int main(int argc, char **argv) {
int t ;
struct sparsematrix A ;
struct opts Options;
SetDefaultOptions(&Options);
A.MMTypeCode[0] = 'D' ;
A.NrProcs = 7 ;
/* Allocate matrix arrays */
A.Pstart = (long *) malloc((A.NrProcs+1)* sizeof(long)) ;
if ( A.Pstart == NULL ){
printf("Error\n") ;
exit(1);
}
for ( t = 0 ; t <= A.NrProcs ; t++ )
A.Pstart[t] = 10*t ;
MMSparseMatrixPrintPstart(&A, stdout, &Options) ;
exit(0) ;
} /* end main */
int main(int argc, char **argv) {
char line[MAX_LINE_LENGTH] ;
struct sparsematrix A ;
struct opts Options;
SetDefaultOptions( &Options );
A.MMTypeCode[0] = 'W' ;
A.MMTypeCode[1] = 'C' ;
A.MMTypeCode[2] = 'R' ;
A.MMTypeCode[3] = 'G' ;
A.ViewType = ViewTypeOriginal;
A.m = 3 ;
A.n = 3 ;
A.NrNzElts = 9 ;
A.NrRowWeights = 3 ;
A.NrColWeights = 3 ;
strcpy(line, "% This is a 3 x 3 matrix\n% with 9 nonzeros\n") ;
A.header = line ;
A.comment = NULL;
MMSparseMatrixPrintHeader(&A, stdout, NULL, &Options);
exit(0) ;
} /* end main */
void main(int argc, char* argv[]) {
struct Options options;
SetDefaultOptions(&options, 0, MAX_FFT_ORDER);
ProcessCommandLine(&options, argc, argv,
"Test forward and inverse floating-point FFT\n");
verbose = options.verbose_;
if (verbose > 255)
DumpOptions(stderr, &options);
if (options.test_mode_) {
struct TestInfo info;
info.real_only_ = options.real_only_;
info.max_fft_order_ = options.max_fft_order_;
info.min_fft_order_ = options.min_fft_order_;
info.do_forward_tests_ = options.do_forward_tests_;
info.do_inverse_tests_ = options.do_inverse_tests_;
/* No known failures */
info.known_failures_ = 0;
info.forward_threshold_ = 138.81;
info.inverse_threshold_ = 138.81;
RunAllTests(&info);
} else {
TestFloatFFT(options.fft_log_size_,
options.signal_type_,
options.signal_value_);
}
}
bool ProductConfigurer::Execute(bool bIsRepair)
{
SetDefaultOptions();
if (!CopyLicFile())
{
CStdString sWarning;
sWarning.Format(_T("Failed to import %s"), m_pProcessController->GetLicFileSpecified().c_str());
m_pProcessController->ReportEvent(EVENTLOG_WARNING_TYPE, ERR_FAILED_TO_COPY_LIC_FILE, sWarning);
return false;
}
InstallDVWordMacroIfFlagSpecifiedOnCommandLine();
return true;
}
int main(int argc, char* argv[]) {
struct Options options;
struct TestInfo info;
int failed_count = 0;
SetDefaultOptions(&options, 1, MAX_FFT_ORDER);
ProcessCommandLine(&options,
argc,
argv,
UsageMessage());
verbose = options.verbose_;
info.real_only_ = options.real_only_;
info.min_fft_order_ = options.min_fft_order_;
info.max_fft_order_ = options.max_fft_order_;
info.do_forward_tests_ = options.do_forward_tests_;
info.do_inverse_tests_ = options.do_inverse_tests_;
/* No known failures */
info.known_failures_ = 0;
SetThresholds(&info);
if (verbose > 255)
DumpOptions(stderr, &options);
if (options.test_mode_) {
failed_count = RunAllTests(&info);
} else {
TestOneFFT(options.fft_log_size_,
options.signal_type_,
options.signal_value_,
&info,
"Float Real FFT");
}
FinishedMessage();
return failed_count > 0 ? 1 : 0;
}
void main(int argc, char* argv[]) {
struct Options options;
SetDefaultOptions(&options, 0, MAX_FFT_ORDER);
ProcessCommandLine(&options, argc, argv,
"Test forward and inverse 32-bit fixed-point FFT\n");
verbose = options.verbose_;
signal_value = options.signal_value_;
if (verbose > 255)
DumpOptions(stderr, &options);
if (options.test_mode_) {
struct TestInfo info;
info.real_only_ = options.real_only_;
info.max_fft_order_ = options.max_fft_order_;
info.min_fft_order_ = options.min_fft_order_;
info.do_forward_tests_ = options.do_forward_tests_;
info.do_inverse_tests_ = options.do_inverse_tests_;
info.known_failures_ = 0;
/*
* These threshold values assume that we're using the default
* signal_value set below.
*/
info.forward_threshold_ = 107.33;
info.inverse_threshold_ = 79.02;
if (!options.signal_value_given_) {
signal_value = 262144; /* 18 bits */
}
RunAllTests(&info);
} else {
TestFFT(options.fft_log_size_,
options.signal_type_,
options.scale_factor_);
}
}
void main(int argc, char* argv[]) {
struct Options options;
struct TestInfo info;
SetDefaultOptions(&options, 1, MAX_FFT_ORDER);
ProcessCommandLine(&options, argc, argv,
"Test forward and inverse real floating-point FFT\n");
verbose = options.verbose_;
if (verbose > 255)
DumpOptions(stderr, &options);
info.real_only_ = options.real_only_;
info.min_fft_order_ = options.min_fft_order_;
info.max_fft_order_ = options.max_fft_order_;
info.do_forward_tests_ = options.do_forward_tests_;
info.do_inverse_tests_ = options.do_inverse_tests_;
/* No known failures */
info.known_failures_ = 0;
#ifdef BIG_FFT_TABLE
info.forward_threshold_ = 136.07;
info.inverse_threshold_ = 140.76;
#else
info.forward_threshold_ = 136.07;
info.inverse_threshold_ = 142.41;
#endif
if (options.test_mode_) {
RunAllTests(&info);
} else {
TestOneFFT(options.fft_log_size_,
options.signal_type_,
options.signal_value_,
&info,
"Float Real FFT");
}
}
void ReadOptions (int argc, char **argv)
{
char *file = NULL;
if (opt != NULL)
puts ("Already read options.\n Over-writing.");
else {
SetDefaultOptions ();
}
// If no filename is given at command line, use default
if (argc == 1 || argv[1][0] == '-')
file = default_optionfile;
// Is the first argument a file rather than an option, if so then read
else
file = argv[1];
if (file != NULL)
ReadOptionsFromFile (file);
ReadOptionsFromCommandLine (argc, argv);
return;
}
int main(int argc, char **argv) {
struct sparsematrix A;
long n, i, j, t, P, weightlo, weighthi, weight0, weight1;
printf("Test SplitMatrixSimple: ");
n = 33; /* n by n dense matrix A, n odd */
A.m = n;
A.n = n;
A.NrNzElts = n*n;
P = 2; /* maximum number of parts */
A.i = (long *) malloc(A.NrNzElts* sizeof(long));
A.j = (long *) malloc(A.NrNzElts* sizeof(long));
A.Pstart = (long *) malloc((P+1)* sizeof(long));
A.RowLambda = (int *)malloc(A.m*sizeof(int));
A.ColLambda = (int *)malloc(A.n*sizeof(int));
A.RowMark = (int *)malloc(A.m*sizeof(int));
A.ColMark = (int *)malloc(A.n*sizeof(int));
if (!SetDefaultOptions(&Options)) {
printf("Error\n");
exit(1);
}
if (A.i == NULL || A.j == NULL || A.Pstart == NULL) {
printf("Error\n");
exit(1);
}
/* Fill matrix with nonzeros */
t= 0;
for (i=0; i<n; i++) {
for (j=0; j<n; j++) {
A.i[t] = i;
A.j[t] = j;
t++;
}
}
A.MMTypeCode[3]='G';
A.NrDummies = 0;
A.Pstart[0] = 0;
A.Pstart[1] = A.NrNzElts;
weightlo = (n*n)/2;
weighthi = (n*n)/2 + 1;
if (!SplitMatrixSimple(&A, 1, 0, weightlo, weighthi, &Options)) {
printf("Error\n");
exit(1);
}
/* Check result */
if (A.Pstart[0] != 0 || A.Pstart[1] != (n*n)/2 ||
A.Pstart[2] != A.NrNzElts) {
printf("Error\n");
exit(1);
}
weight0 = ComputeWeight(&A, A.Pstart[0], A.Pstart[1]-1, NULL, &Options);
weight1 = ComputeWeight(&A, A.Pstart[1], A.Pstart[2]-1, NULL, &Options);
if (weight0 != weightlo || weight1 != weighthi || weight0 < 0 || weight1 < 0) {
printf("Error\n");
exit(1);
}
printf("OK\n");
exit(0);
} /* end main */
int main(int argc, char **argv) {
struct sparsematrix A;
long n, i, j, t, P, weightlo, weighthi, weight0, weight1,
ComVol, MaxOut, MaxIn, MaxCompnts, TotCompnts;
printf("Test SplitMatrixKLFM: ");
n = 33; /* n by n dense matrix A, n odd */
A.m = n;
A.n = n;
A.NrNzElts = n*n;
P = 2; /* maximum number of parts */
A.i = (long *)malloc(A.NrNzElts*sizeof(long));
A.j = (long *)malloc(A.NrNzElts*sizeof(long));
A.Pstart = (long *)malloc((P+1)*sizeof(long));
A.RowLambda = (int *)malloc(A.m*sizeof(int));
A.ColLambda = (int *)malloc(A.n*sizeof(int));
A.RowMark = (int *)malloc(A.m*sizeof(int));
A.ColMark = (int *)malloc(A.n*sizeof(int));
if (!SetDefaultOptions(&Options)) {
printf("Error\n");
exit(1);
}
if (A.i == NULL || A.j == NULL || A.Pstart == NULL) {
printf("Error\n");
exit(1);
}
/* Fill matrix with nonzeros */
t= 0;
for (i=0; i<n; i++) {
for (j=0; j<n; j++) {
A.i[t] = i;
A.j[t] = j;
t++;
}
}
A.MMTypeCode[0]='M'; /* matrix */
A.MMTypeCode[1]='C'; /* coordinate scheme */
A.MMTypeCode[2]='P'; /* pattern only */
A.MMTypeCode[3]='G'; /* general, no symmetry */
A.NrDummies = 0;
A.dummy = NULL;
A.Pstart[0] = 0;
A.Pstart[1] = A.NrNzElts;
weightlo = n*(n-1)/2;
weighthi = n*(n+1)/2;
/* Initialise relevant options */
Options.SplitStrategy = OneDimRow;
Options.Coarsening_NrVertices = 3;
Options.Coarsening_MaxNrVtxInMatch = 4;
Options.Coarsening_StopRatio = 0.1;
Options.Coarsening_VtxMaxFractionOfWeight = 0.1;
Options.Coarsening_MatchingStrategy = MatchRandom;
Options.Coarsening_InprodMatchingOrder = IncreasingDegree;
Options.Coarsening_FineSwitchLevel = 2;
Options.Coarsening_NetScaling = NoNetScaling;
Options.Coarsening_InprodScaling = IpSclMin;
Options.Coarsening_MatchIdenticalFirst = MatchIdNo;
Options.KLFM_InitPart_NrRestarts = 10;
Options.KLFM_InitPart_MaxNrLoops = 10;
Options.KLFM_InitPart_MaxNrNoGainMoves = 50;
Options.KLFM_Refine_MaxNrLoops = 10;
Options.KLFM_Refine_MaxNrNoGainMoves = 50;
if (!SplitMatrixKLFM(&A, 1, 0, ROW, weightlo, weighthi, &Options)) {
printf("Error\n");
exit(1);
}
/* Check nonzeros */
if (A.Pstart[0] != 0 || A.Pstart[1] != n*(n-1)/2 ||
A.Pstart[2] != A.NrNzElts) {
printf("Error\n");
exit(1);
}
/* Check part weights */
weight0 = ComputeWeight(&A, A.Pstart[0], A.Pstart[1]-1, NULL, &Options);
weight1 = ComputeWeight(&A, A.Pstart[1], A.Pstart[2]-1, NULL, &Options);
if (weight0 != weightlo || weight1 != weighthi || weight0 < 0 || weight1 < 0) {
printf("Error\n");
exit(1);
}
/* Check communication */
if (!CalcCom(&A, NULL, ROW, &ComVol, &MaxOut, &MaxIn, &MaxCompnts, &TotCompnts)) {
printf("Error\n");
exit(1);
}
if (ComVol != n) {
printf("Error\n");
exit(1);
}
printf("OK\n");
exit(0);
} /* end main */
long newMondriaan(struct sparsematrix* matrix, int SplitStrategy, struct sparsematrix* output) {
struct opts Options; /* The Mondriaan options */
struct sparsematrix A; /* The Matrix;-) */
long q;
int symmetric;
FILE *File = NULL;
/* Get the parameters from the command line and initialise Options */
SetDefaultOptions(&Options);
if (!SetOptionsFromFile(&Options, "Mondriaan.defaults")) {
fprintf(stderr, "main(): warning, cannot set options from 'Mondriaan.defaults', using default options and creating standard 'Mondriaan.defaults'!\n");
File = fopen("Mondriaan.defaults", "w");
if (File != NULL) {
ExportDefaultOptions(File);
SetDefaultOptions(&Options);
fclose(File);
}
else {
fprintf(stderr, "main(): Unable to create 'Mondriaan.defaults'!\n");
}
}
/* Options.matrix = matrix;*/
Options.P = 2;
Options.eps = 0.03;
if(SplitStrategy == 8) Options.SplitStrategy = MediumGrain;
if(SplitStrategy == 5) Options.SplitStrategy = OneDimCol;
if (!ApplyOptions(&Options)) {
fprintf(stderr, "main(): could not apply given options!\n");
exit(-1);
}
A = *matrix;
/* Remove duplicate nonzeros by adding them */
if (!SparseMatrixRemoveDuplicates(&A)) {
fprintf(stderr, "main(): Unable to remove duplicates!\n");
exit(-1);
}
/* Check if matrix A is already distributed */
if (A.MMTypeCode[0] == 'D') {
/* Matrix will be partitioned again */
fprintf(stderr, "Warning: Matrix '%s' already distributed !\n",
Options.matrix);
fprintf(stderr, " (Ignoring current partitions)\n");
A.NrProcs = 0;
if (A.Pstart != NULL)
free(A.Pstart);
A.Pstart = NULL;
}
/* Register whether the input matrix was symmetric, since the symmetry type code will
be changed by the conversion to full, to the code 'G' for a general matrix */
if (A.m == A.n &&
(A.MMTypeCode[3]=='S' || A.MMTypeCode[3]=='K' || A.MMTypeCode[3]=='H')) {
symmetric = TRUE;
} else {
symmetric = FALSE;
if (Options.SplitStrategy == SFineGrain) {
fprintf(stderr, "Error: Symmetric finegrain can only be used on symmetric input matrices!\n");
exit(-1);
}
}
if (symmetric) {
if (Options.SymmetricMatrix_UseSingleEntry == SingleEntNo)
SparseMatrixSymmetric2Full(&A);
else if (Options.SplitStrategy == SFineGrain)
SparseMatrixSymmetricRandom2Lower(&A);
else if (Options.SymmetricMatrix_SingleEntryType == ETypeRandom)
SparseMatrixSymmetricLower2Random(&A);
}
if (Options.SplitStrategy == SFineGrain && Options.SymmetricMatrix_SingleEntryType == ETypeRandom)
printf("Warning: Symmetric finegrain requires lower-triangular format of symmetric matrix;\n Random single entry type option is overridden.\n");
/* If the matrix is square, add the dummies if requested.
This may lead to an enhanced vector distribution in the case of
an equal distribution of the input and output vectors. */
if (A.m == A.n &&
Options.SquareMatrix_DistributeVectorsEqual == EqVecYes &&
Options.SquareMatrix_DistributeVectorsEqual_AddDummies == DumYes)
AddDummiesToSparseMatrix(&A);
/* Set the number of processors */
A.NrProcs = Options.P;
/* Initialise Pstart with all nonzeros in processor 0 */
A.Pstart = (long *) malloc((A.NrProcs+1) * sizeof(long));
if (A.Pstart == NULL) {
fprintf(stderr, "main(): Not enough memory for Pstart!\n");
exit(-1);
}
A.Pstart[0] = 0;
for (q = 1; q <= A.NrProcs; q++)
A.Pstart[q] = A.NrNzElts;
/**** Distribute the matrix (and time it) ****/
if (!DistributeMatrixMondriaan(&A, A.NrProcs, Options.eps, &Options, 0)) {
fprintf(stderr, "main(): Unable to distribute matrix!\n");
exit(-1);
}
long ComVolumeRow, ComVolumeCol, Dummy;
CalcCom(&A, NULL, ROW, &ComVolumeRow, &Dummy, &Dummy, &Dummy, &Dummy);
CalcCom(&A, NULL, COL, &ComVolumeCol, &Dummy, &Dummy, &Dummy, &Dummy);
long comm_value = ComVolumeRow+ComVolumeCol;
MMInsertProcessorIndices(&A);
*output = copyMatrix(&A);
MMDeleteSparseMatrix(&A);
return comm_value;
} /* end main */
int main(int argc, char **argv) {
/* This test function is geared towards the current SparseMatrix2BiPartHyperGraph
function, and will not work if this function changes the order of the vertices,
nets, or pins. */
struct sparsematrix A;
struct biparthypergraph HG;
long n, i, j, t;
SetDefaultOptions(&Options);
printf("Test SparseMatrix2BiPartHyperGraph: ");
n = 100; /* n by n symmetric tridiagonal matrix A, n even.
The matrix is represented by the nonzeros on or below the diagonal. */
A.m = n;
A.n = n;
A.NrNzElts = 2*n-1;
A.i = (long *) malloc(A.NrNzElts* sizeof(long));
A.j = (long *) malloc(A.NrNzElts* sizeof(long));
A.RowMark = (int *)malloc(A.m*sizeof(int));
A.ColMark = (int *)malloc(A.n*sizeof(int));
A.ReValue = (double *) malloc(A.NrNzElts* sizeof(double));
A.dummy= (int *) malloc(n* sizeof(int));
if ( A.i == NULL || A.j == NULL || A.ReValue == NULL || A.dummy == NULL ){
printf("Error\n");
exit(1);
}
/* Initialise matrix */
A.MMTypeCode[2]='R'; /* real matrix */
A.MMTypeCode[3]='S'; /* symmetric matrix */
Options.SymmetricMatrix_UseSingleEntry = SingleEntYes;
A.NrDummies = n/2;
for (i=0; i<n/2; i++)
A.dummy[i] = TRUE;
for (i=n/2; i<n; i++)
A.dummy[i] = FALSE;
/* Fill matrix with nonzeros */
t= 0;
for (i=0; i<n; i++){
for (j=MAX(0,i-1); j<=i; j++){
A.i[t] = i;
A.j[t] = j;
if (i==j)
A.ReValue[t] = -2.0;
else
A.ReValue[t] = 1.0;
t++;
}
}
Options.SplitStrategy = OneDimRow; /* does not matter, but needs a value
unequal to LocalRatio */
if (!SparseMatrix2BiPartHyperGraph(&A, ROW, &Options, &HG)) {
printf("Error\n");
exit(1);
}
/* Check result values */
if (HG.NrVertices != n ||
HG.NrNets != n ||
HG.NrPins != 2*n-1 ) {
printf("Error\n");
exit(1);
}
/* Check vertex values */
for (t=0; t<n; t++){
if ((t==0 && HG.V[t].vtxwgt != 0) ||
(t>0 && t<n/2 && HG.V[t].vtxwgt != 2) ||
(t>=n/2 && HG.V[t].vtxwgt != 3) ||
(t==0 && HG.V[t].iStart != 0) ||
(t>0 && HG.V[t].iStart != 2*t-1) ||
HG.V[t].iEnd != 2*t+1 ||
HG.Vtx2MatIndex[t] != t ) {
printf("Error\n");
exit(1);
}
}
/* Check net values */
for (t=0; t<n; t++){
if (HG.N[t].netwgt != 1 || /* uniform netweights assumed */
HG.N[t].iStartP0 != 2*t ||
HG.N[t].iStartP1 != HG.N[t].iStartP0 ||
(t<n-1 && HG.N[t].iEnd != 2*t+2) ||
(t==n-1 && HG.N[t].iEnd != 2*n-1) ||
HG.N[t].dir != COL ||
HG.Net2MatIndex[t] != t ) {
printf("Error\n");
exit(1);
}
}
/* Check pin values */
for (t=0; t<2*n-1; t++){
if (HG.VtxAdjncy[t] != t/2 ||
(t%2==0 && HG.NetAdjncy[t] != t/2 ) ||
(t%2==1 && HG.NetAdjncy[t] != t/2 + 1) ||
(t%2==0 && HG.MatReValue[t] != -2.0) ||
(t%2==1 && HG.MatReValue[t] != 1.0) ) {
printf("Error\n");
exit(1);
}
}
printf("OK\n");
exit(0);
} /* end main */
int main(int argc, char **argv) {
/* This function is the main function, called mondriaan. It partitions
a sparse matrix, a dense input vector, and a dense output vector
for the purpose of parallel sparse matrix-vector multiplication u = A*v.
The function tries to minimise the total communication volume
by suitably partitioning the matrix, while keeping the computation
load imbalance withing a user-specified fraction epsilon.
For the resulting communication volume, it tries to balance
the communication work among the processors by suitably
partitioning the vectors. The partitioning of the input and output
vectors can either be done independently, or by imposing
distr(u) = distr(v).
The sparse matrix is read from file using the Matrix Market (MM) format.
Its partitioning is written to file using a variant of the MM format,
where all nonzeros belonging to the same processor are listed together.
If the input file is called
pi.mtx
and 4 processors are required, the following output files are produced:
pi.mtx-P4 containing the matrix distribution
pi.mtx-v4 input vector distribution
pi.mtx-u4 output vector distribution
pi.mtx-C4 Cartesian submatrices corresponding
to the matrix distribution
Furthermore, useful information on the partitioning process
and the result, such as the load imbalance and communication
volume obtained, is written to standard output.
The sparse matrix partitioner can also be used for partitioning
hypergraphs. To do this, Mondriaan must be used in 1D column mode.
Matrix columns then represent hypergraph vertices; they can be given
integer weights. Matrix rows then represent hyperedges; they cannot
be given weights.
The following additional operations are carried out by the main program:
- timing the matrix and vector partitionings
- removing duplicate nonzeros
- (optionally) checking whether the matrix is structurally symmetric
- (optionally) expanding a symmetric matrix stored in lower
triangular format to a fully stored matrix
- (optionally) adding diagonal dummy nonzeros to a square matrix
to make the diagonal completely nonzero (thereby facilitating
the vector distribution with distr(u) = distr(v)).
All partitioning options and parameters can be set using
the file mondriaan.defaults (provided the main function is compiled
with the name mondriaan.) If no such file exists, it is created with
sensible defaults.
If the input matrix has already been partitioned, it will be
partitioned again, from scratch.
*/
struct opts Options; /* The Mondriaan options */
struct sparsematrix A; /* The Matrix;-) */
long MinNrNzElts, MaxNrNzElts; /* minimum, maximum number of nonzero elements
per processor */
double AvgNrNzElts; /* average number of nonzero matrix elements per processor */
long MinWeight, MaxWeight, TotWeight; /* minimum, maximum, total column weight */
double AvgWeight; /* average column weight */
long weight, i, j;
unsigned long int ui;
long int *temp;
long ComU, ComV; /* communication cost for vectors u, v */
long q, nzq; /* processor number and its corresponding number of nonzeros */
int weighted, symmetric; /* boolean registering whether the input matrix
was weighted, symmetric */
long int *u_proc, *v_proc; /* vector distribution for u, v */
char output[MAX_WORD_LENGTH]; /* filename of the output */
char EMMcomment[2000]; /* Comments in EMM one-file output */
long int **row_perms = NULL, **col_perms = NULL; /* Will store local to global index info */
/* Below is used for Extended Matrix-Market one-file output */
long int *inv_count, **row_local2proc, **row_local2index, **col_local2proc, **col_local2index;
FILE *File = NULL;
char stdinName[] = "stdin";
/* Timing variables */
clock_t starttime, endtime;
double cputime;
#ifdef UNIX
struct timeval starttime1, endtime1;
#endif
/* Get the parameters from the command line and initialise Options */
SetDefaultOptions(&Options);
if (!SetOptionsFromFile(&Options, "Mondriaan.defaults")) {
fprintf(stderr, "main(): warning, cannot set options from 'Mondriaan.defaults', using default options and creating standard 'Mondriaan.defaults'!\n");
File = fopen("Mondriaan.defaults", "w");
if (File != NULL) {
ExportDefaultOptions(File);
SetDefaultOptions(&Options);
fclose(File);
}
else {
fprintf(stderr, "main(): Unable to create 'Mondriaan.defaults'!\n");
}
}
if (!GetParameters(&Options, argc, argv)) {
fprintf(stderr, "main(): invalid command line parameters!\n");
exit(-1);
}
if (!ApplyOptions(&Options)) {
fprintf(stderr, "main(): could not apply given options!\n");
exit(-1);
}
/* Read matrix file from disk or standard input. */
if (!strcmp(Options.matrix, "-") || !strcmp(Options.matrix, "stdin")) {
if (!MMReadSparseMatrix(stdin, &A)) {
fprintf(stderr, "main(): Could not read matrix from standard input!\n");
exit(-1);
}
Options.matrix = stdinName;
}
else {
File = fopen(Options.matrix, "r");
if (!File) {
fprintf(stderr, "main(): Unable to open '%s' for reading!\n", Options.matrix);
exit(-1);
}
if (!MMReadSparseMatrix(File, &A)) {
fprintf(stderr, "main(): Could not read matrix!\n");
exit(-1);
}
fclose(File);
}
#ifdef INFO
printf("\nUsing Mondriaan version %s.\n\n", MONDRIAANVERSION);
printf("\n**************************************************************\n");
printf("Problem statistics:\n");
printf(" Matrix: : %s\n",Options.matrix);
printf(" %s %s %s %s %s\n", A.Banner,A.Object,A.Format,A.Field,A.Symmetry);
printf(" m = Nr rows : %ld\n", A.m);
printf(" n = Nr columns : %ld\n", A.n);
printf(" nz = Nr nonzeros : %ld\n", A.NrNzElts);
/* Check if matrix A is structurally symmetric */
if (SparseMatrixStructurallySymmetric(&A))
printf(" Matrix is structurally symmetric\n\n");
else
printf(" Matrix is structurally unsymmetric\n\n");
#endif
/* Remove duplicate nonzeros by adding them */
if (!SparseMatrixRemoveDuplicates(&A)) {
fprintf(stderr, "main(): Unable to remove duplicates!\n");
exit(-1);
}
/* Check if matrix A is already distributed */
if (A.MMTypeCode[0] == 'D') {
/* Matrix will be partitioned again */
fprintf(stderr, "Warning: Matrix '%s' already distributed !\n",
Options.matrix);
fprintf(stderr, " (Ignoring current partitions)\n");
A.NrProcs = 0;
if (A.Pstart != NULL)
free(A.Pstart);
A.Pstart = NULL;
}
/* Check if matrix is weighted (thus representing a hypergraph).
In that case, it must have n column weights (representing vertex weights),
0 row weights and the split strategy must be onedimcol */
if (A.MMTypeCode[0] == 'W' && A.NrColWeights != A.n) {
fprintf(stderr, "main(): Weighted matrix with NrColWeights != n!\n");
exit(-1);
}
if (A.MMTypeCode[0] == 'W' && A.NrRowWeights != 0) {
fprintf(stderr, "Warning: Matrix '%s' has row weights!\n",
Options.matrix);
fprintf(stderr, " Row-weighted column partitioning not yet implemented\n");
fprintf(stderr, " (Ignoring row weights)\n");
A.NrRowWeights = 0;
if (A.RowWeights != NULL)
free(A.RowWeights);
A.RowWeights = NULL;
}
if (A.MMTypeCode[0] == 'W' && Options.SplitStrategy != OneDimCol) {
fprintf(stderr, "Warning: Matrix '%s' is a weighted matrix!\n", Options.matrix);
fprintf(stderr, " must be partitioned by onedimcol strategy\n");
fprintf(stderr, " (Ignoring requested split strategy)\n");
Options.SplitStrategy = OneDimCol;
}
/* Register whether the input matrix was weighted, since the object type code will
be changed by the partitioning procedure to the code 'D' for a distributed matrix */
if (A.MMTypeCode[0] == 'W')
weighted = TRUE;
else
weighted = FALSE;
/* Register whether the input matrix was symmetric, since the symmetry type code will
be changed by the conversion to full, to the code 'G' for a general matrix */
if (A.m == A.n &&
(A.MMTypeCode[3]=='S' || A.MMTypeCode[3]=='K' || A.MMTypeCode[3]=='H')) {
symmetric = TRUE;
} else {
symmetric = FALSE;
if (Options.SplitStrategy == SFineGrain) {
fprintf(stderr, "Error: Symmetric finegrain can only be used on symmetric input matrices!\n");
exit(-1);
}
}
if (symmetric) {
if (Options.SymmetricMatrix_UseSingleEntry == SingleEntNo)
SparseMatrixSymmetric2Full(&A);
else if (Options.SplitStrategy == SFineGrain)
SparseMatrixSymmetricRandom2Lower(&A);
else if (Options.SymmetricMatrix_SingleEntryType == ETypeRandom)
SparseMatrixSymmetricLower2Random(&A);
}
if (Options.SplitStrategy == SFineGrain && Options.SymmetricMatrix_SingleEntryType == ETypeRandom)
printf("Warning: Symmetric finegrain requires lower-triangular format of symmetric matrix;\n Random single entry type option is overridden.\n");
/* If the matrix is square, add the dummies if requested.
This may lead to an enhanced vector distribution in the case of
an equal distribution of the input and output vectors. */
if (A.m == A.n &&
Options.SquareMatrix_DistributeVectorsEqual == EqVecYes &&
Options.SquareMatrix_DistributeVectorsEqual_AddDummies == DumYes)
AddDummiesToSparseMatrix(&A);
/* Set the number of processors */
A.NrProcs = Options.P;
#ifdef INFO
printf(" P = Nr processors : %d\n", A.NrProcs);
printf(" Nr dummies added : %ld\n", A.NrDummies);
printf(" allowed imbalance : %g %c\n", 100*Options.eps, '%');
printf("\n****** Mondriaan matrix distribution ******\n\n");
#endif
/* Initialise Pstart with all nonzeros in processor 0 */
A.Pstart = (long *) malloc((A.NrProcs+1) * sizeof(long));
if (A.Pstart == NULL) {
fprintf(stderr, "main(): Not enough memory for Pstart!\n");
exit(-1);
}
A.Pstart[0] = 0;
for (q = 1; q <= A.NrProcs; q++)
A.Pstart[q] = A.NrNzElts;
/**** Distribute the matrix (and time it) ****/
starttime = clock();
#ifdef UNIX
gettimeofday(&starttime1, NULL);
#endif
if (!DistributeMatrixMondriaan(&A, A.NrProcs, Options.eps, &Options, 0)) {
fprintf(stderr, "main(): Unable to distribute matrix!\n");
exit(-1);
}
endtime = clock();
#ifdef UNIX
gettimeofday(&endtime1, NULL);
#endif
cputime = ((double) (endtime - starttime)) / CLOCKS_PER_SEC;
#ifdef TIME
printf(" matrix distribution CPU-time : %f seconds\n", cputime);
#ifdef UNIX
printf(" matrix distribution elapsed time: %f seconds\n",
(endtime1.tv_sec - starttime1.tv_sec) +
(endtime1.tv_usec - starttime1.tv_usec) / 1000000.0);
#endif
fflush(stdout);
#endif
/* Remove the dummies */
if (A.m == A.n &&
Options.SquareMatrix_DistributeVectorsEqual == EqVecYes &&
Options.SquareMatrix_DistributeVectorsEqual_AddDummies == DumYes)
RemoveDummiesFromSparseMatrix(&A);
/* Print information about the number of matrix elements */
AvgNrNzElts = (double) A.NrNzElts / A.NrProcs;
MinNrNzElts = LONG_MAX;
MaxNrNzElts = LONG_MIN;
for (q = 0; q < A.NrProcs; q++) {
nzq = A.Pstart[q+1] - A.Pstart[q];
if (nzq < MinNrNzElts)
MinNrNzElts = nzq;
if (nzq > MaxNrNzElts)
MaxNrNzElts = nzq;
}
#ifdef INFO
printf(" Nr nonzero matrix elements:\n");
printf(" tot : %ld \n", A.NrNzElts);
printf(" avg = tot/P : %g \n", AvgNrNzElts);
printf(" max : %ld \n", MaxNrNzElts);
printf(" min : %ld \n", MinNrNzElts);
printf(" imbalance : %g %c\n",
100 * (MaxNrNzElts/AvgNrNzElts - 1.0), '%');
#endif
if (weighted){
A.MMTypeCode[0] = 'W'; /* temporarily, needed for computing weights */
TotWeight = ComputeWeight(&A, 0, A.NrNzElts-1, NULL, &Options);
AvgWeight = (double) TotWeight / A.NrProcs;
MinWeight = LONG_MAX;
MaxWeight = LONG_MIN;
for (q = 0; q < A.NrProcs; q++) {
weight = ComputeWeight(&A, A.Pstart[q], A.Pstart[q+1]-1, NULL, &Options);
if (weight < MinWeight)
MinWeight = weight;
if (weight > MaxWeight)
MaxWeight = weight;
}
A.MMTypeCode[0] = 'D'; /* back to distributed matrix */
#ifdef INFO
printf(" Vertex (column) weight:\n");
printf(" tot : %ld \n", TotWeight);
printf(" avg = tot/P : %g \n", AvgWeight);
printf(" max : %ld \n", MaxWeight);
printf(" min : %ld \n", MinWeight);
printf(" imbalance : %g %c\n",
100 * (MaxWeight/AvgWeight - 1.0), '%');
#endif
}
/* Convert randomly represented symmetric matrix to standard
lower triangular form */
if (symmetric &&
Options.SymmetricMatrix_UseSingleEntry == SingleEntYes &&
Options.SymmetricMatrix_SingleEntryType == ETypeRandom)
SparseMatrixSymmetricRandom2Lower(&A);
/* Writing (distributed) matrix info */
if( Options.OutputMode == MultipleFiles ) {
/* Write the distributed matrix to file */
sprintf(output, "%s-P%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
MMWriteSparseMatrix(&A, File, NULL, &Options);
fclose(File);
}
/* Commit to permutation */
if (A.row_perm_inv != NULL && A.col_perm_inv != NULL) {
for( i=0; i<A.NrNzElts; i++ ) {
A.i[ i ] = A.row_perm_inv[ A.i[ i ] ];
A.j[ i ] = A.col_perm_inv[ A.j[ i ] ];
}
}
/* Write permuted matrix */
sprintf(output, "%s-reor-P%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
A.MMTypeCode[0] = 'M';
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
MMWriteSparseMatrix(&A, File, NULL, &Options);
fclose(File);
}
A.MMTypeCode[0] = 'D';
/* Write out block information */
if (A.rowBoundaries) {
if (remembrance_get( A.rowBoundaries )==NULL) {
fprintf(stderr, "main(): Error during read-out of row boundaries\n");
exit( -1 );
}
sprintf(output, "%s-rowblocks%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
remembrance_write(A.rowBoundaries, File);
fclose(File);
}
}
if (Options.SplitStrategy != SFineGrain)
if (A.colBoundaries) {
if (remembrance_get( A.colBoundaries )==NULL) {
fprintf(stderr, "main(): Error during read-out of column boundaries\n");
exit( -1 );
}
sprintf(output, "%s-colblocks%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
remembrance_write(A.colBoundaries, File);
fclose(File);
}
}
/* Undo commit to permutation */
if (A.row_perm_inv != NULL && A.col_perm_inv != NULL) {
for( i=0; i<A.NrNzElts; i++ ) {
A.i[ i ] = A.row_perm[ A.i[ i ] ];
A.j[ i ] = A.col_perm[ A.j[ i ] ];
}
}
/* Write out local matrix format */
row_perms = malloc( (A.NrProcs+1) * sizeof( long int * ) );
col_perms = malloc( (A.NrProcs+1) * sizeof( long int * ) );
if( !row_perms || !col_perms ||
!SparseMatrixOriginal2Local(&A, row_perms, col_perms) ) {
fprintf(stderr, "main(): Unable to transform to local view!");
exit(-1);
}
sprintf(output, "%s-local%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
MMWriteSparseMatrix(&A, File, NULL, &Options);
fclose(File);
}
/* Transform matrix back */
for( i=0; i<A.NrProcs; i++ )
for(j=A.Pstart[i]; j<A.Pstart[i+1]; j++ ) {
A.i[j] = row_perms[i][ A.i[j] ];
A.j[j] = col_perms[i][ A.j[j] ];
}
A.ViewType = ViewTypeOriginal;
/* Write out matrix the entries of which are processor indices. */
if (!MMInsertProcessorIndices(&A)) {
fprintf(stderr, "main(): Unable to write processor indices!\n");
exit(-1);
}
A.MMTypeCode[0] = 'M';
sprintf(output, "%s-I%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
MMWriteSparseMatrix(&A, File, NULL, &Options);
fclose(File);
}
A.MMTypeCode[0] = 'D';
/* Write out permutations. */
if (A.col_perm != NULL) {
sprintf(output, "%s-col%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
WriteVector(A.col_perm, 0, NULL, A.n, 0, File, &Options);
fclose(File);
}
}
if (A.row_perm != NULL) {
sprintf(output, "%s-row%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
WriteVector(A.row_perm, 0, NULL, A.m, 0, File, &Options);
fclose(File);
}
}
/* Free local to global index */
for( i=0; i<A.NrProcs+1; i++ ) {
free( row_perms[i] );
free( col_perms[i] );
}
free( row_perms ); free( col_perms );
} else if (Options.OutputMode == OneFile) {
/* Extended MatrixMarket output */
if (Options.OutputFormat != OutputEMM) {
fprintf(stderr, "main(): Single-file output requires EMM format!\n");
fprintf(stderr, " overriding output format to EMM.\n");
Options.OutputFormat = OutputEMM;
}
sprintf(EMMcomment, "%% File generated by running Mondriaan on A=%s with p=%d and \\epsilon=%f\n", Options.matrix, A.NrProcs, Options.eps);
sprintf(EMMcomment, "%s%% Main matrix corresponds to the distributed version of A\n", EMMcomment);
if (A.row_perm_inv != NULL && A.col_perm_inv != NULL) {
switch( Options.OrderPermutation ) {
case OrderPrefix:
sprintf(EMMcomment, "%s%% This is followed by the (not-distributed) reverse Bordered Block Diagonal permuted matrix 'PAQ',\n", EMMcomment);
break;
case OrderPostfix:
sprintf(EMMcomment, "%s%% This is followed by the (not-distributed) Bordered Block Diagonal permuted matrix 'PAQ',\n", EMMcomment);
break;
default:
sprintf(EMMcomment, "%s%% This is followed by the (not-distributed) Separated Block Diagonal permuted matrix 'PAQ',\n", EMMcomment);
}
sprintf(EMMcomment, "%s%% followed by the row-block boundary vector 'Row-boundaries',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the row-block hierarchy vector 'Row-hierarchy',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the column-block boundary vector 'Column-boundaries',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the column-block hierarchy vector 'Column-hierarchy',\n", EMMcomment);
}
sprintf(EMMcomment, "%s%% followed by a local view of the distributed form of A 'Local-A',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by a mapping of local row indices to global row indices 'LocalRow2Global',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by a mapping of local column indices to global column indices 'LocalCol2Global',\n", EMMcomment);
if (A.row_perm != NULL)
sprintf(EMMcomment, "%s%% followed by the row permutation vector corresponding to P in PAQ 'Row-permutation',\n", EMMcomment);
if (A.col_perm != NULL)
sprintf(EMMcomment, "%s%% followed by the column permutation vector corresponding to Q in PAQ 'Column-permutation',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the input vector distribution 'Input-vector',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the output vector distribution 'Output-vector',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the local lengths of the output vectors 'OutputVectorLengths',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the matrix row index to output vector processor ID mapping 'LocalRow2Processor',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the matrix row index to output vector index mapping 'LocalRow2Index',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the matrix column index to input vector processor ID mapping 'LocalCol2Processor',\n", EMMcomment);
sprintf(EMMcomment, "%s%% followed by the matrix column index to input vector index mapping 'LocalCol2Index'.\n", EMMcomment);
A.comment = EMMcomment;
sprintf(output, "%s-P%d.emm", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) {
fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
exit( -1 );
}
MMWriteSparseMatrix(&A, File, NULL, &Options);
/* Commit to permutation */
if (A.row_perm_inv != NULL && A.col_perm_inv != NULL) {
for( i=0; i<A.NrNzElts; i++ ) {
A.i[ i ] = A.row_perm_inv[ A.i[ i ] ];
A.j[ i ] = A.col_perm_inv[ A.j[ i ] ];
}
}
/* Write permuted matrix */
A.MMTypeCode[0] = 'M';
MMWriteSparseMatrix(&A, File, "PAQ", &Options);
A.MMTypeCode[0] = 'D';
/* Write out block information */
if (A.rowBoundaries) {
if (remembrance_get( A.rowBoundaries )==NULL) {
fprintf(stderr, "main(): Error during read-out of row boundaries\n");
exit( -1 );
}
temp = malloc(A.rowBoundaries->size * sizeof( long int ));
for(ui=0; ui<A.rowBoundaries->size; ui++)
temp[ ui ] = A.rowBoundaries->vector[ ui ].index;
WriteVector(temp, 0, "Row-boundaries", A.rowBoundaries->size, 0, File, &Options);
for(ui=0; ui<A.rowBoundaries->size-1; ui++ )
temp[ ui ] = A.rowBoundaries->vector[ ui ].parent == ULONG_MAX ? 0 : A.rowBoundaries->vector[ ui ].parent + 1;
WriteVector(temp, 1, "Row-hierarchy", A.rowBoundaries->size-1, 0, File, &Options);
free(temp);
}
if (Options.SplitStrategy != SFineGrain) {
if (A.colBoundaries) {
if (remembrance_get( A.colBoundaries )==NULL) {
fprintf(stderr, "main(): Error during read-out of column boundaries\n");
exit( -1 );
}
temp = malloc( A.colBoundaries->size * sizeof( long int ) );
for(ui=0; ui<A.colBoundaries->size; ui++)
temp[ ui ] = A.colBoundaries->vector[ ui ].index;
WriteVector(temp, 0, "Col-boundaries", A.colBoundaries->size, 0, File, &Options);
for(ui=0; ui<A.colBoundaries->size-1; ui++ )
temp[ ui ] = A.colBoundaries->vector[ ui ].parent == ULONG_MAX ? 0 : A.colBoundaries->vector[ ui ].parent + 1;
WriteVector(temp, 1, "Col-hierarchy", A.colBoundaries->size-1, 0, File, &Options);
free(temp);
}
}
/* Write out local matrix format */
row_perms = malloc( (A.NrProcs+1) * sizeof( long int * ) );
col_perms = malloc( (A.NrProcs+1) * sizeof( long int * ) );
if( !row_perms || !col_perms ||
!SparseMatrixOriginal2Local(&A, row_perms, col_perms) ) {
fprintf(stderr, "main(): Unable to transform to local view!");
exit(-1);
}
MMWriteSparseMatrix(&A, File, "Local-A", &Options);
/* Transform matrix back from Local to global view */
for( i=0; i<A.NrProcs; i++ )
for(j=A.Pstart[i]; j<A.Pstart[i+1]; j++ ) {
A.i[j] = row_perms[i][ A.i[j] ];
A.j[j] = col_perms[i][ A.j[j] ];
}
A.ViewType = ViewTypeOriginal;
/* Permutate row_perms and col_perms so that they point
* to global indices (corresponding to A) instead of
* permuted indices (corresponding to PAQ).
for( i=0; i<A.NrProcs; i++ ) {
for( j=0; j<row_perms[A.NrProcs][i]; j++ )
row_perms[i][j] = row_perms[i][j];
for( j=0; j<col_perms[A.NrProcs][i]; j++ )
col_perms[i][j] = col_perms[i][j];
} */
WriteVectorCollection(row_perms, "LocalRow2Global", A.NrProcs, row_perms[A.NrProcs], File);
WriteVectorCollection(col_perms, "LocalCol2Global", A.NrProcs, col_perms[A.NrProcs], File);
/* Write out matrix the entries of which are processor indices */
/*if (!MMInsertProcessorIndices(&A)) {
fprintf(stderr, "main(): Unable to write processor indices!\n");
exit(-1);
}
A.MMTypeCode[0] = 'M';
tempchar = A.MMTypeCode[2];
A.MMTypeCode[2] = 'I';
MMWriteSparseMatrix(&A, File, "Global-A", &Options);
A.MMTypeCode[0] = 'D';
A.MMTypeCode[2] = tempchar;*/
/* Write out permutations */
if (A.row_perm != NULL) {
WriteVector(A.row_perm, 0, "Row-permutation", A.m, 0, File, &Options);
}
if (A.col_perm != NULL) {
WriteVector(A.col_perm, 0, "Column-permutation", A.n, 0, File, &Options);
}
} else if (Options.OutputMode == DIMACS) {
/* For DIMACS we only need the vector distribution. */
} else {
fprintf(stderr, "main(): Unknown output mode!\n" );
exit( -1 );
}
/* Convert symmetrically partitioned, symmetric matrix to full form
for vector distribution */
if (symmetric &&
Options.SymmetricMatrix_UseSingleEntry == SingleEntYes) {
/* Some of the above transformations can make A arbitrary symmetric
instead of lower-traingular */
SparseMatrixSymmetricRandom2Lower(&A);
SparseMatrixSymmetric2Full(&A); /* now A.MMTypeCode[3]='G' */
/* Print information about the number of matrix elements */
AvgNrNzElts = (double) A.NrNzElts / A.NrProcs;
MinNrNzElts = LONG_MAX;
MaxNrNzElts = LONG_MIN;
for (q = 0; q < A.NrProcs; q++) {
nzq = A.Pstart[q+1] - A.Pstart[q];
if (nzq < MinNrNzElts)
MinNrNzElts = nzq;
if (nzq > MaxNrNzElts)
MaxNrNzElts = nzq;
}
#ifdef INFO
printf(" After conversion of distributed symmetric matrix to"
" distributed full matrix\n");
printf(" Nr matrix elements:\n");
printf(" tot : %ld \n", A.NrNzElts);
printf(" avg = tot/P : %g \n", AvgNrNzElts);
printf(" max : %ld \n", MaxNrNzElts);
printf(" min : %ld \n", MinNrNzElts);
printf(" imbalance : %g %c\n",
100 * (MaxNrNzElts / AvgNrNzElts - 1.0), '%');
/* vertex weight information is not printed, since it is not
meaningful here */
#endif
}
#ifdef INFO
printf("\n****** Mondriaan vector distribution ******\n");
#endif
/* Allocate memory for processor numbers for the vector components,
for u = A*v with A an m by n matrix */
u_proc = (long int *) malloc(A.m * sizeof(long int));
v_proc = (long int *) malloc(A.n * sizeof(long int));
if (u_proc == NULL || v_proc == NULL) {
fprintf(stderr, "main(): Not enough memory for vectors u_proc and v_proc!\n");
exit(-1);
}
/**** Distribute the vector (and time it) ****/
starttime = clock();
#ifdef UNIX
gettimeofday(&starttime1, NULL);
#endif
if (A.m == A.n && Options.SquareMatrix_DistributeVectorsEqual == EqVecYes) {
/* Distribute the vectors equally */
if (symmetric &&
Options.SymmetricMatrix_UseSingleEntry == SingleEntYes) {
/* Distribute v independently, and then use the same distribution for u */
ComV = DistributeVec(&A, v_proc, ROW, &Options);
if (ComV < 0) {
fprintf(stderr, "main(): Unable to distribute vectors!\n");
exit(-1);
}
for (i = 0; i < A.m; i++)
u_proc[i] = v_proc[i];
ComU = ComV;
#ifdef INFO
printf("\nCommunication results for u are the same as for v\n");
printf("but with sends and receives reversed\n\n");
#endif
} else {
/* Distribute u and v together */
ComU = DistributeVecOrigEq(&A, u_proc, v_proc, &Options);
if (ComU < 0) {
fprintf(stderr, "main(): Unable to distribute vectors!\n");
exit(-1);
}
ComV = 0;
}
} else {
/* Distribute the vectors independently */
ComV = DistributeVec(&A, v_proc, ROW, &Options);
if (ComV < 0) {
fprintf(stderr, "main(): Unable to distribute input vector!\n");
exit(-1);
}
ComU = DistributeVec(&A, u_proc, COL, &Options);
if (ComU < 0) {
fprintf(stderr, "main(): Unable to distribute output vector!\n");
exit(-1);
}
}
endtime = clock();
#ifdef UNIX
gettimeofday(&endtime1, NULL);
#endif
cputime = ((double) (endtime - starttime)) / CLOCKS_PER_SEC;
#ifdef TIME
printf(" vector distribution CPU-time : %f seconds\n", cputime);
#ifdef UNIX
printf(" vector distribution elapsed time : %f seconds\n",
(endtime1.tv_sec - starttime1.tv_sec) +
(endtime1.tv_usec - starttime1.tv_usec) / 1000000.0);
#endif
fflush(stdout);
#endif
#ifdef INFO
printf("\nCommunication cost for u = A v: %ld (%g) \n", ComU + ComV,
((double) ComU / A.m + (double) ComV / A.n) /2);
fflush(stdout);
#endif
/* Write vector info */
if (Options.OutputMode == MultipleFiles) {
/* Write the vector distribution to file */
sprintf(output, "%s-u%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
WriteVectorDistribution(u_proc, NULL, A.m, A.NrProcs, File, &Options);
fclose(File);
}
sprintf(output, "%s-v%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
WriteVectorDistribution(v_proc, NULL, A.n, A.NrProcs, File, &Options);
fclose(File);
}
} else if (Options.OutputMode == OneFile) {
/* Write the vector distribution to file */
WriteVectorDistribution(v_proc, "Input-vector", A.n, A.NrProcs, File, &Options);
WriteVectorDistribution(u_proc, "Output-vector", A.m, A.NrProcs, File, &Options);
/* LocalRow2Processor */
inv_count = NULL;
row_local2proc = row_local2index = col_local2proc = col_local2index = NULL;
if(!SparseMatrixLocal2Vector(&A, row_perms, u_proc, &inv_count, &row_local2proc, &row_local2index, 0)) {
fprintf(stderr, "Error during derivation of local index arrays for SpMV multiplication (row direction)\n");
exit(-1);
}
WriteVector(inv_count, 0, "OutputVectorLengths", A.NrProcs, 0, File, &Options);
WriteVectorCollection(row_local2proc, "LocalRow2Processor", A.NrProcs, row_perms[A.NrProcs], File);
WriteVectorCollection(row_local2index, "LocalRow2Index", A.NrProcs, row_perms[A.NrProcs], File);
for( i=0; i<A.NrProcs; i++ ) {
free( row_local2proc[i] );
free( row_local2index[i] );
}
free( row_local2proc ); free( row_local2index );
/* LocalCol2Processor */
free(inv_count);
if(!SparseMatrixLocal2Vector(&A, col_perms, v_proc, &inv_count, &col_local2proc, &col_local2index, 1)) {
fprintf(stderr, "Error during derivation of local index arrays for SpMV multiplication (column direction)\n");
exit(-1);
}
WriteVector(inv_count, 0, "InputVectorLengths", A.NrProcs, 0, File, &Options);
WriteVectorCollection(col_local2proc, "LocalCol2Processor", A.NrProcs, col_perms[A.NrProcs], File);
WriteVectorCollection(col_local2index, "LocalCol2Index", A.NrProcs, col_perms[A.NrProcs], File);
for( i=0; i<A.NrProcs; i++ ) {
free( col_local2proc[i] );
free( col_local2index[i] );
}
free( col_local2proc ); free( col_local2index );
free(inv_count);
/* Also free local to global index */
for( i=0; i<A.NrProcs+1; i++ ) {
free( row_perms[i] );
free( col_perms[i] );
}
free( row_perms ); free( col_perms );
fclose(File);
} else if (Options.OutputMode == DIMACS) {
if (A.m != A.n || Options.SquareMatrix_DistributeVectorsEqual != EqVecYes) {
fprintf(stderr, "main(): Unequal vector distributions in DIMACS mode!\n");
}
/* Only write the vector distribution to disk. */
sprintf(output, "%s%d.part", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) {
fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
} else {
for (i = 0; i < A.m; i++) fprintf(File, "%ld\n", u_proc[i]);
fclose(File);
}
}
if (Options.OutputMode == MultipleFiles) {
/* Write the index sets of the Cartesian submatrices to file */
sprintf(output, "%s-C%d", Options.matrix, A.NrProcs);
File = fopen(output, "w");
if (!File) fprintf(stderr, "main(): Unable to open '%s' for writing!\n", output);
else {
MMWriteCartesianSubmatrices(&A, File);
fclose(File);
}
}
/* Free memory */
MMDeleteSparseMatrix(&A);
free(v_proc);
free(u_proc);
/* Exit :) */
exit(0);
} /* end main */
int main(int argc, char **argv) {
struct opts Options;
int i;
time_t CurrentTime;
clock_t TotalClock = clock();
/* Apply Mondriaan options. */
SetDefaultOptions(&Options);
if (!SetOptionsFromFile(&Options, "Mondriaan.defaults")) {
fprintf(stderr, "main(): warning, cannot set options from 'Mondriaan.defaults', using default options!\n");
}
if (!ParseCommandLineOptions(&Options, argc, argv)) {
fprintf(stderr, "main(): invalid command line parameters!\n");
exit(EXIT_FAILURE);
}
if (!ApplyOptions(&Options)) {
fprintf(stderr, "main(): could not apply given options!\n");
exit(EXIT_FAILURE);
}
if (NumMatrices <= 0 || Matrices == NULL) {
fprintf(stderr, "main(): Invalid number of supplied matrices or samples!\n");
exit(EXIT_FAILURE);
}
/* Start profiling ... */
fprintf(stderr, "Profiling Mondriaan for %d matrices, %d samples, %s processors, and %f imbalance.\n", NumMatrices, NumSamples, argv[1], Options.eps);
for (i = 0; i < NumMatrices*NumNumProcs; ++i) {
int j;
fprintf(stderr, "[% 4d/%d] (% 4ld) %s ", i + 1, NumMatrices*NumNumProcs, Matrices[i].P, Matrices[i].File);
fflush(stderr);
if (!SetupAttributes(&Matrices[i])) {
fprintf(stderr, "main(): Cannot setup attributes!\n");
exit(EXIT_FAILURE);
}
Options.P = Matrices[i].P;
/* Take the requested number of samples. */
for (j = 0; j < NumSamples; ++j) {
struct sparsematrix A;
long int *UAssign, *VAssign, Symmetric;
FILE *File;
long l;
int k;
clock_t Clock;
double Duration;
long MaxNz, MinNz;
long MaxComU, MaxComV, ComVolU, ComVolV;
fprintf(stderr, ".");
fflush(stderr);
/* Read matrix from disk. */
File = fopen(Matrices[i].File, "r");
if (!File) {
fprintf(stderr, "main(): Could not open '%s' for reading!\n", Matrices[i].File);
exit(EXIT_FAILURE);
}
if (!MMReadSparseMatrix(File, &A)) {
fprintf(stderr, "main(): Could not read matrix!\n");
exit(EXIT_FAILURE);
}
fclose(File);
/* Remove double zeroes. */
if (!SparseMatrixRemoveDuplicates(&A)) exit(EXIT_FAILURE);
/* Check symmetry. */
if (A.m == A.n && (A.MMTypeCode[3] == 'S' || A.MMTypeCode[3] == 'K' || A.MMTypeCode[3] == 'H')) Symmetric = TRUE;
else Symmetric = FALSE;
if (Symmetric)
{
if (Options.SymmetricMatrix_UseSingleEntry == SingleEntNo) SparseMatrixSymmetric2Full(&A);
else if (Options.SymmetricMatrix_SingleEntryType == ETypeRandom) SparseMatrixSymmetricLower2Random(&A);
}
/* Add dummies if requested. */
if (A.m == A.n && Options.SquareMatrix_DistributeVectorsEqual == EqVecYes && Options.SquareMatrix_DistributeVectorsEqual_AddDummies == DumYes) AddDummiesToSparseMatrix(&A);
/* Initialise processor array. */
A.NrProcs = Options.P;
A.Pstart = (long *)malloc((A.NrProcs + 1)*sizeof(long));
if (A.Pstart == NULL) {
fprintf(stderr, "main(): Cannot allocate processor array!\n");
exit(EXIT_FAILURE);
}
A.Pstart[0] = 0;
for (k = 1; k <= A.NrProcs; ++k) {
A.Pstart[k] = A.NrNzElts;
}
/* Distribute the processors among the matrix entries. */
SetRandomSeed(Options.Seed = 137*j + 12345);
/* ==== Start Mondriaan */
Clock = clock();
if (!DistributeMatrixMondriaan(&A, Options.P, Options.eps, &Options, NULL)) {
fprintf(stderr, "main(): Unable to distribute matrix!\n");
exit(EXIT_FAILURE);
}
/* Remove dummies. */
if (A.m == A.n && Options.SquareMatrix_DistributeVectorsEqual == EqVecYes && Options.SquareMatrix_DistributeVectorsEqual_AddDummies == DumYes) RemoveDummiesFromSparseMatrix(&A);
/* Convert randomly represented matrix to lower triangular form. */
if (Symmetric && Options.SymmetricMatrix_UseSingleEntry == SingleEntYes && Options.SymmetricMatrix_SingleEntryType == ETypeRandom) SparseMatrixSymmetricRandom2Lower(&A);
/* Distribute vectors. */
UAssign = (long int *)malloc(A.m*sizeof(long int));
VAssign = (long int *)malloc(A.n*sizeof(long int));
if (UAssign == NULL || VAssign == NULL) {
fprintf(stderr, "main(): Cannot allocate vertex assign arrays!\n");
exit(EXIT_FAILURE);
}
/* Convert symmetrically partitioned matrix to full form. */
if (Symmetric && Options.SymmetricMatrix_UseSingleEntry == SingleEntYes) SparseMatrixSymmetric2Full(&A);
if (A.m == A.n && Options.SquareMatrix_DistributeVectorsEqual == EqVecYes) {
if (Symmetric && Options.SymmetricMatrix_UseSingleEntry == SingleEntYes) {
MaxComV = DistributeVec(&A, VAssign, ROW, &Options);
if (MaxComV < 0) {
fprintf(stderr, "main(): Unable to distribute vector!\n");
exit(EXIT_FAILURE);
}
for (k = 0; k < A.m; k++) {
UAssign[k] = VAssign[k];
}
MaxComU = MaxComV;
}
else {
MaxComU = DistributeVecOrigEq(&A, UAssign, VAssign, &Options);
if (MaxComU < 0) {
fprintf(stderr, "main(): Unable to distribute vector!\n");
exit(EXIT_FAILURE);
}
MaxComV = 0;
}
}
else {
MaxComV = DistributeVec(&A, VAssign, ROW, &Options);
MaxComU = DistributeVec(&A, UAssign, COL, &Options);
if (MaxComV < 0 || MaxComU < 0) {
fprintf(stderr, "main(): Unable to distribute vector!\n");
exit(EXIT_FAILURE);
}
}
/* ==== Stop Mondriaan */
/* Calculate duration. */
Duration = (double)(clock() - Clock)/(double)CLOCKS_PER_SEC;
/* Determine minimum and maximum number of assigned nonzeroes. */
MaxNz = MinNz = A.Pstart[1] - A.Pstart[0];
for (k = 1; k < A.NrProcs; ++k) {
l = A.Pstart[k + 1] - A.Pstart[k];
if (l > MaxNz) MaxNz = l;
if (l < MinNz) MinNz = l;
}
/* Calculate communication volume. */
if (!CalcCom(&A, VAssign, ROW, &ComVolV, &l, &l, &l, &l) ||
!CalcCom(&A, UAssign, COL, &ComVolU, &l, &l, &l, &l)) {
fprintf(stderr, "main(): Unable to calculate communication volume!\n");
exit(EXIT_FAILURE);
}
/* Store attributes. */
Matrices[i].NumNz = A.NrNzElts;
Matrices[i].Rows = A.m;
Matrices[i].Cols = A.n;
Matrices[i].Attributes[0].Data[j] = Duration;
Matrices[i].Attributes[1].Data[j] = (double)(Options.P*MaxNz - A.NrNzElts)/(double)A.NrNzElts;
Matrices[i].Attributes[2].Data[j] = (double)(MaxComV + MaxComU);
Matrices[i].Attributes[3].Data[j] = (double)(ComVolV + ComVolU);
/* Free memory. */
MMDeleteSparseMatrix(&A);
free(UAssign);
free(VAssign);
}
/* Average attributes. */
if (!AverageAndFreeAttributes(&Matrices[i])) {
fprintf(stderr, "main(): Cannot setup attributes!\n");
exit(EXIT_FAILURE);
}
fprintf(stderr, "\n");
}
/* Write accumulated data to stdout. */
fprintf(stderr, "Finished profiling, writing data ...\n");
printf("%% Profiled Mondriaan for %d matrices, %d samples, and %f imbalance.\n", NumMatrices, NumSamples, Options.eps);
printf("\\documentclass[a4paper, 10pt]{article}\n\n");
printf("\\usepackage{lscape}\n");
printf("\\usepackage{longtable}\n\n");
printf("\\author{\\texttt{Profile.c}}\n");
CurrentTime = time(NULL);
printf("\\date{%s}\n", asctime(localtime(&CurrentTime)));
printf("\\title{Profiling Mondriaan %s with %d %s}\n\n", MONDRIAANVERSION, NumMatrices, NumMatrices > 1 ? "matrices" : "matrix");
printf("\\begin{document}\n\n");
printf("\\maketitle\n\n");
printf("\\section{Results}\n\n");
printf("Used Mondriaan version %s to distribute %d matrices (listed in table \\ref{MondriaanMatrices}) over %d processors with maximum imbalance %f, taking the average of %d samples. The used options can be found in table \\ref{MondriaanSettings} en the numerical results in table \\ref{MondriaanResults}.\n", MONDRIAANVERSION, NumMatrices, Options.P, Options.eps, NumSamples);
printf("This took %.1f minutes in total.\n\n", (double)(clock() - TotalClock)/(60.0*(double)CLOCKS_PER_SEC));
/* Export options. */
printf("\\begin{table}\n");
printf("\\caption{Mondriaan configuration.}\n");
printf("\\label{MondriaanSettings}\n");
printf("\\begin{center}\n");
if (!ExportOptionsToLaTeX(stdout, &Options)) {
fprintf(stderr, "main(): Unable to create option table!\n");
exit(EXIT_FAILURE);
}
printf("\\end{center}\n");
printf("\\end{table}\n\n");
/* Export list of test matrices. */
printf("\\begin{table}\n");
printf("\\caption{%d tested matrices.}\n", NumMatrices);
printf("\\label{MondriaanMatrices}\n");
printf("\\begin{center}\n");
printf("\\begin{tabular}{l|lll}\nFile & $\\mathit{nz}$ & $m$ & $n$ \\\\\n\\hline\n");
for (i = 0; i < NumMatrices*NumNumProcs; i += NumNumProcs) {
printf("\\verb|%s| & %ld & %ld & %ld \\\\\n", Matrices[i].File, Matrices[i].NumNz, Matrices[i].Rows, Matrices[i].Cols);
}
printf("\\hline\n");
printf("\\end{tabular}\n");
printf("\\end{center}\n");
printf("\\end{table}\n\n");
/* Export test data. */
printf("\\begin{landscape}\n");
printf("\\begin{longtable}{lrrrrr}");
printf("\\caption[Profile results]{Profile results for %d matrices.}\n", NumMatrices);
printf("\\label{MondriaanResults} \\\\\n\n");
printf("\\multicolumn{1}{c}{File} & \\multicolumn{1}{c}{$p$} & \\multicolumn{1}{c}{Time (s)} & \\multicolumn{1}{c}{$\\varepsilon$} & \\multicolumn{1}{c}{Max. com.} & \\multicolumn{1}{c}{Com. vol.} \\\\ \\hline\n");
printf("\\endfirsthead\n\n");
printf("\\multicolumn{6}{c}{\\tablename\\ \\thetable{} -- continued from previous page.} \\\\\n");
printf("\\multicolumn{1}{c}{File} & \\multicolumn{1}{c}{$p$} & \\multicolumn{1}{c}{Time (s)} & \\multicolumn{1}{c}{$\\varepsilon$} & \\multicolumn{1}{c}{Max. com.} & \\multicolumn{1}{c}{Com. vol.} \\\\ \\hline\n");
printf("\\endhead\n\n");
printf("\\multicolumn{6}{c}{Continued on next page.} \\\\\n");
printf("\\endfoot\n\n");
printf("\\hline\n\\endlastfoot\n\n");
for (i = 0; i < NumMatrices*NumNumProcs; i += NumNumProcs) {
int j;
for (j = 0; j < NumNumProcs; ++j) {
char Tmp[256];
const struct sMatrixData *Mat = &Matrices[i + j];
if (j == 0) printf("\\verb|%s| & %ld", Mat->File, Mat->P);
else printf(" & %ld", Mat->P);
/*
int k;
for (k = 0; k < NUM_ATTRIBUTES; ++k) {
char Tmp[256];
DoubleToLaTeX(Tmp, Mat->Attributes[k].Average, 3);
printf(" & $%s \\pm ", Tmp);
DoubleToLaTeX(Tmp, sqrt(Mat->Attributes[k].Variance), 1);
printf("%s$", Tmp);
}
*/
DoubleToLaTeX(Tmp, Mat->Attributes[0].Average, 3);
printf(" & $%s \\pm ", Tmp);
DoubleToLaTeX(Tmp, sqrt(Mat->Attributes[0].Variance), 1);
printf("%s$", Tmp);
DoubleToLaTeX(Tmp, Mat->Attributes[1].Average, 3);
printf(" & $%s$", Tmp);
printf(" & $%ld \\pm %ld$", (long)Mat->Attributes[2].Average, (long)sqrt(Mat->Attributes[2].Variance));
printf(" & $%ld \\pm %ld$", (long)Mat->Attributes[3].Average, (long)sqrt(Mat->Attributes[3].Variance));
printf(" \\\\\n");
}
printf("\\hline\n");
}
printf("\n\\end{longtable}\n");
printf("\\end{landscape}\n\n");
printf("\\end{document}\n\n");
/* Append raw data. */
printf("Raw data:\n");
for (i = 0; i < NumMatrices*NumNumProcs; ++i) {
int j;
printf("%s\t%ld\t%ld\t%ld\t%ld\t", Matrices[i].File, Matrices[i].NumNz, Matrices[i].Rows, Matrices[i].Cols, Matrices[i].P);
for (j = 0; j < NUM_ATTRIBUTES - 1; ++j) printf("%e\t%e\t", Matrices[i].Attributes[j].Average, sqrt(Matrices[i].Attributes[j].Variance));
printf("%e\t%e\n", Matrices[i].Attributes[j].Average, sqrt(Matrices[i].Attributes[j].Variance));
}
printf("\n");
free(Matrices);
fprintf(stderr, "Done!\n");
exit(EXIT_SUCCESS);
}
