examples/smpi/NAS/DT-trace/dt.c

   1 /*************************************************************************
   2  *                                                                       *
   3  *        N  A  S     P A R A L L E L     B E N C H M A R K S  3.3       *
   4  *                                                                       *
   5  *                                  D T                                  *
   6  *                                                                       *
   7  *************************************************************************
   8  *                                                                       *
   9  *   This benchmark is part of the NAS Parallel Benchmark 3.3 suite.     *
  10  *                                                                       *
  11  *   Permission to use, copy, distribute and modify this software        *
  12  *   for any purpose with or without fee is hereby granted.  We          *
  13  *   request, however, that all derived work reference the NAS           *
  14  *   Parallel Benchmarks 3.3. This software is provided "as is"          *
  15  *   without express or implied warranty.                                *
  16  *                                                                       *
  17  *   Information on NPB 3.3, including the technical report, the         *
  18  *   original specifications, source code, results and information       *
  19  *   on how to submit new results, is available at:                      *
  20  *                                                                       *
  21  *          http:  www.nas.nasa.gov/Software/NPB                         *
  22  *                                                                       *
  23  *   Send comments or suggestions to  npb@nas.nasa.gov                   *
  24  *   Send bug reports to              npb-bugs@nas.nasa.gov              *
  25  *                                                                       *
  26  *         NAS Parallel Benchmarks Group                                 *
  27  *         NASA Ames Research Center                                     *
  28  *         Mail Stop: T27A-1                                             *
  29  *         Moffett Field, CA   94035-1000                                *
  30  *                                                                       *
  31  *         E-mail:  npb@nas.nasa.gov                                     *
  32  *         Fax:     (650) 604-3957                                       *
  33  *                                                                       *
  34  *************************************************************************
  35  *                                                                       *
  36  *   Author: M. Frumkin                                                  *                                               *
  37  *                                                                       *
  38  *************************************************************************/
  39
  40 #include <stdlib.h>
  41 #include <stdio.h>
  42 #include <string.h>
  43
  44 #include "mpi.h"
  45 #include "npbparams.h"
  46
  47 #include "instr/instr.h" //TRACE_
  48
  49 #ifndef CLASS
  50 #define CLASS 'S'
  51 #define NUM_PROCS            1
  52 #endif
  53
  54 //int      passed_verification;
  55 extern double randlc( double *X, double *A );
  56 extern
  57 void c_print_results( char   *name,
  58                       char   class,
  59                       int    n1,
  60                       int    n2,
  61                       int    n3,
  62                       int    niter,
  63                       int    nprocs_compiled,
  64                       int    nprocs_total,
  65                       double t,
  66                       double mops,
  67                       char   *optype,
  68                       int    passed_verification,
  69                       char   *npbversion,
  70                       char   *compiletime,
  71                       char   *mpicc,
  72                       char   *clink,
  73                       char   *cmpi_lib,
  74                       char   *cmpi_inc,
  75                       char   *cflags,
  76                       char   *clinkflags );
  77
  78 void    timer_clear( int n );
  79 void    timer_start( int n );
  80 void    timer_stop( int n );
  81 double  timer_read( int n );
  82 int timer_on=0,timers_tot=64;
  83
  84 int verify(char *bmname,double rnm2){
  85     double verify_value=0.0;
  86     double epsilon=1.0E-8;
  87     char cls=CLASS;
  88     int verified=-1;
  89     if (cls != 'U') {
  90        if(cls=='S') {
  91          if(strstr(bmname,"BH")){
  92            verify_value=30892725.0;
  93          }else if(strstr(bmname,"WH")){
  94            verify_value=67349758.0;
  95          }else if(strstr(bmname,"SH")){
  96            verify_value=58875767.0;
  97          }else{
  98            fprintf(stderr,"No such benchmark as %s.\n",bmname);
  99          }
 100          verified = 0;
 101        }else if(cls=='W') {
 102          if(strstr(bmname,"BH")){
 103            verify_value = 4102461.0;
 104          }else if(strstr(bmname,"WH")){
 105            verify_value = 204280762.0;
 106          }else if(strstr(bmname,"SH")){
 107            verify_value = 186944764.0;
 108          }else{
 109            fprintf(stderr,"No such benchmark as %s.\n",bmname);
 110          }
 111          verified = 0;
 112        }else if(cls=='A') {
 113          if(strstr(bmname,"BH")){
 114            verify_value = 17809491.0;
 115          }else if(strstr(bmname,"WH")){
 116            verify_value = 1289925229.0;
 117          }else if(strstr(bmname,"SH")){
 118            verify_value = 610856482.0;
 119          }else{
 120            fprintf(stderr,"No such benchmark as %s.\n",bmname);
 121          }
 122          verified = 0;
 123        }else if(cls=='B') {
 124          if(strstr(bmname,"BH")){
 125            verify_value = 4317114.0;
 126          }else if(strstr(bmname,"WH")){
 127            verify_value = 7877279917.0;
 128          }else if(strstr(bmname,"SH")){
 129            verify_value = 1836863082.0;
 130          }else{
 131            fprintf(stderr,"No such benchmark as %s.\n",bmname);
 132            verified = 0;
 133          }
 134        }else if(cls=='C') {
 135          if(strstr(bmname,"BH")){
 136            verify_value = 0.0;
 137          }else if(strstr(bmname,"WH")){
 138            verify_value = 0.0;
 139          }else if(strstr(bmname,"SH")){
 140            verify_value = 0.0;
 141          }else{
 142            fprintf(stderr,"No such benchmark as %s.\n",bmname);
 143            verified = -1;
 144          }
 145        }else if(cls=='D') {
 146          if(strstr(bmname,"BH")){
 147            verify_value = 0.0;
 148          }else if(strstr(bmname,"WH")){
 149            verify_value = 0.0;
 150          }else if(strstr(bmname,"SH")){
 151            verify_value = 0.0;
 152          }else{
 153            fprintf(stderr,"No such benchmark as %s.\n",bmname);
 154          }
 155          verified = -1;
 156        }else{
 157          fprintf(stderr,"No such class as %c.\n",cls);
 158        }
 159        fprintf(stderr," %s L2 Norm = %f\n",bmname,rnm2);
 160        if(verified==-1){
 161          fprintf(stderr," No verification was performed.\n");
 162        }else if( rnm2 - verify_value < epsilon &&
 163                  rnm2 - verify_value > -epsilon) {  /* abs here does not work on ALTIX */
 164           verified = 1;
 165           fprintf(stderr," Deviation = %f\n",(rnm2 - verify_value));
 166        }else{
 167          verified = 0;
 168          fprintf(stderr," The correct verification value = %f\n",verify_value);
 169          fprintf(stderr," Got value = %f\n",rnm2);
 170        }
 171     }else{
 172        verified = -1;
 173     }
 174     return  verified;
 175   }
 176
 177 int ipowMod(int a,long long int n,int md){
 178   int seed=1,q=a,r=1;
 179   if(n<0){
 180     fprintf(stderr,"ipowMod: exponent must be nonnegative exp=%lld\n",n);
 181     n=-n; /* temp fix */
 182 /*    return 1; */
 183   }
 184   if(md<=0){
 185     fprintf(stderr,"ipowMod: module must be positive mod=%d",md);
 186     return 1;
 187   }
 188   if(n==0) return 1;
 189   while(n>1){
 190     int n2 = n/2;
 191     if (n2*2==n){
 192        seed = (q*q)%md;
 193        q=seed;
 194        n = n2;
 195     }else{
 196        seed = (r*q)%md;
 197        r=seed;
 198        n = n-1;
 199     }
 200   }
 201   seed = (r*q)%md;
 202   return seed;
 203 }
 204
 205 #include "DGraph.h"
 206 DGraph *buildSH(char cls){
 207 /*
 208   Nodes of the graph must be topologically sorted
 209   to avoid MPI deadlock.
 210 */
 211   DGraph *dg;
 212   int numSources=NUM_SOURCES; /* must be power of 2 */
 213   int numOfLayers=0,tmpS=numSources>>1;
 214   int firstLayerNode=0;
 215   DGArc *ar=NULL;
 216   DGNode *nd=NULL;
 217   int mask=0x0,ndid=0,ndoff=0;
 218   int i=0,j=0;
 219   char nm[BLOCK_SIZE];
 220
 221   sprintf(nm,"DT_SH.%c",cls);
 222   dg=newDGraph(nm);
 223
 224   while(tmpS>1){
 225     numOfLayers++;
 226     tmpS>>=1;
 227   }
 228   for(i=0;i<numSources;i++){
 229     sprintf(nm,"Source.%d",i);
 230     nd=newNode(nm);
 231     AttachNode(dg,nd);
 232   }
 233   for(j=0;j<numOfLayers;j++){
 234     mask=0x00000001<<j;
 235     for(i=0;i<numSources;i++){
 236       sprintf(nm,"Comparator.%d",(i+j*firstLayerNode));
 237       nd=newNode(nm);
 238       AttachNode(dg,nd);
 239       ndoff=i&(~mask);
 240       ndid=firstLayerNode+ndoff;
 241       ar=newArc(dg->node[ndid],nd);
 242       AttachArc(dg,ar);
 243       ndoff+=mask;
 244       ndid=firstLayerNode+ndoff;
 245       ar=newArc(dg->node[ndid],nd);
 246       AttachArc(dg,ar);
 247     }
 248     firstLayerNode+=numSources;
 249   }
 250   mask=0x00000001<<numOfLayers;
 251   for(i=0;i<numSources;i++){
 252     sprintf(nm,"Sink.%d",i);
 253     nd=newNode(nm);
 254     AttachNode(dg,nd);
 255     ndoff=i&(~mask);
 256     ndid=firstLayerNode+ndoff;
 257     ar=newArc(dg->node[ndid],nd);
 258     AttachArc(dg,ar);
 259     ndoff+=mask;
 260     ndid=firstLayerNode+ndoff;
 261     ar=newArc(dg->node[ndid],nd);
 262     AttachArc(dg,ar);
 263   }
 264 return dg;
 265 }
 266 DGraph *buildWH(char cls){
 267 /*
 268   Nodes of the graph must be topologically sorted
 269   to avoid MPI deadlock.
 270 */
 271   int i=0,j=0;
 272   int numSources=NUM_SOURCES,maxInDeg=4;
 273   int numLayerNodes=numSources,firstLayerNode=0;
 274   int totComparators=0;
 275   int numPrevLayerNodes=numLayerNodes;
 276   int id=0,sid=0;
 277   DGraph *dg;
 278   DGNode *nd=NULL,*source=NULL,*tmp=NULL,*snd=NULL;
 279   DGArc *ar=NULL;
 280   char nm[BLOCK_SIZE];
 281
 282   sprintf(nm,"DT_WH.%c",cls);
 283   dg=newDGraph(nm);
 284
 285   for(i=0;i<numSources;i++){
 286     sprintf(nm,"Sink.%d",i);
 287     nd=newNode(nm);
 288     AttachNode(dg,nd);
 289   }
 290   totComparators=0;
 291   numPrevLayerNodes=numLayerNodes;
 292   while(numLayerNodes>maxInDeg){
 293     numLayerNodes=numLayerNodes/maxInDeg;
 294     if(numLayerNodes*maxInDeg<numPrevLayerNodes)numLayerNodes++;
 295     for(i=0;i<numLayerNodes;i++){
 296       sprintf(nm,"Comparator.%d",totComparators);
 297       totComparators++;
 298       nd=newNode(nm);
 299       id=AttachNode(dg,nd);
 300       for(j=0;j<maxInDeg;j++){
 301         sid=i*maxInDeg+j;
 302         if(sid>=numPrevLayerNodes) break;
 303         snd=dg->node[firstLayerNode+sid];
 304         ar=newArc(dg->node[id],snd);
 305         AttachArc(dg,ar);
 306       }
 307     }
 308     firstLayerNode+=numPrevLayerNodes;
 309     numPrevLayerNodes=numLayerNodes;
 310   }
 311   source=newNode("Source");
 312   AttachNode(dg,source);
 313   for(i=0;i<numPrevLayerNodes;i++){
 314     nd=dg->node[firstLayerNode+i];
 315     ar=newArc(source,nd);
 316     AttachArc(dg,ar);
 317   }
 318
 319   for(i=0;i<dg->numNodes/2;i++){  /* Topological sorting */
 320     tmp=dg->node[i];
 321     dg->node[i]=dg->node[dg->numNodes-1-i];
 322     dg->node[i]->id=i;
 323     dg->node[dg->numNodes-1-i]=tmp;
 324     dg->node[dg->numNodes-1-i]->id=dg->numNodes-1-i;
 325   }
 326 return dg;
 327 }
 328 DGraph *buildBH(char cls){
 329 /*
 330   Nodes of the graph must be topologically sorted
 331   to avoid MPI deadlock.
 332 */
 333   int i=0,j=0;
 334   int numSources=NUM_SOURCES,maxInDeg=4;
 335   int numLayerNodes=numSources,firstLayerNode=0;
 336   DGraph *dg;
 337   DGNode *nd=NULL, *snd=NULL, *sink=NULL;
 338   DGArc *ar=NULL;
 339   int totComparators=0;
 340   int numPrevLayerNodes=numLayerNodes;
 341   int id=0, sid=0;
 342   char nm[BLOCK_SIZE];
 343
 344   sprintf(nm,"DT_BH.%c",cls);
 345   dg=newDGraph(nm);
 346
 347   for(i=0;i<numSources;i++){
 348     sprintf(nm,"Source.%d",i);
 349     nd=newNode(nm);
 350     AttachNode(dg,nd);
 351   }
 352   while(numLayerNodes>maxInDeg){
 353     numLayerNodes=numLayerNodes/maxInDeg;
 354     if(numLayerNodes*maxInDeg<numPrevLayerNodes)numLayerNodes++;
 355     for(i=0;i<numLayerNodes;i++){
 356       sprintf(nm,"Comparator.%d",totComparators);
 357       totComparators++;
 358       nd=newNode(nm);
 359       id=AttachNode(dg,nd);
 360       for(j=0;j<maxInDeg;j++){
 361         sid=i*maxInDeg+j;
 362         if(sid>=numPrevLayerNodes) break;
 363         snd=dg->node[firstLayerNode+sid];
 364         ar=newArc(snd,dg->node[id]);
 365         AttachArc(dg,ar);
 366       }
 367     }
 368     firstLayerNode+=numPrevLayerNodes;
 369     numPrevLayerNodes=numLayerNodes;
 370   }
 371   sink=newNode("Sink");
 372   AttachNode(dg,sink);
 373   for(i=0;i<numPrevLayerNodes;i++){
 374     nd=dg->node[firstLayerNode+i];
 375     ar=newArc(nd,sink);
 376     AttachArc(dg,ar);
 377   }
 378 return dg;
 379 }
 380
 381 typedef struct{
 382   int len;
 383   double* val;
 384 } Arr;
 385 Arr *newArr(int len){
 386   Arr *arr=(Arr *)SMPI_SHARED_MALLOC(sizeof(Arr));
 387   arr->len=len;
 388   arr->val=(double *)SMPI_SHARED_MALLOC(len*sizeof(double));
 389   return arr;
 390 }
 391 void arrShow(Arr* a){
 392   if(!a) fprintf(stderr,"-- NULL array\n");
 393   else{
 394     fprintf(stderr,"-- length=%d\n",a->len);
 395   }
 396 }
 397 double CheckVal(Arr *feat){
 398   double csum=0.0;
 399   int i=0;
 400   for(i=0;i<feat->len;i++){
 401     csum+=feat->val[i]*feat->val[i]/feat->len; /* The truncation does not work since
 402                                                   result will be 0 for large len  */
 403   }
 404    return csum;
 405 }
 406 int GetFNumDPar(int* mean, int* stdev){
 407   *mean=NUM_SAMPLES;
 408   *stdev=STD_DEVIATION;
 409   return 0;
 410 }
 411 int GetFeatureNum(char *mbname,int id){
 412   double tran=314159265.0;
 413   double A=2*id+1;
 414   double denom=randlc(&tran,&A);
 415   char cval='S';
 416   int mean=NUM_SAMPLES,stdev=128;
 417   int rtfs=0,len=0;
 418   GetFNumDPar(&mean,&stdev);
 419   rtfs=ipowMod((int)(1/denom)*(int)cval,(long long int) (2*id+1),2*stdev);
 420   if(rtfs<0) rtfs=-rtfs;
 421   len=mean-stdev+rtfs;
 422   return len;
 423 }
 424 Arr* RandomFeatures(char *bmname,int fdim,int id){
 425   int len=GetFeatureNum(bmname,id)*fdim;
 426   Arr* feat=newArr(len);
 427   int nxg=2,nyg=2,nzg=2,nfg=5;
 428   int nx=421,ny=419,nz=1427,nf=3527;
 429   long long int expon=(len*(id+1))%3141592;
 430   int seedx=ipowMod(nxg,expon,nx),
 431       seedy=ipowMod(nyg,expon,ny),
 432       seedz=ipowMod(nzg,expon,nz),
 433       seedf=ipowMod(nfg,expon,nf);
 434   int i=0;
 435   if(timer_on){
 436     timer_clear(id+1);
 437     timer_start(id+1);
 438   }
 439   for(i=0;i<len;i+=fdim){
 440     seedx=(seedx*nxg)%nx;
 441     seedy=(seedy*nyg)%ny;
 442     seedz=(seedz*nzg)%nz;
 443     seedf=(seedf*nfg)%nf;
 444     feat->val[i]=seedx;
 445     feat->val[i+1]=seedy;
 446     feat->val[i+2]=seedz;
 447     feat->val[i+3]=seedf;
 448   }
 449   if(timer_on){
 450     timer_stop(id+1);
 451     fprintf(stderr,"** RandomFeatures time in node %d = %f\n",id,timer_read(id+1));
 452   }
 453   return feat;
 454 }
 455 void Resample(Arr *a,int blen){
 456     long long int i=0,j=0,jlo=0,jhi=0;
 457     double avval=0.0;
 458     double *nval=(double *)SMPI_SHARED_MALLOC(blen*sizeof(double));
 459     Arr *tmp=newArr(10);
 460     for(i=0;i<blen;i++) nval[i]=0.0;
 461     for(i=1;i<a->len-1;i++){
 462       jlo=(int)(0.5*(2*i-1)*(blen/a->len));
 463       jhi=(int)(0.5*(2*i+1)*(blen/a->len));
 464
 465       avval=a->val[i]/(jhi-jlo+1);
 466       for(j=jlo;j<=jhi;j++){
 467         nval[j]+=avval;
 468       }
 469     }
 470     nval[0]=a->val[0];
 471     nval[blen-1]=a->val[a->len-1];
 472     SMPI_SHARED_FREE(a->val);
 473     a->val=nval;
 474     a->len=blen;
 475 }
 476 #define fielddim 4
 477 Arr* WindowFilter(Arr *a, Arr* b,int w){
 478   int i=0,j=0,k=0;
 479   double rms0=0.0,rms1=0.0,rmsm1=0.0;
 480   double weight=((double) (w+1))/(w+2);
 481
 482   w+=1;
 483   if(timer_on){
 484     timer_clear(w);
 485     timer_start(w);
 486   }
 487   if(a->len<b->len) Resample(a,b->len);
 488   if(a->len>b->len) Resample(b,a->len);
 489   for(i=fielddim;i<a->len-fielddim;i+=fielddim){
 490     rms0=(a->val[i]-b->val[i])*(a->val[i]-b->val[i])
 491         +(a->val[i+1]-b->val[i+1])*(a->val[i+1]-b->val[i+1])
 492         +(a->val[i+2]-b->val[i+2])*(a->val[i+2]-b->val[i+2])
 493         +(a->val[i+3]-b->val[i+3])*(a->val[i+3]-b->val[i+3]);
 494     j=i+fielddim;
 495     rms1=(a->val[j]-b->val[j])*(a->val[j]-b->val[j])
 496         +(a->val[j+1]-b->val[j+1])*(a->val[j+1]-b->val[j+1])
 497         +(a->val[j+2]-b->val[j+2])*(a->val[j+2]-b->val[j+2])
 498         +(a->val[j+3]-b->val[j+3])*(a->val[j+3]-b->val[j+3]);
 499     j=i-fielddim;
 500     rmsm1=(a->val[j]-b->val[j])*(a->val[j]-b->val[j])
 501          +(a->val[j+1]-b->val[j+1])*(a->val[j+1]-b->val[j+1])
 502          +(a->val[j+2]-b->val[j+2])*(a->val[j+2]-b->val[j+2])
 503          +(a->val[j+3]-b->val[j+3])*(a->val[j+3]-b->val[j+3]);
 504     k=0;
 505     if(rms1<rms0){
 506       k=1;
 507       rms0=rms1;
 508     }
 509     if(rmsm1<rms0) k=-1;
 510     if(k==0){
 511       j=i+fielddim;
 512       a->val[i]=weight*b->val[i];
 513       a->val[i+1]=weight*b->val[i+1];
 514       a->val[i+2]=weight*b->val[i+2];
 515       a->val[i+3]=weight*b->val[i+3];
 516     }else if(k==1){
 517       j=i+fielddim;
 518       a->val[i]=weight*b->val[j];
 519       a->val[i+1]=weight*b->val[j+1];
 520       a->val[i+2]=weight*b->val[j+2];
 521       a->val[i+3]=weight*b->val[j+3];
 522     }else { /*if(k==-1)*/
 523       j=i-fielddim;
 524       a->val[i]=weight*b->val[j];
 525       a->val[i+1]=weight*b->val[j+1];
 526       a->val[i+2]=weight*b->val[j+2];
 527       a->val[i+3]=weight*b->val[j+3];
 528     }
 529   }
 530   if(timer_on){
 531     timer_stop(w);
 532     fprintf(stderr,"** WindowFilter time in node %d = %f\n",(w-1),timer_read(w));
 533   }
 534   return a;
 535 }
 536
 537 int SendResults(DGraph *dg,DGNode *nd,Arr *feat){
 538   int i=0,tag=0;
 539   DGArc *ar=NULL;
 540   DGNode *head=NULL;
 541   if(!feat) return 0;
 542   for(i=0;i<nd->outDegree;i++){
 543     ar=nd->outArc[i];
 544     if(ar->tail!=nd) continue;
 545     head=ar->head;
 546     tag=ar->id;
 547     if(head->address!=nd->address){
 548       MPI_Send(&feat->len,1,MPI_INT,head->address,tag,MPI_COMM_WORLD);
 549       MPI_Send(feat->val,feat->len,MPI_DOUBLE,head->address,tag,MPI_COMM_WORLD);
 550     }
 551   }
 552   return 1;
 553 }
 554 Arr* CombineStreams(DGraph *dg,DGNode *nd){
 555   Arr *resfeat=newArr(NUM_SAMPLES*fielddim);
 556   int i=0,len=0,tag=0;
 557   DGArc *ar=NULL;
 558   DGNode *tail=NULL;
 559   MPI_Status status;
 560   Arr *feat=NULL,*featp=NULL;
 561
 562   if(nd->inDegree==0) return NULL;
 563   for(i=0;i<nd->inDegree;i++){
 564     ar=nd->inArc[i];
 565     if(ar->head!=nd) continue;
 566     tail=ar->tail;
 567     if(tail->address!=nd->address){
 568       len=0;
 569       tag=ar->id;
 570       MPI_Recv(&len,1,MPI_INT,tail->address,tag,MPI_COMM_WORLD,&status);
 571       feat=newArr(len);
 572       MPI_Recv(feat->val,feat->len,MPI_DOUBLE,tail->address,tag,MPI_COMM_WORLD,&status);
 573       resfeat=WindowFilter(resfeat,feat,nd->id);
 574       SMPI_SHARED_FREE(feat);
 575     }else{
 576       featp=(Arr *)tail->feat;
 577       feat=newArr(featp->len);
 578       memcpy(feat->val,featp->val,featp->len*sizeof(double));
 579       resfeat=WindowFilter(resfeat,feat,nd->id);
 580       SMPI_SHARED_FREE(feat);
 581     }
 582   }
 583   for(i=0;i<resfeat->len;i++) resfeat->val[i]=((int)resfeat->val[i])/nd->inDegree;
 584   nd->feat=resfeat;
 585   return nd->feat;
 586 }
 587 double Reduce(Arr *a,int w){
 588   double retv=0.0;
 589   if(timer_on){
 590     timer_clear(w);
 591     timer_start(w);
 592   }
 593   retv=(int)(w*CheckVal(a));/* The casting needed for node
 594                                and array dependent verifcation */
 595   if(timer_on){
 596     timer_stop(w);
 597     fprintf(stderr,"** Reduce time in node %d = %f\n",(w-1),timer_read(w));
 598   }
 599   return retv;
 600 }
 601
 602 double ReduceStreams(DGraph *dg,DGNode *nd){
 603   double csum=0.0;
 604   int i=0,len=0,tag=0;
 605   DGArc *ar=NULL;
 606   DGNode *tail=NULL;
 607   Arr *feat=NULL;
 608   double retv=0.0;
 609
 610   for(i=0;i<nd->inDegree;i++){
 611     ar=nd->inArc[i];
 612     if(ar->head!=nd) continue;
 613     tail=ar->tail;
 614     if(tail->address!=nd->address){
 615       MPI_Status status;
 616       len=0;
 617       tag=ar->id;
 618       MPI_Recv(&len,1,MPI_INT,tail->address,tag,MPI_COMM_WORLD,&status);
 619       feat=newArr(len);
 620       MPI_Recv(feat->val,feat->len,MPI_DOUBLE,tail->address,tag,MPI_COMM_WORLD,&status);
 621       csum+=Reduce(feat,(nd->id+1));
 622       SMPI_SHARED_FREE(feat);
 623     }else{
 624       csum+=Reduce(tail->feat,(nd->id+1));
 625     }
 626   }
 627   if(nd->inDegree>0)csum=(((long long int)csum)/nd->inDegree);
 628   retv=(nd->id+1)*csum;
 629   return retv;
 630 }
 631
 632 int ProcessNodes(DGraph *dg,int me){
 633   double chksum=0.0;
 634   Arr *feat=NULL;
 635   int i=0,verified=0,tag;
 636   DGNode *nd=NULL;
 637   double rchksum=0.0;
 638   MPI_Status status;
 639
 640   for(i=0;i<dg->numNodes;i++){
 641     nd=dg->node[i];
 642     if(nd->address!=me) continue;
 643     if(strstr(nd->name,"Source")){
 644       nd->feat=RandomFeatures(dg->name,fielddim,nd->id);
 645       SendResults(dg,nd,nd->feat);
 646     }else if(strstr(nd->name,"Sink")){
 647       chksum=ReduceStreams(dg,nd);
 648       tag=dg->numArcs+nd->id; /* make these to avoid clash with arc tags */
 649       MPI_Send(&chksum,1,MPI_DOUBLE,0,tag,MPI_COMM_WORLD);
 650     }else{
 651       feat=CombineStreams(dg,nd);
 652       SendResults(dg,nd,feat);
 653     }
 654   }
 655   if(me==0){ /* Report node */
 656     rchksum=0.0;
 657     chksum=0.0;
 658     for(i=0;i<dg->numNodes;i++){
 659       nd=dg->node[i];
 660       if(!strstr(nd->name,"Sink")) continue;
 661        tag=dg->numArcs+nd->id; /* make these to avoid clash with arc tags */
 662       MPI_Recv(&rchksum,1,MPI_DOUBLE,nd->address,tag,MPI_COMM_WORLD,&status);
 663       chksum+=rchksum;
 664     }
 665     verified=verify(dg->name,chksum);
 666   }
 667 return verified;
 668 }
 669
 670 int main(int argc,char **argv ){
 671   int my_rank,comm_size;
 672   int i;
 673   DGraph *dg=NULL;
 674   int verified=0, featnum=0;
 675   double bytes_sent=2.0,tot_time=0.0;
 676
 677
 678
 679     MPI_Init( &argc, &argv );
 680     MPI_Comm_rank( MPI_COMM_WORLD, &my_rank );
 681     MPI_Comm_size( MPI_COMM_WORLD, &comm_size );
 682     TRACE_smpi_set_category ("start");
 683
 684      if(argc!=2||
 685                 (  strncmp(argv[1],"BH",2)!=0
 686                  &&strncmp(argv[1],"WH",2)!=0
 687                  &&strncmp(argv[1],"SH",2)!=0
 688                 )
 689       ){
 690       if(my_rank==0){
 691         fprintf(stderr,"** Usage: mpirun -np N ../bin/dt.S GraphName\n");
 692         fprintf(stderr,"** Where \n   - N is integer number of MPI processes\n");
 693         fprintf(stderr,"   - S is the class S, W, or A \n");
 694         fprintf(stderr,"   - GraphName is the communication graph name BH, WH, or SH.\n");
 695         fprintf(stderr,"   - the number of MPI processes N should not be be less than \n");
 696         fprintf(stderr,"     the number of nodes in the graph\n");
 697       }
 698       MPI_Finalize();
 699       exit(0);
 700     }
 701    if(strncmp(argv[1],"BH",2)==0){
 702       dg=buildBH(CLASS);
 703     }else if(strncmp(argv[1],"WH",2)==0){
 704       dg=buildWH(CLASS);
 705     }else if(strncmp(argv[1],"SH",2)==0){
 706       dg=buildSH(CLASS);
 707     }
 708
 709     if(timer_on&&dg->numNodes+1>timers_tot){
 710       timer_on=0;
 711       if(my_rank==0)
 712         fprintf(stderr,"Not enough timers. Node timeing is off. \n");
 713     }
 714     if(dg->numNodes>comm_size){
 715       if(my_rank==0){
 716         fprintf(stderr,"**  The number of MPI processes should not be less than \n");
 717         fprintf(stderr,"**  the number of nodes in the graph\n");
 718         fprintf(stderr,"**  Number of MPI processes = %d\n",comm_size);
 719         fprintf(stderr,"**  Number nodes in the graph = %d\n",dg->numNodes);
 720       }
 721       MPI_Finalize();
 722       exit(0);
 723     }
 724     for(i=0;i<dg->numNodes;i++){
 725       dg->node[i]->address=i;
 726     }
 727     if( my_rank == 0 ){
 728       printf( "\n\n NAS Parallel Benchmarks 3.3 -- DT Benchmark\n\n" );
 729       graphShow(dg,0);
 730       timer_clear(0);
 731       timer_start(0);
 732     }
 733
 734     TRACE_smpi_set_category ("dt");
 735     verified=ProcessNodes(dg,my_rank);
 736     TRACE_smpi_set_category ("finalize");
 737
 738     featnum=NUM_SAMPLES*fielddim;
 739     bytes_sent=featnum*dg->numArcs;
 740     bytes_sent/=1048576;
 741     if(my_rank==0){
 742       timer_stop(0);
 743       tot_time=timer_read(0);
 744       c_print_results( dg->name,
 745                        CLASS,
 746                        featnum,
 747                        0,
 748                        0,
 749                        dg->numNodes,
 750                        0,
 751                        comm_size,
 752                        tot_time,
 753                        bytes_sent/tot_time,
 754                        "bytes transmitted",
 755                        verified,
 756                        NPBVERSION,
 757                        COMPILETIME,
 758                        MPICC,
 759                        CLINK,
 760                        CMPI_LIB,
 761                        CMPI_INC,
 762                        CFLAGS,
 763                        CLINKFLAGS );
 764     }
 765     MPI_Finalize();
 766   return 1;
 767 }