extern "C" magma_int_t
magma_zrowentries(
magma_z_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_index_t *length=NULL;
magma_index_t i,j, maxrowlength=0;
// check whether matrix on CPU
if ( A->memory_location == Magma_CPU ) {
// CSR
if ( A->storage_type == Magma_CSR ) {
CHECK( magma_index_malloc_cpu( &length, A->num_rows));
for( i=0; i<A->num_rows; i++ ) {
length[i] = A->row[i+1]-A->row[i];
if (length[i] > maxrowlength)
maxrowlength = length[i];
}
A->max_nnz_row = maxrowlength;
}
// Dense
else if ( A->storage_type == Magma_DENSE ) {
CHECK( magma_index_malloc_cpu( &length, A->num_rows));
for( i=0; i<A->num_rows; i++ ) {
length[i] = 0;
for( j=0; j<A->num_cols; j++ ) {
if ( MAGMA_Z_REAL( A->val[i*A->num_cols + j] ) != 0. )
length[i]++;
}
if (length[i] > maxrowlength)
maxrowlength = length[i];
}
A->max_nnz_row = maxrowlength;
}
} // end CPU case
else {
printf("error: matrix not on CPU.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
magma_free( length );
return info;
}
magma_int_t
magma_zmLdiagadd(
magma_z_matrix *L,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_z_matrix LL={Magma_CSR};
if( L->row[1]==1 ){ // lower triangular with unit diagonal
//printf("L lower triangular.\n");
LL.diagorder_type = Magma_UNITY;
CHECK( magma_zmconvert( *L, &LL, Magma_CSR, Magma_CSRL, queue ));
}
else if ( L->row[1]==0 ){ // strictly lower triangular
//printf("L strictly lower triangular.\n");
CHECK( magma_zmtransfer( *L, &LL, Magma_CPU, Magma_CPU, queue ));
magma_free_cpu( LL.col );
magma_free_cpu( LL.val );
LL.nnz = L->nnz+L->num_rows;
CHECK( magma_zmalloc_cpu( &LL.val, LL.nnz ));
CHECK( magma_index_malloc_cpu( &LL.col, LL.nnz ));
magma_int_t z=0;
for( magma_int_t i=0; i<L->num_rows; i++){
LL.row[i] = z;
for( magma_int_t j=L->row[i]; j<L->row[i+1]; j++){
LL.val[z] = L->val[j];
LL.col[z] = L->col[j];
z++;
}
// add unit diagonal
LL.val[z] = MAGMA_Z_MAKE(1.0, 0.0);
LL.col[z] = i;
z++;
}
LL.row[LL.num_rows] = z;
LL.nnz = z;
}
else {
printf("error: L neither lower nor strictly lower triangular!\n");
}
magma_zmfree( L, queue );
CHECK( magma_zmtransfer(LL, L, Magma_CPU, Magma_CPU, queue ));
cleanup:
if( info != 0 ){
magma_zmfree( L, queue );
}
magma_zmfree( &LL, queue );
return info;
}
magma_int_t
magma_ddomainoverlap(
magma_index_t num_rows,
magma_int_t *num_indices,
magma_index_t *rowptr,
magma_index_t *colidx,
magma_index_t *x,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t blocksize=128;
magma_int_t row=0, col=0, num_ind=0, offset=0;
magma_index_t *tmp_x;
CHECK( magma_index_malloc_cpu( &tmp_x, blocksize ));
for(magma_int_t i=0; i<blocksize; i++ ){
tmp_x[i] = -1;
}
for(magma_int_t i=0; i<num_rows; i++){
row = i;
for(magma_int_t j=rowptr[row]; j<rowptr[row+1]; j++){
col = colidx[j];
int doubleitem = 0;
for(magma_int_t k=0; k<blocksize; k++){
if( tmp_x[k] == col )
doubleitem = 1;
}
if( doubleitem == 0 ){
tmp_x[num_ind] = col;
num_ind++;
(*num_indices)++;
}
if( num_ind == blocksize || j == rowptr[num_rows]-1 ){
magma_dindexcopy( num_ind, offset, tmp_x, x, queue );
offset=offset+num_ind;
num_ind = 0;
break;
}
}
}
cleanup:
magma_free_cpu( tmp_x );
return info;
}
magma_int_t
magma_zinitguess(
magma_z_matrix A,
magma_z_matrix *L,
magma_z_matrix *U,
magma_queue_t queue )
{
magma_int_t info = 0;
magmaDoubleComplex one = MAGMA_Z_MAKE( 1.0, 0.0 );
magma_z_matrix hAL={Magma_CSR}, hAU={Magma_CSR}, dAL={Magma_CSR},
dAU={Magma_CSR}, dALU={Magma_CSR}, hALU={Magma_CSR}, hD={Magma_CSR},
dD={Magma_CSR}, dL={Magma_CSR}, hL={Magma_CSR};
magma_int_t i,j;
magma_int_t offdiags = 0;
magma_index_t *diag_offset;
magmaDoubleComplex *diag_vals=NULL;
// need only lower triangular
hAL.diagorder_type = Magma_VALUE;
CHECK( magma_zmconvert( A, &hAL, Magma_CSR, Magma_CSRL, queue ));
//magma_zmconvert( hAL, &hALCOO, Magma_CSR, Magma_CSRCOO );
// need only upper triangular
//magma_zmconvert( A, &hAU, Magma_CSR, Magma_CSRU );
CHECK( magma_z_cucsrtranspose( hAL, &hAU, queue ));
//magma_zmconvert( hAU, &hAUCOO, Magma_CSR, Magma_CSRCOO );
CHECK( magma_zmtransfer( hAL, &dAL, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_z_spmm( one, dAL, dAU, &dALU, queue ));
CHECK( magma_zmtransfer( dALU, &hALU, Magma_DEV, Magma_CPU, queue ));
magma_zmfree( &dAU, queue);
magma_zmfree( &dALU, queue);
CHECK( magma_zmalloc_cpu( &diag_vals, offdiags+1 ));
CHECK( magma_index_malloc_cpu( &diag_offset, offdiags+1 ));
diag_offset[0] = 0;
diag_vals[0] = MAGMA_Z_MAKE( 1.0, 0.0 );
CHECK( magma_zmgenerator( hALU.num_rows, offdiags, diag_offset, diag_vals, &hD, queue ));
magma_zmfree( &hALU, queue );
for(i=0; i<hALU.num_rows; i++){
for(j=hALU.row[i]; j<hALU.row[i+1]; j++){
if( hALU.col[j] == i ){
//printf("%d %d %d == %d -> %f -->", i, j, hALU.col[j], i, hALU.val[j]);
hD.val[i] = MAGMA_Z_MAKE(
1.0 / sqrt(fabs(MAGMA_Z_REAL(hALU.val[j]))) , 0.0 );
//printf("insert %f at %d\n", hD.val[i], i);
}
}
}
CHECK( magma_zmtransfer( hD, &dD, Magma_CPU, Magma_DEV, queue ));
magma_zmfree( &hD, queue);
CHECK( magma_z_spmm( one, dD, dAL, &dL, queue ));
magma_zmfree( &dAL, queue );
magma_zmfree( &dD, queue );
/*
// check for diagonal = 1
magma_z_matrix dLt={Magma_CSR}, dLL={Magma_CSR}, LL={Magma_CSR};
CHECK( magma_z_cucsrtranspose( dL, &dLt ));
CHECK( magma_zcuspmm( dL, dLt, &dLL ));
CHECK( magma_zmtransfer( dLL, &LL, Magma_DEV, Magma_CPU ));
//for(i=0; i < hALU.num_rows; i++) {
for(i=0; i < 100; i++) {
for(j=hALU.row[i]; j < hALU.row[i+1]; j++) {
if( hALU.col[j] == i ){
printf("%d %d -> %f -->", i, i, LL.val[j]);
}
}
}
*/
CHECK( magma_zmtransfer( dL, &hL, Magma_DEV, Magma_CPU, queue ));
CHECK( magma_zmconvert( hL, L, Magma_CSR, Magma_CSRCOO, queue ));
cleanup:
if( info !=0 ){
magma_zmfree( L, queue );
magma_zmfree( U, queue );
}
magma_zmfree( &dAU, queue);
magma_zmfree( &dALU, queue);
magma_zmfree( &dL, queue );
magma_zmfree( &hL, queue );
magma_zmfree( &dAL, queue );
magma_zmfree( &dD, queue );
magma_zmfree( &hD, queue);
magma_zmfree( &hALU, queue );
return info;
}
magma_int_t
magma_zilures(
magma_z_matrix A,
magma_z_matrix L,
magma_z_matrix U,
magma_z_matrix *LU,
real_Double_t *res,
real_Double_t *nonlinres,
magma_queue_t queue )
{
magma_int_t info = 0;
magmaDoubleComplex tmp;
real_Double_t tmp2;
magma_int_t i, j, k;
magmaDoubleComplex one = MAGMA_Z_MAKE( 1.0, 0.0 );
magma_z_matrix LL={Magma_CSR}, L_d={Magma_CSR}, U_d={Magma_CSR}, LU_d={Magma_CSR};
if( L.row[1]==1 ){ // lower triangular with unit diagonal
//printf("L lower triangular.\n");
LL.diagorder_type = Magma_UNITY;
CHECK( magma_zmconvert( L, &LL, Magma_CSR, Magma_CSRL, queue ));
}
else if ( L.row[1]==0 ){ // strictly lower triangular
//printf("L strictly lower triangular.\n");
CHECK( magma_zmtransfer( L, &LL, Magma_CPU, Magma_CPU, queue ));
magma_free_cpu( LL.col );
magma_free_cpu( LL.val );
LL.nnz = L.nnz+L.num_rows;
CHECK( magma_zmalloc_cpu( &LL.val, LL.nnz ));
CHECK( magma_index_malloc_cpu( &LL.col, LL.nnz ));
magma_int_t z=0;
for (i=0; i < L.num_rows; i++) {
LL.row[i] = z;
for (j=L.row[i]; j < L.row[i+1]; j++) {
LL.val[z] = L.val[j];
LL.col[z] = L.col[j];
z++;
}
// add unit diagonal
LL.val[z] = MAGMA_Z_MAKE(1.0, 0.0);
LL.col[z] = i;
z++;
}
LL.row[LL.num_rows] = z;
}
else {
printf("error: L neither lower nor strictly lower triangular!\n");
}
CHECK( magma_zmtransfer( LL, &L_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_zmtransfer( U, &U_d, Magma_CPU, Magma_DEV, queue ));
magma_zmfree( &LL, queue );
CHECK( magma_z_spmm( one, L_d, U_d, &LU_d, queue ));
CHECK( magma_zmtransfer(LU_d, LU, Magma_DEV, Magma_CPU, queue ));
magma_zmfree( &L_d, queue );
magma_zmfree( &U_d, queue );
magma_zmfree( &LU_d, queue );
// compute Frobenius norm of A-LU
for(i=0; i<A.num_rows; i++){
for(j=A.row[i]; j<A.row[i+1]; j++){
magma_index_t lcol = A.col[j];
for(k=LU->row[i]; k<LU->row[i+1]; k++){
if( LU->col[k] == lcol ){
tmp = MAGMA_Z_MAKE(
MAGMA_Z_REAL( LU->val[k] )- MAGMA_Z_REAL( A.val[j] )
, 0.0 );
LU->val[k] = tmp;
tmp2 = (real_Double_t) fabs( MAGMA_Z_REAL(tmp) );
(*nonlinres) = (*nonlinres) + tmp2*tmp2;
}
}
}
}
for(i=0; i<LU->num_rows; i++){
for(j=LU->row[i]; j<LU->row[i+1]; j++){
tmp2 = (real_Double_t) fabs( MAGMA_Z_REAL(LU->val[j]) );
(*res) = (*res) + tmp2* tmp2;
}
}
(*res) = sqrt((*res));
(*nonlinres) = sqrt((*nonlinres));
cleanup:
if( info !=0 ){
magma_zmfree( LU, queue );
}
magma_zmfree( &LL, queue );
magma_zmfree( &L_d, queue );
magma_zmfree( &U_d, queue );
magma_zmfree( &LU_d, queue );
return info;
}
magma_int_t
magma_cmgenerator(
magma_int_t n,
magma_int_t offdiags,
magma_index_t *diag_offset,
magmaFloatComplex *diag_vals,
magma_c_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_c_matrix B={Magma_CSR};
B.val = NULL;
B.col = NULL;
B.row = NULL;
B.rowidx = NULL;
B.blockinfo = NULL;
B.diag = NULL;
B.dval = NULL;
B.dcol = NULL;
B.drow = NULL;
B.drowidx = NULL;
B.ddiag = NULL;
B.list = NULL;
B.dlist = NULL;
B.num_rows = n;
B.num_cols = n;
B.fill_mode = MagmaFull;
B.memory_location = Magma_CPU;
B.storage_type = Magma_ELLPACKT;
B.max_nnz_row = (2*offdiags+1);
CHECK( magma_cmalloc_cpu( &B.val, B.max_nnz_row*n ));
CHECK( magma_index_malloc_cpu( &B.col, B.max_nnz_row*n ));
for( int i=0; i<n; i++ ) { // stride over rows
// stride over the number of nonzeros in each row
// left of diagonal
for( int j=0; j<offdiags; j++ ) {
B.val[ i*B.max_nnz_row + j ] = diag_vals[ offdiags - j ];
B.col[ i*B.max_nnz_row + j ] = -1 * diag_offset[ offdiags-j ] + i;
}
// elements on the diagonal
B.val[ i*B.max_nnz_row + offdiags ] = diag_vals[ 0 ];
B.col[ i*B.max_nnz_row + offdiags ] = i;
// right of diagonal
for( int j=0; j<offdiags; j++ ) {
B.val[ i*B.max_nnz_row + j + offdiags +1 ] = diag_vals[ j+1 ];
B.col[ i*B.max_nnz_row + j + offdiags +1 ] = diag_offset[ j+1 ] + i;
}
}
// set invalid entries to zero
for( int i=0; i<n; i++ ) { // stride over rows
for( int j=0; j<B.max_nnz_row; j++ ) { // nonzeros in every row
if ( (B.col[i*B.max_nnz_row + j] < 0) ||
(B.col[i*B.max_nnz_row + j] >= n) ) {
B.val[ i*B.max_nnz_row + j ] = MAGMA_C_MAKE( 0.0, 0.0 );
}
}
}
B.nnz = 0;
for( int i=0; i<n; i++ ) { // stride over rows
for( int j=0; j<B.max_nnz_row; j++ ) { // nonzeros in every row
if ( MAGMA_C_REAL( B.val[i*B.max_nnz_row + j]) != 0.0 )
B.nnz++;
}
}
B.true_nnz = B.nnz;
// converting it to CSR will remove the invalit entries completely
CHECK( magma_cmconvert( B, A, Magma_ELLPACKT, Magma_CSR, queue ));
cleanup:
if( info != 0 ){
magma_cmfree( &B, queue );
}
return info;
}
/**
Purpose
-------
Computes the diameter of a sparse matrix and stores the value in diameter.
Arguments
---------
@param[in,out]
A magma_z_matrix*
sparse matrix
@param[in]
queue magma_queue_t
Queue to execute in.
@ingroup magmasparse_zaux
********************************************************************/
extern "C" magma_int_t
magma_zdiameter(
magma_z_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_index_t i, j, tmp, *dim=NULL, maxdim=0;
// check whether matrix on CPU
if ( A->memory_location == Magma_CPU ) {
// CSR
if ( A->storage_type == Magma_CSR ) {
CHECK( magma_index_malloc_cpu( &dim, A->num_rows));
for( i=0; i<A->num_rows; i++ ) {
dim[i] = 0;
for( j=A->row[i]; j<A->row[i+1]; j++ ) {
// if ( MAGMA_Z_REAL(A->val[j]) > THRESHOLD ) {
tmp = abs( i - A->col[j] );
if ( tmp > dim[i] )
dim[i] = tmp;
// }
}
if ( dim[i] > maxdim )
maxdim = dim[i];
}
A->diameter = maxdim;
}
// Dense
else if ( A->storage_type == Magma_DENSE ) {
magma_index_t i, j, tmp, *dim, maxdim=0;
CHECK( magma_index_malloc_cpu( &dim, A->num_rows));
for( i=0; i<A->num_rows; i++ ) {
dim[i] = 0;
for( j=0; j<A->num_cols; j++ ) {
if ( MAGMA_Z_REAL( A->val[i*A->num_cols + j] ) != 0.0 ) {
tmp = abs( i -j );
if ( tmp > dim[i] )
dim[i] = tmp;
}
}
if ( dim[i] > maxdim )
maxdim = dim[i];
}
A->diameter = maxdim;
}
// ELLPACK
else if ( A->storage_type == Magma_ELL ) {
CHECK( magma_index_malloc_cpu( &dim, A->num_rows));
for( i=0; i<A->num_rows; i++ ) {
dim[i] = 0;
for( j=i*A->max_nnz_row; j<(i+1)*A->max_nnz_row; j++ ) {
if ( MAGMA_Z_REAL( A->val[j] ) > THRESHOLD ) {
tmp = abs( i - A->col[j] );
if ( tmp > dim[i] )
dim[i] = tmp;
}
}
if ( dim[i] > maxdim )
maxdim = dim[i];
}
A->diameter = maxdim;
}
// ELL
else if ( A->storage_type == Magma_ELL ) {
printf("error:format not supported.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
} // end CPU case
else {
printf("error: matrix not on CPU.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
magma_free( &dim );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
magma_d_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
precond_par.solver = Magma_JACOBI;
magma_dsolverinfo_init( &solver_par, &precond_par );
double one = MAGMA_D_MAKE(1.0, 0.0);
double zero = MAGMA_D_MAKE(0.0, 0.0);
magma_d_sparse_matrix A, B, B_d;
magma_d_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=10; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
double *diag_vals;
magma_dmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_D_MAKE( 4.0, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_dmscale( &A, Magma_UNITDIAG );
magma_d_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_mfree(&B_d);
magma_d_mfree(&B);
magma_d_mfree(&A);
magma_d_vfree(&x);
magma_d_vfree(&b);
TESTING_FINALIZE();
return 0;
}
magma_int_t
magma_dsymbolic_ilu(
const magma_int_t levfill, /* level of fill */
const magma_int_t n, /* order of matrix */
magma_int_t *nzl, /* input-output */
magma_int_t *nzu, /* input-output */
const mwIndex *ia,
const mwIndex *ja, /* input */
mwIndex *ial,
mwIndex *jal, /* output lower factor structure */
mwIndex *iau,
mwIndex *jau) /* output upper factor structure */
{
magma_int_t info = 0;
magma_int_t i;
magma_index_t *lnklst=NULL;
magma_index_t *curlev=NULL;
magma_index_t *levels=NULL;
magma_index_t *iwork=NULL;
magma_int_t knzl = 0;
magma_int_t knzu = 0;
CHECK( magma_index_malloc_cpu( &lnklst, n ));
CHECK( magma_index_malloc_cpu( &curlev, n ));
CHECK( magma_index_malloc_cpu( &levels, *nzu ));
CHECK( magma_index_malloc_cpu( &iwork, n ));
for(magma_int_t t=0; t<n; t++){
lnklst[t] = 0;
curlev[t] = 0;
iwork[t] = 0;
}
for(magma_int_t t=0; t<*nzu; t++){
levels[t] = 0;
}
ial[0] = 0;
iau[0] = 0;
for (i=0; i<n; i++) {
//printf("check line %d\n", i);
magma_int_t first, next, j;
/* copy column indices of row into workspace and sort them */
magma_int_t len = ia[i+1] - ia[i];
next = 0;
for (j=ia[i]; j<ia[i+1]; j++)
iwork[next++] = ja[j];
magma_dshell_sort(len, iwork);
//printf("check2 line %d\n", i);
/* construct implied linked list for row */
first = iwork[0];
curlev[first] = 0;
for (j=0; j<=len-2; j++)
{
lnklst[iwork[j]] = iwork[j+1];
curlev[iwork[j]] = 0;
}
// printf("check3 line %d iwork[len-1]:%d\n", i, iwork[len-1]);
lnklst[iwork[len-1]] = n;
curlev[iwork[len-1]] = 0;
/* merge with rows in U */
// printf("check4 line %d lnklst[iwork[len-1]]:%d\n", i, lnklst[iwork[len-1]]);
next = first;
// printf("next:%d (!<) first:%d\n", next, i);
while (next < i)
{
// printf("check line %d while %d\n", i, next);
magma_int_t oldlst = next;
magma_int_t nxtlst = lnklst[next];
magma_int_t row = next;
magma_int_t ii;
/* scan row */
for (ii=iau[row]+1; ii<iau[row+1]; /*nop*/)
{
if (jau[ii] < nxtlst)
{
/* new fill-in */
magma_int_t newlev = curlev[row] + levels[ii] + 1;
if (newlev <= levfill)
{
lnklst[oldlst] = jau[ii];
lnklst[jau[ii]] = nxtlst;
oldlst = jau[ii];
curlev[jau[ii]] = newlev;
}
ii++;
}
else if (jau[ii] == nxtlst)
{
magma_int_t newlev;
oldlst = nxtlst;
nxtlst = lnklst[oldlst];
newlev = curlev[row] + levels[ii] + 1;
curlev[jau[ii]] = min( curlev[jau[ii]], newlev );
ii++;
}
else /* (jau[ii] > nxtlst) */
{
oldlst = nxtlst;
nxtlst = lnklst[oldlst];
}
}
next = lnklst[next];
}
/* gather the pattern magma_int_to L and U */
// printf("check line5 %d\n", i);
next = first;
while (next < i)
{
if (knzl >= *nzl) {
printf("ILU: STORAGE parameter value %d<%d too small.\n", int(*nzl), int(knzl));
printf("Increase STORAGE parameter.\n");
info = -1;
goto cleanup;
}
jal[knzl++] = next;
next = lnklst[next];
}
ial[i+1] = knzl;
// printf("check line6 %d\n", i);
if (next != i)
{
printf("ILU structurally singular.\n");
/*
assert(knzu < *nzu);
levels[knzu] = 2*n;
jau[knzu++] = i;
*/
}
// printf("check line7 %d\n", i);
// printf("next:%d n:%d \n", next, n);
while (next < n)
{
if (knzu >= *nzu) {
printf("ILU: STORAGE parameter value %d < %d too small.\n", int(*nzu), int(knzu));
printf("Increase STORAGE parameter.\n");
info = -1;
goto cleanup;
}
// printf("1 knzu:%d next:%d \n", knzu, next );
levels[knzu] = curlev[next];
// printf("2 knzu:%d next:%d \n", knzu, next );
jau[knzu++] = next;
// printf("3 knzu:%d next:%d \n", knzu, next );
next = lnklst[next];
// printf("4 next:%d n:%d \n", next, n);
}
iau[i+1] = knzu;
}
*nzl = knzl;
*nzu = knzu;
// printf("ende\n");
#if 0
printf( "Actual nnz for ILU: %d\n", *nzl + *nzu );
#endif
cleanup:
magma_free_cpu(lnklst);
magma_free_cpu(curlev);
magma_free_cpu(levels);
magma_free_cpu(iwork);
return info;
}
extern "C" magma_int_t
magma_zmslice(
magma_int_t num_slices,
magma_int_t slice,
magma_z_matrix A,
magma_z_matrix *B,
magma_z_matrix *ALOC,
magma_z_matrix *ANLOC,
magma_index_t *comm_i,
magmaDoubleComplex *comm_v,
magma_int_t *start,
magma_int_t *end,
magma_queue_t queue )
{
magma_int_t info = 0;
if( A.num_rows != A.num_cols ){
printf("%% error: only supported for square matrices.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
if ( A.memory_location == Magma_CPU
&& A.storage_type == Magma_CSR ){
CHECK( magma_zmconvert( A, B, Magma_CSR, Magma_CSR, queue ) );
magma_free_cpu( B->col );
magma_free_cpu( B->val );
CHECK( magma_zmconvert( A, ALOC, Magma_CSR, Magma_CSR, queue ) );
magma_free_cpu( ALOC->col );
magma_free_cpu( ALOC->row );
magma_free_cpu( ALOC->val );
CHECK( magma_zmconvert( A, ANLOC, Magma_CSR, Magma_CSR, queue ) );
magma_free_cpu( ANLOC->col );
magma_free_cpu( ANLOC->row );
magma_free_cpu( ANLOC->val );
magma_int_t i,j,k, nnz, nnz_loc=0, loc_row = 0, nnz_nloc = 0;
magma_index_t col;
magma_int_t size = magma_ceildiv( A.num_rows, num_slices );
magma_int_t lstart = slice*size;
magma_int_t lend = min( (slice+1)*size, A.num_rows );
// correct size for last slice
size = lend-lstart;
CHECK( magma_index_malloc_cpu( &ALOC->row, size+1 ) );
CHECK( magma_index_malloc_cpu( &ANLOC->row, size+1 ) );
// count elements for slice - identity for rest
nnz = A.row[ lend ] - A.row[ lstart ] + ( A.num_rows - size );
CHECK( magma_index_malloc_cpu( &B->col, nnz ) );
CHECK( magma_zmalloc_cpu( &B->val, nnz ) );
// for the communication plan
for( i=0; i<A.num_rows; i++ ) {
comm_i[i] = 0;
comm_v[i] = MAGMA_Z_ZERO;
}
k=0;
B->row[i] = 0;
ALOC->row[0] = 0;
ANLOC->row[0] = 0;
// identity above slice
for( i=0; i<lstart; i++ ) {
B->row[i+1] = B->row[i]+1;
B->val[k] = MAGMA_Z_ONE;
B->col[k] = i;
k++;
}
// slice
for( i=lstart; i<lend; i++ ) {
B->row[i+1] = B->row[i] + (A.row[i+1]-A.row[i]);
for( j=A.row[i]; j<A.row[i+1]; j++ ){
B->val[k] = A.val[j];
col = A.col[j];
B->col[k] = col;
// communication plan
if( col<lstart || col>=lend ){
comm_i[ col ] = 1;
comm_v[ col ] = comm_v[ col ]
+ MAGMA_Z_MAKE( MAGMA_Z_ABS( A.val[j] ), 0.0 );
nnz_nloc++;
} else {
nnz_loc++;
}
k++;
}
loc_row++;
ALOC->row[ loc_row ] = nnz_loc;
ANLOC->row[ loc_row ] = nnz_nloc;
}
CHECK( magma_index_malloc_cpu( &ALOC->col, nnz_loc ) );
CHECK( magma_zmalloc_cpu( &ALOC->val, nnz_loc ) );
ALOC->num_rows = size;
ALOC->num_cols = size;
ALOC->nnz = nnz_loc;
CHECK( magma_index_malloc_cpu( &ANLOC->col, nnz_nloc ) );
CHECK( magma_zmalloc_cpu( &ANLOC->val, nnz_nloc ) );
ANLOC->num_rows = size;
ANLOC->num_cols = A.num_cols;
ANLOC->nnz = nnz_nloc;
nnz_loc = 0;
nnz_nloc = 0;
// local/nonlocal matrix
for( i=lstart; i<lend; i++ ) {
for( j=A.row[i]; j<A.row[i+1]; j++ ){
col = A.col[j];
// insert only in local part in ALOC, nonlocal in ANLOC
if( col<lstart || col>=lend ){
ANLOC->val[ nnz_nloc ] = A.val[j];
ANLOC->col[ nnz_nloc ] = col;
nnz_nloc++;
} else {
ALOC->val[ nnz_loc ] = A.val[j];
ALOC->col[ nnz_loc ] = col-lstart;
nnz_loc++;
}
}
}
// identity below slice
for( i=lend; i<A.num_rows; i++ ) {
B->row[i+1] = B->row[i]+1;
B->val[k] = MAGMA_Z_ONE;
B->col[k] = i;
k++;
}
B->nnz = k;
*start = lstart;
*end = lend;
}
else {
printf("error: mslice only supported for CSR matrices on the CPU: %d %d.\n",
int(A.memory_location), int(A.storage_type) );
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
return info;
}
magma_int_t
magma_d_csr_mtx(
magma_d_matrix *A,
const char *filename,
magma_queue_t queue )
{
char buffer[ 1024 ];
magma_int_t info = 0;
int csr_compressor = 0; // checks for zeros in original file
magma_d_matrix B={Magma_CSR};
magma_index_t *coo_col = NULL;
magma_index_t *coo_row = NULL;
double *coo_val = NULL;
double *new_val = NULL;
magma_index_t* new_row = NULL;
magma_index_t* new_col = NULL;
magma_int_t symmetric = 0;
std::vector< std::pair< magma_index_t, double > > rowval;
FILE *fid = NULL;
MM_typecode matcode;
fid = fopen(filename, "r");
if (fid == NULL) {
printf("%% Unable to open file %s\n", filename);
info = MAGMA_ERR_NOT_FOUND;
goto cleanup;
}
printf("%% Reading sparse matrix from file (%s):", filename);
fflush(stdout);
if (mm_read_banner(fid, &matcode) != 0) {
printf("\n%% Could not process Matrix Market banner: %s.\n", matcode);
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
if (!mm_is_valid(matcode)) {
printf("\n%% Invalid Matrix Market file.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
if ( ! ( ( mm_is_real(matcode) ||
mm_is_integer(matcode) ||
mm_is_pattern(matcode) ||
mm_is_real(matcode) ) &&
mm_is_coordinate(matcode) &&
mm_is_sparse(matcode) ) )
{
mm_snprintf_typecode( buffer, sizeof(buffer), matcode );
printf("\n%% Sorry, MAGMA-sparse does not support Market Market type: [%s]\n", buffer );
printf("%% Only real-valued or pattern coordinate matrices are supported.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
magma_index_t num_rows, num_cols, num_nonzeros;
if (mm_read_mtx_crd_size(fid, &num_rows, &num_cols, &num_nonzeros) != 0) {
info = MAGMA_ERR_UNKNOWN;
goto cleanup;
}
A->storage_type = Magma_CSR;
A->memory_location = Magma_CPU;
A->num_rows = num_rows;
A->num_cols = num_cols;
A->nnz = num_nonzeros;
A->fill_mode = MagmaFull;
CHECK( magma_index_malloc_cpu( &coo_col, A->nnz ) );
CHECK( magma_index_malloc_cpu( &coo_row, A->nnz ) );
CHECK( magma_dmalloc_cpu( &coo_val, A->nnz ) );
if (mm_is_real(matcode) || mm_is_integer(matcode)) {
for(magma_int_t i = 0; i < A->nnz; ++i) {
magma_index_t ROW, COL;
double VAL; // always read in a double and convert later if necessary
fscanf(fid, " %d %d %lf \n", &ROW, &COL, &VAL);
if ( VAL == 0 )
csr_compressor = 1;
coo_row[i] = ROW - 1;
coo_col[i] = COL - 1;
coo_val[i] = MAGMA_D_MAKE( VAL, 0.);
}
} else if (mm_is_pattern(matcode) ) {
for(magma_int_t i = 0; i < A->nnz; ++i) {
magma_index_t ROW, COL;
fscanf(fid, " %d %d \n", &ROW, &COL );
coo_row[i] = ROW - 1;
coo_col[i] = COL - 1;
coo_val[i] = MAGMA_D_MAKE( 1.0, 0.);
}
} else if (mm_is_real(matcode) ){
for(magma_int_t i = 0; i < A->nnz; ++i) {
magma_index_t ROW, COL;
double VAL, VALC; // always read in a double and convert later if necessary
fscanf(fid, " %d %d %lf %lf\n", &ROW, &COL, &VAL, &VALC);
coo_row[i] = ROW - 1;
coo_col[i] = COL - 1;
coo_val[i] = MAGMA_D_MAKE( VAL, VALC);
}
// printf(" ...successfully read real matrix... ");
} else {
printf("\n%% Unrecognized data type\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
fclose(fid);
fid = NULL;
printf(" done. Converting to CSR:");
fflush(stdout);
A->sym = Magma_GENERAL;
if( mm_is_symmetric(matcode) ) {
symmetric = 1;
}
if ( mm_is_symmetric(matcode) || mm_is_symmetric(matcode) ) {
// duplicate off diagonal entries
printf("\n%% Detected symmetric case.");
A->sym = Magma_SYMMETRIC;
magma_index_t off_diagonals = 0;
for(magma_int_t i = 0; i < A->nnz; ++i) {
if (coo_row[i] != coo_col[i])
++off_diagonals;
}
magma_index_t true_nonzeros = 2*off_diagonals + (A->nnz - off_diagonals);
//printf("%% total number of nonzeros: %d\n%%", int(A->nnz));
CHECK( magma_index_malloc_cpu( &new_row, true_nonzeros ));
CHECK( magma_index_malloc_cpu( &new_col, true_nonzeros ));
CHECK( magma_dmalloc_cpu( &new_val, true_nonzeros ));
magma_index_t ptr = 0;
for(magma_int_t i = 0; i < A->nnz; ++i) {
if (coo_row[i] != coo_col[i]) {
new_row[ptr] = coo_row[i];
new_col[ptr] = coo_col[i];
new_val[ptr] = coo_val[i];
ptr++;
new_col[ptr] = coo_row[i];
new_row[ptr] = coo_col[i];
new_val[ptr] = (symmetric == 0) ? coo_val[i] : conj(coo_val[i]);
ptr++;
} else {
new_row[ptr] = coo_row[i];
new_col[ptr] = coo_col[i];
new_val[ptr] = coo_val[i];
ptr++;
}
}
magma_free_cpu(coo_row);
magma_free_cpu(coo_col);
magma_free_cpu(coo_val);
coo_row = new_row;
coo_col = new_col;
coo_val = new_val;
A->nnz = true_nonzeros;
//printf("total number of nonzeros: %d\n", A->nnz);
} // end symmetric case
CHECK( magma_dmalloc_cpu( &A->val, A->nnz ));
CHECK( magma_index_malloc_cpu( &A->col, A->nnz ));
CHECK( magma_index_malloc_cpu( &A->row, A->num_rows+1 ));
// original code from Nathan Bell and Michael Garland
for (magma_index_t i = 0; i < num_rows; i++)
(A->row)[i] = 0;
for (magma_index_t i = 0; i < A->nnz; i++)
(A->row)[coo_row[i]]++;
// cumulative sum the nnz per row to get row[]
magma_int_t cumsum;
cumsum = 0;
for(magma_int_t i = 0; i < num_rows; i++) {
magma_index_t temp = (A->row)[i];
(A->row)[i] = cumsum;
cumsum += temp;
}
(A->row)[num_rows] = A->nnz;
// write Aj,Ax into Bj,Bx
for(magma_int_t i = 0; i < A->nnz; i++) {
magma_index_t row_ = coo_row[i];
magma_index_t dest = (A->row)[row_];
(A->col)[dest] = coo_col[i];
(A->val)[dest] = coo_val[i];
(A->row)[row_]++;
}
magma_free_cpu(coo_row);
magma_free_cpu(coo_col);
magma_free_cpu(coo_val);
coo_row = NULL;
coo_col = NULL;
coo_val = NULL;
int last;
last = 0;
for(int i = 0; i <= num_rows; i++) {
int temp = (A->row)[i];
(A->row)[i] = last;
last = temp;
}
(A->row)[A->num_rows] = A->nnz;
// sort column indices within each row
// copy into vector of pairs (column index, value), sort by column index, then copy back
for (magma_index_t k=0; k < A->num_rows; ++k) {
int kk = (A->row)[k];
int len = (A->row)[k+1] - (A->row)[k];
rowval.resize( len );
for( int i=0; i < len; ++i ) {
rowval[i] = std::make_pair( (A->col)[kk+i], (A->val)[kk+i] );
}
std::sort( rowval.begin(), rowval.end(), compare_first );
for( int i=0; i < len; ++i ) {
(A->col)[kk+i] = rowval[i].first;
(A->val)[kk+i] = rowval[i].second;
}
}
if ( csr_compressor > 0) { // run the CSR compressor to remove zeros
//printf("removing zeros: ");
CHECK( magma_dmtransfer( *A, &B, Magma_CPU, Magma_CPU, queue ));
CHECK( magma_d_csr_compressor(
&(A->val), &(A->row), &(A->col),
&B.val, &B.row, &B.col, &B.num_rows, queue ));
B.nnz = B.row[num_rows];
//printf(" remaining nonzeros:%d ", B.nnz);
magma_free_cpu( A->val );
magma_free_cpu( A->row );
magma_free_cpu( A->col );
CHECK( magma_dmtransfer( B, A, Magma_CPU, Magma_CPU, queue ));
//printf("done.\n");
}
A->true_nnz = A->nnz;
printf(" done.\n");
cleanup:
if ( fid != NULL ) {
fclose( fid );
fid = NULL;
}
magma_dmfree( &B, queue );
magma_free_cpu(coo_row);
magma_free_cpu(coo_col);
magma_free_cpu(coo_val);
return info;
}
extern "C" magma_int_t
magma_ccsrsplit(
magma_int_t bsize,
magma_c_sparse_matrix A,
magma_c_sparse_matrix *D,
magma_c_sparse_matrix *R,
magma_queue_t queue )
{
if ( A.memory_location == Magma_CPU &&
( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO ) ) {
magma_int_t i, k, j, nnz_diag, nnz_offd;
magma_int_t stat_cpu = 0, stat_dev = 0;
D->val = NULL;
D->col = NULL;
D->row = NULL;
D->rowidx = NULL;
D->blockinfo = NULL;
D->diag = NULL;
D->dval = NULL;
D->dcol = NULL;
D->drow = NULL;
D->drowidx = NULL;
D->ddiag = NULL;
R->val = NULL;
R->col = NULL;
R->row = NULL;
R->rowidx = NULL;
R->blockinfo = NULL;
R->diag = NULL;
R->dval = NULL;
R->dcol = NULL;
R->drow = NULL;
R->drowidx = NULL;
R->ddiag = NULL;
nnz_diag = nnz_offd = 0;
// Count the new number of nonzeroes in the two matrices
for( i=0; i<A.num_rows; i+=bsize )
for( k=i; k<min(A.num_rows,i+bsize); k++ )
for( j=A.row[k]; j<A.row[k+1]; j++ )
if ( A.col[j] < i )
nnz_offd++;
else if ( A.col[j] < i+bsize )
nnz_diag++;
else
nnz_offd++;
// Allocate memory for the new matrices
D->storage_type = Magma_CSRD;
D->memory_location = A.memory_location;
D->num_rows = A.num_rows;
D->num_cols = A.num_cols;
D->nnz = nnz_diag;
R->storage_type = Magma_CSR;
R->memory_location = A.memory_location;
R->num_rows = A.num_rows;
R->num_cols = A.num_cols;
R->nnz = nnz_offd;
stat_cpu += magma_cmalloc_cpu( &D->val, nnz_diag );
stat_cpu += magma_index_malloc_cpu( &D->row, A.num_rows+1 );
stat_cpu += magma_index_malloc_cpu( &D->col, nnz_diag );
stat_cpu += magma_cmalloc_cpu( &R->val, nnz_offd );
stat_cpu += magma_index_malloc_cpu( &R->row, A.num_rows+1 );
stat_cpu += magma_index_malloc_cpu( &R->col, nnz_offd );
if( stat_cpu != 0 ){
magma_c_mfree( D, queue );
magma_c_mfree( R, queue );
return MAGMA_ERR_HOST_ALLOC;
}
// Fill up the new sparse matrices
D->row[0] = 0;
R->row[0] = 0;
nnz_offd = nnz_diag = 0;
for( i=0; i<A.num_rows; i+=bsize) {
for( k=i; k<min(A.num_rows,i+bsize); k++ ) {
D->row[k+1] = D->row[k];
R->row[k+1] = R->row[k];
for( j=A.row[k]; j<A.row[k+1]; j++ ) {
if ( A.col[j] < i ) {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
else if ( A.col[j] < i+bsize ) {
// larger than diagonal remain as before
if ( A.col[j]>k ) {
D->val[nnz_diag] = A.val[ j ];
D->col[nnz_diag] = A.col[ j ];
D->row[k+1]++;
}
// diagonal is written first
else if ( A.col[j]==k ) {
D->val[D->row[k]] = A.val[ j ];
D->col[D->row[k]] = A.col[ j ];
D->row[k+1]++;
}
// smaller than diagonal are shifted one to the right
// to have room for the diagonal
else {
D->val[nnz_diag+1] = A.val[ j ];
D->col[nnz_diag+1] = A.col[ j ];
D->row[k+1]++;
}
nnz_diag++;
}
else {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
}
}
}
return MAGMA_SUCCESS;
}
else {
magma_c_sparse_matrix Ah, ACSR, DCSR, RCSR, Dh, Rh;
magma_c_mtransfer( A, &Ah, A.memory_location, Magma_CPU, queue );
magma_c_mconvert( Ah, &ACSR, A.storage_type, Magma_CSR, queue );
magma_ccsrsplit( bsize, ACSR, &DCSR, &RCSR, queue );
magma_c_mconvert( DCSR, &Dh, Magma_CSR, A.storage_type, queue );
magma_c_mconvert( RCSR, &Rh, Magma_CSR, A.storage_type, queue );
magma_c_mtransfer( Dh, D, Magma_CPU, A.memory_location, queue );
magma_c_mtransfer( Rh, R, Magma_CPU, A.memory_location, queue );
magma_c_mfree( &Ah, queue );
magma_c_mfree( &ACSR, queue );
magma_c_mfree( &Dh, queue );
magma_c_mfree( &DCSR, queue );
magma_c_mfree( &Rh, queue );
magma_c_mfree( &RCSR, queue );
return MAGMA_SUCCESS;
}
}
extern "C" magma_int_t
magma_dmtransfer(
magma_d_matrix A,
magma_d_matrix *B,
magma_location_t src,
magma_location_t dst,
magma_queue_t queue )
{
magma_int_t info = 0;
B->val = NULL;
B->diag = NULL;
B->row = NULL;
B->rowidx = NULL;
B->col = NULL;
B->blockinfo = NULL;
B->dval = NULL;
B->ddiag = NULL;
B->drow = NULL;
B->drowidx = NULL;
B->dcol = NULL;
B->diag = NULL;
B->ddiag = NULL;
B->list = NULL;
B->dlist = NULL;
// first case: copy matrix from host to device
if ( src == Magma_CPU && dst == Magma_DEV ) {
//CSR-type
if ( A.storage_type == Magma_CSR || A.storage_type == Magma_CUCSR
|| A.storage_type == Magma_CSC
|| A.storage_type == Magma_CSRD
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
// data transfer
magma_dsetvector( A.nnz, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.num_rows + 1, A.row, 1, B->drow, 1, queue );
magma_index_setvector( A.nnz, A.col, 1, B->dcol, 1, queue );
}
//COO-type
else if ( A.storage_type == Magma_COO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drowidx, A.nnz ));
// data transfer
magma_dsetvector( A.nnz, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.nnz, A.col, 1, B->dcol, 1, queue );
magma_index_setvector( A.nnz, A.rowidx, 1, B->drowidx, 1, queue );
}
//CSRCOO-type
else if ( A.storage_type == Magma_CSRCOO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drowidx, A.nnz ));
// data transfer
magma_dsetvector( A.nnz, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.num_rows + 1, A.row, 1, B->drow, 1, queue );
magma_index_setvector( A.nnz, A.col, 1, B->dcol, 1, queue );
magma_index_setvector( A.nnz, A.rowidx, 1, B->drowidx, 1, queue );
}
//ELL/ELLPACKT-type
else if ( A.storage_type == Magma_ELLPACKT || A.storage_type == Magma_ELL ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dsetvector( A.num_rows * A.max_nnz_row, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.num_rows * A.max_nnz_row, A.col, 1, B->dcol, 1, queue );
}
//ELLD-type
else if ( A.storage_type == Magma_ELLD ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dsetvector( A.num_rows * A.max_nnz_row, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.num_rows * A.max_nnz_row, A.col, 1, B->dcol, 1, queue );
}
//ELLRT-type
else if ( A.storage_type == Magma_ELLRT ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->alignment = A.alignment;
magma_int_t rowlength = magma_roundup( A.max_nnz_row, A.alignment );
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * rowlength ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * rowlength ));
CHECK( magma_index_malloc( &B->drow, A.num_rows ));
// data transfer
magma_dsetvector( A.num_rows * rowlength, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.num_rows * rowlength, A.col, 1, B->dcol, 1, queue );
magma_index_setvector( A.num_rows, A.row, 1, B->drow, 1, queue );
}
//SELLP-type
else if ( A.storage_type == Magma_SELLP ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.numblocks + 1 ));
// data transfer
magma_dsetvector( A.nnz, A.val, 1, B->dval, 1, queue );
magma_index_setvector( A.nnz, A.col, 1, B->dcol, 1, queue );
magma_index_setvector( A.numblocks + 1, A.row, 1, B->drow, 1, queue );
}
//BCSR-type
else if ( A.storage_type == Magma_BCSR ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
magma_int_t size_b = A.blocksize;
//magma_int_t c_blocks = ceil( (float)A.num_cols / (float)size_b );
// max number of blocks per row
//magma_int_t r_blocks = ceil( (float)A.num_rows / (float)size_b );
magma_int_t r_blocks = magma_ceildiv( A.num_rows, size_b );
// max number of blocks per column
// memory allocation
CHECK( magma_dmalloc( &B->dval, size_b * size_b * A.numblocks ));
CHECK( magma_index_malloc( &B->drow, r_blocks + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.numblocks ));
// data transfer
magma_dsetvector( size_b * size_b * A.numblocks, A.val, 1, B->dval, 1, queue );
magma_index_setvector( r_blocks + 1, A.row, 1, B->drow, 1, queue );
magma_index_setvector( A.numblocks, A.col, 1, B->dcol, 1, queue );
}
//DENSE-type
else if ( A.storage_type == Magma_DENSE ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->major = A.major;
B->ld = A.ld;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.num_cols ));
// data transfer
magma_dsetvector( A.num_rows * A.num_cols, A.val, 1, B->dval, 1, queue );
}
}
// second case: copy matrix from host to host
else if ( src == Magma_CPU && dst == Magma_CPU ) {
//CSR-type
if ( A.storage_type == Magma_CSR || A.storage_type == Magma_CUCSR
|| A.storage_type == Magma_CSC
|| A.storage_type == Magma_CSRD
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
// data transfer
for( magma_int_t i=0; i<A.nnz; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
}
for( magma_int_t i=0; i<A.num_rows+1; i++ ) {
B->row[i] = A.row[i];
}
}
//COO-type
else if ( A.storage_type == Magma_COO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->rowidx, A.nnz ));
// data transfer
for( magma_int_t i=0; i<A.nnz; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
B->rowidx[i] = A.rowidx[i];
}
}
//CSRCOO-type
else if ( A.storage_type == Magma_CSRCOO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->rowidx, A.nnz ));
// data transfer
for( magma_int_t i=0; i<A.nnz; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
B->rowidx[i] = A.rowidx[i];
}
for( magma_int_t i=0; i<A.num_rows+1; i++ ) {
B->row[i] = A.row[i];
}
}
//ELL/ELLPACKT-type
else if ( A.storage_type == Magma_ELLPACKT || A.storage_type == Magma_ELL ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc_cpu( &B->col, A.num_rows * A.max_nnz_row ));
// data transfer
for( magma_int_t i=0; i<A.num_rows*A.max_nnz_row; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
}
}
//ELLD-type
else if ( A.storage_type == Magma_ELLD ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc_cpu( &B->col, A.num_rows * A.max_nnz_row ));
// data transfer
for( magma_int_t i=0; i<A.num_rows*A.max_nnz_row; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
}
}
//ELLRT-type
else if ( A.storage_type == Magma_ELLRT ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->alignment = A.alignment;
//int threads_per_row = A.alignment;
//int rowlength = magma_roundup( A.max_nnz_row, threads_per_row );
magma_int_t rowlength = magma_roundup( A.max_nnz_row, A.alignment );
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, rowlength * A.num_rows ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows ));
CHECK( magma_index_malloc_cpu( &B->col, rowlength * A.num_rows ));
// data transfer
for( magma_int_t i=0; i<A.num_rows*rowlength; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
}
for( magma_int_t i=0; i<A.num_rows; i++ ) {
B->row[i] = A.row[i];
}
}
//SELLP-type
else if ( A.storage_type == Magma_SELLP ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->alignment = A.alignment;
B->numblocks = A.numblocks;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.numblocks + 1 ));
// data transfer
for( magma_int_t i=0; i<A.nnz; i++ ) {
B->val[i] = A.val[i];
B->col[i] = A.col[i];
}
for( magma_int_t i=0; i<A.numblocks+1; i++ ) {
B->row[i] = A.row[i];
}
}
//BCSR-type
else if ( A.storage_type == Magma_BCSR ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
magma_int_t size_b = A.blocksize;
//magma_int_t c_blocks = ceil( (float)A.num_cols / (float)size_b );
// max number of blocks per row
//magma_int_t r_blocks = ceil( (float)A.num_rows / (float)size_b );
magma_int_t r_blocks = magma_ceildiv( A.num_rows, size_b );
// max number of blocks per column
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, size_b * size_b * A.numblocks ));
CHECK( magma_index_malloc_cpu( &B->row, r_blocks + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.numblocks ));
// data transfer
//magma_dsetvector( size_b * size_b * A.numblocks, A.val, 1, B->dval, 1, queue );
for( magma_int_t i=0; i<size_b*size_b*A.numblocks; i++ ) {
B->dval[i] = A.val[i];
}
//magma_index_setvector( r_blocks + 1, A.row, 1, B->drow, 1, queue );
for( magma_int_t i=0; i<r_blocks+1; i++ ) {
B->drow[i] = A.row[i];
}
//magma_index_setvector( A.numblocks, A.col, 1, B->dcol, 1, queue );
for( magma_int_t i=0; i<A.numblocks; i++ ) {
B->dcol[i] = A.col[i];
}
}
//DENSE-type
else if ( A.storage_type == Magma_DENSE ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->major = A.major;
B->ld = A.ld;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.num_cols ));
// data transfer
for( magma_int_t i=0; i<A.num_rows*A.num_cols; i++ ) {
B->val[i] = A.val[i];
}
}
}
// third case: copy matrix from device to host
else if ( src == Magma_DEV && dst == Magma_CPU ) {
//CSR-type
if ( A.storage_type == Magma_CSR || A.storage_type == Magma_CUCSR
|| A.storage_type == Magma_CSC
|| A.storage_type == Magma_CSRD
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
// data transfer
magma_dgetvector( A.nnz, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.num_rows + 1, A.drow, 1, B->row, 1, queue );
magma_index_getvector( A.nnz, A.dcol, 1, B->col, 1, queue );
}
//COO-type
else if ( A.storage_type == Magma_COO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->rowidx, A.nnz ));
// data transfer
magma_dgetvector( A.nnz, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.nnz, A.dcol, 1, B->col, 1, queue );
magma_index_getvector( A.nnz, A.drowidx, 1, B->rowidx, 1, queue );
}
//CSRCOO-type
else if ( A.storage_type == Magma_CSRCOO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->rowidx, A.nnz ));
// data transfer
magma_dgetvector( A.nnz, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.num_rows + 1, A.drow, 1, B->row, 1, queue );
magma_index_getvector( A.nnz, A.dcol, 1, B->col, 1, queue );
magma_index_getvector( A.nnz, A.drowidx, 1, B->rowidx, 1, queue );
}
//ELL/ELLPACKT-type
else if ( A.storage_type == Magma_ELLPACKT || A.storage_type == Magma_ELL ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc_cpu( &B->col, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dgetvector( A.num_rows * A.max_nnz_row, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.num_rows * A.max_nnz_row, A.dcol, 1, B->col, 1, queue );
}
//ELLD-type
else if ( A.storage_type == Magma_ELLD ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc_cpu( &B->col, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dgetvector( A.num_rows * A.max_nnz_row, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.num_rows * A.max_nnz_row, A.dcol, 1, B->col, 1, queue );
}
//ELLRT-type
else if ( A.storage_type == Magma_ELLRT ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->alignment = A.alignment;
//int threads_per_row = A.alignment;
//int rowlength = magma_roundup( A.max_nnz_row, threads_per_row );
magma_int_t rowlength = magma_roundup( A.max_nnz_row, A.alignment );
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, rowlength * A.num_rows ));
CHECK( magma_index_malloc_cpu( &B->row, A.num_rows ));
CHECK( magma_index_malloc_cpu( &B->col, rowlength * A.num_rows ));
// data transfer
magma_dgetvector( A.num_rows * rowlength, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.num_rows * rowlength, A.dcol, 1, B->col, 1, queue );
magma_index_getvector( A.num_rows, A.drow, 1, B->row, 1, queue );
}
//SELLP-type
else if ( A.storage_type == Magma_SELLP ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->col, A.nnz ));
CHECK( magma_index_malloc_cpu( &B->row, A.numblocks + 1 ));
// data transfer
magma_dgetvector( A.nnz, A.dval, 1, B->val, 1, queue );
magma_index_getvector( A.nnz, A.dcol, 1, B->col, 1, queue );
magma_index_getvector( A.numblocks + 1, A.drow, 1, B->row, 1, queue );
}
//BCSR-type
else if ( A.storage_type == Magma_BCSR ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
magma_int_t size_b = A.blocksize;
//magma_int_t c_blocks = ceil( (float)A.num_cols / (float)size_b );
// max number of blocks per row
//magma_int_t r_blocks = ceil( (float)A.num_rows / (float)size_b );
magma_int_t r_blocks = magma_ceildiv( A.num_rows, size_b );
// max number of blocks per column
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, size_b * size_b * A.numblocks ));
CHECK( magma_index_malloc_cpu( &B->row, r_blocks + 1 ));
CHECK( magma_index_malloc_cpu( &B->col, A.numblocks ));
// data transfer
magma_dgetvector( size_b * size_b * A.numblocks, A.dval, 1, B->val, 1, queue );
magma_index_getvector( r_blocks + 1, A.drow, 1, B->row, 1, queue );
magma_index_getvector( A.numblocks, A.dcol, 1, B->col, 1, queue );
}
//DENSE-type
else if ( A.storage_type == Magma_DENSE ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_CPU;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->major = A.major;
B->ld = A.ld;
// memory allocation
CHECK( magma_dmalloc_cpu( &B->val, A.num_rows * A.num_cols ));
// data transfer
magma_dgetvector( A.num_rows * A.num_cols, A.dval, 1, B->val, 1, queue );
}
}
// fourth case: copy matrix from device to device
else if ( src == Magma_DEV && dst == Magma_DEV ) {
//CSR-type
if ( A.storage_type == Magma_CSR || A.storage_type == Magma_CUCSR
|| A.storage_type == Magma_CSC
|| A.storage_type == Magma_CSRD
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
// data transfer
magma_dcopyvector( A.nnz, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.num_rows + 1, A.drow, 1, B->drow, 1, queue );
magma_index_copyvector( A.nnz, A.dcol, 1, B->dcol, 1, queue );
}
//COO-type
else if ( A.storage_type == Magma_COO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drowidx, A.nnz ));
// data transfer
magma_dcopyvector( A.nnz, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.nnz, A.dcol, 1, B->dcol, 1, queue );
magma_index_copyvector( A.nnz, A.drowidx, 1, B->drowidx, 1, queue );
}
//CSRCOO-type
else if ( A.storage_type == Magma_CSRCOO ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drowidx, A.nnz ));
// data transfer
magma_dcopyvector( A.nnz, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.num_rows + 1, A.drow, 1, B->drow, 1, queue );
magma_index_copyvector( A.nnz, A.dcol, 1, B->dcol, 1, queue );
magma_index_copyvector( A.nnz, A.drowidx, 1, B->drowidx, 1, queue );
}
//ELL/ELLPACKT-type
else if ( A.storage_type == Magma_ELLPACKT || A.storage_type == Magma_ELL ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dcopyvector( A.num_rows * A.max_nnz_row, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.num_rows * A.max_nnz_row, A.dcol, 1, B->dcol, 1, queue );
}
//ELLD-type
else if ( A.storage_type == Magma_ELLD ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.max_nnz_row ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * A.max_nnz_row ));
// data transfer
magma_dcopyvector( A.num_rows * A.max_nnz_row, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.num_rows * A.max_nnz_row, A.dcol, 1, B->dcol, 1, queue );
}
//ELLRT-type
else if ( A.storage_type == Magma_ELLRT ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->alignment = A.alignment;
//int threads_per_row = A.alignment;
//int rowlength = magma_roundup( A.max_nnz_row, threads_per_row );
magma_int_t rowlength = magma_roundup( A.max_nnz_row, A.alignment );
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * rowlength ));
CHECK( magma_index_malloc( &B->dcol, A.num_rows * rowlength ));
CHECK( magma_index_malloc( &B->drow, A.num_rows ));
// data transfer
magma_dcopyvector( A.num_rows * rowlength, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.num_rows * rowlength, A.dcol, 1, B->dcol, 1, queue );
magma_index_copyvector( A.num_rows, A.drow, 1, B->drow, 1, queue );
}
//SELLP-type
else if ( A.storage_type == Magma_SELLP ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.nnz ));
CHECK( magma_index_malloc( &B->dcol, A.nnz ));
CHECK( magma_index_malloc( &B->drow, A.numblocks + 1 ));
// data transfer
magma_dcopyvector( A.nnz, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( A.nnz, A.dcol, 1, B->dcol, 1, queue );
magma_index_copyvector( A.numblocks + 1, A.drow, 1, B->drow, 1, queue );
}
//BCSR-type
else if ( A.storage_type == Magma_BCSR ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->blocksize = A.blocksize;
B->numblocks = A.numblocks;
B->alignment = A.alignment;
magma_int_t size_b = A.blocksize;
//magma_int_t c_blocks = ceil( (float)A.num_cols / (float)size_b );
// max number of blocks per row
//magma_int_t r_blocks = ceil( (float)A.num_rows / (float)size_b );
magma_int_t r_blocks = magma_ceildiv( A.num_rows, size_b );
// max number of blocks per column
// memory allocation
CHECK( magma_dmalloc( &B->dval, size_b * size_b * A.numblocks ));
CHECK( magma_index_malloc( &B->drow, r_blocks + 1 ));
CHECK( magma_index_malloc( &B->dcol, A.numblocks ));
// data transfer
magma_dcopyvector( size_b * size_b * A.numblocks, A.dval, 1, B->dval, 1, queue );
magma_index_copyvector( r_blocks + 1, A.drow, 1, B->drow, 1, queue );
magma_index_copyvector( A.numblocks, A.dcol, 1, B->dcol, 1, queue );
}
//DENSE-type
else if ( A.storage_type == Magma_DENSE ) {
// fill in information for B
B->storage_type = A.storage_type;
B->memory_location = Magma_DEV;
B->sym = A.sym;
B->diagorder_type = A.diagorder_type;
B->fill_mode = A.fill_mode;
B->num_rows = A.num_rows;
B->num_cols = A.num_cols;
B->nnz = A.nnz; B->true_nnz = A.true_nnz;
B->max_nnz_row = A.max_nnz_row;
B->diameter = A.diameter;
B->major = A.major;
B->ld = A.ld;
// memory allocation
CHECK( magma_dmalloc( &B->dval, A.num_rows * A.num_cols ));
// data transfer
magma_dcopyvector( A.num_rows * A.num_cols, A.dval, 1, B->dval, 1, queue );
}
}
cleanup:
if( info != 0 ){
magma_dmfree( B, queue );
}
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_z_solver_par solver_par;
magma_z_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.restart = 30;
solver_par.num_eigenvalues = 0;
solver_par.ortho = Magma_CGS;
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_z_sparse_matrix A, B, B_d;
magma_z_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=100; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
magmaDoubleComplex *diag_vals;
magma_zmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_Z_MAKE( 4.1, 0.0 );
diag_vals[1] = MAGMA_Z_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_Z_MAKE( -1.0, 0.0 );
magma_zmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_zmscale( &A, Magma_UNITDIAG );
magma_z_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// test CG ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zcg_res( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test BICGSTAB ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zbicgstab( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PBICGSTAB Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
/*
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);printf("here\n");
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test GMRES ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zgmres( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PGMRES Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);*/
// test PGMRES ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
printf("all tests passed.\n");
magma_z_mfree(&B_d);
magma_z_mfree(&B);
magma_z_mfree(&A);
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
c_transpose_csr(
magma_int_t n_rows,
magma_int_t n_cols,
magma_int_t nnz,
magmaFloatComplex *val,
magma_index_t *row,
magma_index_t *col,
magma_int_t *new_n_rows,
magma_int_t *new_n_cols,
magma_int_t *new_nnz,
magmaFloatComplex **new_val,
magma_index_t **new_row,
magma_index_t **new_col,
magma_queue_t queue )
{
nnz = row[n_rows];
*new_n_rows = n_cols;
*new_n_cols = n_rows;
*new_nnz = nnz;
magmaFloatComplex ** valtemp;
magma_index_t ** coltemp;
valtemp = (magmaFloatComplex**)malloc((n_rows)*sizeof(magmaFloatComplex*));
coltemp =(magma_index_t**)malloc((n_rows)*sizeof(magma_index_t*));
// temporary 2-dimensional arrays valtemp/coltemp
// where val[i] is the array with the values of the i-th column of the matrix
magma_index_t *nnztemp;
magma_index_malloc_cpu( &nnztemp, n_rows );
for( magma_int_t i=0; i<n_rows; i++ )
nnztemp[i]=0;
for( magma_int_t i=0; i<nnz; i++ )
nnztemp[col[i]]++;
for( magma_int_t i=0; i<n_rows; i++ ){
valtemp[i] =
(magmaFloatComplex*)malloc((nnztemp[i])*sizeof(magmaFloatComplex));
coltemp[i] = (magma_index_t*)malloc(nnztemp[i]*sizeof(magma_index_t));
}
for( magma_int_t i=0; i<n_rows; i++ )
nnztemp[i]=0;
for( magma_int_t j=0; j<n_rows; j++ ){
for( magma_int_t i=row[j]; i<row[j+1]; i++ ){
valtemp[col[i]][nnztemp[col[i]]]=val[i];
coltemp[col[i]][nnztemp[col[i]]]=j;
nnztemp[col[i]]++;
}
}
//csr structure for transposed matrix
*new_val = new magmaFloatComplex[nnz];
*new_row = new magma_index_t[n_rows+1];
*new_col = new magma_index_t[nnz];
//fill the transposed csr structure
magma_int_t nnztmp=0;
(*new_row)[0]=0;
for( magma_int_t j=0; j<n_rows; j++ ){
for( magma_int_t i=0; i<nnztemp[j]; i++ ){
(*new_val)[nnztmp]=valtemp[j][i];
(*new_col)[nnztmp]=coltemp[j][i];
nnztmp++;
}
(*new_row)[j+1]=nnztmp;
}
//usually the temporary memory should be freed afterwards
//however, it does not work
/*
for( magma_int_t j=0; j<n_rows; j++ ){
free(valtemp[j]);
free(coltemp[j]);
}
free(valtemp);free(coltemp);
printf("check9\n");
fflush(stdout);
*/
magma_free_cpu( nnztemp );
return MAGMA_SUCCESS;
}
magma_int_t
magma_dsymbilu(
magma_d_matrix *A,
magma_int_t levels,
magma_d_matrix *L,
magma_d_matrix *U,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_d_matrix A_copy={Magma_CSR}, B={Magma_CSR};
magma_d_matrix hA={Magma_CSR}, CSRCOOA={Magma_CSR};
if( A->memory_location == Magma_CPU && A->storage_type == Magma_CSR ){
CHECK( magma_dmtransfer( *A, &A_copy, Magma_CPU, Magma_CPU, queue ));
CHECK( magma_dmtransfer( *A, &B, Magma_CPU, Magma_CPU, queue ));
// possibility to scale to unit diagonal
//magma_dmscale( &B, Magma_UNITDIAG );
CHECK( magma_dmconvert( B, L, Magma_CSR, Magma_CSR , queue));
CHECK( magma_dmconvert( B, U, Magma_CSR, Magma_CSR, queue ));
magma_int_t num_lnnz = (levels > 0 ) ? B.nnz/2*(2*levels+50) : B.nnz;
magma_int_t num_unnz = (levels > 0 ) ? B.nnz/2*(2*levels+50) : B.nnz;
magma_free_cpu( L->col );
magma_free_cpu( U->col );
CHECK( magma_index_malloc_cpu( &L->col, num_lnnz ));
CHECK( magma_index_malloc_cpu( &U->col, num_unnz ));
magma_dsymbolic_ilu( levels, A->num_rows, &num_lnnz, &num_unnz, B.row, B.col,
L->row, L->col, U->row, U->col );
L->nnz = num_lnnz;
U->nnz = num_unnz;
magma_free_cpu( L->val );
magma_free_cpu( U->val );
CHECK( magma_dmalloc_cpu( &L->val, L->nnz ));
CHECK( magma_dmalloc_cpu( &U->val, U->nnz ));
for( magma_int_t i=0; i<L->nnz; i++ )
L->val[i] = MAGMA_D_MAKE( 0.0, 0.0 );
for( magma_int_t i=0; i<U->nnz; i++ )
U->val[i] = MAGMA_D_MAKE( 0.0, 0.0 );
// take the original values (scaled) as initial guess for L
for(magma_int_t i=0; i<L->num_rows; i++){
for(magma_int_t j=B.row[i]; j<B.row[i+1]; j++){
magma_index_t lcol = B.col[j];
for(magma_int_t k=L->row[i]; k<L->row[i+1]; k++){
if( L->col[k] == lcol ){
L->val[k] = B.val[j];
}
}
}
}
// take the original values (scaled) as initial guess for U
for(magma_int_t i=0; i<U->num_rows; i++){
for(magma_int_t j=B.row[i]; j<B.row[i+1]; j++){
magma_index_t lcol = B.col[j];
for(magma_int_t k=U->row[i]; k<U->row[i+1]; k++){
if( U->col[k] == lcol ){
U->val[k] = B.val[j];
}
}
}
}
magma_dmfree( &B, queue );
// fill A with the new structure;
magma_free_cpu( A->col );
magma_free_cpu( A->val );
CHECK( magma_index_malloc_cpu( &A->col, L->nnz+U->nnz ));
CHECK( magma_dmalloc_cpu( &A->val, L->nnz+U->nnz ));
A->nnz = L->nnz+U->nnz;
magma_int_t z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
A->row[i] = z;
for(magma_int_t j=L->row[i]; j<L->row[i+1]; j++){
A->col[z] = L->col[j];
A->val[z] = L->val[j];
z++;
}
for(magma_int_t j=U->row[i]; j<U->row[i+1]; j++){
A->col[z] = U->col[j];
A->val[z] = U->val[j];
z++;
}
}
A->row[A->num_rows] = z;
// reset the values of A to the original entries
for(magma_int_t i=0; i<A->num_rows; i++){
for(magma_int_t j=A_copy.row[i]; j<A_copy.row[i+1]; j++){
magma_index_t lcol = A_copy.col[j];
for(magma_int_t k=A->row[i]; k<A->row[i+1]; k++){
if( A->col[k] == lcol ){
A->val[k] = A_copy.val[j];
}
}
}
}
}
else {
magma_storage_t A_storage = A->storage_type;
magma_location_t A_location = A->memory_location;
CHECK( magma_dmtransfer( *A, &hA, A->memory_location, Magma_CPU, queue ));
CHECK( magma_dmconvert( hA, &CSRCOOA, hA.storage_type, Magma_CSR, queue ));
CHECK( magma_dsymbilu( &CSRCOOA, levels, L, U, queue ));
magma_dmfree( &hA, queue );
magma_dmfree( A, queue );
CHECK( magma_dmconvert( CSRCOOA, &hA, Magma_CSR, A_storage, queue ));
CHECK( magma_dmtransfer( hA, A, Magma_CPU, A_location, queue ));
}
cleanup:
if( info != 0 ){
magma_dmfree( L, queue );
magma_dmfree( U, queue );
}
magma_dmfree( &A_copy, queue );
magma_dmfree( &B, queue );
magma_dmfree( &hA, queue );
magma_dmfree( &CSRCOOA, queue );
return info;
}
extern "C" magma_int_t
c_transpose_csr(
magma_int_t n_rows,
magma_int_t n_cols,
magma_int_t nnz,
magmaFloatComplex *values,
magma_index_t *rowptr,
magma_index_t *colind,
magma_int_t *new_n_rows,
magma_int_t *new_n_cols,
magma_int_t *new_nnz,
magmaFloatComplex **new_values,
magma_index_t **new_rowptr,
magma_index_t **new_colind,
magma_queue_t queue )
{
magma_int_t info = 0;
// easier to keep names straight if we convert CSR to CSC,
// which is the same as tranposing CSR.
magmaFloatComplex *csc_values=NULL;
magma_index_t *csc_colptr=NULL, *csc_rowind=NULL;
// i, j are actual row & col indices (0 <= i < nrows, 0 <= j < ncols).
// k is index into col and values (0 <= k < nnz).
magma_int_t i, j, k, total, tmp;
CHECK( magma_cmalloc_cpu( &csc_values, nnz ) );
CHECK( magma_index_malloc_cpu( &csc_colptr, n_cols + 1 ) );
CHECK( magma_index_malloc_cpu( &csc_rowind, nnz ) );
// example matrix
// [ x x 0 x ]
// [ x 0 x x ]
// [ x x 0 0 ]
// rowptr = [ 0 3 6, 8 ]
// colind = [ 0 1 3; 0 2 3; 0 1 ]
// sum up nnz in each original column
// colptr = [ 3 2 1 2, X ]
for( j=0; j < n_cols; ++j ) {
csc_colptr[ j ] = 0;
}
for( k=0; k < nnz; ++k ) {
csc_colptr[ colind[k] ]++;
}
// running sum to convert to new colptr
// colptr = [ 0 3 5 6, 8 ]
total = 0;
for( j=0; j < n_cols; ++j ) {
tmp = csc_colptr[ j ];
csc_colptr[ j ] = total;
total += tmp;
}
csc_colptr[ n_cols ] = total;
assert( total == nnz );
// copy row indices and values
// this increments colptr until it effectively shifts left one
// colptr = [ 3 5 6 8, 8 ]
// rowind = [ 0 1 2; 0 2; 1; 0 1 ]
for( i=0; i < n_rows; ++i ) {
for( k=rowptr[i]; k < rowptr[i+1]; ++k ) {
j = colind[k];
csc_rowind[ csc_colptr[ j ] ] = i;
csc_values[ csc_colptr[ j ] ] = values[k];
csc_colptr[ j ]++;
}
}
assert( csc_colptr[ n_cols-1 ] == nnz );
// shift colptr right one
// colptr = [ 0 3 5 6, 8 ]
for( j=n_cols-1; j > 0; --j ) {
csc_colptr[j] = csc_colptr[j-1];
}
csc_colptr[0] = 0;
// save into output variables
*new_n_rows = n_cols;
*new_n_cols = n_rows;
*new_nnz = nnz;
*new_values = csc_values;
*new_rowptr = csc_colptr;
*new_colind = csc_rowind;
cleanup:
magma_free_cpu( csc_values );
magma_free_cpu( csc_colptr );
magma_free_cpu( csc_rowind );
return info;
}
extern "C" magma_int_t
magma_dmlumerge(
magma_d_sparse_matrix L,
magma_d_sparse_matrix U,
magma_d_sparse_matrix *A,
magma_queue_t queue ){
if( L.storage_type == Magma_CSR && U.storage_type == Magma_CSR ){
if( L.memory_location == Magma_CPU && U.memory_location == Magma_CPU ){
magma_d_mtransfer( L, A, Magma_CPU, Magma_CPU, queue );
magma_free_cpu( A->col );
magma_free_cpu( A->val );
// make sure it is strictly lower triangular
magma_int_t z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
z++;
}
}
A->nnz = z;
// fill A with the new structure;
magma_int_t stat_cpu = 0;
stat_cpu += magma_index_malloc_cpu( &A->col, A->nnz );
stat_cpu += magma_dmalloc_cpu( &A->val, A->nnz );
if( stat_cpu != 0 ){
magma_d_mfree( A, queue );
printf("error: memory allocation.\n");
return MAGMA_ERR_HOST_ALLOC;
}
z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
A->row[i] = z;
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
A->col[z] = L.col[j];
A->val[z] = L.val[j];
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
A->col[z] = U.col[j];
A->val[z] = U.val[j];
z++;
}
}
A->row[A->num_rows] = z;
A->nnz = z;
return MAGMA_SUCCESS;
}
else{
printf("error: matrix not on CPU.\n");
return MAGMA_SUCCESS;
}
}
else{
printf("error: matrix not on CPU.\n");
return MAGMA_SUCCESS;
}
}
magma_int_t
magma_cm_27stencil(
magma_int_t n,
magma_c_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i,j,k;
magma_c_matrix hA={Magma_CSR};
// generate matrix of desired structure and size (3d 27-point stencil)
magma_int_t nn = n*n*n;
magma_int_t offdiags = 13;
magma_index_t *diag_offset=NULL;
magmaFloatComplex *diag_vals=NULL;
CHECK( magma_cmalloc_cpu( &diag_vals, offdiags+1 ));
CHECK( magma_index_malloc_cpu( &diag_offset, offdiags+1 ));
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n-1;
diag_offset[3] = n;
diag_offset[4] = n+1;
diag_offset[5] = n*n-n-1;
diag_offset[6] = n*n-n;
diag_offset[7] = n*n-n+1;
diag_offset[8] = n*n-1;
diag_offset[9] = n*n;
diag_offset[10] = n*n+1;
diag_offset[11] = n*n+n-1;
diag_offset[12] = n*n+n;
diag_offset[13] = n*n+n+1;
diag_vals[0] = MAGMA_C_MAKE( 26.0, 0.0 );
diag_vals[1] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[3] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[4] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[5] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[6] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[7] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[8] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[9] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[10] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[11] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[12] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[13] = MAGMA_C_MAKE( -1.0, 0.0 );
CHECK( magma_cmgenerator( nn, offdiags, diag_offset, diag_vals, &hA, queue ));
// now set some entries to zero (boundary...)
for( i=0; i < n*n; i++ ) {
for( j=0; j < n; j++ ) {
magma_index_t row = i*n+j;
for( k=hA.row[row]; k<hA.row[row+1]; k++) {
if ((hA.col[k] == row-1 ||
hA.col[k] == row-n-1 ||
hA.col[k] == row+n-1 ||
hA.col[k] == row-n*n+n-1 ||
hA.col[k] == row+n*n-n-1 ||
hA.col[k] == row-n*n-1 ||
hA.col[k] == row+n*n-1 ||
hA.col[k] == row-n*n-n-1 ||
hA.col[k] == row+n*n+n-1 ) && (row+1)%n == 1 )
hA.val[k] = MAGMA_C_MAKE( 0.0, 0.0 );
if ((hA.col[k] == row+1 ||
hA.col[k] == row-n+1 ||
hA.col[k] == row+n+1 ||
hA.col[k] == row-n*n+n+1 ||
hA.col[k] == row+n*n-n+1 ||
hA.col[k] == row-n*n+1 ||
hA.col[k] == row+n*n+1 ||
hA.col[k] == row-n*n-n+1 ||
hA.col[k] == row+n*n+n+1 ) && (row)%n ==n-1 )
hA.val[k] = MAGMA_C_MAKE( 0.0, 0.0 );
}
}
}
hA.true_nnz = hA.nnz;
CHECK( magma_cmconvert( hA, A, Magma_CSR, Magma_CSR, queue ));
cleanup:
magma_free_cpu( diag_vals );
magma_free_cpu( diag_offset );
magma_cmfree( &hA, queue );
return info;
}
extern "C" magma_int_t
magma_smlumerge(
magma_s_matrix L,
magma_s_matrix U,
magma_s_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
if( L.storage_type == Magma_CSR && U.storage_type == Magma_CSR ){
if( L.memory_location == Magma_CPU && U.memory_location == Magma_CPU ){
CHECK( magma_smtransfer( L, A, Magma_CPU, Magma_CPU, queue ));
magma_free_cpu( A->col );
magma_free_cpu( A->val );
// make sure it is strictly lower triangular
magma_int_t z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
z++;
}
}
A->nnz = z;
// fill A with the new structure;
CHECK( magma_index_malloc_cpu( &A->col, A->nnz ));
CHECK( magma_smalloc_cpu( &A->val, A->nnz ));
z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
A->row[i] = z;
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
A->col[z] = L.col[j];
A->val[z] = L.val[j];
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
A->col[z] = U.col[j];
A->val[z] = U.val[j];
z++;
}
}
A->row[A->num_rows] = z;
A->nnz = z;
}
else{
printf("error: matrix not on CPU.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
}
else{
printf("error: matrix in wrong format.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
if( info != 0 ){
magma_smfree( A, queue );
}
return info;
}
magma_int_t
magma_cm_5stencil(
magma_int_t n,
magma_c_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i,j,k;
magma_c_matrix hA={Magma_CSR};
// generate matrix of desired structure and size (2d 5-point stencil)
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset=NULL;
magmaFloatComplex *diag_vals=NULL;
CHECK( magma_cmalloc_cpu( &diag_vals, offdiags+1 ));
CHECK( magma_index_malloc_cpu( &diag_offset, offdiags+1 ));
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
#define COMPLEX
#ifdef COMPLEX
// complex case
diag_vals[0] = MAGMA_C_MAKE( 4.0, 4.0 );
diag_vals[1] = MAGMA_C_MAKE( -1.0, -1.0 );
diag_vals[2] = MAGMA_C_MAKE( -1.0, -1.0 );
#else
// real case
diag_vals[0] = MAGMA_C_MAKE( 4.0, 0.0 );
diag_vals[1] = MAGMA_C_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_C_MAKE( -1.0, 0.0 );
#endif
CHECK( magma_cmgenerator( nn, offdiags, diag_offset, diag_vals, &hA, queue ));
// now set some entries to zero (boundary...)
for( i=0; i<n; i++ ) {
for( j=0; j<n; j++ ) {
magma_index_t row = i*n+j;
for( k=hA.row[row]; k<hA.row[row+1]; k++) {
if ((hA.col[k] == row-1 ) && (row+1)%n == 1 )
hA.val[k] = MAGMA_C_MAKE( 0.0, 0.0 );
if ((hA.col[k] == row+1 ) && (row)%n ==n-1 )
hA.val[k] = MAGMA_C_MAKE( 0.0, 0.0 );
}
}
}
CHECK( magma_cmconvert( hA, A, Magma_CSR, Magma_CSR, queue ));
magma_cmcsrcompressor( A, queue );
A->true_nnz = A->nnz;
cleanup:
magma_free_cpu( diag_vals );
magma_free_cpu( diag_offset );
magma_cmfree( &hA, queue );
return info;
}
extern "C" magma_int_t
magma_ccsrsplit(
magma_int_t bsize,
magma_c_matrix A,
magma_c_matrix *D,
magma_c_matrix *R,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i, k, j, nnz_diag, nnz_offd;
D->val = NULL;
D->col = NULL;
D->row = NULL;
D->rowidx = NULL;
D->blockinfo = NULL;
D->diag = NULL;
D->dval = NULL;
D->dcol = NULL;
D->drow = NULL;
D->drowidx = NULL;
D->ddiag = NULL;
R->val = NULL;
R->col = NULL;
R->row = NULL;
R->rowidx = NULL;
R->blockinfo = NULL;
R->diag = NULL;
R->dval = NULL;
R->dcol = NULL;
R->drow = NULL;
R->drowidx = NULL;
R->ddiag = NULL;
if ( A.memory_location == Magma_CPU &&
( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO ) ) {
nnz_diag = nnz_offd = 0;
// Count the new number of nonzeroes in the two matrices
for( i=0; i<A.num_rows; i+=bsize ){
for( k=i; k<min(A.num_rows,i+bsize); k++ ){
int check = 0;
for( j=A.row[k]; j<A.row[k+1]; j++ ){
if ( A.col[j] < i )
nnz_offd++;
else if ( A.col[j] < i+bsize ){
if( A.col[j] == k ){
check = 1;
}
nnz_diag++;
}
else
nnz_offd++;
}
if( check == 0 ){
printf("error: matrix contains zero on diagonal at (%d,%d).\n", i, i);
info = -1;
goto cleanup;
}
}
}
// Allocate memory for the new matrices
D->storage_type = Magma_CSRD;
D->memory_location = A.memory_location;
D->num_rows = A.num_rows;
D->num_cols = A.num_cols;
D->nnz = nnz_diag;
R->storage_type = Magma_CSR;
R->memory_location = A.memory_location;
R->num_rows = A.num_rows;
R->num_cols = A.num_cols;
R->nnz = nnz_offd;
CHECK( magma_cmalloc_cpu( &D->val, nnz_diag ));
CHECK( magma_index_malloc_cpu( &D->row, A.num_rows+1 ));
CHECK( magma_index_malloc_cpu( &D->col, nnz_diag ));
CHECK( magma_cmalloc_cpu( &R->val, nnz_offd ));
CHECK( magma_index_malloc_cpu( &R->row, A.num_rows+1 ));
CHECK( magma_index_malloc_cpu( &R->col, nnz_offd ));
// Fill up the new sparse matrices
D->row[0] = 0;
R->row[0] = 0;
nnz_offd = nnz_diag = 0;
for( i=0; i<A.num_rows; i+=bsize) {
for( k=i; k<min(A.num_rows,i+bsize); k++ ) {
D->row[k+1] = D->row[k];
R->row[k+1] = R->row[k];
for( j=A.row[k]; j<A.row[k+1]; j++ ) {
if ( A.col[j] < i ) {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
else if ( A.col[j] < i+bsize ) {
// larger than diagonal remain as before
if ( A.col[j]>k ) {
D->val[nnz_diag] = A.val[ j ];
D->col[nnz_diag] = A.col[ j ];
D->row[k+1]++;
}
// diagonal is written first
else if ( A.col[j]==k ) {
D->val[D->row[k]] = A.val[ j ];
D->col[D->row[k]] = A.col[ j ];
D->row[k+1]++;
}
// smaller than diagonal are shifted one to the right
// to have room for the diagonal
else {
D->val[nnz_diag+1] = A.val[ j ];
D->col[nnz_diag+1] = A.col[ j ];
D->row[k+1]++;
}
nnz_diag++;
}
else {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
}
}
}
}
else {
magma_c_matrix Ah={Magma_CSR}, ACSR={Magma_CSR}, DCSR={Magma_CSR}, RCSR={Magma_CSR}, Dh={Magma_CSR}, Rh={Magma_CSR};
CHECK( magma_cmtransfer( A, &Ah, A.memory_location, Magma_CPU, queue ));
CHECK( magma_cmconvert( Ah, &ACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_ccsrsplit( bsize, ACSR, &DCSR, &RCSR, queue ));
CHECK( magma_cmconvert( DCSR, &Dh, Magma_CSR, A.storage_type, queue ));
CHECK( magma_cmconvert( RCSR, &Rh, Magma_CSR, A.storage_type, queue ));
CHECK( magma_cmtransfer( Dh, D, Magma_CPU, A.memory_location, queue ));
CHECK( magma_cmtransfer( Rh, R, Magma_CPU, A.memory_location, queue ));
magma_cmfree( &Ah, queue );
magma_cmfree( &ACSR, queue );
magma_cmfree( &Dh, queue );
magma_cmfree( &DCSR, queue );
magma_cmfree( &Rh, queue );
magma_cmfree( &RCSR, queue );
}
cleanup:
if( info != 0 ){
magma_cmfree( D, queue );
magma_cmfree( R, queue );
}
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_d_matrix Z={Magma_CSR}, A={Magma_CSR}, AT={Magma_CSR},
A2={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_index_t *comm_i=NULL;
double *comm_v=NULL;
magma_int_t start, end;
int i=1;
CHECK( magma_dparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_dm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
CHECK( magma_d_csr_mtx( &Z, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(Z.num_rows), int(Z.num_cols), int(Z.nnz) );
// slice matrix
CHECK( magma_index_malloc_cpu( &comm_i, Z.num_rows ) );
CHECK( magma_dmalloc_cpu( &comm_v, Z.num_rows ) );
CHECK( magma_dmslice( 1, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 1, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 8, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
// scale matrix
CHECK( magma_dmscale( &Z, zopts.scaling, queue ));
// remove nonzeros in matrix
CHECK( magma_dmcsrcompressor( &Z, queue ));
// convert to be non-symmetric
CHECK( magma_dmconvert( Z, &A, Magma_CSR, Magma_CSRL, queue ));
// transpose
CHECK( magma_dmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_dmconvert( AT, &B, Magma_CSR, zopts.output_format, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&B, queue );
CHECK( magma_dmcsrcompressor_gpu( &B_d, queue ));
CHECK( magma_dmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_dmfree(&B_d, queue );
CHECK( magma_dmconvert( B, &AT, zopts.output_format,Magma_CSR, queue ));
magma_dmfree(&B, queue );
// transpose back
CHECK( magma_dmtranspose( AT, &A2, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmdiff( A, A2, &res, queue));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester: ok\n");
else
printf("%% tester: failed\n");
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
comm_i=NULL;
comm_v=NULL;
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
i++;
}
cleanup:
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
magma_dmfree(&AT, queue );
magma_dmfree(&A, queue );
magma_dmfree(&B, queue );
magma_dmfree(&B_d, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
