int main(int argc, char **argv)
{
#define test_A(i,j) test_A[(size_t)(j)*N+(i)]
#define test_A2(i,j) test_A2[(size_t)(j)*N+(i)]
int N,NB,w,LDA,BB;
size_t memsize; //bytes
int iam, nprocs, mydevice;
int ICTXT, nprow, npcol, myprow, mypcol;
int i_one = 1, i_zero = 0, i_negone = -1;
double d_one = 1.0, d_zero = 0.0, d_negone = -1.0;
int IASEED = 100;
/* printf("N=?\n");
scanf("%ld",&N);
printf("NB=?\n");
scanf("%d", &NB);
printf("width of Y panel=?\n");
scanf("%ld",&w);
*/
if(argc < 4){
printf("invalid arguments N NB memsize(M)\n");
exit(1);
}
N = atoi(argv[1]);
NB = atoi(argv[2]);
memsize = (size_t)atoi(argv[3])*1024*1024;
BB = (N + NB - 1) / NB;
w = memsize/sizeof(double)/BB/NB/NB - 1;
assert(w > 0);
LDA = N + 0; //padding
int do_io = (N <= NSIZE);
double llttime;
double gflops;
nprow = npcol = 1;
blacs_pinfo_(&iam, &nprocs);
blacs_get_(&i_negone, &i_zero, &ICTXT);
blacs_gridinit_(&ICTXT, "R", &nprow, &npcol);
blacs_gridinfo_(&ICTXT, &nprow, &npcol, &myprow, &mypcol);
#ifdef USE_MIC
#ifdef __INTEL_OFFLOAD
printf("offload compilation enabled\ninitialize each MIC\n");
offload_init(&iam, &mydevice);
#pragma offload target(mic:0)
{
mkl_peak_mem_usage(MKL_PEAK_MEM_ENABLE);
}
#else
if(isroot)
printf("offload compilation not enabled\n");
exit(0);
#endif
#else
#ifdef USE_CUBLASV2
{
cublasStatus_t cuStatus;
for(int r = 0; r < OOC_NTHREADS; r++){
cuStatus = cublasCreate(&worker_handle[r]);
assert(cuStatus == CUBLAS_STATUS_SUCCESS);
}
}
#else
cublasInit();
#endif
#endif
double *test_A = (double*)memalign(64,(size_t)LDA*N*sizeof(double)); // for chol
#ifdef VERIFY
double *test_A2 = (double*)memalign(64,(size_t)LDA*N*sizeof(double)); // for verify
#endif
/*Initialize A */
int i,j;
printf("Initialize A ... "); fflush(stdout);
llttime = MPI_Wtime();
pdmatgen(&ICTXT, "Symm", "Diag", &N,
&N, &NB, &NB,
test_A, &LDA, &i_zero, &i_zero,
&IASEED, &i_zero, &N, &i_zero, &N,
&myprow, &mypcol, &nprow, &npcol);
llttime = MPI_Wtime() - llttime;
printf("time %lf\n", llttime);
/*print test_A*/
if(do_io){
printf("Original A=\n\n");
matprint(test_A, N, LDA, 'A');
}
/*Use directed unblocked Cholesky factorization*/
/*
t1 = clock();
Test_dpotrf(test_A2,N);
t2 = clock();
printf ("time for unblocked Cholesky factorization on host %f \n",
((float) (t2 - t1)) / CLOCKS_PER_SEC);
*/
/*print test_A*/
/*
if(do_io){
printf("Unblocked result:\n\n");
matprint(test_A2,N,'L');
}
*/
/*Use tile algorithm*/
Quark *quark = QUARK_New(OOC_NTHREADS);
QUARK_DOT_DAG_Enable(quark, 0);
#ifdef USE_MIC
// mklmem(NB);
printf("QUARK MIC affinity binding\n");
QUARK_bind(quark);
printf("offload warm up\n");
warmup(quark);
#endif
QUARK_DOT_DAG_Enable(quark, quark_getenv_int("QUARK_DOT_DAG_ENABLE", 0));
printf("LLT start %lf\n", MPI_Wtime());
llttime = Cholesky(quark,test_A,N,NB,LDA,memsize);
printf("LLT end %lf\n", MPI_Wtime());
QUARK_Delete(quark);
#ifdef USE_MIC
offload_destroy();
#else
#ifdef USE_CUBLASV2
{
cublasStatus_t cuStatus;
for(int r = 0; r < OOC_NTHREADS; r++){
cuStatus = cublasDestroy(worker_handle[r]);
assert(cuStatus == CUBLAS_STATUS_SUCCESS);
}
}
#else
cublasShutdown();
#endif
#endif
gflops = (double) N;
gflops = gflops/3.0 + 0.5;
gflops = gflops*(double)(N)*(double)(N);
gflops = gflops/llttime/1024.0/1024.0/1024.0;
printf ("N NB memsize(MB) quark_pthreads time Gflops\n%d %d %lf %d %lf %lf\n",
N, NB, (double)memsize/1024/1024, OOC_NTHREADS, llttime, gflops);
#ifdef USE_MIC
#pragma offload target(mic:0)
{
memsize = mkl_peak_mem_usage(MKL_PEAK_MEM_RESET);
}
printf("mkl_peak_mem_usage %lf MB\n", (double)memsize/1024.0/1024.0);
#endif
/*Update and print L*/
if(do_io){
printf("L:\n\n");
matprint(test_A,N,LDA,'L');
}
#ifdef VERIFY
printf("Verify... ");
llttime = MPI_Wtime();
/*
* ------------------------
* check difference betwen
* test_A and test_A2
* ------------------------
*/
/*
{
double maxerr = 0;
double maxerr2 = 0;
for (j = 0; j < N; j++)
{
for (i = j; i < N; i++)
{
double err = (test_A (i, j) - test_A2 (i, j));
err = ABS (err);
maxerr = MAX (err, maxerr);
maxerr2 = maxerr2 + err * err;
};
};
maxerr2 = sqrt (ABS (maxerr2));
printf ("max difference between test_A and test_A2 %lf \n", maxerr);
printf ("L2 difference between test_A and test_A2 %lf \n", maxerr2);
};
*/
/*
* ------------------
* over-write test_A2
* ------------------
*/
pdmatgen(&ICTXT, "Symm", "Diag", &N,
&N, &NB, &NB,
test_A2, &LDA, &i_zero,
&i_zero, &IASEED, &i_zero, &N, &i_zero, &N,
&myprow, &mypcol, &nprow, &npcol);
/*
* ---------------------------------------
* after solve, test_A2 should be identity
* ---------------------------------------
*/
// test_A = chol(B) = L;
// test_A2 = B
// solve L*L'*X = B
// if L is correct, X is identity */
{
int uplo = 'L';
const char *uplo_char = ((uplo == (int) 'U')
|| (uplo == (int) 'u')) ? "U" : "L";
int info = 0;
int nrhs = N;
int LDA = N;
int ldb = N;
dpotrs(uplo_char, &N, &nrhs, test_A, &LDA, test_A2, &ldb, &info);
assert (info == 0);
}
{
double maxerr = 0;
double maxerr2 = 0;
for (j = 0; j < N; j++)
{
for (i = 0; i < N; i++)
{
double eyeij = (i == j) ? 1.0 : 0.0;
double err = (test_A2 (i, j) - eyeij);
err = ABS (err);
maxerr = MAX (maxerr, err);
maxerr2 = maxerr2 + err * err;
};
};
maxerr2 = sqrt (ABS (maxerr2));
printf("time %lf\n", MPI_Wtime() - llttime);
printf ("max error %lf \n", maxerr);
printf ("max L2 error %lf \n", maxerr2);
}
#endif
free(test_A);test_A=NULL;
#ifdef VERIFY
free(test_A2);test_A2=NULL;
#endif
blacs_gridexit_(&ICTXT);
blacs_exit_(&i_zero);
return 0;
#undef test_A
#undef test_A2
}
/*==== MAIN FUNCTION =================================================*/
int main( int argc, char *argv[] ){
/* ==== Declarations =================================================== */
/* File variables */
FILE *fin;
/* Matrix descriptors */
MDESC descA, descB, descC, descA_local, descB_local;
/* Local scalars */
MKL_INT iam, nprocs, ictxt, myrow, mycol, nprow, npcol;
MKL_INT n, nb, mp, nq, lld, lld_local;
MKL_INT i, j, info;
int n_int, nb_int, nprow_int, npcol_int;
double thresh, diffnorm, anorm, bnorm, residual, eps;
/* Local arrays */
double *A_local, *B_local, *A, *B, *C, *work;
MKL_INT iwork[ 4 ];
/* ==== Executable statements ========================================== */
/* Get information about how many processes are used for program execution
and number of current process */
blacs_pinfo_( &iam, &nprocs );
/* Init temporary 1D process grid */
blacs_get_( &i_negone, &i_zero, &ictxt );
blacs_gridinit_( &ictxt, "C", &nprocs, &i_one );
/* Open input file */
if ( iam == 0 ) {
fin = fopen( "../in/pblas3ex.in", "r" );
if ( fin == NULL ) {
printf( "Error while open input file." );
return 2;
}
}
/* Read data and send it to all processes */
if ( iam == 0 ) {
/* Read parameters */
fscanf( fin, "%d n, dimension of vectors, must be > 0 ", &n_int );
fscanf( fin, "%d nb, size of blocks, must be > 0 ", &nb_int );
fscanf( fin, "%d p, number of rows in the process grid, must be > 0", &nprow_int );
fscanf( fin, "%d q, number of columns in the process grid, must be > 0, p*q = number of processes", &npcol_int );
fscanf( fin, "%lf threshold for residual check (to switch off check set it < 0.0) ", &thresh );
n = (MKL_INT) n_int;
nb = (MKL_INT) nb_int;
nprow = (MKL_INT) nprow_int;
npcol = (MKL_INT) npcol_int;
/* Check if all parameters are correct */
if( ( n<=0 )||( nb<=0 )||( nprow<=0 )||( npcol<=0 )||( nprow*npcol != nprocs ) ) {
printf( "One or several input parameters has incorrect value. Limitations:\n" );
printf( "n > 0, nb > 0, p > 0, q > 0 - integer\n" );
printf( "p*q = number of processes\n" );
printf( "threshold - double (set to negative to swicth off check)\n");
return 2;
}
/* Pack data into array and send it to other processes */
iwork[ 0 ] = n;
iwork[ 1 ] = nb;
iwork[ 2 ] = nprow;
iwork[ 3 ] = npcol;
igebs2d_( &ictxt, "All", " ", &i_four, &i_one, iwork, &i_four );
dgebs2d_( &ictxt, "All", " ", &i_one, &i_one, &thresh, &i_one );
} else {
/* Recieve and unpack data */
igebr2d_( &ictxt, "All", " ", &i_four, &i_one, iwork, &i_four, &i_zero, &i_zero );
dgebr2d_( &ictxt, "All", " ", &i_one, &i_one, &thresh, &i_one, &i_zero, &i_zero );
n = iwork[ 0 ];
nb = iwork[ 1 ];
nprow = iwork[ 2 ];
npcol = iwork[ 3 ];
}
if ( iam == 0 ) { fclose( fin ); }
/* Destroy temporary process grid */
blacs_gridexit_( &ictxt );
/* Init workind 2D process grid */
blacs_get_( &i_negone, &i_zero, &ictxt );
blacs_gridinit_( &ictxt, "R", &nprow, &npcol );
blacs_gridinfo_( &ictxt, &nprow, &npcol, &myrow, &mycol );
/* Create on process 0 two matrices: A - orthonormal, B -random */
if ( ( myrow == 0 ) && ( mycol == 0 ) ){
/* Allocate arrays */
A_local = (double*) calloc( n*n, sizeof( double ) );
B_local = (double*) calloc( n*n, sizeof( double ) );
/* Set arrays */
for ( i=0; i<n; i++ ){
for ( j=0; j<n; j++ ){
B_local[ i+n*j ] = one*rand()/RAND_MAX;
}
B_local[ i+n*i ] += two;
}
for ( j=0; j<n; j++ ){
for ( i=0; i<n; i++ ){
if ( j < n-1 ){
if ( i <= j ){
A_local[ i+n*j ] = one / sqrt( ( double )( (j+1)*(j+2) ) );
} else if ( i == j+1 ) {
A_local[ i+n*j ] = -one / sqrt( one + one/( double )(j+1) );
} else {
A_local[ i+n*j ] = zero;
}
} else {
A_local[ i+n*(n-1) ] = one / sqrt( ( double )n );
}
}
}
/* Print information of task */
printf( "=== START OF EXAMPLE ===================\n" );
printf( "Matrix-matrix multiplication: A*B = C\n\n" );
printf( "/ 1/q_1 ........ 1/q_n-1 1/q_n \\ \n" );
printf( "| . | \n" );
printf( "| `. : : | \n" );
printf( "| -1/q_1 `. : : | \n" );
printf( "| . `. : : | = A \n" );
printf( "| 0 `. ` | \n" );
printf( "| : `. `. 1/q_n-1 1/q_n | \n" );
printf( "| : `. `. | \n" );
printf( "\\ 0 .... 0 -(n-1)/q_n-1 1/q_n / \n\n" );
printf( "q_i = sqrt( i^2 + i ), i=1..n-1, q_n = sqrt( n )\n\n" );
printf( "A - n*n real matrix (orthonormal) \n" );
printf( "B - random n*n real matrix\n\n" );
printf( "n = %d, nb = %d; %dx%d - process grid\n\n", n, nb, nprow, npcol );
printf( "=== PROGRESS ===========================\n" );
} else {
/* Other processes don't contain parts of initial arrays */
A_local = NULL;
B_local = NULL;
}
/* Compute precise length of local pieces and allocate array on
each process for parts of distributed vectors */
mp = numroc_( &n, &nb, &myrow, &i_zero, &nprow );
nq = numroc_( &n, &nb, &mycol, &i_zero, &npcol );
A = (double*) calloc( mp*nq, sizeof( double ) );
B = (double*) calloc( mp*nq, sizeof( double ) );
C = (double*) calloc( mp*nq, sizeof( double ) );
/* Compute leading dimensions */
lld_local = MAX( numroc_( &n, &n, &myrow, &i_zero, &nprow ), 1 );
lld = MAX( mp, 1 );
/* Initialize descriptors for initial arrays located on 0 process */
descinit_( descA_local, &n, &n, &n, &n, &i_zero, &i_zero, &ictxt, &lld_local, &info );
descinit_( descB_local, &n, &n, &n, &n, &i_zero, &i_zero, &ictxt, &lld_local, &info );
/* Initialize descriptors for distributed arrays */
descinit_( descA, &n, &n, &nb, &nb, &i_zero, &i_zero, &ictxt, &lld, &info );
descinit_( descB, &n, &n, &nb, &nb, &i_zero, &i_zero, &ictxt, &lld, &info );
descinit_( descC, &n, &n, &nb, &nb, &i_zero, &i_zero, &ictxt, &lld, &info );
/* Distribute matrices from 0 process over process grid */
pdgeadd_( &trans, &n, &n, &one, A_local, &i_one, &i_one, descA_local, &zero, A, &i_one, &i_one, descA );
pdgeadd_( &trans, &n, &n, &one, B_local, &i_one, &i_one, descB_local, &zero, B, &i_one, &i_one, descB );
if( iam == 0 ){ printf( ".. Arrays are distributed ( p?geadd ) ..\n" ); }
/* Destroy arrays on 0 process - they are not necessary anymore */
if( ( myrow == 0 ) && ( mycol == 0 ) ){
free( A_local );
free( B_local );
}
/* Compute norm of A and B */
work = (double*) calloc( mp, sizeof( double ) );
anorm = pdlange_( "I", &n, &n, A, &i_one, &i_one, descA, work );
bnorm = pdlange_( "I", &n, &n, B, &i_one, &i_one, descB, work );
if( iam == 0 ){ printf( ".. Norms of A and B are computed ( p?lange ) ..\n" ); }
/* Compute product C = A*B */
pdgemm_( "N", "N", &n, &n, &n, &one, A, &i_one, &i_one, descA, B, &i_one, &i_one, descB,
&zero, C, &i_one, &i_one, descC );
if( iam == 0 ){ printf( ".. Multiplication A*B=C is done ( p?gemm ) ..\n" ); }
/* Compute difference B - inv_A*C (inv_A = transpose(A) because A is orthonormal) */
pdgemm_( "T", "N", &n, &n, &n, &one, A, &i_one, &i_one, descA, C, &i_one, &i_one, descC,
&negone, B, &i_one, &i_one, descB );
if( iam == 0 ){ printf( ".. Difference is computed ( p?gemm ) ..\n" ); }
/* Compute norm of B - inv_A*C (which is contained in B) */
diffnorm = pdlange_( "I", &n, &n, B, &i_one, &i_one, descB, work );
free( work );
if( iam == 0 ){ printf( ".. Norms of the difference B-inv_A*C is computed ( p?lange ) ..\n" ); }
/* Print results */
if( iam == 0 ){
printf( ".. Solutions are compared ..\n" );
printf( "== Results ==\n" );
printf( "||A|| = %03.11f\n", anorm );
printf( "||B|| = %03.11f\n", bnorm );
printf( "=== END OF EXAMPLE =====================\n" );
}
/* Compute machine epsilon */
eps = pdlamch_( &ictxt, "e" );
/* Compute residual */
residual = diffnorm /( two*anorm*bnorm*eps );
/* Destroy arrays */
free( A );
free( B );
free( C );
/* Destroy process grid */
blacs_gridexit_( &ictxt );
blacs_exit_( &i_zero );
/* Check if residual passed or failed the threshold */
if ( ( iam == 0 ) && ( thresh >= zero ) && !( residual <= thresh ) ){
printf( "FAILED. Residual = %05.16f\n", residual );
return 1;
} else {
return 0;
}
/*========================================================================
====== End of PBLAS Level 3 example ====================================
======================================================================*/
}
int main()
{
const MKL_INT m = 1000;
const MKL_INT k = 100000;
const MKL_INT n = 1000;
const MKL_INT nb = 100;
const MKL_INT nprow = 2;
const MKL_INT npcol = 2;
MKL_INT iam, nprocs, ictxt, myrow, mycol;
MDESC descA, descB, descC, descA_local, descB_local, descC_local;
MKL_INT info;
MKL_INT a_m_local, a_n_local, b_m_local, b_n_local, c_m_local, c_n_local;
MKL_INT a_lld, b_lld, c_lld;
blacs_pinfo_( &iam, &nprocs );
blacs_get_( &i_negone, &i_zero, &ictxt );
blacs_gridinit_( &ictxt, "R", &nprow, &npcol );
blacs_gridinfo_( &ictxt, &nprow, &npcol, &myrow, &mycol );
double *a = 0;
double *b = 0;
double *c = 0;
if (iam==0)
{
a = gen_a(m, k);
b = gen_b(k, n);
c = (double*)calloc(m*n, sizeof(double));
puts("a=");
print(a, m, k);
puts("b=");
print(b, k, n);
}
a_m_local = numroc_( &m, &nb, &myrow, &i_zero, &nprow );
a_n_local = numroc_( &k, &nb, &mycol, &i_zero, &npcol );
b_m_local = numroc_( &k, &nb, &myrow, &i_zero, &nprow );
b_n_local = numroc_( &n, &nb, &mycol, &i_zero, &npcol );
c_m_local = numroc_( &m, &nb, &myrow, &i_zero, &nprow );
c_n_local = numroc_( &n, &nb, &mycol, &i_zero, &npcol );
double *A = (double*) calloc( a_m_local * a_n_local, sizeof( double ) );
double *B = (double*) calloc( b_m_local * b_n_local, sizeof( double ) );
double *C = (double*) calloc( c_m_local * c_n_local, sizeof( double ) );
a_lld = MAX( a_m_local, 1 );
b_lld = MAX( b_m_local, 1 );
c_lld = MAX( c_m_local, 1 );
if (iam==0)
{
printf("a_m_local = %d\ta_n_local = %d\tb_m_local = %d\tb_n_local = %d\tc_m_local = %d\tc_n_local = %d\n", a_m_local, a_n_local, b_m_local, b_n_local,
c_m_local, c_n_local);
printf("a_lld = %d\tb_lld = %d\tc_lld = %d\n", a_lld, b_lld, c_lld);
}
descinit_( descA_local, &m, &k, &m, &k, &i_zero, &i_zero, &ictxt, &m, &info );
descinit_( descB_local, &k, &n, &k, &n, &i_zero, &i_zero, &ictxt, &k, &info );
descinit_( descC_local, &m, &n, &m, &n, &i_zero, &i_zero, &ictxt, &m, &info );
descinit_( descA, &m, &k, &nb, &nb, &i_zero, &i_zero, &ictxt, &a_lld, &info );
descinit_( descB, &k, &n, &nb, &nb, &i_zero, &i_zero, &ictxt, &b_lld, &info );
descinit_( descC, &m, &n, &nb, &nb, &i_zero, &i_zero, &ictxt, &c_lld, &info );
printf("Rank %d: start distribute data\n", iam);
pdgeadd_( &trans, &m, &k, &one, a, &i_one, &i_one, descA_local, &zero, A, &i_one, &i_one, descA );
pdgeadd_( &trans, &k, &n, &one, b, &i_one, &i_one, descB_local, &zero, B, &i_one, &i_one, descB );
printf("Rank %d: finished distribute data\n", iam);
if (iam==0)
{
puts("a");
print(A, a_m_local, a_n_local);
puts("b");
print(B, b_m_local, b_n_local);
}
pdgemm_( "N", "N", &m, &n, &k, &one, A, &i_one, &i_one, descA, B, &i_one, &i_one, descB,
&zero, C, &i_one, &i_one, descC );
printf("Rank %d: finished dgemm\n", iam);
if (iam == 0)
{
puts("c");
print(C, c_m_local, c_n_local);
}
pdgeadd_( &trans, &m, &n, &one, C, &i_one, &i_one, descC, &zero, c, &i_one, &i_one, descC_local);
if (iam==0)
{
puts("global c");
print(c, m, n);
}
free(A);
free(B);
free(C);
if (iam==0)
{
free(a);
free(b);
free(c);
}
blacs_gridexit_( &ictxt );
blacs_exit_( &i_zero );
}
int MAIN__(int argc, char** argv) {
int num; // number of data
int dim; // dimension of each data
int nprow=4; // number of row
int npcol=1; // number of columnn
int zero=0, one=1; // constant value
int ictxt,myrow,mycol,pnum,pdim,info;
char ifilename[LEN_FILENAME];
char ofilename[LEN_FILENAME];
int myproc, nprocs;
Cblacs_pinfo(&myproc, &nprocs);
Cblacs_setup(&myproc, &nprocs);
Cblacs_get(-1,0,&ictxt);
nprow = nprocs;
npcol = 1; // fixed
char order[] = "Row";
Cblacs_gridinit(&ictxt, order, nprow, npcol);
Cblacs_gridinfo(ictxt, &nprow, &npcol, &myrow, &mycol);
if (DEBUG_MODE) {
printf("ConTxt = %d\n", ictxt);
printf("nprocs=%d, nprow=%d, npcol=%d\n", nprocs, nprow, npcol);
printf("nprocs=%d, myrow=%d, mycol=%d\n", nprocs, myrow, mycol);
}
get_option(argc, argv, ifilename, ofilename, &num, &dim);
// 0. cosinedist(ij) = 1 - V(i)V(j)/(Length(V(i))*Length(V(j)))
// 1. calculate submatrix size
int bsize = num / nprow; // blocking factor
pnum = num / nprow;
pdim = dim;
if ( myrow < (num/bsize)%nprow) {
pnum += bsize;
}
else if ( myrow == (num/bsize)%nprow) {
pnum += (num % bsize);
}
else {
}
if(DEBUG_MODE)
printf("myproc=%d: pnum=%d, pdim=%d, bsize=%d\n", myproc, pnum, pdim, bsize);
int desc_input[9], desc_v[9], desc_ip[9], desc_n[9], desc_result[9];
descinit_(desc_input, &num, &dim, &num, &dim, &zero, &zero, &ictxt, &num, &info);
descinit_(desc_v, &num, &dim, &bsize, &pdim, &zero, &zero, &ictxt, &pnum, &info);
descinit_(desc_ip, &num, &num, &bsize, &num, &zero, &zero, &ictxt, &pnum, &info);
descinit_(desc_n, &num, &one, &bsize, &one, &zero, &zero, &ictxt, &pnum, &info);
descinit_(desc_result, &num, &num, &num, &num, &zero, &zero, &ictxt, &num, &info);
// 2. read input data
double* input;
if (myproc == 0) {
input = (double*)malloc(sizeof(double)*num*dim);
memset(input, 0, sizeof(double)*num*dim);
read_data(ifilename, num, dim, input);
printArray("input", myproc, input, num, dim);
}
// 3. distribute input data array
double* V = (double*)malloc(sizeof(double)*pnum*pdim);
memset(V, 0, sizeof(double)*pnum*pdim);
Cpdgemr2d(num, dim, input, 1, 1, desc_input, V, 1, 1, desc_v, ictxt);
printArray("V", myproc, V, pnum, pdim);
// 4. InnerProduct = VV'
double* InnerProduct = (double*)malloc(sizeof(double)*pnum*num);
memset(InnerProduct, 0, sizeof(double)*pnum*num);
char transa = 'N', transb = 'T';
int m = num, n = num, k = dim;
int lda = num, ldb = num, ldc = num;
double alpha = 1.0f, beta = 0.0f;
pdgemm_(&transa, &transb, &m, &n, &k, &alpha, V, &one, &one, desc_v, V, &one, &one, desc_v, &beta, InnerProduct, &one, &one, desc_ip);
printArray("InnerProduct", myproc, InnerProduct, pnum, num);
// 5. Norm of each vector
double* Norm = (double*)malloc(sizeof(double)*pnum);
for (int i = 0; i < pnum; i++) {
int n = ((myproc*bsize)+(i/bsize)*(nprocs-1)*bsize+i)*pnum + i;
Norm[i] = sqrt(InnerProduct[n]);
}
printArray("Norm", myproc, Norm, 1, pnum);
// 6. Norm product matrix
double* NormProduct = (double*)malloc(sizeof(double)*pnum*num);
memset(NormProduct, 0, sizeof(double)*pnum*num);
char uplo = 'U';
n = num;
alpha = 1.0f;
int incx = 1;
lda = num;
pdsyr_(&uplo, &n, &alpha, Norm, &one, &one, desc_n, &incx, NormProduct, &one, &one, desc_ip);
printArray("NormProduct", myproc, NormProduct, pnum, num);
// 7. CosineDistance(ij) = 1-InnerProduct(ij)/NormProduct(ij)
double* CosineDistance = (double*)malloc(sizeof(double)*pnum*num);
memset(CosineDistance, 0, sizeof(double)*pnum*num);
for (int j = 0; j < num; j++) {
for (int i = 0; i < pnum; i++) {
int n = ((myproc*bsize)+i+(i/bsize)*(nprocs-1)*bsize)*pnum+i;
int p = i+j*pnum;
if (p<=n) {
CosineDistance[p] = 0.0;
}
else {
CosineDistance[p] = 1 - InnerProduct[p]/NormProduct[p];
}
}
}
printArray("CosineDistance", myproc, CosineDistance, pnum, num);
// 8. gather result
double* result;
if ( myproc == 0 ) {
result = (double*)malloc(sizeof(double)*num*num);
memset(result, 0, sizeof(double)*num*num);
}
Cpdgemr2d(num, num, CosineDistance, 1, 1, desc_ip, result, 1, 1, desc_result, ictxt);
// 9. output to file
if ( myproc == 0 ) {
output_results(ofilename, result, num, num);
}
// a. cleanup memory
free(V);
free(InnerProduct);
free(Norm);
free(NormProduct);
free(CosineDistance);
if ( myproc == 0 ) {
free(input);
free(result);
}
blacs_exit_(&zero);
return 0;
}
