)+pizda(2,1)
8526         g_corr5_loc(l-1)=g_corr5_loc(l-1)
8527      &   +ekont*(scalar2(AEAb2(1,1,2),Ub2der(1,l))
8528      &   +0.5d0*scalar2(vv(1),Dtobr2(1,j)))
8529 C Cartesian gradient
8530         do iii=1,2
8531           do kkk=1,5
8532             do lll=1,3
8533               call matmat2(AEAderx(1,1,lll,kkk,iii,2),auxmat(1,1),
8534      &          pizda(1,1))
8535               vv(1)=pizda(1,1)-pizda(2,2)
8536               vv(2)=pizda(1,2)+pizda(2,1)
8537               derx(lll,kkk,iii)=derx(lll,kkk,iii)
8538      &         +scalar2(AEAb2derx(1,lll,kkk,iii,1,2),Ub2(1,l))
8539      &         +0.5d0*scalar2(vv(1),Dtobr2(1,j))
8540             enddo
8541           enddo
8542         enddo
8543 cd        goto 1112
8544 C Contribution from graph IV
8545 cd1110    continue
8546         call transpose2(EE(1,1,itl),auxmat(1,1))
8547         call matmat2(auxmat(1,1),AEA(1,1,2),pizda(1,1))
8548         vv(1)=pizda(1,1)+pizda(2,2)
8549         vv(2)=pizda(2,1)-pizda(1,2)
8550         eello5_4=scalar2(AEAb1(1,2,2),b1(1,itl))
8551      &   -0.5d0*scalar2(vv(1),Ctobr(1,l))
8552 C Explicit gradient in virtual-dihedral angles.
8553         g_corr5_loc(l-1)=g_corr5_loc(l-1)
8554      &   -0.5d0*ekont*scalar2(vv(1),Ctobrder(1,l))
8555         call matmat2(auxmat(1,1),AEAderg(1,1,2),pizda(1,1))
8556         vv(1)=pizda(1,1)+pizda(2,2)
8557         vv(2)=pizda(2,1)-pizda(1,2)
8558         g_corr5_loc(k-1)=g_corr5_loc(k-1)
8559      &   +ekont*(scalar2(AEAb1derg(1,2,2),b1(1,itl))
8560      &   -0.5d0*scalar2(vv(1),Ctobr(1,l)))
8561 C Cartesian gradient
8562         do iii=1,2
8563           do kkk=1,5
8564             do lll=1,3
8565               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,2),
8566      &          pizda(1,1))
8567               vv(1)=pizda(1,1)+pizda(2,2)
8568               vv(2)=pizda(2,1)-pizda(1,2)
8569               derx(lll,kkk,iii)=derx(lll,kkk,iii)
8570      &         +scalar2(AEAb1derx(1,lll,kkk,iii,2,2),b1(1,itl))
8571      &         -0.5d0*scalar2(vv(1),Ctobr(1,l))
8572             enddo
8573           enddo
8574         enddo
8575       else
8576 C Antiparallel orientation
8577 C Contribution from graph III
8578 c        goto 1110
8579         call transpose2(EUg(1,1,j),auxmat(1,1))
8580         call matmat2(AEA(1,1,2),auxmat(1,1),pizda(1,1))
8581         vv(1)=pizda(1,1)-pizda(2,2)
8582         vv(2)=pizda(1,2)+pizda(2,1)
8583         eello5_3=scalar2(AEAb2(1,1,2),Ub2(1,j))
8584      &   +0.5d0*scalar2(vv(1),Dtobr2(1,l))
8585 C Explicit gradient in virtual-dihedral angles.
8586         g_corr5_loc(l-1)=g_corr5_loc(l-1)
8587      &   +ekont*(scalar2(AEAb2derg(1,2,1,2),Ub2(1,j))
8588      &   +0.5d0*scalar2(vv(1),Dtobr2der(1,l)))
8589         call matmat2(AEAderg(1,1,2),auxmat(1,1),pizda(1,1))
8590         vv(1)=pizda(1,1)-pizda(2,2)
8591         vv(2)=pizda(1,2)+pizda(2,1)
8592         g_corr5_loc(k-1)=g_corr5_loc(k-1)
8593      &   +ekont*(scalar2(AEAb2derg(1,1,1,2),Ub2(1,j))
8594      &   +0.5d0*scalar2(vv(1),Dtobr2(1,l)))
8595         call transpose2(EUgder(1,1,j),auxmat1(1,1))
8596         call matmat2(AEA(1,1,2),auxmat1(1,1),pizda(1,1))
8597         vv(1)=pizda(1,1)-pizda(2,2)
8598         vv(2)=pizda(1,2)+pizda(2,1)
8599         g_corr5_loc(j-1)=g_corr5_loc(j-1)
8600      &   +ekont*(scalar2(AEAb2(1,1,2),Ub2der(1,j))
8601      &   +0.5d0*scalar2(vv(1),Dtobr2(1,l)))
8602 C Cartesian gradient
8603         do iii=1,2
8604           do kkk=1,5
8605             do lll=1,3
8606               call matmat2(AEAderx(1,1,lll,kkk,iii,2),auxmat(1,1),
8607      &          pizda(1,1))
8608               vv(1)=pizda(1,1)-pizda(2,2)
8609               vv(2)=pizda(1,2)+pizda(2,1)
8610               derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)
8611      &         +scalar2(AEAb2derx(1,lll,kkk,iii,1,2),Ub2(1,j))
8612      &         +0.5d0*scalar2(vv(1),Dtobr2(1,l))
8613             enddo
8614           enddo
8615         enddo
8616 cd        goto 1112
8617 C Contribution from graph IV
8618 1110    continue
8619         call transpose2(EE(1,1,itj),auxmat(1,1))
8620         call matmat2(auxmat(1,1),AEA(1,1,2),pizda(1,1))
8621         vv(1)=pizda(1,1)+pizda(2,2)
8622         vv(2)=pizda(2,1)-pizda(1,2)
8623         eello5_4=scalar2(AEAb1(1,2,2),b1(1,itj))
8624      &   -0.5d0*scalar2(vv(1),Ctobr(1,j))
8625 C Explicit gradient in virtual-dihedral angles.
8626         g_corr5_loc(j-1)=g_corr5_loc(j-1)
8627      &   -0.5d0*ekont*scalar2(vv(1),Ctobrder(1,j))
8628         call matmat2(auxmat(1,1),AEAderg(1,1,2),pizda(1,1))
8629         vv(1)=pizda(1,1)+pizda(2,2)
8630         vv(2)=pizda(2,1)-pizda(1,2)
8631         g_corr5_loc(k-1)=g_corr5_loc(k-1)
8632      &   +ekont*(scalar2(AEAb1derg(1,2,2),b1(1,itj))
8633      &   -0.5d0*scalar2(vv(1),Ctobr(1,j)))
8634 C Cartesian gradient
8635         do iii=1,2
8636           do kkk=1,5
8637             do lll=1,3
8638               call matmat2(auxmat(1,1),AEAderx(1,1,lll,kkk,iii,2),
8639      &          pizda(1,1))
8640               vv(1)=pizda(1,1)+pizda(2,2)
8641               vv(2)=pizda(2,1)-pizda(1,2)
8642               derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)
8643      &         +scalar2(AEAb1derx(1,lll,kkk,iii,2,2),b1(1,itj))
8644      &         -0.5d0*scalar2(vv(1),Ctobr(1,j))
8645             enddo
8646           enddo
8647         enddo
8648       endif
8649 1112  continue
8650       eel5=eello5_1+eello5_2+eello5_3+eello5_4
8651 cd      if (i.eq.2 .and. j.eq.8 .and. k.eq.3 .and. l.eq.7) then
8652 cd        write (2,*) 'ijkl',i,j,k,l
8653 cd        write (2,*) 'eello5_1',eello5_1,' eello5_2',eello5_2,
8654 cd     &     ' eello5_3',eello5_3,' eello5_4',eello5_4
8655 cd      endif
8656 cd      write(iout,*) 'eello5_1',eello5_1,' eel5_1_num',16*eel5_1_num
8657 cd      write(iout,*) 'eello5_2',eello5_2,' eel5_2_num',16*eel5_2_num
8658 cd      write(iout,*) 'eello5_3',eello5_3,' eel5_3_num',16*eel5_3_num
8659 cd      write(iout,*) 'eello5_4',eello5_4,' eel5_4_num',16*eel5_4_num
8660       if (j.lt.nres-1) then
8661         j1=j+1
8662         j2=j-1
8663       else
8664         j1=j-1
8665         j2=j-2
8666       endif
8667       if (l.lt.nres-1) then
8668         l1=l+1
8669         l2=l-1
8670       else
8671         l1=l-1
8672         l2=l-2
8673       endif
8674 cd      eij=1.0d0
8675 cd      ekl=1.0d0
8676 cd      ekont=1.0d0
8677 cd      write (2,*) 'eij',eij,' ekl',ekl,' ekont',ekont
8678 C 2/11/08 AL Gradients over DC's connecting interacting sites will be
8679 C        summed up outside the subrouine as for the other subroutines 
8680 C        handling long-range interactions. The old code is commented out
8681 C        with "cgrad" to keep track of changes.
8682       do ll=1,3
8683 cgrad        ggg1(ll)=eel5*g_contij(ll,1)
8684 cgrad        ggg2(ll)=eel5*g_contij(ll,2)
8685         gradcorr5ij=eel5*g_contij(ll,1)+ekont*derx(ll,1,1)
8686         gradcorr5kl=eel5*g_contij(ll,2)+ekont*derx(ll,1,2)
8687 c        write (iout,'(a,3i3,a,5f8.3,2i3,a,5f8.3,a,f8.3)') 
8688 c     &   "ecorr5",ll,i,j," derx",derx(ll,2,1),derx(ll,3,1),derx(ll,4,1),
8689 c     &   derx(ll,5,1),k,l," derx",derx(ll,2,2),derx(ll,3,2),
8690 c     &   derx(ll,4,2),derx(ll,5,2)," ekont",ekont
8691 c        write (iout,'(a,3i3,a,3f8.3,2i3,a,3f8.3)') 
8692 c     &   "ecorr5",ll,i,j," gradcorr5",g_contij(ll,1),derx(ll,1,1),
8693 c     &   gradcorr5ij,
8694 c     &   k,l," gradcorr5",g_contij(ll,2),derx(ll,1,2),gradcorr5kl
8695 cold        ghalf=0.5d0*eel5*ekl*gacont_hbr(ll,jj,i)
8696 cgrad        ghalf=0.5d0*ggg1(ll)
8697 cd        ghalf=0.0d0
8698         gradcorr5(ll,i)=gradcorr5(ll,i)+ekont*derx(ll,2,1)
8699         gradcorr5(ll,i+1)=gradcorr5(ll,i+1)+ekont*derx(ll,3,1)
8700         gradcorr5(ll,j)=gradcorr5(ll,j)+ekont*derx(ll,4,1)
8701         gradcorr5(ll,j1)=gradcorr5(ll,j1)+ekont*derx(ll,5,1)
8702         gradcorr5_long(ll,j)=gradcorr5_long(ll,j)+gradcorr5ij
8703         gradcorr5_long(ll,i)=gradcorr5_long(ll,i)-gradcorr5ij
8704 cold        ghalf=0.5d0*eel5*eij*gacont_hbr(ll,kk,k)
8705 cgrad        ghalf=0.5d0*ggg2(ll)
8706 cd        ghalf=0.0d0
8707         gradcorr5(ll,k)=gradcorr5(ll,k)+ghalf+ekont*derx(ll,2,2)
8708         gradcorr5(ll,k+1)=gradcorr5(ll,k+1)+ekont*derx(ll,3,2)
8709         gradcorr5(ll,l)=gradcorr5(ll,l)+ghalf+ekont*derx(ll,4,2)
8710         gradcorr5(ll,l1)=gradcorr5(ll,l1)+ekont*derx(ll,5,2)
8711         gradcorr5_long(ll,l)=gradcorr5_long(ll,l)+gradcorr5kl
8712         gradcorr5_long(ll,k)=gradcorr5_long(ll,k)-gradcorr5kl
8713       enddo
8714 cd      goto 1112
8715 cgrad      do m=i+1,j-1
8716 cgrad        do ll=1,3
8717 cold          gradcorr5(ll,m)=gradcorr5(ll,m)+eel5*ekl*gacont_hbr(ll,jj,i)
8718 cgrad          gradcorr5(ll,m)=gradcorr5(ll,m)+ggg1(ll)
8719 cgrad        enddo
8720 cgrad      enddo
8721 cgrad      do m=k+1,l-1
8722 cgrad        do ll=1,3
8723 cold          gradcorr5(ll,m)=gradcorr5(ll,m)+eel5*eij*gacont_hbr(ll,kk,k)
8724 cgrad          gradcorr5(ll,m)=gradcorr5(ll,m)+ggg2(ll)
8725 cgrad        enddo
8726 cgrad      enddo
8727 c1112  continue
8728 cgrad      do m=i+2,j2
8729 cgrad        do ll=1,3
8730 cgrad          gradcorr5(ll,m)=gradcorr5(ll,m)+ekont*derx(ll,1,1)
8731 cgrad        enddo
8732 cgrad      enddo
8733 cgrad      do m=k+2,l2
8734 cgrad        do ll=1,3
8735 cgrad          gradcorr5(ll,m)=gradcorr5(ll,m)+ekont*derx(ll,1,2)
8736 cgrad        enddo
8737 cgrad      enddo 
8738 cd      do iii=1,nres-3
8739 cd        write (2,*) iii,g_corr5_loc(iii)
8740 cd      enddo
8741       eello5=ekont*eel5
8742 cd      write (2,*) 'ekont',ekont
8743 cd      write (iout,*) 'eello5',ekont*eel5
8744       return
8745       end
8746 c--------------------------------------------------------------------------
8747       double precision function eello6(i,j,k,l,jj,kk)
8748       implicit real*8 (a-h,o-z)
8749       include 'DIMENSIONS'
8750       include 'COMMON.IOUNITS'
8751       include 'COMMON.CHAIN'
8752       include 'COMMON.DERIV'
8753       include 'COMMON.INTERACT'
8754       include 'COMMON.CONTACTS'
8755       include 'COMMON.TORSION'
8756       include 'COMMON.VAR'
8757       include 'COMMON.GEO'
8758       include 'COMMON.FFIELD'
8759       double precision ggg1(3),ggg2(3)
8760 cd      if (i.ne.1 .or. j.ne.3 .or. k.ne.2 .or. l.ne.4) then
8761 cd        eello6=0.0d0
8762 cd        return
8763 cd      endif
8764 cd      write (iout,*)
8765 cd     &   'EELLO6: Contacts have occurred for peptide groups',i,j,
8766 cd     &   ' and',k,l
8767       eello6_1=0.0d0
8768       eello6_2=0.0d0
8769       eello6_3=0.0d0
8770       eello6_4=0.0d0
8771       eello6_5=0.0d0
8772       eello6_6=0.0d0
8773 cd      call checkint6(i,j,k,l,jj,kk,eel6_1_num,eel6_2_num,
8774 cd     &   eel6_3_num,eel6_4_num,eel6_5_num,eel6_6_num)
8775       do iii=1,2
8776         do kkk=1,5
8777           do lll=1,3
8778             derx(lll,kkk,iii)=0.0d0
8779           enddo
8780         enddo
8781       enddo
8782 cd      eij=facont_hb(jj,i)
8783 cd      ekl=facont_hb(kk,k)
8784 cd      ekont=eij*ekl
8785 cd      eij=1.0d0
8786 cd      ekl=1.0d0
8787 cd      ekont=1.0d0
8788       if (l.eq.j+1) then
8789         eello6_1=eello6_graph1(i,j,k,l,1,.false.)
8790         eello6_2=eello6_graph1(j,i,l,k,2,.false.)
8791         eello6_3=eello6_graph2(i,j,k,l,jj,kk,.false.)
8792         eello6_4=eello6_graph4(i,j,k,l,jj,kk,1,.false.)
8793         eello6_5=eello6_graph4(j,i,l,k,jj,kk,2,.false.)
8794         eello6_6=eello6_graph3(i,j,k,l,jj,kk,.false.)
8795       else
8796         eello6_1=eello6_graph1(i,j,k,l,1,.false.)
8797         eello6_2=eello6_graph1(l,k,j,i,2,.true.)
8798         eello6_3=eello6_graph2(i,l,k,j,jj,kk,.true.)
8799         eello6_4=eello6_graph4(i,j,k,l,jj,kk,1,.false.)
8800         if (wturn6.eq.0.0d0 .or. j.ne.i+4) then
8801           eello6_5=eello6_graph4(l,k,j,i,kk,jj,2,.true.)
8802         else
8803           eello6_5=0.0d0
8804         endif
8805         eello6_6=eello6_graph3(i,l,k,j,jj,kk,.true.)
8806       endif
8807 C If turn contributions are considered, they will be handled separately.
8808       eel6=eello6_1+eello6_2+eello6_3+eello6_4+eello6_5+eello6_6
8809 cd      write(iout,*) 'eello6_1',eello6_1!,' eel6_1_num',16*eel6_1_num
8810 cd      write(iout,*) 'eello6_2',eello6_2!,' eel6_2_num',16*eel6_2_num
8811 cd      write(iout,*) 'eello6_3',eello6_3!,' eel6_3_num',16*eel6_3_num
8812 cd      write(iout,*) 'eello6_4',eello6_4!,' eel6_4_num',16*eel6_4_num
8813 cd      write(iout,*) 'eello6_5',eello6_5!,' eel6_5_num',16*eel6_5_num
8814 cd      write(iout,*) 'eello6_6',eello6_6!,' eel6_6_num',16*eel6_6_num
8815 cd      goto 1112
8816       if (j.lt.nres-1) then
8817         j1=j+1
8818         j2=j-1
8819       else
8820         j1=j-1
8821         j2=j-2
8822       endif
8823       if (l.lt.nres-1) then
8824         l1=l+1
8825         l2=l-1
8826       else
8827         l1=l-1
8828         l2=l-2
8829       endif
8830       do ll=1,3
8831 cgrad        ggg1(ll)=eel6*g_contij(ll,1)
8832 cgrad        ggg2(ll)=eel6*g_contij(ll,2)
8833 cold        ghalf=0.5d0*eel6*ekl*gacont_hbr(ll,jj,i)
8834 cgrad        ghalf=0.5d0*ggg1(ll)
8835 cd        ghalf=0.0d0
8836         gradcorr6ij=eel6*g_contij(ll,1)+ekont*derx(ll,1,1)
8837         gradcorr6kl=eel6*g_contij(ll,2)+ekont*derx(ll,1,2)
8838         gradcorr6(ll,i)=gradcorr6(ll,i)+ekont*derx(ll,2,1)
8839         gradcorr6(ll,i+1)=gradcorr6(ll,i+1)+ekont*derx(ll,3,1)
8840         gradcorr6(ll,j)=gradcorr6(ll,j)+ekont*derx(ll,4,1)
8841         gradcorr6(ll,j1)=gradcorr6(ll,j1)+ekont*derx(ll,5,1)
8842         gradcorr6_long(ll,j)=gradcorr6_long(ll,j)+gradcorr6ij
8843         gradcorr6_long(ll,i)=gradcorr6_long(ll,i)-gradcorr6ij
8844 cgrad        ghalf=0.5d0*ggg2(ll)
8845 cold        ghalf=0.5d0*eel6*eij*gacont_hbr(ll,kk,k)
8846 cd        ghalf=0.0d0
8847         gradcorr6(ll,k)=gradcorr6(ll,k)+ekont*derx(ll,2,2)
8848         gradcorr6(ll,k+1)=gradcorr6(ll,k+1)+ekont*derx(ll,3,2)
8849         gradcorr6(ll,l)=gradcorr6(ll,l)+ekont*derx(ll,4,2)
8850         gradcorr6(ll,l1)=gradcorr6(ll,l1)+ekont*derx(ll,5,2)
8851         gradcorr6_long(ll,l)=gradcorr6_long(ll,l)+gradcorr6kl
8852         gradcorr6_long(ll,k)=gradcorr6_long(ll,k)-gradcorr6kl
8853       enddo
8854 cd      goto 1112
8855 cgrad      do m=i+1,j-1
8856 cgrad        do ll=1,3
8857 cold          gradcorr6(ll,m)=gradcorr6(ll,m)+eel6*ekl*gacont_hbr(ll,jj,i)
8858 cgrad          gradcorr6(ll,m)=gradcorr6(ll,m)+ggg1(ll)
8859 cgrad        enddo
8860 cgrad      enddo
8861 cgrad      do m=k+1,l-1
8862 cgrad        do ll=1,3
8863 cold          gradcorr6(ll,m)=gradcorr6(ll,m)+eel6*eij*gacont_hbr(ll,kk,k)
8864 cgrad          gradcorr6(ll,m)=gradcorr6(ll,m)+ggg2(ll)
8865 cgrad        enddo
8866 cgrad      enddo
8867 cgrad1112  continue
8868 cgrad      do m=i+2,j2
8869 cgrad        do ll=1,3
8870 cgrad          gradcorr6(ll,m)=gradcorr6(ll,m)+ekont*derx(ll,1,1)
8871 cgrad        enddo
8872 cgrad      enddo
8873 cgrad      do m=k+2,l2
8874 cgrad        do ll=1,3
8875 cgrad          gradcorr6(ll,m)=gradcorr6(ll,m)+ekont*derx(ll,1,2)
8876 cgrad        enddo
8877 cgrad      enddo 
8878 cd      do iii=1,nres-3
8879 cd        write (2,*) iii,g_corr6_loc(iii)
8880 cd      enddo
8881       eello6=ekont*eel6
8882 cd      write (2,*) 'ekont',ekont
8883 cd      write (iout,*) 'eello6',ekont*eel6
8884       return
8885       end
8886 c--------------------------------------------------------------------------
8887       double precision function eello6_graph1(i,j,k,l,imat,swap)
8888       implicit real*8 (a-h,o-z)
8889       include 'DIMENSIONS'
8890       include 'COMMON.IOUNITS'
8891       include 'COMMON.CHAIN'
8892       include 'COMMON.DERIV'
8893       include 'COMMON.INTERACT'
8894       include 'COMMON.CONTACTS'
8895       include 'COMMON.TORSION'
8896       include 'COMMON.VAR'
8897       include 'COMMON.GEO'
8898       double precision vv(2),vv1(2),pizda(2,2),auxmat(2,2),pizda1(2,2)
8899       logical swap
8900       logical lprn
8901       common /kutas/ lprn
8902 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
8903 C                                              
8904 C      Parallel       Antiparallel
8905 C                                             
8906 C          o             o         
8907 C         /l\           /j\
8908 C        /   \         /   \
8909 C       /| o |         | o |\
8910 C     \ j|/k\|  /   \  |/k\|l /   
8911 C      \ /   \ /     \ /   \ /    
8912 C       o     o       o     o                
8913 C       i             i                     
8914 C
8915 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
8916       itk=itortyp(itype(k))
8917       s1= scalar2(AEAb1(1,2,imat),CUgb2(1,i))
8918       s2=-scalar2(AEAb2(1,1,imat),Ug2Db1t(1,k))
8919       s3= scalar2(AEAb2(1,1,imat),CUgb2(1,k))
8920       call transpose2(EUgC(1,1,k),auxmat(1,1))
8921       call matmat2(AEA(1,1,imat),auxmat(1,1),pizda1(1,1))
8922       vv1(1)=pizda1(1,1)-pizda1(2,2)
8923       vv1(2)=pizda1(1,2)+pizda1(2,1)
8924       s4=0.5d0*scalar2(vv1(1),Dtobr2(1,i))
8925       vv(1)=AEAb1(1,2,imat)*b1(1,itk)-AEAb1(2,2,imat)*b1(2,itk)
8926       vv(2)=AEAb1(1,2,imat)*b1(2,itk)+AEAb1(2,2,imat)*b1(1,itk)
8927       s5=scalar2(vv(1),Dtobr2(1,i))
8928 cd      write (2,*) 's1',s1,' s2',s2,' s3',s3,' s4', s4,' s5',s5
8929       eello6_graph1=-0.5d0*(s1+s2+s3+s4+s5)
8930       if (i.gt.1) g_corr6_loc(i-1)=g_corr6_loc(i-1)
8931      & -0.5d0*ekont*(scalar2(AEAb1(1,2,imat),CUgb2der(1,i))
8932      & -scalar2(AEAb2derg(1,2,1,imat),Ug2Db1t(1,k))
8933      & +scalar2(AEAb2derg(1,2,1,imat),CUgb2(1,k))
8934      & +0.5d0*scalar2(vv1(1),Dtobr2der(1,i))
8935      & +scalar2(vv(1),Dtobr2der(1,i)))
8936       call matmat2(AEAderg(1,1,imat),auxmat(1,1),pizda1(1,1))
8937       vv1(1)=pizda1(1,1)-pizda1(2,2)
8938       vv1(2)=pizda1(1,2)+pizda1(2,1)
8939       vv(1)=AEAb1derg(1,2,imat)*b1(1,itk)-AEAb1derg(2,2,imat)*b1(2,itk)
8940       vv(2)=AEAb1derg(1,2,imat)*b1(2,itk)+AEAb1derg(2,2,imat)*b1(1,itk)
8941       if (l.eq.j+1) then
8942         g_corr6_loc(l-1)=g_corr6_loc(l-1)
8943      & +ekont*(-0.5d0*(scalar2(AEAb1derg(1,2,imat),CUgb2(1,i))
8944      & -scalar2(AEAb2derg(1,1,1,imat),Ug2Db1t(1,k))
8945      & +scalar2(AEAb2derg(1,1,1,imat),CUgb2(1,k))
8946      & +0.5d0*scalar2(vv1(1),Dtobr2(1,i))+scalar2(vv(1),Dtobr2(1,i))))
8947       else
8948         g_corr6_loc(j-1)=g_corr6_loc(j-1)
8949      & +ekont*(-0.5d0*(scalar2(AEAb1derg(1,2,imat),CUgb2(1,i))
8950      & -scalar2(AEAb2derg(1,1,1,imat),Ug2Db1t(1,k))
8951      & +scalar2(AEAb2derg(1,1,1,imat),CUgb2(1,k))
8952      & +0.5d0*scalar2(vv1(1),Dtobr2(1,i))+scalar2(vv(1),Dtobr2(1,i))))
8953       endif
8954       call transpose2(EUgCder(1,1,k),auxmat(1,1))
8955       call matmat2(AEA(1,1,imat),auxmat(1,1),pizda1(1,1))
8956       vv1(1)=pizda1(1,1)-pizda1(2,2)
8957       vv1(2)=pizda1(1,2)+pizda1(2,1)
8958       if (k.gt.1) g_corr6_loc(k-1)=g_corr6_loc(k-1)
8959      & +ekont*(-0.5d0*(-scalar2(AEAb2(1,1,imat),Ug2Db1tder(1,k))
8960      & +scalar2(AEAb2(1,1,imat),CUgb2der(1,k))
8961      & +0.5d0*scalar2(vv1(1),Dtobr2(1,i))))
8962       do iii=1,2
8963         if (swap) then
8964           ind=3-iii
8965         else
8966           ind=iii
8967         endif
8968         do kkk=1,5
8969           do lll=1,3
8970             s1= scalar2(AEAb1derx(1,lll,kkk,iii,2,imat),CUgb2(1,i))
8971             s2=-scalar2(AEAb2derx(1,lll,kkk,iii,1,imat),Ug2Db1t(1,k))
8972             s3= scalar2(AEAb2derx(1,lll,kkk,iii,1,imat),CUgb2(1,k))
8973             call transpose2(EUgC(1,1,k),auxmat(1,1))
8974             call matmat2(AEAderx(1,1,lll,kkk,iii,imat),auxmat(1,1),
8975      &        pizda1(1,1))
8976             vv1(1)=pizda1(1,1)-pizda1(2,2)
8977             vv1(2)=pizda1(1,2)+pizda1(2,1)
8978             s4=0.5d0*scalar2(vv1(1),Dtobr2(1,i))
8979             vv(1)=AEAb1derx(1,lll,kkk,iii,2,imat)*b1(1,itk)
8980      &       -AEAb1derx(2,lll,kkk,iii,2,imat)*b1(2,itk)
8981             vv(2)=AEAb1derx(1,lll,kkk,iii,2,imat)*b1(2,itk)
8982      &       +AEAb1derx(2,lll,kkk,iii,2,imat)*b1(1,itk)
8983             s5=scalar2(vv(1),Dtobr2(1,i))
8984             derx(lll,kkk,ind)=derx(lll,kkk,ind)-0.5d0*(s1+s2+s3+s4+s5)
8985           enddo
8986         enddo
8987       enddo
8988       return
8989       end
8990 c----------------------------------------------------------------------------
8991       double precision function eello6_graph2(i,j,k,l,jj,kk,swap)
8992       implicit real*8 (a-h,o-z)
8993       include 'DIMENSIONS'
8994       include 'COMMON.IOUNITS'
8995       include 'COMMON.CHAIN'
8996       include 'COMMON.DERIV'
8997       include 'COMMON.INTERACT'
8998       include 'COMMON.CONTACTS'
8999       include 'COMMON.TORSION'
9000       include 'COMMON.VAR'
9001       include 'COMMON.GEO'
9002       logical swap
9003       double precision vv(2),pizda(2,2),auxmat(2,2),auxvec(2),
9004      & auxvec1(2),auxvec2(2),auxmat1(2,2)
9005       logical lprn
9006       common /kutas/ lprn
9007 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9008 C                                                                              C
9009 C      Parallel       Antiparallel                                             C
9010 C                                                                              C
9011 C          o             o                                                     C
9012 C     \   /l\           /j\   /                                                C
9013 C      \ /   \         /   \ /                                                 C
9014 C       o| o |         | o |o                                                  C                
9015 C     \ j|/k\|      \  |/k\|l                                                  C
9016 C      \ /   \       \ /   \                                                   C
9017 C       o             o                                                        C
9018 C       i             i                                                        C 
9019 C                                                                              C           
9020 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9021 cd      write (2,*) 'eello6_graph2: i,',i,' j',j,' k',k,' l',l
9022 C AL 7/4/01 s1 would occur in the sixth-order moment, 
9023 C           but not in a cluster cumulant
9024 #ifdef MOMENT
9025       s1=dip(1,jj,i)*dip(1,kk,k)
9026 #endif
9027       call matvec2(ADtEA1(1,1,1),Ub2(1,k),auxvec(1))
9028       s2=-0.5d0*scalar2(Ub2(1,i),auxvec(1))
9029       call matvec2(ADtEA(1,1,2),Ub2(1,l),auxvec1(1))
9030       s3=-0.5d0*scalar2(Ub2(1,j),auxvec1(1))
9031       call transpose2(EUg(1,1,k),auxmat(1,1))
9032       call matmat2(ADtEA1(1,1,1),auxmat(1,1),pizda(1,1))
9033       vv(1)=pizda(1,1)-pizda(2,2)
9034       vv(2)=pizda(1,2)+pizda(2,1)
9035       s4=-0.25d0*scalar2(vv(1),Dtobr2(1,i))
9036 cd      write (2,*) 'eello6_graph2:','s1',s1,' s2',s2,' s3',s3,' s4',s4
9037 #ifdef MOMENT
9038       eello6_graph2=-(s1+s2+s3+s4)
9039 #else
9040       eello6_graph2=-(s2+s3+s4)
9041 #endif
9042 c      eello6_graph2=-s3
9043 C Derivatives in gamma(i-1)
9044       if (i.gt.1) then
9045 #ifdef MOMENT
9046         s1=dipderg(1,jj,i)*dip(1,kk,k)
9047 #endif
9048         s2=-0.5d0*scalar2(Ub2der(1,i),auxvec(1))
9049         call matvec2(ADtEAderg(1,1,1,2),Ub2(1,l),auxvec2(1))
9050         s3=-0.5d0*scalar2(Ub2(1,j),auxvec2(1))
9051         s4=-0.25d0*scalar2(vv(1),Dtobr2der(1,i))
9052 #ifdef MOMENT
9053         g_corr6_loc(i-1)=g_corr6_loc(i-1)-ekont*(s1+s2+s3+s4)
9054 #else
9055         g_corr6_loc(i-1)=g_corr6_loc(i-1)-ekont*(s2+s3+s4)
9056 #endif
9057 c        g_corr6_loc(i-1)=g_corr6_loc(i-1)-s3
9058       endif
9059 C Derivatives in gamma(k-1)
9060 #ifdef MOMENT
9061       s1=dip(1,jj,i)*dipderg(1,kk,k)
9062 #endif
9063       call matvec2(ADtEA1(1,1,1),Ub2der(1,k),auxvec2(1))
9064       s2=-0.5d0*scalar2(Ub2(1,i),auxvec2(1))
9065       call matvec2(ADtEAderg(1,1,2,2),Ub2(1,l),auxvec2(1))
9066       s3=-0.5d0*scalar2(Ub2(1,j),auxvec2(1))
9067       call transpose2(EUgder(1,1,k),auxmat1(1,1))
9068       call matmat2(ADtEA1(1,1,1),auxmat1(1,1),pizda(1,1))
9069       vv(1)=pizda(1,1)-pizda(2,2)
9070       vv(2)=pizda(1,2)+pizda(2,1)
9071       s4=-0.25d0*scalar2(vv(1),Dtobr2(1,i))
9072 #ifdef MOMENT
9073       g_corr6_loc(k-1)=g_corr6_loc(k-1)-ekont*(s1+s2+s3+s4)
9074 #else
9075       g_corr6_loc(k-1)=g_corr6_loc(k-1)-ekont*(s2+s3+s4)
9076 #endif
9077 c      g_corr6_loc(k-1)=g_corr6_loc(k-1)-s3
9078 C Derivatives in gamma(j-1) or gamma(l-1)
9079       if (j.gt.1) then
9080 #ifdef MOMENT
9081         s1=dipderg(3,jj,i)*dip(1,kk,k) 
9082 #endif
9083         call matvec2(ADtEA1derg(1,1,1,1),Ub2(1,k),auxvec2(1))
9084         s2=-0.5d0*scalar2(Ub2(1,i),auxvec2(1))
9085         s3=-0.5d0*scalar2(Ub2der(1,j),auxvec1(1))
9086         call matmat2(ADtEA1derg(1,1,1,1),auxmat(1,1),pizda(1,1))
9087         vv(1)=pizda(1,1)-pizda(2,2)
9088         vv(2)=pizda(1,2)+pizda(2,1)
9089         s4=-0.25d0*scalar2(vv(1),Dtobr2(1,i))
9090 #ifdef MOMENT
9091         if (swap) then
9092           g_corr6_loc(l-1)=g_corr6_loc(l-1)-ekont*s1
9093         else
9094           g_corr6_loc(j-1)=g_corr6_loc(j-1)-ekont*s1
9095         endif
9096 #endif
9097         g_corr6_loc(j-1)=g_corr6_loc(j-1)-ekont*(s2+s3+s4)
9098 c        g_corr6_loc(j-1)=g_corr6_loc(j-1)-s3
9099       endif
9100 C Derivatives in gamma(l-1) or gamma(j-1)
9101       if (l.gt.1) then 
9102 #ifdef MOMENT
9103         s1=dip(1,jj,i)*dipderg(3,kk,k)
9104 #endif
9105         call matvec2(ADtEA1derg(1,1,2,1),Ub2(1,k),auxvec2(1))
9106         s2=-0.5d0*scalar2(Ub2(1,i),auxvec2(1))
9107         call matvec2(ADtEA(1,1,2),Ub2der(1,l),auxvec2(1))
9108         s3=-0.5d0*scalar2(Ub2(1,j),auxvec2(1))
9109         call matmat2(ADtEA1derg(1,1,2,1),auxmat(1,1),pizda(1,1))
9110         vv(1)=pizda(1,1)-pizda(2,2)
9111         vv(2)=pizda(1,2)+pizda(2,1)
9112         s4=-0.25d0*scalar2(vv(1),Dtobr2(1,i))
9113 #ifdef MOMENT
9114         if (swap) then
9115           g_corr6_loc(j-1)=g_corr6_loc(j-1)-ekont*s1
9116         else
9117           g_corr6_loc(l-1)=g_corr6_loc(l-1)-ekont*s1
9118         endif
9119 #endif
9120         g_corr6_loc(l-1)=g_corr6_loc(l-1)-ekont*(s2+s3+s4)
9121 c        g_corr6_loc(l-1)=g_corr6_loc(l-1)-s3
9122       endif
9123 C Cartesian derivatives.
9124       if (lprn) then
9125         write (2,*) 'In eello6_graph2'
9126         do iii=1,2
9127           write (2,*) 'iii=',iii
9128           do kkk=1,5
9129             write (2,*) 'kkk=',kkk
9130             do jjj=1,2
9131               write (2,'(3(2f10.5),5x)') 
9132      &        ((ADtEA1derx(jjj,mmm,lll,kkk,iii,1),mmm=1,2),lll=1,3)
9133             enddo
9134           enddo
9135         enddo
9136       endif
9137       do iii=1,2
9138         do kkk=1,5
9139           do lll=1,3
9140 #ifdef MOMENT
9141             if (iii.eq.1) then
9142               s1=dipderx(lll,kkk,1,jj,i)*dip(1,kk,k)
9143             else
9144               s1=dip(1,jj,i)*dipderx(lll,kkk,1,kk,k)
9145             endif
9146 #endif
9147             call matvec2(ADtEA1derx(1,1,lll,kkk,iii,1),Ub2(1,k),
9148      &        auxvec(1))
9149             s2=-0.5d0*scalar2(Ub2(1,i),auxvec(1))
9150             call matvec2(ADtEAderx(1,1,lll,kkk,iii,2),Ub2(1,l),
9151      &        auxvec(1))
9152             s3=-0.5d0*scalar2(Ub2(1,j),auxvec(1))
9153             call transpose2(EUg(1,1,k),auxmat(1,1))
9154             call matmat2(ADtEA1derx(1,1,lll,kkk,iii,1),auxmat(1,1),
9155      &        pizda(1,1))
9156             vv(1)=pizda(1,1)-pizda(2,2)
9157             vv(2)=pizda(1,2)+pizda(2,1)
9158             s4=-0.25d0*scalar2(vv(1),Dtobr2(1,i))
9159 cd            write (2,*) 's1',s1,' s2',s2,' s3',s3,' s4',s4
9160 #ifdef MOMENT
9161             derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s1+s2+s4)
9162 #else
9163             derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s2+s4)
9164 #endif
9165             if (swap) then
9166               derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)-s3
9167             else
9168               derx(lll,kkk,iii)=derx(lll,kkk,iii)-s3
9169             endif
9170           enddo
9171         enddo
9172       enddo
9173       return
9174       end
9175 c----------------------------------------------------------------------------
9176       double precision function eello6_graph3(i,j,k,l,jj,kk,swap)
9177       implicit real*8 (a-h,o-z)
9178       include 'DIMENSIONS'
9179       include 'COMMON.IOUNITS'
9180       include 'COMMON.CHAIN'
9181       include 'COMMON.DERIV'
9182       include 'COMMON.INTERACT'
9183       include 'COMMON.CONTACTS'
9184       include 'COMMON.TORSION'
9185       include 'COMMON.VAR'
9186       include 'COMMON.GEO'
9187       double precision vv(2),pizda(2,2),auxmat(2,2),auxvec(2)
9188       logical swap
9189 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9190 C                                                                              C 
9191 C      Parallel       Antiparallel                                             C
9192 C                                                                              C
9193 C          o             o                                                     C 
9194 C         /l\   /   \   /j\                                                    C 
9195 C        /   \ /     \ /   \                                                   C
9196 C       /| o |o       o| o |\                                                  C
9197 C       j|/k\|  /      |/k\|l /                                                C
9198 C        /   \ /       /   \ /                                                 C
9199 C       /     o       /     o                                                  C
9200 C       i             i                                                        C
9201 C                                                                              C
9202 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9203 C
9204 C 4/7/01 AL Component s1 was removed, because it pertains to the respective 
9205 C           energy moment and not to the cluster cumulant.
9206       iti=itortyp(itype(i))
9207       if (j.lt.nres-1) then
9208         itj1=itortyp(itype(j+1))
9209       else
9210         itj1=ntortyp+1
9211       endif
9212       itk=itortyp(itype(k))
9213       itk1=itortyp(itype(k+1))
9214       if (l.lt.nres-1) then
9215         itl1=itortyp(itype(l+1))
9216       else
9217         itl1=ntortyp+1
9218       endif
9219 #ifdef MOMENT
9220       s1=dip(4,jj,i)*dip(4,kk,k)
9221 #endif
9222       call matvec2(AECA(1,1,1),b1(1,itk1),auxvec(1))
9223       s2=0.5d0*scalar2(b1(1,itk),auxvec(1))
9224       call matvec2(AECA(1,1,2),b1(1,itl1),auxvec(1))
9225       s3=0.5d0*scalar2(b1(1,itj1),auxvec(1))
9226       call transpose2(EE(1,1,itk),auxmat(1,1))
9227       call matmat2(auxmat(1,1),AECA(1,1,1),pizda(1,1))
9228       vv(1)=pizda(1,1)+pizda(2,2)
9229       vv(2)=pizda(2,1)-pizda(1,2)
9230       s4=-0.25d0*scalar2(vv(1),Ctobr(1,k))
9231 cd      write (2,*) 'eello6_graph3:','s1',s1,' s2',s2,' s3',s3,' s4',s4,
9232 cd     & "sum",-(s2+s3+s4)
9233 #ifdef MOMENT
9234       eello6_graph3=-(s1+s2+s3+s4)
9235 #else
9236       eello6_graph3=-(s2+s3+s4)
9237 #endif
9238 c      eello6_graph3=-s4
9239 C Derivatives in gamma(k-1)
9240       call matvec2(AECAderg(1,1,2),b1(1,itl1),auxvec(1))
9241       s3=0.5d0*scalar2(b1(1,itj1),auxvec(1))
9242       s4=-0.25d0*scalar2(vv(1),Ctobrder(1,k))
9243       g_corr6_loc(k-1)=g_corr6_loc(k-1)-ekont*(s3+s4)
9244 C Derivatives in gamma(l-1)
9245       call matvec2(AECAderg(1,1,1),b1(1,itk1),auxvec(1))
9246       s2=0.5d0*scalar2(b1(1,itk),auxvec(1))
9247       call matmat2(auxmat(1,1),AECAderg(1,1,1),pizda(1,1))
9248       vv(1)=pizda(1,1)+pizda(2,2)
9249       vv(2)=pizda(2,1)-pizda(1,2)
9250       s4=-0.25d0*scalar2(vv(1),Ctobr(1,k))
9251       g_corr6_loc(l-1)=g_corr6_loc(l-1)-ekont*(s2+s4) 
9252 C Cartesian derivatives.
9253       do iii=1,2
9254         do kkk=1,5
9255           do lll=1,3
9256 #ifdef MOMENT
9257             if (iii.eq.1) then
9258               s1=dipderx(lll,kkk,4,jj,i)*dip(4,kk,k)
9259             else
9260               s1=dip(4,jj,i)*dipderx(lll,kkk,4,kk,k)
9261             endif
9262 #endif
9263             call matvec2(AECAderx(1,1,lll,kkk,iii,1),b1(1,itk1),
9264      &        auxvec(1))
9265             s2=0.5d0*scalar2(b1(1,itk),auxvec(1))
9266             call matvec2(AECAderx(1,1,lll,kkk,iii,2),b1(1,itl1),
9267      &        auxvec(1))
9268             s3=0.5d0*scalar2(b1(1,itj1),auxvec(1))
9269             call matmat2(auxmat(1,1),AECAderx(1,1,lll,kkk,iii,1),
9270      &        pizda(1,1))
9271             vv(1)=pizda(1,1)+pizda(2,2)
9272             vv(2)=pizda(2,1)-pizda(1,2)
9273             s4=-0.25d0*scalar2(vv(1),Ctobr(1,k))
9274 #ifdef MOMENT
9275             derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s1+s2+s4)
9276 #else
9277             derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s2+s4)
9278 #endif
9279             if (swap) then
9280               derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)-s3
9281             else
9282               derx(lll,kkk,iii)=derx(lll,kkk,iii)-s3
9283             endif
9284 c            derx(lll,kkk,iii)=derx(lll,kkk,iii)-s4
9285           enddo
9286         enddo
9287       enddo
9288       return
9289       end
9290 c----------------------------------------------------------------------------
9291       double precision function eello6_graph4(i,j,k,l,jj,kk,imat,swap)
9292       implicit real*8 (a-h,o-z)
9293       include 'DIMENSIONS'
9294       include 'COMMON.IOUNITS'
9295       include 'COMMON.CHAIN'
9296       include 'COMMON.DERIV'
9297       include 'COMMON.INTERACT'
9298       include 'COMMON.CONTACTS'
9299       include 'COMMON.TORSION'
9300       include 'COMMON.VAR'
9301       include 'COMMON.GEO'
9302       include 'COMMON.FFIELD'
9303       double precision vv(2),pizda(2,2),auxmat(2,2),auxvec(2),
9304      & auxvec1(2),auxmat1(2,2)
9305       logical swap
9306 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9307 C                                                                              C                       
9308 C      Parallel       Antiparallel                                             C
9309 C                                                                              C
9310 C          o             o                                                     C
9311 C         /l\   /   \   /j\                                                    C
9312 C        /   \ /     \ /   \                                                   C
9313 C       /| o |o       o| o |\                                                  C
9314 C     \ j|/k\|      \  |/k\|l                                                  C
9315 C      \ /   \       \ /   \                                                   C 
9316 C       o     \       o     \                                                  C
9317 C       i             i                                                        C
9318 C                                                                              C 
9319 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
9320 C
9321 C 4/7/01 AL Component s1 was removed, because it pertains to the respective 
9322 C           energy moment and not to the cluster cumulant.
9323 cd      write (2,*) 'eello_graph4: wturn6',wturn6
9324       iti=itortyp(itype(i))
9325       itj=itortyp(itype(j))
9326       if (j.lt.nres-1) then
9327         itj1=itortyp(itype(j+1))
9328       else
9329         itj1=ntortyp+1
9330       endif
9331       itk=itortyp(itype(k))
9332       if (k.lt.nres-1) then
9333         itk1=itortyp(itype(k+1))
9334       else
9335         itk1=ntortyp+1
9336       endif
9337       itl=itortyp(itype(l))
9338       if (l.lt.nres-1) then
9339         itl1=itortyp(itype(l+1))
9340       else
9341         itl1=ntortyp+1
9342       endif
9343 cd      write (2,*) 'eello6_graph4:','i',i,' j',j,' k',k,' l',l
9344 cd      write (2,*) 'iti',iti,' itj',itj,' itj1',itj1,' itk',itk,
9345 cd     & ' itl',itl,' itl1',itl1
9346 #ifdef MOMENT
9347       if (imat.eq.1) then
9348         s1=dip(3,jj,i)*dip(3,kk,k)
9349       else
9350         s1=dip(2,jj,j)*dip(2,kk,l)
9351       endif
9352 #endif
9353       call matvec2(AECA(1,1,imat),Ub2(1,k),auxvec(1))
9354       s2=0.5d0*scalar2(Ub2(1,i),auxvec(1))
9355       if (j.eq.l+1) then
9356         call matvec2(ADtEA1(1,1,3-imat),b1(1,itj1),auxvec1(1))
9357         s3=-0.5d0*scalar2(b1(1,itj),auxvec1(1))
9358       else
9359         call matvec2(ADtEA1(1,1,3-imat),b1(1,itl1),auxvec1(1))
9360         s3=-0.5d0*scalar2(b1(1,itl),auxvec1(1))
9361       endif
9362       call transpose2(EUg(1,1,k),auxmat(1,1))
9363       call matmat2(AECA(1,1,imat),auxmat(1,1),pizda(1,1))
9364       vv(1)=pizda(1,1)-pizda(2,2)
9365       vv(2)=pizda(2,1)+pizda(1,2)
9366       s4=0.25d0*scalar2(vv(1),Dtobr2(1,i))
9367 cd      write (2,*) 'eello6_graph4:','s1',s1,' s2',s2,' s3',s3,' s4',s4
9368 #ifdef MOMENT
9369       eello6_graph4=-(s1+s2+s3+s4)
9370 #else
9371       eello6_graph4=-(s2+s3+s4)
9372 #endif
9373 C Derivatives in gamma(i-1)
9374       if (i.gt.1) then
9375 #ifdef MOMENT
9376         if (imat.eq.1) then
9377           s1=dipderg(2,jj,i)*dip(3,kk,k)
9378         else
9379           s1=dipderg(4,jj,j)*dip(2,kk,l)
9380         endif
9381 #endif
9382         s2=0.5d0*scalar2(Ub2der(1,i),auxvec(1))
9383         if (j.eq.l+1) then
9384           call matvec2(ADtEA1derg(1,1,1,3-imat),b1(1,itj1),auxvec1(1))
9385           s3=-0.5d0*scalar2(b1(1,itj),auxvec1(1))
9386         else
9387           call matvec2(ADtEA1derg(1,1,1,3-imat),b1(1,itl1),auxvec1(1))
9388           s3=-0.5d0*scalar2(b1(1,itl),auxvec1(1))
9389         endif
9390         s4=0.25d0*scalar2(vv(1),Dtobr2der(1,i))
9391         if (wturn6.gt.0.0d0 .and. k.eq.l+4 .and. i.eq.j+2) then
9392 cd          write (2,*) 'turn6 derivatives'
9393 #ifdef MOMENT
9394           gel_loc_turn6(i-1)=gel_loc_turn6(i-1)-ekont*(s1+s2+s3+s4)
9395 #else
9396           gel_loc_turn6(i-1)=gel_loc_turn6(i-1)-ekont*(s2+s3+s4)
9397 #endif
9398         else
9399 #ifdef MOMENT
9400           g_corr6_loc(i-1)=g_corr6_loc(i-1)-ekont*(s1+s2+s3+s4)
9401 #else
9402           g_corr6_loc(i-1)=g_corr6_loc(i-1)-ekont*(s2+s3+s4)
9403 #endif
9404         endif
9405       endif
9406 C Derivatives in gamma(k-1)
9407 #ifdef MOMENT
9408       if (imat.eq.1) then
9409         s1=dip(3,jj,i)*dipderg(2,kk,k)
9410       else
9411         s1=dip(2,jj,j)*dipderg(4,kk,l)
9412       endif
9413 #endif
9414       call matvec2(AECA(1,1,imat),Ub2der(1,k),auxvec1(1))
9415       s2=0.5d0*scalar2(Ub2(1,i),auxvec1(1))
9416       if (j.eq.l+1) then
9417         call matvec2(ADtEA1derg(1,1,2,3-imat),b1(1,itj1),auxvec1(1))
9418         s3=-0.5d0*scalar2(b1(1,itj),auxvec1(1))
9419       else
9420         call matvec2(ADtEA1derg(1,1,2,3-imat),b1(1,itl1),auxvec1(1))
9421         s3=-0.5d0*scalar2(b1(1,itl),auxvec1(1))
9422       endif
9423       call transpose2(EUgder(1,1,k),auxmat1(1,1))
9424       call matmat2(AECA(1,1,imat),auxmat1(1,1),pizda(1,1))
9425       vv(1)=pizda(1,1)-pizda(2,2)
9426       vv(2)=pizda(2,1)+pizda(1,2)
9427       s4=0.25d0*scalar2(vv(1),Dtobr2(1,i))
9428       if (wturn6.gt.0.0d0 .and. k.eq.l+4 .and. i.eq.j+2) then
9429 #ifdef MOMENT
9430         gel_loc_turn6(k-1)=gel_loc_turn6(k-1)-ekont*(s1+s2+s3+s4)
9431 #else
9432         gel_loc_turn6(k-1)=gel_loc_turn6(k-1)-ekont*(s2+s3+s4)
9433 #endif
9434       else
9435 #ifdef MOMENT
9436         g_corr6_loc(k-1)=g_corr6_loc(k-1)-ekont*(s1+s2+s3+s4)
9437 #else
9438         g_corr6_loc(k-1)=g_corr6_loc(k-1)-ekont*(s2+s3+s4)
9439 #endif
9440       endif
9441 C Derivatives in gamma(j-1) or gamma(l-1)
9442       if (l.eq.j+1 .and. l.gt.1) then
9443         call matvec2(AECAderg(1,1,imat),Ub2(1,k),auxvec(1))
9444         s2=0.5d0*scalar2(Ub2(1,i),auxvec(1))
9445         call matmat2(AECAderg(1,1,imat),auxmat(1,1),pizda(1,1))
9446         vv(1)=pizda(1,1)-pizda(2,2)
9447         vv(2)=pizda(2,1)+pizda(1,2)
9448         s4=0.25d0*scalar2(vv(1),Dtobr2(1,i))
9449         g_corr6_loc(l-1)=g_corr6_loc(l-1)-ekont*(s2+s4)
9450       else if (j.gt.1) then
9451         call matvec2(AECAderg(1,1,imat),Ub2(1,k),auxvec(1))
9452         s2=0.5d0*scalar2(Ub2(1,i),auxvec(1))
9453         call matmat2(AECAderg(1,1,imat),auxmat(1,1),pizda(1,1))
9454         vv(1)=pizda(1,1)-pizda(2,2)
9455         vv(2)=pizda(2,1)+pizda(1,2)
9456         s4=0.25d0*scalar2(vv(1),Dtobr2(1,i))
9457         if (wturn6.gt.0.0d0 .and. k.eq.l+4 .and. i.eq.j+2) then
9458           gel_loc_turn6(j-1)=gel_loc_turn6(j-1)-ekont*(s2+s4)
9459         else
9460           g_corr6_loc(j-1)=g_corr6_loc(j-1)-ekont*(s2+s4)
9461         endif
9462       endif
9463 C Cartesian derivatives.
9464       do iii=1,2
9465         do kkk=1,5
9466           do lll=1,3
9467 #ifdef MOMENT
9468             if (iii.eq.1) then
9469               if (imat.eq.1) then
9470                 s1=dipderx(lll,kkk,3,jj,i)*dip(3,kk,k)
9471               else
9472                 s1=dipderx(lll,kkk,2,jj,j)*dip(2,kk,l)
9473               endif
9474             else
9475               if (imat.eq.1) then
9476                 s1=dip(3,jj,i)*dipderx(lll,kkk,3,kk,k)
9477               else
9478                 s1=dip(2,jj,j)*dipderx(lll,kkk,2,kk,l)
9479               endif
9480             endif
9481 #endif
9482             call matvec2(AECAderx(1,1,lll,kkk,iii,imat),Ub2(1,k),
9483      &        auxvec(1))
9484             s2=0.5d0*scalar2(Ub2(1,i),auxvec(1))
9485             if (j.eq.l+1) then
9486               call matvec2(ADtEA1derx(1,1,lll,kkk,iii,3-imat),
9487      &          b1(1,itj1),auxvec(1))
9488               s3=-0.5d0*scalar2(b1(1,itj),auxvec(1))
9489             else
9490               call matvec2(ADtEA1derx(1,1,lll,kkk,iii,3-imat),
9491      &          b1(1,itl1),auxvec(1))
9492               s3=-0.5d0*scalar2(b1(1,itl),auxvec(1))
9493             endif
9494             call matmat2(AECAderx(1,1,lll,kkk,iii,imat),auxmat(1,1),
9495      &        pizda(1,1))
9496             vv(1)=pizda(1,1)-pizda(2,2)
9497             vv(2)=pizda(2,1)+pizda(1,2)
9498             s4=0.25d0*scalar2(vv(1),Dtobr2(1,i))
9499             if (swap) then
9500               if (wturn6.gt.0.0d0 .and. k.eq.l+4 .and. i.eq.j+2) then
9501 #ifdef MOMENT
9502                 derx_turn(lll,kkk,3-iii)=derx_turn(lll,kkk,3-iii)
9503      &             -(s1+s2+s4)
9504 #else
9505                 derx_turn(lll,kkk,3-iii)=derx_turn(lll,kkk,3-iii)
9506      &             -(s2+s4)
9507 #endif
9508                 derx_turn(lll,kkk,iii)=derx_turn(lll,kkk,iii)-s3
9509               else
9510 #ifdef MOMENT
9511                 derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)-(s1+s2+s4)
9512 #else
9513                 derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)-(s2+s4)
9514 #endif
9515                 derx(lll,kkk,iii)=derx(lll,kkk,iii)-s3
9516               endif
9517             else
9518 #ifdef MOMENT
9519               derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s1+s2+s4)
9520 #else
9521               derx(lll,kkk,iii)=derx(lll,kkk,iii)-(s2+s4)
9522 #endif
9523               if (l.eq.j+1) then
9524                 derx(lll,kkk,iii)=derx(lll,kkk,iii)-s3
9525               else 
9526                 derx(lll,kkk,3-iii)=derx(lll,kkk,3-iii)-s3
9527               endif
9528             endif 
9529           enddo
9530         enddo
9531       enddo
9532       return
9533       end
9534 c----------------------------------------------------------------------------
9535       double precision function eello_turn6(i,jj,kk)
9536       implicit real*8 (a-h,o-z)
9537       include 'DIMENSIONS'
9538       include 'COMMON.IOUNITS'
9539       include 'COMMON.CHAIN'
9540       include 'COMMON.DERIV'
9541       include 'COMMON.INTERACT'
9542       include 'COMMON.CONTACTS'
9543       include 'COMMON.TORSION'
9544       include 'COMMON.VAR'
9545       include 'COMMON.GEO'
9546       double precision vtemp1(2),vtemp2(2),vtemp3(2),vtemp4(2),
9547      &  atemp(2,2),auxmat(2,2),achuj_temp(2,2),gtemp(2,2),gvec(2),
9548      &  ggg1(3),ggg2(3)
9549       double precision vtemp1d(2),vtemp2d(2),vtemp3d(2),vtemp4d(2),
9550      &  atempd(2,2),auxmatd(2,2),achuj_tempd(2,2),gtempd(2,2),gvecd(2)
9551 C 4/7/01 AL Components s1, s8, and s13 were removed, because they pertain to
9552 C           the respective energy moment and not to the cluster cumulant.
9553       s1=0.0d0
9554       s8=0.0d0
9555       s13=0.0d0
9556 c
9557       eello_turn6=0.0d0
9558       j=i+4
9559       k=i+1
9560       l=i+3
9561       iti=itortyp(itype(i))
9562       itk=itortyp(itype(k))
9563       itk1=itortyp(itype(k+1))
9564       itl=itortyp(itype(l))
9565       itj=itortyp(itype(j))
9566 cd      write (2,*) 'itk',itk,' itk1',itk1,' itl',itl,' itj',itj
9567 cd      write (2,*) 'i',i,' k',k,' j',j,' l',l
9568 cd      if (i.ne.1 .or. j.ne.3 .or. k.ne.2 .or. l.ne.4) then
9569 cd        eello6=0.0d0
9570 cd        return
9571 cd      endif
9572 cd      write (iout,*)
9573 cd     &   'EELLO6: Contacts have occurred for peptide groups',i,j,
9574 cd     &   ' and',k,l
9575 cd      call checkint_turn6(i,jj,kk,eel_turn6_num)
9576       do iii=1,2
9577         do kkk=1,5
9578           do lll=1,3
9579             derx_turn(lll,kkk,iii)=0.0d0
9580           enddo
9581         enddo
9582       enddo
9583 cd      eij=1.0d0
9584 cd      ekl=1.0d0
9585 cd      ekont=1.0d0
9586       eello6_5=eello6_graph4(l,k,j,i,kk,jj,2,.true.)
9587 cd      eello6_5=0.0d0
9588 cd      write (2,*) 'eello6_5',eello6_5
9589 #ifdef MOMENT
9590       call transpose2(AEA(1,1,1),auxmat(1,1))
9591       call matmat2(EUg(1,1,i+1),auxmat(1,1),auxmat(1,1))
9592       ss1=scalar2(Ub2(1,i+2),b1(1,itl))
9593       s1 = (auxmat(1,1)+auxmat(2,2))*ss1
9594 #endif
9595       call matvec2(EUg(1,1,i+2),b1(1,itl),vtemp1(1))
9596       call matvec2(AEA(1,1,1),vtemp1(1),vtemp1(1))
9597       s2 = scalar2(b1(1,itk),vtemp1(1))
9598 #ifdef MOMENT
9599       call transpose2(AEA(1,1,2),atemp(1,1))
9600       call matmat2(atemp(1,1),EUg(1,1,i+4),atemp(1,1))
9601       call matvec2(Ug2(1,1,i+2),dd(1,1,itk1),vtemp2(1))
9602       s8 = -(atemp(1,1)+atemp(2,2))*scalar2(cc(1,1,itl),vtemp2(1))
9603 #endif
9604       call matmat2(EUg(1,1,i+3),AEA(1,1,2),auxmat(1,1))
9605       call matvec2(auxmat(1,1),Ub2(1,i+4),vtemp3(1))
9606       s12 = scalar2(Ub2(1,i+2),vtemp3(1))
9607 #ifdef MOMENT
9608       call transpose2(a_chuj(1,1,kk,i+1),achuj_temp(1,1))
9609       call matmat2(achuj_temp(1,1),EUg(1,1,i+2),gtemp(1,1))
9610       call matmat2(gtemp(1,1),EUg(1,1,i+3),gtemp(1,1)) 
9611       call matvec2(a_chuj(1,1,jj,i),Ub2(1,i+4),vtemp4(1)) 
9612       ss13 = scalar2(b1(1,itk),vtemp4(1))
9613       s13 = (gtemp(1,1)+gtemp(2,2))*ss13
9614 #endif
9615 c      write (2,*) 's1,s2,s8,s12,s13',s1,s2,s8,s12,s13
9616 c      s1=0.0d0
9617 c      s2=0.0d0
9618 c      s8=0.0d0
9619 c      s12=0.0d0
9620 c      s13=0.0d0
9621       eel_turn6 = eello6_5 - 0.5d0*(s1+s2+s12+s8+s13)
9622 C Derivatives in gamma(i+2)
9623       s1d =0.0d0
9624       s8d =0.0d0
9625 #ifdef MOMENT
9626       call transpose2(AEA(1,1,1),auxmatd(1,1))
9627       call matmat2(EUgder(1,1,i+1),auxmatd(1,1),auxmatd(1,1))
9628       s1d = (auxmatd(1,1)+auxmatd(2,2))*ss1
9629       call transpose2(AEAderg(1,1,2),atempd(1,1))
9630       call matmat2(atempd(1,1),EUg(1,1,i+4),atempd(1,1))
9631       s8d = -(atempd(1,1)+atempd(2,2))*scalar2(cc(1,1,itl),vtemp2(1))
9632 #endif
9633       call matmat2(EUg(1,1,i+3),AEAderg(1,1,2),auxmatd(1,1))
9634       call matvec2(auxmatd(1,1),Ub2(1,i+4),vtemp3d(1))
9635       s12d = scalar2(Ub2(1,i+2),vtemp3d(1))
9636 c      s1d=0.0d0
9637 c      s2d=0.0d0
9638 c      s8d=0.0d0
9639 c      s12d=0.0d0
9640 c      s13d=0.0d0
9641       gel_loc_turn6(i)=gel_loc_turn6(i)-0.5d0*ekont*(s1d+s8d+s12d)
9642 C Derivatives in gamma(i+3)
9643 #ifdef MOMENT
9644       call transpose2(AEA(1,1,1),auxmatd(1,1))
9645       call matmat2(EUg(1,1,i+1),auxmatd(1,1),auxmatd(1,1))
9646       ss1d=scalar2(Ub2der(1,i+2),b1(1,itl))
9647       s1d = (auxmatd(1,1)+auxmatd(2,2))*ss1d
9648 #endif
9649       call matvec2(EUgder(1,1,i+2),b1(1,itl),vtemp1d(1))
9650       call matvec2(AEA(1,1,1),vtemp1d(1),vtemp1d(1))
9651       s2d = scalar2(b1(1,itk),vtemp1d(1))
9652 #ifdef MOMENT
9653       call matvec2(Ug2der(1,1,i+2),dd(1,1,itk1),vtemp2d(1))
9654       s8d = -(atemp(1,1)+atemp(2,2))*scalar2(cc(1,1,itl),vtemp2d(1))
9655 #endif
9656       s12d = scalar2(Ub2der(1,i+2),vtemp3(1))
9657 #ifdef MOMENT
9658       call matmat2(achuj_temp(1,1),EUgder(1,1,i+2),gtempd(1,1))
9659       call matmat2(gtempd(1,1),EUg(1,1,i+3),gtempd(1,1)) 
9660       s13d = (gtempd(1,1)+gtempd(2,2))*ss13
9661 #endif
9662 c      s1d=0.0d0
9663 c      s2d=0.0d0
9664 c      s8d=0.0d0
9665 c      s12d=0.0d0
9666 c      s13d=0.0d0
9667 #ifdef MOMENT
9668       gel_loc_turn6(i+1)=gel_loc_turn6(i+1)
9669      &               -0.5d0*ekont*(s1d+s2d+s8d+s12d+s13d)
9670 #else
9671       gel_loc_turn6(i+1)=gel_loc_turn6(i+1)
9672      &               -0.5d0*ekont*(s2d+s12d)
9673 #endif
9674 C Derivatives in gamma(i+4)
9675       call matmat2(EUgder(1,1,i+3),AEA(1,1,2),auxmatd(1,1))
9676       call matvec2(auxmatd(1,1),Ub2(1,i+4),vtemp3d(1))
9677       s12d = scalar2(Ub2(1,i+2),vtemp3d(1))
9678 #ifdef MOMENT
9679       call matmat2(achuj_temp(1,1),EUg(1,1,i+2),gtempd(1,1))
9680       call matmat2(gtempd(1,1),EUgder(1,1,i+3),gtempd(1,1)) 
9681       s13d = (gtempd(1,1)+gtempd(2,2))*ss13
9682 #endif
9683 c      s1d=0.0d0
9684 c      s2d=0.0d0
9685 c      s8d=0.0d0
9686 C      s12d=0.0d0
9687 c      s13d=0.0d0
9688 #ifdef MOMENT
9689       gel_loc_turn6(i+2)=gel_loc_turn6(i+2)-0.5d0*ekont*(s12d+s13d)
9690 #else
9691       gel_loc_turn6(i+2)=gel_loc_turn6(i+2)-0.5d0*ekont*(s12d)
9692 #endif
9693 C Derivatives in gamma(i+5)
9694 #ifdef MOMENT
9695       call transpose2(AEAderg(1,1,1),auxmatd(1,1))
9696       call matmat2(EUg(1,1,i+1),auxmatd(1,1),auxmatd(1,1))
9697       s1d = (auxmatd(1,1)+auxmatd(2,2))*ss1
9698 #endif
9699       call matvec2(EUg(1,1,i+2),b1(1,itl),vtemp1d(1))
9700       call matvec2(AEAderg(1,1,1),vtemp1d(1),vtemp1d(1))
9701       s2d = scalar2(b1(1,itk),vtemp1d(1))
9702 #ifdef MOMENT
9703       call transpose2(AEA(1,1,2),atempd(1,1))
9704       call matmat2(atempd(1,1),EUgder(1,1,i+4),atempd(1,1))
9705       s8d = -(atempd(1,1)+atempd(2,2))*scalar2(cc(1,1,itl),vtemp2(1))
9706 #endif
9707       call matvec2(auxmat(1,1),Ub2der(1,i+4),vtemp3d(1))
9708       s12d = scalar2(Ub2(1,i+2),vtemp3d(1))
9709 #ifdef MOMENT
9710       call matvec2(a_chuj(1,1,jj,i),Ub2der(1,i+4),vtemp4d(1)) 
9711       ss13d = scalar2(b1(1,itk),vtemp4d(1))
9712       s13d = (gtemp(1,1)+gtemp(2,2))*ss13d
9713 #endif
9714 c      s1d=0.0d0
9715 c      s2d=0.0d0
9716 c      s8d=0.0d0
9717 c      s12d=0.0d0
9718 c      s13d=0.0d0
9719 #ifdef MOMENT
9720       gel_loc_turn6(i+3)=gel_loc_turn6(i+3)
9721      &               -0.5d0*ekont*(s1d+s2d+s8d+s12d+s13d)
9722 #else
9723       gel_loc_turn6(i+3)=gel_loc_turn6(i+3)
9724      &               -0.5d0*ekont*(s2d+s12d)
9725 #endif
9726 C Cartesian derivatives
9727       do iii=1,2
9728         do kkk=1,5
9729           do lll=1,3
9730 #ifdef MOMENT
9731             call transpose2(AEAderx(1,1,lll,kkk,iii,1),auxmatd(1,1))
9732             call matmat2(EUg(1,1,i+1),auxmatd(1,1),auxmatd(1,1))
9733             s1d = (auxmatd(1,1)+auxmatd(2,2))*ss1
9734 #endif
9735             call matvec2(EUg(1,1,i+2),b1(1,itl),vtemp1(1))
9736             call matvec2(AEAderx(1,1,lll,kkk,iii,1),vtemp1(1),
9737      &          vtemp1d(1))
9738             s2d = scalar2(b1(1,itk),vtemp1d(1))
9739 #ifdef MOMENT
9740             call transpose2(AEAderx(1,1,lll,kkk,iii,2),atempd(1,1))
9741             call matmat2(atempd(1,1),EUg(1,1,i+4),atempd(1,1))
9742             s8d = -(atempd(1,1)+atempd(2,2))*
9743      &           scalar2(cc(1,1,itl),vtemp2(1))
9744 #endif
9745             call matmat2(EUg(1,1,i+3),AEAderx(1,1,lll,kkk,iii,2),
9746      &           auxmatd(1,1))
9747             call matvec2(auxmatd(1,1),Ub2(1,i+4),vtemp3d(1))
9748             s12d = scalar2(Ub2(1,i+2),vtemp3d(1))
9749 c      s1d=0.0d0
9750 c      s2d=0.0d0
9751 c      s8d=0.0d0
9752 c      s12d=0.0d0
9753 c      s13d=0.0d0
9754 #ifdef MOMENT
9755             derx_turn(lll,kkk,iii) = derx_turn(lll,kkk,iii) 
9756      &        - 0.5d0*(s1d+s2d)
9757 #else
9758             derx_turn(lll,kkk,iii) = derx_turn(lll,kkk,iii) 
9759      &        - 0.5d0*s2d
9760 #endif
9761 #ifdef MOMENT
9762             derx_turn(lll,kkk,3-iii) = derx_turn(lll,kkk,3-iii) 
9763      &        - 0.5d0*(s8d+s12d)
9764 #else
9765             derx_turn(lll,kkk,3-iii) = derx_turn(lll,kkk,3-iii) 
9766      &        - 0.5d0*s12d
9767 #endif
9768           enddo
9769         enddo
9770       enddo
9771 #ifdef MOMENT
9772       do kkk=1,5
9773         do lll=1,3
9774           call transpose2(a_chuj_der(1,1,lll,kkk,kk,i+1),
9775      &      achuj_tempd(1,1))
9776           call matmat2(achuj_tempd(1,1),EUg(1,1,i+2),gtempd(1,1))
9777           call matmat2(gtempd(1,1),EUg(1,1,i+3),gtempd(1,1)) 
9778           s13d=(gtempd(1,1)+gtempd(2,2))*ss13
9779           derx_turn(lll,kkk,2) = derx_turn(lll,kkk,2)-0.5d0*s13d
9780           call matvec2(a_chuj_der(1,1,lll,kkk,jj,i),Ub2(1,i+4),
9781      &      vtemp4d(1)) 
9782           ss13d = scalar2(b1(1,itk),vtemp4d(1))
9783           s13d = (gtemp(1,1)+gtemp(2,2))*ss13d
9784           derx_turn(lll,kkk,1) = derx_turn(lll,kkk,1)-0.5d0*s13d
9785         enddo
9786       enddo
9787 #endif
9788 cd      write(iout,*) 'eel6_turn6',eel_turn6,' eel_turn6_num',
9789 cd     &  16*eel_turn6_num
9790 cd      goto 1112
9791       if (j.lt.nres-1) then
9792         j1=j+1
9793         j2=j-1
9794       else
9795         j1=j-1
9796         j2=j-2
9797       endif
9798       if (l.lt.nres-1) then
9799         l1=l+1
9800         l2=l-1
9801       else
9802         l1=l-1
9803         l2=l-2
9804       endif
9805       do ll=1,3
9806 cgrad        ggg1(ll)=eel_turn6*g_contij(ll,1)
9807 cgrad        ggg2(ll)=eel_turn6*g_contij(ll,2)
9808 cgrad        ghalf=0.5d0*ggg1(ll)
9809 cd        ghalf=0.0d0
9810         gturn6ij=eel_turn6*g_contij(ll,1)+ekont*derx_turn(ll,1,1)
9811         gturn6kl=eel_turn6*g_contij(ll,2)+ekont*derx_turn(ll,1,2)
9812         gcorr6_turn(ll,i)=gcorr6_turn(ll,i)!+ghalf
9813      &    +ekont*derx_turn(ll,2,1)
9814         gcorr6_turn(ll,i+1)=gcorr6_turn(ll,i+1)+ekont*derx_turn(ll,3,1)
9815         gcorr6_turn(ll,j)=gcorr6_turn(ll,j)!+ghalf
9816      &    +ekont*derx_turn(ll,4,1)
9817         gcorr6_turn(ll,j1)=gcorr6_turn(ll,j1)+ekont*derx_turn(ll,5,1)
9818         gcorr6_turn_long(ll,j)=gcorr6_turn_long(ll,j)+gturn6ij
9819         gcorr6_turn_long(ll,i)=gcorr6_turn_long(ll,i)-gturn6ij
9820 cgrad        ghalf=0.5d0*ggg2(ll)
9821 cd        ghalf=0.0d0
9822         gcorr6_turn(ll,k)=gcorr6_turn(ll,k)!+ghalf
9823      &    +ekont*derx_turn(ll,2,2)
9824         gcorr6_turn(ll,k+1)=gcorr6_turn(ll,k+1)+ekont*derx_turn(ll,3,2)
9825         gcorr6_turn(ll,l)=gcorr6_turn(ll,l)!+ghalf
9826      &    +ekont*derx_turn(ll,4,2)
9827         gcorr6_turn(ll,l1)=gcorr6_turn(ll,l1)+ekont*derx_turn(ll,5,2)
9828         gcorr6_turn_long(ll,l)=gcorr6_turn_long(ll,l)+gturn6kl
9829         gcorr6_turn_long(ll,k)=gcorr6_turn_long(ll,k)-gturn6kl
9830       enddo
9831 cd      goto 1112
9832 cgrad      do m=i+1,j-1
9833 cgrad        do ll=1,3
9834 cgrad          gcorr6_turn(ll,m)=gcorr6_turn(ll,m)+ggg1(ll)
9835 cgrad        enddo
9836 cgrad      enddo
9837 cgrad      do m=k+1,l-1
9838 cgrad        do ll=1,3
9839 cgrad          gcorr6_turn(ll,m)=gcorr6_turn(ll,m)+ggg2(ll)
9840 cgrad        enddo
9841 cgrad      enddo
9842 cgrad1112  continue
9843 cgrad      do m=i+2,j2
9844 cgrad        do ll=1,3
9845 cgrad          gcorr6_turn(ll,m)=gcorr6_turn(ll,m)+ekont*derx_turn(ll,1,1)
9846 cgrad        enddo
9847 cgrad      enddo
9848 cgrad      do m=k+2,l2
9849 cgrad        do ll=1,3
9850 cgrad          gcorr6_turn(ll,m)=gcorr6_turn(ll,m)+ekont*derx_turn(ll,1,2)
9851 cgrad        enddo
9852 cgrad      enddo 
9853 cd      do iii=1,nres-3
9854 cd        write (2,*) iii,g_corr6_loc(iii)
9855 cd      enddo
9856       eello_turn6=ekont*eel_turn6
9857 cd      write (2,*) 'ekont',ekont
9858 cd      write (2,*) 'eel_turn6',ekont*eel_turn6
9859       return
9860       end
9861
9862 C-----------------------------------------------------------------------------
9863       double precision function scalar(u,v)
9864 !DIR$ INLINEALWAYS scalar
9865 #ifndef OSF
9866 cDEC$ ATTRIBUTES FORCEINLINE::scalar
9867 #endif
9868       implicit none
9869       double precision u(3),v(3)
9870 cd      double precision sc
9871 cd      integer i
9872 cd      sc=0.0d0
9873 cd      do i=1,3
9874 cd        sc=sc+u(i)*v(i)
9875 cd      enddo
9876 cd      scalar=sc
9877
9878       scalar=u(1)*v(1)+u(2)*v(2)+u(3)*v(3)
9879       return
9880       end
9881 crc-------------------------------------------------
9882       SUBROUTINE MATVEC2(A1,V1,V2)
9883 !DIR$ INLINEALWAYS MATVEC2
9884 #ifndef OSF
9885 cDEC$ ATTRIBUTES FORCEINLINE::MATVEC2
9886 #endif
9887       implicit real*8 (a-h,o-z)
9888       include 'DIMENSIONS'
9889       DIMENSION A1(2,2),V1(2),V2(2)
9890 c      DO 1 I=1,2
9891 c        VI=0.0
9892 c        DO 3 K=1,2
9893 c    3     VI=VI+A1(I,K)*V1(K)
9894 c        Vaux(I)=VI
9895 c    1 CONTINUE
9896
9897       vaux1=a1(1,1)*v1(1)+a1(1,2)*v1(2)
9898       vaux2=a1(2,1)*v1(1)+a1(2,2)*v1(2)
9899
9900       v2(1)=vaux1
9901       v2(2)=vaux2
9902       END
9903 C---------------------------------------
9904       SUBROUTINE MATMAT2(A1,A2,A3)
9905 #ifndef OSF
9906 cDEC$ ATTRIBUTES FORCEINLINE::MATMAT2  
9907 #endif
9908       implicit real*8 (a-h,o-z)
9909       include 'DIMENSIONS'
9910       DIMENSION A1(2,2),A2(2,2),A3(2,2)
9911 c      DIMENSION AI3(2,2)
9912 c        DO  J=1,2
9913 c          A3IJ=0.0
9914 c          DO K=1,2
9915 c           A3IJ=A3IJ+A1(I,K)*A2(K,J)
9916 c          enddo
9917 c          A3(I,J)=A3IJ
9918 c       enddo
9919 c      enddo
9920
9921       ai3_11=a1(1,1)*a2(1,1)+a1(1,2)*a2(2,1)
9922       ai3_12=a1(1,1)*a2(1,2)+a1(1,2)*a2(2,2)
9923       ai3_21=a1(2,1)*a2(1,1)+a1(2,2)*a2(2,1)
9924       ai3_22=a1(2,1)*a2(1,2)+a1(2,2)*a2(2,2)
9925
9926       A3(1,1)=AI3_11
9927       A3(2,1)=AI3_21
9928       A3(1,2)=AI3_12
9929       A3(2,2)=AI3_22
9930       END
9931
9932 c-------------------------------------------------------------------------
9933       double precision function scalar2(u,v)
9934 !DIR$ INLINEALWAYS scalar2
9935       implicit none
9936       double precision u(2),v(2)
9937       double precision sc
9938       integer i
9939       scalar2=u(1)*v(1)+u(2)*v(2)
9940       return
9941       end
9942
9943 C-----------------------------------------------------------------------------
9944
9945       subroutine transpose2(a,at)
9946 !DIR$ INLINEALWAYS transpose2
9947 #ifndef OSF
9948 cDEC$ ATTRIBUTES FORCEINLINE::transpose2
9949 #endif
9950       implicit none
9951       double precision a(2,2),at(2,2)
9952       at(1,1)=a(1,1)
9953       at(1,2)=a(2,1)
9954       at(2,1)=a(1,2)
9955       at(2,2)=a(2,2)
9956       return
9957       end
9958 c--------------------------------------------------------------------------
9959       subroutine transpose(n,a,at)
9960       implicit none
9961       integer n,i,j
9962       double precision a(n,n),at(n,n)
9963       do i=1,n
9964         do j=1,n
9965           at(j,i)=a(i,j)
9966         enddo
9967       enddo
9968       return
9969       end
9970 C---------------------------------------------------------------------------
9971       subroutine prodmat3(a1,a2,kk,transp,prod)
9972 !DIR$ INLINEALWAYS prodmat3
9973 #ifndef OSF
9974 cDEC$ ATTRIBUTES FORCEINLINE::prodmat3
9975 #endif
9976       implicit none
9977       integer i,j
9978       double precision a1(2,2),a2(2,2),a2t(2,2),kk(2,2),prod(2,2)
9979       logical transp
9980 crc      double precision auxmat(2,2),prod_(2,2)
9981
9982       if (transp) then
9983 crc        call transpose2(kk(1,1),auxmat(1,1))
9984 crc        call matmat2(a1(1,1),auxmat(1,1),auxmat(1,1))
9985 crc        call matmat2(auxmat(1,1),a2(1,1),prod_(1,1)) 
9986         
9987            prod(1,1)=(a1(1,1)*kk(1,1)+a1(1,2)*kk(1,2))*a2(1,1)
9988      & +(a1(1,1)*kk(2,1)+a1(1,2)*kk(2,2))*a2(2,1)
9989            prod(1,2)=(a1(1,1)*kk(1,1)+a1(1,2)*kk(1,2))*a2(1,2)
9990      & +(a1(1,1)*kk(2,1)+a1(1,2)*kk(2,2))*a2(2,2)
9991            prod(2,1)=(a1(2,1)*kk(1,1)+a1(2,2)*kk(1,2))*a2(1,1)
9992      & +(a1(2,1)*kk(2,1)+a1(2,2)*kk(2,2))*a2(2,1)
9993            prod(2,2)=(a1(2,1)*kk(1,1)+a1(2,2)*kk(1,2))*a2(1,2)
9994      & +(a1(2,1)*kk(2,1)+a1(2,2)*kk(2,2))*a2(2,2)
9995
9996       else
9997 crc        call matmat2(a1(1,1),kk(1,1),auxmat(1,1))
9998 crc        call matmat2(auxmat(1,1),a2(1,1),prod_(1,1))
9999
10000            prod(1,1)=(a1(1,1)*kk(1,1)+a1(1,2)*kk(2,1))*a2(1,1)
10001      &  +(a1(1,1)*kk(1,2)+a1(1,2)*kk(2,2))*a2(2,1)
10002            prod(1,2)=(a1(1,1)*kk(1,1)+a1(1,2)*kk(2,1))*a2(1,2)
10003      &  +(a1(1,1)*kk(1,2)+a1(1,2)*kk(2,2))*a2(2,2)
10004            prod(2,1)=(a1(2,1)*kk(1,1)+a1(2,2)*kk(2,1))*a2(1,1)
10005      &  +(a1(2,1)*kk(1,2)+a1(2,2)*kk(2,2))*a2(2,1)
10006            prod(2,2)=(a1(2,1)*kk(1,1)+a1(2,2)*kk(2,1))*a2(1,2)
10007      &  +(a1(2,1)*kk(1,2)+a1(2,2)*kk(2,2))*a2(2,2)
10008
10009       endif
10010 c      call transpose2(a2(1,1),a2t(1,1))
10011
10012 crc      print *,transp
10013 crc      print *,((prod_(i,j),i=1,2),j=1,2)
10014 crc      print *,((prod(i,j),i=1,2),j=1,2)
10015
10016       return
10017       end
10018