int
main (
int argc,
char * argv[])
{
SCOTCH_Num vertnbr; /* Number of vertices */
SCOTCH_Graph grafdat; /* Source graph */
SCOTCH_Ordering ordedat; /* Graph ordering */
SCOTCH_Num * permtab; /* Permutation array */
SCOTCH_Strat stradat; /* Ordering strategy */
SCOTCH_Num straval;
char * straptr;
int flagval;
Clock runtime[2]; /* Timing variables */
int i, j;
errorProg ("gord");
if ((argc >= 2) && (argv[1][0] == '?')) { /* If need for help */
usagePrint (stdout, C_usageList);
return (0);
}
flagval = C_FLAGNONE; /* Default behavior */
straval = 0; /* No strategy flags */
straptr = NULL;
SCOTCH_stratInit (&stradat);
for (i = 0; i < C_FILENBR; i ++) /* Set default stream pointers */
C_fileTab[i].pntr = (C_fileTab[i].mode[0] == 'r') ? stdin : stdout;
for (i = 1; i < argc; i ++) { /* Loop for all option codes */
if ((argv[i][0] != '-') || (argv[i][1] == '\0') || (argv[i][1] == '.')) { /* If found a file name */
if (C_fileNum < C_FILEARGNBR) /* File name has been given */
C_fileTab[C_fileNum ++].name = argv[i];
else
errorPrint ("main: too many file names given");
}
else { /* If found an option name */
switch (argv[i][1]) {
case 'C' :
case 'c' : /* Strategy selection parameters */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'B' :
case 'b' :
straval |= SCOTCH_STRATBALANCE;
break;
case 'Q' :
case 'q' :
straval |= SCOTCH_STRATQUALITY;
break;
case 'S' :
case 's' :
straval |= SCOTCH_STRATSPEED;
break;
case 'T' :
case 't' :
straval |= SCOTCH_STRATSAFETY;
break;
default :
errorPrint ("main: invalid strategy selection option '%c'", argv[i][j]);
}
}
break;
case 'H' : /* Give the usage message */
case 'h' :
usagePrint (stdout, C_usageList);
return (0);
case 'M' : /* Output separator mapping */
case 'm' :
flagval |= C_FLAGMAPOUT;
if (argv[i][2] != '\0')
C_filenamemapout = &argv[i][2];
break;
case 'O' : /* Ordering strategy */
case 'o' :
straptr = &argv[i][2];
SCOTCH_stratExit (&stradat);
SCOTCH_stratInit (&stradat);
SCOTCH_stratGraphOrder (&stradat, straptr);
break;
case 'T' : /* Output separator tree */
case 't' :
flagval |= C_FLAGTREOUT;
if (argv[i][2] != '\0')
C_filenametreout = &argv[i][2];
break;
case 'V' :
fprintf (stderr, "gord, version " SCOTCH_VERSION_STRING "\n");
fprintf (stderr, "Copyright 2004,2007,2008,2010-2012 IPB, Universite de Bordeaux, INRIA & CNRS, France\n");
fprintf (stderr, "This software is libre/free software under CeCILL-C -- see the user's manual for more information\n");
return (0);
case 'v' : /* Output control info */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'S' :
case 's' :
flagval |= C_FLAGVERBSTR;
break;
case 'T' :
case 't' :
flagval |= C_FLAGVERBTIM;
break;
default :
errorPrint ("main: unprocessed parameter '%c' in '%s'", argv[i][j], argv[i]);
}
}
break;
default :
errorPrint ("main: unprocessed option '%s'", argv[i]);
}
}
}
fileBlockOpen (C_fileTab, C_FILENBR); /* Open all files */
clockInit (&runtime[0]);
clockStart (&runtime[0]);
SCOTCH_graphInit (&grafdat); /* Create graph structure */
SCOTCH_graphLoad (&grafdat, C_filepntrsrcinp, -1, 2); /* Read source graph */
SCOTCH_graphSize (&grafdat, &vertnbr, NULL); /* Get graph characteristics */
if (straval != 0) {
if (straptr != NULL)
errorPrint ("main: options '-c' and '-o' are exclusive");
SCOTCH_stratGraphOrderBuild (&stradat, straval, 0, 0.2);
}
clockStop (&runtime[0]); /* Get input time */
clockInit (&runtime[1]);
clockStart (&runtime[1]);
if ((permtab = (SCOTCH_Num *) memAlloc (vertnbr * sizeof (SCOTCH_Num))) == NULL) {
errorPrint ("main: out of memory");
return (1);
}
SCOTCH_graphOrderInit (&grafdat, &ordedat, permtab, NULL, NULL, NULL, NULL); /* Create ordering */
SCOTCH_graphOrderCompute (&grafdat, &ordedat, &stradat); /* Perform ordering */
clockStop (&runtime[1]); /* Get ordering time */
#ifdef SCOTCH_DEBUG_ALL
if (SCOTCH_graphOrderCheck (&grafdat, &ordedat) != 0)
return (1);
#endif /* SCOTCH_DEBUG_ALL */
clockStart (&runtime[0]);
SCOTCH_graphOrderSave (&grafdat, &ordedat, C_filepntrordout); /* Write ordering */
if (flagval & C_FLAGMAPOUT) /* If mapping wanted */
SCOTCH_graphOrderSaveMap (&grafdat, &ordedat, C_filepntrmapout); /* Write mapping */
if (flagval & C_FLAGTREOUT) /* If separator tree wanted */
SCOTCH_graphOrderSaveTree (&grafdat, &ordedat, C_filepntrtreout); /* Write tree */
clockStop (&runtime[0]); /* Get output time */
if (flagval & C_FLAGVERBSTR) {
fprintf (C_filepntrlogout, "S\tStrat=");
SCOTCH_stratSave (&stradat, C_filepntrlogout);
putc ('\n', C_filepntrlogout);
}
if (flagval & C_FLAGVERBTIM) {
fprintf (C_filepntrlogout, "T\tOrder\t\t%g\nT\tI/O\t\t%g\nT\tTotal\t\t%g\n",
(double) clockVal (&runtime[1]),
(double) clockVal (&runtime[0]),
(double) clockVal (&runtime[0]) +
(double) clockVal (&runtime[1]));
}
fileBlockClose (C_fileTab, C_FILENBR); /* Always close explicitely to end eventual (un)compression tasks */
SCOTCH_graphOrderExit (&grafdat, &ordedat);
SCOTCH_stratExit (&stradat);
SCOTCH_graphExit (&grafdat);
memFree (permtab);
#ifdef COMMON_PTHREAD
pthread_exit ((void *) 0); /* Allow potential (un)compression tasks to complete */
#endif /* COMMON_PTHREAD */
return (0);
}
void kass(int levelk,
int rat,
SymbolMatrix * symbptr,
PASTIX_INT baseval,
PASTIX_INT vertnbr,
PASTIX_INT edgenbr,
PASTIX_INT * verttab,
PASTIX_INT * edgetab,
Order * orderptr,
MPI_Comm pastix_comm)
{
PASTIX_INT snodenbr;
PASTIX_INT *snodetab = NULL;
PASTIX_INT *treetab = NULL;
PASTIX_INT *ia = NULL;
PASTIX_INT *ja = NULL;
PASTIX_INT i, j, n;
PASTIX_INT ind;
csptr mat;
PASTIX_INT *tmpj = NULL;
PASTIX_INT *perm = NULL;
PASTIX_INT *iperm = NULL;
PASTIX_INT newcblknbr;
PASTIX_INT *newrangtab = NULL;
Dof dofstr;
Clock timer1;
double nnzS;
int procnum;
(void)edgenbr;
MPI_Comm_rank(pastix_comm,&procnum);
#ifdef DEBUG_KASS
print_one("--- kass begin ---\n");
#endif
/* graphData (graphptr, */
/* (SCOTCH_Num * )&baseval, */
/* (SCOTCH_Num * )&vertnbr, */
/* (SCOTCH_Num **)&verttab, */
/* NULL, NULL, NULL, */
/* (SCOTCH_Num * )&edgenbr, */
/* (SCOTCH_Num **)&edgetab, */
/* NULL); */
n = vertnbr;
ia = verttab;
ja = edgetab;
perm = orderptr->permtab;
iperm = orderptr->peritab;
/*** Convert Fortran to C numbering ***/
if(baseval == 1)
{
for(i=0;i<=n;i++)
ia[i]--;
for(i=0;i<n;i++)
for(j=ia[i];j<ia[i+1];j++)
ja[j]--;
for(i=0;i<n;i++)
orderptr->permtab[i]--;
for(i=0;i<n;i++)
orderptr->peritab[i]--;
}
MALLOC_INTERN(treetab, n, PASTIX_INT);
#ifndef SCOTCH_SNODE
/*if(rat != -1 )*/
{
/***** FIND THE SUPERNODE PARTITION FROM SCRATCH ********/
/*** Find the supernodes of the direct factorization ***/
MALLOC_INTERN(snodetab, n+1, PASTIX_INT);
clockInit(&timer1);
clockStart(&timer1);
find_supernodes(n, ia, ja, perm, iperm, &snodenbr, snodetab, treetab);
clockStop(&timer1);
print_one("Time to find the supernode (direct) %.3g s \n", clockVal(&timer1));
/*memfree(treetab);*/
print_one("Number of supernode for direct factorization %ld \n", (long)snodenbr);
}
#else
/*else*/
{
/***** USE THE SUPERNODE PARTITION OF SCOTCH ********/
snodenbr = orderptr->cblknbr;
MALLOC_INTERN(snodetab, n+1, PASTIX_INT);
memCpy(snodetab, orderptr->rangtab, sizeof(PASTIX_INT)*(snodenbr+1));
print_one("Number of column block found in scotch (direct) %ld \n", (long)snodenbr);
}
#endif
/****************************************/
/* Convert the graph */
/****************************************/
MALLOC_INTERN(mat, 1, struct SparRow);
initCS(mat, n);
MALLOC_INTERN(tmpj, n, PASTIX_INT);
/**** Convert and permute the matrix in sparrow form ****/
/**** The diagonal is not present in the CSR matrix, we have to put it in the matrix ***/
bzero(tmpj, sizeof(PASTIX_INT)*n);
for(i=0;i<n;i++)
{
/*** THE GRAPH DOES NOT CONTAIN THE DIAGONAL WE ADD IT ***/
tmpj[0] = i;
ind = 1;
for(j=ia[i];j<ia[i+1];j++)
tmpj[ind++] = ja[j];
mat->nnzrow[i] = ind;
MALLOC_INTERN(mat->ja[i], ind, PASTIX_INT);
memCpy(mat->ja[i], tmpj, sizeof(PASTIX_INT)*ind);
mat->ma[i] = NULL;
}
CS_Perm(mat, perm);
/*** Reorder the matrix ***/
sort_row(mat);
memFree(tmpj);
/***** COMPUTE THE SYMBOL MATRIX OF ILU(K) WITH AMALGAMATION *****/
kass_symbol(mat, levelk, (double)(rat)/100.0, perm,
iperm, snodenbr, snodetab, treetab, &newcblknbr, &newrangtab,
symbptr, pastix_comm);
cleanCS(mat);
memFree(mat);
memFree(treetab);
dofInit(&dofstr);
dofConstant(&dofstr, 0, symbptr->nodenbr, 1);
nnzS = recursive_sum(0, symbptr->cblknbr-1, nnz, symbptr, &dofstr);
print_one("Number of non zero in the non patched symbol matrix = %g, fillrate1 %.3g \n",
nnzS+n, (nnzS+n)/(ia[n]/2.0 +n));
dofExit(&dofstr);
if(symbolCheck(symbptr) != 0)
{
errorPrint("SymbolCheck after kass_symbol.");
ASSERT(0, MOD_KASS);
}
if(levelk != -1)
{
/********************************************************/
/** ADD BLOCKS IN ORDER TO GET A REAL ELIMINATION TREE **/
/********************************************************/
Patch_SymbolMatrix(symbptr);
}
dofInit(&dofstr);
dofConstant(&dofstr, 0, symbptr->nodenbr, 1);
nnzS = recursive_sum(0, symbptr->cblknbr-1, nnz, symbptr, &dofstr);
dofExit(&dofstr);
print_one("Number of block in final symbol matrix = %ld \n", (long)symbptr->bloknbr);
print_one("Number of non zero in final symbol matrix = %g, fillrate2 %.3g \n", nnzS+n, (nnzS+n)/(ia[n]/2.0 +n));
if(symbolCheck(symbptr) != 0)
{
errorPrint("SymbolCheck after Patch_SymbolMatrix.");
ASSERT(0, MOD_KASS);
}
#ifdef DEBUG_KASS
print_one("--- kass end ---\n");
#endif
memFree(snodetab);
orderptr->cblknbr = newcblknbr;
memFree(orderptr->rangtab);
orderptr->rangtab = newrangtab;
}
void kass_symbol(csptr mat, PASTIX_INT levelk, double rat, PASTIX_INT *perm, PASTIX_INT *iperm, PASTIX_INT snodenbr, PASTIX_INT *snodetab, PASTIX_INT *streetab, PASTIX_INT *cblknbr, PASTIX_INT **rangtab, SymbolMatrix *symbmtx, MPI_Comm pastix_comm)
{
/**************************************************************************************/
/* This function computes a symbolic factorization ILU(k) given a CSR matrix and an */
/* ordering. Then it computes a block partition of the factor to get BLAS3 */
/* efficiency */
/* NOTE: the CSC matrix is given symmetrized and without the diagonal */
/**************************************************************************************/
PASTIX_INT i, j;
PASTIX_INT nnzL;
PASTIX_INT *iperm2 = NULL;
PASTIX_INT *treetab = NULL;
PASTIX_INT n;
csptr P;
Clock timer1;
int procnum;
MPI_Comm_rank(pastix_comm,&procnum);
n = mat->n;
MALLOC_INTERN(iperm2, n, PASTIX_INT);
/*compact_graph(mat, NULL, NULL, NULL);*/
/*** Compute the ILU(k) pattern of the quotient matrix ***/
MALLOC_INTERN(P, 1, struct SparRow);
initCS(P, n);
print_one("Level of fill = %ld\nAmalgamation ratio = %d \n", (long)levelk, (int)(rat*100));
clockInit(&timer1);
clockStart(&timer1);
if(levelk == -1)
{
/***** FACTORISATION DIRECT *******/
/***** (Re)compute also the streetab (usefull when SCOTCH_SNODE
is active) ***/
SF_Direct(mat, snodenbr, snodetab, streetab, P);
clockStop(&timer1);
print_one("Time to compute scalar symbolic direct factorization %.3g s \n", clockVal(&timer1));
#ifdef DEBUG_KASS
print_one("non-zeros in P = %ld \n", (long)CSnnz(P));
#endif
nnzL = 0;
for(i=0;i<P->n;i++)
{
PASTIX_INT ncol;
ncol = snodetab[i+1]-snodetab[i];
nnzL += (ncol*(ncol+1))/2;
#ifdef DEBUG_KASS
ASSERT(P->nnzrow[i] >= ncol, MOD_KASS);
if(P->nnzrow[i] >= n)
fprintf(stderr,"P->nnzrow[%ld] = %ld \n", (long)i, (long)P->nnzrow[i]);
ASSERT(P->nnzrow[i] < n, MOD_KASS);
#endif
nnzL += (P->nnzrow[i]-ncol)*ncol;
}
#ifdef DEBUG_KASS
print_one("NNZL = %ld \n", (long)nnzL);
#endif
}
else
{
/***** FACTORISATION INCOMPLETE *******/
nnzL = SF_level(2, mat, levelk, P);
clockStop(&timer1);
print_one("Time to compute scalar symbolic factorization of ILU(%ld) %.3g s \n",
(long)levelk, clockVal(&timer1));
}
print_one("Scalar nnza = %ld nnzlk = %ld, fillrate0 = %.3g \n",
(long)( CSnnz(mat) + n)/2, (long)nnzL, (double)nnzL/(double)( (CSnnz(mat)+n)/2.0 ));
/** Sort the rows of the symbolic matrix */
sort_row(P);
clockInit(&timer1);
clockStart(&timer1);
if(levelk != -1)
{
/********************************/
/** Compute the "k-supernodes" **/
/********************************/
#ifdef KS
assert(levelk >= 0);
KSupernodes(P, rat, snodenbr, snodetab, cblknbr, rangtab);
#else
#ifdef SCOTCH_SNODE
if(rat == -1)
assert(0); /** do not have treetab with this version of Scotch **/
#endif
MALLOC_INTERN(treetab, P->n, PASTIX_INT);
for(j=0;j<snodenbr;j++)
{
for(i=snodetab[j];i<snodetab[j+1]-1;i++)
treetab[i] = i+1;
/*** Version generale ****/
if(streetab[j] == -1 || streetab[j] == j)
treetab[i] = -1;
else
treetab[i]=snodetab[streetab[j]];
/*** Version restricted inside the supernode (like KSupernodes) ***/
/*treetab[snodetab[j+1]-1] = -1;*/ /** this should give the same results than
KSupernodes **/
}
/** NEW ILUK + DIRECT **/
amalgamate(rat, P, -1, NULL, treetab, cblknbr, rangtab, iperm2, pastix_comm);
memFree(treetab);
for(i=0;i<n;i++)
iperm2[i] = iperm[iperm2[i]];
memcpy(iperm, iperm2, sizeof(PASTIX_INT)*n);
for(i=0;i<n;i++)
perm[iperm[i]] = i;
#endif
}
else{
/*if(0)*/
{
amalgamate(rat, P, snodenbr, snodetab, streetab, cblknbr,
rangtab, iperm2, pastix_comm);
/** iperm2 is the iperm vector of P **/
for(i=0;i<n;i++)
iperm2[i] = iperm[iperm2[i]];
memcpy(iperm, iperm2, sizeof(PASTIX_INT)*n);
for(i=0;i<n;i++)
perm[iperm[i]] = i;
}
/*else
{
fprintf(stderr, "RAT = 0 SKIP amalgamation \n");
*cblknbr = snodenbr;
MALLOC_INTERN(*rangtab, snodenbr+1, PASTIX_INT);
memcpy(*rangtab, snodetab, sizeof(PASTIX_INT)*(snodenbr+1));
}*/
}
clockStop(&timer1);
print_one("Time to compute the amalgamation of supernodes %.3g s\n", clockVal(&timer1));
print_one("Number of cblk in the amalgamated symbol matrix = %ld \n", (long)*cblknbr);
Build_SymbolMatrix(P, *cblknbr, *rangtab, symbmtx);
print_one("Number of block in the non patched symbol matrix = %ld \n", (long)symbmtx->bloknbr);
memFree(iperm2);
cleanCS(P);
memFree(P);
}
int
main (
int argc,
char * argv[])
{
SCOTCH_Dgraph grafdat;
SCOTCH_Dordering ordedat;
SCOTCH_Strat stradat;
SCOTCH_Num straval;
char * straptr;
int flagval;
int procglbnbr;
int proclocnum;
int protglbnum; /* Root process */
Clock runtime[2]; /* Timing variables */
double reduloctab[12]; /* 3 * (min, max, sum) */
double reduglbtab[12];
MPI_Datatype redutype;
MPI_Op reduop;
int i, j;
#ifdef SCOTCH_PTHREAD
int thrdlvlreqval;
int thrdlvlproval;
#endif /* SCOTCH_PTHREAD */
errorProg ("dgord");
#ifdef SCOTCH_PTHREAD
thrdlvlreqval = MPI_THREAD_MULTIPLE;
if (MPI_Init_thread (&argc, &argv, thrdlvlreqval, &thrdlvlproval) != MPI_SUCCESS)
errorPrint ("main: Cannot initialize (1)");
if (thrdlvlreqval > thrdlvlproval)
errorPrint ("main: MPI implementation is not thread-safe: recompile without SCOTCH_PTHREAD");
#else /* SCOTCH_PTHREAD */
if (MPI_Init (&argc, &argv) != MPI_SUCCESS)
errorPrint ("main: Cannot initialize (2)");
#endif /* SCOTCH_PTHREAD */
MPI_Comm_size (MPI_COMM_WORLD, &procglbnbr); /* Get communicator data */
MPI_Comm_rank (MPI_COMM_WORLD, &proclocnum);
protglbnum = 0; /* Assume root process is process 0 */
if ((argc >= 2) && (argv[1][0] == '?')) { /* If need for help */
usagePrint (stdout, C_usageList);
return (0);
}
SCOTCH_randomProc (proclocnum); /* Record process number to initialize pseudo-random seed */
flagval = C_FLAGNONE; /* Default behavior */
straval = 0; /* No strategy flags */
straptr = NULL;
SCOTCH_stratInit (&stradat);
fileBlockInit (C_fileTab, C_FILENBR); /* Set default stream pointers */
for (i = 1; i < argc; i ++) { /* Loop for all option codes */
if ((argv[i][0] != '-') || (argv[i][1] == '\0') || (argv[i][1] == '.')) { /* If found a file name */
if (C_fileNum < C_FILEARGNBR) /* File name has been given */
fileBlockName (C_fileTab, C_fileNum ++) = argv[i];
else
errorPrint ("main: too many file names given");
}
else { /* If found an option name */
switch (argv[i][1]) {
case 'B' :
case 'b' :
flagval |= C_FLAGBLOCK;
break;
case 'C' :
case 'c' : /* Strategy selection parameters */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'B' :
case 'b' :
straval |= SCOTCH_STRATBALANCE;
break;
case 'Q' :
case 'q' :
straval |= SCOTCH_STRATQUALITY;
break;
case 'S' :
case 's' :
straval |= SCOTCH_STRATSPEED;
break;
case 'T' :
case 't' :
straval |= SCOTCH_STRATSAFETY;
break;
case 'X' :
case 'x' :
straval |= SCOTCH_STRATSCALABILITY;
break;
default :
errorPrint ("main: invalid strategy selection option '%c'", argv[i][j]);
}
}
break;
#ifdef SCOTCH_DEBUG_ALL
case 'D' :
case 'd' :
flagval |= C_FLAGDEBUG;
break;
#endif /* SCOTCH_DEBUG_ALL */
case 'H' : /* Give the usage message */
case 'h' :
usagePrint (stdout, C_usageList);
return (0);
case 'M' : /* Output separator mapping */
case 'm' :
flagval |= C_FLAGMAPOUT;
if (argv[i][2] != '\0')
C_filenamemapout = &argv[i][2];
break;
case 'O' : /* Ordering strategy */
case 'o' :
straptr = &argv[i][2];
SCOTCH_stratExit (&stradat);
SCOTCH_stratInit (&stradat);
SCOTCH_stratDgraphOrder (&stradat, straptr);
break;
case 'R' : /* Root process (if necessary) */
case 'r' :
protglbnum = atoi (&argv[i][2]);
if ((protglbnum < 0) ||
(protglbnum >= procglbnbr) ||
((protglbnum == 0) && (argv[i][2] != '0')))
errorPrint ("main: invalid root process number");
break;
case 'T' : /* Output separator tree */
case 't' :
flagval |= C_FLAGTREOUT;
if (argv[i][2] != '\0')
C_filenametreout = &argv[i][2];
break;
case 'V' :
fprintf (stderr, "dgord, version " SCOTCH_VERSION_STRING "\n");
fprintf (stderr, "Copyright 2007-2012,2014 IPB, Universite de Bordeaux, INRIA & CNRS, France\n");
fprintf (stderr, "This software is libre/free software under CeCILL-C -- see the user's manual for more information\n");
return (0);
case 'v' : /* Output control info */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'A' :
case 'a' :
#ifdef COMMON_MEMORY_TRACE
flagval |= C_FLAGVERBMEM;
#else /* COMMON_MEMORY_TRACE */
errorPrint ("main: not compiled with COMMON_MEMORY_TRACE");
#endif /* COMMON_MEMORY_TRACE */
break;
case 'S' :
case 's' :
flagval |= C_FLAGVERBSTR;
break;
case 'T' :
case 't' :
flagval |= C_FLAGVERBTIM;
break;
default :
errorPrint ("main: unprocessed parameter '%c' in '%s'", argv[i][j], argv[i]);
}
}
break;
default :
errorPrint ("main: unprocessed option '%s'", argv[i]);
}
}
}
#ifdef SCOTCH_DEBUG_ALL
if ((flagval & C_FLAGDEBUG) != 0) {
fprintf (stderr, "Proc %4d of %d, pid %d\n", proclocnum, procglbnbr, getpid ());
if (proclocnum == protglbnum) { /* Synchronize on keybord input */
char c;
printf ("Waiting for key press...\n");
scanf ("%c", &c);
}
MPI_Barrier (MPI_COMM_WORLD);
}
#endif /* SCOTCH_DEBUG_ALL */
fileBlockOpenDist (C_fileTab, C_FILENBR, procglbnbr, proclocnum, protglbnum); /* Open all files */
clockInit (&runtime[0]);
clockStart (&runtime[0]);
SCOTCH_dgraphInit (&grafdat, MPI_COMM_WORLD);
SCOTCH_dgraphLoad (&grafdat, C_filepntrsrcinp, -1, 0);
if (straval != 0) {
if (straptr != NULL)
errorPrint ("main: options '-c' and '-o' are exclusive");
SCOTCH_stratDgraphOrderBuild (&stradat, straval, (SCOTCH_Num) procglbnbr, 0, 0.2);
}
clockStop (&runtime[0]); /* Get input time */
clockInit (&runtime[1]);
#ifdef SCOTCH_DEBUG_ALL
if ((flagval & C_FLAGDEBUG) != 0)
MPI_Barrier (MPI_COMM_WORLD);
#endif /* SCOTCH_DEBUG_ALL */
clockStart (&runtime[1]);
SCOTCH_dgraphGhst (&grafdat); /* Compute it once for good */
SCOTCH_dgraphOrderInit (&grafdat, &ordedat);
SCOTCH_dgraphOrderCompute (&grafdat, &ordedat, &stradat);
clockStop (&runtime[1]); /* Get ordering time */
#ifdef SCOTCH_DEBUG_ALL
if ((flagval & C_FLAGDEBUG) != 0)
MPI_Barrier (MPI_COMM_WORLD);
#endif /* SCOTCH_DEBUG_ALL */
clockStart (&runtime[0]);
if (proclocnum == protglbnum) {
if ((flagval & C_FLAGBLOCK) == 0)
SCOTCH_dgraphOrderSave (&grafdat, &ordedat, C_filepntrordout);
else
SCOTCH_dgraphOrderSaveBlock (&grafdat, &ordedat, C_filepntrordout);
if ((flagval & C_FLAGMAPOUT) != 0) /* If mapping wanted */
SCOTCH_dgraphOrderSaveMap (&grafdat, &ordedat, C_filepntrmapout); /* Write mapping */
if ((flagval & C_FLAGTREOUT) != 0) /* If separator tree wanted */
SCOTCH_dgraphOrderSaveTree (&grafdat, &ordedat, C_filepntrtreout); /* Write tree */
}
else {
if ((flagval & C_FLAGBLOCK) == 0)
SCOTCH_dgraphOrderSave (&grafdat, &ordedat, NULL);
else
SCOTCH_dgraphOrderSaveBlock (&grafdat, &ordedat, NULL);
if ((flagval & C_FLAGMAPOUT) != 0)
SCOTCH_dgraphOrderSaveMap (&grafdat, &ordedat, NULL);
if ((flagval & C_FLAGTREOUT) != 0)
SCOTCH_dgraphOrderSaveTree (&grafdat, &ordedat, NULL);
}
clockStop (&runtime[0]);
#ifdef SCOTCH_DEBUG_ALL
if ((flagval & C_FLAGDEBUG) != 0)
MPI_Barrier (MPI_COMM_WORLD);
#endif /* SCOTCH_DEBUG_ALL */
MPI_Type_contiguous (3, MPI_DOUBLE, &redutype);
MPI_Type_commit (&redutype);
MPI_Op_create ((MPI_User_function *) dgordStatReduceOp, 1, &reduop);
if ((flagval & C_FLAGVERBTIM) != 0) {
reduloctab[0] =
reduloctab[1] =
reduloctab[2] = (double) clockVal (&runtime[1]);
reduloctab[3] =
reduloctab[4] =
reduloctab[5] = (double) clockVal (&runtime[0]);
reduloctab[6] =
reduloctab[7] =
reduloctab[8] = reduloctab[0] + reduloctab[3];
MPI_Allreduce (&reduloctab[0], &reduglbtab[0], 3, redutype, reduop, MPI_COMM_WORLD);
}
#ifdef COMMON_MEMORY_TRACE
if ((flagval & C_FLAGVERBMEM) != 0) {
reduloctab[9] =
reduloctab[10] =
reduloctab[11] = (double) memMax ();
MPI_Allreduce (&reduloctab[9], &reduglbtab[9], 1, redutype, reduop, MPI_COMM_WORLD);
}
#endif /* COMMON_MEMORY_TRACE */
MPI_Op_free (&reduop);
MPI_Type_free (&redutype);
if (C_filepntrlogout != NULL) {
if ((flagval & C_FLAGVERBSTR) != 0) {
fprintf (C_filepntrlogout, "S\tStrat=");
SCOTCH_stratSave (&stradat, C_filepntrlogout);
putc ('\n', C_filepntrlogout);
}
if ((flagval & C_FLAGVERBTIM) != 0) {
fprintf (C_filepntrlogout, "T\tOrder\tmin=%g\tmax=%g\tavg=%g\nT\tI/O\tmin=%g\tmax=%g\tavg=%g\nT\tTotal\tmin=%g\tmax=%g\tavg=%g\n",
reduglbtab[0], reduglbtab[1], reduglbtab[2] / (double) procglbnbr,
reduglbtab[3], reduglbtab[4], reduglbtab[5] / (double) procglbnbr,
reduglbtab[6], reduglbtab[7], reduglbtab[8] / (double) procglbnbr);
}
#ifdef COMMON_MEMORY_TRACE
if ((flagval & C_FLAGVERBMEM) != 0)
fprintf (C_filepntrlogout, "A\tMemory\tmin=%g\tmax=%g\tavg=%g\n",
reduglbtab[9], reduglbtab[10], reduglbtab[11] / (double) procglbnbr);
#endif /* COMMON_MEMORY_TRACE */
}
fileBlockClose (C_fileTab, C_FILENBR); /* Always close explicitely to end eventual (un)compression tasks */
SCOTCH_dgraphOrderExit (&grafdat, &ordedat);
SCOTCH_dgraphExit (&grafdat);
SCOTCH_stratExit (&stradat);
MPI_Finalize ();
#ifdef COMMON_PTHREAD
pthread_exit ((void *) 0); /* Allow potential (un)compression tasks to complete */
#endif /* COMMON_PTHREAD */
return (0);
}
int
main (
int argc,
char * argv[])
{
SCOTCH_Graph grafdat; /* Source graph */
SCOTCH_Num grafflag; /* Source graph properties */
SCOTCH_Arch archdat; /* Target architecture */
SCOTCH_Strat stradat; /* Mapping strategy */
SCOTCH_Mapping mapdat; /* Mapping data */
Clock runtime[2]; /* Timing variables */
double kbalval; /* Imbalance tolerance value */
int flagval;
SCOTCH_Num straval;
char * straptr;
int i, j;
flagval = C_FLAGNONE; /* Default behavior */
straval = 0; /* No strategy flags */
straptr = NULL;
i = strlen (argv[0]);
if ((i >= 5) && (strncmp (argv[0] + i - 5, "gpart", 5) == 0)) {
flagval |= C_FLAGPART;
C_paraNbr = 1; /* One more parameter */
C_fileNbr = 3; /* One less file to provide */
errorProg ("gpart");
}
else
errorProg ("gmap");
intRandInit ();
if ((argc >= 2) && (argv[1][0] == '?')) { /* If need for help */
usagePrint (stdout, C_usageList);
return (0);
}
grafflag = 0; /* Use vertex and edge weights */
SCOTCH_stratInit (&stradat); /* Set default mapping strategy */
kbalval = 0.01; /* Set default load imbalance value */
for (i = 0; i < C_FILENBR; i ++) /* Set default stream pointers */
C_fileTab[i].pntr = (C_fileTab[i].mode[0] == 'r') ? stdin : stdout;
for (i = 1; i < argc; i ++) { /* Loop for all option codes */
if ((argv[i][0] != '-') || (argv[i][1] == '\0') || (argv[i][1] == '.')) { /* If found a file name */
if (C_paraNum < C_paraNbr) { /* If number of parameters not reached */
if ((C_partNbr = atoi (argv[i])) < 1) /* Get the number of parts */
errorPrint ("main: invalid number of parts (\"%s\")", argv[i]);
C_paraNum ++;
continue; /* Process the other parameters */
}
if (C_fileNum < C_fileNbr) /* A file name has been given */
C_fileTab[C_fileNum ++].name = argv[i];
else
errorPrint ("main: too many file names given");
}
else { /* If found an option name */
switch (argv[i][1]) {
case 'B' :
case 'b' :
flagval |= C_FLAGKBALVAL;
kbalval = atof (&argv[i][2]);
if ((kbalval < 0.0) ||
(kbalval > 1.0) ||
((kbalval == 0.0) &&
((argv[i][2] != '0') && (argv[i][2] != '.')))) {
errorPrint ("main: invalid load imbalance ratio");
}
break;
case 'C' :
case 'c' : /* Strategy selection parameters */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'B' :
case 'b' :
straval |= SCOTCH_STRATBALANCE;
break;
case 'Q' :
case 'q' :
straval |= SCOTCH_STRATQUALITY;
break;
case 'S' :
case 's' :
straval |= SCOTCH_STRATSPEED;
break;
case 'T' :
case 't' :
straval |= SCOTCH_STRATSAFETY;
break;
default :
errorPrint ("main: invalid strategy selection option (\"%c\")", argv[i][j]);
}
}
break;
case 'H' : /* Give the usage message */
case 'h' :
usagePrint (stdout, C_usageList);
return (0);
case 'M' :
case 'm' :
straptr = &argv[i][2];
SCOTCH_stratExit (&stradat);
SCOTCH_stratInit (&stradat);
SCOTCH_stratGraphMap (&stradat, straptr);
break;
case 'S' :
case 's' : /* Source graph parameters */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'E' :
case 'e' :
grafflag |= 2; /* Do not load edge weights */
break;
case 'V' :
case 'v' :
grafflag |= 1; /* Do not load vertex weights */
break;
default :
errorPrint ("main: invalid source graph option (\"%c\")", argv[i][j]);
}
}
break;
case 'V' :
fprintf (stderr, "gmap/gpart, version " SCOTCH_VERSION_STRING "\n");
fprintf (stderr, "Copyright 2004,2007,2008,2010 ENSEIRB, INRIA & CNRS, France\n");
fprintf (stderr, "This software is libre/free software under CeCILL-C -- see the user's manual for more information\n");
return (0);
case 'v' : /* Output control info */
for (j = 2; argv[i][j] != '\0'; j ++) {
switch (argv[i][j]) {
case 'M' :
case 'm' :
flagval |= C_FLAGVERBMAP;
break;
case 'S' :
case 's' :
flagval |= C_FLAGVERBSTR;
break;
case 'T' :
case 't' :
flagval |= C_FLAGVERBTIM;
break;
default :
errorPrint ("main: unprocessed parameter \"%c\" in \"%s\"", argv[i][j], argv[i]);
}
}
break;
default :
errorPrint ("main: unprocessed option (\"%s\")", argv[i]);
}
}
}
if ((flagval & C_FLAGPART) != 0) { /* If program run as the partitioner */
C_fileTab[3].name = C_fileTab[2].name; /* Put provided file names at their right place */
C_fileTab[2].name = C_fileTab[1].name;
C_fileTab[1].name = "-";
}
fileBlockOpen (C_fileTab, C_FILENBR); /* Open all files */
clockInit (&runtime[0]);
clockStart (&runtime[0]);
SCOTCH_graphInit (&grafdat); /* Create graph structure */
SCOTCH_graphLoad (&grafdat, C_filepntrsrcinp, -1, grafflag); /* Read source graph */
SCOTCH_archInit (&archdat); /* Create architecture structure */
if ((flagval & C_FLAGPART) != 0) /* If program run as the partitioner */
SCOTCH_archCmplt (&archdat, C_partNbr); /* Create a complete graph of proper size */
else {
SCOTCH_archLoad (&archdat, C_filepntrtgtinp); /* Read target architecture */
C_partNbr = SCOTCH_archSize (&archdat);
}
if ((straval != 0) || ((flagval & C_FLAGKBALVAL) != 0)) {
if (straptr != NULL)
errorPrint ("main: options '-b' / '-c' and '-m' are exclusive");
SCOTCH_stratGraphMapBuild (&stradat, straval, (SCOTCH_Num) C_partNbr, kbalval);
}
clockStop (&runtime[0]); /* Get input time */
clockInit (&runtime[1]);
clockStart (&runtime[1]);
SCOTCH_graphMapInit (&grafdat, &mapdat, &archdat, NULL);
SCOTCH_graphMapCompute (&grafdat, &mapdat, &stradat); /* Perform mapping */
clockStop (&runtime[1]); /* Get computation time */
clockStart (&runtime[0]);
SCOTCH_graphMapSave (&grafdat, &mapdat, C_filepntrmapout); /* Write mapping */
clockStop (&runtime[0]); /* Get output time */
if (flagval & C_FLAGVERBSTR) {
fprintf (C_filepntrlogout, "S\tStrat=");
SCOTCH_stratSave (&stradat, C_filepntrlogout);
putc ('\n', C_filepntrlogout);
}
if (flagval & C_FLAGVERBTIM) {
fprintf (C_filepntrlogout, "T\tMapping\t\t%g\nT\tI/O\t\t%g\nT\tTotal\t\t%g\n",
(double) clockVal (&runtime[1]),
(double) clockVal (&runtime[0]),
(double) clockVal (&runtime[0]) +
(double) clockVal (&runtime[1]));
}
if (flagval & C_FLAGVERBMAP)
SCOTCH_graphMapView (&grafdat, &mapdat, C_filepntrlogout);
fileBlockClose (C_fileTab, C_FILENBR); /* Always close explicitely to end eventual (un)compression tasks */
SCOTCH_graphMapExit (&grafdat, &mapdat);
SCOTCH_graphExit (&grafdat);
SCOTCH_stratExit (&stradat);
SCOTCH_archExit (&archdat);
#ifdef COMMON_PTHREAD
pthread_exit ((void *) 0); /* Allow potential (un)compression tasks to complete */
#endif /* COMMON_PTHREAD */
return (0);
}
int
main(int argc, char ** argv)
{
unsigned int iseed = (unsigned int)time(NULL);
int n;
int lda;
PASTIX_FLOAT *A;
PASTIX_FLOAT *B;
PASTIX_FLOAT *B_save;
PASTIX_FLOAT *B_res;
CU_FLOAT *d_A;
CU_FLOAT *d_B;
Clock clk;
Clock clk_wt;
PASTIX_FLOAT alpha = 1.0;
double time_CPU;
double time_CUDA;
double time_CUDA_wt;
int ops = n*n;
if (argc != 3)
{
usage(argv[0]);
return 1;
}
READ_INT(n, 1);
READ_INT(lda, 2);
srand (iseed);
MALLOC_INTERN(A, n*lda, PASTIX_FLOAT);
MALLOC_INTERN(B, n*lda, PASTIX_FLOAT);
MALLOC_INTERN(B_save, n*lda, PASTIX_FLOAT);
MALLOC_INTERN(B_res, n*lda, PASTIX_FLOAT);
FILL(A, n*lda);
FILL(B, n*lda);
memcpy(B_save, B, n*lda*sizeof(PASTIX_FLOAT));
clockInit(&(clk));
clockStart(&(clk));
DimTrans(A, lda, n, B);
clockStop(&(clk));
time_CPU = clockVal(&(clk));
PRINT_TIME("GETRA on CPU", time_CPU, ops);
clockInit(&(clk_wt));
clockStart(&(clk_wt));
CUDA_CALL(cudaMalloc((void*)&(d_A),
lda*n*sizeof(PASTIX_FLOAT)));
CUDA_CALL(cudaMemcpy((void*)d_A, A,
lda*n*sizeof(PASTIX_FLOAT),
cudaMemcpyHostToDevice));
CUDA_CALL(cudaMalloc((void*)&(d_B),
lda*n*sizeof(PASTIX_FLOAT)));
CUDA_CALL(cudaMemcpy((void*)d_B, B_save,
lda*n*sizeof(PASTIX_FLOAT),
cudaMemcpyHostToDevice));
clockInit(&(clk));
clockStart(&(clk));
getra_cuda(d_A, lda,
d_B, lda, n);
clockStop(&(clk));
CUDA_CALL(cudaMemcpy((void*)B_res, d_B,
lda*n*sizeof(PASTIX_FLOAT),
cudaMemcpyDeviceToHost));
CUDA_CALL(cudaFree(d_A));
CUDA_CALL(cudaFree(d_B));
clockStop(&(clk_wt));
time_CUDA = clockVal(&(clk));
time_CUDA_wt = clockVal(&(clk_wt));
COMPARE_TIME("GETRA on GPU",
time_CUDA, ops, time_CPU);
COMPARE_TIME("GETRA on GPU with transfer",
time_CUDA_wt, ops, time_CPU);
COMPARE_RES(B, B_res);
memFree_null(A);
memFree_null(B);
memFree_null(B_save);
memFree_null(B_res);
return EXIT_SUCCESS;
}